blob: f882280f58f7994bafe80ac464cffba6ce3e0bb0 [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
Arnd Bergmann73184542018-02-16 16:25:53 +010078#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
Peter Zijlstra90572892013-10-07 11:29:20 +010079#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;
Mel Gorman5a1eef72017-08-02 13:31:52 -07001127 flush_tlb_batched_pending(mm);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001128 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 do {
1130 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -08001131 if (pte_none(ptent)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 continue;
Robin Holt51c6f662005-11-13 16:06:42 -08001133 }
Hugh Dickins6f5e6b92006-03-16 23:04:09 -08001134
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001136 struct page *page;
Robin Holt51c6f662005-11-13 16:06:42 -08001137
Linus Torvalds6aab3412005-11-28 14:34:23 -08001138 page = vm_normal_page(vma, addr, ptent);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139 if (unlikely(details) && page) {
1140 /*
1141 * unmap_shared_mapping_pages() wants to
1142 * invalidate cache without truncating:
1143 * unmap shared but keep private pages.
1144 */
1145 if (details->check_mapping &&
Kirill A. Shutemov800d8c62016-07-26 15:26:18 -07001146 details->check_mapping != page_rmapping(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 }
Nick Pigginb5810032005-10-29 18:16:12 -07001149 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -07001150 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 tlb_remove_tlb_entry(tlb, pte, addr);
1152 if (unlikely(!page))
1153 continue;
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001154
1155 if (!PageAnon(page)) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001156 if (pte_dirty(ptent)) {
1157 force_flush = 1;
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001158 set_page_dirty(page);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001159 }
Johannes Weiner4917e5d2009-01-06 14:39:17 -08001160 if (pte_young(ptent) &&
Joe Perches64363aa2013-07-08 16:00:18 -07001161 likely(!(vma->vm_flags & VM_SEQ_READ)))
Nick Pigginbf3f3bc2009-01-06 14:38:55 -08001162 mark_page_accessed(page);
Hugh Dickins6237bcd2005-10-29 18:15:54 -07001163 }
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001164 rss[mm_counter(page)]--;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08001165 page_remove_rmap(page, false);
Hugh Dickins3dc14742009-01-06 14:40:08 -08001166 if (unlikely(page_mapcount(page) < 0))
1167 print_bad_pte(vma, addr, ptent, page);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001168 if (unlikely(__tlb_remove_page(tlb, page))) {
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001169 force_flush = 1;
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001170 pending_page = page;
Will Deaconce9ec372014-10-28 13:16:28 -07001171 addr += PAGE_SIZE;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001172 break;
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001173 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 continue;
1175 }
Kirill A. Shutemovf51ed732017-02-22 15:46:34 -08001176 /* If details->check_mapping, we leave swap entries. */
1177 if (unlikely(details))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 continue;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001179
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001180 entry = pte_to_swp_entry(ptent);
1181 if (!non_swap_entry(entry))
1182 rss[MM_SWAPENTS]--;
1183 else if (is_migration_entry(entry)) {
1184 struct page *page;
Konstantin Khlebnikov9f9f1ac2012-01-20 14:34:24 -08001185
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001186 page = migration_entry_to_page(entry);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001187 rss[mm_counter(page)]--;
KAMEZAWA Hiroyukib084d432010-03-05 13:41:42 -08001188 }
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001189 if (unlikely(!free_swap_and_cache(entry)))
1190 print_bad_pte(vma, addr, ptent, NULL);
Zachary Amsden9888a1c2006-09-30 23:29:31 -07001191 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001192 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsae859762005-10-29 18:16:05 -07001193
KAMEZAWA Hiroyukid559db02010-03-05 13:41:39 -08001194 add_mm_rss_vec(mm, rss);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001195 arch_leave_lazy_mmu_mode();
Robin Holt51c6f662005-11-13 16:06:42 -08001196
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001197 /* Do the actual TLB flush before dropping ptl */
Will Deaconfb7332a2014-10-29 10:03:09 +00001198 if (force_flush)
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001199 tlb_flush_mmu_tlbonly(tlb);
Linus Torvalds1cf35d42014-04-25 16:05:40 -07001200 pte_unmap_unlock(start_pte, ptl);
1201
1202 /*
1203 * If we forced a TLB flush (either due to running out of
1204 * batch buffers or because we needed to flush dirty TLB
1205 * entries before releasing the ptl), free the batched
1206 * memory too. Restart if we didn't do everything.
1207 */
1208 if (force_flush) {
1209 force_flush = 0;
1210 tlb_flush_mmu_free(tlb);
Aneesh Kumar K.Ve9d55e12016-07-26 15:24:09 -07001211 if (pending_page) {
1212 /* remove the page with new size */
1213 __tlb_remove_pte_page(tlb, pending_page);
1214 pending_page = NULL;
1215 }
Linus Torvalds2b047252013-08-15 11:42:25 -07001216 if (addr != end)
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001217 goto again;
1218 }
1219
Robin Holt51c6f662005-11-13 16:06:42 -08001220 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221}
1222
Robin Holt51c6f662005-11-13 16:06:42 -08001223static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001224 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001226 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227{
1228 pmd_t *pmd;
1229 unsigned long next;
1230
1231 pmd = pmd_offset(pud, addr);
1232 do {
1233 next = pmd_addr_end(addr, end);
Dan Williams5c7fb562016-01-15 16:56:52 -08001234 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001235 if (next - addr != HPAGE_PMD_SIZE) {
Hugh Dickins68428392016-05-05 16:22:09 -07001236 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1237 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
Kirill A. Shutemov78ddc532016-01-15 16:52:42 -08001238 split_huge_pmd(vma, pmd, addr);
Shaohua Lif21760b2012-01-12 17:19:16 -08001239 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001240 goto next;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08001241 /* fall through */
1242 }
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001243 /*
1244 * Here there can be other concurrent MADV_DONTNEED or
1245 * trans huge page faults running, and if the pmd is
1246 * none or trans huge it can change under us. This is
1247 * because MADV_DONTNEED holds the mmap_sem in read
1248 * mode.
1249 */
1250 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1251 goto next;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001252 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
Andrea Arcangeli1a5a9902012-03-21 16:33:42 -07001253next:
Peter Zijlstra97a89412011-05-24 17:12:04 -07001254 cond_resched();
1255 } while (pmd++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001256
1257 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258}
1259
Robin Holt51c6f662005-11-13 16:06:42 -08001260static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -07001261 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 unsigned long addr, unsigned long end,
Peter Zijlstra97a89412011-05-24 17:12:04 -07001263 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264{
1265 pud_t *pud;
1266 unsigned long next;
1267
1268 pud = pud_offset(pgd, addr);
1269 do {
1270 next = pud_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001271 if (pud_none_or_clear_bad(pud))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001273 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1274 } while (pud++, addr = next, addr != end);
Robin Holt51c6f662005-11-13 16:06:42 -08001275
1276 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277}
1278
Michal Hockoaac45362016-03-25 14:20:24 -07001279void unmap_page_range(struct mmu_gather *tlb,
Al Viro038c7aa2012-03-05 13:25:09 -05001280 struct vm_area_struct *vma,
1281 unsigned long addr, unsigned long end,
1282 struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001283{
1284 pgd_t *pgd;
1285 unsigned long next;
1286
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 BUG_ON(addr >= end);
1288 tlb_start_vma(tlb, vma);
1289 pgd = pgd_offset(vma->vm_mm, addr);
1290 do {
1291 next = pgd_addr_end(addr, end);
Peter Zijlstra97a89412011-05-24 17:12:04 -07001292 if (pgd_none_or_clear_bad(pgd))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 continue;
Peter Zijlstra97a89412011-05-24 17:12:04 -07001294 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1295 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 tlb_end_vma(tlb, vma);
1297}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
Al Virof5cc4ee2012-03-05 14:14:20 -05001299
1300static void unmap_single_vma(struct mmu_gather *tlb,
1301 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001302 unsigned long end_addr,
Al Virof5cc4ee2012-03-05 14:14:20 -05001303 struct zap_details *details)
1304{
1305 unsigned long start = max(vma->vm_start, start_addr);
1306 unsigned long end;
1307
1308 if (start >= vma->vm_end)
1309 return;
1310 end = min(vma->vm_end, end_addr);
1311 if (end <= vma->vm_start)
1312 return;
1313
Srikar Dronamrajucbc91f72012-04-11 16:05:27 +05301314 if (vma->vm_file)
1315 uprobe_munmap(vma, start, end);
1316
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001317 if (unlikely(vma->vm_flags & VM_PFNMAP))
Suresh Siddha5180da42012-10-08 16:28:29 -07001318 untrack_pfn(vma, 0, 0);
Al Virof5cc4ee2012-03-05 14:14:20 -05001319
1320 if (start != end) {
1321 if (unlikely(is_vm_hugetlb_page(vma))) {
1322 /*
1323 * It is undesirable to test vma->vm_file as it
1324 * should be non-null for valid hugetlb area.
1325 * However, vm_file will be NULL in the error
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001326 * cleanup path of mmap_region. When
Al Virof5cc4ee2012-03-05 14:14:20 -05001327 * hugetlbfs ->mmap method fails,
Davidlohr Bueso7aa6b4a2014-04-07 15:37:01 -07001328 * mmap_region() nullifies vma->vm_file
Al Virof5cc4ee2012-03-05 14:14:20 -05001329 * before calling this function to clean up.
1330 * Since no pte has actually been setup, it is
1331 * safe to do nothing in this case.
1332 */
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001333 if (vma->vm_file) {
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001334 i_mmap_lock_write(vma->vm_file->f_mapping);
Mel Gormand8333522012-07-31 16:46:20 -07001335 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
Davidlohr Bueso83cde9e2014-12-12 16:54:21 -08001336 i_mmap_unlock_write(vma->vm_file->f_mapping);
Aneesh Kumar K.V24669e52012-07-31 16:42:03 -07001337 }
Al Virof5cc4ee2012-03-05 14:14:20 -05001338 } else
1339 unmap_page_range(tlb, vma, start, end, details);
1340 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341}
1342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343/**
1344 * unmap_vmas - unmap a range of memory covered by a list of vma's
Randy Dunlap0164f692011-06-15 15:08:09 -07001345 * @tlb: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 * @vma: the starting vma
1347 * @start_addr: virtual address at which to start unmapping
1348 * @end_addr: virtual address at which to end unmapping
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 *
Hugh Dickins508034a2005-10-29 18:16:30 -07001350 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 * Only addresses between `start' and `end' will be unmapped.
1353 *
1354 * The VMA list must be sorted in ascending virtual address order.
1355 *
1356 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1357 * range after unmap_vmas() returns. So the only responsibility here is to
1358 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1359 * drops the lock and schedules.
1360 */
Al Viro6e8bb012012-03-05 13:41:15 -05001361void unmap_vmas(struct mmu_gather *tlb,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 struct vm_area_struct *vma, unsigned long start_addr,
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001363 unsigned long end_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364{
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001365 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001367 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
Al Virof5cc4ee2012-03-05 14:14:20 -05001368 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001369 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
Andrea Arcangelicddb8a52008-07-28 15:46:29 -07001370 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371}
1372
1373/**
1374 * zap_page_range - remove user pages in a given range
1375 * @vma: vm_area_struct holding the applicable pages
Randy Dunlapeb4546b2012-06-20 12:53:02 -07001376 * @start: starting address of pages to zap
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001378 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001379 *
1380 * Caller must protect the VMA list
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 */
Linus Torvalds7e027b12012-05-06 13:43:15 -07001382void zap_page_range(struct vm_area_struct *vma, unsigned long start,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 unsigned long size, struct zap_details *details)
1384{
1385 struct mm_struct *mm = vma->vm_mm;
Peter Zijlstrad16dfc52011-05-24 17:11:45 -07001386 struct mmu_gather tlb;
Linus Torvalds7e027b12012-05-06 13:43:15 -07001387 unsigned long end = start + size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001390 tlb_gather_mmu(&tlb, mm, start, end);
Hugh Dickins365e9c872005-10-29 18:16:18 -07001391 update_hiwater_rss(mm);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001392 mmu_notifier_invalidate_range_start(mm, start, end);
1393 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001394 unmap_single_vma(&tlb, vma, start, end, details);
Linus Torvalds7e027b12012-05-06 13:43:15 -07001395 mmu_notifier_invalidate_range_end(mm, start, end);
1396 tlb_finish_mmu(&tlb, start, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397}
1398
Jack Steinerc627f9c2008-07-29 22:33:53 -07001399/**
Al Virof5cc4ee2012-03-05 14:14:20 -05001400 * zap_page_range_single - remove user pages in a given range
1401 * @vma: vm_area_struct holding the applicable pages
1402 * @address: starting address of pages to zap
1403 * @size: number of bytes to zap
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08001404 * @details: details of shared cache invalidation
Al Virof5cc4ee2012-03-05 14:14:20 -05001405 *
1406 * The range must fit into one VMA.
1407 */
1408static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1409 unsigned long size, struct zap_details *details)
1410{
1411 struct mm_struct *mm = vma->vm_mm;
1412 struct mmu_gather tlb;
1413 unsigned long end = address + size;
Al Virof5cc4ee2012-03-05 14:14:20 -05001414
1415 lru_add_drain();
Linus Torvalds2b047252013-08-15 11:42:25 -07001416 tlb_gather_mmu(&tlb, mm, address, end);
Al Virof5cc4ee2012-03-05 14:14:20 -05001417 update_hiwater_rss(mm);
1418 mmu_notifier_invalidate_range_start(mm, address, end);
Linus Torvalds4f74d2c2012-05-06 13:54:06 -07001419 unmap_single_vma(&tlb, vma, address, end, details);
Al Virof5cc4ee2012-03-05 14:14:20 -05001420 mmu_notifier_invalidate_range_end(mm, address, end);
1421 tlb_finish_mmu(&tlb, address, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422}
1423
Jack Steinerc627f9c2008-07-29 22:33:53 -07001424/**
1425 * zap_vma_ptes - remove ptes mapping the vma
1426 * @vma: vm_area_struct holding ptes to be zapped
1427 * @address: starting address of pages to zap
1428 * @size: number of bytes to zap
1429 *
1430 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1431 *
1432 * The entire address range must be fully contained within the vma.
1433 *
1434 * Returns 0 if successful.
1435 */
1436int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1437 unsigned long size)
1438{
1439 if (address < vma->vm_start || address + size > vma->vm_end ||
1440 !(vma->vm_flags & VM_PFNMAP))
1441 return -1;
Al Virof5cc4ee2012-03-05 14:14:20 -05001442 zap_page_range_single(vma, address, size, NULL);
Jack Steinerc627f9c2008-07-29 22:33:53 -07001443 return 0;
1444}
1445EXPORT_SYMBOL_GPL(zap_vma_ptes);
1446
Namhyung Kim25ca1d62010-10-26 14:21:59 -07001447pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
Harvey Harrison920c7a52008-02-04 22:29:26 -08001448 spinlock_t **ptl)
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001449{
1450 pgd_t * pgd = pgd_offset(mm, addr);
1451 pud_t * pud = pud_alloc(mm, pgd, addr);
1452 if (pud) {
Trond Myklebust49c91fb2005-11-29 19:27:22 -05001453 pmd_t * pmd = pmd_alloc(mm, pud, addr);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001454 if (pmd) {
1455 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001456 return pte_alloc_map_lock(mm, pmd, addr, ptl);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001457 }
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001458 }
1459 return NULL;
1460}
1461
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462/*
Linus Torvalds238f58d2005-11-29 13:01:56 -08001463 * This is the old fallback for page remapping.
1464 *
1465 * For historical reasons, it only allows reserved pages. Only
1466 * old drivers should use this, and they needed to mark their
1467 * pages reserved for the old functions anyway.
1468 */
Nick Piggin423bad62008-04-28 02:13:01 -07001469static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1470 struct page *page, pgprot_t prot)
Linus Torvalds238f58d2005-11-29 13:01:56 -08001471{
Nick Piggin423bad62008-04-28 02:13:01 -07001472 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001473 int retval;
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001474 pte_t *pte;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001475 spinlock_t *ptl;
1476
Linus Torvalds238f58d2005-11-29 13:01:56 -08001477 retval = -EINVAL;
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001478 if (PageAnon(page))
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001479 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001480 retval = -ENOMEM;
1481 flush_dcache_page(page);
Linus Torvaldsc9cfcdd2005-11-29 14:03:14 -08001482 pte = get_locked_pte(mm, addr, &ptl);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001483 if (!pte)
KAMEZAWA Hiroyuki5b4e6552008-10-18 20:28:10 -07001484 goto out;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001485 retval = -EBUSY;
1486 if (!pte_none(*pte))
1487 goto out_unlock;
1488
1489 /* Ok, finally just insert the thing.. */
1490 get_page(page);
Jerome Marchandeca56ff2016-01-14 15:19:26 -08001491 inc_mm_counter_fast(mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07001492 page_add_file_rmap(page, false);
Linus Torvalds238f58d2005-11-29 13:01:56 -08001493 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1494
1495 retval = 0;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08001496 pte_unmap_unlock(pte, ptl);
1497 return retval;
Linus Torvalds238f58d2005-11-29 13:01:56 -08001498out_unlock:
1499 pte_unmap_unlock(pte, ptl);
1500out:
1501 return retval;
1502}
1503
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001504/**
1505 * vm_insert_page - insert single page into user vma
1506 * @vma: user vma to map to
1507 * @addr: target user address of this page
1508 * @page: source kernel page
1509 *
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001510 * This allows drivers to insert individual pages they've allocated
1511 * into a user vma.
1512 *
1513 * The page has to be a nice clean _individual_ kernel allocation.
1514 * If you allocate a compound page, you need to have marked it as
1515 * such (__GFP_COMP), or manually just split the page up yourself
Nick Piggin8dfcc9b2006-03-22 00:08:05 -08001516 * (see split_page()).
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001517 *
1518 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1519 * took an arbitrary page protection parameter. This doesn't allow
1520 * that. Your vma protection will have to be set up correctly, which
1521 * means that if you want a shared writable mapping, you'd better
1522 * ask for a shared writable mapping!
1523 *
1524 * The page does not need to be reserved.
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001525 *
1526 * Usually this function is called from f_op->mmap() handler
1527 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1528 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1529 * function from other places, for example from page-fault handler.
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001530 */
Nick Piggin423bad62008-04-28 02:13:01 -07001531int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1532 struct page *page)
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001533{
1534 if (addr < vma->vm_start || addr >= vma->vm_end)
1535 return -EFAULT;
1536 if (!page_count(page))
1537 return -EINVAL;
Konstantin Khlebnikov4b6e1e32012-10-08 16:28:40 -07001538 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1539 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1540 BUG_ON(vma->vm_flags & VM_PFNMAP);
1541 vma->vm_flags |= VM_MIXEDMAP;
1542 }
Nick Piggin423bad62008-04-28 02:13:01 -07001543 return insert_page(vma, addr, page, vma->vm_page_prot);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001544}
Linus Torvaldse3c33742005-12-03 20:48:11 -08001545EXPORT_SYMBOL(vm_insert_page);
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001546
Nick Piggin423bad62008-04-28 02:13:01 -07001547static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001548 pfn_t pfn, pgprot_t prot)
Nick Piggin423bad62008-04-28 02:13:01 -07001549{
1550 struct mm_struct *mm = vma->vm_mm;
1551 int retval;
1552 pte_t *pte, entry;
1553 spinlock_t *ptl;
1554
1555 retval = -ENOMEM;
1556 pte = get_locked_pte(mm, addr, &ptl);
1557 if (!pte)
1558 goto out;
1559 retval = -EBUSY;
1560 if (!pte_none(*pte))
1561 goto out_unlock;
1562
1563 /* Ok, finally just insert the thing.. */
Dan Williams01c8f1c2016-01-15 16:56:40 -08001564 if (pfn_t_devmap(pfn))
1565 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1566 else
1567 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
Nick Piggin423bad62008-04-28 02:13:01 -07001568 set_pte_at(mm, addr, pte, entry);
Russell King4b3073e2009-12-18 16:40:18 +00001569 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
Nick Piggin423bad62008-04-28 02:13:01 -07001570
1571 retval = 0;
1572out_unlock:
1573 pte_unmap_unlock(pte, ptl);
1574out:
1575 return retval;
1576}
1577
Nick Piggine0dc0d82007-02-12 00:51:36 -08001578/**
1579 * vm_insert_pfn - insert single pfn into user vma
1580 * @vma: user vma to map to
1581 * @addr: target user address of this page
1582 * @pfn: source kernel pfn
1583 *
Robert P. J. Dayc462f172012-10-08 16:33:43 -07001584 * Similar to vm_insert_page, this allows drivers to insert individual pages
Nick Piggine0dc0d82007-02-12 00:51:36 -08001585 * they've allocated into a user vma. Same comments apply.
1586 *
1587 * This function should only be called from a vm_ops->fault handler, and
1588 * in that case the handler should return NULL.
Nick Piggin0d71d10a2008-07-23 21:27:05 -07001589 *
1590 * vma cannot be a COW mapping.
1591 *
1592 * As this is called only for pages that do not currently exist, we
1593 * do not need to flush old virtual caches or the TLB.
Nick Piggine0dc0d82007-02-12 00:51:36 -08001594 */
1595int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
Nick Piggin423bad62008-04-28 02:13:01 -07001596 unsigned long pfn)
Nick Piggine0dc0d82007-02-12 00:51:36 -08001597{
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001598 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1599}
1600EXPORT_SYMBOL(vm_insert_pfn);
1601
1602/**
1603 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1604 * @vma: user vma to map to
1605 * @addr: target user address of this page
1606 * @pfn: source kernel pfn
1607 * @pgprot: pgprot flags for the inserted page
1608 *
1609 * This is exactly like vm_insert_pfn, except that it allows drivers to
1610 * to override pgprot on a per-page basis.
1611 *
1612 * This only makes sense for IO mappings, and it makes no sense for
1613 * cow mappings. In general, using multiple vmas is preferable;
1614 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1615 * impractical.
1616 */
1617int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1618 unsigned long pfn, pgprot_t pgprot)
1619{
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001620 int ret;
Nick Piggin7e675132008-04-28 02:13:00 -07001621 /*
1622 * Technically, architectures with pte_special can avoid all these
1623 * restrictions (same for remap_pfn_range). However we would like
1624 * consistency in testing and feature parity among all, so we should
1625 * try to keep these invariants in place for everybody.
1626 */
Jared Hulbertb379d792008-04-28 02:12:58 -07001627 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1628 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1629 (VM_PFNMAP|VM_MIXEDMAP));
1630 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1631 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
Nick Piggine0dc0d82007-02-12 00:51:36 -08001632
Nick Piggin423bad62008-04-28 02:13:01 -07001633 if (addr < vma->vm_start || addr >= vma->vm_end)
1634 return -EFAULT;
Dan Williamsf25748e32016-01-15 16:56:43 -08001635 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001636 return -EINVAL;
1637
Dan Williams01c8f1c2016-01-15 16:56:40 -08001638 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001639
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001640 return ret;
Nick Piggine0dc0d82007-02-12 00:51:36 -08001641}
Andy Lutomirski1745cbc2015-12-29 20:12:20 -08001642EXPORT_SYMBOL(vm_insert_pfn_prot);
Nick Piggine0dc0d82007-02-12 00:51:36 -08001643
Nick Piggin423bad62008-04-28 02:13:01 -07001644int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
Dan Williams01c8f1c2016-01-15 16:56:40 -08001645 pfn_t pfn)
Nick Piggin423bad62008-04-28 02:13:01 -07001646{
Dan Williams87744ab2016-10-07 17:00:18 -07001647 pgprot_t pgprot = vma->vm_page_prot;
1648
Nick Piggin423bad62008-04-28 02:13:01 -07001649 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1650
1651 if (addr < vma->vm_start || addr >= vma->vm_end)
1652 return -EFAULT;
Dan Williams87744ab2016-10-07 17:00:18 -07001653 if (track_pfn_insert(vma, &pgprot, pfn))
1654 return -EINVAL;
Nick Piggin423bad62008-04-28 02:13:01 -07001655
1656 /*
1657 * If we don't have pte special, then we have to use the pfn_valid()
1658 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1659 * refcount the page if pfn_valid is true (hence insert_page rather
Hugh Dickins62eede62009-09-21 17:03:34 -07001660 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1661 * without pte special, it would there be refcounted as a normal page.
Nick Piggin423bad62008-04-28 02:13:01 -07001662 */
Dan Williams03fc2da2016-01-26 09:48:05 -08001663 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
Nick Piggin423bad62008-04-28 02:13:01 -07001664 struct page *page;
1665
Dan Williams03fc2da2016-01-26 09:48:05 -08001666 /*
1667 * At this point we are committed to insert_page()
1668 * regardless of whether the caller specified flags that
1669 * result in pfn_t_has_page() == false.
1670 */
1671 page = pfn_to_page(pfn_t_to_pfn(pfn));
Dan Williams87744ab2016-10-07 17:00:18 -07001672 return insert_page(vma, addr, page, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001673 }
Dan Williams87744ab2016-10-07 17:00:18 -07001674 return insert_pfn(vma, addr, pfn, pgprot);
Nick Piggin423bad62008-04-28 02:13:01 -07001675}
1676EXPORT_SYMBOL(vm_insert_mixed);
1677
Linus Torvaldsa145dd42005-11-30 09:35:19 -08001678/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679 * maps a range of physical memory into the requested pages. the old
1680 * mappings are removed. any references to nonexistent pages results
1681 * in null mappings (currently treated as "copy-on-access")
1682 */
1683static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1684 unsigned long addr, unsigned long end,
1685 unsigned long pfn, pgprot_t prot)
1686{
1687 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001688 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689
Hugh Dickinsc74df322005-10-29 18:16:23 -07001690 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 if (!pte)
1692 return -ENOMEM;
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001693 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001694 do {
1695 BUG_ON(!pte_none(*pte));
Nick Piggin7e675132008-04-28 02:13:00 -07001696 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 pfn++;
1698 } while (pte++, addr += PAGE_SIZE, addr != end);
Zachary Amsden6606c3e2006-09-30 23:29:33 -07001699 arch_leave_lazy_mmu_mode();
Hugh Dickinsc74df322005-10-29 18:16:23 -07001700 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 return 0;
1702}
1703
1704static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1705 unsigned long addr, unsigned long end,
1706 unsigned long pfn, pgprot_t prot)
1707{
1708 pmd_t *pmd;
1709 unsigned long next;
1710
1711 pfn -= addr >> PAGE_SHIFT;
1712 pmd = pmd_alloc(mm, pud, addr);
1713 if (!pmd)
1714 return -ENOMEM;
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08001715 VM_BUG_ON(pmd_trans_huge(*pmd));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 do {
1717 next = pmd_addr_end(addr, end);
1718 if (remap_pte_range(mm, pmd, addr, next,
1719 pfn + (addr >> PAGE_SHIFT), prot))
1720 return -ENOMEM;
1721 } while (pmd++, addr = next, addr != end);
1722 return 0;
1723}
1724
1725static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1726 unsigned long addr, unsigned long end,
1727 unsigned long pfn, pgprot_t prot)
1728{
1729 pud_t *pud;
1730 unsigned long next;
1731
1732 pfn -= addr >> PAGE_SHIFT;
1733 pud = pud_alloc(mm, pgd, addr);
1734 if (!pud)
1735 return -ENOMEM;
1736 do {
1737 next = pud_addr_end(addr, end);
1738 if (remap_pmd_range(mm, pud, addr, next,
1739 pfn + (addr >> PAGE_SHIFT), prot))
1740 return -ENOMEM;
1741 } while (pud++, addr = next, addr != end);
1742 return 0;
1743}
1744
Rolf Eike Beerbfa5bf62006-09-25 23:31:22 -07001745/**
1746 * remap_pfn_range - remap kernel memory to userspace
1747 * @vma: user vma to map to
1748 * @addr: target user address to start at
1749 * @pfn: physical address of kernel memory
1750 * @size: size of map area
1751 * @prot: page protection flags for this mapping
1752 *
1753 * Note: this is only safe if the mm semaphore is held when called.
1754 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1756 unsigned long pfn, unsigned long size, pgprot_t prot)
1757{
1758 pgd_t *pgd;
1759 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001760 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 struct mm_struct *mm = vma->vm_mm;
Yongji Xied5957d22016-05-20 16:57:41 -07001762 unsigned long remap_pfn = pfn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 int err;
1764
1765 /*
1766 * Physically remapped pages are special. Tell the
1767 * rest of the world about it:
1768 * VM_IO tells people not to look at these pages
1769 * (accesses can have side effects).
Linus Torvalds6aab3412005-11-28 14:34:23 -08001770 * VM_PFNMAP tells the core MM that the base pages are just
1771 * raw PFN mappings, and do not have a "struct page" associated
1772 * with them.
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001773 * VM_DONTEXPAND
1774 * Disable vma merging and expanding with mremap().
1775 * VM_DONTDUMP
1776 * Omit vma from core dump, even when VM_IO turned off.
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001777 *
1778 * There's a horrible special case to handle copy-on-write
1779 * behaviour that some programs depend on. We mark the "original"
1780 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001781 * See vm_normal_page() for details.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 */
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001783 if (is_cow_mapping(vma->vm_flags)) {
1784 if (addr != vma->vm_start || end != vma->vm_end)
1785 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001786 vma->vm_pgoff = pfn;
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001787 }
1788
Yongji Xied5957d22016-05-20 16:57:41 -07001789 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
Konstantin Khlebnikovb3b9c292012-10-08 16:28:34 -07001790 if (err)
venkatesh.pallipadi@intel.com3c8bb732008-12-18 11:41:27 -08001791 return -EINVAL;
Linus Torvaldsfb155c12005-12-11 19:46:02 -08001792
Konstantin Khlebnikov314e51b2012-10-08 16:29:02 -07001793 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794
1795 BUG_ON(addr >= end);
1796 pfn -= addr >> PAGE_SHIFT;
1797 pgd = pgd_offset(mm, addr);
1798 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 do {
1800 next = pgd_addr_end(addr, end);
1801 err = remap_pud_range(mm, pgd, addr, next,
1802 pfn + (addr >> PAGE_SHIFT), prot);
1803 if (err)
1804 break;
1805 } while (pgd++, addr = next, addr != end);
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001806
1807 if (err)
Yongji Xied5957d22016-05-20 16:57:41 -07001808 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
venkatesh.pallipadi@intel.com2ab64032008-12-18 11:41:29 -08001809
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 return err;
1811}
1812EXPORT_SYMBOL(remap_pfn_range);
1813
Linus Torvaldsb4cbb192013-04-16 13:45:37 -07001814/**
1815 * vm_iomap_memory - remap memory to userspace
1816 * @vma: user vma to map to
1817 * @start: start of area
1818 * @len: size of area
1819 *
1820 * This is a simplified io_remap_pfn_range() for common driver use. The
1821 * driver just needs to give us the physical memory range to be mapped,
1822 * we'll figure out the rest from the vma information.
1823 *
1824 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1825 * whatever write-combining details or similar.
1826 */
1827int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1828{
1829 unsigned long vm_len, pfn, pages;
1830
1831 /* Check that the physical memory area passed in looks valid */
1832 if (start + len < start)
1833 return -EINVAL;
1834 /*
1835 * You *really* shouldn't map things that aren't page-aligned,
1836 * but we've historically allowed it because IO memory might
1837 * just have smaller alignment.
1838 */
1839 len += start & ~PAGE_MASK;
1840 pfn = start >> PAGE_SHIFT;
1841 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1842 if (pfn + pages < pfn)
1843 return -EINVAL;
1844
1845 /* We start the mapping 'vm_pgoff' pages into the area */
1846 if (vma->vm_pgoff > pages)
1847 return -EINVAL;
1848 pfn += vma->vm_pgoff;
1849 pages -= vma->vm_pgoff;
1850
1851 /* Can we fit all of the mapping? */
1852 vm_len = vma->vm_end - vma->vm_start;
1853 if (vm_len >> PAGE_SHIFT > pages)
1854 return -EINVAL;
1855
1856 /* Ok, let it rip */
1857 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1858}
1859EXPORT_SYMBOL(vm_iomap_memory);
1860
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001861static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1862 unsigned long addr, unsigned long end,
1863 pte_fn_t fn, void *data)
1864{
1865 pte_t *pte;
1866 int err;
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001867 pgtable_t token;
Borislav Petkov94909912007-05-06 14:49:17 -07001868 spinlock_t *uninitialized_var(ptl);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001869
1870 pte = (mm == &init_mm) ?
1871 pte_alloc_kernel(pmd, addr) :
1872 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1873 if (!pte)
1874 return -ENOMEM;
1875
1876 BUG_ON(pmd_huge(*pmd));
1877
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001878 arch_enter_lazy_mmu_mode();
1879
Martin Schwidefsky2f569af2008-02-08 04:22:04 -08001880 token = pmd_pgtable(*pmd);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001881
1882 do {
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001883 err = fn(pte++, token, addr, data);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001884 if (err)
1885 break;
Daisuke Nishimurac36987e2009-10-26 16:50:23 -07001886 } while (addr += PAGE_SIZE, addr != end);
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001887
Jeremy Fitzhardinge38e0edb2009-01-06 14:39:21 -08001888 arch_leave_lazy_mmu_mode();
1889
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001890 if (mm != &init_mm)
1891 pte_unmap_unlock(pte-1, ptl);
1892 return err;
1893}
1894
1895static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1896 unsigned long addr, unsigned long end,
1897 pte_fn_t fn, void *data)
1898{
1899 pmd_t *pmd;
1900 unsigned long next;
1901 int err;
1902
Andi Kleenceb86872008-07-23 21:27:50 -07001903 BUG_ON(pud_huge(*pud));
1904
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001905 pmd = pmd_alloc(mm, pud, addr);
1906 if (!pmd)
1907 return -ENOMEM;
1908 do {
1909 next = pmd_addr_end(addr, end);
1910 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1911 if (err)
1912 break;
1913 } while (pmd++, addr = next, addr != end);
1914 return err;
1915}
1916
1917static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1918 unsigned long addr, unsigned long end,
1919 pte_fn_t fn, void *data)
1920{
1921 pud_t *pud;
1922 unsigned long next;
1923 int err;
1924
1925 pud = pud_alloc(mm, pgd, addr);
1926 if (!pud)
1927 return -ENOMEM;
1928 do {
1929 next = pud_addr_end(addr, end);
1930 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1931 if (err)
1932 break;
1933 } while (pud++, addr = next, addr != end);
1934 return err;
1935}
1936
1937/*
1938 * Scan a region of virtual memory, filling in page tables as necessary
1939 * and calling a provided function on each leaf page table.
1940 */
1941int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1942 unsigned long size, pte_fn_t fn, void *data)
1943{
1944 pgd_t *pgd;
1945 unsigned long next;
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001946 unsigned long end = addr + size;
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001947 int err;
1948
Mika Penttilä9cb65bc2016-03-15 14:56:45 -07001949 if (WARN_ON(addr >= end))
1950 return -EINVAL;
1951
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001952 pgd = pgd_offset(mm, addr);
1953 do {
1954 next = pgd_addr_end(addr, end);
1955 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1956 if (err)
1957 break;
1958 } while (pgd++, addr = next, addr != end);
Jeremy Fitzhardinge57250a52010-08-09 17:19:52 -07001959
Jeremy Fitzhardingeaee16b32007-05-06 14:48:54 -07001960 return err;
1961}
1962EXPORT_SYMBOL_GPL(apply_to_page_range);
1963
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964/*
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08001965 * handle_pte_fault chooses page fault handler according to an entry which was
1966 * read non-atomically. Before making any commitment, on those architectures
1967 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1968 * parts, do_swap_page must check under lock before unmapping the pte and
1969 * proceeding (but do_wp_page is only called after already making such a check;
Ryota Ozakia335b2e2011-02-10 13:56:28 +09001970 * and do_anonymous_page can safely check later on).
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001971 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001972static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001973 pte_t *page_table, pte_t orig_pte)
1974{
1975 int same = 1;
1976#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1977 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001978 spinlock_t *ptl = pte_lockptr(mm, pmd);
1979 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001980 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001981 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001982 }
1983#endif
1984 pte_unmap(page_table);
1985 return same;
1986}
1987
Atsushi Nemoto9de455b2006-12-12 17:14:55 +00001988static 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 -08001989{
Dan Williams0abdd7a2014-01-21 15:48:12 -08001990 debug_dma_assert_idle(src);
1991
Linus Torvalds6aab3412005-11-28 14:34:23 -08001992 /*
1993 * If the source page was a PFN mapping, we don't have
1994 * a "struct page" for it. We do a best-effort copy by
1995 * just copying from the original user address. If that
1996 * fails, we just zero-fill it. Live with it.
1997 */
1998 if (unlikely(!src)) {
Cong Wang9b04c5f2011-11-25 23:14:39 +08001999 void *kaddr = kmap_atomic(dst);
Linus Torvalds5d2a2dbbc2005-11-29 14:07:55 -08002000 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2001
2002 /*
2003 * This really shouldn't fail, because the page is there
2004 * in the page tables. But it might just be unreadable,
2005 * in which case we just give up and fill the result with
2006 * zeroes.
2007 */
2008 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
Jan Beulich3ecb01d2010-10-26 14:22:27 -07002009 clear_page(kaddr);
Cong Wang9b04c5f2011-11-25 23:14:39 +08002010 kunmap_atomic(kaddr);
Dmitriy Monakhovc4ec7b02006-10-19 23:29:08 -07002011 flush_dcache_page(dst);
Nick Piggin0ed361d2008-02-04 22:29:34 -08002012 } else
2013 copy_user_highpage(dst, src, va, vma);
Linus Torvalds6aab3412005-11-28 14:34:23 -08002014}
2015
Michal Hockoc20cd452016-01-14 15:20:12 -08002016static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2017{
2018 struct file *vm_file = vma->vm_file;
2019
2020 if (vm_file)
2021 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2022
2023 /*
2024 * Special mappings (e.g. VDSO) do not have any file so fake
2025 * a default GFP_KERNEL for them.
2026 */
2027 return GFP_KERNEL;
2028}
2029
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030/*
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002031 * Notify the address space that the page is about to become writable so that
2032 * it can prohibit this or wait for the page to get into an appropriate state.
2033 *
2034 * We do this without the lock held, so that it can sleep if it needs to.
2035 */
2036static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2037 unsigned long address)
2038{
2039 struct vm_fault vmf;
2040 int ret;
2041
2042 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2043 vmf.pgoff = page->index;
2044 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
Michal Hockoc20cd452016-01-14 15:20:12 -08002045 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002046 vmf.page = page;
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002047 vmf.cow_page = NULL;
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07002048
2049 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2050 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2051 return ret;
2052 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2053 lock_page(page);
2054 if (!page->mapping) {
2055 unlock_page(page);
2056 return 0; /* retry */
2057 }
2058 ret |= VM_FAULT_LOCKED;
2059 } else
2060 VM_BUG_ON_PAGE(!PageLocked(page), page);
2061 return ret;
2062}
2063
2064/*
Shachar Raindel4e047f82015-04-14 15:46:25 -07002065 * Handle write page faults for pages that can be reused in the current vma
2066 *
2067 * This can happen either due to the mapping being with the VM_SHARED flag,
2068 * or due to us being the last reference standing to the page. In either
2069 * case, all we need to do here is to mark the page as writable and update
2070 * any related book-keeping.
2071 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002072static inline int wp_page_reuse(struct fault_env *fe, pte_t orig_pte,
2073 struct page *page, int page_mkwrite, int dirty_shared)
2074 __releases(fe->ptl)
Shachar Raindel4e047f82015-04-14 15:46:25 -07002075{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002076 struct vm_area_struct *vma = fe->vma;
Shachar Raindel4e047f82015-04-14 15:46:25 -07002077 pte_t entry;
2078 /*
2079 * Clear the pages cpupid information as the existing
2080 * information potentially belongs to a now completely
2081 * unrelated process.
2082 */
2083 if (page)
2084 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2085
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002086 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel4e047f82015-04-14 15:46:25 -07002087 entry = pte_mkyoung(orig_pte);
2088 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002089 if (ptep_set_access_flags(vma, fe->address, fe->pte, entry, 1))
2090 update_mmu_cache(vma, fe->address, fe->pte);
2091 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002092
2093 if (dirty_shared) {
2094 struct address_space *mapping;
2095 int dirtied;
2096
2097 if (!page_mkwrite)
2098 lock_page(page);
2099
2100 dirtied = set_page_dirty(page);
2101 VM_BUG_ON_PAGE(PageAnon(page), page);
2102 mapping = page->mapping;
2103 unlock_page(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002104 put_page(page);
Shachar Raindel4e047f82015-04-14 15:46:25 -07002105
2106 if ((dirtied || page_mkwrite) && mapping) {
2107 /*
2108 * Some device drivers do not set page.mapping
2109 * but still dirty their pages
2110 */
2111 balance_dirty_pages_ratelimited(mapping);
2112 }
2113
2114 if (!page_mkwrite)
2115 file_update_time(vma->vm_file);
2116 }
2117
2118 return VM_FAULT_WRITE;
2119}
2120
2121/*
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002122 * Handle the case of a page which we actually need to copy to a new page.
2123 *
2124 * Called with mmap_sem locked and the old page referenced, but
2125 * without the ptl held.
2126 *
2127 * High level logic flow:
2128 *
2129 * - Allocate a page, copy the content of the old page to the new one.
2130 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2131 * - Take the PTL. If the pte changed, bail out and release the allocated page
2132 * - If the pte is still the way we remember it, update the page table and all
2133 * relevant references. This includes dropping the reference the page-table
2134 * held to the old page, as well as updating the rmap.
2135 * - In any case, unlock the PTL and drop the reference we took to the old page.
2136 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002137static int wp_page_copy(struct fault_env *fe, pte_t orig_pte,
2138 struct page *old_page)
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002139{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002140 struct vm_area_struct *vma = fe->vma;
2141 struct mm_struct *mm = vma->vm_mm;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002142 struct page *new_page = NULL;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002143 pte_t entry;
2144 int page_copied = 0;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002145 const unsigned long mmun_start = fe->address & PAGE_MASK;
2146 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002147 struct mem_cgroup *memcg;
2148
2149 if (unlikely(anon_vma_prepare(vma)))
2150 goto oom;
2151
2152 if (is_zero_pfn(pte_pfn(orig_pte))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002153 new_page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002154 if (!new_page)
2155 goto oom;
2156 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002157 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2158 fe->address);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002159 if (!new_page)
2160 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002161 cow_user_page(new_page, old_page, fe->address, vma);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002162 }
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002163
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002164 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002165 goto oom_free_new;
2166
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002167 __SetPageUptodate(new_page);
2168
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002169 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2170
2171 /*
2172 * Re-check the pte - we dropped the lock
2173 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002174 fe->pte = pte_offset_map_lock(mm, fe->pmd, fe->address, &fe->ptl);
2175 if (likely(pte_same(*fe->pte, orig_pte))) {
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002176 if (old_page) {
2177 if (!PageAnon(old_page)) {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002178 dec_mm_counter_fast(mm,
2179 mm_counter_file(old_page));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002180 inc_mm_counter_fast(mm, MM_ANONPAGES);
2181 }
2182 } else {
2183 inc_mm_counter_fast(mm, MM_ANONPAGES);
2184 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002185 flush_cache_page(vma, fe->address, pte_pfn(orig_pte));
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002186 entry = mk_pte(new_page, vma->vm_page_prot);
2187 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2188 /*
2189 * Clear the pte entry and flush it first, before updating the
2190 * pte with the new entry. This will avoid a race condition
2191 * seen in the presence of one thread doing SMC and another
2192 * thread doing COW.
2193 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002194 ptep_clear_flush_notify(vma, fe->address, fe->pte);
2195 page_add_new_anon_rmap(new_page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002196 mem_cgroup_commit_charge(new_page, memcg, false, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002197 lru_cache_add_active_or_unevictable(new_page, vma);
2198 /*
2199 * We call the notify macro here because, when using secondary
2200 * mmu page tables (such as kvm shadow page tables), we want the
2201 * new page to be mapped directly into the secondary page table.
2202 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002203 set_pte_at_notify(mm, fe->address, fe->pte, entry);
2204 update_mmu_cache(vma, fe->address, fe->pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002205 if (old_page) {
2206 /*
2207 * Only after switching the pte to the new page may
2208 * we remove the mapcount here. Otherwise another
2209 * process may come and find the rmap count decremented
2210 * before the pte is switched to the new page, and
2211 * "reuse" the old page writing into it while our pte
2212 * here still points into it and can be read by other
2213 * threads.
2214 *
2215 * The critical issue is to order this
2216 * page_remove_rmap with the ptp_clear_flush above.
2217 * Those stores are ordered by (if nothing else,)
2218 * the barrier present in the atomic_add_negative
2219 * in page_remove_rmap.
2220 *
2221 * Then the TLB flush in ptep_clear_flush ensures that
2222 * no process can access the old page before the
2223 * decremented mapcount is visible. And the old page
2224 * cannot be reused until after the decremented
2225 * mapcount is visible. So transitively, TLBs to
2226 * old page will be flushed before it can be reused.
2227 */
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002228 page_remove_rmap(old_page, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002229 }
2230
2231 /* Free the old page.. */
2232 new_page = old_page;
2233 page_copied = 1;
2234 } else {
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002235 mem_cgroup_cancel_charge(new_page, memcg, false);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002236 }
2237
2238 if (new_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002239 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002240
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002241 pte_unmap_unlock(fe->pte, fe->ptl);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002242 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2243 if (old_page) {
2244 /*
2245 * Don't let another task, with possibly unlocked vma,
2246 * keep the mlocked page.
2247 */
2248 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2249 lock_page(old_page); /* LRU manipulation */
Kirill A. Shutemove90309c2016-01-15 16:54:33 -08002250 if (PageMlocked(old_page))
2251 munlock_vma_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002252 unlock_page(old_page);
2253 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002254 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002255 }
2256 return page_copied ? VM_FAULT_WRITE : 0;
2257oom_free_new:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002258 put_page(new_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002259oom:
2260 if (old_page)
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002261 put_page(old_page);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002262 return VM_FAULT_OOM;
2263}
2264
Boaz Harroshdd906182015-04-15 16:15:11 -07002265/*
2266 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2267 * mapping
2268 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002269static int wp_pfn_shared(struct fault_env *fe, pte_t orig_pte)
Boaz Harroshdd906182015-04-15 16:15:11 -07002270{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002271 struct vm_area_struct *vma = fe->vma;
2272
Boaz Harroshdd906182015-04-15 16:15:11 -07002273 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2274 struct vm_fault vmf = {
2275 .page = NULL,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002276 .pgoff = linear_page_index(vma, fe->address),
2277 .virtual_address =
2278 (void __user *)(fe->address & PAGE_MASK),
Boaz Harroshdd906182015-04-15 16:15:11 -07002279 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2280 };
2281 int ret;
2282
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002283 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002284 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2285 if (ret & VM_FAULT_ERROR)
2286 return ret;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002287 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2288 &fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002289 /*
2290 * We might have raced with another page fault while we
2291 * released the pte_offset_map_lock.
2292 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002293 if (!pte_same(*fe->pte, orig_pte)) {
2294 pte_unmap_unlock(fe->pte, fe->ptl);
Boaz Harroshdd906182015-04-15 16:15:11 -07002295 return 0;
2296 }
2297 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002298 return wp_page_reuse(fe, orig_pte, NULL, 0, 0);
Boaz Harroshdd906182015-04-15 16:15:11 -07002299}
2300
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002301static int wp_page_shared(struct fault_env *fe, pte_t orig_pte,
2302 struct page *old_page)
2303 __releases(fe->ptl)
Shachar Raindel93e478d2015-04-14 15:46:35 -07002304{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002305 struct vm_area_struct *vma = fe->vma;
Shachar Raindel93e478d2015-04-14 15:46:35 -07002306 int page_mkwrite = 0;
2307
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002308 get_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002309
Shachar Raindel93e478d2015-04-14 15:46:35 -07002310 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2311 int tmp;
2312
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002313 pte_unmap_unlock(fe->pte, fe->ptl);
2314 tmp = do_page_mkwrite(vma, old_page, fe->address);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002315 if (unlikely(!tmp || (tmp &
2316 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002317 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002318 return tmp;
2319 }
2320 /*
2321 * Since we dropped the lock we need to revalidate
2322 * the PTE as someone else may have changed it. If
2323 * they did, we just return, as we can count on the
2324 * MMU to tell us if they didn't also make it writable.
2325 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002326 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2327 &fe->ptl);
2328 if (!pte_same(*fe->pte, orig_pte)) {
Shachar Raindel93e478d2015-04-14 15:46:35 -07002329 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002330 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002331 put_page(old_page);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002332 return 0;
2333 }
2334 page_mkwrite = 1;
2335 }
2336
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002337 return wp_page_reuse(fe, orig_pte, old_page, page_mkwrite, 1);
Shachar Raindel93e478d2015-04-14 15:46:35 -07002338}
2339
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002340/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 * This routine handles present pages, when users try to write
2342 * to a shared page. It is done by copying the page to a new address
2343 * and decrementing the shared-page counter for the old page.
2344 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 * Note that this routine assumes that the protection checks have been
2346 * done by the caller (the low-level page fault routine in most cases).
2347 * Thus we can safely just mark it writable once we've done any necessary
2348 * COW.
2349 *
2350 * We also mark the page dirty at this point even though the page will
2351 * change only once the write actually happens. This avoids a few races,
2352 * and potentially makes it more efficient.
2353 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002354 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2355 * but allow concurrent faults), with pte both mapped and locked.
2356 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002358static int do_wp_page(struct fault_env *fe, pte_t orig_pte)
2359 __releases(fe->ptl)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002361 struct vm_area_struct *vma = fe->vma;
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002362 struct page *old_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002364 old_page = vm_normal_page(vma, fe->address, orig_pte);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002365 if (!old_page) {
2366 /*
Peter Feiner64e45502014-10-13 15:55:46 -07002367 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2368 * VM_PFNMAP VMA.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002369 *
2370 * We should not cow pages in a shared writeable mapping.
Boaz Harroshdd906182015-04-15 16:15:11 -07002371 * Just mark the pages writable and/or call ops->pfn_mkwrite.
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002372 */
2373 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2374 (VM_WRITE|VM_SHARED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002375 return wp_pfn_shared(fe, orig_pte);
Shachar Raindel2f38ab22015-04-14 15:46:32 -07002376
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002377 pte_unmap_unlock(fe->pte, fe->ptl);
2378 return wp_page_copy(fe, orig_pte, old_page);
Peter Zijlstra251b97f2008-07-04 09:59:24 -07002379 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002381 /*
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002382 * Take out anonymous pages first, anonymous shared vmas are
2383 * not dirty accountable.
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002384 */
Hugh Dickins9a840892009-09-21 17:02:01 -07002385 if (PageAnon(old_page) && !PageKsm(old_page)) {
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002386 int total_mapcount;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002387 if (!trylock_page(old_page)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002388 get_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002389 pte_unmap_unlock(fe->pte, fe->ptl);
Hugh Dickinsab967d82009-01-06 14:39:33 -08002390 lock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002391 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2392 fe->address, &fe->ptl);
2393 if (!pte_same(*fe->pte, orig_pte)) {
Hugh Dickinsab967d82009-01-06 14:39:33 -08002394 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002395 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002396 put_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002397 return 0;
Hugh Dickinsab967d82009-01-06 14:39:33 -08002398 }
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002399 put_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002400 }
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002401 if (reuse_swap_page(old_page, &total_mapcount)) {
2402 if (total_mapcount == 1) {
2403 /*
2404 * The page is all ours. Move it to
2405 * our anon_vma so the rmap code will
2406 * not search our parent or siblings.
2407 * Protected against the rmap code by
2408 * the page lock.
2409 */
Hugh Dickins5a499732016-07-14 12:07:38 -07002410 page_move_anon_rmap(old_page, vma);
Andrea Arcangeli6d0a07e2016-05-12 15:42:25 -07002411 }
Michel Lespinasseb009c022011-01-13 15:46:07 -08002412 unlock_page(old_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002413 return wp_page_reuse(fe, orig_pte, old_page, 0, 0);
Michel Lespinasseb009c022011-01-13 15:46:07 -08002414 }
Hugh Dickinsab967d82009-01-06 14:39:33 -08002415 unlock_page(old_page);
Peter Zijlstraee6a6452006-09-25 23:31:00 -07002416 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
Peter Zijlstrad08b3852006-09-25 23:30:57 -07002417 (VM_WRITE|VM_SHARED))) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002418 return wp_page_shared(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420
2421 /*
2422 * Ok, we need to copy. Oh, well..
2423 */
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002424 get_page(old_page);
Shachar Raindel28766802015-04-14 15:46:29 -07002425
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002426 pte_unmap_unlock(fe->pte, fe->ptl);
2427 return wp_page_copy(fe, orig_pte, old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428}
2429
Peter Zijlstra97a89412011-05-24 17:12:04 -07002430static void unmap_mapping_range_vma(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 unsigned long start_addr, unsigned long end_addr,
2432 struct zap_details *details)
2433{
Al Virof5cc4ee2012-03-05 14:14:20 -05002434 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435}
2436
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002437static inline void unmap_mapping_range_tree(struct rb_root *root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 struct zap_details *details)
2439{
2440 struct vm_area_struct *vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 pgoff_t vba, vea, zba, zea;
2442
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002443 vma_interval_tree_foreach(vma, root,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 details->first_index, details->last_index) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
2446 vba = vma->vm_pgoff;
Libind6e93212013-07-03 15:01:26 -07002447 vea = vba + vma_pages(vma) - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 zba = details->first_index;
2449 if (zba < vba)
2450 zba = vba;
2451 zea = details->last_index;
2452 if (zea > vea)
2453 zea = vea;
2454
Peter Zijlstra97a89412011-05-24 17:12:04 -07002455 unmap_mapping_range_vma(vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2457 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
Peter Zijlstra97a89412011-05-24 17:12:04 -07002458 details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 }
2460}
2461
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462/**
Kirill A. Shutemov8a5f14a2015-02-10 14:09:49 -08002463 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2464 * address_space corresponding to the specified page range in the underlying
2465 * file.
2466 *
Martin Waitz3d410882005-06-23 22:05:21 -07002467 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 * @holebegin: byte in first page to unmap, relative to the start of
2469 * the underlying file. This will be rounded down to a PAGE_SIZE
npiggin@suse.de25d9e2d2009-08-21 02:35:05 +10002470 * boundary. Note that this is different from truncate_pagecache(), which
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471 * must keep the partial page. In contrast, we must get rid of
2472 * partial pages.
2473 * @holelen: size of prospective hole in bytes. This will be rounded
2474 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2475 * end of the file.
2476 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2477 * but 0 when invalidating pagecache, don't throw away private data.
2478 */
2479void unmap_mapping_range(struct address_space *mapping,
2480 loff_t const holebegin, loff_t const holelen, int even_cows)
2481{
Michal Hockoaac45362016-03-25 14:20:24 -07002482 struct zap_details details = { };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 pgoff_t hba = holebegin >> PAGE_SHIFT;
2484 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2485
2486 /* Check for overflow. */
2487 if (sizeof(holelen) > sizeof(hlen)) {
2488 long long holeend =
2489 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2490 if (holeend & ~(long long)ULONG_MAX)
2491 hlen = ULONG_MAX - hba + 1;
2492 }
2493
2494 details.check_mapping = even_cows? NULL: mapping;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 details.first_index = hba;
2496 details.last_index = hba + hlen - 1;
2497 if (details.last_index < details.first_index)
2498 details.last_index = ULONG_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002500 i_mmap_lock_write(mapping);
Michel Lespinasse6b2dbba2012-10-08 16:31:25 -07002501 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 unmap_mapping_range_tree(&mapping->i_mmap, &details);
Kirill A. Shutemov46c043e2015-09-08 14:59:42 -07002503 i_mmap_unlock_write(mapping);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504}
2505EXPORT_SYMBOL(unmap_mapping_range);
2506
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002508 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2509 * but allow concurrent faults), and pte mapped but not yet locked.
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002510 * We return with pte unmapped and unlocked.
2511 *
2512 * We return with the mmap_sem locked or unlocked in the same cases
2513 * as does filemap_fault().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002515int do_swap_page(struct fault_env *fe, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002516{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002517 struct vm_area_struct *vma = fe->vma;
Hugh Dickins56f31802013-02-22 16:36:10 -08002518 struct page *page, *swapcache;
Johannes Weiner00501b52014-08-08 14:19:20 -07002519 struct mem_cgroup *memcg;
Hugh Dickins65500d22005-10-29 18:15:59 -07002520 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 pte_t pte;
Michel Lespinassed065bd82010-10-26 14:21:57 -07002522 int locked;
Rik van Rielad8c2ee2010-08-09 17:19:48 -07002523 int exclusive = 0;
Nick Piggin83c54072007-07-19 01:47:05 -07002524 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002526 if (!pte_unmap_same(vma->vm_mm, fe->pmd, fe->pte, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002527 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07002528
2529 entry = pte_to_swp_entry(orig_pte);
Andi Kleend1737fd2009-09-16 11:50:06 +02002530 if (unlikely(non_swap_entry(entry))) {
2531 if (is_migration_entry(entry)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002532 migration_entry_wait(vma->vm_mm, fe->pmd, fe->address);
Andi Kleend1737fd2009-09-16 11:50:06 +02002533 } else if (is_hwpoison_entry(entry)) {
2534 ret = VM_FAULT_HWPOISON;
2535 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002536 print_bad_pte(vma, fe->address, orig_pte, NULL);
Hugh Dickinsd99be1a2009-12-14 17:59:04 -08002537 ret = VM_FAULT_SIGBUS;
Andi Kleend1737fd2009-09-16 11:50:06 +02002538 }
Christoph Lameter06972122006-06-23 02:03:35 -07002539 goto out;
2540 }
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002541 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 page = lookup_swap_cache(entry);
2543 if (!page) {
Hugh Dickins02098fe2008-02-04 22:28:42 -08002544 page = swapin_readahead(entry,
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002545 GFP_HIGHUSER_MOVABLE, vma, fe->address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 if (!page) {
2547 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002548 * Back out if somebody else faulted in this pte
2549 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002551 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd,
2552 fe->address, &fe->ptl);
2553 if (likely(pte_same(*fe->pte, orig_pte)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 ret = VM_FAULT_OOM;
Shailabh Nagar0ff92242006-07-14 00:24:37 -07002555 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins65500d22005-10-29 18:15:59 -07002556 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 }
2558
2559 /* Had to read the page from swap area: Major fault */
2560 ret = VM_FAULT_MAJOR;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002561 count_vm_event(PGMAJFAULT);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002562 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
Andi Kleend1737fd2009-09-16 11:50:06 +02002563 } else if (PageHWPoison(page)) {
Wu Fengguang71f72522009-12-16 12:19:58 +01002564 /*
2565 * hwpoisoned dirty swapcache pages are kept for killing
2566 * owner processes (which may be unknown at hwpoison time)
2567 */
Andi Kleend1737fd2009-09-16 11:50:06 +02002568 ret = VM_FAULT_HWPOISON;
2569 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Hugh Dickins56f31802013-02-22 16:36:10 -08002570 swapcache = page;
Andi Kleen4779cb32009-10-14 01:51:41 +02002571 goto out_release;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 }
2573
Hugh Dickins56f31802013-02-22 16:36:10 -08002574 swapcache = page;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002575 locked = lock_page_or_retry(page, vma->vm_mm, fe->flags);
Rik van Riele709ffd2012-05-29 15:06:18 -07002576
Balbir Singh20a10222007-11-14 17:00:33 -08002577 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
Michel Lespinassed065bd82010-10-26 14:21:57 -07002578 if (!locked) {
2579 ret |= VM_FAULT_RETRY;
2580 goto out_release;
2581 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002583 /*
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002584 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2585 * release the swapcache from under us. The page pin, and pte_same
2586 * test below, are not enough to exclude that. Even if it is still
2587 * swapcache, we need to check that the page's swap has not changed.
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002588 */
Hugh Dickins31c4a3d2010-09-19 19:40:22 -07002589 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002590 goto out_page;
2591
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002592 page = ksm_might_need_to_copy(page, vma, fe->address);
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002593 if (unlikely(!page)) {
2594 ret = VM_FAULT_OOM;
2595 page = swapcache;
Hugh Dickinscbf86cf2013-02-22 16:35:08 -08002596 goto out_page;
Hugh Dickins5ad64682009-12-14 17:59:24 -08002597 }
2598
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002599 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2600 &memcg, false)) {
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002601 ret = VM_FAULT_OOM;
Johannes Weinerbc43f752009-04-30 15:08:08 -07002602 goto out_page;
KAMEZAWA Hiroyuki073e5872008-10-18 20:28:08 -07002603 }
2604
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002606 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002608 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2609 &fe->ptl);
2610 if (unlikely(!pte_same(*fe->pte, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07002611 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002612
2613 if (unlikely(!PageUptodate(page))) {
2614 ret = VM_FAULT_SIGBUS;
2615 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616 }
2617
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002618 /*
2619 * The page isn't present yet, go ahead with the fault.
2620 *
2621 * Be careful about the sequence of operations here.
2622 * To get its accounting right, reuse_swap_page() must be called
2623 * while the page is counted on swap but not yet in mapcount i.e.
2624 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2625 * must be called after the swap_free(), or it will never succeed.
KAMEZAWA Hiroyuki8c7c6e342009-01-07 18:08:00 -08002626 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002628 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2629 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 pte = mk_pte(page, vma->vm_page_prot);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002631 if ((fe->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002633 fe->flags &= ~FAULT_FLAG_WRITE;
Andrea Arcangeli9a5b4892010-08-09 17:19:49 -07002634 ret |= VM_FAULT_WRITE;
Kirill A. Shutemovd281ee62016-01-15 16:52:16 -08002635 exclusive = RMAP_EXCLUSIVE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 flush_icache_page(vma, page);
Cyrill Gorcunov179ef712013-08-13 16:00:49 -07002638 if (pte_swp_soft_dirty(orig_pte))
2639 pte = pte_mksoft_dirty(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002640 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
Johannes Weiner00501b52014-08-08 14:19:20 -07002641 if (page == swapcache) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002642 do_page_add_anon_rmap(page, vma, fe->address, exclusive);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002643 mem_cgroup_commit_charge(page, memcg, true, false);
Minchan Kim1a8018f2016-08-02 14:02:25 -07002644 activate_page(page);
Johannes Weiner00501b52014-08-08 14:19:20 -07002645 } else { /* ksm created a completely new copy */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002646 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002647 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002648 lru_cache_add_active_or_unevictable(page, vma);
2649 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002651 swap_free(entry);
Vladimir Davydov5ccc5ab2016-01-20 15:03:10 -08002652 if (mem_cgroup_swap_full(page) ||
2653 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08002654 try_to_free_swap(page);
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002655 unlock_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002656 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002657 /*
2658 * Hold the lock to avoid the swap entry to be reused
2659 * until we take the PT lock for the pte_same() check
2660 * (to avoid false positives from pte_same). For
2661 * further safety release the lock after the swap_free
2662 * so that the swap count won't change under a
2663 * parallel locked swapcache.
2664 */
2665 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002666 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002667 }
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07002668
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002669 if (fe->flags & FAULT_FLAG_WRITE) {
2670 ret |= do_wp_page(fe, pte);
Hugh Dickins61469f12008-03-04 14:29:04 -08002671 if (ret & VM_FAULT_ERROR)
2672 ret &= VM_FAULT_ERROR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 goto out;
2674 }
2675
2676 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002677 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002678unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002679 pte_unmap_unlock(fe->pte, fe->ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680out:
2681 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07002682out_nomap:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002683 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002684 pte_unmap_unlock(fe->pte, fe->ptl);
Johannes Weinerbc43f752009-04-30 15:08:08 -07002685out_page:
Kirill Korotaevb8107482005-05-16 21:53:50 -07002686 unlock_page(page);
Andi Kleen4779cb32009-10-14 01:51:41 +02002687out_release:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002688 put_page(page);
Hugh Dickins56f31802013-02-22 16:36:10 -08002689 if (page != swapcache) {
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002690 unlock_page(swapcache);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002691 put_page(swapcache);
Andrea Arcangeli4969c112010-09-09 16:37:52 -07002692 }
Hugh Dickins65500d22005-10-29 18:15:59 -07002693 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694}
2695
2696/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002697 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2698 * but allow concurrent faults), and pte mapped but not yet locked.
2699 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002701static int do_anonymous_page(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002703 struct vm_area_struct *vma = fe->vma;
Johannes Weiner00501b52014-08-08 14:19:20 -07002704 struct mem_cgroup *memcg;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002705 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03002708 /* File mapping without ->vm_ops ? */
2709 if (vma->vm_flags & VM_SHARED)
2710 return VM_FAULT_SIGBUS;
2711
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002712 /*
2713 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2714 * pte_offset_map() on pmds where a huge pmd might be created
2715 * from a different thread.
2716 *
2717 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2718 * parallel threads are excluded by other means.
2719 *
2720 * Here we only have down_read(mmap_sem).
2721 */
2722 if (pte_alloc(vma->vm_mm, fe->pmd, fe->address))
2723 return VM_FAULT_OOM;
2724
2725 /* See the comment in pte_alloc_one_map() */
2726 if (unlikely(pmd_trans_unstable(fe->pmd)))
2727 return 0;
2728
Linus Torvalds11ac5522010-08-14 11:44:56 -07002729 /* Use the zero-page for reads */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002730 if (!(fe->flags & FAULT_FLAG_WRITE) &&
2731 !mm_forbids_zeropage(vma->vm_mm)) {
2732 entry = pte_mkspecial(pfn_pte(my_zero_pfn(fe->address),
Hugh Dickins62eede62009-09-21 17:03:34 -07002733 vma->vm_page_prot));
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002734 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2735 &fe->ptl);
2736 if (!pte_none(*fe->pte))
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002737 goto unlock;
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002738 /* Deliver the page fault to userland, check inside PT lock */
2739 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002740 pte_unmap_unlock(fe->pte, fe->ptl);
2741 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002742 }
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002743 goto setpte;
2744 }
2745
Nick Piggin557ed1f2007-10-16 01:24:40 -07002746 /* Allocate our own private page. */
Nick Piggin557ed1f2007-10-16 01:24:40 -07002747 if (unlikely(anon_vma_prepare(vma)))
2748 goto oom;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002749 page = alloc_zeroed_user_highpage_movable(vma, fe->address);
Nick Piggin557ed1f2007-10-16 01:24:40 -07002750 if (!page)
2751 goto oom;
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002752
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002753 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
Mel Gormaneb3c24f2015-06-24 16:57:27 -07002754 goto oom_free_page;
2755
Minchan Kim52f37622013-04-29 15:08:15 -07002756 /*
2757 * The memory barrier inside __SetPageUptodate makes sure that
2758 * preceeding stores to the page contents become visible before
2759 * the set_pte_at() write.
2760 */
Nick Piggin0ed361d2008-02-04 22:29:34 -08002761 __SetPageUptodate(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762
Nick Piggin557ed1f2007-10-16 01:24:40 -07002763 entry = mk_pte(page, vma->vm_page_prot);
Hugh Dickins1ac0cb52009-09-21 17:03:29 -07002764 if (vma->vm_flags & VM_WRITE)
2765 entry = pte_mkwrite(pte_mkdirty(entry));
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002766
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002767 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2768 &fe->ptl);
2769 if (!pte_none(*fe->pte))
Nick Piggin557ed1f2007-10-16 01:24:40 -07002770 goto release;
Hugh Dickins9ba69292009-09-21 17:02:20 -07002771
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002772 /* Deliver the page fault to userland, check inside PT lock */
2773 if (userfaultfd_missing(vma)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002774 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002775 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002776 put_page(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002777 return handle_userfault(fe, VM_UFFD_MISSING);
Andrea Arcangeli6b251fc2015-09-04 15:46:20 -07002778 }
2779
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002780 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2781 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002782 mem_cgroup_commit_charge(page, memcg, false, false);
Johannes Weiner00501b52014-08-08 14:19:20 -07002783 lru_cache_add_active_or_unevictable(page, vma);
Hugh Dickinsa13ea5b2009-09-21 17:03:30 -07002784setpte:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002785 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786
2787 /* No need to invalidate - it was non-present before */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002788 update_mmu_cache(vma, fe->address, fe->pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07002789unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002790 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07002791 return 0;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002792release:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08002793 mem_cgroup_cancel_charge(page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002794 put_page(page);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002795 goto unlock;
Balbir Singh8a9f3cc2008-02-07 00:13:53 -08002796oom_free_page:
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002797 put_page(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07002798oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 return VM_FAULT_OOM;
2800}
2801
Paul Cassella9a95f3c2014-08-06 16:07:24 -07002802/*
2803 * The mmap_sem must have been held on entry, and may have been
2804 * released depending on flags and vma->vm_ops->fault() return value.
2805 * See filemap_fault() and __lock_page_retry().
2806 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002807static int __do_fault(struct fault_env *fe, pgoff_t pgoff,
2808 struct page *cow_page, struct page **page, void **entry)
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002809{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002810 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002811 struct vm_fault vmf;
2812 int ret;
2813
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002814 vmf.virtual_address = (void __user *)(fe->address & PAGE_MASK);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002815 vmf.pgoff = pgoff;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002816 vmf.flags = fe->flags;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002817 vmf.page = NULL;
Michal Hockoc20cd452016-01-14 15:20:12 -08002818 vmf.gfp_mask = __get_fault_gfp_mask(vma);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08002819 vmf.cow_page = cow_page;
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002820
2821 ret = vma->vm_ops->fault(vma, &vmf);
2822 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2823 return ret;
Jan Karabc2466e2016-05-12 18:29:19 +02002824 if (ret & VM_FAULT_DAX_LOCKED) {
2825 *entry = vmf.entry;
2826 return ret;
2827 }
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002828
2829 if (unlikely(PageHWPoison(vmf.page))) {
2830 if (ret & VM_FAULT_LOCKED)
2831 unlock_page(vmf.page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03002832 put_page(vmf.page);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07002833 return VM_FAULT_HWPOISON;
2834 }
2835
2836 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2837 lock_page(vmf.page);
2838 else
2839 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2840
2841 *page = vmf.page;
2842 return ret;
2843}
2844
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002845static int pte_alloc_one_map(struct fault_env *fe)
2846{
2847 struct vm_area_struct *vma = fe->vma;
2848
2849 if (!pmd_none(*fe->pmd))
2850 goto map_pte;
2851 if (fe->prealloc_pte) {
2852 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2853 if (unlikely(!pmd_none(*fe->pmd))) {
2854 spin_unlock(fe->ptl);
2855 goto map_pte;
2856 }
2857
2858 atomic_long_inc(&vma->vm_mm->nr_ptes);
2859 pmd_populate(vma->vm_mm, fe->pmd, fe->prealloc_pte);
2860 spin_unlock(fe->ptl);
2861 fe->prealloc_pte = 0;
2862 } else if (unlikely(pte_alloc(vma->vm_mm, fe->pmd, fe->address))) {
2863 return VM_FAULT_OOM;
2864 }
2865map_pte:
2866 /*
2867 * If a huge pmd materialized under us just retry later. Use
2868 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2869 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2870 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2871 * in a different thread of this mm, in turn leading to a misleading
2872 * pmd_trans_huge() retval. All we have to ensure is that it is a
2873 * regular pmd that we can walk with pte_offset_map() and we can do that
2874 * through an atomic read in C, which is what pmd_trans_unstable()
2875 * provides.
2876 */
2877 if (pmd_trans_unstable(fe->pmd) || pmd_devmap(*fe->pmd))
2878 return VM_FAULT_NOPAGE;
2879
2880 fe->pte = pte_offset_map_lock(vma->vm_mm, fe->pmd, fe->address,
2881 &fe->ptl);
2882 return 0;
2883}
2884
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002885#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002886
2887#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2888static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2889 unsigned long haddr)
2890{
2891 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2892 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2893 return false;
2894 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2895 return false;
2896 return true;
2897}
2898
2899static int do_set_pmd(struct fault_env *fe, struct page *page)
2900{
2901 struct vm_area_struct *vma = fe->vma;
2902 bool write = fe->flags & FAULT_FLAG_WRITE;
2903 unsigned long haddr = fe->address & HPAGE_PMD_MASK;
2904 pmd_t entry;
2905 int i, ret;
2906
2907 if (!transhuge_vma_suitable(vma, haddr))
2908 return VM_FAULT_FALLBACK;
2909
2910 ret = VM_FAULT_FALLBACK;
2911 page = compound_head(page);
2912
2913 fe->ptl = pmd_lock(vma->vm_mm, fe->pmd);
2914 if (unlikely(!pmd_none(*fe->pmd)))
2915 goto out;
2916
2917 for (i = 0; i < HPAGE_PMD_NR; i++)
2918 flush_icache_page(vma, page + i);
2919
2920 entry = mk_huge_pmd(page, vma->vm_page_prot);
2921 if (write)
2922 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2923
2924 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2925 page_add_file_rmap(page, true);
2926
2927 set_pmd_at(vma->vm_mm, haddr, fe->pmd, entry);
2928
2929 update_mmu_cache_pmd(vma, haddr, fe->pmd);
2930
2931 /* fault is handled */
2932 ret = 0;
Kirill A. Shutemov95ecedc2016-07-26 15:25:31 -07002933 count_vm_event(THP_FILE_MAPPED);
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002934out:
2935 spin_unlock(fe->ptl);
2936 return ret;
2937}
2938#else
2939static int do_set_pmd(struct fault_env *fe, struct page *page)
2940{
2941 BUILD_BUG();
2942 return 0;
2943}
2944#endif
2945
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002946/**
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002947 * alloc_set_pte - setup new PTE entry for given page and add reverse page
2948 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002949 *
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002950 * @fe: fault environment
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002951 * @memcg: memcg to charge page (only for private mappings)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002952 * @page: page to map
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002953 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002954 * Caller must take care of unlocking fe->ptl, if fe->pte is non-NULL on return.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07002955 *
2956 * Target users are page handler itself and implementations of
2957 * vm_ops->map_pages.
2958 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002959int alloc_set_pte(struct fault_env *fe, struct mem_cgroup *memcg,
2960 struct page *page)
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002961{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002962 struct vm_area_struct *vma = fe->vma;
2963 bool write = fe->flags & FAULT_FLAG_WRITE;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002964 pte_t entry;
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002965 int ret;
2966
Kirill A. Shutemove496cf32016-07-26 15:26:35 -07002967 if (pmd_none(*fe->pmd) && PageTransCompound(page) &&
2968 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002969 /* THP on COW? */
2970 VM_BUG_ON_PAGE(memcg, page);
2971
2972 ret = do_set_pmd(fe, page);
2973 if (ret != VM_FAULT_FALLBACK)
2974 return ret;
2975 }
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002976
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002977 if (!fe->pte) {
Kirill A. Shutemov10102452016-07-26 15:25:29 -07002978 ret = pte_alloc_one_map(fe);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002979 if (ret)
2980 return ret;
2981 }
2982
2983 /* Re-check under ptl */
2984 if (unlikely(!pte_none(*fe->pte)))
2985 return VM_FAULT_NOPAGE;
2986
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002987 flush_icache_page(vma, page);
2988 entry = mk_pte(page, vma->vm_page_prot);
2989 if (write)
2990 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002991 /* copy-on-write page */
2992 if (write && !(vma->vm_flags & VM_SHARED)) {
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002993 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07002994 page_add_new_anon_rmap(page, vma, fe->address, false);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07002995 mem_cgroup_commit_charge(page, memcg, false, false);
2996 lru_cache_add_active_or_unevictable(page, vma);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07002997 } else {
Jerome Marchandeca56ff2016-01-14 15:19:26 -08002998 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
Kirill A. Shutemovdd78fed2016-07-26 15:25:26 -07002999 page_add_file_rmap(page, false);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003000 }
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003001 set_pte_at(vma->vm_mm, fe->address, fe->pte, entry);
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003002
3003 /* no need to invalidate: a not-present page won't be cached */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003004 update_mmu_cache(vma, fe->address, fe->pte);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003005
3006 return 0;
Kirill A. Shutemov3bb97792014-04-03 14:48:16 -07003007}
3008
Kirill A. Shutemov3a910532014-08-06 16:08:07 -07003009static unsigned long fault_around_bytes __read_mostly =
Vinayak Menon23f8a2f2015-12-17 08:35:36 +05303010 rounddown_pow_of_two(4096);
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003011
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003012#ifdef CONFIG_DEBUG_FS
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003013static int fault_around_bytes_get(void *data, u64 *val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003014{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003015 *val = fault_around_bytes;
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003016 return 0;
3017}
3018
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003019/*
3020 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3021 * rounded down to nearest page order. It's what do_fault_around() expects to
3022 * see.
3023 */
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003024static int fault_around_bytes_set(void *data, u64 val)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003025{
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003026 if (val / PAGE_SIZE > PTRS_PER_PTE)
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003027 return -EINVAL;
Andrey Ryabininb4903d62014-07-30 16:08:35 -07003028 if (val > PAGE_SIZE)
3029 fault_around_bytes = rounddown_pow_of_two(val);
3030 else
3031 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003032 return 0;
3033}
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003034DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3035 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003036
3037static int __init fault_around_debugfs(void)
3038{
3039 void *ret;
3040
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003041 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3042 &fault_around_bytes_fops);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003043 if (!ret)
Kirill A. Shutemova9b0f862014-06-04 16:10:54 -07003044 pr_warn("Failed to create fault_around_bytes in debugfs");
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003045 return 0;
3046}
3047late_initcall(fault_around_debugfs);
Kirill A. Shutemov1592eef2014-04-07 15:37:22 -07003048#endif
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003049
Kirill A. Shutemov1fdb4122014-06-04 16:10:55 -07003050/*
3051 * do_fault_around() tries to map few pages around the fault address. The hope
3052 * is that the pages will be needed soon and this will lower the number of
3053 * faults to handle.
3054 *
3055 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3056 * not ready to be mapped: not up-to-date, locked, etc.
3057 *
3058 * This function is called with the page table lock taken. In the split ptlock
3059 * case the page table lock only protects only those entries which belong to
3060 * the page table corresponding to the fault address.
3061 *
3062 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3063 * only once.
3064 *
3065 * fault_around_pages() defines how many pages we'll try to map.
3066 * do_fault_around() expects it to return a power of two less than or equal to
3067 * PTRS_PER_PTE.
3068 *
3069 * The virtual address of the area that we map is naturally aligned to the
3070 * fault_around_pages() value (and therefore to page order). This way it's
3071 * easier to guarantee that we don't cross page table boundaries.
3072 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003073static int do_fault_around(struct fault_env *fe, pgoff_t start_pgoff)
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003074{
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003075 unsigned long address = fe->address, nr_pages, mask;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003076 pgoff_t end_pgoff;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003077 int off, ret = 0;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003078
Jason Low4db0c3c2015-04-15 16:14:08 -07003079 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
Kirill A. Shutemovaecd6f42014-08-06 16:08:05 -07003080 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3081
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003082 fe->address = max(address & mask, fe->vma->vm_start);
3083 off = ((address - fe->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003084 start_pgoff -= off;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003085
3086 /*
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003087 * end_pgoff is either end of page table or end of vma
3088 * or fault_around_pages() from start_pgoff, depending what is nearest.
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003089 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003090 end_pgoff = start_pgoff -
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003091 ((fe->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003092 PTRS_PER_PTE - 1;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003093 end_pgoff = min3(end_pgoff, vma_pages(fe->vma) + fe->vma->vm_pgoff - 1,
3094 start_pgoff + nr_pages - 1);
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003095
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003096 if (pmd_none(*fe->pmd)) {
3097 fe->prealloc_pte = pte_alloc_one(fe->vma->vm_mm, fe->address);
Vegard Nossumc5f88bd2016-08-02 14:02:22 -07003098 if (!fe->prealloc_pte)
3099 goto out;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003100 smp_wmb(); /* See comment in __pte_alloc() */
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003101 }
3102
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003103 fe->vma->vm_ops->map_pages(fe, start_pgoff, end_pgoff);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003104
3105 /* preallocated pagetable is unused: free it */
3106 if (fe->prealloc_pte) {
3107 pte_free(fe->vma->vm_mm, fe->prealloc_pte);
3108 fe->prealloc_pte = 0;
3109 }
3110 /* Huge page is mapped? Page fault is solved */
3111 if (pmd_trans_huge(*fe->pmd)) {
3112 ret = VM_FAULT_NOPAGE;
3113 goto out;
3114 }
3115
3116 /* ->map_pages() haven't done anything useful. Cold page cache? */
3117 if (!fe->pte)
3118 goto out;
3119
3120 /* check if the page fault is solved */
3121 fe->pte -= (fe->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3122 if (!pte_none(*fe->pte))
3123 ret = VM_FAULT_NOPAGE;
3124 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003125out:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003126 fe->address = address;
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003127 fe->pte = NULL;
3128 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003129}
3130
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003131static int do_read_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003132{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003133 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003134 struct page *fault_page;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003135 int ret = 0;
3136
3137 /*
3138 * Let's call ->map_pages() first and use ->fault() as fallback
3139 * if page by the offset is not ready to be mapped (cold cache or
3140 * something).
3141 */
Kirill A. Shutemov9b4bdd22015-02-10 14:09:51 -08003142 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003143 ret = do_fault_around(fe, pgoff);
3144 if (ret)
3145 return ret;
Kirill A. Shutemov8c6e50b2014-04-07 15:37:18 -07003146 }
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003147
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003148 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003149 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3150 return ret;
3151
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003152 ret |= alloc_set_pte(fe, NULL, fault_page);
3153 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003154 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003155 unlock_page(fault_page);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003156 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3157 put_page(fault_page);
Kirill A. Shutemove655fb22014-04-03 14:48:11 -07003158 return ret;
3159}
3160
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003161static int do_cow_fault(struct fault_env *fe, pgoff_t pgoff)
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003162{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003163 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003164 struct page *fault_page, *new_page;
Jan Karabc2466e2016-05-12 18:29:19 +02003165 void *fault_entry;
Johannes Weiner00501b52014-08-08 14:19:20 -07003166 struct mem_cgroup *memcg;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003167 int ret;
3168
3169 if (unlikely(anon_vma_prepare(vma)))
3170 return VM_FAULT_OOM;
3171
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003172 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, fe->address);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003173 if (!new_page)
3174 return VM_FAULT_OOM;
3175
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003176 if (mem_cgroup_try_charge(new_page, vma->vm_mm, GFP_KERNEL,
3177 &memcg, false)) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003178 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003179 return VM_FAULT_OOM;
3180 }
3181
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003182 ret = __do_fault(fe, pgoff, new_page, &fault_page, &fault_entry);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003183 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3184 goto uncharge_out;
3185
Jan Karabc2466e2016-05-12 18:29:19 +02003186 if (!(ret & VM_FAULT_DAX_LOCKED))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003187 copy_user_highpage(new_page, fault_page, fe->address, vma);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003188 __SetPageUptodate(new_page);
3189
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003190 ret |= alloc_set_pte(fe, memcg, new_page);
3191 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003192 pte_unmap_unlock(fe->pte, fe->ptl);
Jan Karabc2466e2016-05-12 18:29:19 +02003193 if (!(ret & VM_FAULT_DAX_LOCKED)) {
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003194 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003195 put_page(fault_page);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003196 } else {
Jan Karabc2466e2016-05-12 18:29:19 +02003197 dax_unlock_mapping_entry(vma->vm_file->f_mapping, pgoff);
Matthew Wilcox2e4cdab2015-02-16 15:58:50 -08003198 }
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003199 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3200 goto uncharge_out;
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003201 return ret;
3202uncharge_out:
Kirill A. Shutemovf627c2f2016-01-15 16:52:20 -08003203 mem_cgroup_cancel_charge(new_page, memcg, false);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003204 put_page(new_page);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003205 return ret;
3206}
3207
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003208static int do_shared_fault(struct fault_env *fe, pgoff_t pgoff)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003209{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003210 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003211 struct page *fault_page;
3212 struct address_space *mapping;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003213 int dirtied = 0;
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003214 int ret, tmp;
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003215
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003216 ret = __do_fault(fe, pgoff, NULL, &fault_page, NULL);
Kirill A. Shutemov7eae74a2014-04-03 14:48:10 -07003217 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003218 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219
3220 /*
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003221 * Check if the backing address space wants to know that the page is
3222 * about to become writable
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223 */
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003224 if (vma->vm_ops->page_mkwrite) {
3225 unlock_page(fault_page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003226 tmp = do_page_mkwrite(vma, fault_page, fe->address);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003227 if (unlikely(!tmp ||
3228 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003229 put_page(fault_page);
Kirill A. Shutemovfb09a462014-04-03 14:48:15 -07003230 return tmp;
3231 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 }
3233
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003234 ret |= alloc_set_pte(fe, NULL, fault_page);
3235 if (fe->pte)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003236 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003237 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3238 VM_FAULT_RETRY))) {
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003239 unlock_page(fault_page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003240 put_page(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003241 return ret;
Peter Zijlstrad08b3852006-09-25 23:30:57 -07003242 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003243
3244 if (set_page_dirty(fault_page))
3245 dirtied = 1;
Andrew Mortond82fa872014-12-18 16:17:29 -08003246 /*
3247 * Take a local copy of the address_space - page.mapping may be zeroed
3248 * by truncate after unlock_page(). The address_space itself remains
3249 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3250 * release semantics to prevent the compiler from undoing this copying.
3251 */
Kirill A. Shutemov1c290f62016-01-15 16:52:07 -08003252 mapping = page_rmapping(fault_page);
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003253 unlock_page(fault_page);
3254 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3255 /*
3256 * Some device drivers do not set page.mapping but still
3257 * dirty their pages
3258 */
3259 balance_dirty_pages_ratelimited(mapping);
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003260 }
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003261
Johannes Weiner74ec6752015-02-10 14:11:28 -08003262 if (!vma->vm_ops->page_mkwrite)
Kirill A. Shutemovf0c6d4d2014-04-03 14:48:13 -07003263 file_update_time(vma->vm_file);
3264
KAMEZAWA Hiroyuki1d65f862011-07-25 17:12:27 -07003265 return ret;
Nick Piggin54cb8822007-07-19 01:46:59 -07003266}
Nick Piggind00806b2007-07-19 01:46:57 -07003267
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003268/*
3269 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3270 * but allow concurrent faults).
3271 * The mmap_sem may have been released depending on flags and our
3272 * return value. See filemap_fault() and __lock_page_or_retry().
3273 */
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003274static int do_fault(struct fault_env *fe)
Nick Piggin54cb8822007-07-19 01:46:59 -07003275{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003276 struct vm_area_struct *vma = fe->vma;
3277 pgoff_t pgoff = linear_page_index(vma, fe->address);
Nick Piggin54cb8822007-07-19 01:46:59 -07003278
Kirill A. Shutemov6b7339f2015-07-06 23:18:37 +03003279 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3280 if (!vma->vm_ops->fault)
3281 return VM_FAULT_SIGBUS;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003282 if (!(fe->flags & FAULT_FLAG_WRITE))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003283 return do_read_fault(fe, pgoff);
Kirill A. Shutemovec47c3b2014-04-03 14:48:12 -07003284 if (!(vma->vm_flags & VM_SHARED))
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003285 return do_cow_fault(fe, pgoff);
3286 return do_shared_fault(fe, pgoff);
Nick Piggin54cb8822007-07-19 01:46:59 -07003287}
3288
Rashika Kheriab19a9932014-04-03 14:48:02 -07003289static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
Rik van Riel04bb2f92013-10-07 11:29:36 +01003290 unsigned long addr, int page_nid,
3291 int *flags)
Mel Gorman9532fec2012-11-15 01:24:32 +00003292{
3293 get_page(page);
3294
3295 count_vm_numa_event(NUMA_HINT_FAULTS);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003296 if (page_nid == numa_node_id()) {
Mel Gorman9532fec2012-11-15 01:24:32 +00003297 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
Rik van Riel04bb2f92013-10-07 11:29:36 +01003298 *flags |= TNF_FAULT_LOCAL;
3299 }
Mel Gorman9532fec2012-11-15 01:24:32 +00003300
3301 return mpol_misplaced(page, vma, addr);
3302}
3303
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003304static int do_numa_page(struct fault_env *fe, pte_t pte)
Mel Gormand10e63f2012-10-25 14:16:31 +02003305{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003306 struct vm_area_struct *vma = fe->vma;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003307 struct page *page = NULL;
Mel Gorman8191acb2013-10-07 11:28:45 +01003308 int page_nid = -1;
Peter Zijlstra90572892013-10-07 11:29:20 +01003309 int last_cpupid;
Peter Zijlstracbee9f82012-10-25 14:16:43 +02003310 int target_nid;
Mel Gormanb8593bf2012-11-21 01:18:23 +00003311 bool migrated = false;
Mel Gormanb191f9b2015-03-25 15:55:40 -07003312 bool was_writable = pte_write(pte);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003313 int flags = 0;
Mel Gormand10e63f2012-10-25 14:16:31 +02003314
3315 /*
3316 * The "pte" at this point cannot be used safely without
3317 * validation through pte_unmap_same(). It's of NUMA type but
3318 * the pfn may be screwed if the read is non atomic.
3319 *
Mel Gorman4d942462015-02-12 14:58:28 -08003320 * We can safely just do a "set_pte_at()", because the old
3321 * page table entry is not accessible, so there would be no
3322 * concurrent hardware modifications to the PTE.
Mel Gormand10e63f2012-10-25 14:16:31 +02003323 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003324 fe->ptl = pte_lockptr(vma->vm_mm, fe->pmd);
3325 spin_lock(fe->ptl);
3326 if (unlikely(!pte_same(*fe->pte, pte))) {
3327 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003328 goto out;
3329 }
3330
Mel Gorman4d942462015-02-12 14:58:28 -08003331 /* Make it present again */
3332 pte = pte_modify(pte, vma->vm_page_prot);
3333 pte = pte_mkyoung(pte);
Mel Gormanb191f9b2015-03-25 15:55:40 -07003334 if (was_writable)
3335 pte = pte_mkwrite(pte);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003336 set_pte_at(vma->vm_mm, fe->address, fe->pte, pte);
3337 update_mmu_cache(vma, fe->address, fe->pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003338
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003339 page = vm_normal_page(vma, fe->address, pte);
Mel Gormand10e63f2012-10-25 14:16:31 +02003340 if (!page) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003341 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gormand10e63f2012-10-25 14:16:31 +02003342 return 0;
3343 }
3344
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003345 /* TODO: handle PTE-mapped THP */
3346 if (PageCompound(page)) {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003347 pte_unmap_unlock(fe->pte, fe->ptl);
Kirill A. Shutemove81c4802016-01-15 16:53:49 -08003348 return 0;
3349 }
3350
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003351 /*
Mel Gormanbea66fb2015-03-25 15:55:37 -07003352 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3353 * much anyway since they can be in shared cache state. This misses
3354 * the case where a mapping is writable but the process never writes
3355 * to it but pte_write gets cleared during protection updates and
3356 * pte_dirty has unpredictable behaviour between PTE scan updates,
3357 * background writeback, dirty balancing and application behaviour.
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003358 */
Rik van Rield59dc7b2016-09-08 21:30:53 -04003359 if (!pte_write(pte))
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003360 flags |= TNF_NO_GROUP;
3361
Rik van Rieldabe1d92013-10-07 11:29:34 +01003362 /*
3363 * Flag if the page is shared between multiple address spaces. This
3364 * is later used when determining whether to group tasks together
3365 */
3366 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3367 flags |= TNF_SHARED;
3368
Peter Zijlstra90572892013-10-07 11:29:20 +01003369 last_cpupid = page_cpupid_last(page);
Mel Gorman8191acb2013-10-07 11:28:45 +01003370 page_nid = page_to_nid(page);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003371 target_nid = numa_migrate_prep(page, vma, fe->address, page_nid,
3372 &flags);
3373 pte_unmap_unlock(fe->pte, fe->ptl);
Mel Gorman4daae3b2012-11-02 11:33:45 +00003374 if (target_nid == -1) {
Mel Gorman4daae3b2012-11-02 11:33:45 +00003375 put_page(page);
3376 goto out;
3377 }
3378
3379 /* Migrate to the requested node */
Mel Gorman1bc115d2013-10-07 11:29:05 +01003380 migrated = migrate_misplaced_page(page, vma, target_nid);
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003381 if (migrated) {
Mel Gorman8191acb2013-10-07 11:28:45 +01003382 page_nid = target_nid;
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003383 flags |= TNF_MIGRATED;
Mel Gorman074c2382015-03-25 15:55:42 -07003384 } else
3385 flags |= TNF_MIGRATE_FAIL;
Mel Gorman4daae3b2012-11-02 11:33:45 +00003386
3387out:
Mel Gorman8191acb2013-10-07 11:28:45 +01003388 if (page_nid != -1)
Peter Zijlstra6688cc02013-10-07 11:29:24 +01003389 task_numa_fault(last_cpupid, page_nid, 1, flags);
Mel Gormand10e63f2012-10-25 14:16:31 +02003390 return 0;
3391}
3392
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003393static int create_huge_pmd(struct fault_env *fe)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003394{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003395 struct vm_area_struct *vma = fe->vma;
Kirill A. Shutemovfb6dd5f2015-09-09 15:39:35 -07003396 if (vma_is_anonymous(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003397 return do_huge_pmd_anonymous_page(fe);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003398 if (vma->vm_ops->pmd_fault)
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003399 return vma->vm_ops->pmd_fault(vma, fe->address, fe->pmd,
3400 fe->flags);
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003401 return VM_FAULT_FALLBACK;
3402}
3403
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003404static int wp_huge_pmd(struct fault_env *fe, pmd_t orig_pmd)
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003405{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003406 if (vma_is_anonymous(fe->vma))
3407 return do_huge_pmd_wp_page(fe, orig_pmd);
3408 if (fe->vma->vm_ops->pmd_fault)
3409 return fe->vma->vm_ops->pmd_fault(fe->vma, fe->address, fe->pmd,
3410 fe->flags);
Kirill A. Shutemovaf9e4d52016-07-26 15:25:40 -07003411
3412 /* COW handled on pte level: split pmd */
3413 VM_BUG_ON_VMA(fe->vma->vm_flags & VM_SHARED, fe->vma);
3414 split_huge_pmd(fe->vma, fe->pmd, fe->address);
3415
Matthew Wilcoxb96375f2015-09-08 14:58:48 -07003416 return VM_FAULT_FALLBACK;
3417}
3418
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003419static inline bool vma_is_accessible(struct vm_area_struct *vma)
3420{
3421 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3422}
3423
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424/*
3425 * These routines also need to handle stuff like marking pages dirty
3426 * and/or accessed for architectures that don't do it in hardware (most
3427 * RISC architectures). The early dirtying is also good on the i386.
3428 *
3429 * There is also a hook called "update_mmu_cache()" that architectures
3430 * with external mmu caches can use to update those (ie the Sparc or
3431 * PowerPC hashed page tables that act as extended TLBs).
3432 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003433 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3434 * concurrent faults).
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003435 *
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003436 * The mmap_sem may have been released depending on flags and our return value.
3437 * See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003439static int handle_pte_fault(struct fault_env *fe)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440{
3441 pte_t entry;
3442
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003443 if (unlikely(pmd_none(*fe->pmd))) {
3444 /*
3445 * Leave __pte_alloc() until later: because vm_ops->fault may
3446 * want to allocate huge page, and if we expose page table
3447 * for an instant, it will be difficult to retract from
3448 * concurrent faults and from rmap lookups.
3449 */
3450 fe->pte = NULL;
3451 } else {
3452 /* See comment in pte_alloc_one_map() */
3453 if (pmd_trans_unstable(fe->pmd) || pmd_devmap(*fe->pmd))
3454 return 0;
3455 /*
3456 * A regular pmd is established and it can't morph into a huge
3457 * pmd from under us anymore at this point because we hold the
3458 * mmap_sem read mode and khugepaged takes it in write mode.
3459 * So now it's safe to run pte_offset_map().
3460 */
3461 fe->pte = pte_offset_map(fe->pmd, fe->address);
3462
3463 entry = *fe->pte;
3464
3465 /*
3466 * some architectures can have larger ptes than wordsize,
3467 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3468 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3469 * atomic accesses. The code below just needs a consistent
3470 * view for the ifs and we later double check anyway with the
3471 * ptl lock held. So here a barrier will do.
3472 */
3473 barrier();
Hugh Dickins65500d22005-10-29 18:15:59 -07003474 if (pte_none(entry)) {
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003475 pte_unmap(fe->pte);
3476 fe->pte = NULL;
Hugh Dickins65500d22005-10-29 18:15:59 -07003477 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 }
3479
Kirill A. Shutemov7267ec002016-07-26 15:25:23 -07003480 if (!fe->pte) {
3481 if (vma_is_anonymous(fe->vma))
3482 return do_anonymous_page(fe);
3483 else
3484 return do_fault(fe);
3485 }
3486
3487 if (!pte_present(entry))
3488 return do_swap_page(fe, entry);
3489
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003490 if (pte_protnone(entry) && vma_is_accessible(fe->vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003491 return do_numa_page(fe, entry);
Mel Gormand10e63f2012-10-25 14:16:31 +02003492
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003493 fe->ptl = pte_lockptr(fe->vma->vm_mm, fe->pmd);
3494 spin_lock(fe->ptl);
3495 if (unlikely(!pte_same(*fe->pte, entry)))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003496 goto unlock;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003497 if (fe->flags & FAULT_FLAG_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 if (!pte_write(entry))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003499 return do_wp_page(fe, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 entry = pte_mkdirty(entry);
3501 }
3502 entry = pte_mkyoung(entry);
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003503 if (ptep_set_access_flags(fe->vma, fe->address, fe->pte, entry,
3504 fe->flags & FAULT_FLAG_WRITE)) {
3505 update_mmu_cache(fe->vma, fe->address, fe->pte);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003506 } else {
3507 /*
3508 * This is needed only for protection faults but the arch code
3509 * is not yet telling us if this is a protection fault or not.
3510 * This still avoids useless tlb flushes for .text page faults
3511 * with threads.
3512 */
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003513 if (fe->flags & FAULT_FLAG_WRITE)
3514 flush_tlb_fix_spurious_fault(fe->vma, fe->address);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07003515 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07003516unlock:
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003517 pte_unmap_unlock(fe->pte, fe->ptl);
Nick Piggin83c54072007-07-19 01:47:05 -07003518 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519}
3520
3521/*
3522 * By the time we get here, we already hold the mm semaphore
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003523 *
3524 * The mmap_sem may have been released depending on flags and our
3525 * return value. See filemap_fault() and __lock_page_or_retry().
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003527static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3528 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003530 struct fault_env fe = {
3531 .vma = vma,
3532 .address = address,
3533 .flags = flags,
3534 };
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003535 struct mm_struct *mm = vma->vm_mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536 pgd_t *pgd;
3537 pud_t *pud;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 pud = pud_alloc(mm, pgd, address);
3541 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003542 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003543 fe.pmd = pmd_alloc(mm, pud, address);
3544 if (!fe.pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07003545 return VM_FAULT_OOM;
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003546 if (pmd_none(*fe.pmd) && transparent_hugepage_enabled(vma)) {
3547 int ret = create_huge_pmd(&fe);
Kirill A. Shutemovc0292552013-09-12 15:14:05 -07003548 if (!(ret & VM_FAULT_FALLBACK))
3549 return ret;
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003550 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003551 pmd_t orig_pmd = *fe.pmd;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003552 int ret;
3553
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003554 barrier();
Dan Williams5c7fb562016-01-15 16:56:52 -08003555 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
Lorenzo Stoakes38e08852016-09-11 23:54:25 +01003556 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003557 return do_huge_pmd_numa_page(&fe, orig_pmd);
Mel Gormand10e63f2012-10-25 14:16:31 +02003558
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003559 if ((fe.flags & FAULT_FLAG_WRITE) &&
3560 !pmd_write(orig_pmd)) {
3561 ret = wp_huge_pmd(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003562 if (!(ret & VM_FAULT_FALLBACK))
3563 return ret;
Will Deacona1dd4502012-12-11 16:01:27 -08003564 } else {
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003565 huge_pmd_set_accessed(&fe, orig_pmd);
Kirill A. Shutemov9845cbb2014-02-25 15:01:42 -08003566 return 0;
David Rientjes1f1d06c2012-05-29 15:06:23 -07003567 }
Andrea Arcangeli71e3aac2011-01-13 15:46:52 -08003568 }
3569 }
3570
Kirill A. Shutemovbae473a2016-07-26 15:25:20 -07003571 return handle_pte_fault(&fe);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572}
3573
Paul Cassella9a95f3c2014-08-06 16:07:24 -07003574/*
3575 * By the time we get here, we already hold the mm semaphore
3576 *
3577 * The mmap_sem may have been released depending on flags and our
3578 * return value. See filemap_fault() and __lock_page_or_retry().
3579 */
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003580int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3581 unsigned int flags)
Johannes Weiner519e5242013-09-12 15:13:42 -07003582{
3583 int ret;
3584
3585 __set_current_state(TASK_RUNNING);
3586
3587 count_vm_event(PGFAULT);
Kirill A. Shutemovdcddffd2016-07-26 15:25:18 -07003588 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
Johannes Weiner519e5242013-09-12 15:13:42 -07003589
3590 /* do counter updates before entering really critical section. */
3591 check_sync_rss_stat(current);
3592
Laurent Dufour3c8381d2017-09-08 16:13:12 -07003593 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3594 flags & FAULT_FLAG_INSTRUCTION,
3595 flags & FAULT_FLAG_REMOTE))
3596 return VM_FAULT_SIGSEGV;
3597
Johannes Weiner519e5242013-09-12 15:13:42 -07003598 /*
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 (unlikely(is_vm_hugetlb_page(vma)))
3606 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3607 else
3608 ret = __handle_mm_fault(vma, address, flags);
Johannes Weiner519e5242013-09-12 15:13:42 -07003609
Johannes Weiner49426422013-10-16 13:46:59 -07003610 if (flags & FAULT_FLAG_USER) {
3611 mem_cgroup_oom_disable();
3612 /*
3613 * The task may have entered a memcg OOM situation but
3614 * if the allocation error was handled gracefully (no
3615 * VM_FAULT_OOM), there is no need to kill anything.
3616 * Just clean up the OOM state peacefully.
3617 */
3618 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3619 mem_cgroup_oom_synchronize(false);
3620 }
Johannes Weiner3812c8c2013-09-12 15:13:44 -07003621
Michal Hocko3f70dc32016-10-07 16:59:06 -07003622 /*
3623 * This mm has been already reaped by the oom reaper and so the
3624 * refault cannot be trusted in general. Anonymous refaults would
3625 * lose data and give a zero page instead e.g. This is especially
3626 * problem for use_mm() because regular tasks will just die and
3627 * the corrupted data will not be visible anywhere while kthread
3628 * will outlive the oom victim and potentially propagate the date
3629 * further.
3630 */
3631 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
Michal Hocko59ee25d2017-08-18 15:16:12 -07003632 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags))) {
3633
3634 /*
3635 * We are going to enforce SIGBUS but the PF path might have
3636 * dropped the mmap_sem already so take it again so that
3637 * we do not break expectations of all arch specific PF paths
3638 * and g-u-p
3639 */
3640 if (ret & VM_FAULT_RETRY)
3641 down_read(&vma->vm_mm->mmap_sem);
Michal Hocko3f70dc32016-10-07 16:59:06 -07003642 ret = VM_FAULT_SIGBUS;
Michal Hocko59ee25d2017-08-18 15:16:12 -07003643 }
Michal Hocko3f70dc32016-10-07 16:59:06 -07003644
Johannes Weiner519e5242013-09-12 15:13:42 -07003645 return ret;
3646}
Jesse Barnese1d6d012014-12-12 16:55:27 -08003647EXPORT_SYMBOL_GPL(handle_mm_fault);
Johannes Weiner519e5242013-09-12 15:13:42 -07003648
Linus Torvalds1da177e2005-04-16 15:20:36 -07003649#ifndef __PAGETABLE_PUD_FOLDED
3650/*
3651 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003652 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003654int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003655{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003656 pud_t *new = pud_alloc_one(mm, address);
3657 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003658 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659
Nick Piggin362a61a2008-05-14 06:37:36 +02003660 smp_wmb(); /* See comment in __pte_alloc */
3661
Hugh Dickins872fec12005-10-29 18:16:21 -07003662 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003663 if (pgd_present(*pgd)) /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003664 pud_free(mm, new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003665 else
3666 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07003667 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003668 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669}
3670#endif /* __PAGETABLE_PUD_FOLDED */
3671
3672#ifndef __PAGETABLE_PMD_FOLDED
3673/*
3674 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07003675 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07003677int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678{
Hugh Dickinsc74df322005-10-29 18:16:23 -07003679 pmd_t *new = pmd_alloc_one(mm, address);
3680 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07003681 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003682
Nick Piggin362a61a2008-05-14 06:37:36 +02003683 smp_wmb(); /* See comment in __pte_alloc */
3684
Hugh Dickins872fec12005-10-29 18:16:21 -07003685 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686#ifndef __ARCH_HAS_4LEVEL_HACK
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003687 if (!pud_present(*pud)) {
3688 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003689 pud_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003690 } else /* Another has populated it */
Benjamin Herrenschmidt5e541972008-02-04 22:29:14 -08003691 pmd_free(mm, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003692#else
3693 if (!pgd_present(*pud)) {
3694 mm_inc_nr_pmds(mm);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003695 pgd_populate(mm, pud, new);
Kirill A. Shutemovdc6c9a32015-02-11 15:26:50 -08003696 } else /* Another has populated it */
3697 pmd_free(mm, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07003699 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07003700 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701}
3702#endif /* __PAGETABLE_PMD_FOLDED */
3703
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003704static int __follow_pte(struct mm_struct *mm, unsigned long address,
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003705 pte_t **ptepp, spinlock_t **ptlp)
3706{
3707 pgd_t *pgd;
3708 pud_t *pud;
3709 pmd_t *pmd;
3710 pte_t *ptep;
3711
3712 pgd = pgd_offset(mm, address);
3713 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3714 goto out;
3715
3716 pud = pud_offset(pgd, address);
3717 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3718 goto out;
3719
3720 pmd = pmd_offset(pud, address);
Andrea Arcangelif66055ab2011-01-13 15:46:54 -08003721 VM_BUG_ON(pmd_trans_huge(*pmd));
Johannes Weinerf8ad0f492009-06-16 15:32:33 -07003722 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3723 goto out;
3724
3725 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3726 if (pmd_huge(*pmd))
3727 goto out;
3728
3729 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3730 if (!ptep)
3731 goto out;
3732 if (!pte_present(*ptep))
3733 goto unlock;
3734 *ptepp = ptep;
3735 return 0;
3736unlock:
3737 pte_unmap_unlock(ptep, *ptlp);
3738out:
3739 return -EINVAL;
3740}
3741
Namhyung Kim1b36ba82010-10-26 14:22:00 -07003742static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3743 pte_t **ptepp, spinlock_t **ptlp)
3744{
3745 int res;
3746
3747 /* (void) is needed to make gcc happy */
3748 (void) __cond_lock(*ptlp,
3749 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3750 return res;
3751}
3752
Johannes Weiner3b6748e2009-06-16 15:32:35 -07003753/**
3754 * follow_pfn - look up PFN at a user virtual address
3755 * @vma: memory mapping
3756 * @address: user virtual address
3757 * @pfn: location to store found PFN
3758 *
3759 * Only IO mappings and raw PFN mappings are allowed.
3760 *
3761 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3762 */
3763int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3764 unsigned long *pfn)
3765{
3766 int ret = -EINVAL;
3767 spinlock_t *ptl;
3768 pte_t *ptep;
3769
3770 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3771 return ret;
3772
3773 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3774 if (ret)
3775 return ret;
3776 *pfn = pte_pfn(*ptep);
3777 pte_unmap_unlock(ptep, ptl);
3778 return 0;
3779}
3780EXPORT_SYMBOL(follow_pfn);
3781
Rik van Riel28b2ee22008-07-23 21:27:05 -07003782#ifdef CONFIG_HAVE_IOREMAP_PROT
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003783int follow_phys(struct vm_area_struct *vma,
3784 unsigned long address, unsigned int flags,
3785 unsigned long *prot, resource_size_t *phys)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003786{
Johannes Weiner03668a42009-06-16 15:32:34 -07003787 int ret = -EINVAL;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003788 pte_t *ptep, pte;
3789 spinlock_t *ptl;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003790
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003791 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3792 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003793
Johannes Weiner03668a42009-06-16 15:32:34 -07003794 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003795 goto out;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003796 pte = *ptep;
Johannes Weiner03668a42009-06-16 15:32:34 -07003797
Rik van Riel28b2ee22008-07-23 21:27:05 -07003798 if ((flags & FOLL_WRITE) && !pte_write(pte))
3799 goto unlock;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003800
3801 *prot = pgprot_val(pte_pgprot(pte));
Johannes Weiner03668a42009-06-16 15:32:34 -07003802 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003803
Johannes Weiner03668a42009-06-16 15:32:34 -07003804 ret = 0;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003805unlock:
3806 pte_unmap_unlock(ptep, ptl);
3807out:
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003808 return ret;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003809}
3810
3811int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3812 void *buf, int len, int write)
3813{
3814 resource_size_t phys_addr;
3815 unsigned long prot = 0;
KOSAKI Motohiro2bc72732009-01-06 14:39:43 -08003816 void __iomem *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003817 int offset = addr & (PAGE_SIZE-1);
3818
venkatesh.pallipadi@intel.comd87fe662008-12-19 13:47:27 -08003819 if (follow_phys(vma, addr, write, &prot, &phys_addr))
Rik van Riel28b2ee22008-07-23 21:27:05 -07003820 return -EINVAL;
3821
Grazvydas Ignotas9cb12d72015-02-12 15:00:19 -08003822 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003823 if (write)
3824 memcpy_toio(maddr + offset, buf, len);
3825 else
3826 memcpy_fromio(buf, maddr + offset, len);
3827 iounmap(maddr);
3828
3829 return len;
3830}
Uwe Kleine-König5a736332013-08-07 13:02:52 +02003831EXPORT_SYMBOL_GPL(generic_access_phys);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003832#endif
3833
David Howells0ec76a12006-09-27 01:50:15 -07003834/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003835 * Access another process' address space as given in mm. If non-NULL, use the
3836 * given task for page fault accounting.
David Howells0ec76a12006-09-27 01:50:15 -07003837 */
Eric W. Biedermane71b4e02016-11-22 12:06:50 -06003838int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003839 unsigned long addr, void *buf, int len, unsigned int gup_flags)
David Howells0ec76a12006-09-27 01:50:15 -07003840{
David Howells0ec76a12006-09-27 01:50:15 -07003841 struct vm_area_struct *vma;
David Howells0ec76a12006-09-27 01:50:15 -07003842 void *old_buf = buf;
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003843 int write = gup_flags & FOLL_WRITE;
David Howells0ec76a12006-09-27 01:50:15 -07003844
David Howells0ec76a12006-09-27 01:50:15 -07003845 down_read(&mm->mmap_sem);
Simon Arlott183ff222007-10-20 01:27:18 +02003846 /* ignore errors, just check how much was successfully transferred */
David Howells0ec76a12006-09-27 01:50:15 -07003847 while (len) {
3848 int bytes, ret, offset;
3849 void *maddr;
Rik van Riel28b2ee22008-07-23 21:27:05 -07003850 struct page *page = NULL;
David Howells0ec76a12006-09-27 01:50:15 -07003851
Dave Hansen1e987792016-02-12 13:01:54 -08003852 ret = get_user_pages_remote(tsk, mm, addr, 1,
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003853 gup_flags, &page, &vma);
Rik van Riel28b2ee22008-07-23 21:27:05 -07003854 if (ret <= 0) {
Rik van Rieldbffcd02014-08-06 16:08:12 -07003855#ifndef CONFIG_HAVE_IOREMAP_PROT
3856 break;
3857#else
Rik van Riel28b2ee22008-07-23 21:27:05 -07003858 /*
3859 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3860 * we can access using slightly different code.
3861 */
Rik van Riel28b2ee22008-07-23 21:27:05 -07003862 vma = find_vma(mm, addr);
Michael Ellermanfe936df2011-04-14 15:22:10 -07003863 if (!vma || vma->vm_start > addr)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003864 break;
3865 if (vma->vm_ops && vma->vm_ops->access)
3866 ret = vma->vm_ops->access(vma, addr, buf,
3867 len, write);
3868 if (ret <= 0)
Rik van Riel28b2ee22008-07-23 21:27:05 -07003869 break;
3870 bytes = ret;
Rik van Rieldbffcd02014-08-06 16:08:12 -07003871#endif
David Howells0ec76a12006-09-27 01:50:15 -07003872 } else {
Rik van Riel28b2ee22008-07-23 21:27:05 -07003873 bytes = len;
3874 offset = addr & (PAGE_SIZE-1);
3875 if (bytes > PAGE_SIZE-offset)
3876 bytes = PAGE_SIZE-offset;
3877
3878 maddr = kmap(page);
3879 if (write) {
3880 copy_to_user_page(vma, page, addr,
3881 maddr + offset, buf, bytes);
3882 set_page_dirty_lock(page);
3883 } else {
3884 copy_from_user_page(vma, page, addr,
3885 buf, maddr + offset, bytes);
3886 }
3887 kunmap(page);
Kirill A. Shutemov09cbfea2016-04-01 15:29:47 +03003888 put_page(page);
David Howells0ec76a12006-09-27 01:50:15 -07003889 }
David Howells0ec76a12006-09-27 01:50:15 -07003890 len -= bytes;
3891 buf += bytes;
3892 addr += bytes;
3893 }
3894 up_read(&mm->mmap_sem);
David Howells0ec76a12006-09-27 01:50:15 -07003895
3896 return buf - old_buf;
3897}
Andi Kleen03252912008-01-30 13:33:18 +01003898
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003899/**
Randy Dunlapae91dbf2011-03-26 13:27:01 -07003900 * access_remote_vm - access another process' address space
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003901 * @mm: the mm_struct of the target address space
3902 * @addr: start address to access
3903 * @buf: source or destination buffer
3904 * @len: number of bytes to transfer
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003905 * @gup_flags: flags modifying lookup behaviour
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003906 *
3907 * The caller must hold a reference on @mm.
3908 */
3909int access_remote_vm(struct mm_struct *mm, unsigned long addr,
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003910 void *buf, int len, unsigned int gup_flags)
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003911{
Lorenzo Stoakes6347e8d2016-10-13 01:20:19 +01003912 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
Stephen Wilson5ddd36b2011-03-13 15:49:20 -04003913}
3914
Andi Kleen03252912008-01-30 13:33:18 +01003915/*
Stephen Wilson206cb632011-03-13 15:49:19 -04003916 * Access another process' address space.
3917 * Source/target buffer must be kernel space,
3918 * Do not walk the page table directly, use get_user_pages
3919 */
3920int access_process_vm(struct task_struct *tsk, unsigned long addr,
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003921 void *buf, int len, unsigned int gup_flags)
Stephen Wilson206cb632011-03-13 15:49:19 -04003922{
3923 struct mm_struct *mm;
3924 int ret;
3925
3926 mm = get_task_mm(tsk);
3927 if (!mm)
3928 return 0;
3929
Lorenzo Stoakesf307ab62016-10-13 01:20:20 +01003930 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
Lorenzo Stoakes442486e2016-10-13 01:20:18 +01003931
Stephen Wilson206cb632011-03-13 15:49:19 -04003932 mmput(mm);
3933
3934 return ret;
3935}
3936
Andi Kleen03252912008-01-30 13:33:18 +01003937/*
3938 * Print the name of a VMA.
3939 */
3940void print_vma_addr(char *prefix, unsigned long ip)
3941{
3942 struct mm_struct *mm = current->mm;
3943 struct vm_area_struct *vma;
3944
Ingo Molnare8bff742008-02-13 20:21:06 +01003945 /*
3946 * Do not print if we are in atomic
3947 * contexts (in exception stacks, etc.):
3948 */
3949 if (preempt_count())
3950 return;
3951
Andi Kleen03252912008-01-30 13:33:18 +01003952 down_read(&mm->mmap_sem);
3953 vma = find_vma(mm, ip);
3954 if (vma && vma->vm_file) {
3955 struct file *f = vma->vm_file;
3956 char *buf = (char *)__get_free_page(GFP_KERNEL);
3957 if (buf) {
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003958 char *p;
Andi Kleen03252912008-01-30 13:33:18 +01003959
Miklos Szeredi9bf39ab2015-06-19 10:29:13 +02003960 p = file_path(f, buf, PAGE_SIZE);
Andi Kleen03252912008-01-30 13:33:18 +01003961 if (IS_ERR(p))
3962 p = "?";
Andy Shevchenko2fbc57c2012-12-17 16:01:23 -08003963 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
Andi Kleen03252912008-01-30 13:33:18 +01003964 vma->vm_start,
3965 vma->vm_end - vma->vm_start);
3966 free_page((unsigned long)buf);
3967 }
3968 }
Jeff Liu51a07e52012-07-31 16:43:18 -07003969 up_read(&mm->mmap_sem);
Andi Kleen03252912008-01-30 13:33:18 +01003970}
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003971
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003972#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
David Hildenbrand9ec23532015-05-11 17:52:07 +02003973void __might_fault(const char *file, int line)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003974{
Peter Zijlstra95156f02009-01-12 13:02:11 +01003975 /*
3976 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3977 * holding the mmap_sem, this is safe because kernel memory doesn't
3978 * get paged out, therefore we'll never actually fault, and the
3979 * below annotations will generate false positives.
3980 */
3981 if (segment_eq(get_fs(), KERNEL_DS))
3982 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003983 if (pagefault_disabled())
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003984 return;
David Hildenbrand9ec23532015-05-11 17:52:07 +02003985 __might_sleep(file, line, 0);
3986#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
Michael S. Tsirkin662bbcb2013-05-26 17:32:23 +03003987 if (current->mm)
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003988 might_lock_read(&current->mm->mmap_sem);
David Hildenbrand9ec23532015-05-11 17:52:07 +02003989#endif
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003990}
David Hildenbrand9ec23532015-05-11 17:52:07 +02003991EXPORT_SYMBOL(__might_fault);
Nick Piggin3ee1afa2008-09-10 13:37:17 +02003992#endif
Andrea Arcangeli47ad8472011-01-13 15:46:47 -08003993
3994#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3995static void clear_gigantic_page(struct page *page,
3996 unsigned long addr,
3997 unsigned int pages_per_huge_page)
3998{
3999 int i;
4000 struct page *p = page;
4001
4002 might_sleep();
4003 for (i = 0; i < pages_per_huge_page;
4004 i++, p = mem_map_next(p, page, i)) {
4005 cond_resched();
4006 clear_user_highpage(p, addr + i * PAGE_SIZE);
4007 }
4008}
4009void clear_huge_page(struct page *page,
4010 unsigned long addr, unsigned int pages_per_huge_page)
4011{
4012 int i;
4013
4014 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4015 clear_gigantic_page(page, addr, pages_per_huge_page);
4016 return;
4017 }
4018
4019 might_sleep();
4020 for (i = 0; i < pages_per_huge_page; i++) {
4021 cond_resched();
4022 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4023 }
4024}
4025
4026static void copy_user_gigantic_page(struct page *dst, struct page *src,
4027 unsigned long addr,
4028 struct vm_area_struct *vma,
4029 unsigned int pages_per_huge_page)
4030{
4031 int i;
4032 struct page *dst_base = dst;
4033 struct page *src_base = src;
4034
4035 for (i = 0; i < pages_per_huge_page; ) {
4036 cond_resched();
4037 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4038
4039 i++;
4040 dst = mem_map_next(dst, dst_base, i);
4041 src = mem_map_next(src, src_base, i);
4042 }
4043}
4044
4045void copy_user_huge_page(struct page *dst, struct page *src,
4046 unsigned long addr, struct vm_area_struct *vma,
4047 unsigned int pages_per_huge_page)
4048{
4049 int i;
4050
4051 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4052 copy_user_gigantic_page(dst, src, addr, vma,
4053 pages_per_huge_page);
4054 return;
4055 }
4056
4057 might_sleep();
4058 for (i = 0; i < pages_per_huge_page; i++) {
4059 cond_resched();
4060 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4061 }
4062}
4063#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004064
Olof Johansson40b64ac2013-12-20 14:28:05 -08004065#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004066
4067static struct kmem_cache *page_ptl_cachep;
4068
4069void __init ptlock_cache_init(void)
4070{
4071 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4072 SLAB_PANIC, NULL);
4073}
4074
Peter Zijlstra539edb52013-11-14 14:31:52 -08004075bool ptlock_alloc(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004076{
4077 spinlock_t *ptl;
4078
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004079 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004080 if (!ptl)
4081 return false;
Peter Zijlstra539edb52013-11-14 14:31:52 -08004082 page->ptl = ptl;
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004083 return true;
4084}
4085
Peter Zijlstra539edb52013-11-14 14:31:52 -08004086void ptlock_free(struct page *page)
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004087{
Kirill A. Shutemovb35f1812014-01-21 15:49:07 -08004088 kmem_cache_free(page_ptl_cachep, page->ptl);
Kirill A. Shutemov49076ec2013-11-14 14:31:51 -08004089}
4090#endif