Root/
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->start = start; |
224 | tlb->end = end; |
225 | tlb->need_flush = 0; |
226 | tlb->local.next = NULL; |
227 | tlb->local.nr = 0; |
228 | tlb->local.max = ARRAY_SIZE(tlb->__pages); |
229 | tlb->active = &tlb->local; |
230 | tlb->batch_count = 0; |
231 | |
232 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE |
233 | tlb->batch = NULL; |
234 | #endif |
235 | } |
236 | |
237 | static void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb) |
238 | { |
239 | tlb->need_flush = 0; |
240 | tlb_flush(tlb); |
241 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE |
242 | tlb_table_flush(tlb); |
243 | #endif |
244 | } |
245 | |
246 | static void tlb_flush_mmu_free(struct mmu_gather *tlb) |
247 | { |
248 | struct mmu_gather_batch *batch; |
249 | |
250 | for (batch = &tlb->local; batch; batch = batch->next) { |
251 | free_pages_and_swap_cache(batch->pages, batch->nr); |
252 | batch->nr = 0; |
253 | } |
254 | tlb->active = &tlb->local; |
255 | } |
256 | |
257 | void tlb_flush_mmu(struct mmu_gather *tlb) |
258 | { |
259 | if (!tlb->need_flush) |
260 | return; |
261 | tlb_flush_mmu_tlbonly(tlb); |
262 | tlb_flush_mmu_free(tlb); |
263 | } |
264 | |
265 | /* tlb_finish_mmu |
266 | * Called at the end of the shootdown operation to free up any resources |
267 | * that were required. |
268 | */ |
269 | void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) |
270 | { |
271 | struct mmu_gather_batch *batch, *next; |
272 | |
273 | tlb_flush_mmu(tlb); |
274 | |
275 | /* keep the page table cache within bounds */ |
276 | check_pgt_cache(); |
277 | |
278 | for (batch = tlb->local.next; batch; batch = next) { |
279 | next = batch->next; |
280 | free_pages((unsigned long)batch, 0); |
281 | } |
282 | tlb->local.next = NULL; |
283 | } |
284 | |
285 | /* __tlb_remove_page |
286 | * Must perform the equivalent to __free_pte(pte_get_and_clear(ptep)), while |
287 | * handling the additional races in SMP caused by other CPUs caching valid |
288 | * mappings in their TLBs. Returns the number of free page slots left. |
289 | * When out of page slots we must call tlb_flush_mmu(). |
290 | */ |
291 | int __tlb_remove_page(struct mmu_gather *tlb, struct page *page) |
292 | { |
293 | struct mmu_gather_batch *batch; |
294 | |
295 | VM_BUG_ON(!tlb->need_flush); |
296 | |
297 | batch = tlb->active; |
298 | batch->pages[batch->nr++] = page; |
299 | if (batch->nr == batch->max) { |
300 | if (!tlb_next_batch(tlb)) |
301 | return 0; |
302 | batch = tlb->active; |
303 | } |
304 | VM_BUG_ON_PAGE(batch->nr > batch->max, page); |
305 | |
306 | return batch->max - batch->nr; |
307 | } |
308 | |
309 | #endif /* HAVE_GENERIC_MMU_GATHER */ |
310 | |
311 | #ifdef CONFIG_HAVE_RCU_TABLE_FREE |
312 | |
313 | /* |
314 | * See the comment near struct mmu_table_batch. |
315 | */ |
316 | |
317 | static void tlb_remove_table_smp_sync(void *arg) |
318 | { |
319 | /* Simply deliver the interrupt */ |
320 | } |
321 | |
322 | static void tlb_remove_table_one(void *table) |
323 | { |
324 | /* |
325 | * This isn't an RCU grace period and hence the page-tables cannot be |
326 | * assumed to be actually RCU-freed. |
327 | * |
328 | * It is however sufficient for software page-table walkers that rely on |
329 | * IRQ disabling. See the comment near struct mmu_table_batch. |
330 | */ |
331 | smp_call_function(tlb_remove_table_smp_sync, NULL, 1); |
332 | __tlb_remove_table(table); |
333 | } |
334 | |
335 | static void tlb_remove_table_rcu(struct rcu_head *head) |
336 | { |
337 | struct mmu_table_batch *batch; |
338 | int i; |
339 | |
340 | batch = container_of(head, struct mmu_table_batch, rcu); |
341 | |
342 | for (i = 0; i < batch->nr; i++) |
343 | __tlb_remove_table(batch->tables[i]); |
344 | |
345 | free_page((unsigned long)batch); |
346 | } |
347 | |
348 | void tlb_table_flush(struct mmu_gather *tlb) |
349 | { |
350 | struct mmu_table_batch **batch = &tlb->batch; |
351 | |
352 | if (*batch) { |
353 | call_rcu_sched(&(*batch)->rcu, tlb_remove_table_rcu); |
354 | *batch = NULL; |
355 | } |
356 | } |
357 | |
358 | void tlb_remove_table(struct mmu_gather *tlb, void *table) |
359 | { |
360 | struct mmu_table_batch **batch = &tlb->batch; |
361 | |
362 | tlb->need_flush = 1; |
363 | |
364 | /* |
365 | * When there's less then two users of this mm there cannot be a |
366 | * concurrent page-table walk. |
367 | */ |
368 | if (atomic_read(&tlb->mm->mm_users) < 2) { |
369 | __tlb_remove_table(table); |
370 | return; |
371 | } |
372 | |
373 | if (*batch == NULL) { |
374 | *batch = (struct mmu_table_batch *)__get_free_page(GFP_NOWAIT | __GFP_NOWARN); |
375 | if (*batch == NULL) { |
376 | tlb_remove_table_one(table); |
377 | return; |
378 | } |
379 | (*batch)->nr = 0; |
380 | } |
381 | (*batch)->tables[(*batch)->nr++] = table; |
382 | if ((*batch)->nr == MAX_TABLE_BATCH) |
383 | tlb_table_flush(tlb); |
384 | } |
385 | |
386 | #endif /* CONFIG_HAVE_RCU_TABLE_FREE */ |
387 | |
388 | /* |
389 | * Note: this doesn't free the actual pages themselves. That |
390 | * has been handled earlier when unmapping all the memory regions. |
391 | */ |
392 | static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd, |
393 | unsigned long addr) |
394 | { |
395 | pgtable_t token = pmd_pgtable(*pmd); |
396 | pmd_clear(pmd); |
397 | pte_free_tlb(tlb, token, addr); |
398 | atomic_long_dec(&tlb->mm->nr_ptes); |
399 | } |
400 | |
401 | static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud, |
402 | unsigned long addr, unsigned long end, |
403 | unsigned long floor, unsigned long ceiling) |
404 | { |
405 | pmd_t *pmd; |
406 | unsigned long next; |
407 | unsigned long start; |
408 | |
409 | start = addr; |
410 | pmd = pmd_offset(pud, addr); |
411 | do { |
412 | next = pmd_addr_end(addr, end); |
413 | if (pmd_none_or_clear_bad(pmd)) |
414 | continue; |
415 | free_pte_range(tlb, pmd, addr); |
416 | } while (pmd++, addr = next, addr != end); |
417 | |
418 | start &= PUD_MASK; |
419 | if (start < floor) |
420 | return; |
421 | if (ceiling) { |
422 | ceiling &= PUD_MASK; |
423 | if (!ceiling) |
424 | return; |
425 | } |
426 | if (end - 1 > ceiling - 1) |
427 | return; |
428 | |
429 | pmd = pmd_offset(pud, start); |
430 | pud_clear(pud); |
431 | pmd_free_tlb(tlb, pmd, start); |
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_flags & (VM_PFNMAP | VM_MIXEDMAP)) |
759 | return NULL; |
760 | if (!is_zero_pfn(pfn)) |
761 | print_bad_pte(vma, addr, pte, NULL); |
762 | return NULL; |
763 | } |
764 | |
765 | /* !HAVE_PTE_SPECIAL case follows: */ |
766 | |
767 | if (unlikely(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))) { |
768 | if (vma->vm_flags & VM_MIXEDMAP) { |
769 | if (!pfn_valid(pfn)) |
770 | return NULL; |
771 | goto out; |
772 | } else { |
773 | unsigned long off; |
774 | off = (addr - vma->vm_start) >> PAGE_SHIFT; |
775 | if (pfn == vma->vm_pgoff + off) |
776 | return NULL; |
777 | if (!is_cow_mapping(vma->vm_flags)) |
778 | return NULL; |
779 | } |
780 | } |
781 | |
782 | if (is_zero_pfn(pfn)) |
783 | return NULL; |
784 | check_pfn: |
785 | if (unlikely(pfn > highest_memmap_pfn)) { |
786 | print_bad_pte(vma, addr, pte, NULL); |
787 | return NULL; |
788 | } |
789 | |
790 | /* |
791 | * NOTE! We still have PageReserved() pages in the page tables. |
792 | * eg. VDSO mappings can cause them to exist. |
793 | */ |
794 | out: |
795 | return pfn_to_page(pfn); |
796 | } |
797 | |
798 | /* |
799 | * copy one vm_area from one task to the other. Assumes the page tables |
800 | * already present in the new task to be cleared in the whole range |
801 | * covered by this vma. |
802 | */ |
803 | |
804 | static inline unsigned long |
805 | copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm, |
806 | pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma, |
807 | unsigned long addr, int *rss) |
808 | { |
809 | unsigned long vm_flags = vma->vm_flags; |
810 | pte_t pte = *src_pte; |
811 | struct page *page; |
812 | |
813 | /* pte contains position in swap or file, so copy. */ |
814 | if (unlikely(!pte_present(pte))) { |
815 | if (!pte_file(pte)) { |
816 | swp_entry_t entry = pte_to_swp_entry(pte); |
817 | |
818 | if (likely(!non_swap_entry(entry))) { |
819 | if (swap_duplicate(entry) < 0) |
820 | return entry.val; |
821 | |
822 | /* make sure dst_mm is on swapoff's mmlist. */ |
823 | if (unlikely(list_empty(&dst_mm->mmlist))) { |
824 | spin_lock(&mmlist_lock); |
825 | if (list_empty(&dst_mm->mmlist)) |
826 | list_add(&dst_mm->mmlist, |
827 | &src_mm->mmlist); |
828 | spin_unlock(&mmlist_lock); |
829 | } |
830 | rss[MM_SWAPENTS]++; |
831 | } else if (is_migration_entry(entry)) { |
832 | page = migration_entry_to_page(entry); |
833 | |
834 | if (PageAnon(page)) |
835 | rss[MM_ANONPAGES]++; |
836 | else |
837 | rss[MM_FILEPAGES]++; |
838 | |
839 | if (is_write_migration_entry(entry) && |
840 | is_cow_mapping(vm_flags)) { |
841 | /* |
842 | * COW mappings require pages in both |
843 | * parent and child to be set to read. |
844 | */ |
845 | make_migration_entry_read(&entry); |
846 | pte = swp_entry_to_pte(entry); |
847 | if (pte_swp_soft_dirty(*src_pte)) |
848 | pte = pte_swp_mksoft_dirty(pte); |
849 | set_pte_at(src_mm, addr, src_pte, pte); |
850 | } |
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_NONLINEAR | |
1025 | VM_PFNMAP | VM_MIXEDMAP))) { |
1026 | if (!vma->anon_vma) |
1027 | return 0; |
1028 | } |
1029 | |
1030 | if (is_vm_hugetlb_page(vma)) |
1031 | return copy_hugetlb_page_range(dst_mm, src_mm, vma); |
1032 | |
1033 | if (unlikely(vma->vm_flags & VM_PFNMAP)) { |
1034 | /* |
1035 | * We do not free on error cases below as remove_vma |
1036 | * gets called on error from higher level routine |
1037 | */ |
1038 | ret = track_pfn_copy(vma); |
1039 | if (ret) |
1040 | return ret; |
1041 | } |
1042 | |
1043 | /* |
1044 | * We need to invalidate the secondary MMU mappings only when |
1045 | * there could be a permission downgrade on the ptes of the |
1046 | * parent mm. And a permission downgrade will only happen if |
1047 | * is_cow_mapping() returns true. |
1048 | */ |
1049 | is_cow = is_cow_mapping(vma->vm_flags); |
1050 | mmun_start = addr; |
1051 | mmun_end = end; |
1052 | if (is_cow) |
1053 | mmu_notifier_invalidate_range_start(src_mm, mmun_start, |
1054 | mmun_end); |
1055 | |
1056 | ret = 0; |
1057 | dst_pgd = pgd_offset(dst_mm, addr); |
1058 | src_pgd = pgd_offset(src_mm, addr); |
1059 | do { |
1060 | next = pgd_addr_end(addr, end); |
1061 | if (pgd_none_or_clear_bad(src_pgd)) |
1062 | continue; |
1063 | if (unlikely(copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd, |
1064 | vma, addr, next))) { |
1065 | ret = -ENOMEM; |
1066 | break; |
1067 | } |
1068 | } while (dst_pgd++, src_pgd++, addr = next, addr != end); |
1069 | |
1070 | if (is_cow) |
1071 | mmu_notifier_invalidate_range_end(src_mm, mmun_start, mmun_end); |
1072 | return ret; |
1073 | } |
1074 | |
1075 | static unsigned long zap_pte_range(struct mmu_gather *tlb, |
1076 | struct vm_area_struct *vma, pmd_t *pmd, |
1077 | unsigned long addr, unsigned long end, |
1078 | struct zap_details *details) |
1079 | { |
1080 | struct mm_struct *mm = tlb->mm; |
1081 | int force_flush = 0; |
1082 | int rss[NR_MM_COUNTERS]; |
1083 | spinlock_t *ptl; |
1084 | pte_t *start_pte; |
1085 | pte_t *pte; |
1086 | |
1087 | again: |
1088 | init_rss_vec(rss); |
1089 | start_pte = pte_offset_map_lock(mm, pmd, addr, &ptl); |
1090 | pte = start_pte; |
1091 | arch_enter_lazy_mmu_mode(); |
1092 | do { |
1093 | pte_t ptent = *pte; |
1094 | if (pte_none(ptent)) { |
1095 | continue; |
1096 | } |
1097 | |
1098 | if (pte_present(ptent)) { |
1099 | struct page *page; |
1100 | |
1101 | page = vm_normal_page(vma, addr, ptent); |
1102 | if (unlikely(details) && page) { |
1103 | /* |
1104 | * unmap_shared_mapping_pages() wants to |
1105 | * invalidate cache without truncating: |
1106 | * unmap shared but keep private pages. |
1107 | */ |
1108 | if (details->check_mapping && |
1109 | details->check_mapping != page->mapping) |
1110 | continue; |
1111 | /* |
1112 | * Each page->index must be checked when |
1113 | * invalidating or truncating nonlinear. |
1114 | */ |
1115 | if (details->nonlinear_vma && |
1116 | (page->index < details->first_index || |
1117 | page->index > details->last_index)) |
1118 | continue; |
1119 | } |
1120 | ptent = ptep_get_and_clear_full(mm, addr, pte, |
1121 | tlb->fullmm); |
1122 | tlb_remove_tlb_entry(tlb, pte, addr); |
1123 | if (unlikely(!page)) |
1124 | continue; |
1125 | if (unlikely(details) && details->nonlinear_vma |
1126 | && linear_page_index(details->nonlinear_vma, |
1127 | addr) != page->index) { |
1128 | pte_t ptfile = pgoff_to_pte(page->index); |
1129 | if (pte_soft_dirty(ptent)) |
1130 | ptfile = pte_file_mksoft_dirty(ptfile); |
1131 | set_pte_at(mm, addr, pte, ptfile); |
1132 | } |
1133 | if (PageAnon(page)) |
1134 | rss[MM_ANONPAGES]--; |
1135 | else { |
1136 | if (pte_dirty(ptent)) { |
1137 | force_flush = 1; |
1138 | set_page_dirty(page); |
1139 | } |
1140 | if (pte_young(ptent) && |
1141 | likely(!(vma->vm_flags & VM_SEQ_READ))) |
1142 | mark_page_accessed(page); |
1143 | rss[MM_FILEPAGES]--; |
1144 | } |
1145 | page_remove_rmap(page); |
1146 | if (unlikely(page_mapcount(page) < 0)) |
1147 | print_bad_pte(vma, addr, ptent, page); |
1148 | if (unlikely(!__tlb_remove_page(tlb, page))) { |
1149 | force_flush = 1; |
1150 | addr += PAGE_SIZE; |
1151 | break; |
1152 | } |
1153 | continue; |
1154 | } |
1155 | /* |
1156 | * If details->check_mapping, we leave swap entries; |
1157 | * if details->nonlinear_vma, we leave file entries. |
1158 | */ |
1159 | if (unlikely(details)) |
1160 | continue; |
1161 | if (pte_file(ptent)) { |
1162 | if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) |
1163 | print_bad_pte(vma, addr, ptent, NULL); |
1164 | } else { |
1165 | swp_entry_t entry = pte_to_swp_entry(ptent); |
1166 | |
1167 | if (!non_swap_entry(entry)) |
1168 | rss[MM_SWAPENTS]--; |
1169 | else if (is_migration_entry(entry)) { |
1170 | struct page *page; |
1171 | |
1172 | page = migration_entry_to_page(entry); |
1173 | |
1174 | if (PageAnon(page)) |
1175 | rss[MM_ANONPAGES]--; |
1176 | else |
1177 | rss[MM_FILEPAGES]--; |
1178 | } |
1179 | if (unlikely(!free_swap_and_cache(entry))) |
1180 | print_bad_pte(vma, addr, ptent, NULL); |
1181 | } |
1182 | pte_clear_not_present_full(mm, addr, pte, tlb->fullmm); |
1183 | } while (pte++, addr += PAGE_SIZE, addr != end); |
1184 | |
1185 | add_mm_rss_vec(mm, rss); |
1186 | arch_leave_lazy_mmu_mode(); |
1187 | |
1188 | /* Do the actual TLB flush before dropping ptl */ |
1189 | if (force_flush) { |
1190 | unsigned long old_end; |
1191 | |
1192 | /* |
1193 | * Flush the TLB just for the previous segment, |
1194 | * then update the range to be the remaining |
1195 | * TLB range. |
1196 | */ |
1197 | old_end = tlb->end; |
1198 | tlb->end = addr; |
1199 | tlb_flush_mmu_tlbonly(tlb); |
1200 | tlb->start = addr; |
1201 | tlb->end = old_end; |
1202 | } |
1203 | pte_unmap_unlock(start_pte, ptl); |
1204 | |
1205 | /* |
1206 | * If we forced a TLB flush (either due to running out of |
1207 | * batch buffers or because we needed to flush dirty TLB |
1208 | * entries before releasing the ptl), free the batched |
1209 | * memory too. Restart if we didn't do everything. |
1210 | */ |
1211 | if (force_flush) { |
1212 | force_flush = 0; |
1213 | tlb_flush_mmu_free(tlb); |
1214 | |
1215 | if (addr != end) |
1216 | goto again; |
1217 | } |
1218 | |
1219 | return addr; |
1220 | } |
1221 | |
1222 | static inline unsigned long zap_pmd_range(struct mmu_gather *tlb, |
1223 | struct vm_area_struct *vma, pud_t *pud, |
1224 | unsigned long addr, unsigned long end, |
1225 | struct zap_details *details) |
1226 | { |
1227 | pmd_t *pmd; |
1228 | unsigned long next; |
1229 | |
1230 | pmd = pmd_offset(pud, addr); |
1231 | do { |
1232 | next = pmd_addr_end(addr, end); |
1233 | if (pmd_trans_huge(*pmd)) { |
1234 | if (next - addr != HPAGE_PMD_SIZE) { |
1235 | #ifdef CONFIG_DEBUG_VM |
1236 | if (!rwsem_is_locked(&tlb->mm->mmap_sem)) { |
1237 | pr_err("%s: mmap_sem is unlocked! addr=0x%lx end=0x%lx vma->vm_start=0x%lx vma->vm_end=0x%lx\n", |
1238 | __func__, addr, end, |
1239 | vma->vm_start, |
1240 | vma->vm_end); |
1241 | BUG(); |
1242 | } |
1243 | #endif |
1244 | split_huge_page_pmd(vma, addr, pmd); |
1245 | } else if (zap_huge_pmd(tlb, vma, pmd, addr)) |
1246 | goto next; |
1247 | /* fall through */ |
1248 | } |
1249 | /* |
1250 | * Here there can be other concurrent MADV_DONTNEED or |
1251 | * trans huge page faults running, and if the pmd is |
1252 | * none or trans huge it can change under us. This is |
1253 | * because MADV_DONTNEED holds the mmap_sem in read |
1254 | * mode. |
1255 | */ |
1256 | if (pmd_none_or_trans_huge_or_clear_bad(pmd)) |
1257 | goto next; |
1258 | next = zap_pte_range(tlb, vma, pmd, addr, next, details); |
1259 | next: |
1260 | cond_resched(); |
1261 | } while (pmd++, addr = next, addr != end); |
1262 | |
1263 | return addr; |
1264 | } |
1265 | |
1266 | static inline unsigned long zap_pud_range(struct mmu_gather *tlb, |
1267 | struct vm_area_struct *vma, pgd_t *pgd, |
1268 | unsigned long addr, unsigned long end, |
1269 | struct zap_details *details) |
1270 | { |
1271 | pud_t *pud; |
1272 | unsigned long next; |
1273 | |
1274 | pud = pud_offset(pgd, addr); |
1275 | do { |
1276 | next = pud_addr_end(addr, end); |
1277 | if (pud_none_or_clear_bad(pud)) |
1278 | continue; |
1279 | next = zap_pmd_range(tlb, vma, pud, addr, next, details); |
1280 | } while (pud++, addr = next, addr != end); |
1281 | |
1282 | return addr; |
1283 | } |
1284 | |
1285 | static void unmap_page_range(struct mmu_gather *tlb, |
1286 | struct vm_area_struct *vma, |
1287 | unsigned long addr, unsigned long end, |
1288 | struct zap_details *details) |
1289 | { |
1290 | pgd_t *pgd; |
1291 | unsigned long next; |
1292 | |
1293 | if (details && !details->check_mapping && !details->nonlinear_vma) |
1294 | details = NULL; |
1295 | |
1296 | BUG_ON(addr >= end); |
1297 | tlb_start_vma(tlb, vma); |
1298 | pgd = pgd_offset(vma->vm_mm, addr); |
1299 | do { |
1300 | next = pgd_addr_end(addr, end); |
1301 | if (pgd_none_or_clear_bad(pgd)) |
1302 | continue; |
1303 | next = zap_pud_range(tlb, vma, pgd, addr, next, details); |
1304 | } while (pgd++, addr = next, addr != end); |
1305 | tlb_end_vma(tlb, vma); |
1306 | } |
1307 | |
1308 | |
1309 | static void unmap_single_vma(struct mmu_gather *tlb, |
1310 | struct vm_area_struct *vma, unsigned long start_addr, |
1311 | unsigned long end_addr, |
1312 | struct zap_details *details) |
1313 | { |
1314 | unsigned long start = max(vma->vm_start, start_addr); |
1315 | unsigned long end; |
1316 | |
1317 | if (start >= vma->vm_end) |
1318 | return; |
1319 | end = min(vma->vm_end, end_addr); |
1320 | if (end <= vma->vm_start) |
1321 | return; |
1322 | |
1323 | if (vma->vm_file) |
1324 | uprobe_munmap(vma, start, end); |
1325 | |
1326 | if (unlikely(vma->vm_flags & VM_PFNMAP)) |
1327 | untrack_pfn(vma, 0, 0); |
1328 | |
1329 | if (start != end) { |
1330 | if (unlikely(is_vm_hugetlb_page(vma))) { |
1331 | /* |
1332 | * It is undesirable to test vma->vm_file as it |
1333 | * should be non-null for valid hugetlb area. |
1334 | * However, vm_file will be NULL in the error |
1335 | * cleanup path of mmap_region. When |
1336 | * hugetlbfs ->mmap method fails, |
1337 | * mmap_region() nullifies vma->vm_file |
1338 | * before calling this function to clean up. |
1339 | * Since no pte has actually been setup, it is |
1340 | * safe to do nothing in this case. |
1341 | */ |
1342 | if (vma->vm_file) { |
1343 | mutex_lock(&vma->vm_file->f_mapping->i_mmap_mutex); |
1344 | __unmap_hugepage_range_final(tlb, vma, start, end, NULL); |
1345 | mutex_unlock(&vma->vm_file->f_mapping->i_mmap_mutex); |
1346 | } |
1347 | } else |
1348 | unmap_page_range(tlb, vma, start, end, details); |
1349 | } |
1350 | } |
1351 | |
1352 | /** |
1353 | * unmap_vmas - unmap a range of memory covered by a list of vma's |
1354 | * @tlb: address of the caller's struct mmu_gather |
1355 | * @vma: the starting vma |
1356 | * @start_addr: virtual address at which to start unmapping |
1357 | * @end_addr: virtual address at which to end unmapping |
1358 | * |
1359 | * Unmap all pages in the vma list. |
1360 | * |
1361 | * Only addresses between `start' and `end' will be unmapped. |
1362 | * |
1363 | * The VMA list must be sorted in ascending virtual address order. |
1364 | * |
1365 | * unmap_vmas() assumes that the caller will flush the whole unmapped address |
1366 | * range after unmap_vmas() returns. So the only responsibility here is to |
1367 | * ensure that any thus-far unmapped pages are flushed before unmap_vmas() |
1368 | * drops the lock and schedules. |
1369 | */ |
1370 | void unmap_vmas(struct mmu_gather *tlb, |
1371 | struct vm_area_struct *vma, unsigned long start_addr, |
1372 | unsigned long end_addr) |
1373 | { |
1374 | struct mm_struct *mm = vma->vm_mm; |
1375 | |
1376 | mmu_notifier_invalidate_range_start(mm, start_addr, end_addr); |
1377 | for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) |
1378 | unmap_single_vma(tlb, vma, start_addr, end_addr, NULL); |
1379 | mmu_notifier_invalidate_range_end(mm, start_addr, end_addr); |
1380 | } |
1381 | |
1382 | /** |
1383 | * zap_page_range - remove user pages in a given range |
1384 | * @vma: vm_area_struct holding the applicable pages |
1385 | * @start: starting address of pages to zap |
1386 | * @size: number of bytes to zap |
1387 | * @details: details of nonlinear truncation or shared cache invalidation |
1388 | * |
1389 | * Caller must protect the VMA list |
1390 | */ |
1391 | void zap_page_range(struct vm_area_struct *vma, unsigned long start, |
1392 | unsigned long size, struct zap_details *details) |
1393 | { |
1394 | struct mm_struct *mm = vma->vm_mm; |
1395 | struct mmu_gather tlb; |
1396 | unsigned long end = start + size; |
1397 | |
1398 | lru_add_drain(); |
1399 | tlb_gather_mmu(&tlb, mm, start, end); |
1400 | update_hiwater_rss(mm); |
1401 | mmu_notifier_invalidate_range_start(mm, start, end); |
1402 | for ( ; vma && vma->vm_start < end; vma = vma->vm_next) |
1403 | unmap_single_vma(&tlb, vma, start, end, details); |
1404 | mmu_notifier_invalidate_range_end(mm, start, end); |
1405 | tlb_finish_mmu(&tlb, start, end); |
1406 | } |
1407 | |
1408 | /** |
1409 | * zap_page_range_single - remove user pages in a given range |
1410 | * @vma: vm_area_struct holding the applicable pages |
1411 | * @address: starting address of pages to zap |
1412 | * @size: number of bytes to zap |
1413 | * @details: details of nonlinear truncation or shared cache invalidation |
1414 | * |
1415 | * The range must fit into one VMA. |
1416 | */ |
1417 | static void zap_page_range_single(struct vm_area_struct *vma, unsigned long address, |
1418 | unsigned long size, struct zap_details *details) |
1419 | { |
1420 | struct mm_struct *mm = vma->vm_mm; |
1421 | struct mmu_gather tlb; |
1422 | unsigned long end = address + size; |
1423 | |
1424 | lru_add_drain(); |
1425 | tlb_gather_mmu(&tlb, mm, address, end); |
1426 | update_hiwater_rss(mm); |
1427 | mmu_notifier_invalidate_range_start(mm, address, end); |
1428 | unmap_single_vma(&tlb, vma, address, end, details); |
1429 | mmu_notifier_invalidate_range_end(mm, address, end); |
1430 | tlb_finish_mmu(&tlb, address, end); |
1431 | } |
1432 | |
1433 | /** |
1434 | * zap_vma_ptes - remove ptes mapping the vma |
1435 | * @vma: vm_area_struct holding ptes to be zapped |
1436 | * @address: starting address of pages to zap |
1437 | * @size: number of bytes to zap |
1438 | * |
1439 | * This function only unmaps ptes assigned to VM_PFNMAP vmas. |
1440 | * |
1441 | * The entire address range must be fully contained within the vma. |
1442 | * |
1443 | * Returns 0 if successful. |
1444 | */ |
1445 | int zap_vma_ptes(struct vm_area_struct *vma, unsigned long address, |
1446 | unsigned long size) |
1447 | { |
1448 | if (address < vma->vm_start || address + size > vma->vm_end || |
1449 | !(vma->vm_flags & VM_PFNMAP)) |
1450 | return -1; |
1451 | zap_page_range_single(vma, address, size, NULL); |
1452 | return 0; |
1453 | } |
1454 | EXPORT_SYMBOL_GPL(zap_vma_ptes); |
1455 | |
1456 | pte_t *__get_locked_pte(struct mm_struct *mm, unsigned long addr, |
1457 | spinlock_t **ptl) |
1458 | { |
1459 | pgd_t * pgd = pgd_offset(mm, addr); |
1460 | pud_t * pud = pud_alloc(mm, pgd, addr); |
1461 | if (pud) { |
1462 | pmd_t * pmd = pmd_alloc(mm, pud, addr); |
1463 | if (pmd) { |
1464 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
1465 | return pte_alloc_map_lock(mm, pmd, addr, ptl); |
1466 | } |
1467 | } |
1468 | return NULL; |
1469 | } |
1470 | |
1471 | /* |
1472 | * This is the old fallback for page remapping. |
1473 | * |
1474 | * For historical reasons, it only allows reserved pages. Only |
1475 | * old drivers should use this, and they needed to mark their |
1476 | * pages reserved for the old functions anyway. |
1477 | */ |
1478 | static int insert_page(struct vm_area_struct *vma, unsigned long addr, |
1479 | struct page *page, pgprot_t prot) |
1480 | { |
1481 | struct mm_struct *mm = vma->vm_mm; |
1482 | int retval; |
1483 | pte_t *pte; |
1484 | spinlock_t *ptl; |
1485 | |
1486 | retval = -EINVAL; |
1487 | if (PageAnon(page)) |
1488 | goto out; |
1489 | retval = -ENOMEM; |
1490 | flush_dcache_page(page); |
1491 | pte = get_locked_pte(mm, addr, &ptl); |
1492 | if (!pte) |
1493 | goto out; |
1494 | retval = -EBUSY; |
1495 | if (!pte_none(*pte)) |
1496 | goto out_unlock; |
1497 | |
1498 | /* Ok, finally just insert the thing.. */ |
1499 | get_page(page); |
1500 | inc_mm_counter_fast(mm, MM_FILEPAGES); |
1501 | page_add_file_rmap(page); |
1502 | set_pte_at(mm, addr, pte, mk_pte(page, prot)); |
1503 | |
1504 | retval = 0; |
1505 | pte_unmap_unlock(pte, ptl); |
1506 | return retval; |
1507 | out_unlock: |
1508 | pte_unmap_unlock(pte, ptl); |
1509 | out: |
1510 | return retval; |
1511 | } |
1512 | |
1513 | /** |
1514 | * vm_insert_page - insert single page into user vma |
1515 | * @vma: user vma to map to |
1516 | * @addr: target user address of this page |
1517 | * @page: source kernel page |
1518 | * |
1519 | * This allows drivers to insert individual pages they've allocated |
1520 | * into a user vma. |
1521 | * |
1522 | * The page has to be a nice clean _individual_ kernel allocation. |
1523 | * If you allocate a compound page, you need to have marked it as |
1524 | * such (__GFP_COMP), or manually just split the page up yourself |
1525 | * (see split_page()). |
1526 | * |
1527 | * NOTE! Traditionally this was done with "remap_pfn_range()" which |
1528 | * took an arbitrary page protection parameter. This doesn't allow |
1529 | * that. Your vma protection will have to be set up correctly, which |
1530 | * means that if you want a shared writable mapping, you'd better |
1531 | * ask for a shared writable mapping! |
1532 | * |
1533 | * The page does not need to be reserved. |
1534 | * |
1535 | * Usually this function is called from f_op->mmap() handler |
1536 | * under mm->mmap_sem write-lock, so it can change vma->vm_flags. |
1537 | * Caller must set VM_MIXEDMAP on vma if it wants to call this |
1538 | * function from other places, for example from page-fault handler. |
1539 | */ |
1540 | int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, |
1541 | struct page *page) |
1542 | { |
1543 | if (addr < vma->vm_start || addr >= vma->vm_end) |
1544 | return -EFAULT; |
1545 | if (!page_count(page)) |
1546 | return -EINVAL; |
1547 | if (!(vma->vm_flags & VM_MIXEDMAP)) { |
1548 | BUG_ON(down_read_trylock(&vma->vm_mm->mmap_sem)); |
1549 | BUG_ON(vma->vm_flags & VM_PFNMAP); |
1550 | vma->vm_flags |= VM_MIXEDMAP; |
1551 | } |
1552 | return insert_page(vma, addr, page, vma->vm_page_prot); |
1553 | } |
1554 | EXPORT_SYMBOL(vm_insert_page); |
1555 | |
1556 | static int insert_pfn(struct vm_area_struct *vma, unsigned long addr, |
1557 | unsigned long pfn, pgprot_t prot) |
1558 | { |
1559 | struct mm_struct *mm = vma->vm_mm; |
1560 | int retval; |
1561 | pte_t *pte, entry; |
1562 | spinlock_t *ptl; |
1563 | |
1564 | retval = -ENOMEM; |
1565 | pte = get_locked_pte(mm, addr, &ptl); |
1566 | if (!pte) |
1567 | goto out; |
1568 | retval = -EBUSY; |
1569 | if (!pte_none(*pte)) |
1570 | goto out_unlock; |
1571 | |
1572 | /* Ok, finally just insert the thing.. */ |
1573 | entry = pte_mkspecial(pfn_pte(pfn, prot)); |
1574 | set_pte_at(mm, addr, pte, entry); |
1575 | update_mmu_cache(vma, addr, pte); /* XXX: why not for insert_page? */ |
1576 | |
1577 | retval = 0; |
1578 | out_unlock: |
1579 | pte_unmap_unlock(pte, ptl); |
1580 | out: |
1581 | return retval; |
1582 | } |
1583 | |
1584 | /** |
1585 | * vm_insert_pfn - insert single pfn into user vma |
1586 | * @vma: user vma to map to |
1587 | * @addr: target user address of this page |
1588 | * @pfn: source kernel pfn |
1589 | * |
1590 | * Similar to vm_insert_page, this allows drivers to insert individual pages |
1591 | * they've allocated into a user vma. Same comments apply. |
1592 | * |
1593 | * This function should only be called from a vm_ops->fault handler, and |
1594 | * in that case the handler should return NULL. |
1595 | * |
1596 | * vma cannot be a COW mapping. |
1597 | * |
1598 | * As this is called only for pages that do not currently exist, we |
1599 | * do not need to flush old virtual caches or the TLB. |
1600 | */ |
1601 | int vm_insert_pfn(struct vm_area_struct *vma, unsigned long addr, |
1602 | unsigned long pfn) |
1603 | { |
1604 | int ret; |
1605 | pgprot_t pgprot = vma->vm_page_prot; |
1606 | /* |
1607 | * Technically, architectures with pte_special can avoid all these |
1608 | * restrictions (same for remap_pfn_range). However we would like |
1609 | * consistency in testing and feature parity among all, so we should |
1610 | * try to keep these invariants in place for everybody. |
1611 | */ |
1612 | BUG_ON(!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))); |
1613 | BUG_ON((vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP)) == |
1614 | (VM_PFNMAP|VM_MIXEDMAP)); |
1615 | BUG_ON((vma->vm_flags & VM_PFNMAP) && is_cow_mapping(vma->vm_flags)); |
1616 | BUG_ON((vma->vm_flags & VM_MIXEDMAP) && pfn_valid(pfn)); |
1617 | |
1618 | if (addr < vma->vm_start || addr >= vma->vm_end) |
1619 | return -EFAULT; |
1620 | if (track_pfn_insert(vma, &pgprot, pfn)) |
1621 | return -EINVAL; |
1622 | |
1623 | ret = insert_pfn(vma, addr, pfn, pgprot); |
1624 | |
1625 | return ret; |
1626 | } |
1627 | EXPORT_SYMBOL(vm_insert_pfn); |
1628 | |
1629 | int vm_insert_mixed(struct vm_area_struct *vma, unsigned long addr, |
1630 | unsigned long pfn) |
1631 | { |
1632 | BUG_ON(!(vma->vm_flags & VM_MIXEDMAP)); |
1633 | |
1634 | if (addr < vma->vm_start || addr >= vma->vm_end) |
1635 | return -EFAULT; |
1636 | |
1637 | /* |
1638 | * If we don't have pte special, then we have to use the pfn_valid() |
1639 | * based VM_MIXEDMAP scheme (see vm_normal_page), and thus we *must* |
1640 | * refcount the page if pfn_valid is true (hence insert_page rather |
1641 | * than insert_pfn). If a zero_pfn were inserted into a VM_MIXEDMAP |
1642 | * without pte special, it would there be refcounted as a normal page. |
1643 | */ |
1644 | if (!HAVE_PTE_SPECIAL && pfn_valid(pfn)) { |
1645 | struct page *page; |
1646 | |
1647 | page = pfn_to_page(pfn); |
1648 | return insert_page(vma, addr, page, vma->vm_page_prot); |
1649 | } |
1650 | return insert_pfn(vma, addr, pfn, vma->vm_page_prot); |
1651 | } |
1652 | EXPORT_SYMBOL(vm_insert_mixed); |
1653 | |
1654 | /* |
1655 | * maps a range of physical memory into the requested pages. the old |
1656 | * mappings are removed. any references to nonexistent pages results |
1657 | * in null mappings (currently treated as "copy-on-access") |
1658 | */ |
1659 | static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd, |
1660 | unsigned long addr, unsigned long end, |
1661 | unsigned long pfn, pgprot_t prot) |
1662 | { |
1663 | pte_t *pte; |
1664 | spinlock_t *ptl; |
1665 | |
1666 | pte = pte_alloc_map_lock(mm, pmd, addr, &ptl); |
1667 | if (!pte) |
1668 | return -ENOMEM; |
1669 | arch_enter_lazy_mmu_mode(); |
1670 | do { |
1671 | BUG_ON(!pte_none(*pte)); |
1672 | set_pte_at(mm, addr, pte, pte_mkspecial(pfn_pte(pfn, prot))); |
1673 | pfn++; |
1674 | } while (pte++, addr += PAGE_SIZE, addr != end); |
1675 | arch_leave_lazy_mmu_mode(); |
1676 | pte_unmap_unlock(pte - 1, ptl); |
1677 | return 0; |
1678 | } |
1679 | |
1680 | static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud, |
1681 | unsigned long addr, unsigned long end, |
1682 | unsigned long pfn, pgprot_t prot) |
1683 | { |
1684 | pmd_t *pmd; |
1685 | unsigned long next; |
1686 | |
1687 | pfn -= addr >> PAGE_SHIFT; |
1688 | pmd = pmd_alloc(mm, pud, addr); |
1689 | if (!pmd) |
1690 | return -ENOMEM; |
1691 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
1692 | do { |
1693 | next = pmd_addr_end(addr, end); |
1694 | if (remap_pte_range(mm, pmd, addr, next, |
1695 | pfn + (addr >> PAGE_SHIFT), prot)) |
1696 | return -ENOMEM; |
1697 | } while (pmd++, addr = next, addr != end); |
1698 | return 0; |
1699 | } |
1700 | |
1701 | static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd, |
1702 | unsigned long addr, unsigned long end, |
1703 | unsigned long pfn, pgprot_t prot) |
1704 | { |
1705 | pud_t *pud; |
1706 | unsigned long next; |
1707 | |
1708 | pfn -= addr >> PAGE_SHIFT; |
1709 | pud = pud_alloc(mm, pgd, addr); |
1710 | if (!pud) |
1711 | return -ENOMEM; |
1712 | do { |
1713 | next = pud_addr_end(addr, end); |
1714 | if (remap_pmd_range(mm, pud, addr, next, |
1715 | pfn + (addr >> PAGE_SHIFT), prot)) |
1716 | return -ENOMEM; |
1717 | } while (pud++, addr = next, addr != end); |
1718 | return 0; |
1719 | } |
1720 | |
1721 | /** |
1722 | * remap_pfn_range - remap kernel memory to userspace |
1723 | * @vma: user vma to map to |
1724 | * @addr: target user address to start at |
1725 | * @pfn: physical address of kernel memory |
1726 | * @size: size of map area |
1727 | * @prot: page protection flags for this mapping |
1728 | * |
1729 | * Note: this is only safe if the mm semaphore is held when called. |
1730 | */ |
1731 | int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr, |
1732 | unsigned long pfn, unsigned long size, pgprot_t prot) |
1733 | { |
1734 | pgd_t *pgd; |
1735 | unsigned long next; |
1736 | unsigned long end = addr + PAGE_ALIGN(size); |
1737 | struct mm_struct *mm = vma->vm_mm; |
1738 | int err; |
1739 | |
1740 | /* |
1741 | * Physically remapped pages are special. Tell the |
1742 | * rest of the world about it: |
1743 | * VM_IO tells people not to look at these pages |
1744 | * (accesses can have side effects). |
1745 | * VM_PFNMAP tells the core MM that the base pages are just |
1746 | * raw PFN mappings, and do not have a "struct page" associated |
1747 | * with them. |
1748 | * VM_DONTEXPAND |
1749 | * Disable vma merging and expanding with mremap(). |
1750 | * VM_DONTDUMP |
1751 | * Omit vma from core dump, even when VM_IO turned off. |
1752 | * |
1753 | * There's a horrible special case to handle copy-on-write |
1754 | * behaviour that some programs depend on. We mark the "original" |
1755 | * un-COW'ed pages by matching them up with "vma->vm_pgoff". |
1756 | * See vm_normal_page() for details. |
1757 | */ |
1758 | if (is_cow_mapping(vma->vm_flags)) { |
1759 | if (addr != vma->vm_start || end != vma->vm_end) |
1760 | return -EINVAL; |
1761 | vma->vm_pgoff = pfn; |
1762 | } |
1763 | |
1764 | err = track_pfn_remap(vma, &prot, pfn, addr, PAGE_ALIGN(size)); |
1765 | if (err) |
1766 | return -EINVAL; |
1767 | |
1768 | vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP; |
1769 | |
1770 | BUG_ON(addr >= end); |
1771 | pfn -= addr >> PAGE_SHIFT; |
1772 | pgd = pgd_offset(mm, addr); |
1773 | flush_cache_range(vma, addr, end); |
1774 | do { |
1775 | next = pgd_addr_end(addr, end); |
1776 | err = remap_pud_range(mm, pgd, addr, next, |
1777 | pfn + (addr >> PAGE_SHIFT), prot); |
1778 | if (err) |
1779 | break; |
1780 | } while (pgd++, addr = next, addr != end); |
1781 | |
1782 | if (err) |
1783 | untrack_pfn(vma, pfn, PAGE_ALIGN(size)); |
1784 | |
1785 | return err; |
1786 | } |
1787 | EXPORT_SYMBOL(remap_pfn_range); |
1788 | |
1789 | /** |
1790 | * vm_iomap_memory - remap memory to userspace |
1791 | * @vma: user vma to map to |
1792 | * @start: start of area |
1793 | * @len: size of area |
1794 | * |
1795 | * This is a simplified io_remap_pfn_range() for common driver use. The |
1796 | * driver just needs to give us the physical memory range to be mapped, |
1797 | * we'll figure out the rest from the vma information. |
1798 | * |
1799 | * NOTE! Some drivers might want to tweak vma->vm_page_prot first to get |
1800 | * whatever write-combining details or similar. |
1801 | */ |
1802 | int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len) |
1803 | { |
1804 | unsigned long vm_len, pfn, pages; |
1805 | |
1806 | /* Check that the physical memory area passed in looks valid */ |
1807 | if (start + len < start) |
1808 | return -EINVAL; |
1809 | /* |
1810 | * You *really* shouldn't map things that aren't page-aligned, |
1811 | * but we've historically allowed it because IO memory might |
1812 | * just have smaller alignment. |
1813 | */ |
1814 | len += start & ~PAGE_MASK; |
1815 | pfn = start >> PAGE_SHIFT; |
1816 | pages = (len + ~PAGE_MASK) >> PAGE_SHIFT; |
1817 | if (pfn + pages < pfn) |
1818 | return -EINVAL; |
1819 | |
1820 | /* We start the mapping 'vm_pgoff' pages into the area */ |
1821 | if (vma->vm_pgoff > pages) |
1822 | return -EINVAL; |
1823 | pfn += vma->vm_pgoff; |
1824 | pages -= vma->vm_pgoff; |
1825 | |
1826 | /* Can we fit all of the mapping? */ |
1827 | vm_len = vma->vm_end - vma->vm_start; |
1828 | if (vm_len >> PAGE_SHIFT > pages) |
1829 | return -EINVAL; |
1830 | |
1831 | /* Ok, let it rip */ |
1832 | return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot); |
1833 | } |
1834 | EXPORT_SYMBOL(vm_iomap_memory); |
1835 | |
1836 | static int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd, |
1837 | unsigned long addr, unsigned long end, |
1838 | pte_fn_t fn, void *data) |
1839 | { |
1840 | pte_t *pte; |
1841 | int err; |
1842 | pgtable_t token; |
1843 | spinlock_t *uninitialized_var(ptl); |
1844 | |
1845 | pte = (mm == &init_mm) ? |
1846 | pte_alloc_kernel(pmd, addr) : |
1847 | pte_alloc_map_lock(mm, pmd, addr, &ptl); |
1848 | if (!pte) |
1849 | return -ENOMEM; |
1850 | |
1851 | BUG_ON(pmd_huge(*pmd)); |
1852 | |
1853 | arch_enter_lazy_mmu_mode(); |
1854 | |
1855 | token = pmd_pgtable(*pmd); |
1856 | |
1857 | do { |
1858 | err = fn(pte++, token, addr, data); |
1859 | if (err) |
1860 | break; |
1861 | } while (addr += PAGE_SIZE, addr != end); |
1862 | |
1863 | arch_leave_lazy_mmu_mode(); |
1864 | |
1865 | if (mm != &init_mm) |
1866 | pte_unmap_unlock(pte-1, ptl); |
1867 | return err; |
1868 | } |
1869 | |
1870 | static int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud, |
1871 | unsigned long addr, unsigned long end, |
1872 | pte_fn_t fn, void *data) |
1873 | { |
1874 | pmd_t *pmd; |
1875 | unsigned long next; |
1876 | int err; |
1877 | |
1878 | BUG_ON(pud_huge(*pud)); |
1879 | |
1880 | pmd = pmd_alloc(mm, pud, addr); |
1881 | if (!pmd) |
1882 | return -ENOMEM; |
1883 | do { |
1884 | next = pmd_addr_end(addr, end); |
1885 | err = apply_to_pte_range(mm, pmd, addr, next, fn, data); |
1886 | if (err) |
1887 | break; |
1888 | } while (pmd++, addr = next, addr != end); |
1889 | return err; |
1890 | } |
1891 | |
1892 | static int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd, |
1893 | unsigned long addr, unsigned long end, |
1894 | pte_fn_t fn, void *data) |
1895 | { |
1896 | pud_t *pud; |
1897 | unsigned long next; |
1898 | int err; |
1899 | |
1900 | pud = pud_alloc(mm, pgd, addr); |
1901 | if (!pud) |
1902 | return -ENOMEM; |
1903 | do { |
1904 | next = pud_addr_end(addr, end); |
1905 | err = apply_to_pmd_range(mm, pud, addr, next, fn, data); |
1906 | if (err) |
1907 | break; |
1908 | } while (pud++, addr = next, addr != end); |
1909 | return err; |
1910 | } |
1911 | |
1912 | /* |
1913 | * Scan a region of virtual memory, filling in page tables as necessary |
1914 | * and calling a provided function on each leaf page table. |
1915 | */ |
1916 | int apply_to_page_range(struct mm_struct *mm, unsigned long addr, |
1917 | unsigned long size, pte_fn_t fn, void *data) |
1918 | { |
1919 | pgd_t *pgd; |
1920 | unsigned long next; |
1921 | unsigned long end = addr + size; |
1922 | int err; |
1923 | |
1924 | BUG_ON(addr >= end); |
1925 | pgd = pgd_offset(mm, addr); |
1926 | do { |
1927 | next = pgd_addr_end(addr, end); |
1928 | err = apply_to_pud_range(mm, pgd, addr, next, fn, data); |
1929 | if (err) |
1930 | break; |
1931 | } while (pgd++, addr = next, addr != end); |
1932 | |
1933 | return err; |
1934 | } |
1935 | EXPORT_SYMBOL_GPL(apply_to_page_range); |
1936 | |
1937 | /* |
1938 | * handle_pte_fault chooses page fault handler according to an entry |
1939 | * which was read non-atomically. Before making any commitment, on |
1940 | * those architectures or configurations (e.g. i386 with PAE) which |
1941 | * might give a mix of unmatched parts, do_swap_page and do_nonlinear_fault |
1942 | * must check under lock before unmapping the pte and proceeding |
1943 | * (but do_wp_page is only called after already making such a check; |
1944 | * and do_anonymous_page can safely check later on). |
1945 | */ |
1946 | static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd, |
1947 | pte_t *page_table, pte_t orig_pte) |
1948 | { |
1949 | int same = 1; |
1950 | #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT) |
1951 | if (sizeof(pte_t) > sizeof(unsigned long)) { |
1952 | spinlock_t *ptl = pte_lockptr(mm, pmd); |
1953 | spin_lock(ptl); |
1954 | same = pte_same(*page_table, orig_pte); |
1955 | spin_unlock(ptl); |
1956 | } |
1957 | #endif |
1958 | pte_unmap(page_table); |
1959 | return same; |
1960 | } |
1961 | |
1962 | static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va, struct vm_area_struct *vma) |
1963 | { |
1964 | debug_dma_assert_idle(src); |
1965 | |
1966 | /* |
1967 | * If the source page was a PFN mapping, we don't have |
1968 | * a "struct page" for it. We do a best-effort copy by |
1969 | * just copying from the original user address. If that |
1970 | * fails, we just zero-fill it. Live with it. |
1971 | */ |
1972 | if (unlikely(!src)) { |
1973 | void *kaddr = kmap_atomic(dst); |
1974 | void __user *uaddr = (void __user *)(va & PAGE_MASK); |
1975 | |
1976 | /* |
1977 | * This really shouldn't fail, because the page is there |
1978 | * in the page tables. But it might just be unreadable, |
1979 | * in which case we just give up and fill the result with |
1980 | * zeroes. |
1981 | */ |
1982 | if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE)) |
1983 | clear_page(kaddr); |
1984 | kunmap_atomic(kaddr); |
1985 | flush_dcache_page(dst); |
1986 | } else |
1987 | copy_user_highpage(dst, src, va, vma); |
1988 | } |
1989 | |
1990 | /* |
1991 | * Notify the address space that the page is about to become writable so that |
1992 | * it can prohibit this or wait for the page to get into an appropriate state. |
1993 | * |
1994 | * We do this without the lock held, so that it can sleep if it needs to. |
1995 | */ |
1996 | static int do_page_mkwrite(struct vm_area_struct *vma, struct page *page, |
1997 | unsigned long address) |
1998 | { |
1999 | struct vm_fault vmf; |
2000 | int ret; |
2001 | |
2002 | vmf.virtual_address = (void __user *)(address & PAGE_MASK); |
2003 | vmf.pgoff = page->index; |
2004 | vmf.flags = FAULT_FLAG_WRITE|FAULT_FLAG_MKWRITE; |
2005 | vmf.page = page; |
2006 | |
2007 | ret = vma->vm_ops->page_mkwrite(vma, &vmf); |
2008 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE))) |
2009 | return ret; |
2010 | if (unlikely(!(ret & VM_FAULT_LOCKED))) { |
2011 | lock_page(page); |
2012 | if (!page->mapping) { |
2013 | unlock_page(page); |
2014 | return 0; /* retry */ |
2015 | } |
2016 | ret |= VM_FAULT_LOCKED; |
2017 | } else |
2018 | VM_BUG_ON_PAGE(!PageLocked(page), page); |
2019 | return ret; |
2020 | } |
2021 | |
2022 | /* |
2023 | * This routine handles present pages, when users try to write |
2024 | * to a shared page. It is done by copying the page to a new address |
2025 | * and decrementing the shared-page counter for the old page. |
2026 | * |
2027 | * Note that this routine assumes that the protection checks have been |
2028 | * done by the caller (the low-level page fault routine in most cases). |
2029 | * Thus we can safely just mark it writable once we've done any necessary |
2030 | * COW. |
2031 | * |
2032 | * We also mark the page dirty at this point even though the page will |
2033 | * change only once the write actually happens. This avoids a few races, |
2034 | * and potentially makes it more efficient. |
2035 | * |
2036 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
2037 | * but allow concurrent faults), with pte both mapped and locked. |
2038 | * We return with mmap_sem still held, but pte unmapped and unlocked. |
2039 | */ |
2040 | static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2041 | unsigned long address, pte_t *page_table, pmd_t *pmd, |
2042 | spinlock_t *ptl, pte_t orig_pte) |
2043 | __releases(ptl) |
2044 | { |
2045 | struct page *old_page, *new_page = NULL; |
2046 | pte_t entry; |
2047 | int ret = 0; |
2048 | int page_mkwrite = 0; |
2049 | struct page *dirty_page = NULL; |
2050 | unsigned long mmun_start = 0; /* For mmu_notifiers */ |
2051 | unsigned long mmun_end = 0; /* For mmu_notifiers */ |
2052 | struct mem_cgroup *memcg; |
2053 | |
2054 | old_page = vm_normal_page(vma, address, orig_pte); |
2055 | if (!old_page) { |
2056 | /* |
2057 | * VM_MIXEDMAP !pfn_valid() case, or VM_SOFTDIRTY clear on a |
2058 | * VM_PFNMAP VMA. |
2059 | * |
2060 | * We should not cow pages in a shared writeable mapping. |
2061 | * Just mark the pages writable as we can't do any dirty |
2062 | * accounting on raw pfn maps. |
2063 | */ |
2064 | if ((vma->vm_flags & (VM_WRITE|VM_SHARED)) == |
2065 | (VM_WRITE|VM_SHARED)) |
2066 | goto reuse; |
2067 | goto gotten; |
2068 | } |
2069 | |
2070 | /* |
2071 | * Take out anonymous pages first, anonymous shared vmas are |
2072 | * not dirty accountable. |
2073 | */ |
2074 | if (PageAnon(old_page) && !PageKsm(old_page)) { |
2075 | if (!trylock_page(old_page)) { |
2076 | page_cache_get(old_page); |
2077 | pte_unmap_unlock(page_table, ptl); |
2078 | lock_page(old_page); |
2079 | page_table = pte_offset_map_lock(mm, pmd, address, |
2080 | &ptl); |
2081 | if (!pte_same(*page_table, orig_pte)) { |
2082 | unlock_page(old_page); |
2083 | goto unlock; |
2084 | } |
2085 | page_cache_release(old_page); |
2086 | } |
2087 | if (reuse_swap_page(old_page)) { |
2088 | /* |
2089 | * The page is all ours. Move it to our anon_vma so |
2090 | * the rmap code will not search our parent or siblings. |
2091 | * Protected against the rmap code by the page lock. |
2092 | */ |
2093 | page_move_anon_rmap(old_page, vma, address); |
2094 | unlock_page(old_page); |
2095 | goto reuse; |
2096 | } |
2097 | unlock_page(old_page); |
2098 | } else if (unlikely((vma->vm_flags & (VM_WRITE|VM_SHARED)) == |
2099 | (VM_WRITE|VM_SHARED))) { |
2100 | /* |
2101 | * Only catch write-faults on shared writable pages, |
2102 | * read-only shared pages can get COWed by |
2103 | * get_user_pages(.write=1, .force=1). |
2104 | */ |
2105 | if (vma->vm_ops && vma->vm_ops->page_mkwrite) { |
2106 | int tmp; |
2107 | page_cache_get(old_page); |
2108 | pte_unmap_unlock(page_table, ptl); |
2109 | tmp = do_page_mkwrite(vma, old_page, address); |
2110 | if (unlikely(!tmp || (tmp & |
2111 | (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { |
2112 | page_cache_release(old_page); |
2113 | return tmp; |
2114 | } |
2115 | /* |
2116 | * Since we dropped the lock we need to revalidate |
2117 | * the PTE as someone else may have changed it. If |
2118 | * they did, we just return, as we can count on the |
2119 | * MMU to tell us if they didn't also make it writable. |
2120 | */ |
2121 | page_table = pte_offset_map_lock(mm, pmd, address, |
2122 | &ptl); |
2123 | if (!pte_same(*page_table, orig_pte)) { |
2124 | unlock_page(old_page); |
2125 | goto unlock; |
2126 | } |
2127 | |
2128 | page_mkwrite = 1; |
2129 | } |
2130 | dirty_page = old_page; |
2131 | get_page(dirty_page); |
2132 | |
2133 | reuse: |
2134 | /* |
2135 | * Clear the pages cpupid information as the existing |
2136 | * information potentially belongs to a now completely |
2137 | * unrelated process. |
2138 | */ |
2139 | if (old_page) |
2140 | page_cpupid_xchg_last(old_page, (1 << LAST_CPUPID_SHIFT) - 1); |
2141 | |
2142 | flush_cache_page(vma, address, pte_pfn(orig_pte)); |
2143 | entry = pte_mkyoung(orig_pte); |
2144 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
2145 | if (ptep_set_access_flags(vma, address, page_table, entry,1)) |
2146 | update_mmu_cache(vma, address, page_table); |
2147 | pte_unmap_unlock(page_table, ptl); |
2148 | ret |= VM_FAULT_WRITE; |
2149 | |
2150 | if (!dirty_page) |
2151 | return ret; |
2152 | |
2153 | /* |
2154 | * Yes, Virginia, this is actually required to prevent a race |
2155 | * with clear_page_dirty_for_io() from clearing the page dirty |
2156 | * bit after it clear all dirty ptes, but before a racing |
2157 | * do_wp_page installs a dirty pte. |
2158 | * |
2159 | * do_shared_fault is protected similarly. |
2160 | */ |
2161 | if (!page_mkwrite) { |
2162 | wait_on_page_locked(dirty_page); |
2163 | set_page_dirty_balance(dirty_page); |
2164 | /* file_update_time outside page_lock */ |
2165 | if (vma->vm_file) |
2166 | file_update_time(vma->vm_file); |
2167 | } |
2168 | put_page(dirty_page); |
2169 | if (page_mkwrite) { |
2170 | struct address_space *mapping = dirty_page->mapping; |
2171 | |
2172 | set_page_dirty(dirty_page); |
2173 | unlock_page(dirty_page); |
2174 | page_cache_release(dirty_page); |
2175 | if (mapping) { |
2176 | /* |
2177 | * Some device drivers do not set page.mapping |
2178 | * but still dirty their pages |
2179 | */ |
2180 | balance_dirty_pages_ratelimited(mapping); |
2181 | } |
2182 | } |
2183 | |
2184 | return ret; |
2185 | } |
2186 | |
2187 | /* |
2188 | * Ok, we need to copy. Oh, well.. |
2189 | */ |
2190 | page_cache_get(old_page); |
2191 | gotten: |
2192 | pte_unmap_unlock(page_table, ptl); |
2193 | |
2194 | if (unlikely(anon_vma_prepare(vma))) |
2195 | goto oom; |
2196 | |
2197 | if (is_zero_pfn(pte_pfn(orig_pte))) { |
2198 | new_page = alloc_zeroed_user_highpage_movable(vma, address); |
2199 | if (!new_page) |
2200 | goto oom; |
2201 | } else { |
2202 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
2203 | if (!new_page) |
2204 | goto oom; |
2205 | cow_user_page(new_page, old_page, address, vma); |
2206 | } |
2207 | __SetPageUptodate(new_page); |
2208 | |
2209 | if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) |
2210 | goto oom_free_new; |
2211 | |
2212 | mmun_start = address & PAGE_MASK; |
2213 | mmun_end = mmun_start + PAGE_SIZE; |
2214 | mmu_notifier_invalidate_range_start(mm, mmun_start, mmun_end); |
2215 | |
2216 | /* |
2217 | * Re-check the pte - we dropped the lock |
2218 | */ |
2219 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); |
2220 | if (likely(pte_same(*page_table, orig_pte))) { |
2221 | if (old_page) { |
2222 | if (!PageAnon(old_page)) { |
2223 | dec_mm_counter_fast(mm, MM_FILEPAGES); |
2224 | inc_mm_counter_fast(mm, MM_ANONPAGES); |
2225 | } |
2226 | } else |
2227 | inc_mm_counter_fast(mm, MM_ANONPAGES); |
2228 | flush_cache_page(vma, address, pte_pfn(orig_pte)); |
2229 | entry = mk_pte(new_page, vma->vm_page_prot); |
2230 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
2231 | /* |
2232 | * Clear the pte entry and flush it first, before updating the |
2233 | * pte with the new entry. This will avoid a race condition |
2234 | * seen in the presence of one thread doing SMC and another |
2235 | * thread doing COW. |
2236 | */ |
2237 | ptep_clear_flush(vma, address, page_table); |
2238 | page_add_new_anon_rmap(new_page, vma, address); |
2239 | mem_cgroup_commit_charge(new_page, memcg, false); |
2240 | lru_cache_add_active_or_unevictable(new_page, vma); |
2241 | /* |
2242 | * We call the notify macro here because, when using secondary |
2243 | * mmu page tables (such as kvm shadow page tables), we want the |
2244 | * new page to be mapped directly into the secondary page table. |
2245 | */ |
2246 | set_pte_at_notify(mm, address, page_table, entry); |
2247 | update_mmu_cache(vma, address, page_table); |
2248 | if (old_page) { |
2249 | /* |
2250 | * Only after switching the pte to the new page may |
2251 | * we remove the mapcount here. Otherwise another |
2252 | * process may come and find the rmap count decremented |
2253 | * before the pte is switched to the new page, and |
2254 | * "reuse" the old page writing into it while our pte |
2255 | * here still points into it and can be read by other |
2256 | * threads. |
2257 | * |
2258 | * The critical issue is to order this |
2259 | * page_remove_rmap with the ptp_clear_flush above. |
2260 | * Those stores are ordered by (if nothing else,) |
2261 | * the barrier present in the atomic_add_negative |
2262 | * in page_remove_rmap. |
2263 | * |
2264 | * Then the TLB flush in ptep_clear_flush ensures that |
2265 | * no process can access the old page before the |
2266 | * decremented mapcount is visible. And the old page |
2267 | * cannot be reused until after the decremented |
2268 | * mapcount is visible. So transitively, TLBs to |
2269 | * old page will be flushed before it can be reused. |
2270 | */ |
2271 | page_remove_rmap(old_page); |
2272 | } |
2273 | |
2274 | /* Free the old page.. */ |
2275 | new_page = old_page; |
2276 | ret |= VM_FAULT_WRITE; |
2277 | } else |
2278 | mem_cgroup_cancel_charge(new_page, memcg); |
2279 | |
2280 | if (new_page) |
2281 | page_cache_release(new_page); |
2282 | unlock: |
2283 | pte_unmap_unlock(page_table, ptl); |
2284 | if (mmun_end > mmun_start) |
2285 | mmu_notifier_invalidate_range_end(mm, mmun_start, mmun_end); |
2286 | if (old_page) { |
2287 | /* |
2288 | * Don't let another task, with possibly unlocked vma, |
2289 | * keep the mlocked page. |
2290 | */ |
2291 | if ((ret & VM_FAULT_WRITE) && (vma->vm_flags & VM_LOCKED)) { |
2292 | lock_page(old_page); /* LRU manipulation */ |
2293 | munlock_vma_page(old_page); |
2294 | unlock_page(old_page); |
2295 | } |
2296 | page_cache_release(old_page); |
2297 | } |
2298 | return ret; |
2299 | oom_free_new: |
2300 | page_cache_release(new_page); |
2301 | oom: |
2302 | if (old_page) |
2303 | page_cache_release(old_page); |
2304 | return VM_FAULT_OOM; |
2305 | } |
2306 | |
2307 | static void unmap_mapping_range_vma(struct vm_area_struct *vma, |
2308 | unsigned long start_addr, unsigned long end_addr, |
2309 | struct zap_details *details) |
2310 | { |
2311 | zap_page_range_single(vma, start_addr, end_addr - start_addr, details); |
2312 | } |
2313 | |
2314 | static inline void unmap_mapping_range_tree(struct rb_root *root, |
2315 | struct zap_details *details) |
2316 | { |
2317 | struct vm_area_struct *vma; |
2318 | pgoff_t vba, vea, zba, zea; |
2319 | |
2320 | vma_interval_tree_foreach(vma, root, |
2321 | details->first_index, details->last_index) { |
2322 | |
2323 | vba = vma->vm_pgoff; |
2324 | vea = vba + vma_pages(vma) - 1; |
2325 | /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */ |
2326 | zba = details->first_index; |
2327 | if (zba < vba) |
2328 | zba = vba; |
2329 | zea = details->last_index; |
2330 | if (zea > vea) |
2331 | zea = vea; |
2332 | |
2333 | unmap_mapping_range_vma(vma, |
2334 | ((zba - vba) << PAGE_SHIFT) + vma->vm_start, |
2335 | ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start, |
2336 | details); |
2337 | } |
2338 | } |
2339 | |
2340 | static inline void unmap_mapping_range_list(struct list_head *head, |
2341 | struct zap_details *details) |
2342 | { |
2343 | struct vm_area_struct *vma; |
2344 | |
2345 | /* |
2346 | * In nonlinear VMAs there is no correspondence between virtual address |
2347 | * offset and file offset. So we must perform an exhaustive search |
2348 | * across *all* the pages in each nonlinear VMA, not just the pages |
2349 | * whose virtual address lies outside the file truncation point. |
2350 | */ |
2351 | list_for_each_entry(vma, head, shared.nonlinear) { |
2352 | details->nonlinear_vma = vma; |
2353 | unmap_mapping_range_vma(vma, vma->vm_start, vma->vm_end, details); |
2354 | } |
2355 | } |
2356 | |
2357 | /** |
2358 | * unmap_mapping_range - unmap the portion of all mmaps in the specified address_space corresponding to the specified page range in the underlying file. |
2359 | * @mapping: the address space containing mmaps to be unmapped. |
2360 | * @holebegin: byte in first page to unmap, relative to the start of |
2361 | * the underlying file. This will be rounded down to a PAGE_SIZE |
2362 | * boundary. Note that this is different from truncate_pagecache(), which |
2363 | * must keep the partial page. In contrast, we must get rid of |
2364 | * partial pages. |
2365 | * @holelen: size of prospective hole in bytes. This will be rounded |
2366 | * up to a PAGE_SIZE boundary. A holelen of zero truncates to the |
2367 | * end of the file. |
2368 | * @even_cows: 1 when truncating a file, unmap even private COWed pages; |
2369 | * but 0 when invalidating pagecache, don't throw away private data. |
2370 | */ |
2371 | void unmap_mapping_range(struct address_space *mapping, |
2372 | loff_t const holebegin, loff_t const holelen, int even_cows) |
2373 | { |
2374 | struct zap_details details; |
2375 | pgoff_t hba = holebegin >> PAGE_SHIFT; |
2376 | pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; |
2377 | |
2378 | /* Check for overflow. */ |
2379 | if (sizeof(holelen) > sizeof(hlen)) { |
2380 | long long holeend = |
2381 | (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT; |
2382 | if (holeend & ~(long long)ULONG_MAX) |
2383 | hlen = ULONG_MAX - hba + 1; |
2384 | } |
2385 | |
2386 | details.check_mapping = even_cows? NULL: mapping; |
2387 | details.nonlinear_vma = NULL; |
2388 | details.first_index = hba; |
2389 | details.last_index = hba + hlen - 1; |
2390 | if (details.last_index < details.first_index) |
2391 | details.last_index = ULONG_MAX; |
2392 | |
2393 | |
2394 | mutex_lock(&mapping->i_mmap_mutex); |
2395 | if (unlikely(!RB_EMPTY_ROOT(&mapping->i_mmap))) |
2396 | unmap_mapping_range_tree(&mapping->i_mmap, &details); |
2397 | if (unlikely(!list_empty(&mapping->i_mmap_nonlinear))) |
2398 | unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details); |
2399 | mutex_unlock(&mapping->i_mmap_mutex); |
2400 | } |
2401 | EXPORT_SYMBOL(unmap_mapping_range); |
2402 | |
2403 | /* |
2404 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
2405 | * but allow concurrent faults), and pte mapped but not yet locked. |
2406 | * We return with pte unmapped and unlocked. |
2407 | * |
2408 | * We return with the mmap_sem locked or unlocked in the same cases |
2409 | * as does filemap_fault(). |
2410 | */ |
2411 | static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2412 | unsigned long address, pte_t *page_table, pmd_t *pmd, |
2413 | unsigned int flags, pte_t orig_pte) |
2414 | { |
2415 | spinlock_t *ptl; |
2416 | struct page *page, *swapcache; |
2417 | struct mem_cgroup *memcg; |
2418 | swp_entry_t entry; |
2419 | pte_t pte; |
2420 | int locked; |
2421 | int exclusive = 0; |
2422 | int ret = 0; |
2423 | |
2424 | if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) |
2425 | goto out; |
2426 | |
2427 | entry = pte_to_swp_entry(orig_pte); |
2428 | if (unlikely(non_swap_entry(entry))) { |
2429 | if (is_migration_entry(entry)) { |
2430 | migration_entry_wait(mm, pmd, address); |
2431 | } else if (is_hwpoison_entry(entry)) { |
2432 | ret = VM_FAULT_HWPOISON; |
2433 | } else { |
2434 | print_bad_pte(vma, address, orig_pte, NULL); |
2435 | ret = VM_FAULT_SIGBUS; |
2436 | } |
2437 | goto out; |
2438 | } |
2439 | delayacct_set_flag(DELAYACCT_PF_SWAPIN); |
2440 | page = lookup_swap_cache(entry); |
2441 | if (!page) { |
2442 | page = swapin_readahead(entry, |
2443 | GFP_HIGHUSER_MOVABLE, vma, address); |
2444 | if (!page) { |
2445 | /* |
2446 | * Back out if somebody else faulted in this pte |
2447 | * while we released the pte lock. |
2448 | */ |
2449 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); |
2450 | if (likely(pte_same(*page_table, orig_pte))) |
2451 | ret = VM_FAULT_OOM; |
2452 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); |
2453 | goto unlock; |
2454 | } |
2455 | |
2456 | /* Had to read the page from swap area: Major fault */ |
2457 | ret = VM_FAULT_MAJOR; |
2458 | count_vm_event(PGMAJFAULT); |
2459 | mem_cgroup_count_vm_event(mm, PGMAJFAULT); |
2460 | } else if (PageHWPoison(page)) { |
2461 | /* |
2462 | * hwpoisoned dirty swapcache pages are kept for killing |
2463 | * owner processes (which may be unknown at hwpoison time) |
2464 | */ |
2465 | ret = VM_FAULT_HWPOISON; |
2466 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); |
2467 | swapcache = page; |
2468 | goto out_release; |
2469 | } |
2470 | |
2471 | swapcache = page; |
2472 | locked = lock_page_or_retry(page, mm, flags); |
2473 | |
2474 | delayacct_clear_flag(DELAYACCT_PF_SWAPIN); |
2475 | if (!locked) { |
2476 | ret |= VM_FAULT_RETRY; |
2477 | goto out_release; |
2478 | } |
2479 | |
2480 | /* |
2481 | * Make sure try_to_free_swap or reuse_swap_page or swapoff did not |
2482 | * release the swapcache from under us. The page pin, and pte_same |
2483 | * test below, are not enough to exclude that. Even if it is still |
2484 | * swapcache, we need to check that the page's swap has not changed. |
2485 | */ |
2486 | if (unlikely(!PageSwapCache(page) || page_private(page) != entry.val)) |
2487 | goto out_page; |
2488 | |
2489 | page = ksm_might_need_to_copy(page, vma, address); |
2490 | if (unlikely(!page)) { |
2491 | ret = VM_FAULT_OOM; |
2492 | page = swapcache; |
2493 | goto out_page; |
2494 | } |
2495 | |
2496 | if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) { |
2497 | ret = VM_FAULT_OOM; |
2498 | goto out_page; |
2499 | } |
2500 | |
2501 | /* |
2502 | * Back out if somebody else already faulted in this pte. |
2503 | */ |
2504 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); |
2505 | if (unlikely(!pte_same(*page_table, orig_pte))) |
2506 | goto out_nomap; |
2507 | |
2508 | if (unlikely(!PageUptodate(page))) { |
2509 | ret = VM_FAULT_SIGBUS; |
2510 | goto out_nomap; |
2511 | } |
2512 | |
2513 | /* |
2514 | * The page isn't present yet, go ahead with the fault. |
2515 | * |
2516 | * Be careful about the sequence of operations here. |
2517 | * To get its accounting right, reuse_swap_page() must be called |
2518 | * while the page is counted on swap but not yet in mapcount i.e. |
2519 | * before page_add_anon_rmap() and swap_free(); try_to_free_swap() |
2520 | * must be called after the swap_free(), or it will never succeed. |
2521 | */ |
2522 | |
2523 | inc_mm_counter_fast(mm, MM_ANONPAGES); |
2524 | dec_mm_counter_fast(mm, MM_SWAPENTS); |
2525 | pte = mk_pte(page, vma->vm_page_prot); |
2526 | if ((flags & FAULT_FLAG_WRITE) && reuse_swap_page(page)) { |
2527 | pte = maybe_mkwrite(pte_mkdirty(pte), vma); |
2528 | flags &= ~FAULT_FLAG_WRITE; |
2529 | ret |= VM_FAULT_WRITE; |
2530 | exclusive = 1; |
2531 | } |
2532 | flush_icache_page(vma, page); |
2533 | if (pte_swp_soft_dirty(orig_pte)) |
2534 | pte = pte_mksoft_dirty(pte); |
2535 | set_pte_at(mm, address, page_table, pte); |
2536 | if (page == swapcache) { |
2537 | do_page_add_anon_rmap(page, vma, address, exclusive); |
2538 | mem_cgroup_commit_charge(page, memcg, true); |
2539 | } else { /* ksm created a completely new copy */ |
2540 | page_add_new_anon_rmap(page, vma, address); |
2541 | mem_cgroup_commit_charge(page, memcg, false); |
2542 | lru_cache_add_active_or_unevictable(page, vma); |
2543 | } |
2544 | |
2545 | swap_free(entry); |
2546 | if (vm_swap_full() || (vma->vm_flags & VM_LOCKED) || PageMlocked(page)) |
2547 | try_to_free_swap(page); |
2548 | unlock_page(page); |
2549 | if (page != swapcache) { |
2550 | /* |
2551 | * Hold the lock to avoid the swap entry to be reused |
2552 | * until we take the PT lock for the pte_same() check |
2553 | * (to avoid false positives from pte_same). For |
2554 | * further safety release the lock after the swap_free |
2555 | * so that the swap count won't change under a |
2556 | * parallel locked swapcache. |
2557 | */ |
2558 | unlock_page(swapcache); |
2559 | page_cache_release(swapcache); |
2560 | } |
2561 | |
2562 | if (flags & FAULT_FLAG_WRITE) { |
2563 | ret |= do_wp_page(mm, vma, address, page_table, pmd, ptl, pte); |
2564 | if (ret & VM_FAULT_ERROR) |
2565 | ret &= VM_FAULT_ERROR; |
2566 | goto out; |
2567 | } |
2568 | |
2569 | /* No need to invalidate - it was non-present before */ |
2570 | update_mmu_cache(vma, address, page_table); |
2571 | unlock: |
2572 | pte_unmap_unlock(page_table, ptl); |
2573 | out: |
2574 | return ret; |
2575 | out_nomap: |
2576 | mem_cgroup_cancel_charge(page, memcg); |
2577 | pte_unmap_unlock(page_table, ptl); |
2578 | out_page: |
2579 | unlock_page(page); |
2580 | out_release: |
2581 | page_cache_release(page); |
2582 | if (page != swapcache) { |
2583 | unlock_page(swapcache); |
2584 | page_cache_release(swapcache); |
2585 | } |
2586 | return ret; |
2587 | } |
2588 | |
2589 | /* |
2590 | * This is like a special single-page "expand_{down|up}wards()", |
2591 | * except we must first make sure that 'address{-|+}PAGE_SIZE' |
2592 | * doesn't hit another vma. |
2593 | */ |
2594 | static inline int check_stack_guard_page(struct vm_area_struct *vma, unsigned long address) |
2595 | { |
2596 | address &= PAGE_MASK; |
2597 | if ((vma->vm_flags & VM_GROWSDOWN) && address == vma->vm_start) { |
2598 | struct vm_area_struct *prev = vma->vm_prev; |
2599 | |
2600 | /* |
2601 | * Is there a mapping abutting this one below? |
2602 | * |
2603 | * That's only ok if it's the same stack mapping |
2604 | * that has gotten split.. |
2605 | */ |
2606 | if (prev && prev->vm_end == address) |
2607 | return prev->vm_flags & VM_GROWSDOWN ? 0 : -ENOMEM; |
2608 | |
2609 | expand_downwards(vma, address - PAGE_SIZE); |
2610 | } |
2611 | if ((vma->vm_flags & VM_GROWSUP) && address + PAGE_SIZE == vma->vm_end) { |
2612 | struct vm_area_struct *next = vma->vm_next; |
2613 | |
2614 | /* As VM_GROWSDOWN but s/below/above/ */ |
2615 | if (next && next->vm_start == address + PAGE_SIZE) |
2616 | return next->vm_flags & VM_GROWSUP ? 0 : -ENOMEM; |
2617 | |
2618 | expand_upwards(vma, address + PAGE_SIZE); |
2619 | } |
2620 | return 0; |
2621 | } |
2622 | |
2623 | /* |
2624 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
2625 | * but allow concurrent faults), and pte mapped but not yet locked. |
2626 | * We return with mmap_sem still held, but pte unmapped and unlocked. |
2627 | */ |
2628 | static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma, |
2629 | unsigned long address, pte_t *page_table, pmd_t *pmd, |
2630 | unsigned int flags) |
2631 | { |
2632 | struct mem_cgroup *memcg; |
2633 | struct page *page; |
2634 | spinlock_t *ptl; |
2635 | pte_t entry; |
2636 | |
2637 | pte_unmap(page_table); |
2638 | |
2639 | /* Check if we need to add a guard page to the stack */ |
2640 | if (check_stack_guard_page(vma, address) < 0) |
2641 | return VM_FAULT_SIGBUS; |
2642 | |
2643 | /* Use the zero-page for reads */ |
2644 | if (!(flags & FAULT_FLAG_WRITE)) { |
2645 | entry = pte_mkspecial(pfn_pte(my_zero_pfn(address), |
2646 | vma->vm_page_prot)); |
2647 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); |
2648 | if (!pte_none(*page_table)) |
2649 | goto unlock; |
2650 | goto setpte; |
2651 | } |
2652 | |
2653 | /* Allocate our own private page. */ |
2654 | if (unlikely(anon_vma_prepare(vma))) |
2655 | goto oom; |
2656 | page = alloc_zeroed_user_highpage_movable(vma, address); |
2657 | if (!page) |
2658 | goto oom; |
2659 | /* |
2660 | * The memory barrier inside __SetPageUptodate makes sure that |
2661 | * preceeding stores to the page contents become visible before |
2662 | * the set_pte_at() write. |
2663 | */ |
2664 | __SetPageUptodate(page); |
2665 | |
2666 | if (mem_cgroup_try_charge(page, mm, GFP_KERNEL, &memcg)) |
2667 | goto oom_free_page; |
2668 | |
2669 | entry = mk_pte(page, vma->vm_page_prot); |
2670 | if (vma->vm_flags & VM_WRITE) |
2671 | entry = pte_mkwrite(pte_mkdirty(entry)); |
2672 | |
2673 | page_table = pte_offset_map_lock(mm, pmd, address, &ptl); |
2674 | if (!pte_none(*page_table)) |
2675 | goto release; |
2676 | |
2677 | inc_mm_counter_fast(mm, MM_ANONPAGES); |
2678 | page_add_new_anon_rmap(page, vma, address); |
2679 | mem_cgroup_commit_charge(page, memcg, false); |
2680 | lru_cache_add_active_or_unevictable(page, vma); |
2681 | setpte: |
2682 | set_pte_at(mm, address, page_table, entry); |
2683 | |
2684 | /* No need to invalidate - it was non-present before */ |
2685 | update_mmu_cache(vma, address, page_table); |
2686 | unlock: |
2687 | pte_unmap_unlock(page_table, ptl); |
2688 | return 0; |
2689 | release: |
2690 | mem_cgroup_cancel_charge(page, memcg); |
2691 | page_cache_release(page); |
2692 | goto unlock; |
2693 | oom_free_page: |
2694 | page_cache_release(page); |
2695 | oom: |
2696 | return VM_FAULT_OOM; |
2697 | } |
2698 | |
2699 | /* |
2700 | * The mmap_sem must have been held on entry, and may have been |
2701 | * released depending on flags and vma->vm_ops->fault() return value. |
2702 | * See filemap_fault() and __lock_page_retry(). |
2703 | */ |
2704 | static int __do_fault(struct vm_area_struct *vma, unsigned long address, |
2705 | pgoff_t pgoff, unsigned int flags, struct page **page) |
2706 | { |
2707 | struct vm_fault vmf; |
2708 | int ret; |
2709 | |
2710 | vmf.virtual_address = (void __user *)(address & PAGE_MASK); |
2711 | vmf.pgoff = pgoff; |
2712 | vmf.flags = flags; |
2713 | vmf.page = NULL; |
2714 | |
2715 | ret = vma->vm_ops->fault(vma, &vmf); |
2716 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
2717 | return ret; |
2718 | |
2719 | if (unlikely(PageHWPoison(vmf.page))) { |
2720 | if (ret & VM_FAULT_LOCKED) |
2721 | unlock_page(vmf.page); |
2722 | page_cache_release(vmf.page); |
2723 | return VM_FAULT_HWPOISON; |
2724 | } |
2725 | |
2726 | if (unlikely(!(ret & VM_FAULT_LOCKED))) |
2727 | lock_page(vmf.page); |
2728 | else |
2729 | VM_BUG_ON_PAGE(!PageLocked(vmf.page), vmf.page); |
2730 | |
2731 | *page = vmf.page; |
2732 | return ret; |
2733 | } |
2734 | |
2735 | /** |
2736 | * do_set_pte - setup new PTE entry for given page and add reverse page mapping. |
2737 | * |
2738 | * @vma: virtual memory area |
2739 | * @address: user virtual address |
2740 | * @page: page to map |
2741 | * @pte: pointer to target page table entry |
2742 | * @write: true, if new entry is writable |
2743 | * @anon: true, if it's anonymous page |
2744 | * |
2745 | * Caller must hold page table lock relevant for @pte. |
2746 | * |
2747 | * Target users are page handler itself and implementations of |
2748 | * vm_ops->map_pages. |
2749 | */ |
2750 | void do_set_pte(struct vm_area_struct *vma, unsigned long address, |
2751 | struct page *page, pte_t *pte, bool write, bool anon) |
2752 | { |
2753 | pte_t entry; |
2754 | |
2755 | flush_icache_page(vma, page); |
2756 | entry = mk_pte(page, vma->vm_page_prot); |
2757 | if (write) |
2758 | entry = maybe_mkwrite(pte_mkdirty(entry), vma); |
2759 | else if (pte_file(*pte) && pte_file_soft_dirty(*pte)) |
2760 | entry = pte_mksoft_dirty(entry); |
2761 | if (anon) { |
2762 | inc_mm_counter_fast(vma->vm_mm, MM_ANONPAGES); |
2763 | page_add_new_anon_rmap(page, vma, address); |
2764 | } else { |
2765 | inc_mm_counter_fast(vma->vm_mm, MM_FILEPAGES); |
2766 | page_add_file_rmap(page); |
2767 | } |
2768 | set_pte_at(vma->vm_mm, address, pte, entry); |
2769 | |
2770 | /* no need to invalidate: a not-present page won't be cached */ |
2771 | update_mmu_cache(vma, address, pte); |
2772 | } |
2773 | |
2774 | static unsigned long fault_around_bytes __read_mostly = |
2775 | rounddown_pow_of_two(65536); |
2776 | |
2777 | #ifdef CONFIG_DEBUG_FS |
2778 | static int fault_around_bytes_get(void *data, u64 *val) |
2779 | { |
2780 | *val = fault_around_bytes; |
2781 | return 0; |
2782 | } |
2783 | |
2784 | /* |
2785 | * fault_around_pages() and fault_around_mask() expects fault_around_bytes |
2786 | * rounded down to nearest page order. It's what do_fault_around() expects to |
2787 | * see. |
2788 | */ |
2789 | static int fault_around_bytes_set(void *data, u64 val) |
2790 | { |
2791 | if (val / PAGE_SIZE > PTRS_PER_PTE) |
2792 | return -EINVAL; |
2793 | if (val > PAGE_SIZE) |
2794 | fault_around_bytes = rounddown_pow_of_two(val); |
2795 | else |
2796 | fault_around_bytes = PAGE_SIZE; /* rounddown_pow_of_two(0) is undefined */ |
2797 | return 0; |
2798 | } |
2799 | DEFINE_SIMPLE_ATTRIBUTE(fault_around_bytes_fops, |
2800 | fault_around_bytes_get, fault_around_bytes_set, "%llu\n"); |
2801 | |
2802 | static int __init fault_around_debugfs(void) |
2803 | { |
2804 | void *ret; |
2805 | |
2806 | ret = debugfs_create_file("fault_around_bytes", 0644, NULL, NULL, |
2807 | &fault_around_bytes_fops); |
2808 | if (!ret) |
2809 | pr_warn("Failed to create fault_around_bytes in debugfs"); |
2810 | return 0; |
2811 | } |
2812 | late_initcall(fault_around_debugfs); |
2813 | #endif |
2814 | |
2815 | /* |
2816 | * do_fault_around() tries to map few pages around the fault address. The hope |
2817 | * is that the pages will be needed soon and this will lower the number of |
2818 | * faults to handle. |
2819 | * |
2820 | * It uses vm_ops->map_pages() to map the pages, which skips the page if it's |
2821 | * not ready to be mapped: not up-to-date, locked, etc. |
2822 | * |
2823 | * This function is called with the page table lock taken. In the split ptlock |
2824 | * case the page table lock only protects only those entries which belong to |
2825 | * the page table corresponding to the fault address. |
2826 | * |
2827 | * This function doesn't cross the VMA boundaries, in order to call map_pages() |
2828 | * only once. |
2829 | * |
2830 | * fault_around_pages() defines how many pages we'll try to map. |
2831 | * do_fault_around() expects it to return a power of two less than or equal to |
2832 | * PTRS_PER_PTE. |
2833 | * |
2834 | * The virtual address of the area that we map is naturally aligned to the |
2835 | * fault_around_pages() value (and therefore to page order). This way it's |
2836 | * easier to guarantee that we don't cross page table boundaries. |
2837 | */ |
2838 | static void do_fault_around(struct vm_area_struct *vma, unsigned long address, |
2839 | pte_t *pte, pgoff_t pgoff, unsigned int flags) |
2840 | { |
2841 | unsigned long start_addr, nr_pages, mask; |
2842 | pgoff_t max_pgoff; |
2843 | struct vm_fault vmf; |
2844 | int off; |
2845 | |
2846 | nr_pages = ACCESS_ONCE(fault_around_bytes) >> PAGE_SHIFT; |
2847 | mask = ~(nr_pages * PAGE_SIZE - 1) & PAGE_MASK; |
2848 | |
2849 | start_addr = max(address & mask, vma->vm_start); |
2850 | off = ((address - start_addr) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1); |
2851 | pte -= off; |
2852 | pgoff -= off; |
2853 | |
2854 | /* |
2855 | * max_pgoff is either end of page table or end of vma |
2856 | * or fault_around_pages() from pgoff, depending what is nearest. |
2857 | */ |
2858 | max_pgoff = pgoff - ((start_addr >> PAGE_SHIFT) & (PTRS_PER_PTE - 1)) + |
2859 | PTRS_PER_PTE - 1; |
2860 | max_pgoff = min3(max_pgoff, vma_pages(vma) + vma->vm_pgoff - 1, |
2861 | pgoff + nr_pages - 1); |
2862 | |
2863 | /* Check if it makes any sense to call ->map_pages */ |
2864 | while (!pte_none(*pte)) { |
2865 | if (++pgoff > max_pgoff) |
2866 | return; |
2867 | start_addr += PAGE_SIZE; |
2868 | if (start_addr >= vma->vm_end) |
2869 | return; |
2870 | pte++; |
2871 | } |
2872 | |
2873 | vmf.virtual_address = (void __user *) start_addr; |
2874 | vmf.pte = pte; |
2875 | vmf.pgoff = pgoff; |
2876 | vmf.max_pgoff = max_pgoff; |
2877 | vmf.flags = flags; |
2878 | vma->vm_ops->map_pages(vma, &vmf); |
2879 | } |
2880 | |
2881 | static int do_read_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
2882 | unsigned long address, pmd_t *pmd, |
2883 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) |
2884 | { |
2885 | struct page *fault_page; |
2886 | spinlock_t *ptl; |
2887 | pte_t *pte; |
2888 | int ret = 0; |
2889 | |
2890 | /* |
2891 | * Let's call ->map_pages() first and use ->fault() as fallback |
2892 | * if page by the offset is not ready to be mapped (cold cache or |
2893 | * something). |
2894 | */ |
2895 | if (vma->vm_ops->map_pages && !(flags & FAULT_FLAG_NONLINEAR) && |
2896 | fault_around_bytes >> PAGE_SHIFT > 1) { |
2897 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2898 | do_fault_around(vma, address, pte, pgoff, flags); |
2899 | if (!pte_same(*pte, orig_pte)) |
2900 | goto unlock_out; |
2901 | pte_unmap_unlock(pte, ptl); |
2902 | } |
2903 | |
2904 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); |
2905 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
2906 | return ret; |
2907 | |
2908 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2909 | if (unlikely(!pte_same(*pte, orig_pte))) { |
2910 | pte_unmap_unlock(pte, ptl); |
2911 | unlock_page(fault_page); |
2912 | page_cache_release(fault_page); |
2913 | return ret; |
2914 | } |
2915 | do_set_pte(vma, address, fault_page, pte, false, false); |
2916 | unlock_page(fault_page); |
2917 | unlock_out: |
2918 | pte_unmap_unlock(pte, ptl); |
2919 | return ret; |
2920 | } |
2921 | |
2922 | static int do_cow_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
2923 | unsigned long address, pmd_t *pmd, |
2924 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) |
2925 | { |
2926 | struct page *fault_page, *new_page; |
2927 | struct mem_cgroup *memcg; |
2928 | spinlock_t *ptl; |
2929 | pte_t *pte; |
2930 | int ret; |
2931 | |
2932 | if (unlikely(anon_vma_prepare(vma))) |
2933 | return VM_FAULT_OOM; |
2934 | |
2935 | new_page = alloc_page_vma(GFP_HIGHUSER_MOVABLE, vma, address); |
2936 | if (!new_page) |
2937 | return VM_FAULT_OOM; |
2938 | |
2939 | if (mem_cgroup_try_charge(new_page, mm, GFP_KERNEL, &memcg)) { |
2940 | page_cache_release(new_page); |
2941 | return VM_FAULT_OOM; |
2942 | } |
2943 | |
2944 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); |
2945 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
2946 | goto uncharge_out; |
2947 | |
2948 | copy_user_highpage(new_page, fault_page, address, vma); |
2949 | __SetPageUptodate(new_page); |
2950 | |
2951 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
2952 | if (unlikely(!pte_same(*pte, orig_pte))) { |
2953 | pte_unmap_unlock(pte, ptl); |
2954 | unlock_page(fault_page); |
2955 | page_cache_release(fault_page); |
2956 | goto uncharge_out; |
2957 | } |
2958 | do_set_pte(vma, address, new_page, pte, true, true); |
2959 | mem_cgroup_commit_charge(new_page, memcg, false); |
2960 | lru_cache_add_active_or_unevictable(new_page, vma); |
2961 | pte_unmap_unlock(pte, ptl); |
2962 | unlock_page(fault_page); |
2963 | page_cache_release(fault_page); |
2964 | return ret; |
2965 | uncharge_out: |
2966 | mem_cgroup_cancel_charge(new_page, memcg); |
2967 | page_cache_release(new_page); |
2968 | return ret; |
2969 | } |
2970 | |
2971 | static int do_shared_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
2972 | unsigned long address, pmd_t *pmd, |
2973 | pgoff_t pgoff, unsigned int flags, pte_t orig_pte) |
2974 | { |
2975 | struct page *fault_page; |
2976 | struct address_space *mapping; |
2977 | spinlock_t *ptl; |
2978 | pte_t *pte; |
2979 | int dirtied = 0; |
2980 | int ret, tmp; |
2981 | |
2982 | ret = __do_fault(vma, address, pgoff, flags, &fault_page); |
2983 | if (unlikely(ret & (VM_FAULT_ERROR | VM_FAULT_NOPAGE | VM_FAULT_RETRY))) |
2984 | return ret; |
2985 | |
2986 | /* |
2987 | * Check if the backing address space wants to know that the page is |
2988 | * about to become writable |
2989 | */ |
2990 | if (vma->vm_ops->page_mkwrite) { |
2991 | unlock_page(fault_page); |
2992 | tmp = do_page_mkwrite(vma, fault_page, address); |
2993 | if (unlikely(!tmp || |
2994 | (tmp & (VM_FAULT_ERROR | VM_FAULT_NOPAGE)))) { |
2995 | page_cache_release(fault_page); |
2996 | return tmp; |
2997 | } |
2998 | } |
2999 | |
3000 | pte = pte_offset_map_lock(mm, pmd, address, &ptl); |
3001 | if (unlikely(!pte_same(*pte, orig_pte))) { |
3002 | pte_unmap_unlock(pte, ptl); |
3003 | unlock_page(fault_page); |
3004 | page_cache_release(fault_page); |
3005 | return ret; |
3006 | } |
3007 | do_set_pte(vma, address, fault_page, pte, true, false); |
3008 | pte_unmap_unlock(pte, ptl); |
3009 | |
3010 | if (set_page_dirty(fault_page)) |
3011 | dirtied = 1; |
3012 | mapping = fault_page->mapping; |
3013 | unlock_page(fault_page); |
3014 | if ((dirtied || vma->vm_ops->page_mkwrite) && mapping) { |
3015 | /* |
3016 | * Some device drivers do not set page.mapping but still |
3017 | * dirty their pages |
3018 | */ |
3019 | balance_dirty_pages_ratelimited(mapping); |
3020 | } |
3021 | |
3022 | /* file_update_time outside page_lock */ |
3023 | if (vma->vm_file && !vma->vm_ops->page_mkwrite) |
3024 | file_update_time(vma->vm_file); |
3025 | |
3026 | return ret; |
3027 | } |
3028 | |
3029 | /* |
3030 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
3031 | * but allow concurrent faults). |
3032 | * The mmap_sem may have been released depending on flags and our |
3033 | * return value. See filemap_fault() and __lock_page_or_retry(). |
3034 | */ |
3035 | static int do_linear_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
3036 | unsigned long address, pte_t *page_table, pmd_t *pmd, |
3037 | unsigned int flags, pte_t orig_pte) |
3038 | { |
3039 | pgoff_t pgoff = (((address & PAGE_MASK) |
3040 | - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff; |
3041 | |
3042 | pte_unmap(page_table); |
3043 | if (!(flags & FAULT_FLAG_WRITE)) |
3044 | return do_read_fault(mm, vma, address, pmd, pgoff, flags, |
3045 | orig_pte); |
3046 | if (!(vma->vm_flags & VM_SHARED)) |
3047 | return do_cow_fault(mm, vma, address, pmd, pgoff, flags, |
3048 | orig_pte); |
3049 | return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); |
3050 | } |
3051 | |
3052 | /* |
3053 | * Fault of a previously existing named mapping. Repopulate the pte |
3054 | * from the encoded file_pte if possible. This enables swappable |
3055 | * nonlinear vmas. |
3056 | * |
3057 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
3058 | * but allow concurrent faults), and pte mapped but not yet locked. |
3059 | * We return with pte unmapped and unlocked. |
3060 | * The mmap_sem may have been released depending on flags and our |
3061 | * return value. See filemap_fault() and __lock_page_or_retry(). |
3062 | */ |
3063 | static int do_nonlinear_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
3064 | unsigned long address, pte_t *page_table, pmd_t *pmd, |
3065 | unsigned int flags, pte_t orig_pte) |
3066 | { |
3067 | pgoff_t pgoff; |
3068 | |
3069 | flags |= FAULT_FLAG_NONLINEAR; |
3070 | |
3071 | if (!pte_unmap_same(mm, pmd, page_table, orig_pte)) |
3072 | return 0; |
3073 | |
3074 | if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) { |
3075 | /* |
3076 | * Page table corrupted: show pte and kill process. |
3077 | */ |
3078 | print_bad_pte(vma, address, orig_pte, NULL); |
3079 | return VM_FAULT_SIGBUS; |
3080 | } |
3081 | |
3082 | pgoff = pte_to_pgoff(orig_pte); |
3083 | if (!(flags & FAULT_FLAG_WRITE)) |
3084 | return do_read_fault(mm, vma, address, pmd, pgoff, flags, |
3085 | orig_pte); |
3086 | if (!(vma->vm_flags & VM_SHARED)) |
3087 | return do_cow_fault(mm, vma, address, pmd, pgoff, flags, |
3088 | orig_pte); |
3089 | return do_shared_fault(mm, vma, address, pmd, pgoff, flags, orig_pte); |
3090 | } |
3091 | |
3092 | static int numa_migrate_prep(struct page *page, struct vm_area_struct *vma, |
3093 | unsigned long addr, int page_nid, |
3094 | int *flags) |
3095 | { |
3096 | get_page(page); |
3097 | |
3098 | count_vm_numa_event(NUMA_HINT_FAULTS); |
3099 | if (page_nid == numa_node_id()) { |
3100 | count_vm_numa_event(NUMA_HINT_FAULTS_LOCAL); |
3101 | *flags |= TNF_FAULT_LOCAL; |
3102 | } |
3103 | |
3104 | return mpol_misplaced(page, vma, addr); |
3105 | } |
3106 | |
3107 | static int do_numa_page(struct mm_struct *mm, struct vm_area_struct *vma, |
3108 | unsigned long addr, pte_t pte, pte_t *ptep, pmd_t *pmd) |
3109 | { |
3110 | struct page *page = NULL; |
3111 | spinlock_t *ptl; |
3112 | int page_nid = -1; |
3113 | int last_cpupid; |
3114 | int target_nid; |
3115 | bool migrated = false; |
3116 | int flags = 0; |
3117 | |
3118 | /* |
3119 | * The "pte" at this point cannot be used safely without |
3120 | * validation through pte_unmap_same(). It's of NUMA type but |
3121 | * the pfn may be screwed if the read is non atomic. |
3122 | * |
3123 | * ptep_modify_prot_start is not called as this is clearing |
3124 | * the _PAGE_NUMA bit and it is not really expected that there |
3125 | * would be concurrent hardware modifications to the PTE. |
3126 | */ |
3127 | ptl = pte_lockptr(mm, pmd); |
3128 | spin_lock(ptl); |
3129 | if (unlikely(!pte_same(*ptep, pte))) { |
3130 | pte_unmap_unlock(ptep, ptl); |
3131 | goto out; |
3132 | } |
3133 | |
3134 | pte = pte_mknonnuma(pte); |
3135 | set_pte_at(mm, addr, ptep, pte); |
3136 | update_mmu_cache(vma, addr, ptep); |
3137 | |
3138 | page = vm_normal_page(vma, addr, pte); |
3139 | if (!page) { |
3140 | pte_unmap_unlock(ptep, ptl); |
3141 | return 0; |
3142 | } |
3143 | BUG_ON(is_zero_pfn(page_to_pfn(page))); |
3144 | |
3145 | /* |
3146 | * Avoid grouping on DSO/COW pages in specific and RO pages |
3147 | * in general, RO pages shouldn't hurt as much anyway since |
3148 | * they can be in shared cache state. |
3149 | */ |
3150 | if (!pte_write(pte)) |
3151 | flags |= TNF_NO_GROUP; |
3152 | |
3153 | /* |
3154 | * Flag if the page is shared between multiple address spaces. This |
3155 | * is later used when determining whether to group tasks together |
3156 | */ |
3157 | if (page_mapcount(page) > 1 && (vma->vm_flags & VM_SHARED)) |
3158 | flags |= TNF_SHARED; |
3159 | |
3160 | last_cpupid = page_cpupid_last(page); |
3161 | page_nid = page_to_nid(page); |
3162 | target_nid = numa_migrate_prep(page, vma, addr, page_nid, &flags); |
3163 | pte_unmap_unlock(ptep, ptl); |
3164 | if (target_nid == -1) { |
3165 | put_page(page); |
3166 | goto out; |
3167 | } |
3168 | |
3169 | /* Migrate to the requested node */ |
3170 | migrated = migrate_misplaced_page(page, vma, target_nid); |
3171 | if (migrated) { |
3172 | page_nid = target_nid; |
3173 | flags |= TNF_MIGRATED; |
3174 | } |
3175 | |
3176 | out: |
3177 | if (page_nid != -1) |
3178 | task_numa_fault(last_cpupid, page_nid, 1, flags); |
3179 | return 0; |
3180 | } |
3181 | |
3182 | /* |
3183 | * These routines also need to handle stuff like marking pages dirty |
3184 | * and/or accessed for architectures that don't do it in hardware (most |
3185 | * RISC architectures). The early dirtying is also good on the i386. |
3186 | * |
3187 | * There is also a hook called "update_mmu_cache()" that architectures |
3188 | * with external mmu caches can use to update those (ie the Sparc or |
3189 | * PowerPC hashed page tables that act as extended TLBs). |
3190 | * |
3191 | * We enter with non-exclusive mmap_sem (to exclude vma changes, |
3192 | * but allow concurrent faults), and pte mapped but not yet locked. |
3193 | * We return with pte unmapped and unlocked. |
3194 | * |
3195 | * The mmap_sem may have been released depending on flags and our |
3196 | * return value. See filemap_fault() and __lock_page_or_retry(). |
3197 | */ |
3198 | static int handle_pte_fault(struct mm_struct *mm, |
3199 | struct vm_area_struct *vma, unsigned long address, |
3200 | pte_t *pte, pmd_t *pmd, unsigned int flags) |
3201 | { |
3202 | pte_t entry; |
3203 | spinlock_t *ptl; |
3204 | |
3205 | entry = ACCESS_ONCE(*pte); |
3206 | if (!pte_present(entry)) { |
3207 | if (pte_none(entry)) { |
3208 | if (vma->vm_ops) { |
3209 | if (likely(vma->vm_ops->fault)) |
3210 | return do_linear_fault(mm, vma, address, |
3211 | pte, pmd, flags, entry); |
3212 | } |
3213 | return do_anonymous_page(mm, vma, address, |
3214 | pte, pmd, flags); |
3215 | } |
3216 | if (pte_file(entry)) |
3217 | return do_nonlinear_fault(mm, vma, address, |
3218 | pte, pmd, flags, entry); |
3219 | return do_swap_page(mm, vma, address, |
3220 | pte, pmd, flags, entry); |
3221 | } |
3222 | |
3223 | if (pte_numa(entry)) |
3224 | return do_numa_page(mm, vma, address, entry, pte, pmd); |
3225 | |
3226 | ptl = pte_lockptr(mm, pmd); |
3227 | spin_lock(ptl); |
3228 | if (unlikely(!pte_same(*pte, entry))) |
3229 | goto unlock; |
3230 | if (flags & FAULT_FLAG_WRITE) { |
3231 | if (!pte_write(entry)) |
3232 | return do_wp_page(mm, vma, address, |
3233 | pte, pmd, ptl, entry); |
3234 | entry = pte_mkdirty(entry); |
3235 | } |
3236 | entry = pte_mkyoung(entry); |
3237 | if (ptep_set_access_flags(vma, address, pte, entry, flags & FAULT_FLAG_WRITE)) { |
3238 | update_mmu_cache(vma, address, pte); |
3239 | } else { |
3240 | /* |
3241 | * This is needed only for protection faults but the arch code |
3242 | * is not yet telling us if this is a protection fault or not. |
3243 | * This still avoids useless tlb flushes for .text page faults |
3244 | * with threads. |
3245 | */ |
3246 | if (flags & FAULT_FLAG_WRITE) |
3247 | flush_tlb_fix_spurious_fault(vma, address); |
3248 | } |
3249 | unlock: |
3250 | pte_unmap_unlock(pte, ptl); |
3251 | return 0; |
3252 | } |
3253 | |
3254 | /* |
3255 | * By the time we get here, we already hold the mm semaphore |
3256 | * |
3257 | * The mmap_sem may have been released depending on flags and our |
3258 | * return value. See filemap_fault() and __lock_page_or_retry(). |
3259 | */ |
3260 | static int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
3261 | unsigned long address, unsigned int flags) |
3262 | { |
3263 | pgd_t *pgd; |
3264 | pud_t *pud; |
3265 | pmd_t *pmd; |
3266 | pte_t *pte; |
3267 | |
3268 | if (unlikely(is_vm_hugetlb_page(vma))) |
3269 | return hugetlb_fault(mm, vma, address, flags); |
3270 | |
3271 | pgd = pgd_offset(mm, address); |
3272 | pud = pud_alloc(mm, pgd, address); |
3273 | if (!pud) |
3274 | return VM_FAULT_OOM; |
3275 | pmd = pmd_alloc(mm, pud, address); |
3276 | if (!pmd) |
3277 | return VM_FAULT_OOM; |
3278 | if (pmd_none(*pmd) && transparent_hugepage_enabled(vma)) { |
3279 | int ret = VM_FAULT_FALLBACK; |
3280 | if (!vma->vm_ops) |
3281 | ret = do_huge_pmd_anonymous_page(mm, vma, address, |
3282 | pmd, flags); |
3283 | if (!(ret & VM_FAULT_FALLBACK)) |
3284 | return ret; |
3285 | } else { |
3286 | pmd_t orig_pmd = *pmd; |
3287 | int ret; |
3288 | |
3289 | barrier(); |
3290 | if (pmd_trans_huge(orig_pmd)) { |
3291 | unsigned int dirty = flags & FAULT_FLAG_WRITE; |
3292 | |
3293 | /* |
3294 | * If the pmd is splitting, return and retry the |
3295 | * the fault. Alternative: wait until the split |
3296 | * is done, and goto retry. |
3297 | */ |
3298 | if (pmd_trans_splitting(orig_pmd)) |
3299 | return 0; |
3300 | |
3301 | if (pmd_numa(orig_pmd)) |
3302 | return do_huge_pmd_numa_page(mm, vma, address, |
3303 | orig_pmd, pmd); |
3304 | |
3305 | if (dirty && !pmd_write(orig_pmd)) { |
3306 | ret = do_huge_pmd_wp_page(mm, vma, address, pmd, |
3307 | orig_pmd); |
3308 | if (!(ret & VM_FAULT_FALLBACK)) |
3309 | return ret; |
3310 | } else { |
3311 | huge_pmd_set_accessed(mm, vma, address, pmd, |
3312 | orig_pmd, dirty); |
3313 | return 0; |
3314 | } |
3315 | } |
3316 | } |
3317 | |
3318 | /* |
3319 | * Use __pte_alloc instead of pte_alloc_map, because we can't |
3320 | * run pte_offset_map on the pmd, if an huge pmd could |
3321 | * materialize from under us from a different thread. |
3322 | */ |
3323 | if (unlikely(pmd_none(*pmd)) && |
3324 | unlikely(__pte_alloc(mm, vma, pmd, address))) |
3325 | return VM_FAULT_OOM; |
3326 | /* if an huge pmd materialized from under us just retry later */ |
3327 | if (unlikely(pmd_trans_huge(*pmd))) |
3328 | return 0; |
3329 | /* |
3330 | * A regular pmd is established and it can't morph into a huge pmd |
3331 | * from under us anymore at this point because we hold the mmap_sem |
3332 | * read mode and khugepaged takes it in write mode. So now it's |
3333 | * safe to run pte_offset_map(). |
3334 | */ |
3335 | pte = pte_offset_map(pmd, address); |
3336 | |
3337 | return handle_pte_fault(mm, vma, address, pte, pmd, flags); |
3338 | } |
3339 | |
3340 | /* |
3341 | * By the time we get here, we already hold the mm semaphore |
3342 | * |
3343 | * The mmap_sem may have been released depending on flags and our |
3344 | * return value. See filemap_fault() and __lock_page_or_retry(). |
3345 | */ |
3346 | int handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma, |
3347 | unsigned long address, unsigned int flags) |
3348 | { |
3349 | int ret; |
3350 | |
3351 | __set_current_state(TASK_RUNNING); |
3352 | |
3353 | count_vm_event(PGFAULT); |
3354 | mem_cgroup_count_vm_event(mm, PGFAULT); |
3355 | |
3356 | /* do counter updates before entering really critical section. */ |
3357 | check_sync_rss_stat(current); |
3358 | |
3359 | /* |
3360 | * Enable the memcg OOM handling for faults triggered in user |
3361 | * space. Kernel faults are handled more gracefully. |
3362 | */ |
3363 | if (flags & FAULT_FLAG_USER) |
3364 | mem_cgroup_oom_enable(); |
3365 | |
3366 | ret = __handle_mm_fault(mm, vma, address, flags); |
3367 | |
3368 | if (flags & FAULT_FLAG_USER) { |
3369 | mem_cgroup_oom_disable(); |
3370 | /* |
3371 | * The task may have entered a memcg OOM situation but |
3372 | * if the allocation error was handled gracefully (no |
3373 | * VM_FAULT_OOM), there is no need to kill anything. |
3374 | * Just clean up the OOM state peacefully. |
3375 | */ |
3376 | if (task_in_memcg_oom(current) && !(ret & VM_FAULT_OOM)) |
3377 | mem_cgroup_oom_synchronize(false); |
3378 | } |
3379 | |
3380 | return ret; |
3381 | } |
3382 | |
3383 | #ifndef __PAGETABLE_PUD_FOLDED |
3384 | /* |
3385 | * Allocate page upper directory. |
3386 | * We've already handled the fast-path in-line. |
3387 | */ |
3388 | int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address) |
3389 | { |
3390 | pud_t *new = pud_alloc_one(mm, address); |
3391 | if (!new) |
3392 | return -ENOMEM; |
3393 | |
3394 | smp_wmb(); /* See comment in __pte_alloc */ |
3395 | |
3396 | spin_lock(&mm->page_table_lock); |
3397 | if (pgd_present(*pgd)) /* Another has populated it */ |
3398 | pud_free(mm, new); |
3399 | else |
3400 | pgd_populate(mm, pgd, new); |
3401 | spin_unlock(&mm->page_table_lock); |
3402 | return 0; |
3403 | } |
3404 | #endif /* __PAGETABLE_PUD_FOLDED */ |
3405 | |
3406 | #ifndef __PAGETABLE_PMD_FOLDED |
3407 | /* |
3408 | * Allocate page middle directory. |
3409 | * We've already handled the fast-path in-line. |
3410 | */ |
3411 | int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address) |
3412 | { |
3413 | pmd_t *new = pmd_alloc_one(mm, address); |
3414 | if (!new) |
3415 | return -ENOMEM; |
3416 | |
3417 | smp_wmb(); /* See comment in __pte_alloc */ |
3418 | |
3419 | spin_lock(&mm->page_table_lock); |
3420 | #ifndef __ARCH_HAS_4LEVEL_HACK |
3421 | if (pud_present(*pud)) /* Another has populated it */ |
3422 | pmd_free(mm, new); |
3423 | else |
3424 | pud_populate(mm, pud, new); |
3425 | #else |
3426 | if (pgd_present(*pud)) /* Another has populated it */ |
3427 | pmd_free(mm, new); |
3428 | else |
3429 | pgd_populate(mm, pud, new); |
3430 | #endif /* __ARCH_HAS_4LEVEL_HACK */ |
3431 | spin_unlock(&mm->page_table_lock); |
3432 | return 0; |
3433 | } |
3434 | #endif /* __PAGETABLE_PMD_FOLDED */ |
3435 | |
3436 | static int __follow_pte(struct mm_struct *mm, unsigned long address, |
3437 | pte_t **ptepp, spinlock_t **ptlp) |
3438 | { |
3439 | pgd_t *pgd; |
3440 | pud_t *pud; |
3441 | pmd_t *pmd; |
3442 | pte_t *ptep; |
3443 | |
3444 | pgd = pgd_offset(mm, address); |
3445 | if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd))) |
3446 | goto out; |
3447 | |
3448 | pud = pud_offset(pgd, address); |
3449 | if (pud_none(*pud) || unlikely(pud_bad(*pud))) |
3450 | goto out; |
3451 | |
3452 | pmd = pmd_offset(pud, address); |
3453 | VM_BUG_ON(pmd_trans_huge(*pmd)); |
3454 | if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd))) |
3455 | goto out; |
3456 | |
3457 | /* We cannot handle huge page PFN maps. Luckily they don't exist. */ |
3458 | if (pmd_huge(*pmd)) |
3459 | goto out; |
3460 | |
3461 | ptep = pte_offset_map_lock(mm, pmd, address, ptlp); |
3462 | if (!ptep) |
3463 | goto out; |
3464 | if (!pte_present(*ptep)) |
3465 | goto unlock; |
3466 | *ptepp = ptep; |
3467 | return 0; |
3468 | unlock: |
3469 | pte_unmap_unlock(ptep, *ptlp); |
3470 | out: |
3471 | return -EINVAL; |
3472 | } |
3473 | |
3474 | static inline int follow_pte(struct mm_struct *mm, unsigned long address, |
3475 | pte_t **ptepp, spinlock_t **ptlp) |
3476 | { |
3477 | int res; |
3478 | |
3479 | /* (void) is needed to make gcc happy */ |
3480 | (void) __cond_lock(*ptlp, |
3481 | !(res = __follow_pte(mm, address, ptepp, ptlp))); |
3482 | return res; |
3483 | } |
3484 | |
3485 | /** |
3486 | * follow_pfn - look up PFN at a user virtual address |
3487 | * @vma: memory mapping |
3488 | * @address: user virtual address |
3489 | * @pfn: location to store found PFN |
3490 | * |
3491 | * Only IO mappings and raw PFN mappings are allowed. |
3492 | * |
3493 | * Returns zero and the pfn at @pfn on success, -ve otherwise. |
3494 | */ |
3495 | int follow_pfn(struct vm_area_struct *vma, unsigned long address, |
3496 | unsigned long *pfn) |
3497 | { |
3498 | int ret = -EINVAL; |
3499 | spinlock_t *ptl; |
3500 | pte_t *ptep; |
3501 | |
3502 | if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) |
3503 | return ret; |
3504 | |
3505 | ret = follow_pte(vma->vm_mm, address, &ptep, &ptl); |
3506 | if (ret) |
3507 | return ret; |
3508 | *pfn = pte_pfn(*ptep); |
3509 | pte_unmap_unlock(ptep, ptl); |
3510 | return 0; |
3511 | } |
3512 | EXPORT_SYMBOL(follow_pfn); |
3513 | |
3514 | #ifdef CONFIG_HAVE_IOREMAP_PROT |
3515 | int follow_phys(struct vm_area_struct *vma, |
3516 | unsigned long address, unsigned int flags, |
3517 | unsigned long *prot, resource_size_t *phys) |
3518 | { |
3519 | int ret = -EINVAL; |
3520 | pte_t *ptep, pte; |
3521 | spinlock_t *ptl; |
3522 | |
3523 | if (!(vma->vm_flags & (VM_IO | VM_PFNMAP))) |
3524 | goto out; |
3525 | |
3526 | if (follow_pte(vma->vm_mm, address, &ptep, &ptl)) |
3527 | goto out; |
3528 | pte = *ptep; |
3529 | |
3530 | if ((flags & FOLL_WRITE) && !pte_write(pte)) |
3531 | goto unlock; |
3532 | |
3533 | *prot = pgprot_val(pte_pgprot(pte)); |
3534 | *phys = (resource_size_t)pte_pfn(pte) << PAGE_SHIFT; |
3535 | |
3536 | ret = 0; |
3537 | unlock: |
3538 | pte_unmap_unlock(ptep, ptl); |
3539 | out: |
3540 | return ret; |
3541 | } |
3542 | |
3543 | int generic_access_phys(struct vm_area_struct *vma, unsigned long addr, |
3544 | void *buf, int len, int write) |
3545 | { |
3546 | resource_size_t phys_addr; |
3547 | unsigned long prot = 0; |
3548 | void __iomem *maddr; |
3549 | int offset = addr & (PAGE_SIZE-1); |
3550 | |
3551 | if (follow_phys(vma, addr, write, &prot, &phys_addr)) |
3552 | return -EINVAL; |
3553 | |
3554 | maddr = ioremap_prot(phys_addr, PAGE_SIZE, prot); |
3555 | if (write) |
3556 | memcpy_toio(maddr + offset, buf, len); |
3557 | else |
3558 | memcpy_fromio(buf, maddr + offset, len); |
3559 | iounmap(maddr); |
3560 | |
3561 | return len; |
3562 | } |
3563 | EXPORT_SYMBOL_GPL(generic_access_phys); |
3564 | #endif |
3565 | |
3566 | /* |
3567 | * Access another process' address space as given in mm. If non-NULL, use the |
3568 | * given task for page fault accounting. |
3569 | */ |
3570 | static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm, |
3571 | unsigned long addr, void *buf, int len, int write) |
3572 | { |
3573 | struct vm_area_struct *vma; |
3574 | void *old_buf = buf; |
3575 | |
3576 | down_read(&mm->mmap_sem); |
3577 | /* ignore errors, just check how much was successfully transferred */ |
3578 | while (len) { |
3579 | int bytes, ret, offset; |
3580 | void *maddr; |
3581 | struct page *page = NULL; |
3582 | |
3583 | ret = get_user_pages(tsk, mm, addr, 1, |
3584 | write, 1, &page, &vma); |
3585 | if (ret <= 0) { |
3586 | #ifndef CONFIG_HAVE_IOREMAP_PROT |
3587 | break; |
3588 | #else |
3589 | /* |
3590 | * Check if this is a VM_IO | VM_PFNMAP VMA, which |
3591 | * we can access using slightly different code. |
3592 | */ |
3593 | vma = find_vma(mm, addr); |
3594 | if (!vma || vma->vm_start > addr) |
3595 | break; |
3596 | if (vma->vm_ops && vma->vm_ops->access) |
3597 | ret = vma->vm_ops->access(vma, addr, buf, |
3598 | len, write); |
3599 | if (ret <= 0) |
3600 | break; |
3601 | bytes = ret; |
3602 | #endif |
3603 | } else { |
3604 | bytes = len; |
3605 | offset = addr & (PAGE_SIZE-1); |
3606 | if (bytes > PAGE_SIZE-offset) |
3607 | bytes = PAGE_SIZE-offset; |
3608 | |
3609 | maddr = kmap(page); |
3610 | if (write) { |
3611 | copy_to_user_page(vma, page, addr, |
3612 | maddr + offset, buf, bytes); |
3613 | set_page_dirty_lock(page); |
3614 | } else { |
3615 | copy_from_user_page(vma, page, addr, |
3616 | buf, maddr + offset, bytes); |
3617 | } |
3618 | kunmap(page); |
3619 | page_cache_release(page); |
3620 | } |
3621 | len -= bytes; |
3622 | buf += bytes; |
3623 | addr += bytes; |
3624 | } |
3625 | up_read(&mm->mmap_sem); |
3626 | |
3627 | return buf - old_buf; |
3628 | } |
3629 | |
3630 | /** |
3631 | * access_remote_vm - access another process' address space |
3632 | * @mm: the mm_struct of the target address space |
3633 | * @addr: start address to access |
3634 | * @buf: source or destination buffer |
3635 | * @len: number of bytes to transfer |
3636 | * @write: whether the access is a write |
3637 | * |
3638 | * The caller must hold a reference on @mm. |
3639 | */ |
3640 | int access_remote_vm(struct mm_struct *mm, unsigned long addr, |
3641 | void *buf, int len, int write) |
3642 | { |
3643 | return __access_remote_vm(NULL, mm, addr, buf, len, write); |
3644 | } |
3645 | |
3646 | /* |
3647 | * Access another process' address space. |
3648 | * Source/target buffer must be kernel space, |
3649 | * Do not walk the page table directly, use get_user_pages |
3650 | */ |
3651 | int access_process_vm(struct task_struct *tsk, unsigned long addr, |
3652 | void *buf, int len, int write) |
3653 | { |
3654 | struct mm_struct *mm; |
3655 | int ret; |
3656 | |
3657 | mm = get_task_mm(tsk); |
3658 | if (!mm) |
3659 | return 0; |
3660 | |
3661 | ret = __access_remote_vm(tsk, mm, addr, buf, len, write); |
3662 | mmput(mm); |
3663 | |
3664 | return ret; |
3665 | } |
3666 | |
3667 | /* |
3668 | * Print the name of a VMA. |
3669 | */ |
3670 | void print_vma_addr(char *prefix, unsigned long ip) |
3671 | { |
3672 | struct mm_struct *mm = current->mm; |
3673 | struct vm_area_struct *vma; |
3674 | |
3675 | /* |
3676 | * Do not print if we are in atomic |
3677 | * contexts (in exception stacks, etc.): |
3678 | */ |
3679 | if (preempt_count()) |
3680 | return; |
3681 | |
3682 | down_read(&mm->mmap_sem); |
3683 | vma = find_vma(mm, ip); |
3684 | if (vma && vma->vm_file) { |
3685 | struct file *f = vma->vm_file; |
3686 | char *buf = (char *)__get_free_page(GFP_KERNEL); |
3687 | if (buf) { |
3688 | char *p; |
3689 | |
3690 | p = d_path(&f->f_path, buf, PAGE_SIZE); |
3691 | if (IS_ERR(p)) |
3692 | p = "?"; |
3693 | printk("%s%s[%lx+%lx]", prefix, kbasename(p), |
3694 | vma->vm_start, |
3695 | vma->vm_end - vma->vm_start); |
3696 | free_page((unsigned long)buf); |
3697 | } |
3698 | } |
3699 | up_read(&mm->mmap_sem); |
3700 | } |
3701 | |
3702 | #if defined(CONFIG_PROVE_LOCKING) || defined(CONFIG_DEBUG_ATOMIC_SLEEP) |
3703 | void might_fault(void) |
3704 | { |
3705 | /* |
3706 | * Some code (nfs/sunrpc) uses socket ops on kernel memory while |
3707 | * holding the mmap_sem, this is safe because kernel memory doesn't |
3708 | * get paged out, therefore we'll never actually fault, and the |
3709 | * below annotations will generate false positives. |
3710 | */ |
3711 | if (segment_eq(get_fs(), KERNEL_DS)) |
3712 | return; |
3713 | |
3714 | /* |
3715 | * it would be nicer only to annotate paths which are not under |
3716 | * pagefault_disable, however that requires a larger audit and |
3717 | * providing helpers like get_user_atomic. |
3718 | */ |
3719 | if (in_atomic()) |
3720 | return; |
3721 | |
3722 | __might_sleep(__FILE__, __LINE__, 0); |
3723 | |
3724 | if (current->mm) |
3725 | might_lock_read(¤t->mm->mmap_sem); |
3726 | } |
3727 | EXPORT_SYMBOL(might_fault); |
3728 | #endif |
3729 | |
3730 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_HUGETLBFS) |
3731 | static void clear_gigantic_page(struct page *page, |
3732 | unsigned long addr, |
3733 | unsigned int pages_per_huge_page) |
3734 | { |
3735 | int i; |
3736 | struct page *p = page; |
3737 | |
3738 | might_sleep(); |
3739 | for (i = 0; i < pages_per_huge_page; |
3740 | i++, p = mem_map_next(p, page, i)) { |
3741 | cond_resched(); |
3742 | clear_user_highpage(p, addr + i * PAGE_SIZE); |
3743 | } |
3744 | } |
3745 | void clear_huge_page(struct page *page, |
3746 | unsigned long addr, unsigned int pages_per_huge_page) |
3747 | { |
3748 | int i; |
3749 | |
3750 | if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { |
3751 | clear_gigantic_page(page, addr, pages_per_huge_page); |
3752 | return; |
3753 | } |
3754 | |
3755 | might_sleep(); |
3756 | for (i = 0; i < pages_per_huge_page; i++) { |
3757 | cond_resched(); |
3758 | clear_user_highpage(page + i, addr + i * PAGE_SIZE); |
3759 | } |
3760 | } |
3761 | |
3762 | static void copy_user_gigantic_page(struct page *dst, struct page *src, |
3763 | unsigned long addr, |
3764 | struct vm_area_struct *vma, |
3765 | unsigned int pages_per_huge_page) |
3766 | { |
3767 | int i; |
3768 | struct page *dst_base = dst; |
3769 | struct page *src_base = src; |
3770 | |
3771 | for (i = 0; i < pages_per_huge_page; ) { |
3772 | cond_resched(); |
3773 | copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma); |
3774 | |
3775 | i++; |
3776 | dst = mem_map_next(dst, dst_base, i); |
3777 | src = mem_map_next(src, src_base, i); |
3778 | } |
3779 | } |
3780 | |
3781 | void copy_user_huge_page(struct page *dst, struct page *src, |
3782 | unsigned long addr, struct vm_area_struct *vma, |
3783 | unsigned int pages_per_huge_page) |
3784 | { |
3785 | int i; |
3786 | |
3787 | if (unlikely(pages_per_huge_page > MAX_ORDER_NR_PAGES)) { |
3788 | copy_user_gigantic_page(dst, src, addr, vma, |
3789 | pages_per_huge_page); |
3790 | return; |
3791 | } |
3792 | |
3793 | might_sleep(); |
3794 | for (i = 0; i < pages_per_huge_page; i++) { |
3795 | cond_resched(); |
3796 | copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma); |
3797 | } |
3798 | } |
3799 | #endif /* CONFIG_TRANSPARENT_HUGEPAGE || CONFIG_HUGETLBFS */ |
3800 | |
3801 | #if USE_SPLIT_PTE_PTLOCKS && ALLOC_SPLIT_PTLOCKS |
3802 | |
3803 | static struct kmem_cache *page_ptl_cachep; |
3804 | |
3805 | void __init ptlock_cache_init(void) |
3806 | { |
3807 | page_ptl_cachep = kmem_cache_create("page->ptl", sizeof(spinlock_t), 0, |
3808 | SLAB_PANIC, NULL); |
3809 | } |
3810 | |
3811 | bool ptlock_alloc(struct page *page) |
3812 | { |
3813 | spinlock_t *ptl; |
3814 | |
3815 | ptl = kmem_cache_alloc(page_ptl_cachep, GFP_KERNEL); |
3816 | if (!ptl) |
3817 | return false; |
3818 | page->ptl = ptl; |
3819 | return true; |
3820 | } |
3821 | |
3822 | void ptlock_free(struct page *page) |
3823 | { |
3824 | kmem_cache_free(page_ptl_cachep, page->ptl); |
3825 | } |
3826 | #endif |
3827 |
Branches:
ben-wpan
ben-wpan-stefan
javiroman/ks7010
jz-2.6.34
jz-2.6.34-rc5
jz-2.6.34-rc6
jz-2.6.34-rc7
jz-2.6.35
jz-2.6.36
jz-2.6.37
jz-2.6.38
jz-2.6.39
jz-3.0
jz-3.1
jz-3.11
jz-3.12
jz-3.13
jz-3.15
jz-3.16
jz-3.18-dt
jz-3.2
jz-3.3
jz-3.4
jz-3.5
jz-3.6
jz-3.6-rc2-pwm
jz-3.9
jz-3.9-clk
jz-3.9-rc8
jz47xx
jz47xx-2.6.38
master
Tags:
od-2011-09-04
od-2011-09-18
v2.6.34-rc5
v2.6.34-rc6
v2.6.34-rc7
v3.9