Loading...
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
67#include <asm/io.h>
68#include <asm/mmu_context.h>
69#include <asm/pgalloc.h>
70#include <asm/uaccess.h>
71#include <asm/tlb.h>
72#include <asm/tlbflush.h>
73#include <asm/pgtable.h>
74
75#include "internal.h"
76
77#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
78#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
79#endif
80
81#ifndef CONFIG_NEED_MULTIPLE_NODES
82/* use the per-pgdat data instead for discontigmem - mbligh */
83unsigned long max_mapnr;
84struct page *mem_map;
85
86EXPORT_SYMBOL(max_mapnr);
87EXPORT_SYMBOL(mem_map);
88#endif
89
90/*
91 * A number of key systems in x86 including ioremap() rely on the assumption
92 * that high_memory defines the upper bound on direct map memory, then end
93 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
94 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
95 * and ZONE_HIGHMEM.
96 */
97void * high_memory;
98
99EXPORT_SYMBOL(high_memory);
100
101/*
102 * Randomize the address space (stacks, mmaps, brk, etc.).
103 *
104 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
105 * as ancient (libc5 based) binaries can segfault. )
106 */
107int randomize_va_space __read_mostly =
108#ifdef CONFIG_COMPAT_BRK
109 1;
110#else
111 2;
112#endif
113
114static int __init disable_randmaps(char *s)
115{
116 randomize_va_space = 0;
117 return 1;
118}
119__setup("norandmaps", disable_randmaps);
120
121unsigned long zero_pfn __read_mostly;
122unsigned long highest_memmap_pfn __read_mostly;
123
124EXPORT_SYMBOL(zero_pfn);
125
126/*
127 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
128 */
129static int __init init_zero_pfn(void)
130{
131 zero_pfn = page_to_pfn(ZERO_PAGE(0));
132 return 0;
133}
134core_initcall(init_zero_pfn);
135
136
137#if defined(SPLIT_RSS_COUNTING)
138
139void sync_mm_rss(struct mm_struct *mm)
140{
141 int i;
142
143 for (i = 0; i < NR_MM_COUNTERS; i++) {
144 if (current->rss_stat.count[i]) {
145 add_mm_counter(mm, i, current->rss_stat.count[i]);
146 current->rss_stat.count[i] = 0;
147 }
148 }
149 current->rss_stat.events = 0;
150}
151
152static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
153{
154 struct task_struct *task = current;
155
156 if (likely(task->mm == mm))
157 task->rss_stat.count[member] += val;
158 else
159 add_mm_counter(mm, member, val);
160}
161#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
162#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
163
164/* sync counter once per 64 page faults */
165#define TASK_RSS_EVENTS_THRESH (64)
166static void check_sync_rss_stat(struct task_struct *task)
167{
168 if (unlikely(task != current))
169 return;
170 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
171 sync_mm_rss(task->mm);
172}
173#else /* SPLIT_RSS_COUNTING */
174
175#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
176#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
177
178static void check_sync_rss_stat(struct task_struct *task)
179{
180}
181
182#endif /* SPLIT_RSS_COUNTING */
183
184#ifdef HAVE_GENERIC_MMU_GATHER
185
186static bool tlb_next_batch(struct mmu_gather *tlb)
187{
188 struct mmu_gather_batch *batch;
189
190 batch = tlb->active;
191 if (batch->next) {
192 tlb->active = batch->next;
193 return true;
194 }
195
196 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
197 return false;
198
199 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
200 if (!batch)
201 return false;
202
203 tlb->batch_count++;
204 batch->next = NULL;
205 batch->nr = 0;
206 batch->max = MAX_GATHER_BATCH;
207
208 tlb->active->next = batch;
209 tlb->active = batch;
210
211 return true;
212}
213
214/* tlb_gather_mmu
215 * Called to initialize an (on-stack) mmu_gather structure for page-table
216 * tear-down from @mm. The @fullmm argument is used when @mm is without
217 * users and we're going to destroy the full address space (exit/execve).
218 */
219void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
220{
221 tlb->mm = mm;
222
223 /* Is it from 0 to ~0? */
224 tlb->fullmm = !(start | (end+1));
225 tlb->need_flush_all = 0;
226 tlb->local.next = NULL;
227 tlb->local.nr = 0;
228 tlb->local.max = ARRAY_SIZE(tlb->__pages);
229 tlb->active = &tlb->local;
230 tlb->batch_count = 0;
231
232#ifdef CONFIG_HAVE_RCU_TABLE_FREE
233 tlb->batch = NULL;
234#endif
235
236 __tlb_reset_range(tlb);
237}
238
239static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
240{
241 if (!tlb->end)
242 return;
243
244 tlb_flush(tlb);
245 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
246#ifdef CONFIG_HAVE_RCU_TABLE_FREE
247 tlb_table_flush(tlb);
248#endif
249 __tlb_reset_range(tlb);
250}
251
252static void tlb_flush_mmu_free(struct mmu_gather *tlb)
253{
254 struct mmu_gather_batch *batch;
255
256 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
257 free_pages_and_swap_cache(batch->pages, batch->nr);
258 batch->nr = 0;
259 }
260 tlb->active = &tlb->local;
261}
262
263void tlb_flush_mmu(struct mmu_gather *tlb)
264{
265 tlb_flush_mmu_tlbonly(tlb);
266 tlb_flush_mmu_free(tlb);
267}
268
269/* tlb_finish_mmu
270 * Called at the end of the shootdown operation to free up any resources
271 * that were required.
272 */
273void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
274{
275 struct mmu_gather_batch *batch, *next;
276
277 tlb_flush_mmu(tlb);
278
279 /* keep the page table cache within bounds */
280 check_pgt_cache();
281
282 for (batch = tlb->local.next; batch; batch = next) {
283 next = batch->next;
284 free_pages((unsigned long)batch, 0);
285 }
286 tlb->local.next = NULL;
287}
288
289/* __tlb_remove_page
290 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
291 * handling the additional races in SMP caused by other CPUs caching valid
292 * mappings in their TLBs. Returns the number of free page slots left.
293 * When out of page slots we must call tlb_flush_mmu().
294 */
295int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
296{
297 struct mmu_gather_batch *batch;
298
299 VM_BUG_ON(!tlb->end);
300
301 batch = tlb->active;
302 batch->pages[batch->nr++] = page;
303 if (batch->nr == batch->max) {
304 if (!tlb_next_batch(tlb))
305 return 0;
306 batch = tlb->active;
307 }
308 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
309
310 return batch->max - batch->nr;
311}
312
313#endif /* HAVE_GENERIC_MMU_GATHER */
314
315#ifdef CONFIG_HAVE_RCU_TABLE_FREE
316
317/*
318 * See the comment near struct mmu_table_batch.
319 */
320
321static void tlb_remove_table_smp_sync(void *arg)
322{
323 /* Simply deliver the interrupt */
324}
325
326static void tlb_remove_table_one(void *table)
327{
328 /*
329 * This isn't an RCU grace period and hence the page-tables cannot be
330 * assumed to be actually RCU-freed.
331 *
332 * It is however sufficient for software page-table walkers that rely on
333 * IRQ disabling. See the comment near struct mmu_table_batch.
334 */
335 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
336 __tlb_remove_table(table);
337}
338
339static void tlb_remove_table_rcu(struct rcu_head *head)
340{
341 struct mmu_table_batch *batch;
342 int i;
343
344 batch = container_of(head, struct mmu_table_batch, rcu);
345
346 for (i = 0; i < batch->nr; i++)
347 __tlb_remove_table(batch->tables[i]);
348
349 free_page((unsigned long)batch);
350}
351
352void tlb_table_flush(struct mmu_gather *tlb)
353{
354 struct mmu_table_batch **batch = &tlb->batch;
355
356 if (*batch) {
357 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
358 *batch = NULL;
359 }
360}
361
362void tlb_remove_table(struct mmu_gather *tlb, void *table)
363{
364 struct mmu_table_batch **batch = &tlb->batch;
365
366 /*
367 * When there's less then two users of this mm there cannot be a
368 * concurrent page-table walk.
369 */
370 if (atomic_read(&tlb->mm->mm_users) < 2) {
371 __tlb_remove_table(table);
372 return;
373 }
374
375 if (*batch == NULL) {
376 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
377 if (*batch == NULL) {
378 tlb_remove_table_one(table);
379 return;
380 }
381 (*batch)->nr = 0;
382 }
383 (*batch)->tables[(*batch)->nr++] = table;
384 if ((*batch)->nr == MAX_TABLE_BATCH)
385 tlb_table_flush(tlb);
386}
387
388#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
389
390/*
391 * Note: this doesn't free the actual pages themselves. That
392 * has been handled earlier when unmapping all the memory regions.
393 */
394static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
395 unsigned long addr)
396{
397 pgtable_t token = pmd_pgtable(*pmd);
398 pmd_clear(pmd);
399 pte_free_tlb(tlb, token, addr);
400 atomic_long_dec(&tlb->mm->nr_ptes);
401}
402
403static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
404 unsigned long addr, unsigned long end,
405 unsigned long floor, unsigned long ceiling)
406{
407 pmd_t *pmd;
408 unsigned long next;
409 unsigned long start;
410
411 start = addr;
412 pmd = pmd_offset(pud, addr);
413 do {
414 next = pmd_addr_end(addr, end);
415 if (pmd_none_or_clear_bad(pmd))
416 continue;
417 free_pte_range(tlb, pmd, addr);
418 } while (pmd++, addr = next, addr != end);
419
420 start &= PUD_MASK;
421 if (start < floor)
422 return;
423 if (ceiling) {
424 ceiling &= PUD_MASK;
425 if (!ceiling)
426 return;
427 }
428 if (end - 1 > ceiling - 1)
429 return;
430
431 pmd = pmd_offset(pud, start);
432 pud_clear(pud);
433 pmd_free_tlb(tlb, pmd, start);
434 mm_dec_nr_pmds(tlb->mm);
435}
436
437static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
438 unsigned long addr, unsigned long end,
439 unsigned long floor, unsigned long ceiling)
440{
441 pud_t *pud;
442 unsigned long next;
443 unsigned long start;
444
445 start = addr;
446 pud = pud_offset(pgd, addr);
447 do {
448 next = pud_addr_end(addr, end);
449 if (pud_none_or_clear_bad(pud))
450 continue;
451 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
452 } while (pud++, addr = next, addr != end);
453
454 start &= PGDIR_MASK;
455 if (start < floor)
456 return;
457 if (ceiling) {
458 ceiling &= PGDIR_MASK;
459 if (!ceiling)
460 return;
461 }
462 if (end - 1 > ceiling - 1)
463 return;
464
465 pud = pud_offset(pgd, start);
466 pgd_clear(pgd);
467 pud_free_tlb(tlb, pud, start);
468}
469
470/*
471 * This function frees user-level page tables of a process.
472 */
473void free_pgd_range(struct mmu_gather *tlb,
474 unsigned long addr, unsigned long end,
475 unsigned long floor, unsigned long ceiling)
476{
477 pgd_t *pgd;
478 unsigned long next;
479
480 /*
481 * The next few lines have given us lots of grief...
482 *
483 * Why are we testing PMD* at this top level? Because often
484 * there will be no work to do at all, and we'd prefer not to
485 * go all the way down to the bottom just to discover that.
486 *
487 * Why all these "- 1"s? Because 0 represents both the bottom
488 * of the address space and the top of it (using -1 for the
489 * top wouldn't help much: the masks would do the wrong thing).
490 * The rule is that addr 0 and floor 0 refer to the bottom of
491 * the address space, but end 0 and ceiling 0 refer to the top
492 * Comparisons need to use "end - 1" and "ceiling - 1" (though
493 * that end 0 case should be mythical).
494 *
495 * Wherever addr is brought up or ceiling brought down, we must
496 * be careful to reject "the opposite 0" before it confuses the
497 * subsequent tests. But what about where end is brought down
498 * by PMD_SIZE below? no, end can't go down to 0 there.
499 *
500 * Whereas we round start (addr) and ceiling down, by different
501 * masks at different levels, in order to test whether a table
502 * now has no other vmas using it, so can be freed, we don't
503 * bother to round floor or end up - the tests don't need that.
504 */
505
506 addr &= PMD_MASK;
507 if (addr < floor) {
508 addr += PMD_SIZE;
509 if (!addr)
510 return;
511 }
512 if (ceiling) {
513 ceiling &= PMD_MASK;
514 if (!ceiling)
515 return;
516 }
517 if (end - 1 > ceiling - 1)
518 end -= PMD_SIZE;
519 if (addr > end - 1)
520 return;
521
522 pgd = pgd_offset(tlb->mm, addr);
523 do {
524 next = pgd_addr_end(addr, end);
525 if (pgd_none_or_clear_bad(pgd))
526 continue;
527 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
528 } while (pgd++, addr = next, addr != end);
529}
530
531void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
532 unsigned long floor, unsigned long ceiling)
533{
534 while (vma) {
535 struct vm_area_struct *next = vma->vm_next;
536 unsigned long addr = vma->vm_start;
537
538 /*
539 * Hide vma from rmap and truncate_pagecache before freeing
540 * pgtables
541 */
542 unlink_anon_vmas(vma);
543 unlink_file_vma(vma);
544
545 if (is_vm_hugetlb_page(vma)) {
546 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
547 floor, next? next->vm_start: ceiling);
548 } else {
549 /*
550 * Optimization: gather nearby vmas into one call down
551 */
552 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
553 && !is_vm_hugetlb_page(next)) {
554 vma = next;
555 next = vma->vm_next;
556 unlink_anon_vmas(vma);
557 unlink_file_vma(vma);
558 }
559 free_pgd_range(tlb, addr, vma->vm_end,
560 floor, next? next->vm_start: ceiling);
561 }
562 vma = next;
563 }
564}
565
566int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
567{
568 spinlock_t *ptl;
569 pgtable_t new = pte_alloc_one(mm, address);
570 if (!new)
571 return -ENOMEM;
572
573 /*
574 * Ensure all pte setup (eg. pte page lock and page clearing) are
575 * visible before the pte is made visible to other CPUs by being
576 * put into page tables.
577 *
578 * The other side of the story is the pointer chasing in the page
579 * table walking code (when walking the page table without locking;
580 * ie. most of the time). Fortunately, these data accesses consist
581 * of a chain of data-dependent loads, meaning most CPUs (alpha
582 * being the notable exception) will already guarantee loads are
583 * seen in-order. See the alpha page table accessors for the
584 * smp_read_barrier_depends() barriers in page table walking code.
585 */
586 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
587
588 ptl = pmd_lock(mm, pmd);
589 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
590 atomic_long_inc(&mm->nr_ptes);
591 pmd_populate(mm, pmd, new);
592 new = NULL;
593 }
594 spin_unlock(ptl);
595 if (new)
596 pte_free(mm, new);
597 return 0;
598}
599
600int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
601{
602 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
603 if (!new)
604 return -ENOMEM;
605
606 smp_wmb(); /* See comment in __pte_alloc */
607
608 spin_lock(&init_mm.page_table_lock);
609 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
610 pmd_populate_kernel(&init_mm, pmd, new);
611 new = NULL;
612 }
613 spin_unlock(&init_mm.page_table_lock);
614 if (new)
615 pte_free_kernel(&init_mm, new);
616 return 0;
617}
618
619static inline void init_rss_vec(int *rss)
620{
621 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
622}
623
624static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
625{
626 int i;
627
628 if (current->mm == mm)
629 sync_mm_rss(mm);
630 for (i = 0; i < NR_MM_COUNTERS; i++)
631 if (rss[i])
632 add_mm_counter(mm, i, rss[i]);
633}
634
635/*
636 * This function is called to print an error when a bad pte
637 * is found. For example, we might have a PFN-mapped pte in
638 * a region that doesn't allow it.
639 *
640 * The calling function must still handle the error.
641 */
642static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
643 pte_t pte, struct page *page)
644{
645 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
646 pud_t *pud = pud_offset(pgd, addr);
647 pmd_t *pmd = pmd_offset(pud, addr);
648 struct address_space *mapping;
649 pgoff_t index;
650 static unsigned long resume;
651 static unsigned long nr_shown;
652 static unsigned long nr_unshown;
653
654 /*
655 * Allow a burst of 60 reports, then keep quiet for that minute;
656 * or allow a steady drip of one report per second.
657 */
658 if (nr_shown == 60) {
659 if (time_before(jiffies, resume)) {
660 nr_unshown++;
661 return;
662 }
663 if (nr_unshown) {
664 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
665 nr_unshown);
666 nr_unshown = 0;
667 }
668 nr_shown = 0;
669 }
670 if (nr_shown++ == 0)
671 resume = jiffies + 60 * HZ;
672
673 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
674 index = linear_page_index(vma, addr);
675
676 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
677 current->comm,
678 (long long)pte_val(pte), (long long)pmd_val(*pmd));
679 if (page)
680 dump_page(page, "bad pte");
681 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
682 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
683 /*
684 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
685 */
686 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
687 vma->vm_file,
688 vma->vm_ops ? vma->vm_ops->fault : NULL,
689 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
690 mapping ? mapping->a_ops->readpage : NULL);
691 dump_stack();
692 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
693}
694
695/*
696 * vm_normal_page -- This function gets the "struct page" associated with a pte.
697 *
698 * "Special" mappings do not wish to be associated with a "struct page" (either
699 * it doesn't exist, or it exists but they don't want to touch it). In this
700 * case, NULL is returned here. "Normal" mappings do have a struct page.
701 *
702 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
703 * pte bit, in which case this function is trivial. Secondly, an architecture
704 * may not have a spare pte bit, which requires a more complicated scheme,
705 * described below.
706 *
707 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
708 * special mapping (even if there are underlying and valid "struct pages").
709 * COWed pages of a VM_PFNMAP are always normal.
710 *
711 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
712 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
713 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
714 * mapping will always honor the rule
715 *
716 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
717 *
718 * And for normal mappings this is false.
719 *
720 * This restricts such mappings to be a linear translation from virtual address
721 * to pfn. To get around this restriction, we allow arbitrary mappings so long
722 * as the vma is not a COW mapping; in that case, we know that all ptes are
723 * special (because none can have been COWed).
724 *
725 *
726 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
727 *
728 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
729 * page" backing, however the difference is that _all_ pages with a struct
730 * page (that is, those where pfn_valid is true) are refcounted and considered
731 * normal pages by the VM. The disadvantage is that pages are refcounted
732 * (which can be slower and simply not an option for some PFNMAP users). The
733 * advantage is that we don't have to follow the strict linearity rule of
734 * PFNMAP mappings in order to support COWable mappings.
735 *
736 */
737#ifdef __HAVE_ARCH_PTE_SPECIAL
738# define HAVE_PTE_SPECIAL 1
739#else
740# define HAVE_PTE_SPECIAL 0
741#endif
742struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
743 pte_t pte)
744{
745 unsigned long pfn = pte_pfn(pte);
746
747 if (HAVE_PTE_SPECIAL) {
748 if (likely(!pte_special(pte)))
749 goto check_pfn;
750 if (vma->vm_ops && vma->vm_ops->find_special_page)
751 return vma->vm_ops->find_special_page(vma, addr);
752 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
753 return NULL;
754 if (!is_zero_pfn(pfn))
755 print_bad_pte(vma, addr, pte, NULL);
756 return NULL;
757 }
758
759 /* !HAVE_PTE_SPECIAL case follows: */
760
761 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
762 if (vma->vm_flags & VM_MIXEDMAP) {
763 if (!pfn_valid(pfn))
764 return NULL;
765 goto out;
766 } else {
767 unsigned long off;
768 off = (addr - vma->vm_start) >> PAGE_SHIFT;
769 if (pfn == vma->vm_pgoff + off)
770 return NULL;
771 if (!is_cow_mapping(vma->vm_flags))
772 return NULL;
773 }
774 }
775
776 if (is_zero_pfn(pfn))
777 return NULL;
778check_pfn:
779 if (unlikely(pfn > highest_memmap_pfn)) {
780 print_bad_pte(vma, addr, pte, NULL);
781 return NULL;
782 }
783
784 /*
785 * NOTE! We still have PageReserved() pages in the page tables.
786 * eg. VDSO mappings can cause them to exist.
787 */
788out:
789 return pfn_to_page(pfn);
790}
791
792#ifdef CONFIG_TRANSPARENT_HUGEPAGE
793struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
794 pmd_t pmd)
795{
796 unsigned long pfn = pmd_pfn(pmd);
797
798 /*
799 * There is no pmd_special() but there may be special pmds, e.g.
800 * in a direct-access (dax) mapping, so let's just replicate the
801 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
802 */
803 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
804 if (vma->vm_flags & VM_MIXEDMAP) {
805 if (!pfn_valid(pfn))
806 return NULL;
807 goto out;
808 } else {
809 unsigned long off;
810 off = (addr - vma->vm_start) >> PAGE_SHIFT;
811 if (pfn == vma->vm_pgoff + off)
812 return NULL;
813 if (!is_cow_mapping(vma->vm_flags))
814 return NULL;
815 }
816 }
817
818 if (is_zero_pfn(pfn))
819 return NULL;
820 if (unlikely(pfn > highest_memmap_pfn))
821 return NULL;
822
823 /*
824 * NOTE! We still have PageReserved() pages in the page tables.
825 * eg. VDSO mappings can cause them to exist.
826 */
827out:
828 return pfn_to_page(pfn);
829}
830#endif
831
832/*
833 * copy one vm_area from one task to the other. Assumes the page tables
834 * already present in the new task to be cleared in the whole range
835 * covered by this vma.
836 */
837
838static inline unsigned long
839copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
840 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
841 unsigned long addr, int *rss)
842{
843 unsigned long vm_flags = vma->vm_flags;
844 pte_t pte = *src_pte;
845 struct page *page;
846
847 /* pte contains position in swap or file, so copy. */
848 if (unlikely(!pte_present(pte))) {
849 swp_entry_t entry = pte_to_swp_entry(pte);
850
851 if (likely(!non_swap_entry(entry))) {
852 if (swap_duplicate(entry) < 0)
853 return entry.val;
854
855 /* make sure dst_mm is on swapoff's mmlist. */
856 if (unlikely(list_empty(&dst_mm->mmlist))) {
857 spin_lock(&mmlist_lock);
858 if (list_empty(&dst_mm->mmlist))
859 list_add(&dst_mm->mmlist,
860 &src_mm->mmlist);
861 spin_unlock(&mmlist_lock);
862 }
863 rss[MM_SWAPENTS]++;
864 } else if (is_migration_entry(entry)) {
865 page = migration_entry_to_page(entry);
866
867 rss[mm_counter(page)]++;
868
869 if (is_write_migration_entry(entry) &&
870 is_cow_mapping(vm_flags)) {
871 /*
872 * COW mappings require pages in both
873 * parent and child to be set to read.
874 */
875 make_migration_entry_read(&entry);
876 pte = swp_entry_to_pte(entry);
877 if (pte_swp_soft_dirty(*src_pte))
878 pte = pte_swp_mksoft_dirty(pte);
879 set_pte_at(src_mm, addr, src_pte, pte);
880 }
881 }
882 goto out_set_pte;
883 }
884
885 /*
886 * If it's a COW mapping, write protect it both
887 * in the parent and the child
888 */
889 if (is_cow_mapping(vm_flags)) {
890 ptep_set_wrprotect(src_mm, addr, src_pte);
891 pte = pte_wrprotect(pte);
892 }
893
894 /*
895 * If it's a shared mapping, mark it clean in
896 * the child
897 */
898 if (vm_flags & VM_SHARED)
899 pte = pte_mkclean(pte);
900 pte = pte_mkold(pte);
901
902 page = vm_normal_page(vma, addr, pte);
903 if (page) {
904 get_page(page);
905 page_dup_rmap(page, false);
906 rss[mm_counter(page)]++;
907 }
908
909out_set_pte:
910 set_pte_at(dst_mm, addr, dst_pte, pte);
911 return 0;
912}
913
914static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
916 unsigned long addr, unsigned long end)
917{
918 pte_t *orig_src_pte, *orig_dst_pte;
919 pte_t *src_pte, *dst_pte;
920 spinlock_t *src_ptl, *dst_ptl;
921 int progress = 0;
922 int rss[NR_MM_COUNTERS];
923 swp_entry_t entry = (swp_entry_t){0};
924
925again:
926 init_rss_vec(rss);
927
928 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
929 if (!dst_pte)
930 return -ENOMEM;
931 src_pte = pte_offset_map(src_pmd, addr);
932 src_ptl = pte_lockptr(src_mm, src_pmd);
933 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
934 orig_src_pte = src_pte;
935 orig_dst_pte = dst_pte;
936 arch_enter_lazy_mmu_mode();
937
938 do {
939 /*
940 * We are holding two locks at this point - either of them
941 * could generate latencies in another task on another CPU.
942 */
943 if (progress >= 32) {
944 progress = 0;
945 if (need_resched() ||
946 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
947 break;
948 }
949 if (pte_none(*src_pte)) {
950 progress++;
951 continue;
952 }
953 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
954 vma, addr, rss);
955 if (entry.val)
956 break;
957 progress += 8;
958 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
959
960 arch_leave_lazy_mmu_mode();
961 spin_unlock(src_ptl);
962 pte_unmap(orig_src_pte);
963 add_mm_rss_vec(dst_mm, rss);
964 pte_unmap_unlock(orig_dst_pte, dst_ptl);
965 cond_resched();
966
967 if (entry.val) {
968 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
969 return -ENOMEM;
970 progress = 0;
971 }
972 if (addr != end)
973 goto again;
974 return 0;
975}
976
977static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
978 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
979 unsigned long addr, unsigned long end)
980{
981 pmd_t *src_pmd, *dst_pmd;
982 unsigned long next;
983
984 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
985 if (!dst_pmd)
986 return -ENOMEM;
987 src_pmd = pmd_offset(src_pud, addr);
988 do {
989 next = pmd_addr_end(addr, end);
990 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
991 int err;
992 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
993 err = copy_huge_pmd(dst_mm, src_mm,
994 dst_pmd, src_pmd, addr, vma);
995 if (err == -ENOMEM)
996 return -ENOMEM;
997 if (!err)
998 continue;
999 /* fall through */
1000 }
1001 if (pmd_none_or_clear_bad(src_pmd))
1002 continue;
1003 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1004 vma, addr, next))
1005 return -ENOMEM;
1006 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1007 return 0;
1008}
1009
1010static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1011 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1012 unsigned long addr, unsigned long end)
1013{
1014 pud_t *src_pud, *dst_pud;
1015 unsigned long next;
1016
1017 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1018 if (!dst_pud)
1019 return -ENOMEM;
1020 src_pud = pud_offset(src_pgd, addr);
1021 do {
1022 next = pud_addr_end(addr, end);
1023 if (pud_none_or_clear_bad(src_pud))
1024 continue;
1025 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1026 vma, addr, next))
1027 return -ENOMEM;
1028 } while (dst_pud++, src_pud++, addr = next, addr != end);
1029 return 0;
1030}
1031
1032int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1033 struct vm_area_struct *vma)
1034{
1035 pgd_t *src_pgd, *dst_pgd;
1036 unsigned long next;
1037 unsigned long addr = vma->vm_start;
1038 unsigned long end = vma->vm_end;
1039 unsigned long mmun_start; /* For mmu_notifiers */
1040 unsigned long mmun_end; /* For mmu_notifiers */
1041 bool is_cow;
1042 int ret;
1043
1044 /*
1045 * Don't copy ptes where a page fault will fill them correctly.
1046 * Fork becomes much lighter when there are big shared or private
1047 * readonly mappings. The tradeoff is that copy_page_range is more
1048 * efficient than faulting.
1049 */
1050 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1051 !vma->anon_vma)
1052 return 0;
1053
1054 if (is_vm_hugetlb_page(vma))
1055 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1056
1057 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1058 /*
1059 * We do not free on error cases below as remove_vma
1060 * gets called on error from higher level routine
1061 */
1062 ret = track_pfn_copy(vma);
1063 if (ret)
1064 return ret;
1065 }
1066
1067 /*
1068 * We need to invalidate the secondary MMU mappings only when
1069 * there could be a permission downgrade on the ptes of the
1070 * parent mm. And a permission downgrade will only happen if
1071 * is_cow_mapping() returns true.
1072 */
1073 is_cow = is_cow_mapping(vma->vm_flags);
1074 mmun_start = addr;
1075 mmun_end = end;
1076 if (is_cow)
1077 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1078 mmun_end);
1079
1080 ret = 0;
1081 dst_pgd = pgd_offset(dst_mm, addr);
1082 src_pgd = pgd_offset(src_mm, addr);
1083 do {
1084 next = pgd_addr_end(addr, end);
1085 if (pgd_none_or_clear_bad(src_pgd))
1086 continue;
1087 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1088 vma, addr, next))) {
1089 ret = -ENOMEM;
1090 break;
1091 }
1092 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1093
1094 if (is_cow)
1095 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1096 return ret;
1097}
1098
1099static unsigned long zap_pte_range(struct mmu_gather *tlb,
1100 struct vm_area_struct *vma, pmd_t *pmd,
1101 unsigned long addr, unsigned long end,
1102 struct zap_details *details)
1103{
1104 struct mm_struct *mm = tlb->mm;
1105 int force_flush = 0;
1106 int rss[NR_MM_COUNTERS];
1107 spinlock_t *ptl;
1108 pte_t *start_pte;
1109 pte_t *pte;
1110 swp_entry_t entry;
1111
1112again:
1113 init_rss_vec(rss);
1114 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1115 pte = start_pte;
1116 arch_enter_lazy_mmu_mode();
1117 do {
1118 pte_t ptent = *pte;
1119 if (pte_none(ptent)) {
1120 continue;
1121 }
1122
1123 if (pte_present(ptent)) {
1124 struct page *page;
1125
1126 page = vm_normal_page(vma, addr, ptent);
1127 if (unlikely(details) && page) {
1128 /*
1129 * unmap_shared_mapping_pages() wants to
1130 * invalidate cache without truncating:
1131 * unmap shared but keep private pages.
1132 */
1133 if (details->check_mapping &&
1134 details->check_mapping != page->mapping)
1135 continue;
1136 }
1137 ptent = ptep_get_and_clear_full(mm, addr, pte,
1138 tlb->fullmm);
1139 tlb_remove_tlb_entry(tlb, pte, addr);
1140 if (unlikely(!page))
1141 continue;
1142
1143 if (!PageAnon(page)) {
1144 if (pte_dirty(ptent)) {
1145 /*
1146 * oom_reaper cannot tear down dirty
1147 * pages
1148 */
1149 if (unlikely(details && details->ignore_dirty))
1150 continue;
1151 force_flush = 1;
1152 set_page_dirty(page);
1153 }
1154 if (pte_young(ptent) &&
1155 likely(!(vma->vm_flags & VM_SEQ_READ)))
1156 mark_page_accessed(page);
1157 }
1158 rss[mm_counter(page)]--;
1159 page_remove_rmap(page, false);
1160 if (unlikely(page_mapcount(page) < 0))
1161 print_bad_pte(vma, addr, ptent, page);
1162 if (unlikely(!__tlb_remove_page(tlb, page))) {
1163 force_flush = 1;
1164 addr += PAGE_SIZE;
1165 break;
1166 }
1167 continue;
1168 }
1169 /* only check swap_entries if explicitly asked for in details */
1170 if (unlikely(details && !details->check_swap_entries))
1171 continue;
1172
1173 entry = pte_to_swp_entry(ptent);
1174 if (!non_swap_entry(entry))
1175 rss[MM_SWAPENTS]--;
1176 else if (is_migration_entry(entry)) {
1177 struct page *page;
1178
1179 page = migration_entry_to_page(entry);
1180 rss[mm_counter(page)]--;
1181 }
1182 if (unlikely(!free_swap_and_cache(entry)))
1183 print_bad_pte(vma, addr, ptent, NULL);
1184 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1185 } while (pte++, addr += PAGE_SIZE, addr != end);
1186
1187 add_mm_rss_vec(mm, rss);
1188 arch_leave_lazy_mmu_mode();
1189
1190 /* Do the actual TLB flush before dropping ptl */
1191 if (force_flush)
1192 tlb_flush_mmu_tlbonly(tlb);
1193 pte_unmap_unlock(start_pte, ptl);
1194
1195 /*
1196 * If we forced a TLB flush (either due to running out of
1197 * batch buffers or because we needed to flush dirty TLB
1198 * entries before releasing the ptl), free the batched
1199 * memory too. Restart if we didn't do everything.
1200 */
1201 if (force_flush) {
1202 force_flush = 0;
1203 tlb_flush_mmu_free(tlb);
1204
1205 if (addr != end)
1206 goto again;
1207 }
1208
1209 return addr;
1210}
1211
1212static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1213 struct vm_area_struct *vma, pud_t *pud,
1214 unsigned long addr, unsigned long end,
1215 struct zap_details *details)
1216{
1217 pmd_t *pmd;
1218 unsigned long next;
1219
1220 pmd = pmd_offset(pud, addr);
1221 do {
1222 next = pmd_addr_end(addr, end);
1223 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1224 if (next - addr != HPAGE_PMD_SIZE) {
1225 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1226 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1227 split_huge_pmd(vma, pmd, addr);
1228 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1229 goto next;
1230 /* fall through */
1231 }
1232 /*
1233 * Here there can be other concurrent MADV_DONTNEED or
1234 * trans huge page faults running, and if the pmd is
1235 * none or trans huge it can change under us. This is
1236 * because MADV_DONTNEED holds the mmap_sem in read
1237 * mode.
1238 */
1239 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1240 goto next;
1241 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1242next:
1243 cond_resched();
1244 } while (pmd++, addr = next, addr != end);
1245
1246 return addr;
1247}
1248
1249static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1250 struct vm_area_struct *vma, pgd_t *pgd,
1251 unsigned long addr, unsigned long end,
1252 struct zap_details *details)
1253{
1254 pud_t *pud;
1255 unsigned long next;
1256
1257 pud = pud_offset(pgd, addr);
1258 do {
1259 next = pud_addr_end(addr, end);
1260 if (pud_none_or_clear_bad(pud))
1261 continue;
1262 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1263 } while (pud++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268void unmap_page_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pgd_t *pgd;
1274 unsigned long next;
1275
1276 BUG_ON(addr >= end);
1277 tlb_start_vma(tlb, vma);
1278 pgd = pgd_offset(vma->vm_mm, addr);
1279 do {
1280 next = pgd_addr_end(addr, end);
1281 if (pgd_none_or_clear_bad(pgd))
1282 continue;
1283 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1284 } while (pgd++, addr = next, addr != end);
1285 tlb_end_vma(tlb, vma);
1286}
1287
1288
1289static void unmap_single_vma(struct mmu_gather *tlb,
1290 struct vm_area_struct *vma, unsigned long start_addr,
1291 unsigned long end_addr,
1292 struct zap_details *details)
1293{
1294 unsigned long start = max(vma->vm_start, start_addr);
1295 unsigned long end;
1296
1297 if (start >= vma->vm_end)
1298 return;
1299 end = min(vma->vm_end, end_addr);
1300 if (end <= vma->vm_start)
1301 return;
1302
1303 if (vma->vm_file)
1304 uprobe_munmap(vma, start, end);
1305
1306 if (unlikely(vma->vm_flags & VM_PFNMAP))
1307 untrack_pfn(vma, 0, 0);
1308
1309 if (start != end) {
1310 if (unlikely(is_vm_hugetlb_page(vma))) {
1311 /*
1312 * It is undesirable to test vma->vm_file as it
1313 * should be non-null for valid hugetlb area.
1314 * However, vm_file will be NULL in the error
1315 * cleanup path of mmap_region. When
1316 * hugetlbfs ->mmap method fails,
1317 * mmap_region() nullifies vma->vm_file
1318 * before calling this function to clean up.
1319 * Since no pte has actually been setup, it is
1320 * safe to do nothing in this case.
1321 */
1322 if (vma->vm_file) {
1323 i_mmap_lock_write(vma->vm_file->f_mapping);
1324 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1325 i_mmap_unlock_write(vma->vm_file->f_mapping);
1326 }
1327 } else
1328 unmap_page_range(tlb, vma, start, end, details);
1329 }
1330}
1331
1332/**
1333 * unmap_vmas - unmap a range of memory covered by a list of vma's
1334 * @tlb: address of the caller's struct mmu_gather
1335 * @vma: the starting vma
1336 * @start_addr: virtual address at which to start unmapping
1337 * @end_addr: virtual address at which to end unmapping
1338 *
1339 * Unmap all pages in the vma list.
1340 *
1341 * Only addresses between `start' and `end' will be unmapped.
1342 *
1343 * The VMA list must be sorted in ascending virtual address order.
1344 *
1345 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1346 * range after unmap_vmas() returns. So the only responsibility here is to
1347 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1348 * drops the lock and schedules.
1349 */
1350void unmap_vmas(struct mmu_gather *tlb,
1351 struct vm_area_struct *vma, unsigned long start_addr,
1352 unsigned long end_addr)
1353{
1354 struct mm_struct *mm = vma->vm_mm;
1355
1356 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1357 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1358 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1359 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1360}
1361
1362/**
1363 * zap_page_range - remove user pages in a given range
1364 * @vma: vm_area_struct holding the applicable pages
1365 * @start: starting address of pages to zap
1366 * @size: number of bytes to zap
1367 * @details: details of shared cache invalidation
1368 *
1369 * Caller must protect the VMA list
1370 */
1371void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1372 unsigned long size, struct zap_details *details)
1373{
1374 struct mm_struct *mm = vma->vm_mm;
1375 struct mmu_gather tlb;
1376 unsigned long end = start + size;
1377
1378 lru_add_drain();
1379 tlb_gather_mmu(&tlb, mm, start, end);
1380 update_hiwater_rss(mm);
1381 mmu_notifier_invalidate_range_start(mm, start, end);
1382 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1383 unmap_single_vma(&tlb, vma, start, end, details);
1384 mmu_notifier_invalidate_range_end(mm, start, end);
1385 tlb_finish_mmu(&tlb, start, end);
1386}
1387
1388/**
1389 * zap_page_range_single - remove user pages in a given range
1390 * @vma: vm_area_struct holding the applicable pages
1391 * @address: starting address of pages to zap
1392 * @size: number of bytes to zap
1393 * @details: details of shared cache invalidation
1394 *
1395 * The range must fit into one VMA.
1396 */
1397static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1398 unsigned long size, struct zap_details *details)
1399{
1400 struct mm_struct *mm = vma->vm_mm;
1401 struct mmu_gather tlb;
1402 unsigned long end = address + size;
1403
1404 lru_add_drain();
1405 tlb_gather_mmu(&tlb, mm, address, end);
1406 update_hiwater_rss(mm);
1407 mmu_notifier_invalidate_range_start(mm, address, end);
1408 unmap_single_vma(&tlb, vma, address, end, details);
1409 mmu_notifier_invalidate_range_end(mm, address, end);
1410 tlb_finish_mmu(&tlb, address, end);
1411}
1412
1413/**
1414 * zap_vma_ptes - remove ptes mapping the vma
1415 * @vma: vm_area_struct holding ptes to be zapped
1416 * @address: starting address of pages to zap
1417 * @size: number of bytes to zap
1418 *
1419 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1420 *
1421 * The entire address range must be fully contained within the vma.
1422 *
1423 * Returns 0 if successful.
1424 */
1425int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1426 unsigned long size)
1427{
1428 if (address < vma->vm_start || address + size > vma->vm_end ||
1429 !(vma->vm_flags & VM_PFNMAP))
1430 return -1;
1431 zap_page_range_single(vma, address, size, NULL);
1432 return 0;
1433}
1434EXPORT_SYMBOL_GPL(zap_vma_ptes);
1435
1436pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1437 spinlock_t **ptl)
1438{
1439 pgd_t * pgd = pgd_offset(mm, addr);
1440 pud_t * pud = pud_alloc(mm, pgd, addr);
1441 if (pud) {
1442 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1443 if (pmd) {
1444 VM_BUG_ON(pmd_trans_huge(*pmd));
1445 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1446 }
1447 }
1448 return NULL;
1449}
1450
1451/*
1452 * This is the old fallback for page remapping.
1453 *
1454 * For historical reasons, it only allows reserved pages. Only
1455 * old drivers should use this, and they needed to mark their
1456 * pages reserved for the old functions anyway.
1457 */
1458static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1459 struct page *page, pgprot_t prot)
1460{
1461 struct mm_struct *mm = vma->vm_mm;
1462 int retval;
1463 pte_t *pte;
1464 spinlock_t *ptl;
1465
1466 retval = -EINVAL;
1467 if (PageAnon(page))
1468 goto out;
1469 retval = -ENOMEM;
1470 flush_dcache_page(page);
1471 pte = get_locked_pte(mm, addr, &ptl);
1472 if (!pte)
1473 goto out;
1474 retval = -EBUSY;
1475 if (!pte_none(*pte))
1476 goto out_unlock;
1477
1478 /* Ok, finally just insert the thing.. */
1479 get_page(page);
1480 inc_mm_counter_fast(mm, mm_counter_file(page));
1481 page_add_file_rmap(page);
1482 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1483
1484 retval = 0;
1485 pte_unmap_unlock(pte, ptl);
1486 return retval;
1487out_unlock:
1488 pte_unmap_unlock(pte, ptl);
1489out:
1490 return retval;
1491}
1492
1493/**
1494 * vm_insert_page - insert single page into user vma
1495 * @vma: user vma to map to
1496 * @addr: target user address of this page
1497 * @page: source kernel page
1498 *
1499 * This allows drivers to insert individual pages they've allocated
1500 * into a user vma.
1501 *
1502 * The page has to be a nice clean _individual_ kernel allocation.
1503 * If you allocate a compound page, you need to have marked it as
1504 * such (__GFP_COMP), or manually just split the page up yourself
1505 * (see split_page()).
1506 *
1507 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1508 * took an arbitrary page protection parameter. This doesn't allow
1509 * that. Your vma protection will have to be set up correctly, which
1510 * means that if you want a shared writable mapping, you'd better
1511 * ask for a shared writable mapping!
1512 *
1513 * The page does not need to be reserved.
1514 *
1515 * Usually this function is called from f_op->mmap() handler
1516 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1517 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1518 * function from other places, for example from page-fault handler.
1519 */
1520int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1521 struct page *page)
1522{
1523 if (addr < vma->vm_start || addr >= vma->vm_end)
1524 return -EFAULT;
1525 if (!page_count(page))
1526 return -EINVAL;
1527 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1528 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1529 BUG_ON(vma->vm_flags & VM_PFNMAP);
1530 vma->vm_flags |= VM_MIXEDMAP;
1531 }
1532 return insert_page(vma, addr, page, vma->vm_page_prot);
1533}
1534EXPORT_SYMBOL(vm_insert_page);
1535
1536static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1537 pfn_t pfn, pgprot_t prot)
1538{
1539 struct mm_struct *mm = vma->vm_mm;
1540 int retval;
1541 pte_t *pte, entry;
1542 spinlock_t *ptl;
1543
1544 retval = -ENOMEM;
1545 pte = get_locked_pte(mm, addr, &ptl);
1546 if (!pte)
1547 goto out;
1548 retval = -EBUSY;
1549 if (!pte_none(*pte))
1550 goto out_unlock;
1551
1552 /* Ok, finally just insert the thing.. */
1553 if (pfn_t_devmap(pfn))
1554 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1555 else
1556 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1557 set_pte_at(mm, addr, pte, entry);
1558 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1559
1560 retval = 0;
1561out_unlock:
1562 pte_unmap_unlock(pte, ptl);
1563out:
1564 return retval;
1565}
1566
1567/**
1568 * vm_insert_pfn - insert single pfn into user vma
1569 * @vma: user vma to map to
1570 * @addr: target user address of this page
1571 * @pfn: source kernel pfn
1572 *
1573 * Similar to vm_insert_page, this allows drivers to insert individual pages
1574 * they've allocated into a user vma. Same comments apply.
1575 *
1576 * This function should only be called from a vm_ops->fault handler, and
1577 * in that case the handler should return NULL.
1578 *
1579 * vma cannot be a COW mapping.
1580 *
1581 * As this is called only for pages that do not currently exist, we
1582 * do not need to flush old virtual caches or the TLB.
1583 */
1584int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1585 unsigned long pfn)
1586{
1587 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1588}
1589EXPORT_SYMBOL(vm_insert_pfn);
1590
1591/**
1592 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1593 * @vma: user vma to map to
1594 * @addr: target user address of this page
1595 * @pfn: source kernel pfn
1596 * @pgprot: pgprot flags for the inserted page
1597 *
1598 * This is exactly like vm_insert_pfn, except that it allows drivers to
1599 * to override pgprot on a per-page basis.
1600 *
1601 * This only makes sense for IO mappings, and it makes no sense for
1602 * cow mappings. In general, using multiple vmas is preferable;
1603 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1604 * impractical.
1605 */
1606int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1607 unsigned long pfn, pgprot_t pgprot)
1608{
1609 int ret;
1610 /*
1611 * Technically, architectures with pte_special can avoid all these
1612 * restrictions (same for remap_pfn_range). However we would like
1613 * consistency in testing and feature parity among all, so we should
1614 * try to keep these invariants in place for everybody.
1615 */
1616 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1617 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1618 (VM_PFNMAP|VM_MIXEDMAP));
1619 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1620 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1621
1622 if (addr < vma->vm_start || addr >= vma->vm_end)
1623 return -EFAULT;
1624 if (track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV)))
1625 return -EINVAL;
1626
1627 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1628
1629 return ret;
1630}
1631EXPORT_SYMBOL(vm_insert_pfn_prot);
1632
1633int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1634 pfn_t pfn)
1635{
1636 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1637
1638 if (addr < vma->vm_start || addr >= vma->vm_end)
1639 return -EFAULT;
1640
1641 /*
1642 * If we don't have pte special, then we have to use the pfn_valid()
1643 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1644 * refcount the page if pfn_valid is true (hence insert_page rather
1645 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1646 * without pte special, it would there be refcounted as a normal page.
1647 */
1648 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1649 struct page *page;
1650
1651 /*
1652 * At this point we are committed to insert_page()
1653 * regardless of whether the caller specified flags that
1654 * result in pfn_t_has_page() == false.
1655 */
1656 page = pfn_to_page(pfn_t_to_pfn(pfn));
1657 return insert_page(vma, addr, page, vma->vm_page_prot);
1658 }
1659 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
1660}
1661EXPORT_SYMBOL(vm_insert_mixed);
1662
1663/*
1664 * maps a range of physical memory into the requested pages. the old
1665 * mappings are removed. any references to nonexistent pages results
1666 * in null mappings (currently treated as "copy-on-access")
1667 */
1668static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1669 unsigned long addr, unsigned long end,
1670 unsigned long pfn, pgprot_t prot)
1671{
1672 pte_t *pte;
1673 spinlock_t *ptl;
1674
1675 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1676 if (!pte)
1677 return -ENOMEM;
1678 arch_enter_lazy_mmu_mode();
1679 do {
1680 BUG_ON(!pte_none(*pte));
1681 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1682 pfn++;
1683 } while (pte++, addr += PAGE_SIZE, addr != end);
1684 arch_leave_lazy_mmu_mode();
1685 pte_unmap_unlock(pte - 1, ptl);
1686 return 0;
1687}
1688
1689static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1690 unsigned long addr, unsigned long end,
1691 unsigned long pfn, pgprot_t prot)
1692{
1693 pmd_t *pmd;
1694 unsigned long next;
1695
1696 pfn -= addr >> PAGE_SHIFT;
1697 pmd = pmd_alloc(mm, pud, addr);
1698 if (!pmd)
1699 return -ENOMEM;
1700 VM_BUG_ON(pmd_trans_huge(*pmd));
1701 do {
1702 next = pmd_addr_end(addr, end);
1703 if (remap_pte_range(mm, pmd, addr, next,
1704 pfn + (addr >> PAGE_SHIFT), prot))
1705 return -ENOMEM;
1706 } while (pmd++, addr = next, addr != end);
1707 return 0;
1708}
1709
1710static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1711 unsigned long addr, unsigned long end,
1712 unsigned long pfn, pgprot_t prot)
1713{
1714 pud_t *pud;
1715 unsigned long next;
1716
1717 pfn -= addr >> PAGE_SHIFT;
1718 pud = pud_alloc(mm, pgd, addr);
1719 if (!pud)
1720 return -ENOMEM;
1721 do {
1722 next = pud_addr_end(addr, end);
1723 if (remap_pmd_range(mm, pud, addr, next,
1724 pfn + (addr >> PAGE_SHIFT), prot))
1725 return -ENOMEM;
1726 } while (pud++, addr = next, addr != end);
1727 return 0;
1728}
1729
1730/**
1731 * remap_pfn_range - remap kernel memory to userspace
1732 * @vma: user vma to map to
1733 * @addr: target user address to start at
1734 * @pfn: physical address of kernel memory
1735 * @size: size of map area
1736 * @prot: page protection flags for this mapping
1737 *
1738 * Note: this is only safe if the mm semaphore is held when called.
1739 */
1740int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1741 unsigned long pfn, unsigned long size, pgprot_t prot)
1742{
1743 pgd_t *pgd;
1744 unsigned long next;
1745 unsigned long end = addr + PAGE_ALIGN(size);
1746 struct mm_struct *mm = vma->vm_mm;
1747 int err;
1748
1749 /*
1750 * Physically remapped pages are special. Tell the
1751 * rest of the world about it:
1752 * VM_IO tells people not to look at these pages
1753 * (accesses can have side effects).
1754 * VM_PFNMAP tells the core MM that the base pages are just
1755 * raw PFN mappings, and do not have a "struct page" associated
1756 * with them.
1757 * VM_DONTEXPAND
1758 * Disable vma merging and expanding with mremap().
1759 * VM_DONTDUMP
1760 * Omit vma from core dump, even when VM_IO turned off.
1761 *
1762 * There's a horrible special case to handle copy-on-write
1763 * behaviour that some programs depend on. We mark the "original"
1764 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1765 * See vm_normal_page() for details.
1766 */
1767 if (is_cow_mapping(vma->vm_flags)) {
1768 if (addr != vma->vm_start || end != vma->vm_end)
1769 return -EINVAL;
1770 vma->vm_pgoff = pfn;
1771 }
1772
1773 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
1774 if (err)
1775 return -EINVAL;
1776
1777 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1778
1779 BUG_ON(addr >= end);
1780 pfn -= addr >> PAGE_SHIFT;
1781 pgd = pgd_offset(mm, addr);
1782 flush_cache_range(vma, addr, end);
1783 do {
1784 next = pgd_addr_end(addr, end);
1785 err = remap_pud_range(mm, pgd, addr, next,
1786 pfn + (addr >> PAGE_SHIFT), prot);
1787 if (err)
1788 break;
1789 } while (pgd++, addr = next, addr != end);
1790
1791 if (err)
1792 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
1793
1794 return err;
1795}
1796EXPORT_SYMBOL(remap_pfn_range);
1797
1798/**
1799 * vm_iomap_memory - remap memory to userspace
1800 * @vma: user vma to map to
1801 * @start: start of area
1802 * @len: size of area
1803 *
1804 * This is a simplified io_remap_pfn_range() for common driver use. The
1805 * driver just needs to give us the physical memory range to be mapped,
1806 * we'll figure out the rest from the vma information.
1807 *
1808 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1809 * whatever write-combining details or similar.
1810 */
1811int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1812{
1813 unsigned long vm_len, pfn, pages;
1814
1815 /* Check that the physical memory area passed in looks valid */
1816 if (start + len < start)
1817 return -EINVAL;
1818 /*
1819 * You *really* shouldn't map things that aren't page-aligned,
1820 * but we've historically allowed it because IO memory might
1821 * just have smaller alignment.
1822 */
1823 len += start & ~PAGE_MASK;
1824 pfn = start >> PAGE_SHIFT;
1825 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1826 if (pfn + pages < pfn)
1827 return -EINVAL;
1828
1829 /* We start the mapping 'vm_pgoff' pages into the area */
1830 if (vma->vm_pgoff > pages)
1831 return -EINVAL;
1832 pfn += vma->vm_pgoff;
1833 pages -= vma->vm_pgoff;
1834
1835 /* Can we fit all of the mapping? */
1836 vm_len = vma->vm_end - vma->vm_start;
1837 if (vm_len >> PAGE_SHIFT > pages)
1838 return -EINVAL;
1839
1840 /* Ok, let it rip */
1841 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1842}
1843EXPORT_SYMBOL(vm_iomap_memory);
1844
1845static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1846 unsigned long addr, unsigned long end,
1847 pte_fn_t fn, void *data)
1848{
1849 pte_t *pte;
1850 int err;
1851 pgtable_t token;
1852 spinlock_t *uninitialized_var(ptl);
1853
1854 pte = (mm == &init_mm) ?
1855 pte_alloc_kernel(pmd, addr) :
1856 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1857 if (!pte)
1858 return -ENOMEM;
1859
1860 BUG_ON(pmd_huge(*pmd));
1861
1862 arch_enter_lazy_mmu_mode();
1863
1864 token = pmd_pgtable(*pmd);
1865
1866 do {
1867 err = fn(pte++, token, addr, data);
1868 if (err)
1869 break;
1870 } while (addr += PAGE_SIZE, addr != end);
1871
1872 arch_leave_lazy_mmu_mode();
1873
1874 if (mm != &init_mm)
1875 pte_unmap_unlock(pte-1, ptl);
1876 return err;
1877}
1878
1879static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1880 unsigned long addr, unsigned long end,
1881 pte_fn_t fn, void *data)
1882{
1883 pmd_t *pmd;
1884 unsigned long next;
1885 int err;
1886
1887 BUG_ON(pud_huge(*pud));
1888
1889 pmd = pmd_alloc(mm, pud, addr);
1890 if (!pmd)
1891 return -ENOMEM;
1892 do {
1893 next = pmd_addr_end(addr, end);
1894 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1895 if (err)
1896 break;
1897 } while (pmd++, addr = next, addr != end);
1898 return err;
1899}
1900
1901static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1902 unsigned long addr, unsigned long end,
1903 pte_fn_t fn, void *data)
1904{
1905 pud_t *pud;
1906 unsigned long next;
1907 int err;
1908
1909 pud = pud_alloc(mm, pgd, addr);
1910 if (!pud)
1911 return -ENOMEM;
1912 do {
1913 next = pud_addr_end(addr, end);
1914 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1915 if (err)
1916 break;
1917 } while (pud++, addr = next, addr != end);
1918 return err;
1919}
1920
1921/*
1922 * Scan a region of virtual memory, filling in page tables as necessary
1923 * and calling a provided function on each leaf page table.
1924 */
1925int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1926 unsigned long size, pte_fn_t fn, void *data)
1927{
1928 pgd_t *pgd;
1929 unsigned long next;
1930 unsigned long end = addr + size;
1931 int err;
1932
1933 if (WARN_ON(addr >= end))
1934 return -EINVAL;
1935
1936 pgd = pgd_offset(mm, addr);
1937 do {
1938 next = pgd_addr_end(addr, end);
1939 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1940 if (err)
1941 break;
1942 } while (pgd++, addr = next, addr != end);
1943
1944 return err;
1945}
1946EXPORT_SYMBOL_GPL(apply_to_page_range);
1947
1948/*
1949 * handle_pte_fault chooses page fault handler according to an entry which was
1950 * read non-atomically. Before making any commitment, on those architectures
1951 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1952 * parts, do_swap_page must check under lock before unmapping the pte and
1953 * proceeding (but do_wp_page is only called after already making such a check;
1954 * and do_anonymous_page can safely check later on).
1955 */
1956static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1957 pte_t *page_table, pte_t orig_pte)
1958{
1959 int same = 1;
1960#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1961 if (sizeof(pte_t) > sizeof(unsigned long)) {
1962 spinlock_t *ptl = pte_lockptr(mm, pmd);
1963 spin_lock(ptl);
1964 same = pte_same(*page_table, orig_pte);
1965 spin_unlock(ptl);
1966 }
1967#endif
1968 pte_unmap(page_table);
1969 return same;
1970}
1971
1972static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1973{
1974 debug_dma_assert_idle(src);
1975
1976 /*
1977 * If the source page was a PFN mapping, we don't have
1978 * a "struct page" for it. We do a best-effort copy by
1979 * just copying from the original user address. If that
1980 * fails, we just zero-fill it. Live with it.
1981 */
1982 if (unlikely(!src)) {
1983 void *kaddr = kmap_atomic(dst);
1984 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1985
1986 /*
1987 * This really shouldn't fail, because the page is there
1988 * in the page tables. But it might just be unreadable,
1989 * in which case we just give up and fill the result with
1990 * zeroes.
1991 */
1992 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1993 clear_page(kaddr);
1994 kunmap_atomic(kaddr);
1995 flush_dcache_page(dst);
1996 } else
1997 copy_user_highpage(dst, src, va, vma);
1998}
1999
2000static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2001{
2002 struct file *vm_file = vma->vm_file;
2003
2004 if (vm_file)
2005 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2006
2007 /*
2008 * Special mappings (e.g. VDSO) do not have any file so fake
2009 * a default GFP_KERNEL for them.
2010 */
2011 return GFP_KERNEL;
2012}
2013
2014/*
2015 * Notify the address space that the page is about to become writable so that
2016 * it can prohibit this or wait for the page to get into an appropriate state.
2017 *
2018 * We do this without the lock held, so that it can sleep if it needs to.
2019 */
2020static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2021 unsigned long address)
2022{
2023 struct vm_fault vmf;
2024 int ret;
2025
2026 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2027 vmf.pgoff = page->index;
2028 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2029 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2030 vmf.page = page;
2031 vmf.cow_page = NULL;
2032
2033 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2034 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2035 return ret;
2036 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2037 lock_page(page);
2038 if (!page->mapping) {
2039 unlock_page(page);
2040 return 0; /* retry */
2041 }
2042 ret |= VM_FAULT_LOCKED;
2043 } else
2044 VM_BUG_ON_PAGE(!PageLocked(page), page);
2045 return ret;
2046}
2047
2048/*
2049 * Handle write page faults for pages that can be reused in the current vma
2050 *
2051 * This can happen either due to the mapping being with the VM_SHARED flag,
2052 * or due to us being the last reference standing to the page. In either
2053 * case, all we need to do here is to mark the page as writable and update
2054 * any related book-keeping.
2055 */
2056static inline int wp_page_reuse(struct mm_struct *mm,
2057 struct vm_area_struct *vma, unsigned long address,
2058 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2059 struct page *page, int page_mkwrite,
2060 int dirty_shared)
2061 __releases(ptl)
2062{
2063 pte_t entry;
2064 /*
2065 * Clear the pages cpupid information as the existing
2066 * information potentially belongs to a now completely
2067 * unrelated process.
2068 */
2069 if (page)
2070 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2071
2072 flush_cache_page(vma, address, pte_pfn(orig_pte));
2073 entry = pte_mkyoung(orig_pte);
2074 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2075 if (ptep_set_access_flags(vma, address, page_table, entry, 1))
2076 update_mmu_cache(vma, address, page_table);
2077 pte_unmap_unlock(page_table, ptl);
2078
2079 if (dirty_shared) {
2080 struct address_space *mapping;
2081 int dirtied;
2082
2083 if (!page_mkwrite)
2084 lock_page(page);
2085
2086 dirtied = set_page_dirty(page);
2087 VM_BUG_ON_PAGE(PageAnon(page), page);
2088 mapping = page->mapping;
2089 unlock_page(page);
2090 put_page(page);
2091
2092 if ((dirtied || page_mkwrite) && mapping) {
2093 /*
2094 * Some device drivers do not set page.mapping
2095 * but still dirty their pages
2096 */
2097 balance_dirty_pages_ratelimited(mapping);
2098 }
2099
2100 if (!page_mkwrite)
2101 file_update_time(vma->vm_file);
2102 }
2103
2104 return VM_FAULT_WRITE;
2105}
2106
2107/*
2108 * Handle the case of a page which we actually need to copy to a new page.
2109 *
2110 * Called with mmap_sem locked and the old page referenced, but
2111 * without the ptl held.
2112 *
2113 * High level logic flow:
2114 *
2115 * - Allocate a page, copy the content of the old page to the new one.
2116 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2117 * - Take the PTL. If the pte changed, bail out and release the allocated page
2118 * - If the pte is still the way we remember it, update the page table and all
2119 * relevant references. This includes dropping the reference the page-table
2120 * held to the old page, as well as updating the rmap.
2121 * - In any case, unlock the PTL and drop the reference we took to the old page.
2122 */
2123static int wp_page_copy(struct mm_struct *mm, struct vm_area_struct *vma,
2124 unsigned long address, pte_t *page_table, pmd_t *pmd,
2125 pte_t orig_pte, struct page *old_page)
2126{
2127 struct page *new_page = NULL;
2128 spinlock_t *ptl = NULL;
2129 pte_t entry;
2130 int page_copied = 0;
2131 const unsigned long mmun_start = address & PAGE_MASK; /* For mmu_notifiers */
2132 const unsigned long mmun_end = mmun_start + PAGE_SIZE; /* For mmu_notifiers */
2133 struct mem_cgroup *memcg;
2134
2135 if (unlikely(anon_vma_prepare(vma)))
2136 goto oom;
2137
2138 if (is_zero_pfn(pte_pfn(orig_pte))) {
2139 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2140 if (!new_page)
2141 goto oom;
2142 } else {
2143 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2144 if (!new_page)
2145 goto oom;
2146 cow_user_page(new_page, old_page, address, vma);
2147 }
2148
2149 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2150 goto oom_free_new;
2151
2152 __SetPageUptodate(new_page);
2153
2154 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2155
2156 /*
2157 * Re-check the pte - we dropped the lock
2158 */
2159 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2160 if (likely(pte_same(*page_table, orig_pte))) {
2161 if (old_page) {
2162 if (!PageAnon(old_page)) {
2163 dec_mm_counter_fast(mm,
2164 mm_counter_file(old_page));
2165 inc_mm_counter_fast(mm, MM_ANONPAGES);
2166 }
2167 } else {
2168 inc_mm_counter_fast(mm, MM_ANONPAGES);
2169 }
2170 flush_cache_page(vma, address, pte_pfn(orig_pte));
2171 entry = mk_pte(new_page, vma->vm_page_prot);
2172 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2173 /*
2174 * Clear the pte entry and flush it first, before updating the
2175 * pte with the new entry. This will avoid a race condition
2176 * seen in the presence of one thread doing SMC and another
2177 * thread doing COW.
2178 */
2179 ptep_clear_flush_notify(vma, address, page_table);
2180 page_add_new_anon_rmap(new_page, vma, address, false);
2181 mem_cgroup_commit_charge(new_page, memcg, false, false);
2182 lru_cache_add_active_or_unevictable(new_page, vma);
2183 /*
2184 * We call the notify macro here because, when using secondary
2185 * mmu page tables (such as kvm shadow page tables), we want the
2186 * new page to be mapped directly into the secondary page table.
2187 */
2188 set_pte_at_notify(mm, address, page_table, entry);
2189 update_mmu_cache(vma, address, page_table);
2190 if (old_page) {
2191 /*
2192 * Only after switching the pte to the new page may
2193 * we remove the mapcount here. Otherwise another
2194 * process may come and find the rmap count decremented
2195 * before the pte is switched to the new page, and
2196 * "reuse" the old page writing into it while our pte
2197 * here still points into it and can be read by other
2198 * threads.
2199 *
2200 * The critical issue is to order this
2201 * page_remove_rmap with the ptp_clear_flush above.
2202 * Those stores are ordered by (if nothing else,)
2203 * the barrier present in the atomic_add_negative
2204 * in page_remove_rmap.
2205 *
2206 * Then the TLB flush in ptep_clear_flush ensures that
2207 * no process can access the old page before the
2208 * decremented mapcount is visible. And the old page
2209 * cannot be reused until after the decremented
2210 * mapcount is visible. So transitively, TLBs to
2211 * old page will be flushed before it can be reused.
2212 */
2213 page_remove_rmap(old_page, false);
2214 }
2215
2216 /* Free the old page.. */
2217 new_page = old_page;
2218 page_copied = 1;
2219 } else {
2220 mem_cgroup_cancel_charge(new_page, memcg, false);
2221 }
2222
2223 if (new_page)
2224 put_page(new_page);
2225
2226 pte_unmap_unlock(page_table, ptl);
2227 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2228 if (old_page) {
2229 /*
2230 * Don't let another task, with possibly unlocked vma,
2231 * keep the mlocked page.
2232 */
2233 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2234 lock_page(old_page); /* LRU manipulation */
2235 if (PageMlocked(old_page))
2236 munlock_vma_page(old_page);
2237 unlock_page(old_page);
2238 }
2239 put_page(old_page);
2240 }
2241 return page_copied ? VM_FAULT_WRITE : 0;
2242oom_free_new:
2243 put_page(new_page);
2244oom:
2245 if (old_page)
2246 put_page(old_page);
2247 return VM_FAULT_OOM;
2248}
2249
2250/*
2251 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2252 * mapping
2253 */
2254static int wp_pfn_shared(struct mm_struct *mm,
2255 struct vm_area_struct *vma, unsigned long address,
2256 pte_t *page_table, spinlock_t *ptl, pte_t orig_pte,
2257 pmd_t *pmd)
2258{
2259 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2260 struct vm_fault vmf = {
2261 .page = NULL,
2262 .pgoff = linear_page_index(vma, address),
2263 .virtual_address = (void __user *)(address & PAGE_MASK),
2264 .flags = FAULT_FLAG_WRITE | FAULT_FLAG_MKWRITE,
2265 };
2266 int ret;
2267
2268 pte_unmap_unlock(page_table, ptl);
2269 ret = vma->vm_ops->pfn_mkwrite(vma, &vmf);
2270 if (ret & VM_FAULT_ERROR)
2271 return ret;
2272 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2273 /*
2274 * We might have raced with another page fault while we
2275 * released the pte_offset_map_lock.
2276 */
2277 if (!pte_same(*page_table, orig_pte)) {
2278 pte_unmap_unlock(page_table, ptl);
2279 return 0;
2280 }
2281 }
2282 return wp_page_reuse(mm, vma, address, page_table, ptl, orig_pte,
2283 NULL, 0, 0);
2284}
2285
2286static int wp_page_shared(struct mm_struct *mm, struct vm_area_struct *vma,
2287 unsigned long address, pte_t *page_table,
2288 pmd_t *pmd, spinlock_t *ptl, pte_t orig_pte,
2289 struct page *old_page)
2290 __releases(ptl)
2291{
2292 int page_mkwrite = 0;
2293
2294 get_page(old_page);
2295
2296 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2297 int tmp;
2298
2299 pte_unmap_unlock(page_table, ptl);
2300 tmp = do_page_mkwrite(vma, old_page, address);
2301 if (unlikely(!tmp || (tmp &
2302 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2303 put_page(old_page);
2304 return tmp;
2305 }
2306 /*
2307 * Since we dropped the lock we need to revalidate
2308 * the PTE as someone else may have changed it. If
2309 * they did, we just return, as we can count on the
2310 * MMU to tell us if they didn't also make it writable.
2311 */
2312 page_table = pte_offset_map_lock(mm, pmd, address,
2313 &ptl);
2314 if (!pte_same(*page_table, orig_pte)) {
2315 unlock_page(old_page);
2316 pte_unmap_unlock(page_table, ptl);
2317 put_page(old_page);
2318 return 0;
2319 }
2320 page_mkwrite = 1;
2321 }
2322
2323 return wp_page_reuse(mm, vma, address, page_table, ptl,
2324 orig_pte, old_page, page_mkwrite, 1);
2325}
2326
2327/*
2328 * This routine handles present pages, when users try to write
2329 * to a shared page. It is done by copying the page to a new address
2330 * and decrementing the shared-page counter for the old page.
2331 *
2332 * Note that this routine assumes that the protection checks have been
2333 * done by the caller (the low-level page fault routine in most cases).
2334 * Thus we can safely just mark it writable once we've done any necessary
2335 * COW.
2336 *
2337 * We also mark the page dirty at this point even though the page will
2338 * change only once the write actually happens. This avoids a few races,
2339 * and potentially makes it more efficient.
2340 *
2341 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2342 * but allow concurrent faults), with pte both mapped and locked.
2343 * We return with mmap_sem still held, but pte unmapped and unlocked.
2344 */
2345static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2346 unsigned long address, pte_t *page_table, pmd_t *pmd,
2347 spinlock_t *ptl, pte_t orig_pte)
2348 __releases(ptl)
2349{
2350 struct page *old_page;
2351
2352 old_page = vm_normal_page(vma, address, orig_pte);
2353 if (!old_page) {
2354 /*
2355 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2356 * VM_PFNMAP VMA.
2357 *
2358 * We should not cow pages in a shared writeable mapping.
2359 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2360 */
2361 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2362 (VM_WRITE|VM_SHARED))
2363 return wp_pfn_shared(mm, vma, address, page_table, ptl,
2364 orig_pte, pmd);
2365
2366 pte_unmap_unlock(page_table, ptl);
2367 return wp_page_copy(mm, vma, address, page_table, pmd,
2368 orig_pte, old_page);
2369 }
2370
2371 /*
2372 * Take out anonymous pages first, anonymous shared vmas are
2373 * not dirty accountable.
2374 */
2375 if (PageAnon(old_page) && !PageKsm(old_page)) {
2376 int total_mapcount;
2377 if (!trylock_page(old_page)) {
2378 get_page(old_page);
2379 pte_unmap_unlock(page_table, ptl);
2380 lock_page(old_page);
2381 page_table = pte_offset_map_lock(mm, pmd, address,
2382 &ptl);
2383 if (!pte_same(*page_table, orig_pte)) {
2384 unlock_page(old_page);
2385 pte_unmap_unlock(page_table, ptl);
2386 put_page(old_page);
2387 return 0;
2388 }
2389 put_page(old_page);
2390 }
2391 if (reuse_swap_page(old_page, &total_mapcount)) {
2392 if (total_mapcount == 1) {
2393 /*
2394 * The page is all ours. Move it to
2395 * our anon_vma so the rmap code will
2396 * not search our parent or siblings.
2397 * Protected against the rmap code by
2398 * the page lock.
2399 */
2400 page_move_anon_rmap(compound_head(old_page),
2401 vma, address);
2402 }
2403 unlock_page(old_page);
2404 return wp_page_reuse(mm, vma, address, page_table, ptl,
2405 orig_pte, old_page, 0, 0);
2406 }
2407 unlock_page(old_page);
2408 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2409 (VM_WRITE|VM_SHARED))) {
2410 return wp_page_shared(mm, vma, address, page_table, pmd,
2411 ptl, orig_pte, old_page);
2412 }
2413
2414 /*
2415 * Ok, we need to copy. Oh, well..
2416 */
2417 get_page(old_page);
2418
2419 pte_unmap_unlock(page_table, ptl);
2420 return wp_page_copy(mm, vma, address, page_table, pmd,
2421 orig_pte, old_page);
2422}
2423
2424static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2425 unsigned long start_addr, unsigned long end_addr,
2426 struct zap_details *details)
2427{
2428 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2429}
2430
2431static inline void unmap_mapping_range_tree(struct rb_root *root,
2432 struct zap_details *details)
2433{
2434 struct vm_area_struct *vma;
2435 pgoff_t vba, vea, zba, zea;
2436
2437 vma_interval_tree_foreach(vma, root,
2438 details->first_index, details->last_index) {
2439
2440 vba = vma->vm_pgoff;
2441 vea = vba + vma_pages(vma) - 1;
2442 zba = details->first_index;
2443 if (zba < vba)
2444 zba = vba;
2445 zea = details->last_index;
2446 if (zea > vea)
2447 zea = vea;
2448
2449 unmap_mapping_range_vma(vma,
2450 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2451 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2452 details);
2453 }
2454}
2455
2456/**
2457 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2458 * address_space corresponding to the specified page range in the underlying
2459 * file.
2460 *
2461 * @mapping: the address space containing mmaps to be unmapped.
2462 * @holebegin: byte in first page to unmap, relative to the start of
2463 * the underlying file. This will be rounded down to a PAGE_SIZE
2464 * boundary. Note that this is different from truncate_pagecache(), which
2465 * must keep the partial page. In contrast, we must get rid of
2466 * partial pages.
2467 * @holelen: size of prospective hole in bytes. This will be rounded
2468 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2469 * end of the file.
2470 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2471 * but 0 when invalidating pagecache, don't throw away private data.
2472 */
2473void unmap_mapping_range(struct address_space *mapping,
2474 loff_t const holebegin, loff_t const holelen, int even_cows)
2475{
2476 struct zap_details details = { };
2477 pgoff_t hba = holebegin >> PAGE_SHIFT;
2478 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2479
2480 /* Check for overflow. */
2481 if (sizeof(holelen) > sizeof(hlen)) {
2482 long long holeend =
2483 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2484 if (holeend & ~(long long)ULONG_MAX)
2485 hlen = ULONG_MAX - hba + 1;
2486 }
2487
2488 details.check_mapping = even_cows? NULL: mapping;
2489 details.first_index = hba;
2490 details.last_index = hba + hlen - 1;
2491 if (details.last_index < details.first_index)
2492 details.last_index = ULONG_MAX;
2493
2494
2495 /* DAX uses i_mmap_lock to serialise file truncate vs page fault */
2496 i_mmap_lock_write(mapping);
2497 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2498 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2499 i_mmap_unlock_write(mapping);
2500}
2501EXPORT_SYMBOL(unmap_mapping_range);
2502
2503/*
2504 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2505 * but allow concurrent faults), and pte mapped but not yet locked.
2506 * We return with pte unmapped and unlocked.
2507 *
2508 * We return with the mmap_sem locked or unlocked in the same cases
2509 * as does filemap_fault().
2510 */
2511static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
2512 unsigned long address, pte_t *page_table, pmd_t *pmd,
2513 unsigned int flags, pte_t orig_pte)
2514{
2515 spinlock_t *ptl;
2516 struct page *page, *swapcache;
2517 struct mem_cgroup *memcg;
2518 swp_entry_t entry;
2519 pte_t pte;
2520 int locked;
2521 int exclusive = 0;
2522 int ret = 0;
2523
2524 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2525 goto out;
2526
2527 entry = pte_to_swp_entry(orig_pte);
2528 if (unlikely(non_swap_entry(entry))) {
2529 if (is_migration_entry(entry)) {
2530 migration_entry_wait(mm, pmd, address);
2531 } else if (is_hwpoison_entry(entry)) {
2532 ret = VM_FAULT_HWPOISON;
2533 } else {
2534 print_bad_pte(vma, address, orig_pte, NULL);
2535 ret = VM_FAULT_SIGBUS;
2536 }
2537 goto out;
2538 }
2539 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2540 page = lookup_swap_cache(entry);
2541 if (!page) {
2542 page = swapin_readahead(entry,
2543 GFP_HIGHUSER_MOVABLE, vma, address);
2544 if (!page) {
2545 /*
2546 * Back out if somebody else faulted in this pte
2547 * while we released the pte lock.
2548 */
2549 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2550 if (likely(pte_same(*page_table, orig_pte)))
2551 ret = VM_FAULT_OOM;
2552 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2553 goto unlock;
2554 }
2555
2556 /* Had to read the page from swap area: Major fault */
2557 ret = VM_FAULT_MAJOR;
2558 count_vm_event(PGMAJFAULT);
2559 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
2560 } else if (PageHWPoison(page)) {
2561 /*
2562 * hwpoisoned dirty swapcache pages are kept for killing
2563 * owner processes (which may be unknown at hwpoison time)
2564 */
2565 ret = VM_FAULT_HWPOISON;
2566 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2567 swapcache = page;
2568 goto out_release;
2569 }
2570
2571 swapcache = page;
2572 locked = lock_page_or_retry(page, mm, flags);
2573
2574 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2575 if (!locked) {
2576 ret |= VM_FAULT_RETRY;
2577 goto out_release;
2578 }
2579
2580 /*
2581 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2582 * release the swapcache from under us. The page pin, and pte_same
2583 * test below, are not enough to exclude that. Even if it is still
2584 * swapcache, we need to check that the page's swap has not changed.
2585 */
2586 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2587 goto out_page;
2588
2589 page = ksm_might_need_to_copy(page, vma, address);
2590 if (unlikely(!page)) {
2591 ret = VM_FAULT_OOM;
2592 page = swapcache;
2593 goto out_page;
2594 }
2595
2596 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false)) {
2597 ret = VM_FAULT_OOM;
2598 goto out_page;
2599 }
2600
2601 /*
2602 * Back out if somebody else already faulted in this pte.
2603 */
2604 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2605 if (unlikely(!pte_same(*page_table, orig_pte)))
2606 goto out_nomap;
2607
2608 if (unlikely(!PageUptodate(page))) {
2609 ret = VM_FAULT_SIGBUS;
2610 goto out_nomap;
2611 }
2612
2613 /*
2614 * The page isn't present yet, go ahead with the fault.
2615 *
2616 * Be careful about the sequence of operations here.
2617 * To get its accounting right, reuse_swap_page() must be called
2618 * while the page is counted on swap but not yet in mapcount i.e.
2619 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2620 * must be called after the swap_free(), or it will never succeed.
2621 */
2622
2623 inc_mm_counter_fast(mm, MM_ANONPAGES);
2624 dec_mm_counter_fast(mm, MM_SWAPENTS);
2625 pte = mk_pte(page, vma->vm_page_prot);
2626 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2627 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2628 flags &= ~FAULT_FLAG_WRITE;
2629 ret |= VM_FAULT_WRITE;
2630 exclusive = RMAP_EXCLUSIVE;
2631 }
2632 flush_icache_page(vma, page);
2633 if (pte_swp_soft_dirty(orig_pte))
2634 pte = pte_mksoft_dirty(pte);
2635 set_pte_at(mm, address, page_table, pte);
2636 if (page == swapcache) {
2637 do_page_add_anon_rmap(page, vma, address, exclusive);
2638 mem_cgroup_commit_charge(page, memcg, true, false);
2639 } else { /* ksm created a completely new copy */
2640 page_add_new_anon_rmap(page, vma, address, false);
2641 mem_cgroup_commit_charge(page, memcg, false, false);
2642 lru_cache_add_active_or_unevictable(page, vma);
2643 }
2644
2645 swap_free(entry);
2646 if (mem_cgroup_swap_full(page) ||
2647 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2648 try_to_free_swap(page);
2649 unlock_page(page);
2650 if (page != swapcache) {
2651 /*
2652 * Hold the lock to avoid the swap entry to be reused
2653 * until we take the PT lock for the pte_same() check
2654 * (to avoid false positives from pte_same). For
2655 * further safety release the lock after the swap_free
2656 * so that the swap count won't change under a
2657 * parallel locked swapcache.
2658 */
2659 unlock_page(swapcache);
2660 put_page(swapcache);
2661 }
2662
2663 if (flags & FAULT_FLAG_WRITE) {
2664 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
2665 if (ret & VM_FAULT_ERROR)
2666 ret &= VM_FAULT_ERROR;
2667 goto out;
2668 }
2669
2670 /* No need to invalidate - it was non-present before */
2671 update_mmu_cache(vma, address, page_table);
2672unlock:
2673 pte_unmap_unlock(page_table, ptl);
2674out:
2675 return ret;
2676out_nomap:
2677 mem_cgroup_cancel_charge(page, memcg, false);
2678 pte_unmap_unlock(page_table, ptl);
2679out_page:
2680 unlock_page(page);
2681out_release:
2682 put_page(page);
2683 if (page != swapcache) {
2684 unlock_page(swapcache);
2685 put_page(swapcache);
2686 }
2687 return ret;
2688}
2689
2690/*
2691 * This is like a special single-page "expand_{down|up}wards()",
2692 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2693 * doesn't hit another vma.
2694 */
2695static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2696{
2697 address &= PAGE_MASK;
2698 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2699 struct vm_area_struct *prev = vma->vm_prev;
2700
2701 /*
2702 * Is there a mapping abutting this one below?
2703 *
2704 * That's only ok if it's the same stack mapping
2705 * that has gotten split..
2706 */
2707 if (prev && prev->vm_end == address)
2708 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2709
2710 return expand_downwards(vma, address - PAGE_SIZE);
2711 }
2712 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2713 struct vm_area_struct *next = vma->vm_next;
2714
2715 /* As VM_GROWSDOWN but s/below/above/ */
2716 if (next && next->vm_start == address + PAGE_SIZE)
2717 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2718
2719 return expand_upwards(vma, address + PAGE_SIZE);
2720 }
2721 return 0;
2722}
2723
2724/*
2725 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2726 * but allow concurrent faults), and pte mapped but not yet locked.
2727 * We return with mmap_sem still held, but pte unmapped and unlocked.
2728 */
2729static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2730 unsigned long address, pte_t *page_table, pmd_t *pmd,
2731 unsigned int flags)
2732{
2733 struct mem_cgroup *memcg;
2734 struct page *page;
2735 spinlock_t *ptl;
2736 pte_t entry;
2737
2738 pte_unmap(page_table);
2739
2740 /* File mapping without ->vm_ops ? */
2741 if (vma->vm_flags & VM_SHARED)
2742 return VM_FAULT_SIGBUS;
2743
2744 /* Check if we need to add a guard page to the stack */
2745 if (check_stack_guard_page(vma, address) < 0)
2746 return VM_FAULT_SIGSEGV;
2747
2748 /* Use the zero-page for reads */
2749 if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) {
2750 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
2751 vma->vm_page_prot));
2752 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2753 if (!pte_none(*page_table))
2754 goto unlock;
2755 /* Deliver the page fault to userland, check inside PT lock */
2756 if (userfaultfd_missing(vma)) {
2757 pte_unmap_unlock(page_table, ptl);
2758 return handle_userfault(vma, address, flags,
2759 VM_UFFD_MISSING);
2760 }
2761 goto setpte;
2762 }
2763
2764 /* Allocate our own private page. */
2765 if (unlikely(anon_vma_prepare(vma)))
2766 goto oom;
2767 page = alloc_zeroed_user_highpage_movable(vma, address);
2768 if (!page)
2769 goto oom;
2770
2771 if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg, false))
2772 goto oom_free_page;
2773
2774 /*
2775 * The memory barrier inside __SetPageUptodate makes sure that
2776 * preceeding stores to the page contents become visible before
2777 * the set_pte_at() write.
2778 */
2779 __SetPageUptodate(page);
2780
2781 entry = mk_pte(page, vma->vm_page_prot);
2782 if (vma->vm_flags & VM_WRITE)
2783 entry = pte_mkwrite(pte_mkdirty(entry));
2784
2785 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2786 if (!pte_none(*page_table))
2787 goto release;
2788
2789 /* Deliver the page fault to userland, check inside PT lock */
2790 if (userfaultfd_missing(vma)) {
2791 pte_unmap_unlock(page_table, ptl);
2792 mem_cgroup_cancel_charge(page, memcg, false);
2793 put_page(page);
2794 return handle_userfault(vma, address, flags,
2795 VM_UFFD_MISSING);
2796 }
2797
2798 inc_mm_counter_fast(mm, MM_ANONPAGES);
2799 page_add_new_anon_rmap(page, vma, address, false);
2800 mem_cgroup_commit_charge(page, memcg, false, false);
2801 lru_cache_add_active_or_unevictable(page, vma);
2802setpte:
2803 set_pte_at(mm, address, page_table, entry);
2804
2805 /* No need to invalidate - it was non-present before */
2806 update_mmu_cache(vma, address, page_table);
2807unlock:
2808 pte_unmap_unlock(page_table, ptl);
2809 return 0;
2810release:
2811 mem_cgroup_cancel_charge(page, memcg, false);
2812 put_page(page);
2813 goto unlock;
2814oom_free_page:
2815 put_page(page);
2816oom:
2817 return VM_FAULT_OOM;
2818}
2819
2820/*
2821 * The mmap_sem must have been held on entry, and may have been
2822 * released depending on flags and vma->vm_ops->fault() return value.
2823 * See filemap_fault() and __lock_page_retry().
2824 */
2825static int __do_fault(struct vm_area_struct *vma, unsigned long address,
2826 pgoff_t pgoff, unsigned int flags,
2827 struct page *cow_page, struct page **page)
2828{
2829 struct vm_fault vmf;
2830 int ret;
2831
2832 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2833 vmf.pgoff = pgoff;
2834 vmf.flags = flags;
2835 vmf.page = NULL;
2836 vmf.gfp_mask = __get_fault_gfp_mask(vma);
2837 vmf.cow_page = cow_page;
2838
2839 ret = vma->vm_ops->fault(vma, &vmf);
2840 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
2841 return ret;
2842 if (!vmf.page)
2843 goto out;
2844
2845 if (unlikely(PageHWPoison(vmf.page))) {
2846 if (ret & VM_FAULT_LOCKED)
2847 unlock_page(vmf.page);
2848 put_page(vmf.page);
2849 return VM_FAULT_HWPOISON;
2850 }
2851
2852 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2853 lock_page(vmf.page);
2854 else
2855 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
2856
2857 out:
2858 *page = vmf.page;
2859 return ret;
2860}
2861
2862/**
2863 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
2864 *
2865 * @vma: virtual memory area
2866 * @address: user virtual address
2867 * @page: page to map
2868 * @pte: pointer to target page table entry
2869 * @write: true, if new entry is writable
2870 * @anon: true, if it's anonymous page
2871 *
2872 * Caller must hold page table lock relevant for @pte.
2873 *
2874 * Target users are page handler itself and implementations of
2875 * vm_ops->map_pages.
2876 */
2877void do_set_pte(struct vm_area_struct *vma, unsigned long address,
2878 struct page *page, pte_t *pte, bool write, bool anon)
2879{
2880 pte_t entry;
2881
2882 flush_icache_page(vma, page);
2883 entry = mk_pte(page, vma->vm_page_prot);
2884 if (write)
2885 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2886 if (anon) {
2887 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2888 page_add_new_anon_rmap(page, vma, address, false);
2889 } else {
2890 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
2891 page_add_file_rmap(page);
2892 }
2893 set_pte_at(vma->vm_mm, address, pte, entry);
2894
2895 /* no need to invalidate: a not-present page won't be cached */
2896 update_mmu_cache(vma, address, pte);
2897}
2898
2899static unsigned long fault_around_bytes __read_mostly =
2900 rounddown_pow_of_two(65536);
2901
2902#ifdef CONFIG_DEBUG_FS
2903static int fault_around_bytes_get(void *data, u64 *val)
2904{
2905 *val = fault_around_bytes;
2906 return 0;
2907}
2908
2909/*
2910 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
2911 * rounded down to nearest page order. It's what do_fault_around() expects to
2912 * see.
2913 */
2914static int fault_around_bytes_set(void *data, u64 val)
2915{
2916 if (val / PAGE_SIZE > PTRS_PER_PTE)
2917 return -EINVAL;
2918 if (val > PAGE_SIZE)
2919 fault_around_bytes = rounddown_pow_of_two(val);
2920 else
2921 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
2922 return 0;
2923}
2924DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
2925 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
2926
2927static int __init fault_around_debugfs(void)
2928{
2929 void *ret;
2930
2931 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
2932 &fault_around_bytes_fops);
2933 if (!ret)
2934 pr_warn("Failed to create fault_around_bytes in debugfs");
2935 return 0;
2936}
2937late_initcall(fault_around_debugfs);
2938#endif
2939
2940/*
2941 * do_fault_around() tries to map few pages around the fault address. The hope
2942 * is that the pages will be needed soon and this will lower the number of
2943 * faults to handle.
2944 *
2945 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
2946 * not ready to be mapped: not up-to-date, locked, etc.
2947 *
2948 * This function is called with the page table lock taken. In the split ptlock
2949 * case the page table lock only protects only those entries which belong to
2950 * the page table corresponding to the fault address.
2951 *
2952 * This function doesn't cross the VMA boundaries, in order to call map_pages()
2953 * only once.
2954 *
2955 * fault_around_pages() defines how many pages we'll try to map.
2956 * do_fault_around() expects it to return a power of two less than or equal to
2957 * PTRS_PER_PTE.
2958 *
2959 * The virtual address of the area that we map is naturally aligned to the
2960 * fault_around_pages() value (and therefore to page order). This way it's
2961 * easier to guarantee that we don't cross page table boundaries.
2962 */
2963static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
2964 pte_t *pte, pgoff_t pgoff, unsigned int flags)
2965{
2966 unsigned long start_addr, nr_pages, mask;
2967 pgoff_t max_pgoff;
2968 struct vm_fault vmf;
2969 int off;
2970
2971 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
2972 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
2973
2974 start_addr = max(address & mask, vma->vm_start);
2975 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
2976 pte -= off;
2977 pgoff -= off;
2978
2979 /*
2980 * max_pgoff is either end of page table or end of vma
2981 * or fault_around_pages() from pgoff, depending what is nearest.
2982 */
2983 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
2984 PTRS_PER_PTE - 1;
2985 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
2986 pgoff + nr_pages - 1);
2987
2988 /* Check if it makes any sense to call ->map_pages */
2989 while (!pte_none(*pte)) {
2990 if (++pgoff > max_pgoff)
2991 return;
2992 start_addr += PAGE_SIZE;
2993 if (start_addr >= vma->vm_end)
2994 return;
2995 pte++;
2996 }
2997
2998 vmf.virtual_address = (void __user *) start_addr;
2999 vmf.pte = pte;
3000 vmf.pgoff = pgoff;
3001 vmf.max_pgoff = max_pgoff;
3002 vmf.flags = flags;
3003 vmf.gfp_mask = __get_fault_gfp_mask(vma);
3004 vma->vm_ops->map_pages(vma, &vmf);
3005}
3006
3007static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3008 unsigned long address, pmd_t *pmd,
3009 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3010{
3011 struct page *fault_page;
3012 spinlock_t *ptl;
3013 pte_t *pte;
3014 int ret = 0;
3015
3016 /*
3017 * Let's call ->map_pages() first and use ->fault() as fallback
3018 * if page by the offset is not ready to be mapped (cold cache or
3019 * something).
3020 */
3021 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3022 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3023 do_fault_around(vma, address, pte, pgoff, flags);
3024 if (!pte_same(*pte, orig_pte))
3025 goto unlock_out;
3026 pte_unmap_unlock(pte, ptl);
3027 }
3028
3029 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3030 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3031 return ret;
3032
3033 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3034 if (unlikely(!pte_same(*pte, orig_pte))) {
3035 pte_unmap_unlock(pte, ptl);
3036 unlock_page(fault_page);
3037 put_page(fault_page);
3038 return ret;
3039 }
3040 do_set_pte(vma, address, fault_page, pte, false, false);
3041 unlock_page(fault_page);
3042unlock_out:
3043 pte_unmap_unlock(pte, ptl);
3044 return ret;
3045}
3046
3047static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3048 unsigned long address, pmd_t *pmd,
3049 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3050{
3051 struct page *fault_page, *new_page;
3052 struct mem_cgroup *memcg;
3053 spinlock_t *ptl;
3054 pte_t *pte;
3055 int ret;
3056
3057 if (unlikely(anon_vma_prepare(vma)))
3058 return VM_FAULT_OOM;
3059
3060 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3061 if (!new_page)
3062 return VM_FAULT_OOM;
3063
3064 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false)) {
3065 put_page(new_page);
3066 return VM_FAULT_OOM;
3067 }
3068
3069 ret = __do_fault(vma, address, pgoff, flags, new_page, &fault_page);
3070 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3071 goto uncharge_out;
3072
3073 if (fault_page)
3074 copy_user_highpage(new_page, fault_page, address, vma);
3075 __SetPageUptodate(new_page);
3076
3077 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3078 if (unlikely(!pte_same(*pte, orig_pte))) {
3079 pte_unmap_unlock(pte, ptl);
3080 if (fault_page) {
3081 unlock_page(fault_page);
3082 put_page(fault_page);
3083 } else {
3084 /*
3085 * The fault handler has no page to lock, so it holds
3086 * i_mmap_lock for read to protect against truncate.
3087 */
3088 i_mmap_unlock_read(vma->vm_file->f_mapping);
3089 }
3090 goto uncharge_out;
3091 }
3092 do_set_pte(vma, address, new_page, pte, true, true);
3093 mem_cgroup_commit_charge(new_page, memcg, false, false);
3094 lru_cache_add_active_or_unevictable(new_page, vma);
3095 pte_unmap_unlock(pte, ptl);
3096 if (fault_page) {
3097 unlock_page(fault_page);
3098 put_page(fault_page);
3099 } else {
3100 /*
3101 * The fault handler has no page to lock, so it holds
3102 * i_mmap_lock for read to protect against truncate.
3103 */
3104 i_mmap_unlock_read(vma->vm_file->f_mapping);
3105 }
3106 return ret;
3107uncharge_out:
3108 mem_cgroup_cancel_charge(new_page, memcg, false);
3109 put_page(new_page);
3110 return ret;
3111}
3112
3113static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3114 unsigned long address, pmd_t *pmd,
3115 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3116{
3117 struct page *fault_page;
3118 struct address_space *mapping;
3119 spinlock_t *ptl;
3120 pte_t *pte;
3121 int dirtied = 0;
3122 int ret, tmp;
3123
3124 ret = __do_fault(vma, address, pgoff, flags, NULL, &fault_page);
3125 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3126 return ret;
3127
3128 /*
3129 * Check if the backing address space wants to know that the page is
3130 * about to become writable
3131 */
3132 if (vma->vm_ops->page_mkwrite) {
3133 unlock_page(fault_page);
3134 tmp = do_page_mkwrite(vma, fault_page, address);
3135 if (unlikely(!tmp ||
3136 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3137 put_page(fault_page);
3138 return tmp;
3139 }
3140 }
3141
3142 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3143 if (unlikely(!pte_same(*pte, orig_pte))) {
3144 pte_unmap_unlock(pte, ptl);
3145 unlock_page(fault_page);
3146 put_page(fault_page);
3147 return ret;
3148 }
3149 do_set_pte(vma, address, fault_page, pte, true, false);
3150 pte_unmap_unlock(pte, ptl);
3151
3152 if (set_page_dirty(fault_page))
3153 dirtied = 1;
3154 /*
3155 * Take a local copy of the address_space - page.mapping may be zeroed
3156 * by truncate after unlock_page(). The address_space itself remains
3157 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
3158 * release semantics to prevent the compiler from undoing this copying.
3159 */
3160 mapping = page_rmapping(fault_page);
3161 unlock_page(fault_page);
3162 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3163 /*
3164 * Some device drivers do not set page.mapping but still
3165 * dirty their pages
3166 */
3167 balance_dirty_pages_ratelimited(mapping);
3168 }
3169
3170 if (!vma->vm_ops->page_mkwrite)
3171 file_update_time(vma->vm_file);
3172
3173 return ret;
3174}
3175
3176/*
3177 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3178 * but allow concurrent faults).
3179 * The mmap_sem may have been released depending on flags and our
3180 * return value. See filemap_fault() and __lock_page_or_retry().
3181 */
3182static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3183 unsigned long address, pte_t *page_table, pmd_t *pmd,
3184 unsigned int flags, pte_t orig_pte)
3185{
3186 pgoff_t pgoff = linear_page_index(vma, address);
3187
3188 pte_unmap(page_table);
3189 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3190 if (!vma->vm_ops->fault)
3191 return VM_FAULT_SIGBUS;
3192 if (!(flags & FAULT_FLAG_WRITE))
3193 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3194 orig_pte);
3195 if (!(vma->vm_flags & VM_SHARED))
3196 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3197 orig_pte);
3198 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3199}
3200
3201static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3202 unsigned long addr, int page_nid,
3203 int *flags)
3204{
3205 get_page(page);
3206
3207 count_vm_numa_event(NUMA_HINT_FAULTS);
3208 if (page_nid == numa_node_id()) {
3209 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3210 *flags |= TNF_FAULT_LOCAL;
3211 }
3212
3213 return mpol_misplaced(page, vma, addr);
3214}
3215
3216static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3217 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3218{
3219 struct page *page = NULL;
3220 spinlock_t *ptl;
3221 int page_nid = -1;
3222 int last_cpupid;
3223 int target_nid;
3224 bool migrated = false;
3225 bool was_writable = pte_write(pte);
3226 int flags = 0;
3227
3228 /* A PROT_NONE fault should not end up here */
3229 BUG_ON(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)));
3230
3231 /*
3232 * The "pte" at this point cannot be used safely without
3233 * validation through pte_unmap_same(). It's of NUMA type but
3234 * the pfn may be screwed if the read is non atomic.
3235 *
3236 * We can safely just do a "set_pte_at()", because the old
3237 * page table entry is not accessible, so there would be no
3238 * concurrent hardware modifications to the PTE.
3239 */
3240 ptl = pte_lockptr(mm, pmd);
3241 spin_lock(ptl);
3242 if (unlikely(!pte_same(*ptep, pte))) {
3243 pte_unmap_unlock(ptep, ptl);
3244 goto out;
3245 }
3246
3247 /* Make it present again */
3248 pte = pte_modify(pte, vma->vm_page_prot);
3249 pte = pte_mkyoung(pte);
3250 if (was_writable)
3251 pte = pte_mkwrite(pte);
3252 set_pte_at(mm, addr, ptep, pte);
3253 update_mmu_cache(vma, addr, ptep);
3254
3255 page = vm_normal_page(vma, addr, pte);
3256 if (!page) {
3257 pte_unmap_unlock(ptep, ptl);
3258 return 0;
3259 }
3260
3261 /* TODO: handle PTE-mapped THP */
3262 if (PageCompound(page)) {
3263 pte_unmap_unlock(ptep, ptl);
3264 return 0;
3265 }
3266
3267 /*
3268 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3269 * much anyway since they can be in shared cache state. This misses
3270 * the case where a mapping is writable but the process never writes
3271 * to it but pte_write gets cleared during protection updates and
3272 * pte_dirty has unpredictable behaviour between PTE scan updates,
3273 * background writeback, dirty balancing and application behaviour.
3274 */
3275 if (!(vma->vm_flags & VM_WRITE))
3276 flags |= TNF_NO_GROUP;
3277
3278 /*
3279 * Flag if the page is shared between multiple address spaces. This
3280 * is later used when determining whether to group tasks together
3281 */
3282 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3283 flags |= TNF_SHARED;
3284
3285 last_cpupid = page_cpupid_last(page);
3286 page_nid = page_to_nid(page);
3287 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3288 pte_unmap_unlock(ptep, ptl);
3289 if (target_nid == -1) {
3290 put_page(page);
3291 goto out;
3292 }
3293
3294 /* Migrate to the requested node */
3295 migrated = migrate_misplaced_page(page, vma, target_nid);
3296 if (migrated) {
3297 page_nid = target_nid;
3298 flags |= TNF_MIGRATED;
3299 } else
3300 flags |= TNF_MIGRATE_FAIL;
3301
3302out:
3303 if (page_nid != -1)
3304 task_numa_fault(last_cpupid, page_nid, 1, flags);
3305 return 0;
3306}
3307
3308static int create_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3309 unsigned long address, pmd_t *pmd, unsigned int flags)
3310{
3311 if (vma_is_anonymous(vma))
3312 return do_huge_pmd_anonymous_page(mm, vma, address, pmd, flags);
3313 if (vma->vm_ops->pmd_fault)
3314 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3315 return VM_FAULT_FALLBACK;
3316}
3317
3318static int wp_huge_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
3319 unsigned long address, pmd_t *pmd, pmd_t orig_pmd,
3320 unsigned int flags)
3321{
3322 if (vma_is_anonymous(vma))
3323 return do_huge_pmd_wp_page(mm, vma, address, pmd, orig_pmd);
3324 if (vma->vm_ops->pmd_fault)
3325 return vma->vm_ops->pmd_fault(vma, address, pmd, flags);
3326 return VM_FAULT_FALLBACK;
3327}
3328
3329/*
3330 * These routines also need to handle stuff like marking pages dirty
3331 * and/or accessed for architectures that don't do it in hardware (most
3332 * RISC architectures). The early dirtying is also good on the i386.
3333 *
3334 * There is also a hook called "update_mmu_cache()" that architectures
3335 * with external mmu caches can use to update those (ie the Sparc or
3336 * PowerPC hashed page tables that act as extended TLBs).
3337 *
3338 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3339 * but allow concurrent faults), and pte mapped but not yet locked.
3340 * We return with pte unmapped and unlocked.
3341 *
3342 * The mmap_sem may have been released depending on flags and our
3343 * return value. See filemap_fault() and __lock_page_or_retry().
3344 */
3345static int handle_pte_fault(struct mm_struct *mm,
3346 struct vm_area_struct *vma, unsigned long address,
3347 pte_t *pte, pmd_t *pmd, unsigned int flags)
3348{
3349 pte_t entry;
3350 spinlock_t *ptl;
3351
3352 /*
3353 * some architectures can have larger ptes than wordsize,
3354 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y,
3355 * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses.
3356 * The code below just needs a consistent view for the ifs and
3357 * we later double check anyway with the ptl lock held. So here
3358 * a barrier will do.
3359 */
3360 entry = *pte;
3361 barrier();
3362 if (!pte_present(entry)) {
3363 if (pte_none(entry)) {
3364 if (vma_is_anonymous(vma))
3365 return do_anonymous_page(mm, vma, address,
3366 pte, pmd, flags);
3367 else
3368 return do_fault(mm, vma, address, pte, pmd,
3369 flags, entry);
3370 }
3371 return do_swap_page(mm, vma, address,
3372 pte, pmd, flags, entry);
3373 }
3374
3375 if (pte_protnone(entry))
3376 return do_numa_page(mm, vma, address, entry, pte, pmd);
3377
3378 ptl = pte_lockptr(mm, pmd);
3379 spin_lock(ptl);
3380 if (unlikely(!pte_same(*pte, entry)))
3381 goto unlock;
3382 if (flags & FAULT_FLAG_WRITE) {
3383 if (!pte_write(entry))
3384 return do_wp_page(mm, vma, address,
3385 pte, pmd, ptl, entry);
3386 entry = pte_mkdirty(entry);
3387 }
3388 entry = pte_mkyoung(entry);
3389 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3390 update_mmu_cache(vma, address, pte);
3391 } else {
3392 /*
3393 * This is needed only for protection faults but the arch code
3394 * is not yet telling us if this is a protection fault or not.
3395 * This still avoids useless tlb flushes for .text page faults
3396 * with threads.
3397 */
3398 if (flags & FAULT_FLAG_WRITE)
3399 flush_tlb_fix_spurious_fault(vma, address);
3400 }
3401unlock:
3402 pte_unmap_unlock(pte, ptl);
3403 return 0;
3404}
3405
3406/*
3407 * By the time we get here, we already hold the mm semaphore
3408 *
3409 * The mmap_sem may have been released depending on flags and our
3410 * return value. See filemap_fault() and __lock_page_or_retry().
3411 */
3412static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3413 unsigned long address, unsigned int flags)
3414{
3415 pgd_t *pgd;
3416 pud_t *pud;
3417 pmd_t *pmd;
3418 pte_t *pte;
3419
3420 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3421 flags & FAULT_FLAG_INSTRUCTION,
3422 flags & FAULT_FLAG_REMOTE))
3423 return VM_FAULT_SIGSEGV;
3424
3425 if (unlikely(is_vm_hugetlb_page(vma)))
3426 return hugetlb_fault(mm, vma, address, flags);
3427
3428 pgd = pgd_offset(mm, address);
3429 pud = pud_alloc(mm, pgd, address);
3430 if (!pud)
3431 return VM_FAULT_OOM;
3432 pmd = pmd_alloc(mm, pud, address);
3433 if (!pmd)
3434 return VM_FAULT_OOM;
3435 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3436 int ret = create_huge_pmd(mm, vma, address, pmd, flags);
3437 if (!(ret & VM_FAULT_FALLBACK))
3438 return ret;
3439 } else {
3440 pmd_t orig_pmd = *pmd;
3441 int ret;
3442
3443 barrier();
3444 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3445 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3446
3447 if (pmd_protnone(orig_pmd))
3448 return do_huge_pmd_numa_page(mm, vma, address,
3449 orig_pmd, pmd);
3450
3451 if (dirty && !pmd_write(orig_pmd)) {
3452 ret = wp_huge_pmd(mm, vma, address, pmd,
3453 orig_pmd, flags);
3454 if (!(ret & VM_FAULT_FALLBACK))
3455 return ret;
3456 } else {
3457 huge_pmd_set_accessed(mm, vma, address, pmd,
3458 orig_pmd, dirty);
3459 return 0;
3460 }
3461 }
3462 }
3463
3464 /*
3465 * Use pte_alloc() instead of pte_alloc_map, because we can't
3466 * run pte_offset_map on the pmd, if an huge pmd could
3467 * materialize from under us from a different thread.
3468 */
3469 if (unlikely(pte_alloc(mm, pmd, address)))
3470 return VM_FAULT_OOM;
3471 /*
3472 * If a huge pmd materialized under us just retry later. Use
3473 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
3474 * didn't become pmd_trans_huge under us and then back to pmd_none, as
3475 * a result of MADV_DONTNEED running immediately after a huge pmd fault
3476 * in a different thread of this mm, in turn leading to a misleading
3477 * pmd_trans_huge() retval. All we have to ensure is that it is a
3478 * regular pmd that we can walk with pte_offset_map() and we can do that
3479 * through an atomic read in C, which is what pmd_trans_unstable()
3480 * provides.
3481 */
3482 if (unlikely(pmd_trans_unstable(pmd) || pmd_devmap(*pmd)))
3483 return 0;
3484 /*
3485 * A regular pmd is established and it can't morph into a huge pmd
3486 * from under us anymore at this point because we hold the mmap_sem
3487 * read mode and khugepaged takes it in write mode. So now it's
3488 * safe to run pte_offset_map().
3489 */
3490 pte = pte_offset_map(pmd, address);
3491
3492 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3493}
3494
3495/*
3496 * By the time we get here, we already hold the mm semaphore
3497 *
3498 * The mmap_sem may have been released depending on flags and our
3499 * return value. See filemap_fault() and __lock_page_or_retry().
3500 */
3501int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3502 unsigned long address, unsigned int flags)
3503{
3504 int ret;
3505
3506 __set_current_state(TASK_RUNNING);
3507
3508 count_vm_event(PGFAULT);
3509 mem_cgroup_count_vm_event(mm, PGFAULT);
3510
3511 /* do counter updates before entering really critical section. */
3512 check_sync_rss_stat(current);
3513
3514 /*
3515 * Enable the memcg OOM handling for faults triggered in user
3516 * space. Kernel faults are handled more gracefully.
3517 */
3518 if (flags & FAULT_FLAG_USER)
3519 mem_cgroup_oom_enable();
3520
3521 ret = __handle_mm_fault(mm, vma, address, flags);
3522
3523 if (flags & FAULT_FLAG_USER) {
3524 mem_cgroup_oom_disable();
3525 /*
3526 * The task may have entered a memcg OOM situation but
3527 * if the allocation error was handled gracefully (no
3528 * VM_FAULT_OOM), there is no need to kill anything.
3529 * Just clean up the OOM state peacefully.
3530 */
3531 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3532 mem_cgroup_oom_synchronize(false);
3533 }
3534
3535 return ret;
3536}
3537EXPORT_SYMBOL_GPL(handle_mm_fault);
3538
3539#ifndef __PAGETABLE_PUD_FOLDED
3540/*
3541 * Allocate page upper directory.
3542 * We've already handled the fast-path in-line.
3543 */
3544int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3545{
3546 pud_t *new = pud_alloc_one(mm, address);
3547 if (!new)
3548 return -ENOMEM;
3549
3550 smp_wmb(); /* See comment in __pte_alloc */
3551
3552 spin_lock(&mm->page_table_lock);
3553 if (pgd_present(*pgd)) /* Another has populated it */
3554 pud_free(mm, new);
3555 else
3556 pgd_populate(mm, pgd, new);
3557 spin_unlock(&mm->page_table_lock);
3558 return 0;
3559}
3560#endif /* __PAGETABLE_PUD_FOLDED */
3561
3562#ifndef __PAGETABLE_PMD_FOLDED
3563/*
3564 * Allocate page middle directory.
3565 * We've already handled the fast-path in-line.
3566 */
3567int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3568{
3569 pmd_t *new = pmd_alloc_one(mm, address);
3570 if (!new)
3571 return -ENOMEM;
3572
3573 smp_wmb(); /* See comment in __pte_alloc */
3574
3575 spin_lock(&mm->page_table_lock);
3576#ifndef __ARCH_HAS_4LEVEL_HACK
3577 if (!pud_present(*pud)) {
3578 mm_inc_nr_pmds(mm);
3579 pud_populate(mm, pud, new);
3580 } else /* Another has populated it */
3581 pmd_free(mm, new);
3582#else
3583 if (!pgd_present(*pud)) {
3584 mm_inc_nr_pmds(mm);
3585 pgd_populate(mm, pud, new);
3586 } else /* Another has populated it */
3587 pmd_free(mm, new);
3588#endif /* __ARCH_HAS_4LEVEL_HACK */
3589 spin_unlock(&mm->page_table_lock);
3590 return 0;
3591}
3592#endif /* __PAGETABLE_PMD_FOLDED */
3593
3594static int __follow_pte(struct mm_struct *mm, unsigned long address,
3595 pte_t **ptepp, spinlock_t **ptlp)
3596{
3597 pgd_t *pgd;
3598 pud_t *pud;
3599 pmd_t *pmd;
3600 pte_t *ptep;
3601
3602 pgd = pgd_offset(mm, address);
3603 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3604 goto out;
3605
3606 pud = pud_offset(pgd, address);
3607 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3608 goto out;
3609
3610 pmd = pmd_offset(pud, address);
3611 VM_BUG_ON(pmd_trans_huge(*pmd));
3612 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3613 goto out;
3614
3615 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
3616 if (pmd_huge(*pmd))
3617 goto out;
3618
3619 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3620 if (!ptep)
3621 goto out;
3622 if (!pte_present(*ptep))
3623 goto unlock;
3624 *ptepp = ptep;
3625 return 0;
3626unlock:
3627 pte_unmap_unlock(ptep, *ptlp);
3628out:
3629 return -EINVAL;
3630}
3631
3632static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3633 pte_t **ptepp, spinlock_t **ptlp)
3634{
3635 int res;
3636
3637 /* (void) is needed to make gcc happy */
3638 (void) __cond_lock(*ptlp,
3639 !(res = __follow_pte(mm, address, ptepp, ptlp)));
3640 return res;
3641}
3642
3643/**
3644 * follow_pfn - look up PFN at a user virtual address
3645 * @vma: memory mapping
3646 * @address: user virtual address
3647 * @pfn: location to store found PFN
3648 *
3649 * Only IO mappings and raw PFN mappings are allowed.
3650 *
3651 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3652 */
3653int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3654 unsigned long *pfn)
3655{
3656 int ret = -EINVAL;
3657 spinlock_t *ptl;
3658 pte_t *ptep;
3659
3660 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3661 return ret;
3662
3663 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3664 if (ret)
3665 return ret;
3666 *pfn = pte_pfn(*ptep);
3667 pte_unmap_unlock(ptep, ptl);
3668 return 0;
3669}
3670EXPORT_SYMBOL(follow_pfn);
3671
3672#ifdef CONFIG_HAVE_IOREMAP_PROT
3673int follow_phys(struct vm_area_struct *vma,
3674 unsigned long address, unsigned int flags,
3675 unsigned long *prot, resource_size_t *phys)
3676{
3677 int ret = -EINVAL;
3678 pte_t *ptep, pte;
3679 spinlock_t *ptl;
3680
3681 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3682 goto out;
3683
3684 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3685 goto out;
3686 pte = *ptep;
3687
3688 if ((flags & FOLL_WRITE) && !pte_write(pte))
3689 goto unlock;
3690
3691 *prot = pgprot_val(pte_pgprot(pte));
3692 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3693
3694 ret = 0;
3695unlock:
3696 pte_unmap_unlock(ptep, ptl);
3697out:
3698 return ret;
3699}
3700
3701int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3702 void *buf, int len, int write)
3703{
3704 resource_size_t phys_addr;
3705 unsigned long prot = 0;
3706 void __iomem *maddr;
3707 int offset = addr & (PAGE_SIZE-1);
3708
3709 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3710 return -EINVAL;
3711
3712 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3713 if (write)
3714 memcpy_toio(maddr + offset, buf, len);
3715 else
3716 memcpy_fromio(buf, maddr + offset, len);
3717 iounmap(maddr);
3718
3719 return len;
3720}
3721EXPORT_SYMBOL_GPL(generic_access_phys);
3722#endif
3723
3724/*
3725 * Access another process' address space as given in mm. If non-NULL, use the
3726 * given task for page fault accounting.
3727 */
3728static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3729 unsigned long addr, void *buf, int len, int write)
3730{
3731 struct vm_area_struct *vma;
3732 void *old_buf = buf;
3733
3734 down_read(&mm->mmap_sem);
3735 /* ignore errors, just check how much was successfully transferred */
3736 while (len) {
3737 int bytes, ret, offset;
3738 void *maddr;
3739 struct page *page = NULL;
3740
3741 ret = get_user_pages_remote(tsk, mm, addr, 1,
3742 write, 1, &page, &vma);
3743 if (ret <= 0) {
3744#ifndef CONFIG_HAVE_IOREMAP_PROT
3745 break;
3746#else
3747 /*
3748 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3749 * we can access using slightly different code.
3750 */
3751 vma = find_vma(mm, addr);
3752 if (!vma || vma->vm_start > addr)
3753 break;
3754 if (vma->vm_ops && vma->vm_ops->access)
3755 ret = vma->vm_ops->access(vma, addr, buf,
3756 len, write);
3757 if (ret <= 0)
3758 break;
3759 bytes = ret;
3760#endif
3761 } else {
3762 bytes = len;
3763 offset = addr & (PAGE_SIZE-1);
3764 if (bytes > PAGE_SIZE-offset)
3765 bytes = PAGE_SIZE-offset;
3766
3767 maddr = kmap(page);
3768 if (write) {
3769 copy_to_user_page(vma, page, addr,
3770 maddr + offset, buf, bytes);
3771 set_page_dirty_lock(page);
3772 } else {
3773 copy_from_user_page(vma, page, addr,
3774 buf, maddr + offset, bytes);
3775 }
3776 kunmap(page);
3777 put_page(page);
3778 }
3779 len -= bytes;
3780 buf += bytes;
3781 addr += bytes;
3782 }
3783 up_read(&mm->mmap_sem);
3784
3785 return buf - old_buf;
3786}
3787
3788/**
3789 * access_remote_vm - access another process' address space
3790 * @mm: the mm_struct of the target address space
3791 * @addr: start address to access
3792 * @buf: source or destination buffer
3793 * @len: number of bytes to transfer
3794 * @write: whether the access is a write
3795 *
3796 * The caller must hold a reference on @mm.
3797 */
3798int access_remote_vm(struct mm_struct *mm, unsigned long addr,
3799 void *buf, int len, int write)
3800{
3801 return __access_remote_vm(NULL, mm, addr, buf, len, write);
3802}
3803
3804/*
3805 * Access another process' address space.
3806 * Source/target buffer must be kernel space,
3807 * Do not walk the page table directly, use get_user_pages
3808 */
3809int access_process_vm(struct task_struct *tsk, unsigned long addr,
3810 void *buf, int len, int write)
3811{
3812 struct mm_struct *mm;
3813 int ret;
3814
3815 mm = get_task_mm(tsk);
3816 if (!mm)
3817 return 0;
3818
3819 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
3820 mmput(mm);
3821
3822 return ret;
3823}
3824
3825/*
3826 * Print the name of a VMA.
3827 */
3828void print_vma_addr(char *prefix, unsigned long ip)
3829{
3830 struct mm_struct *mm = current->mm;
3831 struct vm_area_struct *vma;
3832
3833 /*
3834 * Do not print if we are in atomic
3835 * contexts (in exception stacks, etc.):
3836 */
3837 if (preempt_count())
3838 return;
3839
3840 down_read(&mm->mmap_sem);
3841 vma = find_vma(mm, ip);
3842 if (vma && vma->vm_file) {
3843 struct file *f = vma->vm_file;
3844 char *buf = (char *)__get_free_page(GFP_KERNEL);
3845 if (buf) {
3846 char *p;
3847
3848 p = file_path(f, buf, PAGE_SIZE);
3849 if (IS_ERR(p))
3850 p = "?";
3851 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
3852 vma->vm_start,
3853 vma->vm_end - vma->vm_start);
3854 free_page((unsigned long)buf);
3855 }
3856 }
3857 up_read(&mm->mmap_sem);
3858}
3859
3860#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3861void __might_fault(const char *file, int line)
3862{
3863 /*
3864 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
3865 * holding the mmap_sem, this is safe because kernel memory doesn't
3866 * get paged out, therefore we'll never actually fault, and the
3867 * below annotations will generate false positives.
3868 */
3869 if (segment_eq(get_fs(), KERNEL_DS))
3870 return;
3871 if (pagefault_disabled())
3872 return;
3873 __might_sleep(file, line, 0);
3874#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
3875 if (current->mm)
3876 might_lock_read(¤t->mm->mmap_sem);
3877#endif
3878}
3879EXPORT_SYMBOL(__might_fault);
3880#endif
3881
3882#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
3883static void clear_gigantic_page(struct page *page,
3884 unsigned long addr,
3885 unsigned int pages_per_huge_page)
3886{
3887 int i;
3888 struct page *p = page;
3889
3890 might_sleep();
3891 for (i = 0; i < pages_per_huge_page;
3892 i++, p = mem_map_next(p, page, i)) {
3893 cond_resched();
3894 clear_user_highpage(p, addr + i * PAGE_SIZE);
3895 }
3896}
3897void clear_huge_page(struct page *page,
3898 unsigned long addr, unsigned int pages_per_huge_page)
3899{
3900 int i;
3901
3902 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3903 clear_gigantic_page(page, addr, pages_per_huge_page);
3904 return;
3905 }
3906
3907 might_sleep();
3908 for (i = 0; i < pages_per_huge_page; i++) {
3909 cond_resched();
3910 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
3911 }
3912}
3913
3914static void copy_user_gigantic_page(struct page *dst, struct page *src,
3915 unsigned long addr,
3916 struct vm_area_struct *vma,
3917 unsigned int pages_per_huge_page)
3918{
3919 int i;
3920 struct page *dst_base = dst;
3921 struct page *src_base = src;
3922
3923 for (i = 0; i < pages_per_huge_page; ) {
3924 cond_resched();
3925 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
3926
3927 i++;
3928 dst = mem_map_next(dst, dst_base, i);
3929 src = mem_map_next(src, src_base, i);
3930 }
3931}
3932
3933void copy_user_huge_page(struct page *dst, struct page *src,
3934 unsigned long addr, struct vm_area_struct *vma,
3935 unsigned int pages_per_huge_page)
3936{
3937 int i;
3938
3939 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
3940 copy_user_gigantic_page(dst, src, addr, vma,
3941 pages_per_huge_page);
3942 return;
3943 }
3944
3945 might_sleep();
3946 for (i = 0; i < pages_per_huge_page; i++) {
3947 cond_resched();
3948 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
3949 }
3950}
3951#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
3952
3953#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
3954
3955static struct kmem_cache *page_ptl_cachep;
3956
3957void __init ptlock_cache_init(void)
3958{
3959 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
3960 SLAB_PANIC, NULL);
3961}
3962
3963bool ptlock_alloc(struct page *page)
3964{
3965 spinlock_t *ptl;
3966
3967 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
3968 if (!ptl)
3969 return false;
3970 page->ptl = ptl;
3971 return true;
3972}
3973
3974void ptlock_free(struct page *page)
3975{
3976 kmem_cache_free(page_ptl_cachep, page->ptl);
3977}
3978#endif
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 */
84unsigned long max_mapnr;
85struct page *mem_map;
86
87EXPORT_SYMBOL(max_mapnr);
88EXPORT_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 */
98void * high_memory;
99
100EXPORT_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 */
108int randomize_va_space __read_mostly =
109#ifdef CONFIG_COMPAT_BRK
110 1;
111#else
112 2;
113#endif
114
115static int __init disable_randmaps(char *s)
116{
117 randomize_va_space = 0;
118 return 1;
119}
120__setup("norandmaps", disable_randmaps);
121
122unsigned long zero_pfn __read_mostly;
123unsigned long highest_memmap_pfn __read_mostly;
124
125EXPORT_SYMBOL(zero_pfn);
126
127/*
128 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
129 */
130static int __init init_zero_pfn(void)
131{
132 zero_pfn = page_to_pfn(ZERO_PAGE(0));
133 return 0;
134}
135core_initcall(init_zero_pfn);
136
137
138#if defined(SPLIT_RSS_COUNTING)
139
140void 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
153static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
154{
155 struct task_struct *task = current;
156
157 if (likely(task->mm == mm))
158 task->rss_stat.count[member] += val;
159 else
160 add_mm_counter(mm, member, val);
161}
162#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
163#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
164
165/* sync counter once per 64 page faults */
166#define TASK_RSS_EVENTS_THRESH (64)
167static void check_sync_rss_stat(struct task_struct *task)
168{
169 if (unlikely(task != current))
170 return;
171 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
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
179static 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
187static 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 */
220void 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
241static 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
254static 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
265void 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 */
275void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
276{
277 struct mmu_gather_batch *batch, *next;
278
279 tlb_flush_mmu(tlb);
280
281 /* keep the page table cache within bounds */
282 check_pgt_cache();
283
284 for (batch = tlb->local.next; batch; batch = next) {
285 next = batch->next;
286 free_pages((unsigned long)batch, 0);
287 }
288 tlb->local.next = NULL;
289}
290
291/* __tlb_remove_page
292 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
293 * handling the additional races in SMP caused by other CPUs caching valid
294 * mappings in their TLBs. Returns the number of free page slots left.
295 * When out of page slots we must call tlb_flush_mmu().
296 *returns true if the caller should flush.
297 */
298bool __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
329static void tlb_remove_table_smp_sync(void *arg)
330{
331 /* Simply deliver the interrupt */
332}
333
334static 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
347static 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
360void 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
370void 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 */
402static 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
411static 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
445static 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 */
481void 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
543void 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
578int __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
612int __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
631static inline void init_rss_vec(int *rss)
632{
633 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
634}
635
636static 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 */
654static 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
754struct 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;
790check_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 */
800out:
801 return pfn_to_page(pfn);
802}
803
804#ifdef CONFIG_TRANSPARENT_HUGEPAGE
805struct 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 */
839out:
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
850static inline unsigned long
851copy_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
921out_set_pte:
922 set_pte_at(dst_mm, addr, dst_pte, pte);
923 return 0;
924}
925
926static 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
937again:
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
989static 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
1022static 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
1044int 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
1111static 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);
1125again:
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 /*
1159 * oom_reaper cannot tear down dirty
1160 * pages
1161 */
1162 if (unlikely(details && details->ignore_dirty))
1163 continue;
1164 force_flush = 1;
1165 set_page_dirty(page);
1166 }
1167 if (pte_young(ptent) &&
1168 likely(!(vma->vm_flags & VM_SEQ_READ)))
1169 mark_page_accessed(page);
1170 }
1171 rss[mm_counter(page)]--;
1172 page_remove_rmap(page, false);
1173 if (unlikely(page_mapcount(page) < 0))
1174 print_bad_pte(vma, addr, ptent, page);
1175 if (unlikely(__tlb_remove_page(tlb, page))) {
1176 force_flush = 1;
1177 addr += PAGE_SIZE;
1178 break;
1179 }
1180 continue;
1181 }
1182 /* only check swap_entries if explicitly asked for in details */
1183 if (unlikely(details && !details->check_swap_entries))
1184 continue;
1185
1186 entry = pte_to_swp_entry(ptent);
1187 if (!non_swap_entry(entry))
1188 rss[MM_SWAPENTS]--;
1189 else if (is_migration_entry(entry)) {
1190 struct page *page;
1191
1192 page = migration_entry_to_page(entry);
1193 rss[mm_counter(page)]--;
1194 }
1195 if (unlikely(!free_swap_and_cache(entry)))
1196 print_bad_pte(vma, addr, ptent, NULL);
1197 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1198 } while (pte++, addr += PAGE_SIZE, addr != end);
1199
1200 add_mm_rss_vec(mm, rss);
1201 arch_leave_lazy_mmu_mode();
1202
1203 /* Do the actual TLB flush before dropping ptl */
1204 if (force_flush)
1205 tlb_flush_mmu_tlbonly(tlb);
1206 pte_unmap_unlock(start_pte, ptl);
1207
1208 /*
1209 * If we forced a TLB flush (either due to running out of
1210 * batch buffers or because we needed to flush dirty TLB
1211 * entries before releasing the ptl), free the batched
1212 * memory too. Restart if we didn't do everything.
1213 */
1214 if (force_flush) {
1215 force_flush = 0;
1216 tlb_flush_mmu_free(tlb);
1217 if (addr != end)
1218 goto again;
1219 }
1220
1221 return addr;
1222}
1223
1224static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1225 struct vm_area_struct *vma, pud_t *pud,
1226 unsigned long addr, unsigned long end,
1227 struct zap_details *details)
1228{
1229 pmd_t *pmd;
1230 unsigned long next;
1231
1232 pmd = pmd_offset(pud, addr);
1233 do {
1234 next = pmd_addr_end(addr, end);
1235 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1238 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1239 __split_huge_pmd(vma, pmd, addr, false, NULL);
1240 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1241 goto next;
1242 /* fall through */
1243 }
1244 /*
1245 * Here there can be other concurrent MADV_DONTNEED or
1246 * trans huge page faults running, and if the pmd is
1247 * none or trans huge it can change under us. This is
1248 * because MADV_DONTNEED holds the mmap_sem in read
1249 * mode.
1250 */
1251 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1252 goto next;
1253 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1254next:
1255 cond_resched();
1256 } while (pmd++, addr = next, addr != end);
1257
1258 return addr;
1259}
1260
1261static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1262 struct vm_area_struct *vma, pgd_t *pgd,
1263 unsigned long addr, unsigned long end,
1264 struct zap_details *details)
1265{
1266 pud_t *pud;
1267 unsigned long next;
1268
1269 pud = pud_offset(pgd, addr);
1270 do {
1271 next = pud_addr_end(addr, end);
1272 if (pud_none_or_clear_bad(pud))
1273 continue;
1274 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1275 } while (pud++, addr = next, addr != end);
1276
1277 return addr;
1278}
1279
1280void unmap_page_range(struct mmu_gather *tlb,
1281 struct vm_area_struct *vma,
1282 unsigned long addr, unsigned long end,
1283 struct zap_details *details)
1284{
1285 pgd_t *pgd;
1286 unsigned long next;
1287
1288 BUG_ON(addr >= end);
1289 tlb_start_vma(tlb, vma);
1290 pgd = pgd_offset(vma->vm_mm, addr);
1291 do {
1292 next = pgd_addr_end(addr, end);
1293 if (pgd_none_or_clear_bad(pgd))
1294 continue;
1295 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1296 } while (pgd++, addr = next, addr != end);
1297 tlb_end_vma(tlb, vma);
1298}
1299
1300
1301static void unmap_single_vma(struct mmu_gather *tlb,
1302 struct vm_area_struct *vma, unsigned long start_addr,
1303 unsigned long end_addr,
1304 struct zap_details *details)
1305{
1306 unsigned long start = max(vma->vm_start, start_addr);
1307 unsigned long end;
1308
1309 if (start >= vma->vm_end)
1310 return;
1311 end = min(vma->vm_end, end_addr);
1312 if (end <= vma->vm_start)
1313 return;
1314
1315 if (vma->vm_file)
1316 uprobe_munmap(vma, start, end);
1317
1318 if (unlikely(vma->vm_flags & VM_PFNMAP))
1319 untrack_pfn(vma, 0, 0);
1320
1321 if (start != end) {
1322 if (unlikely(is_vm_hugetlb_page(vma))) {
1323 /*
1324 * It is undesirable to test vma->vm_file as it
1325 * should be non-null for valid hugetlb area.
1326 * However, vm_file will be NULL in the error
1327 * cleanup path of mmap_region. When
1328 * hugetlbfs ->mmap method fails,
1329 * mmap_region() nullifies vma->vm_file
1330 * before calling this function to clean up.
1331 * Since no pte has actually been setup, it is
1332 * safe to do nothing in this case.
1333 */
1334 if (vma->vm_file) {
1335 i_mmap_lock_write(vma->vm_file->f_mapping);
1336 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1337 i_mmap_unlock_write(vma->vm_file->f_mapping);
1338 }
1339 } else
1340 unmap_page_range(tlb, vma, start, end, details);
1341 }
1342}
1343
1344/**
1345 * unmap_vmas - unmap a range of memory covered by a list of vma's
1346 * @tlb: address of the caller's struct mmu_gather
1347 * @vma: the starting vma
1348 * @start_addr: virtual address at which to start unmapping
1349 * @end_addr: virtual address at which to end unmapping
1350 *
1351 * Unmap all pages in the vma list.
1352 *
1353 * Only addresses between `start' and `end' will be unmapped.
1354 *
1355 * The VMA list must be sorted in ascending virtual address order.
1356 *
1357 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1358 * range after unmap_vmas() returns. So the only responsibility here is to
1359 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1360 * drops the lock and schedules.
1361 */
1362void unmap_vmas(struct mmu_gather *tlb,
1363 struct vm_area_struct *vma, unsigned long start_addr,
1364 unsigned long end_addr)
1365{
1366 struct mm_struct *mm = vma->vm_mm;
1367
1368 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1369 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1370 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1371 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1372}
1373
1374/**
1375 * zap_page_range - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @start: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of shared cache invalidation
1380 *
1381 * Caller must protect the VMA list
1382 */
1383void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1384 unsigned long size, struct zap_details *details)
1385{
1386 struct mm_struct *mm = vma->vm_mm;
1387 struct mmu_gather tlb;
1388 unsigned long end = start + size;
1389
1390 lru_add_drain();
1391 tlb_gather_mmu(&tlb, mm, start, end);
1392 update_hiwater_rss(mm);
1393 mmu_notifier_invalidate_range_start(mm, start, end);
1394 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1395 unmap_single_vma(&tlb, vma, start, end, details);
1396 mmu_notifier_invalidate_range_end(mm, start, end);
1397 tlb_finish_mmu(&tlb, start, end);
1398}
1399
1400/**
1401 * zap_page_range_single - remove user pages in a given range
1402 * @vma: vm_area_struct holding the applicable pages
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 * @details: details of shared cache invalidation
1406 *
1407 * The range must fit into one VMA.
1408 */
1409static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1410 unsigned long size, struct zap_details *details)
1411{
1412 struct mm_struct *mm = vma->vm_mm;
1413 struct mmu_gather tlb;
1414 unsigned long end = address + size;
1415
1416 lru_add_drain();
1417 tlb_gather_mmu(&tlb, mm, address, end);
1418 update_hiwater_rss(mm);
1419 mmu_notifier_invalidate_range_start(mm, address, end);
1420 unmap_single_vma(&tlb, vma, address, end, details);
1421 mmu_notifier_invalidate_range_end(mm, address, end);
1422 tlb_finish_mmu(&tlb, address, end);
1423}
1424
1425/**
1426 * zap_vma_ptes - remove ptes mapping the vma
1427 * @vma: vm_area_struct holding ptes to be zapped
1428 * @address: starting address of pages to zap
1429 * @size: number of bytes to zap
1430 *
1431 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1432 *
1433 * The entire address range must be fully contained within the vma.
1434 *
1435 * Returns 0 if successful.
1436 */
1437int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1438 unsigned long size)
1439{
1440 if (address < vma->vm_start || address + size > vma->vm_end ||
1441 !(vma->vm_flags & VM_PFNMAP))
1442 return -1;
1443 zap_page_range_single(vma, address, size, NULL);
1444 return 0;
1445}
1446EXPORT_SYMBOL_GPL(zap_vma_ptes);
1447
1448pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1449 spinlock_t **ptl)
1450{
1451 pgd_t * pgd = pgd_offset(mm, addr);
1452 pud_t * pud = pud_alloc(mm, pgd, addr);
1453 if (pud) {
1454 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1455 if (pmd) {
1456 VM_BUG_ON(pmd_trans_huge(*pmd));
1457 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1458 }
1459 }
1460 return NULL;
1461}
1462
1463/*
1464 * This is the old fallback for page remapping.
1465 *
1466 * For historical reasons, it only allows reserved pages. Only
1467 * old drivers should use this, and they needed to mark their
1468 * pages reserved for the old functions anyway.
1469 */
1470static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1471 struct page *page, pgprot_t prot)
1472{
1473 struct mm_struct *mm = vma->vm_mm;
1474 int retval;
1475 pte_t *pte;
1476 spinlock_t *ptl;
1477
1478 retval = -EINVAL;
1479 if (PageAnon(page))
1480 goto out;
1481 retval = -ENOMEM;
1482 flush_dcache_page(page);
1483 pte = get_locked_pte(mm, addr, &ptl);
1484 if (!pte)
1485 goto out;
1486 retval = -EBUSY;
1487 if (!pte_none(*pte))
1488 goto out_unlock;
1489
1490 /* Ok, finally just insert the thing.. */
1491 get_page(page);
1492 inc_mm_counter_fast(mm, mm_counter_file(page));
1493 page_add_file_rmap(page, false);
1494 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1495
1496 retval = 0;
1497 pte_unmap_unlock(pte, ptl);
1498 return retval;
1499out_unlock:
1500 pte_unmap_unlock(pte, ptl);
1501out:
1502 return retval;
1503}
1504
1505/**
1506 * vm_insert_page - insert single page into user vma
1507 * @vma: user vma to map to
1508 * @addr: target user address of this page
1509 * @page: source kernel page
1510 *
1511 * This allows drivers to insert individual pages they've allocated
1512 * into a user vma.
1513 *
1514 * The page has to be a nice clean _individual_ kernel allocation.
1515 * If you allocate a compound page, you need to have marked it as
1516 * such (__GFP_COMP), or manually just split the page up yourself
1517 * (see split_page()).
1518 *
1519 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1520 * took an arbitrary page protection parameter. This doesn't allow
1521 * that. Your vma protection will have to be set up correctly, which
1522 * means that if you want a shared writable mapping, you'd better
1523 * ask for a shared writable mapping!
1524 *
1525 * The page does not need to be reserved.
1526 *
1527 * Usually this function is called from f_op->mmap() handler
1528 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1529 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1530 * function from other places, for example from page-fault handler.
1531 */
1532int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1533 struct page *page)
1534{
1535 if (addr < vma->vm_start || addr >= vma->vm_end)
1536 return -EFAULT;
1537 if (!page_count(page))
1538 return -EINVAL;
1539 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1540 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1541 BUG_ON(vma->vm_flags & VM_PFNMAP);
1542 vma->vm_flags |= VM_MIXEDMAP;
1543 }
1544 return insert_page(vma, addr, page, vma->vm_page_prot);
1545}
1546EXPORT_SYMBOL(vm_insert_page);
1547
1548static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1549 pfn_t pfn, pgprot_t prot)
1550{
1551 struct mm_struct *mm = vma->vm_mm;
1552 int retval;
1553 pte_t *pte, entry;
1554 spinlock_t *ptl;
1555
1556 retval = -ENOMEM;
1557 pte = get_locked_pte(mm, addr, &ptl);
1558 if (!pte)
1559 goto out;
1560 retval = -EBUSY;
1561 if (!pte_none(*pte))
1562 goto out_unlock;
1563
1564 /* Ok, finally just insert the thing.. */
1565 if (pfn_t_devmap(pfn))
1566 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1567 else
1568 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1569 set_pte_at(mm, addr, pte, entry);
1570 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1571
1572 retval = 0;
1573out_unlock:
1574 pte_unmap_unlock(pte, ptl);
1575out:
1576 return retval;
1577}
1578
1579/**
1580 * vm_insert_pfn - insert single pfn into user vma
1581 * @vma: user vma to map to
1582 * @addr: target user address of this page
1583 * @pfn: source kernel pfn
1584 *
1585 * Similar to vm_insert_page, this allows drivers to insert individual pages
1586 * they've allocated into a user vma. Same comments apply.
1587 *
1588 * This function should only be called from a vm_ops->fault handler, and
1589 * in that case the handler should return NULL.
1590 *
1591 * vma cannot be a COW mapping.
1592 *
1593 * As this is called only for pages that do not currently exist, we
1594 * do not need to flush old virtual caches or the TLB.
1595 */
1596int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1597 unsigned long pfn)
1598{
1599 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1600}
1601EXPORT_SYMBOL(vm_insert_pfn);
1602
1603/**
1604 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1605 * @vma: user vma to map to
1606 * @addr: target user address of this page
1607 * @pfn: source kernel pfn
1608 * @pgprot: pgprot flags for the inserted page
1609 *
1610 * This is exactly like vm_insert_pfn, except that it allows drivers to
1611 * to override pgprot on a per-page basis.
1612 *
1613 * This only makes sense for IO mappings, and it makes no sense for
1614 * cow mappings. In general, using multiple vmas is preferable;
1615 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1616 * impractical.
1617 */
1618int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1619 unsigned long pfn, pgprot_t pgprot)
1620{
1621 int ret;
1622 /*
1623 * Technically, architectures with pte_special can avoid all these
1624 * restrictions (same for remap_pfn_range). However we would like
1625 * consistency in testing and feature parity among all, so we should
1626 * try to keep these invariants in place for everybody.
1627 */
1628 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1629 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1630 (VM_PFNMAP|VM_MIXEDMAP));
1631 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1632 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1633
1634 if (addr < vma->vm_start || addr >= vma->vm_end)
1635 return -EFAULT;
1636
1637 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1638
1639 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1640
1641 return ret;
1642}
1643EXPORT_SYMBOL(vm_insert_pfn_prot);
1644
1645int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1646 pfn_t pfn)
1647{
1648 pgprot_t pgprot = vma->vm_page_prot;
1649
1650 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1651
1652 if (addr < vma->vm_start || addr >= vma->vm_end)
1653 return -EFAULT;
1654
1655 track_pfn_insert(vma, &pgprot, pfn);
1656
1657 /*
1658 * If we don't have pte special, then we have to use the pfn_valid()
1659 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1660 * refcount the page if pfn_valid is true (hence insert_page rather
1661 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1662 * without pte special, it would there be refcounted as a normal page.
1663 */
1664 if (!HAVE_PTE_SPECIAL && !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1665 struct page *page;
1666
1667 /*
1668 * At this point we are committed to insert_page()
1669 * regardless of whether the caller specified flags that
1670 * result in pfn_t_has_page() == false.
1671 */
1672 page = pfn_to_page(pfn_t_to_pfn(pfn));
1673 return insert_page(vma, addr, page, pgprot);
1674 }
1675 return insert_pfn(vma, addr, pfn, pgprot);
1676}
1677EXPORT_SYMBOL(vm_insert_mixed);
1678
1679/*
1680 * maps a range of physical memory into the requested pages. the old
1681 * mappings are removed. any references to nonexistent pages results
1682 * in null mappings (currently treated as "copy-on-access")
1683 */
1684static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1685 unsigned long addr, unsigned long end,
1686 unsigned long pfn, pgprot_t prot)
1687{
1688 pte_t *pte;
1689 spinlock_t *ptl;
1690
1691 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1692 if (!pte)
1693 return -ENOMEM;
1694 arch_enter_lazy_mmu_mode();
1695 do {
1696 BUG_ON(!pte_none(*pte));
1697 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1698 pfn++;
1699 } while (pte++, addr += PAGE_SIZE, addr != end);
1700 arch_leave_lazy_mmu_mode();
1701 pte_unmap_unlock(pte - 1, ptl);
1702 return 0;
1703}
1704
1705static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1706 unsigned long addr, unsigned long end,
1707 unsigned long pfn, pgprot_t prot)
1708{
1709 pmd_t *pmd;
1710 unsigned long next;
1711
1712 pfn -= addr >> PAGE_SHIFT;
1713 pmd = pmd_alloc(mm, pud, addr);
1714 if (!pmd)
1715 return -ENOMEM;
1716 VM_BUG_ON(pmd_trans_huge(*pmd));
1717 do {
1718 next = pmd_addr_end(addr, end);
1719 if (remap_pte_range(mm, pmd, addr, next,
1720 pfn + (addr >> PAGE_SHIFT), prot))
1721 return -ENOMEM;
1722 } while (pmd++, addr = next, addr != end);
1723 return 0;
1724}
1725
1726static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1727 unsigned long addr, unsigned long end,
1728 unsigned long pfn, pgprot_t prot)
1729{
1730 pud_t *pud;
1731 unsigned long next;
1732
1733 pfn -= addr >> PAGE_SHIFT;
1734 pud = pud_alloc(mm, pgd, addr);
1735 if (!pud)
1736 return -ENOMEM;
1737 do {
1738 next = pud_addr_end(addr, end);
1739 if (remap_pmd_range(mm, pud, addr, next,
1740 pfn + (addr >> PAGE_SHIFT), prot))
1741 return -ENOMEM;
1742 } while (pud++, addr = next, addr != end);
1743 return 0;
1744}
1745
1746/**
1747 * remap_pfn_range - remap kernel memory to userspace
1748 * @vma: user vma to map to
1749 * @addr: target user address to start at
1750 * @pfn: physical address of kernel memory
1751 * @size: size of map area
1752 * @prot: page protection flags for this mapping
1753 *
1754 * Note: this is only safe if the mm semaphore is held when called.
1755 */
1756int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1757 unsigned long pfn, unsigned long size, pgprot_t prot)
1758{
1759 pgd_t *pgd;
1760 unsigned long next;
1761 unsigned long end = addr + PAGE_ALIGN(size);
1762 struct mm_struct *mm = vma->vm_mm;
1763 unsigned long remap_pfn = pfn;
1764 int err;
1765
1766 /*
1767 * Physically remapped pages are special. Tell the
1768 * rest of the world about it:
1769 * VM_IO tells people not to look at these pages
1770 * (accesses can have side effects).
1771 * VM_PFNMAP tells the core MM that the base pages are just
1772 * raw PFN mappings, and do not have a "struct page" associated
1773 * with them.
1774 * VM_DONTEXPAND
1775 * Disable vma merging and expanding with mremap().
1776 * VM_DONTDUMP
1777 * Omit vma from core dump, even when VM_IO turned off.
1778 *
1779 * There's a horrible special case to handle copy-on-write
1780 * behaviour that some programs depend on. We mark the "original"
1781 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1782 * See vm_normal_page() for details.
1783 */
1784 if (is_cow_mapping(vma->vm_flags)) {
1785 if (addr != vma->vm_start || end != vma->vm_end)
1786 return -EINVAL;
1787 vma->vm_pgoff = pfn;
1788 }
1789
1790 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1791 if (err)
1792 return -EINVAL;
1793
1794 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1795
1796 BUG_ON(addr >= end);
1797 pfn -= addr >> PAGE_SHIFT;
1798 pgd = pgd_offset(mm, addr);
1799 flush_cache_range(vma, addr, end);
1800 do {
1801 next = pgd_addr_end(addr, end);
1802 err = remap_pud_range(mm, pgd, addr, next,
1803 pfn + (addr >> PAGE_SHIFT), prot);
1804 if (err)
1805 break;
1806 } while (pgd++, addr = next, addr != end);
1807
1808 if (err)
1809 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1810
1811 return err;
1812}
1813EXPORT_SYMBOL(remap_pfn_range);
1814
1815/**
1816 * vm_iomap_memory - remap memory to userspace
1817 * @vma: user vma to map to
1818 * @start: start of area
1819 * @len: size of area
1820 *
1821 * This is a simplified io_remap_pfn_range() for common driver use. The
1822 * driver just needs to give us the physical memory range to be mapped,
1823 * we'll figure out the rest from the vma information.
1824 *
1825 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1826 * whatever write-combining details or similar.
1827 */
1828int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1829{
1830 unsigned long vm_len, pfn, pages;
1831
1832 /* Check that the physical memory area passed in looks valid */
1833 if (start + len < start)
1834 return -EINVAL;
1835 /*
1836 * You *really* shouldn't map things that aren't page-aligned,
1837 * but we've historically allowed it because IO memory might
1838 * just have smaller alignment.
1839 */
1840 len += start & ~PAGE_MASK;
1841 pfn = start >> PAGE_SHIFT;
1842 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1843 if (pfn + pages < pfn)
1844 return -EINVAL;
1845
1846 /* We start the mapping 'vm_pgoff' pages into the area */
1847 if (vma->vm_pgoff > pages)
1848 return -EINVAL;
1849 pfn += vma->vm_pgoff;
1850 pages -= vma->vm_pgoff;
1851
1852 /* Can we fit all of the mapping? */
1853 vm_len = vma->vm_end - vma->vm_start;
1854 if (vm_len >> PAGE_SHIFT > pages)
1855 return -EINVAL;
1856
1857 /* Ok, let it rip */
1858 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1859}
1860EXPORT_SYMBOL(vm_iomap_memory);
1861
1862static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1863 unsigned long addr, unsigned long end,
1864 pte_fn_t fn, void *data)
1865{
1866 pte_t *pte;
1867 int err;
1868 pgtable_t token;
1869 spinlock_t *uninitialized_var(ptl);
1870
1871 pte = (mm == &init_mm) ?
1872 pte_alloc_kernel(pmd, addr) :
1873 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1874 if (!pte)
1875 return -ENOMEM;
1876
1877 BUG_ON(pmd_huge(*pmd));
1878
1879 arch_enter_lazy_mmu_mode();
1880
1881 token = pmd_pgtable(*pmd);
1882
1883 do {
1884 err = fn(pte++, token, addr, data);
1885 if (err)
1886 break;
1887 } while (addr += PAGE_SIZE, addr != end);
1888
1889 arch_leave_lazy_mmu_mode();
1890
1891 if (mm != &init_mm)
1892 pte_unmap_unlock(pte-1, ptl);
1893 return err;
1894}
1895
1896static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1897 unsigned long addr, unsigned long end,
1898 pte_fn_t fn, void *data)
1899{
1900 pmd_t *pmd;
1901 unsigned long next;
1902 int err;
1903
1904 BUG_ON(pud_huge(*pud));
1905
1906 pmd = pmd_alloc(mm, pud, addr);
1907 if (!pmd)
1908 return -ENOMEM;
1909 do {
1910 next = pmd_addr_end(addr, end);
1911 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1912 if (err)
1913 break;
1914 } while (pmd++, addr = next, addr != end);
1915 return err;
1916}
1917
1918static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1919 unsigned long addr, unsigned long end,
1920 pte_fn_t fn, void *data)
1921{
1922 pud_t *pud;
1923 unsigned long next;
1924 int err;
1925
1926 pud = pud_alloc(mm, pgd, addr);
1927 if (!pud)
1928 return -ENOMEM;
1929 do {
1930 next = pud_addr_end(addr, end);
1931 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1932 if (err)
1933 break;
1934 } while (pud++, addr = next, addr != end);
1935 return err;
1936}
1937
1938/*
1939 * Scan a region of virtual memory, filling in page tables as necessary
1940 * and calling a provided function on each leaf page table.
1941 */
1942int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1943 unsigned long size, pte_fn_t fn, void *data)
1944{
1945 pgd_t *pgd;
1946 unsigned long next;
1947 unsigned long end = addr + size;
1948 int err;
1949
1950 if (WARN_ON(addr >= end))
1951 return -EINVAL;
1952
1953 pgd = pgd_offset(mm, addr);
1954 do {
1955 next = pgd_addr_end(addr, end);
1956 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1957 if (err)
1958 break;
1959 } while (pgd++, addr = next, addr != end);
1960
1961 return err;
1962}
1963EXPORT_SYMBOL_GPL(apply_to_page_range);
1964
1965/*
1966 * handle_pte_fault chooses page fault handler according to an entry which was
1967 * read non-atomically. Before making any commitment, on those architectures
1968 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1969 * parts, do_swap_page must check under lock before unmapping the pte and
1970 * proceeding (but do_wp_page is only called after already making such a check;
1971 * and do_anonymous_page can safely check later on).
1972 */
1973static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1974 pte_t *page_table, pte_t orig_pte)
1975{
1976 int same = 1;
1977#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1978 if (sizeof(pte_t) > sizeof(unsigned long)) {
1979 spinlock_t *ptl = pte_lockptr(mm, pmd);
1980 spin_lock(ptl);
1981 same = pte_same(*page_table, orig_pte);
1982 spin_unlock(ptl);
1983 }
1984#endif
1985 pte_unmap(page_table);
1986 return same;
1987}
1988
1989static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
1990{
1991 debug_dma_assert_idle(src);
1992
1993 /*
1994 * If the source page was a PFN mapping, we don't have
1995 * a "struct page" for it. We do a best-effort copy by
1996 * just copying from the original user address. If that
1997 * fails, we just zero-fill it. Live with it.
1998 */
1999 if (unlikely(!src)) {
2000 void *kaddr = kmap_atomic(dst);
2001 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2002
2003 /*
2004 * This really shouldn't fail, because the page is there
2005 * in the page tables. But it might just be unreadable,
2006 * in which case we just give up and fill the result with
2007 * zeroes.
2008 */
2009 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2010 clear_page(kaddr);
2011 kunmap_atomic(kaddr);
2012 flush_dcache_page(dst);
2013 } else
2014 copy_user_highpage(dst, src, va, vma);
2015}
2016
2017static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2018{
2019 struct file *vm_file = vma->vm_file;
2020
2021 if (vm_file)
2022 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2023
2024 /*
2025 * Special mappings (e.g. VDSO) do not have any file so fake
2026 * a default GFP_KERNEL for them.
2027 */
2028 return GFP_KERNEL;
2029}
2030
2031/*
2032 * Notify the address space that the page is about to become writable so that
2033 * it can prohibit this or wait for the page to get into an appropriate state.
2034 *
2035 * We do this without the lock held, so that it can sleep if it needs to.
2036 */
2037static int do_page_mkwrite(struct vm_fault *vmf)
2038{
2039 int ret;
2040 struct page *page = vmf->page;
2041 unsigned int old_flags = vmf->flags;
2042
2043 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2044
2045 ret = vmf->vma->vm_ops->page_mkwrite(vmf->vma, vmf);
2046 /* Restore original flags so that caller is not surprised */
2047 vmf->flags = old_flags;
2048 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2049 return ret;
2050 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2051 lock_page(page);
2052 if (!page->mapping) {
2053 unlock_page(page);
2054 return 0; /* retry */
2055 }
2056 ret |= VM_FAULT_LOCKED;
2057 } else
2058 VM_BUG_ON_PAGE(!PageLocked(page), page);
2059 return ret;
2060}
2061
2062/*
2063 * Handle dirtying of a page in shared file mapping on a write fault.
2064 *
2065 * The function expects the page to be locked and unlocks it.
2066 */
2067static void fault_dirty_shared_page(struct vm_area_struct *vma,
2068 struct page *page)
2069{
2070 struct address_space *mapping;
2071 bool dirtied;
2072 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2073
2074 dirtied = set_page_dirty(page);
2075 VM_BUG_ON_PAGE(PageAnon(page), page);
2076 /*
2077 * Take a local copy of the address_space - page.mapping may be zeroed
2078 * by truncate after unlock_page(). The address_space itself remains
2079 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2080 * release semantics to prevent the compiler from undoing this copying.
2081 */
2082 mapping = page_rmapping(page);
2083 unlock_page(page);
2084
2085 if ((dirtied || page_mkwrite) && mapping) {
2086 /*
2087 * Some device drivers do not set page.mapping
2088 * but still dirty their pages
2089 */
2090 balance_dirty_pages_ratelimited(mapping);
2091 }
2092
2093 if (!page_mkwrite)
2094 file_update_time(vma->vm_file);
2095}
2096
2097/*
2098 * Handle write page faults for pages that can be reused in the current vma
2099 *
2100 * This can happen either due to the mapping being with the VM_SHARED flag,
2101 * or due to us being the last reference standing to the page. In either
2102 * case, all we need to do here is to mark the page as writable and update
2103 * any related book-keeping.
2104 */
2105static inline void wp_page_reuse(struct vm_fault *vmf)
2106 __releases(vmf->ptl)
2107{
2108 struct vm_area_struct *vma = vmf->vma;
2109 struct page *page = vmf->page;
2110 pte_t entry;
2111 /*
2112 * Clear the pages cpupid information as the existing
2113 * information potentially belongs to a now completely
2114 * unrelated process.
2115 */
2116 if (page)
2117 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2118
2119 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2120 entry = pte_mkyoung(vmf->orig_pte);
2121 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2122 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2123 update_mmu_cache(vma, vmf->address, vmf->pte);
2124 pte_unmap_unlock(vmf->pte, vmf->ptl);
2125}
2126
2127/*
2128 * Handle the case of a page which we actually need to copy to a new page.
2129 *
2130 * Called with mmap_sem locked and the old page referenced, but
2131 * without the ptl held.
2132 *
2133 * High level logic flow:
2134 *
2135 * - Allocate a page, copy the content of the old page to the new one.
2136 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2137 * - Take the PTL. If the pte changed, bail out and release the allocated page
2138 * - If the pte is still the way we remember it, update the page table and all
2139 * relevant references. This includes dropping the reference the page-table
2140 * held to the old page, as well as updating the rmap.
2141 * - In any case, unlock the PTL and drop the reference we took to the old page.
2142 */
2143static int wp_page_copy(struct vm_fault *vmf)
2144{
2145 struct vm_area_struct *vma = vmf->vma;
2146 struct mm_struct *mm = vma->vm_mm;
2147 struct page *old_page = vmf->page;
2148 struct page *new_page = NULL;
2149 pte_t entry;
2150 int page_copied = 0;
2151 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2152 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2153 struct mem_cgroup *memcg;
2154
2155 if (unlikely(anon_vma_prepare(vma)))
2156 goto oom;
2157
2158 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2159 new_page = alloc_zeroed_user_highpage_movable(vma,
2160 vmf->address);
2161 if (!new_page)
2162 goto oom;
2163 } else {
2164 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2165 vmf->address);
2166 if (!new_page)
2167 goto oom;
2168 cow_user_page(new_page, old_page, vmf->address, vma);
2169 }
2170
2171 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2172 goto oom_free_new;
2173
2174 __SetPageUptodate(new_page);
2175
2176 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2177
2178 /*
2179 * Re-check the pte - we dropped the lock
2180 */
2181 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2182 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2183 if (old_page) {
2184 if (!PageAnon(old_page)) {
2185 dec_mm_counter_fast(mm,
2186 mm_counter_file(old_page));
2187 inc_mm_counter_fast(mm, MM_ANONPAGES);
2188 }
2189 } else {
2190 inc_mm_counter_fast(mm, MM_ANONPAGES);
2191 }
2192 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2193 entry = mk_pte(new_page, vma->vm_page_prot);
2194 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2195 /*
2196 * Clear the pte entry and flush it first, before updating the
2197 * pte with the new entry. This will avoid a race condition
2198 * seen in the presence of one thread doing SMC and another
2199 * thread doing COW.
2200 */
2201 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2202 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2203 mem_cgroup_commit_charge(new_page, memcg, false, false);
2204 lru_cache_add_active_or_unevictable(new_page, vma);
2205 /*
2206 * We call the notify macro here because, when using secondary
2207 * mmu page tables (such as kvm shadow page tables), we want the
2208 * new page to be mapped directly into the secondary page table.
2209 */
2210 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2211 update_mmu_cache(vma, vmf->address, vmf->pte);
2212 if (old_page) {
2213 /*
2214 * Only after switching the pte to the new page may
2215 * we remove the mapcount here. Otherwise another
2216 * process may come and find the rmap count decremented
2217 * before the pte is switched to the new page, and
2218 * "reuse" the old page writing into it while our pte
2219 * here still points into it and can be read by other
2220 * threads.
2221 *
2222 * The critical issue is to order this
2223 * page_remove_rmap with the ptp_clear_flush above.
2224 * Those stores are ordered by (if nothing else,)
2225 * the barrier present in the atomic_add_negative
2226 * in page_remove_rmap.
2227 *
2228 * Then the TLB flush in ptep_clear_flush ensures that
2229 * no process can access the old page before the
2230 * decremented mapcount is visible. And the old page
2231 * cannot be reused until after the decremented
2232 * mapcount is visible. So transitively, TLBs to
2233 * old page will be flushed before it can be reused.
2234 */
2235 page_remove_rmap(old_page, false);
2236 }
2237
2238 /* Free the old page.. */
2239 new_page = old_page;
2240 page_copied = 1;
2241 } else {
2242 mem_cgroup_cancel_charge(new_page, memcg, false);
2243 }
2244
2245 if (new_page)
2246 put_page(new_page);
2247
2248 pte_unmap_unlock(vmf->pte, vmf->ptl);
2249 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2250 if (old_page) {
2251 /*
2252 * Don't let another task, with possibly unlocked vma,
2253 * keep the mlocked page.
2254 */
2255 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2256 lock_page(old_page); /* LRU manipulation */
2257 if (PageMlocked(old_page))
2258 munlock_vma_page(old_page);
2259 unlock_page(old_page);
2260 }
2261 put_page(old_page);
2262 }
2263 return page_copied ? VM_FAULT_WRITE : 0;
2264oom_free_new:
2265 put_page(new_page);
2266oom:
2267 if (old_page)
2268 put_page(old_page);
2269 return VM_FAULT_OOM;
2270}
2271
2272/**
2273 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2274 * writeable once the page is prepared
2275 *
2276 * @vmf: structure describing the fault
2277 *
2278 * This function handles all that is needed to finish a write page fault in a
2279 * shared mapping due to PTE being read-only once the mapped page is prepared.
2280 * It handles locking of PTE and modifying it. The function returns
2281 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2282 * lock.
2283 *
2284 * The function expects the page to be locked or other protection against
2285 * concurrent faults / writeback (such as DAX radix tree locks).
2286 */
2287int finish_mkwrite_fault(struct vm_fault *vmf)
2288{
2289 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2290 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2291 &vmf->ptl);
2292 /*
2293 * We might have raced with another page fault while we released the
2294 * pte_offset_map_lock.
2295 */
2296 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2297 pte_unmap_unlock(vmf->pte, vmf->ptl);
2298 return VM_FAULT_NOPAGE;
2299 }
2300 wp_page_reuse(vmf);
2301 return 0;
2302}
2303
2304/*
2305 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2306 * mapping
2307 */
2308static int wp_pfn_shared(struct vm_fault *vmf)
2309{
2310 struct vm_area_struct *vma = vmf->vma;
2311
2312 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2313 int ret;
2314
2315 pte_unmap_unlock(vmf->pte, vmf->ptl);
2316 vmf->flags |= FAULT_FLAG_MKWRITE;
2317 ret = vma->vm_ops->pfn_mkwrite(vma, vmf);
2318 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2319 return ret;
2320 return finish_mkwrite_fault(vmf);
2321 }
2322 wp_page_reuse(vmf);
2323 return VM_FAULT_WRITE;
2324}
2325
2326static int wp_page_shared(struct vm_fault *vmf)
2327 __releases(vmf->ptl)
2328{
2329 struct vm_area_struct *vma = vmf->vma;
2330
2331 get_page(vmf->page);
2332
2333 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2334 int tmp;
2335
2336 pte_unmap_unlock(vmf->pte, vmf->ptl);
2337 tmp = do_page_mkwrite(vmf);
2338 if (unlikely(!tmp || (tmp &
2339 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2340 put_page(vmf->page);
2341 return tmp;
2342 }
2343 tmp = finish_mkwrite_fault(vmf);
2344 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2345 unlock_page(vmf->page);
2346 put_page(vmf->page);
2347 return tmp;
2348 }
2349 } else {
2350 wp_page_reuse(vmf);
2351 lock_page(vmf->page);
2352 }
2353 fault_dirty_shared_page(vma, vmf->page);
2354 put_page(vmf->page);
2355
2356 return VM_FAULT_WRITE;
2357}
2358
2359/*
2360 * This routine handles present pages, when users try to write
2361 * to a shared page. It is done by copying the page to a new address
2362 * and decrementing the shared-page counter for the old page.
2363 *
2364 * Note that this routine assumes that the protection checks have been
2365 * done by the caller (the low-level page fault routine in most cases).
2366 * Thus we can safely just mark it writable once we've done any necessary
2367 * COW.
2368 *
2369 * We also mark the page dirty at this point even though the page will
2370 * change only once the write actually happens. This avoids a few races,
2371 * and potentially makes it more efficient.
2372 *
2373 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2374 * but allow concurrent faults), with pte both mapped and locked.
2375 * We return with mmap_sem still held, but pte unmapped and unlocked.
2376 */
2377static int do_wp_page(struct vm_fault *vmf)
2378 __releases(vmf->ptl)
2379{
2380 struct vm_area_struct *vma = vmf->vma;
2381
2382 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2383 if (!vmf->page) {
2384 /*
2385 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2386 * VM_PFNMAP VMA.
2387 *
2388 * We should not cow pages in a shared writeable mapping.
2389 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2390 */
2391 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2392 (VM_WRITE|VM_SHARED))
2393 return wp_pfn_shared(vmf);
2394
2395 pte_unmap_unlock(vmf->pte, vmf->ptl);
2396 return wp_page_copy(vmf);
2397 }
2398
2399 /*
2400 * Take out anonymous pages first, anonymous shared vmas are
2401 * not dirty accountable.
2402 */
2403 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2404 int total_mapcount;
2405 if (!trylock_page(vmf->page)) {
2406 get_page(vmf->page);
2407 pte_unmap_unlock(vmf->pte, vmf->ptl);
2408 lock_page(vmf->page);
2409 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2410 vmf->address, &vmf->ptl);
2411 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2412 unlock_page(vmf->page);
2413 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414 put_page(vmf->page);
2415 return 0;
2416 }
2417 put_page(vmf->page);
2418 }
2419 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2420 if (total_mapcount == 1) {
2421 /*
2422 * The page is all ours. Move it to
2423 * our anon_vma so the rmap code will
2424 * not search our parent or siblings.
2425 * Protected against the rmap code by
2426 * the page lock.
2427 */
2428 page_move_anon_rmap(vmf->page, vma);
2429 }
2430 unlock_page(vmf->page);
2431 wp_page_reuse(vmf);
2432 return VM_FAULT_WRITE;
2433 }
2434 unlock_page(vmf->page);
2435 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2436 (VM_WRITE|VM_SHARED))) {
2437 return wp_page_shared(vmf);
2438 }
2439
2440 /*
2441 * Ok, we need to copy. Oh, well..
2442 */
2443 get_page(vmf->page);
2444
2445 pte_unmap_unlock(vmf->pte, vmf->ptl);
2446 return wp_page_copy(vmf);
2447}
2448
2449static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2450 unsigned long start_addr, unsigned long end_addr,
2451 struct zap_details *details)
2452{
2453 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2454}
2455
2456static inline void unmap_mapping_range_tree(struct rb_root *root,
2457 struct zap_details *details)
2458{
2459 struct vm_area_struct *vma;
2460 pgoff_t vba, vea, zba, zea;
2461
2462 vma_interval_tree_foreach(vma, root,
2463 details->first_index, details->last_index) {
2464
2465 vba = vma->vm_pgoff;
2466 vea = vba + vma_pages(vma) - 1;
2467 zba = details->first_index;
2468 if (zba < vba)
2469 zba = vba;
2470 zea = details->last_index;
2471 if (zea > vea)
2472 zea = vea;
2473
2474 unmap_mapping_range_vma(vma,
2475 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2476 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2477 details);
2478 }
2479}
2480
2481/**
2482 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2483 * address_space corresponding to the specified page range in the underlying
2484 * file.
2485 *
2486 * @mapping: the address space containing mmaps to be unmapped.
2487 * @holebegin: byte in first page to unmap, relative to the start of
2488 * the underlying file. This will be rounded down to a PAGE_SIZE
2489 * boundary. Note that this is different from truncate_pagecache(), which
2490 * must keep the partial page. In contrast, we must get rid of
2491 * partial pages.
2492 * @holelen: size of prospective hole in bytes. This will be rounded
2493 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2494 * end of the file.
2495 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2496 * but 0 when invalidating pagecache, don't throw away private data.
2497 */
2498void unmap_mapping_range(struct address_space *mapping,
2499 loff_t const holebegin, loff_t const holelen, int even_cows)
2500{
2501 struct zap_details details = { };
2502 pgoff_t hba = holebegin >> PAGE_SHIFT;
2503 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2504
2505 /* Check for overflow. */
2506 if (sizeof(holelen) > sizeof(hlen)) {
2507 long long holeend =
2508 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2509 if (holeend & ~(long long)ULONG_MAX)
2510 hlen = ULONG_MAX - hba + 1;
2511 }
2512
2513 details.check_mapping = even_cows? NULL: mapping;
2514 details.first_index = hba;
2515 details.last_index = hba + hlen - 1;
2516 if (details.last_index < details.first_index)
2517 details.last_index = ULONG_MAX;
2518
2519 i_mmap_lock_write(mapping);
2520 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2521 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2522 i_mmap_unlock_write(mapping);
2523}
2524EXPORT_SYMBOL(unmap_mapping_range);
2525
2526/*
2527 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2528 * but allow concurrent faults), and pte mapped but not yet locked.
2529 * We return with pte unmapped and unlocked.
2530 *
2531 * We return with the mmap_sem locked or unlocked in the same cases
2532 * as does filemap_fault().
2533 */
2534int do_swap_page(struct vm_fault *vmf)
2535{
2536 struct vm_area_struct *vma = vmf->vma;
2537 struct page *page, *swapcache;
2538 struct mem_cgroup *memcg;
2539 swp_entry_t entry;
2540 pte_t pte;
2541 int locked;
2542 int exclusive = 0;
2543 int ret = 0;
2544
2545 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2546 goto out;
2547
2548 entry = pte_to_swp_entry(vmf->orig_pte);
2549 if (unlikely(non_swap_entry(entry))) {
2550 if (is_migration_entry(entry)) {
2551 migration_entry_wait(vma->vm_mm, vmf->pmd,
2552 vmf->address);
2553 } else if (is_hwpoison_entry(entry)) {
2554 ret = VM_FAULT_HWPOISON;
2555 } else {
2556 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2557 ret = VM_FAULT_SIGBUS;
2558 }
2559 goto out;
2560 }
2561 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2562 page = lookup_swap_cache(entry);
2563 if (!page) {
2564 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2565 vmf->address);
2566 if (!page) {
2567 /*
2568 * Back out if somebody else faulted in this pte
2569 * while we released the pte lock.
2570 */
2571 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2572 vmf->address, &vmf->ptl);
2573 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2574 ret = VM_FAULT_OOM;
2575 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2576 goto unlock;
2577 }
2578
2579 /* Had to read the page from swap area: Major fault */
2580 ret = VM_FAULT_MAJOR;
2581 count_vm_event(PGMAJFAULT);
2582 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2583 } else if (PageHWPoison(page)) {
2584 /*
2585 * hwpoisoned dirty swapcache pages are kept for killing
2586 * owner processes (which may be unknown at hwpoison time)
2587 */
2588 ret = VM_FAULT_HWPOISON;
2589 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2590 swapcache = page;
2591 goto out_release;
2592 }
2593
2594 swapcache = page;
2595 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2596
2597 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2598 if (!locked) {
2599 ret |= VM_FAULT_RETRY;
2600 goto out_release;
2601 }
2602
2603 /*
2604 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2605 * release the swapcache from under us. The page pin, and pte_same
2606 * test below, are not enough to exclude that. Even if it is still
2607 * swapcache, we need to check that the page's swap has not changed.
2608 */
2609 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2610 goto out_page;
2611
2612 page = ksm_might_need_to_copy(page, vma, vmf->address);
2613 if (unlikely(!page)) {
2614 ret = VM_FAULT_OOM;
2615 page = swapcache;
2616 goto out_page;
2617 }
2618
2619 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2620 &memcg, false)) {
2621 ret = VM_FAULT_OOM;
2622 goto out_page;
2623 }
2624
2625 /*
2626 * Back out if somebody else already faulted in this pte.
2627 */
2628 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2629 &vmf->ptl);
2630 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2631 goto out_nomap;
2632
2633 if (unlikely(!PageUptodate(page))) {
2634 ret = VM_FAULT_SIGBUS;
2635 goto out_nomap;
2636 }
2637
2638 /*
2639 * The page isn't present yet, go ahead with the fault.
2640 *
2641 * Be careful about the sequence of operations here.
2642 * To get its accounting right, reuse_swap_page() must be called
2643 * while the page is counted on swap but not yet in mapcount i.e.
2644 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2645 * must be called after the swap_free(), or it will never succeed.
2646 */
2647
2648 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2649 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2650 pte = mk_pte(page, vma->vm_page_prot);
2651 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2652 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2653 vmf->flags &= ~FAULT_FLAG_WRITE;
2654 ret |= VM_FAULT_WRITE;
2655 exclusive = RMAP_EXCLUSIVE;
2656 }
2657 flush_icache_page(vma, page);
2658 if (pte_swp_soft_dirty(vmf->orig_pte))
2659 pte = pte_mksoft_dirty(pte);
2660 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2661 vmf->orig_pte = pte;
2662 if (page == swapcache) {
2663 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2664 mem_cgroup_commit_charge(page, memcg, true, false);
2665 activate_page(page);
2666 } else { /* ksm created a completely new copy */
2667 page_add_new_anon_rmap(page, vma, vmf->address, false);
2668 mem_cgroup_commit_charge(page, memcg, false, false);
2669 lru_cache_add_active_or_unevictable(page, vma);
2670 }
2671
2672 swap_free(entry);
2673 if (mem_cgroup_swap_full(page) ||
2674 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2675 try_to_free_swap(page);
2676 unlock_page(page);
2677 if (page != swapcache) {
2678 /*
2679 * Hold the lock to avoid the swap entry to be reused
2680 * until we take the PT lock for the pte_same() check
2681 * (to avoid false positives from pte_same). For
2682 * further safety release the lock after the swap_free
2683 * so that the swap count won't change under a
2684 * parallel locked swapcache.
2685 */
2686 unlock_page(swapcache);
2687 put_page(swapcache);
2688 }
2689
2690 if (vmf->flags & FAULT_FLAG_WRITE) {
2691 ret |= do_wp_page(vmf);
2692 if (ret & VM_FAULT_ERROR)
2693 ret &= VM_FAULT_ERROR;
2694 goto out;
2695 }
2696
2697 /* No need to invalidate - it was non-present before */
2698 update_mmu_cache(vma, vmf->address, vmf->pte);
2699unlock:
2700 pte_unmap_unlock(vmf->pte, vmf->ptl);
2701out:
2702 return ret;
2703out_nomap:
2704 mem_cgroup_cancel_charge(page, memcg, false);
2705 pte_unmap_unlock(vmf->pte, vmf->ptl);
2706out_page:
2707 unlock_page(page);
2708out_release:
2709 put_page(page);
2710 if (page != swapcache) {
2711 unlock_page(swapcache);
2712 put_page(swapcache);
2713 }
2714 return ret;
2715}
2716
2717/*
2718 * This is like a special single-page "expand_{down|up}wards()",
2719 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2720 * doesn't hit another vma.
2721 */
2722static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2723{
2724 address &= PAGE_MASK;
2725 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2726 struct vm_area_struct *prev = vma->vm_prev;
2727
2728 /*
2729 * Is there a mapping abutting this one below?
2730 *
2731 * That's only ok if it's the same stack mapping
2732 * that has gotten split..
2733 */
2734 if (prev && prev->vm_end == address)
2735 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2736
2737 return expand_downwards(vma, address - PAGE_SIZE);
2738 }
2739 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2740 struct vm_area_struct *next = vma->vm_next;
2741
2742 /* As VM_GROWSDOWN but s/below/above/ */
2743 if (next && next->vm_start == address + PAGE_SIZE)
2744 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2745
2746 return expand_upwards(vma, address + PAGE_SIZE);
2747 }
2748 return 0;
2749}
2750
2751/*
2752 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2753 * but allow concurrent faults), and pte mapped but not yet locked.
2754 * We return with mmap_sem still held, but pte unmapped and unlocked.
2755 */
2756static int do_anonymous_page(struct vm_fault *vmf)
2757{
2758 struct vm_area_struct *vma = vmf->vma;
2759 struct mem_cgroup *memcg;
2760 struct page *page;
2761 pte_t entry;
2762
2763 /* File mapping without ->vm_ops ? */
2764 if (vma->vm_flags & VM_SHARED)
2765 return VM_FAULT_SIGBUS;
2766
2767 /* Check if we need to add a guard page to the stack */
2768 if (check_stack_guard_page(vma, vmf->address) < 0)
2769 return VM_FAULT_SIGSEGV;
2770
2771 /*
2772 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2773 * pte_offset_map() on pmds where a huge pmd might be created
2774 * from a different thread.
2775 *
2776 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2777 * parallel threads are excluded by other means.
2778 *
2779 * Here we only have down_read(mmap_sem).
2780 */
2781 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2782 return VM_FAULT_OOM;
2783
2784 /* See the comment in pte_alloc_one_map() */
2785 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2786 return 0;
2787
2788 /* Use the zero-page for reads */
2789 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2790 !mm_forbids_zeropage(vma->vm_mm)) {
2791 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2792 vma->vm_page_prot));
2793 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2794 vmf->address, &vmf->ptl);
2795 if (!pte_none(*vmf->pte))
2796 goto unlock;
2797 /* Deliver the page fault to userland, check inside PT lock */
2798 if (userfaultfd_missing(vma)) {
2799 pte_unmap_unlock(vmf->pte, vmf->ptl);
2800 return handle_userfault(vmf, VM_UFFD_MISSING);
2801 }
2802 goto setpte;
2803 }
2804
2805 /* Allocate our own private page. */
2806 if (unlikely(anon_vma_prepare(vma)))
2807 goto oom;
2808 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2809 if (!page)
2810 goto oom;
2811
2812 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2813 goto oom_free_page;
2814
2815 /*
2816 * The memory barrier inside __SetPageUptodate makes sure that
2817 * preceeding stores to the page contents become visible before
2818 * the set_pte_at() write.
2819 */
2820 __SetPageUptodate(page);
2821
2822 entry = mk_pte(page, vma->vm_page_prot);
2823 if (vma->vm_flags & VM_WRITE)
2824 entry = pte_mkwrite(pte_mkdirty(entry));
2825
2826 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2827 &vmf->ptl);
2828 if (!pte_none(*vmf->pte))
2829 goto release;
2830
2831 /* Deliver the page fault to userland, check inside PT lock */
2832 if (userfaultfd_missing(vma)) {
2833 pte_unmap_unlock(vmf->pte, vmf->ptl);
2834 mem_cgroup_cancel_charge(page, memcg, false);
2835 put_page(page);
2836 return handle_userfault(vmf, VM_UFFD_MISSING);
2837 }
2838
2839 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2840 page_add_new_anon_rmap(page, vma, vmf->address, false);
2841 mem_cgroup_commit_charge(page, memcg, false, false);
2842 lru_cache_add_active_or_unevictable(page, vma);
2843setpte:
2844 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2845
2846 /* No need to invalidate - it was non-present before */
2847 update_mmu_cache(vma, vmf->address, vmf->pte);
2848unlock:
2849 pte_unmap_unlock(vmf->pte, vmf->ptl);
2850 return 0;
2851release:
2852 mem_cgroup_cancel_charge(page, memcg, false);
2853 put_page(page);
2854 goto unlock;
2855oom_free_page:
2856 put_page(page);
2857oom:
2858 return VM_FAULT_OOM;
2859}
2860
2861/*
2862 * The mmap_sem must have been held on entry, and may have been
2863 * released depending on flags and vma->vm_ops->fault() return value.
2864 * See filemap_fault() and __lock_page_retry().
2865 */
2866static int __do_fault(struct vm_fault *vmf)
2867{
2868 struct vm_area_struct *vma = vmf->vma;
2869 int ret;
2870
2871 ret = vma->vm_ops->fault(vma, vmf);
2872 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2873 VM_FAULT_DONE_COW)))
2874 return ret;
2875
2876 if (unlikely(PageHWPoison(vmf->page))) {
2877 if (ret & VM_FAULT_LOCKED)
2878 unlock_page(vmf->page);
2879 put_page(vmf->page);
2880 vmf->page = NULL;
2881 return VM_FAULT_HWPOISON;
2882 }
2883
2884 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2885 lock_page(vmf->page);
2886 else
2887 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2888
2889 return ret;
2890}
2891
2892static int pte_alloc_one_map(struct vm_fault *vmf)
2893{
2894 struct vm_area_struct *vma = vmf->vma;
2895
2896 if (!pmd_none(*vmf->pmd))
2897 goto map_pte;
2898 if (vmf->prealloc_pte) {
2899 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2900 if (unlikely(!pmd_none(*vmf->pmd))) {
2901 spin_unlock(vmf->ptl);
2902 goto map_pte;
2903 }
2904
2905 atomic_long_inc(&vma->vm_mm->nr_ptes);
2906 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2907 spin_unlock(vmf->ptl);
2908 vmf->prealloc_pte = 0;
2909 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
2910 return VM_FAULT_OOM;
2911 }
2912map_pte:
2913 /*
2914 * If a huge pmd materialized under us just retry later. Use
2915 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2916 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2917 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2918 * in a different thread of this mm, in turn leading to a misleading
2919 * pmd_trans_huge() retval. All we have to ensure is that it is a
2920 * regular pmd that we can walk with pte_offset_map() and we can do that
2921 * through an atomic read in C, which is what pmd_trans_unstable()
2922 * provides.
2923 */
2924 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
2925 return VM_FAULT_NOPAGE;
2926
2927 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2928 &vmf->ptl);
2929 return 0;
2930}
2931
2932#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2933
2934#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2935static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2936 unsigned long haddr)
2937{
2938 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2939 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2940 return false;
2941 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2942 return false;
2943 return true;
2944}
2945
2946static void deposit_prealloc_pte(struct vm_fault *vmf)
2947{
2948 struct vm_area_struct *vma = vmf->vma;
2949
2950 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2951 /*
2952 * We are going to consume the prealloc table,
2953 * count that as nr_ptes.
2954 */
2955 atomic_long_inc(&vma->vm_mm->nr_ptes);
2956 vmf->prealloc_pte = 0;
2957}
2958
2959static int do_set_pmd(struct vm_fault *vmf, struct page *page)
2960{
2961 struct vm_area_struct *vma = vmf->vma;
2962 bool write = vmf->flags & FAULT_FLAG_WRITE;
2963 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2964 pmd_t entry;
2965 int i, ret;
2966
2967 if (!transhuge_vma_suitable(vma, haddr))
2968 return VM_FAULT_FALLBACK;
2969
2970 ret = VM_FAULT_FALLBACK;
2971 page = compound_head(page);
2972
2973 /*
2974 * Archs like ppc64 need additonal space to store information
2975 * related to pte entry. Use the preallocated table for that.
2976 */
2977 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
2978 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
2979 if (!vmf->prealloc_pte)
2980 return VM_FAULT_OOM;
2981 smp_wmb(); /* See comment in __pte_alloc() */
2982 }
2983
2984 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2985 if (unlikely(!pmd_none(*vmf->pmd)))
2986 goto out;
2987
2988 for (i = 0; i < HPAGE_PMD_NR; i++)
2989 flush_icache_page(vma, page + i);
2990
2991 entry = mk_huge_pmd(page, vma->vm_page_prot);
2992 if (write)
2993 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
2994
2995 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
2996 page_add_file_rmap(page, true);
2997 /*
2998 * deposit and withdraw with pmd lock held
2999 */
3000 if (arch_needs_pgtable_deposit())
3001 deposit_prealloc_pte(vmf);
3002
3003 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3004
3005 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3006
3007 /* fault is handled */
3008 ret = 0;
3009 count_vm_event(THP_FILE_MAPPED);
3010out:
3011 spin_unlock(vmf->ptl);
3012 return ret;
3013}
3014#else
3015static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3016{
3017 BUILD_BUG();
3018 return 0;
3019}
3020#endif
3021
3022/**
3023 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3024 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3025 *
3026 * @vmf: fault environment
3027 * @memcg: memcg to charge page (only for private mappings)
3028 * @page: page to map
3029 *
3030 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3031 * return.
3032 *
3033 * Target users are page handler itself and implementations of
3034 * vm_ops->map_pages.
3035 */
3036int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3037 struct page *page)
3038{
3039 struct vm_area_struct *vma = vmf->vma;
3040 bool write = vmf->flags & FAULT_FLAG_WRITE;
3041 pte_t entry;
3042 int ret;
3043
3044 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3045 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3046 /* THP on COW? */
3047 VM_BUG_ON_PAGE(memcg, page);
3048
3049 ret = do_set_pmd(vmf, page);
3050 if (ret != VM_FAULT_FALLBACK)
3051 return ret;
3052 }
3053
3054 if (!vmf->pte) {
3055 ret = pte_alloc_one_map(vmf);
3056 if (ret)
3057 return ret;
3058 }
3059
3060 /* Re-check under ptl */
3061 if (unlikely(!pte_none(*vmf->pte)))
3062 return VM_FAULT_NOPAGE;
3063
3064 flush_icache_page(vma, page);
3065 entry = mk_pte(page, vma->vm_page_prot);
3066 if (write)
3067 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3068 /* copy-on-write page */
3069 if (write && !(vma->vm_flags & VM_SHARED)) {
3070 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3071 page_add_new_anon_rmap(page, vma, vmf->address, false);
3072 mem_cgroup_commit_charge(page, memcg, false, false);
3073 lru_cache_add_active_or_unevictable(page, vma);
3074 } else {
3075 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3076 page_add_file_rmap(page, false);
3077 }
3078 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3079
3080 /* no need to invalidate: a not-present page won't be cached */
3081 update_mmu_cache(vma, vmf->address, vmf->pte);
3082
3083 return 0;
3084}
3085
3086
3087/**
3088 * finish_fault - finish page fault once we have prepared the page to fault
3089 *
3090 * @vmf: structure describing the fault
3091 *
3092 * This function handles all that is needed to finish a page fault once the
3093 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3094 * given page, adds reverse page mapping, handles memcg charges and LRU
3095 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3096 * error.
3097 *
3098 * The function expects the page to be locked and on success it consumes a
3099 * reference of a page being mapped (for the PTE which maps it).
3100 */
3101int finish_fault(struct vm_fault *vmf)
3102{
3103 struct page *page;
3104 int ret;
3105
3106 /* Did we COW the page? */
3107 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3108 !(vmf->vma->vm_flags & VM_SHARED))
3109 page = vmf->cow_page;
3110 else
3111 page = vmf->page;
3112 ret = alloc_set_pte(vmf, vmf->memcg, page);
3113 if (vmf->pte)
3114 pte_unmap_unlock(vmf->pte, vmf->ptl);
3115 return ret;
3116}
3117
3118static unsigned long fault_around_bytes __read_mostly =
3119 rounddown_pow_of_two(65536);
3120
3121#ifdef CONFIG_DEBUG_FS
3122static int fault_around_bytes_get(void *data, u64 *val)
3123{
3124 *val = fault_around_bytes;
3125 return 0;
3126}
3127
3128/*
3129 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3130 * rounded down to nearest page order. It's what do_fault_around() expects to
3131 * see.
3132 */
3133static int fault_around_bytes_set(void *data, u64 val)
3134{
3135 if (val / PAGE_SIZE > PTRS_PER_PTE)
3136 return -EINVAL;
3137 if (val > PAGE_SIZE)
3138 fault_around_bytes = rounddown_pow_of_two(val);
3139 else
3140 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3141 return 0;
3142}
3143DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3144 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3145
3146static int __init fault_around_debugfs(void)
3147{
3148 void *ret;
3149
3150 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3151 &fault_around_bytes_fops);
3152 if (!ret)
3153 pr_warn("Failed to create fault_around_bytes in debugfs");
3154 return 0;
3155}
3156late_initcall(fault_around_debugfs);
3157#endif
3158
3159/*
3160 * do_fault_around() tries to map few pages around the fault address. The hope
3161 * is that the pages will be needed soon and this will lower the number of
3162 * faults to handle.
3163 *
3164 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3165 * not ready to be mapped: not up-to-date, locked, etc.
3166 *
3167 * This function is called with the page table lock taken. In the split ptlock
3168 * case the page table lock only protects only those entries which belong to
3169 * the page table corresponding to the fault address.
3170 *
3171 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3172 * only once.
3173 *
3174 * fault_around_pages() defines how many pages we'll try to map.
3175 * do_fault_around() expects it to return a power of two less than or equal to
3176 * PTRS_PER_PTE.
3177 *
3178 * The virtual address of the area that we map is naturally aligned to the
3179 * fault_around_pages() value (and therefore to page order). This way it's
3180 * easier to guarantee that we don't cross page table boundaries.
3181 */
3182static int do_fault_around(struct vm_fault *vmf)
3183{
3184 unsigned long address = vmf->address, nr_pages, mask;
3185 pgoff_t start_pgoff = vmf->pgoff;
3186 pgoff_t end_pgoff;
3187 int off, ret = 0;
3188
3189 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3190 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3191
3192 vmf->address = max(address & mask, vmf->vma->vm_start);
3193 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3194 start_pgoff -= off;
3195
3196 /*
3197 * end_pgoff is either end of page table or end of vma
3198 * or fault_around_pages() from start_pgoff, depending what is nearest.
3199 */
3200 end_pgoff = start_pgoff -
3201 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3202 PTRS_PER_PTE - 1;
3203 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3204 start_pgoff + nr_pages - 1);
3205
3206 if (pmd_none(*vmf->pmd)) {
3207 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3208 vmf->address);
3209 if (!vmf->prealloc_pte)
3210 goto out;
3211 smp_wmb(); /* See comment in __pte_alloc() */
3212 }
3213
3214 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3215
3216 /* Huge page is mapped? Page fault is solved */
3217 if (pmd_trans_huge(*vmf->pmd)) {
3218 ret = VM_FAULT_NOPAGE;
3219 goto out;
3220 }
3221
3222 /* ->map_pages() haven't done anything useful. Cold page cache? */
3223 if (!vmf->pte)
3224 goto out;
3225
3226 /* check if the page fault is solved */
3227 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3228 if (!pte_none(*vmf->pte))
3229 ret = VM_FAULT_NOPAGE;
3230 pte_unmap_unlock(vmf->pte, vmf->ptl);
3231out:
3232 vmf->address = address;
3233 vmf->pte = NULL;
3234 return ret;
3235}
3236
3237static int do_read_fault(struct vm_fault *vmf)
3238{
3239 struct vm_area_struct *vma = vmf->vma;
3240 int ret = 0;
3241
3242 /*
3243 * Let's call ->map_pages() first and use ->fault() as fallback
3244 * if page by the offset is not ready to be mapped (cold cache or
3245 * something).
3246 */
3247 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3248 ret = do_fault_around(vmf);
3249 if (ret)
3250 return ret;
3251 }
3252
3253 ret = __do_fault(vmf);
3254 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3255 return ret;
3256
3257 ret |= finish_fault(vmf);
3258 unlock_page(vmf->page);
3259 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3260 put_page(vmf->page);
3261 return ret;
3262}
3263
3264static int do_cow_fault(struct vm_fault *vmf)
3265{
3266 struct vm_area_struct *vma = vmf->vma;
3267 int ret;
3268
3269 if (unlikely(anon_vma_prepare(vma)))
3270 return VM_FAULT_OOM;
3271
3272 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3273 if (!vmf->cow_page)
3274 return VM_FAULT_OOM;
3275
3276 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3277 &vmf->memcg, false)) {
3278 put_page(vmf->cow_page);
3279 return VM_FAULT_OOM;
3280 }
3281
3282 ret = __do_fault(vmf);
3283 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3284 goto uncharge_out;
3285 if (ret & VM_FAULT_DONE_COW)
3286 return ret;
3287
3288 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3289 __SetPageUptodate(vmf->cow_page);
3290
3291 ret |= finish_fault(vmf);
3292 unlock_page(vmf->page);
3293 put_page(vmf->page);
3294 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3295 goto uncharge_out;
3296 return ret;
3297uncharge_out:
3298 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3299 put_page(vmf->cow_page);
3300 return ret;
3301}
3302
3303static int do_shared_fault(struct vm_fault *vmf)
3304{
3305 struct vm_area_struct *vma = vmf->vma;
3306 int ret, tmp;
3307
3308 ret = __do_fault(vmf);
3309 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3310 return ret;
3311
3312 /*
3313 * Check if the backing address space wants to know that the page is
3314 * about to become writable
3315 */
3316 if (vma->vm_ops->page_mkwrite) {
3317 unlock_page(vmf->page);
3318 tmp = do_page_mkwrite(vmf);
3319 if (unlikely(!tmp ||
3320 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3321 put_page(vmf->page);
3322 return tmp;
3323 }
3324 }
3325
3326 ret |= finish_fault(vmf);
3327 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3328 VM_FAULT_RETRY))) {
3329 unlock_page(vmf->page);
3330 put_page(vmf->page);
3331 return ret;
3332 }
3333
3334 fault_dirty_shared_page(vma, vmf->page);
3335 return ret;
3336}
3337
3338/*
3339 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3340 * but allow concurrent faults).
3341 * The mmap_sem may have been released depending on flags and our
3342 * return value. See filemap_fault() and __lock_page_or_retry().
3343 */
3344static int do_fault(struct vm_fault *vmf)
3345{
3346 struct vm_area_struct *vma = vmf->vma;
3347 int ret;
3348
3349 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3350 if (!vma->vm_ops->fault)
3351 ret = VM_FAULT_SIGBUS;
3352 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3353 ret = do_read_fault(vmf);
3354 else if (!(vma->vm_flags & VM_SHARED))
3355 ret = do_cow_fault(vmf);
3356 else
3357 ret = do_shared_fault(vmf);
3358
3359 /* preallocated pagetable is unused: free it */
3360 if (vmf->prealloc_pte) {
3361 pte_free(vma->vm_mm, vmf->prealloc_pte);
3362 vmf->prealloc_pte = 0;
3363 }
3364 return ret;
3365}
3366
3367static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3368 unsigned long addr, int page_nid,
3369 int *flags)
3370{
3371 get_page(page);
3372
3373 count_vm_numa_event(NUMA_HINT_FAULTS);
3374 if (page_nid == numa_node_id()) {
3375 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3376 *flags |= TNF_FAULT_LOCAL;
3377 }
3378
3379 return mpol_misplaced(page, vma, addr);
3380}
3381
3382static int do_numa_page(struct vm_fault *vmf)
3383{
3384 struct vm_area_struct *vma = vmf->vma;
3385 struct page *page = NULL;
3386 int page_nid = -1;
3387 int last_cpupid;
3388 int target_nid;
3389 bool migrated = false;
3390 pte_t pte = vmf->orig_pte;
3391 bool was_writable = pte_write(pte);
3392 int flags = 0;
3393
3394 /*
3395 * The "pte" at this point cannot be used safely without
3396 * validation through pte_unmap_same(). It's of NUMA type but
3397 * the pfn may be screwed if the read is non atomic.
3398 *
3399 * We can safely just do a "set_pte_at()", because the old
3400 * page table entry is not accessible, so there would be no
3401 * concurrent hardware modifications to the PTE.
3402 */
3403 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3404 spin_lock(vmf->ptl);
3405 if (unlikely(!pte_same(*vmf->pte, pte))) {
3406 pte_unmap_unlock(vmf->pte, vmf->ptl);
3407 goto out;
3408 }
3409
3410 /* Make it present again */
3411 pte = pte_modify(pte, vma->vm_page_prot);
3412 pte = pte_mkyoung(pte);
3413 if (was_writable)
3414 pte = pte_mkwrite(pte);
3415 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3416 update_mmu_cache(vma, vmf->address, vmf->pte);
3417
3418 page = vm_normal_page(vma, vmf->address, pte);
3419 if (!page) {
3420 pte_unmap_unlock(vmf->pte, vmf->ptl);
3421 return 0;
3422 }
3423
3424 /* TODO: handle PTE-mapped THP */
3425 if (PageCompound(page)) {
3426 pte_unmap_unlock(vmf->pte, vmf->ptl);
3427 return 0;
3428 }
3429
3430 /*
3431 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3432 * much anyway since they can be in shared cache state. This misses
3433 * the case where a mapping is writable but the process never writes
3434 * to it but pte_write gets cleared during protection updates and
3435 * pte_dirty has unpredictable behaviour between PTE scan updates,
3436 * background writeback, dirty balancing and application behaviour.
3437 */
3438 if (!pte_write(pte))
3439 flags |= TNF_NO_GROUP;
3440
3441 /*
3442 * Flag if the page is shared between multiple address spaces. This
3443 * is later used when determining whether to group tasks together
3444 */
3445 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3446 flags |= TNF_SHARED;
3447
3448 last_cpupid = page_cpupid_last(page);
3449 page_nid = page_to_nid(page);
3450 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3451 &flags);
3452 pte_unmap_unlock(vmf->pte, vmf->ptl);
3453 if (target_nid == -1) {
3454 put_page(page);
3455 goto out;
3456 }
3457
3458 /* Migrate to the requested node */
3459 migrated = migrate_misplaced_page(page, vma, target_nid);
3460 if (migrated) {
3461 page_nid = target_nid;
3462 flags |= TNF_MIGRATED;
3463 } else
3464 flags |= TNF_MIGRATE_FAIL;
3465
3466out:
3467 if (page_nid != -1)
3468 task_numa_fault(last_cpupid, page_nid, 1, flags);
3469 return 0;
3470}
3471
3472static int create_huge_pmd(struct vm_fault *vmf)
3473{
3474 struct vm_area_struct *vma = vmf->vma;
3475 if (vma_is_anonymous(vma))
3476 return do_huge_pmd_anonymous_page(vmf);
3477 if (vma->vm_ops->pmd_fault)
3478 return vma->vm_ops->pmd_fault(vma, vmf->address, vmf->pmd,
3479 vmf->flags);
3480 return VM_FAULT_FALLBACK;
3481}
3482
3483static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3484{
3485 if (vma_is_anonymous(vmf->vma))
3486 return do_huge_pmd_wp_page(vmf, orig_pmd);
3487 if (vmf->vma->vm_ops->pmd_fault)
3488 return vmf->vma->vm_ops->pmd_fault(vmf->vma, vmf->address,
3489 vmf->pmd, vmf->flags);
3490
3491 /* COW handled on pte level: split pmd */
3492 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3493 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3494
3495 return VM_FAULT_FALLBACK;
3496}
3497
3498static inline bool vma_is_accessible(struct vm_area_struct *vma)
3499{
3500 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3501}
3502
3503/*
3504 * These routines also need to handle stuff like marking pages dirty
3505 * and/or accessed for architectures that don't do it in hardware (most
3506 * RISC architectures). The early dirtying is also good on the i386.
3507 *
3508 * There is also a hook called "update_mmu_cache()" that architectures
3509 * with external mmu caches can use to update those (ie the Sparc or
3510 * PowerPC hashed page tables that act as extended TLBs).
3511 *
3512 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3513 * concurrent faults).
3514 *
3515 * The mmap_sem may have been released depending on flags and our return value.
3516 * See filemap_fault() and __lock_page_or_retry().
3517 */
3518static int handle_pte_fault(struct vm_fault *vmf)
3519{
3520 pte_t entry;
3521
3522 if (unlikely(pmd_none(*vmf->pmd))) {
3523 /*
3524 * Leave __pte_alloc() until later: because vm_ops->fault may
3525 * want to allocate huge page, and if we expose page table
3526 * for an instant, it will be difficult to retract from
3527 * concurrent faults and from rmap lookups.
3528 */
3529 vmf->pte = NULL;
3530 } else {
3531 /* See comment in pte_alloc_one_map() */
3532 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3533 return 0;
3534 /*
3535 * A regular pmd is established and it can't morph into a huge
3536 * pmd from under us anymore at this point because we hold the
3537 * mmap_sem read mode and khugepaged takes it in write mode.
3538 * So now it's safe to run pte_offset_map().
3539 */
3540 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3541 vmf->orig_pte = *vmf->pte;
3542
3543 /*
3544 * some architectures can have larger ptes than wordsize,
3545 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3546 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3547 * atomic accesses. The code below just needs a consistent
3548 * view for the ifs and we later double check anyway with the
3549 * ptl lock held. So here a barrier will do.
3550 */
3551 barrier();
3552 if (pte_none(vmf->orig_pte)) {
3553 pte_unmap(vmf->pte);
3554 vmf->pte = NULL;
3555 }
3556 }
3557
3558 if (!vmf->pte) {
3559 if (vma_is_anonymous(vmf->vma))
3560 return do_anonymous_page(vmf);
3561 else
3562 return do_fault(vmf);
3563 }
3564
3565 if (!pte_present(vmf->orig_pte))
3566 return do_swap_page(vmf);
3567
3568 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3569 return do_numa_page(vmf);
3570
3571 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3572 spin_lock(vmf->ptl);
3573 entry = vmf->orig_pte;
3574 if (unlikely(!pte_same(*vmf->pte, entry)))
3575 goto unlock;
3576 if (vmf->flags & FAULT_FLAG_WRITE) {
3577 if (!pte_write(entry))
3578 return do_wp_page(vmf);
3579 entry = pte_mkdirty(entry);
3580 }
3581 entry = pte_mkyoung(entry);
3582 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3583 vmf->flags & FAULT_FLAG_WRITE)) {
3584 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3585 } else {
3586 /*
3587 * This is needed only for protection faults but the arch code
3588 * is not yet telling us if this is a protection fault or not.
3589 * This still avoids useless tlb flushes for .text page faults
3590 * with threads.
3591 */
3592 if (vmf->flags & FAULT_FLAG_WRITE)
3593 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3594 }
3595unlock:
3596 pte_unmap_unlock(vmf->pte, vmf->ptl);
3597 return 0;
3598}
3599
3600/*
3601 * By the time we get here, we already hold the mm semaphore
3602 *
3603 * The mmap_sem may have been released depending on flags and our
3604 * return value. See filemap_fault() and __lock_page_or_retry().
3605 */
3606static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3607 unsigned int flags)
3608{
3609 struct vm_fault vmf = {
3610 .vma = vma,
3611 .address = address & PAGE_MASK,
3612 .flags = flags,
3613 .pgoff = linear_page_index(vma, address),
3614 .gfp_mask = __get_fault_gfp_mask(vma),
3615 };
3616 struct mm_struct *mm = vma->vm_mm;
3617 pgd_t *pgd;
3618 pud_t *pud;
3619
3620 pgd = pgd_offset(mm, address);
3621 pud = pud_alloc(mm, pgd, address);
3622 if (!pud)
3623 return VM_FAULT_OOM;
3624 vmf.pmd = pmd_alloc(mm, pud, address);
3625 if (!vmf.pmd)
3626 return VM_FAULT_OOM;
3627 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3628 int ret = create_huge_pmd(&vmf);
3629 if (!(ret & VM_FAULT_FALLBACK))
3630 return ret;
3631 } else {
3632 pmd_t orig_pmd = *vmf.pmd;
3633 int ret;
3634
3635 barrier();
3636 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3637 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3638 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3639
3640 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3641 !pmd_write(orig_pmd)) {
3642 ret = wp_huge_pmd(&vmf, orig_pmd);
3643 if (!(ret & VM_FAULT_FALLBACK))
3644 return ret;
3645 } else {
3646 huge_pmd_set_accessed(&vmf, orig_pmd);
3647 return 0;
3648 }
3649 }
3650 }
3651
3652 return handle_pte_fault(&vmf);
3653}
3654
3655/*
3656 * By the time we get here, we already hold the mm semaphore
3657 *
3658 * The mmap_sem may have been released depending on flags and our
3659 * return value. See filemap_fault() and __lock_page_or_retry().
3660 */
3661int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3662 unsigned int flags)
3663{
3664 int ret;
3665
3666 __set_current_state(TASK_RUNNING);
3667
3668 count_vm_event(PGFAULT);
3669 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3670
3671 /* do counter updates before entering really critical section. */
3672 check_sync_rss_stat(current);
3673
3674 /*
3675 * Enable the memcg OOM handling for faults triggered in user
3676 * space. Kernel faults are handled more gracefully.
3677 */
3678 if (flags & FAULT_FLAG_USER)
3679 mem_cgroup_oom_enable();
3680
3681 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3682 flags & FAULT_FLAG_INSTRUCTION,
3683 flags & FAULT_FLAG_REMOTE))
3684 return VM_FAULT_SIGSEGV;
3685
3686 if (unlikely(is_vm_hugetlb_page(vma)))
3687 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3688 else
3689 ret = __handle_mm_fault(vma, address, flags);
3690
3691 if (flags & FAULT_FLAG_USER) {
3692 mem_cgroup_oom_disable();
3693 /*
3694 * The task may have entered a memcg OOM situation but
3695 * if the allocation error was handled gracefully (no
3696 * VM_FAULT_OOM), there is no need to kill anything.
3697 * Just clean up the OOM state peacefully.
3698 */
3699 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3700 mem_cgroup_oom_synchronize(false);
3701 }
3702
3703 /*
3704 * This mm has been already reaped by the oom reaper and so the
3705 * refault cannot be trusted in general. Anonymous refaults would
3706 * lose data and give a zero page instead e.g. This is especially
3707 * problem for use_mm() because regular tasks will just die and
3708 * the corrupted data will not be visible anywhere while kthread
3709 * will outlive the oom victim and potentially propagate the date
3710 * further.
3711 */
3712 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3713 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3714 ret = VM_FAULT_SIGBUS;
3715
3716 return ret;
3717}
3718EXPORT_SYMBOL_GPL(handle_mm_fault);
3719
3720#ifndef __PAGETABLE_PUD_FOLDED
3721/*
3722 * Allocate page upper directory.
3723 * We've already handled the fast-path in-line.
3724 */
3725int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3726{
3727 pud_t *new = pud_alloc_one(mm, address);
3728 if (!new)
3729 return -ENOMEM;
3730
3731 smp_wmb(); /* See comment in __pte_alloc */
3732
3733 spin_lock(&mm->page_table_lock);
3734 if (pgd_present(*pgd)) /* Another has populated it */
3735 pud_free(mm, new);
3736 else
3737 pgd_populate(mm, pgd, new);
3738 spin_unlock(&mm->page_table_lock);
3739 return 0;
3740}
3741#endif /* __PAGETABLE_PUD_FOLDED */
3742
3743#ifndef __PAGETABLE_PMD_FOLDED
3744/*
3745 * Allocate page middle directory.
3746 * We've already handled the fast-path in-line.
3747 */
3748int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3749{
3750 pmd_t *new = pmd_alloc_one(mm, address);
3751 if (!new)
3752 return -ENOMEM;
3753
3754 smp_wmb(); /* See comment in __pte_alloc */
3755
3756 spin_lock(&mm->page_table_lock);
3757#ifndef __ARCH_HAS_4LEVEL_HACK
3758 if (!pud_present(*pud)) {
3759 mm_inc_nr_pmds(mm);
3760 pud_populate(mm, pud, new);
3761 } else /* Another has populated it */
3762 pmd_free(mm, new);
3763#else
3764 if (!pgd_present(*pud)) {
3765 mm_inc_nr_pmds(mm);
3766 pgd_populate(mm, pud, new);
3767 } else /* Another has populated it */
3768 pmd_free(mm, new);
3769#endif /* __ARCH_HAS_4LEVEL_HACK */
3770 spin_unlock(&mm->page_table_lock);
3771 return 0;
3772}
3773#endif /* __PAGETABLE_PMD_FOLDED */
3774
3775static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3776 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3777{
3778 pgd_t *pgd;
3779 pud_t *pud;
3780 pmd_t *pmd;
3781 pte_t *ptep;
3782
3783 pgd = pgd_offset(mm, address);
3784 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3785 goto out;
3786
3787 pud = pud_offset(pgd, address);
3788 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3789 goto out;
3790
3791 pmd = pmd_offset(pud, address);
3792 VM_BUG_ON(pmd_trans_huge(*pmd));
3793
3794 if (pmd_huge(*pmd)) {
3795 if (!pmdpp)
3796 goto out;
3797
3798 *ptlp = pmd_lock(mm, pmd);
3799 if (pmd_huge(*pmd)) {
3800 *pmdpp = pmd;
3801 return 0;
3802 }
3803 spin_unlock(*ptlp);
3804 }
3805
3806 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3807 goto out;
3808
3809 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3810 if (!ptep)
3811 goto out;
3812 if (!pte_present(*ptep))
3813 goto unlock;
3814 *ptepp = ptep;
3815 return 0;
3816unlock:
3817 pte_unmap_unlock(ptep, *ptlp);
3818out:
3819 return -EINVAL;
3820}
3821
3822static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3823 pte_t **ptepp, spinlock_t **ptlp)
3824{
3825 int res;
3826
3827 /* (void) is needed to make gcc happy */
3828 (void) __cond_lock(*ptlp,
3829 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
3830 ptlp)));
3831 return res;
3832}
3833
3834int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3835 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3836{
3837 int res;
3838
3839 /* (void) is needed to make gcc happy */
3840 (void) __cond_lock(*ptlp,
3841 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
3842 ptlp)));
3843 return res;
3844}
3845EXPORT_SYMBOL(follow_pte_pmd);
3846
3847/**
3848 * follow_pfn - look up PFN at a user virtual address
3849 * @vma: memory mapping
3850 * @address: user virtual address
3851 * @pfn: location to store found PFN
3852 *
3853 * Only IO mappings and raw PFN mappings are allowed.
3854 *
3855 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3856 */
3857int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3858 unsigned long *pfn)
3859{
3860 int ret = -EINVAL;
3861 spinlock_t *ptl;
3862 pte_t *ptep;
3863
3864 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3865 return ret;
3866
3867 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3868 if (ret)
3869 return ret;
3870 *pfn = pte_pfn(*ptep);
3871 pte_unmap_unlock(ptep, ptl);
3872 return 0;
3873}
3874EXPORT_SYMBOL(follow_pfn);
3875
3876#ifdef CONFIG_HAVE_IOREMAP_PROT
3877int follow_phys(struct vm_area_struct *vma,
3878 unsigned long address, unsigned int flags,
3879 unsigned long *prot, resource_size_t *phys)
3880{
3881 int ret = -EINVAL;
3882 pte_t *ptep, pte;
3883 spinlock_t *ptl;
3884
3885 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3886 goto out;
3887
3888 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3889 goto out;
3890 pte = *ptep;
3891
3892 if ((flags & FOLL_WRITE) && !pte_write(pte))
3893 goto unlock;
3894
3895 *prot = pgprot_val(pte_pgprot(pte));
3896 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3897
3898 ret = 0;
3899unlock:
3900 pte_unmap_unlock(ptep, ptl);
3901out:
3902 return ret;
3903}
3904
3905int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3906 void *buf, int len, int write)
3907{
3908 resource_size_t phys_addr;
3909 unsigned long prot = 0;
3910 void __iomem *maddr;
3911 int offset = addr & (PAGE_SIZE-1);
3912
3913 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3914 return -EINVAL;
3915
3916 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3917 if (write)
3918 memcpy_toio(maddr + offset, buf, len);
3919 else
3920 memcpy_fromio(buf, maddr + offset, len);
3921 iounmap(maddr);
3922
3923 return len;
3924}
3925EXPORT_SYMBOL_GPL(generic_access_phys);
3926#endif
3927
3928/*
3929 * Access another process' address space as given in mm. If non-NULL, use the
3930 * given task for page fault accounting.
3931 */
3932int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3933 unsigned long addr, void *buf, int len, unsigned int gup_flags)
3934{
3935 struct vm_area_struct *vma;
3936 void *old_buf = buf;
3937 int write = gup_flags & FOLL_WRITE;
3938
3939 down_read(&mm->mmap_sem);
3940 /* ignore errors, just check how much was successfully transferred */
3941 while (len) {
3942 int bytes, ret, offset;
3943 void *maddr;
3944 struct page *page = NULL;
3945
3946 ret = get_user_pages_remote(tsk, mm, addr, 1,
3947 gup_flags, &page, &vma, NULL);
3948 if (ret <= 0) {
3949#ifndef CONFIG_HAVE_IOREMAP_PROT
3950 break;
3951#else
3952 /*
3953 * Check if this is a VM_IO | VM_PFNMAP VMA, which
3954 * we can access using slightly different code.
3955 */
3956 vma = find_vma(mm, addr);
3957 if (!vma || vma->vm_start > addr)
3958 break;
3959 if (vma->vm_ops && vma->vm_ops->access)
3960 ret = vma->vm_ops->access(vma, addr, buf,
3961 len, write);
3962 if (ret <= 0)
3963 break;
3964 bytes = ret;
3965#endif
3966 } else {
3967 bytes = len;
3968 offset = addr & (PAGE_SIZE-1);
3969 if (bytes > PAGE_SIZE-offset)
3970 bytes = PAGE_SIZE-offset;
3971
3972 maddr = kmap(page);
3973 if (write) {
3974 copy_to_user_page(vma, page, addr,
3975 maddr + offset, buf, bytes);
3976 set_page_dirty_lock(page);
3977 } else {
3978 copy_from_user_page(vma, page, addr,
3979 buf, maddr + offset, bytes);
3980 }
3981 kunmap(page);
3982 put_page(page);
3983 }
3984 len -= bytes;
3985 buf += bytes;
3986 addr += bytes;
3987 }
3988 up_read(&mm->mmap_sem);
3989
3990 return buf - old_buf;
3991}
3992
3993/**
3994 * access_remote_vm - access another process' address space
3995 * @mm: the mm_struct of the target address space
3996 * @addr: start address to access
3997 * @buf: source or destination buffer
3998 * @len: number of bytes to transfer
3999 * @gup_flags: flags modifying lookup behaviour
4000 *
4001 * The caller must hold a reference on @mm.
4002 */
4003int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4004 void *buf, int len, unsigned int gup_flags)
4005{
4006 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4007}
4008
4009/*
4010 * Access another process' address space.
4011 * Source/target buffer must be kernel space,
4012 * Do not walk the page table directly, use get_user_pages
4013 */
4014int access_process_vm(struct task_struct *tsk, unsigned long addr,
4015 void *buf, int len, unsigned int gup_flags)
4016{
4017 struct mm_struct *mm;
4018 int ret;
4019
4020 mm = get_task_mm(tsk);
4021 if (!mm)
4022 return 0;
4023
4024 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4025
4026 mmput(mm);
4027
4028 return ret;
4029}
4030EXPORT_SYMBOL_GPL(access_process_vm);
4031
4032/*
4033 * Print the name of a VMA.
4034 */
4035void print_vma_addr(char *prefix, unsigned long ip)
4036{
4037 struct mm_struct *mm = current->mm;
4038 struct vm_area_struct *vma;
4039
4040 /*
4041 * Do not print if we are in atomic
4042 * contexts (in exception stacks, etc.):
4043 */
4044 if (preempt_count())
4045 return;
4046
4047 down_read(&mm->mmap_sem);
4048 vma = find_vma(mm, ip);
4049 if (vma && vma->vm_file) {
4050 struct file *f = vma->vm_file;
4051 char *buf = (char *)__get_free_page(GFP_KERNEL);
4052 if (buf) {
4053 char *p;
4054
4055 p = file_path(f, buf, PAGE_SIZE);
4056 if (IS_ERR(p))
4057 p = "?";
4058 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4059 vma->vm_start,
4060 vma->vm_end - vma->vm_start);
4061 free_page((unsigned long)buf);
4062 }
4063 }
4064 up_read(&mm->mmap_sem);
4065}
4066
4067#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4068void __might_fault(const char *file, int line)
4069{
4070 /*
4071 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4072 * holding the mmap_sem, this is safe because kernel memory doesn't
4073 * get paged out, therefore we'll never actually fault, and the
4074 * below annotations will generate false positives.
4075 */
4076 if (segment_eq(get_fs(), KERNEL_DS))
4077 return;
4078 if (pagefault_disabled())
4079 return;
4080 __might_sleep(file, line, 0);
4081#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4082 if (current->mm)
4083 might_lock_read(¤t->mm->mmap_sem);
4084#endif
4085}
4086EXPORT_SYMBOL(__might_fault);
4087#endif
4088
4089#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4090static void clear_gigantic_page(struct page *page,
4091 unsigned long addr,
4092 unsigned int pages_per_huge_page)
4093{
4094 int i;
4095 struct page *p = page;
4096
4097 might_sleep();
4098 for (i = 0; i < pages_per_huge_page;
4099 i++, p = mem_map_next(p, page, i)) {
4100 cond_resched();
4101 clear_user_highpage(p, addr + i * PAGE_SIZE);
4102 }
4103}
4104void clear_huge_page(struct page *page,
4105 unsigned long addr, unsigned int pages_per_huge_page)
4106{
4107 int i;
4108
4109 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4110 clear_gigantic_page(page, addr, pages_per_huge_page);
4111 return;
4112 }
4113
4114 might_sleep();
4115 for (i = 0; i < pages_per_huge_page; i++) {
4116 cond_resched();
4117 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4118 }
4119}
4120
4121static void copy_user_gigantic_page(struct page *dst, struct page *src,
4122 unsigned long addr,
4123 struct vm_area_struct *vma,
4124 unsigned int pages_per_huge_page)
4125{
4126 int i;
4127 struct page *dst_base = dst;
4128 struct page *src_base = src;
4129
4130 for (i = 0; i < pages_per_huge_page; ) {
4131 cond_resched();
4132 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4133
4134 i++;
4135 dst = mem_map_next(dst, dst_base, i);
4136 src = mem_map_next(src, src_base, i);
4137 }
4138}
4139
4140void copy_user_huge_page(struct page *dst, struct page *src,
4141 unsigned long addr, struct vm_area_struct *vma,
4142 unsigned int pages_per_huge_page)
4143{
4144 int i;
4145
4146 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4147 copy_user_gigantic_page(dst, src, addr, vma,
4148 pages_per_huge_page);
4149 return;
4150 }
4151
4152 might_sleep();
4153 for (i = 0; i < pages_per_huge_page; i++) {
4154 cond_resched();
4155 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4156 }
4157}
4158#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4159
4160#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4161
4162static struct kmem_cache *page_ptl_cachep;
4163
4164void __init ptlock_cache_init(void)
4165{
4166 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4167 SLAB_PANIC, NULL);
4168}
4169
4170bool ptlock_alloc(struct page *page)
4171{
4172 spinlock_t *ptl;
4173
4174 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4175 if (!ptl)
4176 return false;
4177 page->ptl = ptl;
4178 return true;
4179}
4180
4181void ptlock_free(struct page *page)
4182{
4183 kmem_cache_free(page_ptl_cachep, page->ptl);
4184}
4185#endif