blob: 098f00d0546170d6a969d4994b27cfab7a4955dc [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/hugetlb.h>
44#include <linux/mman.h>
45#include <linux/swap.h>
46#include <linux/highmem.h>
47#include <linux/pagemap.h>
Hugh Dickins9a840892009-09-21 17:02:01 -070048#include <linux/ksm.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070049#include <linux/rmap.h>
Paul Gortmakerb95f1b312011-10-16 02:01:52 -040050#include <linux/export.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070051#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052#include <linux/init.h>
Dan Williams01c8f1c2016-01-15 16:56:40 -080053#include <linux/pfn_t.h>
Peter Zijlstraedc79b22006-09-25 23:30:58 -070054#include <linux/writeback.h>
Balbir Singh8a9f3cc2008-02-07 00:13:53 -080055#include <linux/memcontrol.h>
Andrea Arcangelicddb8a52008-07-28 15:46:29 -070056#include <linux/mmu_notifier.h>
Hugh Dickins3dc14742009-01-06 14:40:08 -080057#include <linux/kallsyms.h>
58#include <linux/swapops.h>
59#include <linux/elf.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090060#include <linux/gfp.h>
Mel Gorman4daae3b2012-11-02 11:33:45 +000061#include <linux/migrate.h>
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -080062#include <linux/string.h>
Dan Williams0abdd7a2014-01-21 15:48:12 -080063#include <linux/dma-debug.h>
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -070064#include <linux/debugfs.h>
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -070065#include <linux/userfaultfd_k.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070066
Alexey Dobriyan6952b612009-09-18 23:55:55 +040067#include <asm/io.h>
Dave Hansen33a709b2016-02-12 13:02:19 -080068#include <asm/mmu_context.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070069#include <asm/pgalloc.h>
70#include <asm/uaccess.h>
71#include <asm/tlb.h>
72#include <asm/tlbflush.h>
73#include <asm/pgtable.h>
74
Jan Beulich42b77722008-07-23 21:27:10 -070075#include "internal.h"
76
Peter Zijlstra90572892013-10-07 11:29:20 +010077#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
78#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080079#endif
80
Andy Whitcroftd41dee32005-06-23 00:07:54 -070081#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070082/* use the per-pgdat data instead for discontigmem - mbligh */
83unsigned long max_mapnr;
84struct page *mem_map;
85
86EXPORT_SYMBOL(max_mapnr);
87EXPORT_SYMBOL(mem_map);
88#endif
89
Linus Torvalds1da177e2005-04-16 15:20:36 -070090/*
91 * A number of key systems in x86 including ioremap() rely on the assumption
92 * that high_memory defines the upper bound on direct map memory, then end
93 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
95 * and ZONE_HIGHMEM.
96 */
97void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070098
Linus Torvalds1da177e2005-04-16 15:20:36 -070099EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100
Ingo Molnar32a93232008-02-06 22:39:44 +0100101/*
102 * Randomize the address space (stacks, mmaps, brk, etc.).
103 *
104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
105 * as ancient (libc5 based) binaries can segfault. )
106 */
107int randomize_va_space __read_mostly =
108#ifdef CONFIG_COMPAT_BRK
109 1;
110#else
111 2;
112#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100113
114static int __init disable_randmaps(char *s)
115{
116 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800117 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100118}
119__setup("norandmaps", disable_randmaps);
120
Hugh Dickins62eede62009-09-21 17:03:34 -0700121unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700122unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700123
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200124EXPORT_SYMBOL(zero_pfn);
125
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700126/*
127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
128 */
129static int __init init_zero_pfn(void)
130{
131 zero_pfn = page_to_pfn(ZERO_PAGE(0));
132 return 0;
133}
134core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100135
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800136
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800137#if defined(SPLIT_RSS_COUNTING)
138
David Rientjesea48cf72012-03-21 16:34:13 -0700139void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800140{
141 int i;
142
143 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700144 if (current->rss_stat.count[i]) {
145 add_mm_counter(mm, i, current->rss_stat.count[i]);
146 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800147 }
148 }
David Rientjes05af2e12012-03-21 16:34:13 -0700149 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800150}
151
152static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
153{
154 struct task_struct *task = current;
155
156 if (likely(task->mm == mm))
157 task->rss_stat.count[member] += val;
158 else
159 add_mm_counter(mm, member, val);
160}
161#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
162#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
163
164/* sync counter once per 64 page faults */
165#define TASK_RSS_EVENTS_THRESH (64)
166static void check_sync_rss_stat(struct task_struct *task)
167{
168 if (unlikely(task != current))
169 return;
170 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700171 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800172}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700173#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800174
175#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
176#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
177
178static void check_sync_rss_stat(struct task_struct *task)
179{
180}
181
Peter Zijlstra9547d012011-05-24 17:12:14 -0700182#endif /* SPLIT_RSS_COUNTING */
183
184#ifdef HAVE_GENERIC_MMU_GATHER
185
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700186static bool tlb_next_batch(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700187{
188 struct mmu_gather_batch *batch;
189
190 batch = tlb->active;
191 if (batch->next) {
192 tlb->active = batch->next;
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700193 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700194 }
195
Michal Hocko53a59fc2013-01-04 15:35:12 -0800196 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700197 return false;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800198
Peter Zijlstra9547d012011-05-24 17:12:14 -0700199 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
200 if (!batch)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700201 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700202
Michal Hocko53a59fc2013-01-04 15:35:12 -0800203 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700204 batch->next = NULL;
205 batch->nr = 0;
206 batch->max = MAX_GATHER_BATCH;
207
208 tlb->active->next = batch;
209 tlb->active = batch;
210
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700211 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700212}
213
214/* tlb_gather_mmu
215 * Called to initialize an (on-stack) mmu_gather structure for page-table
216 * tear-down from @mm. The @fullmm argument is used when @mm is without
217 * users and we're going to destroy the full address space (exit/execve).
218 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700219void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700220{
221 tlb->mm = mm;
222
Linus Torvalds2b047252013-08-15 11:42:25 -0700223 /* Is it from 0 to ~0? */
224 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700225 tlb->need_flush_all = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700226 tlb->local.next = NULL;
227 tlb->local.nr = 0;
228 tlb->local.max = ARRAY_SIZE(tlb->__pages);
229 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800230 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700231
232#ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800234#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000235
236 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700237}
238
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700239static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700240{
Will Deacon721c21c2015-01-12 19:10:55 +0000241 if (!tlb->end)
242 return;
243
Peter Zijlstra9547d012011-05-24 17:12:14 -0700244 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100245 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700246#ifdef CONFIG_HAVE_RCU_TABLE_FREE
247 tlb_table_flush(tlb);
248#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000249 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700250}
251
252static void tlb_flush_mmu_free(struct mmu_gather *tlb)
253{
254 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700255
Will Deacon721c21c2015-01-12 19:10:55 +0000256 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700257 free_pages_and_swap_cache(batch->pages, batch->nr);
258 batch->nr = 0;
259 }
260 tlb->active = &tlb->local;
261}
262
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700263void tlb_flush_mmu(struct mmu_gather *tlb)
264{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700265 tlb_flush_mmu_tlbonly(tlb);
266 tlb_flush_mmu_free(tlb);
267}
268
Peter Zijlstra9547d012011-05-24 17:12:14 -0700269/* tlb_finish_mmu
270 * Called at the end of the shootdown operation to free up any resources
271 * that were required.
272 */
273void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
274{
275 struct mmu_gather_batch *batch, *next;
276
277 tlb_flush_mmu(tlb);
278
279 /* keep the page table cache within bounds */
280 check_pgt_cache();
281
282 for (batch = tlb->local.next; batch; batch = next) {
283 next = batch->next;
284 free_pages((unsigned long)batch, 0);
285 }
286 tlb->local.next = NULL;
287}
288
289/* __tlb_remove_page
290 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
291 * handling the additional races in SMP caused by other CPUs caching valid
292 * mappings in their TLBs. Returns the number of free page slots left.
293 * When out of page slots we must call tlb_flush_mmu().
294 */
295int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
296{
297 struct mmu_gather_batch *batch;
298
Will Deaconfb7332a2014-10-29 10:03:09 +0000299 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700300
Peter Zijlstra9547d012011-05-24 17:12:14 -0700301 batch = tlb->active;
302 batch->pages[batch->nr++] = page;
303 if (batch->nr == batch->max) {
304 if (!tlb_next_batch(tlb))
305 return 0;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700306 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700307 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800308 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700309
310 return batch->max - batch->nr;
311}
312
313#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800314
Peter Zijlstra26723912011-05-24 17:12:00 -0700315#ifdef CONFIG_HAVE_RCU_TABLE_FREE
316
317/*
318 * See the comment near struct mmu_table_batch.
319 */
320
321static void tlb_remove_table_smp_sync(void *arg)
322{
323 /* Simply deliver the interrupt */
324}
325
326static void tlb_remove_table_one(void *table)
327{
328 /*
329 * This isn't an RCU grace period and hence the page-tables cannot be
330 * assumed to be actually RCU-freed.
331 *
332 * It is however sufficient for software page-table walkers that rely on
333 * IRQ disabling. See the comment near struct mmu_table_batch.
334 */
335 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
336 __tlb_remove_table(table);
337}
338
339static void tlb_remove_table_rcu(struct rcu_head *head)
340{
341 struct mmu_table_batch *batch;
342 int i;
343
344 batch = container_of(head, struct mmu_table_batch, rcu);
345
346 for (i = 0; i < batch->nr; i++)
347 __tlb_remove_table(batch->tables[i]);
348
349 free_page((unsigned long)batch);
350}
351
352void tlb_table_flush(struct mmu_gather *tlb)
353{
354 struct mmu_table_batch **batch = &tlb->batch;
355
356 if (*batch) {
357 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
358 *batch = NULL;
359 }
360}
361
362void tlb_remove_table(struct mmu_gather *tlb, void *table)
363{
364 struct mmu_table_batch **batch = &tlb->batch;
365
Peter Zijlstra26723912011-05-24 17:12:00 -0700366 /*
367 * When there's less then two users of this mm there cannot be a
368 * concurrent page-table walk.
369 */
370 if (atomic_read(&tlb->mm->mm_users) < 2) {
371 __tlb_remove_table(table);
372 return;
373 }
374
375 if (*batch == NULL) {
376 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
377 if (*batch == NULL) {
378 tlb_remove_table_one(table);
379 return;
380 }
381 (*batch)->nr = 0;
382 }
383 (*batch)->tables[(*batch)->nr++] = table;
384 if ((*batch)->nr == MAX_TABLE_BATCH)
385 tlb_table_flush(tlb);
386}
387
Peter Zijlstra9547d012011-05-24 17:12:14 -0700388#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700389
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 * Note: this doesn't free the actual pages themselves. That
392 * has been handled earlier when unmapping all the memory regions.
393 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000394static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
395 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800397 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700398 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000399 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800400 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401}
402
Hugh Dickinse0da3822005-04-19 13:29:15 -0700403static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
404 unsigned long addr, unsigned long end,
405 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
407 pmd_t *pmd;
408 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700409 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
Hugh Dickinse0da3822005-04-19 13:29:15 -0700411 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413 do {
414 next = pmd_addr_end(addr, end);
415 if (pmd_none_or_clear_bad(pmd))
416 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000417 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 } while (pmd++, addr = next, addr != end);
419
Hugh Dickinse0da3822005-04-19 13:29:15 -0700420 start &= PUD_MASK;
421 if (start < floor)
422 return;
423 if (ceiling) {
424 ceiling &= PUD_MASK;
425 if (!ceiling)
426 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700428 if (end - 1 > ceiling - 1)
429 return;
430
431 pmd = pmd_offset(pud, start);
432 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000433 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800434 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435}
436
Hugh Dickinse0da3822005-04-19 13:29:15 -0700437static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
438 unsigned long addr, unsigned long end,
439 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440{
441 pud_t *pud;
442 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700443 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444
Hugh Dickinse0da3822005-04-19 13:29:15 -0700445 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 do {
448 next = pud_addr_end(addr, end);
449 if (pud_none_or_clear_bad(pud))
450 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700451 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 } while (pud++, addr = next, addr != end);
453
Hugh Dickinse0da3822005-04-19 13:29:15 -0700454 start &= PGDIR_MASK;
455 if (start < floor)
456 return;
457 if (ceiling) {
458 ceiling &= PGDIR_MASK;
459 if (!ceiling)
460 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700462 if (end - 1 > ceiling - 1)
463 return;
464
465 pud = pud_offset(pgd, start);
466 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000467 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468}
469
470/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700471 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472 */
Jan Beulich42b77722008-07-23 21:27:10 -0700473void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700474 unsigned long addr, unsigned long end,
475 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700476{
477 pgd_t *pgd;
478 unsigned long next;
479
Hugh Dickinse0da3822005-04-19 13:29:15 -0700480 /*
481 * The next few lines have given us lots of grief...
482 *
483 * Why are we testing PMD* at this top level? Because often
484 * there will be no work to do at all, and we'd prefer not to
485 * go all the way down to the bottom just to discover that.
486 *
487 * Why all these "- 1"s? Because 0 represents both the bottom
488 * of the address space and the top of it (using -1 for the
489 * top wouldn't help much: the masks would do the wrong thing).
490 * The rule is that addr 0 and floor 0 refer to the bottom of
491 * the address space, but end 0 and ceiling 0 refer to the top
492 * Comparisons need to use "end - 1" and "ceiling - 1" (though
493 * that end 0 case should be mythical).
494 *
495 * Wherever addr is brought up or ceiling brought down, we must
496 * be careful to reject "the opposite 0" before it confuses the
497 * subsequent tests. But what about where end is brought down
498 * by PMD_SIZE below? no, end can't go down to 0 there.
499 *
500 * Whereas we round start (addr) and ceiling down, by different
501 * masks at different levels, in order to test whether a table
502 * now has no other vmas using it, so can be freed, we don't
503 * bother to round floor or end up - the tests don't need that.
504 */
505
506 addr &= PMD_MASK;
507 if (addr < floor) {
508 addr += PMD_SIZE;
509 if (!addr)
510 return;
511 }
512 if (ceiling) {
513 ceiling &= PMD_MASK;
514 if (!ceiling)
515 return;
516 }
517 if (end - 1 > ceiling - 1)
518 end -= PMD_SIZE;
519 if (addr > end - 1)
520 return;
521
Jan Beulich42b77722008-07-23 21:27:10 -0700522 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 do {
524 next = pgd_addr_end(addr, end);
525 if (pgd_none_or_clear_bad(pgd))
526 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700527 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700529}
530
Jan Beulich42b77722008-07-23 21:27:10 -0700531void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700532 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700533{
534 while (vma) {
535 struct vm_area_struct *next = vma->vm_next;
536 unsigned long addr = vma->vm_start;
537
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700538 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000539 * Hide vma from rmap and truncate_pagecache before freeing
540 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700541 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800542 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700543 unlink_file_vma(vma);
544
David Gibson9da61ae2006-03-22 00:08:57 -0800545 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700546 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700547 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700548 } else {
549 /*
550 * Optimization: gather nearby vmas into one call down
551 */
552 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800553 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700554 vma = next;
555 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800556 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700557 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700558 }
559 free_pgd_range(tlb, addr, vma->vm_end,
560 floor, next? next->vm_start: ceiling);
561 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700562 vma = next;
563 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564}
565
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -0700566int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800568 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800569 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700570 if (!new)
571 return -ENOMEM;
572
Nick Piggin362a61a2008-05-14 06:37:36 +0200573 /*
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
577 *
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
585 */
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800588 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800590 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800592 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800593 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800594 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800595 if (new)
596 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700597 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598}
599
Hugh Dickins1bb36302005-10-29 18:16:22 -0700600int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700602 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
603 if (!new)
604 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Nick Piggin362a61a2008-05-14 06:37:36 +0200606 smp_wmb(); /* See comment in __pte_alloc */
607
Hugh Dickins1bb36302005-10-29 18:16:22 -0700608 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700610 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800611 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800612 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700613 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800614 if (new)
615 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700616 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617}
618
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800619static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700620{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800621 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
622}
623
624static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
625{
626 int i;
627
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800628 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700629 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800630 for (i = 0; i < NR_MM_COUNTERS; i++)
631 if (rss[i])
632 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700633}
634
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800636 * This function is called to print an error when a bad pte
637 * is found. For example, we might have a PFN-mapped pte in
638 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700639 *
640 * The calling function must still handle the error.
641 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800642static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
643 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700644{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800645 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
646 pud_t *pud = pud_offset(pgd, addr);
647 pmd_t *pmd = pmd_offset(pud, addr);
648 struct address_space *mapping;
649 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800650 static unsigned long resume;
651 static unsigned long nr_shown;
652 static unsigned long nr_unshown;
653
654 /*
655 * Allow a burst of 60 reports, then keep quiet for that minute;
656 * or allow a steady drip of one report per second.
657 */
658 if (nr_shown == 60) {
659 if (time_before(jiffies, resume)) {
660 nr_unshown++;
661 return;
662 }
663 if (nr_unshown) {
Joe Perches11705322016-03-17 14:19:50 -0700664 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
665 nr_unshown);
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800666 nr_unshown = 0;
667 }
668 nr_shown = 0;
669 }
670 if (nr_shown++ == 0)
671 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800672
673 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
674 index = linear_page_index(vma, addr);
675
Joe Perches11705322016-03-17 14:19:50 -0700676 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
677 current->comm,
678 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800679 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800680 dump_page(page, "bad pte");
Joe Perches11705322016-03-17 14:19:50 -0700681 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
682 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800683 /*
684 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
685 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700686 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
687 vma->vm_file,
688 vma->vm_ops ? vma->vm_ops->fault : NULL,
689 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
690 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700691 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030692 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700693}
694
695/*
Nick Piggin7e675132008-04-28 02:13:00 -0700696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800697 *
Nick Piggin7e675132008-04-28 02:13:00 -0700698 * "Special" mappings do not wish to be associated with a "struct page" (either
699 * it doesn't exist, or it exists but they don't want to touch it). In this
700 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700701 *
Nick Piggin7e675132008-04-28 02:13:00 -0700702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
703 * pte bit, in which case this function is trivial. Secondly, an architecture
704 * may not have a spare pte bit, which requires a more complicated scheme,
705 * described below.
706 *
707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
708 * special mapping (even if there are underlying and valid "struct pages").
709 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800710 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
714 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800715 *
716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
717 *
Nick Piggin7e675132008-04-28 02:13:00 -0700718 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700719 *
Nick Piggin7e675132008-04-28 02:13:00 -0700720 * This restricts such mappings to be a linear translation from virtual address
721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
722 * as the vma is not a COW mapping; in that case, we know that all ptes are
723 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700724 *
725 *
Nick Piggin7e675132008-04-28 02:13:00 -0700726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
727 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
729 * page" backing, however the difference is that _all_ pages with a struct
730 * page (that is, those where pfn_valid is true) are refcounted and considered
731 * normal pages by the VM. The disadvantage is that pages are refcounted
732 * (which can be slower and simply not an option for some PFNMAP users). The
733 * advantage is that we don't have to follow the strict linearity rule of
734 * PFNMAP mappings in order to support COWable mappings.
735 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800736 */
Nick Piggin7e675132008-04-28 02:13:00 -0700737#ifdef __HAVE_ARCH_PTE_SPECIAL
738# define HAVE_PTE_SPECIAL 1
739#else
740# define HAVE_PTE_SPECIAL 0
741#endif
742struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
743 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800744{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800745 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700746
747 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700748 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800749 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000750 if (vma->vm_ops && vma->vm_ops->find_special_page)
751 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700752 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
753 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700754 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800755 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700756 return NULL;
757 }
758
759 /* !HAVE_PTE_SPECIAL case follows: */
760
Jared Hulbertb379d792008-04-28 02:12:58 -0700761 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
762 if (vma->vm_flags & VM_MIXEDMAP) {
763 if (!pfn_valid(pfn))
764 return NULL;
765 goto out;
766 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700767 unsigned long off;
768 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700769 if (pfn == vma->vm_pgoff + off)
770 return NULL;
771 if (!is_cow_mapping(vma->vm_flags))
772 return NULL;
773 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800774 }
775
Hugh Dickinsb38af472014-08-29 15:18:44 -0700776 if (is_zero_pfn(pfn))
777 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800778check_pfn:
779 if (unlikely(pfn > highest_memmap_pfn)) {
780 print_bad_pte(vma, addr, pte, NULL);
781 return NULL;
782 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800783
784 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700785 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700786 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800787 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700788out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800789 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800790}
791
792/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * copy one vm_area from one task to the other. Assumes the page tables
794 * already present in the new task to be cleared in the whole range
795 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 */
797
Hugh Dickins570a335b2009-12-14 17:58:46 -0800798static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700800 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700801 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802{
Nick Pigginb5810032005-10-29 18:16:12 -0700803 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 pte_t pte = *src_pte;
805 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
807 /* pte contains position in swap or file, so copy. */
808 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800809 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700810
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800811 if (likely(!non_swap_entry(entry))) {
812 if (swap_duplicate(entry) < 0)
813 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800814
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800815 /* make sure dst_mm is on swapoff's mmlist. */
816 if (unlikely(list_empty(&dst_mm->mmlist))) {
817 spin_lock(&mmlist_lock);
818 if (list_empty(&dst_mm->mmlist))
819 list_add(&dst_mm->mmlist,
820 &src_mm->mmlist);
821 spin_unlock(&mmlist_lock);
822 }
823 rss[MM_SWAPENTS]++;
824 } else if (is_migration_entry(entry)) {
825 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800826
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800827 rss[mm_counter(page)]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800828
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800829 if (is_write_migration_entry(entry) &&
830 is_cow_mapping(vm_flags)) {
831 /*
832 * COW mappings require pages in both
833 * parent and child to be set to read.
834 */
835 make_migration_entry_read(&entry);
836 pte = swp_entry_to_pte(entry);
837 if (pte_swp_soft_dirty(*src_pte))
838 pte = pte_swp_mksoft_dirty(pte);
839 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700840 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700842 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 }
844
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 /*
846 * If it's a COW mapping, write protect it both
847 * in the parent and the child
848 */
Linus Torvalds67121172005-12-11 20:38:17 -0800849 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700851 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 }
853
854 /*
855 * If it's a shared mapping, mark it clean in
856 * the child
857 */
858 if (vm_flags & VM_SHARED)
859 pte = pte_mkclean(pte);
860 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800861
862 page = vm_normal_page(vma, addr, pte);
863 if (page) {
864 get_page(page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800865 page_dup_rmap(page, false);
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800866 rss[mm_counter(page)]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800867 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700868
869out_set_pte:
870 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800871 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872}
873
Jerome Marchand21bda262014-08-06 16:06:56 -0700874static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800875 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
876 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700878 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700880 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700881 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800882 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800883 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884
885again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800886 init_rss_vec(rss);
887
Hugh Dickinsc74df322005-10-29 18:16:23 -0700888 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 if (!dst_pte)
890 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700891 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700892 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700893 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700894 orig_src_pte = src_pte;
895 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700896 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 do {
899 /*
900 * We are holding two locks at this point - either of them
901 * could generate latencies in another task on another CPU.
902 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700903 if (progress >= 32) {
904 progress = 0;
905 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100906 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700907 break;
908 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 if (pte_none(*src_pte)) {
910 progress++;
911 continue;
912 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800913 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
914 vma, addr, rss);
915 if (entry.val)
916 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917 progress += 8;
918 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700920 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700921 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700922 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800923 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700924 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700925 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800926
927 if (entry.val) {
928 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
929 return -ENOMEM;
930 progress = 0;
931 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 if (addr != end)
933 goto again;
934 return 0;
935}
936
937static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
938 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
939 unsigned long addr, unsigned long end)
940{
941 pmd_t *src_pmd, *dst_pmd;
942 unsigned long next;
943
944 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
945 if (!dst_pmd)
946 return -ENOMEM;
947 src_pmd = pmd_offset(src_pud, addr);
948 do {
949 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -0800950 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800951 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -0800952 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800953 err = copy_huge_pmd(dst_mm, src_mm,
954 dst_pmd, src_pmd, addr, vma);
955 if (err == -ENOMEM)
956 return -ENOMEM;
957 if (!err)
958 continue;
959 /* fall through */
960 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 if (pmd_none_or_clear_bad(src_pmd))
962 continue;
963 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
964 vma, addr, next))
965 return -ENOMEM;
966 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
967 return 0;
968}
969
970static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
971 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
972 unsigned long addr, unsigned long end)
973{
974 pud_t *src_pud, *dst_pud;
975 unsigned long next;
976
977 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
978 if (!dst_pud)
979 return -ENOMEM;
980 src_pud = pud_offset(src_pgd, addr);
981 do {
982 next = pud_addr_end(addr, end);
983 if (pud_none_or_clear_bad(src_pud))
984 continue;
985 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
986 vma, addr, next))
987 return -ENOMEM;
988 } while (dst_pud++, src_pud++, addr = next, addr != end);
989 return 0;
990}
991
992int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
993 struct vm_area_struct *vma)
994{
995 pgd_t *src_pgd, *dst_pgd;
996 unsigned long next;
997 unsigned long addr = vma->vm_start;
998 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -0700999 unsigned long mmun_start; /* For mmu_notifiers */
1000 unsigned long mmun_end; /* For mmu_notifiers */
1001 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001002 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003
Nick Piggind9928952005-08-28 16:49:11 +10001004 /*
1005 * Don't copy ptes where a page fault will fill them correctly.
1006 * Fork becomes much lighter when there are big shared or private
1007 * readonly mappings. The tradeoff is that copy_page_range is more
1008 * efficient than faulting.
1009 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001010 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1011 !vma->anon_vma)
1012 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001013
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014 if (is_vm_hugetlb_page(vma))
1015 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1016
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001017 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001018 /*
1019 * We do not free on error cases below as remove_vma
1020 * gets called on error from higher level routine
1021 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001022 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001023 if (ret)
1024 return ret;
1025 }
1026
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001027 /*
1028 * We need to invalidate the secondary MMU mappings only when
1029 * there could be a permission downgrade on the ptes of the
1030 * parent mm. And a permission downgrade will only happen if
1031 * is_cow_mapping() returns true.
1032 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001033 is_cow = is_cow_mapping(vma->vm_flags);
1034 mmun_start = addr;
1035 mmun_end = end;
1036 if (is_cow)
1037 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1038 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001039
1040 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001041 dst_pgd = pgd_offset(dst_mm, addr);
1042 src_pgd = pgd_offset(src_mm, addr);
1043 do {
1044 next = pgd_addr_end(addr, end);
1045 if (pgd_none_or_clear_bad(src_pgd))
1046 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001047 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1048 vma, addr, next))) {
1049 ret = -ENOMEM;
1050 break;
1051 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001053
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001054 if (is_cow)
1055 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001056 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057}
1058
Robin Holt51c6f662005-11-13 16:06:42 -08001059static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001060 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001062 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063{
Nick Pigginb5810032005-10-29 18:16:12 -07001064 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001065 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001066 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001067 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001068 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001069 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001070 swp_entry_t entry;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001071
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001072again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001073 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001074 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1075 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001076 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001077 do {
1078 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001079 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001081 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001082
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001084 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001085
Linus Torvalds6aab3412005-11-28 14:34:23 -08001086 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 if (unlikely(details) && page) {
1088 /*
1089 * unmap_shared_mapping_pages() wants to
1090 * invalidate cache without truncating:
1091 * unmap shared but keep private pages.
1092 */
1093 if (details->check_mapping &&
1094 details->check_mapping != page->mapping)
1095 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 }
Nick Pigginb5810032005-10-29 18:16:12 -07001097 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001098 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 tlb_remove_tlb_entry(tlb, pte, addr);
1100 if (unlikely(!page))
1101 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001102
1103 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001104 if (pte_dirty(ptent)) {
Michal Hockoaac45362016-03-25 14:20:24 -07001105 /*
1106 * oom_reaper cannot tear down dirty
1107 * pages
1108 */
1109 if (unlikely(details && details->ignore_dirty))
1110 continue;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001111 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001112 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001113 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001114 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001115 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001116 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001117 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001118 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001119 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001120 if (unlikely(page_mapcount(page) < 0))
1121 print_bad_pte(vma, addr, ptent, page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001122 if (unlikely(!__tlb_remove_page(tlb, page))) {
1123 force_flush = 1;
Will Deaconce9ec372014-10-28 13:16:28 -07001124 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001125 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001126 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 continue;
1128 }
Michal Hockoaac45362016-03-25 14:20:24 -07001129 /* only check swap_entries if explicitly asked for in details */
1130 if (unlikely(details && !details->check_swap_entries))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001132
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001133 entry = pte_to_swp_entry(ptent);
1134 if (!non_swap_entry(entry))
1135 rss[MM_SWAPENTS]--;
1136 else if (is_migration_entry(entry)) {
1137 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001138
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001139 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001140 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001141 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001142 if (unlikely(!free_swap_and_cache(entry)))
1143 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001144 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001145 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001146
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001147 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001148 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001149
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001150 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001151 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001152 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001153 pte_unmap_unlock(start_pte, ptl);
1154
1155 /*
1156 * If we forced a TLB flush (either due to running out of
1157 * batch buffers or because we needed to flush dirty TLB
1158 * entries before releasing the ptl), free the batched
1159 * memory too. Restart if we didn't do everything.
1160 */
1161 if (force_flush) {
1162 force_flush = 0;
1163 tlb_flush_mmu_free(tlb);
Linus Torvalds2b047252013-08-15 11:42:25 -07001164
1165 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001166 goto again;
1167 }
1168
Robin Holt51c6f662005-11-13 16:06:42 -08001169 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001170}
1171
Robin Holt51c6f662005-11-13 16:06:42 -08001172static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001173 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001175 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176{
1177 pmd_t *pmd;
1178 unsigned long next;
1179
1180 pmd = pmd_offset(pud, addr);
1181 do {
1182 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001183 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001184 if (next - addr != HPAGE_PMD_SIZE) {
David Rientjese0897d72012-06-20 12:53:00 -07001185#ifdef CONFIG_DEBUG_VM
1186 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1187 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1188 __func__, addr, end,
1189 vma->vm_start,
1190 vma->vm_end);
1191 BUG();
1192 }
1193#endif
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001194 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001195 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001196 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001197 /* fall through */
1198 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001199 /*
1200 * Here there can be other concurrent MADV_DONTNEED or
1201 * trans huge page faults running, and if the pmd is
1202 * none or trans huge it can change under us. This is
1203 * because MADV_DONTNEED holds the mmap_sem in read
1204 * mode.
1205 */
1206 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1207 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001208 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001209next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001210 cond_resched();
1211 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001212
1213 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214}
1215
Robin Holt51c6f662005-11-13 16:06:42 -08001216static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001217 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001219 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220{
1221 pud_t *pud;
1222 unsigned long next;
1223
1224 pud = pud_offset(pgd, addr);
1225 do {
1226 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001227 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001229 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1230 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001231
1232 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233}
1234
Michal Hockoaac45362016-03-25 14:20:24 -07001235void unmap_page_range(struct mmu_gather *tlb,
Al Viro038c7aa2012-03-05 13:25:09 -05001236 struct vm_area_struct *vma,
1237 unsigned long addr, unsigned long end,
1238 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239{
1240 pgd_t *pgd;
1241 unsigned long next;
1242
Linus Torvalds1da177e2005-04-16 15:20:36 -07001243 BUG_ON(addr >= end);
1244 tlb_start_vma(tlb, vma);
1245 pgd = pgd_offset(vma->vm_mm, addr);
1246 do {
1247 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001248 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001250 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1251 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 tlb_end_vma(tlb, vma);
1253}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
Al Virof5cc4ee2012-03-05 14:14:20 -05001255
1256static void unmap_single_vma(struct mmu_gather *tlb,
1257 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001258 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001259 struct zap_details *details)
1260{
1261 unsigned long start = max(vma->vm_start, start_addr);
1262 unsigned long end;
1263
1264 if (start >= vma->vm_end)
1265 return;
1266 end = min(vma->vm_end, end_addr);
1267 if (end <= vma->vm_start)
1268 return;
1269
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301270 if (vma->vm_file)
1271 uprobe_munmap(vma, start, end);
1272
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001273 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001274 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001275
1276 if (start != end) {
1277 if (unlikely(is_vm_hugetlb_page(vma))) {
1278 /*
1279 * It is undesirable to test vma->vm_file as it
1280 * should be non-null for valid hugetlb area.
1281 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001282 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001283 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001284 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001285 * before calling this function to clean up.
1286 * Since no pte has actually been setup, it is
1287 * safe to do nothing in this case.
1288 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001289 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001290 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001291 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001292 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001293 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001294 } else
1295 unmap_page_range(tlb, vma, start, end, details);
1296 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297}
1298
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299/**
1300 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001301 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 * @vma: the starting vma
1303 * @start_addr: virtual address at which to start unmapping
1304 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001306 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 * Only addresses between `start' and `end' will be unmapped.
1309 *
1310 * The VMA list must be sorted in ascending virtual address order.
1311 *
1312 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1313 * range after unmap_vmas() returns. So the only responsibility here is to
1314 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1315 * drops the lock and schedules.
1316 */
Al Viro6e8bb012012-03-05 13:41:15 -05001317void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001319 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001321 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001323 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001324 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001325 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001326 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327}
1328
1329/**
1330 * zap_page_range - remove user pages in a given range
1331 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001332 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001333 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001334 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001335 *
1336 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001338void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 unsigned long size, struct zap_details *details)
1340{
1341 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001342 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001343 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001346 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001347 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001348 mmu_notifier_invalidate_range_start(mm, start, end);
1349 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001350 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001351 mmu_notifier_invalidate_range_end(mm, start, end);
1352 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353}
1354
Jack Steinerc627f9c2008-07-29 22:33:53 -07001355/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001356 * zap_page_range_single - remove user pages in a given range
1357 * @vma: vm_area_struct holding the applicable pages
1358 * @address: starting address of pages to zap
1359 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001360 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001361 *
1362 * The range must fit into one VMA.
1363 */
1364static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1365 unsigned long size, struct zap_details *details)
1366{
1367 struct mm_struct *mm = vma->vm_mm;
1368 struct mmu_gather tlb;
1369 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001370
1371 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001372 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001373 update_hiwater_rss(mm);
1374 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001375 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001376 mmu_notifier_invalidate_range_end(mm, address, end);
1377 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378}
1379
Jack Steinerc627f9c2008-07-29 22:33:53 -07001380/**
1381 * zap_vma_ptes - remove ptes mapping the vma
1382 * @vma: vm_area_struct holding ptes to be zapped
1383 * @address: starting address of pages to zap
1384 * @size: number of bytes to zap
1385 *
1386 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1387 *
1388 * The entire address range must be fully contained within the vma.
1389 *
1390 * Returns 0 if successful.
1391 */
1392int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1393 unsigned long size)
1394{
1395 if (address < vma->vm_start || address + size > vma->vm_end ||
1396 !(vma->vm_flags & VM_PFNMAP))
1397 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001398 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001399 return 0;
1400}
1401EXPORT_SYMBOL_GPL(zap_vma_ptes);
1402
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001403pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001404 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001405{
1406 pgd_t * pgd = pgd_offset(mm, addr);
1407 pud_t * pud = pud_alloc(mm, pgd, addr);
1408 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001409 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001410 if (pmd) {
1411 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001412 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001413 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001414 }
1415 return NULL;
1416}
1417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001419 * This is the old fallback for page remapping.
1420 *
1421 * For historical reasons, it only allows reserved pages. Only
1422 * old drivers should use this, and they needed to mark their
1423 * pages reserved for the old functions anyway.
1424 */
Nick Piggin423bad602008-04-28 02:13:01 -07001425static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1426 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001427{
Nick Piggin423bad602008-04-28 02:13:01 -07001428 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001429 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001430 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001431 spinlock_t *ptl;
1432
Linus Torvalds238f58d2005-11-29 13:01:56 -08001433 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001434 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001435 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001436 retval = -ENOMEM;
1437 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001438 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001439 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001440 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001441 retval = -EBUSY;
1442 if (!pte_none(*pte))
1443 goto out_unlock;
1444
1445 /* Ok, finally just insert the thing.. */
1446 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001447 inc_mm_counter_fast(mm, mm_counter_file(page));
Linus Torvalds238f58d2005-11-29 13:01:56 -08001448 page_add_file_rmap(page);
1449 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1450
1451 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001452 pte_unmap_unlock(pte, ptl);
1453 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001454out_unlock:
1455 pte_unmap_unlock(pte, ptl);
1456out:
1457 return retval;
1458}
1459
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001460/**
1461 * vm_insert_page - insert single page into user vma
1462 * @vma: user vma to map to
1463 * @addr: target user address of this page
1464 * @page: source kernel page
1465 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001466 * This allows drivers to insert individual pages they've allocated
1467 * into a user vma.
1468 *
1469 * The page has to be a nice clean _individual_ kernel allocation.
1470 * If you allocate a compound page, you need to have marked it as
1471 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001472 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001473 *
1474 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1475 * took an arbitrary page protection parameter. This doesn't allow
1476 * that. Your vma protection will have to be set up correctly, which
1477 * means that if you want a shared writable mapping, you'd better
1478 * ask for a shared writable mapping!
1479 *
1480 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001481 *
1482 * Usually this function is called from f_op->mmap() handler
1483 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1484 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1485 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001486 */
Nick Piggin423bad602008-04-28 02:13:01 -07001487int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1488 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001489{
1490 if (addr < vma->vm_start || addr >= vma->vm_end)
1491 return -EFAULT;
1492 if (!page_count(page))
1493 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001494 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1495 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1496 BUG_ON(vma->vm_flags & VM_PFNMAP);
1497 vma->vm_flags |= VM_MIXEDMAP;
1498 }
Nick Piggin423bad602008-04-28 02:13:01 -07001499 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001500}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001501EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001502
Nick Piggin423bad602008-04-28 02:13:01 -07001503static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001504 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad602008-04-28 02:13:01 -07001505{
1506 struct mm_struct *mm = vma->vm_mm;
1507 int retval;
1508 pte_t *pte, entry;
1509 spinlock_t *ptl;
1510
1511 retval = -ENOMEM;
1512 pte = get_locked_pte(mm, addr, &ptl);
1513 if (!pte)
1514 goto out;
1515 retval = -EBUSY;
1516 if (!pte_none(*pte))
1517 goto out_unlock;
1518
1519 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001520 if (pfn_t_devmap(pfn))
1521 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1522 else
1523 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad602008-04-28 02:13:01 -07001524 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001525 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad602008-04-28 02:13:01 -07001526
1527 retval = 0;
1528out_unlock:
1529 pte_unmap_unlock(pte, ptl);
1530out:
1531 return retval;
1532}
1533
Nick Piggine0dc0d82007-02-12 00:51:36 -08001534/**
1535 * vm_insert_pfn - insert single pfn into user vma
1536 * @vma: user vma to map to
1537 * @addr: target user address of this page
1538 * @pfn: source kernel pfn
1539 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001540 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001541 * they've allocated into a user vma. Same comments apply.
1542 *
1543 * This function should only be called from a vm_ops->fault handler, and
1544 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001545 *
1546 * vma cannot be a COW mapping.
1547 *
1548 * As this is called only for pages that do not currently exist, we
1549 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001550 */
1551int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad602008-04-28 02:13:01 -07001552 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001553{
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001554 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1555}
1556EXPORT_SYMBOL(vm_insert_pfn);
1557
1558/**
1559 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1560 * @vma: user vma to map to
1561 * @addr: target user address of this page
1562 * @pfn: source kernel pfn
1563 * @pgprot: pgprot flags for the inserted page
1564 *
1565 * This is exactly like vm_insert_pfn, except that it allows drivers to
1566 * to override pgprot on a per-page basis.
1567 *
1568 * This only makes sense for IO mappings, and it makes no sense for
1569 * cow mappings. In general, using multiple vmas is preferable;
1570 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1571 * impractical.
1572 */
1573int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1574 unsigned long pfn, pgprot_t pgprot)
1575{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001576 int ret;
Nick Piggin7e675132008-04-28 02:13:00 -07001577 /*
1578 * Technically, architectures with pte_special can avoid all these
1579 * restrictions (same for remap_pfn_range). However we would like
1580 * consistency in testing and feature parity among all, so we should
1581 * try to keep these invariants in place for everybody.
1582 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001583 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1584 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1585 (VM_PFNMAP|VM_MIXEDMAP));
1586 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1587 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001588
Nick Piggin423bad602008-04-28 02:13:01 -07001589 if (addr < vma->vm_start || addr >= vma->vm_end)
1590 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001591 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001592 return -EINVAL;
1593
Dan Williams01c8f1c2016-01-15 16:56:40 -08001594 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001595
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001596 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001597}
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001598EXPORT_SYMBOL(vm_insert_pfn_prot);
Nick Piggine0dc0d82007-02-12 00:51:36 -08001599
Nick Piggin423bad602008-04-28 02:13:01 -07001600int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001601 pfn_t pfn)
Nick Piggin423bad602008-04-28 02:13:01 -07001602{
1603 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1604
1605 if (addr < vma->vm_start || addr >= vma->vm_end)
1606 return -EFAULT;
1607
1608 /*
1609 * If we don't have pte special, then we have to use the pfn_valid()
1610 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1611 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001612 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1613 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad602008-04-28 02:13:01 -07001614 */
Dan Williams03fc2da2016-01-26 09:48:05 -08001615 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
Nick Piggin423bad602008-04-28 02:13:01 -07001616 struct page *page;
1617
Dan Williams03fc2da2016-01-26 09:48:05 -08001618 /*
1619 * At this point we are committed to insert_page()
1620 * regardless of whether the caller specified flags that
1621 * result in pfn_t_has_page() == false.
1622 */
1623 page = pfn_to_page(pfn_t_to_pfn(pfn));
Nick Piggin423bad602008-04-28 02:13:01 -07001624 return insert_page(vma, addr, page, vma->vm_page_prot);
1625 }
1626 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1627}
1628EXPORT_SYMBOL(vm_insert_mixed);
1629
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001630/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 * maps a range of physical memory into the requested pages. the old
1632 * mappings are removed. any references to nonexistent pages results
1633 * in null mappings (currently treated as "copy-on-access")
1634 */
1635static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1636 unsigned long addr, unsigned long end,
1637 unsigned long pfn, pgprot_t prot)
1638{
1639 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001640 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641
Hugh Dickinsc74df322005-10-29 18:16:23 -07001642 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 if (!pte)
1644 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001645 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 do {
1647 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001648 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 pfn++;
1650 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001651 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001652 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 return 0;
1654}
1655
1656static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1657 unsigned long addr, unsigned long end,
1658 unsigned long pfn, pgprot_t prot)
1659{
1660 pmd_t *pmd;
1661 unsigned long next;
1662
1663 pfn -= addr >> PAGE_SHIFT;
1664 pmd = pmd_alloc(mm, pud, addr);
1665 if (!pmd)
1666 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001667 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668 do {
1669 next = pmd_addr_end(addr, end);
1670 if (remap_pte_range(mm, pmd, addr, next,
1671 pfn + (addr >> PAGE_SHIFT), prot))
1672 return -ENOMEM;
1673 } while (pmd++, addr = next, addr != end);
1674 return 0;
1675}
1676
1677static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1678 unsigned long addr, unsigned long end,
1679 unsigned long pfn, pgprot_t prot)
1680{
1681 pud_t *pud;
1682 unsigned long next;
1683
1684 pfn -= addr >> PAGE_SHIFT;
1685 pud = pud_alloc(mm, pgd, addr);
1686 if (!pud)
1687 return -ENOMEM;
1688 do {
1689 next = pud_addr_end(addr, end);
1690 if (remap_pmd_range(mm, pud, addr, next,
1691 pfn + (addr >> PAGE_SHIFT), prot))
1692 return -ENOMEM;
1693 } while (pud++, addr = next, addr != end);
1694 return 0;
1695}
1696
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001697/**
1698 * remap_pfn_range - remap kernel memory to userspace
1699 * @vma: user vma to map to
1700 * @addr: target user address to start at
1701 * @pfn: physical address of kernel memory
1702 * @size: size of map area
1703 * @prot: page protection flags for this mapping
1704 *
1705 * Note: this is only safe if the mm semaphore is held when called.
1706 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1708 unsigned long pfn, unsigned long size, pgprot_t prot)
1709{
1710 pgd_t *pgd;
1711 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001712 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 struct mm_struct *mm = vma->vm_mm;
1714 int err;
1715
1716 /*
1717 * Physically remapped pages are special. Tell the
1718 * rest of the world about it:
1719 * VM_IO tells people not to look at these pages
1720 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001721 * VM_PFNMAP tells the core MM that the base pages are just
1722 * raw PFN mappings, and do not have a "struct page" associated
1723 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001724 * VM_DONTEXPAND
1725 * Disable vma merging and expanding with mremap().
1726 * VM_DONTDUMP
1727 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001728 *
1729 * There's a horrible special case to handle copy-on-write
1730 * behaviour that some programs depend on. We mark the "original"
1731 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001732 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001734 if (is_cow_mapping(vma->vm_flags)) {
1735 if (addr != vma->vm_start || end != vma->vm_end)
1736 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001737 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001738 }
1739
1740 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1741 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001742 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001743
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001744 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745
1746 BUG_ON(addr >= end);
1747 pfn -= addr >> PAGE_SHIFT;
1748 pgd = pgd_offset(mm, addr);
1749 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 do {
1751 next = pgd_addr_end(addr, end);
1752 err = remap_pud_range(mm, pgd, addr, next,
1753 pfn + (addr >> PAGE_SHIFT), prot);
1754 if (err)
1755 break;
1756 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001757
1758 if (err)
Suresh Siddha5180da42012-10-08 16:28:29 -07001759 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001760
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 return err;
1762}
1763EXPORT_SYMBOL(remap_pfn_range);
1764
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001765/**
1766 * vm_iomap_memory - remap memory to userspace
1767 * @vma: user vma to map to
1768 * @start: start of area
1769 * @len: size of area
1770 *
1771 * This is a simplified io_remap_pfn_range() for common driver use. The
1772 * driver just needs to give us the physical memory range to be mapped,
1773 * we'll figure out the rest from the vma information.
1774 *
1775 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1776 * whatever write-combining details or similar.
1777 */
1778int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1779{
1780 unsigned long vm_len, pfn, pages;
1781
1782 /* Check that the physical memory area passed in looks valid */
1783 if (start + len < start)
1784 return -EINVAL;
1785 /*
1786 * You *really* shouldn't map things that aren't page-aligned,
1787 * but we've historically allowed it because IO memory might
1788 * just have smaller alignment.
1789 */
1790 len += start & ~PAGE_MASK;
1791 pfn = start >> PAGE_SHIFT;
1792 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1793 if (pfn + pages < pfn)
1794 return -EINVAL;
1795
1796 /* We start the mapping 'vm_pgoff' pages into the area */
1797 if (vma->vm_pgoff > pages)
1798 return -EINVAL;
1799 pfn += vma->vm_pgoff;
1800 pages -= vma->vm_pgoff;
1801
1802 /* Can we fit all of the mapping? */
1803 vm_len = vma->vm_end - vma->vm_start;
1804 if (vm_len >> PAGE_SHIFT > pages)
1805 return -EINVAL;
1806
1807 /* Ok, let it rip */
1808 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1809}
1810EXPORT_SYMBOL(vm_iomap_memory);
1811
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001812static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1813 unsigned long addr, unsigned long end,
1814 pte_fn_t fn, void *data)
1815{
1816 pte_t *pte;
1817 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001818 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001819 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001820
1821 pte = (mm == &init_mm) ?
1822 pte_alloc_kernel(pmd, addr) :
1823 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1824 if (!pte)
1825 return -ENOMEM;
1826
1827 BUG_ON(pmd_huge(*pmd));
1828
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001829 arch_enter_lazy_mmu_mode();
1830
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001831 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001832
1833 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001834 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001835 if (err)
1836 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001837 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001838
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001839 arch_leave_lazy_mmu_mode();
1840
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001841 if (mm != &init_mm)
1842 pte_unmap_unlock(pte-1, ptl);
1843 return err;
1844}
1845
1846static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1847 unsigned long addr, unsigned long end,
1848 pte_fn_t fn, void *data)
1849{
1850 pmd_t *pmd;
1851 unsigned long next;
1852 int err;
1853
Andi Kleenceb86872008-07-23 21:27:50 -07001854 BUG_ON(pud_huge(*pud));
1855
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001856 pmd = pmd_alloc(mm, pud, addr);
1857 if (!pmd)
1858 return -ENOMEM;
1859 do {
1860 next = pmd_addr_end(addr, end);
1861 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1862 if (err)
1863 break;
1864 } while (pmd++, addr = next, addr != end);
1865 return err;
1866}
1867
1868static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1869 unsigned long addr, unsigned long end,
1870 pte_fn_t fn, void *data)
1871{
1872 pud_t *pud;
1873 unsigned long next;
1874 int err;
1875
1876 pud = pud_alloc(mm, pgd, addr);
1877 if (!pud)
1878 return -ENOMEM;
1879 do {
1880 next = pud_addr_end(addr, end);
1881 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1882 if (err)
1883 break;
1884 } while (pud++, addr = next, addr != end);
1885 return err;
1886}
1887
1888/*
1889 * Scan a region of virtual memory, filling in page tables as necessary
1890 * and calling a provided function on each leaf page table.
1891 */
1892int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1893 unsigned long size, pte_fn_t fn, void *data)
1894{
1895 pgd_t *pgd;
1896 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001897 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001898 int err;
1899
Mika Penttilä9cb65bc2016-03-15 14:56:45 -07001900 if (WARN_ON(addr >= end))
1901 return -EINVAL;
1902
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001903 pgd = pgd_offset(mm, addr);
1904 do {
1905 next = pgd_addr_end(addr, end);
1906 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1907 if (err)
1908 break;
1909 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001910
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001911 return err;
1912}
1913EXPORT_SYMBOL_GPL(apply_to_page_range);
1914
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001916 * handle_pte_fault chooses page fault handler according to an entry which was
1917 * read non-atomically. Before making any commitment, on those architectures
1918 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1919 * parts, do_swap_page must check under lock before unmapping the pte and
1920 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001921 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001922 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001923static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001924 pte_t *page_table, pte_t orig_pte)
1925{
1926 int same = 1;
1927#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1928 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001929 spinlock_t *ptl = pte_lockptr(mm, pmd);
1930 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001931 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001932 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001933 }
1934#endif
1935 pte_unmap(page_table);
1936 return same;
1937}
1938
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001939static 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 -08001940{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001941 debug_dma_assert_idle(src);
1942
Linus Torvalds6aab3412005-11-28 14:34:23 -08001943 /*
1944 * If the source page was a PFN mapping, we don't have
1945 * a "struct page" for it. We do a best-effort copy by
1946 * just copying from the original user address. If that
1947 * fails, we just zero-fill it. Live with it.
1948 */
1949 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001950 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08001951 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1952
1953 /*
1954 * This really shouldn't fail, because the page is there
1955 * in the page tables. But it might just be unreadable,
1956 * in which case we just give up and fill the result with
1957 * zeroes.
1958 */
1959 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07001960 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08001961 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07001962 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08001963 } else
1964 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08001965}
1966
Michal Hockoc20cd452016-01-14 15:20:12 -08001967static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
1968{
1969 struct file *vm_file = vma->vm_file;
1970
1971 if (vm_file)
1972 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
1973
1974 /*
1975 * Special mappings (e.g. VDSO) do not have any file so fake
1976 * a default GFP_KERNEL for them.
1977 */
1978 return GFP_KERNEL;
1979}
1980
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001982 * Notify the address space that the page is about to become writable so that
1983 * it can prohibit this or wait for the page to get into an appropriate state.
1984 *
1985 * We do this without the lock held, so that it can sleep if it needs to.
1986 */
1987static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
1988 unsigned long address)
1989{
1990 struct vm_fault vmf;
1991 int ret;
1992
1993 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
1994 vmf.pgoff = page->index;
1995 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08001996 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001997 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08001998 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07001999
2000 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2001 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2002 return ret;
2003 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2004 lock_page(page);
2005 if (!page->mapping) {
2006 unlock_page(page);
2007 return 0; /* retry */
2008 }
2009 ret |= VM_FAULT_LOCKED;
2010 } else
2011 VM_BUG_ON_PAGE(!PageLocked(page), page);
2012 return ret;
2013}
2014
2015/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07002016 * Handle write page faults for pages that can be reused in the current vma
2017 *
2018 * This can happen either due to the mapping being with the VM_SHARED flag,
2019 * or due to us being the last reference standing to the page. In either
2020 * case, all we need to do here is to mark the page as writable and update
2021 * any related book-keeping.
2022 */
2023static inline int wp_page_reuse(struct mm_struct *mm,
2024 struct vm_area_struct *vma, unsigned long address,
2025 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2026 struct page *page, int page_mkwrite,
2027 int dirty_shared)
2028 __releases(ptl)
2029{
2030 pte_t entry;
2031 /*
2032 * Clear the pages cpupid information as the existing
2033 * information potentially belongs to a now completely
2034 * unrelated process.
2035 */
2036 if (page)
2037 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2038
2039 flush_cache_page(vma, address, pte_pfn(orig_pte));
2040 entry = pte_mkyoung(orig_pte);
2041 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2042 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2043 update_mmu_cache(vma, address, page_table);
2044 pte_unmap_unlock(page_table, ptl);
2045
2046 if (dirty_shared) {
2047 struct address_space *mapping;
2048 int dirtied;
2049
2050 if (!page_mkwrite)
2051 lock_page(page);
2052
2053 dirtied = set_page_dirty(page);
2054 VM_BUG_ON_PAGE(PageAnon(page), page);
2055 mapping = page->mapping;
2056 unlock_page(page);
2057 page_cache_release(page);
2058
2059 if ((dirtied || page_mkwrite) && mapping) {
2060 /*
2061 * Some device drivers do not set page.mapping
2062 * but still dirty their pages
2063 */
2064 balance_dirty_pages_ratelimited(mapping);
2065 }
2066
2067 if (!page_mkwrite)
2068 file_update_time(vma->vm_file);
2069 }
2070
2071 return VM_FAULT_WRITE;
2072}
2073
2074/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002075 * Handle the case of a page which we actually need to copy to a new page.
2076 *
2077 * Called with mmap_sem locked and the old page referenced, but
2078 * without the ptl held.
2079 *
2080 * High level logic flow:
2081 *
2082 * - Allocate a page, copy the content of the old page to the new one.
2083 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2084 * - Take the PTL. If the pte changed, bail out and release the allocated page
2085 * - If the pte is still the way we remember it, update the page table and all
2086 * relevant references. This includes dropping the reference the page-table
2087 * held to the old page, as well as updating the rmap.
2088 * - In any case, unlock the PTL and drop the reference we took to the old page.
2089 */
2090static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2091 unsigned long address, pte_t *page_table, pmd_t *pmd,
2092 pte_t orig_pte, struct page *old_page)
2093{
2094 struct page *new_page = NULL;
2095 spinlock_t *ptl = NULL;
2096 pte_t entry;
2097 int page_copied = 0;
2098 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2099 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2100 struct mem_cgroup *memcg;
2101
2102 if (unlikely(anon_vma_prepare(vma)))
2103 goto oom;
2104
2105 if (is_zero_pfn(pte_pfn(orig_pte))) {
2106 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2107 if (!new_page)
2108 goto oom;
2109 } else {
2110 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2111 if (!new_page)
2112 goto oom;
2113 cow_user_page(new_page, old_page, address, vma);
2114 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002115
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002116 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002117 goto oom_free_new;
2118
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002119 __SetPageUptodate(new_page);
2120
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002121 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2122
2123 /*
2124 * Re-check the pte - we dropped the lock
2125 */
2126 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2127 if (likely(pte_same(*page_table, orig_pte))) {
2128 if (old_page) {
2129 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002130 dec_mm_counter_fast(mm,
2131 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002132 inc_mm_counter_fast(mm, MM_ANONPAGES);
2133 }
2134 } else {
2135 inc_mm_counter_fast(mm, MM_ANONPAGES);
2136 }
2137 flush_cache_page(vma, address, pte_pfn(orig_pte));
2138 entry = mk_pte(new_page, vma->vm_page_prot);
2139 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2140 /*
2141 * Clear the pte entry and flush it first, before updating the
2142 * pte with the new entry. This will avoid a race condition
2143 * seen in the presence of one thread doing SMC and another
2144 * thread doing COW.
2145 */
2146 ptep_clear_flush_notify(vma, address, page_table);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002147 page_add_new_anon_rmap(new_page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002148 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002149 lru_cache_add_active_or_unevictable(new_page, vma);
2150 /*
2151 * We call the notify macro here because, when using secondary
2152 * mmu page tables (such as kvm shadow page tables), we want the
2153 * new page to be mapped directly into the secondary page table.
2154 */
2155 set_pte_at_notify(mm, address, page_table, entry);
2156 update_mmu_cache(vma, address, page_table);
2157 if (old_page) {
2158 /*
2159 * Only after switching the pte to the new page may
2160 * we remove the mapcount here. Otherwise another
2161 * process may come and find the rmap count decremented
2162 * before the pte is switched to the new page, and
2163 * "reuse" the old page writing into it while our pte
2164 * here still points into it and can be read by other
2165 * threads.
2166 *
2167 * The critical issue is to order this
2168 * page_remove_rmap with the ptp_clear_flush above.
2169 * Those stores are ordered by (if nothing else,)
2170 * the barrier present in the atomic_add_negative
2171 * in page_remove_rmap.
2172 *
2173 * Then the TLB flush in ptep_clear_flush ensures that
2174 * no process can access the old page before the
2175 * decremented mapcount is visible. And the old page
2176 * cannot be reused until after the decremented
2177 * mapcount is visible. So transitively, TLBs to
2178 * old page will be flushed before it can be reused.
2179 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002180 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002181 }
2182
2183 /* Free the old page.. */
2184 new_page = old_page;
2185 page_copied = 1;
2186 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002187 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002188 }
2189
2190 if (new_page)
2191 page_cache_release(new_page);
2192
2193 pte_unmap_unlock(page_table, ptl);
2194 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2195 if (old_page) {
2196 /*
2197 * Don't let another task, with possibly unlocked vma,
2198 * keep the mlocked page.
2199 */
2200 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2201 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002202 if (PageMlocked(old_page))
2203 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002204 unlock_page(old_page);
2205 }
2206 page_cache_release(old_page);
2207 }
2208 return page_copied ? VM_FAULT_WRITE : 0;
2209oom_free_new:
2210 page_cache_release(new_page);
2211oom:
2212 if (old_page)
2213 page_cache_release(old_page);
2214 return VM_FAULT_OOM;
2215}
2216
Boaz Harroshdd906182015-04-15 16:15:11 -07002217/*
2218 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2219 * mapping
2220 */
2221static int wp_pfn_shared(struct mm_struct *mm,
2222 struct vm_area_struct *vma, unsigned long address,
2223 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2224 pmd_t *pmd)
2225{
2226 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2227 struct vm_fault vmf = {
2228 .page = NULL,
2229 .pgoff = linear_page_index(vma, address),
2230 .virtual_address = (void __user *)(address & PAGE_MASK),
2231 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2232 };
2233 int ret;
2234
2235 pte_unmap_unlock(page_table, ptl);
2236 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2237 if (ret & VM_FAULT_ERROR)
2238 return ret;
2239 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2240 /*
2241 * We might have raced with another page fault while we
2242 * released the pte_offset_map_lock.
2243 */
2244 if (!pte_same(*page_table, orig_pte)) {
2245 pte_unmap_unlock(page_table, ptl);
2246 return 0;
2247 }
2248 }
2249 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2250 NULL, 0, 0);
2251}
2252
Shachar Raindel93e478d2015-04-14 15:46:35 -07002253static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2254 unsigned long address, pte_t *page_table,
2255 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2256 struct page *old_page)
2257 __releases(ptl)
2258{
2259 int page_mkwrite = 0;
2260
2261 page_cache_get(old_page);
2262
Shachar Raindel93e478d2015-04-14 15:46:35 -07002263 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2264 int tmp;
2265
2266 pte_unmap_unlock(page_table, ptl);
2267 tmp = do_page_mkwrite(vma, old_page, address);
2268 if (unlikely(!tmp || (tmp &
2269 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2270 page_cache_release(old_page);
2271 return tmp;
2272 }
2273 /*
2274 * Since we dropped the lock we need to revalidate
2275 * the PTE as someone else may have changed it. If
2276 * they did, we just return, as we can count on the
2277 * MMU to tell us if they didn't also make it writable.
2278 */
2279 page_table = pte_offset_map_lock(mm, pmd, address,
2280 &ptl);
2281 if (!pte_same(*page_table, orig_pte)) {
2282 unlock_page(old_page);
2283 pte_unmap_unlock(page_table, ptl);
2284 page_cache_release(old_page);
2285 return 0;
2286 }
2287 page_mkwrite = 1;
2288 }
2289
2290 return wp_page_reuse(mm, vma, address, page_table, ptl,
2291 orig_pte, old_page, page_mkwrite, 1);
2292}
2293
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002294/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 * This routine handles present pages, when users try to write
2296 * to a shared page. It is done by copying the page to a new address
2297 * and decrementing the shared-page counter for the old page.
2298 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 * Note that this routine assumes that the protection checks have been
2300 * done by the caller (the low-level page fault routine in most cases).
2301 * Thus we can safely just mark it writable once we've done any necessary
2302 * COW.
2303 *
2304 * We also mark the page dirty at this point even though the page will
2305 * change only once the write actually happens. This avoids a few races,
2306 * and potentially makes it more efficient.
2307 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002308 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2309 * but allow concurrent faults), with pte both mapped and locked.
2310 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002312static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2313 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002314 spinlock_t *ptl, pte_t orig_pte)
Namhyung Kime6219ec2010-10-26 14:22:00 -07002315 __releases(ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316{
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002317 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
Linus Torvalds6aab3412005-11-28 14:34:23 -08002319 old_page = vm_normal_page(vma, address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002320 if (!old_page) {
2321 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002322 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2323 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002324 *
2325 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002326 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002327 */
2328 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2329 (VM_WRITE|VM_SHARED))
Boaz Harroshdd906182015-04-15 16:15:11 -07002330 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2331 orig_pte, pmd);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002332
2333 pte_unmap_unlock(page_table, ptl);
2334 return wp_page_copy(mm, vma, address, page_table, pmd,
2335 orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002336 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002338 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002339 * Take out anonymous pages first, anonymous shared vmas are
2340 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002341 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002342 if (PageAnon(old_page) && !PageKsm(old_page)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002343 if (!trylock_page(old_page)) {
2344 page_cache_get(old_page);
2345 pte_unmap_unlock(page_table, ptl);
2346 lock_page(old_page);
2347 page_table = pte_offset_map_lock(mm, pmd, address,
2348 &ptl);
2349 if (!pte_same(*page_table, orig_pte)) {
2350 unlock_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002351 pte_unmap_unlock(page_table, ptl);
2352 page_cache_release(old_page);
2353 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002354 }
2355 page_cache_release(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002356 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002357 if (reuse_swap_page(old_page)) {
Rik van Rielc44b6742010-03-05 13:42:09 -08002358 /*
2359 * The page is all ours. Move it to our anon_vma so
2360 * the rmap code will not search our parent or siblings.
2361 * Protected against the rmap code by the page lock.
2362 */
2363 page_move_anon_rmap(old_page, vma, address);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002364 unlock_page(old_page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002365 return wp_page_reuse(mm, vma, address, page_table, ptl,
2366 orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002367 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002368 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002369 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002370 (VM_WRITE|VM_SHARED))) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002371 return wp_page_shared(mm, vma, address, page_table, pmd,
2372 ptl, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374
2375 /*
2376 * Ok, we need to copy. Oh, well..
2377 */
Nick Pigginb5810032005-10-29 18:16:12 -07002378 page_cache_get(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002379
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002380 pte_unmap_unlock(page_table, ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002381 return wp_page_copy(mm, vma, address, page_table, pmd,
2382 orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383}
2384
Peter Zijlstra97a89412011-05-24 17:12:04 -07002385static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 unsigned long start_addr, unsigned long end_addr,
2387 struct zap_details *details)
2388{
Al Virof5cc4ee2012-03-05 14:14:20 -05002389 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390}
2391
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002392static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 struct zap_details *details)
2394{
2395 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 pgoff_t vba, vea, zba, zea;
2397
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002398 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400
2401 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002402 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2404 zba = details->first_index;
2405 if (zba < vba)
2406 zba = vba;
2407 zea = details->last_index;
2408 if (zea > vea)
2409 zea = vea;
2410
Peter Zijlstra97a89412011-05-24 17:12:04 -07002411 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2413 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002414 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 }
2416}
2417
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002419 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2420 * address_space corresponding to the specified page range in the underlying
2421 * file.
2422 *
Martin Waitz3d410882005-06-23 22:05:21 -07002423 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 * @holebegin: byte in first page to unmap, relative to the start of
2425 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002426 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 * must keep the partial page. In contrast, we must get rid of
2428 * partial pages.
2429 * @holelen: size of prospective hole in bytes. This will be rounded
2430 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2431 * end of the file.
2432 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2433 * but 0 when invalidating pagecache, don't throw away private data.
2434 */
2435void unmap_mapping_range(struct address_space *mapping,
2436 loff_t const holebegin, loff_t const holelen, int even_cows)
2437{
Michal Hockoaac45362016-03-25 14:20:24 -07002438 struct zap_details details = { };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439 pgoff_t hba = holebegin >> PAGE_SHIFT;
2440 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2441
2442 /* Check for overflow. */
2443 if (sizeof(holelen) > sizeof(hlen)) {
2444 long long holeend =
2445 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2446 if (holeend & ~(long long)ULONG_MAX)
2447 hlen = ULONG_MAX - hba + 1;
2448 }
2449
2450 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 details.first_index = hba;
2452 details.last_index = hba + hlen - 1;
2453 if (details.last_index < details.first_index)
2454 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Ross Zwisler0f90cc62015-10-15 15:28:32 -07002456
2457 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002458 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002459 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002461 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462}
2463EXPORT_SYMBOL(unmap_mapping_range);
2464
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002466 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2467 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002468 * We return with pte unmapped and unlocked.
2469 *
2470 * We return with the mmap_sem locked or unlocked in the same cases
2471 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002473static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2474 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002475 unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002477 spinlock_t *ptl;
Hugh Dickins56f31802013-02-22 16:36:10 -08002478 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002479 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002480 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002482 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002483 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002484 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002486 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002487 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002488
2489 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002490 if (unlikely(non_swap_entry(entry))) {
2491 if (is_migration_entry(entry)) {
2492 migration_entry_wait(mm, pmd, address);
2493 } else if (is_hwpoison_entry(entry)) {
2494 ret = VM_FAULT_HWPOISON;
2495 } else {
2496 print_bad_pte(vma, address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002497 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002498 }
Christoph Lameter06972122006-06-23 02:03:35 -07002499 goto out;
2500 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002501 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 page = lookup_swap_cache(entry);
2503 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002504 page = swapin_readahead(entry,
2505 GFP_HIGHUSER_MOVABLE, vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506 if (!page) {
2507 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002508 * Back out if somebody else faulted in this pte
2509 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002511 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512 if (likely(pte_same(*page_table, orig_pte)))
2513 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002514 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002515 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516 }
2517
2518 /* Had to read the page from swap area: Major fault */
2519 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002520 count_vm_event(PGMAJFAULT);
Ying Han456f9982011-05-26 16:25:38 -07002521 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002522 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002523 /*
2524 * hwpoisoned dirty swapcache pages are kept for killing
2525 * owner processes (which may be unknown at hwpoison time)
2526 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002527 ret = VM_FAULT_HWPOISON;
2528 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002529 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002530 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 }
2532
Hugh Dickins56f31802013-02-22 16:36:10 -08002533 swapcache = page;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002534 locked = lock_page_or_retry(page, mm, flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002535
Balbir Singh20a10222007-11-14 17:00:33 -08002536 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002537 if (!locked) {
2538 ret |= VM_FAULT_RETRY;
2539 goto out_release;
2540 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002542 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002543 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2544 * release the swapcache from under us. The page pin, and pte_same
2545 * test below, are not enough to exclude that. Even if it is still
2546 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002547 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002548 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002549 goto out_page;
2550
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002551 page = ksm_might_need_to_copy(page, vma, address);
2552 if (unlikely(!page)) {
2553 ret = VM_FAULT_OOM;
2554 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002555 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002556 }
2557
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002558 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002559 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002560 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002561 }
2562
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002564 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002566 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07002567 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002568 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002569
2570 if (unlikely(!PageUptodate(page))) {
2571 ret = VM_FAULT_SIGBUS;
2572 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 }
2574
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002575 /*
2576 * The page isn't present yet, go ahead with the fault.
2577 *
2578 * Be careful about the sequence of operations here.
2579 * To get its accounting right, reuse_swap_page() must be called
2580 * while the page is counted on swap but not yet in mapcount i.e.
2581 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2582 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002583 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002585 inc_mm_counter_fast(mm, MM_ANONPAGES);
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08002586 dec_mm_counter_fast(mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 pte = mk_pte(page, vma->vm_page_prot);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002588 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002590 flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002591 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002592 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002595 if (pte_swp_soft_dirty(orig_pte))
2596 pte = pte_mksoft_dirty(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 set_pte_at(mm, address, page_table, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002598 if (page == swapcache) {
Johannes Weineraf347702013-02-22 16:32:20 -08002599 do_page_add_anon_rmap(page, vma, address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002600 mem_cgroup_commit_charge(page, memcg, true, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002601 } else { /* ksm created a completely new copy */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002602 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002603 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002604 lru_cache_add_active_or_unevictable(page, vma);
2605 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002607 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002608 if (mem_cgroup_swap_full(page) ||
2609 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002610 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002611 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002612 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002613 /*
2614 * Hold the lock to avoid the swap entry to be reused
2615 * until we take the PT lock for the pte_same() check
2616 * (to avoid false positives from pte_same). For
2617 * further safety release the lock after the swap_free
2618 * so that the swap count won't change under a
2619 * parallel locked swapcache.
2620 */
2621 unlock_page(swapcache);
2622 page_cache_release(swapcache);
2623 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002624
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002625 if (flags & FAULT_FLAG_WRITE) {
Hugh Dickins61469f12008-03-04 14:29:04 -08002626 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2627 if (ret & VM_FAULT_ERROR)
2628 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 goto out;
2630 }
2631
2632 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002633 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002634unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002635 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636out:
2637 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002638out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002639 mem_cgroup_cancel_charge(page, memcg, false);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002640 pte_unmap_unlock(page_table, ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002641out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002642 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002643out_release:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002644 page_cache_release(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002645 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002646 unlock_page(swapcache);
2647 page_cache_release(swapcache);
2648 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002649 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650}
2651
2652/*
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002653 * This is like a special single-page "expand_{down|up}wards()",
2654 * except we must first make sure that 'address{-|+}PAGE_SIZE'
Linus Torvalds320b2b82010-08-12 17:54:33 -07002655 * doesn't hit another vma.
Linus Torvalds320b2b82010-08-12 17:54:33 -07002656 */
2657static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2658{
2659 address &= PAGE_MASK;
2660 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002661 struct vm_area_struct *prev = vma->vm_prev;
Linus Torvalds320b2b82010-08-12 17:54:33 -07002662
Linus Torvalds0e8e50e2010-08-20 16:49:40 -07002663 /*
2664 * Is there a mapping abutting this one below?
2665 *
2666 * That's only ok if it's the same stack mapping
2667 * that has gotten split..
2668 */
2669 if (prev && prev->vm_end == address)
2670 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2671
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002672 return expand_downwards(vma, address - PAGE_SIZE);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002673 }
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002674 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2675 struct vm_area_struct *next = vma->vm_next;
2676
2677 /* As VM_GROWSDOWN but s/below/above/ */
2678 if (next && next->vm_start == address + PAGE_SIZE)
2679 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2680
Linus Torvaldsfee7e492015-01-06 13:00:05 -08002681 return expand_upwards(vma, address + PAGE_SIZE);
Luck, Tony8ca3eb02010-08-24 11:44:18 -07002682 }
Linus Torvalds320b2b82010-08-12 17:54:33 -07002683 return 0;
2684}
2685
2686/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002687 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2688 * but allow concurrent faults), and pte mapped but not yet locked.
2689 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002691static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2692 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07002693 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694{
Johannes Weiner00501b52014-08-08 14:19:20 -07002695 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002696 struct page *page;
2697 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699
Linus Torvalds11ac5522010-08-14 11:44:56 -07002700 pte_unmap(page_table);
Linus Torvalds320b2b82010-08-12 17:54:33 -07002701
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002702 /* File mapping without ->vm_ops ? */
2703 if (vma->vm_flags & VM_SHARED)
2704 return VM_FAULT_SIGBUS;
2705
Linus Torvalds11ac5522010-08-14 11:44:56 -07002706 /* Check if we need to add a guard page to the stack */
2707 if (check_stack_guard_page(vma, address) < 0)
Linus Torvalds9c145c562015-01-29 11:15:17 -08002708 return VM_FAULT_SIGSEGV;
Linus Torvalds11ac5522010-08-14 11:44:56 -07002709
2710 /* Use the zero-page for reads */
Dominik Dingel593befa2014-10-23 12:07:44 +02002711 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
Hugh Dickins62eede62009-09-21 17:03:34 -07002712 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2713 vma->vm_page_prot));
Linus Torvalds11ac5522010-08-14 11:44:56 -07002714 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002715 if (!pte_none(*page_table))
2716 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002717 /* Deliver the page fault to userland, check inside PT lock */
2718 if (userfaultfd_missing(vma)) {
2719 pte_unmap_unlock(page_table, ptl);
2720 return handle_userfault(vma, address, flags,
2721 VM_UFFD_MISSING);
2722 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002723 goto setpte;
2724 }
2725
Nick Piggin557ed1f2007-10-16 01:24:40 -07002726 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002727 if (unlikely(anon_vma_prepare(vma)))
2728 goto oom;
2729 page = alloc_zeroed_user_highpage_movable(vma, address);
2730 if (!page)
2731 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002732
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002733 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002734 goto oom_free_page;
2735
Minchan Kim52f37622013-04-29 15:08:15 -07002736 /*
2737 * The memory barrier inside __SetPageUptodate makes sure that
2738 * preceeding stores to the page contents become visible before
2739 * the set_pte_at() write.
2740 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002741 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742
Nick Piggin557ed1f2007-10-16 01:24:40 -07002743 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002744 if (vma->vm_flags & VM_WRITE)
2745 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002746
Nick Piggin557ed1f2007-10-16 01:24:40 -07002747 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Andrea Arcangeli1c2fb7a2009-09-21 17:02:22 -07002748 if (!pte_none(*page_table))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002749 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002750
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002751 /* Deliver the page fault to userland, check inside PT lock */
2752 if (userfaultfd_missing(vma)) {
2753 pte_unmap_unlock(page_table, ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002754 mem_cgroup_cancel_charge(page, memcg, false);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002755 page_cache_release(page);
2756 return handle_userfault(vma, address, flags,
2757 VM_UFFD_MISSING);
2758 }
2759
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -08002760 inc_mm_counter_fast(mm, MM_ANONPAGES);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002761 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002762 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002763 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002764setpte:
Hugh Dickins65500d22005-10-29 18:15:59 -07002765 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766
2767 /* No need to invalidate - it was non-present before */
Russell King4b3073e2009-12-18 16:40:18 +00002768 update_mmu_cache(vma, address, page_table);
Hugh Dickins65500d22005-10-29 18:15:59 -07002769unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002770 pte_unmap_unlock(page_table, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002771 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002772release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002773 mem_cgroup_cancel_charge(page, memcg, false);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002774 page_cache_release(page);
2775 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002776oom_free_page:
Hugh Dickins6dbf6d32008-03-04 14:29:04 -08002777 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002778oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 return VM_FAULT_OOM;
2780}
2781
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002782/*
2783 * The mmap_sem must have been held on entry, and may have been
2784 * released depending on flags and vma->vm_ops->fault() return value.
2785 * See filemap_fault() and __lock_page_retry().
2786 */
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002787static int __do_fault(struct vm_area_struct *vma, unsigned long address,
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002788 pgoff_t pgoff, unsigned int flags,
2789 struct page *cow_page, struct page **page)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002790{
2791 struct vm_fault vmf;
2792 int ret;
2793
2794 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2795 vmf.pgoff = pgoff;
2796 vmf.flags = flags;
2797 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002798 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002799 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002800
2801 ret = vma->vm_ops->fault(vma, &vmf);
2802 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2803 return ret;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002804 if (!vmf.page)
2805 goto out;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002806
2807 if (unlikely(PageHWPoison(vmf.page))) {
2808 if (ret & VM_FAULT_LOCKED)
2809 unlock_page(vmf.page);
2810 page_cache_release(vmf.page);
2811 return VM_FAULT_HWPOISON;
2812 }
2813
2814 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2815 lock_page(vmf.page);
2816 else
2817 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2818
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002819 out:
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002820 *page = vmf.page;
2821 return ret;
2822}
2823
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002824/**
2825 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2826 *
2827 * @vma: virtual memory area
2828 * @address: user virtual address
2829 * @page: page to map
2830 * @pte: pointer to target page table entry
2831 * @write: true, if new entry is writable
2832 * @anon: true, if it's anonymous page
2833 *
2834 * Caller must hold page table lock relevant for @pte.
2835 *
2836 * Target users are page handler itself and implementations of
2837 * vm_ops->map_pages.
2838 */
2839void do_set_pte(struct vm_area_struct *vma, unsigned long address,
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002840 struct page *page, pte_t *pte, bool write, bool anon)
2841{
2842 pte_t entry;
2843
2844 flush_icache_page(vma, page);
2845 entry = mk_pte(page, vma->vm_page_prot);
2846 if (write)
2847 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002848 if (anon) {
2849 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002850 page_add_new_anon_rmap(page, vma, address, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002851 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002852 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002853 page_add_file_rmap(page);
2854 }
2855 set_pte_at(vma->vm_mm, address, pte, entry);
2856
2857 /* no need to invalidate: a not-present page won't be cached */
2858 update_mmu_cache(vma, address, pte);
2859}
2860
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07002861static unsigned long fault_around_bytes __read_mostly =
2862 rounddown_pow_of_two(65536);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002863
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002864#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002865static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002866{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002867 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002868 return 0;
2869}
2870
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002871/*
2872 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2873 * rounded down to nearest page order. It's what do_fault_around() expects to
2874 * see.
2875 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002876static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002877{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002878 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002879 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07002880 if (val > PAGE_SIZE)
2881 fault_around_bytes = rounddown_pow_of_two(val);
2882 else
2883 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002884 return 0;
2885}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002886DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2887 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002888
2889static int __init fault_around_debugfs(void)
2890{
2891 void *ret;
2892
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002893 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2894 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002895 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07002896 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002897 return 0;
2898}
2899late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07002900#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002901
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07002902/*
2903 * do_fault_around() tries to map few pages around the fault address. The hope
2904 * is that the pages will be needed soon and this will lower the number of
2905 * faults to handle.
2906 *
2907 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2908 * not ready to be mapped: not up-to-date, locked, etc.
2909 *
2910 * This function is called with the page table lock taken. In the split ptlock
2911 * case the page table lock only protects only those entries which belong to
2912 * the page table corresponding to the fault address.
2913 *
2914 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2915 * only once.
2916 *
2917 * fault_around_pages() defines how many pages we'll try to map.
2918 * do_fault_around() expects it to return a power of two less than or equal to
2919 * PTRS_PER_PTE.
2920 *
2921 * The virtual address of the area that we map is naturally aligned to the
2922 * fault_around_pages() value (and therefore to page order). This way it's
2923 * easier to guarantee that we don't cross page table boundaries.
2924 */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002925static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2926 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2927{
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002928 unsigned long start_addr, nr_pages, mask;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002929 pgoff_t max_pgoff;
2930 struct vm_fault vmf;
2931 int off;
2932
Jason Low4db0c3c2015-04-15 16:14:08 -07002933 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002934 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2935
2936 start_addr = max(address & mask, vma->vm_start);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002937 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2938 pte -= off;
2939 pgoff -= off;
2940
2941 /*
2942 * max_pgoff is either end of page table or end of vma
Kirill A. Shutemov850e9c62014-06-04 16:10:45 -07002943 * or fault_around_pages() from pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002944 */
2945 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2946 PTRS_PER_PTE - 1;
2947 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07002948 pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002949
2950 /* Check if it makes any sense to call ->map_pages */
2951 while (!pte_none(*pte)) {
2952 if (++pgoff > max_pgoff)
2953 return;
2954 start_addr += PAGE_SIZE;
2955 if (start_addr >= vma->vm_end)
2956 return;
2957 pte++;
2958 }
2959
2960 vmf.virtual_address = (void __user *) start_addr;
2961 vmf.pte = pte;
2962 vmf.pgoff = pgoff;
2963 vmf.max_pgoff = max_pgoff;
2964 vmf.flags = flags;
Michal Hockoc20cd452016-01-14 15:20:12 -08002965 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002966 vma->vm_ops->map_pages(vma, &vmf);
2967}
2968
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002969static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2970 unsigned long address, pmd_t *pmd,
2971 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
2972{
2973 struct page *fault_page;
2974 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002975 pte_t *pte;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002976 int ret = 0;
2977
2978 /*
2979 * Let's call ->map_pages() first and use ->fault() as fallback
2980 * if page by the offset is not ready to be mapped (cold cache or
2981 * something).
2982 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08002983 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002984 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2985 do_fault_around(vma, address, pte, pgoff, flags);
2986 if (!pte_same(*pte, orig_pte))
2987 goto unlock_out;
2988 pte_unmap_unlock(pte, ptl);
2989 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002990
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002991 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07002992 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2993 return ret;
2994
2995 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
2996 if (unlikely(!pte_same(*pte, orig_pte))) {
2997 pte_unmap_unlock(pte, ptl);
2998 unlock_page(fault_page);
2999 page_cache_release(fault_page);
3000 return ret;
3001 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003002 do_set_pte(vma, address, fault_page, pte, false, false);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003003 unlock_page(fault_page);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003004unlock_out:
3005 pte_unmap_unlock(pte, ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003006 return ret;
3007}
3008
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003009static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3010 unsigned long address, pmd_t *pmd,
3011 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3012{
3013 struct page *fault_page, *new_page;
Johannes Weiner00501b52014-08-08 14:19:20 -07003014 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003015 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003016 pte_t *pte;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003017 int ret;
3018
3019 if (unlikely(anon_vma_prepare(vma)))
3020 return VM_FAULT_OOM;
3021
3022 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3023 if (!new_page)
3024 return VM_FAULT_OOM;
3025
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003026 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003027 page_cache_release(new_page);
3028 return VM_FAULT_OOM;
3029 }
3030
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003031 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003032 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3033 goto uncharge_out;
3034
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003035 if (fault_page)
3036 copy_user_highpage(new_page, fault_page, address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003037 __SetPageUptodate(new_page);
3038
3039 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3040 if (unlikely(!pte_same(*pte, orig_pte))) {
3041 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003042 if (fault_page) {
3043 unlock_page(fault_page);
3044 page_cache_release(fault_page);
3045 } else {
3046 /*
3047 * The fault handler has no page to lock, so it holds
Yigal Korman0df9d412015-11-16 14:09:15 +02003048 * i_mmap_lock for read to protect against truncate.
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003049 */
Yigal Korman0df9d412015-11-16 14:09:15 +02003050 i_mmap_unlock_read(vma->vm_file->f_mapping);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003051 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003052 goto uncharge_out;
3053 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003054 do_set_pte(vma, address, new_page, pte, true, true);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003055 mem_cgroup_commit_charge(new_page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07003056 lru_cache_add_active_or_unevictable(new_page, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003057 pte_unmap_unlock(pte, ptl);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003058 if (fault_page) {
3059 unlock_page(fault_page);
3060 page_cache_release(fault_page);
3061 } else {
3062 /*
3063 * The fault handler has no page to lock, so it holds
Yigal Korman0df9d412015-11-16 14:09:15 +02003064 * i_mmap_lock for read to protect against truncate.
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003065 */
Yigal Korman0df9d412015-11-16 14:09:15 +02003066 i_mmap_unlock_read(vma->vm_file->f_mapping);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003067 }
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003068 return ret;
3069uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003070 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003071 page_cache_release(new_page);
3072 return ret;
3073}
3074
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003075static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Hugh Dickins16abfa02007-10-04 16:56:06 +01003076 unsigned long address, pmd_t *pmd,
Nick Piggin54cb8822007-07-19 01:46:59 -07003077 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078{
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003079 struct page *fault_page;
3080 struct address_space *mapping;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003081 spinlock_t *ptl;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003082 pte_t *pte;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003083 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003084 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003085
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003086 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003087 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003088 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089
3090 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003091 * Check if the backing address space wants to know that the page is
3092 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003094 if (vma->vm_ops->page_mkwrite) {
3095 unlock_page(fault_page);
3096 tmp = do_page_mkwrite(vma, fault_page, address);
3097 if (unlikely(!tmp ||
3098 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3099 page_cache_release(fault_page);
3100 return tmp;
3101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102 }
3103
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003104 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3105 if (unlikely(!pte_same(*pte, orig_pte))) {
3106 pte_unmap_unlock(pte, ptl);
3107 unlock_page(fault_page);
3108 page_cache_release(fault_page);
3109 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003110 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003111 do_set_pte(vma, address, fault_page, pte, true, false);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003112 pte_unmap_unlock(pte, ptl);
3113
3114 if (set_page_dirty(fault_page))
3115 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003116 /*
3117 * Take a local copy of the address_space - page.mapping may be zeroed
3118 * by truncate after unlock_page(). The address_space itself remains
3119 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3120 * release semantics to prevent the compiler from undoing this copying.
3121 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003122 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003123 unlock_page(fault_page);
3124 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3125 /*
3126 * Some device drivers do not set page.mapping but still
3127 * dirty their pages
3128 */
3129 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003130 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003131
Johannes Weiner74ec6752015-02-10 14:11:28 -08003132 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003133 file_update_time(vma->vm_file);
3134
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003135 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003136}
Nick Piggind00806b2007-07-19 01:46:57 -07003137
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003138/*
3139 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3140 * but allow concurrent faults).
3141 * The mmap_sem may have been released depending on flags and our
3142 * return value. See filemap_fault() and __lock_page_or_retry().
3143 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003144static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Nick Piggin54cb8822007-07-19 01:46:59 -07003145 unsigned long address, pte_t *page_table, pmd_t *pmd,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003146 unsigned int flags, pte_t orig_pte)
Nick Piggin54cb8822007-07-19 01:46:59 -07003147{
Matthew Wilcox88193f72016-03-15 14:57:28 -07003148 pgoff_t pgoff = linear_page_index(vma, address);
Nick Piggin54cb8822007-07-19 01:46:59 -07003149
Hugh Dickins16abfa02007-10-04 16:56:06 +01003150 pte_unmap(page_table);
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003151 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3152 if (!vma->vm_ops->fault)
3153 return VM_FAULT_SIGBUS;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003154 if (!(flags & FAULT_FLAG_WRITE))
3155 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3156 orig_pte);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003157 if (!(vma->vm_flags & VM_SHARED))
3158 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3159 orig_pte);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003160 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
Nick Piggin54cb8822007-07-19 01:46:59 -07003161}
3162
Rashika Kheriab19a9932014-04-03 14:48:02 -07003163static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003164 unsigned long addr, int page_nid,
3165 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003166{
3167 get_page(page);
3168
3169 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003170 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003171 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003172 *flags |= TNF_FAULT_LOCAL;
3173 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003174
3175 return mpol_misplaced(page, vma, addr);
3176}
3177
Rashika Kheriab19a9932014-04-03 14:48:02 -07003178static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
Mel Gormand10e63f2012-10-25 14:16:31 +02003179 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3180{
Mel Gorman4daae3b2012-11-02 11:33:45 +00003181 struct page *page = NULL;
Mel Gormand10e63f2012-10-25 14:16:31 +02003182 spinlock_t *ptl;
Mel Gorman8191acb2013-10-07 11:28:45 +01003183 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003184 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003185 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003186 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003187 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003188 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003189
Mel Gormanc0e7cad2015-02-12 14:58:41 -08003190 /* A PROT_NONE fault should not end up here */
3191 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3192
Mel Gormand10e63f2012-10-25 14:16:31 +02003193 /*
3194 * The "pte" at this point cannot be used safely without
3195 * validation through pte_unmap_same(). It's of NUMA type but
3196 * the pfn may be screwed if the read is non atomic.
3197 *
Mel Gorman4d942462015-02-12 14:58:28 -08003198 * We can safely just do a "set_pte_at()", because the old
3199 * page table entry is not accessible, so there would be no
3200 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003201 */
3202 ptl = pte_lockptr(mm, pmd);
3203 spin_lock(ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003204 if (unlikely(!pte_same(*ptep, pte))) {
3205 pte_unmap_unlock(ptep, ptl);
3206 goto out;
3207 }
3208
Mel Gorman4d942462015-02-12 14:58:28 -08003209 /* Make it present again */
3210 pte = pte_modify(pte, vma->vm_page_prot);
3211 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003212 if (was_writable)
3213 pte = pte_mkwrite(pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003214 set_pte_at(mm, addr, ptep, pte);
3215 update_mmu_cache(vma, addr, ptep);
3216
3217 page = vm_normal_page(vma, addr, pte);
3218 if (!page) {
3219 pte_unmap_unlock(ptep, ptl);
3220 return 0;
3221 }
3222
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003223 /* TODO: handle PTE-mapped THP */
3224 if (PageCompound(page)) {
3225 pte_unmap_unlock(ptep, ptl);
3226 return 0;
3227 }
3228
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003229 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003230 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3231 * much anyway since they can be in shared cache state. This misses
3232 * the case where a mapping is writable but the process never writes
3233 * to it but pte_write gets cleared during protection updates and
3234 * pte_dirty has unpredictable behaviour between PTE scan updates,
3235 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003236 */
Mel Gormanbea66fb2015-03-25 15:55:37 -07003237 if (!(vma->vm_flags & VM_WRITE))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003238 flags |= TNF_NO_GROUP;
3239
Rik van Rieldabe1d92013-10-07 11:29:34 +01003240 /*
3241 * Flag if the page is shared between multiple address spaces. This
3242 * is later used when determining whether to group tasks together
3243 */
3244 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3245 flags |= TNF_SHARED;
3246
Peter Zijlstra90572892013-10-07 11:29:20 +01003247 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003248 page_nid = page_to_nid(page);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003249 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003250 pte_unmap_unlock(ptep, ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003251 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003252 put_page(page);
3253 goto out;
3254 }
3255
3256 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003257 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003258 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003259 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003260 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003261 } else
3262 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003263
3264out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003265 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003266 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003267 return 0;
3268}
3269
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003270static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3271 unsigned long address, pmd_t *pmd, unsigned int flags)
3272{
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003273 if (vma_is_anonymous(vma))
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003274 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3275 if (vma->vm_ops->pmd_fault)
3276 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3277 return VM_FAULT_FALLBACK;
3278}
3279
3280static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3281 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3282 unsigned int flags)
3283{
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003284 if (vma_is_anonymous(vma))
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003285 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3286 if (vma->vm_ops->pmd_fault)
3287 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3288 return VM_FAULT_FALLBACK;
3289}
3290
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291/*
3292 * These routines also need to handle stuff like marking pages dirty
3293 * and/or accessed for architectures that don't do it in hardware (most
3294 * RISC architectures). The early dirtying is also good on the i386.
3295 *
3296 * There is also a hook called "update_mmu_cache()" that architectures
3297 * with external mmu caches can use to update those (ie the Sparc or
3298 * PowerPC hashed page tables that act as extended TLBs).
3299 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07003300 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3301 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003302 * We return with pte unmapped and unlocked.
3303 *
3304 * The mmap_sem may have been released depending on flags and our
3305 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306 */
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003307static int handle_pte_fault(struct mm_struct *mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003308 struct vm_area_struct *vma, unsigned long address,
3309 pte_t *pte, pmd_t *pmd, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310{
3311 pte_t entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003312 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313
Christian Borntraegere37c6982014-12-07 21:41:33 +01003314 /*
3315 * some architectures can have larger ptes than wordsize,
3316 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3317 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3318 * The code below just needs a consistent view for the ifs and
3319 * we later double check anyway with the ptl lock held. So here
3320 * a barrier will do.
3321 */
3322 entry = *pte;
3323 barrier();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07003325 if (pte_none(entry)) {
Oleg Nesterovb5330622015-09-08 14:58:28 -07003326 if (vma_is_anonymous(vma))
3327 return do_anonymous_page(mm, vma, address,
3328 pte, pmd, flags);
3329 else
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003330 return do_fault(mm, vma, address, pte, pmd,
3331 flags, entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07003332 }
Hugh Dickins65500d22005-10-29 18:15:59 -07003333 return do_swap_page(mm, vma, address,
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003334 pte, pmd, flags, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335 }
3336
Mel Gorman8a0516e2015-02-12 14:58:22 -08003337 if (pte_protnone(entry))
Mel Gormand10e63f2012-10-25 14:16:31 +02003338 return do_numa_page(mm, vma, address, entry, pte, pmd);
3339
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07003340 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003341 spin_lock(ptl);
3342 if (unlikely(!pte_same(*pte, entry)))
3343 goto unlock;
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003344 if (flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003345 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003346 return do_wp_page(mm, vma, address,
3347 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348 entry = pte_mkdirty(entry);
3349 }
3350 entry = pte_mkyoung(entry);
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003351 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
Russell King4b3073e2009-12-18 16:40:18 +00003352 update_mmu_cache(vma, address, pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003353 } else {
3354 /*
3355 * This is needed only for protection faults but the arch code
3356 * is not yet telling us if this is a protection fault or not.
3357 * This still avoids useless tlb flushes for .text page faults
3358 * with threads.
3359 */
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003360 if (flags & FAULT_FLAG_WRITE)
Shaohua Li61c77322010-08-16 09:16:55 +08003361 flush_tlb_fix_spurious_fault(vma, address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003362 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003363unlock:
3364 pte_unmap_unlock(pte, ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003365 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366}
3367
3368/*
3369 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003370 *
3371 * The mmap_sem may have been released depending on flags and our
3372 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003374static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3375 unsigned long address, unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376{
3377 pgd_t *pgd;
3378 pud_t *pud;
3379 pmd_t *pmd;
3380 pte_t *pte;
3381
Dave Hansen1b2ee122016-02-12 13:02:21 -08003382 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
Dave Hansend61172b2016-02-12 13:02:24 -08003383 flags & FAULT_FLAG_INSTRUCTION,
Dave Hansen1b2ee122016-02-12 13:02:21 -08003384 flags & FAULT_FLAG_REMOTE))
Dave Hansen33a709b2016-02-12 13:02:19 -08003385 return VM_FAULT_SIGSEGV;
3386
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01003387 if (unlikely(is_vm_hugetlb_page(vma)))
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003388 return hugetlb_fault(mm, vma, address, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391 pud = pud_alloc(mm, pgd, address);
3392 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003393 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394 pmd = pmd_alloc(mm, pud, address);
3395 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003396 return VM_FAULT_OOM;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003397 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003398 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003399 if (!(ret & VM_FAULT_FALLBACK))
3400 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003401 } else {
3402 pmd_t orig_pmd = *pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003403 int ret;
3404
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003405 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003406 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Will Deacona1dd4502012-12-11 16:01:27 -08003407 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3408
Mel Gorman8a0516e2015-02-12 14:58:22 -08003409 if (pmd_protnone(orig_pmd))
Mel Gorman4daae3b2012-11-02 11:33:45 +00003410 return do_huge_pmd_numa_page(mm, vma, address,
Mel Gormand10e63f2012-10-25 14:16:31 +02003411 orig_pmd, pmd);
3412
Linus Torvalds3d59eeb2012-12-16 14:33:25 -08003413 if (dirty && !pmd_write(orig_pmd)) {
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003414 ret = wp_huge_pmd(mm, vma, address, pmd,
3415 orig_pmd, flags);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003416 if (!(ret & VM_FAULT_FALLBACK))
3417 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003418 } else {
3419 huge_pmd_set_accessed(mm, vma, address, pmd,
3420 orig_pmd, dirty);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003421 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003422 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003423 }
3424 }
3425
3426 /*
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -07003427 * Use pte_alloc() instead of pte_alloc_map, because we can't
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003428 * run pte_offset_map on the pmd, if an huge pmd could
3429 * materialize from under us from a different thread.
3430 */
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -07003431 if (unlikely(pte_alloc(mm, pmd, address)))
Hugh Dickinsc74df322005-10-29 18:16:23 -07003432 return VM_FAULT_OOM;
Andrea Arcangeliad33bb02016-02-26 15:19:28 -08003433 /*
3434 * If a huge pmd materialized under us just retry later. Use
3435 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3436 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3437 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3438 * in a different thread of this mm, in turn leading to a misleading
3439 * pmd_trans_huge() retval. All we have to ensure is that it is a
3440 * regular pmd that we can walk with pte_offset_map() and we can do that
3441 * through an atomic read in C, which is what pmd_trans_unstable()
3442 * provides.
3443 */
3444 if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003445 return 0;
3446 /*
3447 * A regular pmd is established and it can't morph into a huge pmd
3448 * from under us anymore at this point because we hold the mmap_sem
3449 * read mode and khugepaged takes it in write mode. So now it's
3450 * safe to run pte_offset_map().
3451 */
3452 pte = pte_offset_map(pmd, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453
Linus Torvalds30c9f3a2009-04-10 08:43:11 -07003454 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003455}
3456
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003457/*
3458 * By the time we get here, we already hold the mm semaphore
3459 *
3460 * The mmap_sem may have been released depending on flags and our
3461 * return value. See filemap_fault() and __lock_page_or_retry().
3462 */
Johannes Weiner519e5242013-09-12 15:13:42 -07003463int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3464 unsigned long address, unsigned int flags)
3465{
3466 int ret;
3467
3468 __set_current_state(TASK_RUNNING);
3469
3470 count_vm_event(PGFAULT);
3471 mem_cgroup_count_vm_event(mm, PGFAULT);
3472
3473 /* do counter updates before entering really critical section. */
3474 check_sync_rss_stat(current);
3475
3476 /*
3477 * Enable the memcg OOM handling for faults triggered in user
3478 * space. Kernel faults are handled more gracefully.
3479 */
3480 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003481 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003482
3483 ret = __handle_mm_fault(mm, vma, address, flags);
3484
Johannes Weiner49426422013-10-16 13:46:59 -07003485 if (flags & FAULT_FLAG_USER) {
3486 mem_cgroup_oom_disable();
3487 /*
3488 * The task may have entered a memcg OOM situation but
3489 * if the allocation error was handled gracefully (no
3490 * VM_FAULT_OOM), there is no need to kill anything.
3491 * Just clean up the OOM state peacefully.
3492 */
3493 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3494 mem_cgroup_oom_synchronize(false);
3495 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003496
Johannes Weiner519e5242013-09-12 15:13:42 -07003497 return ret;
3498}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003499EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003500
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501#ifndef __PAGETABLE_PUD_FOLDED
3502/*
3503 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003504 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003506int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003508 pud_t *new = pud_alloc_one(mm, address);
3509 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003510 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511
Nick Piggin362a61a2008-05-14 06:37:36 +02003512 smp_wmb(); /* See comment in __pte_alloc */
3513
Hugh Dickins872fec12005-10-29 18:16:21 -07003514 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003515 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003516 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003517 else
3518 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003519 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003520 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521}
3522#endif /* __PAGETABLE_PUD_FOLDED */
3523
3524#ifndef __PAGETABLE_PMD_FOLDED
3525/*
3526 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003527 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003529int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003531 pmd_t *new = pmd_alloc_one(mm, address);
3532 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003533 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534
Nick Piggin362a61a2008-05-14 06:37:36 +02003535 smp_wmb(); /* See comment in __pte_alloc */
3536
Hugh Dickins872fec12005-10-29 18:16:21 -07003537 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003539 if (!pud_present(*pud)) {
3540 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003541 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003542 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003543 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003544#else
3545 if (!pgd_present(*pud)) {
3546 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003547 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003548 } else /* Another has populated it */
3549 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003551 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003552 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553}
3554#endif /* __PAGETABLE_PMD_FOLDED */
3555
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003556static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003557 pte_t **ptepp, spinlock_t **ptlp)
3558{
3559 pgd_t *pgd;
3560 pud_t *pud;
3561 pmd_t *pmd;
3562 pte_t *ptep;
3563
3564 pgd = pgd_offset(mm, address);
3565 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3566 goto out;
3567
3568 pud = pud_offset(pgd, address);
3569 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3570 goto out;
3571
3572 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003573 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003574 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3575 goto out;
3576
3577 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3578 if (pmd_huge(*pmd))
3579 goto out;
3580
3581 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3582 if (!ptep)
3583 goto out;
3584 if (!pte_present(*ptep))
3585 goto unlock;
3586 *ptepp = ptep;
3587 return 0;
3588unlock:
3589 pte_unmap_unlock(ptep, *ptlp);
3590out:
3591 return -EINVAL;
3592}
3593
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003594static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3595 pte_t **ptepp, spinlock_t **ptlp)
3596{
3597 int res;
3598
3599 /* (void) is needed to make gcc happy */
3600 (void) __cond_lock(*ptlp,
3601 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3602 return res;
3603}
3604
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003605/**
3606 * follow_pfn - look up PFN at a user virtual address
3607 * @vma: memory mapping
3608 * @address: user virtual address
3609 * @pfn: location to store found PFN
3610 *
3611 * Only IO mappings and raw PFN mappings are allowed.
3612 *
3613 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3614 */
3615int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3616 unsigned long *pfn)
3617{
3618 int ret = -EINVAL;
3619 spinlock_t *ptl;
3620 pte_t *ptep;
3621
3622 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3623 return ret;
3624
3625 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3626 if (ret)
3627 return ret;
3628 *pfn = pte_pfn(*ptep);
3629 pte_unmap_unlock(ptep, ptl);
3630 return 0;
3631}
3632EXPORT_SYMBOL(follow_pfn);
3633
Rik van Riel28b2ee22008-07-23 21:27:05 -07003634#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003635int follow_phys(struct vm_area_struct *vma,
3636 unsigned long address, unsigned int flags,
3637 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003638{
Johannes Weiner03668a42009-06-16 15:32:34 -07003639 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003640 pte_t *ptep, pte;
3641 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003642
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003643 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3644 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003645
Johannes Weiner03668a42009-06-16 15:32:34 -07003646 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003647 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003648 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003649
Rik van Riel28b2ee22008-07-23 21:27:05 -07003650 if ((flags & FOLL_WRITE) && !pte_write(pte))
3651 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003652
3653 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003654 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003655
Johannes Weiner03668a42009-06-16 15:32:34 -07003656 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003657unlock:
3658 pte_unmap_unlock(ptep, ptl);
3659out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003660 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003661}
3662
3663int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3664 void *buf, int len, int write)
3665{
3666 resource_size_t phys_addr;
3667 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003668 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003669 int offset = addr & (PAGE_SIZE-1);
3670
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003671 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003672 return -EINVAL;
3673
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003674 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003675 if (write)
3676 memcpy_toio(maddr + offset, buf, len);
3677 else
3678 memcpy_fromio(buf, maddr + offset, len);
3679 iounmap(maddr);
3680
3681 return len;
3682}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003683EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003684#endif
3685
David Howells0ec76a12006-09-27 01:50:15 -07003686/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003687 * Access another process' address space as given in mm. If non-NULL, use the
3688 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003689 */
Stephen Wilson206cb632011-03-13 15:49:19 -04003690static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3691 unsigned long addr, void *buf, int len, int write)
David Howells0ec76a12006-09-27 01:50:15 -07003692{
David Howells0ec76a12006-09-27 01:50:15 -07003693 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003694 void *old_buf = buf;
3695
David Howells0ec76a12006-09-27 01:50:15 -07003696 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003697 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003698 while (len) {
3699 int bytes, ret, offset;
3700 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003701 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003702
Dave Hansen1e987792016-02-12 13:01:54 -08003703 ret = get_user_pages_remote(tsk, mm, addr, 1,
David Howells0ec76a12006-09-27 01:50:15 -07003704 write, 1, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003705 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003706#ifndef CONFIG_HAVE_IOREMAP_PROT
3707 break;
3708#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003709 /*
3710 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3711 * we can access using slightly different code.
3712 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003713 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003714 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003715 break;
3716 if (vma->vm_ops && vma->vm_ops->access)
3717 ret = vma->vm_ops->access(vma, addr, buf,
3718 len, write);
3719 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003720 break;
3721 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003722#endif
David Howells0ec76a12006-09-27 01:50:15 -07003723 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003724 bytes = len;
3725 offset = addr & (PAGE_SIZE-1);
3726 if (bytes > PAGE_SIZE-offset)
3727 bytes = PAGE_SIZE-offset;
3728
3729 maddr = kmap(page);
3730 if (write) {
3731 copy_to_user_page(vma, page, addr,
3732 maddr + offset, buf, bytes);
3733 set_page_dirty_lock(page);
3734 } else {
3735 copy_from_user_page(vma, page, addr,
3736 buf, maddr + offset, bytes);
3737 }
3738 kunmap(page);
3739 page_cache_release(page);
David Howells0ec76a12006-09-27 01:50:15 -07003740 }
David Howells0ec76a12006-09-27 01:50:15 -07003741 len -= bytes;
3742 buf += bytes;
3743 addr += bytes;
3744 }
3745 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003746
3747 return buf - old_buf;
3748}
Andi Kleen03252912008-01-30 13:33:18 +01003749
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003750/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003751 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003752 * @mm: the mm_struct of the target address space
3753 * @addr: start address to access
3754 * @buf: source or destination buffer
3755 * @len: number of bytes to transfer
3756 * @write: whether the access is a write
3757 *
3758 * The caller must hold a reference on @mm.
3759 */
3760int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3761 void *buf, int len, int write)
3762{
3763 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3764}
3765
Andi Kleen03252912008-01-30 13:33:18 +01003766/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003767 * Access another process' address space.
3768 * Source/target buffer must be kernel space,
3769 * Do not walk the page table directly, use get_user_pages
3770 */
3771int access_process_vm(struct task_struct *tsk, unsigned long addr,
3772 void *buf, int len, int write)
3773{
3774 struct mm_struct *mm;
3775 int ret;
3776
3777 mm = get_task_mm(tsk);
3778 if (!mm)
3779 return 0;
3780
3781 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3782 mmput(mm);
3783
3784 return ret;
3785}
3786
Andi Kleen03252912008-01-30 13:33:18 +01003787/*
3788 * Print the name of a VMA.
3789 */
3790void print_vma_addr(char *prefix, unsigned long ip)
3791{
3792 struct mm_struct *mm = current->mm;
3793 struct vm_area_struct *vma;
3794
Ingo Molnare8bff742008-02-13 20:21:06 +01003795 /*
3796 * Do not print if we are in atomic
3797 * contexts (in exception stacks, etc.):
3798 */
3799 if (preempt_count())
3800 return;
3801
Andi Kleen03252912008-01-30 13:33:18 +01003802 down_read(&mm->mmap_sem);
3803 vma = find_vma(mm, ip);
3804 if (vma && vma->vm_file) {
3805 struct file *f = vma->vm_file;
3806 char *buf = (char *)__get_free_page(GFP_KERNEL);
3807 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003808 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003809
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02003810 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003811 if (IS_ERR(p))
3812 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003813 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003814 vma->vm_start,
3815 vma->vm_end - vma->vm_start);
3816 free_page((unsigned long)buf);
3817 }
3818 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003819 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003820}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003821
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003822#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02003823void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003824{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003825 /*
3826 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3827 * holding the mmap_sem, this is safe because kernel memory doesn't
3828 * get paged out, therefore we'll never actually fault, and the
3829 * below annotations will generate false positives.
3830 */
3831 if (segment_eq(get_fs(), KERNEL_DS))
3832 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003833 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003834 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003835 __might_sleep(file, line, 0);
3836#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003837 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003838 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02003839#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003840}
David Hildenbrand9ec23532015-05-11 17:52:07 +02003841EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003842#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003843
3844#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3845static void clear_gigantic_page(struct page *page,
3846 unsigned long addr,
3847 unsigned int pages_per_huge_page)
3848{
3849 int i;
3850 struct page *p = page;
3851
3852 might_sleep();
3853 for (i = 0; i < pages_per_huge_page;
3854 i++, p = mem_map_next(p, page, i)) {
3855 cond_resched();
3856 clear_user_highpage(p, addr + i * PAGE_SIZE);
3857 }
3858}
3859void clear_huge_page(struct page *page,
3860 unsigned long addr, unsigned int pages_per_huge_page)
3861{
3862 int i;
3863
3864 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3865 clear_gigantic_page(page, addr, pages_per_huge_page);
3866 return;
3867 }
3868
3869 might_sleep();
3870 for (i = 0; i < pages_per_huge_page; i++) {
3871 cond_resched();
3872 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3873 }
3874}
3875
3876static void copy_user_gigantic_page(struct page *dst, struct page *src,
3877 unsigned long addr,
3878 struct vm_area_struct *vma,
3879 unsigned int pages_per_huge_page)
3880{
3881 int i;
3882 struct page *dst_base = dst;
3883 struct page *src_base = src;
3884
3885 for (i = 0; i < pages_per_huge_page; ) {
3886 cond_resched();
3887 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3888
3889 i++;
3890 dst = mem_map_next(dst, dst_base, i);
3891 src = mem_map_next(src, src_base, i);
3892 }
3893}
3894
3895void copy_user_huge_page(struct page *dst, struct page *src,
3896 unsigned long addr, struct vm_area_struct *vma,
3897 unsigned int pages_per_huge_page)
3898{
3899 int i;
3900
3901 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3902 copy_user_gigantic_page(dst, src, addr, vma,
3903 pages_per_huge_page);
3904 return;
3905 }
3906
3907 might_sleep();
3908 for (i = 0; i < pages_per_huge_page; i++) {
3909 cond_resched();
3910 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3911 }
3912}
3913#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003914
Olof Johansson40b64ac2013-12-20 14:28:05 -08003915#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003916
3917static struct kmem_cache *page_ptl_cachep;
3918
3919void __init ptlock_cache_init(void)
3920{
3921 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3922 SLAB_PANIC, NULL);
3923}
3924
Peter Zijlstra539edb52013-11-14 14:31:52 -08003925bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003926{
3927 spinlock_t *ptl;
3928
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003929 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003930 if (!ptl)
3931 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08003932 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003933 return true;
3934}
3935
Peter Zijlstra539edb52013-11-14 14:31:52 -08003936void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003937{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08003938 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08003939}
3940#endif