1/*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
6
7/*
8 * demand-loading started 01.12.91 - seems it is high on the list of
9 * things wanted, and it should be easy to implement. - Linus
10 */
11
12/*
13 * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14 * pages started 02.12.91, seems to work. - Linus.
15 *
16 * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17 * would have taken more than the 6M I have free, but it worked well as
18 * far as I could see.
19 *
20 * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21 */
22
23/*
24 * Real VM (paging to/from disk) started 18.12.91. Much more work and
25 * thought has to go into this. Oh, well..
26 * 19.12.91 - works, somewhat. Sometimes I get faults, don't know why.
27 * Found it. Everything seems to work now.
28 * 20.12.91 - Ok, making the swap-device changeable like the root.
29 */
30
31/*
32 * 05.04.94 - Multi-page memory management added for v1.1.
33 * Idea by Alex Bligh (alex@cconcepts.co.uk)
34 *
35 * 16.07.99 - Support of BIGMEM added by Gerhard Wichert, Siemens AG
36 * (Gerhard.Wichert@pdb.siemens.de)
37 *
38 * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39 */
40
41#include <linux/kernel_stat.h>
42#include <linux/mm.h>
43#include <linux/sched/mm.h>
44#include <linux/sched/coredump.h>
45#include <linux/sched/numa_balancing.h>
46#include <linux/sched/task.h>
47#include <linux/hugetlb.h>
48#include <linux/mman.h>
49#include <linux/swap.h>
50#include <linux/highmem.h>
51#include <linux/pagemap.h>
52#include <linux/memremap.h>
53#include <linux/ksm.h>
54#include <linux/rmap.h>
55#include <linux/export.h>
56#include <linux/delayacct.h>
57#include <linux/init.h>
58#include <linux/pfn_t.h>
59#include <linux/writeback.h>
60#include <linux/memcontrol.h>
61#include <linux/mmu_notifier.h>
62#include <linux/swapops.h>
63#include <linux/elf.h>
64#include <linux/gfp.h>
65#include <linux/migrate.h>
66#include <linux/string.h>
67#include <linux/dma-debug.h>
68#include <linux/debugfs.h>
69#include <linux/userfaultfd_k.h>
70#include <linux/dax.h>
71#include <linux/oom.h>
72#include <linux/numa.h>
73
74#include <asm/io.h>
75#include <asm/mmu_context.h>
76#include <asm/pgalloc.h>
77#include <linux/uaccess.h>
78#include <asm/tlb.h>
79#include <asm/tlbflush.h>
80#include <asm/pgtable.h>
81
82#include "internal.h"
83
84#if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
85#warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
86#endif
87
88#ifndef CONFIG_NEED_MULTIPLE_NODES
89/* use the per-pgdat data instead for discontigmem - mbligh */
90unsigned long max_mapnr;
91EXPORT_SYMBOL(max_mapnr);
92
93struct page *mem_map;
94EXPORT_SYMBOL(mem_map);
95#endif
96
97/*
98 * A number of key systems in x86 including ioremap() rely on the assumption
99 * that high_memory defines the upper bound on direct map memory, then end
100 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
101 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
102 * and ZONE_HIGHMEM.
103 */
104void *high_memory;
105EXPORT_SYMBOL(high_memory);
106
107/*
108 * Randomize the address space (stacks, mmaps, brk, etc.).
109 *
110 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
111 * as ancient (libc5 based) binaries can segfault. )
112 */
113int randomize_va_space __read_mostly =
114#ifdef CONFIG_COMPAT_BRK
115 1;
116#else
117 2;
118#endif
119
120static int __init disable_randmaps(char *s)
121{
122 randomize_va_space = 0;
123 return 1;
124}
125__setup("norandmaps", disable_randmaps);
126
127unsigned long zero_pfn __read_mostly;
128EXPORT_SYMBOL(zero_pfn);
129
130unsigned long highest_memmap_pfn __read_mostly;
131
132/*
133 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
134 */
135static int __init init_zero_pfn(void)
136{
137 zero_pfn = page_to_pfn(ZERO_PAGE(0));
138 return 0;
139}
140core_initcall(init_zero_pfn);
141
142
143#if defined(SPLIT_RSS_COUNTING)
144
145void sync_mm_rss(struct mm_struct *mm)
146{
147 int i;
148
149 for (i = 0; i < NR_MM_COUNTERS; i++) {
150 if (current->rss_stat.count[i]) {
151 add_mm_counter(mm, i, current->rss_stat.count[i]);
152 current->rss_stat.count[i] = 0;
153 }
154 }
155 current->rss_stat.events = 0;
156}
157
158static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
159{
160 struct task_struct *task = current;
161
162 if (likely(task->mm == mm))
163 task->rss_stat.count[member] += val;
164 else
165 add_mm_counter(mm, member, val);
166}
167#define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
168#define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
169
170/* sync counter once per 64 page faults */
171#define TASK_RSS_EVENTS_THRESH (64)
172static void check_sync_rss_stat(struct task_struct *task)
173{
174 if (unlikely(task != current))
175 return;
176 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
177 sync_mm_rss(task->mm);
178}
179#else /* SPLIT_RSS_COUNTING */
180
181#define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
182#define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
183
184static void check_sync_rss_stat(struct task_struct *task)
185{
186}
187
188#endif /* SPLIT_RSS_COUNTING */
189
190/*
191 * Note: this doesn't free the actual pages themselves. That
192 * has been handled earlier when unmapping all the memory regions.
193 */
194static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
195 unsigned long addr)
196{
197 pgtable_t token = pmd_pgtable(*pmd);
198 pmd_clear(pmd);
199 pte_free_tlb(tlb, token, addr);
200 mm_dec_nr_ptes(tlb->mm);
201}
202
203static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
204 unsigned long addr, unsigned long end,
205 unsigned long floor, unsigned long ceiling)
206{
207 pmd_t *pmd;
208 unsigned long next;
209 unsigned long start;
210
211 start = addr;
212 pmd = pmd_offset(pud, addr);
213 do {
214 next = pmd_addr_end(addr, end);
215 if (pmd_none_or_clear_bad(pmd))
216 continue;
217 free_pte_range(tlb, pmd, addr);
218 } while (pmd++, addr = next, addr != end);
219
220 start &= PUD_MASK;
221 if (start < floor)
222 return;
223 if (ceiling) {
224 ceiling &= PUD_MASK;
225 if (!ceiling)
226 return;
227 }
228 if (end - 1 > ceiling - 1)
229 return;
230
231 pmd = pmd_offset(pud, start);
232 pud_clear(pud);
233 pmd_free_tlb(tlb, pmd, start);
234 mm_dec_nr_pmds(tlb->mm);
235}
236
237static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
238 unsigned long addr, unsigned long end,
239 unsigned long floor, unsigned long ceiling)
240{
241 pud_t *pud;
242 unsigned long next;
243 unsigned long start;
244
245 start = addr;
246 pud = pud_offset(p4d, addr);
247 do {
248 next = pud_addr_end(addr, end);
249 if (pud_none_or_clear_bad(pud))
250 continue;
251 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
252 } while (pud++, addr = next, addr != end);
253
254 start &= P4D_MASK;
255 if (start < floor)
256 return;
257 if (ceiling) {
258 ceiling &= P4D_MASK;
259 if (!ceiling)
260 return;
261 }
262 if (end - 1 > ceiling - 1)
263 return;
264
265 pud = pud_offset(p4d, start);
266 p4d_clear(p4d);
267 pud_free_tlb(tlb, pud, start);
268 mm_dec_nr_puds(tlb->mm);
269}
270
271static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
272 unsigned long addr, unsigned long end,
273 unsigned long floor, unsigned long ceiling)
274{
275 p4d_t *p4d;
276 unsigned long next;
277 unsigned long start;
278
279 start = addr;
280 p4d = p4d_offset(pgd, addr);
281 do {
282 next = p4d_addr_end(addr, end);
283 if (p4d_none_or_clear_bad(p4d))
284 continue;
285 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
286 } while (p4d++, addr = next, addr != end);
287
288 start &= PGDIR_MASK;
289 if (start < floor)
290 return;
291 if (ceiling) {
292 ceiling &= PGDIR_MASK;
293 if (!ceiling)
294 return;
295 }
296 if (end - 1 > ceiling - 1)
297 return;
298
299 p4d = p4d_offset(pgd, start);
300 pgd_clear(pgd);
301 p4d_free_tlb(tlb, p4d, start);
302}
303
304/*
305 * This function frees user-level page tables of a process.
306 */
307void free_pgd_range(struct mmu_gather *tlb,
308 unsigned long addr, unsigned long end,
309 unsigned long floor, unsigned long ceiling)
310{
311 pgd_t *pgd;
312 unsigned long next;
313
314 /*
315 * The next few lines have given us lots of grief...
316 *
317 * Why are we testing PMD* at this top level? Because often
318 * there will be no work to do at all, and we'd prefer not to
319 * go all the way down to the bottom just to discover that.
320 *
321 * Why all these "- 1"s? Because 0 represents both the bottom
322 * of the address space and the top of it (using -1 for the
323 * top wouldn't help much: the masks would do the wrong thing).
324 * The rule is that addr 0 and floor 0 refer to the bottom of
325 * the address space, but end 0 and ceiling 0 refer to the top
326 * Comparisons need to use "end - 1" and "ceiling - 1" (though
327 * that end 0 case should be mythical).
328 *
329 * Wherever addr is brought up or ceiling brought down, we must
330 * be careful to reject "the opposite 0" before it confuses the
331 * subsequent tests. But what about where end is brought down
332 * by PMD_SIZE below? no, end can't go down to 0 there.
333 *
334 * Whereas we round start (addr) and ceiling down, by different
335 * masks at different levels, in order to test whether a table
336 * now has no other vmas using it, so can be freed, we don't
337 * bother to round floor or end up - the tests don't need that.
338 */
339
340 addr &= PMD_MASK;
341 if (addr < floor) {
342 addr += PMD_SIZE;
343 if (!addr)
344 return;
345 }
346 if (ceiling) {
347 ceiling &= PMD_MASK;
348 if (!ceiling)
349 return;
350 }
351 if (end - 1 > ceiling - 1)
352 end -= PMD_SIZE;
353 if (addr > end - 1)
354 return;
355 /*
356 * We add page table cache pages with PAGE_SIZE,
357 * (see pte_free_tlb()), flush the tlb if we need
358 */
359 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
360 pgd = pgd_offset(tlb->mm, addr);
361 do {
362 next = pgd_addr_end(addr, end);
363 if (pgd_none_or_clear_bad(pgd))
364 continue;
365 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
366 } while (pgd++, addr = next, addr != end);
367}
368
369void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
370 unsigned long floor, unsigned long ceiling)
371{
372 while (vma) {
373 struct vm_area_struct *next = vma->vm_next;
374 unsigned long addr = vma->vm_start;
375
376 /*
377 * Hide vma from rmap and truncate_pagecache before freeing
378 * pgtables
379 */
380 unlink_anon_vmas(vma);
381 unlink_file_vma(vma);
382
383 if (is_vm_hugetlb_page(vma)) {
384 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
385 floor, next ? next->vm_start : ceiling);
386 } else {
387 /*
388 * Optimization: gather nearby vmas into one call down
389 */
390 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
391 && !is_vm_hugetlb_page(next)) {
392 vma = next;
393 next = vma->vm_next;
394 unlink_anon_vmas(vma);
395 unlink_file_vma(vma);
396 }
397 free_pgd_range(tlb, addr, vma->vm_end,
398 floor, next ? next->vm_start : ceiling);
399 }
400 vma = next;
401 }
402}
403
404int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
405{
406 spinlock_t *ptl;
407 pgtable_t new = pte_alloc_one(mm);
408 if (!new)
409 return -ENOMEM;
410
411 /*
412 * Ensure all pte setup (eg. pte page lock and page clearing) are
413 * visible before the pte is made visible to other CPUs by being
414 * put into page tables.
415 *
416 * The other side of the story is the pointer chasing in the page
417 * table walking code (when walking the page table without locking;
418 * ie. most of the time). Fortunately, these data accesses consist
419 * of a chain of data-dependent loads, meaning most CPUs (alpha
420 * being the notable exception) will already guarantee loads are
421 * seen in-order. See the alpha page table accessors for the
422 * smp_read_barrier_depends() barriers in page table walking code.
423 */
424 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
425
426 ptl = pmd_lock(mm, pmd);
427 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
428 mm_inc_nr_ptes(mm);
429 pmd_populate(mm, pmd, new);
430 new = NULL;
431 }
432 spin_unlock(ptl);
433 if (new)
434 pte_free(mm, new);
435 return 0;
436}
437
438int __pte_alloc_kernel(pmd_t *pmd)
439{
440 pte_t *new = pte_alloc_one_kernel(&init_mm);
441 if (!new)
442 return -ENOMEM;
443
444 smp_wmb(); /* See comment in __pte_alloc */
445
446 spin_lock(&init_mm.page_table_lock);
447 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
448 pmd_populate_kernel(&init_mm, pmd, new);
449 new = NULL;
450 }
451 spin_unlock(&init_mm.page_table_lock);
452 if (new)
453 pte_free_kernel(&init_mm, new);
454 return 0;
455}
456
457static inline void init_rss_vec(int *rss)
458{
459 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
460}
461
462static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
463{
464 int i;
465
466 if (current->mm == mm)
467 sync_mm_rss(mm);
468 for (i = 0; i < NR_MM_COUNTERS; i++)
469 if (rss[i])
470 add_mm_counter(mm, i, rss[i]);
471}
472
473/*
474 * This function is called to print an error when a bad pte
475 * is found. For example, we might have a PFN-mapped pte in
476 * a region that doesn't allow it.
477 *
478 * The calling function must still handle the error.
479 */
480static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
481 pte_t pte, struct page *page)
482{
483 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
484 p4d_t *p4d = p4d_offset(pgd, addr);
485 pud_t *pud = pud_offset(p4d, addr);
486 pmd_t *pmd = pmd_offset(pud, addr);
487 struct address_space *mapping;
488 pgoff_t index;
489 static unsigned long resume;
490 static unsigned long nr_shown;
491 static unsigned long nr_unshown;
492
493 /*
494 * Allow a burst of 60 reports, then keep quiet for that minute;
495 * or allow a steady drip of one report per second.
496 */
497 if (nr_shown == 60) {
498 if (time_before(jiffies, resume)) {
499 nr_unshown++;
500 return;
501 }
502 if (nr_unshown) {
503 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
504 nr_unshown);
505 nr_unshown = 0;
506 }
507 nr_shown = 0;
508 }
509 if (nr_shown++ == 0)
510 resume = jiffies + 60 * HZ;
511
512 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
513 index = linear_page_index(vma, addr);
514
515 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
516 current->comm,
517 (long long)pte_val(pte), (long long)pmd_val(*pmd));
518 if (page)
519 dump_page(page, "bad pte");
520 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
521 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
522 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
523 vma->vm_file,
524 vma->vm_ops ? vma->vm_ops->fault : NULL,
525 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
526 mapping ? mapping->a_ops->readpage : NULL);
527 dump_stack();
528 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
529}
530
531/*
532 * vm_normal_page -- This function gets the "struct page" associated with a pte.
533 *
534 * "Special" mappings do not wish to be associated with a "struct page" (either
535 * it doesn't exist, or it exists but they don't want to touch it). In this
536 * case, NULL is returned here. "Normal" mappings do have a struct page.
537 *
538 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
539 * pte bit, in which case this function is trivial. Secondly, an architecture
540 * may not have a spare pte bit, which requires a more complicated scheme,
541 * described below.
542 *
543 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
544 * special mapping (even if there are underlying and valid "struct pages").
545 * COWed pages of a VM_PFNMAP are always normal.
546 *
547 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
548 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
549 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
550 * mapping will always honor the rule
551 *
552 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
553 *
554 * And for normal mappings this is false.
555 *
556 * This restricts such mappings to be a linear translation from virtual address
557 * to pfn. To get around this restriction, we allow arbitrary mappings so long
558 * as the vma is not a COW mapping; in that case, we know that all ptes are
559 * special (because none can have been COWed).
560 *
561 *
562 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
563 *
564 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
565 * page" backing, however the difference is that _all_ pages with a struct
566 * page (that is, those where pfn_valid is true) are refcounted and considered
567 * normal pages by the VM. The disadvantage is that pages are refcounted
568 * (which can be slower and simply not an option for some PFNMAP users). The
569 * advantage is that we don't have to follow the strict linearity rule of
570 * PFNMAP mappings in order to support COWable mappings.
571 *
572 */
573struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
574 pte_t pte, bool with_public_device)
575{
576 unsigned long pfn = pte_pfn(pte);
577
578 if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
579 if (likely(!pte_special(pte)))
580 goto check_pfn;
581 if (vma->vm_ops && vma->vm_ops->find_special_page)
582 return vma->vm_ops->find_special_page(vma, addr);
583 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
584 return NULL;
585 if (is_zero_pfn(pfn))
586 return NULL;
587
588 /*
589 * Device public pages are special pages (they are ZONE_DEVICE
590 * pages but different from persistent memory). They behave
591 * allmost like normal pages. The difference is that they are
592 * not on the lru and thus should never be involve with any-
593 * thing that involve lru manipulation (mlock, numa balancing,
594 * ...).
595 *
596 * This is why we still want to return NULL for such page from
597 * vm_normal_page() so that we do not have to special case all
598 * call site of vm_normal_page().
599 */
600 if (likely(pfn <= highest_memmap_pfn)) {
601 struct page *page = pfn_to_page(pfn);
602
603 if (is_device_public_page(page)) {
604 if (with_public_device)
605 return page;
606 return NULL;
607 }
608 }
609
610 if (pte_devmap(pte))
611 return NULL;
612
613 print_bad_pte(vma, addr, pte, NULL);
614 return NULL;
615 }
616
617 /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
618
619 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
620 if (vma->vm_flags & VM_MIXEDMAP) {
621 if (!pfn_valid(pfn))
622 return NULL;
623 goto out;
624 } else {
625 unsigned long off;
626 off = (addr - vma->vm_start) >> PAGE_SHIFT;
627 if (pfn == vma->vm_pgoff + off)
628 return NULL;
629 if (!is_cow_mapping(vma->vm_flags))
630 return NULL;
631 }
632 }
633
634 if (is_zero_pfn(pfn))
635 return NULL;
636
637check_pfn:
638 if (unlikely(pfn > highest_memmap_pfn)) {
639 print_bad_pte(vma, addr, pte, NULL);
640 return NULL;
641 }
642
643 /*
644 * NOTE! We still have PageReserved() pages in the page tables.
645 * eg. VDSO mappings can cause them to exist.
646 */
647out:
648 return pfn_to_page(pfn);
649}
650
651#ifdef CONFIG_TRANSPARENT_HUGEPAGE
652struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
653 pmd_t pmd)
654{
655 unsigned long pfn = pmd_pfn(pmd);
656
657 /*
658 * There is no pmd_special() but there may be special pmds, e.g.
659 * in a direct-access (dax) mapping, so let's just replicate the
660 * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
661 */
662 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
663 if (vma->vm_flags & VM_MIXEDMAP) {
664 if (!pfn_valid(pfn))
665 return NULL;
666 goto out;
667 } else {
668 unsigned long off;
669 off = (addr - vma->vm_start) >> PAGE_SHIFT;
670 if (pfn == vma->vm_pgoff + off)
671 return NULL;
672 if (!is_cow_mapping(vma->vm_flags))
673 return NULL;
674 }
675 }
676
677 if (pmd_devmap(pmd))
678 return NULL;
679 if (is_zero_pfn(pfn))
680 return NULL;
681 if (unlikely(pfn > highest_memmap_pfn))
682 return NULL;
683
684 /*
685 * NOTE! We still have PageReserved() pages in the page tables.
686 * eg. VDSO mappings can cause them to exist.
687 */
688out:
689 return pfn_to_page(pfn);
690}
691#endif
692
693/*
694 * copy one vm_area from one task to the other. Assumes the page tables
695 * already present in the new task to be cleared in the whole range
696 * covered by this vma.
697 */
698
699static inline unsigned long
700copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
701 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
702 unsigned long addr, int *rss)
703{
704 unsigned long vm_flags = vma->vm_flags;
705 pte_t pte = *src_pte;
706 struct page *page;
707
708 /* pte contains position in swap or file, so copy. */
709 if (unlikely(!pte_present(pte))) {
710 swp_entry_t entry = pte_to_swp_entry(pte);
711
712 if (likely(!non_swap_entry(entry))) {
713 if (swap_duplicate(entry) < 0)
714 return entry.val;
715
716 /* make sure dst_mm is on swapoff's mmlist. */
717 if (unlikely(list_empty(&dst_mm->mmlist))) {
718 spin_lock(&mmlist_lock);
719 if (list_empty(&dst_mm->mmlist))
720 list_add(&dst_mm->mmlist,
721 &src_mm->mmlist);
722 spin_unlock(&mmlist_lock);
723 }
724 rss[MM_SWAPENTS]++;
725 } else if (is_migration_entry(entry)) {
726 page = migration_entry_to_page(entry);
727
728 rss[mm_counter(page)]++;
729
730 if (is_write_migration_entry(entry) &&
731 is_cow_mapping(vm_flags)) {
732 /*
733 * COW mappings require pages in both
734 * parent and child to be set to read.
735 */
736 make_migration_entry_read(&entry);
737 pte = swp_entry_to_pte(entry);
738 if (pte_swp_soft_dirty(*src_pte))
739 pte = pte_swp_mksoft_dirty(pte);
740 set_pte_at(src_mm, addr, src_pte, pte);
741 }
742 } else if (is_device_private_entry(entry)) {
743 page = device_private_entry_to_page(entry);
744
745 /*
746 * Update rss count even for unaddressable pages, as
747 * they should treated just like normal pages in this
748 * respect.
749 *
750 * We will likely want to have some new rss counters
751 * for unaddressable pages, at some point. But for now
752 * keep things as they are.
753 */
754 get_page(page);
755 rss[mm_counter(page)]++;
756 page_dup_rmap(page, false);
757
758 /*
759 * We do not preserve soft-dirty information, because so
760 * far, checkpoint/restore is the only feature that
761 * requires that. And checkpoint/restore does not work
762 * when a device driver is involved (you cannot easily
763 * save and restore device driver state).
764 */
765 if (is_write_device_private_entry(entry) &&
766 is_cow_mapping(vm_flags)) {
767 make_device_private_entry_read(&entry);
768 pte = swp_entry_to_pte(entry);
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 page = vm_normal_page(vma, addr, pte);
793 if (page) {
794 get_page(page);
795 page_dup_rmap(page, false);
796 rss[mm_counter(page)]++;
797 } else if (pte_devmap(pte)) {
798 page = pte_page(pte);
799
800 /*
801 * Cache coherent device memory behave like regular page and
802 * not like persistent memory page. For more informations see
803 * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
804 */
805 if (is_device_public_page(page)) {
806 get_page(page);
807 page_dup_rmap(page, false);
808 rss[mm_counter(page)]++;
809 }
810 }
811
812out_set_pte:
813 set_pte_at(dst_mm, addr, dst_pte, pte);
814 return 0;
815}
816
817static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
818 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
819 unsigned long addr, unsigned long end)
820{
821 pte_t *orig_src_pte, *orig_dst_pte;
822 pte_t *src_pte, *dst_pte;
823 spinlock_t *src_ptl, *dst_ptl;
824 int progress = 0;
825 int rss[NR_MM_COUNTERS];
826 swp_entry_t entry = (swp_entry_t){0};
827
828again:
829 init_rss_vec(rss);
830
831 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
832 if (!dst_pte)
833 return -ENOMEM;
834 src_pte = pte_offset_map(src_pmd, addr);
835 src_ptl = pte_lockptr(src_mm, src_pmd);
836 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
837 orig_src_pte = src_pte;
838 orig_dst_pte = dst_pte;
839 arch_enter_lazy_mmu_mode();
840
841 do {
842 /*
843 * We are holding two locks at this point - either of them
844 * could generate latencies in another task on another CPU.
845 */
846 if (progress >= 32) {
847 progress = 0;
848 if (need_resched() ||
849 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
850 break;
851 }
852 if (pte_none(*src_pte)) {
853 progress++;
854 continue;
855 }
856 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
857 vma, addr, rss);
858 if (entry.val)
859 break;
860 progress += 8;
861 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
862
863 arch_leave_lazy_mmu_mode();
864 spin_unlock(src_ptl);
865 pte_unmap(orig_src_pte);
866 add_mm_rss_vec(dst_mm, rss);
867 pte_unmap_unlock(orig_dst_pte, dst_ptl);
868 cond_resched();
869
870 if (entry.val) {
871 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
872 return -ENOMEM;
873 progress = 0;
874 }
875 if (addr != end)
876 goto again;
877 return 0;
878}
879
880static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
881 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
882 unsigned long addr, unsigned long end)
883{
884 pmd_t *src_pmd, *dst_pmd;
885 unsigned long next;
886
887 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
888 if (!dst_pmd)
889 return -ENOMEM;
890 src_pmd = pmd_offset(src_pud, addr);
891 do {
892 next = pmd_addr_end(addr, end);
893 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
894 || pmd_devmap(*src_pmd)) {
895 int err;
896 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
897 err = copy_huge_pmd(dst_mm, src_mm,
898 dst_pmd, src_pmd, addr, vma);
899 if (err == -ENOMEM)
900 return -ENOMEM;
901 if (!err)
902 continue;
903 /* fall through */
904 }
905 if (pmd_none_or_clear_bad(src_pmd))
906 continue;
907 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
908 vma, addr, next))
909 return -ENOMEM;
910 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
911 return 0;
912}
913
914static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
915 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
916 unsigned long addr, unsigned long end)
917{
918 pud_t *src_pud, *dst_pud;
919 unsigned long next;
920
921 dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
922 if (!dst_pud)
923 return -ENOMEM;
924 src_pud = pud_offset(src_p4d, addr);
925 do {
926 next = pud_addr_end(addr, end);
927 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
928 int err;
929
930 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
931 err = copy_huge_pud(dst_mm, src_mm,
932 dst_pud, src_pud, addr, vma);
933 if (err == -ENOMEM)
934 return -ENOMEM;
935 if (!err)
936 continue;
937 /* fall through */
938 }
939 if (pud_none_or_clear_bad(src_pud))
940 continue;
941 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
942 vma, addr, next))
943 return -ENOMEM;
944 } while (dst_pud++, src_pud++, addr = next, addr != end);
945 return 0;
946}
947
948static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
949 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
950 unsigned long addr, unsigned long end)
951{
952 p4d_t *src_p4d, *dst_p4d;
953 unsigned long next;
954
955 dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
956 if (!dst_p4d)
957 return -ENOMEM;
958 src_p4d = p4d_offset(src_pgd, addr);
959 do {
960 next = p4d_addr_end(addr, end);
961 if (p4d_none_or_clear_bad(src_p4d))
962 continue;
963 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
964 vma, addr, next))
965 return -ENOMEM;
966 } while (dst_p4d++, src_p4d++, addr = next, addr != end);
967 return 0;
968}
969
970int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
971 struct vm_area_struct *vma)
972{
973 pgd_t *src_pgd, *dst_pgd;
974 unsigned long next;
975 unsigned long addr = vma->vm_start;
976 unsigned long end = vma->vm_end;
977 struct mmu_notifier_range range;
978 bool is_cow;
979 int ret;
980
981 /*
982 * Don't copy ptes where a page fault will fill them correctly.
983 * Fork becomes much lighter when there are big shared or private
984 * readonly mappings. The tradeoff is that copy_page_range is more
985 * efficient than faulting.
986 */
987 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
988 !vma->anon_vma)
989 return 0;
990
991 if (is_vm_hugetlb_page(vma))
992 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
993
994 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
995 /*
996 * We do not free on error cases below as remove_vma
997 * gets called on error from higher level routine
998 */
999 ret = track_pfn_copy(vma);
1000 if (ret)
1001 return ret;
1002 }
1003
1004 /*
1005 * We need to invalidate the secondary MMU mappings only when
1006 * there could be a permission downgrade on the ptes of the
1007 * parent mm. And a permission downgrade will only happen if
1008 * is_cow_mapping() returns true.
1009 */
1010 is_cow = is_cow_mapping(vma->vm_flags);
1011
1012 if (is_cow) {
1013 mmu_notifier_range_init(&range, src_mm, addr, end);
1014 mmu_notifier_invalidate_range_start(&range);
1015 }
1016
1017 ret = 0;
1018 dst_pgd = pgd_offset(dst_mm, addr);
1019 src_pgd = pgd_offset(src_mm, addr);
1020 do {
1021 next = pgd_addr_end(addr, end);
1022 if (pgd_none_or_clear_bad(src_pgd))
1023 continue;
1024 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1025 vma, addr, next))) {
1026 ret = -ENOMEM;
1027 break;
1028 }
1029 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1030
1031 if (is_cow)
1032 mmu_notifier_invalidate_range_end(&range);
1033 return ret;
1034}
1035
1036static unsigned long zap_pte_range(struct mmu_gather *tlb,
1037 struct vm_area_struct *vma, pmd_t *pmd,
1038 unsigned long addr, unsigned long end,
1039 struct zap_details *details)
1040{
1041 struct mm_struct *mm = tlb->mm;
1042 int force_flush = 0;
1043 int rss[NR_MM_COUNTERS];
1044 spinlock_t *ptl;
1045 pte_t *start_pte;
1046 pte_t *pte;
1047 swp_entry_t entry;
1048
1049 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1050again:
1051 init_rss_vec(rss);
1052 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1053 pte = start_pte;
1054 flush_tlb_batched_pending(mm);
1055 arch_enter_lazy_mmu_mode();
1056 do {
1057 pte_t ptent = *pte;
1058 if (pte_none(ptent))
1059 continue;
1060
1061 if (pte_present(ptent)) {
1062 struct page *page;
1063
1064 page = _vm_normal_page(vma, addr, ptent, true);
1065 if (unlikely(details) && page) {
1066 /*
1067 * unmap_shared_mapping_pages() wants to
1068 * invalidate cache without truncating:
1069 * unmap shared but keep private pages.
1070 */
1071 if (details->check_mapping &&
1072 details->check_mapping != page_rmapping(page))
1073 continue;
1074 }
1075 ptent = ptep_get_and_clear_full(mm, addr, pte,
1076 tlb->fullmm);
1077 tlb_remove_tlb_entry(tlb, pte, addr);
1078 if (unlikely(!page))
1079 continue;
1080
1081 if (!PageAnon(page)) {
1082 if (pte_dirty(ptent)) {
1083 force_flush = 1;
1084 set_page_dirty(page);
1085 }
1086 if (pte_young(ptent) &&
1087 likely(!(vma->vm_flags & VM_SEQ_READ)))
1088 mark_page_accessed(page);
1089 }
1090 rss[mm_counter(page)]--;
1091 page_remove_rmap(page, false);
1092 if (unlikely(page_mapcount(page) < 0))
1093 print_bad_pte(vma, addr, ptent, page);
1094 if (unlikely(__tlb_remove_page(tlb, page))) {
1095 force_flush = 1;
1096 addr += PAGE_SIZE;
1097 break;
1098 }
1099 continue;
1100 }
1101
1102 entry = pte_to_swp_entry(ptent);
1103 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1104 struct page *page = device_private_entry_to_page(entry);
1105
1106 if (unlikely(details && details->check_mapping)) {
1107 /*
1108 * unmap_shared_mapping_pages() wants to
1109 * invalidate cache without truncating:
1110 * unmap shared but keep private pages.
1111 */
1112 if (details->check_mapping !=
1113 page_rmapping(page))
1114 continue;
1115 }
1116
1117 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1118 rss[mm_counter(page)]--;
1119 page_remove_rmap(page, false);
1120 put_page(page);
1121 continue;
1122 }
1123
1124 /* If details->check_mapping, we leave swap entries. */
1125 if (unlikely(details))
1126 continue;
1127
1128 entry = pte_to_swp_entry(ptent);
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_free(tlb);
1159 if (addr != end)
1160 goto again;
1161 }
1162
1163 return addr;
1164}
1165
1166static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1167 struct vm_area_struct *vma, pud_t *pud,
1168 unsigned long addr, unsigned long end,
1169 struct zap_details *details)
1170{
1171 pmd_t *pmd;
1172 unsigned long next;
1173
1174 pmd = pmd_offset(pud, addr);
1175 do {
1176 next = pmd_addr_end(addr, end);
1177 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1178 if (next - addr != HPAGE_PMD_SIZE)
1179 __split_huge_pmd(vma, pmd, addr, false, NULL);
1180 else if (zap_huge_pmd(tlb, vma, pmd, addr))
1181 goto next;
1182 /* fall through */
1183 }
1184 /*
1185 * Here there can be other concurrent MADV_DONTNEED or
1186 * trans huge page faults running, and if the pmd is
1187 * none or trans huge it can change under us. This is
1188 * because MADV_DONTNEED holds the mmap_sem in read
1189 * mode.
1190 */
1191 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1192 goto next;
1193 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1194next:
1195 cond_resched();
1196 } while (pmd++, addr = next, addr != end);
1197
1198 return addr;
1199}
1200
1201static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1202 struct vm_area_struct *vma, p4d_t *p4d,
1203 unsigned long addr, unsigned long end,
1204 struct zap_details *details)
1205{
1206 pud_t *pud;
1207 unsigned long next;
1208
1209 pud = pud_offset(p4d, addr);
1210 do {
1211 next = pud_addr_end(addr, end);
1212 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1213 if (next - addr != HPAGE_PUD_SIZE) {
1214 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1215 split_huge_pud(vma, pud, addr);
1216 } else if (zap_huge_pud(tlb, vma, pud, addr))
1217 goto next;
1218 /* fall through */
1219 }
1220 if (pud_none_or_clear_bad(pud))
1221 continue;
1222 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1223next:
1224 cond_resched();
1225 } while (pud++, addr = next, addr != end);
1226
1227 return addr;
1228}
1229
1230static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1231 struct vm_area_struct *vma, pgd_t *pgd,
1232 unsigned long addr, unsigned long end,
1233 struct zap_details *details)
1234{
1235 p4d_t *p4d;
1236 unsigned long next;
1237
1238 p4d = p4d_offset(pgd, addr);
1239 do {
1240 next = p4d_addr_end(addr, end);
1241 if (p4d_none_or_clear_bad(p4d))
1242 continue;
1243 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1244 } while (p4d++, addr = next, addr != end);
1245
1246 return addr;
1247}
1248
1249void unmap_page_range(struct mmu_gather *tlb,
1250 struct vm_area_struct *vma,
1251 unsigned long addr, unsigned long end,
1252 struct zap_details *details)
1253{
1254 pgd_t *pgd;
1255 unsigned long next;
1256
1257 BUG_ON(addr >= end);
1258 tlb_start_vma(tlb, vma);
1259 pgd = pgd_offset(vma->vm_mm, addr);
1260 do {
1261 next = pgd_addr_end(addr, end);
1262 if (pgd_none_or_clear_bad(pgd))
1263 continue;
1264 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1265 } while (pgd++, addr = next, addr != end);
1266 tlb_end_vma(tlb, vma);
1267}
1268
1269
1270static void unmap_single_vma(struct mmu_gather *tlb,
1271 struct vm_area_struct *vma, unsigned long start_addr,
1272 unsigned long end_addr,
1273 struct zap_details *details)
1274{
1275 unsigned long start = max(vma->vm_start, start_addr);
1276 unsigned long end;
1277
1278 if (start >= vma->vm_end)
1279 return;
1280 end = min(vma->vm_end, end_addr);
1281 if (end <= vma->vm_start)
1282 return;
1283
1284 if (vma->vm_file)
1285 uprobe_munmap(vma, start, end);
1286
1287 if (unlikely(vma->vm_flags & VM_PFNMAP))
1288 untrack_pfn(vma, 0, 0);
1289
1290 if (start != end) {
1291 if (unlikely(is_vm_hugetlb_page(vma))) {
1292 /*
1293 * It is undesirable to test vma->vm_file as it
1294 * should be non-null for valid hugetlb area.
1295 * However, vm_file will be NULL in the error
1296 * cleanup path of mmap_region. When
1297 * hugetlbfs ->mmap method fails,
1298 * mmap_region() nullifies vma->vm_file
1299 * before calling this function to clean up.
1300 * Since no pte has actually been setup, it is
1301 * safe to do nothing in this case.
1302 */
1303 if (vma->vm_file) {
1304 i_mmap_lock_write(vma->vm_file->f_mapping);
1305 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1306 i_mmap_unlock_write(vma->vm_file->f_mapping);
1307 }
1308 } else
1309 unmap_page_range(tlb, vma, start, end, details);
1310 }
1311}
1312
1313/**
1314 * unmap_vmas - unmap a range of memory covered by a list of vma's
1315 * @tlb: address of the caller's struct mmu_gather
1316 * @vma: the starting vma
1317 * @start_addr: virtual address at which to start unmapping
1318 * @end_addr: virtual address at which to end unmapping
1319 *
1320 * Unmap all pages in the vma list.
1321 *
1322 * Only addresses between `start' and `end' will be unmapped.
1323 *
1324 * The VMA list must be sorted in ascending virtual address order.
1325 *
1326 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1327 * range after unmap_vmas() returns. So the only responsibility here is to
1328 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1329 * drops the lock and schedules.
1330 */
1331void unmap_vmas(struct mmu_gather *tlb,
1332 struct vm_area_struct *vma, unsigned long start_addr,
1333 unsigned long end_addr)
1334{
1335 struct mmu_notifier_range range;
1336
1337 mmu_notifier_range_init(&range, vma->vm_mm, start_addr, end_addr);
1338 mmu_notifier_invalidate_range_start(&range);
1339 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1340 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1341 mmu_notifier_invalidate_range_end(&range);
1342}
1343
1344/**
1345 * zap_page_range - remove user pages in a given range
1346 * @vma: vm_area_struct holding the applicable pages
1347 * @start: starting address of pages to zap
1348 * @size: number of bytes to zap
1349 *
1350 * Caller must protect the VMA list
1351 */
1352void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1353 unsigned long size)
1354{
1355 struct mmu_notifier_range range;
1356 struct mmu_gather tlb;
1357
1358 lru_add_drain();
1359 mmu_notifier_range_init(&range, vma->vm_mm, start, start + size);
1360 tlb_gather_mmu(&tlb, vma->vm_mm, start, range.end);
1361 update_hiwater_rss(vma->vm_mm);
1362 mmu_notifier_invalidate_range_start(&range);
1363 for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1364 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1365 mmu_notifier_invalidate_range_end(&range);
1366 tlb_finish_mmu(&tlb, start, range.end);
1367}
1368
1369/**
1370 * zap_page_range_single - remove user pages in a given range
1371 * @vma: vm_area_struct holding the applicable pages
1372 * @address: starting address of pages to zap
1373 * @size: number of bytes to zap
1374 * @details: details of shared cache invalidation
1375 *
1376 * The range must fit into one VMA.
1377 */
1378static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1379 unsigned long size, struct zap_details *details)
1380{
1381 struct mmu_notifier_range range;
1382 struct mmu_gather tlb;
1383
1384 lru_add_drain();
1385 mmu_notifier_range_init(&range, vma->vm_mm, address, address + size);
1386 tlb_gather_mmu(&tlb, vma->vm_mm, address, range.end);
1387 update_hiwater_rss(vma->vm_mm);
1388 mmu_notifier_invalidate_range_start(&range);
1389 unmap_single_vma(&tlb, vma, address, range.end, details);
1390 mmu_notifier_invalidate_range_end(&range);
1391 tlb_finish_mmu(&tlb, address, range.end);
1392}
1393
1394/**
1395 * zap_vma_ptes - remove ptes mapping the vma
1396 * @vma: vm_area_struct holding ptes to be zapped
1397 * @address: starting address of pages to zap
1398 * @size: number of bytes to zap
1399 *
1400 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1401 *
1402 * The entire address range must be fully contained within the vma.
1403 *
1404 */
1405void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1406 unsigned long size)
1407{
1408 if (address < vma->vm_start || address + size > vma->vm_end ||
1409 !(vma->vm_flags & VM_PFNMAP))
1410 return;
1411
1412 zap_page_range_single(vma, address, size, NULL);
1413}
1414EXPORT_SYMBOL_GPL(zap_vma_ptes);
1415
1416pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1417 spinlock_t **ptl)
1418{
1419 pgd_t *pgd;
1420 p4d_t *p4d;
1421 pud_t *pud;
1422 pmd_t *pmd;
1423
1424 pgd = pgd_offset(mm, addr);
1425 p4d = p4d_alloc(mm, pgd, addr);
1426 if (!p4d)
1427 return NULL;
1428 pud = pud_alloc(mm, p4d, addr);
1429 if (!pud)
1430 return NULL;
1431 pmd = pmd_alloc(mm, pud, addr);
1432 if (!pmd)
1433 return NULL;
1434
1435 VM_BUG_ON(pmd_trans_huge(*pmd));
1436 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1437}
1438
1439/*
1440 * This is the old fallback for page remapping.
1441 *
1442 * For historical reasons, it only allows reserved pages. Only
1443 * old drivers should use this, and they needed to mark their
1444 * pages reserved for the old functions anyway.
1445 */
1446static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1447 struct page *page, pgprot_t prot)
1448{
1449 struct mm_struct *mm = vma->vm_mm;
1450 int retval;
1451 pte_t *pte;
1452 spinlock_t *ptl;
1453
1454 retval = -EINVAL;
1455 if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1456 goto out;
1457 retval = -ENOMEM;
1458 flush_dcache_page(page);
1459 pte = get_locked_pte(mm, addr, &ptl);
1460 if (!pte)
1461 goto out;
1462 retval = -EBUSY;
1463 if (!pte_none(*pte))
1464 goto out_unlock;
1465
1466 /* Ok, finally just insert the thing.. */
1467 get_page(page);
1468 inc_mm_counter_fast(mm, mm_counter_file(page));
1469 page_add_file_rmap(page, false);
1470 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1471
1472 retval = 0;
1473 pte_unmap_unlock(pte, ptl);
1474 return retval;
1475out_unlock:
1476 pte_unmap_unlock(pte, ptl);
1477out:
1478 return retval;
1479}
1480
1481/**
1482 * vm_insert_page - insert single page into user vma
1483 * @vma: user vma to map to
1484 * @addr: target user address of this page
1485 * @page: source kernel page
1486 *
1487 * This allows drivers to insert individual pages they've allocated
1488 * into a user vma.
1489 *
1490 * The page has to be a nice clean _individual_ kernel allocation.
1491 * If you allocate a compound page, you need to have marked it as
1492 * such (__GFP_COMP), or manually just split the page up yourself
1493 * (see split_page()).
1494 *
1495 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1496 * took an arbitrary page protection parameter. This doesn't allow
1497 * that. Your vma protection will have to be set up correctly, which
1498 * means that if you want a shared writable mapping, you'd better
1499 * ask for a shared writable mapping!
1500 *
1501 * The page does not need to be reserved.
1502 *
1503 * Usually this function is called from f_op->mmap() handler
1504 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1505 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1506 * function from other places, for example from page-fault handler.
1507 *
1508 * Return: %0 on success, negative error code otherwise.
1509 */
1510int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1511 struct page *page)
1512{
1513 if (addr < vma->vm_start || addr >= vma->vm_end)
1514 return -EFAULT;
1515 if (!page_count(page))
1516 return -EINVAL;
1517 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1518 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1519 BUG_ON(vma->vm_flags & VM_PFNMAP);
1520 vma->vm_flags |= VM_MIXEDMAP;
1521 }
1522 return insert_page(vma, addr, page, vma->vm_page_prot);
1523}
1524EXPORT_SYMBOL(vm_insert_page);
1525
1526static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1527 pfn_t pfn, pgprot_t prot, bool mkwrite)
1528{
1529 struct mm_struct *mm = vma->vm_mm;
1530 pte_t *pte, entry;
1531 spinlock_t *ptl;
1532
1533 pte = get_locked_pte(mm, addr, &ptl);
1534 if (!pte)
1535 return VM_FAULT_OOM;
1536 if (!pte_none(*pte)) {
1537 if (mkwrite) {
1538 /*
1539 * For read faults on private mappings the PFN passed
1540 * in may not match the PFN we have mapped if the
1541 * mapped PFN is a writeable COW page. In the mkwrite
1542 * case we are creating a writable PTE for a shared
1543 * mapping and we expect the PFNs to match. If they
1544 * don't match, we are likely racing with block
1545 * allocation and mapping invalidation so just skip the
1546 * update.
1547 */
1548 if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1549 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1550 goto out_unlock;
1551 }
1552 entry = *pte;
1553 goto out_mkwrite;
1554 } else
1555 goto out_unlock;
1556 }
1557
1558 /* Ok, finally just insert the thing.. */
1559 if (pfn_t_devmap(pfn))
1560 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1561 else
1562 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1563
1564out_mkwrite:
1565 if (mkwrite) {
1566 entry = pte_mkyoung(entry);
1567 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1568 }
1569
1570 set_pte_at(mm, addr, pte, entry);
1571 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1572
1573out_unlock:
1574 pte_unmap_unlock(pte, ptl);
1575 return VM_FAULT_NOPAGE;
1576}
1577
1578/**
1579 * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1580 * @vma: user vma to map to
1581 * @addr: target user address of this page
1582 * @pfn: source kernel pfn
1583 * @pgprot: pgprot flags for the inserted page
1584 *
1585 * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1586 * to override pgprot on a per-page basis.
1587 *
1588 * This only makes sense for IO mappings, and it makes no sense for
1589 * COW mappings. In general, using multiple vmas is preferable;
1590 * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1591 * impractical.
1592 *
1593 * Context: Process context. May allocate using %GFP_KERNEL.
1594 * Return: vm_fault_t value.
1595 */
1596vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1597 unsigned long pfn, pgprot_t pgprot)
1598{
1599 /*
1600 * Technically, architectures with pte_special can avoid all these
1601 * restrictions (same for remap_pfn_range). However we would like
1602 * consistency in testing and feature parity among all, so we should
1603 * try to keep these invariants in place for everybody.
1604 */
1605 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1606 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1607 (VM_PFNMAP|VM_MIXEDMAP));
1608 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1609 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1610
1611 if (addr < vma->vm_start || addr >= vma->vm_end)
1612 return VM_FAULT_SIGBUS;
1613
1614 if (!pfn_modify_allowed(pfn, pgprot))
1615 return VM_FAULT_SIGBUS;
1616
1617 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1618
1619 return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1620 false);
1621}
1622EXPORT_SYMBOL(vmf_insert_pfn_prot);
1623
1624/**
1625 * vmf_insert_pfn - insert single pfn into user vma
1626 * @vma: user vma to map to
1627 * @addr: target user address of this page
1628 * @pfn: source kernel pfn
1629 *
1630 * Similar to vm_insert_page, this allows drivers to insert individual pages
1631 * they've allocated into a user vma. Same comments apply.
1632 *
1633 * This function should only be called from a vm_ops->fault handler, and
1634 * in that case the handler should return the result of this function.
1635 *
1636 * vma cannot be a COW mapping.
1637 *
1638 * As this is called only for pages that do not currently exist, we
1639 * do not need to flush old virtual caches or the TLB.
1640 *
1641 * Context: Process context. May allocate using %GFP_KERNEL.
1642 * Return: vm_fault_t value.
1643 */
1644vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1645 unsigned long pfn)
1646{
1647 return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1648}
1649EXPORT_SYMBOL(vmf_insert_pfn);
1650
1651static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1652{
1653 /* these checks mirror the abort conditions in vm_normal_page */
1654 if (vma->vm_flags & VM_MIXEDMAP)
1655 return true;
1656 if (pfn_t_devmap(pfn))
1657 return true;
1658 if (pfn_t_special(pfn))
1659 return true;
1660 if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1661 return true;
1662 return false;
1663}
1664
1665static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1666 unsigned long addr, pfn_t pfn, bool mkwrite)
1667{
1668 pgprot_t pgprot = vma->vm_page_prot;
1669 int err;
1670
1671 BUG_ON(!vm_mixed_ok(vma, pfn));
1672
1673 if (addr < vma->vm_start || addr >= vma->vm_end)
1674 return VM_FAULT_SIGBUS;
1675
1676 track_pfn_insert(vma, &pgprot, pfn);
1677
1678 if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1679 return VM_FAULT_SIGBUS;
1680
1681 /*
1682 * If we don't have pte special, then we have to use the pfn_valid()
1683 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1684 * refcount the page if pfn_valid is true (hence insert_page rather
1685 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1686 * without pte special, it would there be refcounted as a normal page.
1687 */
1688 if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1689 !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1690 struct page *page;
1691
1692 /*
1693 * At this point we are committed to insert_page()
1694 * regardless of whether the caller specified flags that
1695 * result in pfn_t_has_page() == false.
1696 */
1697 page = pfn_to_page(pfn_t_to_pfn(pfn));
1698 err = insert_page(vma, addr, page, pgprot);
1699 } else {
1700 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1701 }
1702
1703 if (err == -ENOMEM)
1704 return VM_FAULT_OOM;
1705 if (err < 0 && err != -EBUSY)
1706 return VM_FAULT_SIGBUS;
1707
1708 return VM_FAULT_NOPAGE;
1709}
1710
1711vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1712 pfn_t pfn)
1713{
1714 return __vm_insert_mixed(vma, addr, pfn, false);
1715}
1716EXPORT_SYMBOL(vmf_insert_mixed);
1717
1718/*
1719 * If the insertion of PTE failed because someone else already added a
1720 * different entry in the mean time, we treat that as success as we assume
1721 * the same entry was actually inserted.
1722 */
1723vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1724 unsigned long addr, pfn_t pfn)
1725{
1726 return __vm_insert_mixed(vma, addr, pfn, true);
1727}
1728EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1729
1730/*
1731 * maps a range of physical memory into the requested pages. the old
1732 * mappings are removed. any references to nonexistent pages results
1733 * in null mappings (currently treated as "copy-on-access")
1734 */
1735static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1736 unsigned long addr, unsigned long end,
1737 unsigned long pfn, pgprot_t prot)
1738{
1739 pte_t *pte;
1740 spinlock_t *ptl;
1741 int err = 0;
1742
1743 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1744 if (!pte)
1745 return -ENOMEM;
1746 arch_enter_lazy_mmu_mode();
1747 do {
1748 BUG_ON(!pte_none(*pte));
1749 if (!pfn_modify_allowed(pfn, prot)) {
1750 err = -EACCES;
1751 break;
1752 }
1753 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1754 pfn++;
1755 } while (pte++, addr += PAGE_SIZE, addr != end);
1756 arch_leave_lazy_mmu_mode();
1757 pte_unmap_unlock(pte - 1, ptl);
1758 return err;
1759}
1760
1761static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1762 unsigned long addr, unsigned long end,
1763 unsigned long pfn, pgprot_t prot)
1764{
1765 pmd_t *pmd;
1766 unsigned long next;
1767 int err;
1768
1769 pfn -= addr >> PAGE_SHIFT;
1770 pmd = pmd_alloc(mm, pud, addr);
1771 if (!pmd)
1772 return -ENOMEM;
1773 VM_BUG_ON(pmd_trans_huge(*pmd));
1774 do {
1775 next = pmd_addr_end(addr, end);
1776 err = remap_pte_range(mm, pmd, addr, next,
1777 pfn + (addr >> PAGE_SHIFT), prot);
1778 if (err)
1779 return err;
1780 } while (pmd++, addr = next, addr != end);
1781 return 0;
1782}
1783
1784static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1785 unsigned long addr, unsigned long end,
1786 unsigned long pfn, pgprot_t prot)
1787{
1788 pud_t *pud;
1789 unsigned long next;
1790 int err;
1791
1792 pfn -= addr >> PAGE_SHIFT;
1793 pud = pud_alloc(mm, p4d, addr);
1794 if (!pud)
1795 return -ENOMEM;
1796 do {
1797 next = pud_addr_end(addr, end);
1798 err = remap_pmd_range(mm, pud, addr, next,
1799 pfn + (addr >> PAGE_SHIFT), prot);
1800 if (err)
1801 return err;
1802 } while (pud++, addr = next, addr != end);
1803 return 0;
1804}
1805
1806static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1807 unsigned long addr, unsigned long end,
1808 unsigned long pfn, pgprot_t prot)
1809{
1810 p4d_t *p4d;
1811 unsigned long next;
1812 int err;
1813
1814 pfn -= addr >> PAGE_SHIFT;
1815 p4d = p4d_alloc(mm, pgd, addr);
1816 if (!p4d)
1817 return -ENOMEM;
1818 do {
1819 next = p4d_addr_end(addr, end);
1820 err = remap_pud_range(mm, p4d, addr, next,
1821 pfn + (addr >> PAGE_SHIFT), prot);
1822 if (err)
1823 return err;
1824 } while (p4d++, addr = next, addr != end);
1825 return 0;
1826}
1827
1828/**
1829 * remap_pfn_range - remap kernel memory to userspace
1830 * @vma: user vma to map to
1831 * @addr: target user address to start at
1832 * @pfn: physical address of kernel memory
1833 * @size: size of map area
1834 * @prot: page protection flags for this mapping
1835 *
1836 * Note: this is only safe if the mm semaphore is held when called.
1837 *
1838 * Return: %0 on success, negative error code otherwise.
1839 */
1840int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1841 unsigned long pfn, unsigned long size, pgprot_t prot)
1842{
1843 pgd_t *pgd;
1844 unsigned long next;
1845 unsigned long end = addr + PAGE_ALIGN(size);
1846 struct mm_struct *mm = vma->vm_mm;
1847 unsigned long remap_pfn = pfn;
1848 int err;
1849
1850 /*
1851 * Physically remapped pages are special. Tell the
1852 * rest of the world about it:
1853 * VM_IO tells people not to look at these pages
1854 * (accesses can have side effects).
1855 * VM_PFNMAP tells the core MM that the base pages are just
1856 * raw PFN mappings, and do not have a "struct page" associated
1857 * with them.
1858 * VM_DONTEXPAND
1859 * Disable vma merging and expanding with mremap().
1860 * VM_DONTDUMP
1861 * Omit vma from core dump, even when VM_IO turned off.
1862 *
1863 * There's a horrible special case to handle copy-on-write
1864 * behaviour that some programs depend on. We mark the "original"
1865 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1866 * See vm_normal_page() for details.
1867 */
1868 if (is_cow_mapping(vma->vm_flags)) {
1869 if (addr != vma->vm_start || end != vma->vm_end)
1870 return -EINVAL;
1871 vma->vm_pgoff = pfn;
1872 }
1873
1874 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1875 if (err)
1876 return -EINVAL;
1877
1878 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1879
1880 BUG_ON(addr >= end);
1881 pfn -= addr >> PAGE_SHIFT;
1882 pgd = pgd_offset(mm, addr);
1883 flush_cache_range(vma, addr, end);
1884 do {
1885 next = pgd_addr_end(addr, end);
1886 err = remap_p4d_range(mm, pgd, addr, next,
1887 pfn + (addr >> PAGE_SHIFT), prot);
1888 if (err)
1889 break;
1890 } while (pgd++, addr = next, addr != end);
1891
1892 if (err)
1893 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1894
1895 return err;
1896}
1897EXPORT_SYMBOL(remap_pfn_range);
1898
1899/**
1900 * vm_iomap_memory - remap memory to userspace
1901 * @vma: user vma to map to
1902 * @start: start of area
1903 * @len: size of area
1904 *
1905 * This is a simplified io_remap_pfn_range() for common driver use. The
1906 * driver just needs to give us the physical memory range to be mapped,
1907 * we'll figure out the rest from the vma information.
1908 *
1909 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1910 * whatever write-combining details or similar.
1911 *
1912 * Return: %0 on success, negative error code otherwise.
1913 */
1914int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1915{
1916 unsigned long vm_len, pfn, pages;
1917
1918 /* Check that the physical memory area passed in looks valid */
1919 if (start + len < start)
1920 return -EINVAL;
1921 /*
1922 * You *really* shouldn't map things that aren't page-aligned,
1923 * but we've historically allowed it because IO memory might
1924 * just have smaller alignment.
1925 */
1926 len += start & ~PAGE_MASK;
1927 pfn = start >> PAGE_SHIFT;
1928 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1929 if (pfn + pages < pfn)
1930 return -EINVAL;
1931
1932 /* We start the mapping 'vm_pgoff' pages into the area */
1933 if (vma->vm_pgoff > pages)
1934 return -EINVAL;
1935 pfn += vma->vm_pgoff;
1936 pages -= vma->vm_pgoff;
1937
1938 /* Can we fit all of the mapping? */
1939 vm_len = vma->vm_end - vma->vm_start;
1940 if (vm_len >> PAGE_SHIFT > pages)
1941 return -EINVAL;
1942
1943 /* Ok, let it rip */
1944 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1945}
1946EXPORT_SYMBOL(vm_iomap_memory);
1947
1948static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1949 unsigned long addr, unsigned long end,
1950 pte_fn_t fn, void *data)
1951{
1952 pte_t *pte;
1953 int err;
1954 pgtable_t token;
1955 spinlock_t *uninitialized_var(ptl);
1956
1957 pte = (mm == &init_mm) ?
1958 pte_alloc_kernel(pmd, addr) :
1959 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1960 if (!pte)
1961 return -ENOMEM;
1962
1963 BUG_ON(pmd_huge(*pmd));
1964
1965 arch_enter_lazy_mmu_mode();
1966
1967 token = pmd_pgtable(*pmd);
1968
1969 do {
1970 err = fn(pte++, token, addr, data);
1971 if (err)
1972 break;
1973 } while (addr += PAGE_SIZE, addr != end);
1974
1975 arch_leave_lazy_mmu_mode();
1976
1977 if (mm != &init_mm)
1978 pte_unmap_unlock(pte-1, ptl);
1979 return err;
1980}
1981
1982static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1983 unsigned long addr, unsigned long end,
1984 pte_fn_t fn, void *data)
1985{
1986 pmd_t *pmd;
1987 unsigned long next;
1988 int err;
1989
1990 BUG_ON(pud_huge(*pud));
1991
1992 pmd = pmd_alloc(mm, pud, addr);
1993 if (!pmd)
1994 return -ENOMEM;
1995 do {
1996 next = pmd_addr_end(addr, end);
1997 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1998 if (err)
1999 break;
2000 } while (pmd++, addr = next, addr != end);
2001 return err;
2002}
2003
2004static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2005 unsigned long addr, unsigned long end,
2006 pte_fn_t fn, void *data)
2007{
2008 pud_t *pud;
2009 unsigned long next;
2010 int err;
2011
2012 pud = pud_alloc(mm, p4d, addr);
2013 if (!pud)
2014 return -ENOMEM;
2015 do {
2016 next = pud_addr_end(addr, end);
2017 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2018 if (err)
2019 break;
2020 } while (pud++, addr = next, addr != end);
2021 return err;
2022}
2023
2024static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2025 unsigned long addr, unsigned long end,
2026 pte_fn_t fn, void *data)
2027{
2028 p4d_t *p4d;
2029 unsigned long next;
2030 int err;
2031
2032 p4d = p4d_alloc(mm, pgd, addr);
2033 if (!p4d)
2034 return -ENOMEM;
2035 do {
2036 next = p4d_addr_end(addr, end);
2037 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2038 if (err)
2039 break;
2040 } while (p4d++, addr = next, addr != end);
2041 return err;
2042}
2043
2044/*
2045 * Scan a region of virtual memory, filling in page tables as necessary
2046 * and calling a provided function on each leaf page table.
2047 */
2048int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2049 unsigned long size, pte_fn_t fn, void *data)
2050{
2051 pgd_t *pgd;
2052 unsigned long next;
2053 unsigned long end = addr + size;
2054 int err;
2055
2056 if (WARN_ON(addr >= end))
2057 return -EINVAL;
2058
2059 pgd = pgd_offset(mm, addr);
2060 do {
2061 next = pgd_addr_end(addr, end);
2062 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2063 if (err)
2064 break;
2065 } while (pgd++, addr = next, addr != end);
2066
2067 return err;
2068}
2069EXPORT_SYMBOL_GPL(apply_to_page_range);
2070
2071/*
2072 * handle_pte_fault chooses page fault handler according to an entry which was
2073 * read non-atomically. Before making any commitment, on those architectures
2074 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2075 * parts, do_swap_page must check under lock before unmapping the pte and
2076 * proceeding (but do_wp_page is only called after already making such a check;
2077 * and do_anonymous_page can safely check later on).
2078 */
2079static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2080 pte_t *page_table, pte_t orig_pte)
2081{
2082 int same = 1;
2083#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2084 if (sizeof(pte_t) > sizeof(unsigned long)) {
2085 spinlock_t *ptl = pte_lockptr(mm, pmd);
2086 spin_lock(ptl);
2087 same = pte_same(*page_table, orig_pte);
2088 spin_unlock(ptl);
2089 }
2090#endif
2091 pte_unmap(page_table);
2092 return same;
2093}
2094
2095static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2096{
2097 debug_dma_assert_idle(src);
2098
2099 /*
2100 * If the source page was a PFN mapping, we don't have
2101 * a "struct page" for it. We do a best-effort copy by
2102 * just copying from the original user address. If that
2103 * fails, we just zero-fill it. Live with it.
2104 */
2105 if (unlikely(!src)) {
2106 void *kaddr = kmap_atomic(dst);
2107 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2108
2109 /*
2110 * This really shouldn't fail, because the page is there
2111 * in the page tables. But it might just be unreadable,
2112 * in which case we just give up and fill the result with
2113 * zeroes.
2114 */
2115 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2116 clear_page(kaddr);
2117 kunmap_atomic(kaddr);
2118 flush_dcache_page(dst);
2119 } else
2120 copy_user_highpage(dst, src, va, vma);
2121}
2122
2123static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2124{
2125 struct file *vm_file = vma->vm_file;
2126
2127 if (vm_file)
2128 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2129
2130 /*
2131 * Special mappings (e.g. VDSO) do not have any file so fake
2132 * a default GFP_KERNEL for them.
2133 */
2134 return GFP_KERNEL;
2135}
2136
2137/*
2138 * Notify the address space that the page is about to become writable so that
2139 * it can prohibit this or wait for the page to get into an appropriate state.
2140 *
2141 * We do this without the lock held, so that it can sleep if it needs to.
2142 */
2143static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2144{
2145 vm_fault_t ret;
2146 struct page *page = vmf->page;
2147 unsigned int old_flags = vmf->flags;
2148
2149 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2150
2151 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2152 /* Restore original flags so that caller is not surprised */
2153 vmf->flags = old_flags;
2154 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2155 return ret;
2156 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2157 lock_page(page);
2158 if (!page->mapping) {
2159 unlock_page(page);
2160 return 0; /* retry */
2161 }
2162 ret |= VM_FAULT_LOCKED;
2163 } else
2164 VM_BUG_ON_PAGE(!PageLocked(page), page);
2165 return ret;
2166}
2167
2168/*
2169 * Handle dirtying of a page in shared file mapping on a write fault.
2170 *
2171 * The function expects the page to be locked and unlocks it.
2172 */
2173static void fault_dirty_shared_page(struct vm_area_struct *vma,
2174 struct page *page)
2175{
2176 struct address_space *mapping;
2177 bool dirtied;
2178 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2179
2180 dirtied = set_page_dirty(page);
2181 VM_BUG_ON_PAGE(PageAnon(page), page);
2182 /*
2183 * Take a local copy of the address_space - page.mapping may be zeroed
2184 * by truncate after unlock_page(). The address_space itself remains
2185 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2186 * release semantics to prevent the compiler from undoing this copying.
2187 */
2188 mapping = page_rmapping(page);
2189 unlock_page(page);
2190
2191 if ((dirtied || page_mkwrite) && mapping) {
2192 /*
2193 * Some device drivers do not set page.mapping
2194 * but still dirty their pages
2195 */
2196 balance_dirty_pages_ratelimited(mapping);
2197 }
2198
2199 if (!page_mkwrite)
2200 file_update_time(vma->vm_file);
2201}
2202
2203/*
2204 * Handle write page faults for pages that can be reused in the current vma
2205 *
2206 * This can happen either due to the mapping being with the VM_SHARED flag,
2207 * or due to us being the last reference standing to the page. In either
2208 * case, all we need to do here is to mark the page as writable and update
2209 * any related book-keeping.
2210 */
2211static inline void wp_page_reuse(struct vm_fault *vmf)
2212 __releases(vmf->ptl)
2213{
2214 struct vm_area_struct *vma = vmf->vma;
2215 struct page *page = vmf->page;
2216 pte_t entry;
2217 /*
2218 * Clear the pages cpupid information as the existing
2219 * information potentially belongs to a now completely
2220 * unrelated process.
2221 */
2222 if (page)
2223 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2224
2225 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2226 entry = pte_mkyoung(vmf->orig_pte);
2227 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2228 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2229 update_mmu_cache(vma, vmf->address, vmf->pte);
2230 pte_unmap_unlock(vmf->pte, vmf->ptl);
2231}
2232
2233/*
2234 * Handle the case of a page which we actually need to copy to a new page.
2235 *
2236 * Called with mmap_sem locked and the old page referenced, but
2237 * without the ptl held.
2238 *
2239 * High level logic flow:
2240 *
2241 * - Allocate a page, copy the content of the old page to the new one.
2242 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2243 * - Take the PTL. If the pte changed, bail out and release the allocated page
2244 * - If the pte is still the way we remember it, update the page table and all
2245 * relevant references. This includes dropping the reference the page-table
2246 * held to the old page, as well as updating the rmap.
2247 * - In any case, unlock the PTL and drop the reference we took to the old page.
2248 */
2249static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2250{
2251 struct vm_area_struct *vma = vmf->vma;
2252 struct mm_struct *mm = vma->vm_mm;
2253 struct page *old_page = vmf->page;
2254 struct page *new_page = NULL;
2255 pte_t entry;
2256 int page_copied = 0;
2257 struct mem_cgroup *memcg;
2258 struct mmu_notifier_range range;
2259
2260 if (unlikely(anon_vma_prepare(vma)))
2261 goto oom;
2262
2263 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2264 new_page = alloc_zeroed_user_highpage_movable(vma,
2265 vmf->address);
2266 if (!new_page)
2267 goto oom;
2268 } else {
2269 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2270 vmf->address);
2271 if (!new_page)
2272 goto oom;
2273 cow_user_page(new_page, old_page, vmf->address, vma);
2274 }
2275
2276 if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2277 goto oom_free_new;
2278
2279 __SetPageUptodate(new_page);
2280
2281 mmu_notifier_range_init(&range, mm, vmf->address & PAGE_MASK,
2282 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2283 mmu_notifier_invalidate_range_start(&range);
2284
2285 /*
2286 * Re-check the pte - we dropped the lock
2287 */
2288 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2289 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2290 if (old_page) {
2291 if (!PageAnon(old_page)) {
2292 dec_mm_counter_fast(mm,
2293 mm_counter_file(old_page));
2294 inc_mm_counter_fast(mm, MM_ANONPAGES);
2295 }
2296 } else {
2297 inc_mm_counter_fast(mm, MM_ANONPAGES);
2298 }
2299 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2300 entry = mk_pte(new_page, vma->vm_page_prot);
2301 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2302 /*
2303 * Clear the pte entry and flush it first, before updating the
2304 * pte with the new entry. This will avoid a race condition
2305 * seen in the presence of one thread doing SMC and another
2306 * thread doing COW.
2307 */
2308 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2309 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2310 mem_cgroup_commit_charge(new_page, memcg, false, false);
2311 lru_cache_add_active_or_unevictable(new_page, vma);
2312 /*
2313 * We call the notify macro here because, when using secondary
2314 * mmu page tables (such as kvm shadow page tables), we want the
2315 * new page to be mapped directly into the secondary page table.
2316 */
2317 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2318 update_mmu_cache(vma, vmf->address, vmf->pte);
2319 if (old_page) {
2320 /*
2321 * Only after switching the pte to the new page may
2322 * we remove the mapcount here. Otherwise another
2323 * process may come and find the rmap count decremented
2324 * before the pte is switched to the new page, and
2325 * "reuse" the old page writing into it while our pte
2326 * here still points into it and can be read by other
2327 * threads.
2328 *
2329 * The critical issue is to order this
2330 * page_remove_rmap with the ptp_clear_flush above.
2331 * Those stores are ordered by (if nothing else,)
2332 * the barrier present in the atomic_add_negative
2333 * in page_remove_rmap.
2334 *
2335 * Then the TLB flush in ptep_clear_flush ensures that
2336 * no process can access the old page before the
2337 * decremented mapcount is visible. And the old page
2338 * cannot be reused until after the decremented
2339 * mapcount is visible. So transitively, TLBs to
2340 * old page will be flushed before it can be reused.
2341 */
2342 page_remove_rmap(old_page, false);
2343 }
2344
2345 /* Free the old page.. */
2346 new_page = old_page;
2347 page_copied = 1;
2348 } else {
2349 mem_cgroup_cancel_charge(new_page, memcg, false);
2350 }
2351
2352 if (new_page)
2353 put_page(new_page);
2354
2355 pte_unmap_unlock(vmf->pte, vmf->ptl);
2356 /*
2357 * No need to double call mmu_notifier->invalidate_range() callback as
2358 * the above ptep_clear_flush_notify() did already call it.
2359 */
2360 mmu_notifier_invalidate_range_only_end(&range);
2361 if (old_page) {
2362 /*
2363 * Don't let another task, with possibly unlocked vma,
2364 * keep the mlocked page.
2365 */
2366 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2367 lock_page(old_page); /* LRU manipulation */
2368 if (PageMlocked(old_page))
2369 munlock_vma_page(old_page);
2370 unlock_page(old_page);
2371 }
2372 put_page(old_page);
2373 }
2374 return page_copied ? VM_FAULT_WRITE : 0;
2375oom_free_new:
2376 put_page(new_page);
2377oom:
2378 if (old_page)
2379 put_page(old_page);
2380 return VM_FAULT_OOM;
2381}
2382
2383/**
2384 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2385 * writeable once the page is prepared
2386 *
2387 * @vmf: structure describing the fault
2388 *
2389 * This function handles all that is needed to finish a write page fault in a
2390 * shared mapping due to PTE being read-only once the mapped page is prepared.
2391 * It handles locking of PTE and modifying it.
2392 *
2393 * The function expects the page to be locked or other protection against
2394 * concurrent faults / writeback (such as DAX radix tree locks).
2395 *
2396 * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2397 * we acquired PTE lock.
2398 */
2399vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2400{
2401 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2402 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2403 &vmf->ptl);
2404 /*
2405 * We might have raced with another page fault while we released the
2406 * pte_offset_map_lock.
2407 */
2408 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2409 pte_unmap_unlock(vmf->pte, vmf->ptl);
2410 return VM_FAULT_NOPAGE;
2411 }
2412 wp_page_reuse(vmf);
2413 return 0;
2414}
2415
2416/*
2417 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2418 * mapping
2419 */
2420static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2421{
2422 struct vm_area_struct *vma = vmf->vma;
2423
2424 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2425 vm_fault_t ret;
2426
2427 pte_unmap_unlock(vmf->pte, vmf->ptl);
2428 vmf->flags |= FAULT_FLAG_MKWRITE;
2429 ret = vma->vm_ops->pfn_mkwrite(vmf);
2430 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2431 return ret;
2432 return finish_mkwrite_fault(vmf);
2433 }
2434 wp_page_reuse(vmf);
2435 return VM_FAULT_WRITE;
2436}
2437
2438static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2439 __releases(vmf->ptl)
2440{
2441 struct vm_area_struct *vma = vmf->vma;
2442
2443 get_page(vmf->page);
2444
2445 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2446 vm_fault_t tmp;
2447
2448 pte_unmap_unlock(vmf->pte, vmf->ptl);
2449 tmp = do_page_mkwrite(vmf);
2450 if (unlikely(!tmp || (tmp &
2451 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2452 put_page(vmf->page);
2453 return tmp;
2454 }
2455 tmp = finish_mkwrite_fault(vmf);
2456 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2457 unlock_page(vmf->page);
2458 put_page(vmf->page);
2459 return tmp;
2460 }
2461 } else {
2462 wp_page_reuse(vmf);
2463 lock_page(vmf->page);
2464 }
2465 fault_dirty_shared_page(vma, vmf->page);
2466 put_page(vmf->page);
2467
2468 return VM_FAULT_WRITE;
2469}
2470
2471/*
2472 * This routine handles present pages, when users try to write
2473 * to a shared page. It is done by copying the page to a new address
2474 * and decrementing the shared-page counter for the old page.
2475 *
2476 * Note that this routine assumes that the protection checks have been
2477 * done by the caller (the low-level page fault routine in most cases).
2478 * Thus we can safely just mark it writable once we've done any necessary
2479 * COW.
2480 *
2481 * We also mark the page dirty at this point even though the page will
2482 * change only once the write actually happens. This avoids a few races,
2483 * and potentially makes it more efficient.
2484 *
2485 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2486 * but allow concurrent faults), with pte both mapped and locked.
2487 * We return with mmap_sem still held, but pte unmapped and unlocked.
2488 */
2489static vm_fault_t do_wp_page(struct vm_fault *vmf)
2490 __releases(vmf->ptl)
2491{
2492 struct vm_area_struct *vma = vmf->vma;
2493
2494 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2495 if (!vmf->page) {
2496 /*
2497 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2498 * VM_PFNMAP VMA.
2499 *
2500 * We should not cow pages in a shared writeable mapping.
2501 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2502 */
2503 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2504 (VM_WRITE|VM_SHARED))
2505 return wp_pfn_shared(vmf);
2506
2507 pte_unmap_unlock(vmf->pte, vmf->ptl);
2508 return wp_page_copy(vmf);
2509 }
2510
2511 /*
2512 * Take out anonymous pages first, anonymous shared vmas are
2513 * not dirty accountable.
2514 */
2515 if (PageAnon(vmf->page)) {
2516 int total_map_swapcount;
2517 if (PageKsm(vmf->page) && (PageSwapCache(vmf->page) ||
2518 page_count(vmf->page) != 1))
2519 goto copy;
2520 if (!trylock_page(vmf->page)) {
2521 get_page(vmf->page);
2522 pte_unmap_unlock(vmf->pte, vmf->ptl);
2523 lock_page(vmf->page);
2524 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2525 vmf->address, &vmf->ptl);
2526 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2527 unlock_page(vmf->page);
2528 pte_unmap_unlock(vmf->pte, vmf->ptl);
2529 put_page(vmf->page);
2530 return 0;
2531 }
2532 put_page(vmf->page);
2533 }
2534 if (PageKsm(vmf->page)) {
2535 bool reused = reuse_ksm_page(vmf->page, vmf->vma,
2536 vmf->address);
2537 unlock_page(vmf->page);
2538 if (!reused)
2539 goto copy;
2540 wp_page_reuse(vmf);
2541 return VM_FAULT_WRITE;
2542 }
2543 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2544 if (total_map_swapcount == 1) {
2545 /*
2546 * The page is all ours. Move it to
2547 * our anon_vma so the rmap code will
2548 * not search our parent or siblings.
2549 * Protected against the rmap code by
2550 * the page lock.
2551 */
2552 page_move_anon_rmap(vmf->page, vma);
2553 }
2554 unlock_page(vmf->page);
2555 wp_page_reuse(vmf);
2556 return VM_FAULT_WRITE;
2557 }
2558 unlock_page(vmf->page);
2559 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2560 (VM_WRITE|VM_SHARED))) {
2561 return wp_page_shared(vmf);
2562 }
2563copy:
2564 /*
2565 * Ok, we need to copy. Oh, well..
2566 */
2567 get_page(vmf->page);
2568
2569 pte_unmap_unlock(vmf->pte, vmf->ptl);
2570 return wp_page_copy(vmf);
2571}
2572
2573static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2574 unsigned long start_addr, unsigned long end_addr,
2575 struct zap_details *details)
2576{
2577 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2578}
2579
2580static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2581 struct zap_details *details)
2582{
2583 struct vm_area_struct *vma;
2584 pgoff_t vba, vea, zba, zea;
2585
2586 vma_interval_tree_foreach(vma, root,
2587 details->first_index, details->last_index) {
2588
2589 vba = vma->vm_pgoff;
2590 vea = vba + vma_pages(vma) - 1;
2591 zba = details->first_index;
2592 if (zba < vba)
2593 zba = vba;
2594 zea = details->last_index;
2595 if (zea > vea)
2596 zea = vea;
2597
2598 unmap_mapping_range_vma(vma,
2599 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2600 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2601 details);
2602 }
2603}
2604
2605/**
2606 * unmap_mapping_pages() - Unmap pages from processes.
2607 * @mapping: The address space containing pages to be unmapped.
2608 * @start: Index of first page to be unmapped.
2609 * @nr: Number of pages to be unmapped. 0 to unmap to end of file.
2610 * @even_cows: Whether to unmap even private COWed pages.
2611 *
2612 * Unmap the pages in this address space from any userspace process which
2613 * has them mmaped. Generally, you want to remove COWed pages as well when
2614 * a file is being truncated, but not when invalidating pages from the page
2615 * cache.
2616 */
2617void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2618 pgoff_t nr, bool even_cows)
2619{
2620 struct zap_details details = { };
2621
2622 details.check_mapping = even_cows ? NULL : mapping;
2623 details.first_index = start;
2624 details.last_index = start + nr - 1;
2625 if (details.last_index < details.first_index)
2626 details.last_index = ULONG_MAX;
2627
2628 i_mmap_lock_write(mapping);
2629 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2630 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2631 i_mmap_unlock_write(mapping);
2632}
2633
2634/**
2635 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2636 * address_space corresponding to the specified byte range in the underlying
2637 * file.
2638 *
2639 * @mapping: the address space containing mmaps to be unmapped.
2640 * @holebegin: byte in first page to unmap, relative to the start of
2641 * the underlying file. This will be rounded down to a PAGE_SIZE
2642 * boundary. Note that this is different from truncate_pagecache(), which
2643 * must keep the partial page. In contrast, we must get rid of
2644 * partial pages.
2645 * @holelen: size of prospective hole in bytes. This will be rounded
2646 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2647 * end of the file.
2648 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2649 * but 0 when invalidating pagecache, don't throw away private data.
2650 */
2651void unmap_mapping_range(struct address_space *mapping,
2652 loff_t const holebegin, loff_t const holelen, int even_cows)
2653{
2654 pgoff_t hba = holebegin >> PAGE_SHIFT;
2655 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2656
2657 /* Check for overflow. */
2658 if (sizeof(holelen) > sizeof(hlen)) {
2659 long long holeend =
2660 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2661 if (holeend & ~(long long)ULONG_MAX)
2662 hlen = ULONG_MAX - hba + 1;
2663 }
2664
2665 unmap_mapping_pages(mapping, hba, hlen, even_cows);
2666}
2667EXPORT_SYMBOL(unmap_mapping_range);
2668
2669/*
2670 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2671 * but allow concurrent faults), and pte mapped but not yet locked.
2672 * We return with pte unmapped and unlocked.
2673 *
2674 * We return with the mmap_sem locked or unlocked in the same cases
2675 * as does filemap_fault().
2676 */
2677vm_fault_t do_swap_page(struct vm_fault *vmf)
2678{
2679 struct vm_area_struct *vma = vmf->vma;
2680 struct page *page = NULL, *swapcache;
2681 struct mem_cgroup *memcg;
2682 swp_entry_t entry;
2683 pte_t pte;
2684 int locked;
2685 int exclusive = 0;
2686 vm_fault_t ret = 0;
2687
2688 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2689 goto out;
2690
2691 entry = pte_to_swp_entry(vmf->orig_pte);
2692 if (unlikely(non_swap_entry(entry))) {
2693 if (is_migration_entry(entry)) {
2694 migration_entry_wait(vma->vm_mm, vmf->pmd,
2695 vmf->address);
2696 } else if (is_device_private_entry(entry)) {
2697 /*
2698 * For un-addressable device memory we call the pgmap
2699 * fault handler callback. The callback must migrate
2700 * the page back to some CPU accessible page.
2701 */
2702 ret = device_private_entry_fault(vma, vmf->address, entry,
2703 vmf->flags, vmf->pmd);
2704 } else if (is_hwpoison_entry(entry)) {
2705 ret = VM_FAULT_HWPOISON;
2706 } else {
2707 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2708 ret = VM_FAULT_SIGBUS;
2709 }
2710 goto out;
2711 }
2712
2713
2714 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2715 page = lookup_swap_cache(entry, vma, vmf->address);
2716 swapcache = page;
2717
2718 if (!page) {
2719 struct swap_info_struct *si = swp_swap_info(entry);
2720
2721 if (si->flags & SWP_SYNCHRONOUS_IO &&
2722 __swap_count(si, entry) == 1) {
2723 /* skip swapcache */
2724 page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2725 vmf->address);
2726 if (page) {
2727 __SetPageLocked(page);
2728 __SetPageSwapBacked(page);
2729 set_page_private(page, entry.val);
2730 lru_cache_add_anon(page);
2731 swap_readpage(page, true);
2732 }
2733 } else {
2734 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2735 vmf);
2736 swapcache = page;
2737 }
2738
2739 if (!page) {
2740 /*
2741 * Back out if somebody else faulted in this pte
2742 * while we released the pte lock.
2743 */
2744 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2745 vmf->address, &vmf->ptl);
2746 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2747 ret = VM_FAULT_OOM;
2748 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2749 goto unlock;
2750 }
2751
2752 /* Had to read the page from swap area: Major fault */
2753 ret = VM_FAULT_MAJOR;
2754 count_vm_event(PGMAJFAULT);
2755 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2756 } else if (PageHWPoison(page)) {
2757 /*
2758 * hwpoisoned dirty swapcache pages are kept for killing
2759 * owner processes (which may be unknown at hwpoison time)
2760 */
2761 ret = VM_FAULT_HWPOISON;
2762 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2763 goto out_release;
2764 }
2765
2766 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2767
2768 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2769 if (!locked) {
2770 ret |= VM_FAULT_RETRY;
2771 goto out_release;
2772 }
2773
2774 /*
2775 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2776 * release the swapcache from under us. The page pin, and pte_same
2777 * test below, are not enough to exclude that. Even if it is still
2778 * swapcache, we need to check that the page's swap has not changed.
2779 */
2780 if (unlikely((!PageSwapCache(page) ||
2781 page_private(page) != entry.val)) && swapcache)
2782 goto out_page;
2783
2784 page = ksm_might_need_to_copy(page, vma, vmf->address);
2785 if (unlikely(!page)) {
2786 ret = VM_FAULT_OOM;
2787 page = swapcache;
2788 goto out_page;
2789 }
2790
2791 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2792 &memcg, false)) {
2793 ret = VM_FAULT_OOM;
2794 goto out_page;
2795 }
2796
2797 /*
2798 * Back out if somebody else already faulted in this pte.
2799 */
2800 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2801 &vmf->ptl);
2802 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2803 goto out_nomap;
2804
2805 if (unlikely(!PageUptodate(page))) {
2806 ret = VM_FAULT_SIGBUS;
2807 goto out_nomap;
2808 }
2809
2810 /*
2811 * The page isn't present yet, go ahead with the fault.
2812 *
2813 * Be careful about the sequence of operations here.
2814 * To get its accounting right, reuse_swap_page() must be called
2815 * while the page is counted on swap but not yet in mapcount i.e.
2816 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2817 * must be called after the swap_free(), or it will never succeed.
2818 */
2819
2820 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2821 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2822 pte = mk_pte(page, vma->vm_page_prot);
2823 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2824 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2825 vmf->flags &= ~FAULT_FLAG_WRITE;
2826 ret |= VM_FAULT_WRITE;
2827 exclusive = RMAP_EXCLUSIVE;
2828 }
2829 flush_icache_page(vma, page);
2830 if (pte_swp_soft_dirty(vmf->orig_pte))
2831 pte = pte_mksoft_dirty(pte);
2832 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2833 arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2834 vmf->orig_pte = pte;
2835
2836 /* ksm created a completely new copy */
2837 if (unlikely(page != swapcache && swapcache)) {
2838 page_add_new_anon_rmap(page, vma, vmf->address, false);
2839 mem_cgroup_commit_charge(page, memcg, false, false);
2840 lru_cache_add_active_or_unevictable(page, vma);
2841 } else {
2842 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2843 mem_cgroup_commit_charge(page, memcg, true, false);
2844 activate_page(page);
2845 }
2846
2847 swap_free(entry);
2848 if (mem_cgroup_swap_full(page) ||
2849 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2850 try_to_free_swap(page);
2851 unlock_page(page);
2852 if (page != swapcache && swapcache) {
2853 /*
2854 * Hold the lock to avoid the swap entry to be reused
2855 * until we take the PT lock for the pte_same() check
2856 * (to avoid false positives from pte_same). For
2857 * further safety release the lock after the swap_free
2858 * so that the swap count won't change under a
2859 * parallel locked swapcache.
2860 */
2861 unlock_page(swapcache);
2862 put_page(swapcache);
2863 }
2864
2865 if (vmf->flags & FAULT_FLAG_WRITE) {
2866 ret |= do_wp_page(vmf);
2867 if (ret & VM_FAULT_ERROR)
2868 ret &= VM_FAULT_ERROR;
2869 goto out;
2870 }
2871
2872 /* No need to invalidate - it was non-present before */
2873 update_mmu_cache(vma, vmf->address, vmf->pte);
2874unlock:
2875 pte_unmap_unlock(vmf->pte, vmf->ptl);
2876out:
2877 return ret;
2878out_nomap:
2879 mem_cgroup_cancel_charge(page, memcg, false);
2880 pte_unmap_unlock(vmf->pte, vmf->ptl);
2881out_page:
2882 unlock_page(page);
2883out_release:
2884 put_page(page);
2885 if (page != swapcache && swapcache) {
2886 unlock_page(swapcache);
2887 put_page(swapcache);
2888 }
2889 return ret;
2890}
2891
2892/*
2893 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2894 * but allow concurrent faults), and pte mapped but not yet locked.
2895 * We return with mmap_sem still held, but pte unmapped and unlocked.
2896 */
2897static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2898{
2899 struct vm_area_struct *vma = vmf->vma;
2900 struct mem_cgroup *memcg;
2901 struct page *page;
2902 vm_fault_t ret = 0;
2903 pte_t entry;
2904
2905 /* File mapping without ->vm_ops ? */
2906 if (vma->vm_flags & VM_SHARED)
2907 return VM_FAULT_SIGBUS;
2908
2909 /*
2910 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2911 * pte_offset_map() on pmds where a huge pmd might be created
2912 * from a different thread.
2913 *
2914 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2915 * parallel threads are excluded by other means.
2916 *
2917 * Here we only have down_read(mmap_sem).
2918 */
2919 if (pte_alloc(vma->vm_mm, vmf->pmd))
2920 return VM_FAULT_OOM;
2921
2922 /* See the comment in pte_alloc_one_map() */
2923 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2924 return 0;
2925
2926 /* Use the zero-page for reads */
2927 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2928 !mm_forbids_zeropage(vma->vm_mm)) {
2929 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2930 vma->vm_page_prot));
2931 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2932 vmf->address, &vmf->ptl);
2933 if (!pte_none(*vmf->pte))
2934 goto unlock;
2935 ret = check_stable_address_space(vma->vm_mm);
2936 if (ret)
2937 goto unlock;
2938 /* Deliver the page fault to userland, check inside PT lock */
2939 if (userfaultfd_missing(vma)) {
2940 pte_unmap_unlock(vmf->pte, vmf->ptl);
2941 return handle_userfault(vmf, VM_UFFD_MISSING);
2942 }
2943 goto setpte;
2944 }
2945
2946 /* Allocate our own private page. */
2947 if (unlikely(anon_vma_prepare(vma)))
2948 goto oom;
2949 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2950 if (!page)
2951 goto oom;
2952
2953 if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
2954 false))
2955 goto oom_free_page;
2956
2957 /*
2958 * The memory barrier inside __SetPageUptodate makes sure that
2959 * preceeding stores to the page contents become visible before
2960 * the set_pte_at() write.
2961 */
2962 __SetPageUptodate(page);
2963
2964 entry = mk_pte(page, vma->vm_page_prot);
2965 if (vma->vm_flags & VM_WRITE)
2966 entry = pte_mkwrite(pte_mkdirty(entry));
2967
2968 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2969 &vmf->ptl);
2970 if (!pte_none(*vmf->pte))
2971 goto release;
2972
2973 ret = check_stable_address_space(vma->vm_mm);
2974 if (ret)
2975 goto release;
2976
2977 /* Deliver the page fault to userland, check inside PT lock */
2978 if (userfaultfd_missing(vma)) {
2979 pte_unmap_unlock(vmf->pte, vmf->ptl);
2980 mem_cgroup_cancel_charge(page, memcg, false);
2981 put_page(page);
2982 return handle_userfault(vmf, VM_UFFD_MISSING);
2983 }
2984
2985 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2986 page_add_new_anon_rmap(page, vma, vmf->address, false);
2987 mem_cgroup_commit_charge(page, memcg, false, false);
2988 lru_cache_add_active_or_unevictable(page, vma);
2989setpte:
2990 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2991
2992 /* No need to invalidate - it was non-present before */
2993 update_mmu_cache(vma, vmf->address, vmf->pte);
2994unlock:
2995 pte_unmap_unlock(vmf->pte, vmf->ptl);
2996 return ret;
2997release:
2998 mem_cgroup_cancel_charge(page, memcg, false);
2999 put_page(page);
3000 goto unlock;
3001oom_free_page:
3002 put_page(page);
3003oom:
3004 return VM_FAULT_OOM;
3005}
3006
3007/*
3008 * The mmap_sem must have been held on entry, and may have been
3009 * released depending on flags and vma->vm_ops->fault() return value.
3010 * See filemap_fault() and __lock_page_retry().
3011 */
3012static vm_fault_t __do_fault(struct vm_fault *vmf)
3013{
3014 struct vm_area_struct *vma = vmf->vma;
3015 vm_fault_t ret;
3016
3017 /*
3018 * Preallocate pte before we take page_lock because this might lead to
3019 * deadlocks for memcg reclaim which waits for pages under writeback:
3020 * lock_page(A)
3021 * SetPageWriteback(A)
3022 * unlock_page(A)
3023 * lock_page(B)
3024 * lock_page(B)
3025 * pte_alloc_pne
3026 * shrink_page_list
3027 * wait_on_page_writeback(A)
3028 * SetPageWriteback(B)
3029 * unlock_page(B)
3030 * # flush A, B to clear the writeback
3031 */
3032 if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3033 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3034 if (!vmf->prealloc_pte)
3035 return VM_FAULT_OOM;
3036 smp_wmb(); /* See comment in __pte_alloc() */
3037 }
3038
3039 ret = vma->vm_ops->fault(vmf);
3040 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3041 VM_FAULT_DONE_COW)))
3042 return ret;
3043
3044 if (unlikely(PageHWPoison(vmf->page))) {
3045 if (ret & VM_FAULT_LOCKED)
3046 unlock_page(vmf->page);
3047 put_page(vmf->page);
3048 vmf->page = NULL;
3049 return VM_FAULT_HWPOISON;
3050 }
3051
3052 if (unlikely(!(ret & VM_FAULT_LOCKED)))
3053 lock_page(vmf->page);
3054 else
3055 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3056
3057 return ret;
3058}
3059
3060/*
3061 * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3062 * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3063 * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3064 * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3065 */
3066static int pmd_devmap_trans_unstable(pmd_t *pmd)
3067{
3068 return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3069}
3070
3071static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3072{
3073 struct vm_area_struct *vma = vmf->vma;
3074
3075 if (!pmd_none(*vmf->pmd))
3076 goto map_pte;
3077 if (vmf->prealloc_pte) {
3078 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3079 if (unlikely(!pmd_none(*vmf->pmd))) {
3080 spin_unlock(vmf->ptl);
3081 goto map_pte;
3082 }
3083
3084 mm_inc_nr_ptes(vma->vm_mm);
3085 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3086 spin_unlock(vmf->ptl);
3087 vmf->prealloc_pte = NULL;
3088 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd))) {
3089 return VM_FAULT_OOM;
3090 }
3091map_pte:
3092 /*
3093 * If a huge pmd materialized under us just retry later. Use
3094 * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3095 * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3096 * under us and then back to pmd_none, as a result of MADV_DONTNEED
3097 * running immediately after a huge pmd fault in a different thread of
3098 * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3099 * All we have to ensure is that it is a regular pmd that we can walk
3100 * with pte_offset_map() and we can do that through an atomic read in
3101 * C, which is what pmd_trans_unstable() provides.
3102 */
3103 if (pmd_devmap_trans_unstable(vmf->pmd))
3104 return VM_FAULT_NOPAGE;
3105
3106 /*
3107 * At this point we know that our vmf->pmd points to a page of ptes
3108 * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3109 * for the duration of the fault. If a racing MADV_DONTNEED runs and
3110 * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3111 * be valid and we will re-check to make sure the vmf->pte isn't
3112 * pte_none() under vmf->ptl protection when we return to
3113 * alloc_set_pte().
3114 */
3115 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3116 &vmf->ptl);
3117 return 0;
3118}
3119
3120#ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3121
3122#define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3123static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3124 unsigned long haddr)
3125{
3126 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3127 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3128 return false;
3129 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3130 return false;
3131 return true;
3132}
3133
3134static void deposit_prealloc_pte(struct vm_fault *vmf)
3135{
3136 struct vm_area_struct *vma = vmf->vma;
3137
3138 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3139 /*
3140 * We are going to consume the prealloc table,
3141 * count that as nr_ptes.
3142 */
3143 mm_inc_nr_ptes(vma->vm_mm);
3144 vmf->prealloc_pte = NULL;
3145}
3146
3147static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3148{
3149 struct vm_area_struct *vma = vmf->vma;
3150 bool write = vmf->flags & FAULT_FLAG_WRITE;
3151 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3152 pmd_t entry;
3153 int i;
3154 vm_fault_t ret;
3155
3156 if (!transhuge_vma_suitable(vma, haddr))
3157 return VM_FAULT_FALLBACK;
3158
3159 ret = VM_FAULT_FALLBACK;
3160 page = compound_head(page);
3161
3162 /*
3163 * Archs like ppc64 need additonal space to store information
3164 * related to pte entry. Use the preallocated table for that.
3165 */
3166 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3167 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3168 if (!vmf->prealloc_pte)
3169 return VM_FAULT_OOM;
3170 smp_wmb(); /* See comment in __pte_alloc() */
3171 }
3172
3173 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3174 if (unlikely(!pmd_none(*vmf->pmd)))
3175 goto out;
3176
3177 for (i = 0; i < HPAGE_PMD_NR; i++)
3178 flush_icache_page(vma, page + i);
3179
3180 entry = mk_huge_pmd(page, vma->vm_page_prot);
3181 if (write)
3182 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3183
3184 add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3185 page_add_file_rmap(page, true);
3186 /*
3187 * deposit and withdraw with pmd lock held
3188 */
3189 if (arch_needs_pgtable_deposit())
3190 deposit_prealloc_pte(vmf);
3191
3192 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3193
3194 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3195
3196 /* fault is handled */
3197 ret = 0;
3198 count_vm_event(THP_FILE_MAPPED);
3199out:
3200 spin_unlock(vmf->ptl);
3201 return ret;
3202}
3203#else
3204static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3205{
3206 BUILD_BUG();
3207 return 0;
3208}
3209#endif
3210
3211/**
3212 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3213 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3214 *
3215 * @vmf: fault environment
3216 * @memcg: memcg to charge page (only for private mappings)
3217 * @page: page to map
3218 *
3219 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3220 * return.
3221 *
3222 * Target users are page handler itself and implementations of
3223 * vm_ops->map_pages.
3224 *
3225 * Return: %0 on success, %VM_FAULT_ code in case of error.
3226 */
3227vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3228 struct page *page)
3229{
3230 struct vm_area_struct *vma = vmf->vma;
3231 bool write = vmf->flags & FAULT_FLAG_WRITE;
3232 pte_t entry;
3233 vm_fault_t ret;
3234
3235 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3236 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3237 /* THP on COW? */
3238 VM_BUG_ON_PAGE(memcg, page);
3239
3240 ret = do_set_pmd(vmf, page);
3241 if (ret != VM_FAULT_FALLBACK)
3242 return ret;
3243 }
3244
3245 if (!vmf->pte) {
3246 ret = pte_alloc_one_map(vmf);
3247 if (ret)
3248 return ret;
3249 }
3250
3251 /* Re-check under ptl */
3252 if (unlikely(!pte_none(*vmf->pte)))
3253 return VM_FAULT_NOPAGE;
3254
3255 flush_icache_page(vma, page);
3256 entry = mk_pte(page, vma->vm_page_prot);
3257 if (write)
3258 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3259 /* copy-on-write page */
3260 if (write && !(vma->vm_flags & VM_SHARED)) {
3261 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3262 page_add_new_anon_rmap(page, vma, vmf->address, false);
3263 mem_cgroup_commit_charge(page, memcg, false, false);
3264 lru_cache_add_active_or_unevictable(page, vma);
3265 } else {
3266 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3267 page_add_file_rmap(page, false);
3268 }
3269 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3270
3271 /* no need to invalidate: a not-present page won't be cached */
3272 update_mmu_cache(vma, vmf->address, vmf->pte);
3273
3274 return 0;
3275}
3276
3277
3278/**
3279 * finish_fault - finish page fault once we have prepared the page to fault
3280 *
3281 * @vmf: structure describing the fault
3282 *
3283 * This function handles all that is needed to finish a page fault once the
3284 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3285 * given page, adds reverse page mapping, handles memcg charges and LRU
3286 * addition.
3287 *
3288 * The function expects the page to be locked and on success it consumes a
3289 * reference of a page being mapped (for the PTE which maps it).
3290 *
3291 * Return: %0 on success, %VM_FAULT_ code in case of error.
3292 */
3293vm_fault_t finish_fault(struct vm_fault *vmf)
3294{
3295 struct page *page;
3296 vm_fault_t ret = 0;
3297
3298 /* Did we COW the page? */
3299 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3300 !(vmf->vma->vm_flags & VM_SHARED))
3301 page = vmf->cow_page;
3302 else
3303 page = vmf->page;
3304
3305 /*
3306 * check even for read faults because we might have lost our CoWed
3307 * page
3308 */
3309 if (!(vmf->vma->vm_flags & VM_SHARED))
3310 ret = check_stable_address_space(vmf->vma->vm_mm);
3311 if (!ret)
3312 ret = alloc_set_pte(vmf, vmf->memcg, page);
3313 if (vmf->pte)
3314 pte_unmap_unlock(vmf->pte, vmf->ptl);
3315 return ret;
3316}
3317
3318static unsigned long fault_around_bytes __read_mostly =
3319 rounddown_pow_of_two(65536);
3320
3321#ifdef CONFIG_DEBUG_FS
3322static int fault_around_bytes_get(void *data, u64 *val)
3323{
3324 *val = fault_around_bytes;
3325 return 0;
3326}
3327
3328/*
3329 * fault_around_bytes must be rounded down to the nearest page order as it's
3330 * what do_fault_around() expects to see.
3331 */
3332static int fault_around_bytes_set(void *data, u64 val)
3333{
3334 if (val / PAGE_SIZE > PTRS_PER_PTE)
3335 return -EINVAL;
3336 if (val > PAGE_SIZE)
3337 fault_around_bytes = rounddown_pow_of_two(val);
3338 else
3339 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3340 return 0;
3341}
3342DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3343 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3344
3345static int __init fault_around_debugfs(void)
3346{
3347 debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3348 &fault_around_bytes_fops);
3349 return 0;
3350}
3351late_initcall(fault_around_debugfs);
3352#endif
3353
3354/*
3355 * do_fault_around() tries to map few pages around the fault address. The hope
3356 * is that the pages will be needed soon and this will lower the number of
3357 * faults to handle.
3358 *
3359 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3360 * not ready to be mapped: not up-to-date, locked, etc.
3361 *
3362 * This function is called with the page table lock taken. In the split ptlock
3363 * case the page table lock only protects only those entries which belong to
3364 * the page table corresponding to the fault address.
3365 *
3366 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3367 * only once.
3368 *
3369 * fault_around_bytes defines how many bytes we'll try to map.
3370 * do_fault_around() expects it to be set to a power of two less than or equal
3371 * to PTRS_PER_PTE.
3372 *
3373 * The virtual address of the area that we map is naturally aligned to
3374 * fault_around_bytes rounded down to the machine page size
3375 * (and therefore to page order). This way it's easier to guarantee
3376 * that we don't cross page table boundaries.
3377 */
3378static vm_fault_t do_fault_around(struct vm_fault *vmf)
3379{
3380 unsigned long address = vmf->address, nr_pages, mask;
3381 pgoff_t start_pgoff = vmf->pgoff;
3382 pgoff_t end_pgoff;
3383 int off;
3384 vm_fault_t ret = 0;
3385
3386 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3387 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3388
3389 vmf->address = max(address & mask, vmf->vma->vm_start);
3390 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3391 start_pgoff -= off;
3392
3393 /*
3394 * end_pgoff is either the end of the page table, the end of
3395 * the vma or nr_pages from start_pgoff, depending what is nearest.
3396 */
3397 end_pgoff = start_pgoff -
3398 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3399 PTRS_PER_PTE - 1;
3400 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3401 start_pgoff + nr_pages - 1);
3402
3403 if (pmd_none(*vmf->pmd)) {
3404 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3405 if (!vmf->prealloc_pte)
3406 goto out;
3407 smp_wmb(); /* See comment in __pte_alloc() */
3408 }
3409
3410 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3411
3412 /* Huge page is mapped? Page fault is solved */
3413 if (pmd_trans_huge(*vmf->pmd)) {
3414 ret = VM_FAULT_NOPAGE;
3415 goto out;
3416 }
3417
3418 /* ->map_pages() haven't done anything useful. Cold page cache? */
3419 if (!vmf->pte)
3420 goto out;
3421
3422 /* check if the page fault is solved */
3423 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3424 if (!pte_none(*vmf->pte))
3425 ret = VM_FAULT_NOPAGE;
3426 pte_unmap_unlock(vmf->pte, vmf->ptl);
3427out:
3428 vmf->address = address;
3429 vmf->pte = NULL;
3430 return ret;
3431}
3432
3433static vm_fault_t do_read_fault(struct vm_fault *vmf)
3434{
3435 struct vm_area_struct *vma = vmf->vma;
3436 vm_fault_t ret = 0;
3437
3438 /*
3439 * Let's call ->map_pages() first and use ->fault() as fallback
3440 * if page by the offset is not ready to be mapped (cold cache or
3441 * something).
3442 */
3443 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3444 ret = do_fault_around(vmf);
3445 if (ret)
3446 return ret;
3447 }
3448
3449 ret = __do_fault(vmf);
3450 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3451 return ret;
3452
3453 ret |= finish_fault(vmf);
3454 unlock_page(vmf->page);
3455 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3456 put_page(vmf->page);
3457 return ret;
3458}
3459
3460static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3461{
3462 struct vm_area_struct *vma = vmf->vma;
3463 vm_fault_t ret;
3464
3465 if (unlikely(anon_vma_prepare(vma)))
3466 return VM_FAULT_OOM;
3467
3468 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3469 if (!vmf->cow_page)
3470 return VM_FAULT_OOM;
3471
3472 if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3473 &vmf->memcg, false)) {
3474 put_page(vmf->cow_page);
3475 return VM_FAULT_OOM;
3476 }
3477
3478 ret = __do_fault(vmf);
3479 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3480 goto uncharge_out;
3481 if (ret & VM_FAULT_DONE_COW)
3482 return ret;
3483
3484 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3485 __SetPageUptodate(vmf->cow_page);
3486
3487 ret |= finish_fault(vmf);
3488 unlock_page(vmf->page);
3489 put_page(vmf->page);
3490 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3491 goto uncharge_out;
3492 return ret;
3493uncharge_out:
3494 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3495 put_page(vmf->cow_page);
3496 return ret;
3497}
3498
3499static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3500{
3501 struct vm_area_struct *vma = vmf->vma;
3502 vm_fault_t ret, tmp;
3503
3504 ret = __do_fault(vmf);
3505 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3506 return ret;
3507
3508 /*
3509 * Check if the backing address space wants to know that the page is
3510 * about to become writable
3511 */
3512 if (vma->vm_ops->page_mkwrite) {
3513 unlock_page(vmf->page);
3514 tmp = do_page_mkwrite(vmf);
3515 if (unlikely(!tmp ||
3516 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3517 put_page(vmf->page);
3518 return tmp;
3519 }
3520 }
3521
3522 ret |= finish_fault(vmf);
3523 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3524 VM_FAULT_RETRY))) {
3525 unlock_page(vmf->page);
3526 put_page(vmf->page);
3527 return ret;
3528 }
3529
3530 fault_dirty_shared_page(vma, vmf->page);
3531 return ret;
3532}
3533
3534/*
3535 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3536 * but allow concurrent faults).
3537 * The mmap_sem may have been released depending on flags and our
3538 * return value. See filemap_fault() and __lock_page_or_retry().
3539 * If mmap_sem is released, vma may become invalid (for example
3540 * by other thread calling munmap()).
3541 */
3542static vm_fault_t do_fault(struct vm_fault *vmf)
3543{
3544 struct vm_area_struct *vma = vmf->vma;
3545 struct mm_struct *vm_mm = vma->vm_mm;
3546 vm_fault_t ret;
3547
3548 /*
3549 * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
3550 */
3551 if (!vma->vm_ops->fault) {
3552 /*
3553 * If we find a migration pmd entry or a none pmd entry, which
3554 * should never happen, return SIGBUS
3555 */
3556 if (unlikely(!pmd_present(*vmf->pmd)))
3557 ret = VM_FAULT_SIGBUS;
3558 else {
3559 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
3560 vmf->pmd,
3561 vmf->address,
3562 &vmf->ptl);
3563 /*
3564 * Make sure this is not a temporary clearing of pte
3565 * by holding ptl and checking again. A R/M/W update
3566 * of pte involves: take ptl, clearing the pte so that
3567 * we don't have concurrent modification by hardware
3568 * followed by an update.
3569 */
3570 if (unlikely(pte_none(*vmf->pte)))
3571 ret = VM_FAULT_SIGBUS;
3572 else
3573 ret = VM_FAULT_NOPAGE;
3574
3575 pte_unmap_unlock(vmf->pte, vmf->ptl);
3576 }
3577 } else if (!(vmf->flags & FAULT_FLAG_WRITE))
3578 ret = do_read_fault(vmf);
3579 else if (!(vma->vm_flags & VM_SHARED))
3580 ret = do_cow_fault(vmf);
3581 else
3582 ret = do_shared_fault(vmf);
3583
3584 /* preallocated pagetable is unused: free it */
3585 if (vmf->prealloc_pte) {
3586 pte_free(vm_mm, vmf->prealloc_pte);
3587 vmf->prealloc_pte = NULL;
3588 }
3589 return ret;
3590}
3591
3592static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3593 unsigned long addr, int page_nid,
3594 int *flags)
3595{
3596 get_page(page);
3597
3598 count_vm_numa_event(NUMA_HINT_FAULTS);
3599 if (page_nid == numa_node_id()) {
3600 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3601 *flags |= TNF_FAULT_LOCAL;
3602 }
3603
3604 return mpol_misplaced(page,