]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/memory.c
Merge branch 'next/late' with mainline
[mirror_ubuntu-bionic-kernel.git] / mm / memory.c
1 /*
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>
48 #include <linux/ksm.h>
49 #include <linux/rmap.h>
50 #include <linux/export.h>
51 #include <linux/delayacct.h>
52 #include <linux/init.h>
53 #include <linux/pfn_t.h>
54 #include <linux/writeback.h>
55 #include <linux/memcontrol.h>
56 #include <linux/mmu_notifier.h>
57 #include <linux/kallsyms.h>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
60 #include <linux/gfp.h>
61 #include <linux/migrate.h>
62 #include <linux/string.h>
63 #include <linux/dma-debug.h>
64 #include <linux/debugfs.h>
65 #include <linux/userfaultfd_k.h>
66 #include <linux/dax.h>
67
68 #include <asm/io.h>
69 #include <asm/mmu_context.h>
70 #include <asm/pgalloc.h>
71 #include <linux/uaccess.h>
72 #include <asm/tlb.h>
73 #include <asm/tlbflush.h>
74 #include <asm/pgtable.h>
75
76 #include "internal.h"
77
78 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
79 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
80 #endif
81
82 #ifndef CONFIG_NEED_MULTIPLE_NODES
83 /* use the per-pgdat data instead for discontigmem - mbligh */
84 unsigned long max_mapnr;
85 struct page *mem_map;
86
87 EXPORT_SYMBOL(max_mapnr);
88 EXPORT_SYMBOL(mem_map);
89 #endif
90
91 /*
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 */
98 void * high_memory;
99
100 EXPORT_SYMBOL(high_memory);
101
102 /*
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 */
108 int randomize_va_space __read_mostly =
109 #ifdef CONFIG_COMPAT_BRK
110 1;
111 #else
112 2;
113 #endif
114
115 static int __init disable_randmaps(char *s)
116 {
117 randomize_va_space = 0;
118 return 1;
119 }
120 __setup("norandmaps", disable_randmaps);
121
122 unsigned long zero_pfn __read_mostly;
123 unsigned long highest_memmap_pfn __read_mostly;
124
125 EXPORT_SYMBOL(zero_pfn);
126
127 /*
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 */
130 static int __init init_zero_pfn(void)
131 {
132 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 return 0;
134 }
135 core_initcall(init_zero_pfn);
136
137
138 #if defined(SPLIT_RSS_COUNTING)
139
140 void sync_mm_rss(struct mm_struct *mm)
141 {
142 int i;
143
144 for (i = 0; i < NR_MM_COUNTERS; i++) {
145 if (current->rss_stat.count[i]) {
146 add_mm_counter(mm, i, current->rss_stat.count[i]);
147 current->rss_stat.count[i] = 0;
148 }
149 }
150 current->rss_stat.events = 0;
151 }
152
153 static 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)
167 static 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))
172 sync_mm_rss(task->mm);
173 }
174 #else /* SPLIT_RSS_COUNTING */
175
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
179 static void check_sync_rss_stat(struct task_struct *task)
180 {
181 }
182
183 #endif /* SPLIT_RSS_COUNTING */
184
185 #ifdef HAVE_GENERIC_MMU_GATHER
186
187 static bool tlb_next_batch(struct mmu_gather *tlb)
188 {
189 struct mmu_gather_batch *batch;
190
191 batch = tlb->active;
192 if (batch->next) {
193 tlb->active = batch->next;
194 return true;
195 }
196
197 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
198 return false;
199
200 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
201 if (!batch)
202 return false;
203
204 tlb->batch_count++;
205 batch->next = NULL;
206 batch->nr = 0;
207 batch->max = MAX_GATHER_BATCH;
208
209 tlb->active->next = batch;
210 tlb->active = batch;
211
212 return true;
213 }
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 */
220 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
221 {
222 tlb->mm = mm;
223
224 /* Is it from 0 to ~0? */
225 tlb->fullmm = !(start | (end+1));
226 tlb->need_flush_all = 0;
227 tlb->local.next = NULL;
228 tlb->local.nr = 0;
229 tlb->local.max = ARRAY_SIZE(tlb->__pages);
230 tlb->active = &tlb->local;
231 tlb->batch_count = 0;
232
233 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
234 tlb->batch = NULL;
235 #endif
236 tlb->page_size = 0;
237
238 __tlb_reset_range(tlb);
239 }
240
241 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
242 {
243 if (!tlb->end)
244 return;
245
246 tlb_flush(tlb);
247 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
248 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
249 tlb_table_flush(tlb);
250 #endif
251 __tlb_reset_range(tlb);
252 }
253
254 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
255 {
256 struct mmu_gather_batch *batch;
257
258 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
259 free_pages_and_swap_cache(batch->pages, batch->nr);
260 batch->nr = 0;
261 }
262 tlb->active = &tlb->local;
263 }
264
265 void tlb_flush_mmu(struct mmu_gather *tlb)
266 {
267 tlb_flush_mmu_tlbonly(tlb);
268 tlb_flush_mmu_free(tlb);
269 }
270
271 /* tlb_finish_mmu
272 * Called at the end of the shootdown operation to free up any resources
273 * that were required.
274 */
275 void 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().
296 *returns true if the caller should flush.
297 */
298 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
299 {
300 struct mmu_gather_batch *batch;
301
302 VM_BUG_ON(!tlb->end);
303 VM_WARN_ON(tlb->page_size != page_size);
304
305 batch = tlb->active;
306 /*
307 * Add the page and check if we are full. If so
308 * force a flush.
309 */
310 batch->pages[batch->nr++] = page;
311 if (batch->nr == batch->max) {
312 if (!tlb_next_batch(tlb))
313 return true;
314 batch = tlb->active;
315 }
316 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
317
318 return false;
319 }
320
321 #endif /* HAVE_GENERIC_MMU_GATHER */
322
323 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
324
325 /*
326 * See the comment near struct mmu_table_batch.
327 */
328
329 static void tlb_remove_table_smp_sync(void *arg)
330 {
331 /* Simply deliver the interrupt */
332 }
333
334 static void tlb_remove_table_one(void *table)
335 {
336 /*
337 * This isn't an RCU grace period and hence the page-tables cannot be
338 * assumed to be actually RCU-freed.
339 *
340 * It is however sufficient for software page-table walkers that rely on
341 * IRQ disabling. See the comment near struct mmu_table_batch.
342 */
343 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
344 __tlb_remove_table(table);
345 }
346
347 static void tlb_remove_table_rcu(struct rcu_head *head)
348 {
349 struct mmu_table_batch *batch;
350 int i;
351
352 batch = container_of(head, struct mmu_table_batch, rcu);
353
354 for (i = 0; i < batch->nr; i++)
355 __tlb_remove_table(batch->tables[i]);
356
357 free_page((unsigned long)batch);
358 }
359
360 void tlb_table_flush(struct mmu_gather *tlb)
361 {
362 struct mmu_table_batch **batch = &tlb->batch;
363
364 if (*batch) {
365 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
366 *batch = NULL;
367 }
368 }
369
370 void tlb_remove_table(struct mmu_gather *tlb, void *table)
371 {
372 struct mmu_table_batch **batch = &tlb->batch;
373
374 /*
375 * When there's less then two users of this mm there cannot be a
376 * concurrent page-table walk.
377 */
378 if (atomic_read(&tlb->mm->mm_users) < 2) {
379 __tlb_remove_table(table);
380 return;
381 }
382
383 if (*batch == NULL) {
384 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
385 if (*batch == NULL) {
386 tlb_remove_table_one(table);
387 return;
388 }
389 (*batch)->nr = 0;
390 }
391 (*batch)->tables[(*batch)->nr++] = table;
392 if ((*batch)->nr == MAX_TABLE_BATCH)
393 tlb_table_flush(tlb);
394 }
395
396 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
397
398 /*
399 * Note: this doesn't free the actual pages themselves. That
400 * has been handled earlier when unmapping all the memory regions.
401 */
402 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
403 unsigned long addr)
404 {
405 pgtable_t token = pmd_pgtable(*pmd);
406 pmd_clear(pmd);
407 pte_free_tlb(tlb, token, addr);
408 atomic_long_dec(&tlb->mm->nr_ptes);
409 }
410
411 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
412 unsigned long addr, unsigned long end,
413 unsigned long floor, unsigned long ceiling)
414 {
415 pmd_t *pmd;
416 unsigned long next;
417 unsigned long start;
418
419 start = addr;
420 pmd = pmd_offset(pud, addr);
421 do {
422 next = pmd_addr_end(addr, end);
423 if (pmd_none_or_clear_bad(pmd))
424 continue;
425 free_pte_range(tlb, pmd, addr);
426 } while (pmd++, addr = next, addr != end);
427
428 start &= PUD_MASK;
429 if (start < floor)
430 return;
431 if (ceiling) {
432 ceiling &= PUD_MASK;
433 if (!ceiling)
434 return;
435 }
436 if (end - 1 > ceiling - 1)
437 return;
438
439 pmd = pmd_offset(pud, start);
440 pud_clear(pud);
441 pmd_free_tlb(tlb, pmd, start);
442 mm_dec_nr_pmds(tlb->mm);
443 }
444
445 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
446 unsigned long addr, unsigned long end,
447 unsigned long floor, unsigned long ceiling)
448 {
449 pud_t *pud;
450 unsigned long next;
451 unsigned long start;
452
453 start = addr;
454 pud = pud_offset(pgd, addr);
455 do {
456 next = pud_addr_end(addr, end);
457 if (pud_none_or_clear_bad(pud))
458 continue;
459 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
460 } while (pud++, addr = next, addr != end);
461
462 start &= PGDIR_MASK;
463 if (start < floor)
464 return;
465 if (ceiling) {
466 ceiling &= PGDIR_MASK;
467 if (!ceiling)
468 return;
469 }
470 if (end - 1 > ceiling - 1)
471 return;
472
473 pud = pud_offset(pgd, start);
474 pgd_clear(pgd);
475 pud_free_tlb(tlb, pud, start);
476 }
477
478 /*
479 * This function frees user-level page tables of a process.
480 */
481 void free_pgd_range(struct mmu_gather *tlb,
482 unsigned long addr, unsigned long end,
483 unsigned long floor, unsigned long ceiling)
484 {
485 pgd_t *pgd;
486 unsigned long next;
487
488 /*
489 * The next few lines have given us lots of grief...
490 *
491 * Why are we testing PMD* at this top level? Because often
492 * there will be no work to do at all, and we'd prefer not to
493 * go all the way down to the bottom just to discover that.
494 *
495 * Why all these "- 1"s? Because 0 represents both the bottom
496 * of the address space and the top of it (using -1 for the
497 * top wouldn't help much: the masks would do the wrong thing).
498 * The rule is that addr 0 and floor 0 refer to the bottom of
499 * the address space, but end 0 and ceiling 0 refer to the top
500 * Comparisons need to use "end - 1" and "ceiling - 1" (though
501 * that end 0 case should be mythical).
502 *
503 * Wherever addr is brought up or ceiling brought down, we must
504 * be careful to reject "the opposite 0" before it confuses the
505 * subsequent tests. But what about where end is brought down
506 * by PMD_SIZE below? no, end can't go down to 0 there.
507 *
508 * Whereas we round start (addr) and ceiling down, by different
509 * masks at different levels, in order to test whether a table
510 * now has no other vmas using it, so can be freed, we don't
511 * bother to round floor or end up - the tests don't need that.
512 */
513
514 addr &= PMD_MASK;
515 if (addr < floor) {
516 addr += PMD_SIZE;
517 if (!addr)
518 return;
519 }
520 if (ceiling) {
521 ceiling &= PMD_MASK;
522 if (!ceiling)
523 return;
524 }
525 if (end - 1 > ceiling - 1)
526 end -= PMD_SIZE;
527 if (addr > end - 1)
528 return;
529 /*
530 * We add page table cache pages with PAGE_SIZE,
531 * (see pte_free_tlb()), flush the tlb if we need
532 */
533 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
534 pgd = pgd_offset(tlb->mm, addr);
535 do {
536 next = pgd_addr_end(addr, end);
537 if (pgd_none_or_clear_bad(pgd))
538 continue;
539 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
540 } while (pgd++, addr = next, addr != end);
541 }
542
543 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
544 unsigned long floor, unsigned long ceiling)
545 {
546 while (vma) {
547 struct vm_area_struct *next = vma->vm_next;
548 unsigned long addr = vma->vm_start;
549
550 /*
551 * Hide vma from rmap and truncate_pagecache before freeing
552 * pgtables
553 */
554 unlink_anon_vmas(vma);
555 unlink_file_vma(vma);
556
557 if (is_vm_hugetlb_page(vma)) {
558 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
559 floor, next? next->vm_start: ceiling);
560 } else {
561 /*
562 * Optimization: gather nearby vmas into one call down
563 */
564 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
565 && !is_vm_hugetlb_page(next)) {
566 vma = next;
567 next = vma->vm_next;
568 unlink_anon_vmas(vma);
569 unlink_file_vma(vma);
570 }
571 free_pgd_range(tlb, addr, vma->vm_end,
572 floor, next? next->vm_start: ceiling);
573 }
574 vma = next;
575 }
576 }
577
578 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
579 {
580 spinlock_t *ptl;
581 pgtable_t new = pte_alloc_one(mm, address);
582 if (!new)
583 return -ENOMEM;
584
585 /*
586 * Ensure all pte setup (eg. pte page lock and page clearing) are
587 * visible before the pte is made visible to other CPUs by being
588 * put into page tables.
589 *
590 * The other side of the story is the pointer chasing in the page
591 * table walking code (when walking the page table without locking;
592 * ie. most of the time). Fortunately, these data accesses consist
593 * of a chain of data-dependent loads, meaning most CPUs (alpha
594 * being the notable exception) will already guarantee loads are
595 * seen in-order. See the alpha page table accessors for the
596 * smp_read_barrier_depends() barriers in page table walking code.
597 */
598 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
599
600 ptl = pmd_lock(mm, pmd);
601 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
602 atomic_long_inc(&mm->nr_ptes);
603 pmd_populate(mm, pmd, new);
604 new = NULL;
605 }
606 spin_unlock(ptl);
607 if (new)
608 pte_free(mm, new);
609 return 0;
610 }
611
612 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
613 {
614 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
615 if (!new)
616 return -ENOMEM;
617
618 smp_wmb(); /* See comment in __pte_alloc */
619
620 spin_lock(&init_mm.page_table_lock);
621 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
622 pmd_populate_kernel(&init_mm, pmd, new);
623 new = NULL;
624 }
625 spin_unlock(&init_mm.page_table_lock);
626 if (new)
627 pte_free_kernel(&init_mm, new);
628 return 0;
629 }
630
631 static inline void init_rss_vec(int *rss)
632 {
633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634 }
635
636 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
637 {
638 int i;
639
640 if (current->mm == mm)
641 sync_mm_rss(mm);
642 for (i = 0; i < NR_MM_COUNTERS; i++)
643 if (rss[i])
644 add_mm_counter(mm, i, rss[i]);
645 }
646
647 /*
648 * This function is called to print an error when a bad pte
649 * is found. For example, we might have a PFN-mapped pte in
650 * a region that doesn't allow it.
651 *
652 * The calling function must still handle the error.
653 */
654 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
655 pte_t pte, struct page *page)
656 {
657 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
658 pud_t *pud = pud_offset(pgd, addr);
659 pmd_t *pmd = pmd_offset(pud, addr);
660 struct address_space *mapping;
661 pgoff_t index;
662 static unsigned long resume;
663 static unsigned long nr_shown;
664 static unsigned long nr_unshown;
665
666 /*
667 * Allow a burst of 60 reports, then keep quiet for that minute;
668 * or allow a steady drip of one report per second.
669 */
670 if (nr_shown == 60) {
671 if (time_before(jiffies, resume)) {
672 nr_unshown++;
673 return;
674 }
675 if (nr_unshown) {
676 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
677 nr_unshown);
678 nr_unshown = 0;
679 }
680 nr_shown = 0;
681 }
682 if (nr_shown++ == 0)
683 resume = jiffies + 60 * HZ;
684
685 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
686 index = linear_page_index(vma, addr);
687
688 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
689 current->comm,
690 (long long)pte_val(pte), (long long)pmd_val(*pmd));
691 if (page)
692 dump_page(page, "bad pte");
693 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
694 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
695 /*
696 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
697 */
698 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
699 vma->vm_file,
700 vma->vm_ops ? vma->vm_ops->fault : NULL,
701 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
702 mapping ? mapping->a_ops->readpage : NULL);
703 dump_stack();
704 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
705 }
706
707 /*
708 * vm_normal_page -- This function gets the "struct page" associated with a pte.
709 *
710 * "Special" mappings do not wish to be associated with a "struct page" (either
711 * it doesn't exist, or it exists but they don't want to touch it). In this
712 * case, NULL is returned here. "Normal" mappings do have a struct page.
713 *
714 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
715 * pte bit, in which case this function is trivial. Secondly, an architecture
716 * may not have a spare pte bit, which requires a more complicated scheme,
717 * described below.
718 *
719 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
720 * special mapping (even if there are underlying and valid "struct pages").
721 * COWed pages of a VM_PFNMAP are always normal.
722 *
723 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
724 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
725 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
726 * mapping will always honor the rule
727 *
728 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
729 *
730 * And for normal mappings this is false.
731 *
732 * This restricts such mappings to be a linear translation from virtual address
733 * to pfn. To get around this restriction, we allow arbitrary mappings so long
734 * as the vma is not a COW mapping; in that case, we know that all ptes are
735 * special (because none can have been COWed).
736 *
737 *
738 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
739 *
740 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
741 * page" backing, however the difference is that _all_ pages with a struct
742 * page (that is, those where pfn_valid is true) are refcounted and considered
743 * normal pages by the VM. The disadvantage is that pages are refcounted
744 * (which can be slower and simply not an option for some PFNMAP users). The
745 * advantage is that we don't have to follow the strict linearity rule of
746 * PFNMAP mappings in order to support COWable mappings.
747 *
748 */
749 #ifdef __HAVE_ARCH_PTE_SPECIAL
750 # define HAVE_PTE_SPECIAL 1
751 #else
752 # define HAVE_PTE_SPECIAL 0
753 #endif
754 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
755 pte_t pte)
756 {
757 unsigned long pfn = pte_pfn(pte);
758
759 if (HAVE_PTE_SPECIAL) {
760 if (likely(!pte_special(pte)))
761 goto check_pfn;
762 if (vma->vm_ops && vma->vm_ops->find_special_page)
763 return vma->vm_ops->find_special_page(vma, addr);
764 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
765 return NULL;
766 if (!is_zero_pfn(pfn))
767 print_bad_pte(vma, addr, pte, NULL);
768 return NULL;
769 }
770
771 /* !HAVE_PTE_SPECIAL case follows: */
772
773 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
774 if (vma->vm_flags & VM_MIXEDMAP) {
775 if (!pfn_valid(pfn))
776 return NULL;
777 goto out;
778 } else {
779 unsigned long off;
780 off = (addr - vma->vm_start) >> PAGE_SHIFT;
781 if (pfn == vma->vm_pgoff + off)
782 return NULL;
783 if (!is_cow_mapping(vma->vm_flags))
784 return NULL;
785 }
786 }
787
788 if (is_zero_pfn(pfn))
789 return NULL;
790 check_pfn:
791 if (unlikely(pfn > highest_memmap_pfn)) {
792 print_bad_pte(vma, addr, pte, NULL);
793 return NULL;
794 }
795
796 /*
797 * NOTE! We still have PageReserved() pages in the page tables.
798 * eg. VDSO mappings can cause them to exist.
799 */
800 out:
801 return pfn_to_page(pfn);
802 }
803
804 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
805 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
806 pmd_t pmd)
807 {
808 unsigned long pfn = pmd_pfn(pmd);
809
810 /*
811 * There is no pmd_special() but there may be special pmds, e.g.
812 * in a direct-access (dax) mapping, so let's just replicate the
813 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
814 */
815 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
816 if (vma->vm_flags & VM_MIXEDMAP) {
817 if (!pfn_valid(pfn))
818 return NULL;
819 goto out;
820 } else {
821 unsigned long off;
822 off = (addr - vma->vm_start) >> PAGE_SHIFT;
823 if (pfn == vma->vm_pgoff + off)
824 return NULL;
825 if (!is_cow_mapping(vma->vm_flags))
826 return NULL;
827 }
828 }
829
830 if (is_zero_pfn(pfn))
831 return NULL;
832 if (unlikely(pfn > highest_memmap_pfn))
833 return NULL;
834
835 /*
836 * NOTE! We still have PageReserved() pages in the page tables.
837 * eg. VDSO mappings can cause them to exist.
838 */
839 out:
840 return pfn_to_page(pfn);
841 }
842 #endif
843
844 /*
845 * copy one vm_area from one task to the other. Assumes the page tables
846 * already present in the new task to be cleared in the whole range
847 * covered by this vma.
848 */
849
850 static inline unsigned long
851 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
852 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
853 unsigned long addr, int *rss)
854 {
855 unsigned long vm_flags = vma->vm_flags;
856 pte_t pte = *src_pte;
857 struct page *page;
858
859 /* pte contains position in swap or file, so copy. */
860 if (unlikely(!pte_present(pte))) {
861 swp_entry_t entry = pte_to_swp_entry(pte);
862
863 if (likely(!non_swap_entry(entry))) {
864 if (swap_duplicate(entry) < 0)
865 return entry.val;
866
867 /* make sure dst_mm is on swapoff's mmlist. */
868 if (unlikely(list_empty(&dst_mm->mmlist))) {
869 spin_lock(&mmlist_lock);
870 if (list_empty(&dst_mm->mmlist))
871 list_add(&dst_mm->mmlist,
872 &src_mm->mmlist);
873 spin_unlock(&mmlist_lock);
874 }
875 rss[MM_SWAPENTS]++;
876 } else if (is_migration_entry(entry)) {
877 page = migration_entry_to_page(entry);
878
879 rss[mm_counter(page)]++;
880
881 if (is_write_migration_entry(entry) &&
882 is_cow_mapping(vm_flags)) {
883 /*
884 * COW mappings require pages in both
885 * parent and child to be set to read.
886 */
887 make_migration_entry_read(&entry);
888 pte = swp_entry_to_pte(entry);
889 if (pte_swp_soft_dirty(*src_pte))
890 pte = pte_swp_mksoft_dirty(pte);
891 set_pte_at(src_mm, addr, src_pte, pte);
892 }
893 }
894 goto out_set_pte;
895 }
896
897 /*
898 * If it's a COW mapping, write protect it both
899 * in the parent and the child
900 */
901 if (is_cow_mapping(vm_flags)) {
902 ptep_set_wrprotect(src_mm, addr, src_pte);
903 pte = pte_wrprotect(pte);
904 }
905
906 /*
907 * If it's a shared mapping, mark it clean in
908 * the child
909 */
910 if (vm_flags & VM_SHARED)
911 pte = pte_mkclean(pte);
912 pte = pte_mkold(pte);
913
914 page = vm_normal_page(vma, addr, pte);
915 if (page) {
916 get_page(page);
917 page_dup_rmap(page, false);
918 rss[mm_counter(page)]++;
919 }
920
921 out_set_pte:
922 set_pte_at(dst_mm, addr, dst_pte, pte);
923 return 0;
924 }
925
926 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
927 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
928 unsigned long addr, unsigned long end)
929 {
930 pte_t *orig_src_pte, *orig_dst_pte;
931 pte_t *src_pte, *dst_pte;
932 spinlock_t *src_ptl, *dst_ptl;
933 int progress = 0;
934 int rss[NR_MM_COUNTERS];
935 swp_entry_t entry = (swp_entry_t){0};
936
937 again:
938 init_rss_vec(rss);
939
940 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
941 if (!dst_pte)
942 return -ENOMEM;
943 src_pte = pte_offset_map(src_pmd, addr);
944 src_ptl = pte_lockptr(src_mm, src_pmd);
945 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
946 orig_src_pte = src_pte;
947 orig_dst_pte = dst_pte;
948 arch_enter_lazy_mmu_mode();
949
950 do {
951 /*
952 * We are holding two locks at this point - either of them
953 * could generate latencies in another task on another CPU.
954 */
955 if (progress >= 32) {
956 progress = 0;
957 if (need_resched() ||
958 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
959 break;
960 }
961 if (pte_none(*src_pte)) {
962 progress++;
963 continue;
964 }
965 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
966 vma, addr, rss);
967 if (entry.val)
968 break;
969 progress += 8;
970 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
971
972 arch_leave_lazy_mmu_mode();
973 spin_unlock(src_ptl);
974 pte_unmap(orig_src_pte);
975 add_mm_rss_vec(dst_mm, rss);
976 pte_unmap_unlock(orig_dst_pte, dst_ptl);
977 cond_resched();
978
979 if (entry.val) {
980 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
981 return -ENOMEM;
982 progress = 0;
983 }
984 if (addr != end)
985 goto again;
986 return 0;
987 }
988
989 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
990 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
991 unsigned long addr, unsigned long end)
992 {
993 pmd_t *src_pmd, *dst_pmd;
994 unsigned long next;
995
996 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
997 if (!dst_pmd)
998 return -ENOMEM;
999 src_pmd = pmd_offset(src_pud, addr);
1000 do {
1001 next = pmd_addr_end(addr, end);
1002 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1003 int err;
1004 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
1005 err = copy_huge_pmd(dst_mm, src_mm,
1006 dst_pmd, src_pmd, addr, vma);
1007 if (err == -ENOMEM)
1008 return -ENOMEM;
1009 if (!err)
1010 continue;
1011 /* fall through */
1012 }
1013 if (pmd_none_or_clear_bad(src_pmd))
1014 continue;
1015 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1016 vma, addr, next))
1017 return -ENOMEM;
1018 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1019 return 0;
1020 }
1021
1022 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1023 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1024 unsigned long addr, unsigned long end)
1025 {
1026 pud_t *src_pud, *dst_pud;
1027 unsigned long next;
1028
1029 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1030 if (!dst_pud)
1031 return -ENOMEM;
1032 src_pud = pud_offset(src_pgd, addr);
1033 do {
1034 next = pud_addr_end(addr, end);
1035 if (pud_none_or_clear_bad(src_pud))
1036 continue;
1037 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1038 vma, addr, next))
1039 return -ENOMEM;
1040 } while (dst_pud++, src_pud++, addr = next, addr != end);
1041 return 0;
1042 }
1043
1044 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1045 struct vm_area_struct *vma)
1046 {
1047 pgd_t *src_pgd, *dst_pgd;
1048 unsigned long next;
1049 unsigned long addr = vma->vm_start;
1050 unsigned long end = vma->vm_end;
1051 unsigned long mmun_start; /* For mmu_notifiers */
1052 unsigned long mmun_end; /* For mmu_notifiers */
1053 bool is_cow;
1054 int ret;
1055
1056 /*
1057 * Don't copy ptes where a page fault will fill them correctly.
1058 * Fork becomes much lighter when there are big shared or private
1059 * readonly mappings. The tradeoff is that copy_page_range is more
1060 * efficient than faulting.
1061 */
1062 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1063 !vma->anon_vma)
1064 return 0;
1065
1066 if (is_vm_hugetlb_page(vma))
1067 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1068
1069 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1070 /*
1071 * We do not free on error cases below as remove_vma
1072 * gets called on error from higher level routine
1073 */
1074 ret = track_pfn_copy(vma);
1075 if (ret)
1076 return ret;
1077 }
1078
1079 /*
1080 * We need to invalidate the secondary MMU mappings only when
1081 * there could be a permission downgrade on the ptes of the
1082 * parent mm. And a permission downgrade will only happen if
1083 * is_cow_mapping() returns true.
1084 */
1085 is_cow = is_cow_mapping(vma->vm_flags);
1086 mmun_start = addr;
1087 mmun_end = end;
1088 if (is_cow)
1089 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1090 mmun_end);
1091
1092 ret = 0;
1093 dst_pgd = pgd_offset(dst_mm, addr);
1094 src_pgd = pgd_offset(src_mm, addr);
1095 do {
1096 next = pgd_addr_end(addr, end);
1097 if (pgd_none_or_clear_bad(src_pgd))
1098 continue;
1099 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1100 vma, addr, next))) {
1101 ret = -ENOMEM;
1102 break;
1103 }
1104 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1105
1106 if (is_cow)
1107 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1108 return ret;
1109 }
1110
1111 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1112 struct vm_area_struct *vma, pmd_t *pmd,
1113 unsigned long addr, unsigned long end,
1114 struct zap_details *details)
1115 {
1116 struct mm_struct *mm = tlb->mm;
1117 int force_flush = 0;
1118 int rss[NR_MM_COUNTERS];
1119 spinlock_t *ptl;
1120 pte_t *start_pte;
1121 pte_t *pte;
1122 swp_entry_t entry;
1123
1124 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1125 again:
1126 init_rss_vec(rss);
1127 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1128 pte = start_pte;
1129 arch_enter_lazy_mmu_mode();
1130 do {
1131 pte_t ptent = *pte;
1132 if (pte_none(ptent)) {
1133 continue;
1134 }
1135
1136 if (pte_present(ptent)) {
1137 struct page *page;
1138
1139 page = vm_normal_page(vma, addr, ptent);
1140 if (unlikely(details) && page) {
1141 /*
1142 * unmap_shared_mapping_pages() wants to
1143 * invalidate cache without truncating:
1144 * unmap shared but keep private pages.
1145 */
1146 if (details->check_mapping &&
1147 details->check_mapping != page_rmapping(page))
1148 continue;
1149 }
1150 ptent = ptep_get_and_clear_full(mm, addr, pte,
1151 tlb->fullmm);
1152 tlb_remove_tlb_entry(tlb, pte, addr);
1153 if (unlikely(!page))
1154 continue;
1155
1156 if (!PageAnon(page)) {
1157 if (pte_dirty(ptent)) {
1158 force_flush = 1;
1159 set_page_dirty(page);
1160 }
1161 if (pte_young(ptent) &&
1162 likely(!(vma->vm_flags & VM_SEQ_READ)))
1163 mark_page_accessed(page);
1164 }
1165 rss[mm_counter(page)]--;
1166 page_remove_rmap(page, false);
1167 if (unlikely(page_mapcount(page) < 0))
1168 print_bad_pte(vma, addr, ptent, page);
1169 if (unlikely(__tlb_remove_page(tlb, page))) {
1170 force_flush = 1;
1171 addr += PAGE_SIZE;
1172 break;
1173 }
1174 continue;
1175 }
1176 /* If details->check_mapping, we leave swap entries. */
1177 if (unlikely(details))
1178 continue;
1179
1180 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;
1185
1186 page = migration_entry_to_page(entry);
1187 rss[mm_counter(page)]--;
1188 }
1189 if (unlikely(!free_swap_and_cache(entry)))
1190 print_bad_pte(vma, addr, ptent, NULL);
1191 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1192 } while (pte++, addr += PAGE_SIZE, addr != end);
1193
1194 add_mm_rss_vec(mm, rss);
1195 arch_leave_lazy_mmu_mode();
1196
1197 /* Do the actual TLB flush before dropping ptl */
1198 if (force_flush)
1199 tlb_flush_mmu_tlbonly(tlb);
1200 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);
1211 if (addr != end)
1212 goto again;
1213 }
1214
1215 return addr;
1216 }
1217
1218 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1219 struct vm_area_struct *vma, pud_t *pud,
1220 unsigned long addr, unsigned long end,
1221 struct zap_details *details)
1222 {
1223 pmd_t *pmd;
1224 unsigned long next;
1225
1226 pmd = pmd_offset(pud, addr);
1227 do {
1228 next = pmd_addr_end(addr, end);
1229 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1230 if (next - addr != HPAGE_PMD_SIZE) {
1231 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1232 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1233 __split_huge_pmd(vma, pmd, addr, false, NULL);
1234 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1235 goto next;
1236 /* fall through */
1237 }
1238 /*
1239 * Here there can be other concurrent MADV_DONTNEED or
1240 * trans huge page faults running, and if the pmd is
1241 * none or trans huge it can change under us. This is
1242 * because MADV_DONTNEED holds the mmap_sem in read
1243 * mode.
1244 */
1245 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1246 goto next;
1247 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1248 next:
1249 cond_resched();
1250 } while (pmd++, addr = next, addr != end);
1251
1252 return addr;
1253 }
1254
1255 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1256 struct vm_area_struct *vma, pgd_t *pgd,
1257 unsigned long addr, unsigned long end,
1258 struct zap_details *details)
1259 {
1260 pud_t *pud;
1261 unsigned long next;
1262
1263 pud = pud_offset(pgd, addr);
1264 do {
1265 next = pud_addr_end(addr, end);
1266 if (pud_none_or_clear_bad(pud))
1267 continue;
1268 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1269 } while (pud++, addr = next, addr != end);
1270
1271 return addr;
1272 }
1273
1274 void unmap_page_range(struct mmu_gather *tlb,
1275 struct vm_area_struct *vma,
1276 unsigned long addr, unsigned long end,
1277 struct zap_details *details)
1278 {
1279 pgd_t *pgd;
1280 unsigned long next;
1281
1282 BUG_ON(addr >= end);
1283 tlb_start_vma(tlb, vma);
1284 pgd = pgd_offset(vma->vm_mm, addr);
1285 do {
1286 next = pgd_addr_end(addr, end);
1287 if (pgd_none_or_clear_bad(pgd))
1288 continue;
1289 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1290 } while (pgd++, addr = next, addr != end);
1291 tlb_end_vma(tlb, vma);
1292 }
1293
1294
1295 static void unmap_single_vma(struct mmu_gather *tlb,
1296 struct vm_area_struct *vma, unsigned long start_addr,
1297 unsigned long end_addr,
1298 struct zap_details *details)
1299 {
1300 unsigned long start = max(vma->vm_start, start_addr);
1301 unsigned long end;
1302
1303 if (start >= vma->vm_end)
1304 return;
1305 end = min(vma->vm_end, end_addr);
1306 if (end <= vma->vm_start)
1307 return;
1308
1309 if (vma->vm_file)
1310 uprobe_munmap(vma, start, end);
1311
1312 if (unlikely(vma->vm_flags & VM_PFNMAP))
1313 untrack_pfn(vma, 0, 0);
1314
1315 if (start != end) {
1316 if (unlikely(is_vm_hugetlb_page(vma))) {
1317 /*
1318 * It is undesirable to test vma->vm_file as it
1319 * should be non-null for valid hugetlb area.
1320 * However, vm_file will be NULL in the error
1321 * cleanup path of mmap_region. When
1322 * hugetlbfs ->mmap method fails,
1323 * mmap_region() nullifies vma->vm_file
1324 * before calling this function to clean up.
1325 * Since no pte has actually been setup, it is
1326 * safe to do nothing in this case.
1327 */
1328 if (vma->vm_file) {
1329 i_mmap_lock_write(vma->vm_file->f_mapping);
1330 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1331 i_mmap_unlock_write(vma->vm_file->f_mapping);
1332 }
1333 } else
1334 unmap_page_range(tlb, vma, start, end, details);
1335 }
1336 }
1337
1338 /**
1339 * unmap_vmas - unmap a range of memory covered by a list of vma's
1340 * @tlb: address of the caller's struct mmu_gather
1341 * @vma: the starting vma
1342 * @start_addr: virtual address at which to start unmapping
1343 * @end_addr: virtual address at which to end unmapping
1344 *
1345 * Unmap all pages in the vma list.
1346 *
1347 * Only addresses between `start' and `end' will be unmapped.
1348 *
1349 * The VMA list must be sorted in ascending virtual address order.
1350 *
1351 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1352 * range after unmap_vmas() returns. So the only responsibility here is to
1353 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1354 * drops the lock and schedules.
1355 */
1356 void unmap_vmas(struct mmu_gather *tlb,
1357 struct vm_area_struct *vma, unsigned long start_addr,
1358 unsigned long end_addr)
1359 {
1360 struct mm_struct *mm = vma->vm_mm;
1361
1362 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1363 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1364 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1365 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1366 }
1367
1368 /**
1369 * zap_page_range - remove user pages in a given range
1370 * @vma: vm_area_struct holding the applicable pages
1371 * @start: starting address of pages to zap
1372 * @size: number of bytes to zap
1373 *
1374 * Caller must protect the VMA list
1375 */
1376 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1377 unsigned long size)
1378 {
1379 struct mm_struct *mm = vma->vm_mm;
1380 struct mmu_gather tlb;
1381 unsigned long end = start + size;
1382
1383 lru_add_drain();
1384 tlb_gather_mmu(&tlb, mm, start, end);
1385 update_hiwater_rss(mm);
1386 mmu_notifier_invalidate_range_start(mm, start, end);
1387 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1388 unmap_single_vma(&tlb, vma, start, end, NULL);
1389 mmu_notifier_invalidate_range_end(mm, start, end);
1390 tlb_finish_mmu(&tlb, start, end);
1391 }
1392
1393 /**
1394 * zap_page_range_single - remove user pages in a given range
1395 * @vma: vm_area_struct holding the applicable pages
1396 * @address: starting address of pages to zap
1397 * @size: number of bytes to zap
1398 * @details: details of shared cache invalidation
1399 *
1400 * The range must fit into one VMA.
1401 */
1402 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1403 unsigned long size, struct zap_details *details)
1404 {
1405 struct mm_struct *mm = vma->vm_mm;
1406 struct mmu_gather tlb;
1407 unsigned long end = address + size;
1408
1409 lru_add_drain();
1410 tlb_gather_mmu(&tlb, mm, address, end);
1411 update_hiwater_rss(mm);
1412 mmu_notifier_invalidate_range_start(mm, address, end);
1413 unmap_single_vma(&tlb, vma, address, end, details);
1414 mmu_notifier_invalidate_range_end(mm, address, end);
1415 tlb_finish_mmu(&tlb, address, end);
1416 }
1417
1418 /**
1419 * zap_vma_ptes - remove ptes mapping the vma
1420 * @vma: vm_area_struct holding ptes to be zapped
1421 * @address: starting address of pages to zap
1422 * @size: number of bytes to zap
1423 *
1424 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1425 *
1426 * The entire address range must be fully contained within the vma.
1427 *
1428 * Returns 0 if successful.
1429 */
1430 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1431 unsigned long size)
1432 {
1433 if (address < vma->vm_start || address + size > vma->vm_end ||
1434 !(vma->vm_flags & VM_PFNMAP))
1435 return -1;
1436 zap_page_range_single(vma, address, size, NULL);
1437 return 0;
1438 }
1439 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1440
1441 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1442 spinlock_t **ptl)
1443 {
1444 pgd_t * pgd = pgd_offset(mm, addr);
1445 pud_t * pud = pud_alloc(mm, pgd, addr);
1446 if (pud) {
1447 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1448 if (pmd) {
1449 VM_BUG_ON(pmd_trans_huge(*pmd));
1450 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1451 }
1452 }
1453 return NULL;
1454 }
1455
1456 /*
1457 * This is the old fallback for page remapping.
1458 *
1459 * For historical reasons, it only allows reserved pages. Only
1460 * old drivers should use this, and they needed to mark their
1461 * pages reserved for the old functions anyway.
1462 */
1463 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1464 struct page *page, pgprot_t prot)
1465 {
1466 struct mm_struct *mm = vma->vm_mm;
1467 int retval;
1468 pte_t *pte;
1469 spinlock_t *ptl;
1470
1471 retval = -EINVAL;
1472 if (PageAnon(page))
1473 goto out;
1474 retval = -ENOMEM;
1475 flush_dcache_page(page);
1476 pte = get_locked_pte(mm, addr, &ptl);
1477 if (!pte)
1478 goto out;
1479 retval = -EBUSY;
1480 if (!pte_none(*pte))
1481 goto out_unlock;
1482
1483 /* Ok, finally just insert the thing.. */
1484 get_page(page);
1485 inc_mm_counter_fast(mm, mm_counter_file(page));
1486 page_add_file_rmap(page, false);
1487 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1488
1489 retval = 0;
1490 pte_unmap_unlock(pte, ptl);
1491 return retval;
1492 out_unlock:
1493 pte_unmap_unlock(pte, ptl);
1494 out:
1495 return retval;
1496 }
1497
1498 /**
1499 * vm_insert_page - insert single page into user vma
1500 * @vma: user vma to map to
1501 * @addr: target user address of this page
1502 * @page: source kernel page
1503 *
1504 * This allows drivers to insert individual pages they've allocated
1505 * into a user vma.
1506 *
1507 * The page has to be a nice clean _individual_ kernel allocation.
1508 * If you allocate a compound page, you need to have marked it as
1509 * such (__GFP_COMP), or manually just split the page up yourself
1510 * (see split_page()).
1511 *
1512 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1513 * took an arbitrary page protection parameter. This doesn't allow
1514 * that. Your vma protection will have to be set up correctly, which
1515 * means that if you want a shared writable mapping, you'd better
1516 * ask for a shared writable mapping!
1517 *
1518 * The page does not need to be reserved.
1519 *
1520 * Usually this function is called from f_op->mmap() handler
1521 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1522 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1523 * function from other places, for example from page-fault handler.
1524 */
1525 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1526 struct page *page)
1527 {
1528 if (addr < vma->vm_start || addr >= vma->vm_end)
1529 return -EFAULT;
1530 if (!page_count(page))
1531 return -EINVAL;
1532 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1533 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1534 BUG_ON(vma->vm_flags & VM_PFNMAP);
1535 vma->vm_flags |= VM_MIXEDMAP;
1536 }
1537 return insert_page(vma, addr, page, vma->vm_page_prot);
1538 }
1539 EXPORT_SYMBOL(vm_insert_page);
1540
1541 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1542 pfn_t pfn, pgprot_t prot)
1543 {
1544 struct mm_struct *mm = vma->vm_mm;
1545 int retval;
1546 pte_t *pte, entry;
1547 spinlock_t *ptl;
1548
1549 retval = -ENOMEM;
1550 pte = get_locked_pte(mm, addr, &ptl);
1551 if (!pte)
1552 goto out;
1553 retval = -EBUSY;
1554 if (!pte_none(*pte))
1555 goto out_unlock;
1556
1557 /* Ok, finally just insert the thing.. */
1558 if (pfn_t_devmap(pfn))
1559 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1560 else
1561 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1562 set_pte_at(mm, addr, pte, entry);
1563 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1564
1565 retval = 0;
1566 out_unlock:
1567 pte_unmap_unlock(pte, ptl);
1568 out:
1569 return retval;
1570 }
1571
1572 /**
1573 * vm_insert_pfn - insert single pfn into user vma
1574 * @vma: user vma to map to
1575 * @addr: target user address of this page
1576 * @pfn: source kernel pfn
1577 *
1578 * Similar to vm_insert_page, this allows drivers to insert individual pages
1579 * they've allocated into a user vma. Same comments apply.
1580 *
1581 * This function should only be called from a vm_ops->fault handler, and
1582 * in that case the handler should return NULL.
1583 *
1584 * vma cannot be a COW mapping.
1585 *
1586 * As this is called only for pages that do not currently exist, we
1587 * do not need to flush old virtual caches or the TLB.
1588 */
1589 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1590 unsigned long pfn)
1591 {
1592 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1593 }
1594 EXPORT_SYMBOL(vm_insert_pfn);
1595
1596 /**
1597 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1598 * @vma: user vma to map to
1599 * @addr: target user address of this page
1600 * @pfn: source kernel pfn
1601 * @pgprot: pgprot flags for the inserted page
1602 *
1603 * This is exactly like vm_insert_pfn, except that it allows drivers to
1604 * to override pgprot on a per-page basis.
1605 *
1606 * This only makes sense for IO mappings, and it makes no sense for
1607 * cow mappings. In general, using multiple vmas is preferable;
1608 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1609 * impractical.
1610 */
1611 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1612 unsigned long pfn, pgprot_t pgprot)
1613 {
1614 int ret;
1615 /*
1616 * Technically, architectures with pte_special can avoid all these
1617 * restrictions (same for remap_pfn_range). However we would like
1618 * consistency in testing and feature parity among all, so we should
1619 * try to keep these invariants in place for everybody.
1620 */
1621 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1622 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1623 (VM_PFNMAP|VM_MIXEDMAP));
1624 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1625 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1626
1627 if (addr < vma->vm_start || addr >= vma->vm_end)
1628 return -EFAULT;
1629
1630 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1631
1632 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1633
1634 return ret;
1635 }
1636 EXPORT_SYMBOL(vm_insert_pfn_prot);
1637
1638 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1639 pfn_t pfn)
1640 {
1641 pgprot_t pgprot = vma->vm_page_prot;
1642
1643 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1644
1645 if (addr < vma->vm_start || addr >= vma->vm_end)
1646 return -EFAULT;
1647
1648 track_pfn_insert(vma, &pgprot, pfn);
1649
1650 /*
1651 * If we don't have pte special, then we have to use the pfn_valid()
1652 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1653 * refcount the page if pfn_valid is true (hence insert_page rather
1654 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1655 * without pte special, it would there be refcounted as a normal page.
1656 */
1657 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1658 struct page *page;
1659
1660 /*
1661 * At this point we are committed to insert_page()
1662 * regardless of whether the caller specified flags that
1663 * result in pfn_t_has_page() == false.
1664 */
1665 page = pfn_to_page(pfn_t_to_pfn(pfn));
1666 return insert_page(vma, addr, page, pgprot);
1667 }
1668 return insert_pfn(vma, addr, pfn, pgprot);
1669 }
1670 EXPORT_SYMBOL(vm_insert_mixed);
1671
1672 /*
1673 * maps a range of physical memory into the requested pages. the old
1674 * mappings are removed. any references to nonexistent pages results
1675 * in null mappings (currently treated as "copy-on-access")
1676 */
1677 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1678 unsigned long addr, unsigned long end,
1679 unsigned long pfn, pgprot_t prot)
1680 {
1681 pte_t *pte;
1682 spinlock_t *ptl;
1683
1684 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1685 if (!pte)
1686 return -ENOMEM;
1687 arch_enter_lazy_mmu_mode();
1688 do {
1689 BUG_ON(!pte_none(*pte));
1690 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1691 pfn++;
1692 } while (pte++, addr += PAGE_SIZE, addr != end);
1693 arch_leave_lazy_mmu_mode();
1694 pte_unmap_unlock(pte - 1, ptl);
1695 return 0;
1696 }
1697
1698 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1699 unsigned long addr, unsigned long end,
1700 unsigned long pfn, pgprot_t prot)
1701 {
1702 pmd_t *pmd;
1703 unsigned long next;
1704
1705 pfn -= addr >> PAGE_SHIFT;
1706 pmd = pmd_alloc(mm, pud, addr);
1707 if (!pmd)
1708 return -ENOMEM;
1709 VM_BUG_ON(pmd_trans_huge(*pmd));
1710 do {
1711 next = pmd_addr_end(addr, end);
1712 if (remap_pte_range(mm, pmd, addr, next,
1713 pfn + (addr >> PAGE_SHIFT), prot))
1714 return -ENOMEM;
1715 } while (pmd++, addr = next, addr != end);
1716 return 0;
1717 }
1718
1719 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1720 unsigned long addr, unsigned long end,
1721 unsigned long pfn, pgprot_t prot)
1722 {
1723 pud_t *pud;
1724 unsigned long next;
1725
1726 pfn -= addr >> PAGE_SHIFT;
1727 pud = pud_alloc(mm, pgd, addr);
1728 if (!pud)
1729 return -ENOMEM;
1730 do {
1731 next = pud_addr_end(addr, end);
1732 if (remap_pmd_range(mm, pud, addr, next,
1733 pfn + (addr >> PAGE_SHIFT), prot))
1734 return -ENOMEM;
1735 } while (pud++, addr = next, addr != end);
1736 return 0;
1737 }
1738
1739 /**
1740 * remap_pfn_range - remap kernel memory to userspace
1741 * @vma: user vma to map to
1742 * @addr: target user address to start at
1743 * @pfn: physical address of kernel memory
1744 * @size: size of map area
1745 * @prot: page protection flags for this mapping
1746 *
1747 * Note: this is only safe if the mm semaphore is held when called.
1748 */
1749 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1750 unsigned long pfn, unsigned long size, pgprot_t prot)
1751 {
1752 pgd_t *pgd;
1753 unsigned long next;
1754 unsigned long end = addr + PAGE_ALIGN(size);
1755 struct mm_struct *mm = vma->vm_mm;
1756 unsigned long remap_pfn = pfn;
1757 int err;
1758
1759 /*
1760 * Physically remapped pages are special. Tell the
1761 * rest of the world about it:
1762 * VM_IO tells people not to look at these pages
1763 * (accesses can have side effects).
1764 * VM_PFNMAP tells the core MM that the base pages are just
1765 * raw PFN mappings, and do not have a "struct page" associated
1766 * with them.
1767 * VM_DONTEXPAND
1768 * Disable vma merging and expanding with mremap().
1769 * VM_DONTDUMP
1770 * Omit vma from core dump, even when VM_IO turned off.
1771 *
1772 * There's a horrible special case to handle copy-on-write
1773 * behaviour that some programs depend on. We mark the "original"
1774 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1775 * See vm_normal_page() for details.
1776 */
1777 if (is_cow_mapping(vma->vm_flags)) {
1778 if (addr != vma->vm_start || end != vma->vm_end)
1779 return -EINVAL;
1780 vma->vm_pgoff = pfn;
1781 }
1782
1783 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1784 if (err)
1785 return -EINVAL;
1786
1787 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1788
1789 BUG_ON(addr >= end);
1790 pfn -= addr >> PAGE_SHIFT;
1791 pgd = pgd_offset(mm, addr);
1792 flush_cache_range(vma, addr, end);
1793 do {
1794 next = pgd_addr_end(addr, end);
1795 err = remap_pud_range(mm, pgd, addr, next,
1796 pfn + (addr >> PAGE_SHIFT), prot);
1797 if (err)
1798 break;
1799 } while (pgd++, addr = next, addr != end);
1800
1801 if (err)
1802 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1803
1804 return err;
1805 }
1806 EXPORT_SYMBOL(remap_pfn_range);
1807
1808 /**
1809 * vm_iomap_memory - remap memory to userspace
1810 * @vma: user vma to map to
1811 * @start: start of area
1812 * @len: size of area
1813 *
1814 * This is a simplified io_remap_pfn_range() for common driver use. The
1815 * driver just needs to give us the physical memory range to be mapped,
1816 * we'll figure out the rest from the vma information.
1817 *
1818 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1819 * whatever write-combining details or similar.
1820 */
1821 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1822 {
1823 unsigned long vm_len, pfn, pages;
1824
1825 /* Check that the physical memory area passed in looks valid */
1826 if (start + len < start)
1827 return -EINVAL;
1828 /*
1829 * You *really* shouldn't map things that aren't page-aligned,
1830 * but we've historically allowed it because IO memory might
1831 * just have smaller alignment.
1832 */
1833 len += start & ~PAGE_MASK;
1834 pfn = start >> PAGE_SHIFT;
1835 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1836 if (pfn + pages < pfn)
1837 return -EINVAL;
1838
1839 /* We start the mapping 'vm_pgoff' pages into the area */
1840 if (vma->vm_pgoff > pages)
1841 return -EINVAL;
1842 pfn += vma->vm_pgoff;
1843 pages -= vma->vm_pgoff;
1844
1845 /* Can we fit all of the mapping? */
1846 vm_len = vma->vm_end - vma->vm_start;
1847 if (vm_len >> PAGE_SHIFT > pages)
1848 return -EINVAL;
1849
1850 /* Ok, let it rip */
1851 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1852 }
1853 EXPORT_SYMBOL(vm_iomap_memory);
1854
1855 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1856 unsigned long addr, unsigned long end,
1857 pte_fn_t fn, void *data)
1858 {
1859 pte_t *pte;
1860 int err;
1861 pgtable_t token;
1862 spinlock_t *uninitialized_var(ptl);
1863
1864 pte = (mm == &init_mm) ?
1865 pte_alloc_kernel(pmd, addr) :
1866 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1867 if (!pte)
1868 return -ENOMEM;
1869
1870 BUG_ON(pmd_huge(*pmd));
1871
1872 arch_enter_lazy_mmu_mode();
1873
1874 token = pmd_pgtable(*pmd);
1875
1876 do {
1877 err = fn(pte++, token, addr, data);
1878 if (err)
1879 break;
1880 } while (addr += PAGE_SIZE, addr != end);
1881
1882 arch_leave_lazy_mmu_mode();
1883
1884 if (mm != &init_mm)
1885 pte_unmap_unlock(pte-1, ptl);
1886 return err;
1887 }
1888
1889 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1890 unsigned long addr, unsigned long end,
1891 pte_fn_t fn, void *data)
1892 {
1893 pmd_t *pmd;
1894 unsigned long next;
1895 int err;
1896
1897 BUG_ON(pud_huge(*pud));
1898
1899 pmd = pmd_alloc(mm, pud, addr);
1900 if (!pmd)
1901 return -ENOMEM;
1902 do {
1903 next = pmd_addr_end(addr, end);
1904 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1905 if (err)
1906 break;
1907 } while (pmd++, addr = next, addr != end);
1908 return err;
1909 }
1910
1911 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1912 unsigned long addr, unsigned long end,
1913 pte_fn_t fn, void *data)
1914 {
1915 pud_t *pud;
1916 unsigned long next;
1917 int err;
1918
1919 pud = pud_alloc(mm, pgd, addr);
1920 if (!pud)
1921 return -ENOMEM;
1922 do {
1923 next = pud_addr_end(addr, end);
1924 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1925 if (err)
1926 break;
1927 } while (pud++, addr = next, addr != end);
1928 return err;
1929 }
1930
1931 /*
1932 * Scan a region of virtual memory, filling in page tables as necessary
1933 * and calling a provided function on each leaf page table.
1934 */
1935 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1936 unsigned long size, pte_fn_t fn, void *data)
1937 {
1938 pgd_t *pgd;
1939 unsigned long next;
1940 unsigned long end = addr + size;
1941 int err;
1942
1943 if (WARN_ON(addr >= end))
1944 return -EINVAL;
1945
1946 pgd = pgd_offset(mm, addr);
1947 do {
1948 next = pgd_addr_end(addr, end);
1949 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1950 if (err)
1951 break;
1952 } while (pgd++, addr = next, addr != end);
1953
1954 return err;
1955 }
1956 EXPORT_SYMBOL_GPL(apply_to_page_range);
1957
1958 /*
1959 * handle_pte_fault chooses page fault handler according to an entry which was
1960 * read non-atomically. Before making any commitment, on those architectures
1961 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1962 * parts, do_swap_page must check under lock before unmapping the pte and
1963 * proceeding (but do_wp_page is only called after already making such a check;
1964 * and do_anonymous_page can safely check later on).
1965 */
1966 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1967 pte_t *page_table, pte_t orig_pte)
1968 {
1969 int same = 1;
1970 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1971 if (sizeof(pte_t) > sizeof(unsigned long)) {
1972 spinlock_t *ptl = pte_lockptr(mm, pmd);
1973 spin_lock(ptl);
1974 same = pte_same(*page_table, orig_pte);
1975 spin_unlock(ptl);
1976 }
1977 #endif
1978 pte_unmap(page_table);
1979 return same;
1980 }
1981
1982 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1983 {
1984 debug_dma_assert_idle(src);
1985
1986 /*
1987 * If the source page was a PFN mapping, we don't have
1988 * a "struct page" for it. We do a best-effort copy by
1989 * just copying from the original user address. If that
1990 * fails, we just zero-fill it. Live with it.
1991 */
1992 if (unlikely(!src)) {
1993 void *kaddr = kmap_atomic(dst);
1994 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1995
1996 /*
1997 * This really shouldn't fail, because the page is there
1998 * in the page tables. But it might just be unreadable,
1999 * in which case we just give up and fill the result with
2000 * zeroes.
2001 */
2002 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2003 clear_page(kaddr);
2004 kunmap_atomic(kaddr);
2005 flush_dcache_page(dst);
2006 } else
2007 copy_user_highpage(dst, src, va, vma);
2008 }
2009
2010 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2011 {
2012 struct file *vm_file = vma->vm_file;
2013
2014 if (vm_file)
2015 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2016
2017 /*
2018 * Special mappings (e.g. VDSO) do not have any file so fake
2019 * a default GFP_KERNEL for them.
2020 */
2021 return GFP_KERNEL;
2022 }
2023
2024 /*
2025 * Notify the address space that the page is about to become writable so that
2026 * it can prohibit this or wait for the page to get into an appropriate state.
2027 *
2028 * We do this without the lock held, so that it can sleep if it needs to.
2029 */
2030 static int do_page_mkwrite(struct vm_fault *vmf)
2031 {
2032 int ret;
2033 struct page *page = vmf->page;
2034 unsigned int old_flags = vmf->flags;
2035
2036 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2037
2038 ret = vmf->vma->vm_ops->page_mkwrite(vmf->vma, vmf);
2039 /* Restore original flags so that caller is not surprised */
2040 vmf->flags = old_flags;
2041 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2042 return ret;
2043 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2044 lock_page(page);
2045 if (!page->mapping) {
2046 unlock_page(page);
2047 return 0; /* retry */
2048 }
2049 ret |= VM_FAULT_LOCKED;
2050 } else
2051 VM_BUG_ON_PAGE(!PageLocked(page), page);
2052 return ret;
2053 }
2054
2055 /*
2056 * Handle dirtying of a page in shared file mapping on a write fault.
2057 *
2058 * The function expects the page to be locked and unlocks it.
2059 */
2060 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2061 struct page *page)
2062 {
2063 struct address_space *mapping;
2064 bool dirtied;
2065 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2066
2067 dirtied = set_page_dirty(page);
2068 VM_BUG_ON_PAGE(PageAnon(page), page);
2069 /*
2070 * Take a local copy of the address_space - page.mapping may be zeroed
2071 * by truncate after unlock_page(). The address_space itself remains
2072 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2073 * release semantics to prevent the compiler from undoing this copying.
2074 */
2075 mapping = page_rmapping(page);
2076 unlock_page(page);
2077
2078 if ((dirtied || page_mkwrite) && mapping) {
2079 /*
2080 * Some device drivers do not set page.mapping
2081 * but still dirty their pages
2082 */
2083 balance_dirty_pages_ratelimited(mapping);
2084 }
2085
2086 if (!page_mkwrite)
2087 file_update_time(vma->vm_file);
2088 }
2089
2090 /*
2091 * Handle write page faults for pages that can be reused in the current vma
2092 *
2093 * This can happen either due to the mapping being with the VM_SHARED flag,
2094 * or due to us being the last reference standing to the page. In either
2095 * case, all we need to do here is to mark the page as writable and update
2096 * any related book-keeping.
2097 */
2098 static inline void wp_page_reuse(struct vm_fault *vmf)
2099 __releases(vmf->ptl)
2100 {
2101 struct vm_area_struct *vma = vmf->vma;
2102 struct page *page = vmf->page;
2103 pte_t entry;
2104 /*
2105 * Clear the pages cpupid information as the existing
2106 * information potentially belongs to a now completely
2107 * unrelated process.
2108 */
2109 if (page)
2110 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2111
2112 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2113 entry = pte_mkyoung(vmf->orig_pte);
2114 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2115 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2116 update_mmu_cache(vma, vmf->address, vmf->pte);
2117 pte_unmap_unlock(vmf->pte, vmf->ptl);
2118 }
2119
2120 /*
2121 * Handle the case of a page which we actually need to copy to a new page.
2122 *
2123 * Called with mmap_sem locked and the old page referenced, but
2124 * without the ptl held.
2125 *
2126 * High level logic flow:
2127 *
2128 * - Allocate a page, copy the content of the old page to the new one.
2129 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2130 * - Take the PTL. If the pte changed, bail out and release the allocated page
2131 * - If the pte is still the way we remember it, update the page table and all
2132 * relevant references. This includes dropping the reference the page-table
2133 * held to the old page, as well as updating the rmap.
2134 * - In any case, unlock the PTL and drop the reference we took to the old page.
2135 */
2136 static int wp_page_copy(struct vm_fault *vmf)
2137 {
2138 struct vm_area_struct *vma = vmf->vma;
2139 struct mm_struct *mm = vma->vm_mm;
2140 struct page *old_page = vmf->page;
2141 struct page *new_page = NULL;
2142 pte_t entry;
2143 int page_copied = 0;
2144 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2145 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2146 struct mem_cgroup *memcg;
2147
2148 if (unlikely(anon_vma_prepare(vma)))
2149 goto oom;
2150
2151 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2152 new_page = alloc_zeroed_user_highpage_movable(vma,
2153 vmf->address);
2154 if (!new_page)
2155 goto oom;
2156 } else {
2157 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2158 vmf->address);
2159 if (!new_page)
2160 goto oom;
2161 cow_user_page(new_page, old_page, vmf->address, vma);
2162 }
2163
2164 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2165 goto oom_free_new;
2166
2167 __SetPageUptodate(new_page);
2168
2169 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2170
2171 /*
2172 * Re-check the pte - we dropped the lock
2173 */
2174 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2175 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2176 if (old_page) {
2177 if (!PageAnon(old_page)) {
2178 dec_mm_counter_fast(mm,
2179 mm_counter_file(old_page));
2180 inc_mm_counter_fast(mm, MM_ANONPAGES);
2181 }
2182 } else {
2183 inc_mm_counter_fast(mm, MM_ANONPAGES);
2184 }
2185 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2186 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 */
2194 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2195 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2196 mem_cgroup_commit_charge(new_page, memcg, false, false);
2197 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 */
2203 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2204 update_mmu_cache(vma, vmf->address, vmf->pte);
2205 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 */
2228 page_remove_rmap(old_page, false);
2229 }
2230
2231 /* Free the old page.. */
2232 new_page = old_page;
2233 page_copied = 1;
2234 } else {
2235 mem_cgroup_cancel_charge(new_page, memcg, false);
2236 }
2237
2238 if (new_page)
2239 put_page(new_page);
2240
2241 pte_unmap_unlock(vmf->pte, vmf->ptl);
2242 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 */
2250 if (PageMlocked(old_page))
2251 munlock_vma_page(old_page);
2252 unlock_page(old_page);
2253 }
2254 put_page(old_page);
2255 }
2256 return page_copied ? VM_FAULT_WRITE : 0;
2257 oom_free_new:
2258 put_page(new_page);
2259 oom:
2260 if (old_page)
2261 put_page(old_page);
2262 return VM_FAULT_OOM;
2263 }
2264
2265 /**
2266 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2267 * writeable once the page is prepared
2268 *
2269 * @vmf: structure describing the fault
2270 *
2271 * This function handles all that is needed to finish a write page fault in a
2272 * shared mapping due to PTE being read-only once the mapped page is prepared.
2273 * It handles locking of PTE and modifying it. The function returns
2274 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2275 * lock.
2276 *
2277 * The function expects the page to be locked or other protection against
2278 * concurrent faults / writeback (such as DAX radix tree locks).
2279 */
2280 int finish_mkwrite_fault(struct vm_fault *vmf)
2281 {
2282 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2283 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2284 &vmf->ptl);
2285 /*
2286 * We might have raced with another page fault while we released the
2287 * pte_offset_map_lock.
2288 */
2289 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2290 pte_unmap_unlock(vmf->pte, vmf->ptl);
2291 return VM_FAULT_NOPAGE;
2292 }
2293 wp_page_reuse(vmf);
2294 return 0;
2295 }
2296
2297 /*
2298 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2299 * mapping
2300 */
2301 static int wp_pfn_shared(struct vm_fault *vmf)
2302 {
2303 struct vm_area_struct *vma = vmf->vma;
2304
2305 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2306 int ret;
2307
2308 pte_unmap_unlock(vmf->pte, vmf->ptl);
2309 vmf->flags |= FAULT_FLAG_MKWRITE;
2310 ret = vma->vm_ops->pfn_mkwrite(vma, vmf);
2311 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2312 return ret;
2313 return finish_mkwrite_fault(vmf);
2314 }
2315 wp_page_reuse(vmf);
2316 return VM_FAULT_WRITE;
2317 }
2318
2319 static int wp_page_shared(struct vm_fault *vmf)
2320 __releases(vmf->ptl)
2321 {
2322 struct vm_area_struct *vma = vmf->vma;
2323
2324 get_page(vmf->page);
2325
2326 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2327 int tmp;
2328
2329 pte_unmap_unlock(vmf->pte, vmf->ptl);
2330 tmp = do_page_mkwrite(vmf);
2331 if (unlikely(!tmp || (tmp &
2332 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2333 put_page(vmf->page);
2334 return tmp;
2335 }
2336 tmp = finish_mkwrite_fault(vmf);
2337 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2338 unlock_page(vmf->page);
2339 put_page(vmf->page);
2340 return tmp;
2341 }
2342 } else {
2343 wp_page_reuse(vmf);
2344 lock_page(vmf->page);
2345 }
2346 fault_dirty_shared_page(vma, vmf->page);
2347 put_page(vmf->page);
2348
2349 return VM_FAULT_WRITE;
2350 }
2351
2352 /*
2353 * This routine handles present pages, when users try to write
2354 * to a shared page. It is done by copying the page to a new address
2355 * and decrementing the shared-page counter for the old page.
2356 *
2357 * Note that this routine assumes that the protection checks have been
2358 * done by the caller (the low-level page fault routine in most cases).
2359 * Thus we can safely just mark it writable once we've done any necessary
2360 * COW.
2361 *
2362 * We also mark the page dirty at this point even though the page will
2363 * change only once the write actually happens. This avoids a few races,
2364 * and potentially makes it more efficient.
2365 *
2366 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2367 * but allow concurrent faults), with pte both mapped and locked.
2368 * We return with mmap_sem still held, but pte unmapped and unlocked.
2369 */
2370 static int do_wp_page(struct vm_fault *vmf)
2371 __releases(vmf->ptl)
2372 {
2373 struct vm_area_struct *vma = vmf->vma;
2374
2375 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2376 if (!vmf->page) {
2377 /*
2378 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2379 * VM_PFNMAP VMA.
2380 *
2381 * We should not cow pages in a shared writeable mapping.
2382 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2383 */
2384 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2385 (VM_WRITE|VM_SHARED))
2386 return wp_pfn_shared(vmf);
2387
2388 pte_unmap_unlock(vmf->pte, vmf->ptl);
2389 return wp_page_copy(vmf);
2390 }
2391
2392 /*
2393 * Take out anonymous pages first, anonymous shared vmas are
2394 * not dirty accountable.
2395 */
2396 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2397 int total_mapcount;
2398 if (!trylock_page(vmf->page)) {
2399 get_page(vmf->page);
2400 pte_unmap_unlock(vmf->pte, vmf->ptl);
2401 lock_page(vmf->page);
2402 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2403 vmf->address, &vmf->ptl);
2404 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2405 unlock_page(vmf->page);
2406 pte_unmap_unlock(vmf->pte, vmf->ptl);
2407 put_page(vmf->page);
2408 return 0;
2409 }
2410 put_page(vmf->page);
2411 }
2412 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2413 if (total_mapcount == 1) {
2414 /*
2415 * The page is all ours. Move it to
2416 * our anon_vma so the rmap code will
2417 * not search our parent or siblings.
2418 * Protected against the rmap code by
2419 * the page lock.
2420 */
2421 page_move_anon_rmap(vmf->page, vma);
2422 }
2423 unlock_page(vmf->page);
2424 wp_page_reuse(vmf);
2425 return VM_FAULT_WRITE;
2426 }
2427 unlock_page(vmf->page);
2428 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2429 (VM_WRITE|VM_SHARED))) {
2430 return wp_page_shared(vmf);
2431 }
2432
2433 /*
2434 * Ok, we need to copy. Oh, well..
2435 */
2436 get_page(vmf->page);
2437
2438 pte_unmap_unlock(vmf->pte, vmf->ptl);
2439 return wp_page_copy(vmf);
2440 }
2441
2442 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2443 unsigned long start_addr, unsigned long end_addr,
2444 struct zap_details *details)
2445 {
2446 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2447 }
2448
2449 static inline void unmap_mapping_range_tree(struct rb_root *root,
2450 struct zap_details *details)
2451 {
2452 struct vm_area_struct *vma;
2453 pgoff_t vba, vea, zba, zea;
2454
2455 vma_interval_tree_foreach(vma, root,
2456 details->first_index, details->last_index) {
2457
2458 vba = vma->vm_pgoff;
2459 vea = vba + vma_pages(vma) - 1;
2460 zba = details->first_index;
2461 if (zba < vba)
2462 zba = vba;
2463 zea = details->last_index;
2464 if (zea > vea)
2465 zea = vea;
2466
2467 unmap_mapping_range_vma(vma,
2468 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2469 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2470 details);
2471 }
2472 }
2473
2474 /**
2475 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2476 * address_space corresponding to the specified page range in the underlying
2477 * file.
2478 *
2479 * @mapping: the address space containing mmaps to be unmapped.
2480 * @holebegin: byte in first page to unmap, relative to the start of
2481 * the underlying file. This will be rounded down to a PAGE_SIZE
2482 * boundary. Note that this is different from truncate_pagecache(), which
2483 * must keep the partial page. In contrast, we must get rid of
2484 * partial pages.
2485 * @holelen: size of prospective hole in bytes. This will be rounded
2486 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2487 * end of the file.
2488 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2489 * but 0 when invalidating pagecache, don't throw away private data.
2490 */
2491 void unmap_mapping_range(struct address_space *mapping,
2492 loff_t const holebegin, loff_t const holelen, int even_cows)
2493 {
2494 struct zap_details details = { };
2495 pgoff_t hba = holebegin >> PAGE_SHIFT;
2496 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2497
2498 /* Check for overflow. */
2499 if (sizeof(holelen) > sizeof(hlen)) {
2500 long long holeend =
2501 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2502 if (holeend & ~(long long)ULONG_MAX)
2503 hlen = ULONG_MAX - hba + 1;
2504 }
2505
2506 details.check_mapping = even_cows? NULL: mapping;
2507 details.first_index = hba;
2508 details.last_index = hba + hlen - 1;
2509 if (details.last_index < details.first_index)
2510 details.last_index = ULONG_MAX;
2511
2512 i_mmap_lock_write(mapping);
2513 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2514 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2515 i_mmap_unlock_write(mapping);
2516 }
2517 EXPORT_SYMBOL(unmap_mapping_range);
2518
2519 /*
2520 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2521 * but allow concurrent faults), and pte mapped but not yet locked.
2522 * We return with pte unmapped and unlocked.
2523 *
2524 * We return with the mmap_sem locked or unlocked in the same cases
2525 * as does filemap_fault().
2526 */
2527 int do_swap_page(struct vm_fault *vmf)
2528 {
2529 struct vm_area_struct *vma = vmf->vma;
2530 struct page *page, *swapcache;
2531 struct mem_cgroup *memcg;
2532 swp_entry_t entry;
2533 pte_t pte;
2534 int locked;
2535 int exclusive = 0;
2536 int ret = 0;
2537
2538 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2539 goto out;
2540
2541 entry = pte_to_swp_entry(vmf->orig_pte);
2542 if (unlikely(non_swap_entry(entry))) {
2543 if (is_migration_entry(entry)) {
2544 migration_entry_wait(vma->vm_mm, vmf->pmd,
2545 vmf->address);
2546 } else if (is_hwpoison_entry(entry)) {
2547 ret = VM_FAULT_HWPOISON;
2548 } else {
2549 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2550 ret = VM_FAULT_SIGBUS;
2551 }
2552 goto out;
2553 }
2554 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2555 page = lookup_swap_cache(entry);
2556 if (!page) {
2557 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2558 vmf->address);
2559 if (!page) {
2560 /*
2561 * Back out if somebody else faulted in this pte
2562 * while we released the pte lock.
2563 */
2564 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2565 vmf->address, &vmf->ptl);
2566 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2567 ret = VM_FAULT_OOM;
2568 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2569 goto unlock;
2570 }
2571
2572 /* Had to read the page from swap area: Major fault */
2573 ret = VM_FAULT_MAJOR;
2574 count_vm_event(PGMAJFAULT);
2575 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2576 } else if (PageHWPoison(page)) {
2577 /*
2578 * hwpoisoned dirty swapcache pages are kept for killing
2579 * owner processes (which may be unknown at hwpoison time)
2580 */
2581 ret = VM_FAULT_HWPOISON;
2582 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2583 swapcache = page;
2584 goto out_release;
2585 }
2586
2587 swapcache = page;
2588 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2589
2590 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2591 if (!locked) {
2592 ret |= VM_FAULT_RETRY;
2593 goto out_release;
2594 }
2595
2596 /*
2597 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2598 * release the swapcache from under us. The page pin, and pte_same
2599 * test below, are not enough to exclude that. Even if it is still
2600 * swapcache, we need to check that the page's swap has not changed.
2601 */
2602 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2603 goto out_page;
2604
2605 page = ksm_might_need_to_copy(page, vma, vmf->address);
2606 if (unlikely(!page)) {
2607 ret = VM_FAULT_OOM;
2608 page = swapcache;
2609 goto out_page;
2610 }
2611
2612 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2613 &memcg, false)) {
2614 ret = VM_FAULT_OOM;
2615 goto out_page;
2616 }
2617
2618 /*
2619 * Back out if somebody else already faulted in this pte.
2620 */
2621 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2622 &vmf->ptl);
2623 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2624 goto out_nomap;
2625
2626 if (unlikely(!PageUptodate(page))) {
2627 ret = VM_FAULT_SIGBUS;
2628 goto out_nomap;
2629 }
2630
2631 /*
2632 * The page isn't present yet, go ahead with the fault.
2633 *
2634 * Be careful about the sequence of operations here.
2635 * To get its accounting right, reuse_swap_page() must be called
2636 * while the page is counted on swap but not yet in mapcount i.e.
2637 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2638 * must be called after the swap_free(), or it will never succeed.
2639 */
2640
2641 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2642 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2643 pte = mk_pte(page, vma->vm_page_prot);
2644 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2645 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2646 vmf->flags &= ~FAULT_FLAG_WRITE;
2647 ret |= VM_FAULT_WRITE;
2648 exclusive = RMAP_EXCLUSIVE;
2649 }
2650 flush_icache_page(vma, page);
2651 if (pte_swp_soft_dirty(vmf->orig_pte))
2652 pte = pte_mksoft_dirty(pte);
2653 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2654 vmf->orig_pte = pte;
2655 if (page == swapcache) {
2656 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2657 mem_cgroup_commit_charge(page, memcg, true, false);
2658 activate_page(page);
2659 } else { /* ksm created a completely new copy */
2660 page_add_new_anon_rmap(page, vma, vmf->address, false);
2661 mem_cgroup_commit_charge(page, memcg, false, false);
2662 lru_cache_add_active_or_unevictable(page, vma);
2663 }
2664
2665 swap_free(entry);
2666 if (mem_cgroup_swap_full(page) ||
2667 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2668 try_to_free_swap(page);
2669 unlock_page(page);
2670 if (page != swapcache) {
2671 /*
2672 * Hold the lock to avoid the swap entry to be reused
2673 * until we take the PT lock for the pte_same() check
2674 * (to avoid false positives from pte_same). For
2675 * further safety release the lock after the swap_free
2676 * so that the swap count won't change under a
2677 * parallel locked swapcache.
2678 */
2679 unlock_page(swapcache);
2680 put_page(swapcache);
2681 }
2682
2683 if (vmf->flags & FAULT_FLAG_WRITE) {
2684 ret |= do_wp_page(vmf);
2685 if (ret & VM_FAULT_ERROR)
2686 ret &= VM_FAULT_ERROR;
2687 goto out;
2688 }
2689
2690 /* No need to invalidate - it was non-present before */
2691 update_mmu_cache(vma, vmf->address, vmf->pte);
2692 unlock:
2693 pte_unmap_unlock(vmf->pte, vmf->ptl);
2694 out:
2695 return ret;
2696 out_nomap:
2697 mem_cgroup_cancel_charge(page, memcg, false);
2698 pte_unmap_unlock(vmf->pte, vmf->ptl);
2699 out_page:
2700 unlock_page(page);
2701 out_release:
2702 put_page(page);
2703 if (page != swapcache) {
2704 unlock_page(swapcache);
2705 put_page(swapcache);
2706 }
2707 return ret;
2708 }
2709
2710 /*
2711 * This is like a special single-page "expand_{down|up}wards()",
2712 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2713 * doesn't hit another vma.
2714 */
2715 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2716 {
2717 address &= PAGE_MASK;
2718 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2719 struct vm_area_struct *prev = vma->vm_prev;
2720
2721 /*
2722 * Is there a mapping abutting this one below?
2723 *
2724 * That's only ok if it's the same stack mapping
2725 * that has gotten split..
2726 */
2727 if (prev && prev->vm_end == address)
2728 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2729
2730 return expand_downwards(vma, address - PAGE_SIZE);
2731 }
2732 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2733 struct vm_area_struct *next = vma->vm_next;
2734
2735 /* As VM_GROWSDOWN but s/below/above/ */
2736 if (next && next->vm_start == address + PAGE_SIZE)
2737 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2738
2739 return expand_upwards(vma, address + PAGE_SIZE);
2740 }
2741 return 0;
2742 }
2743
2744 /*
2745 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2746 * but allow concurrent faults), and pte mapped but not yet locked.
2747 * We return with mmap_sem still held, but pte unmapped and unlocked.
2748 */
2749 static int do_anonymous_page(struct vm_fault *vmf)
2750 {
2751 struct vm_area_struct *vma = vmf->vma;
2752 struct mem_cgroup *memcg;
2753 struct page *page;
2754 pte_t entry;
2755
2756 /* File mapping without ->vm_ops ? */
2757 if (vma->vm_flags & VM_SHARED)
2758 return VM_FAULT_SIGBUS;
2759
2760 /* Check if we need to add a guard page to the stack */
2761 if (check_stack_guard_page(vma, vmf->address) < 0)
2762 return VM_FAULT_SIGSEGV;
2763
2764 /*
2765 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2766 * pte_offset_map() on pmds where a huge pmd might be created
2767 * from a different thread.
2768 *
2769 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2770 * parallel threads are excluded by other means.
2771 *
2772 * Here we only have down_read(mmap_sem).
2773 */
2774 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2775 return VM_FAULT_OOM;
2776
2777 /* See the comment in pte_alloc_one_map() */
2778 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2779 return 0;
2780
2781 /* Use the zero-page for reads */
2782 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2783 !mm_forbids_zeropage(vma->vm_mm)) {
2784 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2785 vma->vm_page_prot));
2786 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2787 vmf->address, &vmf->ptl);
2788 if (!pte_none(*vmf->pte))
2789 goto unlock;
2790 /* Deliver the page fault to userland, check inside PT lock */
2791 if (userfaultfd_missing(vma)) {
2792 pte_unmap_unlock(vmf->pte, vmf->ptl);
2793 return handle_userfault(vmf, VM_UFFD_MISSING);
2794 }
2795 goto setpte;
2796 }
2797
2798 /* Allocate our own private page. */
2799 if (unlikely(anon_vma_prepare(vma)))
2800 goto oom;
2801 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2802 if (!page)
2803 goto oom;
2804
2805 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2806 goto oom_free_page;
2807
2808 /*
2809 * The memory barrier inside __SetPageUptodate makes sure that
2810 * preceeding stores to the page contents become visible before
2811 * the set_pte_at() write.
2812 */
2813 __SetPageUptodate(page);
2814
2815 entry = mk_pte(page, vma->vm_page_prot);
2816 if (vma->vm_flags & VM_WRITE)
2817 entry = pte_mkwrite(pte_mkdirty(entry));
2818
2819 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2820 &vmf->ptl);
2821 if (!pte_none(*vmf->pte))
2822 goto release;
2823
2824 /* Deliver the page fault to userland, check inside PT lock */
2825 if (userfaultfd_missing(vma)) {
2826 pte_unmap_unlock(vmf->pte, vmf->ptl);
2827 mem_cgroup_cancel_charge(page, memcg, false);
2828 put_page(page);
2829 return handle_userfault(vmf, VM_UFFD_MISSING);
2830 }
2831
2832 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2833 page_add_new_anon_rmap(page, vma, vmf->address, false);
2834 mem_cgroup_commit_charge(page, memcg, false, false);
2835 lru_cache_add_active_or_unevictable(page, vma);
2836 setpte:
2837 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2838
2839 /* No need to invalidate - it was non-present before */
2840 update_mmu_cache(vma, vmf->address, vmf->pte);
2841 unlock:
2842 pte_unmap_unlock(vmf->pte, vmf->ptl);
2843 return 0;
2844 release:
2845 mem_cgroup_cancel_charge(page, memcg, false);
2846 put_page(page);
2847 goto unlock;
2848 oom_free_page:
2849 put_page(page);
2850 oom:
2851 return VM_FAULT_OOM;
2852 }
2853
2854 /*
2855 * The mmap_sem must have been held on entry, and may have been
2856 * released depending on flags and vma->vm_ops->fault() return value.
2857 * See filemap_fault() and __lock_page_retry().
2858 */
2859 static int __do_fault(struct vm_fault *vmf)
2860 {
2861 struct vm_area_struct *vma = vmf->vma;
2862 int ret;
2863
2864 ret = vma->vm_ops->fault(vma, vmf);
2865 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2866 VM_FAULT_DONE_COW)))
2867 return ret;
2868
2869 if (unlikely(PageHWPoison(vmf->page))) {
2870 if (ret & VM_FAULT_LOCKED)
2871 unlock_page(vmf->page);
2872 put_page(vmf->page);
2873 vmf->page = NULL;
2874 return VM_FAULT_HWPOISON;
2875 }
2876
2877 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2878 lock_page(vmf->page);
2879 else
2880 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2881
2882 return ret;
2883 }
2884
2885 static int pte_alloc_one_map(struct vm_fault *vmf)
2886 {
2887 struct vm_area_struct *vma = vmf->vma;
2888
2889 if (!pmd_none(*vmf->pmd))
2890 goto map_pte;
2891 if (vmf->prealloc_pte) {
2892 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2893 if (unlikely(!pmd_none(*vmf->pmd))) {
2894 spin_unlock(vmf->ptl);
2895 goto map_pte;
2896 }
2897
2898 atomic_long_inc(&vma->vm_mm->nr_ptes);
2899 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2900 spin_unlock(vmf->ptl);
2901 vmf->prealloc_pte = 0;
2902 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
2903 return VM_FAULT_OOM;
2904 }
2905 map_pte:
2906 /*
2907 * If a huge pmd materialized under us just retry later. Use
2908 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2909 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2910 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2911 * in a different thread of this mm, in turn leading to a misleading
2912 * pmd_trans_huge() retval. All we have to ensure is that it is a
2913 * regular pmd that we can walk with pte_offset_map() and we can do that
2914 * through an atomic read in C, which is what pmd_trans_unstable()
2915 * provides.
2916 */
2917 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
2918 return VM_FAULT_NOPAGE;
2919
2920 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2921 &vmf->ptl);
2922 return 0;
2923 }
2924
2925 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2926
2927 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2928 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2929 unsigned long haddr)
2930 {
2931 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2932 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2933 return false;
2934 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2935 return false;
2936 return true;
2937 }
2938
2939 static void deposit_prealloc_pte(struct vm_fault *vmf)
2940 {
2941 struct vm_area_struct *vma = vmf->vma;
2942
2943 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2944 /*
2945 * We are going to consume the prealloc table,
2946 * count that as nr_ptes.
2947 */
2948 atomic_long_inc(&vma->vm_mm->nr_ptes);
2949 vmf->prealloc_pte = 0;
2950 }
2951
2952 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
2953 {
2954 struct vm_area_struct *vma = vmf->vma;
2955 bool write = vmf->flags & FAULT_FLAG_WRITE;
2956 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2957 pmd_t entry;
2958 int i, ret;
2959
2960 if (!transhuge_vma_suitable(vma, haddr))
2961 return VM_FAULT_FALLBACK;
2962
2963 ret = VM_FAULT_FALLBACK;
2964 page = compound_head(page);
2965
2966 /*
2967 * Archs like ppc64 need additonal space to store information
2968 * related to pte entry. Use the preallocated table for that.
2969 */
2970 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
2971 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
2972 if (!vmf->prealloc_pte)
2973 return VM_FAULT_OOM;
2974 smp_wmb(); /* See comment in __pte_alloc() */
2975 }
2976
2977 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2978 if (unlikely(!pmd_none(*vmf->pmd)))
2979 goto out;
2980
2981 for (i = 0; i < HPAGE_PMD_NR; i++)
2982 flush_icache_page(vma, page + i);
2983
2984 entry = mk_huge_pmd(page, vma->vm_page_prot);
2985 if (write)
2986 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2987
2988 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2989 page_add_file_rmap(page, true);
2990 /*
2991 * deposit and withdraw with pmd lock held
2992 */
2993 if (arch_needs_pgtable_deposit())
2994 deposit_prealloc_pte(vmf);
2995
2996 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
2997
2998 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
2999
3000 /* fault is handled */
3001 ret = 0;
3002 count_vm_event(THP_FILE_MAPPED);
3003 out:
3004 spin_unlock(vmf->ptl);
3005 return ret;
3006 }
3007 #else
3008 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3009 {
3010 BUILD_BUG();
3011 return 0;
3012 }
3013 #endif
3014
3015 /**
3016 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3017 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3018 *
3019 * @vmf: fault environment
3020 * @memcg: memcg to charge page (only for private mappings)
3021 * @page: page to map
3022 *
3023 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3024 * return.
3025 *
3026 * Target users are page handler itself and implementations of
3027 * vm_ops->map_pages.
3028 */
3029 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3030 struct page *page)
3031 {
3032 struct vm_area_struct *vma = vmf->vma;
3033 bool write = vmf->flags & FAULT_FLAG_WRITE;
3034 pte_t entry;
3035 int ret;
3036
3037 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3038 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3039 /* THP on COW? */
3040 VM_BUG_ON_PAGE(memcg, page);
3041
3042 ret = do_set_pmd(vmf, page);
3043 if (ret != VM_FAULT_FALLBACK)
3044 return ret;
3045 }
3046
3047 if (!vmf->pte) {
3048 ret = pte_alloc_one_map(vmf);
3049 if (ret)
3050 return ret;
3051 }
3052
3053 /* Re-check under ptl */
3054 if (unlikely(!pte_none(*vmf->pte)))
3055 return VM_FAULT_NOPAGE;
3056
3057 flush_icache_page(vma, page);
3058 entry = mk_pte(page, vma->vm_page_prot);
3059 if (write)
3060 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3061 /* copy-on-write page */
3062 if (write && !(vma->vm_flags & VM_SHARED)) {
3063 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3064 page_add_new_anon_rmap(page, vma, vmf->address, false);
3065 mem_cgroup_commit_charge(page, memcg, false, false);
3066 lru_cache_add_active_or_unevictable(page, vma);
3067 } else {
3068 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3069 page_add_file_rmap(page, false);
3070 }
3071 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3072
3073 /* no need to invalidate: a not-present page won't be cached */
3074 update_mmu_cache(vma, vmf->address, vmf->pte);
3075
3076 return 0;
3077 }
3078
3079
3080 /**
3081 * finish_fault - finish page fault once we have prepared the page to fault
3082 *
3083 * @vmf: structure describing the fault
3084 *
3085 * This function handles all that is needed to finish a page fault once the
3086 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3087 * given page, adds reverse page mapping, handles memcg charges and LRU
3088 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3089 * error.
3090 *
3091 * The function expects the page to be locked and on success it consumes a
3092 * reference of a page being mapped (for the PTE which maps it).
3093 */
3094 int finish_fault(struct vm_fault *vmf)
3095 {
3096 struct page *page;
3097 int ret;
3098
3099 /* Did we COW the page? */
3100 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3101 !(vmf->vma->vm_flags & VM_SHARED))
3102 page = vmf->cow_page;
3103 else
3104 page = vmf->page;
3105 ret = alloc_set_pte(vmf, vmf->memcg, page);
3106 if (vmf->pte)
3107 pte_unmap_unlock(vmf->pte, vmf->ptl);
3108 return ret;
3109 }
3110
3111 static unsigned long fault_around_bytes __read_mostly =
3112 rounddown_pow_of_two(65536);
3113
3114 #ifdef CONFIG_DEBUG_FS
3115 static int fault_around_bytes_get(void *data, u64 *val)
3116 {
3117 *val = fault_around_bytes;
3118 return 0;
3119 }
3120
3121 /*
3122 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3123 * rounded down to nearest page order. It's what do_fault_around() expects to
3124 * see.
3125 */
3126 static int fault_around_bytes_set(void *data, u64 val)
3127 {
3128 if (val / PAGE_SIZE > PTRS_PER_PTE)
3129 return -EINVAL;
3130 if (val > PAGE_SIZE)
3131 fault_around_bytes = rounddown_pow_of_two(val);
3132 else
3133 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3134 return 0;
3135 }
3136 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3137 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3138
3139 static int __init fault_around_debugfs(void)
3140 {
3141 void *ret;
3142
3143 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3144 &fault_around_bytes_fops);
3145 if (!ret)
3146 pr_warn("Failed to create fault_around_bytes in debugfs");
3147 return 0;
3148 }
3149 late_initcall(fault_around_debugfs);
3150 #endif
3151
3152 /*
3153 * do_fault_around() tries to map few pages around the fault address. The hope
3154 * is that the pages will be needed soon and this will lower the number of
3155 * faults to handle.
3156 *
3157 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3158 * not ready to be mapped: not up-to-date, locked, etc.
3159 *
3160 * This function is called with the page table lock taken. In the split ptlock
3161 * case the page table lock only protects only those entries which belong to
3162 * the page table corresponding to the fault address.
3163 *
3164 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3165 * only once.
3166 *
3167 * fault_around_pages() defines how many pages we'll try to map.
3168 * do_fault_around() expects it to return a power of two less than or equal to
3169 * PTRS_PER_PTE.
3170 *
3171 * The virtual address of the area that we map is naturally aligned to the
3172 * fault_around_pages() value (and therefore to page order). This way it's
3173 * easier to guarantee that we don't cross page table boundaries.
3174 */
3175 static int do_fault_around(struct vm_fault *vmf)
3176 {
3177 unsigned long address = vmf->address, nr_pages, mask;
3178 pgoff_t start_pgoff = vmf->pgoff;
3179 pgoff_t end_pgoff;
3180 int off, ret = 0;
3181
3182 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3183 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3184
3185 vmf->address = max(address & mask, vmf->vma->vm_start);
3186 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3187 start_pgoff -= off;
3188
3189 /*
3190 * end_pgoff is either end of page table or end of vma
3191 * or fault_around_pages() from start_pgoff, depending what is nearest.
3192 */
3193 end_pgoff = start_pgoff -
3194 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3195 PTRS_PER_PTE - 1;
3196 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3197 start_pgoff + nr_pages - 1);
3198
3199 if (pmd_none(*vmf->pmd)) {
3200 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3201 vmf->address);
3202 if (!vmf->prealloc_pte)
3203 goto out;
3204 smp_wmb(); /* See comment in __pte_alloc() */
3205 }
3206
3207 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3208
3209 /* Huge page is mapped? Page fault is solved */
3210 if (pmd_trans_huge(*vmf->pmd)) {
3211 ret = VM_FAULT_NOPAGE;
3212 goto out;
3213 }
3214
3215 /* ->map_pages() haven't done anything useful. Cold page cache? */
3216 if (!vmf->pte)
3217 goto out;
3218
3219 /* check if the page fault is solved */
3220 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3221 if (!pte_none(*vmf->pte))
3222 ret = VM_FAULT_NOPAGE;
3223 pte_unmap_unlock(vmf->pte, vmf->ptl);
3224 out:
3225 vmf->address = address;
3226 vmf->pte = NULL;
3227 return ret;
3228 }
3229
3230 static int do_read_fault(struct vm_fault *vmf)
3231 {
3232 struct vm_area_struct *vma = vmf->vma;
3233 int ret = 0;
3234
3235 /*
3236 * Let's call ->map_pages() first and use ->fault() as fallback
3237 * if page by the offset is not ready to be mapped (cold cache or
3238 * something).
3239 */
3240 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3241 ret = do_fault_around(vmf);
3242 if (ret)
3243 return ret;
3244 }
3245
3246 ret = __do_fault(vmf);
3247 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3248 return ret;
3249
3250 ret |= finish_fault(vmf);
3251 unlock_page(vmf->page);
3252 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3253 put_page(vmf->page);
3254 return ret;
3255 }
3256
3257 static int do_cow_fault(struct vm_fault *vmf)
3258 {
3259 struct vm_area_struct *vma = vmf->vma;
3260 int ret;
3261
3262 if (unlikely(anon_vma_prepare(vma)))
3263 return VM_FAULT_OOM;
3264
3265 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3266 if (!vmf->cow_page)
3267 return VM_FAULT_OOM;
3268
3269 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3270 &vmf->memcg, false)) {
3271 put_page(vmf->cow_page);
3272 return VM_FAULT_OOM;
3273 }
3274
3275 ret = __do_fault(vmf);
3276 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3277 goto uncharge_out;
3278 if (ret & VM_FAULT_DONE_COW)
3279 return ret;
3280
3281 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3282 __SetPageUptodate(vmf->cow_page);
3283
3284 ret |= finish_fault(vmf);
3285 unlock_page(vmf->page);
3286 put_page(vmf->page);
3287 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3288 goto uncharge_out;
3289 return ret;
3290 uncharge_out:
3291 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3292 put_page(vmf->cow_page);
3293 return ret;
3294 }
3295
3296 static int do_shared_fault(struct vm_fault *vmf)
3297 {
3298 struct vm_area_struct *vma = vmf->vma;
3299 int ret, tmp;
3300
3301 ret = __do_fault(vmf);
3302 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3303 return ret;
3304
3305 /*
3306 * Check if the backing address space wants to know that the page is
3307 * about to become writable
3308 */
3309 if (vma->vm_ops->page_mkwrite) {
3310 unlock_page(vmf->page);
3311 tmp = do_page_mkwrite(vmf);
3312 if (unlikely(!tmp ||
3313 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3314 put_page(vmf->page);
3315 return tmp;
3316 }
3317 }
3318
3319 ret |= finish_fault(vmf);
3320 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3321 VM_FAULT_RETRY))) {
3322 unlock_page(vmf->page);
3323 put_page(vmf->page);
3324 return ret;
3325 }
3326
3327 fault_dirty_shared_page(vma, vmf->page);
3328 return ret;
3329 }
3330
3331 /*
3332 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3333 * but allow concurrent faults).
3334 * The mmap_sem may have been released depending on flags and our
3335 * return value. See filemap_fault() and __lock_page_or_retry().
3336 */
3337 static int do_fault(struct vm_fault *vmf)
3338 {
3339 struct vm_area_struct *vma = vmf->vma;
3340 int ret;
3341
3342 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3343 if (!vma->vm_ops->fault)
3344 ret = VM_FAULT_SIGBUS;
3345 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3346 ret = do_read_fault(vmf);
3347 else if (!(vma->vm_flags & VM_SHARED))
3348 ret = do_cow_fault(vmf);
3349 else
3350 ret = do_shared_fault(vmf);
3351
3352 /* preallocated pagetable is unused: free it */
3353 if (vmf->prealloc_pte) {
3354 pte_free(vma->vm_mm, vmf->prealloc_pte);
3355 vmf->prealloc_pte = 0;
3356 }
3357 return ret;
3358 }
3359
3360 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3361 unsigned long addr, int page_nid,
3362 int *flags)
3363 {
3364 get_page(page);
3365
3366 count_vm_numa_event(NUMA_HINT_FAULTS);
3367 if (page_nid == numa_node_id()) {
3368 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3369 *flags |= TNF_FAULT_LOCAL;
3370 }
3371
3372 return mpol_misplaced(page, vma, addr);
3373 }
3374
3375 static int do_numa_page(struct vm_fault *vmf)
3376 {
3377 struct vm_area_struct *vma = vmf->vma;
3378 struct page *page = NULL;
3379 int page_nid = -1;
3380 int last_cpupid;
3381 int target_nid;
3382 bool migrated = false;
3383 pte_t pte = vmf->orig_pte;
3384 bool was_writable = pte_write(pte);
3385 int flags = 0;
3386
3387 /*
3388 * The "pte" at this point cannot be used safely without
3389 * validation through pte_unmap_same(). It's of NUMA type but
3390 * the pfn may be screwed if the read is non atomic.
3391 *
3392 * We can safely just do a "set_pte_at()", because the old
3393 * page table entry is not accessible, so there would be no
3394 * concurrent hardware modifications to the PTE.
3395 */
3396 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3397 spin_lock(vmf->ptl);
3398 if (unlikely(!pte_same(*vmf->pte, pte))) {
3399 pte_unmap_unlock(vmf->pte, vmf->ptl);
3400 goto out;
3401 }
3402
3403 /* Make it present again */
3404 pte = pte_modify(pte, vma->vm_page_prot);
3405 pte = pte_mkyoung(pte);
3406 if (was_writable)
3407 pte = pte_mkwrite(pte);
3408 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3409 update_mmu_cache(vma, vmf->address, vmf->pte);
3410
3411 page = vm_normal_page(vma, vmf->address, pte);
3412 if (!page) {
3413 pte_unmap_unlock(vmf->pte, vmf->ptl);
3414 return 0;
3415 }
3416
3417 /* TODO: handle PTE-mapped THP */
3418 if (PageCompound(page)) {
3419 pte_unmap_unlock(vmf->pte, vmf->ptl);
3420 return 0;
3421 }
3422
3423 /*
3424 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3425 * much anyway since they can be in shared cache state. This misses
3426 * the case where a mapping is writable but the process never writes
3427 * to it but pte_write gets cleared during protection updates and
3428 * pte_dirty has unpredictable behaviour between PTE scan updates,
3429 * background writeback, dirty balancing and application behaviour.
3430 */
3431 if (!pte_write(pte))
3432 flags |= TNF_NO_GROUP;
3433
3434 /*
3435 * Flag if the page is shared between multiple address spaces. This
3436 * is later used when determining whether to group tasks together
3437 */
3438 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3439 flags |= TNF_SHARED;
3440
3441 last_cpupid = page_cpupid_last(page);
3442 page_nid = page_to_nid(page);
3443 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3444 &flags);
3445 pte_unmap_unlock(vmf->pte, vmf->ptl);
3446 if (target_nid == -1) {
3447 put_page(page);
3448 goto out;
3449 }
3450
3451 /* Migrate to the requested node */
3452 migrated = migrate_misplaced_page(page, vma, target_nid);
3453 if (migrated) {
3454 page_nid = target_nid;
3455 flags |= TNF_MIGRATED;
3456 } else
3457 flags |= TNF_MIGRATE_FAIL;
3458
3459 out:
3460 if (page_nid != -1)
3461 task_numa_fault(last_cpupid, page_nid, 1, flags);
3462 return 0;
3463 }
3464
3465 static int create_huge_pmd(struct vm_fault *vmf)
3466 {
3467 if (vma_is_anonymous(vmf->vma))
3468 return do_huge_pmd_anonymous_page(vmf);
3469 if (vmf->vma->vm_ops->pmd_fault)
3470 return vmf->vma->vm_ops->pmd_fault(vmf);
3471 return VM_FAULT_FALLBACK;
3472 }
3473
3474 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3475 {
3476 if (vma_is_anonymous(vmf->vma))
3477 return do_huge_pmd_wp_page(vmf, orig_pmd);
3478 if (vmf->vma->vm_ops->pmd_fault)
3479 return vmf->vma->vm_ops->pmd_fault(vmf);
3480
3481 /* COW handled on pte level: split pmd */
3482 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3483 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3484
3485 return VM_FAULT_FALLBACK;
3486 }
3487
3488 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3489 {
3490 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3491 }
3492
3493 /*
3494 * These routines also need to handle stuff like marking pages dirty
3495 * and/or accessed for architectures that don't do it in hardware (most
3496 * RISC architectures). The early dirtying is also good on the i386.
3497 *
3498 * There is also a hook called "update_mmu_cache()" that architectures
3499 * with external mmu caches can use to update those (ie the Sparc or
3500 * PowerPC hashed page tables that act as extended TLBs).
3501 *
3502 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3503 * concurrent faults).
3504 *
3505 * The mmap_sem may have been released depending on flags and our return value.
3506 * See filemap_fault() and __lock_page_or_retry().
3507 */
3508 static int handle_pte_fault(struct vm_fault *vmf)
3509 {
3510 pte_t entry;
3511
3512 if (unlikely(pmd_none(*vmf->pmd))) {
3513 /*
3514 * Leave __pte_alloc() until later: because vm_ops->fault may
3515 * want to allocate huge page, and if we expose page table
3516 * for an instant, it will be difficult to retract from
3517 * concurrent faults and from rmap lookups.
3518 */
3519 vmf->pte = NULL;
3520 } else {
3521 /* See comment in pte_alloc_one_map() */
3522 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3523 return 0;
3524 /*
3525 * A regular pmd is established and it can't morph into a huge
3526 * pmd from under us anymore at this point because we hold the
3527 * mmap_sem read mode and khugepaged takes it in write mode.
3528 * So now it's safe to run pte_offset_map().
3529 */
3530 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3531 vmf->orig_pte = *vmf->pte;
3532
3533 /*
3534 * some architectures can have larger ptes than wordsize,
3535 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3536 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3537 * atomic accesses. The code below just needs a consistent
3538 * view for the ifs and we later double check anyway with the
3539 * ptl lock held. So here a barrier will do.
3540 */
3541 barrier();
3542 if (pte_none(vmf->orig_pte)) {
3543 pte_unmap(vmf->pte);
3544 vmf->pte = NULL;
3545 }
3546 }
3547
3548 if (!vmf->pte) {
3549 if (vma_is_anonymous(vmf->vma))
3550 return do_anonymous_page(vmf);
3551 else
3552 return do_fault(vmf);
3553 }
3554
3555 if (!pte_present(vmf->orig_pte))
3556 return do_swap_page(vmf);
3557
3558 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3559 return do_numa_page(vmf);
3560
3561 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3562 spin_lock(vmf->ptl);
3563 entry = vmf->orig_pte;
3564 if (unlikely(!pte_same(*vmf->pte, entry)))
3565 goto unlock;
3566 if (vmf->flags & FAULT_FLAG_WRITE) {
3567 if (!pte_write(entry))
3568 return do_wp_page(vmf);
3569 entry = pte_mkdirty(entry);
3570 }
3571 entry = pte_mkyoung(entry);
3572 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3573 vmf->flags & FAULT_FLAG_WRITE)) {
3574 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3575 } else {
3576 /*
3577 * This is needed only for protection faults but the arch code
3578 * is not yet telling us if this is a protection fault or not.
3579 * This still avoids useless tlb flushes for .text page faults
3580 * with threads.
3581 */
3582 if (vmf->flags & FAULT_FLAG_WRITE)
3583 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3584 }
3585 unlock:
3586 pte_unmap_unlock(vmf->pte, vmf->ptl);
3587 return 0;
3588 }
3589
3590 /*
3591 * By the time we get here, we already hold the mm semaphore
3592 *
3593 * The mmap_sem may have been released depending on flags and our
3594 * return value. See filemap_fault() and __lock_page_or_retry().
3595 */
3596 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3597 unsigned int flags)
3598 {
3599 struct vm_fault vmf = {
3600 .vma = vma,
3601 .address = address & PAGE_MASK,
3602 .flags = flags,
3603 .pgoff = linear_page_index(vma, address),
3604 .gfp_mask = __get_fault_gfp_mask(vma),
3605 };
3606 struct mm_struct *mm = vma->vm_mm;
3607 pgd_t *pgd;
3608 pud_t *pud;
3609
3610 pgd = pgd_offset(mm, address);
3611 pud = pud_alloc(mm, pgd, address);
3612 if (!pud)
3613 return VM_FAULT_OOM;
3614 vmf.pmd = pmd_alloc(mm, pud, address);
3615 if (!vmf.pmd)
3616 return VM_FAULT_OOM;
3617 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3618 int ret = create_huge_pmd(&vmf);
3619 if (!(ret & VM_FAULT_FALLBACK))
3620 return ret;
3621 } else {
3622 pmd_t orig_pmd = *vmf.pmd;
3623 int ret;
3624
3625 barrier();
3626 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3627 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3628 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3629
3630 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3631 !pmd_write(orig_pmd)) {
3632 ret = wp_huge_pmd(&vmf, orig_pmd);
3633 if (!(ret & VM_FAULT_FALLBACK))
3634 return ret;
3635 } else {
3636 huge_pmd_set_accessed(&vmf, orig_pmd);
3637 return 0;
3638 }
3639 }
3640 }
3641
3642 return handle_pte_fault(&vmf);
3643 }
3644
3645 /*
3646 * By the time we get here, we already hold the mm semaphore
3647 *
3648 * The mmap_sem may have been released depending on flags and our
3649 * return value. See filemap_fault() and __lock_page_or_retry().
3650 */
3651 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3652 unsigned int flags)
3653 {
3654 int ret;
3655
3656 __set_current_state(TASK_RUNNING);
3657
3658 count_vm_event(PGFAULT);
3659 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3660
3661 /* do counter updates before entering really critical section. */
3662 check_sync_rss_stat(current);
3663
3664 /*
3665 * Enable the memcg OOM handling for faults triggered in user
3666 * space. Kernel faults are handled more gracefully.
3667 */
3668 if (flags & FAULT_FLAG_USER)
3669 mem_cgroup_oom_enable();
3670
3671 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3672 flags & FAULT_FLAG_INSTRUCTION,
3673 flags & FAULT_FLAG_REMOTE))
3674 return VM_FAULT_SIGSEGV;
3675
3676 if (unlikely(is_vm_hugetlb_page(vma)))
3677 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3678 else
3679 ret = __handle_mm_fault(vma, address, flags);
3680
3681 if (flags & FAULT_FLAG_USER) {
3682 mem_cgroup_oom_disable();
3683 /*
3684 * The task may have entered a memcg OOM situation but
3685 * if the allocation error was handled gracefully (no
3686 * VM_FAULT_OOM), there is no need to kill anything.
3687 * Just clean up the OOM state peacefully.
3688 */
3689 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3690 mem_cgroup_oom_synchronize(false);
3691 }
3692
3693 /*
3694 * This mm has been already reaped by the oom reaper and so the
3695 * refault cannot be trusted in general. Anonymous refaults would
3696 * lose data and give a zero page instead e.g. This is especially
3697 * problem for use_mm() because regular tasks will just die and
3698 * the corrupted data will not be visible anywhere while kthread
3699 * will outlive the oom victim and potentially propagate the date
3700 * further.
3701 */
3702 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3703 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3704 ret = VM_FAULT_SIGBUS;
3705
3706 return ret;
3707 }
3708 EXPORT_SYMBOL_GPL(handle_mm_fault);
3709
3710 #ifndef __PAGETABLE_PUD_FOLDED
3711 /*
3712 * Allocate page upper directory.
3713 * We've already handled the fast-path in-line.
3714 */
3715 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3716 {
3717 pud_t *new = pud_alloc_one(mm, address);
3718 if (!new)
3719 return -ENOMEM;
3720
3721 smp_wmb(); /* See comment in __pte_alloc */
3722
3723 spin_lock(&mm->page_table_lock);
3724 if (pgd_present(*pgd)) /* Another has populated it */
3725 pud_free(mm, new);
3726 else
3727 pgd_populate(mm, pgd, new);
3728 spin_unlock(&mm->page_table_lock);
3729 return 0;
3730 }
3731 #endif /* __PAGETABLE_PUD_FOLDED */
3732
3733 #ifndef __PAGETABLE_PMD_FOLDED
3734 /*
3735 * Allocate page middle directory.
3736 * We've already handled the fast-path in-line.
3737 */
3738 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3739 {
3740 pmd_t *new = pmd_alloc_one(mm, address);
3741 if (!new)
3742 return -ENOMEM;
3743
3744 smp_wmb(); /* See comment in __pte_alloc */
3745
3746 spin_lock(&mm->page_table_lock);
3747 #ifndef __ARCH_HAS_4LEVEL_HACK
3748 if (!pud_present(*pud)) {
3749 mm_inc_nr_pmds(mm);
3750 pud_populate(mm, pud, new);
3751 } else /* Another has populated it */
3752 pmd_free(mm, new);
3753 #else
3754 if (!pgd_present(*pud)) {
3755 mm_inc_nr_pmds(mm);
3756 pgd_populate(mm, pud, new);
3757 } else /* Another has populated it */
3758 pmd_free(mm, new);
3759 #endif /* __ARCH_HAS_4LEVEL_HACK */
3760 spin_unlock(&mm->page_table_lock);
3761 return 0;
3762 }
3763 #endif /* __PAGETABLE_PMD_FOLDED */
3764
3765 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3766 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3767 {
3768 pgd_t *pgd;
3769 pud_t *pud;
3770 pmd_t *pmd;
3771 pte_t *ptep;
3772
3773 pgd = pgd_offset(mm, address);
3774 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3775 goto out;
3776
3777 pud = pud_offset(pgd, address);
3778 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3779 goto out;
3780
3781 pmd = pmd_offset(pud, address);
3782 VM_BUG_ON(pmd_trans_huge(*pmd));
3783
3784 if (pmd_huge(*pmd)) {
3785 if (!pmdpp)
3786 goto out;
3787
3788 *ptlp = pmd_lock(mm, pmd);
3789 if (pmd_huge(*pmd)) {
3790 *pmdpp = pmd;
3791 return 0;
3792 }
3793 spin_unlock(*ptlp);
3794 }
3795
3796 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3797 goto out;
3798
3799 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3800 if (!ptep)
3801 goto out;
3802 if (!pte_present(*ptep))
3803 goto unlock;
3804 *ptepp = ptep;
3805 return 0;
3806 unlock:
3807 pte_unmap_unlock(ptep, *ptlp);
3808 out:
3809 return -EINVAL;
3810 }
3811
3812 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3813 pte_t **ptepp, spinlock_t **ptlp)
3814 {
3815 int res;
3816
3817 /* (void) is needed to make gcc happy */
3818 (void) __cond_lock(*ptlp,
3819 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
3820 ptlp)));
3821 return res;
3822 }
3823
3824 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3825 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3826 {
3827 int res;
3828
3829 /* (void) is needed to make gcc happy */
3830 (void) __cond_lock(*ptlp,
3831 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
3832 ptlp)));
3833 return res;
3834 }
3835 EXPORT_SYMBOL(follow_pte_pmd);
3836
3837 /**
3838 * follow_pfn - look up PFN at a user virtual address
3839 * @vma: memory mapping
3840 * @address: user virtual address
3841 * @pfn: location to store found PFN
3842 *
3843 * Only IO mappings and raw PFN mappings are allowed.
3844 *
3845 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3846 */
3847 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3848 unsigned long *pfn)
3849 {
3850 int ret = -EINVAL;
3851 spinlock_t *ptl;
3852 pte_t *ptep;
3853
3854 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3855 return ret;
3856
3857 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3858 if (ret)
3859 return ret;
3860 *pfn = pte_pfn(*ptep);
3861 pte_unmap_unlock(ptep, ptl);
3862 return 0;
3863 }
3864 EXPORT_SYMBOL(follow_pfn);
3865
3866 #ifdef CONFIG_HAVE_IOREMAP_PROT
3867 int follow_phys(struct vm_area_struct *vma,
3868 unsigned long address, unsigned int flags,
3869 unsigned long *prot, resource_size_t *phys)
3870 {
3871 int ret = -EINVAL;
3872 pte_t *ptep, pte;
3873 spinlock_t *ptl;
3874
3875 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3876 goto out;
3877
3878 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3879 goto out;
3880 pte = *ptep;
3881
3882 if ((flags & FOLL_WRITE) && !pte_write(pte))
3883 goto unlock;
3884
3885 *prot = pgprot_val(pte_pgprot(pte));
3886 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3887
3888 ret = 0;
3889 unlock:
3890 pte_unmap_unlock(ptep, ptl);
3891 out:
3892 return ret;
3893 }
3894
3895 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3896 void *buf, int len, int write)
3897 {
3898 resource_size_t phys_addr;
3899 unsigned long prot = 0;
3900 void __iomem *maddr;
3901 int offset = addr & (PAGE_SIZE-1);
3902
3903 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3904 return -EINVAL;
3905
3906 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3907 if (write)
3908 memcpy_toio(maddr + offset, buf, len);
3909 else
3910 memcpy_fromio(buf, maddr + offset, len);
3911 iounmap(maddr);
3912
3913 return len;
3914 }
3915 EXPORT_SYMBOL_GPL(generic_access_phys);
3916 #endif
3917
3918 /*
3919 * Access another process' address space as given in mm. If non-NULL, use the
3920 * given task for page fault accounting.
3921 */
3922 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3923 unsigned long addr, void *buf, int len, unsigned int gup_flags)
3924 {
3925 struct vm_area_struct *vma;
3926 void *old_buf = buf;
3927 int write = gup_flags & FOLL_WRITE;
3928
3929 down_read(&mm->mmap_sem);
3930 /* ignore errors, just check how much was successfully transferred */
3931 while (len) {
3932 int bytes, ret, offset;
3933 void *maddr;
3934 struct page *page = NULL;
3935
3936 ret = get_user_pages_remote(tsk, mm, addr, 1,
3937 gup_flags, &page, &vma, NULL);
3938 if (ret <= 0) {
3939 #ifndef CONFIG_HAVE_IOREMAP_PROT
3940 break;
3941 #else
3942 /*
3943 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3944 * we can access using slightly different code.
3945 */
3946 vma = find_vma(mm, addr);
3947 if (!vma || vma->vm_start > addr)
3948 break;
3949 if (vma->vm_ops && vma->vm_ops->access)
3950 ret = vma->vm_ops->access(vma, addr, buf,
3951 len, write);
3952 if (ret <= 0)
3953 break;
3954 bytes = ret;
3955 #endif
3956 } else {
3957 bytes = len;
3958 offset = addr & (PAGE_SIZE-1);
3959 if (bytes > PAGE_SIZE-offset)
3960 bytes = PAGE_SIZE-offset;
3961
3962 maddr = kmap(page);
3963 if (write) {
3964 copy_to_user_page(vma, page, addr,
3965 maddr + offset, buf, bytes);
3966 set_page_dirty_lock(page);
3967 } else {
3968 copy_from_user_page(vma, page, addr,
3969 buf, maddr + offset, bytes);
3970 }
3971 kunmap(page);
3972 put_page(page);
3973 }
3974 len -= bytes;
3975 buf += bytes;
3976 addr += bytes;
3977 }
3978 up_read(&mm->mmap_sem);
3979
3980 return buf - old_buf;
3981 }
3982
3983 /**
3984 * access_remote_vm - access another process' address space
3985 * @mm: the mm_struct of the target address space
3986 * @addr: start address to access
3987 * @buf: source or destination buffer
3988 * @len: number of bytes to transfer
3989 * @gup_flags: flags modifying lookup behaviour
3990 *
3991 * The caller must hold a reference on @mm.
3992 */
3993 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3994 void *buf, int len, unsigned int gup_flags)
3995 {
3996 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
3997 }
3998
3999 /*
4000 * Access another process' address space.
4001 * Source/target buffer must be kernel space,
4002 * Do not walk the page table directly, use get_user_pages
4003 */
4004 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4005 void *buf, int len, unsigned int gup_flags)
4006 {
4007 struct mm_struct *mm;
4008 int ret;
4009
4010 mm = get_task_mm(tsk);
4011 if (!mm)
4012 return 0;
4013
4014 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4015
4016 mmput(mm);
4017
4018 return ret;
4019 }
4020 EXPORT_SYMBOL_GPL(access_process_vm);
4021
4022 /*
4023 * Print the name of a VMA.
4024 */
4025 void print_vma_addr(char *prefix, unsigned long ip)
4026 {
4027 struct mm_struct *mm = current->mm;
4028 struct vm_area_struct *vma;
4029
4030 /*
4031 * Do not print if we are in atomic
4032 * contexts (in exception stacks, etc.):
4033 */
4034 if (preempt_count())
4035 return;
4036
4037 down_read(&mm->mmap_sem);
4038 vma = find_vma(mm, ip);
4039 if (vma && vma->vm_file) {
4040 struct file *f = vma->vm_file;
4041 char *buf = (char *)__get_free_page(GFP_KERNEL);
4042 if (buf) {
4043 char *p;
4044
4045 p = file_path(f, buf, PAGE_SIZE);
4046 if (IS_ERR(p))
4047 p = "?";
4048 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4049 vma->vm_start,
4050 vma->vm_end - vma->vm_start);
4051 free_page((unsigned long)buf);
4052 }
4053 }
4054 up_read(&mm->mmap_sem);
4055 }
4056
4057 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4058 void __might_fault(const char *file, int line)
4059 {
4060 /*
4061 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4062 * holding the mmap_sem, this is safe because kernel memory doesn't
4063 * get paged out, therefore we'll never actually fault, and the
4064 * below annotations will generate false positives.
4065 */
4066 if (segment_eq(get_fs(), KERNEL_DS))
4067 return;
4068 if (pagefault_disabled())
4069 return;
4070 __might_sleep(file, line, 0);
4071 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4072 if (current->mm)
4073 might_lock_read(&current->mm->mmap_sem);
4074 #endif
4075 }
4076 EXPORT_SYMBOL(__might_fault);
4077 #endif
4078
4079 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4080 static void clear_gigantic_page(struct page *page,
4081 unsigned long addr,
4082 unsigned int pages_per_huge_page)
4083 {
4084 int i;
4085 struct page *p = page;
4086
4087 might_sleep();
4088 for (i = 0; i < pages_per_huge_page;
4089 i++, p = mem_map_next(p, page, i)) {
4090 cond_resched();
4091 clear_user_highpage(p, addr + i * PAGE_SIZE);
4092 }
4093 }
4094 void clear_huge_page(struct page *page,
4095 unsigned long addr, unsigned int pages_per_huge_page)
4096 {
4097 int i;
4098
4099 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4100 clear_gigantic_page(page, addr, pages_per_huge_page);
4101 return;
4102 }
4103
4104 might_sleep();
4105 for (i = 0; i < pages_per_huge_page; i++) {
4106 cond_resched();
4107 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4108 }
4109 }
4110
4111 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4112 unsigned long addr,
4113 struct vm_area_struct *vma,
4114 unsigned int pages_per_huge_page)
4115 {
4116 int i;
4117 struct page *dst_base = dst;
4118 struct page *src_base = src;
4119
4120 for (i = 0; i < pages_per_huge_page; ) {
4121 cond_resched();
4122 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4123
4124 i++;
4125 dst = mem_map_next(dst, dst_base, i);
4126 src = mem_map_next(src, src_base, i);
4127 }
4128 }
4129
4130 void copy_user_huge_page(struct page *dst, struct page *src,
4131 unsigned long addr, struct vm_area_struct *vma,
4132 unsigned int pages_per_huge_page)
4133 {
4134 int i;
4135
4136 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4137 copy_user_gigantic_page(dst, src, addr, vma,
4138 pages_per_huge_page);
4139 return;
4140 }
4141
4142 might_sleep();
4143 for (i = 0; i < pages_per_huge_page; i++) {
4144 cond_resched();
4145 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4146 }
4147 }
4148
4149 long copy_huge_page_from_user(struct page *dst_page,
4150 const void __user *usr_src,
4151 unsigned int pages_per_huge_page,
4152 bool allow_pagefault)
4153 {
4154 void *src = (void *)usr_src;
4155 void *page_kaddr;
4156 unsigned long i, rc = 0;
4157 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4158
4159 for (i = 0; i < pages_per_huge_page; i++) {
4160 if (allow_pagefault)
4161 page_kaddr = kmap(dst_page + i);
4162 else
4163 page_kaddr = kmap_atomic(dst_page + i);
4164 rc = copy_from_user(page_kaddr,
4165 (const void __user *)(src + i * PAGE_SIZE),
4166 PAGE_SIZE);
4167 if (allow_pagefault)
4168 kunmap(dst_page + i);
4169 else
4170 kunmap_atomic(page_kaddr);
4171
4172 ret_val -= (PAGE_SIZE - rc);
4173 if (rc)
4174 break;
4175
4176 cond_resched();
4177 }
4178 return ret_val;
4179 }
4180 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4181
4182 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4183
4184 static struct kmem_cache *page_ptl_cachep;
4185
4186 void __init ptlock_cache_init(void)
4187 {
4188 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4189 SLAB_PANIC, NULL);
4190 }
4191
4192 bool ptlock_alloc(struct page *page)
4193 {
4194 spinlock_t *ptl;
4195
4196 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4197 if (!ptl)
4198 return false;
4199 page->ptl = ptl;
4200 return true;
4201 }
4202
4203 void ptlock_free(struct page *page)
4204 {
4205 kmem_cache_free(page_ptl_cachep, page->ptl);
4206 }
4207 #endif