Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/mm/memory.c
4 *
5 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
6 */
7
8/*
9 * demand-loading started 01.12.91 - seems it is high on the list of
10 * things wanted, and it should be easy to implement. - Linus
11 */
12
13/*
14 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15 * pages started 02.12.91, seems to work. - Linus.
16 *
17 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18 * would have taken more than the 6M I have free, but it worked well as
19 * far as I could see.
20 *
21 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22 */
23
24/*
25 * Real VM (paging to/from disk) started 18.12.91. Much more work and
26 * thought has to go into this. Oh, well..
27 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
28 * Found it. Everything seems to work now.
29 * 20.12.91 - Ok, making the swap-device changeable like the root.
30 */
31
32/*
33 * 05.04.94 - Multi-page memory management added for v1.1.
34 * Idea by Alex Bligh (alex@cconcepts.co.uk)
35 *
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * (Gerhard.Wichert@pdb.siemens.de)
38 *
39 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40 */
41
42#include <linux/kernel_stat.h>
43#include <linux/mm.h>
44#include <linux/sched/mm.h>
45#include <linux/sched/coredump.h>
46#include <linux/sched/numa_balancing.h>
47#include <linux/sched/task.h>
48#include <linux/hugetlb.h>
49#include <linux/mman.h>
50#include <linux/swap.h>
51#include <linux/highmem.h>
52#include <linux/pagemap.h>
53#include <linux/memremap.h>
54#include <linux/ksm.h>
55#include <linux/rmap.h>
56#include <linux/export.h>
57#include <linux/delayacct.h>
58#include <linux/init.h>
59#include <linux/pfn_t.h>
60#include <linux/writeback.h>
61#include <linux/memcontrol.h>
62#include <linux/mmu_notifier.h>
63#include <linux/swapops.h>
64#include <linux/elf.h>
65#include <linux/gfp.h>
66#include <linux/migrate.h>
67#include <linux/string.h>
68#include <linux/dma-debug.h>
69#include <linux/debugfs.h>
70#include <linux/userfaultfd_k.h>
71#include <linux/dax.h>
72#include <linux/oom.h>
73#include <linux/numa.h>
74
75#include <asm/io.h>
76#include <asm/mmu_context.h>
77#include <asm/pgalloc.h>
78#include <linux/uaccess.h>
79#include <asm/tlb.h>
80#include <asm/tlbflush.h>
81#include <asm/pgtable.h>
82
83#include "internal.h"
84
85#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
86#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
87#endif
88
89#ifndef CONFIG_NEED_MULTIPLE_NODES
90/* use the per-pgdat data instead for discontigmem - mbligh */
91unsigned long max_mapnr;
92EXPORT_SYMBOL(max_mapnr);
93
94struct page *mem_map;
95EXPORT_SYMBOL(mem_map);
96#endif
97
98/*
99 * A number of key systems in x86 including ioremap() rely on the assumption
100 * that high_memory defines the upper bound on direct map memory, then end
101 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
102 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
103 * and ZONE_HIGHMEM.
104 */
105void *high_memory;
106EXPORT_SYMBOL(high_memory);
107
108/*
109 * Randomize the address space (stacks, mmaps, brk, etc.).
110 *
111 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
112 * as ancient (libc5 based) binaries can segfault. )
113 */
114int randomize_va_space __read_mostly =
115#ifdef CONFIG_COMPAT_BRK
116 1;
117#else
118 2;
119#endif
120
121static int __init disable_randmaps(char *s)
122{
123 randomize_va_space = 0;
124 return 1;
125}
126__setup("norandmaps", disable_randmaps);
127
128unsigned long zero_pfn __read_mostly;
129EXPORT_SYMBOL(zero_pfn);
130
131unsigned long highest_memmap_pfn __read_mostly;
132
133/*
134 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
135 */
136static int __init init_zero_pfn(void)
137{
138 zero_pfn = page_to_pfn(ZERO_PAGE(0));
139 return 0;
140}
141core_initcall(init_zero_pfn);
142
143
144#if defined(SPLIT_RSS_COUNTING)
145
146void sync_mm_rss(struct mm_struct *mm)
147{
148 int i;
149
150 for (i = 0; i < NR_MM_COUNTERS; i++) {
151 if (current->rss_stat.count[i]) {
152 add_mm_counter(mm, i, current->rss_stat.count[i]);
153 current->rss_stat.count[i] = 0;
154 }
155 }
156 current->rss_stat.events = 0;
157}
158
159static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
160{
161 struct task_struct *task = current;
162
163 if (likely(task->mm == mm))
164 task->rss_stat.count[member] += val;
165 else
166 add_mm_counter(mm, member, val);
167}
168#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
169#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
170
171/* sync counter once per 64 page faults */
172#define TASK_RSS_EVENTS_THRESH (64)
173static void check_sync_rss_stat(struct task_struct *task)
174{
175 if (unlikely(task != current))
176 return;
177 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
178 sync_mm_rss(task->mm);
179}
180#else /* SPLIT_RSS_COUNTING */
181
182#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
183#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
184
185static void check_sync_rss_stat(struct task_struct *task)
186{
187}
188
189#endif /* SPLIT_RSS_COUNTING */
190
191/*
192 * Note: this doesn't free the actual pages themselves. That
193 * has been handled earlier when unmapping all the memory regions.
194 */
195static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
196 unsigned long addr)
197{
198 pgtable_t token = pmd_pgtable(*pmd);
199 pmd_clear(pmd);
200 pte_free_tlb(tlb, token, addr);
201 mm_dec_nr_ptes(tlb->mm);
202}
203
204static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
207{
208 pmd_t *pmd;
209 unsigned long next;
210 unsigned long start;
211
212 start = addr;
213 pmd = pmd_offset(pud, addr);
214 do {
215 next = pmd_addr_end(addr, end);
216 if (pmd_none_or_clear_bad(pmd))
217 continue;
218 free_pte_range(tlb, pmd, addr);
219 } while (pmd++, addr = next, addr != end);
220
221 start &= PUD_MASK;
222 if (start < floor)
223 return;
224 if (ceiling) {
225 ceiling &= PUD_MASK;
226 if (!ceiling)
227 return;
228 }
229 if (end - 1 > ceiling - 1)
230 return;
231
232 pmd = pmd_offset(pud, start);
233 pud_clear(pud);
234 pmd_free_tlb(tlb, pmd, start);
235 mm_dec_nr_pmds(tlb->mm);
236}
237
238static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
239 unsigned long addr, unsigned long end,
240 unsigned long floor, unsigned long ceiling)
241{
242 pud_t *pud;
243 unsigned long next;
244 unsigned long start;
245
246 start = addr;
247 pud = pud_offset(p4d, addr);
248 do {
249 next = pud_addr_end(addr, end);
250 if (pud_none_or_clear_bad(pud))
251 continue;
252 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
253 } while (pud++, addr = next, addr != end);
254
255 start &= P4D_MASK;
256 if (start < floor)
257 return;
258 if (ceiling) {
259 ceiling &= P4D_MASK;
260 if (!ceiling)
261 return;
262 }
263 if (end - 1 > ceiling - 1)
264 return;
265
266 pud = pud_offset(p4d, start);
267 p4d_clear(p4d);
268 pud_free_tlb(tlb, pud, start);
269 mm_dec_nr_puds(tlb->mm);
270}
271
272static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
273 unsigned long addr, unsigned long end,
274 unsigned long floor, unsigned long ceiling)
275{
276 p4d_t *p4d;
277 unsigned long next;
278 unsigned long start;
279
280 start = addr;
281 p4d = p4d_offset(pgd, addr);
282 do {
283 next = p4d_addr_end(addr, end);
284 if (p4d_none_or_clear_bad(p4d))
285 continue;
286 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
287 } while (p4d++, addr = next, addr != end);
288
289 start &= PGDIR_MASK;
290 if (start < floor)
291 return;
292 if (ceiling) {
293 ceiling &= PGDIR_MASK;
294 if (!ceiling)
295 return;
296 }
297 if (end - 1 > ceiling - 1)
298 return;
299
300 p4d = p4d_offset(pgd, start);
301 pgd_clear(pgd);
302 p4d_free_tlb(tlb, p4d, start);
303}
304
305/*
306 * This function frees user-level page tables of a process.
307 */
308void free_pgd_range(struct mmu_gather *tlb,
309 unsigned long addr, unsigned long end,
310 unsigned long floor, unsigned long ceiling)
311{
312 pgd_t *pgd;
313 unsigned long next;
314
315 /*
316 * The next few lines have given us lots of grief...
317 *
318 * Why are we testing PMD* at this top level? Because often
319 * there will be no work to do at all, and we'd prefer not to
320 * go all the way down to the bottom just to discover that.
321 *
322 * Why all these "- 1"s? Because 0 represents both the bottom
323 * of the address space and the top of it (using -1 for the
324 * top wouldn't help much: the masks would do the wrong thing).
325 * The rule is that addr 0 and floor 0 refer to the bottom of
326 * the address space, but end 0 and ceiling 0 refer to the top
327 * Comparisons need to use "end - 1" and "ceiling - 1" (though
328 * that end 0 case should be mythical).
329 *
330 * Wherever addr is brought up or ceiling brought down, we must
331 * be careful to reject "the opposite 0" before it confuses the
332 * subsequent tests. But what about where end is brought down
333 * by PMD_SIZE below? no, end can't go down to 0 there.
334 *
335 * Whereas we round start (addr) and ceiling down, by different
336 * masks at different levels, in order to test whether a table
337 * now has no other vmas using it, so can be freed, we don't
338 * bother to round floor or end up - the tests don't need that.
339 */
340
341 addr &= PMD_MASK;
342 if (addr < floor) {
343 addr += PMD_SIZE;
344 if (!addr)
345 return;
346 }
347 if (ceiling) {
348 ceiling &= PMD_MASK;
349 if (!ceiling)
350 return;
351 }
352 if (end - 1 > ceiling - 1)
353 end -= PMD_SIZE;
354 if (addr > end - 1)
355 return;
356 /*
357 * We add page table cache pages with PAGE_SIZE,
358 * (see pte_free_tlb()), flush the tlb if we need
359 */
360 tlb_change_page_size(tlb, PAGE_SIZE);
361 pgd = pgd_offset(tlb->mm, addr);
362 do {
363 next = pgd_addr_end(addr, end);
364 if (pgd_none_or_clear_bad(pgd))
365 continue;
366 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
367 } while (pgd++, addr = next, addr != end);
368}
369
370void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
371 unsigned long floor, unsigned long ceiling)
372{
373 while (vma) {
374 struct vm_area_struct *next = vma->vm_next;
375 unsigned long addr = vma->vm_start;
376
377 /*
378 * Hide vma from rmap and truncate_pagecache before freeing
379 * pgtables
380 */
381 unlink_anon_vmas(vma);
382 unlink_file_vma(vma);
383
384 if (is_vm_hugetlb_page(vma)) {
385 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
386 floor, next ? next->vm_start : ceiling);
387 } else {
388 /*
389 * Optimization: gather nearby vmas into one call down
390 */
391 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
392 && !is_vm_hugetlb_page(next)) {
393 vma = next;
394 next = vma->vm_next;
395 unlink_anon_vmas(vma);
396 unlink_file_vma(vma);
397 }
398 free_pgd_range(tlb, addr, vma->vm_end,
399 floor, next ? next->vm_start : ceiling);
400 }
401 vma = next;
402 }
403}
404
405int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
406{
407 spinlock_t *ptl;
408 pgtable_t new = pte_alloc_one(mm);
409 if (!new)
410 return -ENOMEM;
411
412 /*
413 * Ensure all pte setup (eg. pte page lock and page clearing) are
414 * visible before the pte is made visible to other CPUs by being
415 * put into page tables.
416 *
417 * The other side of the story is the pointer chasing in the page
418 * table walking code (when walking the page table without locking;
419 * ie. most of the time). Fortunately, these data accesses consist
420 * of a chain of data-dependent loads, meaning most CPUs (alpha
421 * being the notable exception) will already guarantee loads are
422 * seen in-order. See the alpha page table accessors for the
423 * smp_read_barrier_depends() barriers in page table walking code.
424 */
425 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
426
427 ptl = pmd_lock(mm, pmd);
428 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
429 mm_inc_nr_ptes(mm);
430 pmd_populate(mm, pmd, new);
431 new = NULL;
432 }
433 spin_unlock(ptl);
434 if (new)
435 pte_free(mm, new);
436 return 0;
437}
438
439int __pte_alloc_kernel(pmd_t *pmd)
440{
441 pte_t *new = pte_alloc_one_kernel(&init_mm);
442 if (!new)
443 return -ENOMEM;
444
445 smp_wmb(); /* See comment in __pte_alloc */
446
447 spin_lock(&init_mm.page_table_lock);
448 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
449 pmd_populate_kernel(&init_mm, pmd, new);
450 new = NULL;
451 }
452 spin_unlock(&init_mm.page_table_lock);
453 if (new)
454 pte_free_kernel(&init_mm, new);
455 return 0;
456}
457
458static inline void init_rss_vec(int *rss)
459{
460 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
461}
462
463static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
464{
465 int i;
466
467 if (current->mm == mm)
468 sync_mm_rss(mm);
469 for (i = 0; i < NR_MM_COUNTERS; i++)
470 if (rss[i])
471 add_mm_counter(mm, i, rss[i]);
472}
473
474/*
475 * This function is called to print an error when a bad pte
476 * is found. For example, we might have a PFN-mapped pte in
477 * a region that doesn't allow it.
478 *
479 * The calling function must still handle the error.
480 */
481static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
482 pte_t pte, struct page *page)
483{
484 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
485 p4d_t *p4d = p4d_offset(pgd, addr);
486 pud_t *pud = pud_offset(p4d, addr);
487 pmd_t *pmd = pmd_offset(pud, addr);
488 struct address_space *mapping;
489 pgoff_t index;
490 static unsigned long resume;
491 static unsigned long nr_shown;
492 static unsigned long nr_unshown;
493
494 /*
495 * Allow a burst of 60 reports, then keep quiet for that minute;
496 * or allow a steady drip of one report per second.
497 */
498 if (nr_shown == 60) {
499 if (time_before(jiffies, resume)) {
500 nr_unshown++;
501 return;
502 }
503 if (nr_unshown) {
504 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
505 nr_unshown);
506 nr_unshown = 0;
507 }
508 nr_shown = 0;
509 }
510 if (nr_shown++ == 0)
511 resume = jiffies + 60 * HZ;
512
513 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
514 index = linear_page_index(vma, addr);
515
516 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
517 current->comm,
518 (long long)pte_val(pte), (long long)pmd_val(*pmd));
519 if (page)
520 dump_page(page, "bad pte");
521 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
522 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
523 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
524 vma->vm_file,
525 vma->vm_ops ? vma->vm_ops->fault : NULL,
526 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
527 mapping ? mapping->a_ops->readpage : NULL);
528 dump_stack();
529 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
530}
531
532/*
533 * vm_normal_page -- This function gets the "struct page" associated with a pte.
534 *
535 * "Special" mappings do not wish to be associated with a "struct page" (either
536 * it doesn't exist, or it exists but they don't want to touch it). In this
537 * case, NULL is returned here. "Normal" mappings do have a struct page.
538 *
539 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
540 * pte bit, in which case this function is trivial. Secondly, an architecture
541 * may not have a spare pte bit, which requires a more complicated scheme,
542 * described below.
543 *
544 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
545 * special mapping (even if there are underlying and valid "struct pages").
546 * COWed pages of a VM_PFNMAP are always normal.
547 *
548 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
549 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
550 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
551 * mapping will always honor the rule
552 *
553 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
554 *
555 * And for normal mappings this is false.
556 *
557 * This restricts such mappings to be a linear translation from virtual address
558 * to pfn. To get around this restriction, we allow arbitrary mappings so long
559 * as the vma is not a COW mapping; in that case, we know that all ptes are
560 * special (because none can have been COWed).
561 *
562 *
563 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
564 *
565 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
566 * page" backing, however the difference is that _all_ pages with a struct
567 * page (that is, those where pfn_valid is true) are refcounted and considered
568 * normal pages by the VM. The disadvantage is that pages are refcounted
569 * (which can be slower and simply not an option for some PFNMAP users). The
570 * advantage is that we don't have to follow the strict linearity rule of
571 * PFNMAP mappings in order to support COWable mappings.
572 *
573 */
574struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
575 pte_t pte)
576{
577 unsigned long pfn = pte_pfn(pte);
578
579 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
580 if (likely(!pte_special(pte)))
581 goto check_pfn;
582 if (vma->vm_ops && vma->vm_ops->find_special_page)
583 return vma->vm_ops->find_special_page(vma, addr);
584 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
585 return NULL;
586 if (is_zero_pfn(pfn))
587 return NULL;
588 if (pte_devmap(pte))
589 return NULL;
590
591 print_bad_pte(vma, addr, pte, NULL);
592 return NULL;
593 }
594
595 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
596
597 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
598 if (vma->vm_flags & VM_MIXEDMAP) {
599 if (!pfn_valid(pfn))
600 return NULL;
601 goto out;
602 } else {
603 unsigned long off;
604 off = (addr - vma->vm_start) >> PAGE_SHIFT;
605 if (pfn == vma->vm_pgoff + off)
606 return NULL;
607 if (!is_cow_mapping(vma->vm_flags))
608 return NULL;
609 }
610 }
611
612 if (is_zero_pfn(pfn))
613 return NULL;
614
615check_pfn:
616 if (unlikely(pfn > highest_memmap_pfn)) {
617 print_bad_pte(vma, addr, pte, NULL);
618 return NULL;
619 }
620
621 /*
622 * NOTE! We still have PageReserved() pages in the page tables.
623 * eg. VDSO mappings can cause them to exist.
624 */
625out:
626 return pfn_to_page(pfn);
627}
628
629#ifdef CONFIG_TRANSPARENT_HUGEPAGE
630struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
631 pmd_t pmd)
632{
633 unsigned long pfn = pmd_pfn(pmd);
634
635 /*
636 * There is no pmd_special() but there may be special pmds, e.g.
637 * in a direct-access (dax) mapping, so let's just replicate the
638 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
639 */
640 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
641 if (vma->vm_flags & VM_MIXEDMAP) {
642 if (!pfn_valid(pfn))
643 return NULL;
644 goto out;
645 } else {
646 unsigned long off;
647 off = (addr - vma->vm_start) >> PAGE_SHIFT;
648 if (pfn == vma->vm_pgoff + off)
649 return NULL;
650 if (!is_cow_mapping(vma->vm_flags))
651 return NULL;
652 }
653 }
654
655 if (pmd_devmap(pmd))
656 return NULL;
657 if (is_zero_pfn(pfn))
658 return NULL;
659 if (unlikely(pfn > highest_memmap_pfn))
660 return NULL;
661
662 /*
663 * NOTE! We still have PageReserved() pages in the page tables.
664 * eg. VDSO mappings can cause them to exist.
665 */
666out:
667 return pfn_to_page(pfn);
668}
669#endif
670
671/*
672 * copy one vm_area from one task to the other. Assumes the page tables
673 * already present in the new task to be cleared in the whole range
674 * covered by this vma.
675 */
676
677static inline unsigned long
678copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
679 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
680 unsigned long addr, int *rss)
681{
682 unsigned long vm_flags = vma->vm_flags;
683 pte_t pte = *src_pte;
684 struct page *page;
685
686 /* pte contains position in swap or file, so copy. */
687 if (unlikely(!pte_present(pte))) {
688 swp_entry_t entry = pte_to_swp_entry(pte);
689
690 if (likely(!non_swap_entry(entry))) {
691 if (swap_duplicate(entry) < 0)
692 return entry.val;
693
694 /* make sure dst_mm is on swapoff's mmlist. */
695 if (unlikely(list_empty(&dst_mm->mmlist))) {
696 spin_lock(&mmlist_lock);
697 if (list_empty(&dst_mm->mmlist))
698 list_add(&dst_mm->mmlist,
699 &src_mm->mmlist);
700 spin_unlock(&mmlist_lock);
701 }
702 rss[MM_SWAPENTS]++;
703 } else if (is_migration_entry(entry)) {
704 page = migration_entry_to_page(entry);
705
706 rss[mm_counter(page)]++;
707
708 if (is_write_migration_entry(entry) &&
709 is_cow_mapping(vm_flags)) {
710 /*
711 * COW mappings require pages in both
712 * parent and child to be set to read.
713 */
714 make_migration_entry_read(&entry);
715 pte = swp_entry_to_pte(entry);
716 if (pte_swp_soft_dirty(*src_pte))
717 pte = pte_swp_mksoft_dirty(pte);
718 set_pte_at(src_mm, addr, src_pte, pte);
719 }
720 } else if (is_device_private_entry(entry)) {
721 page = device_private_entry_to_page(entry);
722
723 /*
724 * Update rss count even for unaddressable pages, as
725 * they should treated just like normal pages in this
726 * respect.
727 *
728 * We will likely want to have some new rss counters
729 * for unaddressable pages, at some point. But for now
730 * keep things as they are.
731 */
732 get_page(page);
733 rss[mm_counter(page)]++;
734 page_dup_rmap(page, false);
735
736 /*
737 * We do not preserve soft-dirty information, because so
738 * far, checkpoint/restore is the only feature that
739 * requires that. And checkpoint/restore does not work
740 * when a device driver is involved (you cannot easily
741 * save and restore device driver state).
742 */
743 if (is_write_device_private_entry(entry) &&
744 is_cow_mapping(vm_flags)) {
745 make_device_private_entry_read(&entry);
746 pte = swp_entry_to_pte(entry);
747 set_pte_at(src_mm, addr, src_pte, pte);
748 }
749 }
750 goto out_set_pte;
751 }
752
753 /*
754 * If it's a COW mapping, write protect it both
755 * in the parent and the child
756 */
757 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
758 ptep_set_wrprotect(src_mm, addr, src_pte);
759 pte = pte_wrprotect(pte);
760 }
761
762 /*
763 * If it's a shared mapping, mark it clean in
764 * the child
765 */
766 if (vm_flags & VM_SHARED)
767 pte = pte_mkclean(pte);
768 pte = pte_mkold(pte);
769
770 page = vm_normal_page(vma, addr, pte);
771 if (page) {
772 get_page(page);
773 page_dup_rmap(page, false);
774 rss[mm_counter(page)]++;
775 } else if (pte_devmap(pte)) {
776 page = pte_page(pte);
777 }
778
779out_set_pte:
780 set_pte_at(dst_mm, addr, dst_pte, pte);
781 return 0;
782}
783
784static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
785 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
786 unsigned long addr, unsigned long end)
787{
788 pte_t *orig_src_pte, *orig_dst_pte;
789 pte_t *src_pte, *dst_pte;
790 spinlock_t *src_ptl, *dst_ptl;
791 int progress = 0;
792 int rss[NR_MM_COUNTERS];
793 swp_entry_t entry = (swp_entry_t){0};
794
795again:
796 init_rss_vec(rss);
797
798 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
799 if (!dst_pte)
800 return -ENOMEM;
801 src_pte = pte_offset_map(src_pmd, addr);
802 src_ptl = pte_lockptr(src_mm, src_pmd);
803 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
804 orig_src_pte = src_pte;
805 orig_dst_pte = dst_pte;
806 arch_enter_lazy_mmu_mode();
807
808 do {
809 /*
810 * We are holding two locks at this point - either of them
811 * could generate latencies in another task on another CPU.
812 */
813 if (progress >= 32) {
814 progress = 0;
815 if (need_resched() ||
816 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
817 break;
818 }
819 if (pte_none(*src_pte)) {
820 progress++;
821 continue;
822 }
823 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
824 vma, addr, rss);
825 if (entry.val)
826 break;
827 progress += 8;
828 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
829
830 arch_leave_lazy_mmu_mode();
831 spin_unlock(src_ptl);
832 pte_unmap(orig_src_pte);
833 add_mm_rss_vec(dst_mm, rss);
834 pte_unmap_unlock(orig_dst_pte, dst_ptl);
835 cond_resched();
836
837 if (entry.val) {
838 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
839 return -ENOMEM;
840 progress = 0;
841 }
842 if (addr != end)
843 goto again;
844 return 0;
845}
846
847static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
848 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
849 unsigned long addr, unsigned long end)
850{
851 pmd_t *src_pmd, *dst_pmd;
852 unsigned long next;
853
854 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
855 if (!dst_pmd)
856 return -ENOMEM;
857 src_pmd = pmd_offset(src_pud, addr);
858 do {
859 next = pmd_addr_end(addr, end);
860 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
861 || pmd_devmap(*src_pmd)) {
862 int err;
863 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
864 err = copy_huge_pmd(dst_mm, src_mm,
865 dst_pmd, src_pmd, addr, vma);
866 if (err == -ENOMEM)
867 return -ENOMEM;
868 if (!err)
869 continue;
870 /* fall through */
871 }
872 if (pmd_none_or_clear_bad(src_pmd))
873 continue;
874 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
875 vma, addr, next))
876 return -ENOMEM;
877 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
878 return 0;
879}
880
881static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
882 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
883 unsigned long addr, unsigned long end)
884{
885 pud_t *src_pud, *dst_pud;
886 unsigned long next;
887
888 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
889 if (!dst_pud)
890 return -ENOMEM;
891 src_pud = pud_offset(src_p4d, addr);
892 do {
893 next = pud_addr_end(addr, end);
894 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
895 int err;
896
897 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
898 err = copy_huge_pud(dst_mm, src_mm,
899 dst_pud, src_pud, addr, vma);
900 if (err == -ENOMEM)
901 return -ENOMEM;
902 if (!err)
903 continue;
904 /* fall through */
905 }
906 if (pud_none_or_clear_bad(src_pud))
907 continue;
908 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
909 vma, addr, next))
910 return -ENOMEM;
911 } while (dst_pud++, src_pud++, addr = next, addr != end);
912 return 0;
913}
914
915static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
916 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
917 unsigned long addr, unsigned long end)
918{
919 p4d_t *src_p4d, *dst_p4d;
920 unsigned long next;
921
922 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
923 if (!dst_p4d)
924 return -ENOMEM;
925 src_p4d = p4d_offset(src_pgd, addr);
926 do {
927 next = p4d_addr_end(addr, end);
928 if (p4d_none_or_clear_bad(src_p4d))
929 continue;
930 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
931 vma, addr, next))
932 return -ENOMEM;
933 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
934 return 0;
935}
936
937int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
938 struct vm_area_struct *vma)
939{
940 pgd_t *src_pgd, *dst_pgd;
941 unsigned long next;
942 unsigned long addr = vma->vm_start;
943 unsigned long end = vma->vm_end;
944 struct mmu_notifier_range range;
945 bool is_cow;
946 int ret;
947
948 /*
949 * Don't copy ptes where a page fault will fill them correctly.
950 * Fork becomes much lighter when there are big shared or private
951 * readonly mappings. The tradeoff is that copy_page_range is more
952 * efficient than faulting.
953 */
954 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
955 !vma->anon_vma)
956 return 0;
957
958 if (is_vm_hugetlb_page(vma))
959 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
960
961 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
962 /*
963 * We do not free on error cases below as remove_vma
964 * gets called on error from higher level routine
965 */
966 ret = track_pfn_copy(vma);
967 if (ret)
968 return ret;
969 }
970
971 /*
972 * We need to invalidate the secondary MMU mappings only when
973 * there could be a permission downgrade on the ptes of the
974 * parent mm. And a permission downgrade will only happen if
975 * is_cow_mapping() returns true.
976 */
977 is_cow = is_cow_mapping(vma->vm_flags);
978
979 if (is_cow) {
980 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
981 0, vma, src_mm, addr, end);
982 mmu_notifier_invalidate_range_start(&range);
983 }
984
985 ret = 0;
986 dst_pgd = pgd_offset(dst_mm, addr);
987 src_pgd = pgd_offset(src_mm, addr);
988 do {
989 next = pgd_addr_end(addr, end);
990 if (pgd_none_or_clear_bad(src_pgd))
991 continue;
992 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
993 vma, addr, next))) {
994 ret = -ENOMEM;
995 break;
996 }
997 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
998
999 if (is_cow)
1000 mmu_notifier_invalidate_range_end(&range);
1001 return ret;
1002}
1003
1004static unsigned long zap_pte_range(struct mmu_gather *tlb,
1005 struct vm_area_struct *vma, pmd_t *pmd,
1006 unsigned long addr, unsigned long end,
1007 struct zap_details *details)
1008{
1009 struct mm_struct *mm = tlb->mm;
1010 int force_flush = 0;
1011 int rss[NR_MM_COUNTERS];
1012 spinlock_t *ptl;
1013 pte_t *start_pte;
1014 pte_t *pte;
1015 swp_entry_t entry;
1016
1017 tlb_change_page_size(tlb, PAGE_SIZE);
1018again:
1019 init_rss_vec(rss);
1020 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1021 pte = start_pte;
1022 flush_tlb_batched_pending(mm);
1023 arch_enter_lazy_mmu_mode();
1024 do {
1025 pte_t ptent = *pte;
1026 if (pte_none(ptent))
1027 continue;
1028
1029 if (need_resched())
1030 break;
1031
1032 if (pte_present(ptent)) {
1033 struct page *page;
1034
1035 page = vm_normal_page(vma, addr, ptent);
1036 if (unlikely(details) && page) {
1037 /*
1038 * unmap_shared_mapping_pages() wants to
1039 * invalidate cache without truncating:
1040 * unmap shared but keep private pages.
1041 */
1042 if (details->check_mapping &&
1043 details->check_mapping != page_rmapping(page))
1044 continue;
1045 }
1046 ptent = ptep_get_and_clear_full(mm, addr, pte,
1047 tlb->fullmm);
1048 tlb_remove_tlb_entry(tlb, pte, addr);
1049 if (unlikely(!page))
1050 continue;
1051
1052 if (!PageAnon(page)) {
1053 if (pte_dirty(ptent)) {
1054 force_flush = 1;
1055 set_page_dirty(page);
1056 }
1057 if (pte_young(ptent) &&
1058 likely(!(vma->vm_flags & VM_SEQ_READ)))
1059 mark_page_accessed(page);
1060 }
1061 rss[mm_counter(page)]--;
1062 page_remove_rmap(page, false);
1063 if (unlikely(page_mapcount(page) < 0))
1064 print_bad_pte(vma, addr, ptent, page);
1065 if (unlikely(__tlb_remove_page(tlb, page))) {
1066 force_flush = 1;
1067 addr += PAGE_SIZE;
1068 break;
1069 }
1070 continue;
1071 }
1072
1073 entry = pte_to_swp_entry(ptent);
1074 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1075 struct page *page = device_private_entry_to_page(entry);
1076
1077 if (unlikely(details && details->check_mapping)) {
1078 /*
1079 * unmap_shared_mapping_pages() wants to
1080 * invalidate cache without truncating:
1081 * unmap shared but keep private pages.
1082 */
1083 if (details->check_mapping !=
1084 page_rmapping(page))
1085 continue;
1086 }
1087
1088 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1089 rss[mm_counter(page)]--;
1090 page_remove_rmap(page, false);
1091 put_page(page);
1092 continue;
1093 }
1094
1095 /* If details->check_mapping, we leave swap entries. */
1096 if (unlikely(details))
1097 continue;
1098
1099 if (!non_swap_entry(entry))
1100 rss[MM_SWAPENTS]--;
1101 else if (is_migration_entry(entry)) {
1102 struct page *page;
1103
1104 page = migration_entry_to_page(entry);
1105 rss[mm_counter(page)]--;
1106 }
1107 if (unlikely(!free_swap_and_cache(entry)))
1108 print_bad_pte(vma, addr, ptent, NULL);
1109 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1110 } while (pte++, addr += PAGE_SIZE, addr != end);
1111
1112 add_mm_rss_vec(mm, rss);
1113 arch_leave_lazy_mmu_mode();
1114
1115 /* Do the actual TLB flush before dropping ptl */
1116 if (force_flush)
1117 tlb_flush_mmu_tlbonly(tlb);
1118 pte_unmap_unlock(start_pte, ptl);
1119
1120 /*
1121 * If we forced a TLB flush (either due to running out of
1122 * batch buffers or because we needed to flush dirty TLB
1123 * entries before releasing the ptl), free the batched
1124 * memory too. Restart if we didn't do everything.
1125 */
1126 if (force_flush) {
1127 force_flush = 0;
1128 tlb_flush_mmu(tlb);
1129 }
1130
1131 if (addr != end) {
1132 cond_resched();
1133 goto again;
1134 }
1135
1136 return addr;
1137}
1138
1139static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1140 struct vm_area_struct *vma, pud_t *pud,
1141 unsigned long addr, unsigned long end,
1142 struct zap_details *details)
1143{
1144 pmd_t *pmd;
1145 unsigned long next;
1146
1147 pmd = pmd_offset(pud, addr);
1148 do {
1149 next = pmd_addr_end(addr, end);
1150 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1151 if (next - addr != HPAGE_PMD_SIZE)
1152 __split_huge_pmd(vma, pmd, addr, false, NULL);
1153 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1154 goto next;
1155 /* fall through */
1156 }
1157 /*
1158 * Here there can be other concurrent MADV_DONTNEED or
1159 * trans huge page faults running, and if the pmd is
1160 * none or trans huge it can change under us. This is
1161 * because MADV_DONTNEED holds the mmap_sem in read
1162 * mode.
1163 */
1164 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1165 goto next;
1166 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1167next:
1168 cond_resched();
1169 } while (pmd++, addr = next, addr != end);
1170
1171 return addr;
1172}
1173
1174static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1175 struct vm_area_struct *vma, p4d_t *p4d,
1176 unsigned long addr, unsigned long end,
1177 struct zap_details *details)
1178{
1179 pud_t *pud;
1180 unsigned long next;
1181
1182 pud = pud_offset(p4d, addr);
1183 do {
1184 next = pud_addr_end(addr, end);
1185 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1186 if (next - addr != HPAGE_PUD_SIZE) {
1187 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1188 split_huge_pud(vma, pud, addr);
1189 } else if (zap_huge_pud(tlb, vma, pud, addr))
1190 goto next;
1191 /* fall through */
1192 }
1193 if (pud_none_or_clear_bad(pud))
1194 continue;
1195 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1196next:
1197 cond_resched();
1198 } while (pud++, addr = next, addr != end);
1199
1200 return addr;
1201}
1202
1203static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1204 struct vm_area_struct *vma, pgd_t *pgd,
1205 unsigned long addr, unsigned long end,
1206 struct zap_details *details)
1207{
1208 p4d_t *p4d;
1209 unsigned long next;
1210
1211 p4d = p4d_offset(pgd, addr);
1212 do {
1213 next = p4d_addr_end(addr, end);
1214 if (p4d_none_or_clear_bad(p4d))
1215 continue;
1216 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1217 } while (p4d++, addr = next, addr != end);
1218
1219 return addr;
1220}
1221
1222void unmap_page_range(struct mmu_gather *tlb,
1223 struct vm_area_struct *vma,
1224 unsigned long addr, unsigned long end,
1225 struct zap_details *details)
1226{
1227 pgd_t *pgd;
1228 unsigned long next;
1229
1230 BUG_ON(addr >= end);
1231 tlb_start_vma(tlb, vma);
1232 pgd = pgd_offset(vma->vm_mm, addr);
1233 do {
1234 next = pgd_addr_end(addr, end);
1235 if (pgd_none_or_clear_bad(pgd))
1236 continue;
1237 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1238 } while (pgd++, addr = next, addr != end);
1239 tlb_end_vma(tlb, vma);
1240}
1241
1242
1243static void unmap_single_vma(struct mmu_gather *tlb,
1244 struct vm_area_struct *vma, unsigned long start_addr,
1245 unsigned long end_addr,
1246 struct zap_details *details)
1247{
1248 unsigned long start = max(vma->vm_start, start_addr);
1249 unsigned long end;
1250
1251 if (start >= vma->vm_end)
1252 return;
1253 end = min(vma->vm_end, end_addr);
1254 if (end <= vma->vm_start)
1255 return;
1256
1257 if (vma->vm_file)
1258 uprobe_munmap(vma, start, end);
1259
1260 if (unlikely(vma->vm_flags & VM_PFNMAP))
1261 untrack_pfn(vma, 0, 0);
1262
1263 if (start != end) {
1264 if (unlikely(is_vm_hugetlb_page(vma))) {
1265 /*
1266 * It is undesirable to test vma->vm_file as it
1267 * should be non-null for valid hugetlb area.
1268 * However, vm_file will be NULL in the error
1269 * cleanup path of mmap_region. When
1270 * hugetlbfs ->mmap method fails,
1271 * mmap_region() nullifies vma->vm_file
1272 * before calling this function to clean up.
1273 * Since no pte has actually been setup, it is
1274 * safe to do nothing in this case.
1275 */
1276 if (vma->vm_file) {
1277 i_mmap_lock_write(vma->vm_file->f_mapping);
1278 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1279 i_mmap_unlock_write(vma->vm_file->f_mapping);
1280 }
1281 } else
1282 unmap_page_range(tlb, vma, start, end, details);
1283 }
1284}
1285
1286/**
1287 * unmap_vmas - unmap a range of memory covered by a list of vma's
1288 * @tlb: address of the caller's struct mmu_gather
1289 * @vma: the starting vma
1290 * @start_addr: virtual address at which to start unmapping
1291 * @end_addr: virtual address at which to end unmapping
1292 *
1293 * Unmap all pages in the vma list.
1294 *
1295 * Only addresses between `start' and `end' will be unmapped.
1296 *
1297 * The VMA list must be sorted in ascending virtual address order.
1298 *
1299 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1300 * range after unmap_vmas() returns. So the only responsibility here is to
1301 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1302 * drops the lock and schedules.
1303 */
1304void unmap_vmas(struct mmu_gather *tlb,
1305 struct vm_area_struct *vma, unsigned long start_addr,
1306 unsigned long end_addr)
1307{
1308 struct mmu_notifier_range range;
1309
1310 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1311 start_addr, end_addr);
1312 mmu_notifier_invalidate_range_start(&range);
1313 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1314 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1315 mmu_notifier_invalidate_range_end(&range);
1316}
1317
1318/**
1319 * zap_page_range - remove user pages in a given range
1320 * @vma: vm_area_struct holding the applicable pages
1321 * @start: starting address of pages to zap
1322 * @size: number of bytes to zap
1323 *
1324 * Caller must protect the VMA list
1325 */
1326void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1327 unsigned long size)
1328{
1329 struct mmu_notifier_range range;
1330 struct mmu_gather tlb;
1331
1332 lru_add_drain();
1333 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1334 start, start + size);
1335 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1336 update_hiwater_rss(vma->vm_mm);
1337 mmu_notifier_invalidate_range_start(&range);
1338 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1339 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1340 mmu_notifier_invalidate_range_end(&range);
1341 tlb_finish_mmu(&tlb, start, range.end);
1342}
1343
1344/**
1345 * zap_page_range_single - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
1347 * @address: starting address of pages to zap
1348 * @size: number of bytes to zap
1349 * @details: details of shared cache invalidation
1350 *
1351 * The range must fit into one VMA.
1352 */
1353static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1354 unsigned long size, struct zap_details *details)
1355{
1356 struct mmu_notifier_range range;
1357 struct mmu_gather tlb;
1358
1359 lru_add_drain();
1360 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1361 address, address + size);
1362 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1363 update_hiwater_rss(vma->vm_mm);
1364 mmu_notifier_invalidate_range_start(&range);
1365 unmap_single_vma(&tlb, vma, address, range.end, details);
1366 mmu_notifier_invalidate_range_end(&range);
1367 tlb_finish_mmu(&tlb, address, range.end);
1368}
1369
1370/**
1371 * zap_vma_ptes - remove ptes mapping the vma
1372 * @vma: vm_area_struct holding ptes to be zapped
1373 * @address: starting address of pages to zap
1374 * @size: number of bytes to zap
1375 *
1376 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1377 *
1378 * The entire address range must be fully contained within the vma.
1379 *
1380 */
1381void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1382 unsigned long size)
1383{
1384 if (address < vma->vm_start || address + size > vma->vm_end ||
1385 !(vma->vm_flags & VM_PFNMAP))
1386 return;
1387
1388 zap_page_range_single(vma, address, size, NULL);
1389}
1390EXPORT_SYMBOL_GPL(zap_vma_ptes);
1391
1392pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1393 spinlock_t **ptl)
1394{
1395 pgd_t *pgd;
1396 p4d_t *p4d;
1397 pud_t *pud;
1398 pmd_t *pmd;
1399
1400 pgd = pgd_offset(mm, addr);
1401 p4d = p4d_alloc(mm, pgd, addr);
1402 if (!p4d)
1403 return NULL;
1404 pud = pud_alloc(mm, p4d, addr);
1405 if (!pud)
1406 return NULL;
1407 pmd = pmd_alloc(mm, pud, addr);
1408 if (!pmd)
1409 return NULL;
1410
1411 VM_BUG_ON(pmd_trans_huge(*pmd));
1412 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1413}
1414
1415/*
1416 * This is the old fallback for page remapping.
1417 *
1418 * For historical reasons, it only allows reserved pages. Only
1419 * old drivers should use this, and they needed to mark their
1420 * pages reserved for the old functions anyway.
1421 */
1422static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1423 struct page *page, pgprot_t prot)
1424{
1425 struct mm_struct *mm = vma->vm_mm;
1426 int retval;
1427 pte_t *pte;
1428 spinlock_t *ptl;
1429
1430 retval = -EINVAL;
1431 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1432 goto out;
1433 retval = -ENOMEM;
1434 flush_dcache_page(page);
1435 pte = get_locked_pte(mm, addr, &ptl);
1436 if (!pte)
1437 goto out;
1438 retval = -EBUSY;
1439 if (!pte_none(*pte))
1440 goto out_unlock;
1441
1442 /* Ok, finally just insert the thing.. */
1443 get_page(page);
1444 inc_mm_counter_fast(mm, mm_counter_file(page));
1445 page_add_file_rmap(page, false);
1446 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1447
1448 retval = 0;
1449out_unlock:
1450 pte_unmap_unlock(pte, ptl);
1451out:
1452 return retval;
1453}
1454
1455/**
1456 * vm_insert_page - insert single page into user vma
1457 * @vma: user vma to map to
1458 * @addr: target user address of this page
1459 * @page: source kernel page
1460 *
1461 * This allows drivers to insert individual pages they've allocated
1462 * into a user vma.
1463 *
1464 * The page has to be a nice clean _individual_ kernel allocation.
1465 * If you allocate a compound page, you need to have marked it as
1466 * such (__GFP_COMP), or manually just split the page up yourself
1467 * (see split_page()).
1468 *
1469 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1470 * took an arbitrary page protection parameter. This doesn't allow
1471 * that. Your vma protection will have to be set up correctly, which
1472 * means that if you want a shared writable mapping, you'd better
1473 * ask for a shared writable mapping!
1474 *
1475 * The page does not need to be reserved.
1476 *
1477 * Usually this function is called from f_op->mmap() handler
1478 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1479 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1480 * function from other places, for example from page-fault handler.
1481 *
1482 * Return: %0 on success, negative error code otherwise.
1483 */
1484int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1485 struct page *page)
1486{
1487 if (addr < vma->vm_start || addr >= vma->vm_end)
1488 return -EFAULT;
1489 if (!page_count(page))
1490 return -EINVAL;
1491 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1492 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1493 BUG_ON(vma->vm_flags & VM_PFNMAP);
1494 vma->vm_flags |= VM_MIXEDMAP;
1495 }
1496 return insert_page(vma, addr, page, vma->vm_page_prot);
1497}
1498EXPORT_SYMBOL(vm_insert_page);
1499
1500/*
1501 * __vm_map_pages - maps range of kernel pages into user vma
1502 * @vma: user vma to map to
1503 * @pages: pointer to array of source kernel pages
1504 * @num: number of pages in page array
1505 * @offset: user's requested vm_pgoff
1506 *
1507 * This allows drivers to map range of kernel pages into a user vma.
1508 *
1509 * Return: 0 on success and error code otherwise.
1510 */
1511static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1512 unsigned long num, unsigned long offset)
1513{
1514 unsigned long count = vma_pages(vma);
1515 unsigned long uaddr = vma->vm_start;
1516 int ret, i;
1517
1518 /* Fail if the user requested offset is beyond the end of the object */
1519 if (offset >= num)
1520 return -ENXIO;
1521
1522 /* Fail if the user requested size exceeds available object size */
1523 if (count > num - offset)
1524 return -ENXIO;
1525
1526 for (i = 0; i < count; i++) {
1527 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1528 if (ret < 0)
1529 return ret;
1530 uaddr += PAGE_SIZE;
1531 }
1532
1533 return 0;
1534}
1535
1536/**
1537 * vm_map_pages - maps range of kernel pages starts with non zero offset
1538 * @vma: user vma to map to
1539 * @pages: pointer to array of source kernel pages
1540 * @num: number of pages in page array
1541 *
1542 * Maps an object consisting of @num pages, catering for the user's
1543 * requested vm_pgoff
1544 *
1545 * If we fail to insert any page into the vma, the function will return
1546 * immediately leaving any previously inserted pages present. Callers
1547 * from the mmap handler may immediately return the error as their caller
1548 * will destroy the vma, removing any successfully inserted pages. Other
1549 * callers should make their own arrangements for calling unmap_region().
1550 *
1551 * Context: Process context. Called by mmap handlers.
1552 * Return: 0 on success and error code otherwise.
1553 */
1554int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1555 unsigned long num)
1556{
1557 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1558}
1559EXPORT_SYMBOL(vm_map_pages);
1560
1561/**
1562 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1563 * @vma: user vma to map to
1564 * @pages: pointer to array of source kernel pages
1565 * @num: number of pages in page array
1566 *
1567 * Similar to vm_map_pages(), except that it explicitly sets the offset
1568 * to 0. This function is intended for the drivers that did not consider
1569 * vm_pgoff.
1570 *
1571 * Context: Process context. Called by mmap handlers.
1572 * Return: 0 on success and error code otherwise.
1573 */
1574int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1575 unsigned long num)
1576{
1577 return __vm_map_pages(vma, pages, num, 0);
1578}
1579EXPORT_SYMBOL(vm_map_pages_zero);
1580
1581static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1582 pfn_t pfn, pgprot_t prot, bool mkwrite)
1583{
1584 struct mm_struct *mm = vma->vm_mm;
1585 pte_t *pte, entry;
1586 spinlock_t *ptl;
1587
1588 pte = get_locked_pte(mm, addr, &ptl);
1589 if (!pte)
1590 return VM_FAULT_OOM;
1591 if (!pte_none(*pte)) {
1592 if (mkwrite) {
1593 /*
1594 * For read faults on private mappings the PFN passed
1595 * in may not match the PFN we have mapped if the
1596 * mapped PFN is a writeable COW page. In the mkwrite
1597 * case we are creating a writable PTE for a shared
1598 * mapping and we expect the PFNs to match. If they
1599 * don't match, we are likely racing with block
1600 * allocation and mapping invalidation so just skip the
1601 * update.
1602 */
1603 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1604 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1605 goto out_unlock;
1606 }
1607 entry = pte_mkyoung(*pte);
1608 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1609 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1610 update_mmu_cache(vma, addr, pte);
1611 }
1612 goto out_unlock;
1613 }
1614
1615 /* Ok, finally just insert the thing.. */
1616 if (pfn_t_devmap(pfn))
1617 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1618 else
1619 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1620
1621 if (mkwrite) {
1622 entry = pte_mkyoung(entry);
1623 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1624 }
1625
1626 set_pte_at(mm, addr, pte, entry);
1627 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1628
1629out_unlock:
1630 pte_unmap_unlock(pte, ptl);
1631 return VM_FAULT_NOPAGE;
1632}
1633
1634/**
1635 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1636 * @vma: user vma to map to
1637 * @addr: target user address of this page
1638 * @pfn: source kernel pfn
1639 * @pgprot: pgprot flags for the inserted page
1640 *
1641 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1642 * to override pgprot on a per-page basis.
1643 *
1644 * This only makes sense for IO mappings, and it makes no sense for
1645 * COW mappings. In general, using multiple vmas is preferable;
1646 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1647 * impractical.
1648 *
1649 * Context: Process context. May allocate using %GFP_KERNEL.
1650 * Return: vm_fault_t value.
1651 */
1652vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1653 unsigned long pfn, pgprot_t pgprot)
1654{
1655 /*
1656 * Technically, architectures with pte_special can avoid all these
1657 * restrictions (same for remap_pfn_range). However we would like
1658 * consistency in testing and feature parity among all, so we should
1659 * try to keep these invariants in place for everybody.
1660 */
1661 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1662 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1663 (VM_PFNMAP|VM_MIXEDMAP));
1664 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1665 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1666
1667 if (addr < vma->vm_start || addr >= vma->vm_end)
1668 return VM_FAULT_SIGBUS;
1669
1670 if (!pfn_modify_allowed(pfn, pgprot))
1671 return VM_FAULT_SIGBUS;
1672
1673 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1674
1675 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1676 false);
1677}
1678EXPORT_SYMBOL(vmf_insert_pfn_prot);
1679
1680/**
1681 * vmf_insert_pfn - insert single pfn into user vma
1682 * @vma: user vma to map to
1683 * @addr: target user address of this page
1684 * @pfn: source kernel pfn
1685 *
1686 * Similar to vm_insert_page, this allows drivers to insert individual pages
1687 * they've allocated into a user vma. Same comments apply.
1688 *
1689 * This function should only be called from a vm_ops->fault handler, and
1690 * in that case the handler should return the result of this function.
1691 *
1692 * vma cannot be a COW mapping.
1693 *
1694 * As this is called only for pages that do not currently exist, we
1695 * do not need to flush old virtual caches or the TLB.
1696 *
1697 * Context: Process context. May allocate using %GFP_KERNEL.
1698 * Return: vm_fault_t value.
1699 */
1700vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1701 unsigned long pfn)
1702{
1703 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1704}
1705EXPORT_SYMBOL(vmf_insert_pfn);
1706
1707static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1708{
1709 /* these checks mirror the abort conditions in vm_normal_page */
1710 if (vma->vm_flags & VM_MIXEDMAP)
1711 return true;
1712 if (pfn_t_devmap(pfn))
1713 return true;
1714 if (pfn_t_special(pfn))
1715 return true;
1716 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1717 return true;
1718 return false;
1719}
1720
1721static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1722 unsigned long addr, pfn_t pfn, bool mkwrite)
1723{
1724 pgprot_t pgprot = vma->vm_page_prot;
1725 int err;
1726
1727 BUG_ON(!vm_mixed_ok(vma, pfn));
1728
1729 if (addr < vma->vm_start || addr >= vma->vm_end)
1730 return VM_FAULT_SIGBUS;
1731
1732 track_pfn_insert(vma, &pgprot, pfn);
1733
1734 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1735 return VM_FAULT_SIGBUS;
1736
1737 /*
1738 * If we don't have pte special, then we have to use the pfn_valid()
1739 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1740 * refcount the page if pfn_valid is true (hence insert_page rather
1741 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1742 * without pte special, it would there be refcounted as a normal page.
1743 */
1744 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1745 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1746 struct page *page;
1747
1748 /*
1749 * At this point we are committed to insert_page()
1750 * regardless of whether the caller specified flags that
1751 * result in pfn_t_has_page() == false.
1752 */
1753 page = pfn_to_page(pfn_t_to_pfn(pfn));
1754 err = insert_page(vma, addr, page, pgprot);
1755 } else {
1756 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1757 }
1758
1759 if (err == -ENOMEM)
1760 return VM_FAULT_OOM;
1761 if (err < 0 && err != -EBUSY)
1762 return VM_FAULT_SIGBUS;
1763
1764 return VM_FAULT_NOPAGE;
1765}
1766
1767vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1768 pfn_t pfn)
1769{
1770 return __vm_insert_mixed(vma, addr, pfn, false);
1771}
1772EXPORT_SYMBOL(vmf_insert_mixed);
1773
1774/*
1775 * If the insertion of PTE failed because someone else already added a
1776 * different entry in the mean time, we treat that as success as we assume
1777 * the same entry was actually inserted.
1778 */
1779vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1780 unsigned long addr, pfn_t pfn)
1781{
1782 return __vm_insert_mixed(vma, addr, pfn, true);
1783}
1784EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1785
1786/*
1787 * maps a range of physical memory into the requested pages. the old
1788 * mappings are removed. any references to nonexistent pages results
1789 * in null mappings (currently treated as "copy-on-access")
1790 */
1791static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1792 unsigned long addr, unsigned long end,
1793 unsigned long pfn, pgprot_t prot)
1794{
1795 pte_t *pte;
1796 spinlock_t *ptl;
1797 int err = 0;
1798
1799 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1800 if (!pte)
1801 return -ENOMEM;
1802 arch_enter_lazy_mmu_mode();
1803 do {
1804 BUG_ON(!pte_none(*pte));
1805 if (!pfn_modify_allowed(pfn, prot)) {
1806 err = -EACCES;
1807 break;
1808 }
1809 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1810 pfn++;
1811 } while (pte++, addr += PAGE_SIZE, addr != end);
1812 arch_leave_lazy_mmu_mode();
1813 pte_unmap_unlock(pte - 1, ptl);
1814 return err;
1815}
1816
1817static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1818 unsigned long addr, unsigned long end,
1819 unsigned long pfn, pgprot_t prot)
1820{
1821 pmd_t *pmd;
1822 unsigned long next;
1823 int err;
1824
1825 pfn -= addr >> PAGE_SHIFT;
1826 pmd = pmd_alloc(mm, pud, addr);
1827 if (!pmd)
1828 return -ENOMEM;
1829 VM_BUG_ON(pmd_trans_huge(*pmd));
1830 do {
1831 next = pmd_addr_end(addr, end);
1832 err = remap_pte_range(mm, pmd, addr, next,
1833 pfn + (addr >> PAGE_SHIFT), prot);
1834 if (err)
1835 return err;
1836 } while (pmd++, addr = next, addr != end);
1837 return 0;
1838}
1839
1840static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1841 unsigned long addr, unsigned long end,
1842 unsigned long pfn, pgprot_t prot)
1843{
1844 pud_t *pud;
1845 unsigned long next;
1846 int err;
1847
1848 pfn -= addr >> PAGE_SHIFT;
1849 pud = pud_alloc(mm, p4d, addr);
1850 if (!pud)
1851 return -ENOMEM;
1852 do {
1853 next = pud_addr_end(addr, end);
1854 err = remap_pmd_range(mm, pud, addr, next,
1855 pfn + (addr >> PAGE_SHIFT), prot);
1856 if (err)
1857 return err;
1858 } while (pud++, addr = next, addr != end);
1859 return 0;
1860}
1861
1862static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1863 unsigned long addr, unsigned long end,
1864 unsigned long pfn, pgprot_t prot)
1865{
1866 p4d_t *p4d;
1867 unsigned long next;
1868 int err;
1869
1870 pfn -= addr >> PAGE_SHIFT;
1871 p4d = p4d_alloc(mm, pgd, addr);
1872 if (!p4d)
1873 return -ENOMEM;
1874 do {
1875 next = p4d_addr_end(addr, end);
1876 err = remap_pud_range(mm, p4d, addr, next,
1877 pfn + (addr >> PAGE_SHIFT), prot);
1878 if (err)
1879 return err;
1880 } while (p4d++, addr = next, addr != end);
1881 return 0;
1882}
1883
1884/**
1885 * remap_pfn_range - remap kernel memory to userspace
1886 * @vma: user vma to map to
1887 * @addr: target user address to start at
1888 * @pfn: physical address of kernel memory
1889 * @size: size of map area
1890 * @prot: page protection flags for this mapping
1891 *
1892 * Note: this is only safe if the mm semaphore is held when called.
1893 *
1894 * Return: %0 on success, negative error code otherwise.
1895 */
1896int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1897 unsigned long pfn, unsigned long size, pgprot_t prot)
1898{
1899 pgd_t *pgd;
1900 unsigned long next;
1901 unsigned long end = addr + PAGE_ALIGN(size);
1902 struct mm_struct *mm = vma->vm_mm;
1903 unsigned long remap_pfn = pfn;
1904 int err;
1905
1906 /*
1907 * Physically remapped pages are special. Tell the
1908 * rest of the world about it:
1909 * VM_IO tells people not to look at these pages
1910 * (accesses can have side effects).
1911 * VM_PFNMAP tells the core MM that the base pages are just
1912 * raw PFN mappings, and do not have a "struct page" associated
1913 * with them.
1914 * VM_DONTEXPAND
1915 * Disable vma merging and expanding with mremap().
1916 * VM_DONTDUMP
1917 * Omit vma from core dump, even when VM_IO turned off.
1918 *
1919 * There's a horrible special case to handle copy-on-write
1920 * behaviour that some programs depend on. We mark the "original"
1921 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1922 * See vm_normal_page() for details.
1923 */
1924 if (is_cow_mapping(vma->vm_flags)) {
1925 if (addr != vma->vm_start || end != vma->vm_end)
1926 return -EINVAL;
1927 vma->vm_pgoff = pfn;
1928 }
1929
1930 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1931 if (err)
1932 return -EINVAL;
1933
1934 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1935
1936 BUG_ON(addr >= end);
1937 pfn -= addr >> PAGE_SHIFT;
1938 pgd = pgd_offset(mm, addr);
1939 flush_cache_range(vma, addr, end);
1940 do {
1941 next = pgd_addr_end(addr, end);
1942 err = remap_p4d_range(mm, pgd, addr, next,
1943 pfn + (addr >> PAGE_SHIFT), prot);
1944 if (err)
1945 break;
1946 } while (pgd++, addr = next, addr != end);
1947
1948 if (err)
1949 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1950
1951 return err;
1952}
1953EXPORT_SYMBOL(remap_pfn_range);
1954
1955/**
1956 * vm_iomap_memory - remap memory to userspace
1957 * @vma: user vma to map to
1958 * @start: start of area
1959 * @len: size of area
1960 *
1961 * This is a simplified io_remap_pfn_range() for common driver use. The
1962 * driver just needs to give us the physical memory range to be mapped,
1963 * we'll figure out the rest from the vma information.
1964 *
1965 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1966 * whatever write-combining details or similar.
1967 *
1968 * Return: %0 on success, negative error code otherwise.
1969 */
1970int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1971{
1972 unsigned long vm_len, pfn, pages;
1973
1974 /* Check that the physical memory area passed in looks valid */
1975 if (start + len < start)
1976 return -EINVAL;
1977 /*
1978 * You *really* shouldn't map things that aren't page-aligned,
1979 * but we've historically allowed it because IO memory might
1980 * just have smaller alignment.
1981 */
1982 len += start & ~PAGE_MASK;
1983 pfn = start >> PAGE_SHIFT;
1984 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1985 if (pfn + pages < pfn)
1986 return -EINVAL;
1987
1988 /* We start the mapping 'vm_pgoff' pages into the area */
1989 if (vma->vm_pgoff > pages)
1990 return -EINVAL;
1991 pfn += vma->vm_pgoff;
1992 pages -= vma->vm_pgoff;
1993
1994 /* Can we fit all of the mapping? */
1995 vm_len = vma->vm_end - vma->vm_start;
1996 if (vm_len >> PAGE_SHIFT > pages)
1997 return -EINVAL;
1998
1999 /* Ok, let it rip */
2000 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2001}
2002EXPORT_SYMBOL(vm_iomap_memory);
2003
2004static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2005 unsigned long addr, unsigned long end,
2006 pte_fn_t fn, void *data)
2007{
2008 pte_t *pte;
2009 int err;
2010 spinlock_t *uninitialized_var(ptl);
2011
2012 pte = (mm == &init_mm) ?
2013 pte_alloc_kernel(pmd, addr) :
2014 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2015 if (!pte)
2016 return -ENOMEM;
2017
2018 BUG_ON(pmd_huge(*pmd));
2019
2020 arch_enter_lazy_mmu_mode();
2021
2022 do {
2023 err = fn(pte++, addr, data);
2024 if (err)
2025 break;
2026 } while (addr += PAGE_SIZE, addr != end);
2027
2028 arch_leave_lazy_mmu_mode();
2029
2030 if (mm != &init_mm)
2031 pte_unmap_unlock(pte-1, ptl);
2032 return err;
2033}
2034
2035static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2036 unsigned long addr, unsigned long end,
2037 pte_fn_t fn, void *data)
2038{
2039 pmd_t *pmd;
2040 unsigned long next;
2041 int err;
2042
2043 BUG_ON(pud_huge(*pud));
2044
2045 pmd = pmd_alloc(mm, pud, addr);
2046 if (!pmd)
2047 return -ENOMEM;
2048 do {
2049 next = pmd_addr_end(addr, end);
2050 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2051 if (err)
2052 break;
2053 } while (pmd++, addr = next, addr != end);
2054 return err;
2055}
2056
2057static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2058 unsigned long addr, unsigned long end,
2059 pte_fn_t fn, void *data)
2060{
2061 pud_t *pud;
2062 unsigned long next;
2063 int err;
2064
2065 pud = pud_alloc(mm, p4d, addr);
2066 if (!pud)
2067 return -ENOMEM;
2068 do {
2069 next = pud_addr_end(addr, end);
2070 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2071 if (err)
2072 break;
2073 } while (pud++, addr = next, addr != end);
2074 return err;
2075}
2076
2077static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2078 unsigned long addr, unsigned long end,
2079 pte_fn_t fn, void *data)
2080{
2081 p4d_t *p4d;
2082 unsigned long next;
2083 int err;
2084
2085 p4d = p4d_alloc(mm, pgd, addr);
2086 if (!p4d)
2087 return -ENOMEM;
2088 do {
2089 next = p4d_addr_end(addr, end);
2090 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2091 if (err)
2092 break;
2093 } while (p4d++, addr = next, addr != end);
2094 return err;
2095}
2096
2097/*
2098 * Scan a region of virtual memory, filling in page tables as necessary
2099 * and calling a provided function on each leaf page table.
2100 */
2101int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2102 unsigned long size, pte_fn_t fn, void *data)
2103{
2104 pgd_t *pgd;
2105 unsigned long next;
2106 unsigned long end = addr + size;
2107 int err;
2108
2109 if (WARN_ON(addr >= end))
2110 return -EINVAL;
2111
2112 pgd = pgd_offset(mm, addr);
2113 do {
2114 next = pgd_addr_end(addr, end);
2115 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2116 if (err)
2117 break;
2118 } while (pgd++, addr = next, addr != end);
2119
2120 return err;
2121}
2122EXPORT_SYMBOL_GPL(apply_to_page_range);
2123
2124/*
2125 * handle_pte_fault chooses page fault handler according to an entry which was
2126 * read non-atomically. Before making any commitment, on those architectures
2127 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2128 * parts, do_swap_page must check under lock before unmapping the pte and
2129 * proceeding (but do_wp_page is only called after already making such a check;
2130 * and do_anonymous_page can safely check later on).
2131 */
2132static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2133 pte_t *page_table, pte_t orig_pte)
2134{
2135 int same = 1;
2136#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2137 if (sizeof(pte_t) > sizeof(unsigned long)) {
2138 spinlock_t *ptl = pte_lockptr(mm, pmd);
2139 spin_lock(ptl);
2140 same = pte_same(*page_table, orig_pte);
2141 spin_unlock(ptl);
2142 }
2143#endif
2144 pte_unmap(page_table);
2145 return same;
2146}
2147
2148static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2149{
2150 debug_dma_assert_idle(src);
2151
2152 /*
2153 * If the source page was a PFN mapping, we don't have
2154 * a "struct page" for it. We do a best-effort copy by
2155 * just copying from the original user address. If that
2156 * fails, we just zero-fill it. Live with it.
2157 */
2158 if (unlikely(!src)) {
2159 void *kaddr = kmap_atomic(dst);
2160 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2161
2162 /*
2163 * This really shouldn't fail, because the page is there
2164 * in the page tables. But it might just be unreadable,
2165 * in which case we just give up and fill the result with
2166 * zeroes.
2167 */
2168 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2169 clear_page(kaddr);
2170 kunmap_atomic(kaddr);
2171 flush_dcache_page(dst);
2172 } else
2173 copy_user_highpage(dst, src, va, vma);
2174}
2175
2176static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2177{
2178 struct file *vm_file = vma->vm_file;
2179
2180 if (vm_file)
2181 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2182
2183 /*
2184 * Special mappings (e.g. VDSO) do not have any file so fake
2185 * a default GFP_KERNEL for them.
2186 */
2187 return GFP_KERNEL;
2188}
2189
2190/*
2191 * Notify the address space that the page is about to become writable so that
2192 * it can prohibit this or wait for the page to get into an appropriate state.
2193 *
2194 * We do this without the lock held, so that it can sleep if it needs to.
2195 */
2196static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2197{
2198 vm_fault_t ret;
2199 struct page *page = vmf->page;
2200 unsigned int old_flags = vmf->flags;
2201
2202 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2203
2204 if (vmf->vma->vm_file &&
2205 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2206 return VM_FAULT_SIGBUS;
2207
2208 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2209 /* Restore original flags so that caller is not surprised */
2210 vmf->flags = old_flags;
2211 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2212 return ret;
2213 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2214 lock_page(page);
2215 if (!page->mapping) {
2216 unlock_page(page);
2217 return 0; /* retry */
2218 }
2219 ret |= VM_FAULT_LOCKED;
2220 } else
2221 VM_BUG_ON_PAGE(!PageLocked(page), page);
2222 return ret;
2223}
2224
2225/*
2226 * Handle dirtying of a page in shared file mapping on a write fault.
2227 *
2228 * The function expects the page to be locked and unlocks it.
2229 */
2230static void fault_dirty_shared_page(struct vm_area_struct *vma,
2231 struct page *page)
2232{
2233 struct address_space *mapping;
2234 bool dirtied;
2235 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2236
2237 dirtied = set_page_dirty(page);
2238 VM_BUG_ON_PAGE(PageAnon(page), page);
2239 /*
2240 * Take a local copy of the address_space - page.mapping may be zeroed
2241 * by truncate after unlock_page(). The address_space itself remains
2242 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2243 * release semantics to prevent the compiler from undoing this copying.
2244 */
2245 mapping = page_rmapping(page);
2246 unlock_page(page);
2247
2248 if ((dirtied || page_mkwrite) && mapping) {
2249 /*
2250 * Some device drivers do not set page.mapping
2251 * but still dirty their pages
2252 */
2253 balance_dirty_pages_ratelimited(mapping);
2254 }
2255
2256 if (!page_mkwrite)
2257 file_update_time(vma->vm_file);
2258}
2259
2260/*
2261 * Handle write page faults for pages that can be reused in the current vma
2262 *
2263 * This can happen either due to the mapping being with the VM_SHARED flag,
2264 * or due to us being the last reference standing to the page. In either
2265 * case, all we need to do here is to mark the page as writable and update
2266 * any related book-keeping.
2267 */
2268static inline void wp_page_reuse(struct vm_fault *vmf)
2269 __releases(vmf->ptl)
2270{
2271 struct vm_area_struct *vma = vmf->vma;
2272 struct page *page = vmf->page;
2273 pte_t entry;
2274 /*
2275 * Clear the pages cpupid information as the existing
2276 * information potentially belongs to a now completely
2277 * unrelated process.
2278 */
2279 if (page)
2280 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2281
2282 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2283 entry = pte_mkyoung(vmf->orig_pte);
2284 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2285 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2286 update_mmu_cache(vma, vmf->address, vmf->pte);
2287 pte_unmap_unlock(vmf->pte, vmf->ptl);
2288}
2289
2290/*
2291 * Handle the case of a page which we actually need to copy to a new page.
2292 *
2293 * Called with mmap_sem locked and the old page referenced, but
2294 * without the ptl held.
2295 *
2296 * High level logic flow:
2297 *
2298 * - Allocate a page, copy the content of the old page to the new one.
2299 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2300 * - Take the PTL. If the pte changed, bail out and release the allocated page
2301 * - If the pte is still the way we remember it, update the page table and all
2302 * relevant references. This includes dropping the reference the page-table
2303 * held to the old page, as well as updating the rmap.
2304 * - In any case, unlock the PTL and drop the reference we took to the old page.
2305 */
2306static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2307{
2308 struct vm_area_struct *vma = vmf->vma;
2309 struct mm_struct *mm = vma->vm_mm;
2310 struct page *old_page = vmf->page;
2311 struct page *new_page = NULL;
2312 pte_t entry;
2313 int page_copied = 0;
2314 struct mem_cgroup *memcg;
2315 struct mmu_notifier_range range;
2316
2317 if (unlikely(anon_vma_prepare(vma)))
2318 goto oom;
2319
2320 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2321 new_page = alloc_zeroed_user_highpage_movable(vma,
2322 vmf->address);
2323 if (!new_page)
2324 goto oom;
2325 } else {
2326 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2327 vmf->address);
2328 if (!new_page)
2329 goto oom;
2330 cow_user_page(new_page, old_page, vmf->address, vma);
2331 }
2332
2333 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2334 goto oom_free_new;
2335
2336 __SetPageUptodate(new_page);
2337
2338 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2339 vmf->address & PAGE_MASK,
2340 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2341 mmu_notifier_invalidate_range_start(&range);
2342
2343 /*
2344 * Re-check the pte - we dropped the lock
2345 */
2346 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2347 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2348 if (old_page) {
2349 if (!PageAnon(old_page)) {
2350 dec_mm_counter_fast(mm,
2351 mm_counter_file(old_page));
2352 inc_mm_counter_fast(mm, MM_ANONPAGES);
2353 }
2354 } else {
2355 inc_mm_counter_fast(mm, MM_ANONPAGES);
2356 }
2357 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2358 entry = mk_pte(new_page, vma->vm_page_prot);
2359 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2360 /*
2361 * Clear the pte entry and flush it first, before updating the
2362 * pte with the new entry. This will avoid a race condition
2363 * seen in the presence of one thread doing SMC and another
2364 * thread doing COW.
2365 */
2366 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2367 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2368 mem_cgroup_commit_charge(new_page, memcg, false, false);
2369 lru_cache_add_active_or_unevictable(new_page, vma);
2370 /*
2371 * We call the notify macro here because, when using secondary
2372 * mmu page tables (such as kvm shadow page tables), we want the
2373 * new page to be mapped directly into the secondary page table.
2374 */
2375 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2376 update_mmu_cache(vma, vmf->address, vmf->pte);
2377 if (old_page) {
2378 /*
2379 * Only after switching the pte to the new page may
2380 * we remove the mapcount here. Otherwise another
2381 * process may come and find the rmap count decremented
2382 * before the pte is switched to the new page, and
2383 * "reuse" the old page writing into it while our pte
2384 * here still points into it and can be read by other
2385 * threads.
2386 *
2387 * The critical issue is to order this
2388 * page_remove_rmap with the ptp_clear_flush above.
2389 * Those stores are ordered by (if nothing else,)
2390 * the barrier present in the atomic_add_negative
2391 * in page_remove_rmap.
2392 *
2393 * Then the TLB flush in ptep_clear_flush ensures that
2394 * no process can access the old page before the
2395 * decremented mapcount is visible. And the old page
2396 * cannot be reused until after the decremented
2397 * mapcount is visible. So transitively, TLBs to
2398 * old page will be flushed before it can be reused.
2399 */
2400 page_remove_rmap(old_page, false);
2401 }
2402
2403 /* Free the old page.. */
2404 new_page = old_page;
2405 page_copied = 1;
2406 } else {
2407 mem_cgroup_cancel_charge(new_page, memcg, false);
2408 }
2409
2410 if (new_page)
2411 put_page(new_page);
2412
2413 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414 /*
2415 * No need to double call mmu_notifier->invalidate_range() callback as
2416 * the above ptep_clear_flush_notify() did already call it.
2417 */
2418 mmu_notifier_invalidate_range_only_end(&range);
2419 if (old_page) {
2420 /*
2421 * Don't let another task, with possibly unlocked vma,
2422 * keep the mlocked page.
2423 */
2424 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2425 lock_page(old_page); /* LRU manipulation */
2426 if (PageMlocked(old_page))
2427 munlock_vma_page(old_page);
2428 unlock_page(old_page);
2429 }
2430 put_page(old_page);
2431 }
2432 return page_copied ? VM_FAULT_WRITE : 0;
2433oom_free_new:
2434 put_page(new_page);
2435oom:
2436 if (old_page)
2437 put_page(old_page);
2438 return VM_FAULT_OOM;
2439}
2440
2441/**
2442 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2443 * writeable once the page is prepared
2444 *
2445 * @vmf: structure describing the fault
2446 *
2447 * This function handles all that is needed to finish a write page fault in a
2448 * shared mapping due to PTE being read-only once the mapped page is prepared.
2449 * It handles locking of PTE and modifying it.
2450 *
2451 * The function expects the page to be locked or other protection against
2452 * concurrent faults / writeback (such as DAX radix tree locks).
2453 *
2454 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2455 * we acquired PTE lock.
2456 */
2457vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2458{
2459 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2460 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2461 &vmf->ptl);
2462 /*
2463 * We might have raced with another page fault while we released the
2464 * pte_offset_map_lock.
2465 */
2466 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2467 pte_unmap_unlock(vmf->pte, vmf->ptl);
2468 return VM_FAULT_NOPAGE;
2469 }
2470 wp_page_reuse(vmf);
2471 return 0;
2472}
2473
2474/*
2475 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2476 * mapping
2477 */
2478static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2479{
2480 struct vm_area_struct *vma = vmf->vma;
2481
2482 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2483 vm_fault_t ret;
2484
2485 pte_unmap_unlock(vmf->pte, vmf->ptl);
2486 vmf->flags |= FAULT_FLAG_MKWRITE;
2487 ret = vma->vm_ops->pfn_mkwrite(vmf);
2488 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2489 return ret;
2490 return finish_mkwrite_fault(vmf);
2491 }
2492 wp_page_reuse(vmf);
2493 return VM_FAULT_WRITE;
2494}
2495
2496static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2497 __releases(vmf->ptl)
2498{
2499 struct vm_area_struct *vma = vmf->vma;
2500
2501 get_page(vmf->page);
2502
2503 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2504 vm_fault_t tmp;
2505
2506 pte_unmap_unlock(vmf->pte, vmf->ptl);
2507 tmp = do_page_mkwrite(vmf);
2508 if (unlikely(!tmp || (tmp &
2509 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2510 put_page(vmf->page);
2511 return tmp;
2512 }
2513 tmp = finish_mkwrite_fault(vmf);
2514 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2515 unlock_page(vmf->page);
2516 put_page(vmf->page);
2517 return tmp;
2518 }
2519 } else {
2520 wp_page_reuse(vmf);
2521 lock_page(vmf->page);
2522 }
2523 fault_dirty_shared_page(vma, vmf->page);
2524 put_page(vmf->page);
2525
2526 return VM_FAULT_WRITE;
2527}
2528
2529/*
2530 * This routine handles present pages, when users try to write
2531 * to a shared page. It is done by copying the page to a new address
2532 * and decrementing the shared-page counter for the old page.
2533 *
2534 * Note that this routine assumes that the protection checks have been
2535 * done by the caller (the low-level page fault routine in most cases).
2536 * Thus we can safely just mark it writable once we've done any necessary
2537 * COW.
2538 *
2539 * We also mark the page dirty at this point even though the page will
2540 * change only once the write actually happens. This avoids a few races,
2541 * and potentially makes it more efficient.
2542 *
2543 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2544 * but allow concurrent faults), with pte both mapped and locked.
2545 * We return with mmap_sem still held, but pte unmapped and unlocked.
2546 */
2547static vm_fault_t do_wp_page(struct vm_fault *vmf)
2548 __releases(vmf->ptl)
2549{
2550 struct vm_area_struct *vma = vmf->vma;
2551
2552 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2553 if (!vmf->page) {
2554 /*
2555 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2556 * VM_PFNMAP VMA.
2557 *
2558 * We should not cow pages in a shared writeable mapping.
2559 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2560 */
2561 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2562 (VM_WRITE|VM_SHARED))
2563 return wp_pfn_shared(vmf);
2564
2565 pte_unmap_unlock(vmf->pte, vmf->ptl);
2566 return wp_page_copy(vmf);
2567 }
2568
2569 /*
2570 * Take out anonymous pages first, anonymous shared vmas are
2571 * not dirty accountable.
2572 */
2573 if (PageAnon(vmf->page)) {
2574 int total_map_swapcount;
2575 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2576 page_count(vmf->page) != 1))
2577 goto copy;
2578 if (!trylock_page(vmf->page)) {
2579 get_page(vmf->page);
2580 pte_unmap_unlock(vmf->pte, vmf->ptl);
2581 lock_page(vmf->page);
2582 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2583 vmf->address, &vmf->ptl);
2584 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2585 unlock_page(vmf->page);
2586 pte_unmap_unlock(vmf->pte, vmf->ptl);
2587 put_page(vmf->page);
2588 return 0;
2589 }
2590 put_page(vmf->page);
2591 }
2592 if (PageKsm(vmf->page)) {
2593 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2594 vmf->address);
2595 unlock_page(vmf->page);
2596 if (!reused)
2597 goto copy;
2598 wp_page_reuse(vmf);
2599 return VM_FAULT_WRITE;
2600 }
2601 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2602 if (total_map_swapcount == 1) {
2603 /*
2604 * The page is all ours. Move it to
2605 * our anon_vma so the rmap code will
2606 * not search our parent or siblings.
2607 * Protected against the rmap code by
2608 * the page lock.
2609 */
2610 page_move_anon_rmap(vmf->page, vma);
2611 }
2612 unlock_page(vmf->page);
2613 wp_page_reuse(vmf);
2614 return VM_FAULT_WRITE;
2615 }
2616 unlock_page(vmf->page);
2617 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2618 (VM_WRITE|VM_SHARED))) {
2619 return wp_page_shared(vmf);
2620 }
2621copy:
2622 /*
2623 * Ok, we need to copy. Oh, well..
2624 */
2625 get_page(vmf->page);
2626
2627 pte_unmap_unlock(vmf->pte, vmf->ptl);
2628 return wp_page_copy(vmf);
2629}
2630
2631static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2632 unsigned long start_addr, unsigned long end_addr,
2633 struct zap_details *details)
2634{
2635 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2636}
2637
2638static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2639 struct zap_details *details)
2640{
2641 struct vm_area_struct *vma;
2642 pgoff_t vba, vea, zba, zea;
2643
2644 vma_interval_tree_foreach(vma, root,
2645 details->first_index, details->last_index) {
2646
2647 vba = vma->vm_pgoff;
2648 vea = vba + vma_pages(vma) - 1;
2649 zba = details->first_index;
2650 if (zba < vba)
2651 zba = vba;
2652 zea = details->last_index;
2653 if (zea > vea)
2654 zea = vea;
2655
2656 unmap_mapping_range_vma(vma,
2657 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2658 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2659 details);
2660 }
2661}
2662
2663/**
2664 * unmap_mapping_pages() - Unmap pages from processes.
2665 * @mapping: The address space containing pages to be unmapped.
2666 * @start: Index of first page to be unmapped.
2667 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2668 * @even_cows: Whether to unmap even private COWed pages.
2669 *
2670 * Unmap the pages in this address space from any userspace process which
2671 * has them mmaped. Generally, you want to remove COWed pages as well when
2672 * a file is being truncated, but not when invalidating pages from the page
2673 * cache.
2674 */
2675void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2676 pgoff_t nr, bool even_cows)
2677{
2678 struct zap_details details = { };
2679
2680 details.check_mapping = even_cows ? NULL : mapping;
2681 details.first_index = start;
2682 details.last_index = start + nr - 1;
2683 if (details.last_index < details.first_index)
2684 details.last_index = ULONG_MAX;
2685
2686 i_mmap_lock_write(mapping);
2687 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2688 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2689 i_mmap_unlock_write(mapping);
2690}
2691
2692/**
2693 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2694 * address_space corresponding to the specified byte range in the underlying
2695 * file.
2696 *
2697 * @mapping: the address space containing mmaps to be unmapped.
2698 * @holebegin: byte in first page to unmap, relative to the start of
2699 * the underlying file. This will be rounded down to a PAGE_SIZE
2700 * boundary. Note that this is different from truncate_pagecache(), which
2701 * must keep the partial page. In contrast, we must get rid of
2702 * partial pages.
2703 * @holelen: size of prospective hole in bytes. This will be rounded
2704 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2705 * end of the file.
2706 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2707 * but 0 when invalidating pagecache, don't throw away private data.
2708 */
2709void unmap_mapping_range(struct address_space *mapping,
2710 loff_t const holebegin, loff_t const holelen, int even_cows)
2711{
2712 pgoff_t hba = holebegin >> PAGE_SHIFT;
2713 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2714
2715 /* Check for overflow. */
2716 if (sizeof(holelen) > sizeof(hlen)) {
2717 long long holeend =
2718 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2719 if (holeend & ~(long long)ULONG_MAX)
2720 hlen = ULONG_MAX - hba + 1;
2721 }
2722
2723 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2724}
2725EXPORT_SYMBOL(unmap_mapping_range);
2726
2727/*
2728 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2729 * but allow concurrent faults), and pte mapped but not yet locked.
2730 * We return with pte unmapped and unlocked.
2731 *
2732 * We return with the mmap_sem locked or unlocked in the same cases
2733 * as does filemap_fault().
2734 */
2735vm_fault_t do_swap_page(struct vm_fault *vmf)
2736{
2737 struct vm_area_struct *vma = vmf->vma;
2738 struct page *page = NULL, *swapcache;
2739 struct mem_cgroup *memcg;
2740 swp_entry_t entry;
2741 pte_t pte;
2742 int locked;
2743 int exclusive = 0;
2744 vm_fault_t ret = 0;
2745
2746 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2747 goto out;
2748
2749 entry = pte_to_swp_entry(vmf->orig_pte);
2750 if (unlikely(non_swap_entry(entry))) {
2751 if (is_migration_entry(entry)) {
2752 migration_entry_wait(vma->vm_mm, vmf->pmd,
2753 vmf->address);
2754 } else if (is_device_private_entry(entry)) {
2755 vmf->page = device_private_entry_to_page(entry);
2756 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
2757 } else if (is_hwpoison_entry(entry)) {
2758 ret = VM_FAULT_HWPOISON;
2759 } else {
2760 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2761 ret = VM_FAULT_SIGBUS;
2762 }
2763 goto out;
2764 }
2765
2766
2767 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2768 page = lookup_swap_cache(entry, vma, vmf->address);
2769 swapcache = page;
2770
2771 if (!page) {
2772 struct swap_info_struct *si = swp_swap_info(entry);
2773
2774 if (si->flags & SWP_SYNCHRONOUS_IO &&
2775 __swap_count(entry) == 1) {
2776 /* skip swapcache */
2777 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2778 vmf->address);
2779 if (page) {
2780 __SetPageLocked(page);
2781 __SetPageSwapBacked(page);
2782 set_page_private(page, entry.val);
2783 lru_cache_add_anon(page);
2784 swap_readpage(page, true);
2785 }
2786 } else {
2787 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2788 vmf);
2789 swapcache = page;
2790 }
2791
2792 if (!page) {
2793 /*
2794 * Back out if somebody else faulted in this pte
2795 * while we released the pte lock.
2796 */
2797 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2798 vmf->address, &vmf->ptl);
2799 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2800 ret = VM_FAULT_OOM;
2801 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2802 goto unlock;
2803 }
2804
2805 /* Had to read the page from swap area: Major fault */
2806 ret = VM_FAULT_MAJOR;
2807 count_vm_event(PGMAJFAULT);
2808 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2809 } else if (PageHWPoison(page)) {
2810 /*
2811 * hwpoisoned dirty swapcache pages are kept for killing
2812 * owner processes (which may be unknown at hwpoison time)
2813 */
2814 ret = VM_FAULT_HWPOISON;
2815 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2816 goto out_release;
2817 }
2818
2819 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2820
2821 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2822 if (!locked) {
2823 ret |= VM_FAULT_RETRY;
2824 goto out_release;
2825 }
2826
2827 /*
2828 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2829 * release the swapcache from under us. The page pin, and pte_same
2830 * test below, are not enough to exclude that. Even if it is still
2831 * swapcache, we need to check that the page's swap has not changed.
2832 */
2833 if (unlikely((!PageSwapCache(page) ||
2834 page_private(page) != entry.val)) && swapcache)
2835 goto out_page;
2836
2837 page = ksm_might_need_to_copy(page, vma, vmf->address);
2838 if (unlikely(!page)) {
2839 ret = VM_FAULT_OOM;
2840 page = swapcache;
2841 goto out_page;
2842 }
2843
2844 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2845 &memcg, false)) {
2846 ret = VM_FAULT_OOM;
2847 goto out_page;
2848 }
2849
2850 /*
2851 * Back out if somebody else already faulted in this pte.
2852 */
2853 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2854 &vmf->ptl);
2855 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2856 goto out_nomap;
2857
2858 if (unlikely(!PageUptodate(page))) {
2859 ret = VM_FAULT_SIGBUS;
2860 goto out_nomap;
2861 }
2862
2863 /*
2864 * The page isn't present yet, go ahead with the fault.
2865 *
2866 * Be careful about the sequence of operations here.
2867 * To get its accounting right, reuse_swap_page() must be called
2868 * while the page is counted on swap but not yet in mapcount i.e.
2869 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2870 * must be called after the swap_free(), or it will never succeed.
2871 */
2872
2873 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2874 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2875 pte = mk_pte(page, vma->vm_page_prot);
2876 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2877 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2878 vmf->flags &= ~FAULT_FLAG_WRITE;
2879 ret |= VM_FAULT_WRITE;
2880 exclusive = RMAP_EXCLUSIVE;
2881 }
2882 flush_icache_page(vma, page);
2883 if (pte_swp_soft_dirty(vmf->orig_pte))
2884 pte = pte_mksoft_dirty(pte);
2885 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2886 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2887 vmf->orig_pte = pte;
2888
2889 /* ksm created a completely new copy */
2890 if (unlikely(page != swapcache && swapcache)) {
2891 page_add_new_anon_rmap(page, vma, vmf->address, false);
2892 mem_cgroup_commit_charge(page, memcg, false, false);
2893 lru_cache_add_active_or_unevictable(page, vma);
2894 } else {
2895 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2896 mem_cgroup_commit_charge(page, memcg, true, false);
2897 activate_page(page);
2898 }
2899
2900 swap_free(entry);
2901 if (mem_cgroup_swap_full(page) ||
2902 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2903 try_to_free_swap(page);
2904 unlock_page(page);
2905 if (page != swapcache && swapcache) {
2906 /*
2907 * Hold the lock to avoid the swap entry to be reused
2908 * until we take the PT lock for the pte_same() check
2909 * (to avoid false positives from pte_same). For
2910 * further safety release the lock after the swap_free
2911 * so that the swap count won't change under a
2912 * parallel locked swapcache.
2913 */
2914 unlock_page(swapcache);
2915 put_page(swapcache);
2916 }
2917
2918 if (vmf->flags & FAULT_FLAG_WRITE) {
2919 ret |= do_wp_page(vmf);
2920 if (ret & VM_FAULT_ERROR)
2921 ret &= VM_FAULT_ERROR;
2922 goto out;
2923 }
2924
2925 /* No need to invalidate - it was non-present before */
2926 update_mmu_cache(vma, vmf->address, vmf->pte);
2927unlock:
2928 pte_unmap_unlock(vmf->pte, vmf->ptl);
2929out:
2930 return ret;
2931out_nomap:
2932 mem_cgroup_cancel_charge(page, memcg, false);
2933 pte_unmap_unlock(vmf->pte, vmf->ptl);
2934out_page:
2935 unlock_page(page);
2936out_release:
2937 put_page(page);
2938 if (page != swapcache && swapcache) {
2939 unlock_page(swapcache);
2940 put_page(swapcache);
2941 }
2942 return ret;
2943}
2944
2945/*
2946 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2947 * but allow concurrent faults), and pte mapped but not yet locked.
2948 * We return with mmap_sem still held, but pte unmapped and unlocked.
2949 */
2950static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2951{
2952 struct vm_area_struct *vma = vmf->vma;
2953 struct mem_cgroup *memcg;
2954 struct page *page;
2955 vm_fault_t ret = 0;
2956 pte_t entry;
2957
2958 /* File mapping without ->vm_ops ? */
2959 if (vma->vm_flags & VM_SHARED)
2960 return VM_FAULT_SIGBUS;
2961
2962 /*
2963 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2964 * pte_offset_map() on pmds where a huge pmd might be created
2965 * from a different thread.
2966 *
2967 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2968 * parallel threads are excluded by other means.
2969 *
2970 * Here we only have down_read(mmap_sem).
2971 */
2972 if (pte_alloc(vma->vm_mm, vmf->pmd))
2973 return VM_FAULT_OOM;
2974
2975 /* See the comment in pte_alloc_one_map() */
2976 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2977 return 0;
2978
2979 /* Use the zero-page for reads */
2980 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2981 !mm_forbids_zeropage(vma->vm_mm)) {
2982 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2983 vma->vm_page_prot));
2984 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2985 vmf->address, &vmf->ptl);
2986 if (!pte_none(*vmf->pte))
2987 goto unlock;
2988 ret = check_stable_address_space(vma->vm_mm);
2989 if (ret)
2990 goto unlock;
2991 /* Deliver the page fault to userland, check inside PT lock */
2992 if (userfaultfd_missing(vma)) {
2993 pte_unmap_unlock(vmf->pte, vmf->ptl);
2994 return handle_userfault(vmf, VM_UFFD_MISSING);
2995 }
2996 goto setpte;
2997 }
2998
2999 /* Allocate our own private page. */
3000 if (unlikely(anon_vma_prepare(vma)))
3001 goto oom;
3002 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3003 if (!page)
3004 goto oom;
3005
3006 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3007 false))
3008 goto oom_free_page;
3009
3010 /*
3011 * The memory barrier inside __SetPageUptodate makes sure that
3012 * preceeding stores to the page contents become visible before
3013 * the set_pte_at() write.
3014 */
3015 __SetPageUptodate(page);
3016
3017 entry = mk_pte(page, vma->vm_page_prot);
3018 if (vma->vm_flags & VM_WRITE)
3019 entry = pte_mkwrite(pte_mkdirty(entry));
3020
3021 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3022 &vmf->ptl);
3023 if (!pte_none(*vmf->pte))
3024 goto release;
3025
3026 ret = check_stable_address_space(vma->vm_mm);
3027 if (ret)
3028 goto release;
3029
3030 /* Deliver the page fault to userland, check inside PT lock */
3031 if (userfaultfd_missing(vma)) {
3032 pte_unmap_unlock(vmf->pte, vmf->ptl);
3033 mem_cgroup_cancel_charge(page, memcg, false);
3034 put_page(page);
3035 return handle_userfault(vmf, VM_UFFD_MISSING);
3036 }
3037
3038 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3039 page_add_new_anon_rmap(page, vma, vmf->address, false);
3040 mem_cgroup_commit_charge(page, memcg, false, false);
3041 lru_cache_add_active_or_unevictable(page, vma);
3042setpte:
3043 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3044
3045 /* No need to invalidate - it was non-present before */
3046 update_mmu_cache(vma, vmf->address, vmf->pte);
3047unlock:
3048 pte_unmap_unlock(vmf->pte, vmf->ptl);
3049 return ret;
3050release:
3051 mem_cgroup_cancel_charge(page, memcg, false);
3052 put_page(page);
3053 goto unlock;
3054oom_free_page:
3055 put_page(page);
3056oom:
3057 return VM_FAULT_OOM;
3058}
3059
3060/*
3061 * The mmap_sem must have been held on entry, and may have been
3062 * released depending on flags and vma->vm_ops->fault() return value.
3063 * See filemap_fault() and __lock_page_retry().
3064 */
3065static vm_fault_t __do_fault(struct vm_fault *vmf)
3066{
3067 struct vm_area_struct *vma = vmf->vma;
3068 vm_fault_t ret;
3069
3070 /*
3071 * Preallocate pte before we take page_lock because this might lead to
3072 * deadlocks for memcg reclaim which waits for pages under writeback:
3073 * lock_page(A)
3074 * SetPageWriteback(A)
3075 * unlock_page(A)
3076 * lock_page(B)
3077 * lock_page(B)
3078 * pte_alloc_pne
3079 * shrink_page_list
3080 * wait_on_page_writeback(A)
3081 * SetPageWriteback(B)
3082 * unlock_page(B)
3083 * # flush A, B to clear the writeback
3084 */
3085 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3086 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3087 if (!vmf->prealloc_pte)
3088 return VM_FAULT_OOM;
3089 smp_wmb(); /* See comment in __pte_alloc() */
3090 }
3091
3092 ret = vma->vm_ops->fault(vmf);
3093 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3094 VM_FAULT_DONE_COW)))
3095 return ret;
3096
3097 if (unlikely(PageHWPoison(vmf->page))) {
3098 if (ret & VM_FAULT_LOCKED)
3099 unlock_page(vmf->page);
3100 put_page(vmf->page);
3101 vmf->page = NULL;
3102 return VM_FAULT_HWPOISON;
3103 }
3104
3105 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3106 lock_page(vmf->page);
3107 else
3108 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3109
3110 return ret;
3111}
3112
3113/*
3114 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3115 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3116 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3117 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3118 */
3119static int pmd_devmap_trans_unstable(pmd_t *pmd)
3120{
3121 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3122}
3123
3124static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3125{
3126 struct vm_area_struct *vma = vmf->vma;
3127
3128 if (!pmd_none(*vmf->pmd))
3129 goto map_pte;
3130 if (vmf->prealloc_pte) {
3131 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3132 if (unlikely(!pmd_none(*vmf->pmd))) {
3133 spin_unlock(vmf->ptl);
3134 goto map_pte;
3135 }
3136
3137 mm_inc_nr_ptes(vma->vm_mm);
3138 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3139 spin_unlock(vmf->ptl);
3140 vmf->prealloc_pte = NULL;
3141 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3142 return VM_FAULT_OOM;
3143 }
3144map_pte:
3145 /*
3146 * If a huge pmd materialized under us just retry later. Use
3147 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3148 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3149 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3150 * running immediately after a huge pmd fault in a different thread of
3151 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3152 * All we have to ensure is that it is a regular pmd that we can walk
3153 * with pte_offset_map() and we can do that through an atomic read in
3154 * C, which is what pmd_trans_unstable() provides.
3155 */
3156 if (pmd_devmap_trans_unstable(vmf->pmd))
3157 return VM_FAULT_NOPAGE;
3158
3159 /*
3160 * At this point we know that our vmf->pmd points to a page of ptes
3161 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3162 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3163 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3164 * be valid and we will re-check to make sure the vmf->pte isn't
3165 * pte_none() under vmf->ptl protection when we return to
3166 * alloc_set_pte().
3167 */
3168 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3169 &vmf->ptl);
3170 return 0;
3171}
3172
3173#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3174static void deposit_prealloc_pte(struct vm_fault *vmf)
3175{
3176 struct vm_area_struct *vma = vmf->vma;
3177
3178 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3179 /*
3180 * We are going to consume the prealloc table,
3181 * count that as nr_ptes.
3182 */
3183 mm_inc_nr_ptes(vma->vm_mm);
3184 vmf->prealloc_pte = NULL;
3185}
3186
3187static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3188{
3189 struct vm_area_struct *vma = vmf->vma;
3190 bool write = vmf->flags & FAULT_FLAG_WRITE;
3191 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3192 pmd_t entry;
3193 int i;
3194 vm_fault_t ret;
3195
3196 if (!transhuge_vma_suitable(vma, haddr))
3197 return VM_FAULT_FALLBACK;
3198
3199 ret = VM_FAULT_FALLBACK;
3200 page = compound_head(page);
3201
3202 /*
3203 * Archs like ppc64 need additonal space to store information
3204 * related to pte entry. Use the preallocated table for that.
3205 */
3206 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3207 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3208 if (!vmf->prealloc_pte)
3209 return VM_FAULT_OOM;
3210 smp_wmb(); /* See comment in __pte_alloc() */
3211 }
3212
3213 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3214 if (unlikely(!pmd_none(*vmf->pmd)))
3215 goto out;
3216
3217 for (i = 0; i < HPAGE_PMD_NR; i++)
3218 flush_icache_page(vma, page + i);
3219
3220 entry = mk_huge_pmd(page, vma->vm_page_prot);
3221 if (write)
3222 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3223
3224 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3225 page_add_file_rmap(page, true);
3226 /*
3227 * deposit and withdraw with pmd lock held
3228 */
3229 if (arch_needs_pgtable_deposit())
3230 deposit_prealloc_pte(vmf);
3231
3232 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3233
3234 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3235
3236 /* fault is handled */
3237 ret = 0;
3238 count_vm_event(THP_FILE_MAPPED);
3239out:
3240 spin_unlock(vmf->ptl);
3241 return ret;
3242}
3243#else
3244static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3245{
3246 BUILD_BUG();
3247 return 0;
3248}
3249#endif
3250
3251/**
3252 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3253 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3254 *
3255 * @vmf: fault environment
3256 * @memcg: memcg to charge page (only for private mappings)
3257 * @page: page to map
3258 *
3259 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3260 * return.
3261 *
3262 * Target users are page handler itself and implementations of
3263 * vm_ops->map_pages.
3264 *
3265 * Return: %0 on success, %VM_FAULT_ code in case of error.
3266 */
3267vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3268 struct page *page)
3269{
3270 struct vm_area_struct *vma = vmf->vma;
3271 bool write = vmf->flags & FAULT_FLAG_WRITE;
3272 pte_t entry;
3273 vm_fault_t ret;
3274
3275 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3276 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3277 /* THP on COW? */
3278 VM_BUG_ON_PAGE(memcg, page);
3279
3280 ret = do_set_pmd(vmf, page);
3281 if (ret != VM_FAULT_FALLBACK)
3282 return ret;
3283 }
3284
3285 if (!vmf->pte) {
3286 ret = pte_alloc_one_map(vmf);
3287 if (ret)
3288 return ret;
3289 }
3290
3291 /* Re-check under ptl */
3292 if (unlikely(!pte_none(*vmf->pte)))
3293 return VM_FAULT_NOPAGE;
3294
3295 flush_icache_page(vma, page);
3296 entry = mk_pte(page, vma->vm_page_prot);
3297 if (write)
3298 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3299 /* copy-on-write page */
3300 if (write && !(vma->vm_flags & VM_SHARED)) {
3301 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3302 page_add_new_anon_rmap(page, vma, vmf->address, false);
3303 mem_cgroup_commit_charge(page, memcg, false, false);
3304 lru_cache_add_active_or_unevictable(page, vma);
3305 } else {
3306 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3307 page_add_file_rmap(page, false);
3308 }
3309 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3310
3311 /* no need to invalidate: a not-present page won't be cached */
3312 update_mmu_cache(vma, vmf->address, vmf->pte);
3313
3314 return 0;
3315}
3316
3317
3318/**
3319 * finish_fault - finish page fault once we have prepared the page to fault
3320 *
3321 * @vmf: structure describing the fault
3322 *
3323 * This function handles all that is needed to finish a page fault once the
3324 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3325 * given page, adds reverse page mapping, handles memcg charges and LRU
3326 * addition.
3327 *
3328 * The function expects the page to be locked and on success it consumes a
3329 * reference of a page being mapped (for the PTE which maps it).
3330 *
3331 * Return: %0 on success, %VM_FAULT_ code in case of error.
3332 */
3333vm_fault_t finish_fault(struct vm_fault *vmf)
3334{
3335 struct page *page;
3336 vm_fault_t ret = 0;
3337
3338 /* Did we COW the page? */
3339 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3340 !(vmf->vma->vm_flags & VM_SHARED))
3341 page = vmf->cow_page;
3342 else
3343 page = vmf->page;
3344
3345 /*
3346 * check even for read faults because we might have lost our CoWed
3347 * page
3348 */
3349 if (!(vmf->vma->vm_flags & VM_SHARED))
3350 ret = check_stable_address_space(vmf->vma->vm_mm);
3351 if (!ret)
3352 ret = alloc_set_pte(vmf, vmf->memcg, page);
3353 if (vmf->pte)
3354 pte_unmap_unlock(vmf->pte, vmf->ptl);
3355 return ret;
3356}
3357
3358static unsigned long fault_around_bytes __read_mostly =
3359 rounddown_pow_of_two(65536);
3360
3361#ifdef CONFIG_DEBUG_FS
3362static int fault_around_bytes_get(void *data, u64 *val)
3363{
3364 *val = fault_around_bytes;
3365 return 0;
3366}
3367
3368/*
3369 * fault_around_bytes must be rounded down to the nearest page order as it's
3370 * what do_fault_around() expects to see.
3371 */
3372static int fault_around_bytes_set(void *data, u64 val)
3373{
3374 if (val / PAGE_SIZE > PTRS_PER_PTE)
3375 return -EINVAL;
3376 if (val > PAGE_SIZE)
3377 fault_around_bytes = rounddown_pow_of_two(val);
3378 else
3379 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3380 return 0;
3381}
3382DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3383 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3384
3385static int __init fault_around_debugfs(void)
3386{
3387 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3388 &fault_around_bytes_fops);
3389 return 0;
3390}
3391late_initcall(fault_around_debugfs);
3392#endif
3393
3394/*
3395 * do_fault_around() tries to map few pages around the fault address. The hope
3396 * is that the pages will be needed soon and this will lower the number of
3397 * faults to handle.
3398 *
3399 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3400 * not ready to be mapped: not up-to-date, locked, etc.
3401 *
3402 * This function is called with the page table lock taken. In the split ptlock
3403 * case the page table lock only protects only those entries which belong to
3404 * the page table corresponding to the fault address.
3405 *
3406 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3407 * only once.
3408 *
3409 * fault_around_bytes defines how many bytes we'll try to map.
3410 * do_fault_around() expects it to be set to a power of two less than or equal
3411 * to PTRS_PER_PTE.
3412 *
3413 * The virtual address of the area that we map is naturally aligned to
3414 * fault_around_bytes rounded down to the machine page size
3415 * (and therefore to page order). This way it's easier to guarantee
3416 * that we don't cross page table boundaries.
3417 */
3418static vm_fault_t do_fault_around(struct vm_fault *vmf)
3419{
3420 unsigned long address = vmf->address, nr_pages, mask;
3421 pgoff_t start_pgoff = vmf->pgoff;
3422 pgoff_t end_pgoff;
3423 int off;
3424 vm_fault_t ret = 0;
3425
3426 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3427 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3428
3429 vmf->address = max(address & mask, vmf->vma->vm_start);
3430 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3431 start_pgoff -= off;
3432
3433 /*
3434 * end_pgoff is either the end of the page table, the end of
3435 * the vma or nr_pages from start_pgoff, depending what is nearest.
3436 */
3437 end_pgoff = start_pgoff -
3438 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3439 PTRS_PER_PTE - 1;
3440 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3441 start_pgoff + nr_pages - 1);
3442
3443 if (pmd_none(*vmf->pmd)) {
3444 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3445 if (!vmf->prealloc_pte)
3446 goto out;
3447 smp_wmb(); /* See comment in __pte_alloc() */
3448 }
3449
3450 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3451
3452 /* Huge page is mapped? Page fault is solved */
3453 if (pmd_trans_huge(*vmf->pmd)) {
3454 ret = VM_FAULT_NOPAGE;
3455 goto out;
3456 }
3457
3458 /* ->map_pages() haven't done anything useful. Cold page cache? */
3459 if (!vmf->pte)
3460 goto out;
3461
3462 /* check if the page fault is solved */
3463 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3464 if (!pte_none(*vmf->pte))
3465 ret = VM_FAULT_NOPAGE;
3466 pte_unmap_unlock(vmf->pte, vmf->ptl);
3467out:
3468 vmf->address = address;
3469 vmf->pte = NULL;
3470 return ret;
3471}
3472
3473static vm_fault_t do_read_fault(struct vm_fault *vmf)
3474{
3475 struct vm_area_struct *vma = vmf->vma;
3476 vm_fault_t ret = 0;
3477
3478 /*
3479 * Let's call ->map_pages() first and use ->fault() as fallback
3480 * if page by the offset is not ready to be mapped (cold cache or
3481 * something).
3482 */
3483 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3484 ret = do_fault_around(vmf);
3485 if (ret)
3486 return ret;
3487 }
3488
3489 ret = __do_fault(vmf);
3490 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3491 return ret;
3492
3493 ret |= finish_fault(vmf);
3494 unlock_page(vmf->page);
3495 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3496 put_page(vmf->page);
3497 return ret;
3498}
3499
3500static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3501{
3502 struct vm_area_struct *vma = vmf->vma;
3503 vm_fault_t ret;
3504
3505 if (unlikely(anon_vma_prepare(vma)))
3506 return VM_FAULT_OOM;
3507
3508 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3509 if (!vmf->cow_page)
3510 return VM_FAULT_OOM;
3511
3512 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3513 &vmf->memcg, false)) {
3514 put_page(vmf->cow_page);
3515 return VM_FAULT_OOM;
3516 }
3517
3518 ret = __do_fault(vmf);
3519 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3520 goto uncharge_out;
3521 if (ret & VM_FAULT_DONE_COW)
3522 return ret;
3523
3524 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3525 __SetPageUptodate(vmf->cow_page);
3526
3527 ret |= finish_fault(vmf);
3528 unlock_page(vmf->page);
3529 put_page(vmf->page);
3530 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3531 goto uncharge_out;
3532 return ret;
3533uncharge_out:
3534 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3535 put_page(vmf->cow_page);
3536 return ret;
3537}
3538
3539static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3540{
3541 struct vm_area_struct *vma = vmf->vma;
3542 vm_fault_t ret, tmp;
3543
3544 ret = __do_fault(vmf);
3545 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3546 return ret;
3547
3548 /*
3549 * Check if the backing address space wants to know that the page is
3550 * about to become writable
3551 */
3552 if (vma->vm_ops->page_mkwrite) {
3553 unlock_page(vmf->page);
3554 tmp = do_page_mkwrite(vmf);
3555 if (unlikely(!tmp ||
3556 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3557 put_page(vmf->page);
3558 return tmp;
3559 }
3560 }
3561
3562 ret |= finish_fault(vmf);
3563 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3564 VM_FAULT_RETRY))) {
3565 unlock_page(vmf->page);
3566 put_page(vmf->page);
3567 return ret;
3568 }
3569
3570 fault_dirty_shared_page(vma, vmf->page);
3571 return ret;
3572}
3573
3574/*
3575 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3576 * but allow concurrent faults).
3577 * The mmap_sem may have been released depending on flags and our
3578 * return value. See filemap_fault() and __lock_page_or_retry().
3579 * If mmap_sem is released, vma may become invalid (for example
3580 * by other thread calling munmap()).
3581 */
3582static vm_fault_t do_fault(struct vm_fault *vmf)
3583{
3584 struct vm_area_struct *vma = vmf->vma;
3585 struct mm_struct *vm_mm = vma->vm_mm;
3586 vm_fault_t ret;
3587
3588 /*
3589 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3590 */
3591 if (!vma->vm_ops->fault) {
3592 /*
3593 * If we find a migration pmd entry or a none pmd entry, which
3594 * should never happen, return SIGBUS
3595 */
3596 if (unlikely(!pmd_present(*vmf->pmd)))
3597 ret = VM_FAULT_SIGBUS;
3598 else {
3599 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3600 vmf->pmd,
3601 vmf->address,
3602 &vmf->ptl);
3603 /*
3604 * Make sure this is not a temporary clearing of pte
3605 * by holding ptl and checking again. A R/M/W update
3606 * of pte involves: take ptl, clearing the pte so that
3607 * we don't have concurrent modification by hardware
3608 * followed by an update.
3609 */
3610 if (unlikely(pte_none(*vmf->pte)))
3611 ret = VM_FAULT_SIGBUS;
3612 else
3613 ret = VM_FAULT_NOPAGE;
3614
3615 pte_unmap_unlock(vmf->pte, vmf->ptl);
3616 }
3617 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3618 ret = do_read_fault(vmf);
3619 else if (!(vma->vm_flags & VM_SHARED))
3620 ret = do_cow_fault(vmf);
3621 else
3622 ret = do_shared_fault(vmf);
3623
3624 /* preallocated pagetable is unused: free it */
3625 if (vmf->prealloc_pte) {
3626 pte_free(vm_mm, vmf->prealloc_pte);
3627 vmf->prealloc_pte = NULL;
3628 }
3629 return ret;
3630}
3631
3632static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3633 unsigned long addr, int page_nid,
3634 int *flags)
3635{
3636 get_page(page);
3637
3638 count_vm_numa_event(NUMA_HINT_FAULTS);
3639 if (page_nid == numa_node_id()) {
3640 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3641 *flags |= TNF_FAULT_LOCAL;
3642 }
3643
3644 return mpol_misplaced(page, vma, addr);
3645}
3646
3647static vm_fault_t do_numa_page(struct vm_fault *vmf)
3648{
3649 struct vm_area_struct *vma = vmf->vma;
3650 struct page *page = NULL;
3651 int page_nid = NUMA_NO_NODE;
3652 int last_cpupid;
3653 int target_nid;
3654 bool migrated = false;
3655 pte_t pte, old_pte;
3656 bool was_writable = pte_savedwrite(vmf->orig_pte);
3657 int flags = 0;
3658
3659 /*
3660 * The "pte" at this point cannot be used safely without
3661 * validation through pte_unmap_same(). It's of NUMA type but
3662 * the pfn may be screwed if the read is non atomic.
3663 */
3664 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3665 spin_lock(vmf->ptl);
3666 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3667 pte_unmap_unlock(vmf->pte, vmf->ptl);
3668 goto out;
3669 }
3670
3671 /*
3672 * Make it present again, Depending on how arch implementes non
3673 * accessible ptes, some can allow access by kernel mode.
3674 */
3675 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
3676 pte = pte_modify(old_pte, vma->vm_page_prot);
3677 pte = pte_mkyoung(pte);
3678 if (was_writable)
3679 pte = pte_mkwrite(pte);
3680 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
3681 update_mmu_cache(vma, vmf->address, vmf->pte);
3682
3683 page = vm_normal_page(vma, vmf->address, pte);
3684 if (!page) {
3685 pte_unmap_unlock(vmf->pte, vmf->ptl);
3686 return 0;
3687 }
3688
3689 /* TODO: handle PTE-mapped THP */
3690 if (PageCompound(page)) {
3691 pte_unmap_unlock(vmf->pte, vmf->ptl);
3692 return 0;
3693 }
3694
3695 /*
3696 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3697 * much anyway since they can be in shared cache state. This misses
3698 * the case where a mapping is writable but the process never writes
3699 * to it but pte_write gets cleared during protection updates and
3700 * pte_dirty has unpredictable behaviour between PTE scan updates,
3701 * background writeback, dirty balancing and application behaviour.
3702 */
3703 if (!pte_write(pte))
3704 flags |= TNF_NO_GROUP;
3705
3706 /*
3707 * Flag if the page is shared between multiple address spaces. This
3708 * is later used when determining whether to group tasks together
3709 */
3710 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3711 flags |= TNF_SHARED;
3712
3713 last_cpupid = page_cpupid_last(page);
3714 page_nid = page_to_nid(page);
3715 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3716 &flags);
3717 pte_unmap_unlock(vmf->pte, vmf->ptl);
3718 if (target_nid == NUMA_NO_NODE) {
3719 put_page(page);
3720 goto out;
3721 }
3722
3723 /* Migrate to the requested node */
3724 migrated = migrate_misplaced_page(page, vma, target_nid);
3725 if (migrated) {
3726 page_nid = target_nid;
3727 flags |= TNF_MIGRATED;
3728 } else
3729 flags |= TNF_MIGRATE_FAIL;
3730
3731out:
3732 if (page_nid != NUMA_NO_NODE)
3733 task_numa_fault(last_cpupid, page_nid, 1, flags);
3734 return 0;
3735}
3736
3737static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
3738{
3739 if (vma_is_anonymous(vmf->vma))
3740 return do_huge_pmd_anonymous_page(vmf);
3741 if (vmf->vma->vm_ops->huge_fault)
3742 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3743 return VM_FAULT_FALLBACK;
3744}
3745
3746/* `inline' is required to avoid gcc 4.1.2 build error */
3747static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3748{
3749 if (vma_is_anonymous(vmf->vma))
3750 return do_huge_pmd_wp_page(vmf, orig_pmd);
3751 if (vmf->vma->vm_ops->huge_fault)
3752 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3753
3754 /* COW handled on pte level: split pmd */
3755 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3756 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3757
3758 return VM_FAULT_FALLBACK;
3759}
3760
3761static inline bool vma_is_accessible(struct vm_area_struct *vma)
3762{
3763 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3764}
3765
3766static vm_fault_t create_huge_pud(struct vm_fault *vmf)
3767{
3768#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3769 /* No support for anonymous transparent PUD pages yet */
3770 if (vma_is_anonymous(vmf->vma))
3771 return VM_FAULT_FALLBACK;
3772 if (vmf->vma->vm_ops->huge_fault)
3773 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3774#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3775 return VM_FAULT_FALLBACK;
3776}
3777
3778static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3779{
3780#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3781 /* No support for anonymous transparent PUD pages yet */
3782 if (vma_is_anonymous(vmf->vma))
3783 return VM_FAULT_FALLBACK;
3784 if (vmf->vma->vm_ops->huge_fault)
3785 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3786#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3787 return VM_FAULT_FALLBACK;
3788}
3789
3790/*
3791 * These routines also need to handle stuff like marking pages dirty
3792 * and/or accessed for architectures that don't do it in hardware (most
3793 * RISC architectures). The early dirtying is also good on the i386.
3794 *
3795 * There is also a hook called "update_mmu_cache()" that architectures
3796 * with external mmu caches can use to update those (ie the Sparc or
3797 * PowerPC hashed page tables that act as extended TLBs).
3798 *
3799 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3800 * concurrent faults).
3801 *
3802 * The mmap_sem may have been released depending on flags and our return value.
3803 * See filemap_fault() and __lock_page_or_retry().
3804 */
3805static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
3806{
3807 pte_t entry;
3808
3809 if (unlikely(pmd_none(*vmf->pmd))) {
3810 /*
3811 * Leave __pte_alloc() until later: because vm_ops->fault may
3812 * want to allocate huge page, and if we expose page table
3813 * for an instant, it will be difficult to retract from
3814 * concurrent faults and from rmap lookups.
3815 */
3816 vmf->pte = NULL;
3817 } else {
3818 /* See comment in pte_alloc_one_map() */
3819 if (pmd_devmap_trans_unstable(vmf->pmd))
3820 return 0;
3821 /*
3822 * A regular pmd is established and it can't morph into a huge
3823 * pmd from under us anymore at this point because we hold the
3824 * mmap_sem read mode and khugepaged takes it in write mode.
3825 * So now it's safe to run pte_offset_map().
3826 */
3827 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3828 vmf->orig_pte = *vmf->pte;
3829
3830 /*
3831 * some architectures can have larger ptes than wordsize,
3832 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3833 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
3834 * accesses. The code below just needs a consistent view
3835 * for the ifs and we later double check anyway with the
3836 * ptl lock held. So here a barrier will do.
3837 */
3838 barrier();
3839 if (pte_none(vmf->orig_pte)) {
3840 pte_unmap(vmf->pte);
3841 vmf->pte = NULL;
3842 }
3843 }
3844
3845 if (!vmf->pte) {
3846 if (vma_is_anonymous(vmf->vma))
3847 return do_anonymous_page(vmf);
3848 else
3849 return do_fault(vmf);
3850 }
3851
3852 if (!pte_present(vmf->orig_pte))
3853 return do_swap_page(vmf);
3854
3855 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3856 return do_numa_page(vmf);
3857
3858 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3859 spin_lock(vmf->ptl);
3860 entry = vmf->orig_pte;
3861 if (unlikely(!pte_same(*vmf->pte, entry)))
3862 goto unlock;
3863 if (vmf->flags & FAULT_FLAG_WRITE) {
3864 if (!pte_write(entry))
3865 return do_wp_page(vmf);
3866 entry = pte_mkdirty(entry);
3867 }
3868 entry = pte_mkyoung(entry);
3869 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3870 vmf->flags & FAULT_FLAG_WRITE)) {
3871 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3872 } else {
3873 /*
3874 * This is needed only for protection faults but the arch code
3875 * is not yet telling us if this is a protection fault or not.
3876 * This still avoids useless tlb flushes for .text page faults
3877 * with threads.
3878 */
3879 if (vmf->flags & FAULT_FLAG_WRITE)
3880 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3881 }
3882unlock:
3883 pte_unmap_unlock(vmf->pte, vmf->ptl);
3884 return 0;
3885}
3886
3887/*
3888 * By the time we get here, we already hold the mm semaphore
3889 *
3890 * The mmap_sem may have been released depending on flags and our
3891 * return value. See filemap_fault() and __lock_page_or_retry().
3892 */
3893static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
3894 unsigned long address, unsigned int flags)
3895{
3896 struct vm_fault vmf = {
3897 .vma = vma,
3898 .address = address & PAGE_MASK,
3899 .flags = flags,
3900 .pgoff = linear_page_index(vma, address),
3901 .gfp_mask = __get_fault_gfp_mask(vma),
3902 };
3903 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3904 struct mm_struct *mm = vma->vm_mm;
3905 pgd_t *pgd;
3906 p4d_t *p4d;
3907 vm_fault_t ret;
3908
3909 pgd = pgd_offset(mm, address);
3910 p4d = p4d_alloc(mm, pgd, address);
3911 if (!p4d)
3912 return VM_FAULT_OOM;
3913
3914 vmf.pud = pud_alloc(mm, p4d, address);
3915 if (!vmf.pud)
3916 return VM_FAULT_OOM;
3917 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
3918 ret = create_huge_pud(&vmf);
3919 if (!(ret & VM_FAULT_FALLBACK))
3920 return ret;
3921 } else {
3922 pud_t orig_pud = *vmf.pud;
3923
3924 barrier();
3925 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3926
3927 /* NUMA case for anonymous PUDs would go here */
3928
3929 if (dirty && !pud_write(orig_pud)) {
3930 ret = wp_huge_pud(&vmf, orig_pud);
3931 if (!(ret & VM_FAULT_FALLBACK))
3932 return ret;
3933 } else {
3934 huge_pud_set_accessed(&vmf, orig_pud);
3935 return 0;
3936 }
3937 }
3938 }
3939
3940 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3941 if (!vmf.pmd)
3942 return VM_FAULT_OOM;
3943 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
3944 ret = create_huge_pmd(&vmf);
3945 if (!(ret & VM_FAULT_FALLBACK))
3946 return ret;
3947 } else {
3948 pmd_t orig_pmd = *vmf.pmd;
3949
3950 barrier();
3951 if (unlikely(is_swap_pmd(orig_pmd))) {
3952 VM_BUG_ON(thp_migration_supported() &&
3953 !is_pmd_migration_entry(orig_pmd));
3954 if (is_pmd_migration_entry(orig_pmd))
3955 pmd_migration_entry_wait(mm, vmf.pmd);
3956 return 0;
3957 }
3958 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3959 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3960 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3961
3962 if (dirty && !pmd_write(orig_pmd)) {
3963 ret = wp_huge_pmd(&vmf, orig_pmd);
3964 if (!(ret & VM_FAULT_FALLBACK))
3965 return ret;
3966 } else {
3967 huge_pmd_set_accessed(&vmf, orig_pmd);
3968 return 0;
3969 }
3970 }
3971 }
3972
3973 return handle_pte_fault(&vmf);
3974}
3975
3976/*
3977 * By the time we get here, we already hold the mm semaphore
3978 *
3979 * The mmap_sem may have been released depending on flags and our
3980 * return value. See filemap_fault() and __lock_page_or_retry().
3981 */
3982vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3983 unsigned int flags)
3984{
3985 vm_fault_t ret;
3986
3987 __set_current_state(TASK_RUNNING);
3988
3989 count_vm_event(PGFAULT);
3990 count_memcg_event_mm(vma->vm_mm, PGFAULT);
3991
3992 /* do counter updates before entering really critical section. */
3993 check_sync_rss_stat(current);
3994
3995 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3996 flags & FAULT_FLAG_INSTRUCTION,
3997 flags & FAULT_FLAG_REMOTE))
3998 return VM_FAULT_SIGSEGV;
3999
4000 /*
4001 * Enable the memcg OOM handling for faults triggered in user
4002 * space. Kernel faults are handled more gracefully.
4003 */
4004 if (flags & FAULT_FLAG_USER)
4005 mem_cgroup_enter_user_fault();
4006
4007 if (unlikely(is_vm_hugetlb_page(vma)))
4008 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4009 else
4010 ret = __handle_mm_fault(vma, address, flags);
4011
4012 if (flags & FAULT_FLAG_USER) {
4013 mem_cgroup_exit_user_fault();
4014 /*
4015 * The task may have entered a memcg OOM situation but
4016 * if the allocation error was handled gracefully (no
4017 * VM_FAULT_OOM), there is no need to kill anything.
4018 * Just clean up the OOM state peacefully.
4019 */
4020 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4021 mem_cgroup_oom_synchronize(false);
4022 }
4023
4024 return ret;
4025}
4026EXPORT_SYMBOL_GPL(handle_mm_fault);
4027
4028#ifndef __PAGETABLE_P4D_FOLDED
4029/*
4030 * Allocate p4d page table.
4031 * We've already handled the fast-path in-line.
4032 */
4033int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4034{
4035 p4d_t *new = p4d_alloc_one(mm, address);
4036 if (!new)
4037 return -ENOMEM;
4038
4039 smp_wmb(); /* See comment in __pte_alloc */
4040
4041 spin_lock(&mm->page_table_lock);
4042 if (pgd_present(*pgd)) /* Another has populated it */
4043 p4d_free(mm, new);
4044 else
4045 pgd_populate(mm, pgd, new);
4046 spin_unlock(&mm->page_table_lock);
4047 return 0;
4048}
4049#endif /* __PAGETABLE_P4D_FOLDED */
4050
4051#ifndef __PAGETABLE_PUD_FOLDED
4052/*
4053 * Allocate page upper directory.
4054 * We've already handled the fast-path in-line.
4055 */
4056int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4057{
4058 pud_t *new = pud_alloc_one(mm, address);
4059 if (!new)
4060 return -ENOMEM;
4061
4062 smp_wmb(); /* See comment in __pte_alloc */
4063
4064 spin_lock(&mm->page_table_lock);
4065#ifndef __ARCH_HAS_5LEVEL_HACK
4066 if (!p4d_present(*p4d)) {
4067 mm_inc_nr_puds(mm);
4068 p4d_populate(mm, p4d, new);
4069 } else /* Another has populated it */
4070 pud_free(mm, new);
4071#else
4072 if (!pgd_present(*p4d)) {
4073 mm_inc_nr_puds(mm);
4074 pgd_populate(mm, p4d, new);
4075 } else /* Another has populated it */
4076 pud_free(mm, new);
4077#endif /* __ARCH_HAS_5LEVEL_HACK */
4078 spin_unlock(&mm->page_table_lock);
4079 return 0;
4080}
4081#endif /* __PAGETABLE_PUD_FOLDED */
4082
4083#ifndef __PAGETABLE_PMD_FOLDED
4084/*
4085 * Allocate page middle directory.
4086 * We've already handled the fast-path in-line.
4087 */
4088int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4089{
4090 spinlock_t *ptl;
4091 pmd_t *new = pmd_alloc_one(mm, address);
4092 if (!new)
4093 return -ENOMEM;
4094
4095 smp_wmb(); /* See comment in __pte_alloc */
4096
4097 ptl = pud_lock(mm, pud);
4098#ifndef __ARCH_HAS_4LEVEL_HACK
4099 if (!pud_present(*pud)) {
4100 mm_inc_nr_pmds(mm);
4101 pud_populate(mm, pud, new);
4102 } else /* Another has populated it */
4103 pmd_free(mm, new);
4104#else
4105 if (!pgd_present(*pud)) {
4106 mm_inc_nr_pmds(mm);
4107 pgd_populate(mm, pud, new);
4108 } else /* Another has populated it */
4109 pmd_free(mm, new);
4110#endif /* __ARCH_HAS_4LEVEL_HACK */
4111 spin_unlock(ptl);
4112 return 0;
4113}
4114#endif /* __PAGETABLE_PMD_FOLDED */
4115
4116static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4117 struct mmu_notifier_range *range,
4118 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4119{
4120 pgd_t *pgd;
4121 p4d_t *p4d;
4122 pud_t *pud;
4123 pmd_t *pmd;
4124 pte_t *ptep;
4125
4126 pgd = pgd_offset(mm, address);
4127 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4128 goto out;
4129
4130 p4d = p4d_offset(pgd, address);
4131 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4132 goto out;
4133
4134 pud = pud_offset(p4d, address);
4135 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4136 goto out;
4137
4138 pmd = pmd_offset(pud, address);
4139 VM_BUG_ON(pmd_trans_huge(*pmd));
4140
4141 if (pmd_huge(*pmd)) {
4142 if (!pmdpp)
4143 goto out;
4144
4145 if (range) {
4146 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4147 NULL, mm, address & PMD_MASK,
4148 (address & PMD_MASK) + PMD_SIZE);
4149 mmu_notifier_invalidate_range_start(range);
4150 }
4151 *ptlp = pmd_lock(mm, pmd);
4152 if (pmd_huge(*pmd)) {
4153 *pmdpp = pmd;
4154 return 0;
4155 }
4156 spin_unlock(*ptlp);
4157 if (range)
4158 mmu_notifier_invalidate_range_end(range);
4159 }
4160
4161 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4162 goto out;
4163
4164 if (range) {
4165 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4166 address & PAGE_MASK,
4167 (address & PAGE_MASK) + PAGE_SIZE);
4168 mmu_notifier_invalidate_range_start(range);
4169 }
4170 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4171 if (!pte_present(*ptep))
4172 goto unlock;
4173 *ptepp = ptep;
4174 return 0;
4175unlock:
4176 pte_unmap_unlock(ptep, *ptlp);
4177 if (range)
4178 mmu_notifier_invalidate_range_end(range);
4179out:
4180 return -EINVAL;
4181}
4182
4183static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4184 pte_t **ptepp, spinlock_t **ptlp)
4185{
4186 int res;
4187
4188 /* (void) is needed to make gcc happy */
4189 (void) __cond_lock(*ptlp,
4190 !(res = __follow_pte_pmd(mm, address, NULL,
4191 ptepp, NULL, ptlp)));
4192 return res;
4193}
4194
4195int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4196 struct mmu_notifier_range *range,
4197 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4198{
4199 int res;
4200
4201 /* (void) is needed to make gcc happy */
4202 (void) __cond_lock(*ptlp,
4203 !(res = __follow_pte_pmd(mm, address, range,
4204 ptepp, pmdpp, ptlp)));
4205 return res;
4206}
4207EXPORT_SYMBOL(follow_pte_pmd);
4208
4209/**
4210 * follow_pfn - look up PFN at a user virtual address
4211 * @vma: memory mapping
4212 * @address: user virtual address
4213 * @pfn: location to store found PFN
4214 *
4215 * Only IO mappings and raw PFN mappings are allowed.
4216 *
4217 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4218 */
4219int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4220 unsigned long *pfn)
4221{
4222 int ret = -EINVAL;
4223 spinlock_t *ptl;
4224 pte_t *ptep;
4225
4226 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4227 return ret;
4228
4229 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4230 if (ret)
4231 return ret;
4232 *pfn = pte_pfn(*ptep);
4233 pte_unmap_unlock(ptep, ptl);
4234 return 0;
4235}
4236EXPORT_SYMBOL(follow_pfn);
4237
4238#ifdef CONFIG_HAVE_IOREMAP_PROT
4239int follow_phys(struct vm_area_struct *vma,
4240 unsigned long address, unsigned int flags,
4241 unsigned long *prot, resource_size_t *phys)
4242{
4243 int ret = -EINVAL;
4244 pte_t *ptep, pte;
4245 spinlock_t *ptl;
4246
4247 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4248 goto out;
4249
4250 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4251 goto out;
4252 pte = *ptep;
4253
4254 if ((flags & FOLL_WRITE) && !pte_write(pte))
4255 goto unlock;
4256
4257 *prot = pgprot_val(pte_pgprot(pte));
4258 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4259
4260 ret = 0;
4261unlock:
4262 pte_unmap_unlock(ptep, ptl);
4263out:
4264 return ret;
4265}
4266
4267int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4268 void *buf, int len, int write)
4269{
4270 resource_size_t phys_addr;
4271 unsigned long prot = 0;
4272 void __iomem *maddr;
4273 int offset = addr & (PAGE_SIZE-1);
4274
4275 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4276 return -EINVAL;
4277
4278 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4279 if (!maddr)
4280 return -ENOMEM;
4281
4282 if (write)
4283 memcpy_toio(maddr + offset, buf, len);
4284 else
4285 memcpy_fromio(buf, maddr + offset, len);
4286 iounmap(maddr);
4287
4288 return len;
4289}
4290EXPORT_SYMBOL_GPL(generic_access_phys);
4291#endif
4292
4293/*
4294 * Access another process' address space as given in mm. If non-NULL, use the
4295 * given task for page fault accounting.
4296 */
4297int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4298 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4299{
4300 struct vm_area_struct *vma;
4301 void *old_buf = buf;
4302 int write = gup_flags & FOLL_WRITE;
4303
4304 if (down_read_killable(&mm->mmap_sem))
4305 return 0;
4306
4307 /* ignore errors, just check how much was successfully transferred */
4308 while (len) {
4309 int bytes, ret, offset;
4310 void *maddr;
4311 struct page *page = NULL;
4312
4313 ret = get_user_pages_remote(tsk, mm, addr, 1,
4314 gup_flags, &page, &vma, NULL);
4315 if (ret <= 0) {
4316#ifndef CONFIG_HAVE_IOREMAP_PROT
4317 break;
4318#else
4319 /*
4320 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4321 * we can access using slightly different code.
4322 */
4323 vma = find_vma(mm, addr);
4324 if (!vma || vma->vm_start > addr)
4325 break;
4326 if (vma->vm_ops && vma->vm_ops->access)
4327 ret = vma->vm_ops->access(vma, addr, buf,
4328 len, write);
4329 if (ret <= 0)
4330 break;
4331 bytes = ret;
4332#endif
4333 } else {
4334 bytes = len;
4335 offset = addr & (PAGE_SIZE-1);
4336 if (bytes > PAGE_SIZE-offset)
4337 bytes = PAGE_SIZE-offset;
4338
4339 maddr = kmap(page);
4340 if (write) {
4341 copy_to_user_page(vma, page, addr,
4342 maddr + offset, buf, bytes);
4343 set_page_dirty_lock(page);
4344 } else {
4345 copy_from_user_page(vma, page, addr,
4346 buf, maddr + offset, bytes);
4347 }
4348 kunmap(page);
4349 put_page(page);
4350 }
4351 len -= bytes;
4352 buf += bytes;
4353 addr += bytes;
4354 }
4355 up_read(&mm->mmap_sem);
4356
4357 return buf - old_buf;
4358}
4359
4360/**
4361 * access_remote_vm - access another process' address space
4362 * @mm: the mm_struct of the target address space
4363 * @addr: start address to access
4364 * @buf: source or destination buffer
4365 * @len: number of bytes to transfer
4366 * @gup_flags: flags modifying lookup behaviour
4367 *
4368 * The caller must hold a reference on @mm.
4369 *
4370 * Return: number of bytes copied from source to destination.
4371 */
4372int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4373 void *buf, int len, unsigned int gup_flags)
4374{
4375 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4376}
4377
4378/*
4379 * Access another process' address space.
4380 * Source/target buffer must be kernel space,
4381 * Do not walk the page table directly, use get_user_pages
4382 */
4383int access_process_vm(struct task_struct *tsk, unsigned long addr,
4384 void *buf, int len, unsigned int gup_flags)
4385{
4386 struct mm_struct *mm;
4387 int ret;
4388
4389 mm = get_task_mm(tsk);
4390 if (!mm)
4391 return 0;
4392
4393 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4394
4395 mmput(mm);
4396
4397 return ret;
4398}
4399EXPORT_SYMBOL_GPL(access_process_vm);
4400
4401/*
4402 * Print the name of a VMA.
4403 */
4404void print_vma_addr(char *prefix, unsigned long ip)
4405{
4406 struct mm_struct *mm = current->mm;
4407 struct vm_area_struct *vma;
4408
4409 /*
4410 * we might be running from an atomic context so we cannot sleep
4411 */
4412 if (!down_read_trylock(&mm->mmap_sem))
4413 return;
4414
4415 vma = find_vma(mm, ip);
4416 if (vma && vma->vm_file) {
4417 struct file *f = vma->vm_file;
4418 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4419 if (buf) {
4420 char *p;
4421
4422 p = file_path(f, buf, PAGE_SIZE);
4423 if (IS_ERR(p))
4424 p = "?";
4425 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4426 vma->vm_start,
4427 vma->vm_end - vma->vm_start);
4428 free_page((unsigned long)buf);
4429 }
4430 }
4431 up_read(&mm->mmap_sem);
4432}
4433
4434#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4435void __might_fault(const char *file, int line)
4436{
4437 /*
4438 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4439 * holding the mmap_sem, this is safe because kernel memory doesn't
4440 * get paged out, therefore we'll never actually fault, and the
4441 * below annotations will generate false positives.
4442 */
4443 if (uaccess_kernel())
4444 return;
4445 if (pagefault_disabled())
4446 return;
4447 __might_sleep(file, line, 0);
4448#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4449 if (current->mm)
4450 might_lock_read(¤t->mm->mmap_sem);
4451#endif
4452}
4453EXPORT_SYMBOL(__might_fault);
4454#endif
4455
4456#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4457/*
4458 * Process all subpages of the specified huge page with the specified
4459 * operation. The target subpage will be processed last to keep its
4460 * cache lines hot.
4461 */
4462static inline void process_huge_page(
4463 unsigned long addr_hint, unsigned int pages_per_huge_page,
4464 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4465 void *arg)
4466{
4467 int i, n, base, l;
4468 unsigned long addr = addr_hint &
4469 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4470
4471 /* Process target subpage last to keep its cache lines hot */
4472 might_sleep();
4473 n = (addr_hint - addr) / PAGE_SIZE;
4474 if (2 * n <= pages_per_huge_page) {
4475 /* If target subpage in first half of huge page */
4476 base = 0;
4477 l = n;
4478 /* Process subpages at the end of huge page */
4479 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4480 cond_resched();
4481 process_subpage(addr + i * PAGE_SIZE, i, arg);
4482 }
4483 } else {
4484 /* If target subpage in second half of huge page */
4485 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4486 l = pages_per_huge_page - n;
4487 /* Process subpages at the begin of huge page */
4488 for (i = 0; i < base; i++) {
4489 cond_resched();
4490 process_subpage(addr + i * PAGE_SIZE, i, arg);
4491 }
4492 }
4493 /*
4494 * Process remaining subpages in left-right-left-right pattern
4495 * towards the target subpage
4496 */
4497 for (i = 0; i < l; i++) {
4498 int left_idx = base + i;
4499 int right_idx = base + 2 * l - 1 - i;
4500
4501 cond_resched();
4502 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4503 cond_resched();
4504 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4505 }
4506}
4507
4508static void clear_gigantic_page(struct page *page,
4509 unsigned long addr,
4510 unsigned int pages_per_huge_page)
4511{
4512 int i;
4513 struct page *p = page;
4514
4515 might_sleep();
4516 for (i = 0; i < pages_per_huge_page;
4517 i++, p = mem_map_next(p, page, i)) {
4518 cond_resched();
4519 clear_user_highpage(p, addr + i * PAGE_SIZE);
4520 }
4521}
4522
4523static void clear_subpage(unsigned long addr, int idx, void *arg)
4524{
4525 struct page *page = arg;
4526
4527 clear_user_highpage(page + idx, addr);
4528}
4529
4530void clear_huge_page(struct page *page,
4531 unsigned long addr_hint, unsigned int pages_per_huge_page)
4532{
4533 unsigned long addr = addr_hint &
4534 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4535
4536 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4537 clear_gigantic_page(page, addr, pages_per_huge_page);
4538 return;
4539 }
4540
4541 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4542}
4543
4544static void copy_user_gigantic_page(struct page *dst, struct page *src,
4545 unsigned long addr,
4546 struct vm_area_struct *vma,
4547 unsigned int pages_per_huge_page)
4548{
4549 int i;
4550 struct page *dst_base = dst;
4551 struct page *src_base = src;
4552
4553 for (i = 0; i < pages_per_huge_page; ) {
4554 cond_resched();
4555 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4556
4557 i++;
4558 dst = mem_map_next(dst, dst_base, i);
4559 src = mem_map_next(src, src_base, i);
4560 }
4561}
4562
4563struct copy_subpage_arg {
4564 struct page *dst;
4565 struct page *src;
4566 struct vm_area_struct *vma;
4567};
4568
4569static void copy_subpage(unsigned long addr, int idx, void *arg)
4570{
4571 struct copy_subpage_arg *copy_arg = arg;
4572
4573 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4574 addr, copy_arg->vma);
4575}
4576
4577void copy_user_huge_page(struct page *dst, struct page *src,
4578 unsigned long addr_hint, struct vm_area_struct *vma,
4579 unsigned int pages_per_huge_page)
4580{
4581 unsigned long addr = addr_hint &
4582 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4583 struct copy_subpage_arg arg = {
4584 .dst = dst,
4585 .src = src,
4586 .vma = vma,
4587 };
4588
4589 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4590 copy_user_gigantic_page(dst, src, addr, vma,
4591 pages_per_huge_page);
4592 return;
4593 }
4594
4595 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4596}
4597
4598long copy_huge_page_from_user(struct page *dst_page,
4599 const void __user *usr_src,
4600 unsigned int pages_per_huge_page,
4601 bool allow_pagefault)
4602{
4603 void *src = (void *)usr_src;
4604 void *page_kaddr;
4605 unsigned long i, rc = 0;
4606 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4607
4608 for (i = 0; i < pages_per_huge_page; i++) {
4609 if (allow_pagefault)
4610 page_kaddr = kmap(dst_page + i);
4611 else
4612 page_kaddr = kmap_atomic(dst_page + i);
4613 rc = copy_from_user(page_kaddr,
4614 (const void __user *)(src + i * PAGE_SIZE),
4615 PAGE_SIZE);
4616 if (allow_pagefault)
4617 kunmap(dst_page + i);
4618 else
4619 kunmap_atomic(page_kaddr);
4620
4621 ret_val -= (PAGE_SIZE - rc);
4622 if (rc)
4623 break;
4624
4625 cond_resched();
4626 }
4627 return ret_val;
4628}
4629#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4630
4631#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4632
4633static struct kmem_cache *page_ptl_cachep;
4634
4635void __init ptlock_cache_init(void)
4636{
4637 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4638 SLAB_PANIC, NULL);
4639}
4640
4641bool ptlock_alloc(struct page *page)
4642{
4643 spinlock_t *ptl;
4644
4645 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4646 if (!ptl)
4647 return false;
4648 page->ptl = ptl;
4649 return true;
4650}
4651
4652void ptlock_free(struct page *page)
4653{
4654 kmem_cache_free(page_ptl_cachep, page->ptl);
4655}
4656#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/writeback.h>
54#include <linux/memcontrol.h>
55#include <linux/mmu_notifier.h>
56#include <linux/kallsyms.h>
57#include <linux/swapops.h>
58#include <linux/elf.h>
59#include <linux/gfp.h>
60#include <linux/migrate.h>
61#include <linux/string.h>
62#include <linux/dma-debug.h>
63#include <linux/debugfs.h>
64
65#include <asm/io.h>
66#include <asm/pgalloc.h>
67#include <asm/uaccess.h>
68#include <asm/tlb.h>
69#include <asm/tlbflush.h>
70#include <asm/pgtable.h>
71
72#include "internal.h"
73
74#ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
75#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
76#endif
77
78#ifndef CONFIG_NEED_MULTIPLE_NODES
79/* use the per-pgdat data instead for discontigmem - mbligh */
80unsigned long max_mapnr;
81struct page *mem_map;
82
83EXPORT_SYMBOL(max_mapnr);
84EXPORT_SYMBOL(mem_map);
85#endif
86
87/*
88 * A number of key systems in x86 including ioremap() rely on the assumption
89 * that high_memory defines the upper bound on direct map memory, then end
90 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
91 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
92 * and ZONE_HIGHMEM.
93 */
94void * high_memory;
95
96EXPORT_SYMBOL(high_memory);
97
98/*
99 * Randomize the address space (stacks, mmaps, brk, etc.).
100 *
101 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
102 * as ancient (libc5 based) binaries can segfault. )
103 */
104int randomize_va_space __read_mostly =
105#ifdef CONFIG_COMPAT_BRK
106 1;
107#else
108 2;
109#endif
110
111static int __init disable_randmaps(char *s)
112{
113 randomize_va_space = 0;
114 return 1;
115}
116__setup("norandmaps", disable_randmaps);
117
118unsigned long zero_pfn __read_mostly;
119unsigned long highest_memmap_pfn __read_mostly;
120
121/*
122 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
123 */
124static int __init init_zero_pfn(void)
125{
126 zero_pfn = page_to_pfn(ZERO_PAGE(0));
127 return 0;
128}
129core_initcall(init_zero_pfn);
130
131
132#if defined(SPLIT_RSS_COUNTING)
133
134void sync_mm_rss(struct mm_struct *mm)
135{
136 int i;
137
138 for (i = 0; i < NR_MM_COUNTERS; i++) {
139 if (current->rss_stat.count[i]) {
140 add_mm_counter(mm, i, current->rss_stat.count[i]);
141 current->rss_stat.count[i] = 0;
142 }
143 }
144 current->rss_stat.events = 0;
145}
146
147static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
148{
149 struct task_struct *task = current;
150
151 if (likely(task->mm == mm))
152 task->rss_stat.count[member] += val;
153 else
154 add_mm_counter(mm, member, val);
155}
156#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
157#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
158
159/* sync counter once per 64 page faults */
160#define TASK_RSS_EVENTS_THRESH (64)
161static void check_sync_rss_stat(struct task_struct *task)
162{
163 if (unlikely(task != current))
164 return;
165 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
166 sync_mm_rss(task->mm);
167}
168#else /* SPLIT_RSS_COUNTING */
169
170#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
171#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
172
173static void check_sync_rss_stat(struct task_struct *task)
174{
175}
176
177#endif /* SPLIT_RSS_COUNTING */
178
179#ifdef HAVE_GENERIC_MMU_GATHER
180
181static int tlb_next_batch(struct mmu_gather *tlb)
182{
183 struct mmu_gather_batch *batch;
184
185 batch = tlb->active;
186 if (batch->next) {
187 tlb->active = batch->next;
188 return 1;
189 }
190
191 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
192 return 0;
193
194 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
195 if (!batch)
196 return 0;
197
198 tlb->batch_count++;
199 batch->next = NULL;
200 batch->nr = 0;
201 batch->max = MAX_GATHER_BATCH;
202
203 tlb->active->next = batch;
204 tlb->active = batch;
205
206 return 1;
207}
208
209/* tlb_gather_mmu
210 * Called to initialize an (on-stack) mmu_gather structure for page-table
211 * tear-down from @mm. The @fullmm argument is used when @mm is without
212 * users and we're going to destroy the full address space (exit/execve).
213 */
214void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
215{
216 tlb->mm = mm;
217
218 /* Is it from 0 to ~0? */
219 tlb->fullmm = !(start | (end+1));
220 tlb->need_flush_all = 0;
221 tlb->start = start;
222 tlb->end = end;
223 tlb->need_flush = 0;
224 tlb->local.next = NULL;
225 tlb->local.nr = 0;
226 tlb->local.max = ARRAY_SIZE(tlb->__pages);
227 tlb->active = &tlb->local;
228 tlb->batch_count = 0;
229
230#ifdef CONFIG_HAVE_RCU_TABLE_FREE
231 tlb->batch = NULL;
232#endif
233}
234
235static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
236{
237 tlb->need_flush = 0;
238 tlb_flush(tlb);
239#ifdef CONFIG_HAVE_RCU_TABLE_FREE
240 tlb_table_flush(tlb);
241#endif
242}
243
244static void tlb_flush_mmu_free(struct mmu_gather *tlb)
245{
246 struct mmu_gather_batch *batch;
247
248 for (batch = &tlb->local; batch; batch = batch->next) {
249 free_pages_and_swap_cache(batch->pages, batch->nr);
250 batch->nr = 0;
251 }
252 tlb->active = &tlb->local;
253}
254
255void tlb_flush_mmu(struct mmu_gather *tlb)
256{
257 if (!tlb->need_flush)
258 return;
259 tlb_flush_mmu_tlbonly(tlb);
260 tlb_flush_mmu_free(tlb);
261}
262
263/* tlb_finish_mmu
264 * Called at the end of the shootdown operation to free up any resources
265 * that were required.
266 */
267void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
268{
269 struct mmu_gather_batch *batch, *next;
270
271 tlb_flush_mmu(tlb);
272
273 /* keep the page table cache within bounds */
274 check_pgt_cache();
275
276 for (batch = tlb->local.next; batch; batch = next) {
277 next = batch->next;
278 free_pages((unsigned long)batch, 0);
279 }
280 tlb->local.next = NULL;
281}
282
283/* __tlb_remove_page
284 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
285 * handling the additional races in SMP caused by other CPUs caching valid
286 * mappings in their TLBs. Returns the number of free page slots left.
287 * When out of page slots we must call tlb_flush_mmu().
288 */
289int __tlb_remove_page(struct mmu_gather *tlb, struct page *page)
290{
291 struct mmu_gather_batch *batch;
292
293 VM_BUG_ON(!tlb->need_flush);
294
295 batch = tlb->active;
296 batch->pages[batch->nr++] = page;
297 if (batch->nr == batch->max) {
298 if (!tlb_next_batch(tlb))
299 return 0;
300 batch = tlb->active;
301 }
302 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
303
304 return batch->max - batch->nr;
305}
306
307#endif /* HAVE_GENERIC_MMU_GATHER */
308
309#ifdef CONFIG_HAVE_RCU_TABLE_FREE
310
311/*
312 * See the comment near struct mmu_table_batch.
313 */
314
315static void tlb_remove_table_smp_sync(void *arg)
316{
317 /* Simply deliver the interrupt */
318}
319
320static void tlb_remove_table_one(void *table)
321{
322 /*
323 * This isn't an RCU grace period and hence the page-tables cannot be
324 * assumed to be actually RCU-freed.
325 *
326 * It is however sufficient for software page-table walkers that rely on
327 * IRQ disabling. See the comment near struct mmu_table_batch.
328 */
329 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
330 __tlb_remove_table(table);
331}
332
333static void tlb_remove_table_rcu(struct rcu_head *head)
334{
335 struct mmu_table_batch *batch;
336 int i;
337
338 batch = container_of(head, struct mmu_table_batch, rcu);
339
340 for (i = 0; i < batch->nr; i++)
341 __tlb_remove_table(batch->tables[i]);
342
343 free_page((unsigned long)batch);
344}
345
346void tlb_table_flush(struct mmu_gather *tlb)
347{
348 struct mmu_table_batch **batch = &tlb->batch;
349
350 if (*batch) {
351 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
352 *batch = NULL;
353 }
354}
355
356void tlb_remove_table(struct mmu_gather *tlb, void *table)
357{
358 struct mmu_table_batch **batch = &tlb->batch;
359
360 tlb->need_flush = 1;
361
362 /*
363 * When there's less then two users of this mm there cannot be a
364 * concurrent page-table walk.
365 */
366 if (atomic_read(&tlb->mm->mm_users) < 2) {
367 __tlb_remove_table(table);
368 return;
369 }
370
371 if (*batch == NULL) {
372 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
373 if (*batch == NULL) {
374 tlb_remove_table_one(table);
375 return;
376 }
377 (*batch)->nr = 0;
378 }
379 (*batch)->tables[(*batch)->nr++] = table;
380 if ((*batch)->nr == MAX_TABLE_BATCH)
381 tlb_table_flush(tlb);
382}
383
384#endif /* CONFIG_HAVE_RCU_TABLE_FREE */
385
386/*
387 * Note: this doesn't free the actual pages themselves. That
388 * has been handled earlier when unmapping all the memory regions.
389 */
390static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
391 unsigned long addr)
392{
393 pgtable_t token = pmd_pgtable(*pmd);
394 pmd_clear(pmd);
395 pte_free_tlb(tlb, token, addr);
396 atomic_long_dec(&tlb->mm->nr_ptes);
397}
398
399static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
400 unsigned long addr, unsigned long end,
401 unsigned long floor, unsigned long ceiling)
402{
403 pmd_t *pmd;
404 unsigned long next;
405 unsigned long start;
406
407 start = addr;
408 pmd = pmd_offset(pud, addr);
409 do {
410 next = pmd_addr_end(addr, end);
411 if (pmd_none_or_clear_bad(pmd))
412 continue;
413 free_pte_range(tlb, pmd, addr);
414 } while (pmd++, addr = next, addr != end);
415
416 start &= PUD_MASK;
417 if (start < floor)
418 return;
419 if (ceiling) {
420 ceiling &= PUD_MASK;
421 if (!ceiling)
422 return;
423 }
424 if (end - 1 > ceiling - 1)
425 return;
426
427 pmd = pmd_offset(pud, start);
428 pud_clear(pud);
429 pmd_free_tlb(tlb, pmd, start);
430}
431
432static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
433 unsigned long addr, unsigned long end,
434 unsigned long floor, unsigned long ceiling)
435{
436 pud_t *pud;
437 unsigned long next;
438 unsigned long start;
439
440 start = addr;
441 pud = pud_offset(pgd, addr);
442 do {
443 next = pud_addr_end(addr, end);
444 if (pud_none_or_clear_bad(pud))
445 continue;
446 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
447 } while (pud++, addr = next, addr != end);
448
449 start &= PGDIR_MASK;
450 if (start < floor)
451 return;
452 if (ceiling) {
453 ceiling &= PGDIR_MASK;
454 if (!ceiling)
455 return;
456 }
457 if (end - 1 > ceiling - 1)
458 return;
459
460 pud = pud_offset(pgd, start);
461 pgd_clear(pgd);
462 pud_free_tlb(tlb, pud, start);
463}
464
465/*
466 * This function frees user-level page tables of a process.
467 */
468void free_pgd_range(struct mmu_gather *tlb,
469 unsigned long addr, unsigned long end,
470 unsigned long floor, unsigned long ceiling)
471{
472 pgd_t *pgd;
473 unsigned long next;
474
475 /*
476 * The next few lines have given us lots of grief...
477 *
478 * Why are we testing PMD* at this top level? Because often
479 * there will be no work to do at all, and we'd prefer not to
480 * go all the way down to the bottom just to discover that.
481 *
482 * Why all these "- 1"s? Because 0 represents both the bottom
483 * of the address space and the top of it (using -1 for the
484 * top wouldn't help much: the masks would do the wrong thing).
485 * The rule is that addr 0 and floor 0 refer to the bottom of
486 * the address space, but end 0 and ceiling 0 refer to the top
487 * Comparisons need to use "end - 1" and "ceiling - 1" (though
488 * that end 0 case should be mythical).
489 *
490 * Wherever addr is brought up or ceiling brought down, we must
491 * be careful to reject "the opposite 0" before it confuses the
492 * subsequent tests. But what about where end is brought down
493 * by PMD_SIZE below? no, end can't go down to 0 there.
494 *
495 * Whereas we round start (addr) and ceiling down, by different
496 * masks at different levels, in order to test whether a table
497 * now has no other vmas using it, so can be freed, we don't
498 * bother to round floor or end up - the tests don't need that.
499 */
500
501 addr &= PMD_MASK;
502 if (addr < floor) {
503 addr += PMD_SIZE;
504 if (!addr)
505 return;
506 }
507 if (ceiling) {
508 ceiling &= PMD_MASK;
509 if (!ceiling)
510 return;
511 }
512 if (end - 1 > ceiling - 1)
513 end -= PMD_SIZE;
514 if (addr > end - 1)
515 return;
516
517 pgd = pgd_offset(tlb->mm, addr);
518 do {
519 next = pgd_addr_end(addr, end);
520 if (pgd_none_or_clear_bad(pgd))
521 continue;
522 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
523 } while (pgd++, addr = next, addr != end);
524}
525
526void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
527 unsigned long floor, unsigned long ceiling)
528{
529 while (vma) {
530 struct vm_area_struct *next = vma->vm_next;
531 unsigned long addr = vma->vm_start;
532
533 /*
534 * Hide vma from rmap and truncate_pagecache before freeing
535 * pgtables
536 */
537 unlink_anon_vmas(vma);
538 unlink_file_vma(vma);
539
540 if (is_vm_hugetlb_page(vma)) {
541 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
542 floor, next? next->vm_start: ceiling);
543 } else {
544 /*
545 * Optimization: gather nearby vmas into one call down
546 */
547 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
548 && !is_vm_hugetlb_page(next)) {
549 vma = next;
550 next = vma->vm_next;
551 unlink_anon_vmas(vma);
552 unlink_file_vma(vma);
553 }
554 free_pgd_range(tlb, addr, vma->vm_end,
555 floor, next? next->vm_start: ceiling);
556 }
557 vma = next;
558 }
559}
560
561int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma,
562 pmd_t *pmd, unsigned long address)
563{
564 spinlock_t *ptl;
565 pgtable_t new = pte_alloc_one(mm, address);
566 int wait_split_huge_page;
567 if (!new)
568 return -ENOMEM;
569
570 /*
571 * Ensure all pte setup (eg. pte page lock and page clearing) are
572 * visible before the pte is made visible to other CPUs by being
573 * put into page tables.
574 *
575 * The other side of the story is the pointer chasing in the page
576 * table walking code (when walking the page table without locking;
577 * ie. most of the time). Fortunately, these data accesses consist
578 * of a chain of data-dependent loads, meaning most CPUs (alpha
579 * being the notable exception) will already guarantee loads are
580 * seen in-order. See the alpha page table accessors for the
581 * smp_read_barrier_depends() barriers in page table walking code.
582 */
583 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
584
585 ptl = pmd_lock(mm, pmd);
586 wait_split_huge_page = 0;
587 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
588 atomic_long_inc(&mm->nr_ptes);
589 pmd_populate(mm, pmd, new);
590 new = NULL;
591 } else if (unlikely(pmd_trans_splitting(*pmd)))
592 wait_split_huge_page = 1;
593 spin_unlock(ptl);
594 if (new)
595 pte_free(mm, new);
596 if (wait_split_huge_page)
597 wait_split_huge_page(vma->anon_vma, pmd);
598 return 0;
599}
600
601int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
602{
603 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
604 if (!new)
605 return -ENOMEM;
606
607 smp_wmb(); /* See comment in __pte_alloc */
608
609 spin_lock(&init_mm.page_table_lock);
610 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
611 pmd_populate_kernel(&init_mm, pmd, new);
612 new = NULL;
613 } else
614 VM_BUG_ON(pmd_trans_splitting(*pmd));
615 spin_unlock(&init_mm.page_table_lock);
616 if (new)
617 pte_free_kernel(&init_mm, new);
618 return 0;
619}
620
621static inline void init_rss_vec(int *rss)
622{
623 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
624}
625
626static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
627{
628 int i;
629
630 if (current->mm == mm)
631 sync_mm_rss(mm);
632 for (i = 0; i < NR_MM_COUNTERS; i++)
633 if (rss[i])
634 add_mm_counter(mm, i, rss[i]);
635}
636
637/*
638 * This function is called to print an error when a bad pte
639 * is found. For example, we might have a PFN-mapped pte in
640 * a region that doesn't allow it.
641 *
642 * The calling function must still handle the error.
643 */
644static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
645 pte_t pte, struct page *page)
646{
647 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
648 pud_t *pud = pud_offset(pgd, addr);
649 pmd_t *pmd = pmd_offset(pud, addr);
650 struct address_space *mapping;
651 pgoff_t index;
652 static unsigned long resume;
653 static unsigned long nr_shown;
654 static unsigned long nr_unshown;
655
656 /*
657 * Allow a burst of 60 reports, then keep quiet for that minute;
658 * or allow a steady drip of one report per second.
659 */
660 if (nr_shown == 60) {
661 if (time_before(jiffies, resume)) {
662 nr_unshown++;
663 return;
664 }
665 if (nr_unshown) {
666 printk(KERN_ALERT
667 "BUG: Bad page map: %lu messages suppressed\n",
668 nr_unshown);
669 nr_unshown = 0;
670 }
671 nr_shown = 0;
672 }
673 if (nr_shown++ == 0)
674 resume = jiffies + 60 * HZ;
675
676 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
677 index = linear_page_index(vma, addr);
678
679 printk(KERN_ALERT
680 "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
681 current->comm,
682 (long long)pte_val(pte), (long long)pmd_val(*pmd));
683 if (page)
684 dump_page(page, "bad pte");
685 printk(KERN_ALERT
686 "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
687 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
688 /*
689 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
690 */
691 if (vma->vm_ops)
692 printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n",
693 vma->vm_ops->fault);
694 if (vma->vm_file)
695 printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n",
696 vma->vm_file->f_op->mmap);
697 dump_stack();
698 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
699}
700
701static inline bool is_cow_mapping(vm_flags_t flags)
702{
703 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
704}
705
706/*
707 * vm_normal_page -- This function gets the "struct page" associated with a pte.
708 *
709 * "Special" mappings do not wish to be associated with a "struct page" (either
710 * it doesn't exist, or it exists but they don't want to touch it). In this
711 * case, NULL is returned here. "Normal" mappings do have a struct page.
712 *
713 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
714 * pte bit, in which case this function is trivial. Secondly, an architecture
715 * may not have a spare pte bit, which requires a more complicated scheme,
716 * described below.
717 *
718 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
719 * special mapping (even if there are underlying and valid "struct pages").
720 * COWed pages of a VM_PFNMAP are always normal.
721 *
722 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
723 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
724 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
725 * mapping will always honor the rule
726 *
727 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
728 *
729 * And for normal mappings this is false.
730 *
731 * This restricts such mappings to be a linear translation from virtual address
732 * to pfn. To get around this restriction, we allow arbitrary mappings so long
733 * as the vma is not a COW mapping; in that case, we know that all ptes are
734 * special (because none can have been COWed).
735 *
736 *
737 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
738 *
739 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
740 * page" backing, however the difference is that _all_ pages with a struct
741 * page (that is, those where pfn_valid is true) are refcounted and considered
742 * normal pages by the VM. The disadvantage is that pages are refcounted
743 * (which can be slower and simply not an option for some PFNMAP users). The
744 * advantage is that we don't have to follow the strict linearity rule of
745 * PFNMAP mappings in order to support COWable mappings.
746 *
747 */
748#ifdef __HAVE_ARCH_PTE_SPECIAL
749# define HAVE_PTE_SPECIAL 1
750#else
751# define HAVE_PTE_SPECIAL 0
752#endif
753struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
754 pte_t pte)
755{
756 unsigned long pfn = pte_pfn(pte);
757
758 if (HAVE_PTE_SPECIAL) {
759 if (likely(!pte_special(pte)))
760 goto check_pfn;
761 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
762 return NULL;
763 if (!is_zero_pfn(pfn))
764 print_bad_pte(vma, addr, pte, NULL);
765 return NULL;
766 }
767
768 /* !HAVE_PTE_SPECIAL case follows: */
769
770 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
771 if (vma->vm_flags & VM_MIXEDMAP) {
772 if (!pfn_valid(pfn))
773 return NULL;
774 goto out;
775 } else {
776 unsigned long off;
777 off = (addr - vma->vm_start) >> PAGE_SHIFT;
778 if (pfn == vma->vm_pgoff + off)
779 return NULL;
780 if (!is_cow_mapping(vma->vm_flags))
781 return NULL;
782 }
783 }
784
785 if (is_zero_pfn(pfn))
786 return NULL;
787check_pfn:
788 if (unlikely(pfn > highest_memmap_pfn)) {
789 print_bad_pte(vma, addr, pte, NULL);
790 return NULL;
791 }
792
793 /*
794 * NOTE! We still have PageReserved() pages in the page tables.
795 * eg. VDSO mappings can cause them to exist.
796 */
797out:
798 return pfn_to_page(pfn);
799}
800
801/*
802 * copy one vm_area from one task to the other. Assumes the page tables
803 * already present in the new task to be cleared in the whole range
804 * covered by this vma.
805 */
806
807static inline unsigned long
808copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
809 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
810 unsigned long addr, int *rss)
811{
812 unsigned long vm_flags = vma->vm_flags;
813 pte_t pte = *src_pte;
814 struct page *page;
815
816 /* pte contains position in swap or file, so copy. */
817 if (unlikely(!pte_present(pte))) {
818 if (!pte_file(pte)) {
819 swp_entry_t entry = pte_to_swp_entry(pte);
820
821 if (swap_duplicate(entry) < 0)
822 return entry.val;
823
824 /* make sure dst_mm is on swapoff's mmlist. */
825 if (unlikely(list_empty(&dst_mm->mmlist))) {
826 spin_lock(&mmlist_lock);
827 if (list_empty(&dst_mm->mmlist))
828 list_add(&dst_mm->mmlist,
829 &src_mm->mmlist);
830 spin_unlock(&mmlist_lock);
831 }
832 if (likely(!non_swap_entry(entry)))
833 rss[MM_SWAPENTS]++;
834 else if (is_migration_entry(entry)) {
835 page = migration_entry_to_page(entry);
836
837 if (PageAnon(page))
838 rss[MM_ANONPAGES]++;
839 else
840 rss[MM_FILEPAGES]++;
841
842 if (is_write_migration_entry(entry) &&
843 is_cow_mapping(vm_flags)) {
844 /*
845 * COW mappings require pages in both
846 * parent and child to be set to read.
847 */
848 make_migration_entry_read(&entry);
849 pte = swp_entry_to_pte(entry);
850 if (pte_swp_soft_dirty(*src_pte))
851 pte = pte_swp_mksoft_dirty(pte);
852 set_pte_at(src_mm, addr, src_pte, pte);
853 }
854 }
855 }
856 goto out_set_pte;
857 }
858
859 /*
860 * If it's a COW mapping, write protect it both
861 * in the parent and the child
862 */
863 if (is_cow_mapping(vm_flags)) {
864 ptep_set_wrprotect(src_mm, addr, src_pte);
865 pte = pte_wrprotect(pte);
866 }
867
868 /*
869 * If it's a shared mapping, mark it clean in
870 * the child
871 */
872 if (vm_flags & VM_SHARED)
873 pte = pte_mkclean(pte);
874 pte = pte_mkold(pte);
875
876 page = vm_normal_page(vma, addr, pte);
877 if (page) {
878 get_page(page);
879 page_dup_rmap(page);
880 if (PageAnon(page))
881 rss[MM_ANONPAGES]++;
882 else
883 rss[MM_FILEPAGES]++;
884 }
885
886out_set_pte:
887 set_pte_at(dst_mm, addr, dst_pte, pte);
888 return 0;
889}
890
891int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
892 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
893 unsigned long addr, unsigned long end)
894{
895 pte_t *orig_src_pte, *orig_dst_pte;
896 pte_t *src_pte, *dst_pte;
897 spinlock_t *src_ptl, *dst_ptl;
898 int progress = 0;
899 int rss[NR_MM_COUNTERS];
900 swp_entry_t entry = (swp_entry_t){0};
901
902again:
903 init_rss_vec(rss);
904
905 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
906 if (!dst_pte)
907 return -ENOMEM;
908 src_pte = pte_offset_map(src_pmd, addr);
909 src_ptl = pte_lockptr(src_mm, src_pmd);
910 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
911 orig_src_pte = src_pte;
912 orig_dst_pte = dst_pte;
913 arch_enter_lazy_mmu_mode();
914
915 do {
916 /*
917 * We are holding two locks at this point - either of them
918 * could generate latencies in another task on another CPU.
919 */
920 if (progress >= 32) {
921 progress = 0;
922 if (need_resched() ||
923 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
924 break;
925 }
926 if (pte_none(*src_pte)) {
927 progress++;
928 continue;
929 }
930 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
931 vma, addr, rss);
932 if (entry.val)
933 break;
934 progress += 8;
935 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
936
937 arch_leave_lazy_mmu_mode();
938 spin_unlock(src_ptl);
939 pte_unmap(orig_src_pte);
940 add_mm_rss_vec(dst_mm, rss);
941 pte_unmap_unlock(orig_dst_pte, dst_ptl);
942 cond_resched();
943
944 if (entry.val) {
945 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
946 return -ENOMEM;
947 progress = 0;
948 }
949 if (addr != end)
950 goto again;
951 return 0;
952}
953
954static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
955 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
956 unsigned long addr, unsigned long end)
957{
958 pmd_t *src_pmd, *dst_pmd;
959 unsigned long next;
960
961 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
962 if (!dst_pmd)
963 return -ENOMEM;
964 src_pmd = pmd_offset(src_pud, addr);
965 do {
966 next = pmd_addr_end(addr, end);
967 if (pmd_trans_huge(*src_pmd)) {
968 int err;
969 VM_BUG_ON(next-addr != HPAGE_PMD_SIZE);
970 err = copy_huge_pmd(dst_mm, src_mm,
971 dst_pmd, src_pmd, addr, vma);
972 if (err == -ENOMEM)
973 return -ENOMEM;
974 if (!err)
975 continue;
976 /* fall through */
977 }
978 if (pmd_none_or_clear_bad(src_pmd))
979 continue;
980 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
981 vma, addr, next))
982 return -ENOMEM;
983 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
984 return 0;
985}
986
987static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
988 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
989 unsigned long addr, unsigned long end)
990{
991 pud_t *src_pud, *dst_pud;
992 unsigned long next;
993
994 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
995 if (!dst_pud)
996 return -ENOMEM;
997 src_pud = pud_offset(src_pgd, addr);
998 do {
999 next = pud_addr_end(addr, end);
1000 if (pud_none_or_clear_bad(src_pud))
1001 continue;
1002 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1003 vma, addr, next))
1004 return -ENOMEM;
1005 } while (dst_pud++, src_pud++, addr = next, addr != end);
1006 return 0;
1007}
1008
1009int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1010 struct vm_area_struct *vma)
1011{
1012 pgd_t *src_pgd, *dst_pgd;
1013 unsigned long next;
1014 unsigned long addr = vma->vm_start;
1015 unsigned long end = vma->vm_end;
1016 unsigned long mmun_start; /* For mmu_notifiers */
1017 unsigned long mmun_end; /* For mmu_notifiers */
1018 bool is_cow;
1019 int ret;
1020
1021 /*
1022 * Don't copy ptes where a page fault will fill them correctly.
1023 * Fork becomes much lighter when there are big shared or private
1024 * readonly mappings. The tradeoff is that copy_page_range is more
1025 * efficient than faulting.
1026 */
1027 if (!(vma->vm_flags & (VM_HUGETLB | VM_NONLINEAR |
1028 VM_PFNMAP | VM_MIXEDMAP))) {
1029 if (!vma->anon_vma)
1030 return 0;
1031 }
1032
1033 if (is_vm_hugetlb_page(vma))
1034 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1035
1036 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1037 /*
1038 * We do not free on error cases below as remove_vma
1039 * gets called on error from higher level routine
1040 */
1041 ret = track_pfn_copy(vma);
1042 if (ret)
1043 return ret;
1044 }
1045
1046 /*
1047 * We need to invalidate the secondary MMU mappings only when
1048 * there could be a permission downgrade on the ptes of the
1049 * parent mm. And a permission downgrade will only happen if
1050 * is_cow_mapping() returns true.
1051 */
1052 is_cow = is_cow_mapping(vma->vm_flags);
1053 mmun_start = addr;
1054 mmun_end = end;
1055 if (is_cow)
1056 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1057 mmun_end);
1058
1059 ret = 0;
1060 dst_pgd = pgd_offset(dst_mm, addr);
1061 src_pgd = pgd_offset(src_mm, addr);
1062 do {
1063 next = pgd_addr_end(addr, end);
1064 if (pgd_none_or_clear_bad(src_pgd))
1065 continue;
1066 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1067 vma, addr, next))) {
1068 ret = -ENOMEM;
1069 break;
1070 }
1071 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1072
1073 if (is_cow)
1074 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1075 return ret;
1076}
1077
1078static unsigned long zap_pte_range(struct mmu_gather *tlb,
1079 struct vm_area_struct *vma, pmd_t *pmd,
1080 unsigned long addr, unsigned long end,
1081 struct zap_details *details)
1082{
1083 struct mm_struct *mm = tlb->mm;
1084 int force_flush = 0;
1085 int rss[NR_MM_COUNTERS];
1086 spinlock_t *ptl;
1087 pte_t *start_pte;
1088 pte_t *pte;
1089
1090again:
1091 init_rss_vec(rss);
1092 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1093 pte = start_pte;
1094 arch_enter_lazy_mmu_mode();
1095 do {
1096 pte_t ptent = *pte;
1097 if (pte_none(ptent)) {
1098 continue;
1099 }
1100
1101 if (pte_present(ptent)) {
1102 struct page *page;
1103
1104 page = vm_normal_page(vma, addr, ptent);
1105 if (unlikely(details) && page) {
1106 /*
1107 * unmap_shared_mapping_pages() wants to
1108 * invalidate cache without truncating:
1109 * unmap shared but keep private pages.
1110 */
1111 if (details->check_mapping &&
1112 details->check_mapping != page->mapping)
1113 continue;
1114 /*
1115 * Each page->index must be checked when
1116 * invalidating or truncating nonlinear.
1117 */
1118 if (details->nonlinear_vma &&
1119 (page->index < details->first_index ||
1120 page->index > details->last_index))
1121 continue;
1122 }
1123 ptent = ptep_get_and_clear_full(mm, addr, pte,
1124 tlb->fullmm);
1125 tlb_remove_tlb_entry(tlb, pte, addr);
1126 if (unlikely(!page))
1127 continue;
1128 if (unlikely(details) && details->nonlinear_vma
1129 && linear_page_index(details->nonlinear_vma,
1130 addr) != page->index) {
1131 pte_t ptfile = pgoff_to_pte(page->index);
1132 if (pte_soft_dirty(ptent))
1133 pte_file_mksoft_dirty(ptfile);
1134 set_pte_at(mm, addr, pte, ptfile);
1135 }
1136 if (PageAnon(page))
1137 rss[MM_ANONPAGES]--;
1138 else {
1139 if (pte_dirty(ptent)) {
1140 force_flush = 1;
1141 set_page_dirty(page);
1142 }
1143 if (pte_young(ptent) &&
1144 likely(!(vma->vm_flags & VM_SEQ_READ)))
1145 mark_page_accessed(page);
1146 rss[MM_FILEPAGES]--;
1147 }
1148 page_remove_rmap(page);
1149 if (unlikely(page_mapcount(page) < 0))
1150 print_bad_pte(vma, addr, ptent, page);
1151 if (unlikely(!__tlb_remove_page(tlb, page))) {
1152 force_flush = 1;
1153 break;
1154 }
1155 continue;
1156 }
1157 /*
1158 * If details->check_mapping, we leave swap entries;
1159 * if details->nonlinear_vma, we leave file entries.
1160 */
1161 if (unlikely(details))
1162 continue;
1163 if (pte_file(ptent)) {
1164 if (unlikely(!(vma->vm_flags & VM_NONLINEAR)))
1165 print_bad_pte(vma, addr, ptent, NULL);
1166 } else {
1167 swp_entry_t entry = pte_to_swp_entry(ptent);
1168
1169 if (!non_swap_entry(entry))
1170 rss[MM_SWAPENTS]--;
1171 else if (is_migration_entry(entry)) {
1172 struct page *page;
1173
1174 page = migration_entry_to_page(entry);
1175
1176 if (PageAnon(page))
1177 rss[MM_ANONPAGES]--;
1178 else
1179 rss[MM_FILEPAGES]--;
1180 }
1181 if (unlikely(!free_swap_and_cache(entry)))
1182 print_bad_pte(vma, addr, ptent, NULL);
1183 }
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 unsigned long old_end;
1193
1194 /*
1195 * Flush the TLB just for the previous segment,
1196 * then update the range to be the remaining
1197 * TLB range.
1198 */
1199 old_end = tlb->end;
1200 tlb->end = addr;
1201 tlb_flush_mmu_tlbonly(tlb);
1202 tlb->start = addr;
1203 tlb->end = old_end;
1204 }
1205 pte_unmap_unlock(start_pte, ptl);
1206
1207 /*
1208 * If we forced a TLB flush (either due to running out of
1209 * batch buffers or because we needed to flush dirty TLB
1210 * entries before releasing the ptl), free the batched
1211 * memory too. Restart if we didn't do everything.
1212 */
1213 if (force_flush) {
1214 force_flush = 0;
1215 tlb_flush_mmu_free(tlb);
1216
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)) {
1236 if (next - addr != HPAGE_PMD_SIZE) {
1237#ifdef CONFIG_DEBUG_VM
1238 if (!rwsem_is_locked(&tlb->mm->mmap_sem)) {
1239 pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n",
1240 __func__, addr, end,
1241 vma->vm_start,
1242 vma->vm_end);
1243 BUG();
1244 }
1245#endif
1246 split_huge_page_pmd(vma, addr, pmd);
1247 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1248 goto next;
1249 /* fall through */
1250 }
1251 /*
1252 * Here there can be other concurrent MADV_DONTNEED or
1253 * trans huge page faults running, and if the pmd is
1254 * none or trans huge it can change under us. This is
1255 * because MADV_DONTNEED holds the mmap_sem in read
1256 * mode.
1257 */
1258 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1259 goto next;
1260 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1261next:
1262 cond_resched();
1263 } while (pmd++, addr = next, addr != end);
1264
1265 return addr;
1266}
1267
1268static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1269 struct vm_area_struct *vma, pgd_t *pgd,
1270 unsigned long addr, unsigned long end,
1271 struct zap_details *details)
1272{
1273 pud_t *pud;
1274 unsigned long next;
1275
1276 pud = pud_offset(pgd, addr);
1277 do {
1278 next = pud_addr_end(addr, end);
1279 if (pud_none_or_clear_bad(pud))
1280 continue;
1281 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1282 } while (pud++, addr = next, addr != end);
1283
1284 return addr;
1285}
1286
1287static void unmap_page_range(struct mmu_gather *tlb,
1288 struct vm_area_struct *vma,
1289 unsigned long addr, unsigned long end,
1290 struct zap_details *details)
1291{
1292 pgd_t *pgd;
1293 unsigned long next;
1294
1295 if (details && !details->check_mapping && !details->nonlinear_vma)
1296 details = NULL;
1297
1298 BUG_ON(addr >= end);
1299 mem_cgroup_uncharge_start();
1300 tlb_start_vma(tlb, vma);
1301 pgd = pgd_offset(vma->vm_mm, addr);
1302 do {
1303 next = pgd_addr_end(addr, end);
1304 if (pgd_none_or_clear_bad(pgd))
1305 continue;
1306 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1307 } while (pgd++, addr = next, addr != end);
1308 tlb_end_vma(tlb, vma);
1309 mem_cgroup_uncharge_end();
1310}
1311
1312
1313static void unmap_single_vma(struct mmu_gather *tlb,
1314 struct vm_area_struct *vma, unsigned long start_addr,
1315 unsigned long end_addr,
1316 struct zap_details *details)
1317{
1318 unsigned long start = max(vma->vm_start, start_addr);
1319 unsigned long end;
1320
1321 if (start >= vma->vm_end)
1322 return;
1323 end = min(vma->vm_end, end_addr);
1324 if (end <= vma->vm_start)
1325 return;
1326
1327 if (vma->vm_file)
1328 uprobe_munmap(vma, start, end);
1329
1330 if (unlikely(vma->vm_flags & VM_PFNMAP))
1331 untrack_pfn(vma, 0, 0);
1332
1333 if (start != end) {
1334 if (unlikely(is_vm_hugetlb_page(vma))) {
1335 /*
1336 * It is undesirable to test vma->vm_file as it
1337 * should be non-null for valid hugetlb area.
1338 * However, vm_file will be NULL in the error
1339 * cleanup path of mmap_region. When
1340 * hugetlbfs ->mmap method fails,
1341 * mmap_region() nullifies vma->vm_file
1342 * before calling this function to clean up.
1343 * Since no pte has actually been setup, it is
1344 * safe to do nothing in this case.
1345 */
1346 if (vma->vm_file) {
1347 mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex);
1348 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1349 mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex);
1350 }
1351 } else
1352 unmap_page_range(tlb, vma, start, end, details);
1353 }
1354}
1355
1356/**
1357 * unmap_vmas - unmap a range of memory covered by a list of vma's
1358 * @tlb: address of the caller's struct mmu_gather
1359 * @vma: the starting vma
1360 * @start_addr: virtual address at which to start unmapping
1361 * @end_addr: virtual address at which to end unmapping
1362 *
1363 * Unmap all pages in the vma list.
1364 *
1365 * Only addresses between `start' and `end' will be unmapped.
1366 *
1367 * The VMA list must be sorted in ascending virtual address order.
1368 *
1369 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1370 * range after unmap_vmas() returns. So the only responsibility here is to
1371 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1372 * drops the lock and schedules.
1373 */
1374void unmap_vmas(struct mmu_gather *tlb,
1375 struct vm_area_struct *vma, unsigned long start_addr,
1376 unsigned long end_addr)
1377{
1378 struct mm_struct *mm = vma->vm_mm;
1379
1380 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1381 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1382 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1383 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1384}
1385
1386/**
1387 * zap_page_range - remove user pages in a given range
1388 * @vma: vm_area_struct holding the applicable pages
1389 * @start: starting address of pages to zap
1390 * @size: number of bytes to zap
1391 * @details: details of nonlinear truncation or shared cache invalidation
1392 *
1393 * Caller must protect the VMA list
1394 */
1395void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1396 unsigned long size, struct zap_details *details)
1397{
1398 struct mm_struct *mm = vma->vm_mm;
1399 struct mmu_gather tlb;
1400 unsigned long end = start + size;
1401
1402 lru_add_drain();
1403 tlb_gather_mmu(&tlb, mm, start, end);
1404 update_hiwater_rss(mm);
1405 mmu_notifier_invalidate_range_start(mm, start, end);
1406 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1407 unmap_single_vma(&tlb, vma, start, end, details);
1408 mmu_notifier_invalidate_range_end(mm, start, end);
1409 tlb_finish_mmu(&tlb, start, end);
1410}
1411
1412/**
1413 * zap_page_range_single - remove user pages in a given range
1414 * @vma: vm_area_struct holding the applicable pages
1415 * @address: starting address of pages to zap
1416 * @size: number of bytes to zap
1417 * @details: details of nonlinear truncation or shared cache invalidation
1418 *
1419 * The range must fit into one VMA.
1420 */
1421static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1422 unsigned long size, struct zap_details *details)
1423{
1424 struct mm_struct *mm = vma->vm_mm;
1425 struct mmu_gather tlb;
1426 unsigned long end = address + size;
1427
1428 lru_add_drain();
1429 tlb_gather_mmu(&tlb, mm, address, end);
1430 update_hiwater_rss(mm);
1431 mmu_notifier_invalidate_range_start(mm, address, end);
1432 unmap_single_vma(&tlb, vma, address, end, details);
1433 mmu_notifier_invalidate_range_end(mm, address, end);
1434 tlb_finish_mmu(&tlb, address, end);
1435}
1436
1437/**
1438 * zap_vma_ptes - remove ptes mapping the vma
1439 * @vma: vm_area_struct holding ptes to be zapped
1440 * @address: starting address of pages to zap
1441 * @size: number of bytes to zap
1442 *
1443 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1444 *
1445 * The entire address range must be fully contained within the vma.
1446 *
1447 * Returns 0 if successful.
1448 */
1449int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1450 unsigned long size)
1451{
1452 if (address < vma->vm_start || address + size > vma->vm_end ||
1453 !(vma->vm_flags & VM_PFNMAP))
1454 return -1;
1455 zap_page_range_single(vma, address, size, NULL);
1456 return 0;
1457}
1458EXPORT_SYMBOL_GPL(zap_vma_ptes);
1459
1460/**
1461 * follow_page_mask - look up a page descriptor from a user-virtual address
1462 * @vma: vm_area_struct mapping @address
1463 * @address: virtual address to look up
1464 * @flags: flags modifying lookup behaviour
1465 * @page_mask: on output, *page_mask is set according to the size of the page
1466 *
1467 * @flags can have FOLL_ flags set, defined in <linux/mm.h>
1468 *
1469 * Returns the mapped (struct page *), %NULL if no mapping exists, or
1470 * an error pointer if there is a mapping to something not represented
1471 * by a page descriptor (see also vm_normal_page()).
1472 */
1473struct page *follow_page_mask(struct vm_area_struct *vma,
1474 unsigned long address, unsigned int flags,
1475 unsigned int *page_mask)
1476{
1477 pgd_t *pgd;
1478 pud_t *pud;
1479 pmd_t *pmd;
1480 pte_t *ptep, pte;
1481 spinlock_t *ptl;
1482 struct page *page;
1483 struct mm_struct *mm = vma->vm_mm;
1484
1485 *page_mask = 0;
1486
1487 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
1488 if (!IS_ERR(page)) {
1489 BUG_ON(flags & FOLL_GET);
1490 goto out;
1491 }
1492
1493 page = NULL;
1494 pgd = pgd_offset(mm, address);
1495 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
1496 goto no_page_table;
1497
1498 pud = pud_offset(pgd, address);
1499 if (pud_none(*pud))
1500 goto no_page_table;
1501 if (pud_huge(*pud) && vma->vm_flags & VM_HUGETLB) {
1502 if (flags & FOLL_GET)
1503 goto out;
1504 page = follow_huge_pud(mm, address, pud, flags & FOLL_WRITE);
1505 goto out;
1506 }
1507 if (unlikely(pud_bad(*pud)))
1508 goto no_page_table;
1509
1510 pmd = pmd_offset(pud, address);
1511 if (pmd_none(*pmd))
1512 goto no_page_table;
1513 if (pmd_huge(*pmd) && vma->vm_flags & VM_HUGETLB) {
1514 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1515 if (flags & FOLL_GET) {
1516 /*
1517 * Refcount on tail pages are not well-defined and
1518 * shouldn't be taken. The caller should handle a NULL
1519 * return when trying to follow tail pages.
1520 */
1521 if (PageHead(page))
1522 get_page(page);
1523 else {
1524 page = NULL;
1525 goto out;
1526 }
1527 }
1528 goto out;
1529 }
1530 if ((flags & FOLL_NUMA) && pmd_numa(*pmd))
1531 goto no_page_table;
1532 if (pmd_trans_huge(*pmd)) {
1533 if (flags & FOLL_SPLIT) {
1534 split_huge_page_pmd(vma, address, pmd);
1535 goto split_fallthrough;
1536 }
1537 ptl = pmd_lock(mm, pmd);
1538 if (likely(pmd_trans_huge(*pmd))) {
1539 if (unlikely(pmd_trans_splitting(*pmd))) {
1540 spin_unlock(ptl);
1541 wait_split_huge_page(vma->anon_vma, pmd);
1542 } else {
1543 page = follow_trans_huge_pmd(vma, address,
1544 pmd, flags);
1545 spin_unlock(ptl);
1546 *page_mask = HPAGE_PMD_NR - 1;
1547 goto out;
1548 }
1549 } else
1550 spin_unlock(ptl);
1551 /* fall through */
1552 }
1553split_fallthrough:
1554 if (unlikely(pmd_bad(*pmd)))
1555 goto no_page_table;
1556
1557 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1558
1559 pte = *ptep;
1560 if (!pte_present(pte)) {
1561 swp_entry_t entry;
1562 /*
1563 * KSM's break_ksm() relies upon recognizing a ksm page
1564 * even while it is being migrated, so for that case we
1565 * need migration_entry_wait().
1566 */
1567 if (likely(!(flags & FOLL_MIGRATION)))
1568 goto no_page;
1569 if (pte_none(pte) || pte_file(pte))
1570 goto no_page;
1571 entry = pte_to_swp_entry(pte);
1572 if (!is_migration_entry(entry))
1573 goto no_page;
1574 pte_unmap_unlock(ptep, ptl);
1575 migration_entry_wait(mm, pmd, address);
1576 goto split_fallthrough;
1577 }
1578 if ((flags & FOLL_NUMA) && pte_numa(pte))
1579 goto no_page;
1580 if ((flags & FOLL_WRITE) && !pte_write(pte))
1581 goto unlock;
1582
1583 page = vm_normal_page(vma, address, pte);
1584 if (unlikely(!page)) {
1585 if ((flags & FOLL_DUMP) ||
1586 !is_zero_pfn(pte_pfn(pte)))
1587 goto bad_page;
1588 page = pte_page(pte);
1589 }
1590
1591 if (flags & FOLL_GET)
1592 get_page_foll(page);
1593 if (flags & FOLL_TOUCH) {
1594 if ((flags & FOLL_WRITE) &&
1595 !pte_dirty(pte) && !PageDirty(page))
1596 set_page_dirty(page);
1597 /*
1598 * pte_mkyoung() would be more correct here, but atomic care
1599 * is needed to avoid losing the dirty bit: it is easier to use
1600 * mark_page_accessed().
1601 */
1602 mark_page_accessed(page);
1603 }
1604 if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
1605 /*
1606 * The preliminary mapping check is mainly to avoid the
1607 * pointless overhead of lock_page on the ZERO_PAGE
1608 * which might bounce very badly if there is contention.
1609 *
1610 * If the page is already locked, we don't need to
1611 * handle it now - vmscan will handle it later if and
1612 * when it attempts to reclaim the page.
1613 */
1614 if (page->mapping && trylock_page(page)) {
1615 lru_add_drain(); /* push cached pages to LRU */
1616 /*
1617 * Because we lock page here, and migration is
1618 * blocked by the pte's page reference, and we
1619 * know the page is still mapped, we don't even
1620 * need to check for file-cache page truncation.
1621 */
1622 mlock_vma_page(page);
1623 unlock_page(page);
1624 }
1625 }
1626unlock:
1627 pte_unmap_unlock(ptep, ptl);
1628out:
1629 return page;
1630
1631bad_page:
1632 pte_unmap_unlock(ptep, ptl);
1633 return ERR_PTR(-EFAULT);
1634
1635no_page:
1636 pte_unmap_unlock(ptep, ptl);
1637 if (!pte_none(pte))
1638 return page;
1639
1640no_page_table:
1641 /*
1642 * When core dumping an enormous anonymous area that nobody
1643 * has touched so far, we don't want to allocate unnecessary pages or
1644 * page tables. Return error instead of NULL to skip handle_mm_fault,
1645 * then get_dump_page() will return NULL to leave a hole in the dump.
1646 * But we can only make this optimization where a hole would surely
1647 * be zero-filled if handle_mm_fault() actually did handle it.
1648 */
1649 if ((flags & FOLL_DUMP) &&
1650 (!vma->vm_ops || !vma->vm_ops->fault))
1651 return ERR_PTR(-EFAULT);
1652 return page;
1653}
1654
1655static inline int stack_guard_page(struct vm_area_struct *vma, unsigned long addr)
1656{
1657 return stack_guard_page_start(vma, addr) ||
1658 stack_guard_page_end(vma, addr+PAGE_SIZE);
1659}
1660
1661/**
1662 * __get_user_pages() - pin user pages in memory
1663 * @tsk: task_struct of target task
1664 * @mm: mm_struct of target mm
1665 * @start: starting user address
1666 * @nr_pages: number of pages from start to pin
1667 * @gup_flags: flags modifying pin behaviour
1668 * @pages: array that receives pointers to the pages pinned.
1669 * Should be at least nr_pages long. Or NULL, if caller
1670 * only intends to ensure the pages are faulted in.
1671 * @vmas: array of pointers to vmas corresponding to each page.
1672 * Or NULL if the caller does not require them.
1673 * @nonblocking: whether waiting for disk IO or mmap_sem contention
1674 *
1675 * Returns number of pages pinned. This may be fewer than the number
1676 * requested. If nr_pages is 0 or negative, returns 0. If no pages
1677 * were pinned, returns -errno. Each page returned must be released
1678 * with a put_page() call when it is finished with. vmas will only
1679 * remain valid while mmap_sem is held.
1680 *
1681 * Must be called with mmap_sem held for read or write.
1682 *
1683 * __get_user_pages walks a process's page tables and takes a reference to
1684 * each struct page that each user address corresponds to at a given
1685 * instant. That is, it takes the page that would be accessed if a user
1686 * thread accesses the given user virtual address at that instant.
1687 *
1688 * This does not guarantee that the page exists in the user mappings when
1689 * __get_user_pages returns, and there may even be a completely different
1690 * page there in some cases (eg. if mmapped pagecache has been invalidated
1691 * and subsequently re faulted). However it does guarantee that the page
1692 * won't be freed completely. And mostly callers simply care that the page
1693 * contains data that was valid *at some point in time*. Typically, an IO
1694 * or similar operation cannot guarantee anything stronger anyway because
1695 * locks can't be held over the syscall boundary.
1696 *
1697 * If @gup_flags & FOLL_WRITE == 0, the page must not be written to. If
1698 * the page is written to, set_page_dirty (or set_page_dirty_lock, as
1699 * appropriate) must be called after the page is finished with, and
1700 * before put_page is called.
1701 *
1702 * If @nonblocking != NULL, __get_user_pages will not wait for disk IO
1703 * or mmap_sem contention, and if waiting is needed to pin all pages,
1704 * *@nonblocking will be set to 0.
1705 *
1706 * In most cases, get_user_pages or get_user_pages_fast should be used
1707 * instead of __get_user_pages. __get_user_pages should be used only if
1708 * you need some special @gup_flags.
1709 */
1710long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
1711 unsigned long start, unsigned long nr_pages,
1712 unsigned int gup_flags, struct page **pages,
1713 struct vm_area_struct **vmas, int *nonblocking)
1714{
1715 long i;
1716 unsigned long vm_flags;
1717 unsigned int page_mask;
1718
1719 if (!nr_pages)
1720 return 0;
1721
1722 VM_BUG_ON(!!pages != !!(gup_flags & FOLL_GET));
1723
1724 /*
1725 * If FOLL_FORCE and FOLL_NUMA are both set, handle_mm_fault
1726 * would be called on PROT_NONE ranges. We must never invoke
1727 * handle_mm_fault on PROT_NONE ranges or the NUMA hinting
1728 * page faults would unprotect the PROT_NONE ranges if
1729 * _PAGE_NUMA and _PAGE_PROTNONE are sharing the same pte/pmd
1730 * bitflag. So to avoid that, don't set FOLL_NUMA if
1731 * FOLL_FORCE is set.
1732 */
1733 if (!(gup_flags & FOLL_FORCE))
1734 gup_flags |= FOLL_NUMA;
1735
1736 i = 0;
1737
1738 do {
1739 struct vm_area_struct *vma;
1740
1741 vma = find_extend_vma(mm, start);
1742 if (!vma && in_gate_area(mm, start)) {
1743 unsigned long pg = start & PAGE_MASK;
1744 pgd_t *pgd;
1745 pud_t *pud;
1746 pmd_t *pmd;
1747 pte_t *pte;
1748
1749 /* user gate pages are read-only */
1750 if (gup_flags & FOLL_WRITE)
1751 goto efault;
1752 if (pg > TASK_SIZE)
1753 pgd = pgd_offset_k(pg);
1754 else
1755 pgd = pgd_offset_gate(mm, pg);
1756 BUG_ON(pgd_none(*pgd));
1757 pud = pud_offset(pgd, pg);
1758 BUG_ON(pud_none(*pud));
1759 pmd = pmd_offset(pud, pg);
1760 if (pmd_none(*pmd))
1761 goto efault;
1762 VM_BUG_ON(pmd_trans_huge(*pmd));
1763 pte = pte_offset_map(pmd, pg);
1764 if (pte_none(*pte)) {
1765 pte_unmap(pte);
1766 goto efault;
1767 }
1768 vma = get_gate_vma(mm);
1769 if (pages) {
1770 struct page *page;
1771
1772 page = vm_normal_page(vma, start, *pte);
1773 if (!page) {
1774 if (!(gup_flags & FOLL_DUMP) &&
1775 is_zero_pfn(pte_pfn(*pte)))
1776 page = pte_page(*pte);
1777 else {
1778 pte_unmap(pte);
1779 goto efault;
1780 }
1781 }
1782 pages[i] = page;
1783 get_page(page);
1784 }
1785 pte_unmap(pte);
1786 page_mask = 0;
1787 goto next_page;
1788 }
1789
1790 if (!vma)
1791 goto efault;
1792 vm_flags = vma->vm_flags;
1793 if (vm_flags & (VM_IO | VM_PFNMAP))
1794 goto efault;
1795
1796 if (gup_flags & FOLL_WRITE) {
1797 if (!(vm_flags & VM_WRITE)) {
1798 if (!(gup_flags & FOLL_FORCE))
1799 goto efault;
1800 /*
1801 * We used to let the write,force case do COW
1802 * in a VM_MAYWRITE VM_SHARED !VM_WRITE vma, so
1803 * ptrace could set a breakpoint in a read-only
1804 * mapping of an executable, without corrupting
1805 * the file (yet only when that file had been
1806 * opened for writing!). Anon pages in shared
1807 * mappings are surprising: now just reject it.
1808 */
1809 if (!is_cow_mapping(vm_flags)) {
1810 WARN_ON_ONCE(vm_flags & VM_MAYWRITE);
1811 goto efault;
1812 }
1813 }
1814 } else {
1815 if (!(vm_flags & VM_READ)) {
1816 if (!(gup_flags & FOLL_FORCE))
1817 goto efault;
1818 /*
1819 * Is there actually any vma we can reach here
1820 * which does not have VM_MAYREAD set?
1821 */
1822 if (!(vm_flags & VM_MAYREAD))
1823 goto efault;
1824 }
1825 }
1826
1827 if (is_vm_hugetlb_page(vma)) {
1828 i = follow_hugetlb_page(mm, vma, pages, vmas,
1829 &start, &nr_pages, i, gup_flags);
1830 continue;
1831 }
1832
1833 do {
1834 struct page *page;
1835 unsigned int foll_flags = gup_flags;
1836 unsigned int page_increm;
1837
1838 /*
1839 * If we have a pending SIGKILL, don't keep faulting
1840 * pages and potentially allocating memory.
1841 */
1842 if (unlikely(fatal_signal_pending(current)))
1843 return i ? i : -ERESTARTSYS;
1844
1845 cond_resched();
1846 while (!(page = follow_page_mask(vma, start,
1847 foll_flags, &page_mask))) {
1848 int ret;
1849 unsigned int fault_flags = 0;
1850
1851 /* For mlock, just skip the stack guard page. */
1852 if (foll_flags & FOLL_MLOCK) {
1853 if (stack_guard_page(vma, start))
1854 goto next_page;
1855 }
1856 if (foll_flags & FOLL_WRITE)
1857 fault_flags |= FAULT_FLAG_WRITE;
1858 if (nonblocking)
1859 fault_flags |= FAULT_FLAG_ALLOW_RETRY;
1860 if (foll_flags & FOLL_NOWAIT)
1861 fault_flags |= (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT);
1862
1863 ret = handle_mm_fault(mm, vma, start,
1864 fault_flags);
1865
1866 if (ret & VM_FAULT_ERROR) {
1867 if (ret & VM_FAULT_OOM)
1868 return i ? i : -ENOMEM;
1869 if (ret & (VM_FAULT_HWPOISON |
1870 VM_FAULT_HWPOISON_LARGE)) {
1871 if (i)
1872 return i;
1873 else if (gup_flags & FOLL_HWPOISON)
1874 return -EHWPOISON;
1875 else
1876 return -EFAULT;
1877 }
1878 if (ret & VM_FAULT_SIGBUS)
1879 goto efault;
1880 BUG();
1881 }
1882
1883 if (tsk) {
1884 if (ret & VM_FAULT_MAJOR)
1885 tsk->maj_flt++;
1886 else
1887 tsk->min_flt++;
1888 }
1889
1890 if (ret & VM_FAULT_RETRY) {
1891 if (nonblocking)
1892 *nonblocking = 0;
1893 return i;
1894 }
1895
1896 /*
1897 * The VM_FAULT_WRITE bit tells us that
1898 * do_wp_page has broken COW when necessary,
1899 * even if maybe_mkwrite decided not to set
1900 * pte_write. We can thus safely do subsequent
1901 * page lookups as if they were reads. But only
1902 * do so when looping for pte_write is futile:
1903 * in some cases userspace may also be wanting
1904 * to write to the gotten user page, which a
1905 * read fault here might prevent (a readonly
1906 * page might get reCOWed by userspace write).
1907 */
1908 if ((ret & VM_FAULT_WRITE) &&
1909 !(vma->vm_flags & VM_WRITE))
1910 foll_flags &= ~FOLL_WRITE;
1911
1912 cond_resched();
1913 }
1914 if (IS_ERR(page))
1915 return i ? i : PTR_ERR(page);
1916 if (pages) {
1917 pages[i] = page;
1918
1919 flush_anon_page(vma, page, start);
1920 flush_dcache_page(page);
1921 page_mask = 0;
1922 }
1923next_page:
1924 if (vmas) {
1925 vmas[i] = vma;
1926 page_mask = 0;
1927 }
1928 page_increm = 1 + (~(start >> PAGE_SHIFT) & page_mask);
1929 if (page_increm > nr_pages)
1930 page_increm = nr_pages;
1931 i += page_increm;
1932 start += page_increm * PAGE_SIZE;
1933 nr_pages -= page_increm;
1934 } while (nr_pages && start < vma->vm_end);
1935 } while (nr_pages);
1936 return i;
1937efault:
1938 return i ? : -EFAULT;
1939}
1940EXPORT_SYMBOL(__get_user_pages);
1941
1942/*
1943 * fixup_user_fault() - manually resolve a user page fault
1944 * @tsk: the task_struct to use for page fault accounting, or
1945 * NULL if faults are not to be recorded.
1946 * @mm: mm_struct of target mm
1947 * @address: user address
1948 * @fault_flags:flags to pass down to handle_mm_fault()
1949 *
1950 * This is meant to be called in the specific scenario where for locking reasons
1951 * we try to access user memory in atomic context (within a pagefault_disable()
1952 * section), this returns -EFAULT, and we want to resolve the user fault before
1953 * trying again.
1954 *
1955 * Typically this is meant to be used by the futex code.
1956 *
1957 * The main difference with get_user_pages() is that this function will
1958 * unconditionally call handle_mm_fault() which will in turn perform all the
1959 * necessary SW fixup of the dirty and young bits in the PTE, while
1960 * handle_mm_fault() only guarantees to update these in the struct page.
1961 *
1962 * This is important for some architectures where those bits also gate the
1963 * access permission to the page because they are maintained in software. On
1964 * such architectures, gup() will not be enough to make a subsequent access
1965 * succeed.
1966 *
1967 * This should be called with the mm_sem held for read.
1968 */
1969int fixup_user_fault(struct task_struct *tsk, struct mm_struct *mm,
1970 unsigned long address, unsigned int fault_flags)
1971{
1972 struct vm_area_struct *vma;
1973 vm_flags_t vm_flags;
1974 int ret;
1975
1976 vma = find_extend_vma(mm, address);
1977 if (!vma || address < vma->vm_start)
1978 return -EFAULT;
1979
1980 vm_flags = (fault_flags & FAULT_FLAG_WRITE) ? VM_WRITE : VM_READ;
1981 if (!(vm_flags & vma->vm_flags))
1982 return -EFAULT;
1983
1984 ret = handle_mm_fault(mm, vma, address, fault_flags);
1985 if (ret & VM_FAULT_ERROR) {
1986 if (ret & VM_FAULT_OOM)
1987 return -ENOMEM;
1988 if (ret & (VM_FAULT_HWPOISON | VM_FAULT_HWPOISON_LARGE))
1989 return -EHWPOISON;
1990 if (ret & VM_FAULT_SIGBUS)
1991 return -EFAULT;
1992 BUG();
1993 }
1994 if (tsk) {
1995 if (ret & VM_FAULT_MAJOR)
1996 tsk->maj_flt++;
1997 else
1998 tsk->min_flt++;
1999 }
2000 return 0;
2001}
2002
2003/*
2004 * get_user_pages() - pin user pages in memory
2005 * @tsk: the task_struct to use for page fault accounting, or
2006 * NULL if faults are not to be recorded.
2007 * @mm: mm_struct of target mm
2008 * @start: starting user address
2009 * @nr_pages: number of pages from start to pin
2010 * @write: whether pages will be written to by the caller
2011 * @force: whether to force access even when user mapping is currently
2012 * protected (but never forces write access to shared mapping).
2013 * @pages: array that receives pointers to the pages pinned.
2014 * Should be at least nr_pages long. Or NULL, if caller
2015 * only intends to ensure the pages are faulted in.
2016 * @vmas: array of pointers to vmas corresponding to each page.
2017 * Or NULL if the caller does not require them.
2018 *
2019 * Returns number of pages pinned. This may be fewer than the number
2020 * requested. If nr_pages is 0 or negative, returns 0. If no pages
2021 * were pinned, returns -errno. Each page returned must be released
2022 * with a put_page() call when it is finished with. vmas will only
2023 * remain valid while mmap_sem is held.
2024 *
2025 * Must be called with mmap_sem held for read or write.
2026 *
2027 * get_user_pages walks a process's page tables and takes a reference to
2028 * each struct page that each user address corresponds to at a given
2029 * instant. That is, it takes the page that would be accessed if a user
2030 * thread accesses the given user virtual address at that instant.
2031 *
2032 * This does not guarantee that the page exists in the user mappings when
2033 * get_user_pages returns, and there may even be a completely different
2034 * page there in some cases (eg. if mmapped pagecache has been invalidated
2035 * and subsequently re faulted). However it does guarantee that the page
2036 * won't be freed completely. And mostly callers simply care that the page
2037 * contains data that was valid *at some point in time*. Typically, an IO
2038 * or similar operation cannot guarantee anything stronger anyway because
2039 * locks can't be held over the syscall boundary.
2040 *
2041 * If write=0, the page must not be written to. If the page is written to,
2042 * set_page_dirty (or set_page_dirty_lock, as appropriate) must be called
2043 * after the page is finished with, and before put_page is called.
2044 *
2045 * get_user_pages is typically used for fewer-copy IO operations, to get a
2046 * handle on the memory by some means other than accesses via the user virtual
2047 * addresses. The pages may be submitted for DMA to devices or accessed via
2048 * their kernel linear mapping (via the kmap APIs). Care should be taken to
2049 * use the correct cache flushing APIs.
2050 *
2051 * See also get_user_pages_fast, for performance critical applications.
2052 */
2053long get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
2054 unsigned long start, unsigned long nr_pages, int write,
2055 int force, struct page **pages, struct vm_area_struct **vmas)
2056{
2057 int flags = FOLL_TOUCH;
2058
2059 if (pages)
2060 flags |= FOLL_GET;
2061 if (write)
2062 flags |= FOLL_WRITE;
2063 if (force)
2064 flags |= FOLL_FORCE;
2065
2066 return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
2067 NULL);
2068}
2069EXPORT_SYMBOL(get_user_pages);
2070
2071/**
2072 * get_dump_page() - pin user page in memory while writing it to core dump
2073 * @addr: user address
2074 *
2075 * Returns struct page pointer of user page pinned for dump,
2076 * to be freed afterwards by page_cache_release() or put_page().
2077 *
2078 * Returns NULL on any kind of failure - a hole must then be inserted into
2079 * the corefile, to preserve alignment with its headers; and also returns
2080 * NULL wherever the ZERO_PAGE, or an anonymous pte_none, has been found -
2081 * allowing a hole to be left in the corefile to save diskspace.
2082 *
2083 * Called without mmap_sem, but after all other threads have been killed.
2084 */
2085#ifdef CONFIG_ELF_CORE
2086struct page *get_dump_page(unsigned long addr)
2087{
2088 struct vm_area_struct *vma;
2089 struct page *page;
2090
2091 if (__get_user_pages(current, current->mm, addr, 1,
2092 FOLL_FORCE | FOLL_DUMP | FOLL_GET, &page, &vma,
2093 NULL) < 1)
2094 return NULL;
2095 flush_cache_page(vma, addr, page_to_pfn(page));
2096 return page;
2097}
2098#endif /* CONFIG_ELF_CORE */
2099
2100pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
2101 spinlock_t **ptl)
2102{
2103 pgd_t * pgd = pgd_offset(mm, addr);
2104 pud_t * pud = pud_alloc(mm, pgd, addr);
2105 if (pud) {
2106 pmd_t * pmd = pmd_alloc(mm, pud, addr);
2107 if (pmd) {
2108 VM_BUG_ON(pmd_trans_huge(*pmd));
2109 return pte_alloc_map_lock(mm, pmd, addr, ptl);
2110 }
2111 }
2112 return NULL;
2113}
2114
2115/*
2116 * This is the old fallback for page remapping.
2117 *
2118 * For historical reasons, it only allows reserved pages. Only
2119 * old drivers should use this, and they needed to mark their
2120 * pages reserved for the old functions anyway.
2121 */
2122static int insert_page(struct vm_area_struct *vma, unsigned long addr,
2123 struct page *page, pgprot_t prot)
2124{
2125 struct mm_struct *mm = vma->vm_mm;
2126 int retval;
2127 pte_t *pte;
2128 spinlock_t *ptl;
2129
2130 retval = -EINVAL;
2131 if (PageAnon(page))
2132 goto out;
2133 retval = -ENOMEM;
2134 flush_dcache_page(page);
2135 pte = get_locked_pte(mm, addr, &ptl);
2136 if (!pte)
2137 goto out;
2138 retval = -EBUSY;
2139 if (!pte_none(*pte))
2140 goto out_unlock;
2141
2142 /* Ok, finally just insert the thing.. */
2143 get_page(page);
2144 inc_mm_counter_fast(mm, MM_FILEPAGES);
2145 page_add_file_rmap(page);
2146 set_pte_at(mm, addr, pte, mk_pte(page, prot));
2147
2148 retval = 0;
2149 pte_unmap_unlock(pte, ptl);
2150 return retval;
2151out_unlock:
2152 pte_unmap_unlock(pte, ptl);
2153out:
2154 return retval;
2155}
2156
2157/**
2158 * vm_insert_page - insert single page into user vma
2159 * @vma: user vma to map to
2160 * @addr: target user address of this page
2161 * @page: source kernel page
2162 *
2163 * This allows drivers to insert individual pages they've allocated
2164 * into a user vma.
2165 *
2166 * The page has to be a nice clean _individual_ kernel allocation.
2167 * If you allocate a compound page, you need to have marked it as
2168 * such (__GFP_COMP), or manually just split the page up yourself
2169 * (see split_page()).
2170 *
2171 * NOTE! Traditionally this was done with "remap_pfn_range()" which
2172 * took an arbitrary page protection parameter. This doesn't allow
2173 * that. Your vma protection will have to be set up correctly, which
2174 * means that if you want a shared writable mapping, you'd better
2175 * ask for a shared writable mapping!
2176 *
2177 * The page does not need to be reserved.
2178 *
2179 * Usually this function is called from f_op->mmap() handler
2180 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
2181 * Caller must set VM_MIXEDMAP on vma if it wants to call this
2182 * function from other places, for example from page-fault handler.
2183 */
2184int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
2185 struct page *page)
2186{
2187 if (addr < vma->vm_start || addr >= vma->vm_end)
2188 return -EFAULT;
2189 if (!page_count(page))
2190 return -EINVAL;
2191 if (!(vma->vm_flags & VM_MIXEDMAP)) {
2192 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
2193 BUG_ON(vma->vm_flags & VM_PFNMAP);
2194 vma->vm_flags |= VM_MIXEDMAP;
2195 }
2196 return insert_page(vma, addr, page, vma->vm_page_prot);
2197}
2198EXPORT_SYMBOL(vm_insert_page);
2199
2200static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2201 unsigned long pfn, pgprot_t prot)
2202{
2203 struct mm_struct *mm = vma->vm_mm;
2204 int retval;
2205 pte_t *pte, entry;
2206 spinlock_t *ptl;
2207
2208 retval = -ENOMEM;
2209 pte = get_locked_pte(mm, addr, &ptl);
2210 if (!pte)
2211 goto out;
2212 retval = -EBUSY;
2213 if (!pte_none(*pte))
2214 goto out_unlock;
2215
2216 /* Ok, finally just insert the thing.. */
2217 entry = pte_mkspecial(pfn_pte(pfn, prot));
2218 set_pte_at(mm, addr, pte, entry);
2219 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
2220
2221 retval = 0;
2222out_unlock:
2223 pte_unmap_unlock(pte, ptl);
2224out:
2225 return retval;
2226}
2227
2228/**
2229 * vm_insert_pfn - insert single pfn into user vma
2230 * @vma: user vma to map to
2231 * @addr: target user address of this page
2232 * @pfn: source kernel pfn
2233 *
2234 * Similar to vm_insert_page, this allows drivers to insert individual pages
2235 * they've allocated into a user vma. Same comments apply.
2236 *
2237 * This function should only be called from a vm_ops->fault handler, and
2238 * in that case the handler should return NULL.
2239 *
2240 * vma cannot be a COW mapping.
2241 *
2242 * As this is called only for pages that do not currently exist, we
2243 * do not need to flush old virtual caches or the TLB.
2244 */
2245int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2246 unsigned long pfn)
2247{
2248 int ret;
2249 pgprot_t pgprot = vma->vm_page_prot;
2250 /*
2251 * Technically, architectures with pte_special can avoid all these
2252 * restrictions (same for remap_pfn_range). However we would like
2253 * consistency in testing and feature parity among all, so we should
2254 * try to keep these invariants in place for everybody.
2255 */
2256 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2257 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2258 (VM_PFNMAP|VM_MIXEDMAP));
2259 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2260 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2261
2262 if (addr < vma->vm_start || addr >= vma->vm_end)
2263 return -EFAULT;
2264 if (track_pfn_insert(vma, &pgprot, pfn))
2265 return -EINVAL;
2266
2267 ret = insert_pfn(vma, addr, pfn, pgprot);
2268
2269 return ret;
2270}
2271EXPORT_SYMBOL(vm_insert_pfn);
2272
2273int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2274 unsigned long pfn)
2275{
2276 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
2277
2278 if (addr < vma->vm_start || addr >= vma->vm_end)
2279 return -EFAULT;
2280
2281 /*
2282 * If we don't have pte special, then we have to use the pfn_valid()
2283 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2284 * refcount the page if pfn_valid is true (hence insert_page rather
2285 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
2286 * without pte special, it would there be refcounted as a normal page.
2287 */
2288 if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) {
2289 struct page *page;
2290
2291 page = pfn_to_page(pfn);
2292 return insert_page(vma, addr, page, vma->vm_page_prot);
2293 }
2294 return insert_pfn(vma, addr, pfn, vma->vm_page_prot);
2295}
2296EXPORT_SYMBOL(vm_insert_mixed);
2297
2298/*
2299 * maps a range of physical memory into the requested pages. the old
2300 * mappings are removed. any references to nonexistent pages results
2301 * in null mappings (currently treated as "copy-on-access")
2302 */
2303static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2304 unsigned long addr, unsigned long end,
2305 unsigned long pfn, pgprot_t prot)
2306{
2307 pte_t *pte;
2308 spinlock_t *ptl;
2309
2310 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2311 if (!pte)
2312 return -ENOMEM;
2313 arch_enter_lazy_mmu_mode();
2314 do {
2315 BUG_ON(!pte_none(*pte));
2316 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2317 pfn++;
2318 } while (pte++, addr += PAGE_SIZE, addr != end);
2319 arch_leave_lazy_mmu_mode();
2320 pte_unmap_unlock(pte - 1, ptl);
2321 return 0;
2322}
2323
2324static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2325 unsigned long addr, unsigned long end,
2326 unsigned long pfn, pgprot_t prot)
2327{
2328 pmd_t *pmd;
2329 unsigned long next;
2330
2331 pfn -= addr >> PAGE_SHIFT;
2332 pmd = pmd_alloc(mm, pud, addr);
2333 if (!pmd)
2334 return -ENOMEM;
2335 VM_BUG_ON(pmd_trans_huge(*pmd));
2336 do {
2337 next = pmd_addr_end(addr, end);
2338 if (remap_pte_range(mm, pmd, addr, next,
2339 pfn + (addr >> PAGE_SHIFT), prot))
2340 return -ENOMEM;
2341 } while (pmd++, addr = next, addr != end);
2342 return 0;
2343}
2344
2345static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
2346 unsigned long addr, unsigned long end,
2347 unsigned long pfn, pgprot_t prot)
2348{
2349 pud_t *pud;
2350 unsigned long next;
2351
2352 pfn -= addr >> PAGE_SHIFT;
2353 pud = pud_alloc(mm, pgd, addr);
2354 if (!pud)
2355 return -ENOMEM;
2356 do {
2357 next = pud_addr_end(addr, end);
2358 if (remap_pmd_range(mm, pud, addr, next,
2359 pfn + (addr >> PAGE_SHIFT), prot))
2360 return -ENOMEM;
2361 } while (pud++, addr = next, addr != end);
2362 return 0;
2363}
2364
2365/**
2366 * remap_pfn_range - remap kernel memory to userspace
2367 * @vma: user vma to map to
2368 * @addr: target user address to start at
2369 * @pfn: physical address of kernel memory
2370 * @size: size of map area
2371 * @prot: page protection flags for this mapping
2372 *
2373 * Note: this is only safe if the mm semaphore is held when called.
2374 */
2375int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2376 unsigned long pfn, unsigned long size, pgprot_t prot)
2377{
2378 pgd_t *pgd;
2379 unsigned long next;
2380 unsigned long end = addr + PAGE_ALIGN(size);
2381 struct mm_struct *mm = vma->vm_mm;
2382 int err;
2383
2384 /*
2385 * Physically remapped pages are special. Tell the
2386 * rest of the world about it:
2387 * VM_IO tells people not to look at these pages
2388 * (accesses can have side effects).
2389 * VM_PFNMAP tells the core MM that the base pages are just
2390 * raw PFN mappings, and do not have a "struct page" associated
2391 * with them.
2392 * VM_DONTEXPAND
2393 * Disable vma merging and expanding with mremap().
2394 * VM_DONTDUMP
2395 * Omit vma from core dump, even when VM_IO turned off.
2396 *
2397 * There's a horrible special case to handle copy-on-write
2398 * behaviour that some programs depend on. We mark the "original"
2399 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2400 * See vm_normal_page() for details.
2401 */
2402 if (is_cow_mapping(vma->vm_flags)) {
2403 if (addr != vma->vm_start || end != vma->vm_end)
2404 return -EINVAL;
2405 vma->vm_pgoff = pfn;
2406 }
2407
2408 err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size));
2409 if (err)
2410 return -EINVAL;
2411
2412 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2413
2414 BUG_ON(addr >= end);
2415 pfn -= addr >> PAGE_SHIFT;
2416 pgd = pgd_offset(mm, addr);
2417 flush_cache_range(vma, addr, end);
2418 do {
2419 next = pgd_addr_end(addr, end);
2420 err = remap_pud_range(mm, pgd, addr, next,
2421 pfn + (addr >> PAGE_SHIFT), prot);
2422 if (err)
2423 break;
2424 } while (pgd++, addr = next, addr != end);
2425
2426 if (err)
2427 untrack_pfn(vma, pfn, PAGE_ALIGN(size));
2428
2429 return err;
2430}
2431EXPORT_SYMBOL(remap_pfn_range);
2432
2433/**
2434 * vm_iomap_memory - remap memory to userspace
2435 * @vma: user vma to map to
2436 * @start: start of area
2437 * @len: size of area
2438 *
2439 * This is a simplified io_remap_pfn_range() for common driver use. The
2440 * driver just needs to give us the physical memory range to be mapped,
2441 * we'll figure out the rest from the vma information.
2442 *
2443 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2444 * whatever write-combining details or similar.
2445 */
2446int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2447{
2448 unsigned long vm_len, pfn, pages;
2449
2450 /* Check that the physical memory area passed in looks valid */
2451 if (start + len < start)
2452 return -EINVAL;
2453 /*
2454 * You *really* shouldn't map things that aren't page-aligned,
2455 * but we've historically allowed it because IO memory might
2456 * just have smaller alignment.
2457 */
2458 len += start & ~PAGE_MASK;
2459 pfn = start >> PAGE_SHIFT;
2460 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2461 if (pfn + pages < pfn)
2462 return -EINVAL;
2463
2464 /* We start the mapping 'vm_pgoff' pages into the area */
2465 if (vma->vm_pgoff > pages)
2466 return -EINVAL;
2467 pfn += vma->vm_pgoff;
2468 pages -= vma->vm_pgoff;
2469
2470 /* Can we fit all of the mapping? */
2471 vm_len = vma->vm_end - vma->vm_start;
2472 if (vm_len >> PAGE_SHIFT > pages)
2473 return -EINVAL;
2474
2475 /* Ok, let it rip */
2476 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2477}
2478EXPORT_SYMBOL(vm_iomap_memory);
2479
2480static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2481 unsigned long addr, unsigned long end,
2482 pte_fn_t fn, void *data)
2483{
2484 pte_t *pte;
2485 int err;
2486 pgtable_t token;
2487 spinlock_t *uninitialized_var(ptl);
2488
2489 pte = (mm == &init_mm) ?
2490 pte_alloc_kernel(pmd, addr) :
2491 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2492 if (!pte)
2493 return -ENOMEM;
2494
2495 BUG_ON(pmd_huge(*pmd));
2496
2497 arch_enter_lazy_mmu_mode();
2498
2499 token = pmd_pgtable(*pmd);
2500
2501 do {
2502 err = fn(pte++, token, addr, data);
2503 if (err)
2504 break;
2505 } while (addr += PAGE_SIZE, addr != end);
2506
2507 arch_leave_lazy_mmu_mode();
2508
2509 if (mm != &init_mm)
2510 pte_unmap_unlock(pte-1, ptl);
2511 return err;
2512}
2513
2514static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2515 unsigned long addr, unsigned long end,
2516 pte_fn_t fn, void *data)
2517{
2518 pmd_t *pmd;
2519 unsigned long next;
2520 int err;
2521
2522 BUG_ON(pud_huge(*pud));
2523
2524 pmd = pmd_alloc(mm, pud, addr);
2525 if (!pmd)
2526 return -ENOMEM;
2527 do {
2528 next = pmd_addr_end(addr, end);
2529 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
2530 if (err)
2531 break;
2532 } while (pmd++, addr = next, addr != end);
2533 return err;
2534}
2535
2536static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
2537 unsigned long addr, unsigned long end,
2538 pte_fn_t fn, void *data)
2539{
2540 pud_t *pud;
2541 unsigned long next;
2542 int err;
2543
2544 pud = pud_alloc(mm, pgd, addr);
2545 if (!pud)
2546 return -ENOMEM;
2547 do {
2548 next = pud_addr_end(addr, end);
2549 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2550 if (err)
2551 break;
2552 } while (pud++, addr = next, addr != end);
2553 return err;
2554}
2555
2556/*
2557 * Scan a region of virtual memory, filling in page tables as necessary
2558 * and calling a provided function on each leaf page table.
2559 */
2560int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2561 unsigned long size, pte_fn_t fn, void *data)
2562{
2563 pgd_t *pgd;
2564 unsigned long next;
2565 unsigned long end = addr + size;
2566 int err;
2567
2568 BUG_ON(addr >= end);
2569 pgd = pgd_offset(mm, addr);
2570 do {
2571 next = pgd_addr_end(addr, end);
2572 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
2573 if (err)
2574 break;
2575 } while (pgd++, addr = next, addr != end);
2576
2577 return err;
2578}
2579EXPORT_SYMBOL_GPL(apply_to_page_range);
2580
2581/*
2582 * handle_pte_fault chooses page fault handler according to an entry
2583 * which was read non-atomically. Before making any commitment, on
2584 * those architectures or configurations (e.g. i386 with PAE) which
2585 * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault
2586 * must check under lock before unmapping the pte and proceeding
2587 * (but do_wp_page is only called after already making such a check;
2588 * and do_anonymous_page can safely check later on).
2589 */
2590static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2591 pte_t *page_table, pte_t orig_pte)
2592{
2593 int same = 1;
2594#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2595 if (sizeof(pte_t) > sizeof(unsigned long)) {
2596 spinlock_t *ptl = pte_lockptr(mm, pmd);
2597 spin_lock(ptl);
2598 same = pte_same(*page_table, orig_pte);
2599 spin_unlock(ptl);
2600 }
2601#endif
2602 pte_unmap(page_table);
2603 return same;
2604}
2605
2606static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2607{
2608 debug_dma_assert_idle(src);
2609
2610 /*
2611 * If the source page was a PFN mapping, we don't have
2612 * a "struct page" for it. We do a best-effort copy by
2613 * just copying from the original user address. If that
2614 * fails, we just zero-fill it. Live with it.
2615 */
2616 if (unlikely(!src)) {
2617 void *kaddr = kmap_atomic(dst);
2618 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2619
2620 /*
2621 * This really shouldn't fail, because the page is there
2622 * in the page tables. But it might just be unreadable,
2623 * in which case we just give up and fill the result with
2624 * zeroes.
2625 */
2626 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2627 clear_page(kaddr);
2628 kunmap_atomic(kaddr);
2629 flush_dcache_page(dst);
2630 } else
2631 copy_user_highpage(dst, src, va, vma);
2632}
2633
2634/*
2635 * Notify the address space that the page is about to become writable so that
2636 * it can prohibit this or wait for the page to get into an appropriate state.
2637 *
2638 * We do this without the lock held, so that it can sleep if it needs to.
2639 */
2640static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page,
2641 unsigned long address)
2642{
2643 struct vm_fault vmf;
2644 int ret;
2645
2646 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
2647 vmf.pgoff = page->index;
2648 vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2649 vmf.page = page;
2650
2651 ret = vma->vm_ops->page_mkwrite(vma, &vmf);
2652 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2653 return ret;
2654 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2655 lock_page(page);
2656 if (!page->mapping) {
2657 unlock_page(page);
2658 return 0; /* retry */
2659 }
2660 ret |= VM_FAULT_LOCKED;
2661 } else
2662 VM_BUG_ON_PAGE(!PageLocked(page), page);
2663 return ret;
2664}
2665
2666/*
2667 * This routine handles present pages, when users try to write
2668 * to a shared page. It is done by copying the page to a new address
2669 * and decrementing the shared-page counter for the old page.
2670 *
2671 * Note that this routine assumes that the protection checks have been
2672 * done by the caller (the low-level page fault routine in most cases).
2673 * Thus we can safely just mark it writable once we've done any necessary
2674 * COW.
2675 *
2676 * We also mark the page dirty at this point even though the page will
2677 * change only once the write actually happens. This avoids a few races,
2678 * and potentially makes it more efficient.
2679 *
2680 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2681 * but allow concurrent faults), with pte both mapped and locked.
2682 * We return with mmap_sem still held, but pte unmapped and unlocked.
2683 */
2684static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
2685 unsigned long address, pte_t *page_table, pmd_t *pmd,
2686 spinlock_t *ptl, pte_t orig_pte)
2687 __releases(ptl)
2688{
2689 struct page *old_page, *new_page = NULL;
2690 pte_t entry;
2691 int ret = 0;
2692 int page_mkwrite = 0;
2693 struct page *dirty_page = NULL;
2694 unsigned long mmun_start = 0; /* For mmu_notifiers */
2695 unsigned long mmun_end = 0; /* For mmu_notifiers */
2696
2697 old_page = vm_normal_page(vma, address, orig_pte);
2698 if (!old_page) {
2699 /*
2700 * VM_MIXEDMAP !pfn_valid() case
2701 *
2702 * We should not cow pages in a shared writeable mapping.
2703 * Just mark the pages writable as we can't do any dirty
2704 * accounting on raw pfn maps.
2705 */
2706 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2707 (VM_WRITE|VM_SHARED))
2708 goto reuse;
2709 goto gotten;
2710 }
2711
2712 /*
2713 * Take out anonymous pages first, anonymous shared vmas are
2714 * not dirty accountable.
2715 */
2716 if (PageAnon(old_page) && !PageKsm(old_page)) {
2717 if (!trylock_page(old_page)) {
2718 page_cache_get(old_page);
2719 pte_unmap_unlock(page_table, ptl);
2720 lock_page(old_page);
2721 page_table = pte_offset_map_lock(mm, pmd, address,
2722 &ptl);
2723 if (!pte_same(*page_table, orig_pte)) {
2724 unlock_page(old_page);
2725 goto unlock;
2726 }
2727 page_cache_release(old_page);
2728 }
2729 if (reuse_swap_page(old_page)) {
2730 /*
2731 * The page is all ours. Move it to our anon_vma so
2732 * the rmap code will not search our parent or siblings.
2733 * Protected against the rmap code by the page lock.
2734 */
2735 page_move_anon_rmap(old_page, vma, address);
2736 unlock_page(old_page);
2737 goto reuse;
2738 }
2739 unlock_page(old_page);
2740 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2741 (VM_WRITE|VM_SHARED))) {
2742 /*
2743 * Only catch write-faults on shared writable pages,
2744 * read-only shared pages can get COWed by
2745 * get_user_pages(.write=1, .force=1).
2746 */
2747 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2748 int tmp;
2749 page_cache_get(old_page);
2750 pte_unmap_unlock(page_table, ptl);
2751 tmp = do_page_mkwrite(vma, old_page, address);
2752 if (unlikely(!tmp || (tmp &
2753 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2754 page_cache_release(old_page);
2755 return tmp;
2756 }
2757 /*
2758 * Since we dropped the lock we need to revalidate
2759 * the PTE as someone else may have changed it. If
2760 * they did, we just return, as we can count on the
2761 * MMU to tell us if they didn't also make it writable.
2762 */
2763 page_table = pte_offset_map_lock(mm, pmd, address,
2764 &ptl);
2765 if (!pte_same(*page_table, orig_pte)) {
2766 unlock_page(old_page);
2767 goto unlock;
2768 }
2769
2770 page_mkwrite = 1;
2771 }
2772 dirty_page = old_page;
2773 get_page(dirty_page);
2774
2775reuse:
2776 /*
2777 * Clear the pages cpupid information as the existing
2778 * information potentially belongs to a now completely
2779 * unrelated process.
2780 */
2781 if (old_page)
2782 page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1);
2783
2784 flush_cache_page(vma, address, pte_pfn(orig_pte));
2785 entry = pte_mkyoung(orig_pte);
2786 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2787 if (ptep_set_access_flags(vma, address, page_table, entry,1))
2788 update_mmu_cache(vma, address, page_table);
2789 pte_unmap_unlock(page_table, ptl);
2790 ret |= VM_FAULT_WRITE;
2791
2792 if (!dirty_page)
2793 return ret;
2794
2795 /*
2796 * Yes, Virginia, this is actually required to prevent a race
2797 * with clear_page_dirty_for_io() from clearing the page dirty
2798 * bit after it clear all dirty ptes, but before a racing
2799 * do_wp_page installs a dirty pte.
2800 *
2801 * do_shared_fault is protected similarly.
2802 */
2803 if (!page_mkwrite) {
2804 wait_on_page_locked(dirty_page);
2805 set_page_dirty_balance(dirty_page);
2806 /* file_update_time outside page_lock */
2807 if (vma->vm_file)
2808 file_update_time(vma->vm_file);
2809 }
2810 put_page(dirty_page);
2811 if (page_mkwrite) {
2812 struct address_space *mapping = dirty_page->mapping;
2813
2814 set_page_dirty(dirty_page);
2815 unlock_page(dirty_page);
2816 page_cache_release(dirty_page);
2817 if (mapping) {
2818 /*
2819 * Some device drivers do not set page.mapping
2820 * but still dirty their pages
2821 */
2822 balance_dirty_pages_ratelimited(mapping);
2823 }
2824 }
2825
2826 return ret;
2827 }
2828
2829 /*
2830 * Ok, we need to copy. Oh, well..
2831 */
2832 page_cache_get(old_page);
2833gotten:
2834 pte_unmap_unlock(page_table, ptl);
2835
2836 if (unlikely(anon_vma_prepare(vma)))
2837 goto oom;
2838
2839 if (is_zero_pfn(pte_pfn(orig_pte))) {
2840 new_page = alloc_zeroed_user_highpage_movable(vma, address);
2841 if (!new_page)
2842 goto oom;
2843 } else {
2844 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
2845 if (!new_page)
2846 goto oom;
2847 cow_user_page(new_page, old_page, address, vma);
2848 }
2849 __SetPageUptodate(new_page);
2850
2851 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL))
2852 goto oom_free_new;
2853
2854 mmun_start = address & PAGE_MASK;
2855 mmun_end = mmun_start + PAGE_SIZE;
2856 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2857
2858 /*
2859 * Re-check the pte - we dropped the lock
2860 */
2861 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2862 if (likely(pte_same(*page_table, orig_pte))) {
2863 if (old_page) {
2864 if (!PageAnon(old_page)) {
2865 dec_mm_counter_fast(mm, MM_FILEPAGES);
2866 inc_mm_counter_fast(mm, MM_ANONPAGES);
2867 }
2868 } else
2869 inc_mm_counter_fast(mm, MM_ANONPAGES);
2870 flush_cache_page(vma, address, pte_pfn(orig_pte));
2871 entry = mk_pte(new_page, vma->vm_page_prot);
2872 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2873 /*
2874 * Clear the pte entry and flush it first, before updating the
2875 * pte with the new entry. This will avoid a race condition
2876 * seen in the presence of one thread doing SMC and another
2877 * thread doing COW.
2878 */
2879 ptep_clear_flush(vma, address, page_table);
2880 page_add_new_anon_rmap(new_page, vma, address);
2881 /*
2882 * We call the notify macro here because, when using secondary
2883 * mmu page tables (such as kvm shadow page tables), we want the
2884 * new page to be mapped directly into the secondary page table.
2885 */
2886 set_pte_at_notify(mm, address, page_table, entry);
2887 update_mmu_cache(vma, address, page_table);
2888 if (old_page) {
2889 /*
2890 * Only after switching the pte to the new page may
2891 * we remove the mapcount here. Otherwise another
2892 * process may come and find the rmap count decremented
2893 * before the pte is switched to the new page, and
2894 * "reuse" the old page writing into it while our pte
2895 * here still points into it and can be read by other
2896 * threads.
2897 *
2898 * The critical issue is to order this
2899 * page_remove_rmap with the ptp_clear_flush above.
2900 * Those stores are ordered by (if nothing else,)
2901 * the barrier present in the atomic_add_negative
2902 * in page_remove_rmap.
2903 *
2904 * Then the TLB flush in ptep_clear_flush ensures that
2905 * no process can access the old page before the
2906 * decremented mapcount is visible. And the old page
2907 * cannot be reused until after the decremented
2908 * mapcount is visible. So transitively, TLBs to
2909 * old page will be flushed before it can be reused.
2910 */
2911 page_remove_rmap(old_page);
2912 }
2913
2914 /* Free the old page.. */
2915 new_page = old_page;
2916 ret |= VM_FAULT_WRITE;
2917 } else
2918 mem_cgroup_uncharge_page(new_page);
2919
2920 if (new_page)
2921 page_cache_release(new_page);
2922unlock:
2923 pte_unmap_unlock(page_table, ptl);
2924 if (mmun_end > mmun_start)
2925 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2926 if (old_page) {
2927 /*
2928 * Don't let another task, with possibly unlocked vma,
2929 * keep the mlocked page.
2930 */
2931 if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) {
2932 lock_page(old_page); /* LRU manipulation */
2933 munlock_vma_page(old_page);
2934 unlock_page(old_page);
2935 }
2936 page_cache_release(old_page);
2937 }
2938 return ret;
2939oom_free_new:
2940 page_cache_release(new_page);
2941oom:
2942 if (old_page)
2943 page_cache_release(old_page);
2944 return VM_FAULT_OOM;
2945}
2946
2947static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2948 unsigned long start_addr, unsigned long end_addr,
2949 struct zap_details *details)
2950{
2951 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2952}
2953
2954static inline void unmap_mapping_range_tree(struct rb_root *root,
2955 struct zap_details *details)
2956{
2957 struct vm_area_struct *vma;
2958 pgoff_t vba, vea, zba, zea;
2959
2960 vma_interval_tree_foreach(vma, root,
2961 details->first_index, details->last_index) {
2962
2963 vba = vma->vm_pgoff;
2964 vea = vba + vma_pages(vma) - 1;
2965 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
2966 zba = details->first_index;
2967 if (zba < vba)
2968 zba = vba;
2969 zea = details->last_index;
2970 if (zea > vea)
2971 zea = vea;
2972
2973 unmap_mapping_range_vma(vma,
2974 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2975 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2976 details);
2977 }
2978}
2979
2980static inline void unmap_mapping_range_list(struct list_head *head,
2981 struct zap_details *details)
2982{
2983 struct vm_area_struct *vma;
2984
2985 /*
2986 * In nonlinear VMAs there is no correspondence between virtual address
2987 * offset and file offset. So we must perform an exhaustive search
2988 * across *all* the pages in each nonlinear VMA, not just the pages
2989 * whose virtual address lies outside the file truncation point.
2990 */
2991 list_for_each_entry(vma, head, shared.nonlinear) {
2992 details->nonlinear_vma = vma;
2993 unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details);
2994 }
2995}
2996
2997/**
2998 * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file.
2999 * @mapping: the address space containing mmaps to be unmapped.
3000 * @holebegin: byte in first page to unmap, relative to the start of
3001 * the underlying file. This will be rounded down to a PAGE_SIZE
3002 * boundary. Note that this is different from truncate_pagecache(), which
3003 * must keep the partial page. In contrast, we must get rid of
3004 * partial pages.
3005 * @holelen: size of prospective hole in bytes. This will be rounded
3006 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3007 * end of the file.
3008 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3009 * but 0 when invalidating pagecache, don't throw away private data.
3010 */
3011void unmap_mapping_range(struct address_space *mapping,
3012 loff_t const holebegin, loff_t const holelen, int even_cows)
3013{
3014 struct zap_details details;
3015 pgoff_t hba = holebegin >> PAGE_SHIFT;
3016 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3017
3018 /* Check for overflow. */
3019 if (sizeof(holelen) > sizeof(hlen)) {
3020 long long holeend =
3021 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3022 if (holeend & ~(long long)ULONG_MAX)
3023 hlen = ULONG_MAX - hba + 1;
3024 }
3025
3026 details.check_mapping = even_cows? NULL: mapping;
3027 details.nonlinear_vma = NULL;
3028 details.first_index = hba;
3029 details.last_index = hba + hlen - 1;
3030 if (details.last_index < details.first_index)
3031 details.last_index = ULONG_MAX;
3032
3033
3034 mutex_lock(&mapping->i_mmap_mutex);
3035 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
3036 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3037 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
3038 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
3039 mutex_unlock(&mapping->i_mmap_mutex);
3040}
3041EXPORT_SYMBOL(unmap_mapping_range);
3042
3043/*
3044 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3045 * but allow concurrent faults), and pte mapped but not yet locked.
3046 * We return with mmap_sem still held, but pte unmapped and unlocked.
3047 */
3048static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
3049 unsigned long address, pte_t *page_table, pmd_t *pmd,
3050 unsigned int flags, pte_t orig_pte)
3051{
3052 spinlock_t *ptl;
3053 struct page *page, *swapcache;
3054 swp_entry_t entry;
3055 pte_t pte;
3056 int locked;
3057 struct mem_cgroup *ptr;
3058 int exclusive = 0;
3059 int ret = 0;
3060
3061 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3062 goto out;
3063
3064 entry = pte_to_swp_entry(orig_pte);
3065 if (unlikely(non_swap_entry(entry))) {
3066 if (is_migration_entry(entry)) {
3067 migration_entry_wait(mm, pmd, address);
3068 } else if (is_hwpoison_entry(entry)) {
3069 ret = VM_FAULT_HWPOISON;
3070 } else {
3071 print_bad_pte(vma, address, orig_pte, NULL);
3072 ret = VM_FAULT_SIGBUS;
3073 }
3074 goto out;
3075 }
3076 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3077 page = lookup_swap_cache(entry);
3078 if (!page) {
3079 page = swapin_readahead(entry,
3080 GFP_HIGHUSER_MOVABLE, vma, address);
3081 if (!page) {
3082 /*
3083 * Back out if somebody else faulted in this pte
3084 * while we released the pte lock.
3085 */
3086 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3087 if (likely(pte_same(*page_table, orig_pte)))
3088 ret = VM_FAULT_OOM;
3089 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3090 goto unlock;
3091 }
3092
3093 /* Had to read the page from swap area: Major fault */
3094 ret = VM_FAULT_MAJOR;
3095 count_vm_event(PGMAJFAULT);
3096 mem_cgroup_count_vm_event(mm, PGMAJFAULT);
3097 } else if (PageHWPoison(page)) {
3098 /*
3099 * hwpoisoned dirty swapcache pages are kept for killing
3100 * owner processes (which may be unknown at hwpoison time)
3101 */
3102 ret = VM_FAULT_HWPOISON;
3103 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3104 swapcache = page;
3105 goto out_release;
3106 }
3107
3108 swapcache = page;
3109 locked = lock_page_or_retry(page, mm, flags);
3110
3111 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3112 if (!locked) {
3113 ret |= VM_FAULT_RETRY;
3114 goto out_release;
3115 }
3116
3117 /*
3118 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3119 * release the swapcache from under us. The page pin, and pte_same
3120 * test below, are not enough to exclude that. Even if it is still
3121 * swapcache, we need to check that the page's swap has not changed.
3122 */
3123 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
3124 goto out_page;
3125
3126 page = ksm_might_need_to_copy(page, vma, address);
3127 if (unlikely(!page)) {
3128 ret = VM_FAULT_OOM;
3129 page = swapcache;
3130 goto out_page;
3131 }
3132
3133 if (mem_cgroup_try_charge_swapin(mm, page, GFP_KERNEL, &ptr)) {
3134 ret = VM_FAULT_OOM;
3135 goto out_page;
3136 }
3137
3138 /*
3139 * Back out if somebody else already faulted in this pte.
3140 */
3141 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3142 if (unlikely(!pte_same(*page_table, orig_pte)))
3143 goto out_nomap;
3144
3145 if (unlikely(!PageUptodate(page))) {
3146 ret = VM_FAULT_SIGBUS;
3147 goto out_nomap;
3148 }
3149
3150 /*
3151 * The page isn't present yet, go ahead with the fault.
3152 *
3153 * Be careful about the sequence of operations here.
3154 * To get its accounting right, reuse_swap_page() must be called
3155 * while the page is counted on swap but not yet in mapcount i.e.
3156 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3157 * must be called after the swap_free(), or it will never succeed.
3158 * Because delete_from_swap_page() may be called by reuse_swap_page(),
3159 * mem_cgroup_commit_charge_swapin() may not be able to find swp_entry
3160 * in page->private. In this case, a record in swap_cgroup is silently
3161 * discarded at swap_free().
3162 */
3163
3164 inc_mm_counter_fast(mm, MM_ANONPAGES);
3165 dec_mm_counter_fast(mm, MM_SWAPENTS);
3166 pte = mk_pte(page, vma->vm_page_prot);
3167 if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) {
3168 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3169 flags &= ~FAULT_FLAG_WRITE;
3170 ret |= VM_FAULT_WRITE;
3171 exclusive = 1;
3172 }
3173 flush_icache_page(vma, page);
3174 if (pte_swp_soft_dirty(orig_pte))
3175 pte = pte_mksoft_dirty(pte);
3176 set_pte_at(mm, address, page_table, pte);
3177 if (page == swapcache)
3178 do_page_add_anon_rmap(page, vma, address, exclusive);
3179 else /* ksm created a completely new copy */
3180 page_add_new_anon_rmap(page, vma, address);
3181 /* It's better to call commit-charge after rmap is established */
3182 mem_cgroup_commit_charge_swapin(page, ptr);
3183
3184 swap_free(entry);
3185 if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3186 try_to_free_swap(page);
3187 unlock_page(page);
3188 if (page != swapcache) {
3189 /*
3190 * Hold the lock to avoid the swap entry to be reused
3191 * until we take the PT lock for the pte_same() check
3192 * (to avoid false positives from pte_same). For
3193 * further safety release the lock after the swap_free
3194 * so that the swap count won't change under a
3195 * parallel locked swapcache.
3196 */
3197 unlock_page(swapcache);
3198 page_cache_release(swapcache);
3199 }
3200
3201 if (flags & FAULT_FLAG_WRITE) {
3202 ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte);
3203 if (ret & VM_FAULT_ERROR)
3204 ret &= VM_FAULT_ERROR;
3205 goto out;
3206 }
3207
3208 /* No need to invalidate - it was non-present before */
3209 update_mmu_cache(vma, address, page_table);
3210unlock:
3211 pte_unmap_unlock(page_table, ptl);
3212out:
3213 return ret;
3214out_nomap:
3215 mem_cgroup_cancel_charge_swapin(ptr);
3216 pte_unmap_unlock(page_table, ptl);
3217out_page:
3218 unlock_page(page);
3219out_release:
3220 page_cache_release(page);
3221 if (page != swapcache) {
3222 unlock_page(swapcache);
3223 page_cache_release(swapcache);
3224 }
3225 return ret;
3226}
3227
3228/*
3229 * This is like a special single-page "expand_{down|up}wards()",
3230 * except we must first make sure that 'address{-|+}PAGE_SIZE'
3231 * doesn't hit another vma.
3232 */
3233static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
3234{
3235 address &= PAGE_MASK;
3236 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
3237 struct vm_area_struct *prev = vma->vm_prev;
3238
3239 /*
3240 * Is there a mapping abutting this one below?
3241 *
3242 * That's only ok if it's the same stack mapping
3243 * that has gotten split..
3244 */
3245 if (prev && prev->vm_end == address)
3246 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
3247
3248 expand_downwards(vma, address - PAGE_SIZE);
3249 }
3250 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
3251 struct vm_area_struct *next = vma->vm_next;
3252
3253 /* As VM_GROWSDOWN but s/below/above/ */
3254 if (next && next->vm_start == address + PAGE_SIZE)
3255 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
3256
3257 expand_upwards(vma, address + PAGE_SIZE);
3258 }
3259 return 0;
3260}
3261
3262/*
3263 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3264 * but allow concurrent faults), and pte mapped but not yet locked.
3265 * We return with mmap_sem still held, but pte unmapped and unlocked.
3266 */
3267static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
3268 unsigned long address, pte_t *page_table, pmd_t *pmd,
3269 unsigned int flags)
3270{
3271 struct page *page;
3272 spinlock_t *ptl;
3273 pte_t entry;
3274
3275 pte_unmap(page_table);
3276
3277 /* Check if we need to add a guard page to the stack */
3278 if (check_stack_guard_page(vma, address) < 0)
3279 return VM_FAULT_SIGBUS;
3280
3281 /* Use the zero-page for reads */
3282 if (!(flags & FAULT_FLAG_WRITE)) {
3283 entry = pte_mkspecial(pfn_pte(my_zero_pfn(address),
3284 vma->vm_page_prot));
3285 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3286 if (!pte_none(*page_table))
3287 goto unlock;
3288 goto setpte;
3289 }
3290
3291 /* Allocate our own private page. */
3292 if (unlikely(anon_vma_prepare(vma)))
3293 goto oom;
3294 page = alloc_zeroed_user_highpage_movable(vma, address);
3295 if (!page)
3296 goto oom;
3297 /*
3298 * The memory barrier inside __SetPageUptodate makes sure that
3299 * preceeding stores to the page contents become visible before
3300 * the set_pte_at() write.
3301 */
3302 __SetPageUptodate(page);
3303
3304 if (mem_cgroup_charge_anon(page, mm, GFP_KERNEL))
3305 goto oom_free_page;
3306
3307 entry = mk_pte(page, vma->vm_page_prot);
3308 if (vma->vm_flags & VM_WRITE)
3309 entry = pte_mkwrite(pte_mkdirty(entry));
3310
3311 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
3312 if (!pte_none(*page_table))
3313 goto release;
3314
3315 inc_mm_counter_fast(mm, MM_ANONPAGES);
3316 page_add_new_anon_rmap(page, vma, address);
3317setpte:
3318 set_pte_at(mm, address, page_table, entry);
3319
3320 /* No need to invalidate - it was non-present before */
3321 update_mmu_cache(vma, address, page_table);
3322unlock:
3323 pte_unmap_unlock(page_table, ptl);
3324 return 0;
3325release:
3326 mem_cgroup_uncharge_page(page);
3327 page_cache_release(page);
3328 goto unlock;
3329oom_free_page:
3330 page_cache_release(page);
3331oom:
3332 return VM_FAULT_OOM;
3333}
3334
3335static int __do_fault(struct vm_area_struct *vma, unsigned long address,
3336 pgoff_t pgoff, unsigned int flags, struct page **page)
3337{
3338 struct vm_fault vmf;
3339 int ret;
3340
3341 vmf.virtual_address = (void __user *)(address & PAGE_MASK);
3342 vmf.pgoff = pgoff;
3343 vmf.flags = flags;
3344 vmf.page = NULL;
3345
3346 ret = vma->vm_ops->fault(vma, &vmf);
3347 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3348 return ret;
3349
3350 if (unlikely(PageHWPoison(vmf.page))) {
3351 if (ret & VM_FAULT_LOCKED)
3352 unlock_page(vmf.page);
3353 page_cache_release(vmf.page);
3354 return VM_FAULT_HWPOISON;
3355 }
3356
3357 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3358 lock_page(vmf.page);
3359 else
3360 VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page);
3361
3362 *page = vmf.page;
3363 return ret;
3364}
3365
3366/**
3367 * do_set_pte - setup new PTE entry for given page and add reverse page mapping.
3368 *
3369 * @vma: virtual memory area
3370 * @address: user virtual address
3371 * @page: page to map
3372 * @pte: pointer to target page table entry
3373 * @write: true, if new entry is writable
3374 * @anon: true, if it's anonymous page
3375 *
3376 * Caller must hold page table lock relevant for @pte.
3377 *
3378 * Target users are page handler itself and implementations of
3379 * vm_ops->map_pages.
3380 */
3381void do_set_pte(struct vm_area_struct *vma, unsigned long address,
3382 struct page *page, pte_t *pte, bool write, bool anon)
3383{
3384 pte_t entry;
3385
3386 flush_icache_page(vma, page);
3387 entry = mk_pte(page, vma->vm_page_prot);
3388 if (write)
3389 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3390 else if (pte_file(*pte) && pte_file_soft_dirty(*pte))
3391 pte_mksoft_dirty(entry);
3392 if (anon) {
3393 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3394 page_add_new_anon_rmap(page, vma, address);
3395 } else {
3396 inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES);
3397 page_add_file_rmap(page);
3398 }
3399 set_pte_at(vma->vm_mm, address, pte, entry);
3400
3401 /* no need to invalidate: a not-present page won't be cached */
3402 update_mmu_cache(vma, address, pte);
3403}
3404
3405#define FAULT_AROUND_ORDER 4
3406
3407#ifdef CONFIG_DEBUG_FS
3408static unsigned int fault_around_order = FAULT_AROUND_ORDER;
3409
3410static int fault_around_order_get(void *data, u64 *val)
3411{
3412 *val = fault_around_order;
3413 return 0;
3414}
3415
3416static int fault_around_order_set(void *data, u64 val)
3417{
3418 BUILD_BUG_ON((1UL << FAULT_AROUND_ORDER) > PTRS_PER_PTE);
3419 if (1UL << val > PTRS_PER_PTE)
3420 return -EINVAL;
3421 fault_around_order = val;
3422 return 0;
3423}
3424DEFINE_SIMPLE_ATTRIBUTE(fault_around_order_fops,
3425 fault_around_order_get, fault_around_order_set, "%llu\n");
3426
3427static int __init fault_around_debugfs(void)
3428{
3429 void *ret;
3430
3431 ret = debugfs_create_file("fault_around_order", 0644, NULL, NULL,
3432 &fault_around_order_fops);
3433 if (!ret)
3434 pr_warn("Failed to create fault_around_order in debugfs");
3435 return 0;
3436}
3437late_initcall(fault_around_debugfs);
3438
3439static inline unsigned long fault_around_pages(void)
3440{
3441 return 1UL << fault_around_order;
3442}
3443
3444static inline unsigned long fault_around_mask(void)
3445{
3446 return ~((1UL << (PAGE_SHIFT + fault_around_order)) - 1);
3447}
3448#else
3449static inline unsigned long fault_around_pages(void)
3450{
3451 unsigned long nr_pages;
3452
3453 nr_pages = 1UL << FAULT_AROUND_ORDER;
3454 BUILD_BUG_ON(nr_pages > PTRS_PER_PTE);
3455 return nr_pages;
3456}
3457
3458static inline unsigned long fault_around_mask(void)
3459{
3460 return ~((1UL << (PAGE_SHIFT + FAULT_AROUND_ORDER)) - 1);
3461}
3462#endif
3463
3464static void do_fault_around(struct vm_area_struct *vma, unsigned long address,
3465 pte_t *pte, pgoff_t pgoff, unsigned int flags)
3466{
3467 unsigned long start_addr;
3468 pgoff_t max_pgoff;
3469 struct vm_fault vmf;
3470 int off;
3471
3472 start_addr = max(address & fault_around_mask(), vma->vm_start);
3473 off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3474 pte -= off;
3475 pgoff -= off;
3476
3477 /*
3478 * max_pgoff is either end of page table or end of vma
3479 * or fault_around_pages() from pgoff, depending what is neast.
3480 */
3481 max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3482 PTRS_PER_PTE - 1;
3483 max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1,
3484 pgoff + fault_around_pages() - 1);
3485
3486 /* Check if it makes any sense to call ->map_pages */
3487 while (!pte_none(*pte)) {
3488 if (++pgoff > max_pgoff)
3489 return;
3490 start_addr += PAGE_SIZE;
3491 if (start_addr >= vma->vm_end)
3492 return;
3493 pte++;
3494 }
3495
3496 vmf.virtual_address = (void __user *) start_addr;
3497 vmf.pte = pte;
3498 vmf.pgoff = pgoff;
3499 vmf.max_pgoff = max_pgoff;
3500 vmf.flags = flags;
3501 vma->vm_ops->map_pages(vma, &vmf);
3502}
3503
3504static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3505 unsigned long address, pmd_t *pmd,
3506 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3507{
3508 struct page *fault_page;
3509 spinlock_t *ptl;
3510 pte_t *pte;
3511 int ret = 0;
3512
3513 /*
3514 * Let's call ->map_pages() first and use ->fault() as fallback
3515 * if page by the offset is not ready to be mapped (cold cache or
3516 * something).
3517 */
3518 if (vma->vm_ops->map_pages) {
3519 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3520 do_fault_around(vma, address, pte, pgoff, flags);
3521 if (!pte_same(*pte, orig_pte))
3522 goto unlock_out;
3523 pte_unmap_unlock(pte, ptl);
3524 }
3525
3526 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3527 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3528 return ret;
3529
3530 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3531 if (unlikely(!pte_same(*pte, orig_pte))) {
3532 pte_unmap_unlock(pte, ptl);
3533 unlock_page(fault_page);
3534 page_cache_release(fault_page);
3535 return ret;
3536 }
3537 do_set_pte(vma, address, fault_page, pte, false, false);
3538 unlock_page(fault_page);
3539unlock_out:
3540 pte_unmap_unlock(pte, ptl);
3541 return ret;
3542}
3543
3544static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3545 unsigned long address, pmd_t *pmd,
3546 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3547{
3548 struct page *fault_page, *new_page;
3549 spinlock_t *ptl;
3550 pte_t *pte;
3551 int ret;
3552
3553 if (unlikely(anon_vma_prepare(vma)))
3554 return VM_FAULT_OOM;
3555
3556 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address);
3557 if (!new_page)
3558 return VM_FAULT_OOM;
3559
3560 if (mem_cgroup_charge_anon(new_page, mm, GFP_KERNEL)) {
3561 page_cache_release(new_page);
3562 return VM_FAULT_OOM;
3563 }
3564
3565 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3566 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3567 goto uncharge_out;
3568
3569 copy_user_highpage(new_page, fault_page, address, vma);
3570 __SetPageUptodate(new_page);
3571
3572 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3573 if (unlikely(!pte_same(*pte, orig_pte))) {
3574 pte_unmap_unlock(pte, ptl);
3575 unlock_page(fault_page);
3576 page_cache_release(fault_page);
3577 goto uncharge_out;
3578 }
3579 do_set_pte(vma, address, new_page, pte, true, true);
3580 pte_unmap_unlock(pte, ptl);
3581 unlock_page(fault_page);
3582 page_cache_release(fault_page);
3583 return ret;
3584uncharge_out:
3585 mem_cgroup_uncharge_page(new_page);
3586 page_cache_release(new_page);
3587 return ret;
3588}
3589
3590static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3591 unsigned long address, pmd_t *pmd,
3592 pgoff_t pgoff, unsigned int flags, pte_t orig_pte)
3593{
3594 struct page *fault_page;
3595 struct address_space *mapping;
3596 spinlock_t *ptl;
3597 pte_t *pte;
3598 int dirtied = 0;
3599 int ret, tmp;
3600
3601 ret = __do_fault(vma, address, pgoff, flags, &fault_page);
3602 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3603 return ret;
3604
3605 /*
3606 * Check if the backing address space wants to know that the page is
3607 * about to become writable
3608 */
3609 if (vma->vm_ops->page_mkwrite) {
3610 unlock_page(fault_page);
3611 tmp = do_page_mkwrite(vma, fault_page, address);
3612 if (unlikely(!tmp ||
3613 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3614 page_cache_release(fault_page);
3615 return tmp;
3616 }
3617 }
3618
3619 pte = pte_offset_map_lock(mm, pmd, address, &ptl);
3620 if (unlikely(!pte_same(*pte, orig_pte))) {
3621 pte_unmap_unlock(pte, ptl);
3622 unlock_page(fault_page);
3623 page_cache_release(fault_page);
3624 return ret;
3625 }
3626 do_set_pte(vma, address, fault_page, pte, true, false);
3627 pte_unmap_unlock(pte, ptl);
3628
3629 if (set_page_dirty(fault_page))
3630 dirtied = 1;
3631 mapping = fault_page->mapping;
3632 unlock_page(fault_page);
3633 if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) {
3634 /*
3635 * Some device drivers do not set page.mapping but still
3636 * dirty their pages
3637 */
3638 balance_dirty_pages_ratelimited(mapping);
3639 }
3640
3641 /* file_update_time outside page_lock */
3642 if (vma->vm_file && !vma->vm_ops->page_mkwrite)
3643 file_update_time(vma->vm_file);
3644
3645 return ret;
3646}
3647
3648static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3649 unsigned long address, pte_t *page_table, pmd_t *pmd,
3650 unsigned int flags, pte_t orig_pte)
3651{
3652 pgoff_t pgoff = (((address & PAGE_MASK)
3653 - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
3654
3655 pte_unmap(page_table);
3656 if (!(flags & FAULT_FLAG_WRITE))
3657 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3658 orig_pte);
3659 if (!(vma->vm_flags & VM_SHARED))
3660 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3661 orig_pte);
3662 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3663}
3664
3665/*
3666 * Fault of a previously existing named mapping. Repopulate the pte
3667 * from the encoded file_pte if possible. This enables swappable
3668 * nonlinear vmas.
3669 *
3670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3671 * but allow concurrent faults), and pte mapped but not yet locked.
3672 * We return with mmap_sem still held, but pte unmapped and unlocked.
3673 */
3674static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3675 unsigned long address, pte_t *page_table, pmd_t *pmd,
3676 unsigned int flags, pte_t orig_pte)
3677{
3678 pgoff_t pgoff;
3679
3680 flags |= FAULT_FLAG_NONLINEAR;
3681
3682 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
3683 return 0;
3684
3685 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
3686 /*
3687 * Page table corrupted: show pte and kill process.
3688 */
3689 print_bad_pte(vma, address, orig_pte, NULL);
3690 return VM_FAULT_SIGBUS;
3691 }
3692
3693 pgoff = pte_to_pgoff(orig_pte);
3694 if (!(flags & FAULT_FLAG_WRITE))
3695 return do_read_fault(mm, vma, address, pmd, pgoff, flags,
3696 orig_pte);
3697 if (!(vma->vm_flags & VM_SHARED))
3698 return do_cow_fault(mm, vma, address, pmd, pgoff, flags,
3699 orig_pte);
3700 return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte);
3701}
3702
3703static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3704 unsigned long addr, int page_nid,
3705 int *flags)
3706{
3707 get_page(page);
3708
3709 count_vm_numa_event(NUMA_HINT_FAULTS);
3710 if (page_nid == numa_node_id()) {
3711 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3712 *flags |= TNF_FAULT_LOCAL;
3713 }
3714
3715 return mpol_misplaced(page, vma, addr);
3716}
3717
3718static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma,
3719 unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd)
3720{
3721 struct page *page = NULL;
3722 spinlock_t *ptl;
3723 int page_nid = -1;
3724 int last_cpupid;
3725 int target_nid;
3726 bool migrated = false;
3727 int flags = 0;
3728
3729 /*
3730 * The "pte" at this point cannot be used safely without
3731 * validation through pte_unmap_same(). It's of NUMA type but
3732 * the pfn may be screwed if the read is non atomic.
3733 *
3734 * ptep_modify_prot_start is not called as this is clearing
3735 * the _PAGE_NUMA bit and it is not really expected that there
3736 * would be concurrent hardware modifications to the PTE.
3737 */
3738 ptl = pte_lockptr(mm, pmd);
3739 spin_lock(ptl);
3740 if (unlikely(!pte_same(*ptep, pte))) {
3741 pte_unmap_unlock(ptep, ptl);
3742 goto out;
3743 }
3744
3745 pte = pte_mknonnuma(pte);
3746 set_pte_at(mm, addr, ptep, pte);
3747 update_mmu_cache(vma, addr, ptep);
3748
3749 page = vm_normal_page(vma, addr, pte);
3750 if (!page) {
3751 pte_unmap_unlock(ptep, ptl);
3752 return 0;
3753 }
3754 BUG_ON(is_zero_pfn(page_to_pfn(page)));
3755
3756 /*
3757 * Avoid grouping on DSO/COW pages in specific and RO pages
3758 * in general, RO pages shouldn't hurt as much anyway since
3759 * they can be in shared cache state.
3760 */
3761 if (!pte_write(pte))
3762 flags |= TNF_NO_GROUP;
3763
3764 /*
3765 * Flag if the page is shared between multiple address spaces. This
3766 * is later used when determining whether to group tasks together
3767 */
3768 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3769 flags |= TNF_SHARED;
3770
3771 last_cpupid = page_cpupid_last(page);
3772 page_nid = page_to_nid(page);
3773 target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags);
3774 pte_unmap_unlock(ptep, ptl);
3775 if (target_nid == -1) {
3776 put_page(page);
3777 goto out;
3778 }
3779
3780 /* Migrate to the requested node */
3781 migrated = migrate_misplaced_page(page, vma, target_nid);
3782 if (migrated) {
3783 page_nid = target_nid;
3784 flags |= TNF_MIGRATED;
3785 }
3786
3787out:
3788 if (page_nid != -1)
3789 task_numa_fault(last_cpupid, page_nid, 1, flags);
3790 return 0;
3791}
3792
3793/*
3794 * These routines also need to handle stuff like marking pages dirty
3795 * and/or accessed for architectures that don't do it in hardware (most
3796 * RISC architectures). The early dirtying is also good on the i386.
3797 *
3798 * There is also a hook called "update_mmu_cache()" that architectures
3799 * with external mmu caches can use to update those (ie the Sparc or
3800 * PowerPC hashed page tables that act as extended TLBs).
3801 *
3802 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3803 * but allow concurrent faults), and pte mapped but not yet locked.
3804 * We return with mmap_sem still held, but pte unmapped and unlocked.
3805 */
3806static int handle_pte_fault(struct mm_struct *mm,
3807 struct vm_area_struct *vma, unsigned long address,
3808 pte_t *pte, pmd_t *pmd, unsigned int flags)
3809{
3810 pte_t entry;
3811 spinlock_t *ptl;
3812
3813 entry = *pte;
3814 if (!pte_present(entry)) {
3815 if (pte_none(entry)) {
3816 if (vma->vm_ops) {
3817 if (likely(vma->vm_ops->fault))
3818 return do_linear_fault(mm, vma, address,
3819 pte, pmd, flags, entry);
3820 }
3821 return do_anonymous_page(mm, vma, address,
3822 pte, pmd, flags);
3823 }
3824 if (pte_file(entry))
3825 return do_nonlinear_fault(mm, vma, address,
3826 pte, pmd, flags, entry);
3827 return do_swap_page(mm, vma, address,
3828 pte, pmd, flags, entry);
3829 }
3830
3831 if (pte_numa(entry))
3832 return do_numa_page(mm, vma, address, entry, pte, pmd);
3833
3834 ptl = pte_lockptr(mm, pmd);
3835 spin_lock(ptl);
3836 if (unlikely(!pte_same(*pte, entry)))
3837 goto unlock;
3838 if (flags & FAULT_FLAG_WRITE) {
3839 if (!pte_write(entry))
3840 return do_wp_page(mm, vma, address,
3841 pte, pmd, ptl, entry);
3842 entry = pte_mkdirty(entry);
3843 }
3844 entry = pte_mkyoung(entry);
3845 if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) {
3846 update_mmu_cache(vma, address, pte);
3847 } else {
3848 /*
3849 * This is needed only for protection faults but the arch code
3850 * is not yet telling us if this is a protection fault or not.
3851 * This still avoids useless tlb flushes for .text page faults
3852 * with threads.
3853 */
3854 if (flags & FAULT_FLAG_WRITE)
3855 flush_tlb_fix_spurious_fault(vma, address);
3856 }
3857unlock:
3858 pte_unmap_unlock(pte, ptl);
3859 return 0;
3860}
3861
3862/*
3863 * By the time we get here, we already hold the mm semaphore
3864 */
3865static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3866 unsigned long address, unsigned int flags)
3867{
3868 pgd_t *pgd;
3869 pud_t *pud;
3870 pmd_t *pmd;
3871 pte_t *pte;
3872
3873 if (unlikely(is_vm_hugetlb_page(vma)))
3874 return hugetlb_fault(mm, vma, address, flags);
3875
3876 pgd = pgd_offset(mm, address);
3877 pud = pud_alloc(mm, pgd, address);
3878 if (!pud)
3879 return VM_FAULT_OOM;
3880 pmd = pmd_alloc(mm, pud, address);
3881 if (!pmd)
3882 return VM_FAULT_OOM;
3883 if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) {
3884 int ret = VM_FAULT_FALLBACK;
3885 if (!vma->vm_ops)
3886 ret = do_huge_pmd_anonymous_page(mm, vma, address,
3887 pmd, flags);
3888 if (!(ret & VM_FAULT_FALLBACK))
3889 return ret;
3890 } else {
3891 pmd_t orig_pmd = *pmd;
3892 int ret;
3893
3894 barrier();
3895 if (pmd_trans_huge(orig_pmd)) {
3896 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3897
3898 /*
3899 * If the pmd is splitting, return and retry the
3900 * the fault. Alternative: wait until the split
3901 * is done, and goto retry.
3902 */
3903 if (pmd_trans_splitting(orig_pmd))
3904 return 0;
3905
3906 if (pmd_numa(orig_pmd))
3907 return do_huge_pmd_numa_page(mm, vma, address,
3908 orig_pmd, pmd);
3909
3910 if (dirty && !pmd_write(orig_pmd)) {
3911 ret = do_huge_pmd_wp_page(mm, vma, address, pmd,
3912 orig_pmd);
3913 if (!(ret & VM_FAULT_FALLBACK))
3914 return ret;
3915 } else {
3916 huge_pmd_set_accessed(mm, vma, address, pmd,
3917 orig_pmd, dirty);
3918 return 0;
3919 }
3920 }
3921 }
3922
3923 /* THP should already have been handled */
3924 BUG_ON(pmd_numa(*pmd));
3925
3926 /*
3927 * Use __pte_alloc instead of pte_alloc_map, because we can't
3928 * run pte_offset_map on the pmd, if an huge pmd could
3929 * materialize from under us from a different thread.
3930 */
3931 if (unlikely(pmd_none(*pmd)) &&
3932 unlikely(__pte_alloc(mm, vma, pmd, address)))
3933 return VM_FAULT_OOM;
3934 /* if an huge pmd materialized from under us just retry later */
3935 if (unlikely(pmd_trans_huge(*pmd)))
3936 return 0;
3937 /*
3938 * A regular pmd is established and it can't morph into a huge pmd
3939 * from under us anymore at this point because we hold the mmap_sem
3940 * read mode and khugepaged takes it in write mode. So now it's
3941 * safe to run pte_offset_map().
3942 */
3943 pte = pte_offset_map(pmd, address);
3944
3945 return handle_pte_fault(mm, vma, address, pte, pmd, flags);
3946}
3947
3948int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
3949 unsigned long address, unsigned int flags)
3950{
3951 int ret;
3952
3953 __set_current_state(TASK_RUNNING);
3954
3955 count_vm_event(PGFAULT);
3956 mem_cgroup_count_vm_event(mm, PGFAULT);
3957
3958 /* do counter updates before entering really critical section. */
3959 check_sync_rss_stat(current);
3960
3961 /*
3962 * Enable the memcg OOM handling for faults triggered in user
3963 * space. Kernel faults are handled more gracefully.
3964 */
3965 if (flags & FAULT_FLAG_USER)
3966 mem_cgroup_oom_enable();
3967
3968 ret = __handle_mm_fault(mm, vma, address, flags);
3969
3970 if (flags & FAULT_FLAG_USER) {
3971 mem_cgroup_oom_disable();
3972 /*
3973 * The task may have entered a memcg OOM situation but
3974 * if the allocation error was handled gracefully (no
3975 * VM_FAULT_OOM), there is no need to kill anything.
3976 * Just clean up the OOM state peacefully.
3977 */
3978 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3979 mem_cgroup_oom_synchronize(false);
3980 }
3981
3982 return ret;
3983}
3984
3985#ifndef __PAGETABLE_PUD_FOLDED
3986/*
3987 * Allocate page upper directory.
3988 * We've already handled the fast-path in-line.
3989 */
3990int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3991{
3992 pud_t *new = pud_alloc_one(mm, address);
3993 if (!new)
3994 return -ENOMEM;
3995
3996 smp_wmb(); /* See comment in __pte_alloc */
3997
3998 spin_lock(&mm->page_table_lock);
3999 if (pgd_present(*pgd)) /* Another has populated it */
4000 pud_free(mm, new);
4001 else
4002 pgd_populate(mm, pgd, new);
4003 spin_unlock(&mm->page_table_lock);
4004 return 0;
4005}
4006#endif /* __PAGETABLE_PUD_FOLDED */
4007
4008#ifndef __PAGETABLE_PMD_FOLDED
4009/*
4010 * Allocate page middle directory.
4011 * We've already handled the fast-path in-line.
4012 */
4013int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4014{
4015 pmd_t *new = pmd_alloc_one(mm, address);
4016 if (!new)
4017 return -ENOMEM;
4018
4019 smp_wmb(); /* See comment in __pte_alloc */
4020
4021 spin_lock(&mm->page_table_lock);
4022#ifndef __ARCH_HAS_4LEVEL_HACK
4023 if (pud_present(*pud)) /* Another has populated it */
4024 pmd_free(mm, new);
4025 else
4026 pud_populate(mm, pud, new);
4027#else
4028 if (pgd_present(*pud)) /* Another has populated it */
4029 pmd_free(mm, new);
4030 else
4031 pgd_populate(mm, pud, new);
4032#endif /* __ARCH_HAS_4LEVEL_HACK */
4033 spin_unlock(&mm->page_table_lock);
4034 return 0;
4035}
4036#endif /* __PAGETABLE_PMD_FOLDED */
4037
4038#if !defined(__HAVE_ARCH_GATE_AREA)
4039
4040#if defined(AT_SYSINFO_EHDR)
4041static struct vm_area_struct gate_vma;
4042
4043static int __init gate_vma_init(void)
4044{
4045 gate_vma.vm_mm = NULL;
4046 gate_vma.vm_start = FIXADDR_USER_START;
4047 gate_vma.vm_end = FIXADDR_USER_END;
4048 gate_vma.vm_flags = VM_READ | VM_MAYREAD | VM_EXEC | VM_MAYEXEC;
4049 gate_vma.vm_page_prot = __P101;
4050
4051 return 0;
4052}
4053__initcall(gate_vma_init);
4054#endif
4055
4056struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
4057{
4058#ifdef AT_SYSINFO_EHDR
4059 return &gate_vma;
4060#else
4061 return NULL;
4062#endif
4063}
4064
4065int in_gate_area_no_mm(unsigned long addr)
4066{
4067#ifdef AT_SYSINFO_EHDR
4068 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
4069 return 1;
4070#endif
4071 return 0;
4072}
4073
4074#endif /* __HAVE_ARCH_GATE_AREA */
4075
4076static int __follow_pte(struct mm_struct *mm, unsigned long address,
4077 pte_t **ptepp, spinlock_t **ptlp)
4078{
4079 pgd_t *pgd;
4080 pud_t *pud;
4081 pmd_t *pmd;
4082 pte_t *ptep;
4083
4084 pgd = pgd_offset(mm, address);
4085 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4086 goto out;
4087
4088 pud = pud_offset(pgd, address);
4089 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4090 goto out;
4091
4092 pmd = pmd_offset(pud, address);
4093 VM_BUG_ON(pmd_trans_huge(*pmd));
4094 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4095 goto out;
4096
4097 /* We cannot handle huge page PFN maps. Luckily they don't exist. */
4098 if (pmd_huge(*pmd))
4099 goto out;
4100
4101 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4102 if (!ptep)
4103 goto out;
4104 if (!pte_present(*ptep))
4105 goto unlock;
4106 *ptepp = ptep;
4107 return 0;
4108unlock:
4109 pte_unmap_unlock(ptep, *ptlp);
4110out:
4111 return -EINVAL;
4112}
4113
4114static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4115 pte_t **ptepp, spinlock_t **ptlp)
4116{
4117 int res;
4118
4119 /* (void) is needed to make gcc happy */
4120 (void) __cond_lock(*ptlp,
4121 !(res = __follow_pte(mm, address, ptepp, ptlp)));
4122 return res;
4123}
4124
4125/**
4126 * follow_pfn - look up PFN at a user virtual address
4127 * @vma: memory mapping
4128 * @address: user virtual address
4129 * @pfn: location to store found PFN
4130 *
4131 * Only IO mappings and raw PFN mappings are allowed.
4132 *
4133 * Returns zero and the pfn at @pfn on success, -ve otherwise.
4134 */
4135int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4136 unsigned long *pfn)
4137{
4138 int ret = -EINVAL;
4139 spinlock_t *ptl;
4140 pte_t *ptep;
4141
4142 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4143 return ret;
4144
4145 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4146 if (ret)
4147 return ret;
4148 *pfn = pte_pfn(*ptep);
4149 pte_unmap_unlock(ptep, ptl);
4150 return 0;
4151}
4152EXPORT_SYMBOL(follow_pfn);
4153
4154#ifdef CONFIG_HAVE_IOREMAP_PROT
4155int follow_phys(struct vm_area_struct *vma,
4156 unsigned long address, unsigned int flags,
4157 unsigned long *prot, resource_size_t *phys)
4158{
4159 int ret = -EINVAL;
4160 pte_t *ptep, pte;
4161 spinlock_t *ptl;
4162
4163 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4164 goto out;
4165
4166 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4167 goto out;
4168 pte = *ptep;
4169
4170 if ((flags & FOLL_WRITE) && !pte_write(pte))
4171 goto unlock;
4172
4173 *prot = pgprot_val(pte_pgprot(pte));
4174 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4175
4176 ret = 0;
4177unlock:
4178 pte_unmap_unlock(ptep, ptl);
4179out:
4180 return ret;
4181}
4182
4183int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4184 void *buf, int len, int write)
4185{
4186 resource_size_t phys_addr;
4187 unsigned long prot = 0;
4188 void __iomem *maddr;
4189 int offset = addr & (PAGE_SIZE-1);
4190
4191 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4192 return -EINVAL;
4193
4194 maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot);
4195 if (write)
4196 memcpy_toio(maddr + offset, buf, len);
4197 else
4198 memcpy_fromio(buf, maddr + offset, len);
4199 iounmap(maddr);
4200
4201 return len;
4202}
4203EXPORT_SYMBOL_GPL(generic_access_phys);
4204#endif
4205
4206/*
4207 * Access another process' address space as given in mm. If non-NULL, use the
4208 * given task for page fault accounting.
4209 */
4210static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4211 unsigned long addr, void *buf, int len, int write)
4212{
4213 struct vm_area_struct *vma;
4214 void *old_buf = buf;
4215
4216 down_read(&mm->mmap_sem);
4217 /* ignore errors, just check how much was successfully transferred */
4218 while (len) {
4219 int bytes, ret, offset;
4220 void *maddr;
4221 struct page *page = NULL;
4222
4223 ret = get_user_pages(tsk, mm, addr, 1,
4224 write, 1, &page, &vma);
4225 if (ret <= 0) {
4226 /*
4227 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4228 * we can access using slightly different code.
4229 */
4230#ifdef CONFIG_HAVE_IOREMAP_PROT
4231 vma = find_vma(mm, addr);
4232 if (!vma || vma->vm_start > addr)
4233 break;
4234 if (vma->vm_ops && vma->vm_ops->access)
4235 ret = vma->vm_ops->access(vma, addr, buf,
4236 len, write);
4237 if (ret <= 0)
4238#endif
4239 break;
4240 bytes = ret;
4241 } else {
4242 bytes = len;
4243 offset = addr & (PAGE_SIZE-1);
4244 if (bytes > PAGE_SIZE-offset)
4245 bytes = PAGE_SIZE-offset;
4246
4247 maddr = kmap(page);
4248 if (write) {
4249 copy_to_user_page(vma, page, addr,
4250 maddr + offset, buf, bytes);
4251 set_page_dirty_lock(page);
4252 } else {
4253 copy_from_user_page(vma, page, addr,
4254 buf, maddr + offset, bytes);
4255 }
4256 kunmap(page);
4257 page_cache_release(page);
4258 }
4259 len -= bytes;
4260 buf += bytes;
4261 addr += bytes;
4262 }
4263 up_read(&mm->mmap_sem);
4264
4265 return buf - old_buf;
4266}
4267
4268/**
4269 * access_remote_vm - access another process' address space
4270 * @mm: the mm_struct of the target address space
4271 * @addr: start address to access
4272 * @buf: source or destination buffer
4273 * @len: number of bytes to transfer
4274 * @write: whether the access is a write
4275 *
4276 * The caller must hold a reference on @mm.
4277 */
4278int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4279 void *buf, int len, int write)
4280{
4281 return __access_remote_vm(NULL, mm, addr, buf, len, write);
4282}
4283
4284/*
4285 * Access another process' address space.
4286 * Source/target buffer must be kernel space,
4287 * Do not walk the page table directly, use get_user_pages
4288 */
4289int access_process_vm(struct task_struct *tsk, unsigned long addr,
4290 void *buf, int len, int write)
4291{
4292 struct mm_struct *mm;
4293 int ret;
4294
4295 mm = get_task_mm(tsk);
4296 if (!mm)
4297 return 0;
4298
4299 ret = __access_remote_vm(tsk, mm, addr, buf, len, write);
4300 mmput(mm);
4301
4302 return ret;
4303}
4304
4305/*
4306 * Print the name of a VMA.
4307 */
4308void print_vma_addr(char *prefix, unsigned long ip)
4309{
4310 struct mm_struct *mm = current->mm;
4311 struct vm_area_struct *vma;
4312
4313 /*
4314 * Do not print if we are in atomic
4315 * contexts (in exception stacks, etc.):
4316 */
4317 if (preempt_count())
4318 return;
4319
4320 down_read(&mm->mmap_sem);
4321 vma = find_vma(mm, ip);
4322 if (vma && vma->vm_file) {
4323 struct file *f = vma->vm_file;
4324 char *buf = (char *)__get_free_page(GFP_KERNEL);
4325 if (buf) {
4326 char *p;
4327
4328 p = d_path(&f->f_path, buf, PAGE_SIZE);
4329 if (IS_ERR(p))
4330 p = "?";
4331 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4332 vma->vm_start,
4333 vma->vm_end - vma->vm_start);
4334 free_page((unsigned long)buf);
4335 }
4336 }
4337 up_read(&mm->mmap_sem);
4338}
4339
4340#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4341void might_fault(void)
4342{
4343 /*
4344 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4345 * holding the mmap_sem, this is safe because kernel memory doesn't
4346 * get paged out, therefore we'll never actually fault, and the
4347 * below annotations will generate false positives.
4348 */
4349 if (segment_eq(get_fs(), KERNEL_DS))
4350 return;
4351
4352 /*
4353 * it would be nicer only to annotate paths which are not under
4354 * pagefault_disable, however that requires a larger audit and
4355 * providing helpers like get_user_atomic.
4356 */
4357 if (in_atomic())
4358 return;
4359
4360 __might_sleep(__FILE__, __LINE__, 0);
4361
4362 if (current->mm)
4363 might_lock_read(¤t->mm->mmap_sem);
4364}
4365EXPORT_SYMBOL(might_fault);
4366#endif
4367
4368#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4369static void clear_gigantic_page(struct page *page,
4370 unsigned long addr,
4371 unsigned int pages_per_huge_page)
4372{
4373 int i;
4374 struct page *p = page;
4375
4376 might_sleep();
4377 for (i = 0; i < pages_per_huge_page;
4378 i++, p = mem_map_next(p, page, i)) {
4379 cond_resched();
4380 clear_user_highpage(p, addr + i * PAGE_SIZE);
4381 }
4382}
4383void clear_huge_page(struct page *page,
4384 unsigned long addr, unsigned int pages_per_huge_page)
4385{
4386 int i;
4387
4388 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4389 clear_gigantic_page(page, addr, pages_per_huge_page);
4390 return;
4391 }
4392
4393 might_sleep();
4394 for (i = 0; i < pages_per_huge_page; i++) {
4395 cond_resched();
4396 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4397 }
4398}
4399
4400static void copy_user_gigantic_page(struct page *dst, struct page *src,
4401 unsigned long addr,
4402 struct vm_area_struct *vma,
4403 unsigned int pages_per_huge_page)
4404{
4405 int i;
4406 struct page *dst_base = dst;
4407 struct page *src_base = src;
4408
4409 for (i = 0; i < pages_per_huge_page; ) {
4410 cond_resched();
4411 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4412
4413 i++;
4414 dst = mem_map_next(dst, dst_base, i);
4415 src = mem_map_next(src, src_base, i);
4416 }
4417}
4418
4419void copy_user_huge_page(struct page *dst, struct page *src,
4420 unsigned long addr, struct vm_area_struct *vma,
4421 unsigned int pages_per_huge_page)
4422{
4423 int i;
4424
4425 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4426 copy_user_gigantic_page(dst, src, addr, vma,
4427 pages_per_huge_page);
4428 return;
4429 }
4430
4431 might_sleep();
4432 for (i = 0; i < pages_per_huge_page; i++) {
4433 cond_resched();
4434 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4435 }
4436}
4437#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4438
4439#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4440
4441static struct kmem_cache *page_ptl_cachep;
4442
4443void __init ptlock_cache_init(void)
4444{
4445 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4446 SLAB_PANIC, NULL);
4447}
4448
4449bool ptlock_alloc(struct page *page)
4450{
4451 spinlock_t *ptl;
4452
4453 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4454 if (!ptl)
4455 return false;
4456 page->ptl = ptl;
4457 return true;
4458}
4459
4460void ptlock_free(struct page *page)
4461{
4462 kmem_cache_free(page_ptl_cachep, page->ptl);
4463}
4464#endif