550405fc3b5e6396a1dd1fd974dde3e07a3c9d15
[linux-2.6-microblaze.git] / mm / memory.c
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/mm/memory.c
4  *
5  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
6  */
7
8 /*
9  * demand-loading started 01.12.91 - seems it is high on the list of
10  * things wanted, and it should be easy to implement. - Linus
11  */
12
13 /*
14  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
15  * pages started 02.12.91, seems to work. - Linus.
16  *
17  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
18  * would have taken more than the 6M I have free, but it worked well as
19  * far as I could see.
20  *
21  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
22  */
23
24 /*
25  * Real VM (paging to/from disk) started 18.12.91. Much more work and
26  * thought has to go into this. Oh, well..
27  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
28  *              Found it. Everything seems to work now.
29  * 20.12.91  -  Ok, making the swap-device changeable like the root.
30  */
31
32 /*
33  * 05.04.94  -  Multi-page memory management added for v1.1.
34  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
35  *
36  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
37  *              (Gerhard.Wichert@pdb.siemens.de)
38  *
39  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
40  */
41
42 #include <linux/kernel_stat.h>
43 #include <linux/mm.h>
44 #include <linux/sched/mm.h>
45 #include <linux/sched/coredump.h>
46 #include <linux/sched/numa_balancing.h>
47 #include <linux/sched/task.h>
48 #include <linux/hugetlb.h>
49 #include <linux/mman.h>
50 #include <linux/swap.h>
51 #include <linux/highmem.h>
52 #include <linux/pagemap.h>
53 #include <linux/memremap.h>
54 #include <linux/ksm.h>
55 #include <linux/rmap.h>
56 #include <linux/export.h>
57 #include <linux/delayacct.h>
58 #include <linux/init.h>
59 #include <linux/pfn_t.h>
60 #include <linux/writeback.h>
61 #include <linux/memcontrol.h>
62 #include <linux/mmu_notifier.h>
63 #include <linux/swapops.h>
64 #include <linux/elf.h>
65 #include <linux/gfp.h>
66 #include <linux/migrate.h>
67 #include <linux/string.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72 #include <linux/numa.h>
73 #include <linux/perf_event.h>
74 #include <linux/ptrace.h>
75 #include <linux/vmalloc.h>
76
77 #include <trace/events/kmem.h>
78
79 #include <asm/io.h>
80 #include <asm/mmu_context.h>
81 #include <asm/pgalloc.h>
82 #include <linux/uaccess.h>
83 #include <asm/tlb.h>
84 #include <asm/tlbflush.h>
85
86 #include "pgalloc-track.h"
87 #include "internal.h"
88
89 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
90 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
91 #endif
92
93 #ifndef CONFIG_NEED_MULTIPLE_NODES
94 /* use the per-pgdat data instead for discontigmem - mbligh */
95 unsigned long max_mapnr;
96 EXPORT_SYMBOL(max_mapnr);
97
98 struct page *mem_map;
99 EXPORT_SYMBOL(mem_map);
100 #endif
101
102 /*
103  * A number of key systems in x86 including ioremap() rely on the assumption
104  * that high_memory defines the upper bound on direct map memory, then end
105  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
106  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
107  * and ZONE_HIGHMEM.
108  */
109 void *high_memory;
110 EXPORT_SYMBOL(high_memory);
111
112 /*
113  * Randomize the address space (stacks, mmaps, brk, etc.).
114  *
115  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
116  *   as ancient (libc5 based) binaries can segfault. )
117  */
118 int randomize_va_space __read_mostly =
119 #ifdef CONFIG_COMPAT_BRK
120                                         1;
121 #else
122                                         2;
123 #endif
124
125 #ifndef arch_faults_on_old_pte
126 static inline bool arch_faults_on_old_pte(void)
127 {
128         /*
129          * Those arches which don't have hw access flag feature need to
130          * implement their own helper. By default, "true" means pagefault
131          * will be hit on old pte.
132          */
133         return true;
134 }
135 #endif
136
137 #ifndef arch_wants_old_prefaulted_pte
138 static inline bool arch_wants_old_prefaulted_pte(void)
139 {
140         /*
141          * Transitioning a PTE from 'old' to 'young' can be expensive on
142          * some architectures, even if it's performed in hardware. By
143          * default, "false" means prefaulted entries will be 'young'.
144          */
145         return false;
146 }
147 #endif
148
149 static int __init disable_randmaps(char *s)
150 {
151         randomize_va_space = 0;
152         return 1;
153 }
154 __setup("norandmaps", disable_randmaps);
155
156 unsigned long zero_pfn __read_mostly;
157 EXPORT_SYMBOL(zero_pfn);
158
159 unsigned long highest_memmap_pfn __read_mostly;
160
161 /*
162  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
163  */
164 static int __init init_zero_pfn(void)
165 {
166         zero_pfn = page_to_pfn(ZERO_PAGE(0));
167         return 0;
168 }
169 early_initcall(init_zero_pfn);
170
171 void mm_trace_rss_stat(struct mm_struct *mm, int member, long count)
172 {
173         trace_rss_stat(mm, member, count);
174 }
175
176 #if defined(SPLIT_RSS_COUNTING)
177
178 void sync_mm_rss(struct mm_struct *mm)
179 {
180         int i;
181
182         for (i = 0; i < NR_MM_COUNTERS; i++) {
183                 if (current->rss_stat.count[i]) {
184                         add_mm_counter(mm, i, current->rss_stat.count[i]);
185                         current->rss_stat.count[i] = 0;
186                 }
187         }
188         current->rss_stat.events = 0;
189 }
190
191 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
192 {
193         struct task_struct *task = current;
194
195         if (likely(task->mm == mm))
196                 task->rss_stat.count[member] += val;
197         else
198                 add_mm_counter(mm, member, val);
199 }
200 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
201 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
202
203 /* sync counter once per 64 page faults */
204 #define TASK_RSS_EVENTS_THRESH  (64)
205 static void check_sync_rss_stat(struct task_struct *task)
206 {
207         if (unlikely(task != current))
208                 return;
209         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
210                 sync_mm_rss(task->mm);
211 }
212 #else /* SPLIT_RSS_COUNTING */
213
214 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
215 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
216
217 static void check_sync_rss_stat(struct task_struct *task)
218 {
219 }
220
221 #endif /* SPLIT_RSS_COUNTING */
222
223 /*
224  * Note: this doesn't free the actual pages themselves. That
225  * has been handled earlier when unmapping all the memory regions.
226  */
227 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
228                            unsigned long addr)
229 {
230         pgtable_t token = pmd_pgtable(*pmd);
231         pmd_clear(pmd);
232         pte_free_tlb(tlb, token, addr);
233         mm_dec_nr_ptes(tlb->mm);
234 }
235
236 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
237                                 unsigned long addr, unsigned long end,
238                                 unsigned long floor, unsigned long ceiling)
239 {
240         pmd_t *pmd;
241         unsigned long next;
242         unsigned long start;
243
244         start = addr;
245         pmd = pmd_offset(pud, addr);
246         do {
247                 next = pmd_addr_end(addr, end);
248                 if (pmd_none_or_clear_bad(pmd))
249                         continue;
250                 free_pte_range(tlb, pmd, addr);
251         } while (pmd++, addr = next, addr != end);
252
253         start &= PUD_MASK;
254         if (start < floor)
255                 return;
256         if (ceiling) {
257                 ceiling &= PUD_MASK;
258                 if (!ceiling)
259                         return;
260         }
261         if (end - 1 > ceiling - 1)
262                 return;
263
264         pmd = pmd_offset(pud, start);
265         pud_clear(pud);
266         pmd_free_tlb(tlb, pmd, start);
267         mm_dec_nr_pmds(tlb->mm);
268 }
269
270 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
271                                 unsigned long addr, unsigned long end,
272                                 unsigned long floor, unsigned long ceiling)
273 {
274         pud_t *pud;
275         unsigned long next;
276         unsigned long start;
277
278         start = addr;
279         pud = pud_offset(p4d, addr);
280         do {
281                 next = pud_addr_end(addr, end);
282                 if (pud_none_or_clear_bad(pud))
283                         continue;
284                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
285         } while (pud++, addr = next, addr != end);
286
287         start &= P4D_MASK;
288         if (start < floor)
289                 return;
290         if (ceiling) {
291                 ceiling &= P4D_MASK;
292                 if (!ceiling)
293                         return;
294         }
295         if (end - 1 > ceiling - 1)
296                 return;
297
298         pud = pud_offset(p4d, start);
299         p4d_clear(p4d);
300         pud_free_tlb(tlb, pud, start);
301         mm_dec_nr_puds(tlb->mm);
302 }
303
304 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
305                                 unsigned long addr, unsigned long end,
306                                 unsigned long floor, unsigned long ceiling)
307 {
308         p4d_t *p4d;
309         unsigned long next;
310         unsigned long start;
311
312         start = addr;
313         p4d = p4d_offset(pgd, addr);
314         do {
315                 next = p4d_addr_end(addr, end);
316                 if (p4d_none_or_clear_bad(p4d))
317                         continue;
318                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
319         } while (p4d++, addr = next, addr != end);
320
321         start &= PGDIR_MASK;
322         if (start < floor)
323                 return;
324         if (ceiling) {
325                 ceiling &= PGDIR_MASK;
326                 if (!ceiling)
327                         return;
328         }
329         if (end - 1 > ceiling - 1)
330                 return;
331
332         p4d = p4d_offset(pgd, start);
333         pgd_clear(pgd);
334         p4d_free_tlb(tlb, p4d, start);
335 }
336
337 /*
338  * This function frees user-level page tables of a process.
339  */
340 void free_pgd_range(struct mmu_gather *tlb,
341                         unsigned long addr, unsigned long end,
342                         unsigned long floor, unsigned long ceiling)
343 {
344         pgd_t *pgd;
345         unsigned long next;
346
347         /*
348          * The next few lines have given us lots of grief...
349          *
350          * Why are we testing PMD* at this top level?  Because often
351          * there will be no work to do at all, and we'd prefer not to
352          * go all the way down to the bottom just to discover that.
353          *
354          * Why all these "- 1"s?  Because 0 represents both the bottom
355          * of the address space and the top of it (using -1 for the
356          * top wouldn't help much: the masks would do the wrong thing).
357          * The rule is that addr 0 and floor 0 refer to the bottom of
358          * the address space, but end 0 and ceiling 0 refer to the top
359          * Comparisons need to use "end - 1" and "ceiling - 1" (though
360          * that end 0 case should be mythical).
361          *
362          * Wherever addr is brought up or ceiling brought down, we must
363          * be careful to reject "the opposite 0" before it confuses the
364          * subsequent tests.  But what about where end is brought down
365          * by PMD_SIZE below? no, end can't go down to 0 there.
366          *
367          * Whereas we round start (addr) and ceiling down, by different
368          * masks at different levels, in order to test whether a table
369          * now has no other vmas using it, so can be freed, we don't
370          * bother to round floor or end up - the tests don't need that.
371          */
372
373         addr &= PMD_MASK;
374         if (addr < floor) {
375                 addr += PMD_SIZE;
376                 if (!addr)
377                         return;
378         }
379         if (ceiling) {
380                 ceiling &= PMD_MASK;
381                 if (!ceiling)
382                         return;
383         }
384         if (end - 1 > ceiling - 1)
385                 end -= PMD_SIZE;
386         if (addr > end - 1)
387                 return;
388         /*
389          * We add page table cache pages with PAGE_SIZE,
390          * (see pte_free_tlb()), flush the tlb if we need
391          */
392         tlb_change_page_size(tlb, PAGE_SIZE);
393         pgd = pgd_offset(tlb->mm, addr);
394         do {
395                 next = pgd_addr_end(addr, end);
396                 if (pgd_none_or_clear_bad(pgd))
397                         continue;
398                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
399         } while (pgd++, addr = next, addr != end);
400 }
401
402 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
403                 unsigned long floor, unsigned long ceiling)
404 {
405         while (vma) {
406                 struct vm_area_struct *next = vma->vm_next;
407                 unsigned long addr = vma->vm_start;
408
409                 /*
410                  * Hide vma from rmap and truncate_pagecache before freeing
411                  * pgtables
412                  */
413                 unlink_anon_vmas(vma);
414                 unlink_file_vma(vma);
415
416                 if (is_vm_hugetlb_page(vma)) {
417                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
418                                 floor, next ? next->vm_start : ceiling);
419                 } else {
420                         /*
421                          * Optimization: gather nearby vmas into one call down
422                          */
423                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
424                                && !is_vm_hugetlb_page(next)) {
425                                 vma = next;
426                                 next = vma->vm_next;
427                                 unlink_anon_vmas(vma);
428                                 unlink_file_vma(vma);
429                         }
430                         free_pgd_range(tlb, addr, vma->vm_end,
431                                 floor, next ? next->vm_start : ceiling);
432                 }
433                 vma = next;
434         }
435 }
436
437 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd)
438 {
439         spinlock_t *ptl;
440         pgtable_t new = pte_alloc_one(mm);
441         if (!new)
442                 return -ENOMEM;
443
444         /*
445          * Ensure all pte setup (eg. pte page lock and page clearing) are
446          * visible before the pte is made visible to other CPUs by being
447          * put into page tables.
448          *
449          * The other side of the story is the pointer chasing in the page
450          * table walking code (when walking the page table without locking;
451          * ie. most of the time). Fortunately, these data accesses consist
452          * of a chain of data-dependent loads, meaning most CPUs (alpha
453          * being the notable exception) will already guarantee loads are
454          * seen in-order. See the alpha page table accessors for the
455          * smp_rmb() barriers in page table walking code.
456          */
457         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
458
459         ptl = pmd_lock(mm, pmd);
460         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
461                 mm_inc_nr_ptes(mm);
462                 pmd_populate(mm, pmd, new);
463                 new = NULL;
464         }
465         spin_unlock(ptl);
466         if (new)
467                 pte_free(mm, new);
468         return 0;
469 }
470
471 int __pte_alloc_kernel(pmd_t *pmd)
472 {
473         pte_t *new = pte_alloc_one_kernel(&init_mm);
474         if (!new)
475                 return -ENOMEM;
476
477         smp_wmb(); /* See comment in __pte_alloc */
478
479         spin_lock(&init_mm.page_table_lock);
480         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
481                 pmd_populate_kernel(&init_mm, pmd, new);
482                 new = NULL;
483         }
484         spin_unlock(&init_mm.page_table_lock);
485         if (new)
486                 pte_free_kernel(&init_mm, new);
487         return 0;
488 }
489
490 static inline void init_rss_vec(int *rss)
491 {
492         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
493 }
494
495 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
496 {
497         int i;
498
499         if (current->mm == mm)
500                 sync_mm_rss(mm);
501         for (i = 0; i < NR_MM_COUNTERS; i++)
502                 if (rss[i])
503                         add_mm_counter(mm, i, rss[i]);
504 }
505
506 /*
507  * This function is called to print an error when a bad pte
508  * is found. For example, we might have a PFN-mapped pte in
509  * a region that doesn't allow it.
510  *
511  * The calling function must still handle the error.
512  */
513 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
514                           pte_t pte, struct page *page)
515 {
516         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
517         p4d_t *p4d = p4d_offset(pgd, addr);
518         pud_t *pud = pud_offset(p4d, addr);
519         pmd_t *pmd = pmd_offset(pud, addr);
520         struct address_space *mapping;
521         pgoff_t index;
522         static unsigned long resume;
523         static unsigned long nr_shown;
524         static unsigned long nr_unshown;
525
526         /*
527          * Allow a burst of 60 reports, then keep quiet for that minute;
528          * or allow a steady drip of one report per second.
529          */
530         if (nr_shown == 60) {
531                 if (time_before(jiffies, resume)) {
532                         nr_unshown++;
533                         return;
534                 }
535                 if (nr_unshown) {
536                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
537                                  nr_unshown);
538                         nr_unshown = 0;
539                 }
540                 nr_shown = 0;
541         }
542         if (nr_shown++ == 0)
543                 resume = jiffies + 60 * HZ;
544
545         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
546         index = linear_page_index(vma, addr);
547
548         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
549                  current->comm,
550                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
551         if (page)
552                 dump_page(page, "bad pte");
553         pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
554                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
555         pr_alert("file:%pD fault:%ps mmap:%ps readpage:%ps\n",
556                  vma->vm_file,
557                  vma->vm_ops ? vma->vm_ops->fault : NULL,
558                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
559                  mapping ? mapping->a_ops->readpage : NULL);
560         dump_stack();
561         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
562 }
563
564 /*
565  * vm_normal_page -- This function gets the "struct page" associated with a pte.
566  *
567  * "Special" mappings do not wish to be associated with a "struct page" (either
568  * it doesn't exist, or it exists but they don't want to touch it). In this
569  * case, NULL is returned here. "Normal" mappings do have a struct page.
570  *
571  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
572  * pte bit, in which case this function is trivial. Secondly, an architecture
573  * may not have a spare pte bit, which requires a more complicated scheme,
574  * described below.
575  *
576  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
577  * special mapping (even if there are underlying and valid "struct pages").
578  * COWed pages of a VM_PFNMAP are always normal.
579  *
580  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
581  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
582  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
583  * mapping will always honor the rule
584  *
585  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
586  *
587  * And for normal mappings this is false.
588  *
589  * This restricts such mappings to be a linear translation from virtual address
590  * to pfn. To get around this restriction, we allow arbitrary mappings so long
591  * as the vma is not a COW mapping; in that case, we know that all ptes are
592  * special (because none can have been COWed).
593  *
594  *
595  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
596  *
597  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
598  * page" backing, however the difference is that _all_ pages with a struct
599  * page (that is, those where pfn_valid is true) are refcounted and considered
600  * normal pages by the VM. The disadvantage is that pages are refcounted
601  * (which can be slower and simply not an option for some PFNMAP users). The
602  * advantage is that we don't have to follow the strict linearity rule of
603  * PFNMAP mappings in order to support COWable mappings.
604  *
605  */
606 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
607                             pte_t pte)
608 {
609         unsigned long pfn = pte_pfn(pte);
610
611         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
612                 if (likely(!pte_special(pte)))
613                         goto check_pfn;
614                 if (vma->vm_ops && vma->vm_ops->find_special_page)
615                         return vma->vm_ops->find_special_page(vma, addr);
616                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
617                         return NULL;
618                 if (is_zero_pfn(pfn))
619                         return NULL;
620                 if (pte_devmap(pte))
621                         return NULL;
622
623                 print_bad_pte(vma, addr, pte, NULL);
624                 return NULL;
625         }
626
627         /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
628
629         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
630                 if (vma->vm_flags & VM_MIXEDMAP) {
631                         if (!pfn_valid(pfn))
632                                 return NULL;
633                         goto out;
634                 } else {
635                         unsigned long off;
636                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
637                         if (pfn == vma->vm_pgoff + off)
638                                 return NULL;
639                         if (!is_cow_mapping(vma->vm_flags))
640                                 return NULL;
641                 }
642         }
643
644         if (is_zero_pfn(pfn))
645                 return NULL;
646
647 check_pfn:
648         if (unlikely(pfn > highest_memmap_pfn)) {
649                 print_bad_pte(vma, addr, pte, NULL);
650                 return NULL;
651         }
652
653         /*
654          * NOTE! We still have PageReserved() pages in the page tables.
655          * eg. VDSO mappings can cause them to exist.
656          */
657 out:
658         return pfn_to_page(pfn);
659 }
660
661 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
662 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
663                                 pmd_t pmd)
664 {
665         unsigned long pfn = pmd_pfn(pmd);
666
667         /*
668          * There is no pmd_special() but there may be special pmds, e.g.
669          * in a direct-access (dax) mapping, so let's just replicate the
670          * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
671          */
672         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
673                 if (vma->vm_flags & VM_MIXEDMAP) {
674                         if (!pfn_valid(pfn))
675                                 return NULL;
676                         goto out;
677                 } else {
678                         unsigned long off;
679                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
680                         if (pfn == vma->vm_pgoff + off)
681                                 return NULL;
682                         if (!is_cow_mapping(vma->vm_flags))
683                                 return NULL;
684                 }
685         }
686
687         if (pmd_devmap(pmd))
688                 return NULL;
689         if (is_huge_zero_pmd(pmd))
690                 return NULL;
691         if (unlikely(pfn > highest_memmap_pfn))
692                 return NULL;
693
694         /*
695          * NOTE! We still have PageReserved() pages in the page tables.
696          * eg. VDSO mappings can cause them to exist.
697          */
698 out:
699         return pfn_to_page(pfn);
700 }
701 #endif
702
703 /*
704  * copy one vm_area from one task to the other. Assumes the page tables
705  * already present in the new task to be cleared in the whole range
706  * covered by this vma.
707  */
708
709 static unsigned long
710 copy_nonpresent_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
711                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
712                 unsigned long addr, int *rss)
713 {
714         unsigned long vm_flags = vma->vm_flags;
715         pte_t pte = *src_pte;
716         struct page *page;
717         swp_entry_t entry = pte_to_swp_entry(pte);
718
719         if (likely(!non_swap_entry(entry))) {
720                 if (swap_duplicate(entry) < 0)
721                         return entry.val;
722
723                 /* make sure dst_mm is on swapoff's mmlist. */
724                 if (unlikely(list_empty(&dst_mm->mmlist))) {
725                         spin_lock(&mmlist_lock);
726                         if (list_empty(&dst_mm->mmlist))
727                                 list_add(&dst_mm->mmlist,
728                                                 &src_mm->mmlist);
729                         spin_unlock(&mmlist_lock);
730                 }
731                 rss[MM_SWAPENTS]++;
732         } else if (is_migration_entry(entry)) {
733                 page = migration_entry_to_page(entry);
734
735                 rss[mm_counter(page)]++;
736
737                 if (is_write_migration_entry(entry) &&
738                                 is_cow_mapping(vm_flags)) {
739                         /*
740                          * COW mappings require pages in both
741                          * parent and child to be set to read.
742                          */
743                         make_migration_entry_read(&entry);
744                         pte = swp_entry_to_pte(entry);
745                         if (pte_swp_soft_dirty(*src_pte))
746                                 pte = pte_swp_mksoft_dirty(pte);
747                         if (pte_swp_uffd_wp(*src_pte))
748                                 pte = pte_swp_mkuffd_wp(pte);
749                         set_pte_at(src_mm, addr, src_pte, pte);
750                 }
751         } else if (is_device_private_entry(entry)) {
752                 page = device_private_entry_to_page(entry);
753
754                 /*
755                  * Update rss count even for unaddressable pages, as
756                  * they should treated just like normal pages in this
757                  * respect.
758                  *
759                  * We will likely want to have some new rss counters
760                  * for unaddressable pages, at some point. But for now
761                  * keep things as they are.
762                  */
763                 get_page(page);
764                 rss[mm_counter(page)]++;
765                 page_dup_rmap(page, false);
766
767                 /*
768                  * We do not preserve soft-dirty information, because so
769                  * far, checkpoint/restore is the only feature that
770                  * requires that. And checkpoint/restore does not work
771                  * when a device driver is involved (you cannot easily
772                  * save and restore device driver state).
773                  */
774                 if (is_write_device_private_entry(entry) &&
775                     is_cow_mapping(vm_flags)) {
776                         make_device_private_entry_read(&entry);
777                         pte = swp_entry_to_pte(entry);
778                         if (pte_swp_uffd_wp(*src_pte))
779                                 pte = pte_swp_mkuffd_wp(pte);
780                         set_pte_at(src_mm, addr, src_pte, pte);
781                 }
782         }
783         set_pte_at(dst_mm, addr, dst_pte, pte);
784         return 0;
785 }
786
787 /*
788  * Copy a present and normal page if necessary.
789  *
790  * NOTE! The usual case is that this doesn't need to do
791  * anything, and can just return a positive value. That
792  * will let the caller know that it can just increase
793  * the page refcount and re-use the pte the traditional
794  * way.
795  *
796  * But _if_ we need to copy it because it needs to be
797  * pinned in the parent (and the child should get its own
798  * copy rather than just a reference to the same page),
799  * we'll do that here and return zero to let the caller
800  * know we're done.
801  *
802  * And if we need a pre-allocated page but don't yet have
803  * one, return a negative error to let the preallocation
804  * code know so that it can do so outside the page table
805  * lock.
806  */
807 static inline int
808 copy_present_page(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
809                   pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
810                   struct page **prealloc, pte_t pte, struct page *page)
811 {
812         struct page *new_page;
813
814         /*
815          * What we want to do is to check whether this page may
816          * have been pinned by the parent process.  If so,
817          * instead of wrprotect the pte on both sides, we copy
818          * the page immediately so that we'll always guarantee
819          * the pinned page won't be randomly replaced in the
820          * future.
821          *
822          * The page pinning checks are just "has this mm ever
823          * seen pinning", along with the (inexact) check of
824          * the page count. That might give false positives for
825          * for pinning, but it will work correctly.
826          */
827         if (likely(!page_needs_cow_for_dma(src_vma, page)))
828                 return 1;
829
830         new_page = *prealloc;
831         if (!new_page)
832                 return -EAGAIN;
833
834         /*
835          * We have a prealloc page, all good!  Take it
836          * over and copy the page & arm it.
837          */
838         *prealloc = NULL;
839         copy_user_highpage(new_page, page, addr, src_vma);
840         __SetPageUptodate(new_page);
841         page_add_new_anon_rmap(new_page, dst_vma, addr, false);
842         lru_cache_add_inactive_or_unevictable(new_page, dst_vma);
843         rss[mm_counter(new_page)]++;
844
845         /* All done, just insert the new page copy in the child */
846         pte = mk_pte(new_page, dst_vma->vm_page_prot);
847         pte = maybe_mkwrite(pte_mkdirty(pte), dst_vma);
848         set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
849         return 0;
850 }
851
852 /*
853  * Copy one pte.  Returns 0 if succeeded, or -EAGAIN if one preallocated page
854  * is required to copy this pte.
855  */
856 static inline int
857 copy_present_pte(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
858                  pte_t *dst_pte, pte_t *src_pte, unsigned long addr, int *rss,
859                  struct page **prealloc)
860 {
861         struct mm_struct *src_mm = src_vma->vm_mm;
862         unsigned long vm_flags = src_vma->vm_flags;
863         pte_t pte = *src_pte;
864         struct page *page;
865
866         page = vm_normal_page(src_vma, addr, pte);
867         if (page) {
868                 int retval;
869
870                 retval = copy_present_page(dst_vma, src_vma, dst_pte, src_pte,
871                                            addr, rss, prealloc, pte, page);
872                 if (retval <= 0)
873                         return retval;
874
875                 get_page(page);
876                 page_dup_rmap(page, false);
877                 rss[mm_counter(page)]++;
878         }
879
880         /*
881          * If it's a COW mapping, write protect it both
882          * in the parent and the child
883          */
884         if (is_cow_mapping(vm_flags) && pte_write(pte)) {
885                 ptep_set_wrprotect(src_mm, addr, src_pte);
886                 pte = pte_wrprotect(pte);
887         }
888
889         /*
890          * If it's a shared mapping, mark it clean in
891          * the child
892          */
893         if (vm_flags & VM_SHARED)
894                 pte = pte_mkclean(pte);
895         pte = pte_mkold(pte);
896
897         /*
898          * Make sure the _PAGE_UFFD_WP bit is cleared if the new VMA
899          * does not have the VM_UFFD_WP, which means that the uffd
900          * fork event is not enabled.
901          */
902         if (!(vm_flags & VM_UFFD_WP))
903                 pte = pte_clear_uffd_wp(pte);
904
905         set_pte_at(dst_vma->vm_mm, addr, dst_pte, pte);
906         return 0;
907 }
908
909 static inline struct page *
910 page_copy_prealloc(struct mm_struct *src_mm, struct vm_area_struct *vma,
911                    unsigned long addr)
912 {
913         struct page *new_page;
914
915         new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, addr);
916         if (!new_page)
917                 return NULL;
918
919         if (mem_cgroup_charge(new_page, src_mm, GFP_KERNEL)) {
920                 put_page(new_page);
921                 return NULL;
922         }
923         cgroup_throttle_swaprate(new_page, GFP_KERNEL);
924
925         return new_page;
926 }
927
928 static int
929 copy_pte_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
930                pmd_t *dst_pmd, pmd_t *src_pmd, unsigned long addr,
931                unsigned long end)
932 {
933         struct mm_struct *dst_mm = dst_vma->vm_mm;
934         struct mm_struct *src_mm = src_vma->vm_mm;
935         pte_t *orig_src_pte, *orig_dst_pte;
936         pte_t *src_pte, *dst_pte;
937         spinlock_t *src_ptl, *dst_ptl;
938         int progress, ret = 0;
939         int rss[NR_MM_COUNTERS];
940         swp_entry_t entry = (swp_entry_t){0};
941         struct page *prealloc = NULL;
942
943 again:
944         progress = 0;
945         init_rss_vec(rss);
946
947         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
948         if (!dst_pte) {
949                 ret = -ENOMEM;
950                 goto out;
951         }
952         src_pte = pte_offset_map(src_pmd, addr);
953         src_ptl = pte_lockptr(src_mm, src_pmd);
954         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
955         orig_src_pte = src_pte;
956         orig_dst_pte = dst_pte;
957         arch_enter_lazy_mmu_mode();
958
959         do {
960                 /*
961                  * We are holding two locks at this point - either of them
962                  * could generate latencies in another task on another CPU.
963                  */
964                 if (progress >= 32) {
965                         progress = 0;
966                         if (need_resched() ||
967                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
968                                 break;
969                 }
970                 if (pte_none(*src_pte)) {
971                         progress++;
972                         continue;
973                 }
974                 if (unlikely(!pte_present(*src_pte))) {
975                         entry.val = copy_nonpresent_pte(dst_mm, src_mm,
976                                                         dst_pte, src_pte,
977                                                         src_vma, addr, rss);
978                         if (entry.val)
979                                 break;
980                         progress += 8;
981                         continue;
982                 }
983                 /* copy_present_pte() will clear `*prealloc' if consumed */
984                 ret = copy_present_pte(dst_vma, src_vma, dst_pte, src_pte,
985                                        addr, rss, &prealloc);
986                 /*
987                  * If we need a pre-allocated page for this pte, drop the
988                  * locks, allocate, and try again.
989                  */
990                 if (unlikely(ret == -EAGAIN))
991                         break;
992                 if (unlikely(prealloc)) {
993                         /*
994                          * pre-alloc page cannot be reused by next time so as
995                          * to strictly follow mempolicy (e.g., alloc_page_vma()
996                          * will allocate page according to address).  This
997                          * could only happen if one pinned pte changed.
998                          */
999                         put_page(prealloc);
1000                         prealloc = NULL;
1001                 }
1002                 progress += 8;
1003         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1004
1005         arch_leave_lazy_mmu_mode();
1006         spin_unlock(src_ptl);
1007         pte_unmap(orig_src_pte);
1008         add_mm_rss_vec(dst_mm, rss);
1009         pte_unmap_unlock(orig_dst_pte, dst_ptl);
1010         cond_resched();
1011
1012         if (entry.val) {
1013                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) {
1014                         ret = -ENOMEM;
1015                         goto out;
1016                 }
1017                 entry.val = 0;
1018         } else if (ret) {
1019                 WARN_ON_ONCE(ret != -EAGAIN);
1020                 prealloc = page_copy_prealloc(src_mm, src_vma, addr);
1021                 if (!prealloc)
1022                         return -ENOMEM;
1023                 /* We've captured and resolved the error. Reset, try again. */
1024                 ret = 0;
1025         }
1026         if (addr != end)
1027                 goto again;
1028 out:
1029         if (unlikely(prealloc))
1030                 put_page(prealloc);
1031         return ret;
1032 }
1033
1034 static inline int
1035 copy_pmd_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1036                pud_t *dst_pud, pud_t *src_pud, unsigned long addr,
1037                unsigned long end)
1038 {
1039         struct mm_struct *dst_mm = dst_vma->vm_mm;
1040         struct mm_struct *src_mm = src_vma->vm_mm;
1041         pmd_t *src_pmd, *dst_pmd;
1042         unsigned long next;
1043
1044         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1045         if (!dst_pmd)
1046                 return -ENOMEM;
1047         src_pmd = pmd_offset(src_pud, addr);
1048         do {
1049                 next = pmd_addr_end(addr, end);
1050                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
1051                         || pmd_devmap(*src_pmd)) {
1052                         int err;
1053                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, src_vma);
1054                         err = copy_huge_pmd(dst_mm, src_mm,
1055                                             dst_pmd, src_pmd, addr, src_vma);
1056                         if (err == -ENOMEM)
1057                                 return -ENOMEM;
1058                         if (!err)
1059                                 continue;
1060                         /* fall through */
1061                 }
1062                 if (pmd_none_or_clear_bad(src_pmd))
1063                         continue;
1064                 if (copy_pte_range(dst_vma, src_vma, dst_pmd, src_pmd,
1065                                    addr, next))
1066                         return -ENOMEM;
1067         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1068         return 0;
1069 }
1070
1071 static inline int
1072 copy_pud_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1073                p4d_t *dst_p4d, p4d_t *src_p4d, unsigned long addr,
1074                unsigned long end)
1075 {
1076         struct mm_struct *dst_mm = dst_vma->vm_mm;
1077         struct mm_struct *src_mm = src_vma->vm_mm;
1078         pud_t *src_pud, *dst_pud;
1079         unsigned long next;
1080
1081         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
1082         if (!dst_pud)
1083                 return -ENOMEM;
1084         src_pud = pud_offset(src_p4d, addr);
1085         do {
1086                 next = pud_addr_end(addr, end);
1087                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1088                         int err;
1089
1090                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, src_vma);
1091                         err = copy_huge_pud(dst_mm, src_mm,
1092                                             dst_pud, src_pud, addr, src_vma);
1093                         if (err == -ENOMEM)
1094                                 return -ENOMEM;
1095                         if (!err)
1096                                 continue;
1097                         /* fall through */
1098                 }
1099                 if (pud_none_or_clear_bad(src_pud))
1100                         continue;
1101                 if (copy_pmd_range(dst_vma, src_vma, dst_pud, src_pud,
1102                                    addr, next))
1103                         return -ENOMEM;
1104         } while (dst_pud++, src_pud++, addr = next, addr != end);
1105         return 0;
1106 }
1107
1108 static inline int
1109 copy_p4d_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma,
1110                pgd_t *dst_pgd, pgd_t *src_pgd, unsigned long addr,
1111                unsigned long end)
1112 {
1113         struct mm_struct *dst_mm = dst_vma->vm_mm;
1114         p4d_t *src_p4d, *dst_p4d;
1115         unsigned long next;
1116
1117         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
1118         if (!dst_p4d)
1119                 return -ENOMEM;
1120         src_p4d = p4d_offset(src_pgd, addr);
1121         do {
1122                 next = p4d_addr_end(addr, end);
1123                 if (p4d_none_or_clear_bad(src_p4d))
1124                         continue;
1125                 if (copy_pud_range(dst_vma, src_vma, dst_p4d, src_p4d,
1126                                    addr, next))
1127                         return -ENOMEM;
1128         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
1129         return 0;
1130 }
1131
1132 int
1133 copy_page_range(struct vm_area_struct *dst_vma, struct vm_area_struct *src_vma)
1134 {
1135         pgd_t *src_pgd, *dst_pgd;
1136         unsigned long next;
1137         unsigned long addr = src_vma->vm_start;
1138         unsigned long end = src_vma->vm_end;
1139         struct mm_struct *dst_mm = dst_vma->vm_mm;
1140         struct mm_struct *src_mm = src_vma->vm_mm;
1141         struct mmu_notifier_range range;
1142         bool is_cow;
1143         int ret;
1144
1145         /*
1146          * Don't copy ptes where a page fault will fill them correctly.
1147          * Fork becomes much lighter when there are big shared or private
1148          * readonly mappings. The tradeoff is that copy_page_range is more
1149          * efficient than faulting.
1150          */
1151         if (!(src_vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1152             !src_vma->anon_vma)
1153                 return 0;
1154
1155         if (is_vm_hugetlb_page(src_vma))
1156                 return copy_hugetlb_page_range(dst_mm, src_mm, src_vma);
1157
1158         if (unlikely(src_vma->vm_flags & VM_PFNMAP)) {
1159                 /*
1160                  * We do not free on error cases below as remove_vma
1161                  * gets called on error from higher level routine
1162                  */
1163                 ret = track_pfn_copy(src_vma);
1164                 if (ret)
1165                         return ret;
1166         }
1167
1168         /*
1169          * We need to invalidate the secondary MMU mappings only when
1170          * there could be a permission downgrade on the ptes of the
1171          * parent mm. And a permission downgrade will only happen if
1172          * is_cow_mapping() returns true.
1173          */
1174         is_cow = is_cow_mapping(src_vma->vm_flags);
1175
1176         if (is_cow) {
1177                 mmu_notifier_range_init(&range, MMU_NOTIFY_PROTECTION_PAGE,
1178                                         0, src_vma, src_mm, addr, end);
1179                 mmu_notifier_invalidate_range_start(&range);
1180                 /*
1181                  * Disabling preemption is not needed for the write side, as
1182                  * the read side doesn't spin, but goes to the mmap_lock.
1183                  *
1184                  * Use the raw variant of the seqcount_t write API to avoid
1185                  * lockdep complaining about preemptibility.
1186                  */
1187                 mmap_assert_write_locked(src_mm);
1188                 raw_write_seqcount_begin(&src_mm->write_protect_seq);
1189         }
1190
1191         ret = 0;
1192         dst_pgd = pgd_offset(dst_mm, addr);
1193         src_pgd = pgd_offset(src_mm, addr);
1194         do {
1195                 next = pgd_addr_end(addr, end);
1196                 if (pgd_none_or_clear_bad(src_pgd))
1197                         continue;
1198                 if (unlikely(copy_p4d_range(dst_vma, src_vma, dst_pgd, src_pgd,
1199                                             addr, next))) {
1200                         ret = -ENOMEM;
1201                         break;
1202                 }
1203         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1204
1205         if (is_cow) {
1206                 raw_write_seqcount_end(&src_mm->write_protect_seq);
1207                 mmu_notifier_invalidate_range_end(&range);
1208         }
1209         return ret;
1210 }
1211
1212 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1213                                 struct vm_area_struct *vma, pmd_t *pmd,
1214                                 unsigned long addr, unsigned long end,
1215                                 struct zap_details *details)
1216 {
1217         struct mm_struct *mm = tlb->mm;
1218         int force_flush = 0;
1219         int rss[NR_MM_COUNTERS];
1220         spinlock_t *ptl;
1221         pte_t *start_pte;
1222         pte_t *pte;
1223         swp_entry_t entry;
1224
1225         tlb_change_page_size(tlb, PAGE_SIZE);
1226 again:
1227         init_rss_vec(rss);
1228         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1229         pte = start_pte;
1230         flush_tlb_batched_pending(mm);
1231         arch_enter_lazy_mmu_mode();
1232         do {
1233                 pte_t ptent = *pte;
1234                 if (pte_none(ptent))
1235                         continue;
1236
1237                 if (need_resched())
1238                         break;
1239
1240                 if (pte_present(ptent)) {
1241                         struct page *page;
1242
1243                         page = vm_normal_page(vma, addr, ptent);
1244                         if (unlikely(details) && page) {
1245                                 /*
1246                                  * unmap_shared_mapping_pages() wants to
1247                                  * invalidate cache without truncating:
1248                                  * unmap shared but keep private pages.
1249                                  */
1250                                 if (details->check_mapping &&
1251                                     details->check_mapping != page_rmapping(page))
1252                                         continue;
1253                         }
1254                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1255                                                         tlb->fullmm);
1256                         tlb_remove_tlb_entry(tlb, pte, addr);
1257                         if (unlikely(!page))
1258                                 continue;
1259
1260                         if (!PageAnon(page)) {
1261                                 if (pte_dirty(ptent)) {
1262                                         force_flush = 1;
1263                                         set_page_dirty(page);
1264                                 }
1265                                 if (pte_young(ptent) &&
1266                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1267                                         mark_page_accessed(page);
1268                         }
1269                         rss[mm_counter(page)]--;
1270                         page_remove_rmap(page, false);
1271                         if (unlikely(page_mapcount(page) < 0))
1272                                 print_bad_pte(vma, addr, ptent, page);
1273                         if (unlikely(__tlb_remove_page(tlb, page))) {
1274                                 force_flush = 1;
1275                                 addr += PAGE_SIZE;
1276                                 break;
1277                         }
1278                         continue;
1279                 }
1280
1281                 entry = pte_to_swp_entry(ptent);
1282                 if (is_device_private_entry(entry)) {
1283                         struct page *page = device_private_entry_to_page(entry);
1284
1285                         if (unlikely(details && details->check_mapping)) {
1286                                 /*
1287                                  * unmap_shared_mapping_pages() wants to
1288                                  * invalidate cache without truncating:
1289                                  * unmap shared but keep private pages.
1290                                  */
1291                                 if (details->check_mapping !=
1292                                     page_rmapping(page))
1293                                         continue;
1294                         }
1295
1296                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1297                         rss[mm_counter(page)]--;
1298                         page_remove_rmap(page, false);
1299                         put_page(page);
1300                         continue;
1301                 }
1302
1303                 /* If details->check_mapping, we leave swap entries. */
1304                 if (unlikely(details))
1305                         continue;
1306
1307                 if (!non_swap_entry(entry))
1308                         rss[MM_SWAPENTS]--;
1309                 else if (is_migration_entry(entry)) {
1310                         struct page *page;
1311
1312                         page = migration_entry_to_page(entry);
1313                         rss[mm_counter(page)]--;
1314                 }
1315                 if (unlikely(!free_swap_and_cache(entry)))
1316                         print_bad_pte(vma, addr, ptent, NULL);
1317                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1318         } while (pte++, addr += PAGE_SIZE, addr != end);
1319
1320         add_mm_rss_vec(mm, rss);
1321         arch_leave_lazy_mmu_mode();
1322
1323         /* Do the actual TLB flush before dropping ptl */
1324         if (force_flush)
1325                 tlb_flush_mmu_tlbonly(tlb);
1326         pte_unmap_unlock(start_pte, ptl);
1327
1328         /*
1329          * If we forced a TLB flush (either due to running out of
1330          * batch buffers or because we needed to flush dirty TLB
1331          * entries before releasing the ptl), free the batched
1332          * memory too. Restart if we didn't do everything.
1333          */
1334         if (force_flush) {
1335                 force_flush = 0;
1336                 tlb_flush_mmu(tlb);
1337         }
1338
1339         if (addr != end) {
1340                 cond_resched();
1341                 goto again;
1342         }
1343
1344         return addr;
1345 }
1346
1347 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1348                                 struct vm_area_struct *vma, pud_t *pud,
1349                                 unsigned long addr, unsigned long end,
1350                                 struct zap_details *details)
1351 {
1352         pmd_t *pmd;
1353         unsigned long next;
1354
1355         pmd = pmd_offset(pud, addr);
1356         do {
1357                 next = pmd_addr_end(addr, end);
1358                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1359                         if (next - addr != HPAGE_PMD_SIZE)
1360                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1361                         else if (zap_huge_pmd(tlb, vma, pmd, addr))
1362                                 goto next;
1363                         /* fall through */
1364                 }
1365                 /*
1366                  * Here there can be other concurrent MADV_DONTNEED or
1367                  * trans huge page faults running, and if the pmd is
1368                  * none or trans huge it can change under us. This is
1369                  * because MADV_DONTNEED holds the mmap_lock in read
1370                  * mode.
1371                  */
1372                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1373                         goto next;
1374                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1375 next:
1376                 cond_resched();
1377         } while (pmd++, addr = next, addr != end);
1378
1379         return addr;
1380 }
1381
1382 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1383                                 struct vm_area_struct *vma, p4d_t *p4d,
1384                                 unsigned long addr, unsigned long end,
1385                                 struct zap_details *details)
1386 {
1387         pud_t *pud;
1388         unsigned long next;
1389
1390         pud = pud_offset(p4d, addr);
1391         do {
1392                 next = pud_addr_end(addr, end);
1393                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1394                         if (next - addr != HPAGE_PUD_SIZE) {
1395                                 mmap_assert_locked(tlb->mm);
1396                                 split_huge_pud(vma, pud, addr);
1397                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1398                                 goto next;
1399                         /* fall through */
1400                 }
1401                 if (pud_none_or_clear_bad(pud))
1402                         continue;
1403                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1404 next:
1405                 cond_resched();
1406         } while (pud++, addr = next, addr != end);
1407
1408         return addr;
1409 }
1410
1411 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1412                                 struct vm_area_struct *vma, pgd_t *pgd,
1413                                 unsigned long addr, unsigned long end,
1414                                 struct zap_details *details)
1415 {
1416         p4d_t *p4d;
1417         unsigned long next;
1418
1419         p4d = p4d_offset(pgd, addr);
1420         do {
1421                 next = p4d_addr_end(addr, end);
1422                 if (p4d_none_or_clear_bad(p4d))
1423                         continue;
1424                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1425         } while (p4d++, addr = next, addr != end);
1426
1427         return addr;
1428 }
1429
1430 void unmap_page_range(struct mmu_gather *tlb,
1431                              struct vm_area_struct *vma,
1432                              unsigned long addr, unsigned long end,
1433                              struct zap_details *details)
1434 {
1435         pgd_t *pgd;
1436         unsigned long next;
1437
1438         BUG_ON(addr >= end);
1439         tlb_start_vma(tlb, vma);
1440         pgd = pgd_offset(vma->vm_mm, addr);
1441         do {
1442                 next = pgd_addr_end(addr, end);
1443                 if (pgd_none_or_clear_bad(pgd))
1444                         continue;
1445                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1446         } while (pgd++, addr = next, addr != end);
1447         tlb_end_vma(tlb, vma);
1448 }
1449
1450
1451 static void unmap_single_vma(struct mmu_gather *tlb,
1452                 struct vm_area_struct *vma, unsigned long start_addr,
1453                 unsigned long end_addr,
1454                 struct zap_details *details)
1455 {
1456         unsigned long start = max(vma->vm_start, start_addr);
1457         unsigned long end;
1458
1459         if (start >= vma->vm_end)
1460                 return;
1461         end = min(vma->vm_end, end_addr);
1462         if (end <= vma->vm_start)
1463                 return;
1464
1465         if (vma->vm_file)
1466                 uprobe_munmap(vma, start, end);
1467
1468         if (unlikely(vma->vm_flags & VM_PFNMAP))
1469                 untrack_pfn(vma, 0, 0);
1470
1471         if (start != end) {
1472                 if (unlikely(is_vm_hugetlb_page(vma))) {
1473                         /*
1474                          * It is undesirable to test vma->vm_file as it
1475                          * should be non-null for valid hugetlb area.
1476                          * However, vm_file will be NULL in the error
1477                          * cleanup path of mmap_region. When
1478                          * hugetlbfs ->mmap method fails,
1479                          * mmap_region() nullifies vma->vm_file
1480                          * before calling this function to clean up.
1481                          * Since no pte has actually been setup, it is
1482                          * safe to do nothing in this case.
1483                          */
1484                         if (vma->vm_file) {
1485                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1486                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1487                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1488                         }
1489                 } else
1490                         unmap_page_range(tlb, vma, start, end, details);
1491         }
1492 }
1493
1494 /**
1495  * unmap_vmas - unmap a range of memory covered by a list of vma's
1496  * @tlb: address of the caller's struct mmu_gather
1497  * @vma: the starting vma
1498  * @start_addr: virtual address at which to start unmapping
1499  * @end_addr: virtual address at which to end unmapping
1500  *
1501  * Unmap all pages in the vma list.
1502  *
1503  * Only addresses between `start' and `end' will be unmapped.
1504  *
1505  * The VMA list must be sorted in ascending virtual address order.
1506  *
1507  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1508  * range after unmap_vmas() returns.  So the only responsibility here is to
1509  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1510  * drops the lock and schedules.
1511  */
1512 void unmap_vmas(struct mmu_gather *tlb,
1513                 struct vm_area_struct *vma, unsigned long start_addr,
1514                 unsigned long end_addr)
1515 {
1516         struct mmu_notifier_range range;
1517
1518         mmu_notifier_range_init(&range, MMU_NOTIFY_UNMAP, 0, vma, vma->vm_mm,
1519                                 start_addr, end_addr);
1520         mmu_notifier_invalidate_range_start(&range);
1521         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1522                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1523         mmu_notifier_invalidate_range_end(&range);
1524 }
1525
1526 /**
1527  * zap_page_range - remove user pages in a given range
1528  * @vma: vm_area_struct holding the applicable pages
1529  * @start: starting address of pages to zap
1530  * @size: number of bytes to zap
1531  *
1532  * Caller must protect the VMA list
1533  */
1534 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1535                 unsigned long size)
1536 {
1537         struct mmu_notifier_range range;
1538         struct mmu_gather tlb;
1539
1540         lru_add_drain();
1541         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1542                                 start, start + size);
1543         tlb_gather_mmu(&tlb, vma->vm_mm);
1544         update_hiwater_rss(vma->vm_mm);
1545         mmu_notifier_invalidate_range_start(&range);
1546         for ( ; vma && vma->vm_start < range.end; vma = vma->vm_next)
1547                 unmap_single_vma(&tlb, vma, start, range.end, NULL);
1548         mmu_notifier_invalidate_range_end(&range);
1549         tlb_finish_mmu(&tlb);
1550 }
1551
1552 /**
1553  * zap_page_range_single - remove user pages in a given range
1554  * @vma: vm_area_struct holding the applicable pages
1555  * @address: starting address of pages to zap
1556  * @size: number of bytes to zap
1557  * @details: details of shared cache invalidation
1558  *
1559  * The range must fit into one VMA.
1560  */
1561 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1562                 unsigned long size, struct zap_details *details)
1563 {
1564         struct mmu_notifier_range range;
1565         struct mmu_gather tlb;
1566
1567         lru_add_drain();
1568         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, vma->vm_mm,
1569                                 address, address + size);
1570         tlb_gather_mmu(&tlb, vma->vm_mm);
1571         update_hiwater_rss(vma->vm_mm);
1572         mmu_notifier_invalidate_range_start(&range);
1573         unmap_single_vma(&tlb, vma, address, range.end, details);
1574         mmu_notifier_invalidate_range_end(&range);
1575         tlb_finish_mmu(&tlb);
1576 }
1577
1578 /**
1579  * zap_vma_ptes - remove ptes mapping the vma
1580  * @vma: vm_area_struct holding ptes to be zapped
1581  * @address: starting address of pages to zap
1582  * @size: number of bytes to zap
1583  *
1584  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1585  *
1586  * The entire address range must be fully contained within the vma.
1587  *
1588  */
1589 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1590                 unsigned long size)
1591 {
1592         if (address < vma->vm_start || address + size > vma->vm_end ||
1593                         !(vma->vm_flags & VM_PFNMAP))
1594                 return;
1595
1596         zap_page_range_single(vma, address, size, NULL);
1597 }
1598 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1599
1600 static pmd_t *walk_to_pmd(struct mm_struct *mm, unsigned long addr)
1601 {
1602         pgd_t *pgd;
1603         p4d_t *p4d;
1604         pud_t *pud;
1605         pmd_t *pmd;
1606
1607         pgd = pgd_offset(mm, addr);
1608         p4d = p4d_alloc(mm, pgd, addr);
1609         if (!p4d)
1610                 return NULL;
1611         pud = pud_alloc(mm, p4d, addr);
1612         if (!pud)
1613                 return NULL;
1614         pmd = pmd_alloc(mm, pud, addr);
1615         if (!pmd)
1616                 return NULL;
1617
1618         VM_BUG_ON(pmd_trans_huge(*pmd));
1619         return pmd;
1620 }
1621
1622 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1623                         spinlock_t **ptl)
1624 {
1625         pmd_t *pmd = walk_to_pmd(mm, addr);
1626
1627         if (!pmd)
1628                 return NULL;
1629         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1630 }
1631
1632 static int validate_page_before_insert(struct page *page)
1633 {
1634         if (PageAnon(page) || PageSlab(page) || page_has_type(page))
1635                 return -EINVAL;
1636         flush_dcache_page(page);
1637         return 0;
1638 }
1639
1640 static int insert_page_into_pte_locked(struct mm_struct *mm, pte_t *pte,
1641                         unsigned long addr, struct page *page, pgprot_t prot)
1642 {
1643         if (!pte_none(*pte))
1644                 return -EBUSY;
1645         /* Ok, finally just insert the thing.. */
1646         get_page(page);
1647         inc_mm_counter_fast(mm, mm_counter_file(page));
1648         page_add_file_rmap(page, false);
1649         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1650         return 0;
1651 }
1652
1653 /*
1654  * This is the old fallback for page remapping.
1655  *
1656  * For historical reasons, it only allows reserved pages. Only
1657  * old drivers should use this, and they needed to mark their
1658  * pages reserved for the old functions anyway.
1659  */
1660 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1661                         struct page *page, pgprot_t prot)
1662 {
1663         struct mm_struct *mm = vma->vm_mm;
1664         int retval;
1665         pte_t *pte;
1666         spinlock_t *ptl;
1667
1668         retval = validate_page_before_insert(page);
1669         if (retval)
1670                 goto out;
1671         retval = -ENOMEM;
1672         pte = get_locked_pte(mm, addr, &ptl);
1673         if (!pte)
1674                 goto out;
1675         retval = insert_page_into_pte_locked(mm, pte, addr, page, prot);
1676         pte_unmap_unlock(pte, ptl);
1677 out:
1678         return retval;
1679 }
1680
1681 #ifdef pte_index
1682 static int insert_page_in_batch_locked(struct mm_struct *mm, pte_t *pte,
1683                         unsigned long addr, struct page *page, pgprot_t prot)
1684 {
1685         int err;
1686
1687         if (!page_count(page))
1688                 return -EINVAL;
1689         err = validate_page_before_insert(page);
1690         if (err)
1691                 return err;
1692         return insert_page_into_pte_locked(mm, pte, addr, page, prot);
1693 }
1694
1695 /* insert_pages() amortizes the cost of spinlock operations
1696  * when inserting pages in a loop. Arch *must* define pte_index.
1697  */
1698 static int insert_pages(struct vm_area_struct *vma, unsigned long addr,
1699                         struct page **pages, unsigned long *num, pgprot_t prot)
1700 {
1701         pmd_t *pmd = NULL;
1702         pte_t *start_pte, *pte;
1703         spinlock_t *pte_lock;
1704         struct mm_struct *const mm = vma->vm_mm;
1705         unsigned long curr_page_idx = 0;
1706         unsigned long remaining_pages_total = *num;
1707         unsigned long pages_to_write_in_pmd;
1708         int ret;
1709 more:
1710         ret = -EFAULT;
1711         pmd = walk_to_pmd(mm, addr);
1712         if (!pmd)
1713                 goto out;
1714
1715         pages_to_write_in_pmd = min_t(unsigned long,
1716                 remaining_pages_total, PTRS_PER_PTE - pte_index(addr));
1717
1718         /* Allocate the PTE if necessary; takes PMD lock once only. */
1719         ret = -ENOMEM;
1720         if (pte_alloc(mm, pmd))
1721                 goto out;
1722
1723         while (pages_to_write_in_pmd) {
1724                 int pte_idx = 0;
1725                 const int batch_size = min_t(int, pages_to_write_in_pmd, 8);
1726
1727                 start_pte = pte_offset_map_lock(mm, pmd, addr, &pte_lock);
1728                 for (pte = start_pte; pte_idx < batch_size; ++pte, ++pte_idx) {
1729                         int err = insert_page_in_batch_locked(mm, pte,
1730                                 addr, pages[curr_page_idx], prot);
1731                         if (unlikely(err)) {
1732                                 pte_unmap_unlock(start_pte, pte_lock);
1733                                 ret = err;
1734                                 remaining_pages_total -= pte_idx;
1735                                 goto out;
1736                         }
1737                         addr += PAGE_SIZE;
1738                         ++curr_page_idx;
1739                 }
1740                 pte_unmap_unlock(start_pte, pte_lock);
1741                 pages_to_write_in_pmd -= batch_size;
1742                 remaining_pages_total -= batch_size;
1743         }
1744         if (remaining_pages_total)
1745                 goto more;
1746         ret = 0;
1747 out:
1748         *num = remaining_pages_total;
1749         return ret;
1750 }
1751 #endif  /* ifdef pte_index */
1752
1753 /**
1754  * vm_insert_pages - insert multiple pages into user vma, batching the pmd lock.
1755  * @vma: user vma to map to
1756  * @addr: target start user address of these pages
1757  * @pages: source kernel pages
1758  * @num: in: number of pages to map. out: number of pages that were *not*
1759  * mapped. (0 means all pages were successfully mapped).
1760  *
1761  * Preferred over vm_insert_page() when inserting multiple pages.
1762  *
1763  * In case of error, we may have mapped a subset of the provided
1764  * pages. It is the caller's responsibility to account for this case.
1765  *
1766  * The same restrictions apply as in vm_insert_page().
1767  */
1768 int vm_insert_pages(struct vm_area_struct *vma, unsigned long addr,
1769                         struct page **pages, unsigned long *num)
1770 {
1771 #ifdef pte_index
1772         const unsigned long end_addr = addr + (*num * PAGE_SIZE) - 1;
1773
1774         if (addr < vma->vm_start || end_addr >= vma->vm_end)
1775                 return -EFAULT;
1776         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1777                 BUG_ON(mmap_read_trylock(vma->vm_mm));
1778                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1779                 vma->vm_flags |= VM_MIXEDMAP;
1780         }
1781         /* Defer page refcount checking till we're about to map that page. */
1782         return insert_pages(vma, addr, pages, num, vma->vm_page_prot);
1783 #else
1784         unsigned long idx = 0, pgcount = *num;
1785         int err = -EINVAL;
1786
1787         for (; idx < pgcount; ++idx) {
1788                 err = vm_insert_page(vma, addr + (PAGE_SIZE * idx), pages[idx]);
1789                 if (err)
1790                         break;
1791         }
1792         *num = pgcount - idx;
1793         return err;
1794 #endif  /* ifdef pte_index */
1795 }
1796 EXPORT_SYMBOL(vm_insert_pages);
1797
1798 /**
1799  * vm_insert_page - insert single page into user vma
1800  * @vma: user vma to map to
1801  * @addr: target user address of this page
1802  * @page: source kernel page
1803  *
1804  * This allows drivers to insert individual pages they've allocated
1805  * into a user vma.
1806  *
1807  * The page has to be a nice clean _individual_ kernel allocation.
1808  * If you allocate a compound page, you need to have marked it as
1809  * such (__GFP_COMP), or manually just split the page up yourself
1810  * (see split_page()).
1811  *
1812  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1813  * took an arbitrary page protection parameter. This doesn't allow
1814  * that. Your vma protection will have to be set up correctly, which
1815  * means that if you want a shared writable mapping, you'd better
1816  * ask for a shared writable mapping!
1817  *
1818  * The page does not need to be reserved.
1819  *
1820  * Usually this function is called from f_op->mmap() handler
1821  * under mm->mmap_lock write-lock, so it can change vma->vm_flags.
1822  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1823  * function from other places, for example from page-fault handler.
1824  *
1825  * Return: %0 on success, negative error code otherwise.
1826  */
1827 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1828                         struct page *page)
1829 {
1830         if (addr < vma->vm_start || addr >= vma->vm_end)
1831                 return -EFAULT;
1832         if (!page_count(page))
1833                 return -EINVAL;
1834         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1835                 BUG_ON(mmap_read_trylock(vma->vm_mm));
1836                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1837                 vma->vm_flags |= VM_MIXEDMAP;
1838         }
1839         return insert_page(vma, addr, page, vma->vm_page_prot);
1840 }
1841 EXPORT_SYMBOL(vm_insert_page);
1842
1843 /*
1844  * __vm_map_pages - maps range of kernel pages into user vma
1845  * @vma: user vma to map to
1846  * @pages: pointer to array of source kernel pages
1847  * @num: number of pages in page array
1848  * @offset: user's requested vm_pgoff
1849  *
1850  * This allows drivers to map range of kernel pages into a user vma.
1851  *
1852  * Return: 0 on success and error code otherwise.
1853  */
1854 static int __vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1855                                 unsigned long num, unsigned long offset)
1856 {
1857         unsigned long count = vma_pages(vma);
1858         unsigned long uaddr = vma->vm_start;
1859         int ret, i;
1860
1861         /* Fail if the user requested offset is beyond the end of the object */
1862         if (offset >= num)
1863                 return -ENXIO;
1864
1865         /* Fail if the user requested size exceeds available object size */
1866         if (count > num - offset)
1867                 return -ENXIO;
1868
1869         for (i = 0; i < count; i++) {
1870                 ret = vm_insert_page(vma, uaddr, pages[offset + i]);
1871                 if (ret < 0)
1872                         return ret;
1873                 uaddr += PAGE_SIZE;
1874         }
1875
1876         return 0;
1877 }
1878
1879 /**
1880  * vm_map_pages - maps range of kernel pages starts with non zero offset
1881  * @vma: user vma to map to
1882  * @pages: pointer to array of source kernel pages
1883  * @num: number of pages in page array
1884  *
1885  * Maps an object consisting of @num pages, catering for the user's
1886  * requested vm_pgoff
1887  *
1888  * If we fail to insert any page into the vma, the function will return
1889  * immediately leaving any previously inserted pages present.  Callers
1890  * from the mmap handler may immediately return the error as their caller
1891  * will destroy the vma, removing any successfully inserted pages. Other
1892  * callers should make their own arrangements for calling unmap_region().
1893  *
1894  * Context: Process context. Called by mmap handlers.
1895  * Return: 0 on success and error code otherwise.
1896  */
1897 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
1898                                 unsigned long num)
1899 {
1900         return __vm_map_pages(vma, pages, num, vma->vm_pgoff);
1901 }
1902 EXPORT_SYMBOL(vm_map_pages);
1903
1904 /**
1905  * vm_map_pages_zero - map range of kernel pages starts with zero offset
1906  * @vma: user vma to map to
1907  * @pages: pointer to array of source kernel pages
1908  * @num: number of pages in page array
1909  *
1910  * Similar to vm_map_pages(), except that it explicitly sets the offset
1911  * to 0. This function is intended for the drivers that did not consider
1912  * vm_pgoff.
1913  *
1914  * Context: Process context. Called by mmap handlers.
1915  * Return: 0 on success and error code otherwise.
1916  */
1917 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
1918                                 unsigned long num)
1919 {
1920         return __vm_map_pages(vma, pages, num, 0);
1921 }
1922 EXPORT_SYMBOL(vm_map_pages_zero);
1923
1924 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1925                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1926 {
1927         struct mm_struct *mm = vma->vm_mm;
1928         pte_t *pte, entry;
1929         spinlock_t *ptl;
1930
1931         pte = get_locked_pte(mm, addr, &ptl);
1932         if (!pte)
1933                 return VM_FAULT_OOM;
1934         if (!pte_none(*pte)) {
1935                 if (mkwrite) {
1936                         /*
1937                          * For read faults on private mappings the PFN passed
1938                          * in may not match the PFN we have mapped if the
1939                          * mapped PFN is a writeable COW page.  In the mkwrite
1940                          * case we are creating a writable PTE for a shared
1941                          * mapping and we expect the PFNs to match. If they
1942                          * don't match, we are likely racing with block
1943                          * allocation and mapping invalidation so just skip the
1944                          * update.
1945                          */
1946                         if (pte_pfn(*pte) != pfn_t_to_pfn(pfn)) {
1947                                 WARN_ON_ONCE(!is_zero_pfn(pte_pfn(*pte)));
1948                                 goto out_unlock;
1949                         }
1950                         entry = pte_mkyoung(*pte);
1951                         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1952                         if (ptep_set_access_flags(vma, addr, pte, entry, 1))
1953                                 update_mmu_cache(vma, addr, pte);
1954                 }
1955                 goto out_unlock;
1956         }
1957
1958         /* Ok, finally just insert the thing.. */
1959         if (pfn_t_devmap(pfn))
1960                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1961         else
1962                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1963
1964         if (mkwrite) {
1965                 entry = pte_mkyoung(entry);
1966                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1967         }
1968
1969         set_pte_at(mm, addr, pte, entry);
1970         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1971
1972 out_unlock:
1973         pte_unmap_unlock(pte, ptl);
1974         return VM_FAULT_NOPAGE;
1975 }
1976
1977 /**
1978  * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1979  * @vma: user vma to map to
1980  * @addr: target user address of this page
1981  * @pfn: source kernel pfn
1982  * @pgprot: pgprot flags for the inserted page
1983  *
1984  * This is exactly like vmf_insert_pfn(), except that it allows drivers
1985  * to override pgprot on a per-page basis.
1986  *
1987  * This only makes sense for IO mappings, and it makes no sense for
1988  * COW mappings.  In general, using multiple vmas is preferable;
1989  * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1990  * impractical.
1991  *
1992  * See vmf_insert_mixed_prot() for a discussion of the implication of using
1993  * a value of @pgprot different from that of @vma->vm_page_prot.
1994  *
1995  * Context: Process context.  May allocate using %GFP_KERNEL.
1996  * Return: vm_fault_t value.
1997  */
1998 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1999                         unsigned long pfn, pgprot_t pgprot)
2000 {
2001         /*
2002          * Technically, architectures with pte_special can avoid all these
2003          * restrictions (same for remap_pfn_range).  However we would like
2004          * consistency in testing and feature parity among all, so we should
2005          * try to keep these invariants in place for everybody.
2006          */
2007         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
2008         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
2009                                                 (VM_PFNMAP|VM_MIXEDMAP));
2010         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
2011         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
2012
2013         if (addr < vma->vm_start || addr >= vma->vm_end)
2014                 return VM_FAULT_SIGBUS;
2015
2016         if (!pfn_modify_allowed(pfn, pgprot))
2017                 return VM_FAULT_SIGBUS;
2018
2019         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
2020
2021         return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
2022                         false);
2023 }
2024 EXPORT_SYMBOL(vmf_insert_pfn_prot);
2025
2026 /**
2027  * vmf_insert_pfn - insert single pfn into user vma
2028  * @vma: user vma to map to
2029  * @addr: target user address of this page
2030  * @pfn: source kernel pfn
2031  *
2032  * Similar to vm_insert_page, this allows drivers to insert individual pages
2033  * they've allocated into a user vma. Same comments apply.
2034  *
2035  * This function should only be called from a vm_ops->fault handler, and
2036  * in that case the handler should return the result of this function.
2037  *
2038  * vma cannot be a COW mapping.
2039  *
2040  * As this is called only for pages that do not currently exist, we
2041  * do not need to flush old virtual caches or the TLB.
2042  *
2043  * Context: Process context.  May allocate using %GFP_KERNEL.
2044  * Return: vm_fault_t value.
2045  */
2046 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
2047                         unsigned long pfn)
2048 {
2049         return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
2050 }
2051 EXPORT_SYMBOL(vmf_insert_pfn);
2052
2053 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
2054 {
2055         /* these checks mirror the abort conditions in vm_normal_page */
2056         if (vma->vm_flags & VM_MIXEDMAP)
2057                 return true;
2058         if (pfn_t_devmap(pfn))
2059                 return true;
2060         if (pfn_t_special(pfn))
2061                 return true;
2062         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
2063                 return true;
2064         return false;
2065 }
2066
2067 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
2068                 unsigned long addr, pfn_t pfn, pgprot_t pgprot,
2069                 bool mkwrite)
2070 {
2071         int err;
2072
2073         BUG_ON(!vm_mixed_ok(vma, pfn));
2074
2075         if (addr < vma->vm_start || addr >= vma->vm_end)
2076                 return VM_FAULT_SIGBUS;
2077
2078         track_pfn_insert(vma, &pgprot, pfn);
2079
2080         if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
2081                 return VM_FAULT_SIGBUS;
2082
2083         /*
2084          * If we don't have pte special, then we have to use the pfn_valid()
2085          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
2086          * refcount the page if pfn_valid is true (hence insert_page rather
2087          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
2088          * without pte special, it would there be refcounted as a normal page.
2089          */
2090         if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
2091             !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
2092                 struct page *page;
2093
2094                 /*
2095                  * At this point we are committed to insert_page()
2096                  * regardless of whether the caller specified flags that
2097                  * result in pfn_t_has_page() == false.
2098                  */
2099                 page = pfn_to_page(pfn_t_to_pfn(pfn));
2100                 err = insert_page(vma, addr, page, pgprot);
2101         } else {
2102                 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
2103         }
2104
2105         if (err == -ENOMEM)
2106                 return VM_FAULT_OOM;
2107         if (err < 0 && err != -EBUSY)
2108                 return VM_FAULT_SIGBUS;
2109
2110         return VM_FAULT_NOPAGE;
2111 }
2112
2113 /**
2114  * vmf_insert_mixed_prot - insert single pfn into user vma with specified pgprot
2115  * @vma: user vma to map to
2116  * @addr: target user address of this page
2117  * @pfn: source kernel pfn
2118  * @pgprot: pgprot flags for the inserted page
2119  *
2120  * This is exactly like vmf_insert_mixed(), except that it allows drivers
2121  * to override pgprot on a per-page basis.
2122  *
2123  * Typically this function should be used by drivers to set caching- and
2124  * encryption bits different than those of @vma->vm_page_prot, because
2125  * the caching- or encryption mode may not be known at mmap() time.
2126  * This is ok as long as @vma->vm_page_prot is not used by the core vm
2127  * to set caching and encryption bits for those vmas (except for COW pages).
2128  * This is ensured by core vm only modifying these page table entries using
2129  * functions that don't touch caching- or encryption bits, using pte_modify()
2130  * if needed. (See for example mprotect()).
2131  * Also when new page-table entries are created, this is only done using the
2132  * fault() callback, and never using the value of vma->vm_page_prot,
2133  * except for page-table entries that point to anonymous pages as the result
2134  * of COW.
2135  *
2136  * Context: Process context.  May allocate using %GFP_KERNEL.
2137  * Return: vm_fault_t value.
2138  */
2139 vm_fault_t vmf_insert_mixed_prot(struct vm_area_struct *vma, unsigned long addr,
2140                                  pfn_t pfn, pgprot_t pgprot)
2141 {
2142         return __vm_insert_mixed(vma, addr, pfn, pgprot, false);
2143 }
2144 EXPORT_SYMBOL(vmf_insert_mixed_prot);
2145
2146 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
2147                 pfn_t pfn)
2148 {
2149         return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, false);
2150 }
2151 EXPORT_SYMBOL(vmf_insert_mixed);
2152
2153 /*
2154  *  If the insertion of PTE failed because someone else already added a
2155  *  different entry in the mean time, we treat that as success as we assume
2156  *  the same entry was actually inserted.
2157  */
2158 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
2159                 unsigned long addr, pfn_t pfn)
2160 {
2161         return __vm_insert_mixed(vma, addr, pfn, vma->vm_page_prot, true);
2162 }
2163 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
2164
2165 /*
2166  * maps a range of physical memory into the requested pages. the old
2167  * mappings are removed. any references to nonexistent pages results
2168  * in null mappings (currently treated as "copy-on-access")
2169  */
2170 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
2171                         unsigned long addr, unsigned long end,
2172                         unsigned long pfn, pgprot_t prot)
2173 {
2174         pte_t *pte, *mapped_pte;
2175         spinlock_t *ptl;
2176         int err = 0;
2177
2178         mapped_pte = pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
2179         if (!pte)
2180                 return -ENOMEM;
2181         arch_enter_lazy_mmu_mode();
2182         do {
2183                 BUG_ON(!pte_none(*pte));
2184                 if (!pfn_modify_allowed(pfn, prot)) {
2185                         err = -EACCES;
2186                         break;
2187                 }
2188                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
2189                 pfn++;
2190         } while (pte++, addr += PAGE_SIZE, addr != end);
2191         arch_leave_lazy_mmu_mode();
2192         pte_unmap_unlock(mapped_pte, ptl);
2193         return err;
2194 }
2195
2196 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
2197                         unsigned long addr, unsigned long end,
2198                         unsigned long pfn, pgprot_t prot)
2199 {
2200         pmd_t *pmd;
2201         unsigned long next;
2202         int err;
2203
2204         pfn -= addr >> PAGE_SHIFT;
2205         pmd = pmd_alloc(mm, pud, addr);
2206         if (!pmd)
2207                 return -ENOMEM;
2208         VM_BUG_ON(pmd_trans_huge(*pmd));
2209         do {
2210                 next = pmd_addr_end(addr, end);
2211                 err = remap_pte_range(mm, pmd, addr, next,
2212                                 pfn + (addr >> PAGE_SHIFT), prot);
2213                 if (err)
2214                         return err;
2215         } while (pmd++, addr = next, addr != end);
2216         return 0;
2217 }
2218
2219 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
2220                         unsigned long addr, unsigned long end,
2221                         unsigned long pfn, pgprot_t prot)
2222 {
2223         pud_t *pud;
2224         unsigned long next;
2225         int err;
2226
2227         pfn -= addr >> PAGE_SHIFT;
2228         pud = pud_alloc(mm, p4d, addr);
2229         if (!pud)
2230                 return -ENOMEM;
2231         do {
2232                 next = pud_addr_end(addr, end);
2233                 err = remap_pmd_range(mm, pud, addr, next,
2234                                 pfn + (addr >> PAGE_SHIFT), prot);
2235                 if (err)
2236                         return err;
2237         } while (pud++, addr = next, addr != end);
2238         return 0;
2239 }
2240
2241 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2242                         unsigned long addr, unsigned long end,
2243                         unsigned long pfn, pgprot_t prot)
2244 {
2245         p4d_t *p4d;
2246         unsigned long next;
2247         int err;
2248
2249         pfn -= addr >> PAGE_SHIFT;
2250         p4d = p4d_alloc(mm, pgd, addr);
2251         if (!p4d)
2252                 return -ENOMEM;
2253         do {
2254                 next = p4d_addr_end(addr, end);
2255                 err = remap_pud_range(mm, p4d, addr, next,
2256                                 pfn + (addr >> PAGE_SHIFT), prot);
2257                 if (err)
2258                         return err;
2259         } while (p4d++, addr = next, addr != end);
2260         return 0;
2261 }
2262
2263 /**
2264  * remap_pfn_range - remap kernel memory to userspace
2265  * @vma: user vma to map to
2266  * @addr: target page aligned user address to start at
2267  * @pfn: page frame number of kernel physical memory address
2268  * @size: size of mapping area
2269  * @prot: page protection flags for this mapping
2270  *
2271  * Note: this is only safe if the mm semaphore is held when called.
2272  *
2273  * Return: %0 on success, negative error code otherwise.
2274  */
2275 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
2276                     unsigned long pfn, unsigned long size, pgprot_t prot)
2277 {
2278         pgd_t *pgd;
2279         unsigned long next;
2280         unsigned long end = addr + PAGE_ALIGN(size);
2281         struct mm_struct *mm = vma->vm_mm;
2282         unsigned long remap_pfn = pfn;
2283         int err;
2284
2285         if (WARN_ON_ONCE(!PAGE_ALIGNED(addr)))
2286                 return -EINVAL;
2287
2288         /*
2289          * Physically remapped pages are special. Tell the
2290          * rest of the world about it:
2291          *   VM_IO tells people not to look at these pages
2292          *      (accesses can have side effects).
2293          *   VM_PFNMAP tells the core MM that the base pages are just
2294          *      raw PFN mappings, and do not have a "struct page" associated
2295          *      with them.
2296          *   VM_DONTEXPAND
2297          *      Disable vma merging and expanding with mremap().
2298          *   VM_DONTDUMP
2299          *      Omit vma from core dump, even when VM_IO turned off.
2300          *
2301          * There's a horrible special case to handle copy-on-write
2302          * behaviour that some programs depend on. We mark the "original"
2303          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
2304          * See vm_normal_page() for details.
2305          */
2306         if (is_cow_mapping(vma->vm_flags)) {
2307                 if (addr != vma->vm_start || end != vma->vm_end)
2308                         return -EINVAL;
2309                 vma->vm_pgoff = pfn;
2310         }
2311
2312         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
2313         if (err)
2314                 return -EINVAL;
2315
2316         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
2317
2318         BUG_ON(addr >= end);
2319         pfn -= addr >> PAGE_SHIFT;
2320         pgd = pgd_offset(mm, addr);
2321         flush_cache_range(vma, addr, end);
2322         do {
2323                 next = pgd_addr_end(addr, end);
2324                 err = remap_p4d_range(mm, pgd, addr, next,
2325                                 pfn + (addr >> PAGE_SHIFT), prot);
2326                 if (err)
2327                         break;
2328         } while (pgd++, addr = next, addr != end);
2329
2330         if (err)
2331                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
2332
2333         return err;
2334 }
2335 EXPORT_SYMBOL(remap_pfn_range);
2336
2337 /**
2338  * vm_iomap_memory - remap memory to userspace
2339  * @vma: user vma to map to
2340  * @start: start of the physical memory to be mapped
2341  * @len: size of area
2342  *
2343  * This is a simplified io_remap_pfn_range() for common driver use. The
2344  * driver just needs to give us the physical memory range to be mapped,
2345  * we'll figure out the rest from the vma information.
2346  *
2347  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
2348  * whatever write-combining details or similar.
2349  *
2350  * Return: %0 on success, negative error code otherwise.
2351  */
2352 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
2353 {
2354         unsigned long vm_len, pfn, pages;
2355
2356         /* Check that the physical memory area passed in looks valid */
2357         if (start + len < start)
2358                 return -EINVAL;
2359         /*
2360          * You *really* shouldn't map things that aren't page-aligned,
2361          * but we've historically allowed it because IO memory might
2362          * just have smaller alignment.
2363          */
2364         len += start & ~PAGE_MASK;
2365         pfn = start >> PAGE_SHIFT;
2366         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
2367         if (pfn + pages < pfn)
2368                 return -EINVAL;
2369
2370         /* We start the mapping 'vm_pgoff' pages into the area */
2371         if (vma->vm_pgoff > pages)
2372                 return -EINVAL;
2373         pfn += vma->vm_pgoff;
2374         pages -= vma->vm_pgoff;
2375
2376         /* Can we fit all of the mapping? */
2377         vm_len = vma->vm_end - vma->vm_start;
2378         if (vm_len >> PAGE_SHIFT > pages)
2379                 return -EINVAL;
2380
2381         /* Ok, let it rip */
2382         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
2383 }
2384 EXPORT_SYMBOL(vm_iomap_memory);
2385
2386 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
2387                                      unsigned long addr, unsigned long end,
2388                                      pte_fn_t fn, void *data, bool create,
2389                                      pgtbl_mod_mask *mask)
2390 {
2391         pte_t *pte, *mapped_pte;
2392         int err = 0;
2393         spinlock_t *ptl;
2394
2395         if (create) {
2396                 mapped_pte = pte = (mm == &init_mm) ?
2397                         pte_alloc_kernel_track(pmd, addr, mask) :
2398                         pte_alloc_map_lock(mm, pmd, addr, &ptl);
2399                 if (!pte)
2400                         return -ENOMEM;
2401         } else {
2402                 mapped_pte = pte = (mm == &init_mm) ?
2403                         pte_offset_kernel(pmd, addr) :
2404                         pte_offset_map_lock(mm, pmd, addr, &ptl);
2405         }
2406
2407         BUG_ON(pmd_huge(*pmd));
2408
2409         arch_enter_lazy_mmu_mode();
2410
2411         if (fn) {
2412                 do {
2413                         if (create || !pte_none(*pte)) {
2414                                 err = fn(pte++, addr, data);
2415                                 if (err)
2416                                         break;
2417                         }
2418                 } while (addr += PAGE_SIZE, addr != end);
2419         }
2420         *mask |= PGTBL_PTE_MODIFIED;
2421
2422         arch_leave_lazy_mmu_mode();
2423
2424         if (mm != &init_mm)
2425                 pte_unmap_unlock(mapped_pte, ptl);
2426         return err;
2427 }
2428
2429 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
2430                                      unsigned long addr, unsigned long end,
2431                                      pte_fn_t fn, void *data, bool create,
2432                                      pgtbl_mod_mask *mask)
2433 {
2434         pmd_t *pmd;
2435         unsigned long next;
2436         int err = 0;
2437
2438         BUG_ON(pud_huge(*pud));
2439
2440         if (create) {
2441                 pmd = pmd_alloc_track(mm, pud, addr, mask);
2442                 if (!pmd)
2443                         return -ENOMEM;
2444         } else {
2445                 pmd = pmd_offset(pud, addr);
2446         }
2447         do {
2448                 next = pmd_addr_end(addr, end);
2449                 if (create || !pmd_none_or_clear_bad(pmd)) {
2450                         err = apply_to_pte_range(mm, pmd, addr, next, fn, data,
2451                                                  create, mask);
2452                         if (err)
2453                                 break;
2454                 }
2455         } while (pmd++, addr = next, addr != end);
2456         return err;
2457 }
2458
2459 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
2460                                      unsigned long addr, unsigned long end,
2461                                      pte_fn_t fn, void *data, bool create,
2462                                      pgtbl_mod_mask *mask)
2463 {
2464         pud_t *pud;
2465         unsigned long next;
2466         int err = 0;
2467
2468         if (create) {
2469                 pud = pud_alloc_track(mm, p4d, addr, mask);
2470                 if (!pud)
2471                         return -ENOMEM;
2472         } else {
2473                 pud = pud_offset(p4d, addr);
2474         }
2475         do {
2476                 next = pud_addr_end(addr, end);
2477                 if (create || !pud_none_or_clear_bad(pud)) {
2478                         err = apply_to_pmd_range(mm, pud, addr, next, fn, data,
2479                                                  create, mask);
2480                         if (err)
2481                                 break;
2482                 }
2483         } while (pud++, addr = next, addr != end);
2484         return err;
2485 }
2486
2487 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2488                                      unsigned long addr, unsigned long end,
2489                                      pte_fn_t fn, void *data, bool create,
2490                                      pgtbl_mod_mask *mask)
2491 {
2492         p4d_t *p4d;
2493         unsigned long next;
2494         int err = 0;
2495
2496         if (create) {
2497                 p4d = p4d_alloc_track(mm, pgd, addr, mask);
2498                 if (!p4d)
2499                         return -ENOMEM;
2500         } else {
2501                 p4d = p4d_offset(pgd, addr);
2502         }
2503         do {
2504                 next = p4d_addr_end(addr, end);
2505                 if (create || !p4d_none_or_clear_bad(p4d)) {
2506                         err = apply_to_pud_range(mm, p4d, addr, next, fn, data,
2507                                                  create, mask);
2508                         if (err)
2509                                 break;
2510                 }
2511         } while (p4d++, addr = next, addr != end);
2512         return err;
2513 }
2514
2515 static int __apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2516                                  unsigned long size, pte_fn_t fn,
2517                                  void *data, bool create)
2518 {
2519         pgd_t *pgd;
2520         unsigned long start = addr, next;
2521         unsigned long end = addr + size;
2522         pgtbl_mod_mask mask = 0;
2523         int err = 0;
2524
2525         if (WARN_ON(addr >= end))
2526                 return -EINVAL;
2527
2528         pgd = pgd_offset(mm, addr);
2529         do {
2530                 next = pgd_addr_end(addr, end);
2531                 if (!create && pgd_none_or_clear_bad(pgd))
2532                         continue;
2533                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data, create, &mask);
2534                 if (err)
2535                         break;
2536         } while (pgd++, addr = next, addr != end);
2537
2538         if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
2539                 arch_sync_kernel_mappings(start, start + size);
2540
2541         return err;
2542 }
2543
2544 /*
2545  * Scan a region of virtual memory, filling in page tables as necessary
2546  * and calling a provided function on each leaf page table.
2547  */
2548 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2549                         unsigned long size, pte_fn_t fn, void *data)
2550 {
2551         return __apply_to_page_range(mm, addr, size, fn, data, true);
2552 }
2553 EXPORT_SYMBOL_GPL(apply_to_page_range);
2554
2555 /*
2556  * Scan a region of virtual memory, calling a provided function on
2557  * each leaf page table where it exists.
2558  *
2559  * Unlike apply_to_page_range, this does _not_ fill in page tables
2560  * where they are absent.
2561  */
2562 int apply_to_existing_page_range(struct mm_struct *mm, unsigned long addr,
2563                                  unsigned long size, pte_fn_t fn, void *data)
2564 {
2565         return __apply_to_page_range(mm, addr, size, fn, data, false);
2566 }
2567 EXPORT_SYMBOL_GPL(apply_to_existing_page_range);
2568
2569 /*
2570  * handle_pte_fault chooses page fault handler according to an entry which was
2571  * read non-atomically.  Before making any commitment, on those architectures
2572  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2573  * parts, do_swap_page must check under lock before unmapping the pte and
2574  * proceeding (but do_wp_page is only called after already making such a check;
2575  * and do_anonymous_page can safely check later on).
2576  */
2577 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2578                                 pte_t *page_table, pte_t orig_pte)
2579 {
2580         int same = 1;
2581 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPTION)
2582         if (sizeof(pte_t) > sizeof(unsigned long)) {
2583                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2584                 spin_lock(ptl);
2585                 same = pte_same(*page_table, orig_pte);
2586                 spin_unlock(ptl);
2587         }
2588 #endif
2589         pte_unmap(page_table);
2590         return same;
2591 }
2592
2593 static inline bool cow_user_page(struct page *dst, struct page *src,
2594                                  struct vm_fault *vmf)
2595 {
2596         bool ret;
2597         void *kaddr;
2598         void __user *uaddr;
2599         bool locked = false;
2600         struct vm_area_struct *vma = vmf->vma;
2601         struct mm_struct *mm = vma->vm_mm;
2602         unsigned long addr = vmf->address;
2603
2604         if (likely(src)) {
2605                 copy_user_highpage(dst, src, addr, vma);
2606                 return true;
2607         }
2608
2609         /*
2610          * If the source page was a PFN mapping, we don't have
2611          * a "struct page" for it. We do a best-effort copy by
2612          * just copying from the original user address. If that
2613          * fails, we just zero-fill it. Live with it.
2614          */
2615         kaddr = kmap_atomic(dst);
2616         uaddr = (void __user *)(addr & PAGE_MASK);
2617
2618         /*
2619          * On architectures with software "accessed" bits, we would
2620          * take a double page fault, so mark it accessed here.
2621          */
2622         if (arch_faults_on_old_pte() && !pte_young(vmf->orig_pte)) {
2623                 pte_t entry;
2624
2625                 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2626                 locked = true;
2627                 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2628                         /*
2629                          * Other thread has already handled the fault
2630                          * and update local tlb only
2631                          */
2632                         update_mmu_tlb(vma, addr, vmf->pte);
2633                         ret = false;
2634                         goto pte_unlock;
2635                 }
2636
2637                 entry = pte_mkyoung(vmf->orig_pte);
2638                 if (ptep_set_access_flags(vma, addr, vmf->pte, entry, 0))
2639                         update_mmu_cache(vma, addr, vmf->pte);
2640         }
2641
2642         /*
2643          * This really shouldn't fail, because the page is there
2644          * in the page tables. But it might just be unreadable,
2645          * in which case we just give up and fill the result with
2646          * zeroes.
2647          */
2648         if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2649                 if (locked)
2650                         goto warn;
2651
2652                 /* Re-validate under PTL if the page is still mapped */
2653                 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, addr, &vmf->ptl);
2654                 locked = true;
2655                 if (!likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2656                         /* The PTE changed under us, update local tlb */
2657                         update_mmu_tlb(vma, addr, vmf->pte);
2658                         ret = false;
2659                         goto pte_unlock;
2660                 }
2661
2662                 /*
2663                  * The same page can be mapped back since last copy attempt.
2664                  * Try to copy again under PTL.
2665                  */
2666                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) {
2667                         /*
2668                          * Give a warn in case there can be some obscure
2669                          * use-case
2670                          */
2671 warn:
2672                         WARN_ON_ONCE(1);
2673                         clear_page(kaddr);
2674                 }
2675         }
2676
2677         ret = true;
2678
2679 pte_unlock:
2680         if (locked)
2681                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2682         kunmap_atomic(kaddr);
2683         flush_dcache_page(dst);
2684
2685         return ret;
2686 }
2687
2688 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2689 {
2690         struct file *vm_file = vma->vm_file;
2691
2692         if (vm_file)
2693                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2694
2695         /*
2696          * Special mappings (e.g. VDSO) do not have any file so fake
2697          * a default GFP_KERNEL for them.
2698          */
2699         return GFP_KERNEL;
2700 }
2701
2702 /*
2703  * Notify the address space that the page is about to become writable so that
2704  * it can prohibit this or wait for the page to get into an appropriate state.
2705  *
2706  * We do this without the lock held, so that it can sleep if it needs to.
2707  */
2708 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2709 {
2710         vm_fault_t ret;
2711         struct page *page = vmf->page;
2712         unsigned int old_flags = vmf->flags;
2713
2714         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2715
2716         if (vmf->vma->vm_file &&
2717             IS_SWAPFILE(vmf->vma->vm_file->f_mapping->host))
2718                 return VM_FAULT_SIGBUS;
2719
2720         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2721         /* Restore original flags so that caller is not surprised */
2722         vmf->flags = old_flags;
2723         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2724                 return ret;
2725         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2726                 lock_page(page);
2727                 if (!page->mapping) {
2728                         unlock_page(page);
2729                         return 0; /* retry */
2730                 }
2731                 ret |= VM_FAULT_LOCKED;
2732         } else
2733                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2734         return ret;
2735 }
2736
2737 /*
2738  * Handle dirtying of a page in shared file mapping on a write fault.
2739  *
2740  * The function expects the page to be locked and unlocks it.
2741  */
2742 static vm_fault_t fault_dirty_shared_page(struct vm_fault *vmf)
2743 {
2744         struct vm_area_struct *vma = vmf->vma;
2745         struct address_space *mapping;
2746         struct page *page = vmf->page;
2747         bool dirtied;
2748         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2749
2750         dirtied = set_page_dirty(page);
2751         VM_BUG_ON_PAGE(PageAnon(page), page);
2752         /*
2753          * Take a local copy of the address_space - page.mapping may be zeroed
2754          * by truncate after unlock_page().   The address_space itself remains
2755          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2756          * release semantics to prevent the compiler from undoing this copying.
2757          */
2758         mapping = page_rmapping(page);
2759         unlock_page(page);
2760
2761         if (!page_mkwrite)
2762                 file_update_time(vma->vm_file);
2763
2764         /*
2765          * Throttle page dirtying rate down to writeback speed.
2766          *
2767          * mapping may be NULL here because some device drivers do not
2768          * set page.mapping but still dirty their pages
2769          *
2770          * Drop the mmap_lock before waiting on IO, if we can. The file
2771          * is pinning the mapping, as per above.
2772          */
2773         if ((dirtied || page_mkwrite) && mapping) {
2774                 struct file *fpin;
2775
2776                 fpin = maybe_unlock_mmap_for_io(vmf, NULL);
2777                 balance_dirty_pages_ratelimited(mapping);
2778                 if (fpin) {
2779                         fput(fpin);
2780                         return VM_FAULT_RETRY;
2781                 }
2782         }
2783
2784         return 0;
2785 }
2786
2787 /*
2788  * Handle write page faults for pages that can be reused in the current vma
2789  *
2790  * This can happen either due to the mapping being with the VM_SHARED flag,
2791  * or due to us being the last reference standing to the page. In either
2792  * case, all we need to do here is to mark the page as writable and update
2793  * any related book-keeping.
2794  */
2795 static inline void wp_page_reuse(struct vm_fault *vmf)
2796         __releases(vmf->ptl)
2797 {
2798         struct vm_area_struct *vma = vmf->vma;
2799         struct page *page = vmf->page;
2800         pte_t entry;
2801         /*
2802          * Clear the pages cpupid information as the existing
2803          * information potentially belongs to a now completely
2804          * unrelated process.
2805          */
2806         if (page)
2807                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2808
2809         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2810         entry = pte_mkyoung(vmf->orig_pte);
2811         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2812         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2813                 update_mmu_cache(vma, vmf->address, vmf->pte);
2814         pte_unmap_unlock(vmf->pte, vmf->ptl);
2815         count_vm_event(PGREUSE);
2816 }
2817
2818 /*
2819  * Handle the case of a page which we actually need to copy to a new page.
2820  *
2821  * Called with mmap_lock locked and the old page referenced, but
2822  * without the ptl held.
2823  *
2824  * High level logic flow:
2825  *
2826  * - Allocate a page, copy the content of the old page to the new one.
2827  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2828  * - Take the PTL. If the pte changed, bail out and release the allocated page
2829  * - If the pte is still the way we remember it, update the page table and all
2830  *   relevant references. This includes dropping the reference the page-table
2831  *   held to the old page, as well as updating the rmap.
2832  * - In any case, unlock the PTL and drop the reference we took to the old page.
2833  */
2834 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2835 {
2836         struct vm_area_struct *vma = vmf->vma;
2837         struct mm_struct *mm = vma->vm_mm;
2838         struct page *old_page = vmf->page;
2839         struct page *new_page = NULL;
2840         pte_t entry;
2841         int page_copied = 0;
2842         struct mmu_notifier_range range;
2843
2844         if (unlikely(anon_vma_prepare(vma)))
2845                 goto oom;
2846
2847         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2848                 new_page = alloc_zeroed_user_highpage_movable(vma,
2849                                                               vmf->address);
2850                 if (!new_page)
2851                         goto oom;
2852         } else {
2853                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2854                                 vmf->address);
2855                 if (!new_page)
2856                         goto oom;
2857
2858                 if (!cow_user_page(new_page, old_page, vmf)) {
2859                         /*
2860                          * COW failed, if the fault was solved by other,
2861                          * it's fine. If not, userspace would re-fault on
2862                          * the same address and we will handle the fault
2863                          * from the second attempt.
2864                          */
2865                         put_page(new_page);
2866                         if (old_page)
2867                                 put_page(old_page);
2868                         return 0;
2869                 }
2870         }
2871
2872         if (mem_cgroup_charge(new_page, mm, GFP_KERNEL))
2873                 goto oom_free_new;
2874         cgroup_throttle_swaprate(new_page, GFP_KERNEL);
2875
2876         __SetPageUptodate(new_page);
2877
2878         mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, vma, mm,
2879                                 vmf->address & PAGE_MASK,
2880                                 (vmf->address & PAGE_MASK) + PAGE_SIZE);
2881         mmu_notifier_invalidate_range_start(&range);
2882
2883         /*
2884          * Re-check the pte - we dropped the lock
2885          */
2886         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2887         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2888                 if (old_page) {
2889                         if (!PageAnon(old_page)) {
2890                                 dec_mm_counter_fast(mm,
2891                                                 mm_counter_file(old_page));
2892                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2893                         }
2894                 } else {
2895                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2896                 }
2897                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2898                 entry = mk_pte(new_page, vma->vm_page_prot);
2899                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2900
2901                 /*
2902                  * Clear the pte entry and flush it first, before updating the
2903                  * pte with the new entry, to keep TLBs on different CPUs in
2904                  * sync. This code used to set the new PTE then flush TLBs, but
2905                  * that left a window where the new PTE could be loaded into
2906                  * some TLBs while the old PTE remains in others.
2907                  */
2908                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2909                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2910                 lru_cache_add_inactive_or_unevictable(new_page, vma);
2911                 /*
2912                  * We call the notify macro here because, when using secondary
2913                  * mmu page tables (such as kvm shadow page tables), we want the
2914                  * new page to be mapped directly into the secondary page table.
2915                  */
2916                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2917                 update_mmu_cache(vma, vmf->address, vmf->pte);
2918                 if (old_page) {
2919                         /*
2920                          * Only after switching the pte to the new page may
2921                          * we remove the mapcount here. Otherwise another
2922                          * process may come and find the rmap count decremented
2923                          * before the pte is switched to the new page, and
2924                          * "reuse" the old page writing into it while our pte
2925                          * here still points into it and can be read by other
2926                          * threads.
2927                          *
2928                          * The critical issue is to order this
2929                          * page_remove_rmap with the ptp_clear_flush above.
2930                          * Those stores are ordered by (if nothing else,)
2931                          * the barrier present in the atomic_add_negative
2932                          * in page_remove_rmap.
2933                          *
2934                          * Then the TLB flush in ptep_clear_flush ensures that
2935                          * no process can access the old page before the
2936                          * decremented mapcount is visible. And the old page
2937                          * cannot be reused until after the decremented
2938                          * mapcount is visible. So transitively, TLBs to
2939                          * old page will be flushed before it can be reused.
2940                          */
2941                         page_remove_rmap(old_page, false);
2942                 }
2943
2944                 /* Free the old page.. */
2945                 new_page = old_page;
2946                 page_copied = 1;
2947         } else {
2948                 update_mmu_tlb(vma, vmf->address, vmf->pte);
2949         }
2950
2951         if (new_page)
2952                 put_page(new_page);
2953
2954         pte_unmap_unlock(vmf->pte, vmf->ptl);
2955         /*
2956          * No need to double call mmu_notifier->invalidate_range() callback as
2957          * the above ptep_clear_flush_notify() did already call it.
2958          */
2959         mmu_notifier_invalidate_range_only_end(&range);
2960         if (old_page) {
2961                 /*
2962                  * Don't let another task, with possibly unlocked vma,
2963                  * keep the mlocked page.
2964                  */
2965                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2966                         lock_page(old_page);    /* LRU manipulation */
2967                         if (PageMlocked(old_page))
2968                                 munlock_vma_page(old_page);
2969                         unlock_page(old_page);
2970                 }
2971                 put_page(old_page);
2972         }
2973         return page_copied ? VM_FAULT_WRITE : 0;
2974 oom_free_new:
2975         put_page(new_page);
2976 oom:
2977         if (old_page)
2978                 put_page(old_page);
2979         return VM_FAULT_OOM;
2980 }
2981
2982 /**
2983  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2984  *                        writeable once the page is prepared
2985  *
2986  * @vmf: structure describing the fault
2987  *
2988  * This function handles all that is needed to finish a write page fault in a
2989  * shared mapping due to PTE being read-only once the mapped page is prepared.
2990  * It handles locking of PTE and modifying it.
2991  *
2992  * The function expects the page to be locked or other protection against
2993  * concurrent faults / writeback (such as DAX radix tree locks).
2994  *
2995  * Return: %VM_FAULT_WRITE on success, %0 when PTE got changed before
2996  * we acquired PTE lock.
2997  */
2998 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2999 {
3000         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
3001         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
3002                                        &vmf->ptl);
3003         /*
3004          * We might have raced with another page fault while we released the
3005          * pte_offset_map_lock.
3006          */
3007         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
3008                 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
3009                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3010                 return VM_FAULT_NOPAGE;
3011         }
3012         wp_page_reuse(vmf);
3013         return 0;
3014 }
3015
3016 /*
3017  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
3018  * mapping
3019  */
3020 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
3021 {
3022         struct vm_area_struct *vma = vmf->vma;
3023
3024         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
3025                 vm_fault_t ret;
3026
3027                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3028                 vmf->flags |= FAULT_FLAG_MKWRITE;
3029                 ret = vma->vm_ops->pfn_mkwrite(vmf);
3030                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
3031                         return ret;
3032                 return finish_mkwrite_fault(vmf);
3033         }
3034         wp_page_reuse(vmf);
3035         return VM_FAULT_WRITE;
3036 }
3037
3038 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
3039         __releases(vmf->ptl)
3040 {
3041         struct vm_area_struct *vma = vmf->vma;
3042         vm_fault_t ret = VM_FAULT_WRITE;
3043
3044         get_page(vmf->page);
3045
3046         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
3047                 vm_fault_t tmp;
3048
3049                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3050                 tmp = do_page_mkwrite(vmf);
3051                 if (unlikely(!tmp || (tmp &
3052                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3053                         put_page(vmf->page);
3054                         return tmp;
3055                 }
3056                 tmp = finish_mkwrite_fault(vmf);
3057                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
3058                         unlock_page(vmf->page);
3059                         put_page(vmf->page);
3060                         return tmp;
3061                 }
3062         } else {
3063                 wp_page_reuse(vmf);
3064                 lock_page(vmf->page);
3065         }
3066         ret |= fault_dirty_shared_page(vmf);
3067         put_page(vmf->page);
3068
3069         return ret;
3070 }
3071
3072 /*
3073  * This routine handles present pages, when users try to write
3074  * to a shared page. It is done by copying the page to a new address
3075  * and decrementing the shared-page counter for the old page.
3076  *
3077  * Note that this routine assumes that the protection checks have been
3078  * done by the caller (the low-level page fault routine in most cases).
3079  * Thus we can safely just mark it writable once we've done any necessary
3080  * COW.
3081  *
3082  * We also mark the page dirty at this point even though the page will
3083  * change only once the write actually happens. This avoids a few races,
3084  * and potentially makes it more efficient.
3085  *
3086  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3087  * but allow concurrent faults), with pte both mapped and locked.
3088  * We return with mmap_lock still held, but pte unmapped and unlocked.
3089  */
3090 static vm_fault_t do_wp_page(struct vm_fault *vmf)
3091         __releases(vmf->ptl)
3092 {
3093         struct vm_area_struct *vma = vmf->vma;
3094
3095         if (userfaultfd_pte_wp(vma, *vmf->pte)) {
3096                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3097                 return handle_userfault(vmf, VM_UFFD_WP);
3098         }
3099
3100         /*
3101          * Userfaultfd write-protect can defer flushes. Ensure the TLB
3102          * is flushed in this case before copying.
3103          */
3104         if (unlikely(userfaultfd_wp(vmf->vma) &&
3105                      mm_tlb_flush_pending(vmf->vma->vm_mm)))
3106                 flush_tlb_page(vmf->vma, vmf->address);
3107
3108         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
3109         if (!vmf->page) {
3110                 /*
3111                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
3112                  * VM_PFNMAP VMA.
3113                  *
3114                  * We should not cow pages in a shared writeable mapping.
3115                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
3116                  */
3117                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3118                                      (VM_WRITE|VM_SHARED))
3119                         return wp_pfn_shared(vmf);
3120
3121                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3122                 return wp_page_copy(vmf);
3123         }
3124
3125         /*
3126          * Take out anonymous pages first, anonymous shared vmas are
3127          * not dirty accountable.
3128          */
3129         if (PageAnon(vmf->page)) {
3130                 struct page *page = vmf->page;
3131
3132                 /* PageKsm() doesn't necessarily raise the page refcount */
3133                 if (PageKsm(page) || page_count(page) != 1)
3134                         goto copy;
3135                 if (!trylock_page(page))
3136                         goto copy;
3137                 if (PageKsm(page) || page_mapcount(page) != 1 || page_count(page) != 1) {
3138                         unlock_page(page);
3139                         goto copy;
3140                 }
3141                 /*
3142                  * Ok, we've got the only map reference, and the only
3143                  * page count reference, and the page is locked,
3144                  * it's dark out, and we're wearing sunglasses. Hit it.
3145                  */
3146                 unlock_page(page);
3147                 wp_page_reuse(vmf);
3148                 return VM_FAULT_WRITE;
3149         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
3150                                         (VM_WRITE|VM_SHARED))) {
3151                 return wp_page_shared(vmf);
3152         }
3153 copy:
3154         /*
3155          * Ok, we need to copy. Oh, well..
3156          */
3157         get_page(vmf->page);
3158
3159         pte_unmap_unlock(vmf->pte, vmf->ptl);
3160         return wp_page_copy(vmf);
3161 }
3162
3163 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
3164                 unsigned long start_addr, unsigned long end_addr,
3165                 struct zap_details *details)
3166 {
3167         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
3168 }
3169
3170 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
3171                                             struct zap_details *details)
3172 {
3173         struct vm_area_struct *vma;
3174         pgoff_t vba, vea, zba, zea;
3175
3176         vma_interval_tree_foreach(vma, root,
3177                         details->first_index, details->last_index) {
3178
3179                 vba = vma->vm_pgoff;
3180                 vea = vba + vma_pages(vma) - 1;
3181                 zba = details->first_index;
3182                 if (zba < vba)
3183                         zba = vba;
3184                 zea = details->last_index;
3185                 if (zea > vea)
3186                         zea = vea;
3187
3188                 unmap_mapping_range_vma(vma,
3189                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
3190                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
3191                                 details);
3192         }
3193 }
3194
3195 /**
3196  * unmap_mapping_pages() - Unmap pages from processes.
3197  * @mapping: The address space containing pages to be unmapped.
3198  * @start: Index of first page to be unmapped.
3199  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
3200  * @even_cows: Whether to unmap even private COWed pages.
3201  *
3202  * Unmap the pages in this address space from any userspace process which
3203  * has them mmaped.  Generally, you want to remove COWed pages as well when
3204  * a file is being truncated, but not when invalidating pages from the page
3205  * cache.
3206  */
3207 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
3208                 pgoff_t nr, bool even_cows)
3209 {
3210         struct zap_details details = { };
3211
3212         details.check_mapping = even_cows ? NULL : mapping;
3213         details.first_index = start;
3214         details.last_index = start + nr - 1;
3215         if (details.last_index < details.first_index)
3216                 details.last_index = ULONG_MAX;
3217
3218         i_mmap_lock_write(mapping);
3219         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
3220                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
3221         i_mmap_unlock_write(mapping);
3222 }
3223
3224 /**
3225  * unmap_mapping_range - unmap the portion of all mmaps in the specified
3226  * address_space corresponding to the specified byte range in the underlying
3227  * file.
3228  *
3229  * @mapping: the address space containing mmaps to be unmapped.
3230  * @holebegin: byte in first page to unmap, relative to the start of
3231  * the underlying file.  This will be rounded down to a PAGE_SIZE
3232  * boundary.  Note that this is different from truncate_pagecache(), which
3233  * must keep the partial page.  In contrast, we must get rid of
3234  * partial pages.
3235  * @holelen: size of prospective hole in bytes.  This will be rounded
3236  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
3237  * end of the file.
3238  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
3239  * but 0 when invalidating pagecache, don't throw away private data.
3240  */
3241 void unmap_mapping_range(struct address_space *mapping,
3242                 loff_t const holebegin, loff_t const holelen, int even_cows)
3243 {
3244         pgoff_t hba = holebegin >> PAGE_SHIFT;
3245         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3246
3247         /* Check for overflow. */
3248         if (sizeof(holelen) > sizeof(hlen)) {
3249                 long long holeend =
3250                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
3251                 if (holeend & ~(long long)ULONG_MAX)
3252                         hlen = ULONG_MAX - hba + 1;
3253         }
3254
3255         unmap_mapping_pages(mapping, hba, hlen, even_cows);
3256 }
3257 EXPORT_SYMBOL(unmap_mapping_range);
3258
3259 /*
3260  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3261  * but allow concurrent faults), and pte mapped but not yet locked.
3262  * We return with pte unmapped and unlocked.
3263  *
3264  * We return with the mmap_lock locked or unlocked in the same cases
3265  * as does filemap_fault().
3266  */
3267 vm_fault_t do_swap_page(struct vm_fault *vmf)
3268 {
3269         struct vm_area_struct *vma = vmf->vma;
3270         struct page *page = NULL, *swapcache;
3271         swp_entry_t entry;
3272         pte_t pte;
3273         int locked;
3274         int exclusive = 0;
3275         vm_fault_t ret = 0;
3276         void *shadow = NULL;
3277
3278         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
3279                 goto out;
3280
3281         entry = pte_to_swp_entry(vmf->orig_pte);
3282         if (unlikely(non_swap_entry(entry))) {
3283                 if (is_migration_entry(entry)) {
3284                         migration_entry_wait(vma->vm_mm, vmf->pmd,
3285                                              vmf->address);
3286                 } else if (is_device_private_entry(entry)) {
3287                         vmf->page = device_private_entry_to_page(entry);
3288                         ret = vmf->page->pgmap->ops->migrate_to_ram(vmf);
3289                 } else if (is_hwpoison_entry(entry)) {
3290                         ret = VM_FAULT_HWPOISON;
3291                 } else {
3292                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
3293                         ret = VM_FAULT_SIGBUS;
3294                 }
3295                 goto out;
3296         }
3297
3298
3299         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
3300         page = lookup_swap_cache(entry, vma, vmf->address);
3301         swapcache = page;
3302
3303         if (!page) {
3304                 struct swap_info_struct *si = swp_swap_info(entry);
3305
3306                 if (data_race(si->flags & SWP_SYNCHRONOUS_IO) &&
3307                     __swap_count(entry) == 1) {
3308                         /* skip swapcache */
3309                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
3310                                                         vmf->address);
3311                         if (page) {
3312                                 int err;
3313
3314                                 __SetPageLocked(page);
3315                                 __SetPageSwapBacked(page);
3316                                 set_page_private(page, entry.val);
3317
3318                                 /* Tell memcg to use swap ownership records */
3319                                 SetPageSwapCache(page);
3320                                 err = mem_cgroup_charge(page, vma->vm_mm,
3321                                                         GFP_KERNEL);
3322                                 ClearPageSwapCache(page);
3323                                 if (err) {
3324                                         ret = VM_FAULT_OOM;
3325                                         goto out_page;
3326                                 }
3327
3328                                 shadow = get_shadow_from_swap_cache(entry);
3329                                 if (shadow)
3330                                         workingset_refault(page, shadow);
3331
3332                                 lru_cache_add(page);
3333                                 swap_readpage(page, true);
3334                         }
3335                 } else {
3336                         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
3337                                                 vmf);
3338                         swapcache = page;
3339                 }
3340
3341                 if (!page) {
3342                         /*
3343                          * Back out if somebody else faulted in this pte
3344                          * while we released the pte lock.
3345                          */
3346                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3347                                         vmf->address, &vmf->ptl);
3348                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
3349                                 ret = VM_FAULT_OOM;
3350                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3351                         goto unlock;
3352                 }
3353
3354                 /* Had to read the page from swap area: Major fault */
3355                 ret = VM_FAULT_MAJOR;
3356                 count_vm_event(PGMAJFAULT);
3357                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
3358         } else if (PageHWPoison(page)) {
3359                 /*
3360                  * hwpoisoned dirty swapcache pages are kept for killing
3361                  * owner processes (which may be unknown at hwpoison time)
3362                  */
3363                 ret = VM_FAULT_HWPOISON;
3364                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3365                 goto out_release;
3366         }
3367
3368         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
3369
3370         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
3371         if (!locked) {
3372                 ret |= VM_FAULT_RETRY;
3373                 goto out_release;
3374         }
3375
3376         /*
3377          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
3378          * release the swapcache from under us.  The page pin, and pte_same
3379          * test below, are not enough to exclude that.  Even if it is still
3380          * swapcache, we need to check that the page's swap has not changed.
3381          */
3382         if (unlikely((!PageSwapCache(page) ||
3383                         page_private(page) != entry.val)) && swapcache)
3384                 goto out_page;
3385
3386         page = ksm_might_need_to_copy(page, vma, vmf->address);
3387         if (unlikely(!page)) {
3388                 ret = VM_FAULT_OOM;
3389                 page = swapcache;
3390                 goto out_page;
3391         }
3392
3393         cgroup_throttle_swaprate(page, GFP_KERNEL);
3394
3395         /*
3396          * Back out if somebody else already faulted in this pte.
3397          */
3398         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3399                         &vmf->ptl);
3400         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
3401                 goto out_nomap;
3402
3403         if (unlikely(!PageUptodate(page))) {
3404                 ret = VM_FAULT_SIGBUS;
3405                 goto out_nomap;
3406         }
3407
3408         /*
3409          * The page isn't present yet, go ahead with the fault.
3410          *
3411          * Be careful about the sequence of operations here.
3412          * To get its accounting right, reuse_swap_page() must be called
3413          * while the page is counted on swap but not yet in mapcount i.e.
3414          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
3415          * must be called after the swap_free(), or it will never succeed.
3416          */
3417
3418         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3419         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
3420         pte = mk_pte(page, vma->vm_page_prot);
3421         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
3422                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
3423                 vmf->flags &= ~FAULT_FLAG_WRITE;
3424                 ret |= VM_FAULT_WRITE;
3425                 exclusive = RMAP_EXCLUSIVE;
3426         }
3427         flush_icache_page(vma, page);
3428         if (pte_swp_soft_dirty(vmf->orig_pte))
3429                 pte = pte_mksoft_dirty(pte);
3430         if (pte_swp_uffd_wp(vmf->orig_pte)) {
3431                 pte = pte_mkuffd_wp(pte);
3432                 pte = pte_wrprotect(pte);
3433         }
3434         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
3435         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
3436         vmf->orig_pte = pte;
3437
3438         /* ksm created a completely new copy */
3439         if (unlikely(page != swapcache && swapcache)) {
3440                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3441                 lru_cache_add_inactive_or_unevictable(page, vma);
3442         } else {
3443                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
3444         }
3445
3446         swap_free(entry);
3447         if (mem_cgroup_swap_full(page) ||
3448             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
3449                 try_to_free_swap(page);
3450         unlock_page(page);
3451         if (page != swapcache && swapcache) {
3452                 /*
3453                  * Hold the lock to avoid the swap entry to be reused
3454                  * until we take the PT lock for the pte_same() check
3455                  * (to avoid false positives from pte_same). For
3456                  * further safety release the lock after the swap_free
3457                  * so that the swap count won't change under a
3458                  * parallel locked swapcache.
3459                  */
3460                 unlock_page(swapcache);
3461                 put_page(swapcache);
3462         }
3463
3464         if (vmf->flags & FAULT_FLAG_WRITE) {
3465                 ret |= do_wp_page(vmf);
3466                 if (ret & VM_FAULT_ERROR)
3467                         ret &= VM_FAULT_ERROR;
3468                 goto out;
3469         }
3470
3471         /* No need to invalidate - it was non-present before */
3472         update_mmu_cache(vma, vmf->address, vmf->pte);
3473 unlock:
3474         pte_unmap_unlock(vmf->pte, vmf->ptl);
3475 out:
3476         return ret;
3477 out_nomap:
3478         pte_unmap_unlock(vmf->pte, vmf->ptl);
3479 out_page:
3480         unlock_page(page);
3481 out_release:
3482         put_page(page);
3483         if (page != swapcache && swapcache) {
3484                 unlock_page(swapcache);
3485                 put_page(swapcache);
3486         }
3487         return ret;
3488 }
3489
3490 /*
3491  * We enter with non-exclusive mmap_lock (to exclude vma changes,
3492  * but allow concurrent faults), and pte mapped but not yet locked.
3493  * We return with mmap_lock still held, but pte unmapped and unlocked.
3494  */
3495 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
3496 {
3497         struct vm_area_struct *vma = vmf->vma;
3498         struct page *page;
3499         vm_fault_t ret = 0;
3500         pte_t entry;
3501
3502         /* File mapping without ->vm_ops ? */
3503         if (vma->vm_flags & VM_SHARED)
3504                 return VM_FAULT_SIGBUS;
3505
3506         /*
3507          * Use pte_alloc() instead of pte_alloc_map().  We can't run
3508          * pte_offset_map() on pmds where a huge pmd might be created
3509          * from a different thread.
3510          *
3511          * pte_alloc_map() is safe to use under mmap_write_lock(mm) or when
3512          * parallel threads are excluded by other means.
3513          *
3514          * Here we only have mmap_read_lock(mm).
3515          */
3516         if (pte_alloc(vma->vm_mm, vmf->pmd))
3517                 return VM_FAULT_OOM;
3518
3519         /* See comment in handle_pte_fault() */
3520         if (unlikely(pmd_trans_unstable(vmf->pmd)))
3521                 return 0;
3522
3523         /* Use the zero-page for reads */
3524         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
3525                         !mm_forbids_zeropage(vma->vm_mm)) {
3526                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
3527                                                 vma->vm_page_prot));
3528                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3529                                 vmf->address, &vmf->ptl);
3530                 if (!pte_none(*vmf->pte)) {
3531                         update_mmu_tlb(vma, vmf->address, vmf->pte);
3532                         goto unlock;
3533                 }
3534                 ret = check_stable_address_space(vma->vm_mm);
3535                 if (ret)
3536                         goto unlock;
3537                 /* Deliver the page fault to userland, check inside PT lock */
3538                 if (userfaultfd_missing(vma)) {
3539                         pte_unmap_unlock(vmf->pte, vmf->ptl);
3540                         return handle_userfault(vmf, VM_UFFD_MISSING);
3541                 }
3542                 goto setpte;
3543         }
3544
3545         /* Allocate our own private page. */
3546         if (unlikely(anon_vma_prepare(vma)))
3547                 goto oom;
3548         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
3549         if (!page)
3550                 goto oom;
3551
3552         if (mem_cgroup_charge(page, vma->vm_mm, GFP_KERNEL))
3553                 goto oom_free_page;
3554         cgroup_throttle_swaprate(page, GFP_KERNEL);
3555
3556         /*
3557          * The memory barrier inside __SetPageUptodate makes sure that
3558          * preceding stores to the page contents become visible before
3559          * the set_pte_at() write.
3560          */
3561         __SetPageUptodate(page);
3562
3563         entry = mk_pte(page, vma->vm_page_prot);
3564         if (vma->vm_flags & VM_WRITE)
3565                 entry = pte_mkwrite(pte_mkdirty(entry));
3566
3567         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3568                         &vmf->ptl);
3569         if (!pte_none(*vmf->pte)) {
3570                 update_mmu_cache(vma, vmf->address, vmf->pte);
3571                 goto release;
3572         }
3573
3574         ret = check_stable_address_space(vma->vm_mm);
3575         if (ret)
3576                 goto release;
3577
3578         /* Deliver the page fault to userland, check inside PT lock */
3579         if (userfaultfd_missing(vma)) {
3580                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3581                 put_page(page);
3582                 return handle_userfault(vmf, VM_UFFD_MISSING);
3583         }
3584
3585         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3586         page_add_new_anon_rmap(page, vma, vmf->address, false);
3587         lru_cache_add_inactive_or_unevictable(page, vma);
3588 setpte:
3589         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3590
3591         /* No need to invalidate - it was non-present before */
3592         update_mmu_cache(vma, vmf->address, vmf->pte);
3593 unlock:
3594         pte_unmap_unlock(vmf->pte, vmf->ptl);
3595         return ret;
3596 release:
3597         put_page(page);
3598         goto unlock;
3599 oom_free_page:
3600         put_page(page);
3601 oom:
3602         return VM_FAULT_OOM;
3603 }
3604
3605 /*
3606  * The mmap_lock must have been held on entry, and may have been
3607  * released depending on flags and vma->vm_ops->fault() return value.
3608  * See filemap_fault() and __lock_page_retry().
3609  */
3610 static vm_fault_t __do_fault(struct vm_fault *vmf)
3611 {
3612         struct vm_area_struct *vma = vmf->vma;
3613         vm_fault_t ret;
3614
3615         /*
3616          * Preallocate pte before we take page_lock because this might lead to
3617          * deadlocks for memcg reclaim which waits for pages under writeback:
3618          *                              lock_page(A)
3619          *                              SetPageWriteback(A)
3620          *                              unlock_page(A)
3621          * lock_page(B)
3622          *                              lock_page(B)
3623          * pte_alloc_one
3624          *   shrink_page_list
3625          *     wait_on_page_writeback(A)
3626          *                              SetPageWriteback(B)
3627          *                              unlock_page(B)
3628          *                              # flush A, B to clear the writeback
3629          */
3630         if (pmd_none(*vmf->pmd) && !vmf->prealloc_pte) {
3631                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3632                 if (!vmf->prealloc_pte)
3633                         return VM_FAULT_OOM;
3634                 smp_wmb(); /* See comment in __pte_alloc() */
3635         }
3636
3637         ret = vma->vm_ops->fault(vmf);
3638         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
3639                             VM_FAULT_DONE_COW)))
3640                 return ret;
3641
3642         if (unlikely(PageHWPoison(vmf->page))) {
3643                 if (ret & VM_FAULT_LOCKED)
3644                         unlock_page(vmf->page);
3645                 put_page(vmf->page);
3646                 vmf->page = NULL;
3647                 return VM_FAULT_HWPOISON;
3648         }
3649
3650         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3651                 lock_page(vmf->page);
3652         else
3653                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3654
3655         return ret;
3656 }
3657
3658 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3659 static void deposit_prealloc_pte(struct vm_fault *vmf)
3660 {
3661         struct vm_area_struct *vma = vmf->vma;
3662
3663         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3664         /*
3665          * We are going to consume the prealloc table,
3666          * count that as nr_ptes.
3667          */
3668         mm_inc_nr_ptes(vma->vm_mm);
3669         vmf->prealloc_pte = NULL;
3670 }
3671
3672 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3673 {
3674         struct vm_area_struct *vma = vmf->vma;
3675         bool write = vmf->flags & FAULT_FLAG_WRITE;
3676         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3677         pmd_t entry;
3678         int i;
3679         vm_fault_t ret = VM_FAULT_FALLBACK;
3680
3681         if (!transhuge_vma_suitable(vma, haddr))
3682                 return ret;
3683
3684         page = compound_head(page);
3685         if (compound_order(page) != HPAGE_PMD_ORDER)
3686                 return ret;
3687
3688         /*
3689          * Archs like ppc64 need additonal space to store information
3690          * related to pte entry. Use the preallocated table for that.
3691          */
3692         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3693                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm);
3694                 if (!vmf->prealloc_pte)
3695                         return VM_FAULT_OOM;
3696                 smp_wmb(); /* See comment in __pte_alloc() */
3697         }
3698
3699         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3700         if (unlikely(!pmd_none(*vmf->pmd)))
3701                 goto out;
3702
3703         for (i = 0; i < HPAGE_PMD_NR; i++)
3704                 flush_icache_page(vma, page + i);
3705
3706         entry = mk_huge_pmd(page, vma->vm_page_prot);
3707         if (write)
3708                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3709
3710         add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3711         page_add_file_rmap(page, true);
3712         /*
3713          * deposit and withdraw with pmd lock held
3714          */
3715         if (arch_needs_pgtable_deposit())
3716                 deposit_prealloc_pte(vmf);
3717
3718         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3719
3720         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3721
3722         /* fault is handled */
3723         ret = 0;
3724         count_vm_event(THP_FILE_MAPPED);
3725 out:
3726         spin_unlock(vmf->ptl);
3727         return ret;
3728 }
3729 #else
3730 vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3731 {
3732         return VM_FAULT_FALLBACK;
3733 }
3734 #endif
3735
3736 void do_set_pte(struct vm_fault *vmf, struct page *page, unsigned long addr)
3737 {
3738         struct vm_area_struct *vma = vmf->vma;
3739         bool write = vmf->flags & FAULT_FLAG_WRITE;
3740         bool prefault = vmf->address != addr;
3741         pte_t entry;
3742
3743         flush_icache_page(vma, page);
3744         entry = mk_pte(page, vma->vm_page_prot);
3745
3746         if (prefault && arch_wants_old_prefaulted_pte())
3747                 entry = pte_mkold(entry);
3748
3749         if (write)
3750                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3751         /* copy-on-write page */
3752         if (write && !(vma->vm_flags & VM_SHARED)) {
3753                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3754                 page_add_new_anon_rmap(page, vma, addr, false);
3755                 lru_cache_add_inactive_or_unevictable(page, vma);
3756         } else {
3757                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3758                 page_add_file_rmap(page, false);
3759         }
3760         set_pte_at(vma->vm_mm, addr, vmf->pte, entry);
3761 }
3762
3763 /**
3764  * finish_fault - finish page fault once we have prepared the page to fault
3765  *
3766  * @vmf: structure describing the fault
3767  *
3768  * This function handles all that is needed to finish a page fault once the
3769  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3770  * given page, adds reverse page mapping, handles memcg charges and LRU
3771  * addition.
3772  *
3773  * The function expects the page to be locked and on success it consumes a
3774  * reference of a page being mapped (for the PTE which maps it).
3775  *
3776  * Return: %0 on success, %VM_FAULT_ code in case of error.
3777  */
3778 vm_fault_t finish_fault(struct vm_fault *vmf)
3779 {
3780         struct vm_area_struct *vma = vmf->vma;
3781         struct page *page;
3782         vm_fault_t ret;
3783
3784         /* Did we COW the page? */
3785         if ((vmf->flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
3786                 page = vmf->cow_page;
3787         else
3788                 page = vmf->page;
3789
3790         /*
3791          * check even for read faults because we might have lost our CoWed
3792          * page
3793          */
3794         if (!(vma->vm_flags & VM_SHARED)) {
3795                 ret = check_stable_address_space(vma->vm_mm);
3796                 if (ret)
3797                         return ret;
3798         }
3799
3800         if (pmd_none(*vmf->pmd)) {
3801                 if (PageTransCompound(page)) {
3802                         ret = do_set_pmd(vmf, page);
3803                         if (ret != VM_FAULT_FALLBACK)
3804                                 return ret;
3805                 }
3806
3807                 if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd)))
3808                         return VM_FAULT_OOM;
3809         }
3810
3811         /* See comment in handle_pte_fault() */
3812         if (pmd_devmap_trans_unstable(vmf->pmd))
3813                 return 0;
3814
3815         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
3816                                       vmf->address, &vmf->ptl);
3817         ret = 0;
3818         /* Re-check under ptl */
3819         if (likely(pte_none(*vmf->pte)))
3820                 do_set_pte(vmf, page, vmf->address);
3821         else
3822                 ret = VM_FAULT_NOPAGE;
3823
3824         update_mmu_tlb(vma, vmf->address, vmf->pte);
3825         pte_unmap_unlock(vmf->pte, vmf->ptl);
3826         return ret;
3827 }
3828
3829 static unsigned long fault_around_bytes __read_mostly =
3830         rounddown_pow_of_two(65536);
3831
3832 #ifdef CONFIG_DEBUG_FS
3833 static int fault_around_bytes_get(void *data, u64 *val)
3834 {
3835         *val = fault_around_bytes;
3836         return 0;
3837 }
3838
3839 /*
3840  * fault_around_bytes must be rounded down to the nearest page order as it's
3841  * what do_fault_around() expects to see.
3842  */
3843 static int fault_around_bytes_set(void *data, u64 val)
3844 {
3845         if (val / PAGE_SIZE > PTRS_PER_PTE)
3846                 return -EINVAL;
3847         if (val > PAGE_SIZE)
3848                 fault_around_bytes = rounddown_pow_of_two(val);
3849         else
3850                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3851         return 0;
3852 }
3853 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3854                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3855
3856 static int __init fault_around_debugfs(void)
3857 {
3858         debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3859                                    &fault_around_bytes_fops);
3860         return 0;
3861 }
3862 late_initcall(fault_around_debugfs);
3863 #endif
3864
3865 /*
3866  * do_fault_around() tries to map few pages around the fault address. The hope
3867  * is that the pages will be needed soon and this will lower the number of
3868  * faults to handle.
3869  *
3870  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3871  * not ready to be mapped: not up-to-date, locked, etc.
3872  *
3873  * This function is called with the page table lock taken. In the split ptlock
3874  * case the page table lock only protects only those entries which belong to
3875  * the page table corresponding to the fault address.
3876  *
3877  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3878  * only once.
3879  *
3880  * fault_around_bytes defines how many bytes we'll try to map.
3881  * do_fault_around() expects it to be set to a power of two less than or equal
3882  * to PTRS_PER_PTE.
3883  *
3884  * The virtual address of the area that we map is naturally aligned to
3885  * fault_around_bytes rounded down to the machine page size
3886  * (and therefore to page order).  This way it's easier to guarantee
3887  * that we don't cross page table boundaries.
3888  */
3889 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3890 {
3891         unsigned long address = vmf->address, nr_pages, mask;
3892         pgoff_t start_pgoff = vmf->pgoff;
3893         pgoff_t end_pgoff;
3894         int off;
3895
3896         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3897         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3898
3899         address = max(address & mask, vmf->vma->vm_start);
3900         off = ((vmf->address - address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3901         start_pgoff -= off;
3902
3903         /*
3904          *  end_pgoff is either the end of the page table, the end of
3905          *  the vma or nr_pages from start_pgoff, depending what is nearest.
3906          */
3907         end_pgoff = start_pgoff -
3908                 ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3909                 PTRS_PER_PTE - 1;
3910         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3911                         start_pgoff + nr_pages - 1);
3912
3913         if (pmd_none(*vmf->pmd)) {
3914                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm);
3915                 if (!vmf->prealloc_pte)
3916                         return VM_FAULT_OOM;
3917                 smp_wmb(); /* See comment in __pte_alloc() */
3918         }
3919
3920         return vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3921 }
3922
3923 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3924 {
3925         struct vm_area_struct *vma = vmf->vma;
3926         vm_fault_t ret = 0;
3927
3928         /*
3929          * Let's call ->map_pages() first and use ->fault() as fallback
3930          * if page by the offset is not ready to be mapped (cold cache or
3931          * something).
3932          */
3933         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3934                 ret = do_fault_around(vmf);
3935                 if (ret)
3936                         return ret;
3937         }
3938
3939         ret = __do_fault(vmf);
3940         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3941                 return ret;
3942
3943         ret |= finish_fault(vmf);
3944         unlock_page(vmf->page);
3945         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3946                 put_page(vmf->page);
3947         return ret;
3948 }
3949
3950 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3951 {
3952         struct vm_area_struct *vma = vmf->vma;
3953         vm_fault_t ret;
3954
3955         if (unlikely(anon_vma_prepare(vma)))
3956                 return VM_FAULT_OOM;
3957
3958         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3959         if (!vmf->cow_page)
3960                 return VM_FAULT_OOM;
3961
3962         if (mem_cgroup_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL)) {
3963                 put_page(vmf->cow_page);
3964                 return VM_FAULT_OOM;
3965         }
3966         cgroup_throttle_swaprate(vmf->cow_page, GFP_KERNEL);
3967
3968         ret = __do_fault(vmf);
3969         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3970                 goto uncharge_out;
3971         if (ret & VM_FAULT_DONE_COW)
3972                 return ret;
3973
3974         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3975         __SetPageUptodate(vmf->cow_page);
3976
3977         ret |= finish_fault(vmf);
3978         unlock_page(vmf->page);
3979         put_page(vmf->page);
3980         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3981                 goto uncharge_out;
3982         return ret;
3983 uncharge_out:
3984         put_page(vmf->cow_page);
3985         return ret;
3986 }
3987
3988 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3989 {
3990         struct vm_area_struct *vma = vmf->vma;
3991         vm_fault_t ret, tmp;
3992
3993         ret = __do_fault(vmf);
3994         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3995                 return ret;
3996
3997         /*
3998          * Check if the backing address space wants to know that the page is
3999          * about to become writable
4000          */
4001         if (vma->vm_ops->page_mkwrite) {
4002                 unlock_page(vmf->page);
4003                 tmp = do_page_mkwrite(vmf);
4004                 if (unlikely(!tmp ||
4005                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
4006                         put_page(vmf->page);
4007                         return tmp;
4008                 }
4009         }
4010
4011         ret |= finish_fault(vmf);
4012         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
4013                                         VM_FAULT_RETRY))) {
4014                 unlock_page(vmf->page);
4015                 put_page(vmf->page);
4016                 return ret;
4017         }
4018
4019         ret |= fault_dirty_shared_page(vmf);
4020         return ret;
4021 }
4022
4023 /*
4024  * We enter with non-exclusive mmap_lock (to exclude vma changes,
4025  * but allow concurrent faults).
4026  * The mmap_lock may have been released depending on flags and our
4027  * return value.  See filemap_fault() and __lock_page_or_retry().
4028  * If mmap_lock is released, vma may become invalid (for example
4029  * by other thread calling munmap()).
4030  */
4031 static vm_fault_t do_fault(struct vm_fault *vmf)
4032 {
4033         struct vm_area_struct *vma = vmf->vma;
4034         struct mm_struct *vm_mm = vma->vm_mm;
4035         vm_fault_t ret;
4036
4037         /*
4038          * The VMA was not fully populated on mmap() or missing VM_DONTEXPAND
4039          */
4040         if (!vma->vm_ops->fault) {
4041                 /*
4042                  * If we find a migration pmd entry or a none pmd entry, which
4043                  * should never happen, return SIGBUS
4044                  */
4045                 if (unlikely(!pmd_present(*vmf->pmd)))
4046                         ret = VM_FAULT_SIGBUS;
4047                 else {
4048                         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm,
4049                                                        vmf->pmd,
4050                                                        vmf->address,
4051                                                        &vmf->ptl);
4052                         /*
4053                          * Make sure this is not a temporary clearing of pte
4054                          * by holding ptl and checking again. A R/M/W update
4055                          * of pte involves: take ptl, clearing the pte so that
4056                          * we don't have concurrent modification by hardware
4057                          * followed by an update.
4058                          */
4059                         if (unlikely(pte_none(*vmf->pte)))
4060                                 ret = VM_FAULT_SIGBUS;
4061                         else
4062                                 ret = VM_FAULT_NOPAGE;
4063
4064                         pte_unmap_unlock(vmf->pte, vmf->ptl);
4065                 }
4066         } else if (!(vmf->flags & FAULT_FLAG_WRITE))
4067                 ret = do_read_fault(vmf);
4068         else if (!(vma->vm_flags & VM_SHARED))
4069                 ret = do_cow_fault(vmf);
4070         else
4071                 ret = do_shared_fault(vmf);
4072
4073         /* preallocated pagetable is unused: free it */
4074         if (vmf->prealloc_pte) {
4075                 pte_free(vm_mm, vmf->prealloc_pte);
4076                 vmf->prealloc_pte = NULL;
4077         }
4078         return ret;
4079 }
4080
4081 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
4082                                 unsigned long addr, int page_nid,
4083                                 int *flags)
4084 {
4085         get_page(page);
4086
4087         count_vm_numa_event(NUMA_HINT_FAULTS);
4088         if (page_nid == numa_node_id()) {
4089                 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
4090                 *flags |= TNF_FAULT_LOCAL;
4091         }
4092
4093         return mpol_misplaced(page, vma, addr);
4094 }
4095
4096 static vm_fault_t do_numa_page(struct vm_fault *vmf)
4097 {
4098         struct vm_area_struct *vma = vmf->vma;
4099         struct page *page = NULL;
4100         int page_nid = NUMA_NO_NODE;
4101         int last_cpupid;
4102         int target_nid;
4103         bool migrated = false;
4104         pte_t pte, old_pte;
4105         bool was_writable = pte_savedwrite(vmf->orig_pte);
4106         int flags = 0;
4107
4108         /*
4109          * The "pte" at this point cannot be used safely without
4110          * validation through pte_unmap_same(). It's of NUMA type but
4111          * the pfn may be screwed if the read is non atomic.
4112          */
4113         vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
4114         spin_lock(vmf->ptl);
4115         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
4116                 pte_unmap_unlock(vmf->pte, vmf->ptl);
4117                 goto out;
4118         }
4119
4120         /*
4121          * Make it present again, Depending on how arch implementes non
4122          * accessible ptes, some can allow access by kernel mode.
4123          */
4124         old_pte = ptep_modify_prot_start(vma, vmf->address, vmf->pte);
4125         pte = pte_modify(old_pte, vma->vm_page_prot);
4126         pte = pte_mkyoung(pte);
4127         if (was_writable)
4128                 pte = pte_mkwrite(pte);
4129         ptep_modify_prot_commit(vma, vmf->address, vmf->pte, old_pte, pte);
4130         update_mmu_cache(vma, vmf->address, vmf->pte);
4131
4132         page = vm_normal_page(vma, vmf->address, pte);
4133         if (!page) {
4134                 pte_unmap_unlock(vmf->pte, vmf->ptl);
4135                 return 0;
4136         }
4137
4138         /* TODO: handle PTE-mapped THP */
4139         if (PageCompound(page)) {
4140                 pte_unmap_unlock(vmf->pte, vmf->ptl);
4141                 return 0;
4142         }
4143
4144         /*
4145          * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
4146          * much anyway since they can be in shared cache state. This misses
4147          * the case where a mapping is writable but the process never writes
4148          * to it but pte_write gets cleared during protection updates and
4149          * pte_dirty has unpredictable behaviour between PTE scan updates,
4150          * background writeback, dirty balancing and application behaviour.
4151          */
4152         if (!pte_write(pte))
4153                 flags |= TNF_NO_GROUP;
4154
4155         /*
4156          * Flag if the page is shared between multiple address spaces. This
4157          * is later used when determining whether to group tasks together
4158          */
4159         if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
4160                 flags |= TNF_SHARED;
4161
4162         last_cpupid = page_cpupid_last(page);
4163         page_nid = page_to_nid(page);
4164         target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
4165                         &flags);
4166         pte_unmap_unlock(vmf->pte, vmf->ptl);
4167         if (target_nid == NUMA_NO_NODE) {
4168                 put_page(page);
4169                 goto out;
4170         }
4171
4172         /* Migrate to the requested node */
4173         migrated = migrate_misplaced_page(page, vma, target_nid);
4174         if (migrated) {
4175                 page_nid = target_nid;
4176                 flags |= TNF_MIGRATED;
4177         } else
4178                 flags |= TNF_MIGRATE_FAIL;
4179
4180 out:
4181         if (page_nid != NUMA_NO_NODE)
4182                 task_numa_fault(last_cpupid, page_nid, 1, flags);
4183         return 0;
4184 }
4185
4186 static inline vm_fault_t create_huge_pmd(struct vm_fault *vmf)
4187 {
4188         if (vma_is_anonymous(vmf->vma))
4189                 return do_huge_pmd_anonymous_page(vmf);
4190         if (vmf->vma->vm_ops->huge_fault)
4191                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4192         return VM_FAULT_FALLBACK;
4193 }
4194
4195 /* `inline' is required to avoid gcc 4.1.2 build error */
4196 static inline vm_fault_t wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
4197 {
4198         if (vma_is_anonymous(vmf->vma)) {
4199                 if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
4200                         return handle_userfault(vmf, VM_UFFD_WP);
4201                 return do_huge_pmd_wp_page(vmf, orig_pmd);
4202         }
4203         if (vmf->vma->vm_ops->huge_fault) {
4204                 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
4205
4206                 if (!(ret & VM_FAULT_FALLBACK))
4207                         return ret;
4208         }
4209
4210         /* COW or write-notify handled on pte level: split pmd. */
4211         __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
4212
4213         return VM_FAULT_FALLBACK;
4214 }
4215
4216 static vm_fault_t create_huge_pud(struct vm_fault *vmf)
4217 {
4218 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) &&                     \
4219         defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
4220         /* No support for anonymous transparent PUD pages yet */
4221         if (vma_is_anonymous(vmf->vma))
4222                 goto split;
4223         if (vmf->vma->vm_ops->huge_fault) {
4224                 vm_fault_t ret = vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4225
4226                 if (!(ret & VM_FAULT_FALLBACK))
4227                         return ret;
4228         }
4229 split:
4230         /* COW or write-notify not handled on PUD level: split pud.*/
4231         __split_huge_pud(vmf->vma, vmf->pud, vmf->address);
4232 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4233         return VM_FAULT_FALLBACK;
4234 }
4235
4236 static vm_fault_t wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
4237 {
4238 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
4239         /* No support for anonymous transparent PUD pages yet */
4240         if (vma_is_anonymous(vmf->vma))
4241                 return VM_FAULT_FALLBACK;
4242         if (vmf->vma->vm_ops->huge_fault)
4243                 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
4244 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
4245         return VM_FAULT_FALLBACK;
4246 }
4247
4248 /*
4249  * These routines also need to handle stuff like marking pages dirty
4250  * and/or accessed for architectures that don't do it in hardware (most
4251  * RISC architectures).  The early dirtying is also good on the i386.
4252  *
4253  * There is also a hook called "update_mmu_cache()" that architectures
4254  * with external mmu caches can use to update those (ie the Sparc or
4255  * PowerPC hashed page tables that act as extended TLBs).
4256  *
4257  * We enter with non-exclusive mmap_lock (to exclude vma changes, but allow
4258  * concurrent faults).
4259  *
4260  * The mmap_lock may have been released depending on flags and our return value.
4261  * See filemap_fault() and __lock_page_or_retry().
4262  */
4263 static vm_fault_t handle_pte_fault(struct vm_fault *vmf)
4264 {
4265         pte_t entry;
4266
4267         if (unlikely(pmd_none(*vmf->pmd))) {
4268                 /*
4269                  * Leave __pte_alloc() until later: because vm_ops->fault may
4270                  * want to allocate huge page, and if we expose page table
4271                  * for an instant, it will be difficult to retract from
4272                  * concurrent faults and from rmap lookups.
4273                  */
4274                 vmf->pte = NULL;
4275         } else {
4276                 /*
4277                  * If a huge pmd materialized under us just retry later.  Use
4278                  * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead
4279                  * of pmd_trans_huge() to ensure the pmd didn't become
4280                  * pmd_trans_huge under us and then back to pmd_none, as a
4281                  * result of MADV_DONTNEED running immediately after a huge pmd
4282                  * fault in a different thread of this mm, in turn leading to a
4283                  * misleading pmd_trans_huge() retval. All we have to ensure is
4284                  * that it is a regular pmd that we can walk with
4285                  * pte_offset_map() and we can do that through an atomic read
4286                  * in C, which is what pmd_trans_unstable() provides.
4287                  */
4288                 if (pmd_devmap_trans_unstable(vmf->pmd))
4289                         return 0;
4290                 /*
4291                  * A regular pmd is established and it can't morph into a huge
4292                  * pmd from under us anymore at this point because we hold the
4293                  * mmap_lock read mode and khugepaged takes it in write mode.
4294                  * So now it's safe to run pte_offset_map().
4295                  */
4296                 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
4297                 vmf->orig_pte = *vmf->pte;
4298
4299                 /*
4300                  * some architectures can have larger ptes than wordsize,
4301                  * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
4302                  * CONFIG_32BIT=y, so READ_ONCE cannot guarantee atomic
4303                  * accesses.  The code below just needs a consistent view
4304                  * for the ifs and we later double check anyway with the
4305                  * ptl lock held. So here a barrier will do.
4306                  */
4307                 barrier();
4308                 if (pte_none(vmf->orig_pte)) {
4309                         pte_unmap(vmf->pte);
4310                         vmf->pte = NULL;
4311                 }
4312         }
4313
4314         if (!vmf->pte) {
4315                 if (vma_is_anonymous(vmf->vma))
4316                         return do_anonymous_page(vmf);
4317                 else
4318                         return do_fault(vmf);
4319         }
4320
4321         if (!pte_present(vmf->orig_pte))
4322                 return do_swap_page(vmf);
4323
4324         if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
4325                 return do_numa_page(vmf);
4326
4327         vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
4328         spin_lock(vmf->ptl);
4329         entry = vmf->orig_pte;
4330         if (unlikely(!pte_same(*vmf->pte, entry))) {
4331                 update_mmu_tlb(vmf->vma, vmf->address, vmf->pte);
4332                 goto unlock;
4333         }
4334         if (vmf->flags & FAULT_FLAG_WRITE) {
4335                 if (!pte_write(entry))
4336                         return do_wp_page(vmf);
4337                 entry = pte_mkdirty(entry);
4338         }
4339         entry = pte_mkyoung(entry);
4340         if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
4341                                 vmf->flags & FAULT_FLAG_WRITE)) {
4342                 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
4343         } else {
4344                 /* Skip spurious TLB flush for retried page fault */
4345                 if (vmf->flags & FAULT_FLAG_TRIED)
4346                         goto unlock;
4347                 /*
4348                  * This is needed only for protection faults but the arch code
4349                  * is not yet telling us if this is a protection fault or not.
4350                  * This still avoids useless tlb flushes for .text page faults
4351                  * with threads.
4352                  */
4353                 if (vmf->flags & FAULT_FLAG_WRITE)
4354                         flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
4355         }
4356 unlock:
4357         pte_unmap_unlock(vmf->pte, vmf->ptl);
4358         return 0;
4359 }
4360
4361 /*
4362  * By the time we get here, we already hold the mm semaphore
4363  *
4364  * The mmap_lock may have been released depending on flags and our
4365  * return value.  See filemap_fault() and __lock_page_or_retry().
4366  */
4367 static vm_fault_t __handle_mm_fault(struct vm_area_struct *vma,
4368                 unsigned long address, unsigned int flags)
4369 {
4370         struct vm_fault vmf = {
4371                 .vma = vma,
4372                 .address = address & PAGE_MASK,
4373                 .flags = flags,
4374                 .pgoff = linear_page_index(vma, address),
4375                 .gfp_mask = __get_fault_gfp_mask(vma),
4376         };
4377         unsigned int dirty = flags & FAULT_FLAG_WRITE;
4378         struct mm_struct *mm = vma->vm_mm;
4379         pgd_t *pgd;
4380         p4d_t *p4d;
4381         vm_fault_t ret;
4382
4383         pgd = pgd_offset(mm, address);
4384         p4d = p4d_alloc(mm, pgd, address);
4385         if (!p4d)
4386                 return VM_FAULT_OOM;
4387
4388         vmf.pud = pud_alloc(mm, p4d, address);
4389         if (!vmf.pud)
4390                 return VM_FAULT_OOM;
4391 retry_pud:
4392         if (pud_none(*vmf.pud) && __transparent_hugepage_enabled(vma)) {
4393                 ret = create_huge_pud(&vmf);
4394                 if (!(ret & VM_FAULT_FALLBACK))
4395                         return ret;
4396         } else {
4397                 pud_t orig_pud = *vmf.pud;
4398
4399                 barrier();
4400                 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
4401
4402                         /* NUMA case for anonymous PUDs would go here */
4403
4404                         if (dirty && !pud_write(orig_pud)) {
4405                                 ret = wp_huge_pud(&vmf, orig_pud);
4406                                 if (!(ret & VM_FAULT_FALLBACK))
4407                                         return ret;
4408                         } else {
4409                                 huge_pud_set_accessed(&vmf, orig_pud);
4410                                 return 0;
4411                         }
4412                 }
4413         }
4414
4415         vmf.pmd = pmd_alloc(mm, vmf.pud, address);
4416         if (!vmf.pmd)
4417                 return VM_FAULT_OOM;
4418
4419         /* Huge pud page fault raced with pmd_alloc? */
4420         if (pud_trans_unstable(vmf.pud))
4421                 goto retry_pud;
4422
4423         if (pmd_none(*vmf.pmd) && __transparent_hugepage_enabled(vma)) {
4424                 ret = create_huge_pmd(&vmf);
4425                 if (!(ret & VM_FAULT_FALLBACK))
4426                         return ret;
4427         } else {
4428                 pmd_t orig_pmd = *vmf.pmd;
4429
4430                 barrier();
4431                 if (unlikely(is_swap_pmd(orig_pmd))) {
4432                         VM_BUG_ON(thp_migration_supported() &&
4433                                           !is_pmd_migration_entry(orig_pmd));
4434                         if (is_pmd_migration_entry(orig_pmd))
4435                                 pmd_migration_entry_wait(mm, vmf.pmd);
4436                         return 0;
4437                 }
4438                 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
4439                         if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
4440                                 return do_huge_pmd_numa_page(&vmf, orig_pmd);
4441
4442                         if (dirty && !pmd_write(orig_pmd)) {
4443                                 ret = wp_huge_pmd(&vmf, orig_pmd);
4444                                 if (!(ret & VM_FAULT_FALLBACK))
4445                                         return ret;
4446                         } else {
4447                                 huge_pmd_set_accessed(&vmf, orig_pmd);
4448                                 return 0;
4449                         }
4450                 }
4451         }
4452
4453         return handle_pte_fault(&vmf);
4454 }
4455
4456 /**
4457  * mm_account_fault - Do page fault accountings
4458  *
4459  * @regs: the pt_regs struct pointer.  When set to NULL, will skip accounting
4460  *        of perf event counters, but we'll still do the per-task accounting to
4461  *        the task who triggered this page fault.
4462  * @address: the faulted address.
4463  * @flags: the fault flags.
4464  * @ret: the fault retcode.
4465  *
4466  * This will take care of most of the page fault accountings.  Meanwhile, it
4467  * will also include the PERF_COUNT_SW_PAGE_FAULTS_[MAJ|MIN] perf counter
4468  * updates.  However note that the handling of PERF_COUNT_SW_PAGE_FAULTS should
4469  * still be in per-arch page fault handlers at the entry of page fault.
4470  */
4471 static inline void mm_account_fault(struct pt_regs *regs,
4472                                     unsigned long address, unsigned int flags,
4473                                     vm_fault_t ret)
4474 {
4475         bool major;
4476
4477         /*
4478          * We don't do accounting for some specific faults:
4479          *
4480          * - Unsuccessful faults (e.g. when the address wasn't valid).  That
4481          *   includes arch_vma_access_permitted() failing before reaching here.
4482          *   So this is not a "this many hardware page faults" counter.  We
4483          *   should use the hw profiling for that.
4484          *
4485          * - Incomplete faults (VM_FAULT_RETRY).  They will only be counted
4486          *   once they're completed.
4487          */
4488         if (ret & (VM_FAULT_ERROR | VM_FAULT_RETRY))
4489                 return;
4490
4491         /*
4492          * We define the fault as a major fault when the final successful fault
4493          * is VM_FAULT_MAJOR, or if it retried (which implies that we couldn't
4494          * handle it immediately previously).
4495          */
4496         major = (ret & VM_FAULT_MAJOR) || (flags & FAULT_FLAG_TRIED);
4497
4498         if (major)
4499                 current->maj_flt++;
4500         else
4501                 current->min_flt++;
4502
4503         /*
4504          * If the fault is done for GUP, regs will be NULL.  We only do the
4505          * accounting for the per thread fault counters who triggered the
4506          * fault, and we skip the perf event updates.
4507          */
4508         if (!regs)
4509                 return;
4510
4511         if (major)
4512                 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1, regs, address);
4513         else
4514                 perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1, regs, address);
4515 }
4516
4517 /*
4518  * By the time we get here, we already hold the mm semaphore
4519  *
4520  * The mmap_lock may have been released depending on flags and our
4521  * return value.  See filemap_fault() and __lock_page_or_retry().
4522  */
4523 vm_fault_t handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
4524                            unsigned int flags, struct pt_regs *regs)
4525 {
4526         vm_fault_t ret;
4527
4528         __set_current_state(TASK_RUNNING);
4529
4530         count_vm_event(PGFAULT);
4531         count_memcg_event_mm(vma->vm_mm, PGFAULT);
4532
4533         /* do counter updates before entering really critical section. */
4534         check_sync_rss_stat(current);
4535
4536         if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
4537                                             flags & FAULT_FLAG_INSTRUCTION,
4538                                             flags & FAULT_FLAG_REMOTE))
4539                 return VM_FAULT_SIGSEGV;
4540
4541         /*
4542          * Enable the memcg OOM handling for faults triggered in user
4543          * space.  Kernel faults are handled more gracefully.
4544          */
4545         if (flags & FAULT_FLAG_USER)
4546                 mem_cgroup_enter_user_fault();
4547
4548         if (unlikely(is_vm_hugetlb_page(vma)))
4549                 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
4550         else
4551                 ret = __handle_mm_fault(vma, address, flags);
4552
4553         if (flags & FAULT_FLAG_USER) {
4554                 mem_cgroup_exit_user_fault();
4555                 /*
4556                  * The task may have entered a memcg OOM situation but
4557                  * if the allocation error was handled gracefully (no
4558                  * VM_FAULT_OOM), there is no need to kill anything.
4559                  * Just clean up the OOM state peacefully.
4560                  */
4561                 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
4562                         mem_cgroup_oom_synchronize(false);
4563         }
4564
4565         mm_account_fault(regs, address, flags, ret);
4566
4567         return ret;
4568 }
4569 EXPORT_SYMBOL_GPL(handle_mm_fault);
4570
4571 #ifndef __PAGETABLE_P4D_FOLDED
4572 /*
4573  * Allocate p4d page table.
4574  * We've already handled the fast-path in-line.
4575  */
4576 int __p4d_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
4577 {
4578         p4d_t *new = p4d_alloc_one(mm, address);
4579         if (!new)
4580                 return -ENOMEM;
4581
4582         smp_wmb(); /* See comment in __pte_alloc */
4583
4584         spin_lock(&mm->page_table_lock);
4585         if (pgd_present(*pgd))          /* Another has populated it */
4586                 p4d_free(mm, new);
4587         else
4588                 pgd_populate(mm, pgd, new);
4589         spin_unlock(&mm->page_table_lock);
4590         return 0;
4591 }
4592 #endif /* __PAGETABLE_P4D_FOLDED */
4593
4594 #ifndef __PAGETABLE_PUD_FOLDED
4595 /*
4596  * Allocate page upper directory.
4597  * We've already handled the fast-path in-line.
4598  */
4599 int __pud_alloc(struct mm_struct *mm, p4d_t *p4d, unsigned long address)
4600 {
4601         pud_t *new = pud_alloc_one(mm, address);
4602         if (!new)
4603                 return -ENOMEM;
4604
4605         smp_wmb(); /* See comment in __pte_alloc */
4606
4607         spin_lock(&mm->page_table_lock);
4608         if (!p4d_present(*p4d)) {
4609                 mm_inc_nr_puds(mm);
4610                 p4d_populate(mm, p4d, new);
4611         } else  /* Another has populated it */
4612                 pud_free(mm, new);
4613         spin_unlock(&mm->page_table_lock);
4614         return 0;
4615 }
4616 #endif /* __PAGETABLE_PUD_FOLDED */
4617
4618 #ifndef __PAGETABLE_PMD_FOLDED
4619 /*
4620  * Allocate page middle directory.
4621  * We've already handled the fast-path in-line.
4622  */
4623 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
4624 {
4625         spinlock_t *ptl;
4626         pmd_t *new = pmd_alloc_one(mm, address);
4627         if (!new)
4628                 return -ENOMEM;
4629
4630         smp_wmb(); /* See comment in __pte_alloc */
4631
4632         ptl = pud_lock(mm, pud);
4633         if (!pud_present(*pud)) {
4634                 mm_inc_nr_pmds(mm);
4635                 pud_populate(mm, pud, new);
4636         } else  /* Another has populated it */
4637                 pmd_free(mm, new);
4638         spin_unlock(ptl);
4639         return 0;
4640 }
4641 #endif /* __PAGETABLE_PMD_FOLDED */
4642
4643 int follow_invalidate_pte(struct mm_struct *mm, unsigned long address,
4644                           struct mmu_notifier_range *range, pte_t **ptepp,
4645                           pmd_t **pmdpp, spinlock_t **ptlp)
4646 {
4647         pgd_t *pgd;
4648         p4d_t *p4d;
4649         pud_t *pud;
4650         pmd_t *pmd;
4651         pte_t *ptep;
4652
4653         pgd = pgd_offset(mm, address);
4654         if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
4655                 goto out;
4656
4657         p4d = p4d_offset(pgd, address);
4658         if (p4d_none(*p4d) || unlikely(p4d_bad(*p4d)))
4659                 goto out;
4660
4661         pud = pud_offset(p4d, address);
4662         if (pud_none(*pud) || unlikely(pud_bad(*pud)))
4663                 goto out;
4664
4665         pmd = pmd_offset(pud, address);
4666         VM_BUG_ON(pmd_trans_huge(*pmd));
4667
4668         if (pmd_huge(*pmd)) {
4669                 if (!pmdpp)
4670                         goto out;
4671
4672                 if (range) {
4673                         mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0,
4674                                                 NULL, mm, address & PMD_MASK,
4675                                                 (address & PMD_MASK) + PMD_SIZE);
4676                         mmu_notifier_invalidate_range_start(range);
4677                 }
4678                 *ptlp = pmd_lock(mm, pmd);
4679                 if (pmd_huge(*pmd)) {
4680                         *pmdpp = pmd;
4681                         return 0;
4682                 }
4683                 spin_unlock(*ptlp);
4684                 if (range)
4685                         mmu_notifier_invalidate_range_end(range);
4686         }
4687
4688         if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
4689                 goto out;
4690
4691         if (range) {
4692                 mmu_notifier_range_init(range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
4693                                         address & PAGE_MASK,
4694                                         (address & PAGE_MASK) + PAGE_SIZE);
4695                 mmu_notifier_invalidate_range_start(range);
4696         }
4697         ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
4698         if (!pte_present(*ptep))
4699                 goto unlock;
4700         *ptepp = ptep;
4701         return 0;
4702 unlock:
4703         pte_unmap_unlock(ptep, *ptlp);
4704         if (range)
4705                 mmu_notifier_invalidate_range_end(range);
4706 out:
4707         return -EINVAL;
4708 }
4709
4710 /**
4711  * follow_pte - look up PTE at a user virtual address
4712  * @mm: the mm_struct of the target address space
4713  * @address: user virtual address
4714  * @ptepp: location to store found PTE
4715  * @ptlp: location to store the lock for the PTE
4716  *
4717  * On a successful return, the pointer to the PTE is stored in @ptepp;
4718  * the corresponding lock is taken and its location is stored in @ptlp.
4719  * The contents of the PTE are only stable until @ptlp is released;
4720  * any further use, if any, must be protected against invalidation
4721  * with MMU notifiers.
4722  *
4723  * Only IO mappings and raw PFN mappings are allowed.  The mmap semaphore
4724  * should be taken for read.
4725  *
4726  * KVM uses this function.  While it is arguably less bad than ``follow_pfn``,
4727  * it is not a good general-purpose API.
4728  *
4729  * Return: zero on success, -ve otherwise.
4730  */
4731 int follow_pte(struct mm_struct *mm, unsigned long address,
4732                pte_t **ptepp, spinlock_t **ptlp)
4733 {
4734         return follow_invalidate_pte(mm, address, NULL, ptepp, NULL, ptlp);
4735 }
4736 EXPORT_SYMBOL_GPL(follow_pte);
4737
4738 /**
4739  * follow_pfn - look up PFN at a user virtual address
4740  * @vma: memory mapping
4741  * @address: user virtual address
4742  * @pfn: location to store found PFN
4743  *
4744  * Only IO mappings and raw PFN mappings are allowed.
4745  *
4746  * This function does not allow the caller to read the permissions
4747  * of the PTE.  Do not use it.
4748  *
4749  * Return: zero and the pfn at @pfn on success, -ve otherwise.
4750  */
4751 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
4752         unsigned long *pfn)
4753 {
4754         int ret = -EINVAL;
4755         spinlock_t *ptl;
4756         pte_t *ptep;
4757
4758         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4759                 return ret;
4760
4761         ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
4762         if (ret)
4763                 return ret;
4764         *pfn = pte_pfn(*ptep);
4765         pte_unmap_unlock(ptep, ptl);
4766         return 0;
4767 }
4768 EXPORT_SYMBOL(follow_pfn);
4769
4770 #ifdef CONFIG_HAVE_IOREMAP_PROT
4771 int follow_phys(struct vm_area_struct *vma,
4772                 unsigned long address, unsigned int flags,
4773                 unsigned long *prot, resource_size_t *phys)
4774 {
4775         int ret = -EINVAL;
4776         pte_t *ptep, pte;
4777         spinlock_t *ptl;
4778
4779         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4780                 goto out;
4781
4782         if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
4783                 goto out;
4784         pte = *ptep;
4785
4786         if ((flags & FOLL_WRITE) && !pte_write(pte))
4787                 goto unlock;
4788
4789         *prot = pgprot_val(pte_pgprot(pte));
4790         *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4791
4792         ret = 0;
4793 unlock:
4794         pte_unmap_unlock(ptep, ptl);
4795 out:
4796         return ret;
4797 }
4798
4799 /**
4800  * generic_access_phys - generic implementation for iomem mmap access
4801  * @vma: the vma to access
4802  * @addr: userspace addres, not relative offset within @vma
4803  * @buf: buffer to read/write
4804  * @len: length of transfer
4805  * @write: set to FOLL_WRITE when writing, otherwise reading
4806  *
4807  * This is a generic implementation for &vm_operations_struct.access for an
4808  * iomem mapping. This callback is used by access_process_vm() when the @vma is
4809  * not page based.
4810  */
4811 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
4812                         void *buf, int len, int write)
4813 {
4814         resource_size_t phys_addr;
4815         unsigned long prot = 0;
4816         void __iomem *maddr;
4817         pte_t *ptep, pte;
4818         spinlock_t *ptl;
4819         int offset = offset_in_page(addr);
4820         int ret = -EINVAL;
4821
4822         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
4823                 return -EINVAL;
4824
4825 retry:
4826         if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4827                 return -EINVAL;
4828         pte = *ptep;
4829         pte_unmap_unlock(ptep, ptl);
4830
4831         prot = pgprot_val(pte_pgprot(pte));
4832         phys_addr = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
4833
4834         if ((write & FOLL_WRITE) && !pte_write(pte))
4835                 return -EINVAL;
4836
4837         maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
4838         if (!maddr)
4839                 return -ENOMEM;
4840
4841         if (follow_pte(vma->vm_mm, addr, &ptep, &ptl))
4842                 goto out_unmap;
4843
4844         if (!pte_same(pte, *ptep)) {
4845                 pte_unmap_unlock(ptep, ptl);
4846                 iounmap(maddr);
4847
4848                 goto retry;
4849         }
4850
4851         if (write)
4852                 memcpy_toio(maddr + offset, buf, len);
4853         else
4854                 memcpy_fromio(buf, maddr + offset, len);
4855         ret = len;
4856         pte_unmap_unlock(ptep, ptl);
4857 out_unmap:
4858         iounmap(maddr);
4859
4860         return ret;
4861 }
4862 EXPORT_SYMBOL_GPL(generic_access_phys);
4863 #endif
4864
4865 /*
4866  * Access another process' address space as given in mm.
4867  */
4868 int __access_remote_vm(struct mm_struct *mm, unsigned long addr, void *buf,
4869                        int len, unsigned int gup_flags)
4870 {
4871         struct vm_area_struct *vma;
4872         void *old_buf = buf;
4873         int write = gup_flags & FOLL_WRITE;
4874
4875         if (mmap_read_lock_killable(mm))
4876                 return 0;
4877
4878         /* ignore errors, just check how much was successfully transferred */
4879         while (len) {
4880                 int bytes, ret, offset;
4881                 void *maddr;
4882                 struct page *page = NULL;
4883
4884                 ret = get_user_pages_remote(mm, addr, 1,
4885                                 gup_flags, &page, &vma, NULL);
4886                 if (ret <= 0) {
4887 #ifndef CONFIG_HAVE_IOREMAP_PROT
4888                         break;
4889 #else
4890                         /*
4891                          * Check if this is a VM_IO | VM_PFNMAP VMA, which
4892                          * we can access using slightly different code.
4893                          */
4894                         vma = find_vma(mm, addr);
4895                         if (!vma || vma->vm_start > addr)
4896                                 break;
4897                         if (vma->vm_ops && vma->vm_ops->access)
4898                                 ret = vma->vm_ops->access(vma, addr, buf,
4899                                                           len, write);
4900                         if (ret <= 0)
4901                                 break;
4902                         bytes = ret;
4903 #endif
4904                 } else {
4905                         bytes = len;
4906                         offset = addr & (PAGE_SIZE-1);
4907                         if (bytes > PAGE_SIZE-offset)
4908                                 bytes = PAGE_SIZE-offset;
4909
4910                         maddr = kmap(page);
4911                         if (write) {
4912                                 copy_to_user_page(vma, page, addr,
4913                                                   maddr + offset, buf, bytes);
4914                                 set_page_dirty_lock(page);
4915                         } else {
4916                                 copy_from_user_page(vma, page, addr,
4917                                                     buf, maddr + offset, bytes);
4918                         }
4919                         kunmap(page);
4920                         put_page(page);
4921                 }
4922                 len -= bytes;
4923                 buf += bytes;
4924                 addr += bytes;
4925         }
4926         mmap_read_unlock(mm);
4927
4928         return buf - old_buf;
4929 }
4930
4931 /**
4932  * access_remote_vm - access another process' address space
4933  * @mm:         the mm_struct of the target address space
4934  * @addr:       start address to access
4935  * @buf:        source or destination buffer
4936  * @len:        number of bytes to transfer
4937  * @gup_flags:  flags modifying lookup behaviour
4938  *
4939  * The caller must hold a reference on @mm.
4940  *
4941  * Return: number of bytes copied from source to destination.
4942  */
4943 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4944                 void *buf, int len, unsigned int gup_flags)
4945 {
4946         return __access_remote_vm(mm, addr, buf, len, gup_flags);
4947 }
4948
4949 /*
4950  * Access another process' address space.
4951  * Source/target buffer must be kernel space,
4952  * Do not walk the page table directly, use get_user_pages
4953  */
4954 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4955                 void *buf, int len, unsigned int gup_flags)
4956 {
4957         struct mm_struct *mm;
4958         int ret;
4959
4960         mm = get_task_mm(tsk);
4961         if (!mm)
4962                 return 0;
4963
4964         ret = __access_remote_vm(mm, addr, buf, len, gup_flags);
4965
4966         mmput(mm);
4967
4968         return ret;
4969 }
4970 EXPORT_SYMBOL_GPL(access_process_vm);
4971
4972 /*
4973  * Print the name of a VMA.
4974  */
4975 void print_vma_addr(char *prefix, unsigned long ip)
4976 {
4977         struct mm_struct *mm = current->mm;
4978         struct vm_area_struct *vma;
4979
4980         /*
4981          * we might be running from an atomic context so we cannot sleep
4982          */
4983         if (!mmap_read_trylock(mm))
4984                 return;
4985
4986         vma = find_vma(mm, ip);
4987         if (vma && vma->vm_file) {
4988                 struct file *f = vma->vm_file;
4989                 char *buf = (char *)__get_free_page(GFP_NOWAIT);
4990                 if (buf) {
4991                         char *p;
4992
4993                         p = file_path(f, buf, PAGE_SIZE);
4994                         if (IS_ERR(p))
4995                                 p = "?";
4996                         printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4997                                         vma->vm_start,
4998                                         vma->vm_end - vma->vm_start);
4999                         free_page((unsigned long)buf);
5000                 }
5001         }
5002         mmap_read_unlock(mm);
5003 }
5004
5005 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5006 void __might_fault(const char *file, int line)
5007 {
5008         /*
5009          * Some code (nfs/sunrpc) uses socket ops on kernel memory while
5010          * holding the mmap_lock, this is safe because kernel memory doesn't
5011          * get paged out, therefore we'll never actually fault, and the
5012          * below annotations will generate false positives.
5013          */
5014         if (uaccess_kernel())
5015                 return;
5016         if (pagefault_disabled())
5017                 return;
5018         __might_sleep(file, line, 0);
5019 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
5020         if (current->mm)
5021                 might_lock_read(&current->mm->mmap_lock);
5022 #endif
5023 }
5024 EXPORT_SYMBOL(__might_fault);
5025 #endif
5026
5027 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
5028 /*
5029  * Process all subpages of the specified huge page with the specified
5030  * operation.  The target subpage will be processed last to keep its
5031  * cache lines hot.
5032  */
5033 static inline void process_huge_page(
5034         unsigned long addr_hint, unsigned int pages_per_huge_page,
5035         void (*process_subpage)(unsigned long addr, int idx, void *arg),
5036         void *arg)
5037 {
5038         int i, n, base, l;
5039         unsigned long addr = addr_hint &
5040                 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5041
5042         /* Process target subpage last to keep its cache lines hot */
5043         might_sleep();
5044         n = (addr_hint - addr) / PAGE_SIZE;
5045         if (2 * n <= pages_per_huge_page) {
5046                 /* If target subpage in first half of huge page */
5047                 base = 0;
5048                 l = n;
5049                 /* Process subpages at the end of huge page */
5050                 for (i = pages_per_huge_page - 1; i >= 2 * n; i--) {
5051                         cond_resched();
5052                         process_subpage(addr + i * PAGE_SIZE, i, arg);
5053                 }
5054         } else {
5055                 /* If target subpage in second half of huge page */
5056                 base = pages_per_huge_page - 2 * (pages_per_huge_page - n);
5057                 l = pages_per_huge_page - n;
5058                 /* Process subpages at the begin of huge page */
5059                 for (i = 0; i < base; i++) {
5060                         cond_resched();
5061                         process_subpage(addr + i * PAGE_SIZE, i, arg);
5062                 }
5063         }
5064         /*
5065          * Process remaining subpages in left-right-left-right pattern
5066          * towards the target subpage
5067          */
5068         for (i = 0; i < l; i++) {
5069                 int left_idx = base + i;
5070                 int right_idx = base + 2 * l - 1 - i;
5071
5072                 cond_resched();
5073                 process_subpage(addr + left_idx * PAGE_SIZE, left_idx, arg);
5074                 cond_resched();
5075                 process_subpage(addr + right_idx * PAGE_SIZE, right_idx, arg);
5076         }
5077 }
5078
5079 static void clear_gigantic_page(struct page *page,
5080                                 unsigned long addr,
5081                                 unsigned int pages_per_huge_page)
5082 {
5083         int i;
5084         struct page *p = page;
5085
5086         might_sleep();
5087         for (i = 0; i < pages_per_huge_page;
5088              i++, p = mem_map_next(p, page, i)) {
5089                 cond_resched();
5090                 clear_user_highpage(p, addr + i * PAGE_SIZE);
5091         }
5092 }
5093
5094 static void clear_subpage(unsigned long addr, int idx, void *arg)
5095 {
5096         struct page *page = arg;
5097
5098         clear_user_highpage(page + idx, addr);
5099 }
5100
5101 void clear_huge_page(struct page *page,
5102                      unsigned long addr_hint, unsigned int pages_per_huge_page)
5103 {
5104         unsigned long addr = addr_hint &
5105                 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5106
5107         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5108                 clear_gigantic_page(page, addr, pages_per_huge_page);
5109                 return;
5110         }
5111
5112         process_huge_page(addr_hint, pages_per_huge_page, clear_subpage, page);
5113 }
5114
5115 static void copy_user_gigantic_page(struct page *dst, struct page *src,
5116                                     unsigned long addr,
5117                                     struct vm_area_struct *vma,
5118                                     unsigned int pages_per_huge_page)
5119 {
5120         int i;
5121         struct page *dst_base = dst;
5122         struct page *src_base = src;
5123
5124         for (i = 0; i < pages_per_huge_page; ) {
5125                 cond_resched();
5126                 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
5127
5128                 i++;
5129                 dst = mem_map_next(dst, dst_base, i);
5130                 src = mem_map_next(src, src_base, i);
5131         }
5132 }
5133
5134 struct copy_subpage_arg {
5135         struct page *dst;
5136         struct page *src;
5137         struct vm_area_struct *vma;
5138 };
5139
5140 static void copy_subpage(unsigned long addr, int idx, void *arg)
5141 {
5142         struct copy_subpage_arg *copy_arg = arg;
5143
5144         copy_user_highpage(copy_arg->dst + idx, copy_arg->src + idx,
5145                            addr, copy_arg->vma);
5146 }
5147
5148 void copy_user_huge_page(struct page *dst, struct page *src,
5149                          unsigned long addr_hint, struct vm_area_struct *vma,
5150                          unsigned int pages_per_huge_page)
5151 {
5152         unsigned long addr = addr_hint &
5153                 ~(((unsigned long)pages_per_huge_page << PAGE_SHIFT) - 1);
5154         struct copy_subpage_arg arg = {
5155                 .dst = dst,
5156                 .src = src,
5157                 .vma = vma,
5158         };
5159
5160         if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
5161                 copy_user_gigantic_page(dst, src, addr, vma,
5162                                         pages_per_huge_page);
5163                 return;
5164         }
5165
5166         process_huge_page(addr_hint, pages_per_huge_page, copy_subpage, &arg);
5167 }
5168
5169 long copy_huge_page_from_user(struct page *dst_page,
5170                                 const void __user *usr_src,
5171                                 unsigned int pages_per_huge_page,
5172                                 bool allow_pagefault)
5173 {
5174         void *src = (void *)usr_src;
5175         void *page_kaddr;
5176         unsigned long i, rc = 0;
5177         unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
5178         struct page *subpage = dst_page;
5179
5180         for (i = 0; i < pages_per_huge_page;
5181              i++, subpage = mem_map_next(subpage, dst_page, i)) {
5182                 if (allow_pagefault)
5183                         page_kaddr = kmap(subpage);
5184                 else
5185                         page_kaddr = kmap_atomic(subpage);
5186                 rc = copy_from_user(page_kaddr,
5187                                 (const void __user *)(src + i * PAGE_SIZE),
5188                                 PAGE_SIZE);
5189                 if (allow_pagefault)
5190                         kunmap(subpage);
5191                 else
5192                         kunmap_atomic(page_kaddr);
5193
5194                 ret_val -= (PAGE_SIZE - rc);
5195                 if (rc)
5196                         break;
5197
5198                 cond_resched();
5199         }
5200         return ret_val;
5201 }
5202 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
5203
5204 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
5205
5206 static struct kmem_cache *page_ptl_cachep;
5207
5208 void __init ptlock_cache_init(void)
5209 {
5210         page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
5211                         SLAB_PANIC, NULL);
5212 }
5213
5214 bool ptlock_alloc(struct page *page)
5215 {
5216         spinlock_t *ptl;
5217
5218         ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
5219         if (!ptl)
5220                 return false;
5221         page->ptl = ptl;
5222         return true;
5223 }
5224
5225 void ptlock_free(struct page *page)
5226 {
5227         kmem_cache_free(page_ptl_cachep, page->ptl);
5228 }
5229 #endif