blob: 49d9b42e0d60eb07182a06cd7de9e78db84f4849 [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>
Jan Karabc2466e2016-05-12 18:29:19 +020066#include <linux/dax.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070067
Alexey Dobriyan6952b612009-09-18 23:55:55 +040068#include <asm/io.h>
Dave Hansen33a709b2016-02-12 13:02:19 -080069#include <asm/mmu_context.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070070#include <asm/pgalloc.h>
71#include <asm/uaccess.h>
72#include <asm/tlb.h>
73#include <asm/tlbflush.h>
74#include <asm/pgtable.h>
75
Jan Beulich42b77722008-07-23 21:27:10 -070076#include "internal.h"
77
Peter Zijlstra90572892013-10-07 11:29:20 +010078#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
79#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
Peter Zijlstra75980e92013-02-22 16:34:32 -080080#endif
81
Andy Whitcroftd41dee32005-06-23 00:07:54 -070082#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070083/* use the per-pgdat data instead for discontigmem - mbligh */
84unsigned long max_mapnr;
85struct page *mem_map;
86
87EXPORT_SYMBOL(max_mapnr);
88EXPORT_SYMBOL(mem_map);
89#endif
90
Linus Torvalds1da177e2005-04-16 15:20:36 -070091/*
92 * A number of key systems in x86 including ioremap() rely on the assumption
93 * that high_memory defines the upper bound on direct map memory, then end
94 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
95 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
96 * and ZONE_HIGHMEM.
97 */
98void * high_memory;
Linus Torvalds1da177e2005-04-16 15:20:36 -070099
Linus Torvalds1da177e2005-04-16 15:20:36 -0700100EXPORT_SYMBOL(high_memory);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101
Ingo Molnar32a93232008-02-06 22:39:44 +0100102/*
103 * Randomize the address space (stacks, mmaps, brk, etc.).
104 *
105 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
106 * as ancient (libc5 based) binaries can segfault. )
107 */
108int randomize_va_space __read_mostly =
109#ifdef CONFIG_COMPAT_BRK
110 1;
111#else
112 2;
113#endif
Andi Kleena62eaf12006-02-16 23:41:58 +0100114
115static int __init disable_randmaps(char *s)
116{
117 randomize_va_space = 0;
OGAWA Hirofumi9b410462006-03-31 02:30:33 -0800118 return 1;
Andi Kleena62eaf12006-02-16 23:41:58 +0100119}
120__setup("norandmaps", disable_randmaps);
121
Hugh Dickins62eede62009-09-21 17:03:34 -0700122unsigned long zero_pfn __read_mostly;
Hugh Dickins03f64622009-09-21 17:03:35 -0700123unsigned long highest_memmap_pfn __read_mostly;
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700124
Ard Biesheuvel0b700682014-09-12 22:17:23 +0200125EXPORT_SYMBOL(zero_pfn);
126
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700127/*
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 */
130static int __init init_zero_pfn(void)
131{
132 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 return 0;
134}
135core_initcall(init_zero_pfn);
Andi Kleena62eaf12006-02-16 23:41:58 +0100136
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800137
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800138#if defined(SPLIT_RSS_COUNTING)
139
David Rientjesea48cf72012-03-21 16:34:13 -0700140void sync_mm_rss(struct mm_struct *mm)
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800141{
142 int i;
143
144 for (i = 0; i < NR_MM_COUNTERS; i++) {
David Rientjes05af2e12012-03-21 16:34:13 -0700145 if (current->rss_stat.count[i]) {
146 add_mm_counter(mm, i, current->rss_stat.count[i]);
147 current->rss_stat.count[i] = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800148 }
149 }
David Rientjes05af2e12012-03-21 16:34:13 -0700150 current->rss_stat.events = 0;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800151}
152
153static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
154{
155 struct task_struct *task = current;
156
157 if (likely(task->mm == mm))
158 task->rss_stat.count[member] += val;
159 else
160 add_mm_counter(mm, member, val);
161}
162#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164
165/* sync counter once per 64 page faults */
166#define TASK_RSS_EVENTS_THRESH (64)
167static void check_sync_rss_stat(struct task_struct *task)
168{
169 if (unlikely(task != current))
170 return;
171 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
David Rientjesea48cf72012-03-21 16:34:13 -0700172 sync_mm_rss(task->mm);
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800173}
Peter Zijlstra9547d012011-05-24 17:12:14 -0700174#else /* SPLIT_RSS_COUNTING */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800175
176#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
177#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
178
179static void check_sync_rss_stat(struct task_struct *task)
180{
181}
182
Peter Zijlstra9547d012011-05-24 17:12:14 -0700183#endif /* SPLIT_RSS_COUNTING */
184
185#ifdef HAVE_GENERIC_MMU_GATHER
186
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700187static bool tlb_next_batch(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700188{
189 struct mmu_gather_batch *batch;
190
191 batch = tlb->active;
192 if (batch->next) {
193 tlb->active = batch->next;
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700194 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700195 }
196
Michal Hocko53a59fc2013-01-04 15:35:12 -0800197 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700198 return false;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800199
Peter Zijlstra9547d012011-05-24 17:12:14 -0700200 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
201 if (!batch)
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700202 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700203
Michal Hocko53a59fc2013-01-04 15:35:12 -0800204 tlb->batch_count++;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700205 batch->next = NULL;
206 batch->nr = 0;
207 batch->max = MAX_GATHER_BATCH;
208
209 tlb->active->next = batch;
210 tlb->active = batch;
211
Nicholas Krauseca1d6c72015-09-04 15:48:22 -0700212 return true;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700213}
214
215/* tlb_gather_mmu
216 * Called to initialize an (on-stack) mmu_gather structure for page-table
217 * tear-down from @mm. The @fullmm argument is used when @mm is without
218 * users and we're going to destroy the full address space (exit/execve).
219 */
Linus Torvalds2b047252013-08-15 11:42:25 -0700220void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700221{
222 tlb->mm = mm;
223
Linus Torvalds2b047252013-08-15 11:42:25 -0700224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
Dave Hansen1de14c32013-04-12 16:23:54 -0700226 tlb->need_flush_all = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700227 tlb->local.next = NULL;
228 tlb->local.nr = 0;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
Michal Hocko53a59fc2013-01-04 15:35:12 -0800231 tlb->batch_count = 0;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700232
233#ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 tlb->batch = NULL;
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800235#endif
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700236 tlb->page_size = 0;
Will Deaconfb7332a2014-10-29 10:03:09 +0000237
238 __tlb_reset_range(tlb);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700239}
240
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700241static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700242{
Will Deacon721c21c2015-01-12 19:10:55 +0000243 if (!tlb->end)
244 return;
245
Peter Zijlstra9547d012011-05-24 17:12:14 -0700246 tlb_flush(tlb);
Joerg Roedel34ee6452014-11-13 13:46:09 +1100247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700248#ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
250#endif
Will Deaconfb7332a2014-10-29 10:03:09 +0000251 __tlb_reset_range(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700252}
253
254static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255{
256 struct mmu_gather_batch *batch;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700257
Will Deacon721c21c2015-01-12 19:10:55 +0000258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
Peter Zijlstra9547d012011-05-24 17:12:14 -0700259 free_pages_and_swap_cache(batch->pages, batch->nr);
260 batch->nr = 0;
261 }
262 tlb->active = &tlb->local;
263}
264
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700265void tlb_flush_mmu(struct mmu_gather *tlb)
266{
Linus Torvalds1cf35d42014-04-25 16:05:40 -0700267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
269}
270
Peter Zijlstra9547d012011-05-24 17:12:14 -0700271/* tlb_finish_mmu
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
274 */
275void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276{
277 struct mmu_gather_batch *batch, *next;
278
279 tlb_flush_mmu(tlb);
280
281 /* keep the page table cache within bounds */
282 check_pgt_cache();
283
284 for (batch = tlb->local.next; batch; batch = next) {
285 next = batch->next;
286 free_pages((unsigned long)batch, 0);
287 }
288 tlb->local.next = NULL;
289}
290
291/* __tlb_remove_page
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700296 *returns true if the caller should flush.
Peter Zijlstra9547d012011-05-24 17:12:14 -0700297 */
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700298bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
Peter Zijlstra9547d012011-05-24 17:12:14 -0700299{
300 struct mmu_gather_batch *batch;
301
Will Deaconfb7332a2014-10-29 10:03:09 +0000302 VM_BUG_ON(!tlb->end);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700303
Aneesh Kumar K.Ve77b0852016-07-26 15:24:12 -0700304 if (!tlb->page_size)
305 tlb->page_size = page_size;
306 else {
307 if (page_size != tlb->page_size)
308 return true;
309 }
310
Peter Zijlstra9547d012011-05-24 17:12:14 -0700311 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700312 if (batch->nr == batch->max) {
313 if (!tlb_next_batch(tlb))
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700314 return true;
Shaohua Li0b43c3a2011-07-08 15:39:41 -0700315 batch = tlb->active;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700316 }
Sasha Levin309381fea2014-01-23 15:52:54 -0800317 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
Peter Zijlstra9547d012011-05-24 17:12:14 -0700318
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -0700319 batch->pages[batch->nr++] = page;
320 return false;
Peter Zijlstra9547d012011-05-24 17:12:14 -0700321}
322
323#endif /* HAVE_GENERIC_MMU_GATHER */
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800324
Peter Zijlstra26723912011-05-24 17:12:00 -0700325#ifdef CONFIG_HAVE_RCU_TABLE_FREE
326
327/*
328 * See the comment near struct mmu_table_batch.
329 */
330
331static void tlb_remove_table_smp_sync(void *arg)
332{
333 /* Simply deliver the interrupt */
334}
335
336static void tlb_remove_table_one(void *table)
337{
338 /*
339 * This isn't an RCU grace period and hence the page-tables cannot be
340 * assumed to be actually RCU-freed.
341 *
342 * It is however sufficient for software page-table walkers that rely on
343 * IRQ disabling. See the comment near struct mmu_table_batch.
344 */
345 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
346 __tlb_remove_table(table);
347}
348
349static void tlb_remove_table_rcu(struct rcu_head *head)
350{
351 struct mmu_table_batch *batch;
352 int i;
353
354 batch = container_of(head, struct mmu_table_batch, rcu);
355
356 for (i = 0; i < batch->nr; i++)
357 __tlb_remove_table(batch->tables[i]);
358
359 free_page((unsigned long)batch);
360}
361
362void tlb_table_flush(struct mmu_gather *tlb)
363{
364 struct mmu_table_batch **batch = &tlb->batch;
365
366 if (*batch) {
367 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
368 *batch = NULL;
369 }
370}
371
372void tlb_remove_table(struct mmu_gather *tlb, void *table)
373{
374 struct mmu_table_batch **batch = &tlb->batch;
375
Peter Zijlstra26723912011-05-24 17:12:00 -0700376 /*
377 * When there's less then two users of this mm there cannot be a
378 * concurrent page-table walk.
379 */
380 if (atomic_read(&tlb->mm->mm_users) < 2) {
381 __tlb_remove_table(table);
382 return;
383 }
384
385 if (*batch == NULL) {
386 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
387 if (*batch == NULL) {
388 tlb_remove_table_one(table);
389 return;
390 }
391 (*batch)->nr = 0;
392 }
393 (*batch)->tables[(*batch)->nr++] = table;
394 if ((*batch)->nr == MAX_TABLE_BATCH)
395 tlb_table_flush(tlb);
396}
397
Peter Zijlstra9547d012011-05-24 17:12:14 -0700398#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
Peter Zijlstra26723912011-05-24 17:12:00 -0700399
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401 * Note: this doesn't free the actual pages themselves. That
402 * has been handled earlier when unmapping all the memory regions.
403 */
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000404static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
405 unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406{
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800407 pgtable_t token = pmd_pgtable(*pmd);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700408 pmd_clear(pmd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000409 pte_free_tlb(tlb, token, addr);
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800410 atomic_long_dec(&tlb->mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411}
412
Hugh Dickinse0da3822005-04-19 13:29:15 -0700413static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
414 unsigned long addr, unsigned long end,
415 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416{
417 pmd_t *pmd;
418 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700419 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420
Hugh Dickinse0da3822005-04-19 13:29:15 -0700421 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 do {
424 next = pmd_addr_end(addr, end);
425 if (pmd_none_or_clear_bad(pmd))
426 continue;
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000427 free_pte_range(tlb, pmd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428 } while (pmd++, addr = next, addr != end);
429
Hugh Dickinse0da3822005-04-19 13:29:15 -0700430 start &= PUD_MASK;
431 if (start < floor)
432 return;
433 if (ceiling) {
434 ceiling &= PUD_MASK;
435 if (!ceiling)
436 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700438 if (end - 1 > ceiling - 1)
439 return;
440
441 pmd = pmd_offset(pud, start);
442 pud_clear(pud);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000443 pmd_free_tlb(tlb, pmd, start);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -0800444 mm_dec_nr_pmds(tlb->mm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445}
446
Hugh Dickinse0da3822005-04-19 13:29:15 -0700447static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
448 unsigned long addr, unsigned long end,
449 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
451 pud_t *pud;
452 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700453 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454
Hugh Dickinse0da3822005-04-19 13:29:15 -0700455 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457 do {
458 next = pud_addr_end(addr, end);
459 if (pud_none_or_clear_bad(pud))
460 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700461 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 } while (pud++, addr = next, addr != end);
463
Hugh Dickinse0da3822005-04-19 13:29:15 -0700464 start &= PGDIR_MASK;
465 if (start < floor)
466 return;
467 if (ceiling) {
468 ceiling &= PGDIR_MASK;
469 if (!ceiling)
470 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700472 if (end - 1 > ceiling - 1)
473 return;
474
475 pud = pud_offset(pgd, start);
476 pgd_clear(pgd);
Benjamin Herrenschmidt9e1b32c2009-07-22 15:44:28 +1000477 pud_free_tlb(tlb, pud, start);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478}
479
480/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700481 * This function frees user-level page tables of a process.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482 */
Jan Beulich42b77722008-07-23 21:27:10 -0700483void free_pgd_range(struct mmu_gather *tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700484 unsigned long addr, unsigned long end,
485 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700486{
487 pgd_t *pgd;
488 unsigned long next;
489
Hugh Dickinse0da3822005-04-19 13:29:15 -0700490 /*
491 * The next few lines have given us lots of grief...
492 *
493 * Why are we testing PMD* at this top level? Because often
494 * there will be no work to do at all, and we'd prefer not to
495 * go all the way down to the bottom just to discover that.
496 *
497 * Why all these "- 1"s? Because 0 represents both the bottom
498 * of the address space and the top of it (using -1 for the
499 * top wouldn't help much: the masks would do the wrong thing).
500 * The rule is that addr 0 and floor 0 refer to the bottom of
501 * the address space, but end 0 and ceiling 0 refer to the top
502 * Comparisons need to use "end - 1" and "ceiling - 1" (though
503 * that end 0 case should be mythical).
504 *
505 * Wherever addr is brought up or ceiling brought down, we must
506 * be careful to reject "the opposite 0" before it confuses the
507 * subsequent tests. But what about where end is brought down
508 * by PMD_SIZE below? no, end can't go down to 0 there.
509 *
510 * Whereas we round start (addr) and ceiling down, by different
511 * masks at different levels, in order to test whether a table
512 * now has no other vmas using it, so can be freed, we don't
513 * bother to round floor or end up - the tests don't need that.
514 */
515
516 addr &= PMD_MASK;
517 if (addr < floor) {
518 addr += PMD_SIZE;
519 if (!addr)
520 return;
521 }
522 if (ceiling) {
523 ceiling &= PMD_MASK;
524 if (!ceiling)
525 return;
526 }
527 if (end - 1 > ceiling - 1)
528 end -= PMD_SIZE;
529 if (addr > end - 1)
530 return;
531
Jan Beulich42b77722008-07-23 21:27:10 -0700532 pgd = pgd_offset(tlb->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533 do {
534 next = pgd_addr_end(addr, end);
535 if (pgd_none_or_clear_bad(pgd))
536 continue;
Jan Beulich42b77722008-07-23 21:27:10 -0700537 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700539}
540
Jan Beulich42b77722008-07-23 21:27:10 -0700541void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700542 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700543{
544 while (vma) {
545 struct vm_area_struct *next = vma->vm_next;
546 unsigned long addr = vma->vm_start;
547
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700548 /*
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +1000549 * Hide vma from rmap and truncate_pagecache before freeing
550 * pgtables
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700551 */
Rik van Riel5beb4932010-03-05 13:42:07 -0800552 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700553 unlink_file_vma(vma);
554
David Gibson9da61ae2006-03-22 00:08:57 -0800555 if (is_vm_hugetlb_page(vma)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700556 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700557 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700558 } else {
559 /*
560 * Optimization: gather nearby vmas into one call down
561 */
562 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
David Gibson48669202006-03-22 00:08:58 -0800563 && !is_vm_hugetlb_page(next)) {
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700564 vma = next;
565 next = vma->vm_next;
Rik van Riel5beb4932010-03-05 13:42:07 -0800566 unlink_anon_vmas(vma);
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700567 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700568 }
569 free_pgd_range(tlb, addr, vma->vm_end,
570 floor, next? next->vm_start: ceiling);
571 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700572 vma = next;
573 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574}
575
Kirill A. Shutemov3ed3a4f2016-03-17 14:19:11 -0700576int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577{
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800578 spinlock_t *ptl;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800579 pgtable_t new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700580 if (!new)
581 return -ENOMEM;
582
Nick Piggin362a61a2008-05-14 06:37:36 +0200583 /*
584 * Ensure all pte setup (eg. pte page lock and page clearing) are
585 * visible before the pte is made visible to other CPUs by being
586 * put into page tables.
587 *
588 * The other side of the story is the pointer chasing in the page
589 * table walking code (when walking the page table without locking;
590 * ie. most of the time). Fortunately, these data accesses consist
591 * of a chain of data-dependent loads, meaning most CPUs (alpha
592 * being the notable exception) will already guarantee loads are
593 * seen in-order. See the alpha page table accessors for the
594 * smp_read_barrier_depends() barriers in page table walking code.
595 */
596 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
597
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800598 ptl = pmd_lock(mm, pmd);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800599 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Kirill A. Shutemove1f56c82013-11-14 14:30:48 -0800600 atomic_long_inc(&mm->nr_ptes);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 pmd_populate(mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800602 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800603 }
Kirill A. Shutemovc4088eb2013-11-14 14:31:04 -0800604 spin_unlock(ptl);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800605 if (new)
606 pte_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700607 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608}
609
Hugh Dickins1bb36302005-10-29 18:16:22 -0700610int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700612 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
613 if (!new)
614 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
Nick Piggin362a61a2008-05-14 06:37:36 +0200616 smp_wmb(); /* See comment in __pte_alloc */
617
Hugh Dickins1bb36302005-10-29 18:16:22 -0700618 spin_lock(&init_mm.page_table_lock);
Andrea Arcangeli8ac1f832011-01-13 15:46:43 -0800619 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
Hugh Dickins1bb36302005-10-29 18:16:22 -0700620 pmd_populate_kernel(&init_mm, pmd, new);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800621 new = NULL;
Kirill A. Shutemov4b471e82016-01-15 16:53:39 -0800622 }
Hugh Dickins1bb36302005-10-29 18:16:22 -0700623 spin_unlock(&init_mm.page_table_lock);
Martin Schwidefsky2f569af2008-02-08 04:22:04 -0800624 if (new)
625 pte_free_kernel(&init_mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700626 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627}
628
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800629static inline void init_rss_vec(int *rss)
Hugh Dickinsae859762005-10-29 18:16:05 -0700630{
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800631 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
632}
633
634static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
635{
636 int i;
637
KAMEZAWA Hiroyuki34e55232010-03-05 13:41:40 -0800638 if (current->mm == mm)
David Rientjes05af2e12012-03-21 16:34:13 -0700639 sync_mm_rss(mm);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800640 for (i = 0; i < NR_MM_COUNTERS; i++)
641 if (rss[i])
642 add_mm_counter(mm, i, rss[i]);
Hugh Dickinsae859762005-10-29 18:16:05 -0700643}
644
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645/*
Linus Torvalds6aab3412005-11-28 14:34:23 -0800646 * This function is called to print an error when a bad pte
647 * is found. For example, we might have a PFN-mapped pte in
648 * a region that doesn't allow it.
Nick Pigginb5810032005-10-29 18:16:12 -0700649 *
650 * The calling function must still handle the error.
651 */
Hugh Dickins3dc14742009-01-06 14:40:08 -0800652static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
653 pte_t pte, struct page *page)
Nick Pigginb5810032005-10-29 18:16:12 -0700654{
Hugh Dickins3dc14742009-01-06 14:40:08 -0800655 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
656 pud_t *pud = pud_offset(pgd, addr);
657 pmd_t *pmd = pmd_offset(pud, addr);
658 struct address_space *mapping;
659 pgoff_t index;
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800660 static unsigned long resume;
661 static unsigned long nr_shown;
662 static unsigned long nr_unshown;
663
664 /*
665 * Allow a burst of 60 reports, then keep quiet for that minute;
666 * or allow a steady drip of one report per second.
667 */
668 if (nr_shown == 60) {
669 if (time_before(jiffies, resume)) {
670 nr_unshown++;
671 return;
672 }
673 if (nr_unshown) {
Joe Perches11705322016-03-17 14:19:50 -0700674 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
675 nr_unshown);
Hugh Dickinsd936cf92009-01-06 14:40:12 -0800676 nr_unshown = 0;
677 }
678 nr_shown = 0;
679 }
680 if (nr_shown++ == 0)
681 resume = jiffies + 60 * HZ;
Hugh Dickins3dc14742009-01-06 14:40:08 -0800682
683 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
684 index = linear_page_index(vma, addr);
685
Joe Perches11705322016-03-17 14:19:50 -0700686 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
687 current->comm,
688 (long long)pte_val(pte), (long long)pmd_val(*pmd));
Wu Fengguang718a3822010-03-10 15:20:43 -0800689 if (page)
Dave Hansenf0b791a2014-01-23 15:52:49 -0800690 dump_page(page, "bad pte");
Joe Perches11705322016-03-17 14:19:50 -0700691 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
692 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
Hugh Dickins3dc14742009-01-06 14:40:08 -0800693 /*
694 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
695 */
Konstantin Khlebnikov26825822015-04-15 16:15:08 -0700696 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
697 vma->vm_file,
698 vma->vm_ops ? vma->vm_ops->fault : NULL,
699 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
700 mapping ? mapping->a_ops->readpage : NULL);
Nick Pigginb5810032005-10-29 18:16:12 -0700701 dump_stack();
Rusty Russell373d4d02013-01-21 17:17:39 +1030702 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
Nick Pigginb5810032005-10-29 18:16:12 -0700703}
704
705/*
Nick Piggin7e675132008-04-28 02:13:00 -0700706 * vm_normal_page -- This function gets the "struct page" associated with a pte.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800707 *
Nick Piggin7e675132008-04-28 02:13:00 -0700708 * "Special" mappings do not wish to be associated with a "struct page" (either
709 * it doesn't exist, or it exists but they don't want to touch it). In this
710 * case, NULL is returned here. "Normal" mappings do have a struct page.
Jared Hulbertb379d792008-04-28 02:12:58 -0700711 *
Nick Piggin7e675132008-04-28 02:13:00 -0700712 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
713 * pte bit, in which case this function is trivial. Secondly, an architecture
714 * may not have a spare pte bit, which requires a more complicated scheme,
715 * described below.
716 *
717 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
718 * special mapping (even if there are underlying and valid "struct pages").
719 * COWed pages of a VM_PFNMAP are always normal.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800720 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700721 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
722 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
Nick Piggin7e675132008-04-28 02:13:00 -0700723 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
724 * mapping will always honor the rule
Linus Torvalds6aab3412005-11-28 14:34:23 -0800725 *
726 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
727 *
Nick Piggin7e675132008-04-28 02:13:00 -0700728 * And for normal mappings this is false.
Jared Hulbertb379d792008-04-28 02:12:58 -0700729 *
Nick Piggin7e675132008-04-28 02:13:00 -0700730 * This restricts such mappings to be a linear translation from virtual address
731 * to pfn. To get around this restriction, we allow arbitrary mappings so long
732 * as the vma is not a COW mapping; in that case, we know that all ptes are
733 * special (because none can have been COWed).
Jared Hulbertb379d792008-04-28 02:12:58 -0700734 *
735 *
Nick Piggin7e675132008-04-28 02:13:00 -0700736 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
737 *
Jared Hulbertb379d792008-04-28 02:12:58 -0700738 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
739 * page" backing, however the difference is that _all_ pages with a struct
740 * page (that is, those where pfn_valid is true) are refcounted and considered
741 * normal pages by the VM. The disadvantage is that pages are refcounted
742 * (which can be slower and simply not an option for some PFNMAP users). The
743 * advantage is that we don't have to follow the strict linearity rule of
744 * PFNMAP mappings in order to support COWable mappings.
745 *
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800746 */
Nick Piggin7e675132008-04-28 02:13:00 -0700747#ifdef __HAVE_ARCH_PTE_SPECIAL
748# define HAVE_PTE_SPECIAL 1
749#else
750# define HAVE_PTE_SPECIAL 0
751#endif
752struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
753 pte_t pte)
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800754{
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800755 unsigned long pfn = pte_pfn(pte);
Nick Piggin7e675132008-04-28 02:13:00 -0700756
757 if (HAVE_PTE_SPECIAL) {
Hugh Dickinsb38af472014-08-29 15:18:44 -0700758 if (likely(!pte_special(pte)))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800759 goto check_pfn;
David Vrabel667a0a02014-12-18 14:48:15 +0000760 if (vma->vm_ops && vma->vm_ops->find_special_page)
761 return vma->vm_ops->find_special_page(vma, addr);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -0700762 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
763 return NULL;
Hugh Dickins62eede62009-09-21 17:03:34 -0700764 if (!is_zero_pfn(pfn))
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800765 print_bad_pte(vma, addr, pte, NULL);
Nick Piggin7e675132008-04-28 02:13:00 -0700766 return NULL;
767 }
768
769 /* !HAVE_PTE_SPECIAL case follows: */
770
Jared Hulbertb379d792008-04-28 02:12:58 -0700771 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
772 if (vma->vm_flags & VM_MIXEDMAP) {
773 if (!pfn_valid(pfn))
774 return NULL;
775 goto out;
776 } else {
Nick Piggin7e675132008-04-28 02:13:00 -0700777 unsigned long off;
778 off = (addr - vma->vm_start) >> PAGE_SHIFT;
Jared Hulbertb379d792008-04-28 02:12:58 -0700779 if (pfn == vma->vm_pgoff + off)
780 return NULL;
781 if (!is_cow_mapping(vma->vm_flags))
782 return NULL;
783 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800784 }
785
Hugh Dickinsb38af472014-08-29 15:18:44 -0700786 if (is_zero_pfn(pfn))
787 return NULL;
Hugh Dickins22b31ee2009-01-06 14:40:09 -0800788check_pfn:
789 if (unlikely(pfn > highest_memmap_pfn)) {
790 print_bad_pte(vma, addr, pte, NULL);
791 return NULL;
792 }
Linus Torvalds6aab3412005-11-28 14:34:23 -0800793
794 /*
Nick Piggin7e675132008-04-28 02:13:00 -0700795 * NOTE! We still have PageReserved() pages in the page tables.
Nick Piggin7e675132008-04-28 02:13:00 -0700796 * eg. VDSO mappings can cause them to exist.
Linus Torvalds6aab3412005-11-28 14:34:23 -0800797 */
Jared Hulbertb379d792008-04-28 02:12:58 -0700798out:
Linus Torvalds6aab3412005-11-28 14:34:23 -0800799 return pfn_to_page(pfn);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800800}
801
Gerald Schaefer28093f92016-04-28 16:18:35 -0700802#ifdef CONFIG_TRANSPARENT_HUGEPAGE
803struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
804 pmd_t pmd)
805{
806 unsigned long pfn = pmd_pfn(pmd);
807
808 /*
809 * There is no pmd_special() but there may be special pmds, e.g.
810 * in a direct-access (dax) mapping, so let's just replicate the
811 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
812 */
813 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
814 if (vma->vm_flags & VM_MIXEDMAP) {
815 if (!pfn_valid(pfn))
816 return NULL;
817 goto out;
818 } else {
819 unsigned long off;
820 off = (addr - vma->vm_start) >> PAGE_SHIFT;
821 if (pfn == vma->vm_pgoff + off)
822 return NULL;
823 if (!is_cow_mapping(vma->vm_flags))
824 return NULL;
825 }
826 }
827
828 if (is_zero_pfn(pfn))
829 return NULL;
830 if (unlikely(pfn > highest_memmap_pfn))
831 return NULL;
832
833 /*
834 * NOTE! We still have PageReserved() pages in the page tables.
835 * eg. VDSO mappings can cause them to exist.
836 */
837out:
838 return pfn_to_page(pfn);
839}
840#endif
841
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800842/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 * copy one vm_area from one task to the other. Assumes the page tables
844 * already present in the new task to be cleared in the whole range
845 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 */
847
Hugh Dickins570a335b2009-12-14 17:58:46 -0800848static inline unsigned long
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700850 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700851 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
Nick Pigginb5810032005-10-29 18:16:12 -0700853 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 pte_t pte = *src_pte;
855 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856
857 /* pte contains position in swap or file, so copy. */
858 if (unlikely(!pte_present(pte))) {
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800859 swp_entry_t entry = pte_to_swp_entry(pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700860
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800861 if (likely(!non_swap_entry(entry))) {
862 if (swap_duplicate(entry) < 0)
863 return entry.val;
Hugh Dickins570a335b2009-12-14 17:58:46 -0800864
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800865 /* make sure dst_mm is on swapoff's mmlist. */
866 if (unlikely(list_empty(&dst_mm->mmlist))) {
867 spin_lock(&mmlist_lock);
868 if (list_empty(&dst_mm->mmlist))
869 list_add(&dst_mm->mmlist,
870 &src_mm->mmlist);
871 spin_unlock(&mmlist_lock);
872 }
873 rss[MM_SWAPENTS]++;
874 } else if (is_migration_entry(entry)) {
875 page = migration_entry_to_page(entry);
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800876
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800877 rss[mm_counter(page)]++;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -0800878
Kirill A. Shutemov0661a332015-02-10 14:10:04 -0800879 if (is_write_migration_entry(entry) &&
880 is_cow_mapping(vm_flags)) {
881 /*
882 * COW mappings require pages in both
883 * parent and child to be set to read.
884 */
885 make_migration_entry_read(&entry);
886 pte = swp_entry_to_pte(entry);
887 if (pte_swp_soft_dirty(*src_pte))
888 pte = pte_swp_mksoft_dirty(pte);
889 set_pte_at(src_mm, addr, src_pte, pte);
Christoph Lameter06972122006-06-23 02:03:35 -0700890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700892 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 }
894
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 /*
896 * If it's a COW mapping, write protect it both
897 * in the parent and the child
898 */
Linus Torvalds67121172005-12-11 20:38:17 -0800899 if (is_cow_mapping(vm_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 ptep_set_wrprotect(src_mm, addr, src_pte);
Zachary Amsden3dc90792006-09-30 23:29:30 -0700901 pte = pte_wrprotect(pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903
904 /*
905 * If it's a shared mapping, mark it clean in
906 * the child
907 */
908 if (vm_flags & VM_SHARED)
909 pte = pte_mkclean(pte);
910 pte = pte_mkold(pte);
Linus Torvalds6aab3412005-11-28 14:34:23 -0800911
912 page = vm_normal_page(vma, addr, pte);
913 if (page) {
914 get_page(page);
Kirill A. Shutemov53f92632016-01-15 16:53:42 -0800915 page_dup_rmap(page, false);
Jerome Marchandeca56ff2016-01-14 15:19:26 -0800916 rss[mm_counter(page)]++;
Linus Torvalds6aab3412005-11-28 14:34:23 -0800917 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700918
919out_set_pte:
920 set_pte_at(dst_mm, addr, dst_pte, pte);
Hugh Dickins570a335b2009-12-14 17:58:46 -0800921 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922}
923
Jerome Marchand21bda262014-08-06 16:06:56 -0700924static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -0800925 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
926 unsigned long addr, unsigned long end)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927{
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700928 pte_t *orig_src_pte, *orig_dst_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700930 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700931 int progress = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800932 int rss[NR_MM_COUNTERS];
Hugh Dickins570a335b2009-12-14 17:58:46 -0800933 swp_entry_t entry = (swp_entry_t){0};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
935again:
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800936 init_rss_vec(rss);
937
Hugh Dickinsc74df322005-10-29 18:16:23 -0700938 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939 if (!dst_pte)
940 return -ENOMEM;
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700941 src_pte = pte_offset_map(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700942 src_ptl = pte_lockptr(src_mm, src_pmd);
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700943 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700944 orig_src_pte = src_pte;
945 orig_dst_pte = dst_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700946 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 do {
949 /*
950 * We are holding two locks at this point - either of them
951 * could generate latencies in another task on another CPU.
952 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700953 if (progress >= 32) {
954 progress = 0;
955 if (need_resched() ||
Nick Piggin95c354f2008-01-30 13:31:20 +0100956 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700957 break;
958 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 if (pte_none(*src_pte)) {
960 progress++;
961 continue;
962 }
Hugh Dickins570a335b2009-12-14 17:58:46 -0800963 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
964 vma, addr, rss);
965 if (entry.val)
966 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 progress += 8;
968 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700970 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -0700971 spin_unlock(src_ptl);
Peter Zijlstraece0e2b2010-10-26 14:21:52 -0700972 pte_unmap(orig_src_pte);
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -0800973 add_mm_rss_vec(dst_mm, rss);
Daisuke Nishimurac36987e2009-10-26 16:50:23 -0700974 pte_unmap_unlock(orig_dst_pte, dst_ptl);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700975 cond_resched();
Hugh Dickins570a335b2009-12-14 17:58:46 -0800976
977 if (entry.val) {
978 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
979 return -ENOMEM;
980 progress = 0;
981 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 if (addr != end)
983 goto again;
984 return 0;
985}
986
987static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pmd_t *src_pmd, *dst_pmd;
992 unsigned long next;
993
994 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
995 if (!dst_pmd)
996 return -ENOMEM;
997 src_pmd = pmd_offset(src_pud, addr);
998 do {
999 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001000 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001001 int err;
Andrea Arcangeli14d1a552011-01-13 15:47:15 -08001002 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001003 err = copy_huge_pmd(dst_mm, src_mm,
1004 dst_pmd, src_pmd, addr, vma);
1005 if (err == -ENOMEM)
1006 return -ENOMEM;
1007 if (!err)
1008 continue;
1009 /* fall through */
1010 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011 if (pmd_none_or_clear_bad(src_pmd))
1012 continue;
1013 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1014 vma, addr, next))
1015 return -ENOMEM;
1016 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1017 return 0;
1018}
1019
1020static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1021 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1022 unsigned long addr, unsigned long end)
1023{
1024 pud_t *src_pud, *dst_pud;
1025 unsigned long next;
1026
1027 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1028 if (!dst_pud)
1029 return -ENOMEM;
1030 src_pud = pud_offset(src_pgd, addr);
1031 do {
1032 next = pud_addr_end(addr, end);
1033 if (pud_none_or_clear_bad(src_pud))
1034 continue;
1035 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1036 vma, addr, next))
1037 return -ENOMEM;
1038 } while (dst_pud++, src_pud++, addr = next, addr != end);
1039 return 0;
1040}
1041
1042int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1043 struct vm_area_struct *vma)
1044{
1045 pgd_t *src_pgd, *dst_pgd;
1046 unsigned long next;
1047 unsigned long addr = vma->vm_start;
1048 unsigned long end = vma->vm_end;
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001049 unsigned long mmun_start; /* For mmu_notifiers */
1050 unsigned long mmun_end; /* For mmu_notifiers */
1051 bool is_cow;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001052 int ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053
Nick Piggind9928952005-08-28 16:49:11 +10001054 /*
1055 * Don't copy ptes where a page fault will fill them correctly.
1056 * Fork becomes much lighter when there are big shared or private
1057 * readonly mappings. The tradeoff is that copy_page_range is more
1058 * efficient than faulting.
1059 */
Kirill A. Shutemov0661a332015-02-10 14:10:04 -08001060 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1061 !vma->anon_vma)
1062 return 0;
Nick Piggind9928952005-08-28 16:49:11 +10001063
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (is_vm_hugetlb_page(vma))
1065 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1066
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001067 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001068 /*
1069 * We do not free on error cases below as remove_vma
1070 * gets called on error from higher level routine
1071 */
Suresh Siddha5180da42012-10-08 16:28:29 -07001072 ret = track_pfn_copy(vma);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001073 if (ret)
1074 return ret;
1075 }
1076
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001077 /*
1078 * We need to invalidate the secondary MMU mappings only when
1079 * there could be a permission downgrade on the ptes of the
1080 * parent mm. And a permission downgrade will only happen if
1081 * is_cow_mapping() returns true.
1082 */
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001083 is_cow = is_cow_mapping(vma->vm_flags);
1084 mmun_start = addr;
1085 mmun_end = end;
1086 if (is_cow)
1087 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1088 mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001089
1090 ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 dst_pgd = pgd_offset(dst_mm, addr);
1092 src_pgd = pgd_offset(src_mm, addr);
1093 do {
1094 next = pgd_addr_end(addr, end);
1095 if (pgd_none_or_clear_bad(src_pgd))
1096 continue;
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001097 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1098 vma, addr, next))) {
1099 ret = -ENOMEM;
1100 break;
1101 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001102 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001103
Sagi Grimberg2ec74c32012-10-08 16:33:33 -07001104 if (is_cow)
1105 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001106 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107}
1108
Robin Holt51c6f662005-11-13 16:06:42 -08001109static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001110 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001112 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113{
Nick Pigginb5810032005-10-29 18:16:12 -07001114 struct mm_struct *mm = tlb->mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001115 int force_flush = 0;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001116 int rss[NR_MM_COUNTERS];
Peter Zijlstra97a89412011-05-24 17:12:04 -07001117 spinlock_t *ptl;
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001118 pte_t *start_pte;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001119 pte_t *pte;
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001120 swp_entry_t entry;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001121 struct page *pending_page = NULL;
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001122
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001123again:
Peter Zijlstrae3032972011-05-24 17:12:01 -07001124 init_rss_vec(rss);
Steven Rostedt5f1a1902011-06-15 15:08:23 -07001125 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1126 pte = start_pte;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001127 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 do {
1129 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001130 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001132 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001135 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001136
Linus Torvalds6aab3412005-11-28 14:34:23 -08001137 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 if (unlikely(details) && page) {
1139 /*
1140 * unmap_shared_mapping_pages() wants to
1141 * invalidate cache without truncating:
1142 * unmap shared but keep private pages.
1143 */
1144 if (details->check_mapping &&
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001145 details->check_mapping != page_rmapping(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 }
Nick Pigginb5810032005-10-29 18:16:12 -07001148 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001149 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150 tlb_remove_tlb_entry(tlb, pte, addr);
1151 if (unlikely(!page))
1152 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001153
1154 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001155 if (pte_dirty(ptent)) {
Michal Hockoaac45362016-03-25 14:20:24 -07001156 /*
1157 * oom_reaper cannot tear down dirty
1158 * pages
1159 */
1160 if (unlikely(details && details->ignore_dirty))
1161 continue;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001162 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001163 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001164 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001165 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001166 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001167 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001168 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001169 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001170 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001171 if (unlikely(page_mapcount(page) < 0))
1172 print_bad_pte(vma, addr, ptent, page);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001173 if (unlikely(__tlb_remove_page(tlb, page))) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001174 force_flush = 1;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001175 pending_page = page;
Will Deaconce9ec372014-10-28 13:16:28 -07001176 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001177 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001178 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 continue;
1180 }
Michal Hockoaac45362016-03-25 14:20:24 -07001181 /* only check swap_entries if explicitly asked for in details */
1182 if (unlikely(details && !details->check_swap_entries))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001184
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001185 entry = pte_to_swp_entry(ptent);
1186 if (!non_swap_entry(entry))
1187 rss[MM_SWAPENTS]--;
1188 else if (is_migration_entry(entry)) {
1189 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001190
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001191 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001192 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001193 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001194 if (unlikely(!free_swap_and_cache(entry)))
1195 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001196 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001197 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001198
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001199 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001200 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001201
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001202 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001203 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001204 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001216 if (pending_page) {
1217 /* remove the page with new size */
1218 __tlb_remove_pte_page(tlb, pending_page);
1219 pending_page = NULL;
1220 }
Linus Torvalds2b047252013-08-15 11:42:25 -07001221 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001222 goto again;
1223 }
1224
Robin Holt51c6f662005-11-13 16:06:42 -08001225 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226}
1227
Robin Holt51c6f662005-11-13 16:06:42 -08001228static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001229 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001231 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232{
1233 pmd_t *pmd;
1234 unsigned long next;
1235
1236 pmd = pmd_offset(pud, addr);
1237 do {
1238 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001239 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001240 if (next - addr != HPAGE_PMD_SIZE) {
Hugh Dickins68428392016-05-05 16:22:09 -07001241 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1242 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001243 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001244 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001245 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001246 /* fall through */
1247 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001248 /*
1249 * Here there can be other concurrent MADV_DONTNEED or
1250 * trans huge page faults running, and if the pmd is
1251 * none or trans huge it can change under us. This is
1252 * because MADV_DONTNEED holds the mmap_sem in read
1253 * mode.
1254 */
1255 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1256 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001257 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001258next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001259 cond_resched();
1260 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001261
1262 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263}
1264
Robin Holt51c6f662005-11-13 16:06:42 -08001265static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001266 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001268 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269{
1270 pud_t *pud;
1271 unsigned long next;
1272
1273 pud = pud_offset(pgd, addr);
1274 do {
1275 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001276 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001278 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1279 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001280
1281 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282}
1283
Michal Hockoaac45362016-03-25 14:20:24 -07001284void unmap_page_range(struct mmu_gather *tlb,
Al Viro038c7aa2012-03-05 13:25:09 -05001285 struct vm_area_struct *vma,
1286 unsigned long addr, unsigned long end,
1287 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288{
1289 pgd_t *pgd;
1290 unsigned long next;
1291
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 BUG_ON(addr >= end);
1293 tlb_start_vma(tlb, vma);
1294 pgd = pgd_offset(vma->vm_mm, addr);
1295 do {
1296 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001297 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001299 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1300 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 tlb_end_vma(tlb, vma);
1302}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303
Al Virof5cc4ee2012-03-05 14:14:20 -05001304
1305static void unmap_single_vma(struct mmu_gather *tlb,
1306 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001307 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001308 struct zap_details *details)
1309{
1310 unsigned long start = max(vma->vm_start, start_addr);
1311 unsigned long end;
1312
1313 if (start >= vma->vm_end)
1314 return;
1315 end = min(vma->vm_end, end_addr);
1316 if (end <= vma->vm_start)
1317 return;
1318
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301319 if (vma->vm_file)
1320 uprobe_munmap(vma, start, end);
1321
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001322 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001323 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001324
1325 if (start != end) {
1326 if (unlikely(is_vm_hugetlb_page(vma))) {
1327 /*
1328 * It is undesirable to test vma->vm_file as it
1329 * should be non-null for valid hugetlb area.
1330 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001331 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001332 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001333 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001334 * before calling this function to clean up.
1335 * Since no pte has actually been setup, it is
1336 * safe to do nothing in this case.
1337 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001338 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001339 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001340 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001341 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001342 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001343 } else
1344 unmap_page_range(tlb, vma, start, end, details);
1345 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346}
1347
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348/**
1349 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001350 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 * @vma: the starting vma
1352 * @start_addr: virtual address at which to start unmapping
1353 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001355 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 * Only addresses between `start' and `end' will be unmapped.
1358 *
1359 * The VMA list must be sorted in ascending virtual address order.
1360 *
1361 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1362 * range after unmap_vmas() returns. So the only responsibility here is to
1363 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1364 * drops the lock and schedules.
1365 */
Al Viro6e8bb012012-03-05 13:41:15 -05001366void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001368 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001370 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001372 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001373 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001374 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001375 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376}
1377
1378/**
1379 * zap_page_range - remove user pages in a given range
1380 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001381 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001383 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001384 *
1385 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001387void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 unsigned long size, struct zap_details *details)
1389{
1390 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001391 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001392 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001395 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001396 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001397 mmu_notifier_invalidate_range_start(mm, start, end);
1398 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001399 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001400 mmu_notifier_invalidate_range_end(mm, start, end);
1401 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402}
1403
Jack Steinerc627f9c2008-07-29 22:33:53 -07001404/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001405 * zap_page_range_single - remove user pages in a given range
1406 * @vma: vm_area_struct holding the applicable pages
1407 * @address: starting address of pages to zap
1408 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001409 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001410 *
1411 * The range must fit into one VMA.
1412 */
1413static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1414 unsigned long size, struct zap_details *details)
1415{
1416 struct mm_struct *mm = vma->vm_mm;
1417 struct mmu_gather tlb;
1418 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001419
1420 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001421 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001422 update_hiwater_rss(mm);
1423 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001424 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001425 mmu_notifier_invalidate_range_end(mm, address, end);
1426 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427}
1428
Jack Steinerc627f9c2008-07-29 22:33:53 -07001429/**
1430 * zap_vma_ptes - remove ptes mapping the vma
1431 * @vma: vm_area_struct holding ptes to be zapped
1432 * @address: starting address of pages to zap
1433 * @size: number of bytes to zap
1434 *
1435 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1436 *
1437 * The entire address range must be fully contained within the vma.
1438 *
1439 * Returns 0 if successful.
1440 */
1441int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1442 unsigned long size)
1443{
1444 if (address < vma->vm_start || address + size > vma->vm_end ||
1445 !(vma->vm_flags & VM_PFNMAP))
1446 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001447 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001448 return 0;
1449}
1450EXPORT_SYMBOL_GPL(zap_vma_ptes);
1451
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001452pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001453 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001454{
1455 pgd_t * pgd = pgd_offset(mm, addr);
1456 pud_t * pud = pud_alloc(mm, pgd, addr);
1457 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001458 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001459 if (pmd) {
1460 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001461 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001462 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001463 }
1464 return NULL;
1465}
1466
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001468 * This is the old fallback for page remapping.
1469 *
1470 * For historical reasons, it only allows reserved pages. Only
1471 * old drivers should use this, and they needed to mark their
1472 * pages reserved for the old functions anyway.
1473 */
Nick Piggin423bad62008-04-28 02:13:01 -07001474static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1475 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001476{
Nick Piggin423bad62008-04-28 02:13:01 -07001477 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001478 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001479 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001480 spinlock_t *ptl;
1481
Linus Torvalds238f58d2005-11-29 13:01:56 -08001482 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001483 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001484 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001485 retval = -ENOMEM;
1486 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001487 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001488 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001489 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001490 retval = -EBUSY;
1491 if (!pte_none(*pte))
1492 goto out_unlock;
1493
1494 /* Ok, finally just insert the thing.. */
1495 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001496 inc_mm_counter_fast(mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001497 page_add_file_rmap(page, false);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001498 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1499
1500 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001501 pte_unmap_unlock(pte, ptl);
1502 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001503out_unlock:
1504 pte_unmap_unlock(pte, ptl);
1505out:
1506 return retval;
1507}
1508
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001509/**
1510 * vm_insert_page - insert single page into user vma
1511 * @vma: user vma to map to
1512 * @addr: target user address of this page
1513 * @page: source kernel page
1514 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001515 * This allows drivers to insert individual pages they've allocated
1516 * into a user vma.
1517 *
1518 * The page has to be a nice clean _individual_ kernel allocation.
1519 * If you allocate a compound page, you need to have marked it as
1520 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001521 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001522 *
1523 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1524 * took an arbitrary page protection parameter. This doesn't allow
1525 * that. Your vma protection will have to be set up correctly, which
1526 * means that if you want a shared writable mapping, you'd better
1527 * ask for a shared writable mapping!
1528 *
1529 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001530 *
1531 * Usually this function is called from f_op->mmap() handler
1532 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1533 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1534 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001535 */
Nick Piggin423bad62008-04-28 02:13:01 -07001536int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1537 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001538{
1539 if (addr < vma->vm_start || addr >= vma->vm_end)
1540 return -EFAULT;
1541 if (!page_count(page))
1542 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001543 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1544 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1545 BUG_ON(vma->vm_flags & VM_PFNMAP);
1546 vma->vm_flags |= VM_MIXEDMAP;
1547 }
Nick Piggin423bad62008-04-28 02:13:01 -07001548 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001549}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001550EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001551
Nick Piggin423bad62008-04-28 02:13:01 -07001552static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001553 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad62008-04-28 02:13:01 -07001554{
1555 struct mm_struct *mm = vma->vm_mm;
1556 int retval;
1557 pte_t *pte, entry;
1558 spinlock_t *ptl;
1559
1560 retval = -ENOMEM;
1561 pte = get_locked_pte(mm, addr, &ptl);
1562 if (!pte)
1563 goto out;
1564 retval = -EBUSY;
1565 if (!pte_none(*pte))
1566 goto out_unlock;
1567
1568 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001569 if (pfn_t_devmap(pfn))
1570 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1571 else
1572 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad62008-04-28 02:13:01 -07001573 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001574 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001575
1576 retval = 0;
1577out_unlock:
1578 pte_unmap_unlock(pte, ptl);
1579out:
1580 return retval;
1581}
1582
Nick Piggine0dc0d82007-02-12 00:51:36 -08001583/**
1584 * vm_insert_pfn - insert single pfn into user vma
1585 * @vma: user vma to map to
1586 * @addr: target user address of this page
1587 * @pfn: source kernel pfn
1588 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001589 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001590 * they've allocated into a user vma. Same comments apply.
1591 *
1592 * This function should only be called from a vm_ops->fault handler, and
1593 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001594 *
1595 * vma cannot be a COW mapping.
1596 *
1597 * As this is called only for pages that do not currently exist, we
1598 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001599 */
1600int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001601 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001602{
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001603 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1604}
1605EXPORT_SYMBOL(vm_insert_pfn);
1606
1607/**
1608 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1609 * @vma: user vma to map to
1610 * @addr: target user address of this page
1611 * @pfn: source kernel pfn
1612 * @pgprot: pgprot flags for the inserted page
1613 *
1614 * This is exactly like vm_insert_pfn, except that it allows drivers to
1615 * to override pgprot on a per-page basis.
1616 *
1617 * This only makes sense for IO mappings, and it makes no sense for
1618 * cow mappings. In general, using multiple vmas is preferable;
1619 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1620 * impractical.
1621 */
1622int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1623 unsigned long pfn, pgprot_t pgprot)
1624{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001625 int ret;
Nick Piggin7e675132008-04-28 02:13:00 -07001626 /*
1627 * Technically, architectures with pte_special can avoid all these
1628 * restrictions (same for remap_pfn_range). However we would like
1629 * consistency in testing and feature parity among all, so we should
1630 * try to keep these invariants in place for everybody.
1631 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001632 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1633 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1634 (VM_PFNMAP|VM_MIXEDMAP));
1635 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1636 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001637
Nick Piggin423bad62008-04-28 02:13:01 -07001638 if (addr < vma->vm_start || addr >= vma->vm_end)
1639 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001640 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001641 return -EINVAL;
1642
Dan Williams01c8f1c2016-01-15 16:56:40 -08001643 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001644
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001645 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001646}
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001647EXPORT_SYMBOL(vm_insert_pfn_prot);
Nick Piggine0dc0d82007-02-12 00:51:36 -08001648
Nick Piggin423bad62008-04-28 02:13:01 -07001649int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001650 pfn_t pfn)
Nick Piggin423bad62008-04-28 02:13:01 -07001651{
Dan Williams87744ab2016-10-07 17:00:18 -07001652 pgprot_t pgprot = vma->vm_page_prot;
1653
Nick Piggin423bad62008-04-28 02:13:01 -07001654 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1655
1656 if (addr < vma->vm_start || addr >= vma->vm_end)
1657 return -EFAULT;
Dan Williams87744ab2016-10-07 17:00:18 -07001658 if (track_pfn_insert(vma, &pgprot, pfn))
1659 return -EINVAL;
Nick Piggin423bad62008-04-28 02:13:01 -07001660
1661 /*
1662 * If we don't have pte special, then we have to use the pfn_valid()
1663 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1664 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001665 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1666 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001667 */
Dan Williams03fc2da2016-01-26 09:48:05 -08001668 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
Nick Piggin423bad62008-04-28 02:13:01 -07001669 struct page *page;
1670
Dan Williams03fc2da2016-01-26 09:48:05 -08001671 /*
1672 * At this point we are committed to insert_page()
1673 * regardless of whether the caller specified flags that
1674 * result in pfn_t_has_page() == false.
1675 */
1676 page = pfn_to_page(pfn_t_to_pfn(pfn));
Dan Williams87744ab2016-10-07 17:00:18 -07001677 return insert_page(vma, addr, page, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001678 }
Dan Williams87744ab2016-10-07 17:00:18 -07001679 return insert_pfn(vma, addr, pfn, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001680}
1681EXPORT_SYMBOL(vm_insert_mixed);
1682
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001683/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 * maps a range of physical memory into the requested pages. the old
1685 * mappings are removed. any references to nonexistent pages results
1686 * in null mappings (currently treated as "copy-on-access")
1687 */
1688static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1689 unsigned long addr, unsigned long end,
1690 unsigned long pfn, pgprot_t prot)
1691{
1692 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001693 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694
Hugh Dickinsc74df322005-10-29 18:16:23 -07001695 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 if (!pte)
1697 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001698 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 do {
1700 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001701 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 pfn++;
1703 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001704 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001705 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 return 0;
1707}
1708
1709static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1710 unsigned long addr, unsigned long end,
1711 unsigned long pfn, pgprot_t prot)
1712{
1713 pmd_t *pmd;
1714 unsigned long next;
1715
1716 pfn -= addr >> PAGE_SHIFT;
1717 pmd = pmd_alloc(mm, pud, addr);
1718 if (!pmd)
1719 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001720 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 do {
1722 next = pmd_addr_end(addr, end);
1723 if (remap_pte_range(mm, pmd, addr, next,
1724 pfn + (addr >> PAGE_SHIFT), prot))
1725 return -ENOMEM;
1726 } while (pmd++, addr = next, addr != end);
1727 return 0;
1728}
1729
1730static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1731 unsigned long addr, unsigned long end,
1732 unsigned long pfn, pgprot_t prot)
1733{
1734 pud_t *pud;
1735 unsigned long next;
1736
1737 pfn -= addr >> PAGE_SHIFT;
1738 pud = pud_alloc(mm, pgd, addr);
1739 if (!pud)
1740 return -ENOMEM;
1741 do {
1742 next = pud_addr_end(addr, end);
1743 if (remap_pmd_range(mm, pud, addr, next,
1744 pfn + (addr >> PAGE_SHIFT), prot))
1745 return -ENOMEM;
1746 } while (pud++, addr = next, addr != end);
1747 return 0;
1748}
1749
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001750/**
1751 * remap_pfn_range - remap kernel memory to userspace
1752 * @vma: user vma to map to
1753 * @addr: target user address to start at
1754 * @pfn: physical address of kernel memory
1755 * @size: size of map area
1756 * @prot: page protection flags for this mapping
1757 *
1758 * Note: this is only safe if the mm semaphore is held when called.
1759 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1761 unsigned long pfn, unsigned long size, pgprot_t prot)
1762{
1763 pgd_t *pgd;
1764 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001765 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 struct mm_struct *mm = vma->vm_mm;
Yongji Xied5957d22016-05-20 16:57:41 -07001767 unsigned long remap_pfn = pfn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 int err;
1769
1770 /*
1771 * Physically remapped pages are special. Tell the
1772 * rest of the world about it:
1773 * VM_IO tells people not to look at these pages
1774 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001775 * VM_PFNMAP tells the core MM that the base pages are just
1776 * raw PFN mappings, and do not have a "struct page" associated
1777 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001778 * VM_DONTEXPAND
1779 * Disable vma merging and expanding with mremap().
1780 * VM_DONTDUMP
1781 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001782 *
1783 * There's a horrible special case to handle copy-on-write
1784 * behaviour that some programs depend on. We mark the "original"
1785 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001786 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001788 if (is_cow_mapping(vma->vm_flags)) {
1789 if (addr != vma->vm_start || end != vma->vm_end)
1790 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001791 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001792 }
1793
Yongji Xied5957d22016-05-20 16:57:41 -07001794 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001795 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001796 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001797
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001798 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799
1800 BUG_ON(addr >= end);
1801 pfn -= addr >> PAGE_SHIFT;
1802 pgd = pgd_offset(mm, addr);
1803 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804 do {
1805 next = pgd_addr_end(addr, end);
1806 err = remap_pud_range(mm, pgd, addr, next,
1807 pfn + (addr >> PAGE_SHIFT), prot);
1808 if (err)
1809 break;
1810 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001811
1812 if (err)
Yongji Xied5957d22016-05-20 16:57:41 -07001813 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001814
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 return err;
1816}
1817EXPORT_SYMBOL(remap_pfn_range);
1818
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001819/**
1820 * vm_iomap_memory - remap memory to userspace
1821 * @vma: user vma to map to
1822 * @start: start of area
1823 * @len: size of area
1824 *
1825 * This is a simplified io_remap_pfn_range() for common driver use. The
1826 * driver just needs to give us the physical memory range to be mapped,
1827 * we'll figure out the rest from the vma information.
1828 *
1829 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1830 * whatever write-combining details or similar.
1831 */
1832int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1833{
1834 unsigned long vm_len, pfn, pages;
1835
1836 /* Check that the physical memory area passed in looks valid */
1837 if (start + len < start)
1838 return -EINVAL;
1839 /*
1840 * You *really* shouldn't map things that aren't page-aligned,
1841 * but we've historically allowed it because IO memory might
1842 * just have smaller alignment.
1843 */
1844 len += start & ~PAGE_MASK;
1845 pfn = start >> PAGE_SHIFT;
1846 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1847 if (pfn + pages < pfn)
1848 return -EINVAL;
1849
1850 /* We start the mapping 'vm_pgoff' pages into the area */
1851 if (vma->vm_pgoff > pages)
1852 return -EINVAL;
1853 pfn += vma->vm_pgoff;
1854 pages -= vma->vm_pgoff;
1855
1856 /* Can we fit all of the mapping? */
1857 vm_len = vma->vm_end - vma->vm_start;
1858 if (vm_len >> PAGE_SHIFT > pages)
1859 return -EINVAL;
1860
1861 /* Ok, let it rip */
1862 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1863}
1864EXPORT_SYMBOL(vm_iomap_memory);
1865
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001866static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1867 unsigned long addr, unsigned long end,
1868 pte_fn_t fn, void *data)
1869{
1870 pte_t *pte;
1871 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001872 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001873 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001874
1875 pte = (mm == &init_mm) ?
1876 pte_alloc_kernel(pmd, addr) :
1877 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1878 if (!pte)
1879 return -ENOMEM;
1880
1881 BUG_ON(pmd_huge(*pmd));
1882
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001883 arch_enter_lazy_mmu_mode();
1884
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001885 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001886
1887 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001888 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001889 if (err)
1890 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001891 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001892
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001893 arch_leave_lazy_mmu_mode();
1894
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001895 if (mm != &init_mm)
1896 pte_unmap_unlock(pte-1, ptl);
1897 return err;
1898}
1899
1900static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1901 unsigned long addr, unsigned long end,
1902 pte_fn_t fn, void *data)
1903{
1904 pmd_t *pmd;
1905 unsigned long next;
1906 int err;
1907
Andi Kleenceb86872008-07-23 21:27:50 -07001908 BUG_ON(pud_huge(*pud));
1909
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001910 pmd = pmd_alloc(mm, pud, addr);
1911 if (!pmd)
1912 return -ENOMEM;
1913 do {
1914 next = pmd_addr_end(addr, end);
1915 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1916 if (err)
1917 break;
1918 } while (pmd++, addr = next, addr != end);
1919 return err;
1920}
1921
1922static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1923 unsigned long addr, unsigned long end,
1924 pte_fn_t fn, void *data)
1925{
1926 pud_t *pud;
1927 unsigned long next;
1928 int err;
1929
1930 pud = pud_alloc(mm, pgd, addr);
1931 if (!pud)
1932 return -ENOMEM;
1933 do {
1934 next = pud_addr_end(addr, end);
1935 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1936 if (err)
1937 break;
1938 } while (pud++, addr = next, addr != end);
1939 return err;
1940}
1941
1942/*
1943 * Scan a region of virtual memory, filling in page tables as necessary
1944 * and calling a provided function on each leaf page table.
1945 */
1946int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1947 unsigned long size, pte_fn_t fn, void *data)
1948{
1949 pgd_t *pgd;
1950 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001951 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001952 int err;
1953
Mika Penttilä9cb65bc2016-03-15 14:56:45 -07001954 if (WARN_ON(addr >= end))
1955 return -EINVAL;
1956
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001957 pgd = pgd_offset(mm, addr);
1958 do {
1959 next = pgd_addr_end(addr, end);
1960 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1961 if (err)
1962 break;
1963 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001964
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001965 return err;
1966}
1967EXPORT_SYMBOL_GPL(apply_to_page_range);
1968
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001970 * handle_pte_fault chooses page fault handler according to an entry which was
1971 * read non-atomically. Before making any commitment, on those architectures
1972 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1973 * parts, do_swap_page must check under lock before unmapping the pte and
1974 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001975 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001976 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001977static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001978 pte_t *page_table, pte_t orig_pte)
1979{
1980 int same = 1;
1981#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1982 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001983 spinlock_t *ptl = pte_lockptr(mm, pmd);
1984 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001985 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001986 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001987 }
1988#endif
1989 pte_unmap(page_table);
1990 return same;
1991}
1992
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001993static 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 -08001994{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001995 debug_dma_assert_idle(src);
1996
Linus Torvalds6aab3412005-11-28 14:34:23 -08001997 /*
1998 * If the source page was a PFN mapping, we don't have
1999 * a "struct page" for it. We do a best-effort copy by
2000 * just copying from the original user address. If that
2001 * fails, we just zero-fill it. Live with it.
2002 */
2003 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08002004 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002005 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2006
2007 /*
2008 * This really shouldn't fail, because the page is there
2009 * in the page tables. But it might just be unreadable,
2010 * in which case we just give up and fill the result with
2011 * zeroes.
2012 */
2013 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002014 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002015 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002016 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002017 } else
2018 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002019}
2020
Michal Hockoc20cd452016-01-14 15:20:12 -08002021static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2022{
2023 struct file *vm_file = vma->vm_file;
2024
2025 if (vm_file)
2026 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2027
2028 /*
2029 * Special mappings (e.g. VDSO) do not have any file so fake
2030 * a default GFP_KERNEL for them.
2031 */
2032 return GFP_KERNEL;
2033}
2034
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002036 * Notify the address space that the page is about to become writable so that
2037 * it can prohibit this or wait for the page to get into an appropriate state.
2038 *
2039 * We do this without the lock held, so that it can sleep if it needs to.
2040 */
2041static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2042 unsigned long address)
2043{
2044 struct vm_fault vmf;
2045 int ret;
2046
2047 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2048 vmf.pgoff = page->index;
2049 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08002050 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002051 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002052 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002053
2054 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2055 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2056 return ret;
2057 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2058 lock_page(page);
2059 if (!page->mapping) {
2060 unlock_page(page);
2061 return 0; /* retry */
2062 }
2063 ret |= VM_FAULT_LOCKED;
2064 } else
2065 VM_BUG_ON_PAGE(!PageLocked(page), page);
2066 return ret;
2067}
2068
2069/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07002070 * Handle write page faults for pages that can be reused in the current vma
2071 *
2072 * This can happen either due to the mapping being with the VM_SHARED flag,
2073 * or due to us being the last reference standing to the page. In either
2074 * case, all we need to do here is to mark the page as writable and update
2075 * any related book-keeping.
2076 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002077static inline int wp_page_reuse(struct fault_env *fe, pte_t orig_pte,
2078 struct page *page, int page_mkwrite, int dirty_shared)
2079 __releases(fe->ptl)
Shachar Raindel4e047f82015-04-14 15:46:25 -07002080{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002081 struct vm_area_struct *vma = fe->vma;
Shachar Raindel4e047f82015-04-14 15:46:25 -07002082 pte_t entry;
2083 /*
2084 * Clear the pages cpupid information as the existing
2085 * information potentially belongs to a now completely
2086 * unrelated process.
2087 */
2088 if (page)
2089 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2090
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002091 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel4e047f82015-04-14 15:46:25 -07002092 entry = pte_mkyoung(orig_pte);
2093 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002094 if (ptep_set_access_flags(vma, fe->address, fe->pte, entry, 1))
2095 update_mmu_cache(vma, fe->address, fe->pte);
2096 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002097
2098 if (dirty_shared) {
2099 struct address_space *mapping;
2100 int dirtied;
2101
2102 if (!page_mkwrite)
2103 lock_page(page);
2104
2105 dirtied = set_page_dirty(page);
2106 VM_BUG_ON_PAGE(PageAnon(page), page);
2107 mapping = page->mapping;
2108 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002109 put_page(page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002110
2111 if ((dirtied || page_mkwrite) && mapping) {
2112 /*
2113 * Some device drivers do not set page.mapping
2114 * but still dirty their pages
2115 */
2116 balance_dirty_pages_ratelimited(mapping);
2117 }
2118
2119 if (!page_mkwrite)
2120 file_update_time(vma->vm_file);
2121 }
2122
2123 return VM_FAULT_WRITE;
2124}
2125
2126/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002127 * Handle the case of a page which we actually need to copy to a new page.
2128 *
2129 * Called with mmap_sem locked and the old page referenced, but
2130 * without the ptl held.
2131 *
2132 * High level logic flow:
2133 *
2134 * - Allocate a page, copy the content of the old page to the new one.
2135 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2136 * - Take the PTL. If the pte changed, bail out and release the allocated page
2137 * - If the pte is still the way we remember it, update the page table and all
2138 * relevant references. This includes dropping the reference the page-table
2139 * held to the old page, as well as updating the rmap.
2140 * - In any case, unlock the PTL and drop the reference we took to the old page.
2141 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002142static int wp_page_copy(struct fault_env *fe, pte_t orig_pte,
2143 struct page *old_page)
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002144{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002145 struct vm_area_struct *vma = fe->vma;
2146 struct mm_struct *mm = vma->vm_mm;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002147 struct page *new_page = NULL;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002148 pte_t entry;
2149 int page_copied = 0;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002150 const unsigned long mmun_start = fe->address & PAGE_MASK;
2151 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002152 struct mem_cgroup *memcg;
2153
2154 if (unlikely(anon_vma_prepare(vma)))
2155 goto oom;
2156
2157 if (is_zero_pfn(pte_pfn(orig_pte))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002158 new_page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002159 if (!new_page)
2160 goto oom;
2161 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002162 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2163 fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002164 if (!new_page)
2165 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002166 cow_user_page(new_page, old_page, fe->address, vma);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002167 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002168
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002169 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002170 goto oom_free_new;
2171
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002172 __SetPageUptodate(new_page);
2173
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002174 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2175
2176 /*
2177 * Re-check the pte - we dropped the lock
2178 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002179 fe->pte = pte_offset_map_lock(mm, fe->pmd, fe->address, &fe->ptl);
2180 if (likely(pte_same(*fe->pte, orig_pte))) {
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002181 if (old_page) {
2182 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002183 dec_mm_counter_fast(mm,
2184 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002185 inc_mm_counter_fast(mm, MM_ANONPAGES);
2186 }
2187 } else {
2188 inc_mm_counter_fast(mm, MM_ANONPAGES);
2189 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002190 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002191 entry = mk_pte(new_page, vma->vm_page_prot);
2192 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2193 /*
2194 * Clear the pte entry and flush it first, before updating the
2195 * pte with the new entry. This will avoid a race condition
2196 * seen in the presence of one thread doing SMC and another
2197 * thread doing COW.
2198 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002199 ptep_clear_flush_notify(vma, fe->address, fe->pte);
2200 page_add_new_anon_rmap(new_page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002201 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002202 lru_cache_add_active_or_unevictable(new_page, vma);
2203 /*
2204 * We call the notify macro here because, when using secondary
2205 * mmu page tables (such as kvm shadow page tables), we want the
2206 * new page to be mapped directly into the secondary page table.
2207 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002208 set_pte_at_notify(mm, fe->address, fe->pte, entry);
2209 update_mmu_cache(vma, fe->address, fe->pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002210 if (old_page) {
2211 /*
2212 * Only after switching the pte to the new page may
2213 * we remove the mapcount here. Otherwise another
2214 * process may come and find the rmap count decremented
2215 * before the pte is switched to the new page, and
2216 * "reuse" the old page writing into it while our pte
2217 * here still points into it and can be read by other
2218 * threads.
2219 *
2220 * The critical issue is to order this
2221 * page_remove_rmap with the ptp_clear_flush above.
2222 * Those stores are ordered by (if nothing else,)
2223 * the barrier present in the atomic_add_negative
2224 * in page_remove_rmap.
2225 *
2226 * Then the TLB flush in ptep_clear_flush ensures that
2227 * no process can access the old page before the
2228 * decremented mapcount is visible. And the old page
2229 * cannot be reused until after the decremented
2230 * mapcount is visible. So transitively, TLBs to
2231 * old page will be flushed before it can be reused.
2232 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002233 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002234 }
2235
2236 /* Free the old page.. */
2237 new_page = old_page;
2238 page_copied = 1;
2239 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002240 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002241 }
2242
2243 if (new_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002244 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002245
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002246 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002247 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2248 if (old_page) {
2249 /*
2250 * Don't let another task, with possibly unlocked vma,
2251 * keep the mlocked page.
2252 */
2253 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2254 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002255 if (PageMlocked(old_page))
2256 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002257 unlock_page(old_page);
2258 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002259 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002260 }
2261 return page_copied ? VM_FAULT_WRITE : 0;
2262oom_free_new:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002263 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002264oom:
2265 if (old_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002266 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002267 return VM_FAULT_OOM;
2268}
2269
Boaz Harroshdd906182015-04-15 16:15:11 -07002270/*
2271 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2272 * mapping
2273 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002274static int wp_pfn_shared(struct fault_env *fe, pte_t orig_pte)
Boaz Harroshdd906182015-04-15 16:15:11 -07002275{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002276 struct vm_area_struct *vma = fe->vma;
2277
Boaz Harroshdd906182015-04-15 16:15:11 -07002278 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2279 struct vm_fault vmf = {
2280 .page = NULL,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002281 .pgoff = linear_page_index(vma, fe->address),
2282 .virtual_address =
2283 (void __user *)(fe->address & PAGE_MASK),
Boaz Harroshdd906182015-04-15 16:15:11 -07002284 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2285 };
2286 int ret;
2287
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002288 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002289 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2290 if (ret & VM_FAULT_ERROR)
2291 return ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002292 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2293 &fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002294 /*
2295 * We might have raced with another page fault while we
2296 * released the pte_offset_map_lock.
2297 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002298 if (!pte_same(*fe->pte, orig_pte)) {
2299 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002300 return 0;
2301 }
2302 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002303 return wp_page_reuse(fe, orig_pte, NULL, 0, 0);
Boaz Harroshdd906182015-04-15 16:15:11 -07002304}
2305
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002306static int wp_page_shared(struct fault_env *fe, pte_t orig_pte,
2307 struct page *old_page)
2308 __releases(fe->ptl)
Shachar Raindel93e478d2015-04-14 15:46:35 -07002309{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002310 struct vm_area_struct *vma = fe->vma;
Shachar Raindel93e478d2015-04-14 15:46:35 -07002311 int page_mkwrite = 0;
2312
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002313 get_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002314
Shachar Raindel93e478d2015-04-14 15:46:35 -07002315 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2316 int tmp;
2317
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002318 pte_unmap_unlock(fe->pte, fe->ptl);
2319 tmp = do_page_mkwrite(vma, old_page, fe->address);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002320 if (unlikely(!tmp || (tmp &
2321 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002322 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002323 return tmp;
2324 }
2325 /*
2326 * Since we dropped the lock we need to revalidate
2327 * the PTE as someone else may have changed it. If
2328 * they did, we just return, as we can count on the
2329 * MMU to tell us if they didn't also make it writable.
2330 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002331 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2332 &fe->ptl);
2333 if (!pte_same(*fe->pte, orig_pte)) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002334 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002335 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002336 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002337 return 0;
2338 }
2339 page_mkwrite = 1;
2340 }
2341
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002342 return wp_page_reuse(fe, orig_pte, old_page, page_mkwrite, 1);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002343}
2344
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002345/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 * This routine handles present pages, when users try to write
2347 * to a shared page. It is done by copying the page to a new address
2348 * and decrementing the shared-page counter for the old page.
2349 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 * Note that this routine assumes that the protection checks have been
2351 * done by the caller (the low-level page fault routine in most cases).
2352 * Thus we can safely just mark it writable once we've done any necessary
2353 * COW.
2354 *
2355 * We also mark the page dirty at this point even though the page will
2356 * change only once the write actually happens. This avoids a few races,
2357 * and potentially makes it more efficient.
2358 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002359 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2360 * but allow concurrent faults), with pte both mapped and locked.
2361 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002363static int do_wp_page(struct fault_env *fe, pte_t orig_pte)
2364 __releases(fe->ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002366 struct vm_area_struct *vma = fe->vma;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002367 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002369 old_page = vm_normal_page(vma, fe->address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002370 if (!old_page) {
2371 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002372 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2373 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002374 *
2375 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002376 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002377 */
2378 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2379 (VM_WRITE|VM_SHARED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002380 return wp_pfn_shared(fe, orig_pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002381
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002382 pte_unmap_unlock(fe->pte, fe->ptl);
2383 return wp_page_copy(fe, orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002386 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002387 * Take out anonymous pages first, anonymous shared vmas are
2388 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002389 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002390 if (PageAnon(old_page) && !PageKsm(old_page)) {
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002391 int total_mapcount;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002392 if (!trylock_page(old_page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002393 get_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002394 pte_unmap_unlock(fe->pte, fe->ptl);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002395 lock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002396 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2397 fe->address, &fe->ptl);
2398 if (!pte_same(*fe->pte, orig_pte)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002399 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002400 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002401 put_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002402 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002403 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002404 put_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002405 }
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002406 if (reuse_swap_page(old_page, &total_mapcount)) {
2407 if (total_mapcount == 1) {
2408 /*
2409 * The page is all ours. Move it to
2410 * our anon_vma so the rmap code will
2411 * not search our parent or siblings.
2412 * Protected against the rmap code by
2413 * the page lock.
2414 */
Hugh Dickins5a499732016-07-14 12:07:38 -07002415 page_move_anon_rmap(old_page, vma);
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002416 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002417 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002418 return wp_page_reuse(fe, orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002419 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002420 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002421 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002422 (VM_WRITE|VM_SHARED))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002423 return wp_page_shared(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425
2426 /*
2427 * Ok, we need to copy. Oh, well..
2428 */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002429 get_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002430
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002431 pte_unmap_unlock(fe->pte, fe->ptl);
2432 return wp_page_copy(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433}
2434
Peter Zijlstra97a89412011-05-24 17:12:04 -07002435static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 unsigned long start_addr, unsigned long end_addr,
2437 struct zap_details *details)
2438{
Al Virof5cc4ee2012-03-05 14:14:20 -05002439 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440}
2441
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002442static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 struct zap_details *details)
2444{
2445 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 pgoff_t vba, vea, zba, zea;
2447
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002448 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450
2451 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002452 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 zba = details->first_index;
2454 if (zba < vba)
2455 zba = vba;
2456 zea = details->last_index;
2457 if (zea > vea)
2458 zea = vea;
2459
Peter Zijlstra97a89412011-05-24 17:12:04 -07002460 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2462 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002463 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464 }
2465}
2466
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002468 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2469 * address_space corresponding to the specified page range in the underlying
2470 * file.
2471 *
Martin Waitz3d410882005-06-23 22:05:21 -07002472 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 * @holebegin: byte in first page to unmap, relative to the start of
2474 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002475 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 * must keep the partial page. In contrast, we must get rid of
2477 * partial pages.
2478 * @holelen: size of prospective hole in bytes. This will be rounded
2479 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2480 * end of the file.
2481 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2482 * but 0 when invalidating pagecache, don't throw away private data.
2483 */
2484void unmap_mapping_range(struct address_space *mapping,
2485 loff_t const holebegin, loff_t const holelen, int even_cows)
2486{
Michal Hockoaac45362016-03-25 14:20:24 -07002487 struct zap_details details = { };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 pgoff_t hba = holebegin >> PAGE_SHIFT;
2489 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2490
2491 /* Check for overflow. */
2492 if (sizeof(holelen) > sizeof(hlen)) {
2493 long long holeend =
2494 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2495 if (holeend & ~(long long)ULONG_MAX)
2496 hlen = ULONG_MAX - hba + 1;
2497 }
2498
2499 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 details.first_index = hba;
2501 details.last_index = hba + hlen - 1;
2502 if (details.last_index < details.first_index)
2503 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002505 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002506 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002508 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509}
2510EXPORT_SYMBOL(unmap_mapping_range);
2511
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002513 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2514 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002515 * We return with pte unmapped and unlocked.
2516 *
2517 * We return with the mmap_sem locked or unlocked in the same cases
2518 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002520int do_swap_page(struct fault_env *fe, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002522 struct vm_area_struct *vma = fe->vma;
Hugh Dickins56f31802013-02-22 16:36:10 -08002523 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002524 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002525 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002527 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002528 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002529 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002531 if (!pte_unmap_same(vma->vm_mm, fe->pmd, fe->pte, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002532 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002533
2534 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002535 if (unlikely(non_swap_entry(entry))) {
2536 if (is_migration_entry(entry)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002537 migration_entry_wait(vma->vm_mm, fe->pmd, fe->address);
Andi Kleend1737fd2009-09-16 11:50:06 +02002538 } else if (is_hwpoison_entry(entry)) {
2539 ret = VM_FAULT_HWPOISON;
2540 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002541 print_bad_pte(vma, fe->address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002542 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002543 }
Christoph Lameter06972122006-06-23 02:03:35 -07002544 goto out;
2545 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002546 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 page = lookup_swap_cache(entry);
2548 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002549 page = swapin_readahead(entry,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002550 GFP_HIGHUSER_MOVABLE, vma, fe->address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 if (!page) {
2552 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002553 * Back out if somebody else faulted in this pte
2554 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002556 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2557 fe->address, &fe->ptl);
2558 if (likely(pte_same(*fe->pte, orig_pte)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002560 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002561 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 }
2563
2564 /* Had to read the page from swap area: Major fault */
2565 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002566 count_vm_event(PGMAJFAULT);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002567 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002568 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002569 /*
2570 * hwpoisoned dirty swapcache pages are kept for killing
2571 * owner processes (which may be unknown at hwpoison time)
2572 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002573 ret = VM_FAULT_HWPOISON;
2574 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002575 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002576 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 }
2578
Hugh Dickins56f31802013-02-22 16:36:10 -08002579 swapcache = page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002580 locked = lock_page_or_retry(page, vma->vm_mm, fe->flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002581
Balbir Singh20a10222007-11-14 17:00:33 -08002582 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002583 if (!locked) {
2584 ret |= VM_FAULT_RETRY;
2585 goto out_release;
2586 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002588 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002589 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2590 * release the swapcache from under us. The page pin, and pte_same
2591 * test below, are not enough to exclude that. Even if it is still
2592 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002593 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002594 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002595 goto out_page;
2596
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002597 page = ksm_might_need_to_copy(page, vma, fe->address);
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002598 if (unlikely(!page)) {
2599 ret = VM_FAULT_OOM;
2600 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002601 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002602 }
2603
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002604 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2605 &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002606 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002607 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002608 }
2609
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002611 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002613 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2614 &fe->ptl);
2615 if (unlikely(!pte_same(*fe->pte, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002616 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002617
2618 if (unlikely(!PageUptodate(page))) {
2619 ret = VM_FAULT_SIGBUS;
2620 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621 }
2622
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002623 /*
2624 * The page isn't present yet, go ahead with the fault.
2625 *
2626 * Be careful about the sequence of operations here.
2627 * To get its accounting right, reuse_swap_page() must be called
2628 * while the page is counted on swap but not yet in mapcount i.e.
2629 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2630 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002631 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002633 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2634 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 pte = mk_pte(page, vma->vm_page_prot);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002636 if ((fe->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002638 fe->flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002639 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002640 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002643 if (pte_swp_soft_dirty(orig_pte))
2644 pte = pte_mksoft_dirty(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002645 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002646 if (page == swapcache) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002647 do_page_add_anon_rmap(page, vma, fe->address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002648 mem_cgroup_commit_charge(page, memcg, true, false);
Minchan Kim1a8018f2016-08-02 14:02:25 -07002649 activate_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -07002650 } else { /* ksm created a completely new copy */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002651 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002652 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002653 lru_cache_add_active_or_unevictable(page, vma);
2654 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002655
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002656 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002657 if (mem_cgroup_swap_full(page) ||
2658 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002659 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002660 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002661 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002662 /*
2663 * Hold the lock to avoid the swap entry to be reused
2664 * until we take the PT lock for the pte_same() check
2665 * (to avoid false positives from pte_same). For
2666 * further safety release the lock after the swap_free
2667 * so that the swap count won't change under a
2668 * parallel locked swapcache.
2669 */
2670 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002671 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002672 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002673
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002674 if (fe->flags & FAULT_FLAG_WRITE) {
2675 ret |= do_wp_page(fe, pte);
Hugh Dickins61469f12008-03-04 14:29:04 -08002676 if (ret & VM_FAULT_ERROR)
2677 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 goto out;
2679 }
2680
2681 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002682 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002683unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002684 pte_unmap_unlock(fe->pte, fe->ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002685out:
2686 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002687out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002688 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002689 pte_unmap_unlock(fe->pte, fe->ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002690out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002691 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002692out_release:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002693 put_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002694 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002695 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002696 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002697 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002698 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699}
2700
2701/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002702 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2703 * but allow concurrent faults), and pte mapped but not yet locked.
2704 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002706static int do_anonymous_page(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002708 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -07002709 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002710 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002713 /* File mapping without ->vm_ops ? */
2714 if (vma->vm_flags & VM_SHARED)
2715 return VM_FAULT_SIGBUS;
2716
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002717 /*
2718 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2719 * pte_offset_map() on pmds where a huge pmd might be created
2720 * from a different thread.
2721 *
2722 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2723 * parallel threads are excluded by other means.
2724 *
2725 * Here we only have down_read(mmap_sem).
2726 */
2727 if (pte_alloc(vma->vm_mm, fe->pmd, fe->address))
2728 return VM_FAULT_OOM;
2729
2730 /* See the comment in pte_alloc_one_map() */
2731 if (unlikely(pmd_trans_unstable(fe->pmd)))
2732 return 0;
2733
Linus Torvalds11ac5522010-08-14 11:44:56 -07002734 /* Use the zero-page for reads */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002735 if (!(fe->flags & FAULT_FLAG_WRITE) &&
2736 !mm_forbids_zeropage(vma->vm_mm)) {
2737 entry = pte_mkspecial(pfn_pte(my_zero_pfn(fe->address),
Hugh Dickins62eede62009-09-21 17:03:34 -07002738 vma->vm_page_prot));
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002739 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2740 &fe->ptl);
2741 if (!pte_none(*fe->pte))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002742 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002743 /* Deliver the page fault to userland, check inside PT lock */
2744 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002745 pte_unmap_unlock(fe->pte, fe->ptl);
2746 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002747 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002748 goto setpte;
2749 }
2750
Nick Piggin557ed1f2007-10-16 01:24:40 -07002751 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002752 if (unlikely(anon_vma_prepare(vma)))
2753 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002754 page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002755 if (!page)
2756 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002757
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002758 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002759 goto oom_free_page;
2760
Minchan Kim52f37622013-04-29 15:08:15 -07002761 /*
2762 * The memory barrier inside __SetPageUptodate makes sure that
2763 * preceeding stores to the page contents become visible before
2764 * the set_pte_at() write.
2765 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002766 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
Nick Piggin557ed1f2007-10-16 01:24:40 -07002768 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002769 if (vma->vm_flags & VM_WRITE)
2770 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002771
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002772 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2773 &fe->ptl);
2774 if (!pte_none(*fe->pte))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002775 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002776
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002777 /* Deliver the page fault to userland, check inside PT lock */
2778 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002779 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002780 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002781 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002782 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002783 }
2784
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002785 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2786 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002787 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002788 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002789setpte:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002790 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
2792 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002793 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002794unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002795 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002796 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002797release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002798 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002799 put_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002800 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002801oom_free_page:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002802 put_page(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002803oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 return VM_FAULT_OOM;
2805}
2806
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002807/*
2808 * The mmap_sem must have been held on entry, and may have been
2809 * released depending on flags and vma->vm_ops->fault() return value.
2810 * See filemap_fault() and __lock_page_retry().
2811 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002812static int __do_fault(struct fault_env *fe, pgoff_t pgoff,
2813 struct page *cow_page, struct page **page, void **entry)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002814{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002815 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002816 struct vm_fault vmf;
2817 int ret;
2818
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002819 vmf.virtual_address = (void __user *)(fe->address & PAGE_MASK);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002820 vmf.pgoff = pgoff;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002821 vmf.flags = fe->flags;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002822 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002823 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002824 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002825
2826 ret = vma->vm_ops->fault(vma, &vmf);
2827 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2828 return ret;
Jan Karabc2466e2016-05-12 18:29:19 +02002829 if (ret & VM_FAULT_DAX_LOCKED) {
2830 *entry = vmf.entry;
2831 return ret;
2832 }
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002833
2834 if (unlikely(PageHWPoison(vmf.page))) {
2835 if (ret & VM_FAULT_LOCKED)
2836 unlock_page(vmf.page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002837 put_page(vmf.page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002838 return VM_FAULT_HWPOISON;
2839 }
2840
2841 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2842 lock_page(vmf.page);
2843 else
2844 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2845
2846 *page = vmf.page;
2847 return ret;
2848}
2849
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002850static int pte_alloc_one_map(struct fault_env *fe)
2851{
2852 struct vm_area_struct *vma = fe->vma;
2853
2854 if (!pmd_none(*fe->pmd))
2855 goto map_pte;
2856 if (fe->prealloc_pte) {
2857 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2858 if (unlikely(!pmd_none(*fe->pmd))) {
2859 spin_unlock(fe->ptl);
2860 goto map_pte;
2861 }
2862
2863 atomic_long_inc(&vma->vm_mm->nr_ptes);
2864 pmd_populate(vma->vm_mm, fe->pmd, fe->prealloc_pte);
2865 spin_unlock(fe->ptl);
2866 fe->prealloc_pte = 0;
2867 } else if (unlikely(pte_alloc(vma->vm_mm, fe->pmd, fe->address))) {
2868 return VM_FAULT_OOM;
2869 }
2870map_pte:
2871 /*
2872 * If a huge pmd materialized under us just retry later. Use
2873 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2874 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2875 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2876 * in a different thread of this mm, in turn leading to a misleading
2877 * pmd_trans_huge() retval. All we have to ensure is that it is a
2878 * regular pmd that we can walk with pte_offset_map() and we can do that
2879 * through an atomic read in C, which is what pmd_trans_unstable()
2880 * provides.
2881 */
2882 if (pmd_trans_unstable(fe->pmd) || pmd_devmap(*fe->pmd))
2883 return VM_FAULT_NOPAGE;
2884
2885 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2886 &fe->ptl);
2887 return 0;
2888}
2889
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002890#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002891
2892#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2893static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2894 unsigned long haddr)
2895{
2896 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2897 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2898 return false;
2899 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2900 return false;
2901 return true;
2902}
2903
2904static int do_set_pmd(struct fault_env *fe, struct page *page)
2905{
2906 struct vm_area_struct *vma = fe->vma;
2907 bool write = fe->flags & FAULT_FLAG_WRITE;
2908 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
2909 pmd_t entry;
2910 int i, ret;
2911
2912 if (!transhuge_vma_suitable(vma, haddr))
2913 return VM_FAULT_FALLBACK;
2914
2915 ret = VM_FAULT_FALLBACK;
2916 page = compound_head(page);
2917
2918 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2919 if (unlikely(!pmd_none(*fe->pmd)))
2920 goto out;
2921
2922 for (i = 0; i < HPAGE_PMD_NR; i++)
2923 flush_icache_page(vma, page + i);
2924
2925 entry = mk_huge_pmd(page, vma->vm_page_prot);
2926 if (write)
2927 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2928
2929 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2930 page_add_file_rmap(page, true);
2931
2932 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
2933
2934 update_mmu_cache_pmd(vma, haddr, fe->pmd);
2935
2936 /* fault is handled */
2937 ret = 0;
Kirill A. Shutemov95ecedc2016-07-26 15:25:31 -07002938 count_vm_event(THP_FILE_MAPPED);
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002939out:
2940 spin_unlock(fe->ptl);
2941 return ret;
2942}
2943#else
2944static int do_set_pmd(struct fault_env *fe, struct page *page)
2945{
2946 BUILD_BUG();
2947 return 0;
2948}
2949#endif
2950
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002951/**
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002952 * alloc_set_pte - setup new PTE entry for given page and add reverse page
2953 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002954 *
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002955 * @fe: fault environment
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002956 * @memcg: memcg to charge page (only for private mappings)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002957 * @page: page to map
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002958 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002959 * Caller must take care of unlocking fe->ptl, if fe->pte is non-NULL on return.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002960 *
2961 * Target users are page handler itself and implementations of
2962 * vm_ops->map_pages.
2963 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002964int alloc_set_pte(struct fault_env *fe, struct mem_cgroup *memcg,
2965 struct page *page)
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002966{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002967 struct vm_area_struct *vma = fe->vma;
2968 bool write = fe->flags & FAULT_FLAG_WRITE;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002969 pte_t entry;
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002970 int ret;
2971
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002972 if (pmd_none(*fe->pmd) && PageTransCompound(page) &&
2973 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002974 /* THP on COW? */
2975 VM_BUG_ON_PAGE(memcg, page);
2976
2977 ret = do_set_pmd(fe, page);
2978 if (ret != VM_FAULT_FALLBACK)
2979 return ret;
2980 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002981
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002982 if (!fe->pte) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002983 ret = pte_alloc_one_map(fe);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002984 if (ret)
2985 return ret;
2986 }
2987
2988 /* Re-check under ptl */
2989 if (unlikely(!pte_none(*fe->pte)))
2990 return VM_FAULT_NOPAGE;
2991
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002992 flush_icache_page(vma, page);
2993 entry = mk_pte(page, vma->vm_page_prot);
2994 if (write)
2995 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002996 /* copy-on-write page */
2997 if (write && !(vma->vm_flags & VM_SHARED)) {
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002998 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002999 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003000 mem_cgroup_commit_charge(page, memcg, false, false);
3001 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003002 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08003003 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07003004 page_add_file_rmap(page, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003005 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003006 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003007
3008 /* no need to invalidate: a not-present page won't be cached */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003009 update_mmu_cache(vma, fe->address, fe->pte);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003010
3011 return 0;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003012}
3013
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07003014static unsigned long fault_around_bytes __read_mostly =
Vinayak Menon23f8a2f2015-12-17 08:35:36 +05303015 rounddown_pow_of_two(4096);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003016
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003017#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003018static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003019{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003020 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003021 return 0;
3022}
3023
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003024/*
3025 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3026 * rounded down to nearest page order. It's what do_fault_around() expects to
3027 * see.
3028 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003029static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003030{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003031 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003032 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003033 if (val > PAGE_SIZE)
3034 fault_around_bytes = rounddown_pow_of_two(val);
3035 else
3036 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003037 return 0;
3038}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003039DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3040 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003041
3042static int __init fault_around_debugfs(void)
3043{
3044 void *ret;
3045
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003046 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3047 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003048 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003049 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003050 return 0;
3051}
3052late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003053#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003054
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07003055/*
3056 * do_fault_around() tries to map few pages around the fault address. The hope
3057 * is that the pages will be needed soon and this will lower the number of
3058 * faults to handle.
3059 *
3060 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3061 * not ready to be mapped: not up-to-date, locked, etc.
3062 *
3063 * This function is called with the page table lock taken. In the split ptlock
3064 * case the page table lock only protects only those entries which belong to
3065 * the page table corresponding to the fault address.
3066 *
3067 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3068 * only once.
3069 *
3070 * fault_around_pages() defines how many pages we'll try to map.
3071 * do_fault_around() expects it to return a power of two less than or equal to
3072 * PTRS_PER_PTE.
3073 *
3074 * The virtual address of the area that we map is naturally aligned to the
3075 * fault_around_pages() value (and therefore to page order). This way it's
3076 * easier to guarantee that we don't cross page table boundaries.
3077 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003078static int do_fault_around(struct fault_env *fe, pgoff_t start_pgoff)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003079{
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003080 unsigned long address = fe->address, nr_pages, mask;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003081 pgoff_t end_pgoff;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003082 int off, ret = 0;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003083
Jason Low4db0c3c2015-04-15 16:14:08 -07003084 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07003085 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3086
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003087 fe->address = max(address & mask, fe->vma->vm_start);
3088 off = ((address - fe->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003089 start_pgoff -= off;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003090
3091 /*
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003092 * end_pgoff is either end of page table or end of vma
3093 * or fault_around_pages() from start_pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003094 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003095 end_pgoff = start_pgoff -
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003096 ((fe->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003097 PTRS_PER_PTE - 1;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003098 end_pgoff = min3(end_pgoff, vma_pages(fe->vma) + fe->vma->vm_pgoff - 1,
3099 start_pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003100
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003101 if (pmd_none(*fe->pmd)) {
3102 fe->prealloc_pte = pte_alloc_one(fe->vma->vm_mm, fe->address);
Vegard Nossumc5f88bd2016-08-02 14:02:22 -07003103 if (!fe->prealloc_pte)
3104 goto out;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003105 smp_wmb(); /* See comment in __pte_alloc() */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003106 }
3107
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003108 fe->vma->vm_ops->map_pages(fe, start_pgoff, end_pgoff);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003109
3110 /* preallocated pagetable is unused: free it */
3111 if (fe->prealloc_pte) {
3112 pte_free(fe->vma->vm_mm, fe->prealloc_pte);
3113 fe->prealloc_pte = 0;
3114 }
3115 /* Huge page is mapped? Page fault is solved */
3116 if (pmd_trans_huge(*fe->pmd)) {
3117 ret = VM_FAULT_NOPAGE;
3118 goto out;
3119 }
3120
3121 /* ->map_pages() haven't done anything useful. Cold page cache? */
3122 if (!fe->pte)
3123 goto out;
3124
3125 /* check if the page fault is solved */
3126 fe->pte -= (fe->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3127 if (!pte_none(*fe->pte))
3128 ret = VM_FAULT_NOPAGE;
3129 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003130out:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003131 fe->address = address;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003132 fe->pte = NULL;
3133 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003134}
3135
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003136static int do_read_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003137{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003138 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003139 struct page *fault_page;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003140 int ret = 0;
3141
3142 /*
3143 * Let's call ->map_pages() first and use ->fault() as fallback
3144 * if page by the offset is not ready to be mapped (cold cache or
3145 * something).
3146 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003147 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003148 ret = do_fault_around(fe, pgoff);
3149 if (ret)
3150 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003151 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003152
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003153 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003154 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3155 return ret;
3156
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003157 ret |= alloc_set_pte(fe, NULL, fault_page);
3158 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003159 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003160 unlock_page(fault_page);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003161 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3162 put_page(fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003163 return ret;
3164}
3165
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003166static int do_cow_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003167{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003168 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003169 struct page *fault_page, *new_page;
Jan Karabc2466e2016-05-12 18:29:19 +02003170 void *fault_entry;
Johannes Weiner00501b52014-08-08 14:19:20 -07003171 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003172 int ret;
3173
3174 if (unlikely(anon_vma_prepare(vma)))
3175 return VM_FAULT_OOM;
3176
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003177 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, fe->address);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003178 if (!new_page)
3179 return VM_FAULT_OOM;
3180
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003181 if (mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
3182 &memcg, false)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003183 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003184 return VM_FAULT_OOM;
3185 }
3186
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003187 ret = __do_fault(fe, pgoff, new_page, &fault_page, &fault_entry);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003188 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3189 goto uncharge_out;
3190
Jan Karabc2466e2016-05-12 18:29:19 +02003191 if (!(ret & VM_FAULT_DAX_LOCKED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003192 copy_user_highpage(new_page, fault_page, fe->address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003193 __SetPageUptodate(new_page);
3194
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003195 ret |= alloc_set_pte(fe, memcg, new_page);
3196 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003197 pte_unmap_unlock(fe->pte, fe->ptl);
Jan Karabc2466e2016-05-12 18:29:19 +02003198 if (!(ret & VM_FAULT_DAX_LOCKED)) {
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003199 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003200 put_page(fault_page);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003201 } else {
Jan Karabc2466e2016-05-12 18:29:19 +02003202 dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003203 }
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003204 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3205 goto uncharge_out;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003206 return ret;
3207uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003208 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003209 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003210 return ret;
3211}
3212
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003213static int do_shared_fault(struct fault_env *fe, pgoff_t pgoff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003215 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003216 struct page *fault_page;
3217 struct address_space *mapping;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003218 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003219 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003220
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003221 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003222 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003223 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224
3225 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003226 * Check if the backing address space wants to know that the page is
3227 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003228 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003229 if (vma->vm_ops->page_mkwrite) {
3230 unlock_page(fault_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003231 tmp = do_page_mkwrite(vma, fault_page, fe->address);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003232 if (unlikely(!tmp ||
3233 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003234 put_page(fault_page);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003235 return tmp;
3236 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 }
3238
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003239 ret |= alloc_set_pte(fe, NULL, fault_page);
3240 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003241 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003242 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3243 VM_FAULT_RETRY))) {
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003244 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003245 put_page(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003246 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003247 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003248
3249 if (set_page_dirty(fault_page))
3250 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003251 /*
3252 * Take a local copy of the address_space - page.mapping may be zeroed
3253 * by truncate after unlock_page(). The address_space itself remains
3254 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3255 * release semantics to prevent the compiler from undoing this copying.
3256 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003257 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003258 unlock_page(fault_page);
3259 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3260 /*
3261 * Some device drivers do not set page.mapping but still
3262 * dirty their pages
3263 */
3264 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003265 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003266
Johannes Weiner74ec6752015-02-10 14:11:28 -08003267 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003268 file_update_time(vma->vm_file);
3269
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003270 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003271}
Nick Piggind00806b2007-07-19 01:46:57 -07003272
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003273/*
3274 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3275 * but allow concurrent faults).
3276 * The mmap_sem may have been released depending on flags and our
3277 * return value. See filemap_fault() and __lock_page_or_retry().
3278 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003279static int do_fault(struct fault_env *fe)
Nick Piggin54cb8822007-07-19 01:46:59 -07003280{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003281 struct vm_area_struct *vma = fe->vma;
3282 pgoff_t pgoff = linear_page_index(vma, fe->address);
Nick Piggin54cb8822007-07-19 01:46:59 -07003283
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003284 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3285 if (!vma->vm_ops->fault)
3286 return VM_FAULT_SIGBUS;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003287 if (!(fe->flags & FAULT_FLAG_WRITE))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003288 return do_read_fault(fe, pgoff);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003289 if (!(vma->vm_flags & VM_SHARED))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003290 return do_cow_fault(fe, pgoff);
3291 return do_shared_fault(fe, pgoff);
Nick Piggin54cb8822007-07-19 01:46:59 -07003292}
3293
Rashika Kheriab19a9932014-04-03 14:48:02 -07003294static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003295 unsigned long addr, int page_nid,
3296 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003297{
3298 get_page(page);
3299
3300 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003301 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003302 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003303 *flags |= TNF_FAULT_LOCAL;
3304 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003305
3306 return mpol_misplaced(page, vma, addr);
3307}
3308
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003309static int do_numa_page(struct fault_env *fe, pte_t pte)
Mel Gormand10e63f2012-10-25 14:16:31 +02003310{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003311 struct vm_area_struct *vma = fe->vma;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003312 struct page *page = NULL;
Mel Gorman8191acb2013-10-07 11:28:45 +01003313 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003314 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003315 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003316 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003317 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003318 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003319
3320 /*
3321 * The "pte" at this point cannot be used safely without
3322 * validation through pte_unmap_same(). It's of NUMA type but
3323 * the pfn may be screwed if the read is non atomic.
3324 *
Mel Gorman4d942462015-02-12 14:58:28 -08003325 * We can safely just do a "set_pte_at()", because the old
3326 * page table entry is not accessible, so there would be no
3327 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003328 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003329 fe->ptl = pte_lockptr(vma->vm_mm, fe->pmd);
3330 spin_lock(fe->ptl);
3331 if (unlikely(!pte_same(*fe->pte, pte))) {
3332 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003333 goto out;
3334 }
3335
Mel Gorman4d942462015-02-12 14:58:28 -08003336 /* Make it present again */
3337 pte = pte_modify(pte, vma->vm_page_prot);
3338 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003339 if (was_writable)
3340 pte = pte_mkwrite(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003341 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
3342 update_mmu_cache(vma, fe->address, fe->pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003343
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003344 page = vm_normal_page(vma, fe->address, pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003345 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003346 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003347 return 0;
3348 }
3349
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003350 /* TODO: handle PTE-mapped THP */
3351 if (PageCompound(page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003352 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003353 return 0;
3354 }
3355
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003356 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003357 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3358 * much anyway since they can be in shared cache state. This misses
3359 * the case where a mapping is writable but the process never writes
3360 * to it but pte_write gets cleared during protection updates and
3361 * pte_dirty has unpredictable behaviour between PTE scan updates,
3362 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003363 */
Rik van Rield59dc7b2016-09-08 21:30:53 -04003364 if (!pte_write(pte))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003365 flags |= TNF_NO_GROUP;
3366
Rik van Rieldabe1d92013-10-07 11:29:34 +01003367 /*
3368 * Flag if the page is shared between multiple address spaces. This
3369 * is later used when determining whether to group tasks together
3370 */
3371 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3372 flags |= TNF_SHARED;
3373
Peter Zijlstra90572892013-10-07 11:29:20 +01003374 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003375 page_nid = page_to_nid(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003376 target_nid = numa_migrate_prep(page, vma, fe->address, page_nid,
3377 &flags);
3378 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003379 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003380 put_page(page);
3381 goto out;
3382 }
3383
3384 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003385 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003386 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003387 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003388 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003389 } else
3390 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003391
3392out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003393 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003394 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003395 return 0;
3396}
3397
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003398static int create_huge_pmd(struct fault_env *fe)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003399{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003400 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003401 if (vma_is_anonymous(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003402 return do_huge_pmd_anonymous_page(fe);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003403 if (vma->vm_ops->pmd_fault)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003404 return vma->vm_ops->pmd_fault(vma, fe->address, fe->pmd,
3405 fe->flags);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003406 return VM_FAULT_FALLBACK;
3407}
3408
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003409static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003410{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003411 if (vma_is_anonymous(fe->vma))
3412 return do_huge_pmd_wp_page(fe, orig_pmd);
3413 if (fe->vma->vm_ops->pmd_fault)
3414 return fe->vma->vm_ops->pmd_fault(fe->vma, fe->address, fe->pmd,
3415 fe->flags);
Kirill A. Shutemovaf9e4d52016-07-26 15:25:40 -07003416
3417 /* COW handled on pte level: split pmd */
3418 VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
3419 split_huge_pmd(fe->vma, fe->pmd, fe->address);
3420
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003421 return VM_FAULT_FALLBACK;
3422}
3423
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003424static inline bool vma_is_accessible(struct vm_area_struct *vma)
3425{
3426 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3427}
3428
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429/*
3430 * These routines also need to handle stuff like marking pages dirty
3431 * and/or accessed for architectures that don't do it in hardware (most
3432 * RISC architectures). The early dirtying is also good on the i386.
3433 *
3434 * There is also a hook called "update_mmu_cache()" that architectures
3435 * with external mmu caches can use to update those (ie the Sparc or
3436 * PowerPC hashed page tables that act as extended TLBs).
3437 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003438 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3439 * concurrent faults).
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003440 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003441 * The mmap_sem may have been released depending on flags and our return value.
3442 * See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003444static int handle_pte_fault(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445{
3446 pte_t entry;
3447
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003448 if (unlikely(pmd_none(*fe->pmd))) {
3449 /*
3450 * Leave __pte_alloc() until later: because vm_ops->fault may
3451 * want to allocate huge page, and if we expose page table
3452 * for an instant, it will be difficult to retract from
3453 * concurrent faults and from rmap lookups.
3454 */
3455 fe->pte = NULL;
3456 } else {
3457 /* See comment in pte_alloc_one_map() */
3458 if (pmd_trans_unstable(fe->pmd) || pmd_devmap(*fe->pmd))
3459 return 0;
3460 /*
3461 * A regular pmd is established and it can't morph into a huge
3462 * pmd from under us anymore at this point because we hold the
3463 * mmap_sem read mode and khugepaged takes it in write mode.
3464 * So now it's safe to run pte_offset_map().
3465 */
3466 fe->pte = pte_offset_map(fe->pmd, fe->address);
3467
3468 entry = *fe->pte;
3469
3470 /*
3471 * some architectures can have larger ptes than wordsize,
3472 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3473 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3474 * atomic accesses. The code below just needs a consistent
3475 * view for the ifs and we later double check anyway with the
3476 * ptl lock held. So here a barrier will do.
3477 */
3478 barrier();
Hugh Dickins65500d22005-10-29 18:15:59 -07003479 if (pte_none(entry)) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003480 pte_unmap(fe->pte);
3481 fe->pte = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07003482 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483 }
3484
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003485 if (!fe->pte) {
3486 if (vma_is_anonymous(fe->vma))
3487 return do_anonymous_page(fe);
3488 else
3489 return do_fault(fe);
3490 }
3491
3492 if (!pte_present(entry))
3493 return do_swap_page(fe, entry);
3494
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003495 if (pte_protnone(entry) && vma_is_accessible(fe->vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003496 return do_numa_page(fe, entry);
Mel Gormand10e63f2012-10-25 14:16:31 +02003497
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003498 fe->ptl = pte_lockptr(fe->vma->vm_mm, fe->pmd);
3499 spin_lock(fe->ptl);
3500 if (unlikely(!pte_same(*fe->pte, entry)))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003501 goto unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003502 if (fe->flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 if (!pte_write(entry))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003504 return do_wp_page(fe, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 entry = pte_mkdirty(entry);
3506 }
3507 entry = pte_mkyoung(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003508 if (ptep_set_access_flags(fe->vma, fe->address, fe->pte, entry,
3509 fe->flags & FAULT_FLAG_WRITE)) {
3510 update_mmu_cache(fe->vma, fe->address, fe->pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003511 } else {
3512 /*
3513 * This is needed only for protection faults but the arch code
3514 * is not yet telling us if this is a protection fault or not.
3515 * This still avoids useless tlb flushes for .text page faults
3516 * with threads.
3517 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003518 if (fe->flags & FAULT_FLAG_WRITE)
3519 flush_tlb_fix_spurious_fault(fe->vma, fe->address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003520 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003521unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003522 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003523 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524}
3525
3526/*
3527 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003528 *
3529 * The mmap_sem may have been released depending on flags and our
3530 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003532static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3533 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003535 struct fault_env fe = {
3536 .vma = vma,
3537 .address = address,
3538 .flags = flags,
3539 };
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003540 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 pgd_t *pgd;
3542 pud_t *pud;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 pud = pud_alloc(mm, pgd, address);
3546 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003547 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003548 fe.pmd = pmd_alloc(mm, pud, address);
3549 if (!fe.pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003550 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003551 if (pmd_none(*fe.pmd) && transparent_hugepage_enabled(vma)) {
3552 int ret = create_huge_pmd(&fe);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003553 if (!(ret & VM_FAULT_FALLBACK))
3554 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003555 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003556 pmd_t orig_pmd = *fe.pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003557 int ret;
3558
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003559 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003560 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003561 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003562 return do_huge_pmd_numa_page(&fe, orig_pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02003563
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003564 if ((fe.flags & FAULT_FLAG_WRITE) &&
3565 !pmd_write(orig_pmd)) {
3566 ret = wp_huge_pmd(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003567 if (!(ret & VM_FAULT_FALLBACK))
3568 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003569 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003570 huge_pmd_set_accessed(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003571 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003572 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003573 }
3574 }
3575
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003576 return handle_pte_fault(&fe);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577}
3578
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003579/*
3580 * By the time we get here, we already hold the mm semaphore
3581 *
3582 * The mmap_sem may have been released depending on flags and our
3583 * return value. See filemap_fault() and __lock_page_or_retry().
3584 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003585int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3586 unsigned int flags)
Johannes Weiner519e5242013-09-12 15:13:42 -07003587{
3588 int ret;
3589
3590 __set_current_state(TASK_RUNNING);
3591
3592 count_vm_event(PGFAULT);
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003593 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
Johannes Weiner519e5242013-09-12 15:13:42 -07003594
3595 /* do counter updates before entering really critical section. */
3596 check_sync_rss_stat(current);
3597
3598 /*
3599 * Enable the memcg OOM handling for faults triggered in user
3600 * space. Kernel faults are handled more gracefully.
3601 */
3602 if (flags & FAULT_FLAG_USER)
Johannes Weiner49426422013-10-16 13:46:59 -07003603 mem_cgroup_oom_enable();
Johannes Weiner519e5242013-09-12 15:13:42 -07003604
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003605 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3606 flags & FAULT_FLAG_INSTRUCTION,
3607 flags & FAULT_FLAG_REMOTE))
3608 return VM_FAULT_SIGSEGV;
3609
3610 if (unlikely(is_vm_hugetlb_page(vma)))
3611 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3612 else
3613 ret = __handle_mm_fault(vma, address, flags);
Johannes Weiner519e5242013-09-12 15:13:42 -07003614
Johannes Weiner49426422013-10-16 13:46:59 -07003615 if (flags & FAULT_FLAG_USER) {
3616 mem_cgroup_oom_disable();
3617 /*
3618 * The task may have entered a memcg OOM situation but
3619 * if the allocation error was handled gracefully (no
3620 * VM_FAULT_OOM), there is no need to kill anything.
3621 * Just clean up the OOM state peacefully.
3622 */
3623 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3624 mem_cgroup_oom_synchronize(false);
3625 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003626
Michal Hocko3f70dc32016-10-07 16:59:06 -07003627 /*
3628 * This mm has been already reaped by the oom reaper and so the
3629 * refault cannot be trusted in general. Anonymous refaults would
3630 * lose data and give a zero page instead e.g. This is especially
3631 * problem for use_mm() because regular tasks will just die and
3632 * the corrupted data will not be visible anywhere while kthread
3633 * will outlive the oom victim and potentially propagate the date
3634 * further.
3635 */
3636 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3637 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3638 ret = VM_FAULT_SIGBUS;
3639
Johannes Weiner519e5242013-09-12 15:13:42 -07003640 return ret;
3641}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003642EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003643
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644#ifndef __PAGETABLE_PUD_FOLDED
3645/*
3646 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003647 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003649int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003651 pud_t *new = pud_alloc_one(mm, address);
3652 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003653 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003654
Nick Piggin362a61a2008-05-14 06:37:36 +02003655 smp_wmb(); /* See comment in __pte_alloc */
3656
Hugh Dickins872fec12005-10-29 18:16:21 -07003657 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003658 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003659 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003660 else
3661 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003662 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003663 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664}
3665#endif /* __PAGETABLE_PUD_FOLDED */
3666
3667#ifndef __PAGETABLE_PMD_FOLDED
3668/*
3669 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003670 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003672int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003674 pmd_t *new = pmd_alloc_one(mm, address);
3675 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003676 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677
Nick Piggin362a61a2008-05-14 06:37:36 +02003678 smp_wmb(); /* See comment in __pte_alloc */
3679
Hugh Dickins872fec12005-10-29 18:16:21 -07003680 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003682 if (!pud_present(*pud)) {
3683 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003684 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003685 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003686 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003687#else
3688 if (!pgd_present(*pud)) {
3689 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003690 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003691 } else /* Another has populated it */
3692 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003694 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003695 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003696}
3697#endif /* __PAGETABLE_PMD_FOLDED */
3698
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003699static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003700 pte_t **ptepp, spinlock_t **ptlp)
3701{
3702 pgd_t *pgd;
3703 pud_t *pud;
3704 pmd_t *pmd;
3705 pte_t *ptep;
3706
3707 pgd = pgd_offset(mm, address);
3708 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3709 goto out;
3710
3711 pud = pud_offset(pgd, address);
3712 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3713 goto out;
3714
3715 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003716 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003717 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3718 goto out;
3719
3720 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3721 if (pmd_huge(*pmd))
3722 goto out;
3723
3724 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3725 if (!ptep)
3726 goto out;
3727 if (!pte_present(*ptep))
3728 goto unlock;
3729 *ptepp = ptep;
3730 return 0;
3731unlock:
3732 pte_unmap_unlock(ptep, *ptlp);
3733out:
3734 return -EINVAL;
3735}
3736
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003737static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3738 pte_t **ptepp, spinlock_t **ptlp)
3739{
3740 int res;
3741
3742 /* (void) is needed to make gcc happy */
3743 (void) __cond_lock(*ptlp,
3744 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3745 return res;
3746}
3747
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003748/**
3749 * follow_pfn - look up PFN at a user virtual address
3750 * @vma: memory mapping
3751 * @address: user virtual address
3752 * @pfn: location to store found PFN
3753 *
3754 * Only IO mappings and raw PFN mappings are allowed.
3755 *
3756 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3757 */
3758int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3759 unsigned long *pfn)
3760{
3761 int ret = -EINVAL;
3762 spinlock_t *ptl;
3763 pte_t *ptep;
3764
3765 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3766 return ret;
3767
3768 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3769 if (ret)
3770 return ret;
3771 *pfn = pte_pfn(*ptep);
3772 pte_unmap_unlock(ptep, ptl);
3773 return 0;
3774}
3775EXPORT_SYMBOL(follow_pfn);
3776
Rik van Riel28b2ee22008-07-23 21:27:05 -07003777#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003778int follow_phys(struct vm_area_struct *vma,
3779 unsigned long address, unsigned int flags,
3780 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003781{
Johannes Weiner03668a42009-06-16 15:32:34 -07003782 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003783 pte_t *ptep, pte;
3784 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003785
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003786 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3787 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003788
Johannes Weiner03668a42009-06-16 15:32:34 -07003789 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003790 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003791 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003792
Rik van Riel28b2ee22008-07-23 21:27:05 -07003793 if ((flags & FOLL_WRITE) && !pte_write(pte))
3794 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003795
3796 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003797 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003798
Johannes Weiner03668a42009-06-16 15:32:34 -07003799 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003800unlock:
3801 pte_unmap_unlock(ptep, ptl);
3802out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003803 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003804}
3805
3806int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3807 void *buf, int len, int write)
3808{
3809 resource_size_t phys_addr;
3810 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003811 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003812 int offset = addr & (PAGE_SIZE-1);
3813
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003814 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003815 return -EINVAL;
3816
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003817 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003818 if (write)
3819 memcpy_toio(maddr + offset, buf, len);
3820 else
3821 memcpy_fromio(buf, maddr + offset, len);
3822 iounmap(maddr);
3823
3824 return len;
3825}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003826EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003827#endif
3828
David Howells0ec76a12006-09-27 01:50:15 -07003829/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003830 * Access another process' address space as given in mm. If non-NULL, use the
3831 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003832 */
Eric W. Biedermane71b4e02016-11-22 12:06:50 -06003833int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003834 unsigned long addr, void *buf, int len, unsigned int gup_flags)
David Howells0ec76a12006-09-27 01:50:15 -07003835{
David Howells0ec76a12006-09-27 01:50:15 -07003836 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003837 void *old_buf = buf;
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003838 int write = gup_flags & FOLL_WRITE;
David Howells0ec76a12006-09-27 01:50:15 -07003839
David Howells0ec76a12006-09-27 01:50:15 -07003840 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003841 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003842 while (len) {
3843 int bytes, ret, offset;
3844 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003845 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003846
Dave Hansen1e987792016-02-12 13:01:54 -08003847 ret = get_user_pages_remote(tsk, mm, addr, 1,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003848 gup_flags, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003849 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003850#ifndef CONFIG_HAVE_IOREMAP_PROT
3851 break;
3852#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003853 /*
3854 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3855 * we can access using slightly different code.
3856 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003857 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003858 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003859 break;
3860 if (vma->vm_ops && vma->vm_ops->access)
3861 ret = vma->vm_ops->access(vma, addr, buf,
3862 len, write);
3863 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003864 break;
3865 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003866#endif
David Howells0ec76a12006-09-27 01:50:15 -07003867 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003868 bytes = len;
3869 offset = addr & (PAGE_SIZE-1);
3870 if (bytes > PAGE_SIZE-offset)
3871 bytes = PAGE_SIZE-offset;
3872
3873 maddr = kmap(page);
3874 if (write) {
3875 copy_to_user_page(vma, page, addr,
3876 maddr + offset, buf, bytes);
3877 set_page_dirty_lock(page);
3878 } else {
3879 copy_from_user_page(vma, page, addr,
3880 buf, maddr + offset, bytes);
3881 }
3882 kunmap(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003883 put_page(page);
David Howells0ec76a12006-09-27 01:50:15 -07003884 }
David Howells0ec76a12006-09-27 01:50:15 -07003885 len -= bytes;
3886 buf += bytes;
3887 addr += bytes;
3888 }
3889 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003890
3891 return buf - old_buf;
3892}
Andi Kleen03252912008-01-30 13:33:18 +01003893
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003894/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003895 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003896 * @mm: the mm_struct of the target address space
3897 * @addr: start address to access
3898 * @buf: source or destination buffer
3899 * @len: number of bytes to transfer
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003900 * @gup_flags: flags modifying lookup behaviour
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003901 *
3902 * The caller must hold a reference on @mm.
3903 */
3904int access_remote_vm(struct mm_struct *mm, unsigned long addr,
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003905 void *buf, int len, unsigned int gup_flags)
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003906{
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003907 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003908}
3909
Andi Kleen03252912008-01-30 13:33:18 +01003910/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003911 * Access another process' address space.
3912 * Source/target buffer must be kernel space,
3913 * Do not walk the page table directly, use get_user_pages
3914 */
3915int access_process_vm(struct task_struct *tsk, unsigned long addr,
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003916 void *buf, int len, unsigned int gup_flags)
Stephen Wilson206cb632011-03-13 15:49:19 -04003917{
3918 struct mm_struct *mm;
3919 int ret;
3920
3921 mm = get_task_mm(tsk);
3922 if (!mm)
3923 return 0;
3924
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003925 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003926
Stephen Wilson206cb632011-03-13 15:49:19 -04003927 mmput(mm);
3928
3929 return ret;
3930}
3931
Andi Kleen03252912008-01-30 13:33:18 +01003932/*
3933 * Print the name of a VMA.
3934 */
3935void print_vma_addr(char *prefix, unsigned long ip)
3936{
3937 struct mm_struct *mm = current->mm;
3938 struct vm_area_struct *vma;
3939
Ingo Molnare8bff742008-02-13 20:21:06 +01003940 /*
3941 * Do not print if we are in atomic
3942 * contexts (in exception stacks, etc.):
3943 */
3944 if (preempt_count())
3945 return;
3946
Andi Kleen03252912008-01-30 13:33:18 +01003947 down_read(&mm->mmap_sem);
3948 vma = find_vma(mm, ip);
3949 if (vma && vma->vm_file) {
3950 struct file *f = vma->vm_file;
3951 char *buf = (char *)__get_free_page(GFP_KERNEL);
3952 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003953 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003954
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02003955 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003956 if (IS_ERR(p))
3957 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003958 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003959 vma->vm_start,
3960 vma->vm_end - vma->vm_start);
3961 free_page((unsigned long)buf);
3962 }
3963 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003964 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003965}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003966
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003967#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02003968void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003969{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003970 /*
3971 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3972 * holding the mmap_sem, this is safe because kernel memory doesn't
3973 * get paged out, therefore we'll never actually fault, and the
3974 * below annotations will generate false positives.
3975 */
3976 if (segment_eq(get_fs(), KERNEL_DS))
3977 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003978 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003979 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003980 __might_sleep(file, line, 0);
3981#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003982 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003983 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02003984#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003985}
David Hildenbrand9ec23532015-05-11 17:52:07 +02003986EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003987#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003988
3989#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3990static void clear_gigantic_page(struct page *page,
3991 unsigned long addr,
3992 unsigned int pages_per_huge_page)
3993{
3994 int i;
3995 struct page *p = page;
3996
3997 might_sleep();
3998 for (i = 0; i < pages_per_huge_page;
3999 i++, p = mem_map_next(p, page, i)) {
4000 cond_resched();
4001 clear_user_highpage(p, addr + i * PAGE_SIZE);
4002 }
4003}
4004void clear_huge_page(struct page *page,
4005 unsigned long addr, unsigned int pages_per_huge_page)
4006{
4007 int i;
4008
4009 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4010 clear_gigantic_page(page, addr, pages_per_huge_page);
4011 return;
4012 }
4013
4014 might_sleep();
4015 for (i = 0; i < pages_per_huge_page; i++) {
4016 cond_resched();
4017 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4018 }
4019}
4020
4021static void copy_user_gigantic_page(struct page *dst, struct page *src,
4022 unsigned long addr,
4023 struct vm_area_struct *vma,
4024 unsigned int pages_per_huge_page)
4025{
4026 int i;
4027 struct page *dst_base = dst;
4028 struct page *src_base = src;
4029
4030 for (i = 0; i < pages_per_huge_page; ) {
4031 cond_resched();
4032 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4033
4034 i++;
4035 dst = mem_map_next(dst, dst_base, i);
4036 src = mem_map_next(src, src_base, i);
4037 }
4038}
4039
4040void copy_user_huge_page(struct page *dst, struct page *src,
4041 unsigned long addr, struct vm_area_struct *vma,
4042 unsigned int pages_per_huge_page)
4043{
4044 int i;
4045
4046 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4047 copy_user_gigantic_page(dst, src, addr, vma,
4048 pages_per_huge_page);
4049 return;
4050 }
4051
4052 might_sleep();
4053 for (i = 0; i < pages_per_huge_page; i++) {
4054 cond_resched();
4055 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4056 }
4057}
4058#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004059
Olof Johansson40b64ac2013-12-20 14:28:05 -08004060#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004061
4062static struct kmem_cache *page_ptl_cachep;
4063
4064void __init ptlock_cache_init(void)
4065{
4066 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4067 SLAB_PANIC, NULL);
4068}
4069
Peter Zijlstra539edb52013-11-14 14:31:52 -08004070bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004071{
4072 spinlock_t *ptl;
4073
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004074 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004075 if (!ptl)
4076 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08004077 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004078 return true;
4079}
4080
Peter Zijlstra539edb52013-11-14 14:31:52 -08004081void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004082{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004083 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004084}
4085#endif