projects/linux·blob v5.7 · refs

5020 lines · 135.5 KiB

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 ([email protected])
35 *
36 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
37 * ([email protected])
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 <trace/events/kmem.h>
76
77#include <asm/io.h>
78#include <asm/mmu_context.h>
79#include <asm/pgalloc.h>
80#include <linux/uaccess.h>
81#include <asm/tlb.h>
82#include <asm/tlbflush.h>
83#include <asm/pgtable.h>
84
85#include "internal.h"
86
87#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
88#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
89#endif
90
91#ifndef CONFIG_NEED_MULTIPLE_NODES
92/* use the per-pgdat data instead for discontigmem - mbligh */
93unsigned long max_mapnr;
94EXPORT_SYMBOL(max_mapnr);
95
96struct page *mem_map;
97EXPORT_SYMBOL(mem_map);
98#endif
99
100/*
101 * A number of key systems in x86 including ioremap() rely on the assumption
102 * that high_memory defines the upper bound on direct map memory, then end
103 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
104 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
105 * and ZONE_HIGHMEM.
106 */
107void *high_memory;
108EXPORT_SYMBOL(high_memory);
109
110/*
111 * Randomize the address space (stacks, mmaps, brk, etc.).
112 *
113 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
114 * as ancient (libc5 based) binaries can segfault. )
115 */
116int randomize_va_space __read_mostly =
117#ifdef CONFIG_COMPAT_BRK
118 1;
119#else
120 2;
121#endif
122
123#ifndef arch_faults_on_old_pte
124static inline bool arch_faults_on_old_pte(void)
125{
126 /*
127 * Those arches which don't have hw access flag feature need to
128 * implement their own helper. By default, "true" means pagefault
129 * will be hit on old pte.
130 */
131 return true;
132}
133#endif
134
135static int __init disable_randmaps(char *s)
136{
137 randomize_va_space = 0;
138 return 1;
139}
140__setup("norandmaps", disable_randmaps);
141
142unsigned long zero_pfn __read_mostly;
143EXPORT_SYMBOL(zero_pfn);
144
145unsigned long highest_memmap_pfn __read_mostly;
146
147/*
148 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
149 */
150static int __init init_zero_pfn(void)
151{
152 zero_pfn = page_to_pfn(ZERO_PAGE(0));
153 return 0;
154}
155core_initcall(init_zero_pfn);
156
157void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
158{
159 trace_rss_stat(mm, member, count);
160}
161
162#if defined(SPLIT_RSS_COUNTING)
163
164void sync_mm_rss(struct mm_struct *mm)
165{
166 int i;
167
168 for (i = 0; i < NR_MM_COUNTERS; i++) {
169 if (current->rss_stat.count[i]) {
170 add_mm_counter(mm, i, current->rss_stat.count[i]);
171 current->rss_stat.count[i] = 0;
172 }
173 }
174 current->rss_stat.events = 0;
175}
176
177static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
178{
179 struct task_struct *task = current;
180
181 if (likely(task->mm == mm))
182 task->rss_stat.count[member] += val;
183 else
184 add_mm_counter(mm, member, val);
185}
186#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
187#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
188
189/* sync counter once per 64 page faults */
190#define TASK_RSS_EVENTS_THRESH (64)
191static void check_sync_rss_stat(struct task_struct *task)
192{
193 if (unlikely(task != current))
194 return;
195 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
196 sync_mm_rss(task->mm);
197}
198#else /* SPLIT_RSS_COUNTING */
199
200#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
201#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
202
203static void check_sync_rss_stat(struct task_struct *task)
204{
205}
206
207#endif /* SPLIT_RSS_COUNTING */
208
209/*
210 * Note: this doesn't free the actual pages themselves. That
211 * has been handled earlier when unmapping all the memory regions.
212 */
213static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
214 unsigned long addr)
215{
216 pgtable_t token = pmd_pgtable(*pmd);
217 pmd_clear(pmd);
218 pte_free_tlb(tlb, token, addr);
219 mm_dec_nr_ptes(tlb->mm);
220}
221
222static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
223 unsigned long addr, unsigned long end,
224 unsigned long floor, unsigned long ceiling)
225{
226 pmd_t *pmd;
227 unsigned long next;
228 unsigned long start;
229
230 start = addr;
231 pmd = pmd_offset(pud, addr);
232 do {
233 next = pmd_addr_end(addr, end);
234 if (pmd_none_or_clear_bad(pmd))
235 continue;
236 free_pte_range(tlb, pmd, addr);
237 } while (pmd++, addr = next, addr != end);
238
239 start &= PUD_MASK;
240 if (start < floor)
241 return;
242 if (ceiling) {
243 ceiling &= PUD_MASK;
244 if (!ceiling)
245 return;
246 }
247 if (end - 1 > ceiling - 1)
248 return;
249
250 pmd = pmd_offset(pud, start);
251 pud_clear(pud);
252 pmd_free_tlb(tlb, pmd, start);
253 mm_dec_nr_pmds(tlb->mm);
254}
255
256static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
257 unsigned long addr, unsigned long end,
258 unsigned long floor, unsigned long ceiling)
259{
260 pud_t *pud;
261 unsigned long next;
262 unsigned long start;
263
264 start = addr;
265 pud = pud_offset(p4d, addr);
266 do {
267 next = pud_addr_end(addr, end);
268 if (pud_none_or_clear_bad(pud))
269 continue;
270 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
271 } while (pud++, addr = next, addr != end);
272
273 start &= P4D_MASK;
274 if (start < floor)
275 return;
276 if (ceiling) {
277 ceiling &= P4D_MASK;
278 if (!ceiling)
279 return;
280 }
281 if (end - 1 > ceiling - 1)
282 return;
283
284 pud = pud_offset(p4d, start);
285 p4d_clear(p4d);
286 pud_free_tlb(tlb, pud, start);
287 mm_dec_nr_puds(tlb->mm);
288}
289
290static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
291 unsigned long addr, unsigned long end,
292 unsigned long floor, unsigned long ceiling)
293{
294 p4d_t *p4d;
295 unsigned long next;
296 unsigned long start;
297
298 start = addr;
299 p4d = p4d_offset(pgd, addr);
300 do {
301 next = p4d_addr_end(addr, end);
302 if (p4d_none_or_clear_bad(p4d))
303 continue;
304 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
305 } while (p4d++, addr = next, addr != end);
306
307 start &= PGDIR_MASK;
308 if (start < floor)
309 return;
310 if (ceiling) {
311 ceiling &= PGDIR_MASK;
312 if (!ceiling)
313 return;
314 }
315 if (end - 1 > ceiling - 1)
316 return;
317
318 p4d = p4d_offset(pgd, start);
319 pgd_clear(pgd);
320 p4d_free_tlb(tlb, p4d, start);
321}
322
323/*
324 * This function frees user-level page tables of a process.
325 */
326void free_pgd_range(struct mmu_gather *tlb,
327 unsigned long addr, unsigned long end,
328 unsigned long floor, unsigned long ceiling)
329{
330 pgd_t *pgd;
331 unsigned long next;
332
333 /*
334 * The next few lines have given us lots of grief...
335 *
336 * Why are we testing PMD* at this top level? Because often
337 * there will be no work to do at all, and we'd prefer not to
338 * go all the way down to the bottom just to discover that.
339 *
340 * Why all these "- 1"s? Because 0 represents both the bottom
341 * of the address space and the top of it (using -1 for the
342 * top wouldn't help much: the masks would do the wrong thing).
343 * The rule is that addr 0 and floor 0 refer to the bottom of
344 * the address space, but end 0 and ceiling 0 refer to the top
345 * Comparisons need to use "end - 1" and "ceiling - 1" (though
346 * that end 0 case should be mythical).
347 *
348 * Wherever addr is brought up or ceiling brought down, we must
349 * be careful to reject "the opposite 0" before it confuses the
350 * subsequent tests. But what about where end is brought down
351 * by PMD_SIZE below? no, end can't go down to 0 there.
352 *
353 * Whereas we round start (addr) and ceiling down, by different
354 * masks at different levels, in order to test whether a table
355 * now has no other vmas using it, so can be freed, we don't
356 * bother to round floor or end up - the tests don't need that.
357 */
358
359 addr &= PMD_MASK;
360 if (addr < floor) {
361 addr += PMD_SIZE;
362 if (!addr)
363 return;
364 }
365 if (ceiling) {
366 ceiling &= PMD_MASK;
367 if (!ceiling)
368 return;
369 }
370 if (end - 1 > ceiling - 1)
371 end -= PMD_SIZE;
372 if (addr > end - 1)
373 return;
374 /*
375 * We add page table cache pages with PAGE_SIZE,
376 * (see pte_free_tlb()), flush the tlb if we need
377 */
378 tlb_change_page_size(tlb, PAGE_SIZE);
379 pgd = pgd_offset(tlb->mm, addr);
380 do {
381 next = pgd_addr_end(addr, end);
382 if (pgd_none_or_clear_bad(pgd))
383 continue;
384 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
385 } while (pgd++, addr = next, addr != end);
386}
387
388void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
389 unsigned long floor, unsigned long ceiling)
390{
391 while (vma) {
392 struct vm_area_struct *next = vma->vm_next;
393 unsigned long addr = vma->vm_start;
394
395 /*
396 * Hide vma from rmap and truncate_pagecache before freeing
397 * pgtables
398 */
399 unlink_anon_vmas(vma);
400 unlink_file_vma(vma);
401
402 if (is_vm_hugetlb_page(vma)) {
403 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
404 floor, next ? next->vm_start : ceiling);
405 } else {
406 /*
407 * Optimization: gather nearby vmas into one call down
408 */
409 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
410 && !is_vm_hugetlb_page(next)) {
411 vma = next;
412 next = vma->vm_next;
413 unlink_anon_vmas(vma);
414 unlink_file_vma(vma);
415 }
416 free_pgd_range(tlb, addr, vma->vm_end,
417 floor, next ? next->vm_start : ceiling);
418 }
419 vma = next;
420 }
421}
422
423int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
424{
425 spinlock_t *ptl;
426 pgtable_t new = pte_alloc_one(mm);
427 if (!new)
428 return -ENOMEM;
429
430 /*
431 * Ensure all pte setup (eg. pte page lock and page clearing) are
432 * visible before the pte is made visible to other CPUs by being
433 * put into page tables.
434 *
435 * The other side of the story is the pointer chasing in the page
436 * table walking code (when walking the page table without locking;
437 * ie. most of the time). Fortunately, these data accesses consist
438 * of a chain of data-dependent loads, meaning most CPUs (alpha
439 * being the notable exception) will already guarantee loads are
440 * seen in-order. See the alpha page table accessors for the
441 * smp_read_barrier_depends() barriers in page table walking code.
442 */
443 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
444
445 ptl = pmd_lock(mm, pmd);
446 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
447 mm_inc_nr_ptes(mm);
448 pmd_populate(mm, pmd, new);
449 new = NULL;
450 }
451 spin_unlock(ptl);
452 if (new)
453 pte_free(mm, new);
454 return 0;
455}
456
457int __pte_alloc_kernel(pmd_t *pmd)
458{
459 pte_t *new = pte_alloc_one_kernel(&init_mm);
460 if (!new)
461 return -ENOMEM;
462
463 smp_wmb(); /* See comment in __pte_alloc */
464
465 spin_lock(&init_mm.page_table_lock);
466 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
467 pmd_populate_kernel(&init_mm, pmd, new);
468 new = NULL;
469 }
470 spin_unlock(&init_mm.page_table_lock);
471 if (new)
472 pte_free_kernel(&init_mm, new);
473 return 0;
474}
475
476static inline void init_rss_vec(int *rss)
477{
478 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
479}
480
481static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
482{
483 int i;
484
485 if (current->mm == mm)
486 sync_mm_rss(mm);
487 for (i = 0; i < NR_MM_COUNTERS; i++)
488 if (rss[i])
489 add_mm_counter(mm, i, rss[i]);
490}
491
492/*
493 * This function is called to print an error when a bad pte
494 * is found. For example, we might have a PFN-mapped pte in
495 * a region that doesn't allow it.
496 *
497 * The calling function must still handle the error.
498 */
499static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
500 pte_t pte, struct page *page)
501{
502 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
503 p4d_t *p4d = p4d_offset(pgd, addr);
504 pud_t *pud = pud_offset(p4d, addr);
505 pmd_t *pmd = pmd_offset(pud, addr);
506 struct address_space *mapping;
507 pgoff_t index;
508 static unsigned long resume;
509 static unsigned long nr_shown;
510 static unsigned long nr_unshown;
511
512 /*
513 * Allow a burst of 60 reports, then keep quiet for that minute;
514 * or allow a steady drip of one report per second.
515 */
516 if (nr_shown == 60) {
517 if (time_before(jiffies, resume)) {
518 nr_unshown++;
519 return;
520 }
521 if (nr_unshown) {
522 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
523 nr_unshown);
524 nr_unshown = 0;
525 }
526 nr_shown = 0;
527 }
528 if (nr_shown++ == 0)
529 resume = jiffies + 60 * HZ;
530
531 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
532 index = linear_page_index(vma, addr);
533
534 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
535 current->comm,
536 (long long)pte_val(pte), (long long)pmd_val(*pmd));
537 if (page)
538 dump_page(page, "bad pte");
539 pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
540 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
541 pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
542 vma->vm_file,
543 vma->vm_ops ? vma->vm_ops->fault : NULL,
544 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
545 mapping ? mapping->a_ops->readpage : NULL);
546 dump_stack();
547 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
548}
549
550/*
551 * vm_normal_page -- This function gets the "struct page" associated with a pte.
552 *
553 * "Special" mappings do not wish to be associated with a "struct page" (either
554 * it doesn't exist, or it exists but they don't want to touch it). In this
555 * case, NULL is returned here. "Normal" mappings do have a struct page.
556 *
557 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
558 * pte bit, in which case this function is trivial. Secondly, an architecture
559 * may not have a spare pte bit, which requires a more complicated scheme,
560 * described below.
561 *
562 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
563 * special mapping (even if there are underlying and valid "struct pages").
564 * COWed pages of a VM_PFNMAP are always normal.
565 *
566 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
567 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
568 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
569 * mapping will always honor the rule
570 *
571 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
572 *
573 * And for normal mappings this is false.
574 *
575 * This restricts such mappings to be a linear translation from virtual address
576 * to pfn. To get around this restriction, we allow arbitrary mappings so long
577 * as the vma is not a COW mapping; in that case, we know that all ptes are
578 * special (because none can have been COWed).
579 *
580 *
581 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
582 *
583 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
584 * page" backing, however the difference is that _all_ pages with a struct
585 * page (that is, those where pfn_valid is true) are refcounted and considered
586 * normal pages by the VM. The disadvantage is that pages are refcounted
587 * (which can be slower and simply not an option for some PFNMAP users). The
588 * advantage is that we don't have to follow the strict linearity rule of
589 * PFNMAP mappings in order to support COWable mappings.
590 *
591 */
592struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
593 pte_t pte)
594{
595 unsigned long pfn = pte_pfn(pte);
596
597 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
598 if (likely(!pte_special(pte)))
599 goto check_pfn;
600 if (vma->vm_ops && vma->vm_ops->find_special_page)
601 return vma->vm_ops->find_special_page(vma, addr);
602 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
603 return NULL;
604 if (is_zero_pfn(pfn))
605 return NULL;
606 if (pte_devmap(pte))
607 return NULL;
608
609 print_bad_pte(vma, addr, pte, NULL);
610 return NULL;
611 }
612
613 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
614
615 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
616 if (vma->vm_flags & VM_MIXEDMAP) {
617 if (!pfn_valid(pfn))
618 return NULL;
619 goto out;
620 } else {
621 unsigned long off;
622 off = (addr - vma->vm_start) >> PAGE_SHIFT;
623 if (pfn == vma->vm_pgoff + off)
624 return NULL;
625 if (!is_cow_mapping(vma->vm_flags))
626 return NULL;
627 }
628 }
629
630 if (is_zero_pfn(pfn))
631 return NULL;
632
633check_pfn:
634 if (unlikely(pfn > highest_memmap_pfn)) {
635 print_bad_pte(vma, addr, pte, NULL);
636 return NULL;
637 }
638
639 /*
640 * NOTE! We still have PageReserved() pages in the page tables.
641 * eg. VDSO mappings can cause them to exist.
642 */
643out:
644 return pfn_to_page(pfn);
645}
646
647#ifdef CONFIG_TRANSPARENT_HUGEPAGE
648struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
649 pmd_t pmd)
650{
651 unsigned long pfn = pmd_pfn(pmd);
652
653 /*
654 * There is no pmd_special() but there may be special pmds, e.g.
655 * in a direct-access (dax) mapping, so let's just replicate the
656 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
657 */
658 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
659 if (vma->vm_flags & VM_MIXEDMAP) {
660 if (!pfn_valid(pfn))
661 return NULL;
662 goto out;
663 } else {
664 unsigned long off;
665 off = (addr - vma->vm_start) >> PAGE_SHIFT;
666 if (pfn == vma->vm_pgoff + off)
667 return NULL;
668 if (!is_cow_mapping(vma->vm_flags))
669 return NULL;
670 }
671 }
672
673 if (pmd_devmap(pmd))
674 return NULL;
675 if (is_huge_zero_pmd(pmd))
676 return NULL;
677 if (unlikely(pfn > highest_memmap_pfn))
678 return NULL;
679
680 /*
681 * NOTE! We still have PageReserved() pages in the page tables.
682 * eg. VDSO mappings can cause them to exist.
683 */
684out:
685 return pfn_to_page(pfn);
686}
687#endif
688
689/*
690 * copy one vm_area from one task to the other. Assumes the page tables
691 * already present in the new task to be cleared in the whole range
692 * covered by this vma.
693 */
694
695static inline unsigned long
696copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
697 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
698 unsigned long addr, int *rss)
699{
700 unsigned long vm_flags = vma->vm_flags;
701 pte_t pte = *src_pte;
702 struct page *page;
703
704 /* pte contains position in swap or file, so copy. */
705 if (unlikely(!pte_present(pte))) {
706 swp_entry_t entry = pte_to_swp_entry(pte);
707
708 if (likely(!non_swap_entry(entry))) {
709 if (swap_duplicate(entry) < 0)
710 return entry.val;
711
712 /* make sure dst_mm is on swapoff's mmlist. */
713 if (unlikely(list_empty(&dst_mm->mmlist))) {
714 spin_lock(&mmlist_lock);
715 if (list_empty(&dst_mm->mmlist))
716 list_add(&dst_mm->mmlist,
717 &src_mm->mmlist);
718 spin_unlock(&mmlist_lock);
719 }
720 rss[MM_SWAPENTS]++;
721 } else if (is_migration_entry(entry)) {
722 page = migration_entry_to_page(entry);
723
724 rss[mm_counter(page)]++;
725
726 if (is_write_migration_entry(entry) &&
727 is_cow_mapping(vm_flags)) {
728 /*
729 * COW mappings require pages in both
730 * parent and child to be set to read.
731 */
732 make_migration_entry_read(&entry);
733 pte = swp_entry_to_pte(entry);
734 if (pte_swp_soft_dirty(*src_pte))
735 pte = pte_swp_mksoft_dirty(pte);
736 if (pte_swp_uffd_wp(*src_pte))
737 pte = pte_swp_mkuffd_wp(pte);
738 set_pte_at(src_mm, addr, src_pte, pte);
739 }
740 } else if (is_device_private_entry(entry)) {
741 page = device_private_entry_to_page(entry);
742
743 /*
744 * Update rss count even for unaddressable pages, as
745 * they should treated just like normal pages in this
746 * respect.
747 *
748 * We will likely want to have some new rss counters
749 * for unaddressable pages, at some point. But for now
750 * keep things as they are.
751 */
752 get_page(page);
753 rss[mm_counter(page)]++;
754 page_dup_rmap(page, false);
755
756 /*
757 * We do not preserve soft-dirty information, because so
758 * far, checkpoint/restore is the only feature that
759 * requires that. And checkpoint/restore does not work
760 * when a device driver is involved (you cannot easily
761 * save and restore device driver state).
762 */
763 if (is_write_device_private_entry(entry) &&
764 is_cow_mapping(vm_flags)) {
765 make_device_private_entry_read(&entry);
766 pte = swp_entry_to_pte(entry);
767 if (pte_swp_uffd_wp(*src_pte))
768 pte = pte_swp_mkuffd_wp(pte);
769 set_pte_at(src_mm, addr, src_pte, pte);
770 }
771 }
772 goto out_set_pte;
773 }
774
775 /*
776 * If it's a COW mapping, write protect it both
777 * in the parent and the child
778 */
779 if (is_cow_mapping(vm_flags) && pte_write(pte)) {
780 ptep_set_wrprotect(src_mm, addr, src_pte);
781 pte = pte_wrprotect(pte);
782 }
783
784 /*
785 * If it's a shared mapping, mark it clean in
786 * the child
787 */
788 if (vm_flags & VM_SHARED)
789 pte = pte_mkclean(pte);
790 pte = pte_mkold(pte);
791
792 /*
793 * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
794 * does not have the VM_UFFD_WP, which means that the uffd
795 * fork event is not enabled.
796 */
797 if (!(vm_flags & VM_UFFD_WP))
798 pte = pte_clear_uffd_wp(pte);
799
800 page = vm_normal_page(vma, addr, pte);
801 if (page) {
802 get_page(page);
803 page_dup_rmap(page, false);
804 rss[mm_counter(page)]++;
805 } else if (pte_devmap(pte)) {
806 page = pte_page(pte);
807 }
808
809out_set_pte:
810 set_pte_at(dst_mm, addr, dst_pte, pte);
811 return 0;
812}
813
814static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
815 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
816 unsigned long addr, unsigned long end)
817{
818 pte_t *orig_src_pte, *orig_dst_pte;
819 pte_t *src_pte, *dst_pte;
820 spinlock_t *src_ptl, *dst_ptl;
821 int progress = 0;
822 int rss[NR_MM_COUNTERS];
823 swp_entry_t entry = (swp_entry_t){0};
824
825again:
826 init_rss_vec(rss);
827
828 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
829 if (!dst_pte)
830 return -ENOMEM;
831 src_pte = pte_offset_map(src_pmd, addr);
832 src_ptl = pte_lockptr(src_mm, src_pmd);
833 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
834 orig_src_pte = src_pte;
835 orig_dst_pte = dst_pte;
836 arch_enter_lazy_mmu_mode();
837
838 do {
839 /*
840 * We are holding two locks at this point - either of them
841 * could generate latencies in another task on another CPU.
842 */
843 if (progress >= 32) {
844 progress = 0;
845 if (need_resched() ||
846 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
847 break;
848 }
849 if (pte_none(*src_pte)) {
850 progress++;
851 continue;
852 }
853 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
854 vma, addr, rss);
855 if (entry.val)
856 break;
857 progress += 8;
858 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
859
860 arch_leave_lazy_mmu_mode();
861 spin_unlock(src_ptl);
862 pte_unmap(orig_src_pte);
863 add_mm_rss_vec(dst_mm, rss);
864 pte_unmap_unlock(orig_dst_pte, dst_ptl);
865 cond_resched();
866
867 if (entry.val) {
868 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
869 return -ENOMEM;
870 progress = 0;
871 }
872 if (addr != end)
873 goto again;
874 return 0;
875}
876
877static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
878 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
879 unsigned long addr, unsigned long end)
880{
881 pmd_t *src_pmd, *dst_pmd;
882 unsigned long next;
883
884 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
885 if (!dst_pmd)
886 return -ENOMEM;
887 src_pmd = pmd_offset(src_pud, addr);
888 do {
889 next = pmd_addr_end(addr, end);
890 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
891 || pmd_devmap(*src_pmd)) {
892 int err;
893 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
894 err = copy_huge_pmd(dst_mm, src_mm,
895 dst_pmd, src_pmd, addr, vma);
896 if (err == -ENOMEM)
897 return -ENOMEM;
898 if (!err)
899 continue;
900 /* fall through */
901 }
902 if (pmd_none_or_clear_bad(src_pmd))
903 continue;
904 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
905 vma, addr, next))
906 return -ENOMEM;
907 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
908 return 0;
909}
910
911static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
912 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
913 unsigned long addr, unsigned long end)
914{
915 pud_t *src_pud, *dst_pud;
916 unsigned long next;
917
918 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
919 if (!dst_pud)
920 return -ENOMEM;
921 src_pud = pud_offset(src_p4d, addr);
922 do {
923 next = pud_addr_end(addr, end);
924 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
925 int err;
926
927 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
928 err = copy_huge_pud(dst_mm, src_mm,
929 dst_pud, src_pud, addr, vma);
930 if (err == -ENOMEM)
931 return -ENOMEM;
932 if (!err)
933 continue;
934 /* fall through */
935 }
936 if (pud_none_or_clear_bad(src_pud))
937 continue;
938 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
939 vma, addr, next))
940 return -ENOMEM;
941 } while (dst_pud++, src_pud++, addr = next, addr != end);
942 return 0;
943}
944
945static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
946 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
947 unsigned long addr, unsigned long end)
948{
949 p4d_t *src_p4d, *dst_p4d;
950 unsigned long next;
951
952 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
953 if (!dst_p4d)
954 return -ENOMEM;
955 src_p4d = p4d_offset(src_pgd, addr);
956 do {
957 next = p4d_addr_end(addr, end);
958 if (p4d_none_or_clear_bad(src_p4d))
959 continue;
960 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
961 vma, addr, next))
962 return -ENOMEM;
963 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
964 return 0;
965}
966
967int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
968 struct vm_area_struct *vma)
969{
970 pgd_t *src_pgd, *dst_pgd;
971 unsigned long next;
972 unsigned long addr = vma->vm_start;
973 unsigned long end = vma->vm_end;
974 struct mmu_notifier_range range;
975 bool is_cow;
976 int ret;
977
978 /*
979 * Don't copy ptes where a page fault will fill them correctly.
980 * Fork becomes much lighter when there are big shared or private
981 * readonly mappings. The tradeoff is that copy_page_range is more
982 * efficient than faulting.
983 */
984 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
985 !vma->anon_vma)
986 return 0;
987
988 if (is_vm_hugetlb_page(vma))
989 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
990
991 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
992 /*
993 * We do not free on error cases below as remove_vma
994 * gets called on error from higher level routine
995 */
996 ret = track_pfn_copy(vma);
997 if (ret)
998 return ret;
999 }
1000
1001 /*
1002 * We need to invalidate the secondary MMU mappings only when
1003 * there could be a permission downgrade on the ptes of the
1004 * parent mm. And a permission downgrade will only happen if
1005 * is_cow_mapping() returns true.
1006 */
1007 is_cow = is_cow_mapping(vma->vm_flags);
1008
1009 if (is_cow) {
1010 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1011 0, vma, src_mm, addr, end);
1012 mmu_notifier_invalidate_range_start(&range);
1013 }
1014
1015 ret = 0;
1016 dst_pgd = pgd_offset(dst_mm, addr);
1017 src_pgd = pgd_offset(src_mm, addr);
1018 do {
1019 next = pgd_addr_end(addr, end);
1020 if (pgd_none_or_clear_bad(src_pgd))
1021 continue;
1022 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1023 vma, addr, next))) {
1024 ret = -ENOMEM;
1025 break;
1026 }
1027 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1028
1029 if (is_cow)
1030 mmu_notifier_invalidate_range_end(&range);
1031 return ret;
1032}
1033
1034static unsigned long zap_pte_range(struct mmu_gather *tlb,
1035 struct vm_area_struct *vma, pmd_t *pmd,
1036 unsigned long addr, unsigned long end,
1037 struct zap_details *details)
1038{
1039 struct mm_struct *mm = tlb->mm;
1040 int force_flush = 0;
1041 int rss[NR_MM_COUNTERS];
1042 spinlock_t *ptl;
1043 pte_t *start_pte;
1044 pte_t *pte;
1045 swp_entry_t entry;
1046
1047 tlb_change_page_size(tlb, PAGE_SIZE);
1048again:
1049 init_rss_vec(rss);
1050 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1051 pte = start_pte;
1052 flush_tlb_batched_pending(mm);
1053 arch_enter_lazy_mmu_mode();
1054 do {
1055 pte_t ptent = *pte;
1056 if (pte_none(ptent))
1057 continue;
1058
1059 if (need_resched())
1060 break;
1061
1062 if (pte_present(ptent)) {
1063 struct page *page;
1064
1065 page = vm_normal_page(vma, addr, ptent);
1066 if (unlikely(details) && page) {
1067 /*
1068 * unmap_shared_mapping_pages() wants to
1069 * invalidate cache without truncating:
1070 * unmap shared but keep private pages.
1071 */
1072 if (details->check_mapping &&
1073 details->check_mapping != page_rmapping(page))
1074 continue;
1075 }
1076 ptent = ptep_get_and_clear_full(mm, addr, pte,
1077 tlb->fullmm);
1078 tlb_remove_tlb_entry(tlb, pte, addr);
1079 if (unlikely(!page))
1080 continue;
1081
1082 if (!PageAnon(page)) {
1083 if (pte_dirty(ptent)) {
1084 force_flush = 1;
1085 set_page_dirty(page);
1086 }
1087 if (pte_young(ptent) &&
1088 likely(!(vma->vm_flags & VM_SEQ_READ)))
1089 mark_page_accessed(page);
1090 }
1091 rss[mm_counter(page)]--;
1092 page_remove_rmap(page, false);
1093 if (unlikely(page_mapcount(page) < 0))
1094 print_bad_pte(vma, addr, ptent, page);
1095 if (unlikely(__tlb_remove_page(tlb, page))) {
1096 force_flush = 1;
1097 addr += PAGE_SIZE;
1098 break;
1099 }
1100 continue;
1101 }
1102
1103 entry = pte_to_swp_entry(ptent);
1104 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1105 struct page *page = device_private_entry_to_page(entry);
1106
1107 if (unlikely(details && details->check_mapping)) {
1108 /*
1109 * unmap_shared_mapping_pages() wants to
1110 * invalidate cache without truncating:
1111 * unmap shared but keep private pages.
1112 */
1113 if (details->check_mapping !=
1114 page_rmapping(page))
1115 continue;
1116 }
1117
1118 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1119 rss[mm_counter(page)]--;
1120 page_remove_rmap(page, false);
1121 put_page(page);
1122 continue;
1123 }
1124
1125 /* If details->check_mapping, we leave swap entries. */
1126 if (unlikely(details))
1127 continue;
1128
1129 if (!non_swap_entry(entry))
1130 rss[MM_SWAPENTS]--;
1131 else if (is_migration_entry(entry)) {
1132 struct page *page;
1133
1134 page = migration_entry_to_page(entry);
1135 rss[mm_counter(page)]--;
1136 }
1137 if (unlikely(!free_swap_and_cache(entry)))
1138 print_bad_pte(vma, addr, ptent, NULL);
1139 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1140 } while (pte++, addr += PAGE_SIZE, addr != end);
1141
1142 add_mm_rss_vec(mm, rss);
1143 arch_leave_lazy_mmu_mode();
1144
1145 /* Do the actual TLB flush before dropping ptl */
1146 if (force_flush)
1147 tlb_flush_mmu_tlbonly(tlb);
1148 pte_unmap_unlock(start_pte, ptl);
1149
1150 /*
1151 * If we forced a TLB flush (either due to running out of
1152 * batch buffers or because we needed to flush dirty TLB
1153 * entries before releasing the ptl), free the batched
1154 * memory too. Restart if we didn't do everything.
1155 */
1156 if (force_flush) {
1157 force_flush = 0;
1158 tlb_flush_mmu(tlb);
1159 }
1160
1161 if (addr != end) {
1162 cond_resched();
1163 goto again;
1164 }
1165
1166 return addr;
1167}
1168
1169static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1170 struct vm_area_struct *vma, pud_t *pud,
1171 unsigned long addr, unsigned long end,
1172 struct zap_details *details)
1173{
1174 pmd_t *pmd;
1175 unsigned long next;
1176
1177 pmd = pmd_offset(pud, addr);
1178 do {
1179 next = pmd_addr_end(addr, end);
1180 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1181 if (next - addr != HPAGE_PMD_SIZE)
1182 __split_huge_pmd(vma, pmd, addr, false, NULL);
1183 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1184 goto next;
1185 /* fall through */
1186 }
1187 /*
1188 * Here there can be other concurrent MADV_DONTNEED or
1189 * trans huge page faults running, and if the pmd is
1190 * none or trans huge it can change under us. This is
1191 * because MADV_DONTNEED holds the mmap_sem in read
1192 * mode.
1193 */
1194 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1195 goto next;
1196 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1197next:
1198 cond_resched();
1199 } while (pmd++, addr = next, addr != end);
1200
1201 return addr;
1202}
1203
1204static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1205 struct vm_area_struct *vma, p4d_t *p4d,
1206 unsigned long addr, unsigned long end,
1207 struct zap_details *details)
1208{
1209 pud_t *pud;
1210 unsigned long next;
1211
1212 pud = pud_offset(p4d, addr);
1213 do {
1214 next = pud_addr_end(addr, end);
1215 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1216 if (next - addr != HPAGE_PUD_SIZE) {
1217 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1218 split_huge_pud(vma, pud, addr);
1219 } else if (zap_huge_pud(tlb, vma, pud, addr))
1220 goto next;
1221 /* fall through */
1222 }
1223 if (pud_none_or_clear_bad(pud))
1224 continue;
1225 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1226next:
1227 cond_resched();
1228 } while (pud++, addr = next, addr != end);
1229
1230 return addr;
1231}
1232
1233static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1234 struct vm_area_struct *vma, pgd_t *pgd,
1235 unsigned long addr, unsigned long end,
1236 struct zap_details *details)
1237{
1238 p4d_t *p4d;
1239 unsigned long next;
1240
1241 p4d = p4d_offset(pgd, addr);
1242 do {
1243 next = p4d_addr_end(addr, end);
1244 if (p4d_none_or_clear_bad(p4d))
1245 continue;
1246 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1247 } while (p4d++, addr = next, addr != end);
1248
1249 return addr;
1250}
1251
1252void unmap_page_range(struct mmu_gather *tlb,
1253 struct vm_area_struct *vma,
1254 unsigned long addr, unsigned long end,
1255 struct zap_details *details)
1256{
1257 pgd_t *pgd;
1258 unsigned long next;
1259
1260 BUG_ON(addr >= end);
1261 tlb_start_vma(tlb, vma);
1262 pgd = pgd_offset(vma->vm_mm, addr);
1263 do {
1264 next = pgd_addr_end(addr, end);
1265 if (pgd_none_or_clear_bad(pgd))
1266 continue;
1267 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1268 } while (pgd++, addr = next, addr != end);
1269 tlb_end_vma(tlb, vma);
1270}
1271
1272
1273static void unmap_single_vma(struct mmu_gather *tlb,
1274 struct vm_area_struct *vma, unsigned long start_addr,
1275 unsigned long end_addr,
1276 struct zap_details *details)
1277{
1278 unsigned long start = max(vma->vm_start, start_addr);
1279 unsigned long end;
1280
1281 if (start >= vma->vm_end)
1282 return;
1283 end = min(vma->vm_end, end_addr);
1284 if (end <= vma->vm_start)
1285 return;
1286
1287 if (vma->vm_file)
1288 uprobe_munmap(vma, start, end);
1289
1290 if (unlikely(vma->vm_flags & VM_PFNMAP))
1291 untrack_pfn(vma, 0, 0);
1292
1293 if (start != end) {
1294 if (unlikely(is_vm_hugetlb_page(vma))) {
1295 /*
1296 * It is undesirable to test vma->vm_file as it
1297 * should be non-null for valid hugetlb area.
1298 * However, vm_file will be NULL in the error
1299 * cleanup path of mmap_region. When
1300 * hugetlbfs ->mmap method fails,
1301 * mmap_region() nullifies vma->vm_file
1302 * before calling this function to clean up.
1303 * Since no pte has actually been setup, it is
1304 * safe to do nothing in this case.
1305 */
1306 if (vma->vm_file) {
1307 i_mmap_lock_write(vma->vm_file->f_mapping);
1308 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1309 i_mmap_unlock_write(vma->vm_file->f_mapping);
1310 }
1311 } else
1312 unmap_page_range(tlb, vma, start, end, details);
1313 }
1314}
1315
1316/**
1317 * unmap_vmas - unmap a range of memory covered by a list of vma's
1318 * @tlb: address of the caller's struct mmu_gather
1319 * @vma: the starting vma
1320 * @start_addr: virtual address at which to start unmapping
1321 * @end_addr: virtual address at which to end unmapping
1322 *
1323 * Unmap all pages in the vma list.
1324 *
1325 * Only addresses between `start' and `end' will be unmapped.
1326 *
1327 * The VMA list must be sorted in ascending virtual address order.
1328 *
1329 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1330 * range after unmap_vmas() returns. So the only responsibility here is to
1331 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1332 * drops the lock and schedules.
1333 */
1334void unmap_vmas(struct mmu_gather *tlb,
1335 struct vm_area_struct *vma, unsigned long start_addr,
1336 unsigned long end_addr)
1337{
1338 struct mmu_notifier_range range;
1339
1340 mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1341 start_addr, end_addr);
1342 mmu_notifier_invalidate_range_start(&range);
1343 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1344 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1345 mmu_notifier_invalidate_range_end(&range);
1346}
1347
1348/**
1349 * zap_page_range - remove user pages in a given range
1350 * @vma: vm_area_struct holding the applicable pages
1351 * @start: starting address of pages to zap
1352 * @size: number of bytes to zap
1353 *
1354 * Caller must protect the VMA list
1355 */
1356void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1357 unsigned long size)
1358{
1359 struct mmu_notifier_range range;
1360 struct mmu_gather tlb;
1361
1362 lru_add_drain();
1363 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1364 start, start + size);
1365 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1366 update_hiwater_rss(vma->vm_mm);
1367 mmu_notifier_invalidate_range_start(&range);
1368 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1369 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1370 mmu_notifier_invalidate_range_end(&range);
1371 tlb_finish_mmu(&tlb, start, range.end);
1372}
1373
1374/**
1375 * zap_page_range_single - remove user pages in a given range
1376 * @vma: vm_area_struct holding the applicable pages
1377 * @address: starting address of pages to zap
1378 * @size: number of bytes to zap
1379 * @details: details of shared cache invalidation
1380 *
1381 * The range must fit into one VMA.
1382 */
1383static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1384 unsigned long size, struct zap_details *details)
1385{
1386 struct mmu_notifier_range range;
1387 struct mmu_gather tlb;
1388
1389 lru_add_drain();
1390 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1391 address, address + size);
1392 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1393 update_hiwater_rss(vma->vm_mm);
1394 mmu_notifier_invalidate_range_start(&range);
1395 unmap_single_vma(&tlb, vma, address, range.end, details);
1396 mmu_notifier_invalidate_range_end(&range);
1397 tlb_finish_mmu(&tlb, address, range.end);
1398}
1399
1400/**
1401 * zap_vma_ptes - remove ptes mapping the vma
1402 * @vma: vm_area_struct holding ptes to be zapped
1403 * @address: starting address of pages to zap
1404 * @size: number of bytes to zap
1405 *
1406 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1407 *
1408 * The entire address range must be fully contained within the vma.
1409 *
1410 */
1411void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1412 unsigned long size)
1413{
1414 if (address < vma->vm_start || address + size > vma->vm_end ||
1415 !(vma->vm_flags & VM_PFNMAP))
1416 return;
1417
1418 zap_page_range_single(vma, address, size, NULL);
1419}
1420EXPORT_SYMBOL_GPL(zap_vma_ptes);
1421
1422static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1423{
1424 pgd_t *pgd;
1425 p4d_t *p4d;
1426 pud_t *pud;
1427 pmd_t *pmd;
1428
1429 pgd = pgd_offset(mm, addr);
1430 p4d = p4d_alloc(mm, pgd, addr);
1431 if (!p4d)
1432 return NULL;
1433 pud = pud_alloc(mm, p4d, addr);
1434 if (!pud)
1435 return NULL;
1436 pmd = pmd_alloc(mm, pud, addr);
1437 if (!pmd)
1438 return NULL;
1439
1440 VM_BUG_ON(pmd_trans_huge(*pmd));
1441 return pmd;
1442}
1443
1444pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1445 spinlock_t **ptl)
1446{
1447 pmd_t *pmd = walk_to_pmd(mm, addr);
1448
1449 if (!pmd)
1450 return NULL;
1451 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1452}
1453
1454static int validate_page_before_insert(struct page *page)
1455{
1456 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1457 return -EINVAL;
1458 flush_dcache_page(page);
1459 return 0;
1460}
1461
1462static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1463 unsigned long addr, struct page *page, pgprot_t prot)
1464{
1465 if (!pte_none(*pte))
1466 return -EBUSY;
1467 /* Ok, finally just insert the thing.. */
1468 get_page(page);
1469 inc_mm_counter_fast(mm, mm_counter_file(page));
1470 page_add_file_rmap(page, false);
1471 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1472 return 0;
1473}
1474
1475/*
1476 * This is the old fallback for page remapping.
1477 *
1478 * For historical reasons, it only allows reserved pages. Only
1479 * old drivers should use this, and they needed to mark their
1480 * pages reserved for the old functions anyway.
1481 */
1482static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1483 struct page *page, pgprot_t prot)
1484{
1485 struct mm_struct *mm = vma->vm_mm;
1486 int retval;
1487 pte_t *pte;
1488 spinlock_t *ptl;
1489
1490 retval = validate_page_before_insert(page);
1491 if (retval)
1492 goto out;
1493 retval = -ENOMEM;
1494 pte = get_locked_pte(mm, addr, &ptl);
1495 if (!pte)
1496 goto out;
1497 retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1498 pte_unmap_unlock(pte, ptl);
1499out:
1500 return retval;
1501}
1502
1503#ifdef pte_index
1504static int insert_page_in_batch_locked(struct mm_struct *mm, pmd_t *pmd,
1505 unsigned long addr, struct page *page, pgprot_t prot)
1506{
1507 int err;
1508
1509 if (!page_count(page))
1510 return -EINVAL;
1511 err = validate_page_before_insert(page);
1512 return err ? err : insert_page_into_pte_locked(
1513 mm, pte_offset_map(pmd, addr), addr, page, prot);
1514}
1515
1516/* insert_pages() amortizes the cost of spinlock operations
1517 * when inserting pages in a loop. Arch *must* define pte_index.
1518 */
1519static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1520 struct page **pages, unsigned long *num, pgprot_t prot)
1521{
1522 pmd_t *pmd = NULL;
1523 spinlock_t *pte_lock = NULL;
1524 struct mm_struct *const mm = vma->vm_mm;
1525 unsigned long curr_page_idx = 0;
1526 unsigned long remaining_pages_total = *num;
1527 unsigned long pages_to_write_in_pmd;
1528 int ret;
1529more:
1530 ret = -EFAULT;
1531 pmd = walk_to_pmd(mm, addr);
1532 if (!pmd)
1533 goto out;
1534
1535 pages_to_write_in_pmd = min_t(unsigned long,
1536 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1537
1538 /* Allocate the PTE if necessary; takes PMD lock once only. */
1539 ret = -ENOMEM;
1540 if (pte_alloc(mm, pmd))
1541 goto out;
1542 pte_lock = pte_lockptr(mm, pmd);
1543
1544 while (pages_to_write_in_pmd) {
1545 int pte_idx = 0;
1546 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1547
1548 spin_lock(pte_lock);
1549 for (; pte_idx < batch_size; ++pte_idx) {
1550 int err = insert_page_in_batch_locked(mm, pmd,
1551 addr, pages[curr_page_idx], prot);
1552 if (unlikely(err)) {
1553 spin_unlock(pte_lock);
1554 ret = err;
1555 remaining_pages_total -= pte_idx;
1556 goto out;
1557 }
1558 addr += PAGE_SIZE;
1559 ++curr_page_idx;
1560 }
1561 spin_unlock(pte_lock);
1562 pages_to_write_in_pmd -= batch_size;
1563 remaining_pages_total -= batch_size;
1564 }
1565 if (remaining_pages_total)
1566 goto more;
1567 ret = 0;
1568out:
1569 *num = remaining_pages_total;
1570 return ret;
1571}
1572#endif /* ifdef pte_index */
1573
1574/**
1575 * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1576 * @vma: user vma to map to
1577 * @addr: target start user address of these pages
1578 * @pages: source kernel pages
1579 * @num: in: number of pages to map. out: number of pages that were *not*
1580 * mapped. (0 means all pages were successfully mapped).
1581 *
1582 * Preferred over vm_insert_page() when inserting multiple pages.
1583 *
1584 * In case of error, we may have mapped a subset of the provided
1585 * pages. It is the caller's responsibility to account for this case.
1586 *
1587 * The same restrictions apply as in vm_insert_page().
1588 */
1589int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1590 struct page **pages, unsigned long *num)
1591{
1592#ifdef pte_index
1593 const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1594
1595 if (addr < vma->vm_start || end_addr >= vma->vm_end)
1596 return -EFAULT;
1597 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1598 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1599 BUG_ON(vma->vm_flags & VM_PFNMAP);
1600 vma->vm_flags |= VM_MIXEDMAP;
1601 }
1602 /* Defer page refcount checking till we're about to map that page. */
1603 return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1604#else
1605 unsigned long idx = 0, pgcount = *num;
1606 int err;
1607
1608 for (; idx < pgcount; ++idx) {
1609 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1610 if (err)
1611 break;
1612 }
1613 *num = pgcount - idx;
1614 return err;
1615#endif /* ifdef pte_index */
1616}
1617EXPORT_SYMBOL(vm_insert_pages);
1618
1619/**
1620 * vm_insert_page - insert single page into user vma
1621 * @vma: user vma to map to
1622 * @addr: target user address of this page
1623 * @page: source kernel page
1624 *
1625 * This allows drivers to insert individual pages they've allocated
1626 * into a user vma.
1627 *
1628 * The page has to be a nice clean _individual_ kernel allocation.
1629 * If you allocate a compound page, you need to have marked it as
1630 * such (__GFP_COMP), or manually just split the page up yourself
1631 * (see split_page()).
1632 *
1633 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1634 * took an arbitrary page protection parameter. This doesn't allow
1635 * that. Your vma protection will have to be set up correctly, which
1636 * means that if you want a shared writable mapping, you'd better
1637 * ask for a shared writable mapping!
1638 *
1639 * The page does not need to be reserved.
1640 *
1641 * Usually this function is called from f_op->mmap() handler
1642 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1643 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1644 * function from other places, for example from page-fault handler.
1645 *
1646 * Return: %0 on success, negative error code otherwise.
1647 */
1648int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1649 struct page *page)
1650{
1651 if (addr < vma->vm_start || addr >= vma->vm_end)
1652 return -EFAULT;
1653 if (!page_count(page))
1654 return -EINVAL;
1655 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1656 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1657 BUG_ON(vma->vm_flags & VM_PFNMAP);
1658 vma->vm_flags |= VM_MIXEDMAP;
1659 }
1660 return insert_page(vma, addr, page, vma->vm_page_prot);
1661}
1662EXPORT_SYMBOL(vm_insert_page);
1663
1664/*
1665 * __vm_map_pages - maps range of kernel pages into user vma
1666 * @vma: user vma to map to
1667 * @pages: pointer to array of source kernel pages
1668 * @num: number of pages in page array
1669 * @offset: user's requested vm_pgoff
1670 *
1671 * This allows drivers to map range of kernel pages into a user vma.
1672 *
1673 * Return: 0 on success and error code otherwise.
1674 */
1675static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1676 unsigned long num, unsigned long offset)
1677{
1678 unsigned long count = vma_pages(vma);
1679 unsigned long uaddr = vma->vm_start;
1680 int ret, i;
1681
1682 /* Fail if the user requested offset is beyond the end of the object */
1683 if (offset >= num)
1684 return -ENXIO;
1685
1686 /* Fail if the user requested size exceeds available object size */
1687 if (count > num - offset)
1688 return -ENXIO;
1689
1690 for (i = 0; i < count; i++) {
1691 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1692 if (ret < 0)
1693 return ret;
1694 uaddr += PAGE_SIZE;
1695 }
1696
1697 return 0;
1698}
1699
1700/**
1701 * vm_map_pages - maps range of kernel pages starts with non zero offset
1702 * @vma: user vma to map to
1703 * @pages: pointer to array of source kernel pages
1704 * @num: number of pages in page array
1705 *
1706 * Maps an object consisting of @num pages, catering for the user's
1707 * requested vm_pgoff
1708 *
1709 * If we fail to insert any page into the vma, the function will return
1710 * immediately leaving any previously inserted pages present. Callers
1711 * from the mmap handler may immediately return the error as their caller
1712 * will destroy the vma, removing any successfully inserted pages. Other
1713 * callers should make their own arrangements for calling unmap_region().
1714 *
1715 * Context: Process context. Called by mmap handlers.
1716 * Return: 0 on success and error code otherwise.
1717 */
1718int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1719 unsigned long num)
1720{
1721 return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1722}
1723EXPORT_SYMBOL(vm_map_pages);
1724
1725/**
1726 * vm_map_pages_zero - map range of kernel pages starts with zero offset
1727 * @vma: user vma to map to
1728 * @pages: pointer to array of source kernel pages
1729 * @num: number of pages in page array
1730 *
1731 * Similar to vm_map_pages(), except that it explicitly sets the offset
1732 * to 0. This function is intended for the drivers that did not consider
1733 * vm_pgoff.
1734 *
1735 * Context: Process context. Called by mmap handlers.
1736 * Return: 0 on success and error code otherwise.
1737 */
1738int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1739 unsigned long num)
1740{
1741 return __vm_map_pages(vma, pages, num, 0);
1742}
1743EXPORT_SYMBOL(vm_map_pages_zero);
1744
1745static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1746 pfn_t pfn, pgprot_t prot, bool mkwrite)
1747{
1748 struct mm_struct *mm = vma->vm_mm;
1749 pte_t *pte, entry;
1750 spinlock_t *ptl;
1751
1752 pte = get_locked_pte(mm, addr, &ptl);
1753 if (!pte)
1754 return VM_FAULT_OOM;
1755 if (!pte_none(*pte)) {
1756 if (mkwrite) {
1757 /*
1758 * For read faults on private mappings the PFN passed
1759 * in may not match the PFN we have mapped if the
1760 * mapped PFN is a writeable COW page. In the mkwrite
1761 * case we are creating a writable PTE for a shared
1762 * mapping and we expect the PFNs to match. If they
1763 * don't match, we are likely racing with block
1764 * allocation and mapping invalidation so just skip the
1765 * update.
1766 */
1767 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1768 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1769 goto out_unlock;
1770 }
1771 entry = pte_mkyoung(*pte);
1772 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1773 if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1774 update_mmu_cache(vma, addr, pte);
1775 }
1776 goto out_unlock;
1777 }
1778
1779 /* Ok, finally just insert the thing.. */
1780 if (pfn_t_devmap(pfn))
1781 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1782 else
1783 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1784
1785 if (mkwrite) {
1786 entry = pte_mkyoung(entry);
1787 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1788 }
1789
1790 set_pte_at(mm, addr, pte, entry);
1791 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1792
1793out_unlock:
1794 pte_unmap_unlock(pte, ptl);
1795 return VM_FAULT_NOPAGE;
1796}
1797
1798/**
1799 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1800 * @vma: user vma to map to
1801 * @addr: target user address of this page
1802 * @pfn: source kernel pfn
1803 * @pgprot: pgprot flags for the inserted page
1804 *
1805 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1806 * to override pgprot on a per-page basis.
1807 *
1808 * This only makes sense for IO mappings, and it makes no sense for
1809 * COW mappings. In general, using multiple vmas is preferable;
1810 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1811 * impractical.
1812 *
1813 * See vmf_insert_mixed_prot() for a discussion of the implication of using
1814 * a value of @pgprot different from that of @vma->vm_page_prot.
1815 *
1816 * Context: Process context. May allocate using %GFP_KERNEL.
1817 * Return: vm_fault_t value.
1818 */
1819vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1820 unsigned long pfn, pgprot_t pgprot)
1821{
1822 /*
1823 * Technically, architectures with pte_special can avoid all these
1824 * restrictions (same for remap_pfn_range). However we would like
1825 * consistency in testing and feature parity among all, so we should
1826 * try to keep these invariants in place for everybody.
1827 */
1828 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1829 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1830 (VM_PFNMAP|VM_MIXEDMAP));
1831 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1832 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1833
1834 if (addr < vma->vm_start || addr >= vma->vm_end)
1835 return VM_FAULT_SIGBUS;
1836
1837 if (!pfn_modify_allowed(pfn, pgprot))
1838 return VM_FAULT_SIGBUS;
1839
1840 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1841
1842 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1843 false);
1844}
1845EXPORT_SYMBOL(vmf_insert_pfn_prot);
1846
1847/**
1848 * vmf_insert_pfn - insert single pfn into user vma
1849 * @vma: user vma to map to
1850 * @addr: target user address of this page
1851 * @pfn: source kernel pfn
1852 *
1853 * Similar to vm_insert_page, this allows drivers to insert individual pages
1854 * they've allocated into a user vma. Same comments apply.
1855 *
1856 * This function should only be called from a vm_ops->fault handler, and
1857 * in that case the handler should return the result of this function.
1858 *
1859 * vma cannot be a COW mapping.
1860 *
1861 * As this is called only for pages that do not currently exist, we
1862 * do not need to flush old virtual caches or the TLB.
1863 *
1864 * Context: Process context. May allocate using %GFP_KERNEL.
1865 * Return: vm_fault_t value.
1866 */
1867vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1868 unsigned long pfn)
1869{
1870 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1871}
1872EXPORT_SYMBOL(vmf_insert_pfn);
1873
1874static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1875{
1876 /* these checks mirror the abort conditions in vm_normal_page */
1877 if (vma->vm_flags & VM_MIXEDMAP)
1878 return true;
1879 if (pfn_t_devmap(pfn))
1880 return true;
1881 if (pfn_t_special(pfn))
1882 return true;
1883 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1884 return true;
1885 return false;
1886}
1887
1888static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1889 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
1890 bool mkwrite)
1891{
1892 int err;
1893
1894 BUG_ON(!vm_mixed_ok(vma, pfn));
1895
1896 if (addr < vma->vm_start || addr >= vma->vm_end)
1897 return VM_FAULT_SIGBUS;
1898
1899 track_pfn_insert(vma, &pgprot, pfn);
1900
1901 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1902 return VM_FAULT_SIGBUS;
1903
1904 /*
1905 * If we don't have pte special, then we have to use the pfn_valid()
1906 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1907 * refcount the page if pfn_valid is true (hence insert_page rather
1908 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1909 * without pte special, it would there be refcounted as a normal page.
1910 */
1911 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1912 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1913 struct page *page;
1914
1915 /*
1916 * At this point we are committed to insert_page()
1917 * regardless of whether the caller specified flags that
1918 * result in pfn_t_has_page() == false.
1919 */
1920 page = pfn_to_page(pfn_t_to_pfn(pfn));
1921 err = insert_page(vma, addr, page, pgprot);
1922 } else {
1923 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1924 }
1925
1926 if (err == -ENOMEM)
1927 return VM_FAULT_OOM;
1928 if (err < 0 && err != -EBUSY)
1929 return VM_FAULT_SIGBUS;
1930
1931 return VM_FAULT_NOPAGE;
1932}
1933
1934/**
1935 * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
1936 * @vma: user vma to map to
1937 * @addr: target user address of this page
1938 * @pfn: source kernel pfn
1939 * @pgprot: pgprot flags for the inserted page
1940 *
1941 * This is exactly like vmf_insert_mixed(), except that it allows drivers to
1942 * to override pgprot on a per-page basis.
1943 *
1944 * Typically this function should be used by drivers to set caching- and
1945 * encryption bits different than those of @vma->vm_page_prot, because
1946 * the caching- or encryption mode may not be known at mmap() time.
1947 * This is ok as long as @vma->vm_page_prot is not used by the core vm
1948 * to set caching and encryption bits for those vmas (except for COW pages).
1949 * This is ensured by core vm only modifying these page table entries using
1950 * functions that don't touch caching- or encryption bits, using pte_modify()
1951 * if needed. (See for example mprotect()).
1952 * Also when new page-table entries are created, this is only done using the
1953 * fault() callback, and never using the value of vma->vm_page_prot,
1954 * except for page-table entries that point to anonymous pages as the result
1955 * of COW.
1956 *
1957 * Context: Process context. May allocate using %GFP_KERNEL.
1958 * Return: vm_fault_t value.
1959 */
1960vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
1961 pfn_t pfn, pgprot_t pgprot)
1962{
1963 return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
1964}
1965EXPORT_SYMBOL(vmf_insert_mixed_prot);
1966
1967vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1968 pfn_t pfn)
1969{
1970 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
1971}
1972EXPORT_SYMBOL(vmf_insert_mixed);
1973
1974/*
1975 * If the insertion of PTE failed because someone else already added a
1976 * different entry in the mean time, we treat that as success as we assume
1977 * the same entry was actually inserted.
1978 */
1979vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1980 unsigned long addr, pfn_t pfn)
1981{
1982 return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
1983}
1984EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1985
1986/*
1987 * maps a range of physical memory into the requested pages. the old
1988 * mappings are removed. any references to nonexistent pages results
1989 * in null mappings (currently treated as "copy-on-access")
1990 */
1991static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1992 unsigned long addr, unsigned long end,
1993 unsigned long pfn, pgprot_t prot)
1994{
1995 pte_t *pte;
1996 spinlock_t *ptl;
1997 int err = 0;
1998
1999 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2000 if (!pte)
2001 return -ENOMEM;
2002 arch_enter_lazy_mmu_mode();
2003 do {
2004 BUG_ON(!pte_none(*pte));
2005 if (!pfn_modify_allowed(pfn, prot)) {
2006 err = -EACCES;
2007 break;
2008 }
2009 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2010 pfn++;
2011 } while (pte++, addr += PAGE_SIZE, addr != end);
2012 arch_leave_lazy_mmu_mode();
2013 pte_unmap_unlock(pte - 1, ptl);
2014 return err;
2015}
2016
2017static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2018 unsigned long addr, unsigned long end,
2019 unsigned long pfn, pgprot_t prot)
2020{
2021 pmd_t *pmd;
2022 unsigned long next;
2023 int err;
2024
2025 pfn -= addr >> PAGE_SHIFT;
2026 pmd = pmd_alloc(mm, pud, addr);
2027 if (!pmd)
2028 return -ENOMEM;
2029 VM_BUG_ON(pmd_trans_huge(*pmd));
2030 do {
2031 next = pmd_addr_end(addr, end);
2032 err = remap_pte_range(mm, pmd, addr, next,
2033 pfn + (addr >> PAGE_SHIFT), prot);
2034 if (err)
2035 return err;
2036 } while (pmd++, addr = next, addr != end);
2037 return 0;
2038}
2039
2040static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2041 unsigned long addr, unsigned long end,
2042 unsigned long pfn, pgprot_t prot)
2043{
2044 pud_t *pud;
2045 unsigned long next;
2046 int err;
2047
2048 pfn -= addr >> PAGE_SHIFT;
2049 pud = pud_alloc(mm, p4d, addr);
2050 if (!pud)
2051 return -ENOMEM;
2052 do {
2053 next = pud_addr_end(addr, end);
2054 err = remap_pmd_range(mm, pud, addr, next,
2055 pfn + (addr >> PAGE_SHIFT), prot);
2056 if (err)
2057 return err;
2058 } while (pud++, addr = next, addr != end);
2059 return 0;
2060}
2061
2062static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2063 unsigned long addr, unsigned long end,
2064 unsigned long pfn, pgprot_t prot)
2065{
2066 p4d_t *p4d;
2067 unsigned long next;
2068 int err;
2069
2070 pfn -= addr >> PAGE_SHIFT;
2071 p4d = p4d_alloc(mm, pgd, addr);
2072 if (!p4d)
2073 return -ENOMEM;
2074 do {
2075 next = p4d_addr_end(addr, end);
2076 err = remap_pud_range(mm, p4d, addr, next,
2077 pfn + (addr >> PAGE_SHIFT), prot);
2078 if (err)
2079 return err;
2080 } while (p4d++, addr = next, addr != end);
2081 return 0;
2082}
2083
2084/**
2085 * remap_pfn_range - remap kernel memory to userspace
2086 * @vma: user vma to map to
2087 * @addr: target user address to start at
2088 * @pfn: page frame number of kernel physical memory address
2089 * @size: size of mapping area
2090 * @prot: page protection flags for this mapping
2091 *
2092 * Note: this is only safe if the mm semaphore is held when called.
2093 *
2094 * Return: %0 on success, negative error code otherwise.
2095 */
2096int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2097 unsigned long pfn, unsigned long size, pgprot_t prot)
2098{
2099 pgd_t *pgd;
2100 unsigned long next;
2101 unsigned long end = addr + PAGE_ALIGN(size);
2102 struct mm_struct *mm = vma->vm_mm;
2103 unsigned long remap_pfn = pfn;
2104 int err;
2105
2106 /*
2107 * Physically remapped pages are special. Tell the
2108 * rest of the world about it:
2109 * VM_IO tells people not to look at these pages
2110 * (accesses can have side effects).
2111 * VM_PFNMAP tells the core MM that the base pages are just
2112 * raw PFN mappings, and do not have a "struct page" associated
2113 * with them.
2114 * VM_DONTEXPAND
2115 * Disable vma merging and expanding with mremap().
2116 * VM_DONTDUMP
2117 * Omit vma from core dump, even when VM_IO turned off.
2118 *
2119 * There's a horrible special case to handle copy-on-write
2120 * behaviour that some programs depend on. We mark the "original"
2121 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2122 * See vm_normal_page() for details.
2123 */
2124 if (is_cow_mapping(vma->vm_flags)) {
2125 if (addr != vma->vm_start || end != vma->vm_end)
2126 return -EINVAL;
2127 vma->vm_pgoff = pfn;
2128 }
2129
2130 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2131 if (err)
2132 return -EINVAL;
2133
2134 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2135
2136 BUG_ON(addr >= end);
2137 pfn -= addr >> PAGE_SHIFT;
2138 pgd = pgd_offset(mm, addr);
2139 flush_cache_range(vma, addr, end);
2140 do {
2141 next = pgd_addr_end(addr, end);
2142 err = remap_p4d_range(mm, pgd, addr, next,
2143 pfn + (addr >> PAGE_SHIFT), prot);
2144 if (err)
2145 break;
2146 } while (pgd++, addr = next, addr != end);
2147
2148 if (err)
2149 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2150
2151 return err;
2152}
2153EXPORT_SYMBOL(remap_pfn_range);
2154
2155/**
2156 * vm_iomap_memory - remap memory to userspace
2157 * @vma: user vma to map to
2158 * @start: start of the physical memory to be mapped
2159 * @len: size of area
2160 *
2161 * This is a simplified io_remap_pfn_range() for common driver use. The
2162 * driver just needs to give us the physical memory range to be mapped,
2163 * we'll figure out the rest from the vma information.
2164 *
2165 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2166 * whatever write-combining details or similar.
2167 *
2168 * Return: %0 on success, negative error code otherwise.
2169 */
2170int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2171{
2172 unsigned long vm_len, pfn, pages;
2173
2174 /* Check that the physical memory area passed in looks valid */
2175 if (start + len < start)
2176 return -EINVAL;
2177 /*
2178 * You *really* shouldn't map things that aren't page-aligned,
2179 * but we've historically allowed it because IO memory might
2180 * just have smaller alignment.
2181 */
2182 len += start & ~PAGE_MASK;
2183 pfn = start >> PAGE_SHIFT;
2184 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2185 if (pfn + pages < pfn)
2186 return -EINVAL;
2187
2188 /* We start the mapping 'vm_pgoff' pages into the area */
2189 if (vma->vm_pgoff > pages)
2190 return -EINVAL;
2191 pfn += vma->vm_pgoff;
2192 pages -= vma->vm_pgoff;
2193
2194 /* Can we fit all of the mapping? */
2195 vm_len = vma->vm_end - vma->vm_start;
2196 if (vm_len >> PAGE_SHIFT > pages)
2197 return -EINVAL;
2198
2199 /* Ok, let it rip */
2200 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2201}
2202EXPORT_SYMBOL(vm_iomap_memory);
2203
2204static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2205 unsigned long addr, unsigned long end,
2206 pte_fn_t fn, void *data, bool create)
2207{
2208 pte_t *pte;
2209 int err = 0;
2210 spinlock_t *uninitialized_var(ptl);
2211
2212 if (create) {
2213 pte = (mm == &init_mm) ?
2214 pte_alloc_kernel(pmd, addr) :
2215 pte_alloc_map_lock(mm, pmd, addr, &ptl);
2216 if (!pte)
2217 return -ENOMEM;
2218 } else {
2219 pte = (mm == &init_mm) ?
2220 pte_offset_kernel(pmd, addr) :
2221 pte_offset_map_lock(mm, pmd, addr, &ptl);
2222 }
2223
2224 BUG_ON(pmd_huge(*pmd));
2225
2226 arch_enter_lazy_mmu_mode();
2227
2228 do {
2229 if (create || !pte_none(*pte)) {
2230 err = fn(pte++, addr, data);
2231 if (err)
2232 break;
2233 }
2234 } while (addr += PAGE_SIZE, addr != end);
2235
2236 arch_leave_lazy_mmu_mode();
2237
2238 if (mm != &init_mm)
2239 pte_unmap_unlock(pte-1, ptl);
2240 return err;
2241}
2242
2243static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2244 unsigned long addr, unsigned long end,
2245 pte_fn_t fn, void *data, bool create)
2246{
2247 pmd_t *pmd;
2248 unsigned long next;
2249 int err = 0;
2250
2251 BUG_ON(pud_huge(*pud));
2252
2253 if (create) {
2254 pmd = pmd_alloc(mm, pud, addr);
2255 if (!pmd)
2256 return -ENOMEM;
2257 } else {
2258 pmd = pmd_offset(pud, addr);
2259 }
2260 do {
2261 next = pmd_addr_end(addr, end);
2262 if (create || !pmd_none_or_clear_bad(pmd)) {
2263 err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2264 create);
2265 if (err)
2266 break;
2267 }
2268 } while (pmd++, addr = next, addr != end);
2269 return err;
2270}
2271
2272static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2273 unsigned long addr, unsigned long end,
2274 pte_fn_t fn, void *data, bool create)
2275{
2276 pud_t *pud;
2277 unsigned long next;
2278 int err = 0;
2279
2280 if (create) {
2281 pud = pud_alloc(mm, p4d, addr);
2282 if (!pud)
2283 return -ENOMEM;
2284 } else {
2285 pud = pud_offset(p4d, addr);
2286 }
2287 do {
2288 next = pud_addr_end(addr, end);
2289 if (create || !pud_none_or_clear_bad(pud)) {
2290 err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2291 create);
2292 if (err)
2293 break;
2294 }
2295 } while (pud++, addr = next, addr != end);
2296 return err;
2297}
2298
2299static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2300 unsigned long addr, unsigned long end,
2301 pte_fn_t fn, void *data, bool create)
2302{
2303 p4d_t *p4d;
2304 unsigned long next;
2305 int err = 0;
2306
2307 if (create) {
2308 p4d = p4d_alloc(mm, pgd, addr);
2309 if (!p4d)
2310 return -ENOMEM;
2311 } else {
2312 p4d = p4d_offset(pgd, addr);
2313 }
2314 do {
2315 next = p4d_addr_end(addr, end);
2316 if (create || !p4d_none_or_clear_bad(p4d)) {
2317 err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2318 create);
2319 if (err)
2320 break;
2321 }
2322 } while (p4d++, addr = next, addr != end);
2323 return err;
2324}
2325
2326static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2327 unsigned long size, pte_fn_t fn,
2328 void *data, bool create)
2329{
2330 pgd_t *pgd;
2331 unsigned long next;
2332 unsigned long end = addr + size;
2333 int err = 0;
2334
2335 if (WARN_ON(addr >= end))
2336 return -EINVAL;
2337
2338 pgd = pgd_offset(mm, addr);
2339 do {
2340 next = pgd_addr_end(addr, end);
2341 if (!create && pgd_none_or_clear_bad(pgd))
2342 continue;
2343 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create);
2344 if (err)
2345 break;
2346 } while (pgd++, addr = next, addr != end);
2347
2348 return err;
2349}
2350
2351/*
2352 * Scan a region of virtual memory, filling in page tables as necessary
2353 * and calling a provided function on each leaf page table.
2354 */
2355int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2356 unsigned long size, pte_fn_t fn, void *data)
2357{
2358 return __apply_to_page_range(mm, addr, size, fn, data, true);
2359}
2360EXPORT_SYMBOL_GPL(apply_to_page_range);
2361
2362/*
2363 * Scan a region of virtual memory, calling a provided function on
2364 * each leaf page table where it exists.
2365 *
2366 * Unlike apply_to_page_range, this does _not_ fill in page tables
2367 * where they are absent.
2368 */
2369int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2370 unsigned long size, pte_fn_t fn, void *data)
2371{
2372 return __apply_to_page_range(mm, addr, size, fn, data, false);
2373}
2374EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2375
2376/*
2377 * handle_pte_fault chooses page fault handler according to an entry which was
2378 * read non-atomically. Before making any commitment, on those architectures
2379 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2380 * parts, do_swap_page must check under lock before unmapping the pte and
2381 * proceeding (but do_wp_page is only called after already making such a check;
2382 * and do_anonymous_page can safely check later on).
2383 */
2384static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2385 pte_t *page_table, pte_t orig_pte)
2386{
2387 int same = 1;
2388#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2389 if (sizeof(pte_t) > sizeof(unsigned long)) {
2390 spinlock_t *ptl = pte_lockptr(mm, pmd);
2391 spin_lock(ptl);
2392 same = pte_same(*page_table, orig_pte);
2393 spin_unlock(ptl);
2394 }
2395#endif
2396 pte_unmap(page_table);
2397 return same;
2398}
2399
2400static inline bool cow_user_page(struct page *dst, struct page *src,
2401 struct vm_fault *vmf)
2402{
2403 bool ret;
2404 void *kaddr;
2405 void __user *uaddr;
2406 bool locked = false;
2407 struct vm_area_struct *vma = vmf->vma;
2408 struct mm_struct *mm = vma->vm_mm;
2409 unsigned long addr = vmf->address;
2410
2411 debug_dma_assert_idle(src);
2412
2413 if (likely(src)) {
2414 copy_user_highpage(dst, src, addr, vma);
2415 return true;
2416 }
2417
2418 /*
2419 * If the source page was a PFN mapping, we don't have
2420 * a "struct page" for it. We do a best-effort copy by
2421 * just copying from the original user address. If that
2422 * fails, we just zero-fill it. Live with it.
2423 */
2424 kaddr = kmap_atomic(dst);
2425 uaddr = (void __user *)(addr & PAGE_MASK);
2426
2427 /*
2428 * On architectures with software "accessed" bits, we would
2429 * take a double page fault, so mark it accessed here.
2430 */
2431 if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2432 pte_t entry;
2433
2434 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2435 locked = true;
2436 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2437 /*
2438 * Other thread has already handled the fault
2439 * and we don't need to do anything. If it's
2440 * not the case, the fault will be triggered
2441 * again on the same address.
2442 */
2443 ret = false;
2444 goto pte_unlock;
2445 }
2446
2447 entry = pte_mkyoung(vmf->orig_pte);
2448 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2449 update_mmu_cache(vma, addr, vmf->pte);
2450 }
2451
2452 /*
2453 * This really shouldn't fail, because the page is there
2454 * in the page tables. But it might just be unreadable,
2455 * in which case we just give up and fill the result with
2456 * zeroes.
2457 */
2458 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2459 if (locked)
2460 goto warn;
2461
2462 /* Re-validate under PTL if the page is still mapped */
2463 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2464 locked = true;
2465 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2466 /* The PTE changed under us. Retry page fault. */
2467 ret = false;
2468 goto pte_unlock;
2469 }
2470
2471 /*
2472 * The same page can be mapped back since last copy attampt.
2473 * Try to copy again under PTL.
2474 */
2475 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2476 /*
2477 * Give a warn in case there can be some obscure
2478 * use-case
2479 */
2480warn:
2481 WARN_ON_ONCE(1);
2482 clear_page(kaddr);
2483 }
2484 }
2485
2486 ret = true;
2487
2488pte_unlock:
2489 if (locked)
2490 pte_unmap_unlock(vmf->pte, vmf->ptl);
2491 kunmap_atomic(kaddr);
2492 flush_dcache_page(dst);
2493
2494 return ret;
2495}
2496
2497static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2498{
2499 struct file *vm_file = vma->vm_file;
2500
2501 if (vm_file)
2502 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2503
2504 /*
2505 * Special mappings (e.g. VDSO) do not have any file so fake
2506 * a default GFP_KERNEL for them.
2507 */
2508 return GFP_KERNEL;
2509}
2510
2511/*
2512 * Notify the address space that the page is about to become writable so that
2513 * it can prohibit this or wait for the page to get into an appropriate state.
2514 *
2515 * We do this without the lock held, so that it can sleep if it needs to.
2516 */
2517static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2518{
2519 vm_fault_t ret;
2520 struct page *page = vmf->page;
2521 unsigned int old_flags = vmf->flags;
2522
2523 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2524
2525 if (vmf->vma->vm_file &&
2526 IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2527 return VM_FAULT_SIGBUS;
2528
2529 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2530 /* Restore original flags so that caller is not surprised */
2531 vmf->flags = old_flags;
2532 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2533 return ret;
2534 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2535 lock_page(page);
2536 if (!page->mapping) {
2537 unlock_page(page);
2538 return 0; /* retry */
2539 }
2540 ret |= VM_FAULT_LOCKED;
2541 } else
2542 VM_BUG_ON_PAGE(!PageLocked(page), page);
2543 return ret;
2544}
2545
2546/*
2547 * Handle dirtying of a page in shared file mapping on a write fault.
2548 *
2549 * The function expects the page to be locked and unlocks it.
2550 */
2551static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2552{
2553 struct vm_area_struct *vma = vmf->vma;
2554 struct address_space *mapping;
2555 struct page *page = vmf->page;
2556 bool dirtied;
2557 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2558
2559 dirtied = set_page_dirty(page);
2560 VM_BUG_ON_PAGE(PageAnon(page), page);
2561 /*
2562 * Take a local copy of the address_space - page.mapping may be zeroed
2563 * by truncate after unlock_page(). The address_space itself remains
2564 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2565 * release semantics to prevent the compiler from undoing this copying.
2566 */
2567 mapping = page_rmapping(page);
2568 unlock_page(page);
2569
2570 if (!page_mkwrite)
2571 file_update_time(vma->vm_file);
2572
2573 /*
2574 * Throttle page dirtying rate down to writeback speed.
2575 *
2576 * mapping may be NULL here because some device drivers do not
2577 * set page.mapping but still dirty their pages
2578 *
2579 * Drop the mmap_sem before waiting on IO, if we can. The file
2580 * is pinning the mapping, as per above.
2581 */
2582 if ((dirtied || page_mkwrite) && mapping) {
2583 struct file *fpin;
2584
2585 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2586 balance_dirty_pages_ratelimited(mapping);
2587 if (fpin) {
2588 fput(fpin);
2589 return VM_FAULT_RETRY;
2590 }
2591 }
2592
2593 return 0;
2594}
2595
2596/*
2597 * Handle write page faults for pages that can be reused in the current vma
2598 *
2599 * This can happen either due to the mapping being with the VM_SHARED flag,
2600 * or due to us being the last reference standing to the page. In either
2601 * case, all we need to do here is to mark the page as writable and update
2602 * any related book-keeping.
2603 */
2604static inline void wp_page_reuse(struct vm_fault *vmf)
2605 __releases(vmf->ptl)
2606{
2607 struct vm_area_struct *vma = vmf->vma;
2608 struct page *page = vmf->page;
2609 pte_t entry;
2610 /*
2611 * Clear the pages cpupid information as the existing
2612 * information potentially belongs to a now completely
2613 * unrelated process.
2614 */
2615 if (page)
2616 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2617
2618 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2619 entry = pte_mkyoung(vmf->orig_pte);
2620 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2621 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2622 update_mmu_cache(vma, vmf->address, vmf->pte);
2623 pte_unmap_unlock(vmf->pte, vmf->ptl);
2624}
2625
2626/*
2627 * Handle the case of a page which we actually need to copy to a new page.
2628 *
2629 * Called with mmap_sem locked and the old page referenced, but
2630 * without the ptl held.
2631 *
2632 * High level logic flow:
2633 *
2634 * - Allocate a page, copy the content of the old page to the new one.
2635 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2636 * - Take the PTL. If the pte changed, bail out and release the allocated page
2637 * - If the pte is still the way we remember it, update the page table and all
2638 * relevant references. This includes dropping the reference the page-table
2639 * held to the old page, as well as updating the rmap.
2640 * - In any case, unlock the PTL and drop the reference we took to the old page.
2641 */
2642static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2643{
2644 struct vm_area_struct *vma = vmf->vma;
2645 struct mm_struct *mm = vma->vm_mm;
2646 struct page *old_page = vmf->page;
2647 struct page *new_page = NULL;
2648 pte_t entry;
2649 int page_copied = 0;
2650 struct mem_cgroup *memcg;
2651 struct mmu_notifier_range range;
2652
2653 if (unlikely(anon_vma_prepare(vma)))
2654 goto oom;
2655
2656 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2657 new_page = alloc_zeroed_user_highpage_movable(vma,
2658 vmf->address);
2659 if (!new_page)
2660 goto oom;
2661 } else {
2662 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2663 vmf->address);
2664 if (!new_page)
2665 goto oom;
2666
2667 if (!cow_user_page(new_page, old_page, vmf)) {
2668 /*
2669 * COW failed, if the fault was solved by other,
2670 * it's fine. If not, userspace would re-fault on
2671 * the same address and we will handle the fault
2672 * from the second attempt.
2673 */
2674 put_page(new_page);
2675 if (old_page)
2676 put_page(old_page);
2677 return 0;
2678 }
2679 }
2680
2681 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2682 goto oom_free_new;
2683
2684 __SetPageUptodate(new_page);
2685
2686 mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2687 vmf->address & PAGE_MASK,
2688 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2689 mmu_notifier_invalidate_range_start(&range);
2690
2691 /*
2692 * Re-check the pte - we dropped the lock
2693 */
2694 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2695 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2696 if (old_page) {
2697 if (!PageAnon(old_page)) {
2698 dec_mm_counter_fast(mm,
2699 mm_counter_file(old_page));
2700 inc_mm_counter_fast(mm, MM_ANONPAGES);
2701 }
2702 } else {
2703 inc_mm_counter_fast(mm, MM_ANONPAGES);
2704 }
2705 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2706 entry = mk_pte(new_page, vma->vm_page_prot);
2707 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2708 /*
2709 * Clear the pte entry and flush it first, before updating the
2710 * pte with the new entry. This will avoid a race condition
2711 * seen in the presence of one thread doing SMC and another
2712 * thread doing COW.
2713 */
2714 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2715 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2716 mem_cgroup_commit_charge(new_page, memcg, false, false);
2717 lru_cache_add_active_or_unevictable(new_page, vma);
2718 /*
2719 * We call the notify macro here because, when using secondary
2720 * mmu page tables (such as kvm shadow page tables), we want the
2721 * new page to be mapped directly into the secondary page table.
2722 */
2723 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2724 update_mmu_cache(vma, vmf->address, vmf->pte);
2725 if (old_page) {
2726 /*
2727 * Only after switching the pte to the new page may
2728 * we remove the mapcount here. Otherwise another
2729 * process may come and find the rmap count decremented
2730 * before the pte is switched to the new page, and
2731 * "reuse" the old page writing into it while our pte
2732 * here still points into it and can be read by other
2733 * threads.
2734 *
2735 * The critical issue is to order this
2736 * page_remove_rmap with the ptp_clear_flush above.
2737 * Those stores are ordered by (if nothing else,)
2738 * the barrier present in the atomic_add_negative
2739 * in page_remove_rmap.
2740 *
2741 * Then the TLB flush in ptep_clear_flush ensures that
2742 * no process can access the old page before the
2743 * decremented mapcount is visible. And the old page
2744 * cannot be reused until after the decremented
2745 * mapcount is visible. So transitively, TLBs to
2746 * old page will be flushed before it can be reused.
2747 */
2748 page_remove_rmap(old_page, false);
2749 }
2750
2751 /* Free the old page.. */
2752 new_page = old_page;
2753 page_copied = 1;
2754 } else {
2755 mem_cgroup_cancel_charge(new_page, memcg, false);
2756 }
2757
2758 if (new_page)
2759 put_page(new_page);
2760
2761 pte_unmap_unlock(vmf->pte, vmf->ptl);
2762 /*
2763 * No need to double call mmu_notifier->invalidate_range() callback as
2764 * the above ptep_clear_flush_notify() did already call it.
2765 */
2766 mmu_notifier_invalidate_range_only_end(&range);
2767 if (old_page) {
2768 /*
2769 * Don't let another task, with possibly unlocked vma,
2770 * keep the mlocked page.
2771 */
2772 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2773 lock_page(old_page); /* LRU manipulation */
2774 if (PageMlocked(old_page))
2775 munlock_vma_page(old_page);
2776 unlock_page(old_page);
2777 }
2778 put_page(old_page);
2779 }
2780 return page_copied ? VM_FAULT_WRITE : 0;
2781oom_free_new:
2782 put_page(new_page);
2783oom:
2784 if (old_page)
2785 put_page(old_page);
2786 return VM_FAULT_OOM;
2787}
2788
2789/**
2790 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2791 * writeable once the page is prepared
2792 *
2793 * @vmf: structure describing the fault
2794 *
2795 * This function handles all that is needed to finish a write page fault in a
2796 * shared mapping due to PTE being read-only once the mapped page is prepared.
2797 * It handles locking of PTE and modifying it.
2798 *
2799 * The function expects the page to be locked or other protection against
2800 * concurrent faults / writeback (such as DAX radix tree locks).
2801 *
2802 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2803 * we acquired PTE lock.
2804 */
2805vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2806{
2807 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2808 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2809 &vmf->ptl);
2810 /*
2811 * We might have raced with another page fault while we released the
2812 * pte_offset_map_lock.
2813 */
2814 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2815 pte_unmap_unlock(vmf->pte, vmf->ptl);
2816 return VM_FAULT_NOPAGE;
2817 }
2818 wp_page_reuse(vmf);
2819 return 0;
2820}
2821
2822/*
2823 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2824 * mapping
2825 */
2826static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2827{
2828 struct vm_area_struct *vma = vmf->vma;
2829
2830 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2831 vm_fault_t ret;
2832
2833 pte_unmap_unlock(vmf->pte, vmf->ptl);
2834 vmf->flags |= FAULT_FLAG_MKWRITE;
2835 ret = vma->vm_ops->pfn_mkwrite(vmf);
2836 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2837 return ret;
2838 return finish_mkwrite_fault(vmf);
2839 }
2840 wp_page_reuse(vmf);
2841 return VM_FAULT_WRITE;
2842}
2843
2844static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2845 __releases(vmf->ptl)
2846{
2847 struct vm_area_struct *vma = vmf->vma;
2848 vm_fault_t ret = VM_FAULT_WRITE;
2849
2850 get_page(vmf->page);
2851
2852 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2853 vm_fault_t tmp;
2854
2855 pte_unmap_unlock(vmf->pte, vmf->ptl);
2856 tmp = do_page_mkwrite(vmf);
2857 if (unlikely(!tmp || (tmp &
2858 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2859 put_page(vmf->page);
2860 return tmp;
2861 }
2862 tmp = finish_mkwrite_fault(vmf);
2863 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2864 unlock_page(vmf->page);
2865 put_page(vmf->page);
2866 return tmp;
2867 }
2868 } else {
2869 wp_page_reuse(vmf);
2870 lock_page(vmf->page);
2871 }
2872 ret |= fault_dirty_shared_page(vmf);
2873 put_page(vmf->page);
2874
2875 return ret;
2876}
2877
2878/*
2879 * This routine handles present pages, when users try to write
2880 * to a shared page. It is done by copying the page to a new address
2881 * and decrementing the shared-page counter for the old page.
2882 *
2883 * Note that this routine assumes that the protection checks have been
2884 * done by the caller (the low-level page fault routine in most cases).
2885 * Thus we can safely just mark it writable once we've done any necessary
2886 * COW.
2887 *
2888 * We also mark the page dirty at this point even though the page will
2889 * change only once the write actually happens. This avoids a few races,
2890 * and potentially makes it more efficient.
2891 *
2892 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2893 * but allow concurrent faults), with pte both mapped and locked.
2894 * We return with mmap_sem still held, but pte unmapped and unlocked.
2895 */
2896static vm_fault_t do_wp_page(struct vm_fault *vmf)
2897 __releases(vmf->ptl)
2898{
2899 struct vm_area_struct *vma = vmf->vma;
2900
2901 if (userfaultfd_pte_wp(vma, *vmf->pte)) {
2902 pte_unmap_unlock(vmf->pte, vmf->ptl);
2903 return handle_userfault(vmf, VM_UFFD_WP);
2904 }
2905
2906 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2907 if (!vmf->page) {
2908 /*
2909 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2910 * VM_PFNMAP VMA.
2911 *
2912 * We should not cow pages in a shared writeable mapping.
2913 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2914 */
2915 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2916 (VM_WRITE|VM_SHARED))
2917 return wp_pfn_shared(vmf);
2918
2919 pte_unmap_unlock(vmf->pte, vmf->ptl);
2920 return wp_page_copy(vmf);
2921 }
2922
2923 /*
2924 * Take out anonymous pages first, anonymous shared vmas are
2925 * not dirty accountable.
2926 */
2927 if (PageAnon(vmf->page)) {
2928 int total_map_swapcount;
2929 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2930 page_count(vmf->page) != 1))
2931 goto copy;
2932 if (!trylock_page(vmf->page)) {
2933 get_page(vmf->page);
2934 pte_unmap_unlock(vmf->pte, vmf->ptl);
2935 lock_page(vmf->page);
2936 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2937 vmf->address, &vmf->ptl);
2938 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2939 unlock_page(vmf->page);
2940 pte_unmap_unlock(vmf->pte, vmf->ptl);
2941 put_page(vmf->page);
2942 return 0;
2943 }
2944 put_page(vmf->page);
2945 }
2946 if (PageKsm(vmf->page)) {
2947 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2948 vmf->address);
2949 unlock_page(vmf->page);
2950 if (!reused)
2951 goto copy;
2952 wp_page_reuse(vmf);
2953 return VM_FAULT_WRITE;
2954 }
2955 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2956 if (total_map_swapcount == 1) {
2957 /*
2958 * The page is all ours. Move it to
2959 * our anon_vma so the rmap code will
2960 * not search our parent or siblings.
2961 * Protected against the rmap code by
2962 * the page lock.
2963 */
2964 page_move_anon_rmap(vmf->page, vma);
2965 }
2966 unlock_page(vmf->page);
2967 wp_page_reuse(vmf);
2968 return VM_FAULT_WRITE;
2969 }
2970 unlock_page(vmf->page);
2971 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2972 (VM_WRITE|VM_SHARED))) {
2973 return wp_page_shared(vmf);
2974 }
2975copy:
2976 /*
2977 * Ok, we need to copy. Oh, well..
2978 */
2979 get_page(vmf->page);
2980
2981 pte_unmap_unlock(vmf->pte, vmf->ptl);
2982 return wp_page_copy(vmf);
2983}
2984
2985static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2986 unsigned long start_addr, unsigned long end_addr,
2987 struct zap_details *details)
2988{
2989 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2990}
2991
2992static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2993 struct zap_details *details)
2994{
2995 struct vm_area_struct *vma;
2996 pgoff_t vba, vea, zba, zea;
2997
2998 vma_interval_tree_foreach(vma, root,
2999 details->first_index, details->last_index) {
3000
3001 vba = vma->vm_pgoff;
3002 vea = vba + vma_pages(vma) - 1;
3003 zba = details->first_index;
3004 if (zba < vba)
3005 zba = vba;
3006 zea = details->last_index;
3007 if (zea > vea)
3008 zea = vea;
3009
3010 unmap_mapping_range_vma(vma,
3011 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3012 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3013 details);
3014 }
3015}
3016
3017/**
3018 * unmap_mapping_pages() - Unmap pages from processes.
3019 * @mapping: The address space containing pages to be unmapped.
3020 * @start: Index of first page to be unmapped.
3021 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
3022 * @even_cows: Whether to unmap even private COWed pages.
3023 *
3024 * Unmap the pages in this address space from any userspace process which
3025 * has them mmaped. Generally, you want to remove COWed pages as well when
3026 * a file is being truncated, but not when invalidating pages from the page
3027 * cache.
3028 */
3029void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3030 pgoff_t nr, bool even_cows)
3031{
3032 struct zap_details details = { };
3033
3034 details.check_mapping = even_cows ? NULL : mapping;
3035 details.first_index = start;
3036 details.last_index = start + nr - 1;
3037 if (details.last_index < details.first_index)
3038 details.last_index = ULONG_MAX;
3039
3040 i_mmap_lock_write(mapping);
3041 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3042 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3043 i_mmap_unlock_write(mapping);
3044}
3045
3046/**
3047 * unmap_mapping_range - unmap the portion of all mmaps in the specified
3048 * address_space corresponding to the specified byte range in the underlying
3049 * file.
3050 *
3051 * @mapping: the address space containing mmaps to be unmapped.
3052 * @holebegin: byte in first page to unmap, relative to the start of
3053 * the underlying file. This will be rounded down to a PAGE_SIZE
3054 * boundary. Note that this is different from truncate_pagecache(), which
3055 * must keep the partial page. In contrast, we must get rid of
3056 * partial pages.
3057 * @holelen: size of prospective hole in bytes. This will be rounded
3058 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
3059 * end of the file.
3060 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3061 * but 0 when invalidating pagecache, don't throw away private data.
3062 */
3063void unmap_mapping_range(struct address_space *mapping,
3064 loff_t const holebegin, loff_t const holelen, int even_cows)
3065{
3066 pgoff_t hba = holebegin >> PAGE_SHIFT;
3067 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3068
3069 /* Check for overflow. */
3070 if (sizeof(holelen) > sizeof(hlen)) {
3071 long long holeend =
3072 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3073 if (holeend & ~(long long)ULONG_MAX)
3074 hlen = ULONG_MAX - hba + 1;
3075 }
3076
3077 unmap_mapping_pages(mapping, hba, hlen, even_cows);
3078}
3079EXPORT_SYMBOL(unmap_mapping_range);
3080
3081/*
3082 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3083 * but allow concurrent faults), and pte mapped but not yet locked.
3084 * We return with pte unmapped and unlocked.
3085 *
3086 * We return with the mmap_sem locked or unlocked in the same cases
3087 * as does filemap_fault().
3088 */
3089vm_fault_t do_swap_page(struct vm_fault *vmf)
3090{
3091 struct vm_area_struct *vma = vmf->vma;
3092 struct page *page = NULL, *swapcache;
3093 struct mem_cgroup *memcg;
3094 swp_entry_t entry;
3095 pte_t pte;
3096 int locked;
3097 int exclusive = 0;
3098 vm_fault_t ret = 0;
3099
3100 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3101 goto out;
3102
3103 entry = pte_to_swp_entry(vmf->orig_pte);
3104 if (unlikely(non_swap_entry(entry))) {
3105 if (is_migration_entry(entry)) {
3106 migration_entry_wait(vma->vm_mm, vmf->pmd,
3107 vmf->address);
3108 } else if (is_device_private_entry(entry)) {
3109 vmf->page = device_private_entry_to_page(entry);
3110 ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3111 } else if (is_hwpoison_entry(entry)) {
3112 ret = VM_FAULT_HWPOISON;
3113 } else {
3114 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3115 ret = VM_FAULT_SIGBUS;
3116 }
3117 goto out;
3118 }
3119
3120
3121 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3122 page = lookup_swap_cache(entry, vma, vmf->address);
3123 swapcache = page;
3124
3125 if (!page) {
3126 struct swap_info_struct *si = swp_swap_info(entry);
3127
3128 if (si->flags & SWP_SYNCHRONOUS_IO &&
3129 __swap_count(entry) == 1) {
3130 /* skip swapcache */
3131 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3132 vmf->address);
3133 if (page) {
3134 __SetPageLocked(page);
3135 __SetPageSwapBacked(page);
3136 set_page_private(page, entry.val);
3137 lru_cache_add_anon(page);
3138 swap_readpage(page, true);
3139 }
3140 } else {
3141 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3142 vmf);
3143 swapcache = page;
3144 }
3145
3146 if (!page) {
3147 /*
3148 * Back out if somebody else faulted in this pte
3149 * while we released the pte lock.
3150 */
3151 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3152 vmf->address, &vmf->ptl);
3153 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3154 ret = VM_FAULT_OOM;
3155 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3156 goto unlock;
3157 }
3158
3159 /* Had to read the page from swap area: Major fault */
3160 ret = VM_FAULT_MAJOR;
3161 count_vm_event(PGMAJFAULT);
3162 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3163 } else if (PageHWPoison(page)) {
3164 /*
3165 * hwpoisoned dirty swapcache pages are kept for killing
3166 * owner processes (which may be unknown at hwpoison time)
3167 */
3168 ret = VM_FAULT_HWPOISON;
3169 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3170 goto out_release;
3171 }
3172
3173 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3174
3175 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3176 if (!locked) {
3177 ret |= VM_FAULT_RETRY;
3178 goto out_release;
3179 }
3180
3181 /*
3182 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3183 * release the swapcache from under us. The page pin, and pte_same
3184 * test below, are not enough to exclude that. Even if it is still
3185 * swapcache, we need to check that the page's swap has not changed.
3186 */
3187 if (unlikely((!PageSwapCache(page) ||
3188 page_private(page) != entry.val)) && swapcache)
3189 goto out_page;
3190
3191 page = ksm_might_need_to_copy(page, vma, vmf->address);
3192 if (unlikely(!page)) {
3193 ret = VM_FAULT_OOM;
3194 page = swapcache;
3195 goto out_page;
3196 }
3197
3198 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
3199 &memcg, false)) {
3200 ret = VM_FAULT_OOM;
3201 goto out_page;
3202 }
3203
3204 /*
3205 * Back out if somebody else already faulted in this pte.
3206 */
3207 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3208 &vmf->ptl);
3209 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3210 goto out_nomap;
3211
3212 if (unlikely(!PageUptodate(page))) {
3213 ret = VM_FAULT_SIGBUS;
3214 goto out_nomap;
3215 }
3216
3217 /*
3218 * The page isn't present yet, go ahead with the fault.
3219 *
3220 * Be careful about the sequence of operations here.
3221 * To get its accounting right, reuse_swap_page() must be called
3222 * while the page is counted on swap but not yet in mapcount i.e.
3223 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3224 * must be called after the swap_free(), or it will never succeed.
3225 */
3226
3227 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3228 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3229 pte = mk_pte(page, vma->vm_page_prot);
3230 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3231 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3232 vmf->flags &= ~FAULT_FLAG_WRITE;
3233 ret |= VM_FAULT_WRITE;
3234 exclusive = RMAP_EXCLUSIVE;
3235 }
3236 flush_icache_page(vma, page);
3237 if (pte_swp_soft_dirty(vmf->orig_pte))
3238 pte = pte_mksoft_dirty(pte);
3239 if (pte_swp_uffd_wp(vmf->orig_pte)) {
3240 pte = pte_mkuffd_wp(pte);
3241 pte = pte_wrprotect(pte);
3242 }
3243 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3244 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3245 vmf->orig_pte = pte;
3246
3247 /* ksm created a completely new copy */
3248 if (unlikely(page != swapcache && swapcache)) {
3249 page_add_new_anon_rmap(page, vma, vmf->address, false);
3250 mem_cgroup_commit_charge(page, memcg, false, false);
3251 lru_cache_add_active_or_unevictable(page, vma);
3252 } else {
3253 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3254 mem_cgroup_commit_charge(page, memcg, true, false);
3255 activate_page(page);
3256 }
3257
3258 swap_free(entry);
3259 if (mem_cgroup_swap_full(page) ||
3260 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3261 try_to_free_swap(page);
3262 unlock_page(page);
3263 if (page != swapcache && swapcache) {
3264 /*
3265 * Hold the lock to avoid the swap entry to be reused
3266 * until we take the PT lock for the pte_same() check
3267 * (to avoid false positives from pte_same). For
3268 * further safety release the lock after the swap_free
3269 * so that the swap count won't change under a
3270 * parallel locked swapcache.
3271 */
3272 unlock_page(swapcache);
3273 put_page(swapcache);
3274 }
3275
3276 if (vmf->flags & FAULT_FLAG_WRITE) {
3277 ret |= do_wp_page(vmf);
3278 if (ret & VM_FAULT_ERROR)
3279 ret &= VM_FAULT_ERROR;
3280 goto out;
3281 }
3282
3283 /* No need to invalidate - it was non-present before */
3284 update_mmu_cache(vma, vmf->address, vmf->pte);
3285unlock:
3286 pte_unmap_unlock(vmf->pte, vmf->ptl);
3287out:
3288 return ret;
3289out_nomap:
3290 mem_cgroup_cancel_charge(page, memcg, false);
3291 pte_unmap_unlock(vmf->pte, vmf->ptl);
3292out_page:
3293 unlock_page(page);
3294out_release:
3295 put_page(page);
3296 if (page != swapcache && swapcache) {
3297 unlock_page(swapcache);
3298 put_page(swapcache);
3299 }
3300 return ret;
3301}
3302
3303/*
3304 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3305 * but allow concurrent faults), and pte mapped but not yet locked.
3306 * We return with mmap_sem still held, but pte unmapped and unlocked.
3307 */
3308static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3309{
3310 struct vm_area_struct *vma = vmf->vma;
3311 struct mem_cgroup *memcg;
3312 struct page *page;
3313 vm_fault_t ret = 0;
3314 pte_t entry;
3315
3316 /* File mapping without ->vm_ops ? */
3317 if (vma->vm_flags & VM_SHARED)
3318 return VM_FAULT_SIGBUS;
3319
3320 /*
3321 * Use pte_alloc() instead of pte_alloc_map(). We can't run
3322 * pte_offset_map() on pmds where a huge pmd might be created
3323 * from a different thread.
3324 *
3325 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
3326 * parallel threads are excluded by other means.
3327 *
3328 * Here we only have down_read(mmap_sem).
3329 */
3330 if (pte_alloc(vma->vm_mm, vmf->pmd))
3331 return VM_FAULT_OOM;
3332
3333 /* See the comment in pte_alloc_one_map() */
3334 if (unlikely(pmd_trans_unstable(vmf->pmd)))
3335 return 0;
3336
3337 /* Use the zero-page for reads */
3338 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3339 !mm_forbids_zeropage(vma->vm_mm)) {
3340 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3341 vma->vm_page_prot));
3342 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3343 vmf->address, &vmf->ptl);
3344 if (!pte_none(*vmf->pte))
3345 goto unlock;
3346 ret = check_stable_address_space(vma->vm_mm);
3347 if (ret)
3348 goto unlock;
3349 /* Deliver the page fault to userland, check inside PT lock */
3350 if (userfaultfd_missing(vma)) {
3351 pte_unmap_unlock(vmf->pte, vmf->ptl);
3352 return handle_userfault(vmf, VM_UFFD_MISSING);
3353 }
3354 goto setpte;
3355 }
3356
3357 /* Allocate our own private page. */
3358 if (unlikely(anon_vma_prepare(vma)))
3359 goto oom;
3360 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3361 if (!page)
3362 goto oom;
3363
3364 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
3365 false))
3366 goto oom_free_page;
3367
3368 /*
3369 * The memory barrier inside __SetPageUptodate makes sure that
3370 * preceding stores to the page contents become visible before
3371 * the set_pte_at() write.
3372 */
3373 __SetPageUptodate(page);
3374
3375 entry = mk_pte(page, vma->vm_page_prot);
3376 if (vma->vm_flags & VM_WRITE)
3377 entry = pte_mkwrite(pte_mkdirty(entry));
3378
3379 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3380 &vmf->ptl);
3381 if (!pte_none(*vmf->pte))
3382 goto release;
3383
3384 ret = check_stable_address_space(vma->vm_mm);
3385 if (ret)
3386 goto release;
3387
3388 /* Deliver the page fault to userland, check inside PT lock */
3389 if (userfaultfd_missing(vma)) {
3390 pte_unmap_unlock(vmf->pte, vmf->ptl);
3391 mem_cgroup_cancel_charge(page, memcg, false);
3392 put_page(page);
3393 return handle_userfault(vmf, VM_UFFD_MISSING);
3394 }
3395
3396 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3397 page_add_new_anon_rmap(page, vma, vmf->address, false);
3398 mem_cgroup_commit_charge(page, memcg, false, false);
3399 lru_cache_add_active_or_unevictable(page, vma);
3400setpte:
3401 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3402
3403 /* No need to invalidate - it was non-present before */
3404 update_mmu_cache(vma, vmf->address, vmf->pte);
3405unlock:
3406 pte_unmap_unlock(vmf->pte, vmf->ptl);
3407 return ret;
3408release:
3409 mem_cgroup_cancel_charge(page, memcg, false);
3410 put_page(page);
3411 goto unlock;
3412oom_free_page:
3413 put_page(page);
3414oom:
3415 return VM_FAULT_OOM;
3416}
3417
3418/*
3419 * The mmap_sem must have been held on entry, and may have been
3420 * released depending on flags and vma->vm_ops->fault() return value.
3421 * See filemap_fault() and __lock_page_retry().
3422 */
3423static vm_fault_t __do_fault(struct vm_fault *vmf)
3424{
3425 struct vm_area_struct *vma = vmf->vma;
3426 vm_fault_t ret;
3427
3428 /*
3429 * Preallocate pte before we take page_lock because this might lead to
3430 * deadlocks for memcg reclaim which waits for pages under writeback:
3431 * lock_page(A)
3432 * SetPageWriteback(A)
3433 * unlock_page(A)
3434 * lock_page(B)
3435 * lock_page(B)
3436 * pte_alloc_pne
3437 * shrink_page_list
3438 * wait_on_page_writeback(A)
3439 * SetPageWriteback(B)
3440 * unlock_page(B)
3441 * # flush A, B to clear the writeback
3442 */
3443 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3444 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3445 if (!vmf->prealloc_pte)
3446 return VM_FAULT_OOM;
3447 smp_wmb(); /* See comment in __pte_alloc() */
3448 }
3449
3450 ret = vma->vm_ops->fault(vmf);
3451 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3452 VM_FAULT_DONE_COW)))
3453 return ret;
3454
3455 if (unlikely(PageHWPoison(vmf->page))) {
3456 if (ret & VM_FAULT_LOCKED)
3457 unlock_page(vmf->page);
3458 put_page(vmf->page);
3459 vmf->page = NULL;
3460 return VM_FAULT_HWPOISON;
3461 }
3462
3463 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3464 lock_page(vmf->page);
3465 else
3466 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3467
3468 return ret;
3469}
3470
3471/*
3472 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3473 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3474 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3475 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3476 */
3477static int pmd_devmap_trans_unstable(pmd_t *pmd)
3478{
3479 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3480}
3481
3482static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3483{
3484 struct vm_area_struct *vma = vmf->vma;
3485
3486 if (!pmd_none(*vmf->pmd))
3487 goto map_pte;
3488 if (vmf->prealloc_pte) {
3489 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3490 if (unlikely(!pmd_none(*vmf->pmd))) {
3491 spin_unlock(vmf->ptl);
3492 goto map_pte;
3493 }
3494
3495 mm_inc_nr_ptes(vma->vm_mm);
3496 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3497 spin_unlock(vmf->ptl);
3498 vmf->prealloc_pte = NULL;
3499 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3500 return VM_FAULT_OOM;
3501 }
3502map_pte:
3503 /*
3504 * If a huge pmd materialized under us just retry later. Use
3505 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3506 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3507 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3508 * running immediately after a huge pmd fault in a different thread of
3509 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3510 * All we have to ensure is that it is a regular pmd that we can walk
3511 * with pte_offset_map() and we can do that through an atomic read in
3512 * C, which is what pmd_trans_unstable() provides.
3513 */
3514 if (pmd_devmap_trans_unstable(vmf->pmd))
3515 return VM_FAULT_NOPAGE;
3516
3517 /*
3518 * At this point we know that our vmf->pmd points to a page of ptes
3519 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3520 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3521 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3522 * be valid and we will re-check to make sure the vmf->pte isn't
3523 * pte_none() under vmf->ptl protection when we return to
3524 * alloc_set_pte().
3525 */
3526 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3527 &vmf->ptl);
3528 return 0;
3529}
3530
3531#ifdef CONFIG_TRANSPARENT_HUGEPAGE
3532static void deposit_prealloc_pte(struct vm_fault *vmf)
3533{
3534 struct vm_area_struct *vma = vmf->vma;
3535
3536 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3537 /*
3538 * We are going to consume the prealloc table,
3539 * count that as nr_ptes.
3540 */
3541 mm_inc_nr_ptes(vma->vm_mm);
3542 vmf->prealloc_pte = NULL;
3543}
3544
3545static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3546{
3547 struct vm_area_struct *vma = vmf->vma;
3548 bool write = vmf->flags & FAULT_FLAG_WRITE;
3549 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3550 pmd_t entry;
3551 int i;
3552 vm_fault_t ret;
3553
3554 if (!transhuge_vma_suitable(vma, haddr))
3555 return VM_FAULT_FALLBACK;
3556
3557 ret = VM_FAULT_FALLBACK;
3558 page = compound_head(page);
3559
3560 /*
3561 * Archs like ppc64 need additonal space to store information
3562 * related to pte entry. Use the preallocated table for that.
3563 */
3564 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3565 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3566 if (!vmf->prealloc_pte)
3567 return VM_FAULT_OOM;
3568 smp_wmb(); /* See comment in __pte_alloc() */
3569 }
3570
3571 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3572 if (unlikely(!pmd_none(*vmf->pmd)))
3573 goto out;
3574
3575 for (i = 0; i < HPAGE_PMD_NR; i++)
3576 flush_icache_page(vma, page + i);
3577
3578 entry = mk_huge_pmd(page, vma->vm_page_prot);
3579 if (write)
3580 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3581
3582 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3583 page_add_file_rmap(page, true);
3584 /*
3585 * deposit and withdraw with pmd lock held
3586 */
3587 if (arch_needs_pgtable_deposit())
3588 deposit_prealloc_pte(vmf);
3589
3590 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3591
3592 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3593
3594 /* fault is handled */
3595 ret = 0;
3596 count_vm_event(THP_FILE_MAPPED);
3597out:
3598 spin_unlock(vmf->ptl);
3599 return ret;
3600}
3601#else
3602static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3603{
3604 BUILD_BUG();
3605 return 0;
3606}
3607#endif
3608
3609/**
3610 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3611 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3612 *
3613 * @vmf: fault environment
3614 * @memcg: memcg to charge page (only for private mappings)
3615 * @page: page to map
3616 *
3617 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3618 * return.
3619 *
3620 * Target users are page handler itself and implementations of
3621 * vm_ops->map_pages.
3622 *
3623 * Return: %0 on success, %VM_FAULT_ code in case of error.
3624 */
3625vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3626 struct page *page)
3627{
3628 struct vm_area_struct *vma = vmf->vma;
3629 bool write = vmf->flags & FAULT_FLAG_WRITE;
3630 pte_t entry;
3631 vm_fault_t ret;
3632
3633 if (pmd_none(*vmf->pmd) && PageTransCompound(page)) {
3634 /* THP on COW? */
3635 VM_BUG_ON_PAGE(memcg, page);
3636
3637 ret = do_set_pmd(vmf, page);
3638 if (ret != VM_FAULT_FALLBACK)
3639 return ret;
3640 }
3641
3642 if (!vmf->pte) {
3643 ret = pte_alloc_one_map(vmf);
3644 if (ret)
3645 return ret;
3646 }
3647
3648 /* Re-check under ptl */
3649 if (unlikely(!pte_none(*vmf->pte)))
3650 return VM_FAULT_NOPAGE;
3651
3652 flush_icache_page(vma, page);
3653 entry = mk_pte(page, vma->vm_page_prot);
3654 if (write)
3655 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3656 /* copy-on-write page */
3657 if (write && !(vma->vm_flags & VM_SHARED)) {
3658 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3659 page_add_new_anon_rmap(page, vma, vmf->address, false);
3660 mem_cgroup_commit_charge(page, memcg, false, false);
3661 lru_cache_add_active_or_unevictable(page, vma);
3662 } else {
3663 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3664 page_add_file_rmap(page, false);
3665 }
3666 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3667
3668 /* no need to invalidate: a not-present page won't be cached */
3669 update_mmu_cache(vma, vmf->address, vmf->pte);
3670
3671 return 0;
3672}
3673
3674
3675/**
3676 * finish_fault - finish page fault once we have prepared the page to fault
3677 *
3678 * @vmf: structure describing the fault
3679 *
3680 * This function handles all that is needed to finish a page fault once the
3681 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3682 * given page, adds reverse page mapping, handles memcg charges and LRU
3683 * addition.
3684 *
3685 * The function expects the page to be locked and on success it consumes a
3686 * reference of a page being mapped (for the PTE which maps it).
3687 *
3688 * Return: %0 on success, %VM_FAULT_ code in case of error.
3689 */
3690vm_fault_t finish_fault(struct vm_fault *vmf)
3691{
3692 struct page *page;
3693 vm_fault_t ret = 0;
3694
3695 /* Did we COW the page? */
3696 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3697 !(vmf->vma->vm_flags & VM_SHARED))
3698 page = vmf->cow_page;
3699 else
3700 page = vmf->page;
3701
3702 /*
3703 * check even for read faults because we might have lost our CoWed
3704 * page
3705 */
3706 if (!(vmf->vma->vm_flags & VM_SHARED))
3707 ret = check_stable_address_space(vmf->vma->vm_mm);
3708 if (!ret)
3709 ret = alloc_set_pte(vmf, vmf->memcg, page);
3710 if (vmf->pte)
3711 pte_unmap_unlock(vmf->pte, vmf->ptl);
3712 return ret;
3713}
3714
3715static unsigned long fault_around_bytes __read_mostly =
3716 rounddown_pow_of_two(65536);
3717
3718#ifdef CONFIG_DEBUG_FS
3719static int fault_around_bytes_get(void *data, u64 *val)
3720{
3721 *val = fault_around_bytes;
3722 return 0;
3723}
3724
3725/*
3726 * fault_around_bytes must be rounded down to the nearest page order as it's
3727 * what do_fault_around() expects to see.
3728 */
3729static int fault_around_bytes_set(void *data, u64 val)
3730{
3731 if (val / PAGE_SIZE > PTRS_PER_PTE)
3732 return -EINVAL;
3733 if (val > PAGE_SIZE)
3734 fault_around_bytes = rounddown_pow_of_two(val);
3735 else
3736 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3737 return 0;
3738}
3739DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3740 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3741
3742static int __init fault_around_debugfs(void)
3743{
3744 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3745 &fault_around_bytes_fops);
3746 return 0;
3747}
3748late_initcall(fault_around_debugfs);
3749#endif
3750
3751/*
3752 * do_fault_around() tries to map few pages around the fault address. The hope
3753 * is that the pages will be needed soon and this will lower the number of
3754 * faults to handle.
3755 *
3756 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3757 * not ready to be mapped: not up-to-date, locked, etc.
3758 *
3759 * This function is called with the page table lock taken. In the split ptlock
3760 * case the page table lock only protects only those entries which belong to
3761 * the page table corresponding to the fault address.
3762 *
3763 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3764 * only once.
3765 *
3766 * fault_around_bytes defines how many bytes we'll try to map.
3767 * do_fault_around() expects it to be set to a power of two less than or equal
3768 * to PTRS_PER_PTE.
3769 *
3770 * The virtual address of the area that we map is naturally aligned to
3771 * fault_around_bytes rounded down to the machine page size
3772 * (and therefore to page order). This way it's easier to guarantee
3773 * that we don't cross page table boundaries.
3774 */
3775static vm_fault_t do_fault_around(struct vm_fault *vmf)
3776{
3777 unsigned long address = vmf->address, nr_pages, mask;
3778 pgoff_t start_pgoff = vmf->pgoff;
3779 pgoff_t end_pgoff;
3780 int off;
3781 vm_fault_t ret = 0;
3782
3783 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3784 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3785
3786 vmf->address = max(address & mask, vmf->vma->vm_start);
3787 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3788 start_pgoff -= off;
3789
3790 /*
3791 * end_pgoff is either the end of the page table, the end of
3792 * the vma or nr_pages from start_pgoff, depending what is nearest.
3793 */
3794 end_pgoff = start_pgoff -
3795 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3796 PTRS_PER_PTE - 1;
3797 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3798 start_pgoff + nr_pages - 1);
3799
3800 if (pmd_none(*vmf->pmd)) {
3801 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3802 if (!vmf->prealloc_pte)
3803 goto out;
3804 smp_wmb(); /* See comment in __pte_alloc() */
3805 }
3806
3807 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3808
3809 /* Huge page is mapped? Page fault is solved */
3810 if (pmd_trans_huge(*vmf->pmd)) {
3811 ret = VM_FAULT_NOPAGE;
3812 goto out;
3813 }
3814
3815 /* ->map_pages() haven't done anything useful. Cold page cache? */
3816 if (!vmf->pte)
3817 goto out;
3818
3819 /* check if the page fault is solved */
3820 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3821 if (!pte_none(*vmf->pte))
3822 ret = VM_FAULT_NOPAGE;
3823 pte_unmap_unlock(vmf->pte, vmf->ptl);
3824out:
3825 vmf->address = address;
3826 vmf->pte = NULL;
3827 return ret;
3828}
3829
3830static vm_fault_t do_read_fault(struct vm_fault *vmf)
3831{
3832 struct vm_area_struct *vma = vmf->vma;
3833 vm_fault_t ret = 0;
3834
3835 /*
3836 * Let's call ->map_pages() first and use ->fault() as fallback
3837 * if page by the offset is not ready to be mapped (cold cache or
3838 * something).
3839 */
3840 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3841 ret = do_fault_around(vmf);
3842 if (ret)
3843 return ret;
3844 }
3845
3846 ret = __do_fault(vmf);
3847 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3848 return ret;
3849
3850 ret |= finish_fault(vmf);
3851 unlock_page(vmf->page);
3852 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3853 put_page(vmf->page);
3854 return ret;
3855}
3856
3857static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3858{
3859 struct vm_area_struct *vma = vmf->vma;
3860 vm_fault_t ret;
3861
3862 if (unlikely(anon_vma_prepare(vma)))
3863 return VM_FAULT_OOM;
3864
3865 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3866 if (!vmf->cow_page)
3867 return VM_FAULT_OOM;
3868
3869 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3870 &vmf->memcg, false)) {
3871 put_page(vmf->cow_page);
3872 return VM_FAULT_OOM;
3873 }
3874
3875 ret = __do_fault(vmf);
3876 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3877 goto uncharge_out;
3878 if (ret & VM_FAULT_DONE_COW)
3879 return ret;
3880
3881 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3882 __SetPageUptodate(vmf->cow_page);
3883
3884 ret |= finish_fault(vmf);
3885 unlock_page(vmf->page);
3886 put_page(vmf->page);
3887 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3888 goto uncharge_out;
3889 return ret;
3890uncharge_out:
3891 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3892 put_page(vmf->cow_page);
3893 return ret;
3894}
3895
3896static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3897{
3898 struct vm_area_struct *vma = vmf->vma;
3899 vm_fault_t ret, tmp;
3900
3901 ret = __do_fault(vmf);
3902 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3903 return ret;
3904
3905 /*
3906 * Check if the backing address space wants to know that the page is
3907 * about to become writable
3908 */
3909 if (vma->vm_ops->page_mkwrite) {
3910 unlock_page(vmf->page);
3911 tmp = do_page_mkwrite(vmf);
3912 if (unlikely(!tmp ||
3913 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3914 put_page(vmf->page);
3915 return tmp;
3916 }
3917 }
3918
3919 ret |= finish_fault(vmf);
3920 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3921 VM_FAULT_RETRY))) {
3922 unlock_page(vmf->page);
3923 put_page(vmf->page);
3924 return ret;
3925 }
3926
3927 ret |= fault_dirty_shared_page(vmf);
3928 return ret;
3929}
3930
3931/*
3932 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3933 * but allow concurrent faults).
3934 * The mmap_sem may have been released depending on flags and our
3935 * return value. See filemap_fault() and __lock_page_or_retry().
3936 * If mmap_sem is released, vma may become invalid (for example
3937 * by other thread calling munmap()).
3938 */
3939static vm_fault_t do_fault(struct vm_fault *vmf)
3940{
3941 struct vm_area_struct *vma = vmf->vma;
3942 struct mm_struct *vm_mm = vma->vm_mm;
3943 vm_fault_t ret;
3944
3945 /*
3946 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3947 */
3948 if (!vma->vm_ops->fault) {
3949 /*
3950 * If we find a migration pmd entry or a none pmd entry, which
3951 * should never happen, return SIGBUS
3952 */
3953 if (unlikely(!pmd_present(*vmf->pmd)))
3954 ret = VM_FAULT_SIGBUS;
3955 else {
3956 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3957 vmf->pmd,
3958 vmf->address,
3959 &vmf->ptl);
3960 /*
3961 * Make sure this is not a temporary clearing of pte
3962 * by holding ptl and checking again. A R/M/W update
3963 * of pte involves: take ptl, clearing the pte so that
3964 * we don't have concurrent modification by hardware
3965 * followed by an update.
3966 */
3967 if (unlikely(pte_none(*vmf->pte)))
3968 ret = VM_FAULT_SIGBUS;
3969 else
3970 ret = VM_FAULT_NOPAGE;
3971
3972 pte_unmap_unlock(vmf->pte, vmf->ptl);
3973 }
3974 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3975 ret = do_read_fault(vmf);
3976 else if (!(vma->vm_flags & VM_SHARED))
3977 ret = do_cow_fault(vmf);
3978 else
3979 ret = do_shared_fault(vmf);
3980
3981 /* preallocated pagetable is unused: free it */
3982 if (vmf->prealloc_pte) {
3983 pte_free(vm_mm, vmf->prealloc_pte);
3984 vmf->prealloc_pte = NULL;
3985 }
3986 return ret;
3987}
3988
3989static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3990 unsigned long addr, int page_nid,
3991 int *flags)
3992{
3993 get_page(page);
3994
3995 count_vm_numa_event(NUMA_HINT_FAULTS);
3996 if (page_nid == numa_node_id()) {
3997 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3998 *flags |= TNF_FAULT_LOCAL;
3999 }
4000
4001 return mpol_misplaced(page, vma, addr);
4002}
4003
4004static vm_fault_t do_numa_page(struct vm_fault *vmf)
4005{
4006 struct vm_area_struct *vma = vmf->vma;
4007 struct page *page = NULL;
4008 int page_nid = NUMA_NO_NODE;
4009 int last_cpupid;
4010 int target_nid;
4011 bool migrated = false;
4012 pte_t pte, old_pte;
4013 bool was_writable = pte_savedwrite(vmf->orig_pte);
4014 int flags = 0;
4015
4016 /*
4017 * The "pte" at this point cannot be used safely without
4018 * validation through pte_unmap_same(). It's of NUMA type but
4019 * the pfn may be screwed if the read is non atomic.
4020 */
4021 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4022 spin_lock(vmf->ptl);
4023 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4024 pte_unmap_unlock(vmf->pte, vmf->ptl);
4025 goto out;
4026 }
4027
4028 /*
4029 * Make it present again, Depending on how arch implementes non
4030 * accessible ptes, some can allow access by kernel mode.
4031 */
4032 old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4033 pte = pte_modify(old_pte, vma->vm_page_prot);
4034 pte = pte_mkyoung(pte);
4035 if (was_writable)
4036 pte = pte_mkwrite(pte);
4037 ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4038 update_mmu_cache(vma, vmf->address, vmf->pte);
4039
4040 page = vm_normal_page(vma, vmf->address, pte);
4041 if (!page) {
4042 pte_unmap_unlock(vmf->pte, vmf->ptl);
4043 return 0;
4044 }
4045
4046 /* TODO: handle PTE-mapped THP */
4047 if (PageCompound(page)) {
4048 pte_unmap_unlock(vmf->pte, vmf->ptl);
4049 return 0;
4050 }
4051
4052 /*
4053 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4054 * much anyway since they can be in shared cache state. This misses
4055 * the case where a mapping is writable but the process never writes
4056 * to it but pte_write gets cleared during protection updates and
4057 * pte_dirty has unpredictable behaviour between PTE scan updates,
4058 * background writeback, dirty balancing and application behaviour.
4059 */
4060 if (!pte_write(pte))
4061 flags |= TNF_NO_GROUP;
4062
4063 /*
4064 * Flag if the page is shared between multiple address spaces. This
4065 * is later used when determining whether to group tasks together
4066 */
4067 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4068 flags |= TNF_SHARED;
4069
4070 last_cpupid = page_cpupid_last(page);
4071 page_nid = page_to_nid(page);
4072 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4073 &flags);
4074 pte_unmap_unlock(vmf->pte, vmf->ptl);
4075 if (target_nid == NUMA_NO_NODE) {
4076 put_page(page);
4077 goto out;
4078 }
4079
4080 /* Migrate to the requested node */
4081 migrated = migrate_misplaced_page(page, vma, target_nid);
4082 if (migrated) {
4083 page_nid = target_nid;
4084 flags |= TNF_MIGRATED;
4085 } else
4086 flags |= TNF_MIGRATE_FAIL;
4087
4088out:
4089 if (page_nid != NUMA_NO_NODE)
4090 task_numa_fault(last_cpupid, page_nid, 1, flags);
4091 return 0;
4092}
4093
4094static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4095{
4096 if (vma_is_anonymous(vmf->vma))
4097 return do_huge_pmd_anonymous_page(vmf);
4098 if (vmf->vma->vm_ops->huge_fault)
4099 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4100 return VM_FAULT_FALLBACK;
4101}
4102
4103/* `inline' is required to avoid gcc 4.1.2 build error */
4104static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4105{
4106 if (vma_is_anonymous(vmf->vma)) {
4107 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4108 return handle_userfault(vmf, VM_UFFD_WP);
4109 return do_huge_pmd_wp_page(vmf, orig_pmd);
4110 }
4111 if (vmf->vma->vm_ops->huge_fault) {
4112 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4113
4114 if (!(ret & VM_FAULT_FALLBACK))
4115 return ret;
4116 }
4117
4118 /* COW or write-notify handled on pte level: split pmd. */
4119 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4120
4121 return VM_FAULT_FALLBACK;
4122}
4123
4124static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4125{
4126#if defined(CONFIG_TRANSPARENT_HUGEPAGE) && \
4127 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4128 /* No support for anonymous transparent PUD pages yet */
4129 if (vma_is_anonymous(vmf->vma))
4130 goto split;
4131 if (vmf->vma->vm_ops->huge_fault) {
4132 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4133
4134 if (!(ret & VM_FAULT_FALLBACK))
4135 return ret;
4136 }
4137split:
4138 /* COW or write-notify not handled on PUD level: split pud.*/
4139 __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4140#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4141 return VM_FAULT_FALLBACK;
4142}
4143
4144static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4145{
4146#ifdef CONFIG_TRANSPARENT_HUGEPAGE
4147 /* No support for anonymous transparent PUD pages yet */
4148 if (vma_is_anonymous(vmf->vma))
4149 return VM_FAULT_FALLBACK;
4150 if (vmf->vma->vm_ops->huge_fault)
4151 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4152#endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4153 return VM_FAULT_FALLBACK;
4154}
4155
4156/*
4157 * These routines also need to handle stuff like marking pages dirty
4158 * and/or accessed for architectures that don't do it in hardware (most
4159 * RISC architectures). The early dirtying is also good on the i386.
4160 *
4161 * There is also a hook called "update_mmu_cache()" that architectures
4162 * with external mmu caches can use to update those (ie the Sparc or
4163 * PowerPC hashed page tables that act as extended TLBs).
4164 *
4165 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
4166 * concurrent faults).
4167 *
4168 * The mmap_sem may have been released depending on flags and our return value.
4169 * See filemap_fault() and __lock_page_or_retry().
4170 */
4171static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4172{
4173 pte_t entry;
4174
4175 if (unlikely(pmd_none(*vmf->pmd))) {
4176 /*
4177 * Leave __pte_alloc() until later: because vm_ops->fault may
4178 * want to allocate huge page, and if we expose page table
4179 * for an instant, it will be difficult to retract from
4180 * concurrent faults and from rmap lookups.
4181 */
4182 vmf->pte = NULL;
4183 } else {
4184 /* See comment in pte_alloc_one_map() */
4185 if (pmd_devmap_trans_unstable(vmf->pmd))
4186 return 0;
4187 /*
4188 * A regular pmd is established and it can't morph into a huge
4189 * pmd from under us anymore at this point because we hold the
4190 * mmap_sem read mode and khugepaged takes it in write mode.
4191 * So now it's safe to run pte_offset_map().
4192 */
4193 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4194 vmf->orig_pte = *vmf->pte;
4195
4196 /*
4197 * some architectures can have larger ptes than wordsize,
4198 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4199 * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4200 * accesses. The code below just needs a consistent view
4201 * for the ifs and we later double check anyway with the
4202 * ptl lock held. So here a barrier will do.
4203 */
4204 barrier();
4205 if (pte_none(vmf->orig_pte)) {
4206 pte_unmap(vmf->pte);
4207 vmf->pte = NULL;
4208 }
4209 }
4210
4211 if (!vmf->pte) {
4212 if (vma_is_anonymous(vmf->vma))
4213 return do_anonymous_page(vmf);
4214 else
4215 return do_fault(vmf);
4216 }
4217
4218 if (!pte_present(vmf->orig_pte))
4219 return do_swap_page(vmf);
4220
4221 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4222 return do_numa_page(vmf);
4223
4224 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4225 spin_lock(vmf->ptl);
4226 entry = vmf->orig_pte;
4227 if (unlikely(!pte_same(*vmf->pte, entry)))
4228 goto unlock;
4229 if (vmf->flags & FAULT_FLAG_WRITE) {
4230 if (!pte_write(entry))
4231 return do_wp_page(vmf);
4232 entry = pte_mkdirty(entry);
4233 }
4234 entry = pte_mkyoung(entry);
4235 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4236 vmf->flags & FAULT_FLAG_WRITE)) {
4237 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4238 } else {
4239 /*
4240 * This is needed only for protection faults but the arch code
4241 * is not yet telling us if this is a protection fault or not.
4242 * This still avoids useless tlb flushes for .text page faults
4243 * with threads.
4244 */
4245 if (vmf->flags & FAULT_FLAG_WRITE)
4246 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4247 }
4248unlock:
4249 pte_unmap_unlock(vmf->pte, vmf->ptl);
4250 return 0;
4251}
4252
4253/*
4254 * By the time we get here, we already hold the mm semaphore
4255 *
4256 * The mmap_sem may have been released depending on flags and our
4257 * return value. See filemap_fault() and __lock_page_or_retry().
4258 */
4259static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4260 unsigned long address, unsigned int flags)
4261{
4262 struct vm_fault vmf = {
4263 .vma = vma,
4264 .address = address & PAGE_MASK,
4265 .flags = flags,
4266 .pgoff = linear_page_index(vma, address),
4267 .gfp_mask = __get_fault_gfp_mask(vma),
4268 };
4269 unsigned int dirty = flags & FAULT_FLAG_WRITE;
4270 struct mm_struct *mm = vma->vm_mm;
4271 pgd_t *pgd;
4272 p4d_t *p4d;
4273 vm_fault_t ret;
4274
4275 pgd = pgd_offset(mm, address);
4276 p4d = p4d_alloc(mm, pgd, address);
4277 if (!p4d)
4278 return VM_FAULT_OOM;
4279
4280 vmf.pud = pud_alloc(mm, p4d, address);
4281 if (!vmf.pud)
4282 return VM_FAULT_OOM;
4283retry_pud:
4284 if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4285 ret = create_huge_pud(&vmf);
4286 if (!(ret & VM_FAULT_FALLBACK))
4287 return ret;
4288 } else {
4289 pud_t orig_pud = *vmf.pud;
4290
4291 barrier();
4292 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4293
4294 /* NUMA case for anonymous PUDs would go here */
4295
4296 if (dirty && !pud_write(orig_pud)) {
4297 ret = wp_huge_pud(&vmf, orig_pud);
4298 if (!(ret & VM_FAULT_FALLBACK))
4299 return ret;
4300 } else {
4301 huge_pud_set_accessed(&vmf, orig_pud);
4302 return 0;
4303 }
4304 }
4305 }
4306
4307 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4308 if (!vmf.pmd)
4309 return VM_FAULT_OOM;
4310
4311 /* Huge pud page fault raced with pmd_alloc? */
4312 if (pud_trans_unstable(vmf.pud))
4313 goto retry_pud;
4314
4315 if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4316 ret = create_huge_pmd(&vmf);
4317 if (!(ret & VM_FAULT_FALLBACK))
4318 return ret;
4319 } else {
4320 pmd_t orig_pmd = *vmf.pmd;
4321
4322 barrier();
4323 if (unlikely(is_swap_pmd(orig_pmd))) {
4324 VM_BUG_ON(thp_migration_supported() &&
4325 !is_pmd_migration_entry(orig_pmd));
4326 if (is_pmd_migration_entry(orig_pmd))
4327 pmd_migration_entry_wait(mm, vmf.pmd);
4328 return 0;
4329 }
4330 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4331 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4332 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4333
4334 if (dirty && !pmd_write(orig_pmd)) {
4335 ret = wp_huge_pmd(&vmf, orig_pmd);
4336 if (!(ret & VM_FAULT_FALLBACK))
4337 return ret;
4338 } else {
4339 huge_pmd_set_accessed(&vmf, orig_pmd);
4340 return 0;
4341 }
4342 }
4343 }
4344
4345 return handle_pte_fault(&vmf);
4346}
4347
4348/*
4349 * By the time we get here, we already hold the mm semaphore
4350 *
4351 * The mmap_sem may have been released depending on flags and our
4352 * return value. See filemap_fault() and __lock_page_or_retry().
4353 */
4354vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4355 unsigned int flags)
4356{
4357 vm_fault_t ret;
4358
4359 __set_current_state(TASK_RUNNING);
4360
4361 count_vm_event(PGFAULT);
4362 count_memcg_event_mm(vma->vm_mm, PGFAULT);
4363
4364 /* do counter updates before entering really critical section. */
4365 check_sync_rss_stat(current);
4366
4367 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4368 flags & FAULT_FLAG_INSTRUCTION,
4369 flags & FAULT_FLAG_REMOTE))
4370 return VM_FAULT_SIGSEGV;
4371
4372 /*
4373 * Enable the memcg OOM handling for faults triggered in user
4374 * space. Kernel faults are handled more gracefully.
4375 */
4376 if (flags & FAULT_FLAG_USER)
4377 mem_cgroup_enter_user_fault();
4378
4379 if (unlikely(is_vm_hugetlb_page(vma)))
4380 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4381 else
4382 ret = __handle_mm_fault(vma, address, flags);
4383
4384 if (flags & FAULT_FLAG_USER) {
4385 mem_cgroup_exit_user_fault();
4386 /*
4387 * The task may have entered a memcg OOM situation but
4388 * if the allocation error was handled gracefully (no
4389 * VM_FAULT_OOM), there is no need to kill anything.
4390 * Just clean up the OOM state peacefully.
4391 */
4392 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4393 mem_cgroup_oom_synchronize(false);
4394 }
4395
4396 return ret;
4397}
4398EXPORT_SYMBOL_GPL(handle_mm_fault);
4399
4400#ifndef __PAGETABLE_P4D_FOLDED
4401/*
4402 * Allocate p4d page table.
4403 * We've already handled the fast-path in-line.
4404 */
4405int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4406{
4407 p4d_t *new = p4d_alloc_one(mm, address);
4408 if (!new)
4409 return -ENOMEM;
4410
4411 smp_wmb(); /* See comment in __pte_alloc */
4412
4413 spin_lock(&mm->page_table_lock);
4414 if (pgd_present(*pgd)) /* Another has populated it */
4415 p4d_free(mm, new);
4416 else
4417 pgd_populate(mm, pgd, new);
4418 spin_unlock(&mm->page_table_lock);
4419 return 0;
4420}
4421#endif /* __PAGETABLE_P4D_FOLDED */
4422
4423#ifndef __PAGETABLE_PUD_FOLDED
4424/*
4425 * Allocate page upper directory.
4426 * We've already handled the fast-path in-line.
4427 */
4428int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4429{
4430 pud_t *new = pud_alloc_one(mm, address);
4431 if (!new)
4432 return -ENOMEM;
4433
4434 smp_wmb(); /* See comment in __pte_alloc */
4435
4436 spin_lock(&mm->page_table_lock);
4437#ifndef __ARCH_HAS_5LEVEL_HACK
4438 if (!p4d_present(*p4d)) {
4439 mm_inc_nr_puds(mm);
4440 p4d_populate(mm, p4d, new);
4441 } else /* Another has populated it */
4442 pud_free(mm, new);
4443#else
4444 if (!pgd_present(*p4d)) {
4445 mm_inc_nr_puds(mm);
4446 pgd_populate(mm, p4d, new);
4447 } else /* Another has populated it */
4448 pud_free(mm, new);
4449#endif /* __ARCH_HAS_5LEVEL_HACK */
4450 spin_unlock(&mm->page_table_lock);
4451 return 0;
4452}
4453#endif /* __PAGETABLE_PUD_FOLDED */
4454
4455#ifndef __PAGETABLE_PMD_FOLDED
4456/*
4457 * Allocate page middle directory.
4458 * We've already handled the fast-path in-line.
4459 */
4460int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4461{
4462 spinlock_t *ptl;
4463 pmd_t *new = pmd_alloc_one(mm, address);
4464 if (!new)
4465 return -ENOMEM;
4466
4467 smp_wmb(); /* See comment in __pte_alloc */
4468
4469 ptl = pud_lock(mm, pud);
4470 if (!pud_present(*pud)) {
4471 mm_inc_nr_pmds(mm);
4472 pud_populate(mm, pud, new);
4473 } else /* Another has populated it */
4474 pmd_free(mm, new);
4475 spin_unlock(ptl);
4476 return 0;
4477}
4478#endif /* __PAGETABLE_PMD_FOLDED */
4479
4480static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4481 struct mmu_notifier_range *range,
4482 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4483{
4484 pgd_t *pgd;
4485 p4d_t *p4d;
4486 pud_t *pud;
4487 pmd_t *pmd;
4488 pte_t *ptep;
4489
4490 pgd = pgd_offset(mm, address);
4491 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4492 goto out;
4493
4494 p4d = p4d_offset(pgd, address);
4495 if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4496 goto out;
4497
4498 pud = pud_offset(p4d, address);
4499 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4500 goto out;
4501
4502 pmd = pmd_offset(pud, address);
4503 VM_BUG_ON(pmd_trans_huge(*pmd));
4504
4505 if (pmd_huge(*pmd)) {
4506 if (!pmdpp)
4507 goto out;
4508
4509 if (range) {
4510 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4511 NULL, mm, address & PMD_MASK,
4512 (address & PMD_MASK) + PMD_SIZE);
4513 mmu_notifier_invalidate_range_start(range);
4514 }
4515 *ptlp = pmd_lock(mm, pmd);
4516 if (pmd_huge(*pmd)) {
4517 *pmdpp = pmd;
4518 return 0;
4519 }
4520 spin_unlock(*ptlp);
4521 if (range)
4522 mmu_notifier_invalidate_range_end(range);
4523 }
4524
4525 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4526 goto out;
4527
4528 if (range) {
4529 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4530 address & PAGE_MASK,
4531 (address & PAGE_MASK) + PAGE_SIZE);
4532 mmu_notifier_invalidate_range_start(range);
4533 }
4534 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4535 if (!pte_present(*ptep))
4536 goto unlock;
4537 *ptepp = ptep;
4538 return 0;
4539unlock:
4540 pte_unmap_unlock(ptep, *ptlp);
4541 if (range)
4542 mmu_notifier_invalidate_range_end(range);
4543out:
4544 return -EINVAL;
4545}
4546
4547static inline int follow_pte(struct mm_struct *mm, unsigned long address,
4548 pte_t **ptepp, spinlock_t **ptlp)
4549{
4550 int res;
4551
4552 /* (void) is needed to make gcc happy */
4553 (void) __cond_lock(*ptlp,
4554 !(res = __follow_pte_pmd(mm, address, NULL,
4555 ptepp, NULL, ptlp)));
4556 return res;
4557}
4558
4559int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
4560 struct mmu_notifier_range *range,
4561 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
4562{
4563 int res;
4564
4565 /* (void) is needed to make gcc happy */
4566 (void) __cond_lock(*ptlp,
4567 !(res = __follow_pte_pmd(mm, address, range,
4568 ptepp, pmdpp, ptlp)));
4569 return res;
4570}
4571EXPORT_SYMBOL(follow_pte_pmd);
4572
4573/**
4574 * follow_pfn - look up PFN at a user virtual address
4575 * @vma: memory mapping
4576 * @address: user virtual address
4577 * @pfn: location to store found PFN
4578 *
4579 * Only IO mappings and raw PFN mappings are allowed.
4580 *
4581 * Return: zero and the pfn at @pfn on success, -ve otherwise.
4582 */
4583int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4584 unsigned long *pfn)
4585{
4586 int ret = -EINVAL;
4587 spinlock_t *ptl;
4588 pte_t *ptep;
4589
4590 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4591 return ret;
4592
4593 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4594 if (ret)
4595 return ret;
4596 *pfn = pte_pfn(*ptep);
4597 pte_unmap_unlock(ptep, ptl);
4598 return 0;
4599}
4600EXPORT_SYMBOL(follow_pfn);
4601
4602#ifdef CONFIG_HAVE_IOREMAP_PROT
4603int follow_phys(struct vm_area_struct *vma,
4604 unsigned long address, unsigned int flags,
4605 unsigned long *prot, resource_size_t *phys)
4606{
4607 int ret = -EINVAL;
4608 pte_t *ptep, pte;
4609 spinlock_t *ptl;
4610
4611 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4612 goto out;
4613
4614 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4615 goto out;
4616 pte = *ptep;
4617
4618 if ((flags & FOLL_WRITE) && !pte_write(pte))
4619 goto unlock;
4620
4621 *prot = pgprot_val(pte_pgprot(pte));
4622 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4623
4624 ret = 0;
4625unlock:
4626 pte_unmap_unlock(ptep, ptl);
4627out:
4628 return ret;
4629}
4630
4631int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4632 void *buf, int len, int write)
4633{
4634 resource_size_t phys_addr;
4635 unsigned long prot = 0;
4636 void __iomem *maddr;
4637 int offset = addr & (PAGE_SIZE-1);
4638
4639 if (follow_phys(vma, addr, write, &prot, &phys_addr))
4640 return -EINVAL;
4641
4642 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4643 if (!maddr)
4644 return -ENOMEM;
4645
4646 if (write)
4647 memcpy_toio(maddr + offset, buf, len);
4648 else
4649 memcpy_fromio(buf, maddr + offset, len);
4650 iounmap(maddr);
4651
4652 return len;
4653}
4654EXPORT_SYMBOL_GPL(generic_access_phys);
4655#endif
4656
4657/*
4658 * Access another process' address space as given in mm. If non-NULL, use the
4659 * given task for page fault accounting.
4660 */
4661int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
4662 unsigned long addr, void *buf, int len, unsigned int gup_flags)
4663{
4664 struct vm_area_struct *vma;
4665 void *old_buf = buf;
4666 int write = gup_flags & FOLL_WRITE;
4667
4668 if (down_read_killable(&mm->mmap_sem))
4669 return 0;
4670
4671 /* ignore errors, just check how much was successfully transferred */
4672 while (len) {
4673 int bytes, ret, offset;
4674 void *maddr;
4675 struct page *page = NULL;
4676
4677 ret = get_user_pages_remote(tsk, mm, addr, 1,
4678 gup_flags, &page, &vma, NULL);
4679 if (ret <= 0) {
4680#ifndef CONFIG_HAVE_IOREMAP_PROT
4681 break;
4682#else
4683 /*
4684 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4685 * we can access using slightly different code.
4686 */
4687 vma = find_vma(mm, addr);
4688 if (!vma || vma->vm_start > addr)
4689 break;
4690 if (vma->vm_ops && vma->vm_ops->access)
4691 ret = vma->vm_ops->access(vma, addr, buf,
4692 len, write);
4693 if (ret <= 0)
4694 break;
4695 bytes = ret;
4696#endif
4697 } else {
4698 bytes = len;
4699 offset = addr & (PAGE_SIZE-1);
4700 if (bytes > PAGE_SIZE-offset)
4701 bytes = PAGE_SIZE-offset;
4702
4703 maddr = kmap(page);
4704 if (write) {
4705 copy_to_user_page(vma, page, addr,
4706 maddr + offset, buf, bytes);
4707 set_page_dirty_lock(page);
4708 } else {
4709 copy_from_user_page(vma, page, addr,
4710 buf, maddr + offset, bytes);
4711 }
4712 kunmap(page);
4713 put_page(page);
4714 }
4715 len -= bytes;
4716 buf += bytes;
4717 addr += bytes;
4718 }
4719 up_read(&mm->mmap_sem);
4720
4721 return buf - old_buf;
4722}
4723
4724/**
4725 * access_remote_vm - access another process' address space
4726 * @mm: the mm_struct of the target address space
4727 * @addr: start address to access
4728 * @buf: source or destination buffer
4729 * @len: number of bytes to transfer
4730 * @gup_flags: flags modifying lookup behaviour
4731 *
4732 * The caller must hold a reference on @mm.
4733 *
4734 * Return: number of bytes copied from source to destination.
4735 */
4736int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4737 void *buf, int len, unsigned int gup_flags)
4738{
4739 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4740}
4741
4742/*
4743 * Access another process' address space.
4744 * Source/target buffer must be kernel space,
4745 * Do not walk the page table directly, use get_user_pages
4746 */
4747int access_process_vm(struct task_struct *tsk, unsigned long addr,
4748 void *buf, int len, unsigned int gup_flags)
4749{
4750 struct mm_struct *mm;
4751 int ret;
4752
4753 mm = get_task_mm(tsk);
4754 if (!mm)
4755 return 0;
4756
4757 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4758
4759 mmput(mm);
4760
4761 return ret;
4762}
4763EXPORT_SYMBOL_GPL(access_process_vm);
4764
4765/*
4766 * Print the name of a VMA.
4767 */
4768void print_vma_addr(char *prefix, unsigned long ip)
4769{
4770 struct mm_struct *mm = current->mm;
4771 struct vm_area_struct *vma;
4772
4773 /*
4774 * we might be running from an atomic context so we cannot sleep
4775 */
4776 if (!down_read_trylock(&mm->mmap_sem))
4777 return;
4778
4779 vma = find_vma(mm, ip);
4780 if (vma && vma->vm_file) {
4781 struct file *f = vma->vm_file;
4782 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4783 if (buf) {
4784 char *p;
4785
4786 p = file_path(f, buf, PAGE_SIZE);
4787 if (IS_ERR(p))
4788 p = "?";
4789 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4790 vma->vm_start,
4791 vma->vm_end - vma->vm_start);
4792 free_page((unsigned long)buf);
4793 }
4794 }
4795 up_read(&mm->mmap_sem);
4796}
4797
4798#if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4799void __might_fault(const char *file, int line)
4800{
4801 /*
4802 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4803 * holding the mmap_sem, this is safe because kernel memory doesn't
4804 * get paged out, therefore we'll never actually fault, and the
4805 * below annotations will generate false positives.
4806 */
4807 if (uaccess_kernel())
4808 return;
4809 if (pagefault_disabled())
4810 return;
4811 __might_sleep(file, line, 0);
4812#if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4813 if (current->mm)
4814 might_lock_read(&current->mm->mmap_sem);
4815#endif
4816}
4817EXPORT_SYMBOL(__might_fault);
4818#endif
4819
4820#if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4821/*
4822 * Process all subpages of the specified huge page with the specified
4823 * operation. The target subpage will be processed last to keep its
4824 * cache lines hot.
4825 */
4826static inline void process_huge_page(
4827 unsigned long addr_hint, unsigned int pages_per_huge_page,
4828 void (*process_subpage)(unsigned long addr, int idx, void *arg),
4829 void *arg)
4830{
4831 int i, n, base, l;
4832 unsigned long addr = addr_hint &
4833 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4834
4835 /* Process target subpage last to keep its cache lines hot */
4836 might_sleep();
4837 n = (addr_hint - addr) / PAGE_SIZE;
4838 if (2 * n <= pages_per_huge_page) {
4839 /* If target subpage in first half of huge page */
4840 base = 0;
4841 l = n;
4842 /* Process subpages at the end of huge page */
4843 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
4844 cond_resched();
4845 process_subpage(addr + i * PAGE_SIZE, i, arg);
4846 }
4847 } else {
4848 /* If target subpage in second half of huge page */
4849 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
4850 l = pages_per_huge_page - n;
4851 /* Process subpages at the begin of huge page */
4852 for (i = 0; i < base; i++) {
4853 cond_resched();
4854 process_subpage(addr + i * PAGE_SIZE, i, arg);
4855 }
4856 }
4857 /*
4858 * Process remaining subpages in left-right-left-right pattern
4859 * towards the target subpage
4860 */
4861 for (i = 0; i < l; i++) {
4862 int left_idx = base + i;
4863 int right_idx = base + 2 * l - 1 - i;
4864
4865 cond_resched();
4866 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
4867 cond_resched();
4868 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
4869 }
4870}
4871
4872static void clear_gigantic_page(struct page *page,
4873 unsigned long addr,
4874 unsigned int pages_per_huge_page)
4875{
4876 int i;
4877 struct page *p = page;
4878
4879 might_sleep();
4880 for (i = 0; i < pages_per_huge_page;
4881 i++, p = mem_map_next(p, page, i)) {
4882 cond_resched();
4883 clear_user_highpage(p, addr + i * PAGE_SIZE);
4884 }
4885}
4886
4887static void clear_subpage(unsigned long addr, int idx, void *arg)
4888{
4889 struct page *page = arg;
4890
4891 clear_user_highpage(page + idx, addr);
4892}
4893
4894void clear_huge_page(struct page *page,
4895 unsigned long addr_hint, unsigned int pages_per_huge_page)
4896{
4897 unsigned long addr = addr_hint &
4898 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4899
4900 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4901 clear_gigantic_page(page, addr, pages_per_huge_page);
4902 return;
4903 }
4904
4905 process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
4906}
4907
4908static void copy_user_gigantic_page(struct page *dst, struct page *src,
4909 unsigned long addr,
4910 struct vm_area_struct *vma,
4911 unsigned int pages_per_huge_page)
4912{
4913 int i;
4914 struct page *dst_base = dst;
4915 struct page *src_base = src;
4916
4917 for (i = 0; i < pages_per_huge_page; ) {
4918 cond_resched();
4919 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4920
4921 i++;
4922 dst = mem_map_next(dst, dst_base, i);
4923 src = mem_map_next(src, src_base, i);
4924 }
4925}
4926
4927struct copy_subpage_arg {
4928 struct page *dst;
4929 struct page *src;
4930 struct vm_area_struct *vma;
4931};
4932
4933static void copy_subpage(unsigned long addr, int idx, void *arg)
4934{
4935 struct copy_subpage_arg *copy_arg = arg;
4936
4937 copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
4938 addr, copy_arg->vma);
4939}
4940
4941void copy_user_huge_page(struct page *dst, struct page *src,
4942 unsigned long addr_hint, struct vm_area_struct *vma,
4943 unsigned int pages_per_huge_page)
4944{
4945 unsigned long addr = addr_hint &
4946 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
4947 struct copy_subpage_arg arg = {
4948 .dst = dst,
4949 .src = src,
4950 .vma = vma,
4951 };
4952
4953 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4954 copy_user_gigantic_page(dst, src, addr, vma,
4955 pages_per_huge_page);
4956 return;
4957 }
4958
4959 process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
4960}
4961
4962long copy_huge_page_from_user(struct page *dst_page,
4963 const void __user *usr_src,
4964 unsigned int pages_per_huge_page,
4965 bool allow_pagefault)
4966{
4967 void *src = (void *)usr_src;
4968 void *page_kaddr;
4969 unsigned long i, rc = 0;
4970 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4971
4972 for (i = 0; i < pages_per_huge_page; i++) {
4973 if (allow_pagefault)
4974 page_kaddr = kmap(dst_page + i);
4975 else
4976 page_kaddr = kmap_atomic(dst_page + i);
4977 rc = copy_from_user(page_kaddr,
4978 (const void __user *)(src + i * PAGE_SIZE),
4979 PAGE_SIZE);
4980 if (allow_pagefault)
4981 kunmap(dst_page + i);
4982 else
4983 kunmap_atomic(page_kaddr);
4984
4985 ret_val -= (PAGE_SIZE - rc);
4986 if (rc)
4987 break;
4988
4989 cond_resched();
4990 }
4991 return ret_val;
4992}
4993#endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4994
4995#if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4996
4997static struct kmem_cache *page_ptl_cachep;
4998
4999void __init ptlock_cache_init(void)
5000{
5001 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5002 SLAB_PANIC, NULL);
5003}
5004
5005bool ptlock_alloc(struct page *page)
5006{
5007 spinlock_t *ptl;
5008
5009 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5010 if (!ptl)
5011 return false;
5012 page->ptl = ptl;
5013 return true;
5014}
5015
5016void ptlock_free(struct page *page)
5017{
5018 kmem_cache_free(page_ptl_cachep, page->ptl);
5019}
5020#endif