]>
Commit | Line | Data |
---|---|---|
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/hugetlb.h> | |
44 | #include <linux/mman.h> | |
45 | #include <linux/swap.h> | |
46 | #include <linux/highmem.h> | |
47 | #include <linux/pagemap.h> | |
48 | #include <linux/ksm.h> | |
49 | #include <linux/rmap.h> | |
50 | #include <linux/export.h> | |
51 | #include <linux/delayacct.h> | |
52 | #include <linux/init.h> | |
53 | #include <linux/writeback.h> | |
54 | #include <linux/memcontrol.h> | |
55 | #include <linux/mmu_notifier.h> | |
56 | #include <linux/kallsyms.h> | |
57 | #include <linux/swapops.h> | |
58 | #include <linux/elf.h> | |
59 | #include <linux/gfp.h> | |
60 | #include <linux/migrate.h> | |
61 | #include <linux/string.h> | |
62 | #include <linux/dma-debug.h> | |
63 | #include <linux/debugfs.h> | |
64 | ||
65 | #include <asm/io.h> | |
66 | #include <asm/pgalloc.h> | |
67 | #include <asm/uaccess.h> | |
68 | #include <asm/tlb.h> | |
69 | #include <asm/tlbflush.h> | |
70 | #include <asm/pgtable.h> | |
71 | ||
72 | #include "internal.h" | |
73 | ||
74 | #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS | |
75 | #warning Unfortunate NUMA and NUMA Balancing config, growing page-frame for last_cpupid. | |
76 | #endif | |
77 | ||
78 | #ifndef CONFIG_NEED_MULTIPLE_NODES | |
79 | /* use the per-pgdat data instead for discontigmem - mbligh */ | |
80 | unsigned long max_mapnr; | |
81 | struct page *mem_map; | |
82 | ||
83 | EXPORT_SYMBOL(max_mapnr); | |
84 | EXPORT_SYMBOL(mem_map); | |
85 | #endif | |
86 | ||
87 | /* | |
88 | * A number of key systems in x86 including ioremap() rely on the assumption | |
89 | * that high_memory defines the upper bound on direct map memory, then end | |
90 | * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and | |
91 | * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL | |
92 | * and ZONE_HIGHMEM. | |
93 | */ | |
94 | void * high_memory; | |
95 | ||
96 | EXPORT_SYMBOL(high_memory); | |
97 | ||
98 | /* | |
99 | * Randomize the address space (stacks, mmaps, brk, etc.). | |
100 | * | |
101 | * ( When CONFIG_COMPAT_BRK=y we exclude brk from randomization, | |
102 | * as ancient (libc5 based) binaries can segfault. ) | |
103 | */ | |
104 | int randomize_va_space __read_mostly = | |
105 | #ifdef CONFIG_COMPAT_BRK | |
106 | 1; | |
107 | #else | |
108 | 2; | |
109 | #endif | |
110 | ||
111 | static int __init disable_randmaps(char *s) | |
112 | { | |
113 | randomize_va_space = 0; | |
114 | return 1; | |
115 | } | |
116 | __setup("norandmaps", disable_randmaps); | |
117 | ||
118 | unsigned long zero_pfn __read_mostly; | |
119 | unsigned long highest_memmap_pfn __read_mostly; | |
120 | ||
121 | EXPORT_SYMBOL(zero_pfn); | |
122 | ||
123 | /* | |
124 | * CONFIG_MMU architectures set up ZERO_PAGE in their paging_init() | |
125 | */ | |
126 | static int __init init_zero_pfn(void) | |
127 | { | |
128 | zero_pfn = page_to_pfn(ZERO_PAGE(0)); | |
129 | return 0; | |
130 | } | |
131 | core_initcall(init_zero_pfn); | |
132 | ||
133 | ||
134 | #if defined(SPLIT_RSS_COUNTING) | |
135 | ||
136 | void sync_mm_rss(struct mm_struct *mm) | |
137 | { | |
138 | int i; | |
139 | ||
140 | for (i = 0; i < NR_MM_COUNTERS; i++) { | |
141 | if (current->rss_stat.count[i]) { | |
142 | add_mm_counter(mm, i, current->rss_stat.count[i]); | |
143 | current->rss_stat.count[i] = 0; | |
144 | } | |
145 | } | |
146 | current->rss_stat.events = 0; | |
147 | } | |
148 | ||
149 | static void add_mm_counter_fast(struct mm_struct *mm, int member, int val) | |
150 | { | |
151 | struct task_struct *task = current; | |
152 | ||
153 | if (likely(task->mm == mm)) | |
154 | task->rss_stat.count[member] += val; | |
155 | else | |
156 | add_mm_counter(mm, member, val); | |
157 | } | |
158 | #define inc_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, 1) | |
159 | #define dec_mm_counter_fast(mm, member) add_mm_counter_fast(mm, member, -1) | |
160 | ||
161 | /* sync counter once per 64 page faults */ | |
162 | #define TASK_RSS_EVENTS_THRESH (64) | |
163 | static void check_sync_rss_stat(struct task_struct *task) | |
164 | { | |
165 | if (unlikely(task != current)) | |
166 | return; | |
167 | if (unlikely(task->rss_stat.events++ > TASK_RSS_EVENTS_THRESH)) | |
168 | sync_mm_rss(task->mm); | |
169 | } | |
170 | #else /* SPLIT_RSS_COUNTING */ | |
171 | ||
172 | #define inc_mm_counter_fast(mm, member) inc_mm_counter(mm, member) | |
173 | #define dec_mm_counter_fast(mm, member) dec_mm_counter(mm, member) | |
174 | ||
175 | static void check_sync_rss_stat(struct task_struct *task) | |
176 | { | |
177 | } | |
178 | ||
179 | #endif /* SPLIT_RSS_COUNTING */ | |
180 | ||
181 | #ifdef HAVE_GENERIC_MMU_GATHER | |
182 | ||
183 | static int tlb_next_batch(struct mmu_gather *tlb) | |
184 | { | |
185 | struct mmu_gather_batch *batch; | |
186 | ||
187 | batch = tlb->active; | |
188 | if (batch->next) { | |
189 | tlb->active = batch->next; | |
190 | return 1; | |
191 | } | |
192 | ||
193 | if (tlb->batch_count == MAX_GATHER_BATCH_COUNT) | |
194 | return 0; | |
195 | ||
196 | batch = (void *)__get_free_pages(GFP_NOWAIT | __GFP_NOWARN, 0); | |
197 | if (!batch) | |
198 | return 0; | |
199 | ||
200 | tlb->batch_count++; | |
201 | batch->next = NULL; | |
202 | batch->nr = 0; | |
203 | batch->max = MAX_GATHER_BATCH; | |
204 | ||
205 | tlb->active->next = batch; | |
206 | tlb->active = batch; | |
207 | ||
208 | return 1; | |
209 | } | |
210 | ||
211 | /* tlb_gather_mmu | |
212 | * Called to initialize an (on-stack) mmu_gather structure for page-table | |
213 | * tear-down from @mm. The @fullmm argument is used when @mm is without | |
214 | * users and we're going to destroy the full address space (exit/execve). | |
215 | */ | |
216 | void tlb_gather_mmu(struct mmu_gather *tlb, struct mm_struct *mm, unsigned long start, unsigned long end) | |
217 | { | |
218 | tlb->mm = mm; | |
219 | ||
220 | /* Is it from 0 to ~0? */ | |
221 | tlb->fullmm = !(start | (end+1)); | |
222 | tlb->need_flush_all = 0; | |
223 | tlb->local.next = NULL; | |
224 | tlb->local.nr = 0; | |
225 | tlb->local.max = ARRAY_SIZE(tlb->__pages); | |
226 | tlb->active = &tlb->local; | |
227 | tlb->batch_count = 0; | |
228 | ||
229 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE | |
230 | tlb->batch = NULL; | |
231 | #endif | |
232 | ||
233 | __tlb_reset_range(tlb); | |
234 | } | |
235 | ||
236 | static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb) | |
237 | { | |
238 | if (!tlb->end) | |
239 | return; | |
240 | ||
241 | tlb_flush(tlb); | |
242 | mmu_notifier_invalidate_range(tlb->mm, tlb->start, tlb->end); | |
243 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE | |
244 | tlb_table_flush(tlb); | |
245 | #endif | |
246 | __tlb_reset_range(tlb); | |
247 | } | |
248 | ||
249 | static void tlb_flush_mmu_free(struct mmu_gather *tlb) | |
250 | { | |
251 | struct mmu_gather_batch *batch; | |
252 | ||
253 | for (batch = &tlb->local; batch && batch->nr; batch = batch->next) { | |
254 | free_pages_and_swap_cache(batch->pages, batch->nr); | |
255 | batch->nr = 0; | |
256 | } | |
257 | tlb->active = &tlb->local; | |
258 | } | |
259 | ||
260 | void tlb_flush_mmu(struct mmu_gather *tlb) | |
261 | { | |
262 | tlb_flush_mmu_tlbonly(tlb); | |
263 | tlb_flush_mmu_free(tlb); | |
264 | } | |
265 | ||
266 | /* tlb_finish_mmu | |
267 | * Called at the end of the shootdown operation to free up any resources | |
268 | * that were required. | |
269 | */ | |
270 | void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) | |
271 | { | |
272 | struct mmu_gather_batch *batch, *next; | |
273 | ||
274 | tlb_flush_mmu(tlb); | |
275 | ||
276 | /* keep the page table cache within bounds */ | |
277 | check_pgt_cache(); | |
278 | ||
279 | for (batch = tlb->local.next; batch; batch = next) { | |
280 | next = batch->next; | |
281 | free_pages((unsigned long)batch, 0); | |
282 | } | |
283 | tlb->local.next = NULL; | |
284 | } | |
285 | ||
286 | /* __tlb_remove_page | |
287 | * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while | |
288 | * handling the additional races in SMP caused by other CPUs caching valid | |
289 | * mappings in their TLBs. Returns the number of free page slots left. | |
290 | * When out of page slots we must call tlb_flush_mmu(). | |
291 | */ | |
292 | int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) | |
293 | { | |
294 | struct mmu_gather_batch *batch; | |
295 | ||
296 | VM_BUG_ON(!tlb->end); | |
297 | ||
298 | batch = tlb->active; | |
299 | batch->pages[batch->nr++] = page; | |
300 | if (batch->nr == batch->max) { | |
301 | if (!tlb_next_batch(tlb)) | |
302 | return 0; | |
303 | batch = tlb->active; | |
304 | } | |
305 | VM_BUG_ON_PAGE(batch->nr > batch->max, page); | |
306 | ||
307 | return batch->max - batch->nr; | |
308 | } | |
309 | ||
310 | #endif /* HAVE_GENERIC_MMU_GATHER */ | |
311 | ||
312 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE | |
313 | ||
314 | /* | |
315 | * See the comment near struct mmu_table_batch. | |
316 | */ | |
317 | ||
318 | static void tlb_remove_table_smp_sync(void *arg) | |
319 | { | |
320 | /* Simply deliver the interrupt */ | |
321 | } | |
322 | ||
323 | static void tlb_remove_table_one(void *table) | |
324 | { | |
325 | /* | |
326 | * This isn't an RCU grace period and hence the page-tables cannot be | |
327 | * assumed to be actually RCU-freed. | |
328 | * | |
329 | * It is however sufficient for software page-table walkers that rely on | |
330 | * IRQ disabling. See the comment near struct mmu_table_batch. | |
331 | */ | |
332 | smp_call_function(tlb_remove_table_smp_sync, NULL, 1); | |
333 | __tlb_remove_table(table); | |
334 | } | |
335 | ||
336 | static void tlb_remove_table_rcu(struct rcu_head *head) | |
337 | { | |
338 | struct mmu_table_batch *batch; | |
339 | int i; | |
340 | ||
341 | batch = container_of(head, struct mmu_table_batch, rcu); | |
342 | ||
343 | for (i = 0; i < batch->nr; i++) | |
344 | __tlb_remove_table(batch->tables[i]); | |
345 | ||
346 | free_page((unsigned long)batch); | |
347 | } | |
348 | ||
349 | void tlb_table_flush(struct mmu_gather *tlb) | |
350 | { | |
351 | struct mmu_table_batch **batch = &tlb->batch; | |
352 | ||
353 | if (*batch) { | |
354 | call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); | |
355 | *batch = NULL; | |
356 | } | |
357 | } | |
358 | ||
359 | void tlb_remove_table(struct mmu_gather *tlb, void *table) | |
360 | { | |
361 | struct mmu_table_batch **batch = &tlb->batch; | |
362 | ||
363 | /* | |
364 | * When there's less then two users of this mm there cannot be a | |
365 | * concurrent page-table walk. | |
366 | */ | |
367 | if (atomic_read(&tlb->mm->mm_users) < 2) { | |
368 | __tlb_remove_table(table); | |
369 | return; | |
370 | } | |
371 | ||
372 | if (*batch == NULL) { | |
373 | *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); | |
374 | if (*batch == NULL) { | |
375 | tlb_remove_table_one(table); | |
376 | return; | |
377 | } | |
378 | (*batch)->nr = 0; | |
379 | } | |
380 | (*batch)->tables[(*batch)->nr++] = table; | |
381 | if ((*batch)->nr == MAX_TABLE_BATCH) | |
382 | tlb_table_flush(tlb); | |
383 | } | |
384 | ||
385 | #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ | |
386 | ||
387 | /* | |
388 | * Note: this doesn't free the actual pages themselves. That | |
389 | * has been handled earlier when unmapping all the memory regions. | |
390 | */ | |
391 | static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, | |
392 | unsigned long addr) | |
393 | { | |
394 | pgtable_t token = pmd_pgtable(*pmd); | |
395 | pmd_clear(pmd); | |
396 | pte_free_tlb(tlb, token, addr); | |
397 | atomic_long_dec(&tlb->mm->nr_ptes); | |
398 | } | |
399 | ||
400 | static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, | |
401 | unsigned long addr, unsigned long end, | |
402 | unsigned long floor, unsigned long ceiling) | |
403 | { | |
404 | pmd_t *pmd; | |
405 | unsigned long next; | |
406 | unsigned long start; | |
407 | ||
408 | start = addr; | |
409 | pmd = pmd_offset(pud, addr); | |
410 | do { | |
411 | next = pmd_addr_end(addr, end); | |
412 | if (pmd_none_or_clear_bad(pmd)) | |
413 | continue; | |
414 | free_pte_range(tlb, pmd, addr); | |
415 | } while (pmd++, addr = next, addr != end); | |
416 | ||
417 | start &= PUD_MASK; | |
418 | if (start < floor) | |
419 | return; | |
420 | if (ceiling) { | |
421 | ceiling &= PUD_MASK; | |
422 | if (!ceiling) | |
423 | return; | |
424 | } | |
425 | if (end - 1 > ceiling - 1) | |
426 | return; | |
427 | ||
428 | pmd = pmd_offset(pud, start); | |
429 | pud_clear(pud); | |
430 | pmd_free_tlb(tlb, pmd, start); | |
431 | mm_dec_nr_pmds(tlb->mm); | |
432 | } | |
433 | ||
434 | static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd, | |
435 | unsigned long addr, unsigned long end, | |
436 | unsigned long floor, unsigned long ceiling) | |
437 | { | |
438 | pud_t *pud; | |
439 | unsigned long next; | |
440 | unsigned long start; | |
441 | ||
442 | start = addr; | |
443 | pud = pud_offset(pgd, addr); | |
444 | do { | |
445 | next = pud_addr_end(addr, end); | |
446 | if (pud_none_or_clear_bad(pud)) | |
447 | continue; | |
448 | free_pmd_range(tlb, pud, addr, next, floor, ceiling); | |
449 | } while (pud++, addr = next, addr != end); | |
450 | ||
451 | start &= PGDIR_MASK; | |
452 | if (start < floor) | |
453 | return; | |
454 | if (ceiling) { | |
455 | ceiling &= PGDIR_MASK; | |
456 | if (!ceiling) | |
457 | return; | |
458 | } | |
459 | if (end - 1 > ceiling - 1) | |
460 | return; | |
461 | ||
462 | pud = pud_offset(pgd, start); | |
463 | pgd_clear(pgd); | |
464 | pud_free_tlb(tlb, pud, start); | |
465 | } | |
466 | ||
467 | /* | |
468 | * This function frees user-level page tables of a process. | |
469 | */ | |
470 | void free_pgd_range(struct mmu_gather *tlb, | |
471 | unsigned long addr, unsigned long end, | |
472 | unsigned long floor, unsigned long ceiling) | |
473 | { | |
474 | pgd_t *pgd; | |
475 | unsigned long next; | |
476 | ||
477 | /* | |
478 | * The next few lines have given us lots of grief... | |
479 | * | |
480 | * Why are we testing PMD* at this top level? Because often | |
481 | * there will be no work to do at all, and we'd prefer not to | |
482 | * go all the way down to the bottom just to discover that. | |
483 | * | |
484 | * Why all these "- 1"s? Because 0 represents both the bottom | |
485 | * of the address space and the top of it (using -1 for the | |
486 | * top wouldn't help much: the masks would do the wrong thing). | |
487 | * The rule is that addr 0 and floor 0 refer to the bottom of | |
488 | * the address space, but end 0 and ceiling 0 refer to the top | |
489 | * Comparisons need to use "end - 1" and "ceiling - 1" (though | |
490 | * that end 0 case should be mythical). | |
491 | * | |
492 | * Wherever addr is brought up or ceiling brought down, we must | |
493 | * be careful to reject "the opposite 0" before it confuses the | |
494 | * subsequent tests. But what about where end is brought down | |
495 | * by PMD_SIZE below? no, end can't go down to 0 there. | |
496 | * | |
497 | * Whereas we round start (addr) and ceiling down, by different | |
498 | * masks at different levels, in order to test whether a table | |
499 | * now has no other vmas using it, so can be freed, we don't | |
500 | * bother to round floor or end up - the tests don't need that. | |
501 | */ | |
502 | ||
503 | addr &= PMD_MASK; | |
504 | if (addr < floor) { | |
505 | addr += PMD_SIZE; | |
506 | if (!addr) | |
507 | return; | |
508 | } | |
509 | if (ceiling) { | |
510 | ceiling &= PMD_MASK; | |
511 | if (!ceiling) | |
512 | return; | |
513 | } | |
514 | if (end - 1 > ceiling - 1) | |
515 | end -= PMD_SIZE; | |
516 | if (addr > end - 1) | |
517 | return; | |
518 | ||
519 | pgd = pgd_offset(tlb->mm, addr); | |
520 | do { | |
521 | next = pgd_addr_end(addr, end); | |
522 | if (pgd_none_or_clear_bad(pgd)) | |
523 | continue; | |
524 | free_pud_range(tlb, pgd, addr, next, floor, ceiling); | |
525 | } while (pgd++, addr = next, addr != end); | |
526 | } | |
527 | ||
528 | void free_pgtables(struct mmu_gather *tlb, struct vm_area_struct *vma, | |
529 | unsigned long floor, unsigned long ceiling) | |
530 | { | |
531 | while (vma) { | |
532 | struct vm_area_struct *next = vma->vm_next; | |
533 | unsigned long addr = vma->vm_start; | |
534 | ||
535 | /* | |
536 | * Hide vma from rmap and truncate_pagecache before freeing | |
537 | * pgtables | |
538 | */ | |
539 | unlink_anon_vmas(vma); | |
540 | unlink_file_vma(vma); | |
541 | ||
542 | if (is_vm_hugetlb_page(vma)) { | |
543 | hugetlb_free_pgd_range(tlb, addr, vma->vm_end, | |
544 | floor, next? next->vm_start: ceiling); | |
545 | } else { | |
546 | /* | |
547 | * Optimization: gather nearby vmas into one call down | |
548 | */ | |
549 | while (next && next->vm_start <= vma->vm_end + PMD_SIZE | |
550 | && !is_vm_hugetlb_page(next)) { | |
551 | vma = next; | |
552 | next = vma->vm_next; | |
553 | unlink_anon_vmas(vma); | |
554 | unlink_file_vma(vma); | |
555 | } | |
556 | free_pgd_range(tlb, addr, vma->vm_end, | |
557 | floor, next? next->vm_start: ceiling); | |
558 | } | |
559 | vma = next; | |
560 | } | |
561 | } | |
562 | ||
563 | int __pte_alloc(struct mm_struct *mm, struct vm_area_struct *vma, | |
564 | pmd_t *pmd, unsigned long address) | |
565 | { | |
566 | spinlock_t *ptl; | |
567 | pgtable_t new = pte_alloc_one(mm, address); | |
568 | int wait_split_huge_page; | |
569 | if (!new) | |
570 | return -ENOMEM; | |
571 | ||
572 | /* | |
573 | * Ensure all pte setup (eg. pte page lock and page clearing) are | |
574 | * visible before the pte is made visible to other CPUs by being | |
575 | * put into page tables. | |
576 | * | |
577 | * The other side of the story is the pointer chasing in the page | |
578 | * table walking code (when walking the page table without locking; | |
579 | * ie. most of the time). Fortunately, these data accesses consist | |
580 | * of a chain of data-dependent loads, meaning most CPUs (alpha | |
581 | * being the notable exception) will already guarantee loads are | |
582 | * seen in-order. See the alpha page table accessors for the | |
583 | * smp_read_barrier_depends() barriers in page table walking code. | |
584 | */ | |
585 | smp_wmb(); /* Could be smp_wmb__xxx(before|after)_spin_lock */ | |
586 | ||
587 | ptl = pmd_lock(mm, pmd); | |
588 | wait_split_huge_page = 0; | |
589 | if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ | |
590 | atomic_long_inc(&mm->nr_ptes); | |
591 | pmd_populate(mm, pmd, new); | |
592 | new = NULL; | |
593 | } else if (unlikely(pmd_trans_splitting(*pmd))) | |
594 | wait_split_huge_page = 1; | |
595 | spin_unlock(ptl); | |
596 | if (new) | |
597 | pte_free(mm, new); | |
598 | if (wait_split_huge_page) | |
599 | wait_split_huge_page(vma->anon_vma, pmd); | |
600 | return 0; | |
601 | } | |
602 | ||
603 | int __pte_alloc_kernel(pmd_t *pmd, unsigned long address) | |
604 | { | |
605 | pte_t *new = pte_alloc_one_kernel(&init_mm, address); | |
606 | if (!new) | |
607 | return -ENOMEM; | |
608 | ||
609 | smp_wmb(); /* See comment in __pte_alloc */ | |
610 | ||
611 | spin_lock(&init_mm.page_table_lock); | |
612 | if (likely(pmd_none(*pmd))) { /* Has another populated it ? */ | |
613 | pmd_populate_kernel(&init_mm, pmd, new); | |
614 | new = NULL; | |
615 | } else | |
616 | VM_BUG_ON(pmd_trans_splitting(*pmd)); | |
617 | spin_unlock(&init_mm.page_table_lock); | |
618 | if (new) | |
619 | pte_free_kernel(&init_mm, new); | |
620 | return 0; | |
621 | } | |
622 | ||
623 | static inline void init_rss_vec(int *rss) | |
624 | { | |
625 | memset(rss, 0, sizeof(int) * NR_MM_COUNTERS); | |
626 | } | |
627 | ||
628 | static inline void add_mm_rss_vec(struct mm_struct *mm, int *rss) | |
629 | { | |
630 | int i; | |
631 | ||
632 | if (current->mm == mm) | |
633 | sync_mm_rss(mm); | |
634 | for (i = 0; i < NR_MM_COUNTERS; i++) | |
635 | if (rss[i]) | |
636 | add_mm_counter(mm, i, rss[i]); | |
637 | } | |
638 | ||
639 | /* | |
640 | * This function is called to print an error when a bad pte | |
641 | * is found. For example, we might have a PFN-mapped pte in | |
642 | * a region that doesn't allow it. | |
643 | * | |
644 | * The calling function must still handle the error. | |
645 | */ | |
646 | static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr, | |
647 | pte_t pte, struct page *page) | |
648 | { | |
649 | pgd_t *pgd = pgd_offset(vma->vm_mm, addr); | |
650 | pud_t *pud = pud_offset(pgd, addr); | |
651 | pmd_t *pmd = pmd_offset(pud, addr); | |
652 | struct address_space *mapping; | |
653 | pgoff_t index; | |
654 | static unsigned long resume; | |
655 | static unsigned long nr_shown; | |
656 | static unsigned long nr_unshown; | |
657 | ||
658 | /* | |
659 | * Allow a burst of 60 reports, then keep quiet for that minute; | |
660 | * or allow a steady drip of one report per second. | |
661 | */ | |
662 | if (nr_shown == 60) { | |
663 | if (time_before(jiffies, resume)) { | |
664 | nr_unshown++; | |
665 | return; | |
666 | } | |
667 | if (nr_unshown) { | |
668 | printk(KERN_ALERT | |
669 | "BUG: Bad page map: %lu messages suppressed\n", | |
670 | nr_unshown); | |
671 | nr_unshown = 0; | |
672 | } | |
673 | nr_shown = 0; | |
674 | } | |
675 | if (nr_shown++ == 0) | |
676 | resume = jiffies + 60 * HZ; | |
677 | ||
678 | mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL; | |
679 | index = linear_page_index(vma, addr); | |
680 | ||
681 | printk(KERN_ALERT | |
682 | "BUG: Bad page map in process %s pte:%08llx pmd:%08llx\n", | |
683 | current->comm, | |
684 | (long long)pte_val(pte), (long long)pmd_val(*pmd)); | |
685 | if (page) | |
686 | dump_page(page, "bad pte"); | |
687 | printk(KERN_ALERT | |
688 | "addr:%p vm_flags:%08lx anon_vma:%p mapping:%p index:%lx\n", | |
689 | (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index); | |
690 | /* | |
691 | * Choose text because data symbols depend on CONFIG_KALLSYMS_ALL=y | |
692 | */ | |
693 | if (vma->vm_ops) | |
694 | printk(KERN_ALERT "vma->vm_ops->fault: %pSR\n", | |
695 | vma->vm_ops->fault); | |
696 | if (vma->vm_file) | |
697 | printk(KERN_ALERT "vma->vm_file->f_op->mmap: %pSR\n", | |
698 | vma->vm_file->f_op->mmap); | |
699 | dump_stack(); | |
700 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); | |
701 | } | |
702 | ||
703 | /* | |
704 | * vm_normal_page -- This function gets the "struct page" associated with a pte. | |
705 | * | |
706 | * "Special" mappings do not wish to be associated with a "struct page" (either | |
707 | * it doesn't exist, or it exists but they don't want to touch it). In this | |
708 | * case, NULL is returned here. "Normal" mappings do have a struct page. | |
709 | * | |
710 | * There are 2 broad cases. Firstly, an architecture may define a pte_special() | |
711 | * pte bit, in which case this function is trivial. Secondly, an architecture | |
712 | * may not have a spare pte bit, which requires a more complicated scheme, | |
713 | * described below. | |
714 | * | |
715 | * A raw VM_PFNMAP mapping (ie. one that is not COWed) is always considered a | |
716 | * special mapping (even if there are underlying and valid "struct pages"). | |
717 | * COWed pages of a VM_PFNMAP are always normal. | |
718 | * | |
719 | * The way we recognize COWed pages within VM_PFNMAP mappings is through the | |
720 | * rules set up by "remap_pfn_range()": the vma will have the VM_PFNMAP bit | |
721 | * set, and the vm_pgoff will point to the first PFN mapped: thus every special | |
722 | * mapping will always honor the rule | |
723 | * | |
724 | * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT) | |
725 | * | |
726 | * And for normal mappings this is false. | |
727 | * | |
728 | * This restricts such mappings to be a linear translation from virtual address | |
729 | * to pfn. To get around this restriction, we allow arbitrary mappings so long | |
730 | * as the vma is not a COW mapping; in that case, we know that all ptes are | |
731 | * special (because none can have been COWed). | |
732 | * | |
733 | * | |
734 | * In order to support COW of arbitrary special mappings, we have VM_MIXEDMAP. | |
735 | * | |
736 | * VM_MIXEDMAP mappings can likewise contain memory with or without "struct | |
737 | * page" backing, however the difference is that _all_ pages with a struct | |
738 | * page (that is, those where pfn_valid is true) are refcounted and considered | |
739 | * normal pages by the VM. The disadvantage is that pages are refcounted | |
740 | * (which can be slower and simply not an option for some PFNMAP users). The | |
741 | * advantage is that we don't have to follow the strict linearity rule of | |
742 | * PFNMAP mappings in order to support COWable mappings. | |
743 | * | |
744 | */ | |
745 | #ifdef __HAVE_ARCH_PTE_SPECIAL | |
746 | # define HAVE_PTE_SPECIAL 1 | |
747 | #else | |
748 | # define HAVE_PTE_SPECIAL 0 | |
749 | #endif | |
750 | struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, | |
751 | pte_t pte) | |
752 | { | |
753 | unsigned long pfn = pte_pfn(pte); | |
754 | ||
755 | if (HAVE_PTE_SPECIAL) { | |
756 | if (likely(!pte_special(pte))) | |
757 | goto check_pfn; | |
758 | if (vma->vm_ops && vma->vm_ops->find_special_page) | |
759 | return vma->vm_ops->find_special_page(vma, addr); | |
760 | if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP)) | |
761 | return NULL; | |
762 | if (!is_zero_pfn(pfn)) | |
763 | print_bad_pte(vma, addr, pte, NULL); | |
764 | return NULL; | |
765 | } | |
766 | ||
767 | /* !HAVE_PTE_SPECIAL case follows: */ | |
768 | ||
769 | if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { | |
770 | if (vma->vm_flags & VM_MIXEDMAP) { | |
771 | if (!pfn_valid(pfn)) | |
772 | return NULL; | |
773 | goto out; | |
774 | } else { | |
775 | unsigned long off; | |
776 | off = (addr - vma->vm_start) >> PAGE_SHIFT; | |
777 | if (pfn == vma->vm_pgoff + off) | |
778 | return NULL; | |
779 | if (!is_cow_mapping(vma->vm_flags)) | |
780 | return NULL; | |
781 | } | |
782 | } | |
783 | ||
784 | if (is_zero_pfn(pfn)) | |
785 | return NULL; | |
786 | check_pfn: | |
787 | if (unlikely(pfn > highest_memmap_pfn)) { | |
788 | print_bad_pte(vma, addr, pte, NULL); | |
789 | return NULL; | |
790 | } | |
791 | ||
792 | /* | |
793 | * NOTE! We still have PageReserved() pages in the page tables. | |
794 | * eg. VDSO mappings can cause them to exist. | |
795 | */ | |
796 | out: | |
797 | return pfn_to_page(pfn); | |
798 | } | |
799 | ||
800 | /* | |
801 | * copy one vm_area from one task to the other. Assumes the page tables | |
802 | * already present in the new task to be cleared in the whole range | |
803 | * covered by this vma. | |
804 | */ | |
805 | ||
806 | static inline unsigned long | |
807 | copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
808 | pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, | |
809 | unsigned long addr, int *rss) | |
810 | { | |
811 | unsigned long vm_flags = vma->vm_flags; | |
812 | pte_t pte = *src_pte; | |
813 | struct page *page; | |
814 | ||
815 | /* pte contains position in swap or file, so copy. */ | |
816 | if (unlikely(!pte_present(pte))) { | |
817 | swp_entry_t entry = pte_to_swp_entry(pte); | |
818 | ||
819 | if (likely(!non_swap_entry(entry))) { | |
820 | if (swap_duplicate(entry) < 0) | |
821 | return entry.val; | |
822 | ||
823 | /* make sure dst_mm is on swapoff's mmlist. */ | |
824 | if (unlikely(list_empty(&dst_mm->mmlist))) { | |
825 | spin_lock(&mmlist_lock); | |
826 | if (list_empty(&dst_mm->mmlist)) | |
827 | list_add(&dst_mm->mmlist, | |
828 | &src_mm->mmlist); | |
829 | spin_unlock(&mmlist_lock); | |
830 | } | |
831 | rss[MM_SWAPENTS]++; | |
832 | } else if (is_migration_entry(entry)) { | |
833 | page = migration_entry_to_page(entry); | |
834 | ||
835 | if (PageAnon(page)) | |
836 | rss[MM_ANONPAGES]++; | |
837 | else | |
838 | rss[MM_FILEPAGES]++; | |
839 | ||
840 | if (is_write_migration_entry(entry) && | |
841 | is_cow_mapping(vm_flags)) { | |
842 | /* | |
843 | * COW mappings require pages in both | |
844 | * parent and child to be set to read. | |
845 | */ | |
846 | make_migration_entry_read(&entry); | |
847 | pte = swp_entry_to_pte(entry); | |
848 | if (pte_swp_soft_dirty(*src_pte)) | |
849 | pte = pte_swp_mksoft_dirty(pte); | |
850 | set_pte_at(src_mm, addr, src_pte, pte); | |
851 | } | |
852 | } | |
853 | goto out_set_pte; | |
854 | } | |
855 | ||
856 | /* | |
857 | * If it's a COW mapping, write protect it both | |
858 | * in the parent and the child | |
859 | */ | |
860 | if (is_cow_mapping(vm_flags)) { | |
861 | ptep_set_wrprotect(src_mm, addr, src_pte); | |
862 | pte = pte_wrprotect(pte); | |
863 | } | |
864 | ||
865 | /* | |
866 | * If it's a shared mapping, mark it clean in | |
867 | * the child | |
868 | */ | |
869 | if (vm_flags & VM_SHARED) | |
870 | pte = pte_mkclean(pte); | |
871 | pte = pte_mkold(pte); | |
872 | ||
873 | page = vm_normal_page(vma, addr, pte); | |
874 | if (page) { | |
875 | get_page(page); | |
876 | page_dup_rmap(page); | |
877 | if (PageAnon(page)) | |
878 | rss[MM_ANONPAGES]++; | |
879 | else | |
880 | rss[MM_FILEPAGES]++; | |
881 | } | |
882 | ||
883 | out_set_pte: | |
884 | set_pte_at(dst_mm, addr, dst_pte, pte); | |
885 | return 0; | |
886 | } | |
887 | ||
888 | static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
889 | pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma, | |
890 | unsigned long addr, unsigned long end) | |
891 | { | |
892 | pte_t *orig_src_pte, *orig_dst_pte; | |
893 | pte_t *src_pte, *dst_pte; | |
894 | spinlock_t *src_ptl, *dst_ptl; | |
895 | int progress = 0; | |
896 | int rss[NR_MM_COUNTERS]; | |
897 | swp_entry_t entry = (swp_entry_t){0}; | |
898 | ||
899 | again: | |
900 | init_rss_vec(rss); | |
901 | ||
902 | dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl); | |
903 | if (!dst_pte) | |
904 | return -ENOMEM; | |
905 | src_pte = pte_offset_map(src_pmd, addr); | |
906 | src_ptl = pte_lockptr(src_mm, src_pmd); | |
907 | spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); | |
908 | orig_src_pte = src_pte; | |
909 | orig_dst_pte = dst_pte; | |
910 | arch_enter_lazy_mmu_mode(); | |
911 | ||
912 | do { | |
913 | /* | |
914 | * We are holding two locks at this point - either of them | |
915 | * could generate latencies in another task on another CPU. | |
916 | */ | |
917 | if (progress >= 32) { | |
918 | progress = 0; | |
919 | if (need_resched() || | |
920 | spin_needbreak(src_ptl) || spin_needbreak(dst_ptl)) | |
921 | break; | |
922 | } | |
923 | if (pte_none(*src_pte)) { | |
924 | progress++; | |
925 | continue; | |
926 | } | |
927 | entry.val = copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, | |
928 | vma, addr, rss); | |
929 | if (entry.val) | |
930 | break; | |
931 | progress += 8; | |
932 | } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end); | |
933 | ||
934 | arch_leave_lazy_mmu_mode(); | |
935 | spin_unlock(src_ptl); | |
936 | pte_unmap(orig_src_pte); | |
937 | add_mm_rss_vec(dst_mm, rss); | |
938 | pte_unmap_unlock(orig_dst_pte, dst_ptl); | |
939 | cond_resched(); | |
940 | ||
941 | if (entry.val) { | |
942 | if (add_swap_count_continuation(entry, GFP_KERNEL) < 0) | |
943 | return -ENOMEM; | |
944 | progress = 0; | |
945 | } | |
946 | if (addr != end) | |
947 | goto again; | |
948 | return 0; | |
949 | } | |
950 | ||
951 | static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
952 | pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma, | |
953 | unsigned long addr, unsigned long end) | |
954 | { | |
955 | pmd_t *src_pmd, *dst_pmd; | |
956 | unsigned long next; | |
957 | ||
958 | dst_pmd = pmd_alloc(dst_mm, dst_pud, addr); | |
959 | if (!dst_pmd) | |
960 | return -ENOMEM; | |
961 | src_pmd = pmd_offset(src_pud, addr); | |
962 | do { | |
963 | next = pmd_addr_end(addr, end); | |
964 | if (pmd_trans_huge(*src_pmd)) { | |
965 | int err; | |
966 | VM_BUG_ON(next-addr != HPAGE_PMD_SIZE); | |
967 | err = copy_huge_pmd(dst_mm, src_mm, | |
968 | dst_pmd, src_pmd, addr, vma); | |
969 | if (err == -ENOMEM) | |
970 | return -ENOMEM; | |
971 | if (!err) | |
972 | continue; | |
973 | /* fall through */ | |
974 | } | |
975 | if (pmd_none_or_clear_bad(src_pmd)) | |
976 | continue; | |
977 | if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd, | |
978 | vma, addr, next)) | |
979 | return -ENOMEM; | |
980 | } while (dst_pmd++, src_pmd++, addr = next, addr != end); | |
981 | return 0; | |
982 | } | |
983 | ||
984 | static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
985 | pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma, | |
986 | unsigned long addr, unsigned long end) | |
987 | { | |
988 | pud_t *src_pud, *dst_pud; | |
989 | unsigned long next; | |
990 | ||
991 | dst_pud = pud_alloc(dst_mm, dst_pgd, addr); | |
992 | if (!dst_pud) | |
993 | return -ENOMEM; | |
994 | src_pud = pud_offset(src_pgd, addr); | |
995 | do { | |
996 | next = pud_addr_end(addr, end); | |
997 | if (pud_none_or_clear_bad(src_pud)) | |
998 | continue; | |
999 | if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud, | |
1000 | vma, addr, next)) | |
1001 | return -ENOMEM; | |
1002 | } while (dst_pud++, src_pud++, addr = next, addr != end); | |
1003 | return 0; | |
1004 | } | |
1005 | ||
1006 | int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm, | |
1007 | struct vm_area_struct *vma) | |
1008 | { | |
1009 | pgd_t *src_pgd, *dst_pgd; | |
1010 | unsigned long next; | |
1011 | unsigned long addr = vma->vm_start; | |
1012 | unsigned long end = vma->vm_end; | |
1013 | unsigned long mmun_start; /* For mmu_notifiers */ | |
1014 | unsigned long mmun_end; /* For mmu_notifiers */ | |
1015 | bool is_cow; | |
1016 | int ret; | |
1017 | ||
1018 | /* | |
1019 | * Don't copy ptes where a page fault will fill them correctly. | |
1020 | * Fork becomes much lighter when there are big shared or private | |
1021 | * readonly mappings. The tradeoff is that copy_page_range is more | |
1022 | * efficient than faulting. | |
1023 | */ | |
1024 | if (!(vma->vm_flags & (VM_HUGETLB | VM_PFNMAP | VM_MIXEDMAP)) && | |
1025 | !vma->anon_vma) | |
1026 | return 0; | |
1027 | ||
1028 | if (is_vm_hugetlb_page(vma)) | |
1029 | return copy_hugetlb_page_range(dst_mm, src_mm, vma); | |
1030 | ||
1031 | if (unlikely(vma->vm_flags & VM_PFNMAP)) { | |
1032 | /* | |
1033 | * We do not free on error cases below as remove_vma | |
1034 | * gets called on error from higher level routine | |
1035 | */ | |
1036 | ret = track_pfn_copy(vma); | |
1037 | if (ret) | |
1038 | return ret; | |
1039 | } | |
1040 | ||
1041 | /* | |
1042 | * We need to invalidate the secondary MMU mappings only when | |
1043 | * there could be a permission downgrade on the ptes of the | |
1044 | * parent mm. And a permission downgrade will only happen if | |
1045 | * is_cow_mapping() returns true. | |
1046 | */ | |
1047 | is_cow = is_cow_mapping(vma->vm_flags); | |
1048 | mmun_start = addr; | |
1049 | mmun_end = end; | |
1050 | if (is_cow) | |
1051 | mmu_notifier_invalidate_range_start(src_mm, mmun_start, | |
1052 | mmun_end); | |
1053 | ||
1054 | ret = 0; | |
1055 | dst_pgd = pgd_offset(dst_mm, addr); | |
1056 | src_pgd = pgd_offset(src_mm, addr); | |
1057 | do { | |
1058 | next = pgd_addr_end(addr, end); | |
1059 | if (pgd_none_or_clear_bad(src_pgd)) | |
1060 | continue; | |
1061 | if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, | |
1062 | vma, addr, next))) { | |
1063 | ret = -ENOMEM; | |
1064 | break; | |
1065 | } | |
1066 | } while (dst_pgd++, src_pgd++, addr = next, addr != end); | |
1067 | ||
1068 | if (is_cow) | |
1069 | mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end); | |
1070 | return ret; | |
1071 | } | |
1072 | ||
1073 | static unsigned long zap_pte_range(struct mmu_gather *tlb, | |
1074 | struct vm_area_struct *vma, pmd_t *pmd, | |
1075 | unsigned long addr, unsigned long end, | |
1076 | struct zap_details *details) | |
1077 | { | |
1078 | struct mm_struct *mm = tlb->mm; | |
1079 | int force_flush = 0; | |
1080 | int rss[NR_MM_COUNTERS]; | |
1081 | spinlock_t *ptl; | |
1082 | pte_t *start_pte; | |
1083 | pte_t *pte; | |
1084 | swp_entry_t entry; | |
1085 | ||
1086 | again: | |
1087 | init_rss_vec(rss); | |
1088 | start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); | |
1089 | pte = start_pte; | |
1090 | arch_enter_lazy_mmu_mode(); | |
1091 | do { | |
1092 | pte_t ptent = *pte; | |
1093 | if (pte_none(ptent)) { | |
1094 | continue; | |
1095 | } | |
1096 | ||
1097 | if (pte_present(ptent)) { | |
1098 | struct page *page; | |
1099 | ||
1100 | page = vm_normal_page(vma, addr, ptent); | |
1101 | if (unlikely(details) && page) { | |
1102 | /* | |
1103 | * unmap_shared_mapping_pages() wants to | |
1104 | * invalidate cache without truncating: | |
1105 | * unmap shared but keep private pages. | |
1106 | */ | |
1107 | if (details->check_mapping && | |
1108 | details->check_mapping != page->mapping) | |
1109 | continue; | |
1110 | } | |
1111 | ptent = ptep_get_and_clear_full(mm, addr, pte, | |
1112 | tlb->fullmm); | |
1113 | tlb_remove_tlb_entry(tlb, pte, addr); | |
1114 | if (unlikely(!page)) | |
1115 | continue; | |
1116 | if (PageAnon(page)) | |
1117 | rss[MM_ANONPAGES]--; | |
1118 | else { | |
1119 | if (pte_dirty(ptent)) { | |
1120 | force_flush = 1; | |
1121 | set_page_dirty(page); | |
1122 | } | |
1123 | if (pte_young(ptent) && | |
1124 | likely(!(vma->vm_flags & VM_SEQ_READ))) | |
1125 | mark_page_accessed(page); | |
1126 | rss[MM_FILEPAGES]--; | |
1127 | } | |
1128 | page_remove_rmap(page); | |
1129 | if (unlikely(page_mapcount(page) < 0)) | |
1130 | print_bad_pte(vma, addr, ptent, page); | |
1131 | if (unlikely(!__tlb_remove_page(tlb, page))) { | |
1132 | force_flush = 1; | |
1133 | addr += PAGE_SIZE; | |
1134 | break; | |
1135 | } | |
1136 | continue; | |
1137 | } | |
1138 | /* If details->check_mapping, we leave swap entries. */ | |
1139 | if (unlikely(details)) | |
1140 | continue; | |
1141 | ||
1142 | entry = pte_to_swp_entry(ptent); | |
1143 | if (!non_swap_entry(entry)) | |
1144 | rss[MM_SWAPENTS]--; | |
1145 | else if (is_migration_entry(entry)) { | |
1146 | struct page *page; | |
1147 | ||
1148 | page = migration_entry_to_page(entry); | |
1149 | ||
1150 | if (PageAnon(page)) | |
1151 | rss[MM_ANONPAGES]--; | |
1152 | else | |
1153 | rss[MM_FILEPAGES]--; | |
1154 | } | |
1155 | if (unlikely(!free_swap_and_cache(entry))) | |
1156 | print_bad_pte(vma, addr, ptent, NULL); | |
1157 | pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); | |
1158 | } while (pte++, addr += PAGE_SIZE, addr != end); | |
1159 | ||
1160 | add_mm_rss_vec(mm, rss); | |
1161 | arch_leave_lazy_mmu_mode(); | |
1162 | ||
1163 | /* Do the actual TLB flush before dropping ptl */ | |
1164 | if (force_flush) | |
1165 | tlb_flush_mmu_tlbonly(tlb); | |
1166 | pte_unmap_unlock(start_pte, ptl); | |
1167 | ||
1168 | /* | |
1169 | * If we forced a TLB flush (either due to running out of | |
1170 | * batch buffers or because we needed to flush dirty TLB | |
1171 | * entries before releasing the ptl), free the batched | |
1172 | * memory too. Restart if we didn't do everything. | |
1173 | */ | |
1174 | if (force_flush) { | |
1175 | force_flush = 0; | |
1176 | tlb_flush_mmu_free(tlb); | |
1177 | ||
1178 | if (addr != end) | |
1179 | goto again; | |
1180 | } | |
1181 | ||
1182 | return addr; | |
1183 | } | |
1184 | ||
1185 | static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, | |
1186 | struct vm_area_struct *vma, pud_t *pud, | |
1187 | unsigned long addr, unsigned long end, | |
1188 | struct zap_details *details) | |
1189 | { | |
1190 | pmd_t *pmd; | |
1191 | unsigned long next; | |
1192 | ||
1193 | pmd = pmd_offset(pud, addr); | |
1194 | do { | |
1195 | next = pmd_addr_end(addr, end); | |
1196 | if (pmd_trans_huge(*pmd)) { | |
1197 | if (next - addr != HPAGE_PMD_SIZE) { | |
1198 | #ifdef CONFIG_DEBUG_VM | |
1199 | if (!rwsem_is_locked(&tlb->mm->mmap_sem)) { | |
1200 | pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n", | |
1201 | __func__, addr, end, | |
1202 | vma->vm_start, | |
1203 | vma->vm_end); | |
1204 | BUG(); | |
1205 | } | |
1206 | #endif | |
1207 | split_huge_page_pmd(vma, addr, pmd); | |
1208 | } else if (zap_huge_pmd(tlb, vma, pmd, addr)) | |
1209 | goto next; | |
1210 | /* fall through */ | |
1211 | } | |
1212 | /* | |
1213 | * Here there can be other concurrent MADV_DONTNEED or | |
1214 | * trans huge page faults running, and if the pmd is | |
1215 | * none or trans huge it can change under us. This is | |
1216 | * because MADV_DONTNEED holds the mmap_sem in read | |
1217 | * mode. | |
1218 | */ | |
1219 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) | |
1220 | goto next; | |
1221 | next = zap_pte_range(tlb, vma, pmd, addr, next, details); | |
1222 | next: | |
1223 | cond_resched(); | |
1224 | } while (pmd++, addr = next, addr != end); | |
1225 | ||
1226 | return addr; | |
1227 | } | |
1228 | ||
1229 | static inline unsigned long zap_pud_range(struct mmu_gather *tlb, | |
1230 | struct vm_area_struct *vma, pgd_t *pgd, | |
1231 | unsigned long addr, unsigned long end, | |
1232 | struct zap_details *details) | |
1233 | { | |
1234 | pud_t *pud; | |
1235 | unsigned long next; | |
1236 | ||
1237 | pud = pud_offset(pgd, addr); | |
1238 | do { | |
1239 | next = pud_addr_end(addr, end); | |
1240 | if (pud_none_or_clear_bad(pud)) | |
1241 | continue; | |
1242 | next = zap_pmd_range(tlb, vma, pud, addr, next, details); | |
1243 | } while (pud++, addr = next, addr != end); | |
1244 | ||
1245 | return addr; | |
1246 | } | |
1247 | ||
1248 | static void unmap_page_range(struct mmu_gather *tlb, | |
1249 | struct vm_area_struct *vma, | |
1250 | unsigned long addr, unsigned long end, | |
1251 | struct zap_details *details) | |
1252 | { | |
1253 | pgd_t *pgd; | |
1254 | unsigned long next; | |
1255 | ||
1256 | if (details && !details->check_mapping) | |
1257 | details = NULL; | |
1258 | ||
1259 | BUG_ON(addr >= end); | |
1260 | tlb_start_vma(tlb, vma); | |
1261 | pgd = pgd_offset(vma->vm_mm, addr); | |
1262 | do { | |
1263 | next = pgd_addr_end(addr, end); | |
1264 | if (pgd_none_or_clear_bad(pgd)) | |
1265 | continue; | |
1266 | next = zap_pud_range(tlb, vma, pgd, addr, next, details); | |
1267 | } while (pgd++, addr = next, addr != end); | |
1268 | tlb_end_vma(tlb, vma); | |
1269 | } | |
1270 | ||
1271 | ||
1272 | static void unmap_single_vma(struct mmu_gather *tlb, | |
1273 | struct vm_area_struct *vma, unsigned long start_addr, | |
1274 | unsigned long end_addr, | |
1275 | struct zap_details *details) | |
1276 | { | |
1277 | unsigned long start = max(vma->vm_start, start_addr); | |
1278 | unsigned long end; | |
1279 | ||
1280 | if (start >= vma->vm_end) | |
1281 | return; | |
1282 | end = min(vma->vm_end, end_addr); | |
1283 | if (end <= vma->vm_start) | |
1284 | return; | |
1285 | ||
1286 | if (vma->vm_file) | |
1287 | uprobe_munmap(vma, start, end); | |
1288 | ||
1289 | if (unlikely(vma->vm_flags & VM_PFNMAP)) | |
1290 | untrack_pfn(vma, 0, 0); | |
1291 | ||
1292 | if (start != end) { | |
1293 | if (unlikely(is_vm_hugetlb_page(vma))) { | |
1294 | /* | |
1295 | * It is undesirable to test vma->vm_file as it | |
1296 | * should be non-null for valid hugetlb area. | |
1297 | * However, vm_file will be NULL in the error | |
1298 | * cleanup path of mmap_region. When | |
1299 | * hugetlbfs ->mmap method fails, | |
1300 | * mmap_region() nullifies vma->vm_file | |
1301 | * before calling this function to clean up. | |
1302 | * Since no pte has actually been setup, it is | |
1303 | * safe to do nothing in this case. | |
1304 | */ | |
1305 | if (vma->vm_file) { | |
1306 | i_mmap_lock_write(vma->vm_file->f_mapping); | |
1307 | __unmap_hugepage_range_final(tlb, vma, start, end, NULL); | |
1308 | i_mmap_unlock_write(vma->vm_file->f_mapping); | |
1309 | } | |
1310 | } else | |
1311 | unmap_page_range(tlb, vma, start, end, details); | |
1312 | } | |
1313 | } | |
1314 | ||
1315 | /** | |
1316 | * unmap_vmas - unmap a range of memory covered by a list of vma's | |
1317 | * @tlb: address of the caller's struct mmu_gather | |
1318 | * @vma: the starting vma | |
1319 | * @start_addr: virtual address at which to start unmapping | |
1320 | * @end_addr: virtual address at which to end unmapping | |
1321 | * | |
1322 | * Unmap all pages in the vma list. | |
1323 | * | |
1324 | * Only addresses between `start' and `end' will be unmapped. | |
1325 | * | |
1326 | * The VMA list must be sorted in ascending virtual address order. | |
1327 | * | |
1328 | * unmap_vmas() assumes that the caller will flush the whole unmapped address | |
1329 | * range after unmap_vmas() returns. So the only responsibility here is to | |
1330 | * ensure that any thus-far unmapped pages are flushed before unmap_vmas() | |
1331 | * drops the lock and schedules. | |
1332 | */ | |
1333 | void unmap_vmas(struct mmu_gather *tlb, | |
1334 | struct vm_area_struct *vma, unsigned long start_addr, | |
1335 | unsigned long end_addr) | |
1336 | { | |
1337 | struct mm_struct *mm = vma->vm_mm; | |
1338 | ||
1339 | mmu_notifier_invalidate_range_start(mm, start_addr, end_addr); | |
1340 | for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) | |
1341 | unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); | |
1342 | mmu_notifier_invalidate_range_end(mm, start_addr, end_addr); | |
1343 | } | |
1344 | ||
1345 | /** | |
1346 | * zap_page_range - remove user pages in a given range | |
1347 | * @vma: vm_area_struct holding the applicable pages | |
1348 | * @start: starting address of pages to zap | |
1349 | * @size: number of bytes to zap | |
1350 | * @details: details of shared cache invalidation | |
1351 | * | |
1352 | * Caller must protect the VMA list | |
1353 | */ | |
1354 | void zap_page_range(struct vm_area_struct *vma, unsigned long start, | |
1355 | unsigned long size, struct zap_details *details) | |
1356 | { | |
1357 | struct mm_struct *mm = vma->vm_mm; | |
1358 | struct mmu_gather tlb; | |
1359 | unsigned long end = start + size; | |
1360 | ||
1361 | lru_add_drain(); | |
1362 | tlb_gather_mmu(&tlb, mm, start, end); | |
1363 | update_hiwater_rss(mm); | |
1364 | mmu_notifier_invalidate_range_start(mm, start, end); | |
1365 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) | |
1366 | unmap_single_vma(&tlb, vma, start, end, details); | |
1367 | mmu_notifier_invalidate_range_end(mm, start, end); | |
1368 | tlb_finish_mmu(&tlb, start, end); | |
1369 | } | |
1370 | ||
1371 | /** | |
1372 | * zap_page_range_single - remove user pages in a given range | |
1373 | * @vma: vm_area_struct holding the applicable pages | |
1374 | * @address: starting address of pages to zap | |
1375 | * @size: number of bytes to zap | |
1376 | * @details: details of shared cache invalidation | |
1377 | * | |
1378 | * The range must fit into one VMA. | |
1379 | */ | |
1380 | static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, | |
1381 | unsigned long size, struct zap_details *details) | |
1382 | { | |
1383 | struct mm_struct *mm = vma->vm_mm; | |
1384 | struct mmu_gather tlb; | |
1385 | unsigned long end = address + size; | |
1386 | ||
1387 | lru_add_drain(); | |
1388 | tlb_gather_mmu(&tlb, mm, address, end); | |
1389 | update_hiwater_rss(mm); | |
1390 | mmu_notifier_invalidate_range_start(mm, address, end); | |
1391 | unmap_single_vma(&tlb, vma, address, end, details); | |
1392 | mmu_notifier_invalidate_range_end(mm, address, end); | |
1393 | tlb_finish_mmu(&tlb, address, end); | |
1394 | } | |
1395 | ||
1396 | /** | |
1397 | * zap_vma_ptes - remove ptes mapping the vma | |
1398 | * @vma: vm_area_struct holding ptes to be zapped | |
1399 | * @address: starting address of pages to zap | |
1400 | * @size: number of bytes to zap | |
1401 | * | |
1402 | * This function only unmaps ptes assigned to VM_PFNMAP vmas. | |
1403 | * | |
1404 | * The entire address range must be fully contained within the vma. | |
1405 | * | |
1406 | * Returns 0 if successful. | |
1407 | */ | |
1408 | int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, | |
1409 | unsigned long size) | |
1410 | { | |
1411 | if (address < vma->vm_start || address + size > vma->vm_end || | |
1412 | !(vma->vm_flags & VM_PFNMAP)) | |
1413 | return -1; | |
1414 | zap_page_range_single(vma, address, size, NULL); | |
1415 | return 0; | |
1416 | } | |
1417 | EXPORT_SYMBOL_GPL(zap_vma_ptes); | |
1418 | ||
1419 | pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, | |
1420 | spinlock_t **ptl) | |
1421 | { | |
1422 | pgd_t * pgd = pgd_offset(mm, addr); | |
1423 | pud_t * pud = pud_alloc(mm, pgd, addr); | |
1424 | if (pud) { | |
1425 | pmd_t * pmd = pmd_alloc(mm, pud, addr); | |
1426 | if (pmd) { | |
1427 | VM_BUG_ON(pmd_trans_huge(*pmd)); | |
1428 | return pte_alloc_map_lock(mm, pmd, addr, ptl); | |
1429 | } | |
1430 | } | |
1431 | return NULL; | |
1432 | } | |
1433 | ||
1434 | /* | |
1435 | * This is the old fallback for page remapping. | |
1436 | * | |
1437 | * For historical reasons, it only allows reserved pages. Only | |
1438 | * old drivers should use this, and they needed to mark their | |
1439 | * pages reserved for the old functions anyway. | |
1440 | */ | |
1441 | static int insert_page(struct vm_area_struct *vma, unsigned long addr, | |
1442 | struct page *page, pgprot_t prot) | |
1443 | { | |
1444 | struct mm_struct *mm = vma->vm_mm; | |
1445 | int retval; | |
1446 | pte_t *pte; | |
1447 | spinlock_t *ptl; | |
1448 | ||
1449 | retval = -EINVAL; | |
1450 | if (PageAnon(page)) | |
1451 | goto out; | |
1452 | retval = -ENOMEM; | |
1453 | flush_dcache_page(page); | |
1454 | pte = get_locked_pte(mm, addr, &ptl); | |
1455 | if (!pte) | |
1456 | goto out; | |
1457 | retval = -EBUSY; | |
1458 | if (!pte_none(*pte)) | |
1459 | goto out_unlock; | |
1460 | ||
1461 | /* Ok, finally just insert the thing.. */ | |
1462 | get_page(page); | |
1463 | inc_mm_counter_fast(mm, MM_FILEPAGES); | |
1464 | page_add_file_rmap(page); | |
1465 | set_pte_at(mm, addr, pte, mk_pte(page, prot)); | |
1466 | ||
1467 | retval = 0; | |
1468 | pte_unmap_unlock(pte, ptl); | |
1469 | return retval; | |
1470 | out_unlock: | |
1471 | pte_unmap_unlock(pte, ptl); | |
1472 | out: | |
1473 | return retval; | |
1474 | } | |
1475 | ||
1476 | /** | |
1477 | * vm_insert_page - insert single page into user vma | |
1478 | * @vma: user vma to map to | |
1479 | * @addr: target user address of this page | |
1480 | * @page: source kernel page | |
1481 | * | |
1482 | * This allows drivers to insert individual pages they've allocated | |
1483 | * into a user vma. | |
1484 | * | |
1485 | * The page has to be a nice clean _individual_ kernel allocation. | |
1486 | * If you allocate a compound page, you need to have marked it as | |
1487 | * such (__GFP_COMP), or manually just split the page up yourself | |
1488 | * (see split_page()). | |
1489 | * | |
1490 | * NOTE! Traditionally this was done with "remap_pfn_range()" which | |
1491 | * took an arbitrary page protection parameter. This doesn't allow | |
1492 | * that. Your vma protection will have to be set up correctly, which | |
1493 | * means that if you want a shared writable mapping, you'd better | |
1494 | * ask for a shared writable mapping! | |
1495 | * | |
1496 | * The page does not need to be reserved. | |
1497 | * | |
1498 | * Usually this function is called from f_op->mmap() handler | |
1499 | * under mm->mmap_sem write-lock, so it can change vma->vm_flags. | |
1500 | * Caller must set VM_MIXEDMAP on vma if it wants to call this | |
1501 | * function from other places, for example from page-fault handler. | |
1502 | */ | |
1503 | int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, | |
1504 | struct page *page) | |
1505 | { | |
1506 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
1507 | return -EFAULT; | |
1508 | if (!page_count(page)) | |
1509 | return -EINVAL; | |
1510 | if (!(vma->vm_flags & VM_MIXEDMAP)) { | |
1511 | BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem)); | |
1512 | BUG_ON(vma->vm_flags & VM_PFNMAP); | |
1513 | vma->vm_flags |= VM_MIXEDMAP; | |
1514 | } | |
1515 | return insert_page(vma, addr, page, vma->vm_page_prot); | |
1516 | } | |
1517 | EXPORT_SYMBOL(vm_insert_page); | |
1518 | ||
1519 | static int insert_pfn(struct vm_area_struct *vma, unsigned long addr, | |
1520 | unsigned long pfn, pgprot_t prot) | |
1521 | { | |
1522 | struct mm_struct *mm = vma->vm_mm; | |
1523 | int retval; | |
1524 | pte_t *pte, entry; | |
1525 | spinlock_t *ptl; | |
1526 | ||
1527 | retval = -ENOMEM; | |
1528 | pte = get_locked_pte(mm, addr, &ptl); | |
1529 | if (!pte) | |
1530 | goto out; | |
1531 | retval = -EBUSY; | |
1532 | if (!pte_none(*pte)) | |
1533 | goto out_unlock; | |
1534 | ||
1535 | /* Ok, finally just insert the thing.. */ | |
1536 | entry = pte_mkspecial(pfn_pte(pfn, prot)); | |
1537 | set_pte_at(mm, addr, pte, entry); | |
1538 | update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ | |
1539 | ||
1540 | retval = 0; | |
1541 | out_unlock: | |
1542 | pte_unmap_unlock(pte, ptl); | |
1543 | out: | |
1544 | return retval; | |
1545 | } | |
1546 | ||
1547 | /** | |
1548 | * vm_insert_pfn - insert single pfn into user vma | |
1549 | * @vma: user vma to map to | |
1550 | * @addr: target user address of this page | |
1551 | * @pfn: source kernel pfn | |
1552 | * | |
1553 | * Similar to vm_insert_page, this allows drivers to insert individual pages | |
1554 | * they've allocated into a user vma. Same comments apply. | |
1555 | * | |
1556 | * This function should only be called from a vm_ops->fault handler, and | |
1557 | * in that case the handler should return NULL. | |
1558 | * | |
1559 | * vma cannot be a COW mapping. | |
1560 | * | |
1561 | * As this is called only for pages that do not currently exist, we | |
1562 | * do not need to flush old virtual caches or the TLB. | |
1563 | */ | |
1564 | int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr, | |
1565 | unsigned long pfn) | |
1566 | { | |
1567 | int ret; | |
1568 | pgprot_t pgprot = vma->vm_page_prot; | |
1569 | /* | |
1570 | * Technically, architectures with pte_special can avoid all these | |
1571 | * restrictions (same for remap_pfn_range). However we would like | |
1572 | * consistency in testing and feature parity among all, so we should | |
1573 | * try to keep these invariants in place for everybody. | |
1574 | */ | |
1575 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); | |
1576 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == | |
1577 | (VM_PFNMAP|VM_MIXEDMAP)); | |
1578 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); | |
1579 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); | |
1580 | ||
1581 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
1582 | return -EFAULT; | |
1583 | if (track_pfn_insert(vma, &pgprot, pfn)) | |
1584 | return -EINVAL; | |
1585 | ||
1586 | ret = insert_pfn(vma, addr, pfn, pgprot); | |
1587 | ||
1588 | return ret; | |
1589 | } | |
1590 | EXPORT_SYMBOL(vm_insert_pfn); | |
1591 | ||
1592 | int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr, | |
1593 | unsigned long pfn) | |
1594 | { | |
1595 | BUG_ON(!(vma->vm_flags & VM_MIXEDMAP)); | |
1596 | ||
1597 | if (addr < vma->vm_start || addr >= vma->vm_end) | |
1598 | return -EFAULT; | |
1599 | ||
1600 | /* | |
1601 | * If we don't have pte special, then we have to use the pfn_valid() | |
1602 | * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* | |
1603 | * refcount the page if pfn_valid is true (hence insert_page rather | |
1604 | * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP | |
1605 | * without pte special, it would there be refcounted as a normal page. | |
1606 | */ | |
1607 | if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) { | |
1608 | struct page *page; | |
1609 | ||
1610 | page = pfn_to_page(pfn); | |
1611 | return insert_page(vma, addr, page, vma->vm_page_prot); | |
1612 | } | |
1613 | return insert_pfn(vma, addr, pfn, vma->vm_page_prot); | |
1614 | } | |
1615 | EXPORT_SYMBOL(vm_insert_mixed); | |
1616 | ||
1617 | /* | |
1618 | * maps a range of physical memory into the requested pages. the old | |
1619 | * mappings are removed. any references to nonexistent pages results | |
1620 | * in null mappings (currently treated as "copy-on-access") | |
1621 | */ | |
1622 | static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, | |
1623 | unsigned long addr, unsigned long end, | |
1624 | unsigned long pfn, pgprot_t prot) | |
1625 | { | |
1626 | pte_t *pte; | |
1627 | spinlock_t *ptl; | |
1628 | ||
1629 | pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); | |
1630 | if (!pte) | |
1631 | return -ENOMEM; | |
1632 | arch_enter_lazy_mmu_mode(); | |
1633 | do { | |
1634 | BUG_ON(!pte_none(*pte)); | |
1635 | set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); | |
1636 | pfn++; | |
1637 | } while (pte++, addr += PAGE_SIZE, addr != end); | |
1638 | arch_leave_lazy_mmu_mode(); | |
1639 | pte_unmap_unlock(pte - 1, ptl); | |
1640 | return 0; | |
1641 | } | |
1642 | ||
1643 | static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, | |
1644 | unsigned long addr, unsigned long end, | |
1645 | unsigned long pfn, pgprot_t prot) | |
1646 | { | |
1647 | pmd_t *pmd; | |
1648 | unsigned long next; | |
1649 | ||
1650 | pfn -= addr >> PAGE_SHIFT; | |
1651 | pmd = pmd_alloc(mm, pud, addr); | |
1652 | if (!pmd) | |
1653 | return -ENOMEM; | |
1654 | VM_BUG_ON(pmd_trans_huge(*pmd)); | |
1655 | do { | |
1656 | next = pmd_addr_end(addr, end); | |
1657 | if (remap_pte_range(mm, pmd, addr, next, | |
1658 | pfn + (addr >> PAGE_SHIFT), prot)) | |
1659 | return -ENOMEM; | |
1660 | } while (pmd++, addr = next, addr != end); | |
1661 | return 0; | |
1662 | } | |
1663 | ||
1664 | static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, | |
1665 | unsigned long addr, unsigned long end, | |
1666 | unsigned long pfn, pgprot_t prot) | |
1667 | { | |
1668 | pud_t *pud; | |
1669 | unsigned long next; | |
1670 | ||
1671 | pfn -= addr >> PAGE_SHIFT; | |
1672 | pud = pud_alloc(mm, pgd, addr); | |
1673 | if (!pud) | |
1674 | return -ENOMEM; | |
1675 | do { | |
1676 | next = pud_addr_end(addr, end); | |
1677 | if (remap_pmd_range(mm, pud, addr, next, | |
1678 | pfn + (addr >> PAGE_SHIFT), prot)) | |
1679 | return -ENOMEM; | |
1680 | } while (pud++, addr = next, addr != end); | |
1681 | return 0; | |
1682 | } | |
1683 | ||
1684 | /** | |
1685 | * remap_pfn_range - remap kernel memory to userspace | |
1686 | * @vma: user vma to map to | |
1687 | * @addr: target user address to start at | |
1688 | * @pfn: physical address of kernel memory | |
1689 | * @size: size of map area | |
1690 | * @prot: page protection flags for this mapping | |
1691 | * | |
1692 | * Note: this is only safe if the mm semaphore is held when called. | |
1693 | */ | |
1694 | int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, | |
1695 | unsigned long pfn, unsigned long size, pgprot_t prot) | |
1696 | { | |
1697 | pgd_t *pgd; | |
1698 | unsigned long next; | |
1699 | unsigned long end = addr + PAGE_ALIGN(size); | |
1700 | struct mm_struct *mm = vma->vm_mm; | |
1701 | int err; | |
1702 | ||
1703 | /* | |
1704 | * Physically remapped pages are special. Tell the | |
1705 | * rest of the world about it: | |
1706 | * VM_IO tells people not to look at these pages | |
1707 | * (accesses can have side effects). | |
1708 | * VM_PFNMAP tells the core MM that the base pages are just | |
1709 | * raw PFN mappings, and do not have a "struct page" associated | |
1710 | * with them. | |
1711 | * VM_DONTEXPAND | |
1712 | * Disable vma merging and expanding with mremap(). | |
1713 | * VM_DONTDUMP | |
1714 | * Omit vma from core dump, even when VM_IO turned off. | |
1715 | * | |
1716 | * There's a horrible special case to handle copy-on-write | |
1717 | * behaviour that some programs depend on. We mark the "original" | |
1718 | * un-COW'ed pages by matching them up with "vma->vm_pgoff". | |
1719 | * See vm_normal_page() for details. | |
1720 | */ | |
1721 | if (is_cow_mapping(vma->vm_flags)) { | |
1722 | if (addr != vma->vm_start || end != vma->vm_end) | |
1723 | return -EINVAL; | |
1724 | vma->vm_pgoff = pfn; | |
1725 | } | |
1726 | ||
1727 | err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); | |
1728 | if (err) | |
1729 | return -EINVAL; | |
1730 | ||
1731 | vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; | |
1732 | ||
1733 | BUG_ON(addr >= end); | |
1734 | pfn -= addr >> PAGE_SHIFT; | |
1735 | pgd = pgd_offset(mm, addr); | |
1736 | flush_cache_range(vma, addr, end); | |
1737 | do { | |
1738 | next = pgd_addr_end(addr, end); | |
1739 | err = remap_pud_range(mm, pgd, addr, next, | |
1740 | pfn + (addr >> PAGE_SHIFT), prot); | |
1741 | if (err) | |
1742 | break; | |
1743 | } while (pgd++, addr = next, addr != end); | |
1744 | ||
1745 | if (err) | |
1746 | untrack_pfn(vma, pfn, PAGE_ALIGN(size)); | |
1747 | ||
1748 | return err; | |
1749 | } | |
1750 | EXPORT_SYMBOL(remap_pfn_range); | |
1751 | ||
1752 | /** | |
1753 | * vm_iomap_memory - remap memory to userspace | |
1754 | * @vma: user vma to map to | |
1755 | * @start: start of area | |
1756 | * @len: size of area | |
1757 | * | |
1758 | * This is a simplified io_remap_pfn_range() for common driver use. The | |
1759 | * driver just needs to give us the physical memory range to be mapped, | |
1760 | * we'll figure out the rest from the vma information. | |
1761 | * | |
1762 | * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get | |
1763 | * whatever write-combining details or similar. | |
1764 | */ | |
1765 | int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) | |
1766 | { | |
1767 | unsigned long vm_len, pfn, pages; | |
1768 | ||
1769 | /* Check that the physical memory area passed in looks valid */ | |
1770 | if (start + len < start) | |
1771 | return -EINVAL; | |
1772 | /* | |
1773 | * You *really* shouldn't map things that aren't page-aligned, | |
1774 | * but we've historically allowed it because IO memory might | |
1775 | * just have smaller alignment. | |
1776 | */ | |
1777 | len += start & ~PAGE_MASK; | |
1778 | pfn = start >> PAGE_SHIFT; | |
1779 | pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; | |
1780 | if (pfn + pages < pfn) | |
1781 | return -EINVAL; | |
1782 | ||
1783 | /* We start the mapping 'vm_pgoff' pages into the area */ | |
1784 | if (vma->vm_pgoff > pages) | |
1785 | return -EINVAL; | |
1786 | pfn += vma->vm_pgoff; | |
1787 | pages -= vma->vm_pgoff; | |
1788 | ||
1789 | /* Can we fit all of the mapping? */ | |
1790 | vm_len = vma->vm_end - vma->vm_start; | |
1791 | if (vm_len >> PAGE_SHIFT > pages) | |
1792 | return -EINVAL; | |
1793 | ||
1794 | /* Ok, let it rip */ | |
1795 | return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); | |
1796 | } | |
1797 | EXPORT_SYMBOL(vm_iomap_memory); | |
1798 | ||
1799 | static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, | |
1800 | unsigned long addr, unsigned long end, | |
1801 | pte_fn_t fn, void *data) | |
1802 | { | |
1803 | pte_t *pte; | |
1804 | int err; | |
1805 | pgtable_t token; | |
1806 | spinlock_t *uninitialized_var(ptl); | |
1807 | ||
1808 | pte = (mm == &init_mm) ? | |
1809 | pte_alloc_kernel(pmd, addr) : | |
1810 | pte_alloc_map_lock(mm, pmd, addr, &ptl); | |
1811 | if (!pte) | |
1812 | return -ENOMEM; | |
1813 | ||
1814 | BUG_ON(pmd_huge(*pmd)); | |
1815 | ||
1816 | arch_enter_lazy_mmu_mode(); | |
1817 | ||
1818 | token = pmd_pgtable(*pmd); | |
1819 | ||
1820 | do { | |
1821 | err = fn(pte++, token, addr, data); | |
1822 | if (err) | |
1823 | break; | |
1824 | } while (addr += PAGE_SIZE, addr != end); | |
1825 | ||
1826 | arch_leave_lazy_mmu_mode(); | |
1827 | ||
1828 | if (mm != &init_mm) | |
1829 | pte_unmap_unlock(pte-1, ptl); | |
1830 | return err; | |
1831 | } | |
1832 | ||
1833 | static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, | |
1834 | unsigned long addr, unsigned long end, | |
1835 | pte_fn_t fn, void *data) | |
1836 | { | |
1837 | pmd_t *pmd; | |
1838 | unsigned long next; | |
1839 | int err; | |
1840 | ||
1841 | BUG_ON(pud_huge(*pud)); | |
1842 | ||
1843 | pmd = pmd_alloc(mm, pud, addr); | |
1844 | if (!pmd) | |
1845 | return -ENOMEM; | |
1846 | do { | |
1847 | next = pmd_addr_end(addr, end); | |
1848 | err = apply_to_pte_range(mm, pmd, addr, next, fn, data); | |
1849 | if (err) | |
1850 | break; | |
1851 | } while (pmd++, addr = next, addr != end); | |
1852 | return err; | |
1853 | } | |
1854 | ||
1855 | static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, | |
1856 | unsigned long addr, unsigned long end, | |
1857 | pte_fn_t fn, void *data) | |
1858 | { | |
1859 | pud_t *pud; | |
1860 | unsigned long next; | |
1861 | int err; | |
1862 | ||
1863 | pud = pud_alloc(mm, pgd, addr); | |
1864 | if (!pud) | |
1865 | return -ENOMEM; | |
1866 | do { | |
1867 | next = pud_addr_end(addr, end); | |
1868 | err = apply_to_pmd_range(mm, pud, addr, next, fn, data); | |
1869 | if (err) | |
1870 | break; | |
1871 | } while (pud++, addr = next, addr != end); | |
1872 | return err; | |
1873 | } | |
1874 | ||
1875 | /* | |
1876 | * Scan a region of virtual memory, filling in page tables as necessary | |
1877 | * and calling a provided function on each leaf page table. | |
1878 | */ | |
1879 | int apply_to_page_range(struct mm_struct *mm, unsigned long addr, | |
1880 | unsigned long size, pte_fn_t fn, void *data) | |
1881 | { | |
1882 | pgd_t *pgd; | |
1883 | unsigned long next; | |
1884 | unsigned long end = addr + size; | |
1885 | int err; | |
1886 | ||
1887 | BUG_ON(addr >= end); | |
1888 | pgd = pgd_offset(mm, addr); | |
1889 | do { | |
1890 | next = pgd_addr_end(addr, end); | |
1891 | err = apply_to_pud_range(mm, pgd, addr, next, fn, data); | |
1892 | if (err) | |
1893 | break; | |
1894 | } while (pgd++, addr = next, addr != end); | |
1895 | ||
1896 | return err; | |
1897 | } | |
1898 | EXPORT_SYMBOL_GPL(apply_to_page_range); | |
1899 | ||
1900 | /* | |
1901 | * handle_pte_fault chooses page fault handler according to an entry which was | |
1902 | * read non-atomically. Before making any commitment, on those architectures | |
1903 | * or configurations (e.g. i386 with PAE) which might give a mix of unmatched | |
1904 | * parts, do_swap_page must check under lock before unmapping the pte and | |
1905 | * proceeding (but do_wp_page is only called after already making such a check; | |
1906 | * and do_anonymous_page can safely check later on). | |
1907 | */ | |
1908 | static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, | |
1909 | pte_t *page_table, pte_t orig_pte) | |
1910 | { | |
1911 | int same = 1; | |
1912 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) | |
1913 | if (sizeof(pte_t) > sizeof(unsigned long)) { | |
1914 | spinlock_t *ptl = pte_lockptr(mm, pmd); | |
1915 | spin_lock(ptl); | |
1916 | same = pte_same(*page_table, orig_pte); | |
1917 | spin_unlock(ptl); | |
1918 | } | |
1919 | #endif | |
1920 | pte_unmap(page_table); | |
1921 | return same; | |
1922 | } | |
1923 | ||
1924 | static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) | |
1925 | { | |
1926 | debug_dma_assert_idle(src); | |
1927 | ||
1928 | /* | |
1929 | * If the source page was a PFN mapping, we don't have | |
1930 | * a "struct page" for it. We do a best-effort copy by | |
1931 | * just copying from the original user address. If that | |
1932 | * fails, we just zero-fill it. Live with it. | |
1933 | */ | |
1934 | if (unlikely(!src)) { | |
1935 | void *kaddr = kmap_atomic(dst); | |
1936 | void __user *uaddr = (void __user *)(va & PAGE_MASK); | |
1937 | ||
1938 | /* | |
1939 | * This really shouldn't fail, because the page is there | |
1940 | * in the page tables. But it might just be unreadable, | |
1941 | * in which case we just give up and fill the result with | |
1942 | * zeroes. | |
1943 | */ | |
1944 | if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) | |
1945 | clear_page(kaddr); | |
1946 | kunmap_atomic(kaddr); | |
1947 | flush_dcache_page(dst); | |
1948 | } else | |
1949 | copy_user_highpage(dst, src, va, vma); | |
1950 | } | |
1951 | ||
1952 | /* | |
1953 | * Notify the address space that the page is about to become writable so that | |
1954 | * it can prohibit this or wait for the page to get into an appropriate state. | |
1955 | * | |
1956 | * We do this without the lock held, so that it can sleep if it needs to. | |
1957 | */ | |
1958 | static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page, | |
1959 | unsigned long address) | |
1960 | { | |
1961 | struct vm_fault vmf; | |
1962 | int ret; | |
1963 | ||
1964 | vmf.virtual_address = (void __user *)(address & PAGE_MASK); | |
1965 | vmf.pgoff = page->index; | |
1966 | vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; | |
1967 | vmf.page = page; | |
1968 | ||
1969 | ret = vma->vm_ops->page_mkwrite(vma, &vmf); | |
1970 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) | |
1971 | return ret; | |
1972 | if (unlikely(!(ret & VM_FAULT_LOCKED))) { | |
1973 | lock_page(page); | |
1974 | if (!page->mapping) { | |
1975 | unlock_page(page); | |
1976 | return 0; /* retry */ | |
1977 | } | |
1978 | ret |= VM_FAULT_LOCKED; | |
1979 | } else | |
1980 | VM_BUG_ON_PAGE(!PageLocked(page), page); | |
1981 | return ret; | |
1982 | } | |
1983 | ||
1984 | /* | |
1985 | * This routine handles present pages, when users try to write | |
1986 | * to a shared page. It is done by copying the page to a new address | |
1987 | * and decrementing the shared-page counter for the old page. | |
1988 | * | |
1989 | * Note that this routine assumes that the protection checks have been | |
1990 | * done by the caller (the low-level page fault routine in most cases). | |
1991 | * Thus we can safely just mark it writable once we've done any necessary | |
1992 | * COW. | |
1993 | * | |
1994 | * We also mark the page dirty at this point even though the page will | |
1995 | * change only once the write actually happens. This avoids a few races, | |
1996 | * and potentially makes it more efficient. | |
1997 | * | |
1998 | * We enter with non-exclusive mmap_sem (to exclude vma changes, | |
1999 | * but allow concurrent faults), with pte both mapped and locked. | |
2000 | * We return with mmap_sem still held, but pte unmapped and unlocked. | |
2001 | */ | |
2002 | static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
2003 | unsigned long address, pte_t *page_table, pmd_t *pmd, | |
2004 | spinlock_t *ptl, pte_t orig_pte) | |
2005 | __releases(ptl) | |
2006 | { | |
2007 | struct page *old_page, *new_page = NULL; | |
2008 | pte_t entry; | |
2009 | int ret = 0; | |
2010 | int page_mkwrite = 0; | |
2011 | bool dirty_shared = false; | |
2012 | unsigned long mmun_start = 0; /* For mmu_notifiers */ | |
2013 | unsigned long mmun_end = 0; /* For mmu_notifiers */ | |
2014 | struct mem_cgroup *memcg; | |
2015 | ||
2016 | old_page = vm_normal_page(vma, address, orig_pte); | |
2017 | if (!old_page) { | |
2018 | /* | |
2019 | * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a | |
2020 | * VM_PFNMAP VMA. | |
2021 | * | |
2022 | * We should not cow pages in a shared writeable mapping. | |
2023 | * Just mark the pages writable as we can't do any dirty | |
2024 | * accounting on raw pfn maps. | |
2025 | */ | |
2026 | if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == | |
2027 | (VM_WRITE|VM_SHARED)) | |
2028 | goto reuse; | |
2029 | goto gotten; | |
2030 | } | |
2031 | ||
2032 | /* | |
2033 | * Take out anonymous pages first, anonymous shared vmas are | |
2034 | * not dirty accountable. | |
2035 | */ | |
2036 | if (PageAnon(old_page) && !PageKsm(old_page)) { | |
2037 | if (!trylock_page(old_page)) { | |
2038 | page_cache_get(old_page); | |
2039 | pte_unmap_unlock(page_table, ptl); | |
2040 | lock_page(old_page); | |
2041 | page_table = pte_offset_map_lock(mm, pmd, address, | |
2042 | &ptl); | |
2043 | if (!pte_same(*page_table, orig_pte)) { | |
2044 | unlock_page(old_page); | |
2045 | goto unlock; | |
2046 | } | |
2047 | page_cache_release(old_page); | |
2048 | } | |
2049 | if (reuse_swap_page(old_page)) { | |
2050 | /* | |
2051 | * The page is all ours. Move it to our anon_vma so | |
2052 | * the rmap code will not search our parent or siblings. | |
2053 | * Protected against the rmap code by the page lock. | |
2054 | */ | |
2055 | page_move_anon_rmap(old_page, vma, address); | |
2056 | unlock_page(old_page); | |
2057 | goto reuse; | |
2058 | } | |
2059 | unlock_page(old_page); | |
2060 | } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == | |
2061 | (VM_WRITE|VM_SHARED))) { | |
2062 | page_cache_get(old_page); | |
2063 | /* | |
2064 | * Only catch write-faults on shared writable pages, | |
2065 | * read-only shared pages can get COWed by | |
2066 | * get_user_pages(.write=1, .force=1). | |
2067 | */ | |
2068 | if (vma->vm_ops && vma->vm_ops->page_mkwrite) { | |
2069 | int tmp; | |
2070 | ||
2071 | pte_unmap_unlock(page_table, ptl); | |
2072 | tmp = do_page_mkwrite(vma, old_page, address); | |
2073 | if (unlikely(!tmp || (tmp & | |
2074 | (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { | |
2075 | page_cache_release(old_page); | |
2076 | return tmp; | |
2077 | } | |
2078 | /* | |
2079 | * Since we dropped the lock we need to revalidate | |
2080 | * the PTE as someone else may have changed it. If | |
2081 | * they did, we just return, as we can count on the | |
2082 | * MMU to tell us if they didn't also make it writable. | |
2083 | */ | |
2084 | page_table = pte_offset_map_lock(mm, pmd, address, | |
2085 | &ptl); | |
2086 | if (!pte_same(*page_table, orig_pte)) { | |
2087 | unlock_page(old_page); | |
2088 | goto unlock; | |
2089 | } | |
2090 | page_mkwrite = 1; | |
2091 | } | |
2092 | ||
2093 | dirty_shared = true; | |
2094 | ||
2095 | reuse: | |
2096 | /* | |
2097 | * Clear the pages cpupid information as the existing | |
2098 | * information potentially belongs to a now completely | |
2099 | * unrelated process. | |
2100 | */ | |
2101 | if (old_page) | |
2102 | page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1); | |
2103 | ||
2104 | flush_cache_page(vma, address, pte_pfn(orig_pte)); | |
2105 | entry = pte_mkyoung(orig_pte); | |
2106 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
2107 | if (ptep_set_access_flags(vma, address, page_table, entry,1)) | |
2108 | update_mmu_cache(vma, address, page_table); | |
2109 | pte_unmap_unlock(page_table, ptl); | |
2110 | ret |= VM_FAULT_WRITE; | |
2111 | ||
2112 | if (dirty_shared) { | |
2113 | struct address_space *mapping; | |
2114 | int dirtied; | |
2115 | ||
2116 | if (!page_mkwrite) | |
2117 | lock_page(old_page); | |
2118 | ||
2119 | dirtied = set_page_dirty(old_page); | |
2120 | VM_BUG_ON_PAGE(PageAnon(old_page), old_page); | |
2121 | mapping = old_page->mapping; | |
2122 | unlock_page(old_page); | |
2123 | page_cache_release(old_page); | |
2124 | ||
2125 | if ((dirtied || page_mkwrite) && mapping) { | |
2126 | /* | |
2127 | * Some device drivers do not set page.mapping | |
2128 | * but still dirty their pages | |
2129 | */ | |
2130 | balance_dirty_pages_ratelimited(mapping); | |
2131 | } | |
2132 | ||
2133 | if (!page_mkwrite) | |
2134 | file_update_time(vma->vm_file); | |
2135 | } | |
2136 | ||
2137 | return ret; | |
2138 | } | |
2139 | ||
2140 | /* | |
2141 | * Ok, we need to copy. Oh, well.. | |
2142 | */ | |
2143 | page_cache_get(old_page); | |
2144 | gotten: | |
2145 | pte_unmap_unlock(page_table, ptl); | |
2146 | ||
2147 | if (unlikely(anon_vma_prepare(vma))) | |
2148 | goto oom; | |
2149 | ||
2150 | if (is_zero_pfn(pte_pfn(orig_pte))) { | |
2151 | new_page = alloc_zeroed_user_highpage_movable(vma, address); | |
2152 | if (!new_page) | |
2153 | goto oom; | |
2154 | } else { | |
2155 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); | |
2156 | if (!new_page) | |
2157 | goto oom; | |
2158 | cow_user_page(new_page, old_page, address, vma); | |
2159 | } | |
2160 | __SetPageUptodate(new_page); | |
2161 | ||
2162 | if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) | |
2163 | goto oom_free_new; | |
2164 | ||
2165 | mmun_start = address & PAGE_MASK; | |
2166 | mmun_end = mmun_start + PAGE_SIZE; | |
2167 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); | |
2168 | ||
2169 | /* | |
2170 | * Re-check the pte - we dropped the lock | |
2171 | */ | |
2172 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2173 | if (likely(pte_same(*page_table, orig_pte))) { | |
2174 | if (old_page) { | |
2175 | if (!PageAnon(old_page)) { | |
2176 | dec_mm_counter_fast(mm, MM_FILEPAGES); | |
2177 | inc_mm_counter_fast(mm, MM_ANONPAGES); | |
2178 | } | |
2179 | } else | |
2180 | inc_mm_counter_fast(mm, MM_ANONPAGES); | |
2181 | flush_cache_page(vma, address, pte_pfn(orig_pte)); | |
2182 | entry = mk_pte(new_page, vma->vm_page_prot); | |
2183 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
2184 | /* | |
2185 | * Clear the pte entry and flush it first, before updating the | |
2186 | * pte with the new entry. This will avoid a race condition | |
2187 | * seen in the presence of one thread doing SMC and another | |
2188 | * thread doing COW. | |
2189 | */ | |
2190 | ptep_clear_flush_notify(vma, address, page_table); | |
2191 | page_add_new_anon_rmap(new_page, vma, address); | |
2192 | mem_cgroup_commit_charge(new_page, memcg, false); | |
2193 | lru_cache_add_active_or_unevictable(new_page, vma); | |
2194 | /* | |
2195 | * We call the notify macro here because, when using secondary | |
2196 | * mmu page tables (such as kvm shadow page tables), we want the | |
2197 | * new page to be mapped directly into the secondary page table. | |
2198 | */ | |
2199 | set_pte_at_notify(mm, address, page_table, entry); | |
2200 | update_mmu_cache(vma, address, page_table); | |
2201 | if (old_page) { | |
2202 | /* | |
2203 | * Only after switching the pte to the new page may | |
2204 | * we remove the mapcount here. Otherwise another | |
2205 | * process may come and find the rmap count decremented | |
2206 | * before the pte is switched to the new page, and | |
2207 | * "reuse" the old page writing into it while our pte | |
2208 | * here still points into it and can be read by other | |
2209 | * threads. | |
2210 | * | |
2211 | * The critical issue is to order this | |
2212 | * page_remove_rmap with the ptp_clear_flush above. | |
2213 | * Those stores are ordered by (if nothing else,) | |
2214 | * the barrier present in the atomic_add_negative | |
2215 | * in page_remove_rmap. | |
2216 | * | |
2217 | * Then the TLB flush in ptep_clear_flush ensures that | |
2218 | * no process can access the old page before the | |
2219 | * decremented mapcount is visible. And the old page | |
2220 | * cannot be reused until after the decremented | |
2221 | * mapcount is visible. So transitively, TLBs to | |
2222 | * old page will be flushed before it can be reused. | |
2223 | */ | |
2224 | page_remove_rmap(old_page); | |
2225 | } | |
2226 | ||
2227 | /* Free the old page.. */ | |
2228 | new_page = old_page; | |
2229 | ret |= VM_FAULT_WRITE; | |
2230 | } else | |
2231 | mem_cgroup_cancel_charge(new_page, memcg); | |
2232 | ||
2233 | if (new_page) | |
2234 | page_cache_release(new_page); | |
2235 | unlock: | |
2236 | pte_unmap_unlock(page_table, ptl); | |
2237 | if (mmun_end > mmun_start) | |
2238 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); | |
2239 | if (old_page) { | |
2240 | /* | |
2241 | * Don't let another task, with possibly unlocked vma, | |
2242 | * keep the mlocked page. | |
2243 | */ | |
2244 | if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) { | |
2245 | lock_page(old_page); /* LRU manipulation */ | |
2246 | munlock_vma_page(old_page); | |
2247 | unlock_page(old_page); | |
2248 | } | |
2249 | page_cache_release(old_page); | |
2250 | } | |
2251 | return ret; | |
2252 | oom_free_new: | |
2253 | page_cache_release(new_page); | |
2254 | oom: | |
2255 | if (old_page) | |
2256 | page_cache_release(old_page); | |
2257 | return VM_FAULT_OOM; | |
2258 | } | |
2259 | ||
2260 | static void unmap_mapping_range_vma(struct vm_area_struct *vma, | |
2261 | unsigned long start_addr, unsigned long end_addr, | |
2262 | struct zap_details *details) | |
2263 | { | |
2264 | zap_page_range_single(vma, start_addr, end_addr - start_addr, details); | |
2265 | } | |
2266 | ||
2267 | static inline void unmap_mapping_range_tree(struct rb_root *root, | |
2268 | struct zap_details *details) | |
2269 | { | |
2270 | struct vm_area_struct *vma; | |
2271 | pgoff_t vba, vea, zba, zea; | |
2272 | ||
2273 | vma_interval_tree_foreach(vma, root, | |
2274 | details->first_index, details->last_index) { | |
2275 | ||
2276 | vba = vma->vm_pgoff; | |
2277 | vea = vba + vma_pages(vma) - 1; | |
2278 | /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ | |
2279 | zba = details->first_index; | |
2280 | if (zba < vba) | |
2281 | zba = vba; | |
2282 | zea = details->last_index; | |
2283 | if (zea > vea) | |
2284 | zea = vea; | |
2285 | ||
2286 | unmap_mapping_range_vma(vma, | |
2287 | ((zba - vba) << PAGE_SHIFT) + vma->vm_start, | |
2288 | ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, | |
2289 | details); | |
2290 | } | |
2291 | } | |
2292 | ||
2293 | /** | |
2294 | * unmap_mapping_range - unmap the portion of all mmaps in the specified | |
2295 | * address_space corresponding to the specified page range in the underlying | |
2296 | * file. | |
2297 | * | |
2298 | * @mapping: the address space containing mmaps to be unmapped. | |
2299 | * @holebegin: byte in first page to unmap, relative to the start of | |
2300 | * the underlying file. This will be rounded down to a PAGE_SIZE | |
2301 | * boundary. Note that this is different from truncate_pagecache(), which | |
2302 | * must keep the partial page. In contrast, we must get rid of | |
2303 | * partial pages. | |
2304 | * @holelen: size of prospective hole in bytes. This will be rounded | |
2305 | * up to a PAGE_SIZE boundary. A holelen of zero truncates to the | |
2306 | * end of the file. | |
2307 | * @even_cows: 1 when truncating a file, unmap even private COWed pages; | |
2308 | * but 0 when invalidating pagecache, don't throw away private data. | |
2309 | */ | |
2310 | void unmap_mapping_range(struct address_space *mapping, | |
2311 | loff_t const holebegin, loff_t const holelen, int even_cows) | |
2312 | { | |
2313 | struct zap_details details; | |
2314 | pgoff_t hba = holebegin >> PAGE_SHIFT; | |
2315 | pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2316 | ||
2317 | /* Check for overflow. */ | |
2318 | if (sizeof(holelen) > sizeof(hlen)) { | |
2319 | long long holeend = | |
2320 | (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; | |
2321 | if (holeend & ~(long long)ULONG_MAX) | |
2322 | hlen = ULONG_MAX - hba + 1; | |
2323 | } | |
2324 | ||
2325 | details.check_mapping = even_cows? NULL: mapping; | |
2326 | details.first_index = hba; | |
2327 | details.last_index = hba + hlen - 1; | |
2328 | if (details.last_index < details.first_index) | |
2329 | details.last_index = ULONG_MAX; | |
2330 | ||
2331 | ||
2332 | i_mmap_lock_write(mapping); | |
2333 | if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap))) | |
2334 | unmap_mapping_range_tree(&mapping->i_mmap, &details); | |
2335 | i_mmap_unlock_write(mapping); | |
2336 | } | |
2337 | EXPORT_SYMBOL(unmap_mapping_range); | |
2338 | ||
2339 | /* | |
2340 | * We enter with non-exclusive mmap_sem (to exclude vma changes, | |
2341 | * but allow concurrent faults), and pte mapped but not yet locked. | |
2342 | * We return with pte unmapped and unlocked. | |
2343 | * | |
2344 | * We return with the mmap_sem locked or unlocked in the same cases | |
2345 | * as does filemap_fault(). | |
2346 | */ | |
2347 | static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
2348 | unsigned long address, pte_t *page_table, pmd_t *pmd, | |
2349 | unsigned int flags, pte_t orig_pte) | |
2350 | { | |
2351 | spinlock_t *ptl; | |
2352 | struct page *page, *swapcache; | |
2353 | struct mem_cgroup *memcg; | |
2354 | swp_entry_t entry; | |
2355 | pte_t pte; | |
2356 | int locked; | |
2357 | int exclusive = 0; | |
2358 | int ret = 0; | |
2359 | ||
2360 | if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) | |
2361 | goto out; | |
2362 | ||
2363 | entry = pte_to_swp_entry(orig_pte); | |
2364 | if (unlikely(non_swap_entry(entry))) { | |
2365 | if (is_migration_entry(entry)) { | |
2366 | migration_entry_wait(mm, pmd, address); | |
2367 | } else if (is_hwpoison_entry(entry)) { | |
2368 | ret = VM_FAULT_HWPOISON; | |
2369 | } else { | |
2370 | print_bad_pte(vma, address, orig_pte, NULL); | |
2371 | ret = VM_FAULT_SIGBUS; | |
2372 | } | |
2373 | goto out; | |
2374 | } | |
2375 | delayacct_set_flag(DELAYACCT_PF_SWAPIN); | |
2376 | page = lookup_swap_cache(entry); | |
2377 | if (!page) { | |
2378 | page = swapin_readahead(entry, | |
2379 | GFP_HIGHUSER_MOVABLE, vma, address); | |
2380 | if (!page) { | |
2381 | /* | |
2382 | * Back out if somebody else faulted in this pte | |
2383 | * while we released the pte lock. | |
2384 | */ | |
2385 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2386 | if (likely(pte_same(*page_table, orig_pte))) | |
2387 | ret = VM_FAULT_OOM; | |
2388 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); | |
2389 | goto unlock; | |
2390 | } | |
2391 | ||
2392 | /* Had to read the page from swap area: Major fault */ | |
2393 | ret = VM_FAULT_MAJOR; | |
2394 | count_vm_event(PGMAJFAULT); | |
2395 | mem_cgroup_count_vm_event(mm, PGMAJFAULT); | |
2396 | } else if (PageHWPoison(page)) { | |
2397 | /* | |
2398 | * hwpoisoned dirty swapcache pages are kept for killing | |
2399 | * owner processes (which may be unknown at hwpoison time) | |
2400 | */ | |
2401 | ret = VM_FAULT_HWPOISON; | |
2402 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); | |
2403 | swapcache = page; | |
2404 | goto out_release; | |
2405 | } | |
2406 | ||
2407 | swapcache = page; | |
2408 | locked = lock_page_or_retry(page, mm, flags); | |
2409 | ||
2410 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); | |
2411 | if (!locked) { | |
2412 | ret |= VM_FAULT_RETRY; | |
2413 | goto out_release; | |
2414 | } | |
2415 | ||
2416 | /* | |
2417 | * Make sure try_to_free_swap or reuse_swap_page or swapoff did not | |
2418 | * release the swapcache from under us. The page pin, and pte_same | |
2419 | * test below, are not enough to exclude that. Even if it is still | |
2420 | * swapcache, we need to check that the page's swap has not changed. | |
2421 | */ | |
2422 | if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) | |
2423 | goto out_page; | |
2424 | ||
2425 | page = ksm_might_need_to_copy(page, vma, address); | |
2426 | if (unlikely(!page)) { | |
2427 | ret = VM_FAULT_OOM; | |
2428 | page = swapcache; | |
2429 | goto out_page; | |
2430 | } | |
2431 | ||
2432 | if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) { | |
2433 | ret = VM_FAULT_OOM; | |
2434 | goto out_page; | |
2435 | } | |
2436 | ||
2437 | /* | |
2438 | * Back out if somebody else already faulted in this pte. | |
2439 | */ | |
2440 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2441 | if (unlikely(!pte_same(*page_table, orig_pte))) | |
2442 | goto out_nomap; | |
2443 | ||
2444 | if (unlikely(!PageUptodate(page))) { | |
2445 | ret = VM_FAULT_SIGBUS; | |
2446 | goto out_nomap; | |
2447 | } | |
2448 | ||
2449 | /* | |
2450 | * The page isn't present yet, go ahead with the fault. | |
2451 | * | |
2452 | * Be careful about the sequence of operations here. | |
2453 | * To get its accounting right, reuse_swap_page() must be called | |
2454 | * while the page is counted on swap but not yet in mapcount i.e. | |
2455 | * before page_add_anon_rmap() and swap_free(); try_to_free_swap() | |
2456 | * must be called after the swap_free(), or it will never succeed. | |
2457 | */ | |
2458 | ||
2459 | inc_mm_counter_fast(mm, MM_ANONPAGES); | |
2460 | dec_mm_counter_fast(mm, MM_SWAPENTS); | |
2461 | pte = mk_pte(page, vma->vm_page_prot); | |
2462 | if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { | |
2463 | pte = maybe_mkwrite(pte_mkdirty(pte), vma); | |
2464 | flags &= ~FAULT_FLAG_WRITE; | |
2465 | ret |= VM_FAULT_WRITE; | |
2466 | exclusive = 1; | |
2467 | } | |
2468 | flush_icache_page(vma, page); | |
2469 | if (pte_swp_soft_dirty(orig_pte)) | |
2470 | pte = pte_mksoft_dirty(pte); | |
2471 | set_pte_at(mm, address, page_table, pte); | |
2472 | if (page == swapcache) { | |
2473 | do_page_add_anon_rmap(page, vma, address, exclusive); | |
2474 | mem_cgroup_commit_charge(page, memcg, true); | |
2475 | } else { /* ksm created a completely new copy */ | |
2476 | page_add_new_anon_rmap(page, vma, address); | |
2477 | mem_cgroup_commit_charge(page, memcg, false); | |
2478 | lru_cache_add_active_or_unevictable(page, vma); | |
2479 | } | |
2480 | ||
2481 | swap_free(entry); | |
2482 | if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) | |
2483 | try_to_free_swap(page); | |
2484 | unlock_page(page); | |
2485 | if (page != swapcache) { | |
2486 | /* | |
2487 | * Hold the lock to avoid the swap entry to be reused | |
2488 | * until we take the PT lock for the pte_same() check | |
2489 | * (to avoid false positives from pte_same). For | |
2490 | * further safety release the lock after the swap_free | |
2491 | * so that the swap count won't change under a | |
2492 | * parallel locked swapcache. | |
2493 | */ | |
2494 | unlock_page(swapcache); | |
2495 | page_cache_release(swapcache); | |
2496 | } | |
2497 | ||
2498 | if (flags & FAULT_FLAG_WRITE) { | |
2499 | ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); | |
2500 | if (ret & VM_FAULT_ERROR) | |
2501 | ret &= VM_FAULT_ERROR; | |
2502 | goto out; | |
2503 | } | |
2504 | ||
2505 | /* No need to invalidate - it was non-present before */ | |
2506 | update_mmu_cache(vma, address, page_table); | |
2507 | unlock: | |
2508 | pte_unmap_unlock(page_table, ptl); | |
2509 | out: | |
2510 | return ret; | |
2511 | out_nomap: | |
2512 | mem_cgroup_cancel_charge(page, memcg); | |
2513 | pte_unmap_unlock(page_table, ptl); | |
2514 | out_page: | |
2515 | unlock_page(page); | |
2516 | out_release: | |
2517 | page_cache_release(page); | |
2518 | if (page != swapcache) { | |
2519 | unlock_page(swapcache); | |
2520 | page_cache_release(swapcache); | |
2521 | } | |
2522 | return ret; | |
2523 | } | |
2524 | ||
2525 | /* | |
2526 | * This is like a special single-page "expand_{down|up}wards()", | |
2527 | * except we must first make sure that 'address{-|+}PAGE_SIZE' | |
2528 | * doesn't hit another vma. | |
2529 | */ | |
2530 | static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) | |
2531 | { | |
2532 | address &= PAGE_MASK; | |
2533 | if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { | |
2534 | struct vm_area_struct *prev = vma->vm_prev; | |
2535 | ||
2536 | /* | |
2537 | * Is there a mapping abutting this one below? | |
2538 | * | |
2539 | * That's only ok if it's the same stack mapping | |
2540 | * that has gotten split.. | |
2541 | */ | |
2542 | if (prev && prev->vm_end == address) | |
2543 | return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; | |
2544 | ||
2545 | return expand_downwards(vma, address - PAGE_SIZE); | |
2546 | } | |
2547 | if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { | |
2548 | struct vm_area_struct *next = vma->vm_next; | |
2549 | ||
2550 | /* As VM_GROWSDOWN but s/below/above/ */ | |
2551 | if (next && next->vm_start == address + PAGE_SIZE) | |
2552 | return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; | |
2553 | ||
2554 | return expand_upwards(vma, address + PAGE_SIZE); | |
2555 | } | |
2556 | return 0; | |
2557 | } | |
2558 | ||
2559 | /* | |
2560 | * We enter with non-exclusive mmap_sem (to exclude vma changes, | |
2561 | * but allow concurrent faults), and pte mapped but not yet locked. | |
2562 | * We return with mmap_sem still held, but pte unmapped and unlocked. | |
2563 | */ | |
2564 | static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
2565 | unsigned long address, pte_t *page_table, pmd_t *pmd, | |
2566 | unsigned int flags) | |
2567 | { | |
2568 | struct mem_cgroup *memcg; | |
2569 | struct page *page; | |
2570 | spinlock_t *ptl; | |
2571 | pte_t entry; | |
2572 | ||
2573 | pte_unmap(page_table); | |
2574 | ||
2575 | /* Check if we need to add a guard page to the stack */ | |
2576 | if (check_stack_guard_page(vma, address) < 0) | |
2577 | return VM_FAULT_SIGSEGV; | |
2578 | ||
2579 | /* Use the zero-page for reads */ | |
2580 | if (!(flags & FAULT_FLAG_WRITE) && !mm_forbids_zeropage(mm)) { | |
2581 | entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), | |
2582 | vma->vm_page_prot)); | |
2583 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2584 | if (!pte_none(*page_table)) | |
2585 | goto unlock; | |
2586 | goto setpte; | |
2587 | } | |
2588 | ||
2589 | /* Allocate our own private page. */ | |
2590 | if (unlikely(anon_vma_prepare(vma))) | |
2591 | goto oom; | |
2592 | page = alloc_zeroed_user_highpage_movable(vma, address); | |
2593 | if (!page) | |
2594 | goto oom; | |
2595 | /* | |
2596 | * The memory barrier inside __SetPageUptodate makes sure that | |
2597 | * preceeding stores to the page contents become visible before | |
2598 | * the set_pte_at() write. | |
2599 | */ | |
2600 | __SetPageUptodate(page); | |
2601 | ||
2602 | if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) | |
2603 | goto oom_free_page; | |
2604 | ||
2605 | entry = mk_pte(page, vma->vm_page_prot); | |
2606 | if (vma->vm_flags & VM_WRITE) | |
2607 | entry = pte_mkwrite(pte_mkdirty(entry)); | |
2608 | ||
2609 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2610 | if (!pte_none(*page_table)) | |
2611 | goto release; | |
2612 | ||
2613 | inc_mm_counter_fast(mm, MM_ANONPAGES); | |
2614 | page_add_new_anon_rmap(page, vma, address); | |
2615 | mem_cgroup_commit_charge(page, memcg, false); | |
2616 | lru_cache_add_active_or_unevictable(page, vma); | |
2617 | setpte: | |
2618 | set_pte_at(mm, address, page_table, entry); | |
2619 | ||
2620 | /* No need to invalidate - it was non-present before */ | |
2621 | update_mmu_cache(vma, address, page_table); | |
2622 | unlock: | |
2623 | pte_unmap_unlock(page_table, ptl); | |
2624 | return 0; | |
2625 | release: | |
2626 | mem_cgroup_cancel_charge(page, memcg); | |
2627 | page_cache_release(page); | |
2628 | goto unlock; | |
2629 | oom_free_page: | |
2630 | page_cache_release(page); | |
2631 | oom: | |
2632 | return VM_FAULT_OOM; | |
2633 | } | |
2634 | ||
2635 | /* | |
2636 | * The mmap_sem must have been held on entry, and may have been | |
2637 | * released depending on flags and vma->vm_ops->fault() return value. | |
2638 | * See filemap_fault() and __lock_page_retry(). | |
2639 | */ | |
2640 | static int __do_fault(struct vm_area_struct *vma, unsigned long address, | |
2641 | pgoff_t pgoff, unsigned int flags, struct page **page) | |
2642 | { | |
2643 | struct vm_fault vmf; | |
2644 | int ret; | |
2645 | ||
2646 | vmf.virtual_address = (void __user *)(address & PAGE_MASK); | |
2647 | vmf.pgoff = pgoff; | |
2648 | vmf.flags = flags; | |
2649 | vmf.page = NULL; | |
2650 | ||
2651 | ret = vma->vm_ops->fault(vma, &vmf); | |
2652 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) | |
2653 | return ret; | |
2654 | ||
2655 | if (unlikely(PageHWPoison(vmf.page))) { | |
2656 | if (ret & VM_FAULT_LOCKED) | |
2657 | unlock_page(vmf.page); | |
2658 | page_cache_release(vmf.page); | |
2659 | return VM_FAULT_HWPOISON; | |
2660 | } | |
2661 | ||
2662 | if (unlikely(!(ret & VM_FAULT_LOCKED))) | |
2663 | lock_page(vmf.page); | |
2664 | else | |
2665 | VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page); | |
2666 | ||
2667 | *page = vmf.page; | |
2668 | return ret; | |
2669 | } | |
2670 | ||
2671 | /** | |
2672 | * do_set_pte - setup new PTE entry for given page and add reverse page mapping. | |
2673 | * | |
2674 | * @vma: virtual memory area | |
2675 | * @address: user virtual address | |
2676 | * @page: page to map | |
2677 | * @pte: pointer to target page table entry | |
2678 | * @write: true, if new entry is writable | |
2679 | * @anon: true, if it's anonymous page | |
2680 | * | |
2681 | * Caller must hold page table lock relevant for @pte. | |
2682 | * | |
2683 | * Target users are page handler itself and implementations of | |
2684 | * vm_ops->map_pages. | |
2685 | */ | |
2686 | void do_set_pte(struct vm_area_struct *vma, unsigned long address, | |
2687 | struct page *page, pte_t *pte, bool write, bool anon) | |
2688 | { | |
2689 | pte_t entry; | |
2690 | ||
2691 | flush_icache_page(vma, page); | |
2692 | entry = mk_pte(page, vma->vm_page_prot); | |
2693 | if (write) | |
2694 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); | |
2695 | if (anon) { | |
2696 | inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); | |
2697 | page_add_new_anon_rmap(page, vma, address); | |
2698 | } else { | |
2699 | inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES); | |
2700 | page_add_file_rmap(page); | |
2701 | } | |
2702 | set_pte_at(vma->vm_mm, address, pte, entry); | |
2703 | ||
2704 | /* no need to invalidate: a not-present page won't be cached */ | |
2705 | update_mmu_cache(vma, address, pte); | |
2706 | } | |
2707 | ||
2708 | static unsigned long fault_around_bytes __read_mostly = | |
2709 | rounddown_pow_of_two(65536); | |
2710 | ||
2711 | #ifdef CONFIG_DEBUG_FS | |
2712 | static int fault_around_bytes_get(void *data, u64 *val) | |
2713 | { | |
2714 | *val = fault_around_bytes; | |
2715 | return 0; | |
2716 | } | |
2717 | ||
2718 | /* | |
2719 | * fault_around_pages() and fault_around_mask() expects fault_around_bytes | |
2720 | * rounded down to nearest page order. It's what do_fault_around() expects to | |
2721 | * see. | |
2722 | */ | |
2723 | static int fault_around_bytes_set(void *data, u64 val) | |
2724 | { | |
2725 | if (val / PAGE_SIZE > PTRS_PER_PTE) | |
2726 | return -EINVAL; | |
2727 | if (val > PAGE_SIZE) | |
2728 | fault_around_bytes = rounddown_pow_of_two(val); | |
2729 | else | |
2730 | fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ | |
2731 | return 0; | |
2732 | } | |
2733 | DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops, | |
2734 | fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); | |
2735 | ||
2736 | static int __init fault_around_debugfs(void) | |
2737 | { | |
2738 | void *ret; | |
2739 | ||
2740 | ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL, | |
2741 | &fault_around_bytes_fops); | |
2742 | if (!ret) | |
2743 | pr_warn("Failed to create fault_around_bytes in debugfs"); | |
2744 | return 0; | |
2745 | } | |
2746 | late_initcall(fault_around_debugfs); | |
2747 | #endif | |
2748 | ||
2749 | /* | |
2750 | * do_fault_around() tries to map few pages around the fault address. The hope | |
2751 | * is that the pages will be needed soon and this will lower the number of | |
2752 | * faults to handle. | |
2753 | * | |
2754 | * It uses vm_ops->map_pages() to map the pages, which skips the page if it's | |
2755 | * not ready to be mapped: not up-to-date, locked, etc. | |
2756 | * | |
2757 | * This function is called with the page table lock taken. In the split ptlock | |
2758 | * case the page table lock only protects only those entries which belong to | |
2759 | * the page table corresponding to the fault address. | |
2760 | * | |
2761 | * This function doesn't cross the VMA boundaries, in order to call map_pages() | |
2762 | * only once. | |
2763 | * | |
2764 | * fault_around_pages() defines how many pages we'll try to map. | |
2765 | * do_fault_around() expects it to return a power of two less than or equal to | |
2766 | * PTRS_PER_PTE. | |
2767 | * | |
2768 | * The virtual address of the area that we map is naturally aligned to the | |
2769 | * fault_around_pages() value (and therefore to page order). This way it's | |
2770 | * easier to guarantee that we don't cross page table boundaries. | |
2771 | */ | |
2772 | static void do_fault_around(struct vm_area_struct *vma, unsigned long address, | |
2773 | pte_t *pte, pgoff_t pgoff, unsigned int flags) | |
2774 | { | |
2775 | unsigned long start_addr, nr_pages, mask; | |
2776 | pgoff_t max_pgoff; | |
2777 | struct vm_fault vmf; | |
2778 | int off; | |
2779 | ||
2780 | nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT; | |
2781 | mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; | |
2782 | ||
2783 | start_addr = max(address & mask, vma->vm_start); | |
2784 | off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); | |
2785 | pte -= off; | |
2786 | pgoff -= off; | |
2787 | ||
2788 | /* | |
2789 | * max_pgoff is either end of page table or end of vma | |
2790 | * or fault_around_pages() from pgoff, depending what is nearest. | |
2791 | */ | |
2792 | max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + | |
2793 | PTRS_PER_PTE - 1; | |
2794 | max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1, | |
2795 | pgoff + nr_pages - 1); | |
2796 | ||
2797 | /* Check if it makes any sense to call ->map_pages */ | |
2798 | while (!pte_none(*pte)) { | |
2799 | if (++pgoff > max_pgoff) | |
2800 | return; | |
2801 | start_addr += PAGE_SIZE; | |
2802 | if (start_addr >= vma->vm_end) | |
2803 | return; | |
2804 | pte++; | |
2805 | } | |
2806 | ||
2807 | vmf.virtual_address = (void __user *) start_addr; | |
2808 | vmf.pte = pte; | |
2809 | vmf.pgoff = pgoff; | |
2810 | vmf.max_pgoff = max_pgoff; | |
2811 | vmf.flags = flags; | |
2812 | vma->vm_ops->map_pages(vma, &vmf); | |
2813 | } | |
2814 | ||
2815 | static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
2816 | unsigned long address, pmd_t *pmd, | |
2817 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) | |
2818 | { | |
2819 | struct page *fault_page; | |
2820 | spinlock_t *ptl; | |
2821 | pte_t *pte; | |
2822 | int ret = 0; | |
2823 | ||
2824 | /* | |
2825 | * Let's call ->map_pages() first and use ->fault() as fallback | |
2826 | * if page by the offset is not ready to be mapped (cold cache or | |
2827 | * something). | |
2828 | */ | |
2829 | if (vma->vm_ops->map_pages && fault_around_bytes >> PAGE_SHIFT > 1) { | |
2830 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2831 | do_fault_around(vma, address, pte, pgoff, flags); | |
2832 | if (!pte_same(*pte, orig_pte)) | |
2833 | goto unlock_out; | |
2834 | pte_unmap_unlock(pte, ptl); | |
2835 | } | |
2836 | ||
2837 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); | |
2838 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) | |
2839 | return ret; | |
2840 | ||
2841 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2842 | if (unlikely(!pte_same(*pte, orig_pte))) { | |
2843 | pte_unmap_unlock(pte, ptl); | |
2844 | unlock_page(fault_page); | |
2845 | page_cache_release(fault_page); | |
2846 | return ret; | |
2847 | } | |
2848 | do_set_pte(vma, address, fault_page, pte, false, false); | |
2849 | unlock_page(fault_page); | |
2850 | unlock_out: | |
2851 | pte_unmap_unlock(pte, ptl); | |
2852 | return ret; | |
2853 | } | |
2854 | ||
2855 | static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
2856 | unsigned long address, pmd_t *pmd, | |
2857 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) | |
2858 | { | |
2859 | struct page *fault_page, *new_page; | |
2860 | struct mem_cgroup *memcg; | |
2861 | spinlock_t *ptl; | |
2862 | pte_t *pte; | |
2863 | int ret; | |
2864 | ||
2865 | if (unlikely(anon_vma_prepare(vma))) | |
2866 | return VM_FAULT_OOM; | |
2867 | ||
2868 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); | |
2869 | if (!new_page) | |
2870 | return VM_FAULT_OOM; | |
2871 | ||
2872 | if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) { | |
2873 | page_cache_release(new_page); | |
2874 | return VM_FAULT_OOM; | |
2875 | } | |
2876 | ||
2877 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); | |
2878 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) | |
2879 | goto uncharge_out; | |
2880 | ||
2881 | copy_user_highpage(new_page, fault_page, address, vma); | |
2882 | __SetPageUptodate(new_page); | |
2883 | ||
2884 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2885 | if (unlikely(!pte_same(*pte, orig_pte))) { | |
2886 | pte_unmap_unlock(pte, ptl); | |
2887 | unlock_page(fault_page); | |
2888 | page_cache_release(fault_page); | |
2889 | goto uncharge_out; | |
2890 | } | |
2891 | do_set_pte(vma, address, new_page, pte, true, true); | |
2892 | mem_cgroup_commit_charge(new_page, memcg, false); | |
2893 | lru_cache_add_active_or_unevictable(new_page, vma); | |
2894 | pte_unmap_unlock(pte, ptl); | |
2895 | unlock_page(fault_page); | |
2896 | page_cache_release(fault_page); | |
2897 | return ret; | |
2898 | uncharge_out: | |
2899 | mem_cgroup_cancel_charge(new_page, memcg); | |
2900 | page_cache_release(new_page); | |
2901 | return ret; | |
2902 | } | |
2903 | ||
2904 | static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
2905 | unsigned long address, pmd_t *pmd, | |
2906 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) | |
2907 | { | |
2908 | struct page *fault_page; | |
2909 | struct address_space *mapping; | |
2910 | spinlock_t *ptl; | |
2911 | pte_t *pte; | |
2912 | int dirtied = 0; | |
2913 | int ret, tmp; | |
2914 | ||
2915 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); | |
2916 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) | |
2917 | return ret; | |
2918 | ||
2919 | /* | |
2920 | * Check if the backing address space wants to know that the page is | |
2921 | * about to become writable | |
2922 | */ | |
2923 | if (vma->vm_ops->page_mkwrite) { | |
2924 | unlock_page(fault_page); | |
2925 | tmp = do_page_mkwrite(vma, fault_page, address); | |
2926 | if (unlikely(!tmp || | |
2927 | (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { | |
2928 | page_cache_release(fault_page); | |
2929 | return tmp; | |
2930 | } | |
2931 | } | |
2932 | ||
2933 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); | |
2934 | if (unlikely(!pte_same(*pte, orig_pte))) { | |
2935 | pte_unmap_unlock(pte, ptl); | |
2936 | unlock_page(fault_page); | |
2937 | page_cache_release(fault_page); | |
2938 | return ret; | |
2939 | } | |
2940 | do_set_pte(vma, address, fault_page, pte, true, false); | |
2941 | pte_unmap_unlock(pte, ptl); | |
2942 | ||
2943 | if (set_page_dirty(fault_page)) | |
2944 | dirtied = 1; | |
2945 | /* | |
2946 | * Take a local copy of the address_space - page.mapping may be zeroed | |
2947 | * by truncate after unlock_page(). The address_space itself remains | |
2948 | * pinned by vma->vm_file's reference. We rely on unlock_page()'s | |
2949 | * release semantics to prevent the compiler from undoing this copying. | |
2950 | */ | |
2951 | mapping = fault_page->mapping; | |
2952 | unlock_page(fault_page); | |
2953 | if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) { | |
2954 | /* | |
2955 | * Some device drivers do not set page.mapping but still | |
2956 | * dirty their pages | |
2957 | */ | |
2958 | balance_dirty_pages_ratelimited(mapping); | |
2959 | } | |
2960 | ||
2961 | if (!vma->vm_ops->page_mkwrite) | |
2962 | file_update_time(vma->vm_file); | |
2963 | ||
2964 | return ret; | |
2965 | } | |
2966 | ||
2967 | /* | |
2968 | * We enter with non-exclusive mmap_sem (to exclude vma changes, | |
2969 | * but allow concurrent faults). | |
2970 | * The mmap_sem may have been released depending on flags and our | |
2971 | * return value. See filemap_fault() and __lock_page_or_retry(). | |
2972 | */ | |
2973 | static int do_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
2974 | unsigned long address, pte_t *page_table, pmd_t *pmd, | |
2975 | unsigned int flags, pte_t orig_pte) | |
2976 | { | |
2977 | pgoff_t pgoff = (((address & PAGE_MASK) | |
2978 | - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; | |
2979 | ||
2980 | pte_unmap(page_table); | |
2981 | if (!(flags & FAULT_FLAG_WRITE)) | |
2982 | return do_read_fault(mm, vma, address, pmd, pgoff, flags, | |
2983 | orig_pte); | |
2984 | if (!(vma->vm_flags & VM_SHARED)) | |
2985 | return do_cow_fault(mm, vma, address, pmd, pgoff, flags, | |
2986 | orig_pte); | |
2987 | return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); | |
2988 | } | |
2989 | ||
2990 | static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, | |
2991 | unsigned long addr, int page_nid, | |
2992 | int *flags) | |
2993 | { | |
2994 | get_page(page); | |
2995 | ||
2996 | count_vm_numa_event(NUMA_HINT_FAULTS); | |
2997 | if (page_nid == numa_node_id()) { | |
2998 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); | |
2999 | *flags |= TNF_FAULT_LOCAL; | |
3000 | } | |
3001 | ||
3002 | return mpol_misplaced(page, vma, addr); | |
3003 | } | |
3004 | ||
3005 | static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, | |
3006 | unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd) | |
3007 | { | |
3008 | struct page *page = NULL; | |
3009 | spinlock_t *ptl; | |
3010 | int page_nid = -1; | |
3011 | int last_cpupid; | |
3012 | int target_nid; | |
3013 | bool migrated = false; | |
3014 | int flags = 0; | |
3015 | ||
3016 | /* | |
3017 | * The "pte" at this point cannot be used safely without | |
3018 | * validation through pte_unmap_same(). It's of NUMA type but | |
3019 | * the pfn may be screwed if the read is non atomic. | |
3020 | * | |
3021 | * We can safely just do a "set_pte_at()", because the old | |
3022 | * page table entry is not accessible, so there would be no | |
3023 | * concurrent hardware modifications to the PTE. | |
3024 | */ | |
3025 | ptl = pte_lockptr(mm, pmd); | |
3026 | spin_lock(ptl); | |
3027 | if (unlikely(!pte_same(*ptep, pte))) { | |
3028 | pte_unmap_unlock(ptep, ptl); | |
3029 | goto out; | |
3030 | } | |
3031 | ||
3032 | /* Make it present again */ | |
3033 | pte = pte_modify(pte, vma->vm_page_prot); | |
3034 | pte = pte_mkyoung(pte); | |
3035 | set_pte_at(mm, addr, ptep, pte); | |
3036 | update_mmu_cache(vma, addr, ptep); | |
3037 | ||
3038 | page = vm_normal_page(vma, addr, pte); | |
3039 | if (!page) { | |
3040 | pte_unmap_unlock(ptep, ptl); | |
3041 | return 0; | |
3042 | } | |
3043 | ||
3044 | /* | |
3045 | * Avoid grouping on DSO/COW pages in specific and RO pages | |
3046 | * in general, RO pages shouldn't hurt as much anyway since | |
3047 | * they can be in shared cache state. | |
3048 | */ | |
3049 | if (!pte_write(pte)) | |
3050 | flags |= TNF_NO_GROUP; | |
3051 | ||
3052 | /* | |
3053 | * Flag if the page is shared between multiple address spaces. This | |
3054 | * is later used when determining whether to group tasks together | |
3055 | */ | |
3056 | if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) | |
3057 | flags |= TNF_SHARED; | |
3058 | ||
3059 | last_cpupid = page_cpupid_last(page); | |
3060 | page_nid = page_to_nid(page); | |
3061 | target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags); | |
3062 | pte_unmap_unlock(ptep, ptl); | |
3063 | if (target_nid == -1) { | |
3064 | put_page(page); | |
3065 | goto out; | |
3066 | } | |
3067 | ||
3068 | /* Migrate to the requested node */ | |
3069 | migrated = migrate_misplaced_page(page, vma, target_nid); | |
3070 | if (migrated) { | |
3071 | page_nid = target_nid; | |
3072 | flags |= TNF_MIGRATED; | |
3073 | } | |
3074 | ||
3075 | out: | |
3076 | if (page_nid != -1) | |
3077 | task_numa_fault(last_cpupid, page_nid, 1, flags); | |
3078 | return 0; | |
3079 | } | |
3080 | ||
3081 | /* | |
3082 | * These routines also need to handle stuff like marking pages dirty | |
3083 | * and/or accessed for architectures that don't do it in hardware (most | |
3084 | * RISC architectures). The early dirtying is also good on the i386. | |
3085 | * | |
3086 | * There is also a hook called "update_mmu_cache()" that architectures | |
3087 | * with external mmu caches can use to update those (ie the Sparc or | |
3088 | * PowerPC hashed page tables that act as extended TLBs). | |
3089 | * | |
3090 | * We enter with non-exclusive mmap_sem (to exclude vma changes, | |
3091 | * but allow concurrent faults), and pte mapped but not yet locked. | |
3092 | * We return with pte unmapped and unlocked. | |
3093 | * | |
3094 | * The mmap_sem may have been released depending on flags and our | |
3095 | * return value. See filemap_fault() and __lock_page_or_retry(). | |
3096 | */ | |
3097 | static int handle_pte_fault(struct mm_struct *mm, | |
3098 | struct vm_area_struct *vma, unsigned long address, | |
3099 | pte_t *pte, pmd_t *pmd, unsigned int flags) | |
3100 | { | |
3101 | pte_t entry; | |
3102 | spinlock_t *ptl; | |
3103 | ||
3104 | /* | |
3105 | * some architectures can have larger ptes than wordsize, | |
3106 | * e.g.ppc44x-defconfig has CONFIG_PTE_64BIT=y and CONFIG_32BIT=y, | |
3107 | * so READ_ONCE or ACCESS_ONCE cannot guarantee atomic accesses. | |
3108 | * The code below just needs a consistent view for the ifs and | |
3109 | * we later double check anyway with the ptl lock held. So here | |
3110 | * a barrier will do. | |
3111 | */ | |
3112 | entry = *pte; | |
3113 | barrier(); | |
3114 | if (!pte_present(entry)) { | |
3115 | if (pte_none(entry)) { | |
3116 | if (vma->vm_ops) { | |
3117 | if (likely(vma->vm_ops->fault)) | |
3118 | return do_fault(mm, vma, address, pte, | |
3119 | pmd, flags, entry); | |
3120 | } | |
3121 | return do_anonymous_page(mm, vma, address, | |
3122 | pte, pmd, flags); | |
3123 | } | |
3124 | return do_swap_page(mm, vma, address, | |
3125 | pte, pmd, flags, entry); | |
3126 | } | |
3127 | ||
3128 | if (pte_protnone(entry)) | |
3129 | return do_numa_page(mm, vma, address, entry, pte, pmd); | |
3130 | ||
3131 | ptl = pte_lockptr(mm, pmd); | |
3132 | spin_lock(ptl); | |
3133 | if (unlikely(!pte_same(*pte, entry))) | |
3134 | goto unlock; | |
3135 | if (flags & FAULT_FLAG_WRITE) { | |
3136 | if (!pte_write(entry)) | |
3137 | return do_wp_page(mm, vma, address, | |
3138 | pte, pmd, ptl, entry); | |
3139 | entry = pte_mkdirty(entry); | |
3140 | } | |
3141 | entry = pte_mkyoung(entry); | |
3142 | if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) { | |
3143 | update_mmu_cache(vma, address, pte); | |
3144 | } else { | |
3145 | /* | |
3146 | * This is needed only for protection faults but the arch code | |
3147 | * is not yet telling us if this is a protection fault or not. | |
3148 | * This still avoids useless tlb flushes for .text page faults | |
3149 | * with threads. | |
3150 | */ | |
3151 | if (flags & FAULT_FLAG_WRITE) | |
3152 | flush_tlb_fix_spurious_fault(vma, address); | |
3153 | } | |
3154 | unlock: | |
3155 | pte_unmap_unlock(pte, ptl); | |
3156 | return 0; | |
3157 | } | |
3158 | ||
3159 | /* | |
3160 | * By the time we get here, we already hold the mm semaphore | |
3161 | * | |
3162 | * The mmap_sem may have been released depending on flags and our | |
3163 | * return value. See filemap_fault() and __lock_page_or_retry(). | |
3164 | */ | |
3165 | static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
3166 | unsigned long address, unsigned int flags) | |
3167 | { | |
3168 | pgd_t *pgd; | |
3169 | pud_t *pud; | |
3170 | pmd_t *pmd; | |
3171 | pte_t *pte; | |
3172 | ||
3173 | if (unlikely(is_vm_hugetlb_page(vma))) | |
3174 | return hugetlb_fault(mm, vma, address, flags); | |
3175 | ||
3176 | pgd = pgd_offset(mm, address); | |
3177 | pud = pud_alloc(mm, pgd, address); | |
3178 | if (!pud) | |
3179 | return VM_FAULT_OOM; | |
3180 | pmd = pmd_alloc(mm, pud, address); | |
3181 | if (!pmd) | |
3182 | return VM_FAULT_OOM; | |
3183 | if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { | |
3184 | int ret = VM_FAULT_FALLBACK; | |
3185 | if (!vma->vm_ops) | |
3186 | ret = do_huge_pmd_anonymous_page(mm, vma, address, | |
3187 | pmd, flags); | |
3188 | if (!(ret & VM_FAULT_FALLBACK)) | |
3189 | return ret; | |
3190 | } else { | |
3191 | pmd_t orig_pmd = *pmd; | |
3192 | int ret; | |
3193 | ||
3194 | barrier(); | |
3195 | if (pmd_trans_huge(orig_pmd)) { | |
3196 | unsigned int dirty = flags & FAULT_FLAG_WRITE; | |
3197 | ||
3198 | /* | |
3199 | * If the pmd is splitting, return and retry the | |
3200 | * the fault. Alternative: wait until the split | |
3201 | * is done, and goto retry. | |
3202 | */ | |
3203 | if (pmd_trans_splitting(orig_pmd)) | |
3204 | return 0; | |
3205 | ||
3206 | if (pmd_protnone(orig_pmd)) | |
3207 | return do_huge_pmd_numa_page(mm, vma, address, | |
3208 | orig_pmd, pmd); | |
3209 | ||
3210 | if (dirty && !pmd_write(orig_pmd)) { | |
3211 | ret = do_huge_pmd_wp_page(mm, vma, address, pmd, | |
3212 | orig_pmd); | |
3213 | if (!(ret & VM_FAULT_FALLBACK)) | |
3214 | return ret; | |
3215 | } else { | |
3216 | huge_pmd_set_accessed(mm, vma, address, pmd, | |
3217 | orig_pmd, dirty); | |
3218 | return 0; | |
3219 | } | |
3220 | } | |
3221 | } | |
3222 | ||
3223 | /* | |
3224 | * Use __pte_alloc instead of pte_alloc_map, because we can't | |
3225 | * run pte_offset_map on the pmd, if an huge pmd could | |
3226 | * materialize from under us from a different thread. | |
3227 | */ | |
3228 | if (unlikely(pmd_none(*pmd)) && | |
3229 | unlikely(__pte_alloc(mm, vma, pmd, address))) | |
3230 | return VM_FAULT_OOM; | |
3231 | /* if an huge pmd materialized from under us just retry later */ | |
3232 | if (unlikely(pmd_trans_huge(*pmd))) | |
3233 | return 0; | |
3234 | /* | |
3235 | * A regular pmd is established and it can't morph into a huge pmd | |
3236 | * from under us anymore at this point because we hold the mmap_sem | |
3237 | * read mode and khugepaged takes it in write mode. So now it's | |
3238 | * safe to run pte_offset_map(). | |
3239 | */ | |
3240 | pte = pte_offset_map(pmd, address); | |
3241 | ||
3242 | return handle_pte_fault(mm, vma, address, pte, pmd, flags); | |
3243 | } | |
3244 | ||
3245 | /* | |
3246 | * By the time we get here, we already hold the mm semaphore | |
3247 | * | |
3248 | * The mmap_sem may have been released depending on flags and our | |
3249 | * return value. See filemap_fault() and __lock_page_or_retry(). | |
3250 | */ | |
3251 | int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, | |
3252 | unsigned long address, unsigned int flags) | |
3253 | { | |
3254 | int ret; | |
3255 | ||
3256 | __set_current_state(TASK_RUNNING); | |
3257 | ||
3258 | count_vm_event(PGFAULT); | |
3259 | mem_cgroup_count_vm_event(mm, PGFAULT); | |
3260 | ||
3261 | /* do counter updates before entering really critical section. */ | |
3262 | check_sync_rss_stat(current); | |
3263 | ||
3264 | /* | |
3265 | * Enable the memcg OOM handling for faults triggered in user | |
3266 | * space. Kernel faults are handled more gracefully. | |
3267 | */ | |
3268 | if (flags & FAULT_FLAG_USER) | |
3269 | mem_cgroup_oom_enable(); | |
3270 | ||
3271 | ret = __handle_mm_fault(mm, vma, address, flags); | |
3272 | ||
3273 | if (flags & FAULT_FLAG_USER) { | |
3274 | mem_cgroup_oom_disable(); | |
3275 | /* | |
3276 | * The task may have entered a memcg OOM situation but | |
3277 | * if the allocation error was handled gracefully (no | |
3278 | * VM_FAULT_OOM), there is no need to kill anything. | |
3279 | * Just clean up the OOM state peacefully. | |
3280 | */ | |
3281 | if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) | |
3282 | mem_cgroup_oom_synchronize(false); | |
3283 | } | |
3284 | ||
3285 | return ret; | |
3286 | } | |
3287 | EXPORT_SYMBOL_GPL(handle_mm_fault); | |
3288 | ||
3289 | #ifndef __PAGETABLE_PUD_FOLDED | |
3290 | /* | |
3291 | * Allocate page upper directory. | |
3292 | * We've already handled the fast-path in-line. | |
3293 | */ | |
3294 | int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) | |
3295 | { | |
3296 | pud_t *new = pud_alloc_one(mm, address); | |
3297 | if (!new) | |
3298 | return -ENOMEM; | |
3299 | ||
3300 | smp_wmb(); /* See comment in __pte_alloc */ | |
3301 | ||
3302 | spin_lock(&mm->page_table_lock); | |
3303 | if (pgd_present(*pgd)) /* Another has populated it */ | |
3304 | pud_free(mm, new); | |
3305 | else | |
3306 | pgd_populate(mm, pgd, new); | |
3307 | spin_unlock(&mm->page_table_lock); | |
3308 | return 0; | |
3309 | } | |
3310 | #endif /* __PAGETABLE_PUD_FOLDED */ | |
3311 | ||
3312 | #ifndef __PAGETABLE_PMD_FOLDED | |
3313 | /* | |
3314 | * Allocate page middle directory. | |
3315 | * We've already handled the fast-path in-line. | |
3316 | */ | |
3317 | int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) | |
3318 | { | |
3319 | pmd_t *new = pmd_alloc_one(mm, address); | |
3320 | if (!new) | |
3321 | return -ENOMEM; | |
3322 | ||
3323 | smp_wmb(); /* See comment in __pte_alloc */ | |
3324 | ||
3325 | spin_lock(&mm->page_table_lock); | |
3326 | #ifndef __ARCH_HAS_4LEVEL_HACK | |
3327 | if (!pud_present(*pud)) { | |
3328 | mm_inc_nr_pmds(mm); | |
3329 | pud_populate(mm, pud, new); | |
3330 | } else /* Another has populated it */ | |
3331 | pmd_free(mm, new); | |
3332 | #else | |
3333 | if (!pgd_present(*pud)) { | |
3334 | mm_inc_nr_pmds(mm); | |
3335 | pgd_populate(mm, pud, new); | |
3336 | } else /* Another has populated it */ | |
3337 | pmd_free(mm, new); | |
3338 | #endif /* __ARCH_HAS_4LEVEL_HACK */ | |
3339 | spin_unlock(&mm->page_table_lock); | |
3340 | return 0; | |
3341 | } | |
3342 | #endif /* __PAGETABLE_PMD_FOLDED */ | |
3343 | ||
3344 | static int __follow_pte(struct mm_struct *mm, unsigned long address, | |
3345 | pte_t **ptepp, spinlock_t **ptlp) | |
3346 | { | |
3347 | pgd_t *pgd; | |
3348 | pud_t *pud; | |
3349 | pmd_t *pmd; | |
3350 | pte_t *ptep; | |
3351 | ||
3352 | pgd = pgd_offset(mm, address); | |
3353 | if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) | |
3354 | goto out; | |
3355 | ||
3356 | pud = pud_offset(pgd, address); | |
3357 | if (pud_none(*pud) || unlikely(pud_bad(*pud))) | |
3358 | goto out; | |
3359 | ||
3360 | pmd = pmd_offset(pud, address); | |
3361 | VM_BUG_ON(pmd_trans_huge(*pmd)); | |
3362 | if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) | |
3363 | goto out; | |
3364 | ||
3365 | /* We cannot handle huge page PFN maps. Luckily they don't exist. */ | |
3366 | if (pmd_huge(*pmd)) | |
3367 | goto out; | |
3368 | ||
3369 | ptep = pte_offset_map_lock(mm, pmd, address, ptlp); | |
3370 | if (!ptep) | |
3371 | goto out; | |
3372 | if (!pte_present(*ptep)) | |
3373 | goto unlock; | |
3374 | *ptepp = ptep; | |
3375 | return 0; | |
3376 | unlock: | |
3377 | pte_unmap_unlock(ptep, *ptlp); | |
3378 | out: | |
3379 | return -EINVAL; | |
3380 | } | |
3381 | ||
3382 | static inline int follow_pte(struct mm_struct *mm, unsigned long address, | |
3383 | pte_t **ptepp, spinlock_t **ptlp) | |
3384 | { | |
3385 | int res; | |
3386 | ||
3387 | /* (void) is needed to make gcc happy */ | |
3388 | (void) __cond_lock(*ptlp, | |
3389 | !(res = __follow_pte(mm, address, ptepp, ptlp))); | |
3390 | return res; | |
3391 | } | |
3392 | ||
3393 | /** | |
3394 | * follow_pfn - look up PFN at a user virtual address | |
3395 | * @vma: memory mapping | |
3396 | * @address: user virtual address | |
3397 | * @pfn: location to store found PFN | |
3398 | * | |
3399 | * Only IO mappings and raw PFN mappings are allowed. | |
3400 | * | |
3401 | * Returns zero and the pfn at @pfn on success, -ve otherwise. | |
3402 | */ | |
3403 | int follow_pfn(struct vm_area_struct *vma, unsigned long address, | |
3404 | unsigned long *pfn) | |
3405 | { | |
3406 | int ret = -EINVAL; | |
3407 | spinlock_t *ptl; | |
3408 | pte_t *ptep; | |
3409 | ||
3410 | if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) | |
3411 | return ret; | |
3412 | ||
3413 | ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); | |
3414 | if (ret) | |
3415 | return ret; | |
3416 | *pfn = pte_pfn(*ptep); | |
3417 | pte_unmap_unlock(ptep, ptl); | |
3418 | return 0; | |
3419 | } | |
3420 | EXPORT_SYMBOL(follow_pfn); | |
3421 | ||
3422 | #ifdef CONFIG_HAVE_IOREMAP_PROT | |
3423 | int follow_phys(struct vm_area_struct *vma, | |
3424 | unsigned long address, unsigned int flags, | |
3425 | unsigned long *prot, resource_size_t *phys) | |
3426 | { | |
3427 | int ret = -EINVAL; | |
3428 | pte_t *ptep, pte; | |
3429 | spinlock_t *ptl; | |
3430 | ||
3431 | if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) | |
3432 | goto out; | |
3433 | ||
3434 | if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) | |
3435 | goto out; | |
3436 | pte = *ptep; | |
3437 | ||
3438 | if ((flags & FOLL_WRITE) && !pte_write(pte)) | |
3439 | goto unlock; | |
3440 | ||
3441 | *prot = pgprot_val(pte_pgprot(pte)); | |
3442 | *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; | |
3443 | ||
3444 | ret = 0; | |
3445 | unlock: | |
3446 | pte_unmap_unlock(ptep, ptl); | |
3447 | out: | |
3448 | return ret; | |
3449 | } | |
3450 | ||
3451 | int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, | |
3452 | void *buf, int len, int write) | |
3453 | { | |
3454 | resource_size_t phys_addr; | |
3455 | unsigned long prot = 0; | |
3456 | void __iomem *maddr; | |
3457 | int offset = addr & (PAGE_SIZE-1); | |
3458 | ||
3459 | if (follow_phys(vma, addr, write, &prot, &phys_addr)) | |
3460 | return -EINVAL; | |
3461 | ||
3462 | maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot); | |
3463 | if (write) | |
3464 | memcpy_toio(maddr + offset, buf, len); | |
3465 | else | |
3466 | memcpy_fromio(buf, maddr + offset, len); | |
3467 | iounmap(maddr); | |
3468 | ||
3469 | return len; | |
3470 | } | |
3471 | EXPORT_SYMBOL_GPL(generic_access_phys); | |
3472 | #endif | |
3473 | ||
3474 | /* | |
3475 | * Access another process' address space as given in mm. If non-NULL, use the | |
3476 | * given task for page fault accounting. | |
3477 | */ | |
3478 | static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, | |
3479 | unsigned long addr, void *buf, int len, int write) | |
3480 | { | |
3481 | struct vm_area_struct *vma; | |
3482 | void *old_buf = buf; | |
3483 | ||
3484 | down_read(&mm->mmap_sem); | |
3485 | /* ignore errors, just check how much was successfully transferred */ | |
3486 | while (len) { | |
3487 | int bytes, ret, offset; | |
3488 | void *maddr; | |
3489 | struct page *page = NULL; | |
3490 | ||
3491 | ret = get_user_pages(tsk, mm, addr, 1, | |
3492 | write, 1, &page, &vma); | |
3493 | if (ret <= 0) { | |
3494 | #ifndef CONFIG_HAVE_IOREMAP_PROT | |
3495 | break; | |
3496 | #else | |
3497 | /* | |
3498 | * Check if this is a VM_IO | VM_PFNMAP VMA, which | |
3499 | * we can access using slightly different code. | |
3500 | */ | |
3501 | vma = find_vma(mm, addr); | |
3502 | if (!vma || vma->vm_start > addr) | |
3503 | break; | |
3504 | if (vma->vm_ops && vma->vm_ops->access) | |
3505 | ret = vma->vm_ops->access(vma, addr, buf, | |
3506 | len, write); | |
3507 | if (ret <= 0) | |
3508 | break; | |
3509 | bytes = ret; | |
3510 | #endif | |
3511 | } else { | |
3512 | bytes = len; | |
3513 | offset = addr & (PAGE_SIZE-1); | |
3514 | if (bytes > PAGE_SIZE-offset) | |
3515 | bytes = PAGE_SIZE-offset; | |
3516 | ||
3517 | maddr = kmap(page); | |
3518 | if (write) { | |
3519 | copy_to_user_page(vma, page, addr, | |
3520 | maddr + offset, buf, bytes); | |
3521 | set_page_dirty_lock(page); | |
3522 | } else { | |
3523 | copy_from_user_page(vma, page, addr, | |
3524 | buf, maddr + offset, bytes); | |
3525 | } | |
3526 | kunmap(page); | |
3527 | page_cache_release(page); | |
3528 | } | |
3529 | len -= bytes; | |
3530 | buf += bytes; | |
3531 | addr += bytes; | |
3532 | } | |
3533 | up_read(&mm->mmap_sem); | |
3534 | ||
3535 | return buf - old_buf; | |
3536 | } | |
3537 | ||
3538 | /** | |
3539 | * access_remote_vm - access another process' address space | |
3540 | * @mm: the mm_struct of the target address space | |
3541 | * @addr: start address to access | |
3542 | * @buf: source or destination buffer | |
3543 | * @len: number of bytes to transfer | |
3544 | * @write: whether the access is a write | |
3545 | * | |
3546 | * The caller must hold a reference on @mm. | |
3547 | */ | |
3548 | int access_remote_vm(struct mm_struct *mm, unsigned long addr, | |
3549 | void *buf, int len, int write) | |
3550 | { | |
3551 | return __access_remote_vm(NULL, mm, addr, buf, len, write); | |
3552 | } | |
3553 | ||
3554 | /* | |
3555 | * Access another process' address space. | |
3556 | * Source/target buffer must be kernel space, | |
3557 | * Do not walk the page table directly, use get_user_pages | |
3558 | */ | |
3559 | int access_process_vm(struct task_struct *tsk, unsigned long addr, | |
3560 | void *buf, int len, int write) | |
3561 | { | |
3562 | struct mm_struct *mm; | |
3563 | int ret; | |
3564 | ||
3565 | mm = get_task_mm(tsk); | |
3566 | if (!mm) | |
3567 | return 0; | |
3568 | ||
3569 | ret = __access_remote_vm(tsk, mm, addr, buf, len, write); | |
3570 | mmput(mm); | |
3571 | ||
3572 | return ret; | |
3573 | } | |
3574 | ||
3575 | /* | |
3576 | * Print the name of a VMA. | |
3577 | */ | |
3578 | void print_vma_addr(char *prefix, unsigned long ip) | |
3579 | { | |
3580 | struct mm_struct *mm = current->mm; | |
3581 | struct vm_area_struct *vma; | |
3582 | ||
3583 | /* | |
3584 | * Do not print if we are in atomic | |
3585 | * contexts (in exception stacks, etc.): | |
3586 | */ | |
3587 | if (preempt_count()) | |
3588 | return; | |
3589 | ||
3590 | down_read(&mm->mmap_sem); | |
3591 | vma = find_vma(mm, ip); | |
3592 | if (vma && vma->vm_file) { | |
3593 | struct file *f = vma->vm_file; | |
3594 | char *buf = (char *)__get_free_page(GFP_KERNEL); | |
3595 | if (buf) { | |
3596 | char *p; | |
3597 | ||
3598 | p = d_path(&f->f_path, buf, PAGE_SIZE); | |
3599 | if (IS_ERR(p)) | |
3600 | p = "?"; | |
3601 | printk("%s%s[%lx+%lx]", prefix, kbasename(p), | |
3602 | vma->vm_start, | |
3603 | vma->vm_end - vma->vm_start); | |
3604 | free_page((unsigned long)buf); | |
3605 | } | |
3606 | } | |
3607 | up_read(&mm->mmap_sem); | |
3608 | } | |
3609 | ||
3610 | #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) | |
3611 | void might_fault(void) | |
3612 | { | |
3613 | /* | |
3614 | * Some code (nfs/sunrpc) uses socket ops on kernel memory while | |
3615 | * holding the mmap_sem, this is safe because kernel memory doesn't | |
3616 | * get paged out, therefore we'll never actually fault, and the | |
3617 | * below annotations will generate false positives. | |
3618 | */ | |
3619 | if (segment_eq(get_fs(), KERNEL_DS)) | |
3620 | return; | |
3621 | ||
3622 | /* | |
3623 | * it would be nicer only to annotate paths which are not under | |
3624 | * pagefault_disable, however that requires a larger audit and | |
3625 | * providing helpers like get_user_atomic. | |
3626 | */ | |
3627 | if (in_atomic()) | |
3628 | return; | |
3629 | ||
3630 | __might_sleep(__FILE__, __LINE__, 0); | |
3631 | ||
3632 | if (current->mm) | |
3633 | might_lock_read(¤t->mm->mmap_sem); | |
3634 | } | |
3635 | EXPORT_SYMBOL(might_fault); | |
3636 | #endif | |
3637 | ||
3638 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) | |
3639 | static void clear_gigantic_page(struct page *page, | |
3640 | unsigned long addr, | |
3641 | unsigned int pages_per_huge_page) | |
3642 | { | |
3643 | int i; | |
3644 | struct page *p = page; | |
3645 | ||
3646 | might_sleep(); | |
3647 | for (i = 0; i < pages_per_huge_page; | |
3648 | i++, p = mem_map_next(p, page, i)) { | |
3649 | cond_resched(); | |
3650 | clear_user_highpage(p, addr + i * PAGE_SIZE); | |
3651 | } | |
3652 | } | |
3653 | void clear_huge_page(struct page *page, | |
3654 | unsigned long addr, unsigned int pages_per_huge_page) | |
3655 | { | |
3656 | int i; | |
3657 | ||
3658 | if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { | |
3659 | clear_gigantic_page(page, addr, pages_per_huge_page); | |
3660 | return; | |
3661 | } | |
3662 | ||
3663 | might_sleep(); | |
3664 | for (i = 0; i < pages_per_huge_page; i++) { | |
3665 | cond_resched(); | |
3666 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); | |
3667 | } | |
3668 | } | |
3669 | ||
3670 | static void copy_user_gigantic_page(struct page *dst, struct page *src, | |
3671 | unsigned long addr, | |
3672 | struct vm_area_struct *vma, | |
3673 | unsigned int pages_per_huge_page) | |
3674 | { | |
3675 | int i; | |
3676 | struct page *dst_base = dst; | |
3677 | struct page *src_base = src; | |
3678 | ||
3679 | for (i = 0; i < pages_per_huge_page; ) { | |
3680 | cond_resched(); | |
3681 | copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); | |
3682 | ||
3683 | i++; | |
3684 | dst = mem_map_next(dst, dst_base, i); | |
3685 | src = mem_map_next(src, src_base, i); | |
3686 | } | |
3687 | } | |
3688 | ||
3689 | void copy_user_huge_page(struct page *dst, struct page *src, | |
3690 | unsigned long addr, struct vm_area_struct *vma, | |
3691 | unsigned int pages_per_huge_page) | |
3692 | { | |
3693 | int i; | |
3694 | ||
3695 | if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { | |
3696 | copy_user_gigantic_page(dst, src, addr, vma, | |
3697 | pages_per_huge_page); | |
3698 | return; | |
3699 | } | |
3700 | ||
3701 | might_sleep(); | |
3702 | for (i = 0; i < pages_per_huge_page; i++) { | |
3703 | cond_resched(); | |
3704 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); | |
3705 | } | |
3706 | } | |
3707 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ | |
3708 | ||
3709 | #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS | |
3710 | ||
3711 | static struct kmem_cache *page_ptl_cachep; | |
3712 | ||
3713 | void __init ptlock_cache_init(void) | |
3714 | { | |
3715 | page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, | |
3716 | SLAB_PANIC, NULL); | |
3717 | } | |
3718 | ||
3719 | bool ptlock_alloc(struct page *page) | |
3720 | { | |
3721 | spinlock_t *ptl; | |
3722 | ||
3723 | ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); | |
3724 | if (!ptl) | |
3725 | return false; | |
3726 | page->ptl = ptl; | |
3727 | return true; | |
3728 | } | |
3729 | ||
3730 | void ptlock_free(struct page *page) | |
3731 | { | |
3732 | kmem_cache_free(page_ptl_cachep, page->ptl); | |
3733 | } | |
3734 | #endif |