blob: 388dcf9aa283c83ee78dcf22503de253e812a27d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
Hugh Dickins9a840892009-09-21 17:02:01 -070048#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/rmap.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040050#include <linux/export.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070053#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080054#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070055#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080056#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090059#include <linux/gfp.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000060#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080061#include <linux/string.h>
Dan Williams0abdd7a2014-01-21 15:48:12 -080062#include <linux/dma-debug.h>
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -070063#include <linux/debugfs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070064
Alexey Dobriyan6952b612009-09-18 23:55:55 +040065#include <asm/io.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
Jan Beulich42b77722008-07-23 21:27:10 -070072#include "internal.h"
73
Peter Zijlstra90572892013-10-07 11:29:20 +010074#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080076#endif
77
Andy Whitcroftd41dee32005-06-23 00:07:54 -070078#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070079/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
Linus Torvalds1da177e2005-04-16 15:20:36 -070087/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070095
Linus Torvalds1da177e2005-04-16 15:20:36 -070096EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -070097
Ingo Molnar32a93232008-02-06 22:39:44 +010098/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800114 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100115}
116__setup("norandmaps", disable_randmaps);
117
Hugh Dickins62eede62009-09-21 17:03:34 -0700118unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700119unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700120
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200121EXPORT_SYMBOL(zero_pfn);
122
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700123/*
124 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
125 */
126static int __init init_zero_pfn(void)
127{
128 zero_pfn = page_to_pfn(ZERO_PAGE(0));
129 return 0;
130}
131core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100132
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800133
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800134#if defined(SPLIT_RSS_COUNTING)
135
David Rientjesea48cf72012-03-21 16:34:13 -0700136void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800137{
138 int i;
139
140 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700141 if (current->rss_stat.count[i]) {
142 add_mm_counter(mm, i, current->rss_stat.count[i]);
143 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800144 }
145 }
David Rientjes05af2e12012-03-21 16:34:13 -0700146 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800147}
148
149static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
150{
151 struct task_struct *task = current;
152
153 if (likely(task->mm == mm))
154 task->rss_stat.count[member] += val;
155 else
156 add_mm_counter(mm, member, val);
157}
158#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
159#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
160
161/* sync counter once per 64 page faults */
162#define TASK_RSS_EVENTS_THRESH (64)
163static void check_sync_rss_stat(struct task_struct *task)
164{
165 if (unlikely(task != current))
166 return;
167 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700168 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800169}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700170#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800171
172#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
173#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
174
175static void check_sync_rss_stat(struct task_struct *task)
176{
177}
178
Peter Zijlstra9547d012011-05-24 17:12:14 -0700179#endif /* SPLIT_RSS_COUNTING */
180
181#ifdef HAVE_GENERIC_MMU_GATHER
182
183static int tlb_next_batch(struct mmu_gather *tlb)
184{
185 struct mmu_gather_batch *batch;
186
187 batch = tlb->active;
188 if (batch->next) {
189 tlb->active = batch->next;
190 return 1;
191 }
192
Michal Hocko53a59fc2013-01-04 15:35:12 -0800193 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
194 return 0;
195
Peter Zijlstra9547d012011-05-24 17:12:14 -0700196 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
197 if (!batch)
198 return 0;
199
Michal Hocko53a59fc2013-01-04 15:35:12 -0800200 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700201 batch->next = NULL;
202 batch->nr = 0;
203 batch->max = MAX_GATHER_BATCH;
204
205 tlb->active->next = batch;
206 tlb->active = batch;
207
208 return 1;
209}
210
211/* tlb_gather_mmu
212 * Called to initialize an (on-stack) mmu_gather structure for page-table
213 * tear-down from @mm. The @fullmm argument is used when @mm is without
214 * users and we're going to destroy the full address space (exit/execve).
215 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700216void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700217{
218 tlb->mm = mm;
219
Linus Torvalds2b047252013-08-15 11:42:25 -0700220 /* Is it from 0 to ~0? */
221 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700222 tlb->need_flush_all = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700223 tlb->local.next = NULL;
224 tlb->local.nr = 0;
225 tlb->local.max = ARRAY_SIZE(tlb->__pages);
226 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800227 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700228
229#ifdef CONFIG_HAVE_RCU_TABLE_FREE
230 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800231#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000232
233 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700234}
235
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700236static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700237{
Will Deacon721c21c2015-01-12 19:10:55 +0000238 if (!tlb->end)
239 return;
240
Peter Zijlstra9547d012011-05-24 17:12:14 -0700241 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100242 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700243#ifdef CONFIG_HAVE_RCU_TABLE_FREE
244 tlb_table_flush(tlb);
245#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000246 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700247}
248
249static void tlb_flush_mmu_free(struct mmu_gather *tlb)
250{
251 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700252
Will Deacon721c21c2015-01-12 19:10:55 +0000253 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700254 free_pages_and_swap_cache(batch->pages, batch->nr);
255 batch->nr = 0;
256 }
257 tlb->active = &tlb->local;
258}
259
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700260void tlb_flush_mmu(struct mmu_gather *tlb)
261{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700262 tlb_flush_mmu_tlbonly(tlb);
263 tlb_flush_mmu_free(tlb);
264}
265
Peter Zijlstra9547d012011-05-24 17:12:14 -0700266/* tlb_finish_mmu
267 * Called at the end of the shootdown operation to free up any resources
268 * that were required.
269 */
270void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
271{
272 struct mmu_gather_batch *batch, *next;
273
274 tlb_flush_mmu(tlb);
275
276 /* keep the page table cache within bounds */
277 check_pgt_cache();
278
279 for (batch = tlb->local.next; batch; batch = next) {
280 next = batch->next;
281 free_pages((unsigned long)batch, 0);
282 }
283 tlb->local.next = NULL;
284}
285
286/* __tlb_remove_page
287 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
288 * handling the additional races in SMP caused by other CPUs caching valid
289 * mappings in their TLBs. Returns the number of free page slots left.
290 * When out of page slots we must call tlb_flush_mmu().
291 */
292int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
293{
294 struct mmu_gather_batch *batch;
295
Will Deaconfb7332a2014-10-29 10:03:09 +0000296 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297
Peter Zijlstra9547d012011-05-24 17:12:14 -0700298 batch = tlb->active;
299 batch->pages[batch->nr++] = page;
300 if (batch->nr == batch->max) {
301 if (!tlb_next_batch(tlb))
302 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700303 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700304 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800305 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700306
307 return batch->max - batch->nr;
308}
309
310#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800311
Peter Zijlstra26723912011-05-24 17:12:00 -0700312#ifdef CONFIG_HAVE_RCU_TABLE_FREE
313
314/*
315 * See the comment near struct mmu_table_batch.
316 */
317
318static void tlb_remove_table_smp_sync(void *arg)
319{
320 /* Simply deliver the interrupt */
321}
322
323static void tlb_remove_table_one(void *table)
324{
325 /*
326 * This isn't an RCU grace period and hence the page-tables cannot be
327 * assumed to be actually RCU-freed.
328 *
329 * It is however sufficient for software page-table walkers that rely on
330 * IRQ disabling. See the comment near struct mmu_table_batch.
331 */
332 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
333 __tlb_remove_table(table);
334}
335
336static void tlb_remove_table_rcu(struct rcu_head *head)
337{
338 struct mmu_table_batch *batch;
339 int i;
340
341 batch = container_of(head, struct mmu_table_batch, rcu);
342
343 for (i = 0; i < batch->nr; i++)
344 __tlb_remove_table(batch->tables[i]);
345
346 free_page((unsigned long)batch);
347}
348
349void tlb_table_flush(struct mmu_gather *tlb)
350{
351 struct mmu_table_batch **batch = &tlb->batch;
352
353 if (*batch) {
354 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
355 *batch = NULL;
356 }
357}
358
359void tlb_remove_table(struct mmu_gather *tlb, void *table)
360{
361 struct mmu_table_batch **batch = &tlb->batch;
362
Peter Zijlstra26723912011-05-24 17:12:00 -0700363 /*
364 * When there's less then two users of this mm there cannot be a
365 * concurrent page-table walk.
366 */
367 if (atomic_read(&tlb->mm->mm_users) < 2) {
368 __tlb_remove_table(table);
369 return;
370 }
371
372 if (*batch == NULL) {
373 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
374 if (*batch == NULL) {
375 tlb_remove_table_one(table);
376 return;
377 }
378 (*batch)->nr = 0;
379 }
380 (*batch)->tables[(*batch)->nr++] = table;
381 if ((*batch)->nr == MAX_TABLE_BATCH)
382 tlb_table_flush(tlb);
383}
384
Peter Zijlstra9547d012011-05-24 17:12:14 -0700385#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700386
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 * Note: this doesn't free the actual pages themselves. That
389 * has been handled earlier when unmapping all the memory regions.
390 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000391static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
392 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700393{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800394 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700395 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000396 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800397 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398}
399
Hugh Dickinse0da3822005-04-19 13:29:15 -0700400static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
401 unsigned long addr, unsigned long end,
402 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403{
404 pmd_t *pmd;
405 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700406 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410 do {
411 next = pmd_addr_end(addr, end);
412 if (pmd_none_or_clear_bad(pmd))
413 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000414 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 } while (pmd++, addr = next, addr != end);
416
Hugh Dickinse0da3822005-04-19 13:29:15 -0700417 start &= PUD_MASK;
418 if (start < floor)
419 return;
420 if (ceiling) {
421 ceiling &= PUD_MASK;
422 if (!ceiling)
423 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700425 if (end - 1 > ceiling - 1)
426 return;
427
428 pmd = pmd_offset(pud, start);
429 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000430 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800431 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432}
433
Hugh Dickinse0da3822005-04-19 13:29:15 -0700434static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
435 unsigned long addr, unsigned long end,
436 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437{
438 pud_t *pud;
439 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700440 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700441
Hugh Dickinse0da3822005-04-19 13:29:15 -0700442 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 do {
445 next = pud_addr_end(addr, end);
446 if (pud_none_or_clear_bad(pud))
447 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700448 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 } while (pud++, addr = next, addr != end);
450
Hugh Dickinse0da3822005-04-19 13:29:15 -0700451 start &= PGDIR_MASK;
452 if (start < floor)
453 return;
454 if (ceiling) {
455 ceiling &= PGDIR_MASK;
456 if (!ceiling)
457 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700459 if (end - 1 > ceiling - 1)
460 return;
461
462 pud = pud_offset(pgd, start);
463 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000464 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465}
466
467/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700468 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 */
Jan Beulich42b77722008-07-23 21:27:10 -0700470void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 unsigned long addr, unsigned long end,
472 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473{
474 pgd_t *pgd;
475 unsigned long next;
476
Hugh Dickinse0da3822005-04-19 13:29:15 -0700477 /*
478 * The next few lines have given us lots of grief...
479 *
480 * Why are we testing PMD* at this top level? Because often
481 * there will be no work to do at all, and we'd prefer not to
482 * go all the way down to the bottom just to discover that.
483 *
484 * Why all these "- 1"s? Because 0 represents both the bottom
485 * of the address space and the top of it (using -1 for the
486 * top wouldn't help much: the masks would do the wrong thing).
487 * The rule is that addr 0 and floor 0 refer to the bottom of
488 * the address space, but end 0 and ceiling 0 refer to the top
489 * Comparisons need to use "end - 1" and "ceiling - 1" (though
490 * that end 0 case should be mythical).
491 *
492 * Wherever addr is brought up or ceiling brought down, we must
493 * be careful to reject "the opposite 0" before it confuses the
494 * subsequent tests. But what about where end is brought down
495 * by PMD_SIZE below? no, end can't go down to 0 there.
496 *
497 * Whereas we round start (addr) and ceiling down, by different
498 * masks at different levels, in order to test whether a table
499 * now has no other vmas using it, so can be freed, we don't
500 * bother to round floor or end up - the tests don't need that.
501 */
502
503 addr &= PMD_MASK;
504 if (addr < floor) {
505 addr += PMD_SIZE;
506 if (!addr)
507 return;
508 }
509 if (ceiling) {
510 ceiling &= PMD_MASK;
511 if (!ceiling)
512 return;
513 }
514 if (end - 1 > ceiling - 1)
515 end -= PMD_SIZE;
516 if (addr > end - 1)
517 return;
518
Jan Beulich42b77722008-07-23 21:27:10 -0700519 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 do {
521 next = pgd_addr_end(addr, end);
522 if (pgd_none_or_clear_bad(pgd))
523 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700524 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700526}
527
Jan Beulich42b77722008-07-23 21:27:10 -0700528void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700529 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700530{
531 while (vma) {
532 struct vm_area_struct *next = vma->vm_next;
533 unsigned long addr = vma->vm_start;
534
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700535 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000536 * Hide vma from rmap and truncate_pagecache before freeing
537 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700538 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800539 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700540 unlink_file_vma(vma);
541
David Gibson9da61ae2006-03-22 00:08:57 -0800542 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700543 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700544 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700545 } else {
546 /*
547 * Optimization: gather nearby vmas into one call down
548 */
549 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800550 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700551 vma = next;
552 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800553 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700554 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700555 }
556 free_pgd_range(tlb, addr, vma->vm_end,
557 floor, next? next->vm_start: ceiling);
558 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700559 vma = next;
560 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800563int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
564 pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700565{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800566 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800567 pgtable_t new = pte_alloc_one(mm, address);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800568 int wait_split_huge_page;
Hugh Dickins1bb36302005-10-29 18:16:22 -0700569 if (!new)
570 return -ENOMEM;
571
Nick Piggin362a61a2008-05-14 06:37:36 +0200572 /*
573 * Ensure all pte setup (eg. pte page lock and page clearing) are
574 * visible before the pte is made visible to other CPUs by being
575 * put into page tables.
576 *
577 * The other side of the story is the pointer chasing in the page
578 * table walking code (when walking the page table without locking;
579 * ie. most of the time). Fortunately, these data accesses consist
580 * of a chain of data-dependent loads, meaning most CPUs (alpha
581 * being the notable exception) will already guarantee loads are
582 * seen in-order. See the alpha page table accessors for the
583 * smp_read_barrier_depends() barriers in page table walking code.
584 */
585 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
586
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800587 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800588 wait_split_huge_page = 0;
589 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;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800593 } else if (unlikely(pmd_trans_splitting(*pmd)))
594 wait_split_huge_page = 1;
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800595 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800596 if (new)
597 pte_free(mm, new);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800598 if (wait_split_huge_page)
599 wait_split_huge_page(vma->anon_vma, pmd);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700600 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601}
602
Hugh Dickins1bb36302005-10-29 18:16:22 -0700603int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700605 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
606 if (!new)
607 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608
Nick Piggin362a61a2008-05-14 06:37:36 +0200609 smp_wmb(); /* See comment in __pte_alloc */
610
Hugh Dickins1bb36302005-10-29 18:16:22 -0700611 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800612 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800614 new = NULL;
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800615 } else
616 VM_BUG_ON(pmd_trans_splitting(*pmd));
Hugh Dickins1bb36302005-10-29 18:16:22 -0700617 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800618 if (new)
619 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700620 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621}
622
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800623static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700624{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800625 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
626}
627
628static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
629{
630 int i;
631
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800632 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700633 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800634 for (i = 0; i < NR_MM_COUNTERS; i++)
635 if (rss[i])
636 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700637}
638
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800640 * This function is called to print an error when a bad pte
641 * is found. For example, we might have a PFN-mapped pte in
642 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700643 *
644 * The calling function must still handle the error.
645 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800646static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
647 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700648{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800649 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
650 pud_t *pud = pud_offset(pgd, addr);
651 pmd_t *pmd = pmd_offset(pud, addr);
652 struct address_space *mapping;
653 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800654 static unsigned long resume;
655 static unsigned long nr_shown;
656 static unsigned long nr_unshown;
657
658 /*
659 * Allow a burst of 60 reports, then keep quiet for that minute;
660 * or allow a steady drip of one report per second.
661 */
662 if (nr_shown == 60) {
663 if (time_before(jiffies, resume)) {
664 nr_unshown++;
665 return;
666 }
667 if (nr_unshown) {
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800668 printk(KERN_ALERT
669 "BUG: Bad page map: %lu messages suppressed\n",
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800670 nr_unshown);
671 nr_unshown = 0;
672 }
673 nr_shown = 0;
674 }
675 if (nr_shown++ == 0)
676 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800677
678 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
679 index = linear_page_index(vma, addr);
680
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800681 printk(KERN_ALERT
682 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
Hugh Dickins3dc14742009-01-06 14:40:08 -0800683 current->comm,
684 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800685 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800686 dump_page(page, "bad pte");
Hugh Dickins1e9e6362009-01-06 14:40:13 -0800687 printk(KERN_ALERT
Hugh Dickins3dc14742009-01-06 14:40:08 -0800688 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
689 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
690 /*
691 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
692 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700693 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
694 vma->vm_file,
695 vma->vm_ops ? vma->vm_ops->fault : NULL,
696 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
697 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700698 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030699 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700700}
701
702/*
Nick Piggin7e675132008-04-28 02:13:00 -0700703 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800704 *
Nick Piggin7e675132008-04-28 02:13:00 -0700705 * "Special" mappings do not wish to be associated with a "struct page" (either
706 * it doesn't exist, or it exists but they don't want to touch it). In this
707 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700708 *
Nick Piggin7e675132008-04-28 02:13:00 -0700709 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
710 * pte bit, in which case this function is trivial. Secondly, an architecture
711 * may not have a spare pte bit, which requires a more complicated scheme,
712 * described below.
713 *
714 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
715 * special mapping (even if there are underlying and valid "struct pages").
716 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800717 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700718 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
719 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700720 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
721 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800722 *
723 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
724 *
Nick Piggin7e675132008-04-28 02:13:00 -0700725 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700726 *
Nick Piggin7e675132008-04-28 02:13:00 -0700727 * This restricts such mappings to be a linear translation from virtual address
728 * to pfn. To get around this restriction, we allow arbitrary mappings so long
729 * as the vma is not a COW mapping; in that case, we know that all ptes are
730 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700731 *
732 *
Nick Piggin7e675132008-04-28 02:13:00 -0700733 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
734 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700735 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
736 * page" backing, however the difference is that _all_ pages with a struct
737 * page (that is, those where pfn_valid is true) are refcounted and considered
738 * normal pages by the VM. The disadvantage is that pages are refcounted
739 * (which can be slower and simply not an option for some PFNMAP users). The
740 * advantage is that we don't have to follow the strict linearity rule of
741 * PFNMAP mappings in order to support COWable mappings.
742 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800743 */
Nick Piggin7e675132008-04-28 02:13:00 -0700744#ifdef __HAVE_ARCH_PTE_SPECIAL
745# define HAVE_PTE_SPECIAL 1
746#else
747# define HAVE_PTE_SPECIAL 0
748#endif
749struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
750 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800751{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800752 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700753
754 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700755 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800756 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000757 if (vma->vm_ops && vma->vm_ops->find_special_page)
758 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700759 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
760 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700761 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800762 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700763 return NULL;
764 }
765
766 /* !HAVE_PTE_SPECIAL case follows: */
767
Jared Hulbertb379d792008-04-28 02:12:58 -0700768 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
769 if (vma->vm_flags & VM_MIXEDMAP) {
770 if (!pfn_valid(pfn))
771 return NULL;
772 goto out;
773 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700774 unsigned long off;
775 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700776 if (pfn == vma->vm_pgoff + off)
777 return NULL;
778 if (!is_cow_mapping(vma->vm_flags))
779 return NULL;
780 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800781 }
782
Hugh Dickinsb38af472014-08-29 15:18:44 -0700783 if (is_zero_pfn(pfn))
784 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800785check_pfn:
786 if (unlikely(pfn > highest_memmap_pfn)) {
787 print_bad_pte(vma, addr, pte, NULL);
788 return NULL;
789 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800790
791 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700792 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700793 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800794 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700795out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800796 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800797}
798
799/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 * copy one vm_area from one task to the other. Assumes the page tables
801 * already present in the new task to be cleared in the whole range
802 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 */
804
Hugh Dickins570a335b2009-12-14 17:58:46 -0800805static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700807 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700808 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Nick Pigginb5810032005-10-29 18:16:12 -0700810 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 pte_t pte = *src_pte;
812 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
814 /* pte contains position in swap or file, so copy. */
815 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800816 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700817
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800818 if (likely(!non_swap_entry(entry))) {
819 if (swap_duplicate(entry) < 0)
820 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800821
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800822 /* make sure dst_mm is on swapoff's mmlist. */
823 if (unlikely(list_empty(&dst_mm->mmlist))) {
824 spin_lock(&mmlist_lock);
825 if (list_empty(&dst_mm->mmlist))
826 list_add(&dst_mm->mmlist,
827 &src_mm->mmlist);
828 spin_unlock(&mmlist_lock);
829 }
830 rss[MM_SWAPENTS]++;
831 } else if (is_migration_entry(entry)) {
832 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800833
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800834 if (PageAnon(page))
835 rss[MM_ANONPAGES]++;
836 else
837 rss[MM_FILEPAGES]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800838
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800839 if (is_write_migration_entry(entry) &&
840 is_cow_mapping(vm_flags)) {
841 /*
842 * COW mappings require pages in both
843 * parent and child to be set to read.
844 */
845 make_migration_entry_read(&entry);
846 pte = swp_entry_to_pte(entry);
847 if (pte_swp_soft_dirty(*src_pte))
848 pte = pte_swp_mksoft_dirty(pte);
849 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700850 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700852 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700853 }
854
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 /*
856 * If it's a COW mapping, write protect it both
857 * in the parent and the child
858 */
Linus Torvalds67121172005-12-11 20:38:17 -0800859 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700861 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
863
864 /*
865 * If it's a shared mapping, mark it clean in
866 * the child
867 */
868 if (vm_flags & VM_SHARED)
869 pte = pte_mkclean(pte);
870 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800871
872 page = vm_normal_page(vma, addr, pte);
873 if (page) {
874 get_page(page);
Hugh Dickins21333b22009-09-21 17:01:59 -0700875 page_dup_rmap(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800876 if (PageAnon(page))
877 rss[MM_ANONPAGES]++;
878 else
879 rss[MM_FILEPAGES]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800880 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700881
882out_set_pte:
883 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800884 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885}
886
Jerome Marchand21bda262014-08-06 16:06:56 -0700887static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800888 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
889 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700891 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700893 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700894 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800895 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800896 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
898again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800899 init_rss_vec(rss);
900
Hugh Dickinsc74df322005-10-29 18:16:23 -0700901 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 if (!dst_pte)
903 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700904 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700905 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700906 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700907 orig_src_pte = src_pte;
908 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700909 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 do {
912 /*
913 * We are holding two locks at this point - either of them
914 * could generate latencies in another task on another CPU.
915 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700916 if (progress >= 32) {
917 progress = 0;
918 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100919 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700920 break;
921 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 if (pte_none(*src_pte)) {
923 progress++;
924 continue;
925 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800926 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
927 vma, addr, rss);
928 if (entry.val)
929 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 progress += 8;
931 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700933 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700934 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700935 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800936 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700937 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700938 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800939
940 if (entry.val) {
941 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
942 return -ENOMEM;
943 progress = 0;
944 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 if (addr != end)
946 goto again;
947 return 0;
948}
949
950static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
951 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
952 unsigned long addr, unsigned long end)
953{
954 pmd_t *src_pmd, *dst_pmd;
955 unsigned long next;
956
957 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
958 if (!dst_pmd)
959 return -ENOMEM;
960 src_pmd = pmd_offset(src_pud, addr);
961 do {
962 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800963 if (pmd_trans_huge(*src_pmd)) {
964 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800965 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800966 err = copy_huge_pmd(dst_mm, src_mm,
967 dst_pmd, src_pmd, addr, vma);
968 if (err == -ENOMEM)
969 return -ENOMEM;
970 if (!err)
971 continue;
972 /* fall through */
973 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 if (pmd_none_or_clear_bad(src_pmd))
975 continue;
976 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
977 vma, addr, next))
978 return -ENOMEM;
979 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
980 return 0;
981}
982
983static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
984 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
985 unsigned long addr, unsigned long end)
986{
987 pud_t *src_pud, *dst_pud;
988 unsigned long next;
989
990 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
991 if (!dst_pud)
992 return -ENOMEM;
993 src_pud = pud_offset(src_pgd, addr);
994 do {
995 next = pud_addr_end(addr, end);
996 if (pud_none_or_clear_bad(src_pud))
997 continue;
998 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
999 vma, addr, next))
1000 return -ENOMEM;
1001 } while (dst_pud++, src_pud++, addr = next, addr != end);
1002 return 0;
1003}
1004
1005int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1006 struct vm_area_struct *vma)
1007{
1008 pgd_t *src_pgd, *dst_pgd;
1009 unsigned long next;
1010 unsigned long addr = vma->vm_start;
1011 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001012 unsigned long mmun_start; /* For mmu_notifiers */
1013 unsigned long mmun_end; /* For mmu_notifiers */
1014 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001015 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001016
Nick Piggind9928952005-08-28 16:49:11 +10001017 /*
1018 * Don't copy ptes where a page fault will fill them correctly.
1019 * Fork becomes much lighter when there are big shared or private
1020 * readonly mappings. The tradeoff is that copy_page_range is more
1021 * efficient than faulting.
1022 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001023 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1024 !vma->anon_vma)
1025 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001026
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 if (is_vm_hugetlb_page(vma))
1028 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1029
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001030 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001031 /*
1032 * We do not free on error cases below as remove_vma
1033 * gets called on error from higher level routine
1034 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001035 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001036 if (ret)
1037 return ret;
1038 }
1039
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001040 /*
1041 * We need to invalidate the secondary MMU mappings only when
1042 * there could be a permission downgrade on the ptes of the
1043 * parent mm. And a permission downgrade will only happen if
1044 * is_cow_mapping() returns true.
1045 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001046 is_cow = is_cow_mapping(vma->vm_flags);
1047 mmun_start = addr;
1048 mmun_end = end;
1049 if (is_cow)
1050 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1051 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001052
1053 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054 dst_pgd = pgd_offset(dst_mm, addr);
1055 src_pgd = pgd_offset(src_mm, addr);
1056 do {
1057 next = pgd_addr_end(addr, end);
1058 if (pgd_none_or_clear_bad(src_pgd))
1059 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001060 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1061 vma, addr, next))) {
1062 ret = -ENOMEM;
1063 break;
1064 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001066
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001067 if (is_cow)
1068 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001069 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070}
1071
Robin Holt51c6f662005-11-13 16:06:42 -08001072static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001073 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001075 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001076{
Nick Pigginb5810032005-10-29 18:16:12 -07001077 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001078 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001079 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001080 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001081 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001082 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001083 swp_entry_t entry;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001084
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001085again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001086 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001087 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1088 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001089 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 do {
1091 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001092 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001094 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001095
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001097 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001098
Linus Torvalds6aab3412005-11-28 14:34:23 -08001099 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001100 if (unlikely(details) && page) {
1101 /*
1102 * unmap_shared_mapping_pages() wants to
1103 * invalidate cache without truncating:
1104 * unmap shared but keep private pages.
1105 */
1106 if (details->check_mapping &&
1107 details->check_mapping != page->mapping)
1108 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 }
Nick Pigginb5810032005-10-29 18:16:12 -07001110 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001111 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 tlb_remove_tlb_entry(tlb, pte, addr);
1113 if (unlikely(!page))
1114 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 if (PageAnon(page))
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001116 rss[MM_ANONPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001117 else {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001118 if (pte_dirty(ptent)) {
1119 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001120 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001121 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001122 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001123 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001124 mark_page_accessed(page);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001125 rss[MM_FILEPAGES]--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001126 }
Hugh Dickinsedc315f2009-01-06 14:40:11 -08001127 page_remove_rmap(page);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001128 if (unlikely(page_mapcount(page) < 0))
1129 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001130 if (unlikely(!__tlb_remove_page(tlb, page))) {
1131 force_flush = 1;
Will Deaconce9ec372014-10-28 13:16:28 -07001132 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001133 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001134 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 continue;
1136 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001137 /* If details->check_mapping, we leave swap entries. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 if (unlikely(details))
1139 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001140
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001141 entry = pte_to_swp_entry(ptent);
1142 if (!non_swap_entry(entry))
1143 rss[MM_SWAPENTS]--;
1144 else if (is_migration_entry(entry)) {
1145 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001146
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001147 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001148
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001149 if (PageAnon(page))
1150 rss[MM_ANONPAGES]--;
1151 else
1152 rss[MM_FILEPAGES]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001153 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001154 if (unlikely(!free_swap_and_cache(entry)))
1155 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001156 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001157 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001158
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001159 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001160 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001161
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001162 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001163 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001164 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001165 pte_unmap_unlock(start_pte, ptl);
1166
1167 /*
1168 * If we forced a TLB flush (either due to running out of
1169 * batch buffers or because we needed to flush dirty TLB
1170 * entries before releasing the ptl), free the batched
1171 * memory too. Restart if we didn't do everything.
1172 */
1173 if (force_flush) {
1174 force_flush = 0;
1175 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001176
1177 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001178 goto again;
1179 }
1180
Robin Holt51c6f662005-11-13 16:06:42 -08001181 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001182}
1183
Robin Holt51c6f662005-11-13 16:06:42 -08001184static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001185 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001186 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001187 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188{
1189 pmd_t *pmd;
1190 unsigned long next;
1191
1192 pmd = pmd_offset(pud, addr);
1193 do {
1194 next = pmd_addr_end(addr, end);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001195 if (pmd_trans_huge(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001196 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001197#ifdef CONFIG_DEBUG_VM
1198 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1199 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1200 __func__, addr, end,
1201 vma->vm_start,
1202 vma->vm_end);
1203 BUG();
1204 }
1205#endif
Kirill A. Shutemove1803772012-12-12 13:50:59 -08001206 split_huge_page_pmd(vma, addr, pmd);
Shaohua Lif21760b2012-01-12 17:19:16 -08001207 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001208 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001209 /* fall through */
1210 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001211 /*
1212 * Here there can be other concurrent MADV_DONTNEED or
1213 * trans huge page faults running, and if the pmd is
1214 * none or trans huge it can change under us. This is
1215 * because MADV_DONTNEED holds the mmap_sem in read
1216 * mode.
1217 */
1218 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1219 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001220 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001221next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001222 cond_resched();
1223 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001224
1225 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226}
1227
Robin Holt51c6f662005-11-13 16:06:42 -08001228static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001229 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001231 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232{
1233 pud_t *pud;
1234 unsigned long next;
1235
1236 pud = pud_offset(pgd, addr);
1237 do {
1238 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001239 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001241 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1242 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001243
1244 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245}
1246
Al Viro038c7aa2012-03-05 13:25:09 -05001247static void unmap_page_range(struct mmu_gather *tlb,
1248 struct vm_area_struct *vma,
1249 unsigned long addr, unsigned long end,
1250 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
1252 pgd_t *pgd;
1253 unsigned long next;
1254
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001255 if (details && !details->check_mapping)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 details = NULL;
1257
1258 BUG_ON(addr >= end);
1259 tlb_start_vma(tlb, vma);
1260 pgd = pgd_offset(vma->vm_mm, addr);
1261 do {
1262 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001263 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001265 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1266 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 tlb_end_vma(tlb, vma);
1268}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269
Al Virof5cc4ee2012-03-05 14:14:20 -05001270
1271static void unmap_single_vma(struct mmu_gather *tlb,
1272 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001273 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001274 struct zap_details *details)
1275{
1276 unsigned long start = max(vma->vm_start, start_addr);
1277 unsigned long end;
1278
1279 if (start >= vma->vm_end)
1280 return;
1281 end = min(vma->vm_end, end_addr);
1282 if (end <= vma->vm_start)
1283 return;
1284
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301285 if (vma->vm_file)
1286 uprobe_munmap(vma, start, end);
1287
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001288 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001289 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001290
1291 if (start != end) {
1292 if (unlikely(is_vm_hugetlb_page(vma))) {
1293 /*
1294 * It is undesirable to test vma->vm_file as it
1295 * should be non-null for valid hugetlb area.
1296 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001297 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001298 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001299 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001300 * before calling this function to clean up.
1301 * Since no pte has actually been setup, it is
1302 * safe to do nothing in this case.
1303 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001304 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001305 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001306 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001307 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001308 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001309 } else
1310 unmap_page_range(tlb, vma, start, end, details);
1311 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312}
1313
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314/**
1315 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001316 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 * @vma: the starting vma
1318 * @start_addr: virtual address at which to start unmapping
1319 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001321 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 * Only addresses between `start' and `end' will be unmapped.
1324 *
1325 * The VMA list must be sorted in ascending virtual address order.
1326 *
1327 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1328 * range after unmap_vmas() returns. So the only responsibility here is to
1329 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1330 * drops the lock and schedules.
1331 */
Al Viro6e8bb012012-03-05 13:41:15 -05001332void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001334 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001336 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001338 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001339 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001340 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001341 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342}
1343
1344/**
1345 * zap_page_range - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001347 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001349 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001350 *
1351 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001353void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 unsigned long size, struct zap_details *details)
1355{
1356 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001357 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001358 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001361 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001362 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001363 mmu_notifier_invalidate_range_start(mm, start, end);
1364 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001365 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001366 mmu_notifier_invalidate_range_end(mm, start, end);
1367 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368}
1369
Jack Steinerc627f9c2008-07-29 22:33:53 -07001370/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001371 * zap_page_range_single - remove user pages in a given range
1372 * @vma: vm_area_struct holding the applicable pages
1373 * @address: starting address of pages to zap
1374 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001375 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001376 *
1377 * The range must fit into one VMA.
1378 */
1379static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1380 unsigned long size, struct zap_details *details)
1381{
1382 struct mm_struct *mm = vma->vm_mm;
1383 struct mmu_gather tlb;
1384 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001385
1386 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001387 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001388 update_hiwater_rss(mm);
1389 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001390 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001391 mmu_notifier_invalidate_range_end(mm, address, end);
1392 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393}
1394
Jack Steinerc627f9c2008-07-29 22:33:53 -07001395/**
1396 * zap_vma_ptes - remove ptes mapping the vma
1397 * @vma: vm_area_struct holding ptes to be zapped
1398 * @address: starting address of pages to zap
1399 * @size: number of bytes to zap
1400 *
1401 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1402 *
1403 * The entire address range must be fully contained within the vma.
1404 *
1405 * Returns 0 if successful.
1406 */
1407int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1408 unsigned long size)
1409{
1410 if (address < vma->vm_start || address + size > vma->vm_end ||
1411 !(vma->vm_flags & VM_PFNMAP))
1412 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001413 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001414 return 0;
1415}
1416EXPORT_SYMBOL_GPL(zap_vma_ptes);
1417
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001418pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001419 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001420{
1421 pgd_t * pgd = pgd_offset(mm, addr);
1422 pud_t * pud = pud_alloc(mm, pgd, addr);
1423 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001424 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001425 if (pmd) {
1426 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001427 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001428 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001429 }
1430 return NULL;
1431}
1432
Linus Torvalds1da177e2005-04-16 15:20:36 -07001433/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001434 * This is the old fallback for page remapping.
1435 *
1436 * For historical reasons, it only allows reserved pages. Only
1437 * old drivers should use this, and they needed to mark their
1438 * pages reserved for the old functions anyway.
1439 */
Nick Piggin423bad62008-04-28 02:13:01 -07001440static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1441 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001442{
Nick Piggin423bad62008-04-28 02:13:01 -07001443 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001444 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001445 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001446 spinlock_t *ptl;
1447
Linus Torvalds238f58d2005-11-29 13:01:56 -08001448 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001449 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001450 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001451 retval = -ENOMEM;
1452 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001453 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001454 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001455 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001456 retval = -EBUSY;
1457 if (!pte_none(*pte))
1458 goto out_unlock;
1459
1460 /* Ok, finally just insert the thing.. */
1461 get_page(page);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08001462 inc_mm_counter_fast(mm, MM_FILEPAGES);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001463 page_add_file_rmap(page);
1464 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1465
1466 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001467 pte_unmap_unlock(pte, ptl);
1468 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001469out_unlock:
1470 pte_unmap_unlock(pte, ptl);
1471out:
1472 return retval;
1473}
1474
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001475/**
1476 * vm_insert_page - insert single page into user vma
1477 * @vma: user vma to map to
1478 * @addr: target user address of this page
1479 * @page: source kernel page
1480 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001481 * This allows drivers to insert individual pages they've allocated
1482 * into a user vma.
1483 *
1484 * The page has to be a nice clean _individual_ kernel allocation.
1485 * If you allocate a compound page, you need to have marked it as
1486 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001487 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001488 *
1489 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1490 * took an arbitrary page protection parameter. This doesn't allow
1491 * that. Your vma protection will have to be set up correctly, which
1492 * means that if you want a shared writable mapping, you'd better
1493 * ask for a shared writable mapping!
1494 *
1495 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001496 *
1497 * Usually this function is called from f_op->mmap() handler
1498 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1499 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1500 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001501 */
Nick Piggin423bad62008-04-28 02:13:01 -07001502int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1503 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001504{
1505 if (addr < vma->vm_start || addr >= vma->vm_end)
1506 return -EFAULT;
1507 if (!page_count(page))
1508 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001509 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1510 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1511 BUG_ON(vma->vm_flags & VM_PFNMAP);
1512 vma->vm_flags |= VM_MIXEDMAP;
1513 }
Nick Piggin423bad62008-04-28 02:13:01 -07001514 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001515}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001516EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001517
Nick Piggin423bad62008-04-28 02:13:01 -07001518static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1519 unsigned long pfn, pgprot_t prot)
1520{
1521 struct mm_struct *mm = vma->vm_mm;
1522 int retval;
1523 pte_t *pte, entry;
1524 spinlock_t *ptl;
1525
1526 retval = -ENOMEM;
1527 pte = get_locked_pte(mm, addr, &ptl);
1528 if (!pte)
1529 goto out;
1530 retval = -EBUSY;
1531 if (!pte_none(*pte))
1532 goto out_unlock;
1533
1534 /* Ok, finally just insert the thing.. */
1535 entry = pte_mkspecial(pfn_pte(pfn, prot));
1536 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001537 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001538
1539 retval = 0;
1540out_unlock:
1541 pte_unmap_unlock(pte, ptl);
1542out:
1543 return retval;
1544}
1545
Nick Piggine0dc0d82007-02-12 00:51:36 -08001546/**
1547 * vm_insert_pfn - insert single pfn into user vma
1548 * @vma: user vma to map to
1549 * @addr: target user address of this page
1550 * @pfn: source kernel pfn
1551 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001552 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001553 * they've allocated into a user vma. Same comments apply.
1554 *
1555 * This function should only be called from a vm_ops->fault handler, and
1556 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001557 *
1558 * vma cannot be a COW mapping.
1559 *
1560 * As this is called only for pages that do not currently exist, we
1561 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001562 */
1563int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001564 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001565{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001566 int ret;
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001567 pgprot_t pgprot = vma->vm_page_prot;
Nick Piggin7e675132008-04-28 02:13:00 -07001568 /*
1569 * Technically, architectures with pte_special can avoid all these
1570 * restrictions (same for remap_pfn_range). However we would like
1571 * consistency in testing and feature parity among all, so we should
1572 * try to keep these invariants in place for everybody.
1573 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001574 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1575 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1576 (VM_PFNMAP|VM_MIXEDMAP));
1577 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1578 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001579
Nick Piggin423bad62008-04-28 02:13:01 -07001580 if (addr < vma->vm_start || addr >= vma->vm_end)
1581 return -EFAULT;
Suresh Siddha5180da42012-10-08 16:28:29 -07001582 if (track_pfn_insert(vma, &pgprot, pfn))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001583 return -EINVAL;
1584
venkatesh.pallipadi@intel.come4b866e2009-01-09 16:13:11 -08001585 ret = insert_pfn(vma, addr, pfn, pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001586
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001587 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001588}
1589EXPORT_SYMBOL(vm_insert_pfn);
1590
Nick Piggin423bad62008-04-28 02:13:01 -07001591int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1592 unsigned long pfn)
1593{
1594 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1595
1596 if (addr < vma->vm_start || addr >= vma->vm_end)
1597 return -EFAULT;
1598
1599 /*
1600 * If we don't have pte special, then we have to use the pfn_valid()
1601 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1602 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001603 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1604 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001605 */
1606 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
1607 struct page *page;
1608
1609 page = pfn_to_page(pfn);
1610 return insert_page(vma, addr, page, vma->vm_page_prot);
1611 }
1612 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1613}
1614EXPORT_SYMBOL(vm_insert_mixed);
1615
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001616/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 * maps a range of physical memory into the requested pages. the old
1618 * mappings are removed. any references to nonexistent pages results
1619 * in null mappings (currently treated as "copy-on-access")
1620 */
1621static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1622 unsigned long addr, unsigned long end,
1623 unsigned long pfn, pgprot_t prot)
1624{
1625 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001626 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Hugh Dickinsc74df322005-10-29 18:16:23 -07001628 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 if (!pte)
1630 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001631 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 do {
1633 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001634 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 pfn++;
1636 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001637 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001638 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 return 0;
1640}
1641
1642static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1643 unsigned long addr, unsigned long end,
1644 unsigned long pfn, pgprot_t prot)
1645{
1646 pmd_t *pmd;
1647 unsigned long next;
1648
1649 pfn -= addr >> PAGE_SHIFT;
1650 pmd = pmd_alloc(mm, pud, addr);
1651 if (!pmd)
1652 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001653 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 do {
1655 next = pmd_addr_end(addr, end);
1656 if (remap_pte_range(mm, pmd, addr, next,
1657 pfn + (addr >> PAGE_SHIFT), prot))
1658 return -ENOMEM;
1659 } while (pmd++, addr = next, addr != end);
1660 return 0;
1661}
1662
1663static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1664 unsigned long addr, unsigned long end,
1665 unsigned long pfn, pgprot_t prot)
1666{
1667 pud_t *pud;
1668 unsigned long next;
1669
1670 pfn -= addr >> PAGE_SHIFT;
1671 pud = pud_alloc(mm, pgd, addr);
1672 if (!pud)
1673 return -ENOMEM;
1674 do {
1675 next = pud_addr_end(addr, end);
1676 if (remap_pmd_range(mm, pud, addr, next,
1677 pfn + (addr >> PAGE_SHIFT), prot))
1678 return -ENOMEM;
1679 } while (pud++, addr = next, addr != end);
1680 return 0;
1681}
1682
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001683/**
1684 * remap_pfn_range - remap kernel memory to userspace
1685 * @vma: user vma to map to
1686 * @addr: target user address to start at
1687 * @pfn: physical address of kernel memory
1688 * @size: size of map area
1689 * @prot: page protection flags for this mapping
1690 *
1691 * Note: this is only safe if the mm semaphore is held when called.
1692 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1694 unsigned long pfn, unsigned long size, pgprot_t prot)
1695{
1696 pgd_t *pgd;
1697 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001698 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 struct mm_struct *mm = vma->vm_mm;
1700 int err;
1701
1702 /*
1703 * Physically remapped pages are special. Tell the
1704 * rest of the world about it:
1705 * VM_IO tells people not to look at these pages
1706 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001707 * VM_PFNMAP tells the core MM that the base pages are just
1708 * raw PFN mappings, and do not have a "struct page" associated
1709 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001710 * VM_DONTEXPAND
1711 * Disable vma merging and expanding with mremap().
1712 * VM_DONTDUMP
1713 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001714 *
1715 * There's a horrible special case to handle copy-on-write
1716 * behaviour that some programs depend on. We mark the "original"
1717 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001718 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001720 if (is_cow_mapping(vma->vm_flags)) {
1721 if (addr != vma->vm_start || end != vma->vm_end)
1722 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001723 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001724 }
1725
1726 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1727 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001728 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001729
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001730 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731
1732 BUG_ON(addr >= end);
1733 pfn -= addr >> PAGE_SHIFT;
1734 pgd = pgd_offset(mm, addr);
1735 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 do {
1737 next = pgd_addr_end(addr, end);
1738 err = remap_pud_range(mm, pgd, addr, next,
1739 pfn + (addr >> PAGE_SHIFT), prot);
1740 if (err)
1741 break;
1742 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001743
1744 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001745 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001746
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 return err;
1748}
1749EXPORT_SYMBOL(remap_pfn_range);
1750
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001751/**
1752 * vm_iomap_memory - remap memory to userspace
1753 * @vma: user vma to map to
1754 * @start: start of area
1755 * @len: size of area
1756 *
1757 * This is a simplified io_remap_pfn_range() for common driver use. The
1758 * driver just needs to give us the physical memory range to be mapped,
1759 * we'll figure out the rest from the vma information.
1760 *
1761 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1762 * whatever write-combining details or similar.
1763 */
1764int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1765{
1766 unsigned long vm_len, pfn, pages;
1767
1768 /* Check that the physical memory area passed in looks valid */
1769 if (start + len < start)
1770 return -EINVAL;
1771 /*
1772 * You *really* shouldn't map things that aren't page-aligned,
1773 * but we've historically allowed it because IO memory might
1774 * just have smaller alignment.
1775 */
1776 len += start & ~PAGE_MASK;
1777 pfn = start >> PAGE_SHIFT;
1778 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1779 if (pfn + pages < pfn)
1780 return -EINVAL;
1781
1782 /* We start the mapping 'vm_pgoff' pages into the area */
1783 if (vma->vm_pgoff > pages)
1784 return -EINVAL;
1785 pfn += vma->vm_pgoff;
1786 pages -= vma->vm_pgoff;
1787
1788 /* Can we fit all of the mapping? */
1789 vm_len = vma->vm_end - vma->vm_start;
1790 if (vm_len >> PAGE_SHIFT > pages)
1791 return -EINVAL;
1792
1793 /* Ok, let it rip */
1794 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1795}
1796EXPORT_SYMBOL(vm_iomap_memory);
1797
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001798static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1799 unsigned long addr, unsigned long end,
1800 pte_fn_t fn, void *data)
1801{
1802 pte_t *pte;
1803 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001804 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001805 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001806
1807 pte = (mm == &init_mm) ?
1808 pte_alloc_kernel(pmd, addr) :
1809 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1810 if (!pte)
1811 return -ENOMEM;
1812
1813 BUG_ON(pmd_huge(*pmd));
1814
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001815 arch_enter_lazy_mmu_mode();
1816
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001817 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001818
1819 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001820 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001821 if (err)
1822 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001823 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001824
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001825 arch_leave_lazy_mmu_mode();
1826
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001827 if (mm != &init_mm)
1828 pte_unmap_unlock(pte-1, ptl);
1829 return err;
1830}
1831
1832static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1833 unsigned long addr, unsigned long end,
1834 pte_fn_t fn, void *data)
1835{
1836 pmd_t *pmd;
1837 unsigned long next;
1838 int err;
1839
Andi Kleenceb86872008-07-23 21:27:50 -07001840 BUG_ON(pud_huge(*pud));
1841
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001842 pmd = pmd_alloc(mm, pud, addr);
1843 if (!pmd)
1844 return -ENOMEM;
1845 do {
1846 next = pmd_addr_end(addr, end);
1847 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1848 if (err)
1849 break;
1850 } while (pmd++, addr = next, addr != end);
1851 return err;
1852}
1853
1854static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1855 unsigned long addr, unsigned long end,
1856 pte_fn_t fn, void *data)
1857{
1858 pud_t *pud;
1859 unsigned long next;
1860 int err;
1861
1862 pud = pud_alloc(mm, pgd, addr);
1863 if (!pud)
1864 return -ENOMEM;
1865 do {
1866 next = pud_addr_end(addr, end);
1867 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1868 if (err)
1869 break;
1870 } while (pud++, addr = next, addr != end);
1871 return err;
1872}
1873
1874/*
1875 * Scan a region of virtual memory, filling in page tables as necessary
1876 * and calling a provided function on each leaf page table.
1877 */
1878int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1879 unsigned long size, pte_fn_t fn, void *data)
1880{
1881 pgd_t *pgd;
1882 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001883 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001884 int err;
1885
1886 BUG_ON(addr >= end);
1887 pgd = pgd_offset(mm, addr);
1888 do {
1889 next = pgd_addr_end(addr, end);
1890 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1891 if (err)
1892 break;
1893 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001894
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001895 return err;
1896}
1897EXPORT_SYMBOL_GPL(apply_to_page_range);
1898
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001900 * handle_pte_fault chooses page fault handler according to an entry which was
1901 * read non-atomically. Before making any commitment, on those architectures
1902 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1903 * parts, do_swap_page must check under lock before unmapping the pte and
1904 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001905 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001906 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001907static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001908 pte_t *page_table, pte_t orig_pte)
1909{
1910 int same = 1;
1911#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1912 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001913 spinlock_t *ptl = pte_lockptr(mm, pmd);
1914 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001915 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001916 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001917 }
1918#endif
1919 pte_unmap(page_table);
1920 return same;
1921}
1922
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001923static 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 -08001924{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001925 debug_dma_assert_idle(src);
1926
Linus Torvalds6aab3412005-11-28 14:34:23 -08001927 /*
1928 * If the source page was a PFN mapping, we don't have
1929 * a "struct page" for it. We do a best-effort copy by
1930 * just copying from the original user address. If that
1931 * fails, we just zero-fill it. Live with it.
1932 */
1933 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001934 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001935 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1936
1937 /*
1938 * This really shouldn't fail, because the page is there
1939 * in the page tables. But it might just be unreadable,
1940 * in which case we just give up and fill the result with
1941 * zeroes.
1942 */
1943 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001944 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001945 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001946 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001947 } else
1948 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001949}
1950
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001952 * Notify the address space that the page is about to become writable so that
1953 * it can prohibit this or wait for the page to get into an appropriate state.
1954 *
1955 * We do this without the lock held, so that it can sleep if it needs to.
1956 */
1957static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1958 unsigned long address)
1959{
1960 struct vm_fault vmf;
1961 int ret;
1962
1963 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1964 vmf.pgoff = page->index;
1965 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
1966 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08001967 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001968
1969 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
1970 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
1971 return ret;
1972 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
1973 lock_page(page);
1974 if (!page->mapping) {
1975 unlock_page(page);
1976 return 0; /* retry */
1977 }
1978 ret |= VM_FAULT_LOCKED;
1979 } else
1980 VM_BUG_ON_PAGE(!PageLocked(page), page);
1981 return ret;
1982}
1983
1984/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07001985 * Handle write page faults for pages that can be reused in the current vma
1986 *
1987 * This can happen either due to the mapping being with the VM_SHARED flag,
1988 * or due to us being the last reference standing to the page. In either
1989 * case, all we need to do here is to mark the page as writable and update
1990 * any related book-keeping.
1991 */
1992static inline int wp_page_reuse(struct mm_struct *mm,
1993 struct vm_area_struct *vma, unsigned long address,
1994 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
1995 struct page *page, int page_mkwrite,
1996 int dirty_shared)
1997 __releases(ptl)
1998{
1999 pte_t entry;
2000 /*
2001 * Clear the pages cpupid information as the existing
2002 * information potentially belongs to a now completely
2003 * unrelated process.
2004 */
2005 if (page)
2006 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2007
2008 flush_cache_page(vma, address, pte_pfn(orig_pte));
2009 entry = pte_mkyoung(orig_pte);
2010 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2011 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2012 update_mmu_cache(vma, address, page_table);
2013 pte_unmap_unlock(page_table, ptl);
2014
2015 if (dirty_shared) {
2016 struct address_space *mapping;
2017 int dirtied;
2018
2019 if (!page_mkwrite)
2020 lock_page(page);
2021
2022 dirtied = set_page_dirty(page);
2023 VM_BUG_ON_PAGE(PageAnon(page), page);
2024 mapping = page->mapping;
2025 unlock_page(page);
2026 page_cache_release(page);
2027
2028 if ((dirtied || page_mkwrite) && mapping) {
2029 /*
2030 * Some device drivers do not set page.mapping
2031 * but still dirty their pages
2032 */
2033 balance_dirty_pages_ratelimited(mapping);
2034 }
2035
2036 if (!page_mkwrite)
2037 file_update_time(vma->vm_file);
2038 }
2039
2040 return VM_FAULT_WRITE;
2041}
2042
2043/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002044 * Handle the case of a page which we actually need to copy to a new page.
2045 *
2046 * Called with mmap_sem locked and the old page referenced, but
2047 * without the ptl held.
2048 *
2049 * High level logic flow:
2050 *
2051 * - Allocate a page, copy the content of the old page to the new one.
2052 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2053 * - Take the PTL. If the pte changed, bail out and release the allocated page
2054 * - If the pte is still the way we remember it, update the page table and all
2055 * relevant references. This includes dropping the reference the page-table
2056 * held to the old page, as well as updating the rmap.
2057 * - In any case, unlock the PTL and drop the reference we took to the old page.
2058 */
2059static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2060 unsigned long address, pte_t *page_table, pmd_t *pmd,
2061 pte_t orig_pte, struct page *old_page)
2062{
2063 struct page *new_page = NULL;
2064 spinlock_t *ptl = NULL;
2065 pte_t entry;
2066 int page_copied = 0;
2067 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2068 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2069 struct mem_cgroup *memcg;
2070
2071 if (unlikely(anon_vma_prepare(vma)))
2072 goto oom;
2073
2074 if (is_zero_pfn(pte_pfn(orig_pte))) {
2075 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2076 if (!new_page)
2077 goto oom;
2078 } else {
2079 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2080 if (!new_page)
2081 goto oom;
2082 cow_user_page(new_page, old_page, address, vma);
2083 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002084
2085 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg))
2086 goto oom_free_new;
2087
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002088 __SetPageUptodate(new_page);
2089
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002090 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2091
2092 /*
2093 * Re-check the pte - we dropped the lock
2094 */
2095 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2096 if (likely(pte_same(*page_table, orig_pte))) {
2097 if (old_page) {
2098 if (!PageAnon(old_page)) {
2099 dec_mm_counter_fast(mm, MM_FILEPAGES);
2100 inc_mm_counter_fast(mm, MM_ANONPAGES);
2101 }
2102 } else {
2103 inc_mm_counter_fast(mm, MM_ANONPAGES);
2104 }
2105 flush_cache_page(vma, address, pte_pfn(orig_pte));
2106 entry = mk_pte(new_page, vma->vm_page_prot);
2107 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2108 /*
2109 * Clear the pte entry and flush it first, before updating the
2110 * pte with the new entry. This will avoid a race condition
2111 * seen in the presence of one thread doing SMC and another
2112 * thread doing COW.
2113 */
2114 ptep_clear_flush_notify(vma, address, page_table);
2115 page_add_new_anon_rmap(new_page, vma, address);
2116 mem_cgroup_commit_charge(new_page, memcg, false);
2117 lru_cache_add_active_or_unevictable(new_page, vma);
2118 /*
2119 * We call the notify macro here because, when using secondary
2120 * mmu page tables (such as kvm shadow page tables), we want the
2121 * new page to be mapped directly into the secondary page table.
2122 */
2123 set_pte_at_notify(mm, address, page_table, entry);
2124 update_mmu_cache(vma, address, page_table);
2125 if (old_page) {
2126 /*
2127 * Only after switching the pte to the new page may
2128 * we remove the mapcount here. Otherwise another
2129 * process may come and find the rmap count decremented
2130 * before the pte is switched to the new page, and
2131 * "reuse" the old page writing into it while our pte
2132 * here still points into it and can be read by other
2133 * threads.
2134 *
2135 * The critical issue is to order this
2136 * page_remove_rmap with the ptp_clear_flush above.
2137 * Those stores are ordered by (if nothing else,)
2138 * the barrier present in the atomic_add_negative
2139 * in page_remove_rmap.
2140 *
2141 * Then the TLB flush in ptep_clear_flush ensures that
2142 * no process can access the old page before the
2143 * decremented mapcount is visible. And the old page
2144 * cannot be reused until after the decremented
2145 * mapcount is visible. So transitively, TLBs to
2146 * old page will be flushed before it can be reused.
2147 */
2148 page_remove_rmap(old_page);
2149 }
2150
2151 /* Free the old page.. */
2152 new_page = old_page;
2153 page_copied = 1;
2154 } else {
2155 mem_cgroup_cancel_charge(new_page, memcg);
2156 }
2157
2158 if (new_page)
2159 page_cache_release(new_page);
2160
2161 pte_unmap_unlock(page_table, ptl);
2162 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2163 if (old_page) {
2164 /*
2165 * Don't let another task, with possibly unlocked vma,
2166 * keep the mlocked page.
2167 */
2168 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2169 lock_page(old_page); /* LRU manipulation */
2170 munlock_vma_page(old_page);
2171 unlock_page(old_page);
2172 }
2173 page_cache_release(old_page);
2174 }
2175 return page_copied ? VM_FAULT_WRITE : 0;
2176oom_free_new:
2177 page_cache_release(new_page);
2178oom:
2179 if (old_page)
2180 page_cache_release(old_page);
2181 return VM_FAULT_OOM;
2182}
2183
Boaz Harroshdd906182015-04-15 16:15:11 -07002184/*
2185 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2186 * mapping
2187 */
2188static int wp_pfn_shared(struct mm_struct *mm,
2189 struct vm_area_struct *vma, unsigned long address,
2190 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2191 pmd_t *pmd)
2192{
2193 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2194 struct vm_fault vmf = {
2195 .page = NULL,
2196 .pgoff = linear_page_index(vma, address),
2197 .virtual_address = (void __user *)(address & PAGE_MASK),
2198 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2199 };
2200 int ret;
2201
2202 pte_unmap_unlock(page_table, ptl);
2203 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2204 if (ret & VM_FAULT_ERROR)
2205 return ret;
2206 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2207 /*
2208 * We might have raced with another page fault while we
2209 * released the pte_offset_map_lock.
2210 */
2211 if (!pte_same(*page_table, orig_pte)) {
2212 pte_unmap_unlock(page_table, ptl);
2213 return 0;
2214 }
2215 }
2216 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2217 NULL, 0, 0);
2218}
2219
Shachar Raindel93e478d2015-04-14 15:46:35 -07002220static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2221 unsigned long address, pte_t *page_table,
2222 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2223 struct page *old_page)
2224 __releases(ptl)
2225{
2226 int page_mkwrite = 0;
2227
2228 page_cache_get(old_page);
2229
2230 /*
2231 * Only catch write-faults on shared writable pages,
2232 * read-only shared pages can get COWed by
2233 * get_user_pages(.write=1, .force=1).
2234 */
2235 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2236 int tmp;
2237
2238 pte_unmap_unlock(page_table, ptl);
2239 tmp = do_page_mkwrite(vma, old_page, address);
2240 if (unlikely(!tmp || (tmp &
2241 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2242 page_cache_release(old_page);
2243 return tmp;
2244 }
2245 /*
2246 * Since we dropped the lock we need to revalidate
2247 * the PTE as someone else may have changed it. If
2248 * they did, we just return, as we can count on the
2249 * MMU to tell us if they didn't also make it writable.
2250 */
2251 page_table = pte_offset_map_lock(mm, pmd, address,
2252 &ptl);
2253 if (!pte_same(*page_table, orig_pte)) {
2254 unlock_page(old_page);
2255 pte_unmap_unlock(page_table, ptl);
2256 page_cache_release(old_page);
2257 return 0;
2258 }
2259 page_mkwrite = 1;
2260 }
2261
2262 return wp_page_reuse(mm, vma, address, page_table, ptl,
2263 orig_pte, old_page, page_mkwrite, 1);
2264}
2265
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002266/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 * This routine handles present pages, when users try to write
2268 * to a shared page. It is done by copying the page to a new address
2269 * and decrementing the shared-page counter for the old page.
2270 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 * Note that this routine assumes that the protection checks have been
2272 * done by the caller (the low-level page fault routine in most cases).
2273 * Thus we can safely just mark it writable once we've done any necessary
2274 * COW.
2275 *
2276 * We also mark the page dirty at this point even though the page will
2277 * change only once the write actually happens. This avoids a few races,
2278 * and potentially makes it more efficient.
2279 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002280 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2281 * but allow concurrent faults), with pte both mapped and locked.
2282 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002284static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2285 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002286 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002287 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288{
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002289 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290
Linus Torvalds6aab3412005-11-28 14:34:23 -08002291 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002292 if (!old_page) {
2293 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002294 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2295 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002296 *
2297 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002298 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002299 */
2300 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2301 (VM_WRITE|VM_SHARED))
Boaz Harroshdd906182015-04-15 16:15:11 -07002302 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2303 orig_pte, pmd);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002304
2305 pte_unmap_unlock(page_table, ptl);
2306 return wp_page_copy(mm, vma, address, page_table, pmd,
2307 orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002308 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002310 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002311 * Take out anonymous pages first, anonymous shared vmas are
2312 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002313 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002314 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002315 if (!trylock_page(old_page)) {
2316 page_cache_get(old_page);
2317 pte_unmap_unlock(page_table, ptl);
2318 lock_page(old_page);
2319 page_table = pte_offset_map_lock(mm, pmd, address,
2320 &ptl);
2321 if (!pte_same(*page_table, orig_pte)) {
2322 unlock_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002323 pte_unmap_unlock(page_table, ptl);
2324 page_cache_release(old_page);
2325 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002326 }
2327 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002328 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002329 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002330 /*
2331 * The page is all ours. Move it to our anon_vma so
2332 * the rmap code will not search our parent or siblings.
2333 * Protected against the rmap code by the page lock.
2334 */
2335 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002336 unlock_page(old_page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002337 return wp_page_reuse(mm, vma, address, page_table, ptl,
2338 orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002339 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002340 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002341 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002342 (VM_WRITE|VM_SHARED))) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002343 return wp_page_shared(mm, vma, address, page_table, pmd,
2344 ptl, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
2347 /*
2348 * Ok, we need to copy. Oh, well..
2349 */
Nick Pigginb5810032005-10-29 18:16:12 -07002350 page_cache_get(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002351
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002352 pte_unmap_unlock(page_table, ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002353 return wp_page_copy(mm, vma, address, page_table, pmd,
2354 orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355}
2356
Peter Zijlstra97a89412011-05-24 17:12:04 -07002357static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 unsigned long start_addr, unsigned long end_addr,
2359 struct zap_details *details)
2360{
Al Virof5cc4ee2012-03-05 14:14:20 -05002361 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362}
2363
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002364static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 struct zap_details *details)
2366{
2367 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 pgoff_t vba, vea, zba, zea;
2369
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002370 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372
2373 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002374 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2376 zba = details->first_index;
2377 if (zba < vba)
2378 zba = vba;
2379 zea = details->last_index;
2380 if (zea > vea)
2381 zea = vea;
2382
Peter Zijlstra97a89412011-05-24 17:12:04 -07002383 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2385 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002386 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387 }
2388}
2389
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002391 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2392 * address_space corresponding to the specified page range in the underlying
2393 * file.
2394 *
Martin Waitz3d410882005-06-23 22:05:21 -07002395 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 * @holebegin: byte in first page to unmap, relative to the start of
2397 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002398 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 * must keep the partial page. In contrast, we must get rid of
2400 * partial pages.
2401 * @holelen: size of prospective hole in bytes. This will be rounded
2402 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2403 * end of the file.
2404 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2405 * but 0 when invalidating pagecache, don't throw away private data.
2406 */
2407void unmap_mapping_range(struct address_space *mapping,
2408 loff_t const holebegin, loff_t const holelen, int even_cows)
2409{
2410 struct zap_details details;
2411 pgoff_t hba = holebegin >> PAGE_SHIFT;
2412 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2413
2414 /* Check for overflow. */
2415 if (sizeof(holelen) > sizeof(hlen)) {
2416 long long holeend =
2417 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2418 if (holeend & ~(long long)ULONG_MAX)
2419 hlen = ULONG_MAX - hba + 1;
2420 }
2421
2422 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423 details.first_index = hba;
2424 details.last_index = hba + hlen - 1;
2425 if (details.last_index < details.first_index)
2426 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427
Peter Zijlstra97a89412011-05-24 17:12:04 -07002428
Matthew Wilcox283307c2015-02-16 15:58:46 -08002429 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
Kirill A. Shutemov48ec8332014-12-22 21:01:54 +02002430 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002431 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov48ec8332014-12-22 21:01:54 +02002433 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434}
2435EXPORT_SYMBOL(unmap_mapping_range);
2436
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002438 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2439 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002440 * We return with pte unmapped and unlocked.
2441 *
2442 * We return with the mmap_sem locked or unlocked in the same cases
2443 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002445static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2446 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002447 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002449 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002450 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002451 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002452 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002454 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002455 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002456 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002458 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002459 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002460
2461 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002462 if (unlikely(non_swap_entry(entry))) {
2463 if (is_migration_entry(entry)) {
2464 migration_entry_wait(mm, pmd, address);
2465 } else if (is_hwpoison_entry(entry)) {
2466 ret = VM_FAULT_HWPOISON;
2467 } else {
2468 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002469 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002470 }
Christoph Lameter06972122006-06-23 02:03:35 -07002471 goto out;
2472 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002473 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 page = lookup_swap_cache(entry);
2475 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002476 page = swapin_readahead(entry,
2477 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 if (!page) {
2479 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002480 * Back out if somebody else faulted in this pte
2481 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002483 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 if (likely(pte_same(*page_table, orig_pte)))
2485 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002486 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002487 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 }
2489
2490 /* Had to read the page from swap area: Major fault */
2491 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002492 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002493 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002494 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002495 /*
2496 * hwpoisoned dirty swapcache pages are kept for killing
2497 * owner processes (which may be unknown at hwpoison time)
2498 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002499 ret = VM_FAULT_HWPOISON;
2500 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002501 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002502 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 }
2504
Hugh Dickins56f31802013-02-22 16:36:10 -08002505 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002506 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002507
Balbir Singh20a10222007-11-14 17:00:33 -08002508 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002509 if (!locked) {
2510 ret |= VM_FAULT_RETRY;
2511 goto out_release;
2512 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002514 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002515 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2516 * release the swapcache from under us. The page pin, and pte_same
2517 * test below, are not enough to exclude that. Even if it is still
2518 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002519 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002520 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002521 goto out_page;
2522
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002523 page = ksm_might_need_to_copy(page, vma, address);
2524 if (unlikely(!page)) {
2525 ret = VM_FAULT_OOM;
2526 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002527 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002528 }
2529
Johannes Weiner00501b52014-08-08 14:19:20 -07002530 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002531 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002532 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002533 }
2534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002536 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002538 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002539 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002540 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002541
2542 if (unlikely(!PageUptodate(page))) {
2543 ret = VM_FAULT_SIGBUS;
2544 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 }
2546
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002547 /*
2548 * The page isn't present yet, go ahead with the fault.
2549 *
2550 * Be careful about the sequence of operations here.
2551 * To get its accounting right, reuse_swap_page() must be called
2552 * while the page is counted on swap but not yet in mapcount i.e.
2553 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2554 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002555 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002557 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002558 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002560 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002562 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002563 ret |= VM_FAULT_WRITE;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002564 exclusive = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002567 if (pte_swp_soft_dirty(orig_pte))
2568 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002570 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002571 do_page_add_anon_rmap(page, vma, address, exclusive);
Johannes Weiner00501b52014-08-08 14:19:20 -07002572 mem_cgroup_commit_charge(page, memcg, true);
2573 } else { /* ksm created a completely new copy */
Hugh Dickins56f31802013-02-22 16:36:10 -08002574 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002575 mem_cgroup_commit_charge(page, memcg, false);
2576 lru_cache_add_active_or_unevictable(page, vma);
2577 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002579 swap_free(entry);
Nick Pigginb291f002008-10-18 20:26:44 -07002580 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002581 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002582 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002583 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002584 /*
2585 * Hold the lock to avoid the swap entry to be reused
2586 * until we take the PT lock for the pte_same() check
2587 * (to avoid false positives from pte_same). For
2588 * further safety release the lock after the swap_free
2589 * so that the swap count won't change under a
2590 * parallel locked swapcache.
2591 */
2592 unlock_page(swapcache);
2593 page_cache_release(swapcache);
2594 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002595
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002596 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002597 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2598 if (ret & VM_FAULT_ERROR)
2599 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600 goto out;
2601 }
2602
2603 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002604 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002605unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002606 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607out:
2608 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002609out_nomap:
Johannes Weiner00501b52014-08-08 14:19:20 -07002610 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002611 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002612out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002613 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002614out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002615 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002616 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002617 unlock_page(swapcache);
2618 page_cache_release(swapcache);
2619 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002620 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621}
2622
2623/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002624 * This is like a special single-page "expand_{down|up}wards()",
2625 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002626 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002627 */
2628static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2629{
2630 address &= PAGE_MASK;
2631 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002632 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002633
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002634 /*
2635 * Is there a mapping abutting this one below?
2636 *
2637 * That's only ok if it's the same stack mapping
2638 * that has gotten split..
2639 */
2640 if (prev && prev->vm_end == address)
2641 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2642
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002643 return expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002644 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002645 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2646 struct vm_area_struct *next = vma->vm_next;
2647
2648 /* As VM_GROWSDOWN but s/below/above/ */
2649 if (next && next->vm_start == address + PAGE_SIZE)
2650 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2651
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002652 return expand_upwards(vma, address + PAGE_SIZE);
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002653 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002654 return 0;
2655}
2656
2657/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002658 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2659 * but allow concurrent faults), and pte mapped but not yet locked.
2660 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002662static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2663 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002664 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665{
Johannes Weiner00501b52014-08-08 14:19:20 -07002666 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002667 struct page *page;
2668 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002669 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670
Linus Torvalds11ac5522010-08-14 11:44:56 -07002671 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002672
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002673 /* File mapping without ->vm_ops ? */
2674 if (vma->vm_flags & VM_SHARED)
2675 return VM_FAULT_SIGBUS;
2676
Linus Torvalds11ac5522010-08-14 11:44:56 -07002677 /* Check if we need to add a guard page to the stack */
2678 if (check_stack_guard_page(vma, address) < 0)
Linus Torvalds9c145c562015-01-29 11:15:17 -08002679 return VM_FAULT_SIGSEGV;
Linus Torvalds11ac5522010-08-14 11:44:56 -07002680
2681 /* Use the zero-page for reads */
Dominik Dingel593befa2014-10-23 12:07:44 +02002682 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
Hugh Dickins62eede62009-09-21 17:03:34 -07002683 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2684 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002685 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002686 if (!pte_none(*page_table))
2687 goto unlock;
2688 goto setpte;
2689 }
2690
Nick Piggin557ed1f2007-10-16 01:24:40 -07002691 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002692 if (unlikely(anon_vma_prepare(vma)))
2693 goto oom;
2694 page = alloc_zeroed_user_highpage_movable(vma, address);
2695 if (!page)
2696 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002697
2698 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg))
2699 goto oom_free_page;
2700
Minchan Kim52f37622013-04-29 15:08:15 -07002701 /*
2702 * The memory barrier inside __SetPageUptodate makes sure that
2703 * preceeding stores to the page contents become visible before
2704 * the set_pte_at() write.
2705 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002706 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
Nick Piggin557ed1f2007-10-16 01:24:40 -07002708 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002709 if (vma->vm_flags & VM_WRITE)
2710 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002711
Nick Piggin557ed1f2007-10-16 01:24:40 -07002712 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002713 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002714 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002715
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002716 inc_mm_counter_fast(mm, MM_ANONPAGES);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002717 page_add_new_anon_rmap(page, vma, address);
Johannes Weiner00501b52014-08-08 14:19:20 -07002718 mem_cgroup_commit_charge(page, memcg, false);
2719 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002720setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002721 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722
2723 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002724 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002725unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002726 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002727 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002728release:
Johannes Weiner00501b52014-08-08 14:19:20 -07002729 mem_cgroup_cancel_charge(page, memcg);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002730 page_cache_release(page);
2731 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002732oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002733 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002734oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 return VM_FAULT_OOM;
2736}
2737
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002738/*
2739 * The mmap_sem must have been held on entry, and may have been
2740 * released depending on flags and vma->vm_ops->fault() return value.
2741 * See filemap_fault() and __lock_page_retry().
2742 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002743static int __do_fault(struct vm_area_struct *vma, unsigned long address,
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002744 pgoff_t pgoff, unsigned int flags,
2745 struct page *cow_page, struct page **page)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002746{
2747 struct vm_fault vmf;
2748 int ret;
2749
2750 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2751 vmf.pgoff = pgoff;
2752 vmf.flags = flags;
2753 vmf.page = NULL;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002754 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002755
2756 ret = vma->vm_ops->fault(vma, &vmf);
2757 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2758 return ret;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002759 if (!vmf.page)
2760 goto out;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002761
2762 if (unlikely(PageHWPoison(vmf.page))) {
2763 if (ret & VM_FAULT_LOCKED)
2764 unlock_page(vmf.page);
2765 page_cache_release(vmf.page);
2766 return VM_FAULT_HWPOISON;
2767 }
2768
2769 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2770 lock_page(vmf.page);
2771 else
2772 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2773
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002774 out:
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002775 *page = vmf.page;
2776 return ret;
2777}
2778
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002779/**
2780 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2781 *
2782 * @vma: virtual memory area
2783 * @address: user virtual address
2784 * @page: page to map
2785 * @pte: pointer to target page table entry
2786 * @write: true, if new entry is writable
2787 * @anon: true, if it's anonymous page
2788 *
2789 * Caller must hold page table lock relevant for @pte.
2790 *
2791 * Target users are page handler itself and implementations of
2792 * vm_ops->map_pages.
2793 */
2794void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002795 struct page *page, pte_t *pte, bool write, bool anon)
2796{
2797 pte_t entry;
2798
2799 flush_icache_page(vma, page);
2800 entry = mk_pte(page, vma->vm_page_prot);
2801 if (write)
2802 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002803 if (anon) {
2804 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2805 page_add_new_anon_rmap(page, vma, address);
2806 } else {
2807 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
2808 page_add_file_rmap(page);
2809 }
2810 set_pte_at(vma->vm_mm, address, pte, entry);
2811
2812 /* no need to invalidate: a not-present page won't be cached */
2813 update_mmu_cache(vma, address, pte);
2814}
2815
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002816static unsigned long fault_around_bytes __read_mostly =
2817 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002818
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002819#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002820static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002821{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002822 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002823 return 0;
2824}
2825
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002826/*
2827 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2828 * rounded down to nearest page order. It's what do_fault_around() expects to
2829 * see.
2830 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002831static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002832{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002833 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002834 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002835 if (val > PAGE_SIZE)
2836 fault_around_bytes = rounddown_pow_of_two(val);
2837 else
2838 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002839 return 0;
2840}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002841DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2842 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002843
2844static int __init fault_around_debugfs(void)
2845{
2846 void *ret;
2847
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002848 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2849 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002850 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002851 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002852 return 0;
2853}
2854late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002855#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002856
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002857/*
2858 * do_fault_around() tries to map few pages around the fault address. The hope
2859 * is that the pages will be needed soon and this will lower the number of
2860 * faults to handle.
2861 *
2862 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2863 * not ready to be mapped: not up-to-date, locked, etc.
2864 *
2865 * This function is called with the page table lock taken. In the split ptlock
2866 * case the page table lock only protects only those entries which belong to
2867 * the page table corresponding to the fault address.
2868 *
2869 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2870 * only once.
2871 *
2872 * fault_around_pages() defines how many pages we'll try to map.
2873 * do_fault_around() expects it to return a power of two less than or equal to
2874 * PTRS_PER_PTE.
2875 *
2876 * The virtual address of the area that we map is naturally aligned to the
2877 * fault_around_pages() value (and therefore to page order). This way it's
2878 * easier to guarantee that we don't cross page table boundaries.
2879 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002880static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2881 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2882{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002883 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002884 pgoff_t max_pgoff;
2885 struct vm_fault vmf;
2886 int off;
2887
Jason Low4db0c3c2015-04-15 16:14:08 -07002888 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002889 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2890
2891 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002892 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2893 pte -= off;
2894 pgoff -= off;
2895
2896 /*
2897 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002898 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002899 */
2900 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2901 PTRS_PER_PTE - 1;
2902 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002903 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002904
2905 /* Check if it makes any sense to call ->map_pages */
2906 while (!pte_none(*pte)) {
2907 if (++pgoff > max_pgoff)
2908 return;
2909 start_addr += PAGE_SIZE;
2910 if (start_addr >= vma->vm_end)
2911 return;
2912 pte++;
2913 }
2914
2915 vmf.virtual_address = (void __user *) start_addr;
2916 vmf.pte = pte;
2917 vmf.pgoff = pgoff;
2918 vmf.max_pgoff = max_pgoff;
2919 vmf.flags = flags;
2920 vma->vm_ops->map_pages(vma, &vmf);
2921}
2922
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002923static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2924 unsigned long address, pmd_t *pmd,
2925 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2926{
2927 struct page *fault_page;
2928 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002929 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002930 int ret = 0;
2931
2932 /*
2933 * Let's call ->map_pages() first and use ->fault() as fallback
2934 * if page by the offset is not ready to be mapped (cold cache or
2935 * something).
2936 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08002937 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002938 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2939 do_fault_around(vma, address, pte, pgoff, flags);
2940 if (!pte_same(*pte, orig_pte))
2941 goto unlock_out;
2942 pte_unmap_unlock(pte, ptl);
2943 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002944
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002945 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002946 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2947 return ret;
2948
2949 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2950 if (unlikely(!pte_same(*pte, orig_pte))) {
2951 pte_unmap_unlock(pte, ptl);
2952 unlock_page(fault_page);
2953 page_cache_release(fault_page);
2954 return ret;
2955 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002956 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002957 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002958unlock_out:
2959 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002960 return ret;
2961}
2962
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002963static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2964 unsigned long address, pmd_t *pmd,
2965 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2966{
2967 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07002968 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002969 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002970 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002971 int ret;
2972
2973 if (unlikely(anon_vma_prepare(vma)))
2974 return VM_FAULT_OOM;
2975
2976 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2977 if (!new_page)
2978 return VM_FAULT_OOM;
2979
Johannes Weiner00501b52014-08-08 14:19:20 -07002980 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002981 page_cache_release(new_page);
2982 return VM_FAULT_OOM;
2983 }
2984
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002985 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002986 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2987 goto uncharge_out;
2988
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002989 if (fault_page)
2990 copy_user_highpage(new_page, fault_page, address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07002991 __SetPageUptodate(new_page);
2992
2993 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2994 if (unlikely(!pte_same(*pte, orig_pte))) {
2995 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002996 if (fault_page) {
2997 unlock_page(fault_page);
2998 page_cache_release(fault_page);
2999 } else {
3000 /*
3001 * The fault handler has no page to lock, so it holds
3002 * i_mmap_lock for read to protect against truncate.
3003 */
3004 i_mmap_unlock_read(vma->vm_file->f_mapping);
3005 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003006 goto uncharge_out;
3007 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003008 do_set_pte(vma, address, new_page, pte, true, true);
Johannes Weiner00501b52014-08-08 14:19:20 -07003009 mem_cgroup_commit_charge(new_page, memcg, false);
3010 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003011 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003012 if (fault_page) {
3013 unlock_page(fault_page);
3014 page_cache_release(fault_page);
3015 } else {
3016 /*
3017 * The fault handler has no page to lock, so it holds
3018 * i_mmap_lock for read to protect against truncate.
3019 */
3020 i_mmap_unlock_read(vma->vm_file->f_mapping);
3021 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003022 return ret;
3023uncharge_out:
Johannes Weiner00501b52014-08-08 14:19:20 -07003024 mem_cgroup_cancel_charge(new_page, memcg);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003025 page_cache_release(new_page);
3026 return ret;
3027}
3028
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003029static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003030 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003031 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003033 struct page *fault_page;
3034 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003035 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003036 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003037 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003038 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003039
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003040 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003041 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003042 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043
3044 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003045 * Check if the backing address space wants to know that the page is
3046 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003048 if (vma->vm_ops->page_mkwrite) {
3049 unlock_page(fault_page);
3050 tmp = do_page_mkwrite(vma, fault_page, address);
3051 if (unlikely(!tmp ||
3052 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3053 page_cache_release(fault_page);
3054 return tmp;
3055 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 }
3057
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003058 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3059 if (unlikely(!pte_same(*pte, orig_pte))) {
3060 pte_unmap_unlock(pte, ptl);
3061 unlock_page(fault_page);
3062 page_cache_release(fault_page);
3063 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003064 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003065 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003066 pte_unmap_unlock(pte, ptl);
3067
3068 if (set_page_dirty(fault_page))
3069 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003070 /*
3071 * Take a local copy of the address_space - page.mapping may be zeroed
3072 * by truncate after unlock_page(). The address_space itself remains
3073 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3074 * release semantics to prevent the compiler from undoing this copying.
3075 */
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003076 mapping = fault_page->mapping;
3077 unlock_page(fault_page);
3078 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3079 /*
3080 * Some device drivers do not set page.mapping but still
3081 * dirty their pages
3082 */
3083 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003084 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003085
Johannes Weiner74ec6752015-02-10 14:11:28 -08003086 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003087 file_update_time(vma->vm_file);
3088
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003089 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003090}
Nick Piggind00806b2007-07-19 01:46:57 -07003091
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003092/*
3093 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3094 * but allow concurrent faults).
3095 * The mmap_sem may have been released depending on flags and our
3096 * return value. See filemap_fault() and __lock_page_or_retry().
3097 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003098static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Nick Piggin54cb8822007-07-19 01:46:59 -07003099 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003100 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003101{
3102 pgoff_t pgoff = (((address & PAGE_MASK)
Dean Nelson0da7e012007-10-16 01:24:45 -07003103 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
Nick Piggin54cb8822007-07-19 01:46:59 -07003104
Hugh Dickins16abfa02007-10-04 16:56:06 +01003105 pte_unmap(page_table);
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003106 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3107 if (!vma->vm_ops->fault)
3108 return VM_FAULT_SIGBUS;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003109 if (!(flags & FAULT_FLAG_WRITE))
3110 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3111 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003112 if (!(vma->vm_flags & VM_SHARED))
3113 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3114 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003115 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003116}
3117
Rashika Kheriab19a9932014-04-03 14:48:02 -07003118static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003119 unsigned long addr, int page_nid,
3120 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003121{
3122 get_page(page);
3123
3124 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003125 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003126 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003127 *flags |= TNF_FAULT_LOCAL;
3128 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003129
3130 return mpol_misplaced(page, vma, addr);
3131}
3132
Rashika Kheriab19a9932014-04-03 14:48:02 -07003133static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gormand10e63f2012-10-25 14:16:31 +02003134 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3135{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003136 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003137 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003138 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003139 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003140 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003141 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003142 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003143 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003144
Mel Gormanc0e7cad2015-02-12 14:58:41 -08003145 /* A PROT_NONE fault should not end up here */
3146 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3147
Mel Gormand10e63f2012-10-25 14:16:31 +02003148 /*
3149 * The "pte" at this point cannot be used safely without
3150 * validation through pte_unmap_same(). It's of NUMA type but
3151 * the pfn may be screwed if the read is non atomic.
3152 *
Mel Gorman4d942462015-02-12 14:58:28 -08003153 * We can safely just do a "set_pte_at()", because the old
3154 * page table entry is not accessible, so there would be no
3155 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003156 */
3157 ptl = pte_lockptr(mm, pmd);
3158 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003159 if (unlikely(!pte_same(*ptep, pte))) {
3160 pte_unmap_unlock(ptep, ptl);
3161 goto out;
3162 }
3163
Mel Gorman4d942462015-02-12 14:58:28 -08003164 /* Make it present again */
3165 pte = pte_modify(pte, vma->vm_page_prot);
3166 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003167 if (was_writable)
3168 pte = pte_mkwrite(pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003169 set_pte_at(mm, addr, ptep, pte);
3170 update_mmu_cache(vma, addr, ptep);
3171
3172 page = vm_normal_page(vma, addr, pte);
3173 if (!page) {
3174 pte_unmap_unlock(ptep, ptl);
3175 return 0;
3176 }
3177
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003178 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003179 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3180 * much anyway since they can be in shared cache state. This misses
3181 * the case where a mapping is writable but the process never writes
3182 * to it but pte_write gets cleared during protection updates and
3183 * pte_dirty has unpredictable behaviour between PTE scan updates,
3184 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003185 */
Mel Gormanbea66fb2015-03-25 15:55:37 -07003186 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003187 flags |= TNF_NO_GROUP;
3188
Rik van Rieldabe1d92013-10-07 11:29:34 +01003189 /*
3190 * Flag if the page is shared between multiple address spaces. This
3191 * is later used when determining whether to group tasks together
3192 */
3193 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3194 flags |= TNF_SHARED;
3195
Peter Zijlstra90572892013-10-07 11:29:20 +01003196 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003197 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003198 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003199 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003200 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003201 put_page(page);
3202 goto out;
3203 }
3204
3205 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003206 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003207 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003208 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003209 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003210 } else
3211 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003212
3213out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003214 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003215 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003216 return 0;
3217}
3218
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219/*
3220 * These routines also need to handle stuff like marking pages dirty
3221 * and/or accessed for architectures that don't do it in hardware (most
3222 * RISC architectures). The early dirtying is also good on the i386.
3223 *
3224 * There is also a hook called "update_mmu_cache()" that architectures
3225 * with external mmu caches can use to update those (ie the Sparc or
3226 * PowerPC hashed page tables that act as extended TLBs).
3227 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003228 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3229 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003230 * We return with pte unmapped and unlocked.
3231 *
3232 * The mmap_sem may have been released depending on flags and our
3233 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003235static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003236 struct vm_area_struct *vma, unsigned long address,
3237 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238{
3239 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003240 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241
Christian Borntraegere37c6982014-12-07 21:41:33 +01003242 /*
3243 * some architectures can have larger ptes than wordsize,
3244 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3245 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3246 * The code below just needs a consistent view for the ifs and
3247 * we later double check anyway with the ptl lock held. So here
3248 * a barrier will do.
3249 */
3250 entry = *pte;
3251 barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003253 if (pte_none(entry)) {
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003254 if (vma->vm_ops)
3255 return do_fault(mm, vma, address, pte, pmd,
3256 flags, entry);
3257
3258 return do_anonymous_page(mm, vma, address, pte, pmd,
3259 flags);
Hugh Dickins65500d22005-10-29 18:15:59 -07003260 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003261 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003262 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003263 }
3264
Mel Gorman8a0516e2015-02-12 14:58:22 -08003265 if (pte_protnone(entry))
Mel Gormand10e63f2012-10-25 14:16:31 +02003266 return do_numa_page(mm, vma, address, entry, pte, pmd);
3267
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003268 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003269 spin_lock(ptl);
3270 if (unlikely(!pte_same(*pte, entry)))
3271 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003272 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003274 return do_wp_page(mm, vma, address,
3275 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003276 entry = pte_mkdirty(entry);
3277 }
3278 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003279 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003280 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003281 } else {
3282 /*
3283 * This is needed only for protection faults but the arch code
3284 * is not yet telling us if this is a protection fault or not.
3285 * This still avoids useless tlb flushes for .text page faults
3286 * with threads.
3287 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003288 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003289 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003290 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003291unlock:
3292 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003293 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294}
3295
3296/*
3297 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003298 *
3299 * The mmap_sem may have been released depending on flags and our
3300 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003302static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3303 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304{
3305 pgd_t *pgd;
3306 pud_t *pud;
3307 pmd_t *pmd;
3308 pte_t *pte;
3309
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003310 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003311 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 pud = pud_alloc(mm, pgd, address);
3315 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003316 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317 pmd = pmd_alloc(mm, pud, address);
3318 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003319 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003320 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003321 int ret = VM_FAULT_FALLBACK;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003322 if (!vma->vm_ops)
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003323 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3324 pmd, flags);
3325 if (!(ret & VM_FAULT_FALLBACK))
3326 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003327 } else {
3328 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003329 int ret;
3330
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003331 barrier();
3332 if (pmd_trans_huge(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003333 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3334
Linus Torvaldse53289c2013-01-09 08:36:54 -08003335 /*
3336 * If the pmd is splitting, return and retry the
3337 * the fault. Alternative: wait until the split
3338 * is done, and goto retry.
3339 */
3340 if (pmd_trans_splitting(orig_pmd))
3341 return 0;
3342
Mel Gorman8a0516e2015-02-12 14:58:22 -08003343 if (pmd_protnone(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003344 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003345 orig_pmd, pmd);
3346
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003347 if (dirty && !pmd_write(orig_pmd)) {
David Rientjes1f1d06c2012-05-29 15:06:23 -07003348 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3349 orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003350 if (!(ret & VM_FAULT_FALLBACK))
3351 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003352 } else {
3353 huge_pmd_set_accessed(mm, vma, address, pmd,
3354 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003355 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003356 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003357 }
3358 }
3359
3360 /*
3361 * Use __pte_alloc instead of pte_alloc_map, because we can't
3362 * run pte_offset_map on the pmd, if an huge pmd could
3363 * materialize from under us from a different thread.
3364 */
Mel Gorman4fd01772011-10-12 21:06:51 +02003365 if (unlikely(pmd_none(*pmd)) &&
3366 unlikely(__pte_alloc(mm, vma, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003367 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003368 /* if an huge pmd materialized from under us just retry later */
3369 if (unlikely(pmd_trans_huge(*pmd)))
3370 return 0;
3371 /*
3372 * A regular pmd is established and it can't morph into a huge pmd
3373 * from under us anymore at this point because we hold the mmap_sem
3374 * read mode and khugepaged takes it in write mode. So now it's
3375 * safe to run pte_offset_map().
3376 */
3377 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003379 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380}
3381
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003382/*
3383 * By the time we get here, we already hold the mm semaphore
3384 *
3385 * The mmap_sem may have been released depending on flags and our
3386 * return value. See filemap_fault() and __lock_page_or_retry().
3387 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003388int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3389 unsigned long address, unsigned int flags)
3390{
3391 int ret;
3392
3393 __set_current_state(TASK_RUNNING);
3394
3395 count_vm_event(PGFAULT);
3396 mem_cgroup_count_vm_event(mm, PGFAULT);
3397
3398 /* do counter updates before entering really critical section. */
3399 check_sync_rss_stat(current);
3400
3401 /*
3402 * Enable the memcg OOM handling for faults triggered in user
3403 * space. Kernel faults are handled more gracefully.
3404 */
3405 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003406 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003407
3408 ret = __handle_mm_fault(mm, vma, address, flags);
3409
Johannes Weiner49426422013-10-16 13:46:59 -07003410 if (flags & FAULT_FLAG_USER) {
3411 mem_cgroup_oom_disable();
3412 /*
3413 * The task may have entered a memcg OOM situation but
3414 * if the allocation error was handled gracefully (no
3415 * VM_FAULT_OOM), there is no need to kill anything.
3416 * Just clean up the OOM state peacefully.
3417 */
3418 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3419 mem_cgroup_oom_synchronize(false);
3420 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003421
Johannes Weiner519e5242013-09-12 15:13:42 -07003422 return ret;
3423}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003424EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003425
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426#ifndef __PAGETABLE_PUD_FOLDED
3427/*
3428 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003429 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003431int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003433 pud_t *new = pud_alloc_one(mm, address);
3434 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003435 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436
Nick Piggin362a61a2008-05-14 06:37:36 +02003437 smp_wmb(); /* See comment in __pte_alloc */
3438
Hugh Dickins872fec12005-10-29 18:16:21 -07003439 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003440 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003441 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003442 else
3443 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003444 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003445 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446}
3447#endif /* __PAGETABLE_PUD_FOLDED */
3448
3449#ifndef __PAGETABLE_PMD_FOLDED
3450/*
3451 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003452 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003454int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003456 pmd_t *new = pmd_alloc_one(mm, address);
3457 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003458 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459
Nick Piggin362a61a2008-05-14 06:37:36 +02003460 smp_wmb(); /* See comment in __pte_alloc */
3461
Hugh Dickins872fec12005-10-29 18:16:21 -07003462 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003464 if (!pud_present(*pud)) {
3465 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003466 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003467 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003468 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003469#else
3470 if (!pgd_present(*pud)) {
3471 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003472 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003473 } else /* Another has populated it */
3474 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003476 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003477 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478}
3479#endif /* __PAGETABLE_PMD_FOLDED */
3480
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003481static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003482 pte_t **ptepp, spinlock_t **ptlp)
3483{
3484 pgd_t *pgd;
3485 pud_t *pud;
3486 pmd_t *pmd;
3487 pte_t *ptep;
3488
3489 pgd = pgd_offset(mm, address);
3490 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3491 goto out;
3492
3493 pud = pud_offset(pgd, address);
3494 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3495 goto out;
3496
3497 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003498 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003499 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3500 goto out;
3501
3502 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3503 if (pmd_huge(*pmd))
3504 goto out;
3505
3506 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3507 if (!ptep)
3508 goto out;
3509 if (!pte_present(*ptep))
3510 goto unlock;
3511 *ptepp = ptep;
3512 return 0;
3513unlock:
3514 pte_unmap_unlock(ptep, *ptlp);
3515out:
3516 return -EINVAL;
3517}
3518
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003519static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3520 pte_t **ptepp, spinlock_t **ptlp)
3521{
3522 int res;
3523
3524 /* (void) is needed to make gcc happy */
3525 (void) __cond_lock(*ptlp,
3526 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3527 return res;
3528}
3529
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003530/**
3531 * follow_pfn - look up PFN at a user virtual address
3532 * @vma: memory mapping
3533 * @address: user virtual address
3534 * @pfn: location to store found PFN
3535 *
3536 * Only IO mappings and raw PFN mappings are allowed.
3537 *
3538 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3539 */
3540int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3541 unsigned long *pfn)
3542{
3543 int ret = -EINVAL;
3544 spinlock_t *ptl;
3545 pte_t *ptep;
3546
3547 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3548 return ret;
3549
3550 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3551 if (ret)
3552 return ret;
3553 *pfn = pte_pfn(*ptep);
3554 pte_unmap_unlock(ptep, ptl);
3555 return 0;
3556}
3557EXPORT_SYMBOL(follow_pfn);
3558
Rik van Riel28b2ee22008-07-23 21:27:05 -07003559#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003560int follow_phys(struct vm_area_struct *vma,
3561 unsigned long address, unsigned int flags,
3562 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003563{
Johannes Weiner03668a42009-06-16 15:32:34 -07003564 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003565 pte_t *ptep, pte;
3566 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003567
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003568 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3569 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003570
Johannes Weiner03668a42009-06-16 15:32:34 -07003571 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003572 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003573 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003574
Rik van Riel28b2ee22008-07-23 21:27:05 -07003575 if ((flags & FOLL_WRITE) && !pte_write(pte))
3576 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003577
3578 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003579 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003580
Johannes Weiner03668a42009-06-16 15:32:34 -07003581 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003582unlock:
3583 pte_unmap_unlock(ptep, ptl);
3584out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003585 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003586}
3587
3588int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3589 void *buf, int len, int write)
3590{
3591 resource_size_t phys_addr;
3592 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003593 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003594 int offset = addr & (PAGE_SIZE-1);
3595
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003596 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003597 return -EINVAL;
3598
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003599 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003600 if (write)
3601 memcpy_toio(maddr + offset, buf, len);
3602 else
3603 memcpy_fromio(buf, maddr + offset, len);
3604 iounmap(maddr);
3605
3606 return len;
3607}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003608EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003609#endif
3610
David Howells0ec76a12006-09-27 01:50:15 -07003611/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003612 * Access another process' address space as given in mm. If non-NULL, use the
3613 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003614 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003615static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3616 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003617{
David Howells0ec76a12006-09-27 01:50:15 -07003618 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003619 void *old_buf = buf;
3620
David Howells0ec76a12006-09-27 01:50:15 -07003621 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003622 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003623 while (len) {
3624 int bytes, ret, offset;
3625 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003626 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003627
3628 ret = get_user_pages(tsk, mm, addr, 1,
3629 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003630 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003631#ifndef CONFIG_HAVE_IOREMAP_PROT
3632 break;
3633#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003634 /*
3635 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3636 * we can access using slightly different code.
3637 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003638 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003639 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003640 break;
3641 if (vma->vm_ops && vma->vm_ops->access)
3642 ret = vma->vm_ops->access(vma, addr, buf,
3643 len, write);
3644 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003645 break;
3646 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003647#endif
David Howells0ec76a12006-09-27 01:50:15 -07003648 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003649 bytes = len;
3650 offset = addr & (PAGE_SIZE-1);
3651 if (bytes > PAGE_SIZE-offset)
3652 bytes = PAGE_SIZE-offset;
3653
3654 maddr = kmap(page);
3655 if (write) {
3656 copy_to_user_page(vma, page, addr,
3657 maddr + offset, buf, bytes);
3658 set_page_dirty_lock(page);
3659 } else {
3660 copy_from_user_page(vma, page, addr,
3661 buf, maddr + offset, bytes);
3662 }
3663 kunmap(page);
3664 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003665 }
David Howells0ec76a12006-09-27 01:50:15 -07003666 len -= bytes;
3667 buf += bytes;
3668 addr += bytes;
3669 }
3670 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003671
3672 return buf - old_buf;
3673}
Andi Kleen03252912008-01-30 13:33:18 +01003674
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003675/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003676 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003677 * @mm: the mm_struct of the target address space
3678 * @addr: start address to access
3679 * @buf: source or destination buffer
3680 * @len: number of bytes to transfer
3681 * @write: whether the access is a write
3682 *
3683 * The caller must hold a reference on @mm.
3684 */
3685int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3686 void *buf, int len, int write)
3687{
3688 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3689}
3690
Andi Kleen03252912008-01-30 13:33:18 +01003691/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003692 * Access another process' address space.
3693 * Source/target buffer must be kernel space,
3694 * Do not walk the page table directly, use get_user_pages
3695 */
3696int access_process_vm(struct task_struct *tsk, unsigned long addr,
3697 void *buf, int len, int write)
3698{
3699 struct mm_struct *mm;
3700 int ret;
3701
3702 mm = get_task_mm(tsk);
3703 if (!mm)
3704 return 0;
3705
3706 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3707 mmput(mm);
3708
3709 return ret;
3710}
3711
Andi Kleen03252912008-01-30 13:33:18 +01003712/*
3713 * Print the name of a VMA.
3714 */
3715void print_vma_addr(char *prefix, unsigned long ip)
3716{
3717 struct mm_struct *mm = current->mm;
3718 struct vm_area_struct *vma;
3719
Ingo Molnare8bff742008-02-13 20:21:06 +01003720 /*
3721 * Do not print if we are in atomic
3722 * contexts (in exception stacks, etc.):
3723 */
3724 if (preempt_count())
3725 return;
3726
Andi Kleen03252912008-01-30 13:33:18 +01003727 down_read(&mm->mmap_sem);
3728 vma = find_vma(mm, ip);
3729 if (vma && vma->vm_file) {
3730 struct file *f = vma->vm_file;
3731 char *buf = (char *)__get_free_page(GFP_KERNEL);
3732 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003733 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003734
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02003735 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003736 if (IS_ERR(p))
3737 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003738 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003739 vma->vm_start,
3740 vma->vm_end - vma->vm_start);
3741 free_page((unsigned long)buf);
3742 }
3743 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003744 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003745}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003746
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003747#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02003748void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003749{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003750 /*
3751 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3752 * holding the mmap_sem, this is safe because kernel memory doesn't
3753 * get paged out, therefore we'll never actually fault, and the
3754 * below annotations will generate false positives.
3755 */
3756 if (segment_eq(get_fs(), KERNEL_DS))
3757 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003758 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003759 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003760 __might_sleep(file, line, 0);
3761#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003762 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003763 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02003764#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003765}
David Hildenbrand9ec23532015-05-11 17:52:07 +02003766EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003767#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003768
3769#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3770static void clear_gigantic_page(struct page *page,
3771 unsigned long addr,
3772 unsigned int pages_per_huge_page)
3773{
3774 int i;
3775 struct page *p = page;
3776
3777 might_sleep();
3778 for (i = 0; i < pages_per_huge_page;
3779 i++, p = mem_map_next(p, page, i)) {
3780 cond_resched();
3781 clear_user_highpage(p, addr + i * PAGE_SIZE);
3782 }
3783}
3784void clear_huge_page(struct page *page,
3785 unsigned long addr, unsigned int pages_per_huge_page)
3786{
3787 int i;
3788
3789 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3790 clear_gigantic_page(page, addr, pages_per_huge_page);
3791 return;
3792 }
3793
3794 might_sleep();
3795 for (i = 0; i < pages_per_huge_page; i++) {
3796 cond_resched();
3797 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3798 }
3799}
3800
3801static void copy_user_gigantic_page(struct page *dst, struct page *src,
3802 unsigned long addr,
3803 struct vm_area_struct *vma,
3804 unsigned int pages_per_huge_page)
3805{
3806 int i;
3807 struct page *dst_base = dst;
3808 struct page *src_base = src;
3809
3810 for (i = 0; i < pages_per_huge_page; ) {
3811 cond_resched();
3812 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3813
3814 i++;
3815 dst = mem_map_next(dst, dst_base, i);
3816 src = mem_map_next(src, src_base, i);
3817 }
3818}
3819
3820void copy_user_huge_page(struct page *dst, struct page *src,
3821 unsigned long addr, struct vm_area_struct *vma,
3822 unsigned int pages_per_huge_page)
3823{
3824 int i;
3825
3826 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3827 copy_user_gigantic_page(dst, src, addr, vma,
3828 pages_per_huge_page);
3829 return;
3830 }
3831
3832 might_sleep();
3833 for (i = 0; i < pages_per_huge_page; i++) {
3834 cond_resched();
3835 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3836 }
3837}
3838#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003839
Olof Johansson40b64ac2013-12-20 14:28:05 -08003840#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003841
3842static struct kmem_cache *page_ptl_cachep;
3843
3844void __init ptlock_cache_init(void)
3845{
3846 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3847 SLAB_PANIC, NULL);
3848}
3849
Peter Zijlstra539edb52013-11-14 14:31:52 -08003850bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003851{
3852 spinlock_t *ptl;
3853
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003854 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003855 if (!ptl)
3856 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003857 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003858 return true;
3859}
3860
Peter Zijlstra539edb52013-11-14 14:31:52 -08003861void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003862{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003863 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003864}
3865#endif