mm/memory.c: recheck page table entry with page table lock held
[muen/linux.git] / mm / memory.c
1 /*
2  *  linux/mm/memory.c
3  *
4  *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
5  */
6
7 /*
8  * demand-loading started 01.12.91 - seems it is high on the list of
9  * things wanted, and it should be easy to implement. - Linus
10  */
11
12 /*
13  * Ok, demand-loading was easy, shared pages a little bit tricker. Shared
14  * pages started 02.12.91, seems to work. - Linus.
15  *
16  * Tested sharing by executing about 30 /bin/sh: under the old kernel it
17  * would have taken more than the 6M I have free, but it worked well as
18  * far as I could see.
19  *
20  * Also corrected some "invalidate()"s - I wasn't doing enough of them.
21  */
22
23 /*
24  * Real VM (paging to/from disk) started 18.12.91. Much more work and
25  * thought has to go into this. Oh, well..
26  * 19.12.91  -  works, somewhat. Sometimes I get faults, don't know why.
27  *              Found it. Everything seems to work now.
28  * 20.12.91  -  Ok, making the swap-device changeable like the root.
29  */
30
31 /*
32  * 05.04.94  -  Multi-page memory management added for v1.1.
33  *              Idea by Alex Bligh (alex@cconcepts.co.uk)
34  *
35  * 16.07.99  -  Support of BIGMEM added by Gerhard Wichert, Siemens AG
36  *              (Gerhard.Wichert@pdb.siemens.de)
37  *
38  * Aug/Sep 2004 Changed to four level page tables (Andi Kleen)
39  */
40
41 #include <linux/kernel_stat.h>
42 #include <linux/mm.h>
43 #include <linux/sched/mm.h>
44 #include <linux/sched/coredump.h>
45 #include <linux/sched/numa_balancing.h>
46 #include <linux/sched/task.h>
47 #include <linux/hugetlb.h>
48 #include <linux/mman.h>
49 #include <linux/swap.h>
50 #include <linux/highmem.h>
51 #include <linux/pagemap.h>
52 #include <linux/memremap.h>
53 #include <linux/ksm.h>
54 #include <linux/rmap.h>
55 #include <linux/export.h>
56 #include <linux/delayacct.h>
57 #include <linux/init.h>
58 #include <linux/pfn_t.h>
59 #include <linux/writeback.h>
60 #include <linux/memcontrol.h>
61 #include <linux/mmu_notifier.h>
62 #include <linux/swapops.h>
63 #include <linux/elf.h>
64 #include <linux/gfp.h>
65 #include <linux/migrate.h>
66 #include <linux/string.h>
67 #include <linux/dma-debug.h>
68 #include <linux/debugfs.h>
69 #include <linux/userfaultfd_k.h>
70 #include <linux/dax.h>
71 #include <linux/oom.h>
72
73 #include <asm/io.h>
74 #include <asm/mmu_context.h>
75 #include <asm/pgalloc.h>
76 #include <linux/uaccess.h>
77 #include <asm/tlb.h>
78 #include <asm/tlbflush.h>
79 #include <asm/pgtable.h>
80
81 #include "internal.h"
82
83 #if defined(LAST_CPUPID_NOT_IN_PAGE_FLAGS) && !defined(CONFIG_COMPILE_TEST)
84 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
85 #endif
86
87 #ifndef CONFIG_NEED_MULTIPLE_NODES
88 /* use the per-pgdat data instead for discontigmem - mbligh */
89 unsigned long max_mapnr;
90 EXPORT_SYMBOL(max_mapnr);
91
92 struct page *mem_map;
93 EXPORT_SYMBOL(mem_map);
94 #endif
95
96 /*
97  * A number of key systems in x86 including ioremap() rely on the assumption
98  * that high_memory defines the upper bound on direct map memory, then end
99  * of ZONE_NORMAL.  Under CONFIG_DISCONTIG this means that max_low_pfn and
100  * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
101  * and ZONE_HIGHMEM.
102  */
103 void *high_memory;
104 EXPORT_SYMBOL(high_memory);
105
106 /*
107  * Randomize the address space (stacks, mmaps, brk, etc.).
108  *
109  * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
110  *   as ancient (libc5 based) binaries can segfault. )
111  */
112 int randomize_va_space __read_mostly =
113 #ifdef CONFIG_COMPAT_BRK
114                                         1;
115 #else
116                                         2;
117 #endif
118
119 static int __init disable_randmaps(char *s)
120 {
121         randomize_va_space = 0;
122         return 1;
123 }
124 __setup("norandmaps", disable_randmaps);
125
126 unsigned long zero_pfn __read_mostly;
127 EXPORT_SYMBOL(zero_pfn);
128
129 unsigned long highest_memmap_pfn __read_mostly;
130
131 /*
132  * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
133  */
134 static int __init init_zero_pfn(void)
135 {
136         zero_pfn = page_to_pfn(ZERO_PAGE(0));
137         return 0;
138 }
139 core_initcall(init_zero_pfn);
140
141
142 #if defined(SPLIT_RSS_COUNTING)
143
144 void sync_mm_rss(struct mm_struct *mm)
145 {
146         int i;
147
148         for (i = 0; i < NR_MM_COUNTERS; i++) {
149                 if (current->rss_stat.count[i]) {
150                         add_mm_counter(mm, i, current->rss_stat.count[i]);
151                         current->rss_stat.count[i] = 0;
152                 }
153         }
154         current->rss_stat.events = 0;
155 }
156
157 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
158 {
159         struct task_struct *task = current;
160
161         if (likely(task->mm == mm))
162                 task->rss_stat.count[member] += val;
163         else
164                 add_mm_counter(mm, member, val);
165 }
166 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
167 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
168
169 /* sync counter once per 64 page faults */
170 #define TASK_RSS_EVENTS_THRESH  (64)
171 static void check_sync_rss_stat(struct task_struct *task)
172 {
173         if (unlikely(task != current))
174                 return;
175         if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
176                 sync_mm_rss(task->mm);
177 }
178 #else /* SPLIT_RSS_COUNTING */
179
180 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
181 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
182
183 static void check_sync_rss_stat(struct task_struct *task)
184 {
185 }
186
187 #endif /* SPLIT_RSS_COUNTING */
188
189 /*
190  * Note: this doesn't free the actual pages themselves. That
191  * has been handled earlier when unmapping all the memory regions.
192  */
193 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
194                            unsigned long addr)
195 {
196         pgtable_t token = pmd_pgtable(*pmd);
197         pmd_clear(pmd);
198         pte_free_tlb(tlb, token, addr);
199         mm_dec_nr_ptes(tlb->mm);
200 }
201
202 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
203                                 unsigned long addr, unsigned long end,
204                                 unsigned long floor, unsigned long ceiling)
205 {
206         pmd_t *pmd;
207         unsigned long next;
208         unsigned long start;
209
210         start = addr;
211         pmd = pmd_offset(pud, addr);
212         do {
213                 next = pmd_addr_end(addr, end);
214                 if (pmd_none_or_clear_bad(pmd))
215                         continue;
216                 free_pte_range(tlb, pmd, addr);
217         } while (pmd++, addr = next, addr != end);
218
219         start &= PUD_MASK;
220         if (start < floor)
221                 return;
222         if (ceiling) {
223                 ceiling &= PUD_MASK;
224                 if (!ceiling)
225                         return;
226         }
227         if (end - 1 > ceiling - 1)
228                 return;
229
230         pmd = pmd_offset(pud, start);
231         pud_clear(pud);
232         pmd_free_tlb(tlb, pmd, start);
233         mm_dec_nr_pmds(tlb->mm);
234 }
235
236 static inline void free_pud_range(struct mmu_gather *tlb, p4d_t *p4d,
237                                 unsigned long addr, unsigned long end,
238                                 unsigned long floor, unsigned long ceiling)
239 {
240         pud_t *pud;
241         unsigned long next;
242         unsigned long start;
243
244         start = addr;
245         pud = pud_offset(p4d, addr);
246         do {
247                 next = pud_addr_end(addr, end);
248                 if (pud_none_or_clear_bad(pud))
249                         continue;
250                 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
251         } while (pud++, addr = next, addr != end);
252
253         start &= P4D_MASK;
254         if (start < floor)
255                 return;
256         if (ceiling) {
257                 ceiling &= P4D_MASK;
258                 if (!ceiling)
259                         return;
260         }
261         if (end - 1 > ceiling - 1)
262                 return;
263
264         pud = pud_offset(p4d, start);
265         p4d_clear(p4d);
266         pud_free_tlb(tlb, pud, start);
267         mm_dec_nr_puds(tlb->mm);
268 }
269
270 static inline void free_p4d_range(struct mmu_gather *tlb, pgd_t *pgd,
271                                 unsigned long addr, unsigned long end,
272                                 unsigned long floor, unsigned long ceiling)
273 {
274         p4d_t *p4d;
275         unsigned long next;
276         unsigned long start;
277
278         start = addr;
279         p4d = p4d_offset(pgd, addr);
280         do {
281                 next = p4d_addr_end(addr, end);
282                 if (p4d_none_or_clear_bad(p4d))
283                         continue;
284                 free_pud_range(tlb, p4d, addr, next, floor, ceiling);
285         } while (p4d++, addr = next, addr != end);
286
287         start &= PGDIR_MASK;
288         if (start < floor)
289                 return;
290         if (ceiling) {
291                 ceiling &= PGDIR_MASK;
292                 if (!ceiling)
293                         return;
294         }
295         if (end - 1 > ceiling - 1)
296                 return;
297
298         p4d = p4d_offset(pgd, start);
299         pgd_clear(pgd);
300         p4d_free_tlb(tlb, p4d, start);
301 }
302
303 /*
304  * This function frees user-level page tables of a process.
305  */
306 void free_pgd_range(struct mmu_gather *tlb,
307                         unsigned long addr, unsigned long end,
308                         unsigned long floor, unsigned long ceiling)
309 {
310         pgd_t *pgd;
311         unsigned long next;
312
313         /*
314          * The next few lines have given us lots of grief...
315          *
316          * Why are we testing PMD* at this top level?  Because often
317          * there will be no work to do at all, and we'd prefer not to
318          * go all the way down to the bottom just to discover that.
319          *
320          * Why all these "- 1"s?  Because 0 represents both the bottom
321          * of the address space and the top of it (using -1 for the
322          * top wouldn't help much: the masks would do the wrong thing).
323          * The rule is that addr 0 and floor 0 refer to the bottom of
324          * the address space, but end 0 and ceiling 0 refer to the top
325          * Comparisons need to use "end - 1" and "ceiling - 1" (though
326          * that end 0 case should be mythical).
327          *
328          * Wherever addr is brought up or ceiling brought down, we must
329          * be careful to reject "the opposite 0" before it confuses the
330          * subsequent tests.  But what about where end is brought down
331          * by PMD_SIZE below? no, end can't go down to 0 there.
332          *
333          * Whereas we round start (addr) and ceiling down, by different
334          * masks at different levels, in order to test whether a table
335          * now has no other vmas using it, so can be freed, we don't
336          * bother to round floor or end up - the tests don't need that.
337          */
338
339         addr &= PMD_MASK;
340         if (addr < floor) {
341                 addr += PMD_SIZE;
342                 if (!addr)
343                         return;
344         }
345         if (ceiling) {
346                 ceiling &= PMD_MASK;
347                 if (!ceiling)
348                         return;
349         }
350         if (end - 1 > ceiling - 1)
351                 end -= PMD_SIZE;
352         if (addr > end - 1)
353                 return;
354         /*
355          * We add page table cache pages with PAGE_SIZE,
356          * (see pte_free_tlb()), flush the tlb if we need
357          */
358         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
359         pgd = pgd_offset(tlb->mm, addr);
360         do {
361                 next = pgd_addr_end(addr, end);
362                 if (pgd_none_or_clear_bad(pgd))
363                         continue;
364                 free_p4d_range(tlb, pgd, addr, next, floor, ceiling);
365         } while (pgd++, addr = next, addr != end);
366 }
367
368 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
369                 unsigned long floor, unsigned long ceiling)
370 {
371         while (vma) {
372                 struct vm_area_struct *next = vma->vm_next;
373                 unsigned long addr = vma->vm_start;
374
375                 /*
376                  * Hide vma from rmap and truncate_pagecache before freeing
377                  * pgtables
378                  */
379                 unlink_anon_vmas(vma);
380                 unlink_file_vma(vma);
381
382                 if (is_vm_hugetlb_page(vma)) {
383                         hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
384                                 floor, next ? next->vm_start : ceiling);
385                 } else {
386                         /*
387                          * Optimization: gather nearby vmas into one call down
388                          */
389                         while (next && next->vm_start <= vma->vm_end + PMD_SIZE
390                                && !is_vm_hugetlb_page(next)) {
391                                 vma = next;
392                                 next = vma->vm_next;
393                                 unlink_anon_vmas(vma);
394                                 unlink_file_vma(vma);
395                         }
396                         free_pgd_range(tlb, addr, vma->vm_end,
397                                 floor, next ? next->vm_start : ceiling);
398                 }
399                 vma = next;
400         }
401 }
402
403 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
404 {
405         spinlock_t *ptl;
406         pgtable_t new = pte_alloc_one(mm, address);
407         if (!new)
408                 return -ENOMEM;
409
410         /*
411          * Ensure all pte setup (eg. pte page lock and page clearing) are
412          * visible before the pte is made visible to other CPUs by being
413          * put into page tables.
414          *
415          * The other side of the story is the pointer chasing in the page
416          * table walking code (when walking the page table without locking;
417          * ie. most of the time). Fortunately, these data accesses consist
418          * of a chain of data-dependent loads, meaning most CPUs (alpha
419          * being the notable exception) will already guarantee loads are
420          * seen in-order. See the alpha page table accessors for the
421          * smp_read_barrier_depends() barriers in page table walking code.
422          */
423         smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
424
425         ptl = pmd_lock(mm, pmd);
426         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
427                 mm_inc_nr_ptes(mm);
428                 pmd_populate(mm, pmd, new);
429                 new = NULL;
430         }
431         spin_unlock(ptl);
432         if (new)
433                 pte_free(mm, new);
434         return 0;
435 }
436
437 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
438 {
439         pte_t *new = pte_alloc_one_kernel(&init_mm, address);
440         if (!new)
441                 return -ENOMEM;
442
443         smp_wmb(); /* See comment in __pte_alloc */
444
445         spin_lock(&init_mm.page_table_lock);
446         if (likely(pmd_none(*pmd))) {   /* Has another populated it ? */
447                 pmd_populate_kernel(&init_mm, pmd, new);
448                 new = NULL;
449         }
450         spin_unlock(&init_mm.page_table_lock);
451         if (new)
452                 pte_free_kernel(&init_mm, new);
453         return 0;
454 }
455
456 static inline void init_rss_vec(int *rss)
457 {
458         memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
459 }
460
461 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
462 {
463         int i;
464
465         if (current->mm == mm)
466                 sync_mm_rss(mm);
467         for (i = 0; i < NR_MM_COUNTERS; i++)
468                 if (rss[i])
469                         add_mm_counter(mm, i, rss[i]);
470 }
471
472 /*
473  * This function is called to print an error when a bad pte
474  * is found. For example, we might have a PFN-mapped pte in
475  * a region that doesn't allow it.
476  *
477  * The calling function must still handle the error.
478  */
479 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
480                           pte_t pte, struct page *page)
481 {
482         pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
483         p4d_t *p4d = p4d_offset(pgd, addr);
484         pud_t *pud = pud_offset(p4d, addr);
485         pmd_t *pmd = pmd_offset(pud, addr);
486         struct address_space *mapping;
487         pgoff_t index;
488         static unsigned long resume;
489         static unsigned long nr_shown;
490         static unsigned long nr_unshown;
491
492         /*
493          * Allow a burst of 60 reports, then keep quiet for that minute;
494          * or allow a steady drip of one report per second.
495          */
496         if (nr_shown == 60) {
497                 if (time_before(jiffies, resume)) {
498                         nr_unshown++;
499                         return;
500                 }
501                 if (nr_unshown) {
502                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
503                                  nr_unshown);
504                         nr_unshown = 0;
505                 }
506                 nr_shown = 0;
507         }
508         if (nr_shown++ == 0)
509                 resume = jiffies + 60 * HZ;
510
511         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
512         index = linear_page_index(vma, addr);
513
514         pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
515                  current->comm,
516                  (long long)pte_val(pte), (long long)pmd_val(*pmd));
517         if (page)
518                 dump_page(page, "bad pte");
519         pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
520                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
521         pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
522                  vma->vm_file,
523                  vma->vm_ops ? vma->vm_ops->fault : NULL,
524                  vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
525                  mapping ? mapping->a_ops->readpage : NULL);
526         dump_stack();
527         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
528 }
529
530 /*
531  * vm_normal_page -- This function gets the "struct page" associated with a pte.
532  *
533  * "Special" mappings do not wish to be associated with a "struct page" (either
534  * it doesn't exist, or it exists but they don't want to touch it). In this
535  * case, NULL is returned here. "Normal" mappings do have a struct page.
536  *
537  * There are 2 broad cases. Firstly, an architecture may define a pte_special()
538  * pte bit, in which case this function is trivial. Secondly, an architecture
539  * may not have a spare pte bit, which requires a more complicated scheme,
540  * described below.
541  *
542  * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
543  * special mapping (even if there are underlying and valid "struct pages").
544  * COWed pages of a VM_PFNMAP are always normal.
545  *
546  * The way we recognize COWed pages within VM_PFNMAP mappings is through the
547  * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
548  * set, and the vm_pgoff will point to the first PFN mapped: thus every special
549  * mapping will always honor the rule
550  *
551  *      pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
552  *
553  * And for normal mappings this is false.
554  *
555  * This restricts such mappings to be a linear translation from virtual address
556  * to pfn. To get around this restriction, we allow arbitrary mappings so long
557  * as the vma is not a COW mapping; in that case, we know that all ptes are
558  * special (because none can have been COWed).
559  *
560  *
561  * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
562  *
563  * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
564  * page" backing, however the difference is that _all_ pages with a struct
565  * page (that is, those where pfn_valid is true) are refcounted and considered
566  * normal pages by the VM. The disadvantage is that pages are refcounted
567  * (which can be slower and simply not an option for some PFNMAP users). The
568  * advantage is that we don't have to follow the strict linearity rule of
569  * PFNMAP mappings in order to support COWable mappings.
570  *
571  */
572 struct page *_vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
573                              pte_t pte, bool with_public_device)
574 {
575         unsigned long pfn = pte_pfn(pte);
576
577         if (IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL)) {
578                 if (likely(!pte_special(pte)))
579                         goto check_pfn;
580                 if (vma->vm_ops && vma->vm_ops->find_special_page)
581                         return vma->vm_ops->find_special_page(vma, addr);
582                 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
583                         return NULL;
584                 if (is_zero_pfn(pfn))
585                         return NULL;
586
587                 /*
588                  * Device public pages are special pages (they are ZONE_DEVICE
589                  * pages but different from persistent memory). They behave
590                  * allmost like normal pages. The difference is that they are
591                  * not on the lru and thus should never be involve with any-
592                  * thing that involve lru manipulation (mlock, numa balancing,
593                  * ...).
594                  *
595                  * This is why we still want to return NULL for such page from
596                  * vm_normal_page() so that we do not have to special case all
597                  * call site of vm_normal_page().
598                  */
599                 if (likely(pfn <= highest_memmap_pfn)) {
600                         struct page *page = pfn_to_page(pfn);
601
602                         if (is_device_public_page(page)) {
603                                 if (with_public_device)
604                                         return page;
605                                 return NULL;
606                         }
607                 }
608
609                 if (pte_devmap(pte))
610                         return NULL;
611
612                 print_bad_pte(vma, addr, pte, NULL);
613                 return NULL;
614         }
615
616         /* !CONFIG_ARCH_HAS_PTE_SPECIAL case follows: */
617
618         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
619                 if (vma->vm_flags & VM_MIXEDMAP) {
620                         if (!pfn_valid(pfn))
621                                 return NULL;
622                         goto out;
623                 } else {
624                         unsigned long off;
625                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
626                         if (pfn == vma->vm_pgoff + off)
627                                 return NULL;
628                         if (!is_cow_mapping(vma->vm_flags))
629                                 return NULL;
630                 }
631         }
632
633         if (is_zero_pfn(pfn))
634                 return NULL;
635
636 check_pfn:
637         if (unlikely(pfn > highest_memmap_pfn)) {
638                 print_bad_pte(vma, addr, pte, NULL);
639                 return NULL;
640         }
641
642         /*
643          * NOTE! We still have PageReserved() pages in the page tables.
644          * eg. VDSO mappings can cause them to exist.
645          */
646 out:
647         return pfn_to_page(pfn);
648 }
649
650 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
651 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
652                                 pmd_t pmd)
653 {
654         unsigned long pfn = pmd_pfn(pmd);
655
656         /*
657          * There is no pmd_special() but there may be special pmds, e.g.
658          * in a direct-access (dax) mapping, so let's just replicate the
659          * !CONFIG_ARCH_HAS_PTE_SPECIAL case from vm_normal_page() here.
660          */
661         if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
662                 if (vma->vm_flags & VM_MIXEDMAP) {
663                         if (!pfn_valid(pfn))
664                                 return NULL;
665                         goto out;
666                 } else {
667                         unsigned long off;
668                         off = (addr - vma->vm_start) >> PAGE_SHIFT;
669                         if (pfn == vma->vm_pgoff + off)
670                                 return NULL;
671                         if (!is_cow_mapping(vma->vm_flags))
672                                 return NULL;
673                 }
674         }
675
676         if (pmd_devmap(pmd))
677                 return NULL;
678         if (is_zero_pfn(pfn))
679                 return NULL;
680         if (unlikely(pfn > highest_memmap_pfn))
681                 return NULL;
682
683         /*
684          * NOTE! We still have PageReserved() pages in the page tables.
685          * eg. VDSO mappings can cause them to exist.
686          */
687 out:
688         return pfn_to_page(pfn);
689 }
690 #endif
691
692 /*
693  * copy one vm_area from one task to the other. Assumes the page tables
694  * already present in the new task to be cleared in the whole range
695  * covered by this vma.
696  */
697
698 static inline unsigned long
699 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
700                 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
701                 unsigned long addr, int *rss)
702 {
703         unsigned long vm_flags = vma->vm_flags;
704         pte_t pte = *src_pte;
705         struct page *page;
706
707         /* pte contains position in swap or file, so copy. */
708         if (unlikely(!pte_present(pte))) {
709                 swp_entry_t entry = pte_to_swp_entry(pte);
710
711                 if (likely(!non_swap_entry(entry))) {
712                         if (swap_duplicate(entry) < 0)
713                                 return entry.val;
714
715                         /* make sure dst_mm is on swapoff's mmlist. */
716                         if (unlikely(list_empty(&dst_mm->mmlist))) {
717                                 spin_lock(&mmlist_lock);
718                                 if (list_empty(&dst_mm->mmlist))
719                                         list_add(&dst_mm->mmlist,
720                                                         &src_mm->mmlist);
721                                 spin_unlock(&mmlist_lock);
722                         }
723                         rss[MM_SWAPENTS]++;
724                 } else if (is_migration_entry(entry)) {
725                         page = migration_entry_to_page(entry);
726
727                         rss[mm_counter(page)]++;
728
729                         if (is_write_migration_entry(entry) &&
730                                         is_cow_mapping(vm_flags)) {
731                                 /*
732                                  * COW mappings require pages in both
733                                  * parent and child to be set to read.
734                                  */
735                                 make_migration_entry_read(&entry);
736                                 pte = swp_entry_to_pte(entry);
737                                 if (pte_swp_soft_dirty(*src_pte))
738                                         pte = pte_swp_mksoft_dirty(pte);
739                                 set_pte_at(src_mm, addr, src_pte, pte);
740                         }
741                 } else if (is_device_private_entry(entry)) {
742                         page = device_private_entry_to_page(entry);
743
744                         /*
745                          * Update rss count even for unaddressable pages, as
746                          * they should treated just like normal pages in this
747                          * respect.
748                          *
749                          * We will likely want to have some new rss counters
750                          * for unaddressable pages, at some point. But for now
751                          * keep things as they are.
752                          */
753                         get_page(page);
754                         rss[mm_counter(page)]++;
755                         page_dup_rmap(page, false);
756
757                         /*
758                          * We do not preserve soft-dirty information, because so
759                          * far, checkpoint/restore is the only feature that
760                          * requires that. And checkpoint/restore does not work
761                          * when a device driver is involved (you cannot easily
762                          * save and restore device driver state).
763                          */
764                         if (is_write_device_private_entry(entry) &&
765                             is_cow_mapping(vm_flags)) {
766                                 make_device_private_entry_read(&entry);
767                                 pte = swp_entry_to_pte(entry);
768                                 set_pte_at(src_mm, addr, src_pte, pte);
769                         }
770                 }
771                 goto out_set_pte;
772         }
773
774         /*
775          * If it's a COW mapping, write protect it both
776          * in the parent and the child
777          */
778         if (is_cow_mapping(vm_flags) && pte_write(pte)) {
779                 ptep_set_wrprotect(src_mm, addr, src_pte);
780                 pte = pte_wrprotect(pte);
781         }
782
783         /*
784          * If it's a shared mapping, mark it clean in
785          * the child
786          */
787         if (vm_flags & VM_SHARED)
788                 pte = pte_mkclean(pte);
789         pte = pte_mkold(pte);
790
791         page = vm_normal_page(vma, addr, pte);
792         if (page) {
793                 get_page(page);
794                 page_dup_rmap(page, false);
795                 rss[mm_counter(page)]++;
796         } else if (pte_devmap(pte)) {
797                 page = pte_page(pte);
798
799                 /*
800                  * Cache coherent device memory behave like regular page and
801                  * not like persistent memory page. For more informations see
802                  * MEMORY_DEVICE_CACHE_COHERENT in memory_hotplug.h
803                  */
804                 if (is_device_public_page(page)) {
805                         get_page(page);
806                         page_dup_rmap(page, false);
807                         rss[mm_counter(page)]++;
808                 }
809         }
810
811 out_set_pte:
812         set_pte_at(dst_mm, addr, dst_pte, pte);
813         return 0;
814 }
815
816 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
817                    pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
818                    unsigned long addr, unsigned long end)
819 {
820         pte_t *orig_src_pte, *orig_dst_pte;
821         pte_t *src_pte, *dst_pte;
822         spinlock_t *src_ptl, *dst_ptl;
823         int progress = 0;
824         int rss[NR_MM_COUNTERS];
825         swp_entry_t entry = (swp_entry_t){0};
826
827 again:
828         init_rss_vec(rss);
829
830         dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
831         if (!dst_pte)
832                 return -ENOMEM;
833         src_pte = pte_offset_map(src_pmd, addr);
834         src_ptl = pte_lockptr(src_mm, src_pmd);
835         spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
836         orig_src_pte = src_pte;
837         orig_dst_pte = dst_pte;
838         arch_enter_lazy_mmu_mode();
839
840         do {
841                 /*
842                  * We are holding two locks at this point - either of them
843                  * could generate latencies in another task on another CPU.
844                  */
845                 if (progress >= 32) {
846                         progress = 0;
847                         if (need_resched() ||
848                             spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
849                                 break;
850                 }
851                 if (pte_none(*src_pte)) {
852                         progress++;
853                         continue;
854                 }
855                 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
856                                                         vma, addr, rss);
857                 if (entry.val)
858                         break;
859                 progress += 8;
860         } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
861
862         arch_leave_lazy_mmu_mode();
863         spin_unlock(src_ptl);
864         pte_unmap(orig_src_pte);
865         add_mm_rss_vec(dst_mm, rss);
866         pte_unmap_unlock(orig_dst_pte, dst_ptl);
867         cond_resched();
868
869         if (entry.val) {
870                 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
871                         return -ENOMEM;
872                 progress = 0;
873         }
874         if (addr != end)
875                 goto again;
876         return 0;
877 }
878
879 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
880                 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
881                 unsigned long addr, unsigned long end)
882 {
883         pmd_t *src_pmd, *dst_pmd;
884         unsigned long next;
885
886         dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
887         if (!dst_pmd)
888                 return -ENOMEM;
889         src_pmd = pmd_offset(src_pud, addr);
890         do {
891                 next = pmd_addr_end(addr, end);
892                 if (is_swap_pmd(*src_pmd) || pmd_trans_huge(*src_pmd)
893                         || pmd_devmap(*src_pmd)) {
894                         int err;
895                         VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
896                         err = copy_huge_pmd(dst_mm, src_mm,
897                                             dst_pmd, src_pmd, addr, vma);
898                         if (err == -ENOMEM)
899                                 return -ENOMEM;
900                         if (!err)
901                                 continue;
902                         /* fall through */
903                 }
904                 if (pmd_none_or_clear_bad(src_pmd))
905                         continue;
906                 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
907                                                 vma, addr, next))
908                         return -ENOMEM;
909         } while (dst_pmd++, src_pmd++, addr = next, addr != end);
910         return 0;
911 }
912
913 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
914                 p4d_t *dst_p4d, p4d_t *src_p4d, struct vm_area_struct *vma,
915                 unsigned long addr, unsigned long end)
916 {
917         pud_t *src_pud, *dst_pud;
918         unsigned long next;
919
920         dst_pud = pud_alloc(dst_mm, dst_p4d, addr);
921         if (!dst_pud)
922                 return -ENOMEM;
923         src_pud = pud_offset(src_p4d, addr);
924         do {
925                 next = pud_addr_end(addr, end);
926                 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
927                         int err;
928
929                         VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
930                         err = copy_huge_pud(dst_mm, src_mm,
931                                             dst_pud, src_pud, addr, vma);
932                         if (err == -ENOMEM)
933                                 return -ENOMEM;
934                         if (!err)
935                                 continue;
936                         /* fall through */
937                 }
938                 if (pud_none_or_clear_bad(src_pud))
939                         continue;
940                 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
941                                                 vma, addr, next))
942                         return -ENOMEM;
943         } while (dst_pud++, src_pud++, addr = next, addr != end);
944         return 0;
945 }
946
947 static inline int copy_p4d_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
948                 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
949                 unsigned long addr, unsigned long end)
950 {
951         p4d_t *src_p4d, *dst_p4d;
952         unsigned long next;
953
954         dst_p4d = p4d_alloc(dst_mm, dst_pgd, addr);
955         if (!dst_p4d)
956                 return -ENOMEM;
957         src_p4d = p4d_offset(src_pgd, addr);
958         do {
959                 next = p4d_addr_end(addr, end);
960                 if (p4d_none_or_clear_bad(src_p4d))
961                         continue;
962                 if (copy_pud_range(dst_mm, src_mm, dst_p4d, src_p4d,
963                                                 vma, addr, next))
964                         return -ENOMEM;
965         } while (dst_p4d++, src_p4d++, addr = next, addr != end);
966         return 0;
967 }
968
969 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
970                 struct vm_area_struct *vma)
971 {
972         pgd_t *src_pgd, *dst_pgd;
973         unsigned long next;
974         unsigned long addr = vma->vm_start;
975         unsigned long end = vma->vm_end;
976         unsigned long mmun_start;       /* For mmu_notifiers */
977         unsigned long mmun_end;         /* For mmu_notifiers */
978         bool is_cow;
979         int ret;
980
981         /*
982          * Don't copy ptes where a page fault will fill them correctly.
983          * Fork becomes much lighter when there are big shared or private
984          * readonly mappings. The tradeoff is that copy_page_range is more
985          * efficient than faulting.
986          */
987         if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
988                         !vma->anon_vma)
989                 return 0;
990
991         if (is_vm_hugetlb_page(vma))
992                 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
993
994         if (unlikely(vma->vm_flags & VM_PFNMAP)) {
995                 /*
996                  * We do not free on error cases below as remove_vma
997                  * gets called on error from higher level routine
998                  */
999                 ret = track_pfn_copy(vma);
1000                 if (ret)
1001                         return ret;
1002         }
1003
1004         /*
1005          * We need to invalidate the secondary MMU mappings only when
1006          * there could be a permission downgrade on the ptes of the
1007          * parent mm. And a permission downgrade will only happen if
1008          * is_cow_mapping() returns true.
1009          */
1010         is_cow = is_cow_mapping(vma->vm_flags);
1011         mmun_start = addr;
1012         mmun_end   = end;
1013         if (is_cow)
1014                 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1015                                                     mmun_end);
1016
1017         ret = 0;
1018         dst_pgd = pgd_offset(dst_mm, addr);
1019         src_pgd = pgd_offset(src_mm, addr);
1020         do {
1021                 next = pgd_addr_end(addr, end);
1022                 if (pgd_none_or_clear_bad(src_pgd))
1023                         continue;
1024                 if (unlikely(copy_p4d_range(dst_mm, src_mm, dst_pgd, src_pgd,
1025                                             vma, addr, next))) {
1026                         ret = -ENOMEM;
1027                         break;
1028                 }
1029         } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1030
1031         if (is_cow)
1032                 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1033         return ret;
1034 }
1035
1036 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1037                                 struct vm_area_struct *vma, pmd_t *pmd,
1038                                 unsigned long addr, unsigned long end,
1039                                 struct zap_details *details)
1040 {
1041         struct mm_struct *mm = tlb->mm;
1042         int force_flush = 0;
1043         int rss[NR_MM_COUNTERS];
1044         spinlock_t *ptl;
1045         pte_t *start_pte;
1046         pte_t *pte;
1047         swp_entry_t entry;
1048
1049         tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1050 again:
1051         init_rss_vec(rss);
1052         start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1053         pte = start_pte;
1054         flush_tlb_batched_pending(mm);
1055         arch_enter_lazy_mmu_mode();
1056         do {
1057                 pte_t ptent = *pte;
1058                 if (pte_none(ptent))
1059                         continue;
1060
1061                 if (pte_present(ptent)) {
1062                         struct page *page;
1063
1064                         page = _vm_normal_page(vma, addr, ptent, true);
1065                         if (unlikely(details) && page) {
1066                                 /*
1067                                  * unmap_shared_mapping_pages() wants to
1068                                  * invalidate cache without truncating:
1069                                  * unmap shared but keep private pages.
1070                                  */
1071                                 if (details->check_mapping &&
1072                                     details->check_mapping != page_rmapping(page))
1073                                         continue;
1074                         }
1075                         ptent = ptep_get_and_clear_full(mm, addr, pte,
1076                                                         tlb->fullmm);
1077                         tlb_remove_tlb_entry(tlb, pte, addr);
1078                         if (unlikely(!page))
1079                                 continue;
1080
1081                         if (!PageAnon(page)) {
1082                                 if (pte_dirty(ptent)) {
1083                                         force_flush = 1;
1084                                         set_page_dirty(page);
1085                                 }
1086                                 if (pte_young(ptent) &&
1087                                     likely(!(vma->vm_flags & VM_SEQ_READ)))
1088                                         mark_page_accessed(page);
1089                         }
1090                         rss[mm_counter(page)]--;
1091                         page_remove_rmap(page, false);
1092                         if (unlikely(page_mapcount(page) < 0))
1093                                 print_bad_pte(vma, addr, ptent, page);
1094                         if (unlikely(__tlb_remove_page(tlb, page))) {
1095                                 force_flush = 1;
1096                                 addr += PAGE_SIZE;
1097                                 break;
1098                         }
1099                         continue;
1100                 }
1101
1102                 entry = pte_to_swp_entry(ptent);
1103                 if (non_swap_entry(entry) && is_device_private_entry(entry)) {
1104                         struct page *page = device_private_entry_to_page(entry);
1105
1106                         if (unlikely(details && details->check_mapping)) {
1107                                 /*
1108                                  * unmap_shared_mapping_pages() wants to
1109                                  * invalidate cache without truncating:
1110                                  * unmap shared but keep private pages.
1111                                  */
1112                                 if (details->check_mapping !=
1113                                     page_rmapping(page))
1114                                         continue;
1115                         }
1116
1117                         pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1118                         rss[mm_counter(page)]--;
1119                         page_remove_rmap(page, false);
1120                         put_page(page);
1121                         continue;
1122                 }
1123
1124                 /* If details->check_mapping, we leave swap entries. */
1125                 if (unlikely(details))
1126                         continue;
1127
1128                 entry = pte_to_swp_entry(ptent);
1129                 if (!non_swap_entry(entry))
1130                         rss[MM_SWAPENTS]--;
1131                 else if (is_migration_entry(entry)) {
1132                         struct page *page;
1133
1134                         page = migration_entry_to_page(entry);
1135                         rss[mm_counter(page)]--;
1136                 }
1137                 if (unlikely(!free_swap_and_cache(entry)))
1138                         print_bad_pte(vma, addr, ptent, NULL);
1139                 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1140         } while (pte++, addr += PAGE_SIZE, addr != end);
1141
1142         add_mm_rss_vec(mm, rss);
1143         arch_leave_lazy_mmu_mode();
1144
1145         /* Do the actual TLB flush before dropping ptl */
1146         if (force_flush)
1147                 tlb_flush_mmu_tlbonly(tlb);
1148         pte_unmap_unlock(start_pte, ptl);
1149
1150         /*
1151          * If we forced a TLB flush (either due to running out of
1152          * batch buffers or because we needed to flush dirty TLB
1153          * entries before releasing the ptl), free the batched
1154          * memory too. Restart if we didn't do everything.
1155          */
1156         if (force_flush) {
1157                 force_flush = 0;
1158                 tlb_flush_mmu_free(tlb);
1159                 if (addr != end)
1160                         goto again;
1161         }
1162
1163         return addr;
1164 }
1165
1166 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1167                                 struct vm_area_struct *vma, pud_t *pud,
1168                                 unsigned long addr, unsigned long end,
1169                                 struct zap_details *details)
1170 {
1171         pmd_t *pmd;
1172         unsigned long next;
1173
1174         pmd = pmd_offset(pud, addr);
1175         do {
1176                 next = pmd_addr_end(addr, end);
1177                 if (is_swap_pmd(*pmd) || pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1178                         if (next - addr != HPAGE_PMD_SIZE)
1179                                 __split_huge_pmd(vma, pmd, addr, false, NULL);
1180                         else if (zap_huge_pmd(tlb, vma, pmd, addr))
1181                                 goto next;
1182                         /* fall through */
1183                 }
1184                 /*
1185                  * Here there can be other concurrent MADV_DONTNEED or
1186                  * trans huge page faults running, and if the pmd is
1187                  * none or trans huge it can change under us. This is
1188                  * because MADV_DONTNEED holds the mmap_sem in read
1189                  * mode.
1190                  */
1191                 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1192                         goto next;
1193                 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1194 next:
1195                 cond_resched();
1196         } while (pmd++, addr = next, addr != end);
1197
1198         return addr;
1199 }
1200
1201 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1202                                 struct vm_area_struct *vma, p4d_t *p4d,
1203                                 unsigned long addr, unsigned long end,
1204                                 struct zap_details *details)
1205 {
1206         pud_t *pud;
1207         unsigned long next;
1208
1209         pud = pud_offset(p4d, addr);
1210         do {
1211                 next = pud_addr_end(addr, end);
1212                 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1213                         if (next - addr != HPAGE_PUD_SIZE) {
1214                                 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1215                                 split_huge_pud(vma, pud, addr);
1216                         } else if (zap_huge_pud(tlb, vma, pud, addr))
1217                                 goto next;
1218                         /* fall through */
1219                 }
1220                 if (pud_none_or_clear_bad(pud))
1221                         continue;
1222                 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1223 next:
1224                 cond_resched();
1225         } while (pud++, addr = next, addr != end);
1226
1227         return addr;
1228 }
1229
1230 static inline unsigned long zap_p4d_range(struct mmu_gather *tlb,
1231                                 struct vm_area_struct *vma, pgd_t *pgd,
1232                                 unsigned long addr, unsigned long end,
1233                                 struct zap_details *details)
1234 {
1235         p4d_t *p4d;
1236         unsigned long next;
1237
1238         p4d = p4d_offset(pgd, addr);
1239         do {
1240                 next = p4d_addr_end(addr, end);
1241                 if (p4d_none_or_clear_bad(p4d))
1242                         continue;
1243                 next = zap_pud_range(tlb, vma, p4d, addr, next, details);
1244         } while (p4d++, addr = next, addr != end);
1245
1246         return addr;
1247 }
1248
1249 void unmap_page_range(struct mmu_gather *tlb,
1250                              struct vm_area_struct *vma,
1251                              unsigned long addr, unsigned long end,
1252                              struct zap_details *details)
1253 {
1254         pgd_t *pgd;
1255         unsigned long next;
1256
1257         BUG_ON(addr >= end);
1258         tlb_start_vma(tlb, vma);
1259         pgd = pgd_offset(vma->vm_mm, addr);
1260         do {
1261                 next = pgd_addr_end(addr, end);
1262                 if (pgd_none_or_clear_bad(pgd))
1263                         continue;
1264                 next = zap_p4d_range(tlb, vma, pgd, addr, next, details);
1265         } while (pgd++, addr = next, addr != end);
1266         tlb_end_vma(tlb, vma);
1267 }
1268
1269
1270 static void unmap_single_vma(struct mmu_gather *tlb,
1271                 struct vm_area_struct *vma, unsigned long start_addr,
1272                 unsigned long end_addr,
1273                 struct zap_details *details)
1274 {
1275         unsigned long start = max(vma->vm_start, start_addr);
1276         unsigned long end;
1277
1278         if (start >= vma->vm_end)
1279                 return;
1280         end = min(vma->vm_end, end_addr);
1281         if (end <= vma->vm_start)
1282                 return;
1283
1284         if (vma->vm_file)
1285                 uprobe_munmap(vma, start, end);
1286
1287         if (unlikely(vma->vm_flags & VM_PFNMAP))
1288                 untrack_pfn(vma, 0, 0);
1289
1290         if (start != end) {
1291                 if (unlikely(is_vm_hugetlb_page(vma))) {
1292                         /*
1293                          * It is undesirable to test vma->vm_file as it
1294                          * should be non-null for valid hugetlb area.
1295                          * However, vm_file will be NULL in the error
1296                          * cleanup path of mmap_region. When
1297                          * hugetlbfs ->mmap method fails,
1298                          * mmap_region() nullifies vma->vm_file
1299                          * before calling this function to clean up.
1300                          * Since no pte has actually been setup, it is
1301                          * safe to do nothing in this case.
1302                          */
1303                         if (vma->vm_file) {
1304                                 i_mmap_lock_write(vma->vm_file->f_mapping);
1305                                 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1306                                 i_mmap_unlock_write(vma->vm_file->f_mapping);
1307                         }
1308                 } else
1309                         unmap_page_range(tlb, vma, start, end, details);
1310         }
1311 }
1312
1313 /**
1314  * unmap_vmas - unmap a range of memory covered by a list of vma's
1315  * @tlb: address of the caller's struct mmu_gather
1316  * @vma: the starting vma
1317  * @start_addr: virtual address at which to start unmapping
1318  * @end_addr: virtual address at which to end unmapping
1319  *
1320  * Unmap all pages in the vma list.
1321  *
1322  * Only addresses between `start' and `end' will be unmapped.
1323  *
1324  * The VMA list must be sorted in ascending virtual address order.
1325  *
1326  * unmap_vmas() assumes that the caller will flush the whole unmapped address
1327  * range after unmap_vmas() returns.  So the only responsibility here is to
1328  * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1329  * drops the lock and schedules.
1330  */
1331 void unmap_vmas(struct mmu_gather *tlb,
1332                 struct vm_area_struct *vma, unsigned long start_addr,
1333                 unsigned long end_addr)
1334 {
1335         struct mm_struct *mm = vma->vm_mm;
1336
1337         mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1338         for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1339                 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1340         mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1341 }
1342
1343 /**
1344  * zap_page_range - remove user pages in a given range
1345  * @vma: vm_area_struct holding the applicable pages
1346  * @start: starting address of pages to zap
1347  * @size: number of bytes to zap
1348  *
1349  * Caller must protect the VMA list
1350  */
1351 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1352                 unsigned long size)
1353 {
1354         struct mm_struct *mm = vma->vm_mm;
1355         struct mmu_gather tlb;
1356         unsigned long end = start + size;
1357
1358         lru_add_drain();
1359         tlb_gather_mmu(&tlb, mm, start, end);
1360         update_hiwater_rss(mm);
1361         mmu_notifier_invalidate_range_start(mm, start, end);
1362         for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1363                 unmap_single_vma(&tlb, vma, start, end, NULL);
1364         mmu_notifier_invalidate_range_end(mm, start, end);
1365         tlb_finish_mmu(&tlb, start, end);
1366 }
1367
1368 /**
1369  * zap_page_range_single - remove user pages in a given range
1370  * @vma: vm_area_struct holding the applicable pages
1371  * @address: starting address of pages to zap
1372  * @size: number of bytes to zap
1373  * @details: details of shared cache invalidation
1374  *
1375  * The range must fit into one VMA.
1376  */
1377 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1378                 unsigned long size, struct zap_details *details)
1379 {
1380         struct mm_struct *mm = vma->vm_mm;
1381         struct mmu_gather tlb;
1382         unsigned long end = address + size;
1383
1384         lru_add_drain();
1385         tlb_gather_mmu(&tlb, mm, address, end);
1386         update_hiwater_rss(mm);
1387         mmu_notifier_invalidate_range_start(mm, address, end);
1388         unmap_single_vma(&tlb, vma, address, end, details);
1389         mmu_notifier_invalidate_range_end(mm, address, end);
1390         tlb_finish_mmu(&tlb, address, end);
1391 }
1392
1393 /**
1394  * zap_vma_ptes - remove ptes mapping the vma
1395  * @vma: vm_area_struct holding ptes to be zapped
1396  * @address: starting address of pages to zap
1397  * @size: number of bytes to zap
1398  *
1399  * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1400  *
1401  * The entire address range must be fully contained within the vma.
1402  *
1403  */
1404 void zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1405                 unsigned long size)
1406 {
1407         if (address < vma->vm_start || address + size > vma->vm_end ||
1408                         !(vma->vm_flags & VM_PFNMAP))
1409                 return;
1410
1411         zap_page_range_single(vma, address, size, NULL);
1412 }
1413 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1414
1415 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1416                         spinlock_t **ptl)
1417 {
1418         pgd_t *pgd;
1419         p4d_t *p4d;
1420         pud_t *pud;
1421         pmd_t *pmd;
1422
1423         pgd = pgd_offset(mm, addr);
1424         p4d = p4d_alloc(mm, pgd, addr);
1425         if (!p4d)
1426                 return NULL;
1427         pud = pud_alloc(mm, p4d, addr);
1428         if (!pud)
1429                 return NULL;
1430         pmd = pmd_alloc(mm, pud, addr);
1431         if (!pmd)
1432                 return NULL;
1433
1434         VM_BUG_ON(pmd_trans_huge(*pmd));
1435         return pte_alloc_map_lock(mm, pmd, addr, ptl);
1436 }
1437
1438 /*
1439  * This is the old fallback for page remapping.
1440  *
1441  * For historical reasons, it only allows reserved pages. Only
1442  * old drivers should use this, and they needed to mark their
1443  * pages reserved for the old functions anyway.
1444  */
1445 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1446                         struct page *page, pgprot_t prot)
1447 {
1448         struct mm_struct *mm = vma->vm_mm;
1449         int retval;
1450         pte_t *pte;
1451         spinlock_t *ptl;
1452
1453         retval = -EINVAL;
1454         if (PageAnon(page))
1455                 goto out;
1456         retval = -ENOMEM;
1457         flush_dcache_page(page);
1458         pte = get_locked_pte(mm, addr, &ptl);
1459         if (!pte)
1460                 goto out;
1461         retval = -EBUSY;
1462         if (!pte_none(*pte))
1463                 goto out_unlock;
1464
1465         /* Ok, finally just insert the thing.. */
1466         get_page(page);
1467         inc_mm_counter_fast(mm, mm_counter_file(page));
1468         page_add_file_rmap(page, false);
1469         set_pte_at(mm, addr, pte, mk_pte(page, prot));
1470
1471         retval = 0;
1472         pte_unmap_unlock(pte, ptl);
1473         return retval;
1474 out_unlock:
1475         pte_unmap_unlock(pte, ptl);
1476 out:
1477         return retval;
1478 }
1479
1480 /**
1481  * vm_insert_page - insert single page into user vma
1482  * @vma: user vma to map to
1483  * @addr: target user address of this page
1484  * @page: source kernel page
1485  *
1486  * This allows drivers to insert individual pages they've allocated
1487  * into a user vma.
1488  *
1489  * The page has to be a nice clean _individual_ kernel allocation.
1490  * If you allocate a compound page, you need to have marked it as
1491  * such (__GFP_COMP), or manually just split the page up yourself
1492  * (see split_page()).
1493  *
1494  * NOTE! Traditionally this was done with "remap_pfn_range()" which
1495  * took an arbitrary page protection parameter. This doesn't allow
1496  * that. Your vma protection will have to be set up correctly, which
1497  * means that if you want a shared writable mapping, you'd better
1498  * ask for a shared writable mapping!
1499  *
1500  * The page does not need to be reserved.
1501  *
1502  * Usually this function is called from f_op->mmap() handler
1503  * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1504  * Caller must set VM_MIXEDMAP on vma if it wants to call this
1505  * function from other places, for example from page-fault handler.
1506  */
1507 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1508                         struct page *page)
1509 {
1510         if (addr < vma->vm_start || addr >= vma->vm_end)
1511                 return -EFAULT;
1512         if (!page_count(page))
1513                 return -EINVAL;
1514         if (!(vma->vm_flags & VM_MIXEDMAP)) {
1515                 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1516                 BUG_ON(vma->vm_flags & VM_PFNMAP);
1517                 vma->vm_flags |= VM_MIXEDMAP;
1518         }
1519         return insert_page(vma, addr, page, vma->vm_page_prot);
1520 }
1521 EXPORT_SYMBOL(vm_insert_page);
1522
1523 static vm_fault_t insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1524                         pfn_t pfn, pgprot_t prot, bool mkwrite)
1525 {
1526         struct mm_struct *mm = vma->vm_mm;
1527         pte_t *pte, entry;
1528         spinlock_t *ptl;
1529
1530         pte = get_locked_pte(mm, addr, &ptl);
1531         if (!pte)
1532                 return VM_FAULT_OOM;
1533         if (!pte_none(*pte)) {
1534                 if (mkwrite) {
1535                         /*
1536                          * For read faults on private mappings the PFN passed
1537                          * in may not match the PFN we have mapped if the
1538                          * mapped PFN is a writeable COW page.  In the mkwrite
1539                          * case we are creating a writable PTE for a shared
1540                          * mapping and we expect the PFNs to match.
1541                          */
1542                         if (WARN_ON_ONCE(pte_pfn(*pte) != pfn_t_to_pfn(pfn)))
1543                                 goto out_unlock;
1544                         entry = *pte;
1545                         goto out_mkwrite;
1546                 } else
1547                         goto out_unlock;
1548         }
1549
1550         /* Ok, finally just insert the thing.. */
1551         if (pfn_t_devmap(pfn))
1552                 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1553         else
1554                 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1555
1556 out_mkwrite:
1557         if (mkwrite) {
1558                 entry = pte_mkyoung(entry);
1559                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1560         }
1561
1562         set_pte_at(mm, addr, pte, entry);
1563         update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1564
1565 out_unlock:
1566         pte_unmap_unlock(pte, ptl);
1567         return VM_FAULT_NOPAGE;
1568 }
1569
1570 /**
1571  * vmf_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1572  * @vma: user vma to map to
1573  * @addr: target user address of this page
1574  * @pfn: source kernel pfn
1575  * @pgprot: pgprot flags for the inserted page
1576  *
1577  * This is exactly like vmf_insert_pfn(), except that it allows drivers to
1578  * to override pgprot on a per-page basis.
1579  *
1580  * This only makes sense for IO mappings, and it makes no sense for
1581  * COW mappings.  In general, using multiple vmas is preferable;
1582  * vmf_insert_pfn_prot should only be used if using multiple VMAs is
1583  * impractical.
1584  *
1585  * Context: Process context.  May allocate using %GFP_KERNEL.
1586  * Return: vm_fault_t value.
1587  */
1588 vm_fault_t vmf_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1589                         unsigned long pfn, pgprot_t pgprot)
1590 {
1591         /*
1592          * Technically, architectures with pte_special can avoid all these
1593          * restrictions (same for remap_pfn_range).  However we would like
1594          * consistency in testing and feature parity among all, so we should
1595          * try to keep these invariants in place for everybody.
1596          */
1597         BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1598         BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1599                                                 (VM_PFNMAP|VM_MIXEDMAP));
1600         BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1601         BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1602
1603         if (addr < vma->vm_start || addr >= vma->vm_end)
1604                 return VM_FAULT_SIGBUS;
1605
1606         if (!pfn_modify_allowed(pfn, pgprot))
1607                 return VM_FAULT_SIGBUS;
1608
1609         track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1610
1611         return insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot,
1612                         false);
1613 }
1614 EXPORT_SYMBOL(vmf_insert_pfn_prot);
1615
1616 /**
1617  * vmf_insert_pfn - insert single pfn into user vma
1618  * @vma: user vma to map to
1619  * @addr: target user address of this page
1620  * @pfn: source kernel pfn
1621  *
1622  * Similar to vm_insert_page, this allows drivers to insert individual pages
1623  * they've allocated into a user vma. Same comments apply.
1624  *
1625  * This function should only be called from a vm_ops->fault handler, and
1626  * in that case the handler should return the result of this function.
1627  *
1628  * vma cannot be a COW mapping.
1629  *
1630  * As this is called only for pages that do not currently exist, we
1631  * do not need to flush old virtual caches or the TLB.
1632  *
1633  * Context: Process context.  May allocate using %GFP_KERNEL.
1634  * Return: vm_fault_t value.
1635  */
1636 vm_fault_t vmf_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1637                         unsigned long pfn)
1638 {
1639         return vmf_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1640 }
1641 EXPORT_SYMBOL(vmf_insert_pfn);
1642
1643 static bool vm_mixed_ok(struct vm_area_struct *vma, pfn_t pfn)
1644 {
1645         /* these checks mirror the abort conditions in vm_normal_page */
1646         if (vma->vm_flags & VM_MIXEDMAP)
1647                 return true;
1648         if (pfn_t_devmap(pfn))
1649                 return true;
1650         if (pfn_t_special(pfn))
1651                 return true;
1652         if (is_zero_pfn(pfn_t_to_pfn(pfn)))
1653                 return true;
1654         return false;
1655 }
1656
1657 static vm_fault_t __vm_insert_mixed(struct vm_area_struct *vma,
1658                 unsigned long addr, pfn_t pfn, bool mkwrite)
1659 {
1660         pgprot_t pgprot = vma->vm_page_prot;
1661         int err;
1662
1663         BUG_ON(!vm_mixed_ok(vma, pfn));
1664
1665         if (addr < vma->vm_start || addr >= vma->vm_end)
1666                 return VM_FAULT_SIGBUS;
1667
1668         track_pfn_insert(vma, &pgprot, pfn);
1669
1670         if (!pfn_modify_allowed(pfn_t_to_pfn(pfn), pgprot))
1671                 return VM_FAULT_SIGBUS;
1672
1673         /*
1674          * If we don't have pte special, then we have to use the pfn_valid()
1675          * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1676          * refcount the page if pfn_valid is true (hence insert_page rather
1677          * than insert_pfn).  If a zero_pfn were inserted into a VM_MIXEDMAP
1678          * without pte special, it would there be refcounted as a normal page.
1679          */
1680         if (!IS_ENABLED(CONFIG_ARCH_HAS_PTE_SPECIAL) &&
1681             !pfn_t_devmap(pfn) && pfn_t_valid(pfn)) {
1682                 struct page *page;
1683
1684                 /*
1685                  * At this point we are committed to insert_page()
1686                  * regardless of whether the caller specified flags that
1687                  * result in pfn_t_has_page() == false.
1688                  */
1689                 page = pfn_to_page(pfn_t_to_pfn(pfn));
1690                 err = insert_page(vma, addr, page, pgprot);
1691         } else {
1692                 return insert_pfn(vma, addr, pfn, pgprot, mkwrite);
1693         }
1694
1695         if (err == -ENOMEM)
1696                 return VM_FAULT_OOM;
1697         if (err < 0 && err != -EBUSY)
1698                 return VM_FAULT_SIGBUS;
1699
1700         return VM_FAULT_NOPAGE;
1701 }
1702
1703 vm_fault_t vmf_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1704                 pfn_t pfn)
1705 {
1706         return __vm_insert_mixed(vma, addr, pfn, false);
1707 }
1708 EXPORT_SYMBOL(vmf_insert_mixed);
1709
1710 /*
1711  *  If the insertion of PTE failed because someone else already added a
1712  *  different entry in the mean time, we treat that as success as we assume
1713  *  the same entry was actually inserted.
1714  */
1715 vm_fault_t vmf_insert_mixed_mkwrite(struct vm_area_struct *vma,
1716                 unsigned long addr, pfn_t pfn)
1717 {
1718         return __vm_insert_mixed(vma, addr, pfn, true);
1719 }
1720 EXPORT_SYMBOL(vmf_insert_mixed_mkwrite);
1721
1722 /*
1723  * maps a range of physical memory into the requested pages. the old
1724  * mappings are removed. any references to nonexistent pages results
1725  * in null mappings (currently treated as "copy-on-access")
1726  */
1727 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1728                         unsigned long addr, unsigned long end,
1729                         unsigned long pfn, pgprot_t prot)
1730 {
1731         pte_t *pte;
1732         spinlock_t *ptl;
1733         int err = 0;
1734
1735         pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1736         if (!pte)
1737                 return -ENOMEM;
1738         arch_enter_lazy_mmu_mode();
1739         do {
1740                 BUG_ON(!pte_none(*pte));
1741                 if (!pfn_modify_allowed(pfn, prot)) {
1742                         err = -EACCES;
1743                         break;
1744                 }
1745                 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1746                 pfn++;
1747         } while (pte++, addr += PAGE_SIZE, addr != end);
1748         arch_leave_lazy_mmu_mode();
1749         pte_unmap_unlock(pte - 1, ptl);
1750         return err;
1751 }
1752
1753 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1754                         unsigned long addr, unsigned long end,
1755                         unsigned long pfn, pgprot_t prot)
1756 {
1757         pmd_t *pmd;
1758         unsigned long next;
1759         int err;
1760
1761         pfn -= addr >> PAGE_SHIFT;
1762         pmd = pmd_alloc(mm, pud, addr);
1763         if (!pmd)
1764                 return -ENOMEM;
1765         VM_BUG_ON(pmd_trans_huge(*pmd));
1766         do {
1767                 next = pmd_addr_end(addr, end);
1768                 err = remap_pte_range(mm, pmd, addr, next,
1769                                 pfn + (addr >> PAGE_SHIFT), prot);
1770                 if (err)
1771                         return err;
1772         } while (pmd++, addr = next, addr != end);
1773         return 0;
1774 }
1775
1776 static inline int remap_pud_range(struct mm_struct *mm, p4d_t *p4d,
1777                         unsigned long addr, unsigned long end,
1778                         unsigned long pfn, pgprot_t prot)
1779 {
1780         pud_t *pud;
1781         unsigned long next;
1782         int err;
1783
1784         pfn -= addr >> PAGE_SHIFT;
1785         pud = pud_alloc(mm, p4d, addr);
1786         if (!pud)
1787                 return -ENOMEM;
1788         do {
1789                 next = pud_addr_end(addr, end);
1790                 err = remap_pmd_range(mm, pud, addr, next,
1791                                 pfn + (addr >> PAGE_SHIFT), prot);
1792                 if (err)
1793                         return err;
1794         } while (pud++, addr = next, addr != end);
1795         return 0;
1796 }
1797
1798 static inline int remap_p4d_range(struct mm_struct *mm, pgd_t *pgd,
1799                         unsigned long addr, unsigned long end,
1800                         unsigned long pfn, pgprot_t prot)
1801 {
1802         p4d_t *p4d;
1803         unsigned long next;
1804         int err;
1805
1806         pfn -= addr >> PAGE_SHIFT;
1807         p4d = p4d_alloc(mm, pgd, addr);
1808         if (!p4d)
1809                 return -ENOMEM;
1810         do {
1811                 next = p4d_addr_end(addr, end);
1812                 err = remap_pud_range(mm, p4d, addr, next,
1813                                 pfn + (addr >> PAGE_SHIFT), prot);
1814                 if (err)
1815                         return err;
1816         } while (p4d++, addr = next, addr != end);
1817         return 0;
1818 }
1819
1820 /**
1821  * remap_pfn_range - remap kernel memory to userspace
1822  * @vma: user vma to map to
1823  * @addr: target user address to start at
1824  * @pfn: physical address of kernel memory
1825  * @size: size of map area
1826  * @prot: page protection flags for this mapping
1827  *
1828  *  Note: this is only safe if the mm semaphore is held when called.
1829  */
1830 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1831                     unsigned long pfn, unsigned long size, pgprot_t prot)
1832 {
1833         pgd_t *pgd;
1834         unsigned long next;
1835         unsigned long end = addr + PAGE_ALIGN(size);
1836         struct mm_struct *mm = vma->vm_mm;
1837         unsigned long remap_pfn = pfn;
1838         int err;
1839
1840         /*
1841          * Physically remapped pages are special. Tell the
1842          * rest of the world about it:
1843          *   VM_IO tells people not to look at these pages
1844          *      (accesses can have side effects).
1845          *   VM_PFNMAP tells the core MM that the base pages are just
1846          *      raw PFN mappings, and do not have a "struct page" associated
1847          *      with them.
1848          *   VM_DONTEXPAND
1849          *      Disable vma merging and expanding with mremap().
1850          *   VM_DONTDUMP
1851          *      Omit vma from core dump, even when VM_IO turned off.
1852          *
1853          * There's a horrible special case to handle copy-on-write
1854          * behaviour that some programs depend on. We mark the "original"
1855          * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1856          * See vm_normal_page() for details.
1857          */
1858         if (is_cow_mapping(vma->vm_flags)) {
1859                 if (addr != vma->vm_start || end != vma->vm_end)
1860                         return -EINVAL;
1861                 vma->vm_pgoff = pfn;
1862         }
1863
1864         err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1865         if (err)
1866                 return -EINVAL;
1867
1868         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1869
1870         BUG_ON(addr >= end);
1871         pfn -= addr >> PAGE_SHIFT;
1872         pgd = pgd_offset(mm, addr);
1873         flush_cache_range(vma, addr, end);
1874         do {
1875                 next = pgd_addr_end(addr, end);
1876                 err = remap_p4d_range(mm, pgd, addr, next,
1877                                 pfn + (addr >> PAGE_SHIFT), prot);
1878                 if (err)
1879                         break;
1880         } while (pgd++, addr = next, addr != end);
1881
1882         if (err)
1883                 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1884
1885         return err;
1886 }
1887 EXPORT_SYMBOL(remap_pfn_range);
1888
1889 /**
1890  * vm_iomap_memory - remap memory to userspace
1891  * @vma: user vma to map to
1892  * @start: start of area
1893  * @len: size of area
1894  *
1895  * This is a simplified io_remap_pfn_range() for common driver use. The
1896  * driver just needs to give us the physical memory range to be mapped,
1897  * we'll figure out the rest from the vma information.
1898  *
1899  * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1900  * whatever write-combining details or similar.
1901  */
1902 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1903 {
1904         unsigned long vm_len, pfn, pages;
1905
1906         /* Check that the physical memory area passed in looks valid */
1907         if (start + len < start)
1908                 return -EINVAL;
1909         /*
1910          * You *really* shouldn't map things that aren't page-aligned,
1911          * but we've historically allowed it because IO memory might
1912          * just have smaller alignment.
1913          */
1914         len += start & ~PAGE_MASK;
1915         pfn = start >> PAGE_SHIFT;
1916         pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1917         if (pfn + pages < pfn)
1918                 return -EINVAL;
1919
1920         /* We start the mapping 'vm_pgoff' pages into the area */
1921         if (vma->vm_pgoff > pages)
1922                 return -EINVAL;
1923         pfn += vma->vm_pgoff;
1924         pages -= vma->vm_pgoff;
1925
1926         /* Can we fit all of the mapping? */
1927         vm_len = vma->vm_end - vma->vm_start;
1928         if (vm_len >> PAGE_SHIFT > pages)
1929                 return -EINVAL;
1930
1931         /* Ok, let it rip */
1932         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1933 }
1934 EXPORT_SYMBOL(vm_iomap_memory);
1935
1936 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1937                                      unsigned long addr, unsigned long end,
1938                                      pte_fn_t fn, void *data)
1939 {
1940         pte_t *pte;
1941         int err;
1942         pgtable_t token;
1943         spinlock_t *uninitialized_var(ptl);
1944
1945         pte = (mm == &init_mm) ?
1946                 pte_alloc_kernel(pmd, addr) :
1947                 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1948         if (!pte)
1949                 return -ENOMEM;
1950
1951         BUG_ON(pmd_huge(*pmd));
1952
1953         arch_enter_lazy_mmu_mode();
1954
1955         token = pmd_pgtable(*pmd);
1956
1957         do {
1958                 err = fn(pte++, token, addr, data);
1959                 if (err)
1960                         break;
1961         } while (addr += PAGE_SIZE, addr != end);
1962
1963         arch_leave_lazy_mmu_mode();
1964
1965         if (mm != &init_mm)
1966                 pte_unmap_unlock(pte-1, ptl);
1967         return err;
1968 }
1969
1970 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1971                                      unsigned long addr, unsigned long end,
1972                                      pte_fn_t fn, void *data)
1973 {
1974         pmd_t *pmd;
1975         unsigned long next;
1976         int err;
1977
1978         BUG_ON(pud_huge(*pud));
1979
1980         pmd = pmd_alloc(mm, pud, addr);
1981         if (!pmd)
1982                 return -ENOMEM;
1983         do {
1984                 next = pmd_addr_end(addr, end);
1985                 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1986                 if (err)
1987                         break;
1988         } while (pmd++, addr = next, addr != end);
1989         return err;
1990 }
1991
1992 static int apply_to_pud_range(struct mm_struct *mm, p4d_t *p4d,
1993                                      unsigned long addr, unsigned long end,
1994                                      pte_fn_t fn, void *data)
1995 {
1996         pud_t *pud;
1997         unsigned long next;
1998         int err;
1999
2000         pud = pud_alloc(mm, p4d, addr);
2001         if (!pud)
2002                 return -ENOMEM;
2003         do {
2004                 next = pud_addr_end(addr, end);
2005                 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
2006                 if (err)
2007                         break;
2008         } while (pud++, addr = next, addr != end);
2009         return err;
2010 }
2011
2012 static int apply_to_p4d_range(struct mm_struct *mm, pgd_t *pgd,
2013                                      unsigned long addr, unsigned long end,
2014                                      pte_fn_t fn, void *data)
2015 {
2016         p4d_t *p4d;
2017         unsigned long next;
2018         int err;
2019
2020         p4d = p4d_alloc(mm, pgd, addr);
2021         if (!p4d)
2022                 return -ENOMEM;
2023         do {
2024                 next = p4d_addr_end(addr, end);
2025                 err = apply_to_pud_range(mm, p4d, addr, next, fn, data);
2026                 if (err)
2027                         break;
2028         } while (p4d++, addr = next, addr != end);
2029         return err;
2030 }
2031
2032 /*
2033  * Scan a region of virtual memory, filling in page tables as necessary
2034  * and calling a provided function on each leaf page table.
2035  */
2036 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
2037                         unsigned long size, pte_fn_t fn, void *data)
2038 {
2039         pgd_t *pgd;
2040         unsigned long next;
2041         unsigned long end = addr + size;
2042         int err;
2043
2044         if (WARN_ON(addr >= end))
2045                 return -EINVAL;
2046
2047         pgd = pgd_offset(mm, addr);
2048         do {
2049                 next = pgd_addr_end(addr, end);
2050                 err = apply_to_p4d_range(mm, pgd, addr, next, fn, data);
2051                 if (err)
2052                         break;
2053         } while (pgd++, addr = next, addr != end);
2054
2055         return err;
2056 }
2057 EXPORT_SYMBOL_GPL(apply_to_page_range);
2058
2059 /*
2060  * handle_pte_fault chooses page fault handler according to an entry which was
2061  * read non-atomically.  Before making any commitment, on those architectures
2062  * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
2063  * parts, do_swap_page must check under lock before unmapping the pte and
2064  * proceeding (but do_wp_page is only called after already making such a check;
2065  * and do_anonymous_page can safely check later on).
2066  */
2067 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
2068                                 pte_t *page_table, pte_t orig_pte)
2069 {
2070         int same = 1;
2071 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
2072         if (sizeof(pte_t) > sizeof(unsigned long)) {
2073                 spinlock_t *ptl = pte_lockptr(mm, pmd);
2074                 spin_lock(ptl);
2075                 same = pte_same(*page_table, orig_pte);
2076                 spin_unlock(ptl);
2077         }
2078 #endif
2079         pte_unmap(page_table);
2080         return same;
2081 }
2082
2083 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2084 {
2085         debug_dma_assert_idle(src);
2086
2087         /*
2088          * If the source page was a PFN mapping, we don't have
2089          * a "struct page" for it. We do a best-effort copy by
2090          * just copying from the original user address. If that
2091          * fails, we just zero-fill it. Live with it.
2092          */
2093         if (unlikely(!src)) {
2094                 void *kaddr = kmap_atomic(dst);
2095                 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2096
2097                 /*
2098                  * This really shouldn't fail, because the page is there
2099                  * in the page tables. But it might just be unreadable,
2100                  * in which case we just give up and fill the result with
2101                  * zeroes.
2102                  */
2103                 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2104                         clear_page(kaddr);
2105                 kunmap_atomic(kaddr);
2106                 flush_dcache_page(dst);
2107         } else
2108                 copy_user_highpage(dst, src, va, vma);
2109 }
2110
2111 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2112 {
2113         struct file *vm_file = vma->vm_file;
2114
2115         if (vm_file)
2116                 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2117
2118         /*
2119          * Special mappings (e.g. VDSO) do not have any file so fake
2120          * a default GFP_KERNEL for them.
2121          */
2122         return GFP_KERNEL;
2123 }
2124
2125 /*
2126  * Notify the address space that the page is about to become writable so that
2127  * it can prohibit this or wait for the page to get into an appropriate state.
2128  *
2129  * We do this without the lock held, so that it can sleep if it needs to.
2130  */
2131 static vm_fault_t do_page_mkwrite(struct vm_fault *vmf)
2132 {
2133         vm_fault_t ret;
2134         struct page *page = vmf->page;
2135         unsigned int old_flags = vmf->flags;
2136
2137         vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2138
2139         ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2140         /* Restore original flags so that caller is not surprised */
2141         vmf->flags = old_flags;
2142         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2143                 return ret;
2144         if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2145                 lock_page(page);
2146                 if (!page->mapping) {
2147                         unlock_page(page);
2148                         return 0; /* retry */
2149                 }
2150                 ret |= VM_FAULT_LOCKED;
2151         } else
2152                 VM_BUG_ON_PAGE(!PageLocked(page), page);
2153         return ret;
2154 }
2155
2156 /*
2157  * Handle dirtying of a page in shared file mapping on a write fault.
2158  *
2159  * The function expects the page to be locked and unlocks it.
2160  */
2161 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2162                                     struct page *page)
2163 {
2164         struct address_space *mapping;
2165         bool dirtied;
2166         bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2167
2168         dirtied = set_page_dirty(page);
2169         VM_BUG_ON_PAGE(PageAnon(page), page);
2170         /*
2171          * Take a local copy of the address_space - page.mapping may be zeroed
2172          * by truncate after unlock_page().   The address_space itself remains
2173          * pinned by vma->vm_file's reference.  We rely on unlock_page()'s
2174          * release semantics to prevent the compiler from undoing this copying.
2175          */
2176         mapping = page_rmapping(page);
2177         unlock_page(page);
2178
2179         if ((dirtied || page_mkwrite) && mapping) {
2180                 /*
2181                  * Some device drivers do not set page.mapping
2182                  * but still dirty their pages
2183                  */
2184                 balance_dirty_pages_ratelimited(mapping);
2185         }
2186
2187         if (!page_mkwrite)
2188                 file_update_time(vma->vm_file);
2189 }
2190
2191 /*
2192  * Handle write page faults for pages that can be reused in the current vma
2193  *
2194  * This can happen either due to the mapping being with the VM_SHARED flag,
2195  * or due to us being the last reference standing to the page. In either
2196  * case, all we need to do here is to mark the page as writable and update
2197  * any related book-keeping.
2198  */
2199 static inline void wp_page_reuse(struct vm_fault *vmf)
2200         __releases(vmf->ptl)
2201 {
2202         struct vm_area_struct *vma = vmf->vma;
2203         struct page *page = vmf->page;
2204         pte_t entry;
2205         /*
2206          * Clear the pages cpupid information as the existing
2207          * information potentially belongs to a now completely
2208          * unrelated process.
2209          */
2210         if (page)
2211                 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2212
2213         flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2214         entry = pte_mkyoung(vmf->orig_pte);
2215         entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2216         if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2217                 update_mmu_cache(vma, vmf->address, vmf->pte);
2218         pte_unmap_unlock(vmf->pte, vmf->ptl);
2219 }
2220
2221 /*
2222  * Handle the case of a page which we actually need to copy to a new page.
2223  *
2224  * Called with mmap_sem locked and the old page referenced, but
2225  * without the ptl held.
2226  *
2227  * High level logic flow:
2228  *
2229  * - Allocate a page, copy the content of the old page to the new one.
2230  * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2231  * - Take the PTL. If the pte changed, bail out and release the allocated page
2232  * - If the pte is still the way we remember it, update the page table and all
2233  *   relevant references. This includes dropping the reference the page-table
2234  *   held to the old page, as well as updating the rmap.
2235  * - In any case, unlock the PTL and drop the reference we took to the old page.
2236  */
2237 static vm_fault_t wp_page_copy(struct vm_fault *vmf)
2238 {
2239         struct vm_area_struct *vma = vmf->vma;
2240         struct mm_struct *mm = vma->vm_mm;
2241         struct page *old_page = vmf->page;
2242         struct page *new_page = NULL;
2243         pte_t entry;
2244         int page_copied = 0;
2245         const unsigned long mmun_start = vmf->address & PAGE_MASK;
2246         const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2247         struct mem_cgroup *memcg;
2248
2249         if (unlikely(anon_vma_prepare(vma)))
2250                 goto oom;
2251
2252         if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2253                 new_page = alloc_zeroed_user_highpage_movable(vma,
2254                                                               vmf->address);
2255                 if (!new_page)
2256                         goto oom;
2257         } else {
2258                 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2259                                 vmf->address);
2260                 if (!new_page)
2261                         goto oom;
2262                 cow_user_page(new_page, old_page, vmf->address, vma);
2263         }
2264
2265         if (mem_cgroup_try_charge_delay(new_page, mm, GFP_KERNEL, &memcg, false))
2266                 goto oom_free_new;
2267
2268         __SetPageUptodate(new_page);
2269
2270         mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2271
2272         /*
2273          * Re-check the pte - we dropped the lock
2274          */
2275         vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2276         if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2277                 if (old_page) {
2278                         if (!PageAnon(old_page)) {
2279                                 dec_mm_counter_fast(mm,
2280                                                 mm_counter_file(old_page));
2281                                 inc_mm_counter_fast(mm, MM_ANONPAGES);
2282                         }
2283                 } else {
2284                         inc_mm_counter_fast(mm, MM_ANONPAGES);
2285                 }
2286                 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2287                 entry = mk_pte(new_page, vma->vm_page_prot);
2288                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2289                 /*
2290                  * Clear the pte entry and flush it first, before updating the
2291                  * pte with the new entry. This will avoid a race condition
2292                  * seen in the presence of one thread doing SMC and another
2293                  * thread doing COW.
2294                  */
2295                 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2296                 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2297                 mem_cgroup_commit_charge(new_page, memcg, false, false);
2298                 lru_cache_add_active_or_unevictable(new_page, vma);
2299                 /*
2300                  * We call the notify macro here because, when using secondary
2301                  * mmu page tables (such as kvm shadow page tables), we want the
2302                  * new page to be mapped directly into the secondary page table.
2303                  */
2304                 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2305                 update_mmu_cache(vma, vmf->address, vmf->pte);
2306                 if (old_page) {
2307                         /*
2308                          * Only after switching the pte to the new page may
2309                          * we remove the mapcount here. Otherwise another
2310                          * process may come and find the rmap count decremented
2311                          * before the pte is switched to the new page, and
2312                          * "reuse" the old page writing into it while our pte
2313                          * here still points into it and can be read by other
2314                          * threads.
2315                          *
2316                          * The critical issue is to order this
2317                          * page_remove_rmap with the ptp_clear_flush above.
2318                          * Those stores are ordered by (if nothing else,)
2319                          * the barrier present in the atomic_add_negative
2320                          * in page_remove_rmap.
2321                          *
2322                          * Then the TLB flush in ptep_clear_flush ensures that
2323                          * no process can access the old page before the
2324                          * decremented mapcount is visible. And the old page
2325                          * cannot be reused until after the decremented
2326                          * mapcount is visible. So transitively, TLBs to
2327                          * old page will be flushed before it can be reused.
2328                          */
2329                         page_remove_rmap(old_page, false);
2330                 }
2331
2332                 /* Free the old page.. */
2333                 new_page = old_page;
2334                 page_copied = 1;
2335         } else {
2336                 mem_cgroup_cancel_charge(new_page, memcg, false);
2337         }
2338
2339         if (new_page)
2340                 put_page(new_page);
2341
2342         pte_unmap_unlock(vmf->pte, vmf->ptl);
2343         /*
2344          * No need to double call mmu_notifier->invalidate_range() callback as
2345          * the above ptep_clear_flush_notify() did already call it.
2346          */
2347         mmu_notifier_invalidate_range_only_end(mm, mmun_start, mmun_end);
2348         if (old_page) {
2349                 /*
2350                  * Don't let another task, with possibly unlocked vma,
2351                  * keep the mlocked page.
2352                  */
2353                 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2354                         lock_page(old_page);    /* LRU manipulation */
2355                         if (PageMlocked(old_page))
2356                                 munlock_vma_page(old_page);
2357                         unlock_page(old_page);
2358                 }
2359                 put_page(old_page);
2360         }
2361         return page_copied ? VM_FAULT_WRITE : 0;
2362 oom_free_new:
2363         put_page(new_page);
2364 oom:
2365         if (old_page)
2366                 put_page(old_page);
2367         return VM_FAULT_OOM;
2368 }
2369
2370 /**
2371  * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2372  *                        writeable once the page is prepared
2373  *
2374  * @vmf: structure describing the fault
2375  *
2376  * This function handles all that is needed to finish a write page fault in a
2377  * shared mapping due to PTE being read-only once the mapped page is prepared.
2378  * It handles locking of PTE and modifying it. The function returns
2379  * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2380  * lock.
2381  *
2382  * The function expects the page to be locked or other protection against
2383  * concurrent faults / writeback (such as DAX radix tree locks).
2384  */
2385 vm_fault_t finish_mkwrite_fault(struct vm_fault *vmf)
2386 {
2387         WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2388         vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2389                                        &vmf->ptl);
2390         /*
2391          * We might have raced with another page fault while we released the
2392          * pte_offset_map_lock.
2393          */
2394         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2395                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2396                 return VM_FAULT_NOPAGE;
2397         }
2398         wp_page_reuse(vmf);
2399         return 0;
2400 }
2401
2402 /*
2403  * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2404  * mapping
2405  */
2406 static vm_fault_t wp_pfn_shared(struct vm_fault *vmf)
2407 {
2408         struct vm_area_struct *vma = vmf->vma;
2409
2410         if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2411                 vm_fault_t ret;
2412
2413                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2414                 vmf->flags |= FAULT_FLAG_MKWRITE;
2415                 ret = vma->vm_ops->pfn_mkwrite(vmf);
2416                 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2417                         return ret;
2418                 return finish_mkwrite_fault(vmf);
2419         }
2420         wp_page_reuse(vmf);
2421         return VM_FAULT_WRITE;
2422 }
2423
2424 static vm_fault_t wp_page_shared(struct vm_fault *vmf)
2425         __releases(vmf->ptl)
2426 {
2427         struct vm_area_struct *vma = vmf->vma;
2428
2429         get_page(vmf->page);
2430
2431         if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2432                 vm_fault_t tmp;
2433
2434                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2435                 tmp = do_page_mkwrite(vmf);
2436                 if (unlikely(!tmp || (tmp &
2437                                       (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2438                         put_page(vmf->page);
2439                         return tmp;
2440                 }
2441                 tmp = finish_mkwrite_fault(vmf);
2442                 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2443                         unlock_page(vmf->page);
2444                         put_page(vmf->page);
2445                         return tmp;
2446                 }
2447         } else {
2448                 wp_page_reuse(vmf);
2449                 lock_page(vmf->page);
2450         }
2451         fault_dirty_shared_page(vma, vmf->page);
2452         put_page(vmf->page);
2453
2454         return VM_FAULT_WRITE;
2455 }
2456
2457 /*
2458  * This routine handles present pages, when users try to write
2459  * to a shared page. It is done by copying the page to a new address
2460  * and decrementing the shared-page counter for the old page.
2461  *
2462  * Note that this routine assumes that the protection checks have been
2463  * done by the caller (the low-level page fault routine in most cases).
2464  * Thus we can safely just mark it writable once we've done any necessary
2465  * COW.
2466  *
2467  * We also mark the page dirty at this point even though the page will
2468  * change only once the write actually happens. This avoids a few races,
2469  * and potentially makes it more efficient.
2470  *
2471  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2472  * but allow concurrent faults), with pte both mapped and locked.
2473  * We return with mmap_sem still held, but pte unmapped and unlocked.
2474  */
2475 static vm_fault_t do_wp_page(struct vm_fault *vmf)
2476         __releases(vmf->ptl)
2477 {
2478         struct vm_area_struct *vma = vmf->vma;
2479
2480         vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2481         if (!vmf->page) {
2482                 /*
2483                  * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2484                  * VM_PFNMAP VMA.
2485                  *
2486                  * We should not cow pages in a shared writeable mapping.
2487                  * Just mark the pages writable and/or call ops->pfn_mkwrite.
2488                  */
2489                 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2490                                      (VM_WRITE|VM_SHARED))
2491                         return wp_pfn_shared(vmf);
2492
2493                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2494                 return wp_page_copy(vmf);
2495         }
2496
2497         /*
2498          * Take out anonymous pages first, anonymous shared vmas are
2499          * not dirty accountable.
2500          */
2501         if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2502                 int total_map_swapcount;
2503                 if (!trylock_page(vmf->page)) {
2504                         get_page(vmf->page);
2505                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2506                         lock_page(vmf->page);
2507                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2508                                         vmf->address, &vmf->ptl);
2509                         if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2510                                 unlock_page(vmf->page);
2511                                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2512                                 put_page(vmf->page);
2513                                 return 0;
2514                         }
2515                         put_page(vmf->page);
2516                 }
2517                 if (reuse_swap_page(vmf->page, &total_map_swapcount)) {
2518                         if (total_map_swapcount == 1) {
2519                                 /*
2520                                  * The page is all ours. Move it to
2521                                  * our anon_vma so the rmap code will
2522                                  * not search our parent or siblings.
2523                                  * Protected against the rmap code by
2524                                  * the page lock.
2525                                  */
2526                                 page_move_anon_rmap(vmf->page, vma);
2527                         }
2528                         unlock_page(vmf->page);
2529                         wp_page_reuse(vmf);
2530                         return VM_FAULT_WRITE;
2531                 }
2532                 unlock_page(vmf->page);
2533         } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2534                                         (VM_WRITE|VM_SHARED))) {
2535                 return wp_page_shared(vmf);
2536         }
2537
2538         /*
2539          * Ok, we need to copy. Oh, well..
2540          */
2541         get_page(vmf->page);
2542
2543         pte_unmap_unlock(vmf->pte, vmf->ptl);
2544         return wp_page_copy(vmf);
2545 }
2546
2547 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2548                 unsigned long start_addr, unsigned long end_addr,
2549                 struct zap_details *details)
2550 {
2551         zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2552 }
2553
2554 static inline void unmap_mapping_range_tree(struct rb_root_cached *root,
2555                                             struct zap_details *details)
2556 {
2557         struct vm_area_struct *vma;
2558         pgoff_t vba, vea, zba, zea;
2559
2560         vma_interval_tree_foreach(vma, root,
2561                         details->first_index, details->last_index) {
2562
2563                 vba = vma->vm_pgoff;
2564                 vea = vba + vma_pages(vma) - 1;
2565                 zba = details->first_index;
2566                 if (zba < vba)
2567                         zba = vba;
2568                 zea = details->last_index;
2569                 if (zea > vea)
2570                         zea = vea;
2571
2572                 unmap_mapping_range_vma(vma,
2573                         ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2574                         ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2575                                 details);
2576         }
2577 }
2578
2579 /**
2580  * unmap_mapping_pages() - Unmap pages from processes.
2581  * @mapping: The address space containing pages to be unmapped.
2582  * @start: Index of first page to be unmapped.
2583  * @nr: Number of pages to be unmapped.  0 to unmap to end of file.
2584  * @even_cows: Whether to unmap even private COWed pages.
2585  *
2586  * Unmap the pages in this address space from any userspace process which
2587  * has them mmaped.  Generally, you want to remove COWed pages as well when
2588  * a file is being truncated, but not when invalidating pages from the page
2589  * cache.
2590  */
2591 void unmap_mapping_pages(struct address_space *mapping, pgoff_t start,
2592                 pgoff_t nr, bool even_cows)
2593 {
2594         struct zap_details details = { };
2595
2596         details.check_mapping = even_cows ? NULL : mapping;
2597         details.first_index = start;
2598         details.last_index = start + nr - 1;
2599         if (details.last_index < details.first_index)
2600                 details.last_index = ULONG_MAX;
2601
2602         i_mmap_lock_write(mapping);
2603         if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap.rb_root)))
2604                 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2605         i_mmap_unlock_write(mapping);
2606 }
2607
2608 /**
2609  * unmap_mapping_range - unmap the portion of all mmaps in the specified
2610  * address_space corresponding to the specified byte range in the underlying
2611  * file.
2612  *
2613  * @mapping: the address space containing mmaps to be unmapped.
2614  * @holebegin: byte in first page to unmap, relative to the start of
2615  * the underlying file.  This will be rounded down to a PAGE_SIZE
2616  * boundary.  Note that this is different from truncate_pagecache(), which
2617  * must keep the partial page.  In contrast, we must get rid of
2618  * partial pages.
2619  * @holelen: size of prospective hole in bytes.  This will be rounded
2620  * up to a PAGE_SIZE boundary.  A holelen of zero truncates to the
2621  * end of the file.
2622  * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2623  * but 0 when invalidating pagecache, don't throw away private data.
2624  */
2625 void unmap_mapping_range(struct address_space *mapping,
2626                 loff_t const holebegin, loff_t const holelen, int even_cows)
2627 {
2628         pgoff_t hba = holebegin >> PAGE_SHIFT;
2629         pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2630
2631         /* Check for overflow. */
2632         if (sizeof(holelen) > sizeof(hlen)) {
2633                 long long holeend =
2634                         (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2635                 if (holeend & ~(long long)ULONG_MAX)
2636                         hlen = ULONG_MAX - hba + 1;
2637         }
2638
2639         unmap_mapping_pages(mapping, hba, hlen, even_cows);
2640 }
2641 EXPORT_SYMBOL(unmap_mapping_range);
2642
2643 /*
2644  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2645  * but allow concurrent faults), and pte mapped but not yet locked.
2646  * We return with pte unmapped and unlocked.
2647  *
2648  * We return with the mmap_sem locked or unlocked in the same cases
2649  * as does filemap_fault().
2650  */
2651 vm_fault_t do_swap_page(struct vm_fault *vmf)
2652 {
2653         struct vm_area_struct *vma = vmf->vma;
2654         struct page *page = NULL, *swapcache;
2655         struct mem_cgroup *memcg;
2656         swp_entry_t entry;
2657         pte_t pte;
2658         int locked;
2659         int exclusive = 0;
2660         vm_fault_t ret = 0;
2661
2662         if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2663                 goto out;
2664
2665         entry = pte_to_swp_entry(vmf->orig_pte);
2666         if (unlikely(non_swap_entry(entry))) {
2667                 if (is_migration_entry(entry)) {
2668                         migration_entry_wait(vma->vm_mm, vmf->pmd,
2669                                              vmf->address);
2670                 } else if (is_device_private_entry(entry)) {
2671                         /*
2672                          * For un-addressable device memory we call the pgmap
2673                          * fault handler callback. The callback must migrate
2674                          * the page back to some CPU accessible page.
2675                          */
2676                         ret = device_private_entry_fault(vma, vmf->address, entry,
2677                                                  vmf->flags, vmf->pmd);
2678                 } else if (is_hwpoison_entry(entry)) {
2679                         ret = VM_FAULT_HWPOISON;
2680                 } else {
2681                         print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2682                         ret = VM_FAULT_SIGBUS;
2683                 }
2684                 goto out;
2685         }
2686
2687
2688         delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2689         page = lookup_swap_cache(entry, vma, vmf->address);
2690         swapcache = page;
2691
2692         if (!page) {
2693                 struct swap_info_struct *si = swp_swap_info(entry);
2694
2695                 if (si->flags & SWP_SYNCHRONOUS_IO &&
2696                                 __swap_count(si, entry) == 1) {
2697                         /* skip swapcache */
2698                         page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2699                                                         vmf->address);
2700                         if (page) {
2701                                 __SetPageLocked(page);
2702                                 __SetPageSwapBacked(page);
2703                                 set_page_private(page, entry.val);
2704                                 lru_cache_add_anon(page);
2705                                 swap_readpage(page, true);
2706                         }
2707                 } else {
2708                         page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE,
2709                                                 vmf);
2710                         swapcache = page;
2711                 }
2712
2713                 if (!page) {
2714                         /*
2715                          * Back out if somebody else faulted in this pte
2716                          * while we released the pte lock.
2717                          */
2718                         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2719                                         vmf->address, &vmf->ptl);
2720                         if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2721                                 ret = VM_FAULT_OOM;
2722                         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2723                         goto unlock;
2724                 }
2725
2726                 /* Had to read the page from swap area: Major fault */
2727                 ret = VM_FAULT_MAJOR;
2728                 count_vm_event(PGMAJFAULT);
2729                 count_memcg_event_mm(vma->vm_mm, PGMAJFAULT);
2730         } else if (PageHWPoison(page)) {
2731                 /*
2732                  * hwpoisoned dirty swapcache pages are kept for killing
2733                  * owner processes (which may be unknown at hwpoison time)
2734                  */
2735                 ret = VM_FAULT_HWPOISON;
2736                 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2737                 goto out_release;
2738         }
2739
2740         locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2741
2742         delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2743         if (!locked) {
2744                 ret |= VM_FAULT_RETRY;
2745                 goto out_release;
2746         }
2747
2748         /*
2749          * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2750          * release the swapcache from under us.  The page pin, and pte_same
2751          * test below, are not enough to exclude that.  Even if it is still
2752          * swapcache, we need to check that the page's swap has not changed.
2753          */
2754         if (unlikely((!PageSwapCache(page) ||
2755                         page_private(page) != entry.val)) && swapcache)
2756                 goto out_page;
2757
2758         page = ksm_might_need_to_copy(page, vma, vmf->address);
2759         if (unlikely(!page)) {
2760                 ret = VM_FAULT_OOM;
2761                 page = swapcache;
2762                 goto out_page;
2763         }
2764
2765         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL,
2766                                         &memcg, false)) {
2767                 ret = VM_FAULT_OOM;
2768                 goto out_page;
2769         }
2770
2771         /*
2772          * Back out if somebody else already faulted in this pte.
2773          */
2774         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2775                         &vmf->ptl);
2776         if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2777                 goto out_nomap;
2778
2779         if (unlikely(!PageUptodate(page))) {
2780                 ret = VM_FAULT_SIGBUS;
2781                 goto out_nomap;
2782         }
2783
2784         /*
2785          * The page isn't present yet, go ahead with the fault.
2786          *
2787          * Be careful about the sequence of operations here.
2788          * To get its accounting right, reuse_swap_page() must be called
2789          * while the page is counted on swap but not yet in mapcount i.e.
2790          * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2791          * must be called after the swap_free(), or it will never succeed.
2792          */
2793
2794         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2795         dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2796         pte = mk_pte(page, vma->vm_page_prot);
2797         if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2798                 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2799                 vmf->flags &= ~FAULT_FLAG_WRITE;
2800                 ret |= VM_FAULT_WRITE;
2801                 exclusive = RMAP_EXCLUSIVE;
2802         }
2803         flush_icache_page(vma, page);
2804         if (pte_swp_soft_dirty(vmf->orig_pte))
2805                 pte = pte_mksoft_dirty(pte);
2806         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2807         arch_do_swap_page(vma->vm_mm, vma, vmf->address, pte, vmf->orig_pte);
2808         vmf->orig_pte = pte;
2809
2810         /* ksm created a completely new copy */
2811         if (unlikely(page != swapcache && swapcache)) {
2812                 page_add_new_anon_rmap(page, vma, vmf->address, false);
2813                 mem_cgroup_commit_charge(page, memcg, false, false);
2814                 lru_cache_add_active_or_unevictable(page, vma);
2815         } else {
2816                 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2817                 mem_cgroup_commit_charge(page, memcg, true, false);
2818                 activate_page(page);
2819         }
2820
2821         swap_free(entry);
2822         if (mem_cgroup_swap_full(page) ||
2823             (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2824                 try_to_free_swap(page);
2825         unlock_page(page);
2826         if (page != swapcache && swapcache) {
2827                 /*
2828                  * Hold the lock to avoid the swap entry to be reused
2829                  * until we take the PT lock for the pte_same() check
2830                  * (to avoid false positives from pte_same). For
2831                  * further safety release the lock after the swap_free
2832                  * so that the swap count won't change under a
2833                  * parallel locked swapcache.
2834                  */
2835                 unlock_page(swapcache);
2836                 put_page(swapcache);
2837         }
2838
2839         if (vmf->flags & FAULT_FLAG_WRITE) {
2840                 ret |= do_wp_page(vmf);
2841                 if (ret & VM_FAULT_ERROR)
2842                         ret &= VM_FAULT_ERROR;
2843                 goto out;
2844         }
2845
2846         /* No need to invalidate - it was non-present before */
2847         update_mmu_cache(vma, vmf->address, vmf->pte);
2848 unlock:
2849         pte_unmap_unlock(vmf->pte, vmf->ptl);
2850 out:
2851         return ret;
2852 out_nomap:
2853         mem_cgroup_cancel_charge(page, memcg, false);
2854         pte_unmap_unlock(vmf->pte, vmf->ptl);
2855 out_page:
2856         unlock_page(page);
2857 out_release:
2858         put_page(page);
2859         if (page != swapcache && swapcache) {
2860                 unlock_page(swapcache);
2861                 put_page(swapcache);
2862         }
2863         return ret;
2864 }
2865
2866 /*
2867  * We enter with non-exclusive mmap_sem (to exclude vma changes,
2868  * but allow concurrent faults), and pte mapped but not yet locked.
2869  * We return with mmap_sem still held, but pte unmapped and unlocked.
2870  */
2871 static vm_fault_t do_anonymous_page(struct vm_fault *vmf)
2872 {
2873         struct vm_area_struct *vma = vmf->vma;
2874         struct mem_cgroup *memcg;
2875         struct page *page;
2876         vm_fault_t ret = 0;
2877         pte_t entry;
2878
2879         /* File mapping without ->vm_ops ? */
2880         if (vma->vm_flags & VM_SHARED)
2881                 return VM_FAULT_SIGBUS;
2882
2883         /*
2884          * Use pte_alloc() instead of pte_alloc_map().  We can't run
2885          * pte_offset_map() on pmds where a huge pmd might be created
2886          * from a different thread.
2887          *
2888          * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2889          * parallel threads are excluded by other means.
2890          *
2891          * Here we only have down_read(mmap_sem).
2892          */
2893         if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2894                 return VM_FAULT_OOM;
2895
2896         /* See the comment in pte_alloc_one_map() */
2897         if (unlikely(pmd_trans_unstable(vmf->pmd)))
2898                 return 0;
2899
2900         /* Use the zero-page for reads */
2901         if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2902                         !mm_forbids_zeropage(vma->vm_mm)) {
2903                 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2904                                                 vma->vm_page_prot));
2905                 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2906                                 vmf->address, &vmf->ptl);
2907                 if (!pte_none(*vmf->pte))
2908                         goto unlock;
2909                 ret = check_stable_address_space(vma->vm_mm);
2910                 if (ret)
2911                         goto unlock;
2912                 /* Deliver the page fault to userland, check inside PT lock */
2913                 if (userfaultfd_missing(vma)) {
2914                         pte_unmap_unlock(vmf->pte, vmf->ptl);
2915                         return handle_userfault(vmf, VM_UFFD_MISSING);
2916                 }
2917                 goto setpte;
2918         }
2919
2920         /* Allocate our own private page. */
2921         if (unlikely(anon_vma_prepare(vma)))
2922                 goto oom;
2923         page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2924         if (!page)
2925                 goto oom;
2926
2927         if (mem_cgroup_try_charge_delay(page, vma->vm_mm, GFP_KERNEL, &memcg,
2928                                         false))
2929                 goto oom_free_page;
2930
2931         /*
2932          * The memory barrier inside __SetPageUptodate makes sure that
2933          * preceeding stores to the page contents become visible before
2934          * the set_pte_at() write.
2935          */
2936         __SetPageUptodate(page);
2937
2938         entry = mk_pte(page, vma->vm_page_prot);
2939         if (vma->vm_flags & VM_WRITE)
2940                 entry = pte_mkwrite(pte_mkdirty(entry));
2941
2942         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2943                         &vmf->ptl);
2944         if (!pte_none(*vmf->pte))
2945                 goto release;
2946
2947         ret = check_stable_address_space(vma->vm_mm);
2948         if (ret)
2949                 goto release;
2950
2951         /* Deliver the page fault to userland, check inside PT lock */
2952         if (userfaultfd_missing(vma)) {
2953                 pte_unmap_unlock(vmf->pte, vmf->ptl);
2954                 mem_cgroup_cancel_charge(page, memcg, false);
2955                 put_page(page);
2956                 return handle_userfault(vmf, VM_UFFD_MISSING);
2957         }
2958
2959         inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2960         page_add_new_anon_rmap(page, vma, vmf->address, false);
2961         mem_cgroup_commit_charge(page, memcg, false, false);
2962         lru_cache_add_active_or_unevictable(page, vma);
2963 setpte:
2964         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2965
2966         /* No need to invalidate - it was non-present before */
2967         update_mmu_cache(vma, vmf->address, vmf->pte);
2968 unlock:
2969         pte_unmap_unlock(vmf->pte, vmf->ptl);
2970         return ret;
2971 release:
2972         mem_cgroup_cancel_charge(page, memcg, false);
2973         put_page(page);
2974         goto unlock;
2975 oom_free_page:
2976         put_page(page);
2977 oom:
2978         return VM_FAULT_OOM;
2979 }
2980
2981 /*
2982  * The mmap_sem must have been held on entry, and may have been
2983  * released depending on flags and vma->vm_ops->fault() return value.
2984  * See filemap_fault() and __lock_page_retry().
2985  */
2986 static vm_fault_t __do_fault(struct vm_fault *vmf)
2987 {
2988         struct vm_area_struct *vma = vmf->vma;
2989         vm_fault_t ret;
2990
2991         ret = vma->vm_ops->fault(vmf);
2992         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2993                             VM_FAULT_DONE_COW)))
2994                 return ret;
2995
2996         if (unlikely(PageHWPoison(vmf->page))) {
2997                 if (ret & VM_FAULT_LOCKED)
2998                         unlock_page(vmf->page);
2999                 put_page(vmf->page);
3000                 vmf->page = NULL;
3001                 return VM_FAULT_HWPOISON;
3002         }
3003
3004         if (unlikely(!(ret & VM_FAULT_LOCKED)))
3005                 lock_page(vmf->page);
3006         else
3007                 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
3008
3009         return ret;
3010 }
3011
3012 /*
3013  * The ordering of these checks is important for pmds with _PAGE_DEVMAP set.
3014  * If we check pmd_trans_unstable() first we will trip the bad_pmd() check
3015  * inside of pmd_none_or_trans_huge_or_clear_bad(). This will end up correctly
3016  * returning 1 but not before it spams dmesg with the pmd_clear_bad() output.
3017  */
3018 static int pmd_devmap_trans_unstable(pmd_t *pmd)
3019 {
3020         return pmd_devmap(*pmd) || pmd_trans_unstable(pmd);
3021 }
3022
3023 static vm_fault_t pte_alloc_one_map(struct vm_fault *vmf)
3024 {
3025         struct vm_area_struct *vma = vmf->vma;
3026
3027         if (!pmd_none(*vmf->pmd))
3028                 goto map_pte;
3029         if (vmf->prealloc_pte) {
3030                 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3031                 if (unlikely(!pmd_none(*vmf->pmd))) {
3032                         spin_unlock(vmf->ptl);
3033                         goto map_pte;
3034                 }
3035
3036                 mm_inc_nr_ptes(vma->vm_mm);
3037                 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3038                 spin_unlock(vmf->ptl);
3039                 vmf->prealloc_pte = NULL;
3040         } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
3041                 return VM_FAULT_OOM;
3042         }
3043 map_pte:
3044         /*
3045          * If a huge pmd materialized under us just retry later.  Use
3046          * pmd_trans_unstable() via pmd_devmap_trans_unstable() instead of
3047          * pmd_trans_huge() to ensure the pmd didn't become pmd_trans_huge
3048          * under us and then back to pmd_none, as a result of MADV_DONTNEED
3049          * running immediately after a huge pmd fault in a different thread of
3050          * this mm, in turn leading to a misleading pmd_trans_huge() retval.
3051          * All we have to ensure is that it is a regular pmd that we can walk
3052          * with pte_offset_map() and we can do that through an atomic read in
3053          * C, which is what pmd_trans_unstable() provides.
3054          */
3055         if (pmd_devmap_trans_unstable(vmf->pmd))
3056                 return VM_FAULT_NOPAGE;
3057
3058         /*
3059          * At this point we know that our vmf->pmd points to a page of ptes
3060          * and it cannot become pmd_none(), pmd_devmap() or pmd_trans_huge()
3061          * for the duration of the fault.  If a racing MADV_DONTNEED runs and
3062          * we zap the ptes pointed to by our vmf->pmd, the vmf->ptl will still
3063          * be valid and we will re-check to make sure the vmf->pte isn't
3064          * pte_none() under vmf->ptl protection when we return to
3065          * alloc_set_pte().
3066          */
3067         vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
3068                         &vmf->ptl);
3069         return 0;
3070 }
3071
3072 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
3073
3074 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
3075 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
3076                 unsigned long haddr)
3077 {
3078         if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
3079                         (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
3080                 return false;
3081         if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
3082                 return false;
3083         return true;
3084 }
3085
3086 static void deposit_prealloc_pte(struct vm_fault *vmf)
3087 {
3088         struct vm_area_struct *vma = vmf->vma;
3089
3090         pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
3091         /*
3092          * We are going to consume the prealloc table,
3093          * count that as nr_ptes.
3094          */
3095         mm_inc_nr_ptes(vma->vm_mm);
3096         vmf->prealloc_pte = NULL;
3097 }
3098
3099 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3100 {
3101         struct vm_area_struct *vma = vmf->vma;
3102         bool write = vmf->flags & FAULT_FLAG_WRITE;
3103         unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
3104         pmd_t entry;
3105         int i;
3106         vm_fault_t ret;
3107
3108         if (!transhuge_vma_suitable(vma, haddr))
3109                 return VM_FAULT_FALLBACK;
3110
3111         ret = VM_FAULT_FALLBACK;
3112         page = compound_head(page);
3113
3114         /*
3115          * Archs like ppc64 need additonal space to store information
3116          * related to pte entry. Use the preallocated table for that.
3117          */
3118         if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
3119                 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
3120                 if (!vmf->prealloc_pte)
3121                         return VM_FAULT_OOM;
3122                 smp_wmb(); /* See comment in __pte_alloc() */
3123         }
3124
3125         vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3126         if (unlikely(!pmd_none(*vmf->pmd)))
3127                 goto out;
3128
3129         for (i = 0; i < HPAGE_PMD_NR; i++)
3130                 flush_icache_page(vma, page + i);
3131
3132         entry = mk_huge_pmd(page, vma->vm_page_prot);
3133         if (write)
3134                 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3135
3136         add_mm_counter(vma->vm_mm, mm_counter_file(page), HPAGE_PMD_NR);
3137         page_add_file_rmap(page, true);
3138         /*
3139          * deposit and withdraw with pmd lock held
3140          */
3141         if (arch_needs_pgtable_deposit())
3142                 deposit_prealloc_pte(vmf);
3143
3144         set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3145
3146         update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3147
3148         /* fault is handled */
3149         ret = 0;
3150         count_vm_event(THP_FILE_MAPPED);
3151 out:
3152         spin_unlock(vmf->ptl);
3153         return ret;
3154 }
3155 #else
3156 static vm_fault_t do_set_pmd(struct vm_fault *vmf, struct page *page)
3157 {
3158         BUILD_BUG();
3159         return 0;
3160 }
3161 #endif
3162
3163 /**
3164  * alloc_set_pte - setup new PTE entry for given page and add reverse page
3165  * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3166  *
3167  * @vmf: fault environment
3168  * @memcg: memcg to charge page (only for private mappings)
3169  * @page: page to map
3170  *
3171  * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3172  * return.
3173  *
3174  * Target users are page handler itself and implementations of
3175  * vm_ops->map_pages.
3176  */
3177 vm_fault_t alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3178                 struct page *page)
3179 {
3180         struct vm_area_struct *vma = vmf->vma;
3181         bool write = vmf->flags & FAULT_FLAG_WRITE;
3182         pte_t entry;
3183         vm_fault_t ret;
3184
3185         if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3186                         IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3187                 /* THP on COW? */
3188                 VM_BUG_ON_PAGE(memcg, page);
3189
3190                 ret = do_set_pmd(vmf, page);
3191                 if (ret != VM_FAULT_FALLBACK)
3192                         return ret;
3193         }
3194
3195         if (!vmf->pte) {
3196                 ret = pte_alloc_one_map(vmf);
3197                 if (ret)
3198                         return ret;
3199         }
3200
3201         /* Re-check under ptl */
3202         if (unlikely(!pte_none(*vmf->pte)))
3203                 return VM_FAULT_NOPAGE;
3204
3205         flush_icache_page(vma, page);
3206         entry = mk_pte(page, vma->vm_page_prot);
3207         if (write)
3208                 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3209         /* copy-on-write page */
3210         if (write && !(vma->vm_flags & VM_SHARED)) {
3211                 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3212                 page_add_new_anon_rmap(page, vma, vmf->address, false);
3213                 mem_cgroup_commit_charge(page, memcg, false, false);
3214                 lru_cache_add_active_or_unevictable(page, vma);
3215         } else {
3216                 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3217                 page_add_file_rmap(page, false);
3218         }
3219         set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3220
3221         /* no need to invalidate: a not-present page won't be cached */
3222         update_mmu_cache(vma, vmf->address, vmf->pte);
3223
3224         return 0;
3225 }
3226
3227
3228 /**
3229  * finish_fault - finish page fault once we have prepared the page to fault
3230  *
3231  * @vmf: structure describing the fault
3232  *
3233  * This function handles all that is needed to finish a page fault once the
3234  * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3235  * given page, adds reverse page mapping, handles memcg charges and LRU
3236  * addition. The function returns 0 on success, VM_FAULT_ code in case of
3237  * error.
3238  *
3239  * The function expects the page to be locked and on success it consumes a
3240  * reference of a page being mapped (for the PTE which maps it).
3241  */
3242 vm_fault_t finish_fault(struct vm_fault *vmf)
3243 {
3244         struct page *page;
3245         vm_fault_t ret = 0;
3246
3247         /* Did we COW the page? */
3248         if ((vmf->flags & FAULT_FLAG_WRITE) &&
3249             !(vmf->vma->vm_flags & VM_SHARED))
3250                 page = vmf->cow_page;
3251         else
3252                 page = vmf->page;
3253
3254         /*
3255          * check even for read faults because we might have lost our CoWed
3256          * page
3257          */
3258         if (!(vmf->vma->vm_flags & VM_SHARED))
3259                 ret = check_stable_address_space(vmf->vma->vm_mm);
3260         if (!ret)
3261                 ret = alloc_set_pte(vmf, vmf->memcg, page);
3262         if (vmf->pte)
3263                 pte_unmap_unlock(vmf->pte, vmf->ptl);
3264         return ret;
3265 }
3266
3267 static unsigned long fault_around_bytes __read_mostly =
3268         rounddown_pow_of_two(65536);
3269
3270 #ifdef CONFIG_DEBUG_FS
3271 static int fault_around_bytes_get(void *data, u64 *val)
3272 {
3273         *val = fault_around_bytes;
3274         return 0;
3275 }
3276
3277 /*
3278  * fault_around_bytes must be rounded down to the nearest page order as it's
3279  * what do_fault_around() expects to see.
3280  */
3281 static int fault_around_bytes_set(void *data, u64 val)
3282 {
3283         if (val / PAGE_SIZE > PTRS_PER_PTE)
3284                 return -EINVAL;
3285         if (val > PAGE_SIZE)
3286                 fault_around_bytes = rounddown_pow_of_two(val);
3287         else
3288                 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3289         return 0;
3290 }
3291 DEFINE_DEBUGFS_ATTRIBUTE(fault_around_bytes_fops,
3292                 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3293
3294 static int __init fault_around_debugfs(void)
3295 {
3296         void *ret;
3297
3298         ret = debugfs_create_file_unsafe("fault_around_bytes", 0644, NULL, NULL,
3299                         &fault_around_bytes_fops);
3300         if (!ret)
3301                 pr_warn("Failed to create fault_around_bytes in debugfs");
3302         return 0;
3303 }
3304 late_initcall(fault_around_debugfs);
3305 #endif
3306
3307 /*
3308  * do_fault_around() tries to map few pages around the fault address. The hope
3309  * is that the pages will be needed soon and this will lower the number of
3310  * faults to handle.
3311  *
3312  * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3313  * not ready to be mapped: not up-to-date, locked, etc.
3314  *
3315  * This function is called with the page table lock taken. In the split ptlock
3316  * case the page table lock only protects only those entries which belong to
3317  * the page table corresponding to the fault address.
3318  *
3319  * This function doesn't cross the VMA boundaries, in order to call map_pages()
3320  * only once.
3321  *
3322  * fault_around_bytes defines how many bytes we'll try to map.
3323  * do_fault_around() expects it to be set to a power of two less than or equal
3324  * to PTRS_PER_PTE.
3325  *
3326  * The virtual address of the area that we map is naturally aligned to
3327  * fault_around_bytes rounded down to the machine page size
3328  * (and therefore to page order).  This way it's easier to guarantee
3329  * that we don't cross page table boundaries.
3330  */
3331 static vm_fault_t do_fault_around(struct vm_fault *vmf)
3332 {
3333         unsigned long address = vmf->address, nr_pages, mask;
3334         pgoff_t start_pgoff = vmf->pgoff;
3335         pgoff_t end_pgoff;
3336         int off;
3337         vm_fault_t ret = 0;
3338
3339         nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3340         mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3341
3342         vmf->address = max(address & mask, vmf->vma->vm_start);
3343         off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3344         start_pgoff -= off;
3345
3346         /*
3347          *  end_pgoff is either the end of the page table, the end of
3348          *  the vma or nr_pages from start_pgoff, depending what is nearest.
3349          */
3350         end_pgoff = start_pgoff -
3351                 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3352                 PTRS_PER_PTE - 1;
3353         end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3354                         start_pgoff + nr_pages - 1);
3355
3356         if (pmd_none(*vmf->pmd)) {
3357                 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3358                                                   vmf->address);
3359                 if (!vmf->prealloc_pte)
3360                         goto out;
3361                 smp_wmb(); /* See comment in __pte_alloc() */
3362         }
3363
3364         vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3365
3366         /* Huge page is mapped? Page fault is solved */
3367         if (pmd_trans_huge(*vmf->pmd)) {
3368                 ret = VM_FAULT_NOPAGE;
3369                 goto out;
3370         }
3371
3372         /* ->map_pages() haven't done anything useful. Cold page cache? */
3373         if (!vmf->pte)
3374                 goto out;
3375
3376         /* check if the page fault is solved */
3377         vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3378         if (!pte_none(*vmf->pte))
3379                 ret = VM_FAULT_NOPAGE;
3380         pte_unmap_unlock(vmf->pte, vmf->ptl);
3381 out:
3382         vmf->address = address;
3383         vmf->pte = NULL;
3384         return ret;
3385 }
3386
3387 static vm_fault_t do_read_fault(struct vm_fault *vmf)
3388 {
3389         struct vm_area_struct *vma = vmf->vma;
3390         vm_fault_t ret = 0;
3391
3392         /*
3393          * Let's call ->map_pages() first and use ->fault() as fallback
3394          * if page by the offset is not ready to be mapped (cold cache or
3395          * something).
3396          */
3397         if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3398                 ret = do_fault_around(vmf);
3399                 if (ret)
3400                         return ret;
3401         }
3402
3403         ret = __do_fault(vmf);
3404         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3405                 return ret;
3406
3407         ret |= finish_fault(vmf);
3408         unlock_page(vmf->page);
3409         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3410                 put_page(vmf->page);
3411         return ret;
3412 }
3413
3414 static vm_fault_t do_cow_fault(struct vm_fault *vmf)
3415 {
3416         struct vm_area_struct *vma = vmf->vma;
3417         vm_fault_t ret;
3418
3419         if (unlikely(anon_vma_prepare(vma)))
3420                 return VM_FAULT_OOM;
3421
3422         vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3423         if (!vmf->cow_page)
3424                 return VM_FAULT_OOM;
3425
3426         if (mem_cgroup_try_charge_delay(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3427                                 &vmf->memcg, false)) {
3428                 put_page(vmf->cow_page);
3429                 return VM_FAULT_OOM;
3430         }
3431
3432         ret = __do_fault(vmf);
3433         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3434                 goto uncharge_out;
3435         if (ret & VM_FAULT_DONE_COW)
3436                 return ret;
3437
3438         copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3439         __SetPageUptodate(vmf->cow_page);
3440
3441         ret |= finish_fault(vmf);
3442         unlock_page(vmf->page);
3443         put_page(vmf->page);
3444         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3445                 goto uncharge_out;
3446         return ret;
3447 uncharge_out:
3448         mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3449         put_page(vmf->cow_page);
3450         return ret;
3451 }
3452
3453 static vm_fault_t do_shared_fault(struct vm_fault *vmf)
3454 {
3455         struct vm_area_struct *vma = vmf->vma;
3456         vm_fault_t ret, tmp;
3457
3458         ret = __do_fault(vmf);
3459         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3460                 return ret;
3461
3462         /*
3463          * Check if the backing address space wants to know that the page is
3464          * about to become writable
3465          */
3466         if (vma->vm_ops->page_mkwrite) {
3467                 unlock_page(vmf->page);
3468                 tmp = do_page_mkwrite(vmf);
3469                 if (unlikely(!tmp ||
3470                                 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3471                         put_page(vmf->page);
3472                         return tmp;
3473                 }
3474         }
3475
3476         ret |= finish_fault(vmf);
3477         if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3478                            &n