]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/memory.c
Merge tag 'for-linus-4.11-rc1-tag' of git://git.kernel.org/pub/scm/linux/kernel/git...
[mirror_ubuntu-bionic-kernel.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/ksm.h>
53 #include <linux/rmap.h>
54 #include <linux/export.h>
55 #include <linux/delayacct.h>
56 #include <linux/init.h>
57 #include <linux/pfn_t.h>
58 #include <linux/writeback.h>
59 #include <linux/memcontrol.h>
60 #include <linux/mmu_notifier.h>
61 #include <linux/kallsyms.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
72 #include <asm/io.h>
73 #include <asm/mmu_context.h>
74 #include <asm/pgalloc.h>
75 #include <linux/uaccess.h>
76 #include <asm/tlb.h>
77 #include <asm/tlbflush.h>
78 #include <asm/pgtable.h>
79
80 #include "internal.h"
81
82 #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
83 #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid.
84 #endif
85
86 #ifndef CONFIG_NEED_MULTIPLE_NODES
87 /* use the per-pgdat data instead for discontigmem - mbligh */
88 unsigned long max_mapnr;
89 EXPORT_SYMBOL(max_mapnr);
90
91 struct page *mem_map;
92 EXPORT_SYMBOL(mem_map);
93 #endif
94
95 /*
96 * A number of key systems in x86 including ioremap() rely on the assumption
97 * that high_memory defines the upper bound on direct map memory, then end
98 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
99 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
100 * and ZONE_HIGHMEM.
101 */
102 void *high_memory;
103 EXPORT_SYMBOL(high_memory);
104
105 /*
106 * Randomize the address space (stacks, mmaps, brk, etc.).
107 *
108 * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization,
109 * as ancient (libc5 based) binaries can segfault. )
110 */
111 int randomize_va_space __read_mostly =
112 #ifdef CONFIG_COMPAT_BRK
113 1;
114 #else
115 2;
116 #endif
117
118 static int __init disable_randmaps(char *s)
119 {
120 randomize_va_space = 0;
121 return 1;
122 }
123 __setup("norandmaps", disable_randmaps);
124
125 unsigned long zero_pfn __read_mostly;
126 EXPORT_SYMBOL(zero_pfn);
127
128 unsigned long highest_memmap_pfn __read_mostly;
129
130 /*
131 * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init()
132 */
133 static int __init init_zero_pfn(void)
134 {
135 zero_pfn = page_to_pfn(ZERO_PAGE(0));
136 return 0;
137 }
138 core_initcall(init_zero_pfn);
139
140
141 #if defined(SPLIT_RSS_COUNTING)
142
143 void sync_mm_rss(struct mm_struct *mm)
144 {
145 int i;
146
147 for (i = 0; i < NR_MM_COUNTERS; i++) {
148 if (current->rss_stat.count[i]) {
149 add_mm_counter(mm, i, current->rss_stat.count[i]);
150 current->rss_stat.count[i] = 0;
151 }
152 }
153 current->rss_stat.events = 0;
154 }
155
156 static void add_mm_counter_fast(struct mm_struct *mm, int member, int val)
157 {
158 struct task_struct *task = current;
159
160 if (likely(task->mm == mm))
161 task->rss_stat.count[member] += val;
162 else
163 add_mm_counter(mm, member, val);
164 }
165 #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1)
166 #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1)
167
168 /* sync counter once per 64 page faults */
169 #define TASK_RSS_EVENTS_THRESH (64)
170 static void check_sync_rss_stat(struct task_struct *task)
171 {
172 if (unlikely(task != current))
173 return;
174 if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH))
175 sync_mm_rss(task->mm);
176 }
177 #else /* SPLIT_RSS_COUNTING */
178
179 #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member)
180 #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member)
181
182 static void check_sync_rss_stat(struct task_struct *task)
183 {
184 }
185
186 #endif /* SPLIT_RSS_COUNTING */
187
188 #ifdef HAVE_GENERIC_MMU_GATHER
189
190 static bool tlb_next_batch(struct mmu_gather *tlb)
191 {
192 struct mmu_gather_batch *batch;
193
194 batch = tlb->active;
195 if (batch->next) {
196 tlb->active = batch->next;
197 return true;
198 }
199
200 if (tlb->batch_count == MAX_GATHER_BATCH_COUNT)
201 return false;
202
203 batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0);
204 if (!batch)
205 return false;
206
207 tlb->batch_count++;
208 batch->next = NULL;
209 batch->nr = 0;
210 batch->max = MAX_GATHER_BATCH;
211
212 tlb->active->next = batch;
213 tlb->active = batch;
214
215 return true;
216 }
217
218 /* tlb_gather_mmu
219 * Called to initialize an (on-stack) mmu_gather structure for page-table
220 * tear-down from @mm. The @fullmm argument is used when @mm is without
221 * users and we're going to destroy the full address space (exit/execve).
222 */
223 void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end)
224 {
225 tlb->mm = mm;
226
227 /* Is it from 0 to ~0? */
228 tlb->fullmm = !(start | (end+1));
229 tlb->need_flush_all = 0;
230 tlb->local.next = NULL;
231 tlb->local.nr = 0;
232 tlb->local.max = ARRAY_SIZE(tlb->__pages);
233 tlb->active = &tlb->local;
234 tlb->batch_count = 0;
235
236 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
237 tlb->batch = NULL;
238 #endif
239 tlb->page_size = 0;
240
241 __tlb_reset_range(tlb);
242 }
243
244 static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
245 {
246 if (!tlb->end)
247 return;
248
249 tlb_flush(tlb);
250 mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end);
251 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
252 tlb_table_flush(tlb);
253 #endif
254 __tlb_reset_range(tlb);
255 }
256
257 static void tlb_flush_mmu_free(struct mmu_gather *tlb)
258 {
259 struct mmu_gather_batch *batch;
260
261 for (batch = &tlb->local; batch && batch->nr; batch = batch->next) {
262 free_pages_and_swap_cache(batch->pages, batch->nr);
263 batch->nr = 0;
264 }
265 tlb->active = &tlb->local;
266 }
267
268 void tlb_flush_mmu(struct mmu_gather *tlb)
269 {
270 tlb_flush_mmu_tlbonly(tlb);
271 tlb_flush_mmu_free(tlb);
272 }
273
274 /* tlb_finish_mmu
275 * Called at the end of the shootdown operation to free up any resources
276 * that were required.
277 */
278 void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end)
279 {
280 struct mmu_gather_batch *batch, *next;
281
282 tlb_flush_mmu(tlb);
283
284 /* keep the page table cache within bounds */
285 check_pgt_cache();
286
287 for (batch = tlb->local.next; batch; batch = next) {
288 next = batch->next;
289 free_pages((unsigned long)batch, 0);
290 }
291 tlb->local.next = NULL;
292 }
293
294 /* __tlb_remove_page
295 * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while
296 * handling the additional races in SMP caused by other CPUs caching valid
297 * mappings in their TLBs. Returns the number of free page slots left.
298 * When out of page slots we must call tlb_flush_mmu().
299 *returns true if the caller should flush.
300 */
301 bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_size)
302 {
303 struct mmu_gather_batch *batch;
304
305 VM_BUG_ON(!tlb->end);
306 VM_WARN_ON(tlb->page_size != page_size);
307
308 batch = tlb->active;
309 /*
310 * Add the page and check if we are full. If so
311 * force a flush.
312 */
313 batch->pages[batch->nr++] = page;
314 if (batch->nr == batch->max) {
315 if (!tlb_next_batch(tlb))
316 return true;
317 batch = tlb->active;
318 }
319 VM_BUG_ON_PAGE(batch->nr > batch->max, page);
320
321 return false;
322 }
323
324 #endif /* HAVE_GENERIC_MMU_GATHER */
325
326 #ifdef CONFIG_HAVE_RCU_TABLE_FREE
327
328 /*
329 * See the comment near struct mmu_table_batch.
330 */
331
332 static void tlb_remove_table_smp_sync(void *arg)
333 {
334 /* Simply deliver the interrupt */
335 }
336
337 static void tlb_remove_table_one(void *table)
338 {
339 /*
340 * This isn't an RCU grace period and hence the page-tables cannot be
341 * assumed to be actually RCU-freed.
342 *
343 * It is however sufficient for software page-table walkers that rely on
344 * IRQ disabling. See the comment near struct mmu_table_batch.
345 */
346 smp_call_function(tlb_remove_table_smp_sync, NULL, 1);
347 __tlb_remove_table(table);
348 }
349
350 static void tlb_remove_table_rcu(struct rcu_head *head)
351 {
352 struct mmu_table_batch *batch;
353 int i;
354
355 batch = container_of(head, struct mmu_table_batch, rcu);
356
357 for (i = 0; i < batch->nr; i++)
358 __tlb_remove_table(batch->tables[i]);
359
360 free_page((unsigned long)batch);
361 }
362
363 void tlb_table_flush(struct mmu_gather *tlb)
364 {
365 struct mmu_table_batch **batch = &tlb->batch;
366
367 if (*batch) {
368 call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu);
369 *batch = NULL;
370 }
371 }
372
373 void tlb_remove_table(struct mmu_gather *tlb, void *table)
374 {
375 struct mmu_table_batch **batch = &tlb->batch;
376
377 /*
378 * When there's less then two users of this mm there cannot be a
379 * concurrent page-table walk.
380 */
381 if (atomic_read(&tlb->mm->mm_users) < 2) {
382 __tlb_remove_table(table);
383 return;
384 }
385
386 if (*batch == NULL) {
387 *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN);
388 if (*batch == NULL) {
389 tlb_remove_table_one(table);
390 return;
391 }
392 (*batch)->nr = 0;
393 }
394 (*batch)->tables[(*batch)->nr++] = table;
395 if ((*batch)->nr == MAX_TABLE_BATCH)
396 tlb_table_flush(tlb);
397 }
398
399 #endif /* CONFIG_HAVE_RCU_TABLE_FREE */
400
401 /*
402 * Note: this doesn't free the actual pages themselves. That
403 * has been handled earlier when unmapping all the memory regions.
404 */
405 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd,
406 unsigned long addr)
407 {
408 pgtable_t token = pmd_pgtable(*pmd);
409 pmd_clear(pmd);
410 pte_free_tlb(tlb, token, addr);
411 atomic_long_dec(&tlb->mm->nr_ptes);
412 }
413
414 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
415 unsigned long addr, unsigned long end,
416 unsigned long floor, unsigned long ceiling)
417 {
418 pmd_t *pmd;
419 unsigned long next;
420 unsigned long start;
421
422 start = addr;
423 pmd = pmd_offset(pud, addr);
424 do {
425 next = pmd_addr_end(addr, end);
426 if (pmd_none_or_clear_bad(pmd))
427 continue;
428 free_pte_range(tlb, pmd, addr);
429 } while (pmd++, addr = next, addr != end);
430
431 start &= PUD_MASK;
432 if (start < floor)
433 return;
434 if (ceiling) {
435 ceiling &= PUD_MASK;
436 if (!ceiling)
437 return;
438 }
439 if (end - 1 > ceiling - 1)
440 return;
441
442 pmd = pmd_offset(pud, start);
443 pud_clear(pud);
444 pmd_free_tlb(tlb, pmd, start);
445 mm_dec_nr_pmds(tlb->mm);
446 }
447
448 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
449 unsigned long addr, unsigned long end,
450 unsigned long floor, unsigned long ceiling)
451 {
452 pud_t *pud;
453 unsigned long next;
454 unsigned long start;
455
456 start = addr;
457 pud = pud_offset(pgd, addr);
458 do {
459 next = pud_addr_end(addr, end);
460 if (pud_none_or_clear_bad(pud))
461 continue;
462 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
463 } while (pud++, addr = next, addr != end);
464
465 start &= PGDIR_MASK;
466 if (start < floor)
467 return;
468 if (ceiling) {
469 ceiling &= PGDIR_MASK;
470 if (!ceiling)
471 return;
472 }
473 if (end - 1 > ceiling - 1)
474 return;
475
476 pud = pud_offset(pgd, start);
477 pgd_clear(pgd);
478 pud_free_tlb(tlb, pud, start);
479 }
480
481 /*
482 * This function frees user-level page tables of a process.
483 */
484 void free_pgd_range(struct mmu_gather *tlb,
485 unsigned long addr, unsigned long end,
486 unsigned long floor, unsigned long ceiling)
487 {
488 pgd_t *pgd;
489 unsigned long next;
490
491 /*
492 * The next few lines have given us lots of grief...
493 *
494 * Why are we testing PMD* at this top level? Because often
495 * there will be no work to do at all, and we'd prefer not to
496 * go all the way down to the bottom just to discover that.
497 *
498 * Why all these "- 1"s? Because 0 represents both the bottom
499 * of the address space and the top of it (using -1 for the
500 * top wouldn't help much: the masks would do the wrong thing).
501 * The rule is that addr 0 and floor 0 refer to the bottom of
502 * the address space, but end 0 and ceiling 0 refer to the top
503 * Comparisons need to use "end - 1" and "ceiling - 1" (though
504 * that end 0 case should be mythical).
505 *
506 * Wherever addr is brought up or ceiling brought down, we must
507 * be careful to reject "the opposite 0" before it confuses the
508 * subsequent tests. But what about where end is brought down
509 * by PMD_SIZE below? no, end can't go down to 0 there.
510 *
511 * Whereas we round start (addr) and ceiling down, by different
512 * masks at different levels, in order to test whether a table
513 * now has no other vmas using it, so can be freed, we don't
514 * bother to round floor or end up - the tests don't need that.
515 */
516
517 addr &= PMD_MASK;
518 if (addr < floor) {
519 addr += PMD_SIZE;
520 if (!addr)
521 return;
522 }
523 if (ceiling) {
524 ceiling &= PMD_MASK;
525 if (!ceiling)
526 return;
527 }
528 if (end - 1 > ceiling - 1)
529 end -= PMD_SIZE;
530 if (addr > end - 1)
531 return;
532 /*
533 * We add page table cache pages with PAGE_SIZE,
534 * (see pte_free_tlb()), flush the tlb if we need
535 */
536 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
537 pgd = pgd_offset(tlb->mm, addr);
538 do {
539 next = pgd_addr_end(addr, end);
540 if (pgd_none_or_clear_bad(pgd))
541 continue;
542 free_pud_range(tlb, pgd, addr, next, floor, ceiling);
543 } while (pgd++, addr = next, addr != end);
544 }
545
546 void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma,
547 unsigned long floor, unsigned long ceiling)
548 {
549 while (vma) {
550 struct vm_area_struct *next = vma->vm_next;
551 unsigned long addr = vma->vm_start;
552
553 /*
554 * Hide vma from rmap and truncate_pagecache before freeing
555 * pgtables
556 */
557 unlink_anon_vmas(vma);
558 unlink_file_vma(vma);
559
560 if (is_vm_hugetlb_page(vma)) {
561 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
562 floor, next ? next->vm_start : ceiling);
563 } else {
564 /*
565 * Optimization: gather nearby vmas into one call down
566 */
567 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
568 && !is_vm_hugetlb_page(next)) {
569 vma = next;
570 next = vma->vm_next;
571 unlink_anon_vmas(vma);
572 unlink_file_vma(vma);
573 }
574 free_pgd_range(tlb, addr, vma->vm_end,
575 floor, next ? next->vm_start : ceiling);
576 }
577 vma = next;
578 }
579 }
580
581 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
582 {
583 spinlock_t *ptl;
584 pgtable_t new = pte_alloc_one(mm, address);
585 if (!new)
586 return -ENOMEM;
587
588 /*
589 * Ensure all pte setup (eg. pte page lock and page clearing) are
590 * visible before the pte is made visible to other CPUs by being
591 * put into page tables.
592 *
593 * The other side of the story is the pointer chasing in the page
594 * table walking code (when walking the page table without locking;
595 * ie. most of the time). Fortunately, these data accesses consist
596 * of a chain of data-dependent loads, meaning most CPUs (alpha
597 * being the notable exception) will already guarantee loads are
598 * seen in-order. See the alpha page table accessors for the
599 * smp_read_barrier_depends() barriers in page table walking code.
600 */
601 smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */
602
603 ptl = pmd_lock(mm, pmd);
604 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
605 atomic_long_inc(&mm->nr_ptes);
606 pmd_populate(mm, pmd, new);
607 new = NULL;
608 }
609 spin_unlock(ptl);
610 if (new)
611 pte_free(mm, new);
612 return 0;
613 }
614
615 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
616 {
617 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
618 if (!new)
619 return -ENOMEM;
620
621 smp_wmb(); /* See comment in __pte_alloc */
622
623 spin_lock(&init_mm.page_table_lock);
624 if (likely(pmd_none(*pmd))) { /* Has another populated it ? */
625 pmd_populate_kernel(&init_mm, pmd, new);
626 new = NULL;
627 }
628 spin_unlock(&init_mm.page_table_lock);
629 if (new)
630 pte_free_kernel(&init_mm, new);
631 return 0;
632 }
633
634 static inline void init_rss_vec(int *rss)
635 {
636 memset(rss, 0, sizeof(int) * NR_MM_COUNTERS);
637 }
638
639 static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss)
640 {
641 int i;
642
643 if (current->mm == mm)
644 sync_mm_rss(mm);
645 for (i = 0; i < NR_MM_COUNTERS; i++)
646 if (rss[i])
647 add_mm_counter(mm, i, rss[i]);
648 }
649
650 /*
651 * This function is called to print an error when a bad pte
652 * is found. For example, we might have a PFN-mapped pte in
653 * a region that doesn't allow it.
654 *
655 * The calling function must still handle the error.
656 */
657 static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
658 pte_t pte, struct page *page)
659 {
660 pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
661 pud_t *pud = pud_offset(pgd, addr);
662 pmd_t *pmd = pmd_offset(pud, addr);
663 struct address_space *mapping;
664 pgoff_t index;
665 static unsigned long resume;
666 static unsigned long nr_shown;
667 static unsigned long nr_unshown;
668
669 /*
670 * Allow a burst of 60 reports, then keep quiet for that minute;
671 * or allow a steady drip of one report per second.
672 */
673 if (nr_shown == 60) {
674 if (time_before(jiffies, resume)) {
675 nr_unshown++;
676 return;
677 }
678 if (nr_unshown) {
679 pr_alert("BUG: Bad page map: %lu messages suppressed\n",
680 nr_unshown);
681 nr_unshown = 0;
682 }
683 nr_shown = 0;
684 }
685 if (nr_shown++ == 0)
686 resume = jiffies + 60 * HZ;
687
688 mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
689 index = linear_page_index(vma, addr);
690
691 pr_alert("BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n",
692 current->comm,
693 (long long)pte_val(pte), (long long)pmd_val(*pmd));
694 if (page)
695 dump_page(page, "bad pte");
696 pr_alert("addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n",
697 (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
698 /*
699 * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y
700 */
701 pr_alert("file:%pD fault:%pf mmap:%pf readpage:%pf\n",
702 vma->vm_file,
703 vma->vm_ops ? vma->vm_ops->fault : NULL,
704 vma->vm_file ? vma->vm_file->f_op->mmap : NULL,
705 mapping ? mapping->a_ops->readpage : NULL);
706 dump_stack();
707 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
708 }
709
710 /*
711 * vm_normal_page -- This function gets the "struct page" associated with a pte.
712 *
713 * "Special" mappings do not wish to be associated with a "struct page" (either
714 * it doesn't exist, or it exists but they don't want to touch it). In this
715 * case, NULL is returned here. "Normal" mappings do have a struct page.
716 *
717 * There are 2 broad cases. Firstly, an architecture may define a pte_special()
718 * pte bit, in which case this function is trivial. Secondly, an architecture
719 * may not have a spare pte bit, which requires a more complicated scheme,
720 * described below.
721 *
722 * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a
723 * special mapping (even if there are underlying and valid "struct pages").
724 * COWed pages of a VM_PFNMAP are always normal.
725 *
726 * The way we recognize COWed pages within VM_PFNMAP mappings is through the
727 * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit
728 * set, and the vm_pgoff will point to the first PFN mapped: thus every special
729 * mapping will always honor the rule
730 *
731 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
732 *
733 * And for normal mappings this is false.
734 *
735 * This restricts such mappings to be a linear translation from virtual address
736 * to pfn. To get around this restriction, we allow arbitrary mappings so long
737 * as the vma is not a COW mapping; in that case, we know that all ptes are
738 * special (because none can have been COWed).
739 *
740 *
741 * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP.
742 *
743 * VM_MIXEDMAP mappings can likewise contain memory with or without "struct
744 * page" backing, however the difference is that _all_ pages with a struct
745 * page (that is, those where pfn_valid is true) are refcounted and considered
746 * normal pages by the VM. The disadvantage is that pages are refcounted
747 * (which can be slower and simply not an option for some PFNMAP users). The
748 * advantage is that we don't have to follow the strict linearity rule of
749 * PFNMAP mappings in order to support COWable mappings.
750 *
751 */
752 #ifdef __HAVE_ARCH_PTE_SPECIAL
753 # define HAVE_PTE_SPECIAL 1
754 #else
755 # define HAVE_PTE_SPECIAL 0
756 #endif
757 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr,
758 pte_t pte)
759 {
760 unsigned long pfn = pte_pfn(pte);
761
762 if (HAVE_PTE_SPECIAL) {
763 if (likely(!pte_special(pte)))
764 goto check_pfn;
765 if (vma->vm_ops && vma->vm_ops->find_special_page)
766 return vma->vm_ops->find_special_page(vma, addr);
767 if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
768 return NULL;
769 if (!is_zero_pfn(pfn))
770 print_bad_pte(vma, addr, pte, NULL);
771 return NULL;
772 }
773
774 /* !HAVE_PTE_SPECIAL case follows: */
775
776 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
777 if (vma->vm_flags & VM_MIXEDMAP) {
778 if (!pfn_valid(pfn))
779 return NULL;
780 goto out;
781 } else {
782 unsigned long off;
783 off = (addr - vma->vm_start) >> PAGE_SHIFT;
784 if (pfn == vma->vm_pgoff + off)
785 return NULL;
786 if (!is_cow_mapping(vma->vm_flags))
787 return NULL;
788 }
789 }
790
791 if (is_zero_pfn(pfn))
792 return NULL;
793 check_pfn:
794 if (unlikely(pfn > highest_memmap_pfn)) {
795 print_bad_pte(vma, addr, pte, NULL);
796 return NULL;
797 }
798
799 /*
800 * NOTE! We still have PageReserved() pages in the page tables.
801 * eg. VDSO mappings can cause them to exist.
802 */
803 out:
804 return pfn_to_page(pfn);
805 }
806
807 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
808 struct page *vm_normal_page_pmd(struct vm_area_struct *vma, unsigned long addr,
809 pmd_t pmd)
810 {
811 unsigned long pfn = pmd_pfn(pmd);
812
813 /*
814 * There is no pmd_special() but there may be special pmds, e.g.
815 * in a direct-access (dax) mapping, so let's just replicate the
816 * !HAVE_PTE_SPECIAL case from vm_normal_page() here.
817 */
818 if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) {
819 if (vma->vm_flags & VM_MIXEDMAP) {
820 if (!pfn_valid(pfn))
821 return NULL;
822 goto out;
823 } else {
824 unsigned long off;
825 off = (addr - vma->vm_start) >> PAGE_SHIFT;
826 if (pfn == vma->vm_pgoff + off)
827 return NULL;
828 if (!is_cow_mapping(vma->vm_flags))
829 return NULL;
830 }
831 }
832
833 if (is_zero_pfn(pfn))
834 return NULL;
835 if (unlikely(pfn > highest_memmap_pfn))
836 return NULL;
837
838 /*
839 * NOTE! We still have PageReserved() pages in the page tables.
840 * eg. VDSO mappings can cause them to exist.
841 */
842 out:
843 return pfn_to_page(pfn);
844 }
845 #endif
846
847 /*
848 * copy one vm_area from one task to the other. Assumes the page tables
849 * already present in the new task to be cleared in the whole range
850 * covered by this vma.
851 */
852
853 static inline unsigned long
854 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
855 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
856 unsigned long addr, int *rss)
857 {
858 unsigned long vm_flags = vma->vm_flags;
859 pte_t pte = *src_pte;
860 struct page *page;
861
862 /* pte contains position in swap or file, so copy. */
863 if (unlikely(!pte_present(pte))) {
864 swp_entry_t entry = pte_to_swp_entry(pte);
865
866 if (likely(!non_swap_entry(entry))) {
867 if (swap_duplicate(entry) < 0)
868 return entry.val;
869
870 /* make sure dst_mm is on swapoff's mmlist. */
871 if (unlikely(list_empty(&dst_mm->mmlist))) {
872 spin_lock(&mmlist_lock);
873 if (list_empty(&dst_mm->mmlist))
874 list_add(&dst_mm->mmlist,
875 &src_mm->mmlist);
876 spin_unlock(&mmlist_lock);
877 }
878 rss[MM_SWAPENTS]++;
879 } else if (is_migration_entry(entry)) {
880 page = migration_entry_to_page(entry);
881
882 rss[mm_counter(page)]++;
883
884 if (is_write_migration_entry(entry) &&
885 is_cow_mapping(vm_flags)) {
886 /*
887 * COW mappings require pages in both
888 * parent and child to be set to read.
889 */
890 make_migration_entry_read(&entry);
891 pte = swp_entry_to_pte(entry);
892 if (pte_swp_soft_dirty(*src_pte))
893 pte = pte_swp_mksoft_dirty(pte);
894 set_pte_at(src_mm, addr, src_pte, pte);
895 }
896 }
897 goto out_set_pte;
898 }
899
900 /*
901 * If it's a COW mapping, write protect it both
902 * in the parent and the child
903 */
904 if (is_cow_mapping(vm_flags)) {
905 ptep_set_wrprotect(src_mm, addr, src_pte);
906 pte = pte_wrprotect(pte);
907 }
908
909 /*
910 * If it's a shared mapping, mark it clean in
911 * the child
912 */
913 if (vm_flags & VM_SHARED)
914 pte = pte_mkclean(pte);
915 pte = pte_mkold(pte);
916
917 page = vm_normal_page(vma, addr, pte);
918 if (page) {
919 get_page(page);
920 page_dup_rmap(page, false);
921 rss[mm_counter(page)]++;
922 }
923
924 out_set_pte:
925 set_pte_at(dst_mm, addr, dst_pte, pte);
926 return 0;
927 }
928
929 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
930 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
931 unsigned long addr, unsigned long end)
932 {
933 pte_t *orig_src_pte, *orig_dst_pte;
934 pte_t *src_pte, *dst_pte;
935 spinlock_t *src_ptl, *dst_ptl;
936 int progress = 0;
937 int rss[NR_MM_COUNTERS];
938 swp_entry_t entry = (swp_entry_t){0};
939
940 again:
941 init_rss_vec(rss);
942
943 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
944 if (!dst_pte)
945 return -ENOMEM;
946 src_pte = pte_offset_map(src_pmd, addr);
947 src_ptl = pte_lockptr(src_mm, src_pmd);
948 spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING);
949 orig_src_pte = src_pte;
950 orig_dst_pte = dst_pte;
951 arch_enter_lazy_mmu_mode();
952
953 do {
954 /*
955 * We are holding two locks at this point - either of them
956 * could generate latencies in another task on another CPU.
957 */
958 if (progress >= 32) {
959 progress = 0;
960 if (need_resched() ||
961 spin_needbreak(src_ptl) || spin_needbreak(dst_ptl))
962 break;
963 }
964 if (pte_none(*src_pte)) {
965 progress++;
966 continue;
967 }
968 entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte,
969 vma, addr, rss);
970 if (entry.val)
971 break;
972 progress += 8;
973 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
974
975 arch_leave_lazy_mmu_mode();
976 spin_unlock(src_ptl);
977 pte_unmap(orig_src_pte);
978 add_mm_rss_vec(dst_mm, rss);
979 pte_unmap_unlock(orig_dst_pte, dst_ptl);
980 cond_resched();
981
982 if (entry.val) {
983 if (add_swap_count_continuation(entry, GFP_KERNEL) < 0)
984 return -ENOMEM;
985 progress = 0;
986 }
987 if (addr != end)
988 goto again;
989 return 0;
990 }
991
992 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
993 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
994 unsigned long addr, unsigned long end)
995 {
996 pmd_t *src_pmd, *dst_pmd;
997 unsigned long next;
998
999 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
1000 if (!dst_pmd)
1001 return -ENOMEM;
1002 src_pmd = pmd_offset(src_pud, addr);
1003 do {
1004 next = pmd_addr_end(addr, end);
1005 if (pmd_trans_huge(*src_pmd) || pmd_devmap(*src_pmd)) {
1006 int err;
1007 VM_BUG_ON_VMA(next-addr != HPAGE_PMD_SIZE, vma);
1008 err = copy_huge_pmd(dst_mm, src_mm,
1009 dst_pmd, src_pmd, addr, vma);
1010 if (err == -ENOMEM)
1011 return -ENOMEM;
1012 if (!err)
1013 continue;
1014 /* fall through */
1015 }
1016 if (pmd_none_or_clear_bad(src_pmd))
1017 continue;
1018 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
1019 vma, addr, next))
1020 return -ENOMEM;
1021 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
1022 return 0;
1023 }
1024
1025 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1026 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
1027 unsigned long addr, unsigned long end)
1028 {
1029 pud_t *src_pud, *dst_pud;
1030 unsigned long next;
1031
1032 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
1033 if (!dst_pud)
1034 return -ENOMEM;
1035 src_pud = pud_offset(src_pgd, addr);
1036 do {
1037 next = pud_addr_end(addr, end);
1038 if (pud_trans_huge(*src_pud) || pud_devmap(*src_pud)) {
1039 int err;
1040
1041 VM_BUG_ON_VMA(next-addr != HPAGE_PUD_SIZE, vma);
1042 err = copy_huge_pud(dst_mm, src_mm,
1043 dst_pud, src_pud, addr, vma);
1044 if (err == -ENOMEM)
1045 return -ENOMEM;
1046 if (!err)
1047 continue;
1048 /* fall through */
1049 }
1050 if (pud_none_or_clear_bad(src_pud))
1051 continue;
1052 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
1053 vma, addr, next))
1054 return -ENOMEM;
1055 } while (dst_pud++, src_pud++, addr = next, addr != end);
1056 return 0;
1057 }
1058
1059 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
1060 struct vm_area_struct *vma)
1061 {
1062 pgd_t *src_pgd, *dst_pgd;
1063 unsigned long next;
1064 unsigned long addr = vma->vm_start;
1065 unsigned long end = vma->vm_end;
1066 unsigned long mmun_start; /* For mmu_notifiers */
1067 unsigned long mmun_end; /* For mmu_notifiers */
1068 bool is_cow;
1069 int ret;
1070
1071 /*
1072 * Don't copy ptes where a page fault will fill them correctly.
1073 * Fork becomes much lighter when there are big shared or private
1074 * readonly mappings. The tradeoff is that copy_page_range is more
1075 * efficient than faulting.
1076 */
1077 if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) &&
1078 !vma->anon_vma)
1079 return 0;
1080
1081 if (is_vm_hugetlb_page(vma))
1082 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
1083
1084 if (unlikely(vma->vm_flags & VM_PFNMAP)) {
1085 /*
1086 * We do not free on error cases below as remove_vma
1087 * gets called on error from higher level routine
1088 */
1089 ret = track_pfn_copy(vma);
1090 if (ret)
1091 return ret;
1092 }
1093
1094 /*
1095 * We need to invalidate the secondary MMU mappings only when
1096 * there could be a permission downgrade on the ptes of the
1097 * parent mm. And a permission downgrade will only happen if
1098 * is_cow_mapping() returns true.
1099 */
1100 is_cow = is_cow_mapping(vma->vm_flags);
1101 mmun_start = addr;
1102 mmun_end = end;
1103 if (is_cow)
1104 mmu_notifier_invalidate_range_start(src_mm, mmun_start,
1105 mmun_end);
1106
1107 ret = 0;
1108 dst_pgd = pgd_offset(dst_mm, addr);
1109 src_pgd = pgd_offset(src_mm, addr);
1110 do {
1111 next = pgd_addr_end(addr, end);
1112 if (pgd_none_or_clear_bad(src_pgd))
1113 continue;
1114 if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
1115 vma, addr, next))) {
1116 ret = -ENOMEM;
1117 break;
1118 }
1119 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
1120
1121 if (is_cow)
1122 mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end);
1123 return ret;
1124 }
1125
1126 static unsigned long zap_pte_range(struct mmu_gather *tlb,
1127 struct vm_area_struct *vma, pmd_t *pmd,
1128 unsigned long addr, unsigned long end,
1129 struct zap_details *details)
1130 {
1131 struct mm_struct *mm = tlb->mm;
1132 int force_flush = 0;
1133 int rss[NR_MM_COUNTERS];
1134 spinlock_t *ptl;
1135 pte_t *start_pte;
1136 pte_t *pte;
1137 swp_entry_t entry;
1138
1139 tlb_remove_check_page_size_change(tlb, PAGE_SIZE);
1140 again:
1141 init_rss_vec(rss);
1142 start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1143 pte = start_pte;
1144 arch_enter_lazy_mmu_mode();
1145 do {
1146 pte_t ptent = *pte;
1147 if (pte_none(ptent))
1148 continue;
1149
1150 if (pte_present(ptent)) {
1151 struct page *page;
1152
1153 page = vm_normal_page(vma, addr, ptent);
1154 if (unlikely(details) && page) {
1155 /*
1156 * unmap_shared_mapping_pages() wants to
1157 * invalidate cache without truncating:
1158 * unmap shared but keep private pages.
1159 */
1160 if (details->check_mapping &&
1161 details->check_mapping != page_rmapping(page))
1162 continue;
1163 }
1164 ptent = ptep_get_and_clear_full(mm, addr, pte,
1165 tlb->fullmm);
1166 tlb_remove_tlb_entry(tlb, pte, addr);
1167 if (unlikely(!page))
1168 continue;
1169
1170 if (!PageAnon(page)) {
1171 if (pte_dirty(ptent)) {
1172 force_flush = 1;
1173 set_page_dirty(page);
1174 }
1175 if (pte_young(ptent) &&
1176 likely(!(vma->vm_flags & VM_SEQ_READ)))
1177 mark_page_accessed(page);
1178 }
1179 rss[mm_counter(page)]--;
1180 page_remove_rmap(page, false);
1181 if (unlikely(page_mapcount(page) < 0))
1182 print_bad_pte(vma, addr, ptent, page);
1183 if (unlikely(__tlb_remove_page(tlb, page))) {
1184 force_flush = 1;
1185 addr += PAGE_SIZE;
1186 break;
1187 }
1188 continue;
1189 }
1190 /* If details->check_mapping, we leave swap entries. */
1191 if (unlikely(details))
1192 continue;
1193
1194 entry = pte_to_swp_entry(ptent);
1195 if (!non_swap_entry(entry))
1196 rss[MM_SWAPENTS]--;
1197 else if (is_migration_entry(entry)) {
1198 struct page *page;
1199
1200 page = migration_entry_to_page(entry);
1201 rss[mm_counter(page)]--;
1202 }
1203 if (unlikely(!free_swap_and_cache(entry)))
1204 print_bad_pte(vma, addr, ptent, NULL);
1205 pte_clear_not_present_full(mm, addr, pte, tlb->fullmm);
1206 } while (pte++, addr += PAGE_SIZE, addr != end);
1207
1208 add_mm_rss_vec(mm, rss);
1209 arch_leave_lazy_mmu_mode();
1210
1211 /* Do the actual TLB flush before dropping ptl */
1212 if (force_flush)
1213 tlb_flush_mmu_tlbonly(tlb);
1214 pte_unmap_unlock(start_pte, ptl);
1215
1216 /*
1217 * If we forced a TLB flush (either due to running out of
1218 * batch buffers or because we needed to flush dirty TLB
1219 * entries before releasing the ptl), free the batched
1220 * memory too. Restart if we didn't do everything.
1221 */
1222 if (force_flush) {
1223 force_flush = 0;
1224 tlb_flush_mmu_free(tlb);
1225 if (addr != end)
1226 goto again;
1227 }
1228
1229 return addr;
1230 }
1231
1232 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
1233 struct vm_area_struct *vma, pud_t *pud,
1234 unsigned long addr, unsigned long end,
1235 struct zap_details *details)
1236 {
1237 pmd_t *pmd;
1238 unsigned long next;
1239
1240 pmd = pmd_offset(pud, addr);
1241 do {
1242 next = pmd_addr_end(addr, end);
1243 if (pmd_trans_huge(*pmd) || pmd_devmap(*pmd)) {
1244 if (next - addr != HPAGE_PMD_SIZE) {
1245 VM_BUG_ON_VMA(vma_is_anonymous(vma) &&
1246 !rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1247 __split_huge_pmd(vma, pmd, addr, false, NULL);
1248 } else if (zap_huge_pmd(tlb, vma, pmd, addr))
1249 goto next;
1250 /* fall through */
1251 }
1252 /*
1253 * Here there can be other concurrent MADV_DONTNEED or
1254 * trans huge page faults running, and if the pmd is
1255 * none or trans huge it can change under us. This is
1256 * because MADV_DONTNEED holds the mmap_sem in read
1257 * mode.
1258 */
1259 if (pmd_none_or_trans_huge_or_clear_bad(pmd))
1260 goto next;
1261 next = zap_pte_range(tlb, vma, pmd, addr, next, details);
1262 next:
1263 cond_resched();
1264 } while (pmd++, addr = next, addr != end);
1265
1266 return addr;
1267 }
1268
1269 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
1270 struct vm_area_struct *vma, pgd_t *pgd,
1271 unsigned long addr, unsigned long end,
1272 struct zap_details *details)
1273 {
1274 pud_t *pud;
1275 unsigned long next;
1276
1277 pud = pud_offset(pgd, addr);
1278 do {
1279 next = pud_addr_end(addr, end);
1280 if (pud_trans_huge(*pud) || pud_devmap(*pud)) {
1281 if (next - addr != HPAGE_PUD_SIZE) {
1282 VM_BUG_ON_VMA(!rwsem_is_locked(&tlb->mm->mmap_sem), vma);
1283 split_huge_pud(vma, pud, addr);
1284 } else if (zap_huge_pud(tlb, vma, pud, addr))
1285 goto next;
1286 /* fall through */
1287 }
1288 if (pud_none_or_clear_bad(pud))
1289 continue;
1290 next = zap_pmd_range(tlb, vma, pud, addr, next, details);
1291 next:
1292 cond_resched();
1293 } while (pud++, addr = next, addr != end);
1294
1295 return addr;
1296 }
1297
1298 void unmap_page_range(struct mmu_gather *tlb,
1299 struct vm_area_struct *vma,
1300 unsigned long addr, unsigned long end,
1301 struct zap_details *details)
1302 {
1303 pgd_t *pgd;
1304 unsigned long next;
1305
1306 BUG_ON(addr >= end);
1307 tlb_start_vma(tlb, vma);
1308 pgd = pgd_offset(vma->vm_mm, addr);
1309 do {
1310 next = pgd_addr_end(addr, end);
1311 if (pgd_none_or_clear_bad(pgd))
1312 continue;
1313 next = zap_pud_range(tlb, vma, pgd, addr, next, details);
1314 } while (pgd++, addr = next, addr != end);
1315 tlb_end_vma(tlb, vma);
1316 }
1317
1318
1319 static void unmap_single_vma(struct mmu_gather *tlb,
1320 struct vm_area_struct *vma, unsigned long start_addr,
1321 unsigned long end_addr,
1322 struct zap_details *details)
1323 {
1324 unsigned long start = max(vma->vm_start, start_addr);
1325 unsigned long end;
1326
1327 if (start >= vma->vm_end)
1328 return;
1329 end = min(vma->vm_end, end_addr);
1330 if (end <= vma->vm_start)
1331 return;
1332
1333 if (vma->vm_file)
1334 uprobe_munmap(vma, start, end);
1335
1336 if (unlikely(vma->vm_flags & VM_PFNMAP))
1337 untrack_pfn(vma, 0, 0);
1338
1339 if (start != end) {
1340 if (unlikely(is_vm_hugetlb_page(vma))) {
1341 /*
1342 * It is undesirable to test vma->vm_file as it
1343 * should be non-null for valid hugetlb area.
1344 * However, vm_file will be NULL in the error
1345 * cleanup path of mmap_region. When
1346 * hugetlbfs ->mmap method fails,
1347 * mmap_region() nullifies vma->vm_file
1348 * before calling this function to clean up.
1349 * Since no pte has actually been setup, it is
1350 * safe to do nothing in this case.
1351 */
1352 if (vma->vm_file) {
1353 i_mmap_lock_write(vma->vm_file->f_mapping);
1354 __unmap_hugepage_range_final(tlb, vma, start, end, NULL);
1355 i_mmap_unlock_write(vma->vm_file->f_mapping);
1356 }
1357 } else
1358 unmap_page_range(tlb, vma, start, end, details);
1359 }
1360 }
1361
1362 /**
1363 * unmap_vmas - unmap a range of memory covered by a list of vma's
1364 * @tlb: address of the caller's struct mmu_gather
1365 * @vma: the starting vma
1366 * @start_addr: virtual address at which to start unmapping
1367 * @end_addr: virtual address at which to end unmapping
1368 *
1369 * Unmap all pages in the vma list.
1370 *
1371 * Only addresses between `start' and `end' will be unmapped.
1372 *
1373 * The VMA list must be sorted in ascending virtual address order.
1374 *
1375 * unmap_vmas() assumes that the caller will flush the whole unmapped address
1376 * range after unmap_vmas() returns. So the only responsibility here is to
1377 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
1378 * drops the lock and schedules.
1379 */
1380 void unmap_vmas(struct mmu_gather *tlb,
1381 struct vm_area_struct *vma, unsigned long start_addr,
1382 unsigned long end_addr)
1383 {
1384 struct mm_struct *mm = vma->vm_mm;
1385
1386 mmu_notifier_invalidate_range_start(mm, start_addr, end_addr);
1387 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next)
1388 unmap_single_vma(tlb, vma, start_addr, end_addr, NULL);
1389 mmu_notifier_invalidate_range_end(mm, start_addr, end_addr);
1390 }
1391
1392 /**
1393 * zap_page_range - remove user pages in a given range
1394 * @vma: vm_area_struct holding the applicable pages
1395 * @start: starting address of pages to zap
1396 * @size: number of bytes to zap
1397 *
1398 * Caller must protect the VMA list
1399 */
1400 void zap_page_range(struct vm_area_struct *vma, unsigned long start,
1401 unsigned long size)
1402 {
1403 struct mm_struct *mm = vma->vm_mm;
1404 struct mmu_gather tlb;
1405 unsigned long end = start + size;
1406
1407 lru_add_drain();
1408 tlb_gather_mmu(&tlb, mm, start, end);
1409 update_hiwater_rss(mm);
1410 mmu_notifier_invalidate_range_start(mm, start, end);
1411 for ( ; vma && vma->vm_start < end; vma = vma->vm_next)
1412 unmap_single_vma(&tlb, vma, start, end, NULL);
1413 mmu_notifier_invalidate_range_end(mm, start, end);
1414 tlb_finish_mmu(&tlb, start, end);
1415 }
1416
1417 /**
1418 * zap_page_range_single - remove user pages in a given range
1419 * @vma: vm_area_struct holding the applicable pages
1420 * @address: starting address of pages to zap
1421 * @size: number of bytes to zap
1422 * @details: details of shared cache invalidation
1423 *
1424 * The range must fit into one VMA.
1425 */
1426 static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address,
1427 unsigned long size, struct zap_details *details)
1428 {
1429 struct mm_struct *mm = vma->vm_mm;
1430 struct mmu_gather tlb;
1431 unsigned long end = address + size;
1432
1433 lru_add_drain();
1434 tlb_gather_mmu(&tlb, mm, address, end);
1435 update_hiwater_rss(mm);
1436 mmu_notifier_invalidate_range_start(mm, address, end);
1437 unmap_single_vma(&tlb, vma, address, end, details);
1438 mmu_notifier_invalidate_range_end(mm, address, end);
1439 tlb_finish_mmu(&tlb, address, end);
1440 }
1441
1442 /**
1443 * zap_vma_ptes - remove ptes mapping the vma
1444 * @vma: vm_area_struct holding ptes to be zapped
1445 * @address: starting address of pages to zap
1446 * @size: number of bytes to zap
1447 *
1448 * This function only unmaps ptes assigned to VM_PFNMAP vmas.
1449 *
1450 * The entire address range must be fully contained within the vma.
1451 *
1452 * Returns 0 if successful.
1453 */
1454 int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address,
1455 unsigned long size)
1456 {
1457 if (address < vma->vm_start || address + size > vma->vm_end ||
1458 !(vma->vm_flags & VM_PFNMAP))
1459 return -1;
1460 zap_page_range_single(vma, address, size, NULL);
1461 return 0;
1462 }
1463 EXPORT_SYMBOL_GPL(zap_vma_ptes);
1464
1465 pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr,
1466 spinlock_t **ptl)
1467 {
1468 pgd_t *pgd = pgd_offset(mm, addr);
1469 pud_t *pud = pud_alloc(mm, pgd, addr);
1470 if (pud) {
1471 pmd_t *pmd = pmd_alloc(mm, pud, addr);
1472 if (pmd) {
1473 VM_BUG_ON(pmd_trans_huge(*pmd));
1474 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1475 }
1476 }
1477 return NULL;
1478 }
1479
1480 /*
1481 * This is the old fallback for page remapping.
1482 *
1483 * For historical reasons, it only allows reserved pages. Only
1484 * old drivers should use this, and they needed to mark their
1485 * pages reserved for the old functions anyway.
1486 */
1487 static int insert_page(struct vm_area_struct *vma, unsigned long addr,
1488 struct page *page, pgprot_t prot)
1489 {
1490 struct mm_struct *mm = vma->vm_mm;
1491 int retval;
1492 pte_t *pte;
1493 spinlock_t *ptl;
1494
1495 retval = -EINVAL;
1496 if (PageAnon(page))
1497 goto out;
1498 retval = -ENOMEM;
1499 flush_dcache_page(page);
1500 pte = get_locked_pte(mm, addr, &ptl);
1501 if (!pte)
1502 goto out;
1503 retval = -EBUSY;
1504 if (!pte_none(*pte))
1505 goto out_unlock;
1506
1507 /* Ok, finally just insert the thing.. */
1508 get_page(page);
1509 inc_mm_counter_fast(mm, mm_counter_file(page));
1510 page_add_file_rmap(page, false);
1511 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1512
1513 retval = 0;
1514 pte_unmap_unlock(pte, ptl);
1515 return retval;
1516 out_unlock:
1517 pte_unmap_unlock(pte, ptl);
1518 out:
1519 return retval;
1520 }
1521
1522 /**
1523 * vm_insert_page - insert single page into user vma
1524 * @vma: user vma to map to
1525 * @addr: target user address of this page
1526 * @page: source kernel page
1527 *
1528 * This allows drivers to insert individual pages they've allocated
1529 * into a user vma.
1530 *
1531 * The page has to be a nice clean _individual_ kernel allocation.
1532 * If you allocate a compound page, you need to have marked it as
1533 * such (__GFP_COMP), or manually just split the page up yourself
1534 * (see split_page()).
1535 *
1536 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1537 * took an arbitrary page protection parameter. This doesn't allow
1538 * that. Your vma protection will have to be set up correctly, which
1539 * means that if you want a shared writable mapping, you'd better
1540 * ask for a shared writable mapping!
1541 *
1542 * The page does not need to be reserved.
1543 *
1544 * Usually this function is called from f_op->mmap() handler
1545 * under mm->mmap_sem write-lock, so it can change vma->vm_flags.
1546 * Caller must set VM_MIXEDMAP on vma if it wants to call this
1547 * function from other places, for example from page-fault handler.
1548 */
1549 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
1550 struct page *page)
1551 {
1552 if (addr < vma->vm_start || addr >= vma->vm_end)
1553 return -EFAULT;
1554 if (!page_count(page))
1555 return -EINVAL;
1556 if (!(vma->vm_flags & VM_MIXEDMAP)) {
1557 BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem));
1558 BUG_ON(vma->vm_flags & VM_PFNMAP);
1559 vma->vm_flags |= VM_MIXEDMAP;
1560 }
1561 return insert_page(vma, addr, page, vma->vm_page_prot);
1562 }
1563 EXPORT_SYMBOL(vm_insert_page);
1564
1565 static int insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1566 pfn_t pfn, pgprot_t prot)
1567 {
1568 struct mm_struct *mm = vma->vm_mm;
1569 int retval;
1570 pte_t *pte, entry;
1571 spinlock_t *ptl;
1572
1573 retval = -ENOMEM;
1574 pte = get_locked_pte(mm, addr, &ptl);
1575 if (!pte)
1576 goto out;
1577 retval = -EBUSY;
1578 if (!pte_none(*pte))
1579 goto out_unlock;
1580
1581 /* Ok, finally just insert the thing.. */
1582 if (pfn_t_devmap(pfn))
1583 entry = pte_mkdevmap(pfn_t_pte(pfn, prot));
1584 else
1585 entry = pte_mkspecial(pfn_t_pte(pfn, prot));
1586 set_pte_at(mm, addr, pte, entry);
1587 update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */
1588
1589 retval = 0;
1590 out_unlock:
1591 pte_unmap_unlock(pte, ptl);
1592 out:
1593 return retval;
1594 }
1595
1596 /**
1597 * vm_insert_pfn - insert single pfn into user vma
1598 * @vma: user vma to map to
1599 * @addr: target user address of this page
1600 * @pfn: source kernel pfn
1601 *
1602 * Similar to vm_insert_page, this allows drivers to insert individual pages
1603 * they've allocated into a user vma. Same comments apply.
1604 *
1605 * This function should only be called from a vm_ops->fault handler, and
1606 * in that case the handler should return NULL.
1607 *
1608 * vma cannot be a COW mapping.
1609 *
1610 * As this is called only for pages that do not currently exist, we
1611 * do not need to flush old virtual caches or the TLB.
1612 */
1613 int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr,
1614 unsigned long pfn)
1615 {
1616 return vm_insert_pfn_prot(vma, addr, pfn, vma->vm_page_prot);
1617 }
1618 EXPORT_SYMBOL(vm_insert_pfn);
1619
1620 /**
1621 * vm_insert_pfn_prot - insert single pfn into user vma with specified pgprot
1622 * @vma: user vma to map to
1623 * @addr: target user address of this page
1624 * @pfn: source kernel pfn
1625 * @pgprot: pgprot flags for the inserted page
1626 *
1627 * This is exactly like vm_insert_pfn, except that it allows drivers to
1628 * to override pgprot on a per-page basis.
1629 *
1630 * This only makes sense for IO mappings, and it makes no sense for
1631 * cow mappings. In general, using multiple vmas is preferable;
1632 * vm_insert_pfn_prot should only be used if using multiple VMAs is
1633 * impractical.
1634 */
1635 int vm_insert_pfn_prot(struct vm_area_struct *vma, unsigned long addr,
1636 unsigned long pfn, pgprot_t pgprot)
1637 {
1638 int ret;
1639 /*
1640 * Technically, architectures with pte_special can avoid all these
1641 * restrictions (same for remap_pfn_range). However we would like
1642 * consistency in testing and feature parity among all, so we should
1643 * try to keep these invariants in place for everybody.
1644 */
1645 BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)));
1646 BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) ==
1647 (VM_PFNMAP|VM_MIXEDMAP));
1648 BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags));
1649 BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn));
1650
1651 if (addr < vma->vm_start || addr >= vma->vm_end)
1652 return -EFAULT;
1653
1654 track_pfn_insert(vma, &pgprot, __pfn_to_pfn_t(pfn, PFN_DEV));
1655
1656 ret = insert_pfn(vma, addr, __pfn_to_pfn_t(pfn, PFN_DEV), pgprot);
1657
1658 return ret;
1659 }
1660 EXPORT_SYMBOL(vm_insert_pfn_prot);
1661
1662 int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr,
1663 pfn_t pfn)
1664 {
1665 pgprot_t pgprot = vma->vm_page_prot;
1666
1667 BUG_ON(!(vma->vm_flags & VM_MIXEDMAP));
1668
1669 if (addr < vma->vm_start || addr >= vma->vm_end)
1670 return -EFAULT;
1671
1672 track_pfn_insert(vma, &pgprot, pfn);
1673
1674 /*
1675 * If we don't have pte special, then we have to use the pfn_valid()
1676 * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must*
1677 * refcount the page if pfn_valid is true (hence insert_page rather
1678 * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP
1679 * without pte special, it would there be refcounted as a normal page.
1680 */
1681 if (!HAVE_PTE_SPECIAL && !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 return insert_page(vma, addr, page, pgprot);
1691 }
1692 return insert_pfn(vma, addr, pfn, pgprot);
1693 }
1694 EXPORT_SYMBOL(vm_insert_mixed);
1695
1696 /*
1697 * maps a range of physical memory into the requested pages. the old
1698 * mappings are removed. any references to nonexistent pages results
1699 * in null mappings (currently treated as "copy-on-access")
1700 */
1701 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1702 unsigned long addr, unsigned long end,
1703 unsigned long pfn, pgprot_t prot)
1704 {
1705 pte_t *pte;
1706 spinlock_t *ptl;
1707
1708 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1709 if (!pte)
1710 return -ENOMEM;
1711 arch_enter_lazy_mmu_mode();
1712 do {
1713 BUG_ON(!pte_none(*pte));
1714 set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot)));
1715 pfn++;
1716 } while (pte++, addr += PAGE_SIZE, addr != end);
1717 arch_leave_lazy_mmu_mode();
1718 pte_unmap_unlock(pte - 1, ptl);
1719 return 0;
1720 }
1721
1722 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1723 unsigned long addr, unsigned long end,
1724 unsigned long pfn, pgprot_t prot)
1725 {
1726 pmd_t *pmd;
1727 unsigned long next;
1728
1729 pfn -= addr >> PAGE_SHIFT;
1730 pmd = pmd_alloc(mm, pud, addr);
1731 if (!pmd)
1732 return -ENOMEM;
1733 VM_BUG_ON(pmd_trans_huge(*pmd));
1734 do {
1735 next = pmd_addr_end(addr, end);
1736 if (remap_pte_range(mm, pmd, addr, next,
1737 pfn + (addr >> PAGE_SHIFT), prot))
1738 return -ENOMEM;
1739 } while (pmd++, addr = next, addr != end);
1740 return 0;
1741 }
1742
1743 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1744 unsigned long addr, unsigned long end,
1745 unsigned long pfn, pgprot_t prot)
1746 {
1747 pud_t *pud;
1748 unsigned long next;
1749
1750 pfn -= addr >> PAGE_SHIFT;
1751 pud = pud_alloc(mm, pgd, addr);
1752 if (!pud)
1753 return -ENOMEM;
1754 do {
1755 next = pud_addr_end(addr, end);
1756 if (remap_pmd_range(mm, pud, addr, next,
1757 pfn + (addr >> PAGE_SHIFT), prot))
1758 return -ENOMEM;
1759 } while (pud++, addr = next, addr != end);
1760 return 0;
1761 }
1762
1763 /**
1764 * remap_pfn_range - remap kernel memory to userspace
1765 * @vma: user vma to map to
1766 * @addr: target user address to start at
1767 * @pfn: physical address of kernel memory
1768 * @size: size of map area
1769 * @prot: page protection flags for this mapping
1770 *
1771 * Note: this is only safe if the mm semaphore is held when called.
1772 */
1773 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1774 unsigned long pfn, unsigned long size, pgprot_t prot)
1775 {
1776 pgd_t *pgd;
1777 unsigned long next;
1778 unsigned long end = addr + PAGE_ALIGN(size);
1779 struct mm_struct *mm = vma->vm_mm;
1780 unsigned long remap_pfn = pfn;
1781 int err;
1782
1783 /*
1784 * Physically remapped pages are special. Tell the
1785 * rest of the world about it:
1786 * VM_IO tells people not to look at these pages
1787 * (accesses can have side effects).
1788 * VM_PFNMAP tells the core MM that the base pages are just
1789 * raw PFN mappings, and do not have a "struct page" associated
1790 * with them.
1791 * VM_DONTEXPAND
1792 * Disable vma merging and expanding with mremap().
1793 * VM_DONTDUMP
1794 * Omit vma from core dump, even when VM_IO turned off.
1795 *
1796 * There's a horrible special case to handle copy-on-write
1797 * behaviour that some programs depend on. We mark the "original"
1798 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1799 * See vm_normal_page() for details.
1800 */
1801 if (is_cow_mapping(vma->vm_flags)) {
1802 if (addr != vma->vm_start || end != vma->vm_end)
1803 return -EINVAL;
1804 vma->vm_pgoff = pfn;
1805 }
1806
1807 err = track_pfn_remap(vma, &prot, remap_pfn, addr, PAGE_ALIGN(size));
1808 if (err)
1809 return -EINVAL;
1810
1811 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1812
1813 BUG_ON(addr >= end);
1814 pfn -= addr >> PAGE_SHIFT;
1815 pgd = pgd_offset(mm, addr);
1816 flush_cache_range(vma, addr, end);
1817 do {
1818 next = pgd_addr_end(addr, end);
1819 err = remap_pud_range(mm, pgd, addr, next,
1820 pfn + (addr >> PAGE_SHIFT), prot);
1821 if (err)
1822 break;
1823 } while (pgd++, addr = next, addr != end);
1824
1825 if (err)
1826 untrack_pfn(vma, remap_pfn, PAGE_ALIGN(size));
1827
1828 return err;
1829 }
1830 EXPORT_SYMBOL(remap_pfn_range);
1831
1832 /**
1833 * vm_iomap_memory - remap memory to userspace
1834 * @vma: user vma to map to
1835 * @start: start of area
1836 * @len: size of area
1837 *
1838 * This is a simplified io_remap_pfn_range() for common driver use. The
1839 * driver just needs to give us the physical memory range to be mapped,
1840 * we'll figure out the rest from the vma information.
1841 *
1842 * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get
1843 * whatever write-combining details or similar.
1844 */
1845 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1846 {
1847 unsigned long vm_len, pfn, pages;
1848
1849 /* Check that the physical memory area passed in looks valid */
1850 if (start + len < start)
1851 return -EINVAL;
1852 /*
1853 * You *really* shouldn't map things that aren't page-aligned,
1854 * but we've historically allowed it because IO memory might
1855 * just have smaller alignment.
1856 */
1857 len += start & ~PAGE_MASK;
1858 pfn = start >> PAGE_SHIFT;
1859 pages = (len + ~PAGE_MASK) >> PAGE_SHIFT;
1860 if (pfn + pages < pfn)
1861 return -EINVAL;
1862
1863 /* We start the mapping 'vm_pgoff' pages into the area */
1864 if (vma->vm_pgoff > pages)
1865 return -EINVAL;
1866 pfn += vma->vm_pgoff;
1867 pages -= vma->vm_pgoff;
1868
1869 /* Can we fit all of the mapping? */
1870 vm_len = vma->vm_end - vma->vm_start;
1871 if (vm_len >> PAGE_SHIFT > pages)
1872 return -EINVAL;
1873
1874 /* Ok, let it rip */
1875 return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1876 }
1877 EXPORT_SYMBOL(vm_iomap_memory);
1878
1879 static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1880 unsigned long addr, unsigned long end,
1881 pte_fn_t fn, void *data)
1882 {
1883 pte_t *pte;
1884 int err;
1885 pgtable_t token;
1886 spinlock_t *uninitialized_var(ptl);
1887
1888 pte = (mm == &init_mm) ?
1889 pte_alloc_kernel(pmd, addr) :
1890 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1891 if (!pte)
1892 return -ENOMEM;
1893
1894 BUG_ON(pmd_huge(*pmd));
1895
1896 arch_enter_lazy_mmu_mode();
1897
1898 token = pmd_pgtable(*pmd);
1899
1900 do {
1901 err = fn(pte++, token, addr, data);
1902 if (err)
1903 break;
1904 } while (addr += PAGE_SIZE, addr != end);
1905
1906 arch_leave_lazy_mmu_mode();
1907
1908 if (mm != &init_mm)
1909 pte_unmap_unlock(pte-1, ptl);
1910 return err;
1911 }
1912
1913 static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1914 unsigned long addr, unsigned long end,
1915 pte_fn_t fn, void *data)
1916 {
1917 pmd_t *pmd;
1918 unsigned long next;
1919 int err;
1920
1921 BUG_ON(pud_huge(*pud));
1922
1923 pmd = pmd_alloc(mm, pud, addr);
1924 if (!pmd)
1925 return -ENOMEM;
1926 do {
1927 next = pmd_addr_end(addr, end);
1928 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1929 if (err)
1930 break;
1931 } while (pmd++, addr = next, addr != end);
1932 return err;
1933 }
1934
1935 static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1936 unsigned long addr, unsigned long end,
1937 pte_fn_t fn, void *data)
1938 {
1939 pud_t *pud;
1940 unsigned long next;
1941 int err;
1942
1943 pud = pud_alloc(mm, pgd, addr);
1944 if (!pud)
1945 return -ENOMEM;
1946 do {
1947 next = pud_addr_end(addr, end);
1948 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1949 if (err)
1950 break;
1951 } while (pud++, addr = next, addr != end);
1952 return err;
1953 }
1954
1955 /*
1956 * Scan a region of virtual memory, filling in page tables as necessary
1957 * and calling a provided function on each leaf page table.
1958 */
1959 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1960 unsigned long size, pte_fn_t fn, void *data)
1961 {
1962 pgd_t *pgd;
1963 unsigned long next;
1964 unsigned long end = addr + size;
1965 int err;
1966
1967 if (WARN_ON(addr >= end))
1968 return -EINVAL;
1969
1970 pgd = pgd_offset(mm, addr);
1971 do {
1972 next = pgd_addr_end(addr, end);
1973 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1974 if (err)
1975 break;
1976 } while (pgd++, addr = next, addr != end);
1977
1978 return err;
1979 }
1980 EXPORT_SYMBOL_GPL(apply_to_page_range);
1981
1982 /*
1983 * handle_pte_fault chooses page fault handler according to an entry which was
1984 * read non-atomically. Before making any commitment, on those architectures
1985 * or configurations (e.g. i386 with PAE) which might give a mix of unmatched
1986 * parts, do_swap_page must check under lock before unmapping the pte and
1987 * proceeding (but do_wp_page is only called after already making such a check;
1988 * and do_anonymous_page can safely check later on).
1989 */
1990 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1991 pte_t *page_table, pte_t orig_pte)
1992 {
1993 int same = 1;
1994 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1995 if (sizeof(pte_t) > sizeof(unsigned long)) {
1996 spinlock_t *ptl = pte_lockptr(mm, pmd);
1997 spin_lock(ptl);
1998 same = pte_same(*page_table, orig_pte);
1999 spin_unlock(ptl);
2000 }
2001 #endif
2002 pte_unmap(page_table);
2003 return same;
2004 }
2005
2006 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma)
2007 {
2008 debug_dma_assert_idle(src);
2009
2010 /*
2011 * If the source page was a PFN mapping, we don't have
2012 * a "struct page" for it. We do a best-effort copy by
2013 * just copying from the original user address. If that
2014 * fails, we just zero-fill it. Live with it.
2015 */
2016 if (unlikely(!src)) {
2017 void *kaddr = kmap_atomic(dst);
2018 void __user *uaddr = (void __user *)(va & PAGE_MASK);
2019
2020 /*
2021 * This really shouldn't fail, because the page is there
2022 * in the page tables. But it might just be unreadable,
2023 * in which case we just give up and fill the result with
2024 * zeroes.
2025 */
2026 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
2027 clear_page(kaddr);
2028 kunmap_atomic(kaddr);
2029 flush_dcache_page(dst);
2030 } else
2031 copy_user_highpage(dst, src, va, vma);
2032 }
2033
2034 static gfp_t __get_fault_gfp_mask(struct vm_area_struct *vma)
2035 {
2036 struct file *vm_file = vma->vm_file;
2037
2038 if (vm_file)
2039 return mapping_gfp_mask(vm_file->f_mapping) | __GFP_FS | __GFP_IO;
2040
2041 /*
2042 * Special mappings (e.g. VDSO) do not have any file so fake
2043 * a default GFP_KERNEL for them.
2044 */
2045 return GFP_KERNEL;
2046 }
2047
2048 /*
2049 * Notify the address space that the page is about to become writable so that
2050 * it can prohibit this or wait for the page to get into an appropriate state.
2051 *
2052 * We do this without the lock held, so that it can sleep if it needs to.
2053 */
2054 static int do_page_mkwrite(struct vm_fault *vmf)
2055 {
2056 int ret;
2057 struct page *page = vmf->page;
2058 unsigned int old_flags = vmf->flags;
2059
2060 vmf->flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE;
2061
2062 ret = vmf->vma->vm_ops->page_mkwrite(vmf);
2063 /* Restore original flags so that caller is not surprised */
2064 vmf->flags = old_flags;
2065 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))
2066 return ret;
2067 if (unlikely(!(ret & VM_FAULT_LOCKED))) {
2068 lock_page(page);
2069 if (!page->mapping) {
2070 unlock_page(page);
2071 return 0; /* retry */
2072 }
2073 ret |= VM_FAULT_LOCKED;
2074 } else
2075 VM_BUG_ON_PAGE(!PageLocked(page), page);
2076 return ret;
2077 }
2078
2079 /*
2080 * Handle dirtying of a page in shared file mapping on a write fault.
2081 *
2082 * The function expects the page to be locked and unlocks it.
2083 */
2084 static void fault_dirty_shared_page(struct vm_area_struct *vma,
2085 struct page *page)
2086 {
2087 struct address_space *mapping;
2088 bool dirtied;
2089 bool page_mkwrite = vma->vm_ops && vma->vm_ops->page_mkwrite;
2090
2091 dirtied = set_page_dirty(page);
2092 VM_BUG_ON_PAGE(PageAnon(page), page);
2093 /*
2094 * Take a local copy of the address_space - page.mapping may be zeroed
2095 * by truncate after unlock_page(). The address_space itself remains
2096 * pinned by vma->vm_file's reference. We rely on unlock_page()'s
2097 * release semantics to prevent the compiler from undoing this copying.
2098 */
2099 mapping = page_rmapping(page);
2100 unlock_page(page);
2101
2102 if ((dirtied || page_mkwrite) && mapping) {
2103 /*
2104 * Some device drivers do not set page.mapping
2105 * but still dirty their pages
2106 */
2107 balance_dirty_pages_ratelimited(mapping);
2108 }
2109
2110 if (!page_mkwrite)
2111 file_update_time(vma->vm_file);
2112 }
2113
2114 /*
2115 * Handle write page faults for pages that can be reused in the current vma
2116 *
2117 * This can happen either due to the mapping being with the VM_SHARED flag,
2118 * or due to us being the last reference standing to the page. In either
2119 * case, all we need to do here is to mark the page as writable and update
2120 * any related book-keeping.
2121 */
2122 static inline void wp_page_reuse(struct vm_fault *vmf)
2123 __releases(vmf->ptl)
2124 {
2125 struct vm_area_struct *vma = vmf->vma;
2126 struct page *page = vmf->page;
2127 pte_t entry;
2128 /*
2129 * Clear the pages cpupid information as the existing
2130 * information potentially belongs to a now completely
2131 * unrelated process.
2132 */
2133 if (page)
2134 page_cpupid_xchg_last(page, (1 << LAST_CPUPID_SHIFT) - 1);
2135
2136 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2137 entry = pte_mkyoung(vmf->orig_pte);
2138 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2139 if (ptep_set_access_flags(vma, vmf->address, vmf->pte, entry, 1))
2140 update_mmu_cache(vma, vmf->address, vmf->pte);
2141 pte_unmap_unlock(vmf->pte, vmf->ptl);
2142 }
2143
2144 /*
2145 * Handle the case of a page which we actually need to copy to a new page.
2146 *
2147 * Called with mmap_sem locked and the old page referenced, but
2148 * without the ptl held.
2149 *
2150 * High level logic flow:
2151 *
2152 * - Allocate a page, copy the content of the old page to the new one.
2153 * - Handle book keeping and accounting - cgroups, mmu-notifiers, etc.
2154 * - Take the PTL. If the pte changed, bail out and release the allocated page
2155 * - If the pte is still the way we remember it, update the page table and all
2156 * relevant references. This includes dropping the reference the page-table
2157 * held to the old page, as well as updating the rmap.
2158 * - In any case, unlock the PTL and drop the reference we took to the old page.
2159 */
2160 static int wp_page_copy(struct vm_fault *vmf)
2161 {
2162 struct vm_area_struct *vma = vmf->vma;
2163 struct mm_struct *mm = vma->vm_mm;
2164 struct page *old_page = vmf->page;
2165 struct page *new_page = NULL;
2166 pte_t entry;
2167 int page_copied = 0;
2168 const unsigned long mmun_start = vmf->address & PAGE_MASK;
2169 const unsigned long mmun_end = mmun_start + PAGE_SIZE;
2170 struct mem_cgroup *memcg;
2171
2172 if (unlikely(anon_vma_prepare(vma)))
2173 goto oom;
2174
2175 if (is_zero_pfn(pte_pfn(vmf->orig_pte))) {
2176 new_page = alloc_zeroed_user_highpage_movable(vma,
2177 vmf->address);
2178 if (!new_page)
2179 goto oom;
2180 } else {
2181 new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma,
2182 vmf->address);
2183 if (!new_page)
2184 goto oom;
2185 cow_user_page(new_page, old_page, vmf->address, vma);
2186 }
2187
2188 if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg, false))
2189 goto oom_free_new;
2190
2191 __SetPageUptodate(new_page);
2192
2193 mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end);
2194
2195 /*
2196 * Re-check the pte - we dropped the lock
2197 */
2198 vmf->pte = pte_offset_map_lock(mm, vmf->pmd, vmf->address, &vmf->ptl);
2199 if (likely(pte_same(*vmf->pte, vmf->orig_pte))) {
2200 if (old_page) {
2201 if (!PageAnon(old_page)) {
2202 dec_mm_counter_fast(mm,
2203 mm_counter_file(old_page));
2204 inc_mm_counter_fast(mm, MM_ANONPAGES);
2205 }
2206 } else {
2207 inc_mm_counter_fast(mm, MM_ANONPAGES);
2208 }
2209 flush_cache_page(vma, vmf->address, pte_pfn(vmf->orig_pte));
2210 entry = mk_pte(new_page, vma->vm_page_prot);
2211 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2212 /*
2213 * Clear the pte entry and flush it first, before updating the
2214 * pte with the new entry. This will avoid a race condition
2215 * seen in the presence of one thread doing SMC and another
2216 * thread doing COW.
2217 */
2218 ptep_clear_flush_notify(vma, vmf->address, vmf->pte);
2219 page_add_new_anon_rmap(new_page, vma, vmf->address, false);
2220 mem_cgroup_commit_charge(new_page, memcg, false, false);
2221 lru_cache_add_active_or_unevictable(new_page, vma);
2222 /*
2223 * We call the notify macro here because, when using secondary
2224 * mmu page tables (such as kvm shadow page tables), we want the
2225 * new page to be mapped directly into the secondary page table.
2226 */
2227 set_pte_at_notify(mm, vmf->address, vmf->pte, entry);
2228 update_mmu_cache(vma, vmf->address, vmf->pte);
2229 if (old_page) {
2230 /*
2231 * Only after switching the pte to the new page may
2232 * we remove the mapcount here. Otherwise another
2233 * process may come and find the rmap count decremented
2234 * before the pte is switched to the new page, and
2235 * "reuse" the old page writing into it while our pte
2236 * here still points into it and can be read by other
2237 * threads.
2238 *
2239 * The critical issue is to order this
2240 * page_remove_rmap with the ptp_clear_flush above.
2241 * Those stores are ordered by (if nothing else,)
2242 * the barrier present in the atomic_add_negative
2243 * in page_remove_rmap.
2244 *
2245 * Then the TLB flush in ptep_clear_flush ensures that
2246 * no process can access the old page before the
2247 * decremented mapcount is visible. And the old page
2248 * cannot be reused until after the decremented
2249 * mapcount is visible. So transitively, TLBs to
2250 * old page will be flushed before it can be reused.
2251 */
2252 page_remove_rmap(old_page, false);
2253 }
2254
2255 /* Free the old page.. */
2256 new_page = old_page;
2257 page_copied = 1;
2258 } else {
2259 mem_cgroup_cancel_charge(new_page, memcg, false);
2260 }
2261
2262 if (new_page)
2263 put_page(new_page);
2264
2265 pte_unmap_unlock(vmf->pte, vmf->ptl);
2266 mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end);
2267 if (old_page) {
2268 /*
2269 * Don't let another task, with possibly unlocked vma,
2270 * keep the mlocked page.
2271 */
2272 if (page_copied && (vma->vm_flags & VM_LOCKED)) {
2273 lock_page(old_page); /* LRU manipulation */
2274 if (PageMlocked(old_page))
2275 munlock_vma_page(old_page);
2276 unlock_page(old_page);
2277 }
2278 put_page(old_page);
2279 }
2280 return page_copied ? VM_FAULT_WRITE : 0;
2281 oom_free_new:
2282 put_page(new_page);
2283 oom:
2284 if (old_page)
2285 put_page(old_page);
2286 return VM_FAULT_OOM;
2287 }
2288
2289 /**
2290 * finish_mkwrite_fault - finish page fault for a shared mapping, making PTE
2291 * writeable once the page is prepared
2292 *
2293 * @vmf: structure describing the fault
2294 *
2295 * This function handles all that is needed to finish a write page fault in a
2296 * shared mapping due to PTE being read-only once the mapped page is prepared.
2297 * It handles locking of PTE and modifying it. The function returns
2298 * VM_FAULT_WRITE on success, 0 when PTE got changed before we acquired PTE
2299 * lock.
2300 *
2301 * The function expects the page to be locked or other protection against
2302 * concurrent faults / writeback (such as DAX radix tree locks).
2303 */
2304 int finish_mkwrite_fault(struct vm_fault *vmf)
2305 {
2306 WARN_ON_ONCE(!(vmf->vma->vm_flags & VM_SHARED));
2307 vmf->pte = pte_offset_map_lock(vmf->vma->vm_mm, vmf->pmd, vmf->address,
2308 &vmf->ptl);
2309 /*
2310 * We might have raced with another page fault while we released the
2311 * pte_offset_map_lock.
2312 */
2313 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2314 pte_unmap_unlock(vmf->pte, vmf->ptl);
2315 return VM_FAULT_NOPAGE;
2316 }
2317 wp_page_reuse(vmf);
2318 return 0;
2319 }
2320
2321 /*
2322 * Handle write page faults for VM_MIXEDMAP or VM_PFNMAP for a VM_SHARED
2323 * mapping
2324 */
2325 static int wp_pfn_shared(struct vm_fault *vmf)
2326 {
2327 struct vm_area_struct *vma = vmf->vma;
2328
2329 if (vma->vm_ops && vma->vm_ops->pfn_mkwrite) {
2330 int ret;
2331
2332 pte_unmap_unlock(vmf->pte, vmf->ptl);
2333 vmf->flags |= FAULT_FLAG_MKWRITE;
2334 ret = vma->vm_ops->pfn_mkwrite(vmf);
2335 if (ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))
2336 return ret;
2337 return finish_mkwrite_fault(vmf);
2338 }
2339 wp_page_reuse(vmf);
2340 return VM_FAULT_WRITE;
2341 }
2342
2343 static int wp_page_shared(struct vm_fault *vmf)
2344 __releases(vmf->ptl)
2345 {
2346 struct vm_area_struct *vma = vmf->vma;
2347
2348 get_page(vmf->page);
2349
2350 if (vma->vm_ops && vma->vm_ops->page_mkwrite) {
2351 int tmp;
2352
2353 pte_unmap_unlock(vmf->pte, vmf->ptl);
2354 tmp = do_page_mkwrite(vmf);
2355 if (unlikely(!tmp || (tmp &
2356 (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
2357 put_page(vmf->page);
2358 return tmp;
2359 }
2360 tmp = finish_mkwrite_fault(vmf);
2361 if (unlikely(tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) {
2362 unlock_page(vmf->page);
2363 put_page(vmf->page);
2364 return tmp;
2365 }
2366 } else {
2367 wp_page_reuse(vmf);
2368 lock_page(vmf->page);
2369 }
2370 fault_dirty_shared_page(vma, vmf->page);
2371 put_page(vmf->page);
2372
2373 return VM_FAULT_WRITE;
2374 }
2375
2376 /*
2377 * This routine handles present pages, when users try to write
2378 * to a shared page. It is done by copying the page to a new address
2379 * and decrementing the shared-page counter for the old page.
2380 *
2381 * Note that this routine assumes that the protection checks have been
2382 * done by the caller (the low-level page fault routine in most cases).
2383 * Thus we can safely just mark it writable once we've done any necessary
2384 * COW.
2385 *
2386 * We also mark the page dirty at this point even though the page will
2387 * change only once the write actually happens. This avoids a few races,
2388 * and potentially makes it more efficient.
2389 *
2390 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2391 * but allow concurrent faults), with pte both mapped and locked.
2392 * We return with mmap_sem still held, but pte unmapped and unlocked.
2393 */
2394 static int do_wp_page(struct vm_fault *vmf)
2395 __releases(vmf->ptl)
2396 {
2397 struct vm_area_struct *vma = vmf->vma;
2398
2399 vmf->page = vm_normal_page(vma, vmf->address, vmf->orig_pte);
2400 if (!vmf->page) {
2401 /*
2402 * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a
2403 * VM_PFNMAP VMA.
2404 *
2405 * We should not cow pages in a shared writeable mapping.
2406 * Just mark the pages writable and/or call ops->pfn_mkwrite.
2407 */
2408 if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2409 (VM_WRITE|VM_SHARED))
2410 return wp_pfn_shared(vmf);
2411
2412 pte_unmap_unlock(vmf->pte, vmf->ptl);
2413 return wp_page_copy(vmf);
2414 }
2415
2416 /*
2417 * Take out anonymous pages first, anonymous shared vmas are
2418 * not dirty accountable.
2419 */
2420 if (PageAnon(vmf->page) && !PageKsm(vmf->page)) {
2421 int total_mapcount;
2422 if (!trylock_page(vmf->page)) {
2423 get_page(vmf->page);
2424 pte_unmap_unlock(vmf->pte, vmf->ptl);
2425 lock_page(vmf->page);
2426 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2427 vmf->address, &vmf->ptl);
2428 if (!pte_same(*vmf->pte, vmf->orig_pte)) {
2429 unlock_page(vmf->page);
2430 pte_unmap_unlock(vmf->pte, vmf->ptl);
2431 put_page(vmf->page);
2432 return 0;
2433 }
2434 put_page(vmf->page);
2435 }
2436 if (reuse_swap_page(vmf->page, &total_mapcount)) {
2437 if (total_mapcount == 1) {
2438 /*
2439 * The page is all ours. Move it to
2440 * our anon_vma so the rmap code will
2441 * not search our parent or siblings.
2442 * Protected against the rmap code by
2443 * the page lock.
2444 */
2445 page_move_anon_rmap(vmf->page, vma);
2446 }
2447 unlock_page(vmf->page);
2448 wp_page_reuse(vmf);
2449 return VM_FAULT_WRITE;
2450 }
2451 unlock_page(vmf->page);
2452 } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) ==
2453 (VM_WRITE|VM_SHARED))) {
2454 return wp_page_shared(vmf);
2455 }
2456
2457 /*
2458 * Ok, we need to copy. Oh, well..
2459 */
2460 get_page(vmf->page);
2461
2462 pte_unmap_unlock(vmf->pte, vmf->ptl);
2463 return wp_page_copy(vmf);
2464 }
2465
2466 static void unmap_mapping_range_vma(struct vm_area_struct *vma,
2467 unsigned long start_addr, unsigned long end_addr,
2468 struct zap_details *details)
2469 {
2470 zap_page_range_single(vma, start_addr, end_addr - start_addr, details);
2471 }
2472
2473 static inline void unmap_mapping_range_tree(struct rb_root *root,
2474 struct zap_details *details)
2475 {
2476 struct vm_area_struct *vma;
2477 pgoff_t vba, vea, zba, zea;
2478
2479 vma_interval_tree_foreach(vma, root,
2480 details->first_index, details->last_index) {
2481
2482 vba = vma->vm_pgoff;
2483 vea = vba + vma_pages(vma) - 1;
2484 zba = details->first_index;
2485 if (zba < vba)
2486 zba = vba;
2487 zea = details->last_index;
2488 if (zea > vea)
2489 zea = vea;
2490
2491 unmap_mapping_range_vma(vma,
2492 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
2493 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
2494 details);
2495 }
2496 }
2497
2498 /**
2499 * unmap_mapping_range - unmap the portion of all mmaps in the specified
2500 * address_space corresponding to the specified page range in the underlying
2501 * file.
2502 *
2503 * @mapping: the address space containing mmaps to be unmapped.
2504 * @holebegin: byte in first page to unmap, relative to the start of
2505 * the underlying file. This will be rounded down to a PAGE_SIZE
2506 * boundary. Note that this is different from truncate_pagecache(), which
2507 * must keep the partial page. In contrast, we must get rid of
2508 * partial pages.
2509 * @holelen: size of prospective hole in bytes. This will be rounded
2510 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
2511 * end of the file.
2512 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
2513 * but 0 when invalidating pagecache, don't throw away private data.
2514 */
2515 void unmap_mapping_range(struct address_space *mapping,
2516 loff_t const holebegin, loff_t const holelen, int even_cows)
2517 {
2518 struct zap_details details = { };
2519 pgoff_t hba = holebegin >> PAGE_SHIFT;
2520 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2521
2522 /* Check for overflow. */
2523 if (sizeof(holelen) > sizeof(hlen)) {
2524 long long holeend =
2525 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
2526 if (holeend & ~(long long)ULONG_MAX)
2527 hlen = ULONG_MAX - hba + 1;
2528 }
2529
2530 details.check_mapping = even_cows ? NULL : mapping;
2531 details.first_index = hba;
2532 details.last_index = hba + hlen - 1;
2533 if (details.last_index < details.first_index)
2534 details.last_index = ULONG_MAX;
2535
2536 i_mmap_lock_write(mapping);
2537 if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap)))
2538 unmap_mapping_range_tree(&mapping->i_mmap, &details);
2539 i_mmap_unlock_write(mapping);
2540 }
2541 EXPORT_SYMBOL(unmap_mapping_range);
2542
2543 /*
2544 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2545 * but allow concurrent faults), and pte mapped but not yet locked.
2546 * We return with pte unmapped and unlocked.
2547 *
2548 * We return with the mmap_sem locked or unlocked in the same cases
2549 * as does filemap_fault().
2550 */
2551 int do_swap_page(struct vm_fault *vmf)
2552 {
2553 struct vm_area_struct *vma = vmf->vma;
2554 struct page *page, *swapcache;
2555 struct mem_cgroup *memcg;
2556 swp_entry_t entry;
2557 pte_t pte;
2558 int locked;
2559 int exclusive = 0;
2560 int ret = 0;
2561
2562 if (!pte_unmap_same(vma->vm_mm, vmf->pmd, vmf->pte, vmf->orig_pte))
2563 goto out;
2564
2565 entry = pte_to_swp_entry(vmf->orig_pte);
2566 if (unlikely(non_swap_entry(entry))) {
2567 if (is_migration_entry(entry)) {
2568 migration_entry_wait(vma->vm_mm, vmf->pmd,
2569 vmf->address);
2570 } else if (is_hwpoison_entry(entry)) {
2571 ret = VM_FAULT_HWPOISON;
2572 } else {
2573 print_bad_pte(vma, vmf->address, vmf->orig_pte, NULL);
2574 ret = VM_FAULT_SIGBUS;
2575 }
2576 goto out;
2577 }
2578 delayacct_set_flag(DELAYACCT_PF_SWAPIN);
2579 page = lookup_swap_cache(entry);
2580 if (!page) {
2581 page = swapin_readahead(entry, GFP_HIGHUSER_MOVABLE, vma,
2582 vmf->address);
2583 if (!page) {
2584 /*
2585 * Back out if somebody else faulted in this pte
2586 * while we released the pte lock.
2587 */
2588 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2589 vmf->address, &vmf->ptl);
2590 if (likely(pte_same(*vmf->pte, vmf->orig_pte)))
2591 ret = VM_FAULT_OOM;
2592 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2593 goto unlock;
2594 }
2595
2596 /* Had to read the page from swap area: Major fault */
2597 ret = VM_FAULT_MAJOR;
2598 count_vm_event(PGMAJFAULT);
2599 mem_cgroup_count_vm_event(vma->vm_mm, PGMAJFAULT);
2600 } else if (PageHWPoison(page)) {
2601 /*
2602 * hwpoisoned dirty swapcache pages are kept for killing
2603 * owner processes (which may be unknown at hwpoison time)
2604 */
2605 ret = VM_FAULT_HWPOISON;
2606 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2607 swapcache = page;
2608 goto out_release;
2609 }
2610
2611 swapcache = page;
2612 locked = lock_page_or_retry(page, vma->vm_mm, vmf->flags);
2613
2614 delayacct_clear_flag(DELAYACCT_PF_SWAPIN);
2615 if (!locked) {
2616 ret |= VM_FAULT_RETRY;
2617 goto out_release;
2618 }
2619
2620 /*
2621 * Make sure try_to_free_swap or reuse_swap_page or swapoff did not
2622 * release the swapcache from under us. The page pin, and pte_same
2623 * test below, are not enough to exclude that. Even if it is still
2624 * swapcache, we need to check that the page's swap has not changed.
2625 */
2626 if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val))
2627 goto out_page;
2628
2629 page = ksm_might_need_to_copy(page, vma, vmf->address);
2630 if (unlikely(!page)) {
2631 ret = VM_FAULT_OOM;
2632 page = swapcache;
2633 goto out_page;
2634 }
2635
2636 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL,
2637 &memcg, false)) {
2638 ret = VM_FAULT_OOM;
2639 goto out_page;
2640 }
2641
2642 /*
2643 * Back out if somebody else already faulted in this pte.
2644 */
2645 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2646 &vmf->ptl);
2647 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte)))
2648 goto out_nomap;
2649
2650 if (unlikely(!PageUptodate(page))) {
2651 ret = VM_FAULT_SIGBUS;
2652 goto out_nomap;
2653 }
2654
2655 /*
2656 * The page isn't present yet, go ahead with the fault.
2657 *
2658 * Be careful about the sequence of operations here.
2659 * To get its accounting right, reuse_swap_page() must be called
2660 * while the page is counted on swap but not yet in mapcount i.e.
2661 * before page_add_anon_rmap() and swap_free(); try_to_free_swap()
2662 * must be called after the swap_free(), or it will never succeed.
2663 */
2664
2665 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2666 dec_mm_counter_fast(vma->vm_mm, MM_SWAPENTS);
2667 pte = mk_pte(page, vma->vm_page_prot);
2668 if ((vmf->flags & FAULT_FLAG_WRITE) && reuse_swap_page(page, NULL)) {
2669 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2670 vmf->flags &= ~FAULT_FLAG_WRITE;
2671 ret |= VM_FAULT_WRITE;
2672 exclusive = RMAP_EXCLUSIVE;
2673 }
2674 flush_icache_page(vma, page);
2675 if (pte_swp_soft_dirty(vmf->orig_pte))
2676 pte = pte_mksoft_dirty(pte);
2677 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, pte);
2678 vmf->orig_pte = pte;
2679 if (page == swapcache) {
2680 do_page_add_anon_rmap(page, vma, vmf->address, exclusive);
2681 mem_cgroup_commit_charge(page, memcg, true, false);
2682 activate_page(page);
2683 } else { /* ksm created a completely new copy */
2684 page_add_new_anon_rmap(page, vma, vmf->address, false);
2685 mem_cgroup_commit_charge(page, memcg, false, false);
2686 lru_cache_add_active_or_unevictable(page, vma);
2687 }
2688
2689 swap_free(entry);
2690 if (mem_cgroup_swap_full(page) ||
2691 (vma->vm_flags & VM_LOCKED) || PageMlocked(page))
2692 try_to_free_swap(page);
2693 unlock_page(page);
2694 if (page != swapcache) {
2695 /*
2696 * Hold the lock to avoid the swap entry to be reused
2697 * until we take the PT lock for the pte_same() check
2698 * (to avoid false positives from pte_same). For
2699 * further safety release the lock after the swap_free
2700 * so that the swap count won't change under a
2701 * parallel locked swapcache.
2702 */
2703 unlock_page(swapcache);
2704 put_page(swapcache);
2705 }
2706
2707 if (vmf->flags & FAULT_FLAG_WRITE) {
2708 ret |= do_wp_page(vmf);
2709 if (ret & VM_FAULT_ERROR)
2710 ret &= VM_FAULT_ERROR;
2711 goto out;
2712 }
2713
2714 /* No need to invalidate - it was non-present before */
2715 update_mmu_cache(vma, vmf->address, vmf->pte);
2716 unlock:
2717 pte_unmap_unlock(vmf->pte, vmf->ptl);
2718 out:
2719 return ret;
2720 out_nomap:
2721 mem_cgroup_cancel_charge(page, memcg, false);
2722 pte_unmap_unlock(vmf->pte, vmf->ptl);
2723 out_page:
2724 unlock_page(page);
2725 out_release:
2726 put_page(page);
2727 if (page != swapcache) {
2728 unlock_page(swapcache);
2729 put_page(swapcache);
2730 }
2731 return ret;
2732 }
2733
2734 /*
2735 * This is like a special single-page "expand_{down|up}wards()",
2736 * except we must first make sure that 'address{-|+}PAGE_SIZE'
2737 * doesn't hit another vma.
2738 */
2739 static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address)
2740 {
2741 address &= PAGE_MASK;
2742 if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) {
2743 struct vm_area_struct *prev = vma->vm_prev;
2744
2745 /*
2746 * Is there a mapping abutting this one below?
2747 *
2748 * That's only ok if it's the same stack mapping
2749 * that has gotten split..
2750 */
2751 if (prev && prev->vm_end == address)
2752 return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM;
2753
2754 return expand_downwards(vma, address - PAGE_SIZE);
2755 }
2756 if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) {
2757 struct vm_area_struct *next = vma->vm_next;
2758
2759 /* As VM_GROWSDOWN but s/below/above/ */
2760 if (next && next->vm_start == address + PAGE_SIZE)
2761 return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM;
2762
2763 return expand_upwards(vma, address + PAGE_SIZE);
2764 }
2765 return 0;
2766 }
2767
2768 /*
2769 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2770 * but allow concurrent faults), and pte mapped but not yet locked.
2771 * We return with mmap_sem still held, but pte unmapped and unlocked.
2772 */
2773 static int do_anonymous_page(struct vm_fault *vmf)
2774 {
2775 struct vm_area_struct *vma = vmf->vma;
2776 struct mem_cgroup *memcg;
2777 struct page *page;
2778 pte_t entry;
2779
2780 /* File mapping without ->vm_ops ? */
2781 if (vma->vm_flags & VM_SHARED)
2782 return VM_FAULT_SIGBUS;
2783
2784 /* Check if we need to add a guard page to the stack */
2785 if (check_stack_guard_page(vma, vmf->address) < 0)
2786 return VM_FAULT_SIGSEGV;
2787
2788 /*
2789 * Use pte_alloc() instead of pte_alloc_map(). We can't run
2790 * pte_offset_map() on pmds where a huge pmd might be created
2791 * from a different thread.
2792 *
2793 * pte_alloc_map() is safe to use under down_write(mmap_sem) or when
2794 * parallel threads are excluded by other means.
2795 *
2796 * Here we only have down_read(mmap_sem).
2797 */
2798 if (pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))
2799 return VM_FAULT_OOM;
2800
2801 /* See the comment in pte_alloc_one_map() */
2802 if (unlikely(pmd_trans_unstable(vmf->pmd)))
2803 return 0;
2804
2805 /* Use the zero-page for reads */
2806 if (!(vmf->flags & FAULT_FLAG_WRITE) &&
2807 !mm_forbids_zeropage(vma->vm_mm)) {
2808 entry = pte_mkspecial(pfn_pte(my_zero_pfn(vmf->address),
2809 vma->vm_page_prot));
2810 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
2811 vmf->address, &vmf->ptl);
2812 if (!pte_none(*vmf->pte))
2813 goto unlock;
2814 /* Deliver the page fault to userland, check inside PT lock */
2815 if (userfaultfd_missing(vma)) {
2816 pte_unmap_unlock(vmf->pte, vmf->ptl);
2817 return handle_userfault(vmf, VM_UFFD_MISSING);
2818 }
2819 goto setpte;
2820 }
2821
2822 /* Allocate our own private page. */
2823 if (unlikely(anon_vma_prepare(vma)))
2824 goto oom;
2825 page = alloc_zeroed_user_highpage_movable(vma, vmf->address);
2826 if (!page)
2827 goto oom;
2828
2829 if (mem_cgroup_try_charge(page, vma->vm_mm, GFP_KERNEL, &memcg, false))
2830 goto oom_free_page;
2831
2832 /*
2833 * The memory barrier inside __SetPageUptodate makes sure that
2834 * preceeding stores to the page contents become visible before
2835 * the set_pte_at() write.
2836 */
2837 __SetPageUptodate(page);
2838
2839 entry = mk_pte(page, vma->vm_page_prot);
2840 if (vma->vm_flags & VM_WRITE)
2841 entry = pte_mkwrite(pte_mkdirty(entry));
2842
2843 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2844 &vmf->ptl);
2845 if (!pte_none(*vmf->pte))
2846 goto release;
2847
2848 /* Deliver the page fault to userland, check inside PT lock */
2849 if (userfaultfd_missing(vma)) {
2850 pte_unmap_unlock(vmf->pte, vmf->ptl);
2851 mem_cgroup_cancel_charge(page, memcg, false);
2852 put_page(page);
2853 return handle_userfault(vmf, VM_UFFD_MISSING);
2854 }
2855
2856 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
2857 page_add_new_anon_rmap(page, vma, vmf->address, false);
2858 mem_cgroup_commit_charge(page, memcg, false, false);
2859 lru_cache_add_active_or_unevictable(page, vma);
2860 setpte:
2861 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
2862
2863 /* No need to invalidate - it was non-present before */
2864 update_mmu_cache(vma, vmf->address, vmf->pte);
2865 unlock:
2866 pte_unmap_unlock(vmf->pte, vmf->ptl);
2867 return 0;
2868 release:
2869 mem_cgroup_cancel_charge(page, memcg, false);
2870 put_page(page);
2871 goto unlock;
2872 oom_free_page:
2873 put_page(page);
2874 oom:
2875 return VM_FAULT_OOM;
2876 }
2877
2878 /*
2879 * The mmap_sem must have been held on entry, and may have been
2880 * released depending on flags and vma->vm_ops->fault() return value.
2881 * See filemap_fault() and __lock_page_retry().
2882 */
2883 static int __do_fault(struct vm_fault *vmf)
2884 {
2885 struct vm_area_struct *vma = vmf->vma;
2886 int ret;
2887
2888 ret = vma->vm_ops->fault(vmf);
2889 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY |
2890 VM_FAULT_DONE_COW)))
2891 return ret;
2892
2893 if (unlikely(PageHWPoison(vmf->page))) {
2894 if (ret & VM_FAULT_LOCKED)
2895 unlock_page(vmf->page);
2896 put_page(vmf->page);
2897 vmf->page = NULL;
2898 return VM_FAULT_HWPOISON;
2899 }
2900
2901 if (unlikely(!(ret & VM_FAULT_LOCKED)))
2902 lock_page(vmf->page);
2903 else
2904 VM_BUG_ON_PAGE(!PageLocked(vmf->page), vmf->page);
2905
2906 return ret;
2907 }
2908
2909 static int pte_alloc_one_map(struct vm_fault *vmf)
2910 {
2911 struct vm_area_struct *vma = vmf->vma;
2912
2913 if (!pmd_none(*vmf->pmd))
2914 goto map_pte;
2915 if (vmf->prealloc_pte) {
2916 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
2917 if (unlikely(!pmd_none(*vmf->pmd))) {
2918 spin_unlock(vmf->ptl);
2919 goto map_pte;
2920 }
2921
2922 atomic_long_inc(&vma->vm_mm->nr_ptes);
2923 pmd_populate(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2924 spin_unlock(vmf->ptl);
2925 vmf->prealloc_pte = NULL;
2926 } else if (unlikely(pte_alloc(vma->vm_mm, vmf->pmd, vmf->address))) {
2927 return VM_FAULT_OOM;
2928 }
2929 map_pte:
2930 /*
2931 * If a huge pmd materialized under us just retry later. Use
2932 * pmd_trans_unstable() instead of pmd_trans_huge() to ensure the pmd
2933 * didn't become pmd_trans_huge under us and then back to pmd_none, as
2934 * a result of MADV_DONTNEED running immediately after a huge pmd fault
2935 * in a different thread of this mm, in turn leading to a misleading
2936 * pmd_trans_huge() retval. All we have to ensure is that it is a
2937 * regular pmd that we can walk with pte_offset_map() and we can do that
2938 * through an atomic read in C, which is what pmd_trans_unstable()
2939 * provides.
2940 */
2941 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
2942 return VM_FAULT_NOPAGE;
2943
2944 vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd, vmf->address,
2945 &vmf->ptl);
2946 return 0;
2947 }
2948
2949 #ifdef CONFIG_TRANSPARENT_HUGE_PAGECACHE
2950
2951 #define HPAGE_CACHE_INDEX_MASK (HPAGE_PMD_NR - 1)
2952 static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
2953 unsigned long haddr)
2954 {
2955 if (((vma->vm_start >> PAGE_SHIFT) & HPAGE_CACHE_INDEX_MASK) !=
2956 (vma->vm_pgoff & HPAGE_CACHE_INDEX_MASK))
2957 return false;
2958 if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
2959 return false;
2960 return true;
2961 }
2962
2963 static void deposit_prealloc_pte(struct vm_fault *vmf)
2964 {
2965 struct vm_area_struct *vma = vmf->vma;
2966
2967 pgtable_trans_huge_deposit(vma->vm_mm, vmf->pmd, vmf->prealloc_pte);
2968 /*
2969 * We are going to consume the prealloc table,
2970 * count that as nr_ptes.
2971 */
2972 atomic_long_inc(&vma->vm_mm->nr_ptes);
2973 vmf->prealloc_pte = NULL;
2974 }
2975
2976 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
2977 {
2978 struct vm_area_struct *vma = vmf->vma;
2979 bool write = vmf->flags & FAULT_FLAG_WRITE;
2980 unsigned long haddr = vmf->address & HPAGE_PMD_MASK;
2981 pmd_t entry;
2982 int i, ret;
2983
2984 if (!transhuge_vma_suitable(vma, haddr))
2985 return VM_FAULT_FALLBACK;
2986
2987 ret = VM_FAULT_FALLBACK;
2988 page = compound_head(page);
2989
2990 /*
2991 * Archs like ppc64 need additonal space to store information
2992 * related to pte entry. Use the preallocated table for that.
2993 */
2994 if (arch_needs_pgtable_deposit() && !vmf->prealloc_pte) {
2995 vmf->prealloc_pte = pte_alloc_one(vma->vm_mm, vmf->address);
2996 if (!vmf->prealloc_pte)
2997 return VM_FAULT_OOM;
2998 smp_wmb(); /* See comment in __pte_alloc() */
2999 }
3000
3001 vmf->ptl = pmd_lock(vma->vm_mm, vmf->pmd);
3002 if (unlikely(!pmd_none(*vmf->pmd)))
3003 goto out;
3004
3005 for (i = 0; i < HPAGE_PMD_NR; i++)
3006 flush_icache_page(vma, page + i);
3007
3008 entry = mk_huge_pmd(page, vma->vm_page_prot);
3009 if (write)
3010 entry = maybe_pmd_mkwrite(pmd_mkdirty(entry), vma);
3011
3012 add_mm_counter(vma->vm_mm, MM_FILEPAGES, HPAGE_PMD_NR);
3013 page_add_file_rmap(page, true);
3014 /*
3015 * deposit and withdraw with pmd lock held
3016 */
3017 if (arch_needs_pgtable_deposit())
3018 deposit_prealloc_pte(vmf);
3019
3020 set_pmd_at(vma->vm_mm, haddr, vmf->pmd, entry);
3021
3022 update_mmu_cache_pmd(vma, haddr, vmf->pmd);
3023
3024 /* fault is handled */
3025 ret = 0;
3026 count_vm_event(THP_FILE_MAPPED);
3027 out:
3028 spin_unlock(vmf->ptl);
3029 return ret;
3030 }
3031 #else
3032 static int do_set_pmd(struct vm_fault *vmf, struct page *page)
3033 {
3034 BUILD_BUG();
3035 return 0;
3036 }
3037 #endif
3038
3039 /**
3040 * alloc_set_pte - setup new PTE entry for given page and add reverse page
3041 * mapping. If needed, the fucntion allocates page table or use pre-allocated.
3042 *
3043 * @vmf: fault environment
3044 * @memcg: memcg to charge page (only for private mappings)
3045 * @page: page to map
3046 *
3047 * Caller must take care of unlocking vmf->ptl, if vmf->pte is non-NULL on
3048 * return.
3049 *
3050 * Target users are page handler itself and implementations of
3051 * vm_ops->map_pages.
3052 */
3053 int alloc_set_pte(struct vm_fault *vmf, struct mem_cgroup *memcg,
3054 struct page *page)
3055 {
3056 struct vm_area_struct *vma = vmf->vma;
3057 bool write = vmf->flags & FAULT_FLAG_WRITE;
3058 pte_t entry;
3059 int ret;
3060
3061 if (pmd_none(*vmf->pmd) && PageTransCompound(page) &&
3062 IS_ENABLED(CONFIG_TRANSPARENT_HUGE_PAGECACHE)) {
3063 /* THP on COW? */
3064 VM_BUG_ON_PAGE(memcg, page);
3065
3066 ret = do_set_pmd(vmf, page);
3067 if (ret != VM_FAULT_FALLBACK)
3068 return ret;
3069 }
3070
3071 if (!vmf->pte) {
3072 ret = pte_alloc_one_map(vmf);
3073 if (ret)
3074 return ret;
3075 }
3076
3077 /* Re-check under ptl */
3078 if (unlikely(!pte_none(*vmf->pte)))
3079 return VM_FAULT_NOPAGE;
3080
3081 flush_icache_page(vma, page);
3082 entry = mk_pte(page, vma->vm_page_prot);
3083 if (write)
3084 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
3085 /* copy-on-write page */
3086 if (write && !(vma->vm_flags & VM_SHARED)) {
3087 inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES);
3088 page_add_new_anon_rmap(page, vma, vmf->address, false);
3089 mem_cgroup_commit_charge(page, memcg, false, false);
3090 lru_cache_add_active_or_unevictable(page, vma);
3091 } else {
3092 inc_mm_counter_fast(vma->vm_mm, mm_counter_file(page));
3093 page_add_file_rmap(page, false);
3094 }
3095 set_pte_at(vma->vm_mm, vmf->address, vmf->pte, entry);
3096
3097 /* no need to invalidate: a not-present page won't be cached */
3098 update_mmu_cache(vma, vmf->address, vmf->pte);
3099
3100 return 0;
3101 }
3102
3103
3104 /**
3105 * finish_fault - finish page fault once we have prepared the page to fault
3106 *
3107 * @vmf: structure describing the fault
3108 *
3109 * This function handles all that is needed to finish a page fault once the
3110 * page to fault in is prepared. It handles locking of PTEs, inserts PTE for
3111 * given page, adds reverse page mapping, handles memcg charges and LRU
3112 * addition. The function returns 0 on success, VM_FAULT_ code in case of
3113 * error.
3114 *
3115 * The function expects the page to be locked and on success it consumes a
3116 * reference of a page being mapped (for the PTE which maps it).
3117 */
3118 int finish_fault(struct vm_fault *vmf)
3119 {
3120 struct page *page;
3121 int ret;
3122
3123 /* Did we COW the page? */
3124 if ((vmf->flags & FAULT_FLAG_WRITE) &&
3125 !(vmf->vma->vm_flags & VM_SHARED))
3126 page = vmf->cow_page;
3127 else
3128 page = vmf->page;
3129 ret = alloc_set_pte(vmf, vmf->memcg, page);
3130 if (vmf->pte)
3131 pte_unmap_unlock(vmf->pte, vmf->ptl);
3132 return ret;
3133 }
3134
3135 static unsigned long fault_around_bytes __read_mostly =
3136 rounddown_pow_of_two(65536);
3137
3138 #ifdef CONFIG_DEBUG_FS
3139 static int fault_around_bytes_get(void *data, u64 *val)
3140 {
3141 *val = fault_around_bytes;
3142 return 0;
3143 }
3144
3145 /*
3146 * fault_around_pages() and fault_around_mask() expects fault_around_bytes
3147 * rounded down to nearest page order. It's what do_fault_around() expects to
3148 * see.
3149 */
3150 static int fault_around_bytes_set(void *data, u64 val)
3151 {
3152 if (val / PAGE_SIZE > PTRS_PER_PTE)
3153 return -EINVAL;
3154 if (val > PAGE_SIZE)
3155 fault_around_bytes = rounddown_pow_of_two(val);
3156 else
3157 fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */
3158 return 0;
3159 }
3160 DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops,
3161 fault_around_bytes_get, fault_around_bytes_set, "%llu\n");
3162
3163 static int __init fault_around_debugfs(void)
3164 {
3165 void *ret;
3166
3167 ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL,
3168 &fault_around_bytes_fops);
3169 if (!ret)
3170 pr_warn("Failed to create fault_around_bytes in debugfs");
3171 return 0;
3172 }
3173 late_initcall(fault_around_debugfs);
3174 #endif
3175
3176 /*
3177 * do_fault_around() tries to map few pages around the fault address. The hope
3178 * is that the pages will be needed soon and this will lower the number of
3179 * faults to handle.
3180 *
3181 * It uses vm_ops->map_pages() to map the pages, which skips the page if it's
3182 * not ready to be mapped: not up-to-date, locked, etc.
3183 *
3184 * This function is called with the page table lock taken. In the split ptlock
3185 * case the page table lock only protects only those entries which belong to
3186 * the page table corresponding to the fault address.
3187 *
3188 * This function doesn't cross the VMA boundaries, in order to call map_pages()
3189 * only once.
3190 *
3191 * fault_around_pages() defines how many pages we'll try to map.
3192 * do_fault_around() expects it to return a power of two less than or equal to
3193 * PTRS_PER_PTE.
3194 *
3195 * The virtual address of the area that we map is naturally aligned to the
3196 * fault_around_pages() value (and therefore to page order). This way it's
3197 * easier to guarantee that we don't cross page table boundaries.
3198 */
3199 static int do_fault_around(struct vm_fault *vmf)
3200 {
3201 unsigned long address = vmf->address, nr_pages, mask;
3202 pgoff_t start_pgoff = vmf->pgoff;
3203 pgoff_t end_pgoff;
3204 int off, ret = 0;
3205
3206 nr_pages = READ_ONCE(fault_around_bytes) >> PAGE_SHIFT;
3207 mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK;
3208
3209 vmf->address = max(address & mask, vmf->vma->vm_start);
3210 off = ((address - vmf->address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1);
3211 start_pgoff -= off;
3212
3213 /*
3214 * end_pgoff is either end of page table or end of vma
3215 * or fault_around_pages() from start_pgoff, depending what is nearest.
3216 */
3217 end_pgoff = start_pgoff -
3218 ((vmf->address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) +
3219 PTRS_PER_PTE - 1;
3220 end_pgoff = min3(end_pgoff, vma_pages(vmf->vma) + vmf->vma->vm_pgoff - 1,
3221 start_pgoff + nr_pages - 1);
3222
3223 if (pmd_none(*vmf->pmd)) {
3224 vmf->prealloc_pte = pte_alloc_one(vmf->vma->vm_mm,
3225 vmf->address);
3226 if (!vmf->prealloc_pte)
3227 goto out;
3228 smp_wmb(); /* See comment in __pte_alloc() */
3229 }
3230
3231 vmf->vma->vm_ops->map_pages(vmf, start_pgoff, end_pgoff);
3232
3233 /* Huge page is mapped? Page fault is solved */
3234 if (pmd_trans_huge(*vmf->pmd)) {
3235 ret = VM_FAULT_NOPAGE;
3236 goto out;
3237 }
3238
3239 /* ->map_pages() haven't done anything useful. Cold page cache? */
3240 if (!vmf->pte)
3241 goto out;
3242
3243 /* check if the page fault is solved */
3244 vmf->pte -= (vmf->address >> PAGE_SHIFT) - (address >> PAGE_SHIFT);
3245 if (!pte_none(*vmf->pte))
3246 ret = VM_FAULT_NOPAGE;
3247 pte_unmap_unlock(vmf->pte, vmf->ptl);
3248 out:
3249 vmf->address = address;
3250 vmf->pte = NULL;
3251 return ret;
3252 }
3253
3254 static int do_read_fault(struct vm_fault *vmf)
3255 {
3256 struct vm_area_struct *vma = vmf->vma;
3257 int ret = 0;
3258
3259 /*
3260 * Let's call ->map_pages() first and use ->fault() as fallback
3261 * if page by the offset is not ready to be mapped (cold cache or
3262 * something).
3263 */
3264 if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) {
3265 ret = do_fault_around(vmf);
3266 if (ret)
3267 return ret;
3268 }
3269
3270 ret = __do_fault(vmf);
3271 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3272 return ret;
3273
3274 ret |= finish_fault(vmf);
3275 unlock_page(vmf->page);
3276 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3277 put_page(vmf->page);
3278 return ret;
3279 }
3280
3281 static int do_cow_fault(struct vm_fault *vmf)
3282 {
3283 struct vm_area_struct *vma = vmf->vma;
3284 int ret;
3285
3286 if (unlikely(anon_vma_prepare(vma)))
3287 return VM_FAULT_OOM;
3288
3289 vmf->cow_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, vmf->address);
3290 if (!vmf->cow_page)
3291 return VM_FAULT_OOM;
3292
3293 if (mem_cgroup_try_charge(vmf->cow_page, vma->vm_mm, GFP_KERNEL,
3294 &vmf->memcg, false)) {
3295 put_page(vmf->cow_page);
3296 return VM_FAULT_OOM;
3297 }
3298
3299 ret = __do_fault(vmf);
3300 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3301 goto uncharge_out;
3302 if (ret & VM_FAULT_DONE_COW)
3303 return ret;
3304
3305 copy_user_highpage(vmf->cow_page, vmf->page, vmf->address, vma);
3306 __SetPageUptodate(vmf->cow_page);
3307
3308 ret |= finish_fault(vmf);
3309 unlock_page(vmf->page);
3310 put_page(vmf->page);
3311 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3312 goto uncharge_out;
3313 return ret;
3314 uncharge_out:
3315 mem_cgroup_cancel_charge(vmf->cow_page, vmf->memcg, false);
3316 put_page(vmf->cow_page);
3317 return ret;
3318 }
3319
3320 static int do_shared_fault(struct vm_fault *vmf)
3321 {
3322 struct vm_area_struct *vma = vmf->vma;
3323 int ret, tmp;
3324
3325 ret = __do_fault(vmf);
3326 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY)))
3327 return ret;
3328
3329 /*
3330 * Check if the backing address space wants to know that the page is
3331 * about to become writable
3332 */
3333 if (vma->vm_ops->page_mkwrite) {
3334 unlock_page(vmf->page);
3335 tmp = do_page_mkwrite(vmf);
3336 if (unlikely(!tmp ||
3337 (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) {
3338 put_page(vmf->page);
3339 return tmp;
3340 }
3341 }
3342
3343 ret |= finish_fault(vmf);
3344 if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE |
3345 VM_FAULT_RETRY))) {
3346 unlock_page(vmf->page);
3347 put_page(vmf->page);
3348 return ret;
3349 }
3350
3351 fault_dirty_shared_page(vma, vmf->page);
3352 return ret;
3353 }
3354
3355 /*
3356 * We enter with non-exclusive mmap_sem (to exclude vma changes,
3357 * but allow concurrent faults).
3358 * The mmap_sem may have been released depending on flags and our
3359 * return value. See filemap_fault() and __lock_page_or_retry().
3360 */
3361 static int do_fault(struct vm_fault *vmf)
3362 {
3363 struct vm_area_struct *vma = vmf->vma;
3364 int ret;
3365
3366 /* The VMA was not fully populated on mmap() or missing VM_DONTEXPAND */
3367 if (!vma->vm_ops->fault)
3368 ret = VM_FAULT_SIGBUS;
3369 else if (!(vmf->flags & FAULT_FLAG_WRITE))
3370 ret = do_read_fault(vmf);
3371 else if (!(vma->vm_flags & VM_SHARED))
3372 ret = do_cow_fault(vmf);
3373 else
3374 ret = do_shared_fault(vmf);
3375
3376 /* preallocated pagetable is unused: free it */
3377 if (vmf->prealloc_pte) {
3378 pte_free(vma->vm_mm, vmf->prealloc_pte);
3379 vmf->prealloc_pte = NULL;
3380 }
3381 return ret;
3382 }
3383
3384 static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma,
3385 unsigned long addr, int page_nid,
3386 int *flags)
3387 {
3388 get_page(page);
3389
3390 count_vm_numa_event(NUMA_HINT_FAULTS);
3391 if (page_nid == numa_node_id()) {
3392 count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL);
3393 *flags |= TNF_FAULT_LOCAL;
3394 }
3395
3396 return mpol_misplaced(page, vma, addr);
3397 }
3398
3399 static int do_numa_page(struct vm_fault *vmf)
3400 {
3401 struct vm_area_struct *vma = vmf->vma;
3402 struct page *page = NULL;
3403 int page_nid = -1;
3404 int last_cpupid;
3405 int target_nid;
3406 bool migrated = false;
3407 pte_t pte;
3408 bool was_writable = pte_savedwrite(vmf->orig_pte);
3409 int flags = 0;
3410
3411 /*
3412 * The "pte" at this point cannot be used safely without
3413 * validation through pte_unmap_same(). It's of NUMA type but
3414 * the pfn may be screwed if the read is non atomic.
3415 */
3416 vmf->ptl = pte_lockptr(vma->vm_mm, vmf->pmd);
3417 spin_lock(vmf->ptl);
3418 if (unlikely(!pte_same(*vmf->pte, vmf->orig_pte))) {
3419 pte_unmap_unlock(vmf->pte, vmf->ptl);
3420 goto out;
3421 }
3422
3423 /*
3424 * Make it present again, Depending on how arch implementes non
3425 * accessible ptes, some can allow access by kernel mode.
3426 */
3427 pte = ptep_modify_prot_start(vma->vm_mm, vmf->address, vmf->pte);
3428 pte = pte_modify(pte, vma->vm_page_prot);
3429 pte = pte_mkyoung(pte);
3430 if (was_writable)
3431 pte = pte_mkwrite(pte);
3432 ptep_modify_prot_commit(vma->vm_mm, vmf->address, vmf->pte, pte);
3433 update_mmu_cache(vma, vmf->address, vmf->pte);
3434
3435 page = vm_normal_page(vma, vmf->address, pte);
3436 if (!page) {
3437 pte_unmap_unlock(vmf->pte, vmf->ptl);
3438 return 0;
3439 }
3440
3441 /* TODO: handle PTE-mapped THP */
3442 if (PageCompound(page)) {
3443 pte_unmap_unlock(vmf->pte, vmf->ptl);
3444 return 0;
3445 }
3446
3447 /*
3448 * Avoid grouping on RO pages in general. RO pages shouldn't hurt as
3449 * much anyway since they can be in shared cache state. This misses
3450 * the case where a mapping is writable but the process never writes
3451 * to it but pte_write gets cleared during protection updates and
3452 * pte_dirty has unpredictable behaviour between PTE scan updates,
3453 * background writeback, dirty balancing and application behaviour.
3454 */
3455 if (!pte_write(pte))
3456 flags |= TNF_NO_GROUP;
3457
3458 /*
3459 * Flag if the page is shared between multiple address spaces. This
3460 * is later used when determining whether to group tasks together
3461 */
3462 if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED))
3463 flags |= TNF_SHARED;
3464
3465 last_cpupid = page_cpupid_last(page);
3466 page_nid = page_to_nid(page);
3467 target_nid = numa_migrate_prep(page, vma, vmf->address, page_nid,
3468 &flags);
3469 pte_unmap_unlock(vmf->pte, vmf->ptl);
3470 if (target_nid == -1) {
3471 put_page(page);
3472 goto out;
3473 }
3474
3475 /* Migrate to the requested node */
3476 migrated = migrate_misplaced_page(page, vma, target_nid);
3477 if (migrated) {
3478 page_nid = target_nid;
3479 flags |= TNF_MIGRATED;
3480 } else
3481 flags |= TNF_MIGRATE_FAIL;
3482
3483 out:
3484 if (page_nid != -1)
3485 task_numa_fault(last_cpupid, page_nid, 1, flags);
3486 return 0;
3487 }
3488
3489 static int create_huge_pmd(struct vm_fault *vmf)
3490 {
3491 if (vma_is_anonymous(vmf->vma))
3492 return do_huge_pmd_anonymous_page(vmf);
3493 if (vmf->vma->vm_ops->huge_fault)
3494 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3495 return VM_FAULT_FALLBACK;
3496 }
3497
3498 static int wp_huge_pmd(struct vm_fault *vmf, pmd_t orig_pmd)
3499 {
3500 if (vma_is_anonymous(vmf->vma))
3501 return do_huge_pmd_wp_page(vmf, orig_pmd);
3502 if (vmf->vma->vm_ops->huge_fault)
3503 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PMD);
3504
3505 /* COW handled on pte level: split pmd */
3506 VM_BUG_ON_VMA(vmf->vma->vm_flags & VM_SHARED, vmf->vma);
3507 __split_huge_pmd(vmf->vma, vmf->pmd, vmf->address, false, NULL);
3508
3509 return VM_FAULT_FALLBACK;
3510 }
3511
3512 static inline bool vma_is_accessible(struct vm_area_struct *vma)
3513 {
3514 return vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE);
3515 }
3516
3517 static int create_huge_pud(struct vm_fault *vmf)
3518 {
3519 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3520 /* No support for anonymous transparent PUD pages yet */
3521 if (vma_is_anonymous(vmf->vma))
3522 return VM_FAULT_FALLBACK;
3523 if (vmf->vma->vm_ops->huge_fault)
3524 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3525 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3526 return VM_FAULT_FALLBACK;
3527 }
3528
3529 static int wp_huge_pud(struct vm_fault *vmf, pud_t orig_pud)
3530 {
3531 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
3532 /* No support for anonymous transparent PUD pages yet */
3533 if (vma_is_anonymous(vmf->vma))
3534 return VM_FAULT_FALLBACK;
3535 if (vmf->vma->vm_ops->huge_fault)
3536 return vmf->vma->vm_ops->huge_fault(vmf, PE_SIZE_PUD);
3537 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
3538 return VM_FAULT_FALLBACK;
3539 }
3540
3541 /*
3542 * These routines also need to handle stuff like marking pages dirty
3543 * and/or accessed for architectures that don't do it in hardware (most
3544 * RISC architectures). The early dirtying is also good on the i386.
3545 *
3546 * There is also a hook called "update_mmu_cache()" that architectures
3547 * with external mmu caches can use to update those (ie the Sparc or
3548 * PowerPC hashed page tables that act as extended TLBs).
3549 *
3550 * We enter with non-exclusive mmap_sem (to exclude vma changes, but allow
3551 * concurrent faults).
3552 *
3553 * The mmap_sem may have been released depending on flags and our return value.
3554 * See filemap_fault() and __lock_page_or_retry().
3555 */
3556 static int handle_pte_fault(struct vm_fault *vmf)
3557 {
3558 pte_t entry;
3559
3560 if (unlikely(pmd_none(*vmf->pmd))) {
3561 /*
3562 * Leave __pte_alloc() until later: because vm_ops->fault may
3563 * want to allocate huge page, and if we expose page table
3564 * for an instant, it will be difficult to retract from
3565 * concurrent faults and from rmap lookups.
3566 */
3567 vmf->pte = NULL;
3568 } else {
3569 /* See comment in pte_alloc_one_map() */
3570 if (pmd_trans_unstable(vmf->pmd) || pmd_devmap(*vmf->pmd))
3571 return 0;
3572 /*
3573 * A regular pmd is established and it can't morph into a huge
3574 * pmd from under us anymore at this point because we hold the
3575 * mmap_sem read mode and khugepaged takes it in write mode.
3576 * So now it's safe to run pte_offset_map().
3577 */
3578 vmf->pte = pte_offset_map(vmf->pmd, vmf->address);
3579 vmf->orig_pte = *vmf->pte;
3580
3581 /*
3582 * some architectures can have larger ptes than wordsize,
3583 * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and
3584 * CONFIG_32BIT=y, so READ_ONCE or ACCESS_ONCE cannot guarantee
3585 * atomic accesses. The code below just needs a consistent
3586 * view for the ifs and we later double check anyway with the
3587 * ptl lock held. So here a barrier will do.
3588 */
3589 barrier();
3590 if (pte_none(vmf->orig_pte)) {
3591 pte_unmap(vmf->pte);
3592 vmf->pte = NULL;
3593 }
3594 }
3595
3596 if (!vmf->pte) {
3597 if (vma_is_anonymous(vmf->vma))
3598 return do_anonymous_page(vmf);
3599 else
3600 return do_fault(vmf);
3601 }
3602
3603 if (!pte_present(vmf->orig_pte))
3604 return do_swap_page(vmf);
3605
3606 if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
3607 return do_numa_page(vmf);
3608
3609 vmf->ptl = pte_lockptr(vmf->vma->vm_mm, vmf->pmd);
3610 spin_lock(vmf->ptl);
3611 entry = vmf->orig_pte;
3612 if (unlikely(!pte_same(*vmf->pte, entry)))
3613 goto unlock;
3614 if (vmf->flags & FAULT_FLAG_WRITE) {
3615 if (!pte_write(entry))
3616 return do_wp_page(vmf);
3617 entry = pte_mkdirty(entry);
3618 }
3619 entry = pte_mkyoung(entry);
3620 if (ptep_set_access_flags(vmf->vma, vmf->address, vmf->pte, entry,
3621 vmf->flags & FAULT_FLAG_WRITE)) {
3622 update_mmu_cache(vmf->vma, vmf->address, vmf->pte);
3623 } else {
3624 /*
3625 * This is needed only for protection faults but the arch code
3626 * is not yet telling us if this is a protection fault or not.
3627 * This still avoids useless tlb flushes for .text page faults
3628 * with threads.
3629 */
3630 if (vmf->flags & FAULT_FLAG_WRITE)
3631 flush_tlb_fix_spurious_fault(vmf->vma, vmf->address);
3632 }
3633 unlock:
3634 pte_unmap_unlock(vmf->pte, vmf->ptl);
3635 return 0;
3636 }
3637
3638 /*
3639 * By the time we get here, we already hold the mm semaphore
3640 *
3641 * The mmap_sem may have been released depending on flags and our
3642 * return value. See filemap_fault() and __lock_page_or_retry().
3643 */
3644 static int __handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3645 unsigned int flags)
3646 {
3647 struct vm_fault vmf = {
3648 .vma = vma,
3649 .address = address & PAGE_MASK,
3650 .flags = flags,
3651 .pgoff = linear_page_index(vma, address),
3652 .gfp_mask = __get_fault_gfp_mask(vma),
3653 };
3654 struct mm_struct *mm = vma->vm_mm;
3655 pgd_t *pgd;
3656 int ret;
3657
3658 pgd = pgd_offset(mm, address);
3659
3660 vmf.pud = pud_alloc(mm, pgd, address);
3661 if (!vmf.pud)
3662 return VM_FAULT_OOM;
3663 if (pud_none(*vmf.pud) && transparent_hugepage_enabled(vma)) {
3664 ret = create_huge_pud(&vmf);
3665 if (!(ret & VM_FAULT_FALLBACK))
3666 return ret;
3667 } else {
3668 pud_t orig_pud = *vmf.pud;
3669
3670 barrier();
3671 if (pud_trans_huge(orig_pud) || pud_devmap(orig_pud)) {
3672 unsigned int dirty = flags & FAULT_FLAG_WRITE;
3673
3674 /* NUMA case for anonymous PUDs would go here */
3675
3676 if (dirty && !pud_write(orig_pud)) {
3677 ret = wp_huge_pud(&vmf, orig_pud);
3678 if (!(ret & VM_FAULT_FALLBACK))
3679 return ret;
3680 } else {
3681 huge_pud_set_accessed(&vmf, orig_pud);
3682 return 0;
3683 }
3684 }
3685 }
3686
3687 vmf.pmd = pmd_alloc(mm, vmf.pud, address);
3688 if (!vmf.pmd)
3689 return VM_FAULT_OOM;
3690 if (pmd_none(*vmf.pmd) && transparent_hugepage_enabled(vma)) {
3691 ret = create_huge_pmd(&vmf);
3692 if (!(ret & VM_FAULT_FALLBACK))
3693 return ret;
3694 } else {
3695 pmd_t orig_pmd = *vmf.pmd;
3696
3697 barrier();
3698 if (pmd_trans_huge(orig_pmd) || pmd_devmap(orig_pmd)) {
3699 if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
3700 return do_huge_pmd_numa_page(&vmf, orig_pmd);
3701
3702 if ((vmf.flags & FAULT_FLAG_WRITE) &&
3703 !pmd_write(orig_pmd)) {
3704 ret = wp_huge_pmd(&vmf, orig_pmd);
3705 if (!(ret & VM_FAULT_FALLBACK))
3706 return ret;
3707 } else {
3708 huge_pmd_set_accessed(&vmf, orig_pmd);
3709 return 0;
3710 }
3711 }
3712 }
3713
3714 return handle_pte_fault(&vmf);
3715 }
3716
3717 /*
3718 * By the time we get here, we already hold the mm semaphore
3719 *
3720 * The mmap_sem may have been released depending on flags and our
3721 * return value. See filemap_fault() and __lock_page_or_retry().
3722 */
3723 int handle_mm_fault(struct vm_area_struct *vma, unsigned long address,
3724 unsigned int flags)
3725 {
3726 int ret;
3727
3728 __set_current_state(TASK_RUNNING);
3729
3730 count_vm_event(PGFAULT);
3731 mem_cgroup_count_vm_event(vma->vm_mm, PGFAULT);
3732
3733 /* do counter updates before entering really critical section. */
3734 check_sync_rss_stat(current);
3735
3736 /*
3737 * Enable the memcg OOM handling for faults triggered in user
3738 * space. Kernel faults are handled more gracefully.
3739 */
3740 if (flags & FAULT_FLAG_USER)
3741 mem_cgroup_oom_enable();
3742
3743 if (!arch_vma_access_permitted(vma, flags & FAULT_FLAG_WRITE,
3744 flags & FAULT_FLAG_INSTRUCTION,
3745 flags & FAULT_FLAG_REMOTE))
3746 return VM_FAULT_SIGSEGV;
3747
3748 if (unlikely(is_vm_hugetlb_page(vma)))
3749 ret = hugetlb_fault(vma->vm_mm, vma, address, flags);
3750 else
3751 ret = __handle_mm_fault(vma, address, flags);
3752
3753 if (flags & FAULT_FLAG_USER) {
3754 mem_cgroup_oom_disable();
3755 /*
3756 * The task may have entered a memcg OOM situation but
3757 * if the allocation error was handled gracefully (no
3758 * VM_FAULT_OOM), there is no need to kill anything.
3759 * Just clean up the OOM state peacefully.
3760 */
3761 if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM))
3762 mem_cgroup_oom_synchronize(false);
3763 }
3764
3765 /*
3766 * This mm has been already reaped by the oom reaper and so the
3767 * refault cannot be trusted in general. Anonymous refaults would
3768 * lose data and give a zero page instead e.g. This is especially
3769 * problem for use_mm() because regular tasks will just die and
3770 * the corrupted data will not be visible anywhere while kthread
3771 * will outlive the oom victim and potentially propagate the date
3772 * further.
3773 */
3774 if (unlikely((current->flags & PF_KTHREAD) && !(ret & VM_FAULT_ERROR)
3775 && test_bit(MMF_UNSTABLE, &vma->vm_mm->flags)))
3776 ret = VM_FAULT_SIGBUS;
3777
3778 return ret;
3779 }
3780 EXPORT_SYMBOL_GPL(handle_mm_fault);
3781
3782 #ifndef __PAGETABLE_PUD_FOLDED
3783 /*
3784 * Allocate page upper directory.
3785 * We've already handled the fast-path in-line.
3786 */
3787 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
3788 {
3789 pud_t *new = pud_alloc_one(mm, address);
3790 if (!new)
3791 return -ENOMEM;
3792
3793 smp_wmb(); /* See comment in __pte_alloc */
3794
3795 spin_lock(&mm->page_table_lock);
3796 if (pgd_present(*pgd)) /* Another has populated it */
3797 pud_free(mm, new);
3798 else
3799 pgd_populate(mm, pgd, new);
3800 spin_unlock(&mm->page_table_lock);
3801 return 0;
3802 }
3803 #endif /* __PAGETABLE_PUD_FOLDED */
3804
3805 #ifndef __PAGETABLE_PMD_FOLDED
3806 /*
3807 * Allocate page middle directory.
3808 * We've already handled the fast-path in-line.
3809 */
3810 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
3811 {
3812 spinlock_t *ptl;
3813 pmd_t *new = pmd_alloc_one(mm, address);
3814 if (!new)
3815 return -ENOMEM;
3816
3817 smp_wmb(); /* See comment in __pte_alloc */
3818
3819 ptl = pud_lock(mm, pud);
3820 #ifndef __ARCH_HAS_4LEVEL_HACK
3821 if (!pud_present(*pud)) {
3822 mm_inc_nr_pmds(mm);
3823 pud_populate(mm, pud, new);
3824 } else /* Another has populated it */
3825 pmd_free(mm, new);
3826 #else
3827 if (!pgd_present(*pud)) {
3828 mm_inc_nr_pmds(mm);
3829 pgd_populate(mm, pud, new);
3830 } else /* Another has populated it */
3831 pmd_free(mm, new);
3832 #endif /* __ARCH_HAS_4LEVEL_HACK */
3833 spin_unlock(ptl);
3834 return 0;
3835 }
3836 #endif /* __PAGETABLE_PMD_FOLDED */
3837
3838 static int __follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3839 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3840 {
3841 pgd_t *pgd;
3842 pud_t *pud;
3843 pmd_t *pmd;
3844 pte_t *ptep;
3845
3846 pgd = pgd_offset(mm, address);
3847 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
3848 goto out;
3849
3850 pud = pud_offset(pgd, address);
3851 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
3852 goto out;
3853
3854 pmd = pmd_offset(pud, address);
3855 VM_BUG_ON(pmd_trans_huge(*pmd));
3856
3857 if (pmd_huge(*pmd)) {
3858 if (!pmdpp)
3859 goto out;
3860
3861 *ptlp = pmd_lock(mm, pmd);
3862 if (pmd_huge(*pmd)) {
3863 *pmdpp = pmd;
3864 return 0;
3865 }
3866 spin_unlock(*ptlp);
3867 }
3868
3869 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
3870 goto out;
3871
3872 ptep = pte_offset_map_lock(mm, pmd, address, ptlp);
3873 if (!ptep)
3874 goto out;
3875 if (!pte_present(*ptep))
3876 goto unlock;
3877 *ptepp = ptep;
3878 return 0;
3879 unlock:
3880 pte_unmap_unlock(ptep, *ptlp);
3881 out:
3882 return -EINVAL;
3883 }
3884
3885 static inline int follow_pte(struct mm_struct *mm, unsigned long address,
3886 pte_t **ptepp, spinlock_t **ptlp)
3887 {
3888 int res;
3889
3890 /* (void) is needed to make gcc happy */
3891 (void) __cond_lock(*ptlp,
3892 !(res = __follow_pte_pmd(mm, address, ptepp, NULL,
3893 ptlp)));
3894 return res;
3895 }
3896
3897 int follow_pte_pmd(struct mm_struct *mm, unsigned long address,
3898 pte_t **ptepp, pmd_t **pmdpp, spinlock_t **ptlp)
3899 {
3900 int res;
3901
3902 /* (void) is needed to make gcc happy */
3903 (void) __cond_lock(*ptlp,
3904 !(res = __follow_pte_pmd(mm, address, ptepp, pmdpp,
3905 ptlp)));
3906 return res;
3907 }
3908 EXPORT_SYMBOL(follow_pte_pmd);
3909
3910 /**
3911 * follow_pfn - look up PFN at a user virtual address
3912 * @vma: memory mapping
3913 * @address: user virtual address
3914 * @pfn: location to store found PFN
3915 *
3916 * Only IO mappings and raw PFN mappings are allowed.
3917 *
3918 * Returns zero and the pfn at @pfn on success, -ve otherwise.
3919 */
3920 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
3921 unsigned long *pfn)
3922 {
3923 int ret = -EINVAL;
3924 spinlock_t *ptl;
3925 pte_t *ptep;
3926
3927 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3928 return ret;
3929
3930 ret = follow_pte(vma->vm_mm, address, &ptep, &ptl);
3931 if (ret)
3932 return ret;
3933 *pfn = pte_pfn(*ptep);
3934 pte_unmap_unlock(ptep, ptl);
3935 return 0;
3936 }
3937 EXPORT_SYMBOL(follow_pfn);
3938
3939 #ifdef CONFIG_HAVE_IOREMAP_PROT
3940 int follow_phys(struct vm_area_struct *vma,
3941 unsigned long address, unsigned int flags,
3942 unsigned long *prot, resource_size_t *phys)
3943 {
3944 int ret = -EINVAL;
3945 pte_t *ptep, pte;
3946 spinlock_t *ptl;
3947
3948 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
3949 goto out;
3950
3951 if (follow_pte(vma->vm_mm, address, &ptep, &ptl))
3952 goto out;
3953 pte = *ptep;
3954
3955 if ((flags & FOLL_WRITE) && !pte_write(pte))
3956 goto unlock;
3957
3958 *prot = pgprot_val(pte_pgprot(pte));
3959 *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT;
3960
3961 ret = 0;
3962 unlock:
3963 pte_unmap_unlock(ptep, ptl);
3964 out:
3965 return ret;
3966 }
3967
3968 int generic_access_phys(struct vm_area_struct *vma, unsigned long addr,
3969 void *buf, int len, int write)
3970 {
3971 resource_size_t phys_addr;
3972 unsigned long prot = 0;
3973 void __iomem *maddr;
3974 int offset = addr & (PAGE_SIZE-1);
3975
3976 if (follow_phys(vma, addr, write, &prot, &phys_addr))
3977 return -EINVAL;
3978
3979 maddr = ioremap_prot(phys_addr, PAGE_ALIGN(len + offset), prot);
3980 if (write)
3981 memcpy_toio(maddr + offset, buf, len);
3982 else
3983 memcpy_fromio(buf, maddr + offset, len);
3984 iounmap(maddr);
3985
3986 return len;
3987 }
3988 EXPORT_SYMBOL_GPL(generic_access_phys);
3989 #endif
3990
3991 /*
3992 * Access another process' address space as given in mm. If non-NULL, use the
3993 * given task for page fault accounting.
3994 */
3995 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
3996 unsigned long addr, void *buf, int len, unsigned int gup_flags)
3997 {
3998 struct vm_area_struct *vma;
3999 void *old_buf = buf;
4000 int write = gup_flags & FOLL_WRITE;
4001
4002 down_read(&mm->mmap_sem);
4003 /* ignore errors, just check how much was successfully transferred */
4004 while (len) {
4005 int bytes, ret, offset;
4006 void *maddr;
4007 struct page *page = NULL;
4008
4009 ret = get_user_pages_remote(tsk, mm, addr, 1,
4010 gup_flags, &page, &vma, NULL);
4011 if (ret <= 0) {
4012 #ifndef CONFIG_HAVE_IOREMAP_PROT
4013 break;
4014 #else
4015 /*
4016 * Check if this is a VM_IO | VM_PFNMAP VMA, which
4017 * we can access using slightly different code.
4018 */
4019 vma = find_vma(mm, addr);
4020 if (!vma || vma->vm_start > addr)
4021 break;
4022 if (vma->vm_ops && vma->vm_ops->access)
4023 ret = vma->vm_ops->access(vma, addr, buf,
4024 len, write);
4025 if (ret <= 0)
4026 break;
4027 bytes = ret;
4028 #endif
4029 } else {
4030 bytes = len;
4031 offset = addr & (PAGE_SIZE-1);
4032 if (bytes > PAGE_SIZE-offset)
4033 bytes = PAGE_SIZE-offset;
4034
4035 maddr = kmap(page);
4036 if (write) {
4037 copy_to_user_page(vma, page, addr,
4038 maddr + offset, buf, bytes);
4039 set_page_dirty_lock(page);
4040 } else {
4041 copy_from_user_page(vma, page, addr,
4042 buf, maddr + offset, bytes);
4043 }
4044 kunmap(page);
4045 put_page(page);
4046 }
4047 len -= bytes;
4048 buf += bytes;
4049 addr += bytes;
4050 }
4051 up_read(&mm->mmap_sem);
4052
4053 return buf - old_buf;
4054 }
4055
4056 /**
4057 * access_remote_vm - access another process' address space
4058 * @mm: the mm_struct of the target address space
4059 * @addr: start address to access
4060 * @buf: source or destination buffer
4061 * @len: number of bytes to transfer
4062 * @gup_flags: flags modifying lookup behaviour
4063 *
4064 * The caller must hold a reference on @mm.
4065 */
4066 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
4067 void *buf, int len, unsigned int gup_flags)
4068 {
4069 return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
4070 }
4071
4072 /*
4073 * Access another process' address space.
4074 * Source/target buffer must be kernel space,
4075 * Do not walk the page table directly, use get_user_pages
4076 */
4077 int access_process_vm(struct task_struct *tsk, unsigned long addr,
4078 void *buf, int len, unsigned int gup_flags)
4079 {
4080 struct mm_struct *mm;
4081 int ret;
4082
4083 mm = get_task_mm(tsk);
4084 if (!mm)
4085 return 0;
4086
4087 ret = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
4088
4089 mmput(mm);
4090
4091 return ret;
4092 }
4093 EXPORT_SYMBOL_GPL(access_process_vm);
4094
4095 /*
4096 * Print the name of a VMA.
4097 */
4098 void print_vma_addr(char *prefix, unsigned long ip)
4099 {
4100 struct mm_struct *mm = current->mm;
4101 struct vm_area_struct *vma;
4102
4103 /*
4104 * Do not print if we are in atomic
4105 * contexts (in exception stacks, etc.):
4106 */
4107 if (preempt_count())
4108 return;
4109
4110 down_read(&mm->mmap_sem);
4111 vma = find_vma(mm, ip);
4112 if (vma && vma->vm_file) {
4113 struct file *f = vma->vm_file;
4114 char *buf = (char *)__get_free_page(GFP_KERNEL);
4115 if (buf) {
4116 char *p;
4117
4118 p = file_path(f, buf, PAGE_SIZE);
4119 if (IS_ERR(p))
4120 p = "?";
4121 printk("%s%s[%lx+%lx]", prefix, kbasename(p),
4122 vma->vm_start,
4123 vma->vm_end - vma->vm_start);
4124 free_page((unsigned long)buf);
4125 }
4126 }
4127 up_read(&mm->mmap_sem);
4128 }
4129
4130 #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4131 void __might_fault(const char *file, int line)
4132 {
4133 /*
4134 * Some code (nfs/sunrpc) uses socket ops on kernel memory while
4135 * holding the mmap_sem, this is safe because kernel memory doesn't
4136 * get paged out, therefore we'll never actually fault, and the
4137 * below annotations will generate false positives.
4138 */
4139 if (segment_eq(get_fs(), KERNEL_DS))
4140 return;
4141 if (pagefault_disabled())
4142 return;
4143 __might_sleep(file, line, 0);
4144 #if defined(CONFIG_DEBUG_ATOMIC_SLEEP)
4145 if (current->mm)
4146 might_lock_read(&current->mm->mmap_sem);
4147 #endif
4148 }
4149 EXPORT_SYMBOL(__might_fault);
4150 #endif
4151
4152 #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS)
4153 static void clear_gigantic_page(struct page *page,
4154 unsigned long addr,
4155 unsigned int pages_per_huge_page)
4156 {
4157 int i;
4158 struct page *p = page;
4159
4160 might_sleep();
4161 for (i = 0; i < pages_per_huge_page;
4162 i++, p = mem_map_next(p, page, i)) {
4163 cond_resched();
4164 clear_user_highpage(p, addr + i * PAGE_SIZE);
4165 }
4166 }
4167 void clear_huge_page(struct page *page,
4168 unsigned long addr, unsigned int pages_per_huge_page)
4169 {
4170 int i;
4171
4172 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4173 clear_gigantic_page(page, addr, pages_per_huge_page);
4174 return;
4175 }
4176
4177 might_sleep();
4178 for (i = 0; i < pages_per_huge_page; i++) {
4179 cond_resched();
4180 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
4181 }
4182 }
4183
4184 static void copy_user_gigantic_page(struct page *dst, struct page *src,
4185 unsigned long addr,
4186 struct vm_area_struct *vma,
4187 unsigned int pages_per_huge_page)
4188 {
4189 int i;
4190 struct page *dst_base = dst;
4191 struct page *src_base = src;
4192
4193 for (i = 0; i < pages_per_huge_page; ) {
4194 cond_resched();
4195 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
4196
4197 i++;
4198 dst = mem_map_next(dst, dst_base, i);
4199 src = mem_map_next(src, src_base, i);
4200 }
4201 }
4202
4203 void copy_user_huge_page(struct page *dst, struct page *src,
4204 unsigned long addr, struct vm_area_struct *vma,
4205 unsigned int pages_per_huge_page)
4206 {
4207 int i;
4208
4209 if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) {
4210 copy_user_gigantic_page(dst, src, addr, vma,
4211 pages_per_huge_page);
4212 return;
4213 }
4214
4215 might_sleep();
4216 for (i = 0; i < pages_per_huge_page; i++) {
4217 cond_resched();
4218 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
4219 }
4220 }
4221
4222 long copy_huge_page_from_user(struct page *dst_page,
4223 const void __user *usr_src,
4224 unsigned int pages_per_huge_page,
4225 bool allow_pagefault)
4226 {
4227 void *src = (void *)usr_src;
4228 void *page_kaddr;
4229 unsigned long i, rc = 0;
4230 unsigned long ret_val = pages_per_huge_page * PAGE_SIZE;
4231
4232 for (i = 0; i < pages_per_huge_page; i++) {
4233 if (allow_pagefault)
4234 page_kaddr = kmap(dst_page + i);
4235 else
4236 page_kaddr = kmap_atomic(dst_page + i);
4237 rc = copy_from_user(page_kaddr,
4238 (const void __user *)(src + i * PAGE_SIZE),
4239 PAGE_SIZE);
4240 if (allow_pagefault)
4241 kunmap(dst_page + i);
4242 else
4243 kunmap_atomic(page_kaddr);
4244
4245 ret_val -= (PAGE_SIZE - rc);
4246 if (rc)
4247 break;
4248
4249 cond_resched();
4250 }
4251 return ret_val;
4252 }
4253 #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */
4254
4255 #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS
4256
4257 static struct kmem_cache *page_ptl_cachep;
4258
4259 void __init ptlock_cache_init(void)
4260 {
4261 page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0,
4262 SLAB_PANIC, NULL);
4263 }
4264
4265 bool ptlock_alloc(struct page *page)
4266 {
4267 spinlock_t *ptl;
4268
4269 ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL);
4270 if (!ptl)
4271 return false;
4272 page->ptl = ptl;
4273 return true;
4274 }
4275
4276 void ptlock_free(struct page *page)
4277 {
4278 kmem_cache_free(page_ptl_cachep, page->ptl);
4279 }
4280 #endif