]> bbs.cooldavid.org Git - net-next-2.6.git/blame - mm/memory.c
[PATCH] unpaged: COW on VM_UNPAGED
[net-next-2.6.git] / mm / memory.c
CommitLineData
1da177e4
LT
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/rmap.h>
49#include <linux/module.h>
50#include <linux/init.h>
51
52#include <asm/pgalloc.h>
53#include <asm/uaccess.h>
54#include <asm/tlb.h>
55#include <asm/tlbflush.h>
56#include <asm/pgtable.h>
57
58#include <linux/swapops.h>
59#include <linux/elf.h>
60
d41dee36 61#ifndef CONFIG_NEED_MULTIPLE_NODES
1da177e4
LT
62/* use the per-pgdat data instead for discontigmem - mbligh */
63unsigned long max_mapnr;
64struct page *mem_map;
65
66EXPORT_SYMBOL(max_mapnr);
67EXPORT_SYMBOL(mem_map);
68#endif
69
70unsigned long num_physpages;
71/*
72 * A number of key systems in x86 including ioremap() rely on the assumption
73 * that high_memory defines the upper bound on direct map memory, then end
74 * of ZONE_NORMAL. Under CONFIG_DISCONTIG this means that max_low_pfn and
75 * highstart_pfn must be the same; there must be no gap between ZONE_NORMAL
76 * and ZONE_HIGHMEM.
77 */
78void * high_memory;
79unsigned long vmalloc_earlyreserve;
80
81EXPORT_SYMBOL(num_physpages);
82EXPORT_SYMBOL(high_memory);
83EXPORT_SYMBOL(vmalloc_earlyreserve);
84
85/*
86 * If a p?d_bad entry is found while walking page tables, report
87 * the error, before resetting entry to p?d_none. Usually (but
88 * very seldom) called out from the p?d_none_or_clear_bad macros.
89 */
90
91void pgd_clear_bad(pgd_t *pgd)
92{
93 pgd_ERROR(*pgd);
94 pgd_clear(pgd);
95}
96
97void pud_clear_bad(pud_t *pud)
98{
99 pud_ERROR(*pud);
100 pud_clear(pud);
101}
102
103void pmd_clear_bad(pmd_t *pmd)
104{
105 pmd_ERROR(*pmd);
106 pmd_clear(pmd);
107}
108
109/*
110 * Note: this doesn't free the actual pages themselves. That
111 * has been handled earlier when unmapping all the memory regions.
112 */
e0da382c 113static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
1da177e4 114{
e0da382c
HD
115 struct page *page = pmd_page(*pmd);
116 pmd_clear(pmd);
4c21e2f2 117 pte_lock_deinit(page);
e0da382c
HD
118 pte_free_tlb(tlb, page);
119 dec_page_state(nr_page_table_pages);
120 tlb->mm->nr_ptes--;
1da177e4
LT
121}
122
e0da382c
HD
123static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
124 unsigned long addr, unsigned long end,
125 unsigned long floor, unsigned long ceiling)
1da177e4
LT
126{
127 pmd_t *pmd;
128 unsigned long next;
e0da382c 129 unsigned long start;
1da177e4 130
e0da382c 131 start = addr;
1da177e4 132 pmd = pmd_offset(pud, addr);
1da177e4
LT
133 do {
134 next = pmd_addr_end(addr, end);
135 if (pmd_none_or_clear_bad(pmd))
136 continue;
e0da382c 137 free_pte_range(tlb, pmd);
1da177e4
LT
138 } while (pmd++, addr = next, addr != end);
139
e0da382c
HD
140 start &= PUD_MASK;
141 if (start < floor)
142 return;
143 if (ceiling) {
144 ceiling &= PUD_MASK;
145 if (!ceiling)
146 return;
1da177e4 147 }
e0da382c
HD
148 if (end - 1 > ceiling - 1)
149 return;
150
151 pmd = pmd_offset(pud, start);
152 pud_clear(pud);
153 pmd_free_tlb(tlb, pmd);
1da177e4
LT
154}
155
e0da382c
HD
156static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
157 unsigned long addr, unsigned long end,
158 unsigned long floor, unsigned long ceiling)
1da177e4
LT
159{
160 pud_t *pud;
161 unsigned long next;
e0da382c 162 unsigned long start;
1da177e4 163
e0da382c 164 start = addr;
1da177e4 165 pud = pud_offset(pgd, addr);
1da177e4
LT
166 do {
167 next = pud_addr_end(addr, end);
168 if (pud_none_or_clear_bad(pud))
169 continue;
e0da382c 170 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
1da177e4
LT
171 } while (pud++, addr = next, addr != end);
172
e0da382c
HD
173 start &= PGDIR_MASK;
174 if (start < floor)
175 return;
176 if (ceiling) {
177 ceiling &= PGDIR_MASK;
178 if (!ceiling)
179 return;
1da177e4 180 }
e0da382c
HD
181 if (end - 1 > ceiling - 1)
182 return;
183
184 pud = pud_offset(pgd, start);
185 pgd_clear(pgd);
186 pud_free_tlb(tlb, pud);
1da177e4
LT
187}
188
189/*
e0da382c
HD
190 * This function frees user-level page tables of a process.
191 *
1da177e4
LT
192 * Must be called with pagetable lock held.
193 */
3bf5ee95 194void free_pgd_range(struct mmu_gather **tlb,
e0da382c
HD
195 unsigned long addr, unsigned long end,
196 unsigned long floor, unsigned long ceiling)
1da177e4
LT
197{
198 pgd_t *pgd;
199 unsigned long next;
e0da382c
HD
200 unsigned long start;
201
202 /*
203 * The next few lines have given us lots of grief...
204 *
205 * Why are we testing PMD* at this top level? Because often
206 * there will be no work to do at all, and we'd prefer not to
207 * go all the way down to the bottom just to discover that.
208 *
209 * Why all these "- 1"s? Because 0 represents both the bottom
210 * of the address space and the top of it (using -1 for the
211 * top wouldn't help much: the masks would do the wrong thing).
212 * The rule is that addr 0 and floor 0 refer to the bottom of
213 * the address space, but end 0 and ceiling 0 refer to the top
214 * Comparisons need to use "end - 1" and "ceiling - 1" (though
215 * that end 0 case should be mythical).
216 *
217 * Wherever addr is brought up or ceiling brought down, we must
218 * be careful to reject "the opposite 0" before it confuses the
219 * subsequent tests. But what about where end is brought down
220 * by PMD_SIZE below? no, end can't go down to 0 there.
221 *
222 * Whereas we round start (addr) and ceiling down, by different
223 * masks at different levels, in order to test whether a table
224 * now has no other vmas using it, so can be freed, we don't
225 * bother to round floor or end up - the tests don't need that.
226 */
1da177e4 227
e0da382c
HD
228 addr &= PMD_MASK;
229 if (addr < floor) {
230 addr += PMD_SIZE;
231 if (!addr)
232 return;
233 }
234 if (ceiling) {
235 ceiling &= PMD_MASK;
236 if (!ceiling)
237 return;
238 }
239 if (end - 1 > ceiling - 1)
240 end -= PMD_SIZE;
241 if (addr > end - 1)
242 return;
243
244 start = addr;
3bf5ee95 245 pgd = pgd_offset((*tlb)->mm, addr);
1da177e4
LT
246 do {
247 next = pgd_addr_end(addr, end);
248 if (pgd_none_or_clear_bad(pgd))
249 continue;
3bf5ee95 250 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
1da177e4 251 } while (pgd++, addr = next, addr != end);
e0da382c 252
4d6ddfa9 253 if (!(*tlb)->fullmm)
3bf5ee95 254 flush_tlb_pgtables((*tlb)->mm, start, end);
e0da382c
HD
255}
256
257void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
3bf5ee95 258 unsigned long floor, unsigned long ceiling)
e0da382c
HD
259{
260 while (vma) {
261 struct vm_area_struct *next = vma->vm_next;
262 unsigned long addr = vma->vm_start;
263
8f4f8c16
HD
264 /*
265 * Hide vma from rmap and vmtruncate before freeing pgtables
266 */
267 anon_vma_unlink(vma);
268 unlink_file_vma(vma);
269
3bf5ee95
HD
270 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
e0da382c 272 floor, next? next->vm_start: ceiling);
3bf5ee95
HD
273 } else {
274 /*
275 * Optimization: gather nearby vmas into one call down
276 */
277 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
278 && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
279 HPAGE_SIZE)) {
280 vma = next;
281 next = vma->vm_next;
8f4f8c16
HD
282 anon_vma_unlink(vma);
283 unlink_file_vma(vma);
3bf5ee95
HD
284 }
285 free_pgd_range(tlb, addr, vma->vm_end,
286 floor, next? next->vm_start: ceiling);
287 }
e0da382c
HD
288 vma = next;
289 }
1da177e4
LT
290}
291
1bb3630e 292int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
1da177e4 293{
c74df32c 294 struct page *new = pte_alloc_one(mm, address);
1bb3630e
HD
295 if (!new)
296 return -ENOMEM;
297
4c21e2f2 298 pte_lock_init(new);
c74df32c 299 spin_lock(&mm->page_table_lock);
4c21e2f2
HD
300 if (pmd_present(*pmd)) { /* Another has populated it */
301 pte_lock_deinit(new);
1bb3630e 302 pte_free(new);
4c21e2f2 303 } else {
1da177e4
LT
304 mm->nr_ptes++;
305 inc_page_state(nr_page_table_pages);
306 pmd_populate(mm, pmd, new);
307 }
c74df32c 308 spin_unlock(&mm->page_table_lock);
1bb3630e 309 return 0;
1da177e4
LT
310}
311
1bb3630e 312int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
1da177e4 313{
1bb3630e
HD
314 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315 if (!new)
316 return -ENOMEM;
317
318 spin_lock(&init_mm.page_table_lock);
319 if (pmd_present(*pmd)) /* Another has populated it */
320 pte_free_kernel(new);
321 else
322 pmd_populate_kernel(&init_mm, pmd, new);
323 spin_unlock(&init_mm.page_table_lock);
324 return 0;
1da177e4
LT
325}
326
ae859762
HD
327static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
328{
329 if (file_rss)
330 add_mm_counter(mm, file_rss, file_rss);
331 if (anon_rss)
332 add_mm_counter(mm, anon_rss, anon_rss);
333}
334
b5810039
NP
335/*
336 * This function is called to print an error when a pte in a
0b14c179 337 * !VM_UNPAGED region is found pointing to an invalid pfn (which
b5810039
NP
338 * is an error.
339 *
340 * The calling function must still handle the error.
341 */
342void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
343{
344 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
345 "vm_flags = %lx, vaddr = %lx\n",
346 (long long)pte_val(pte),
347 (vma->vm_mm == current->mm ? current->comm : "???"),
348 vma->vm_flags, vaddr);
349 dump_stack();
350}
351
1da177e4
LT
352/*
353 * copy one vm_area from one task to the other. Assumes the page tables
354 * already present in the new task to be cleared in the whole range
355 * covered by this vma.
1da177e4
LT
356 */
357
8c103762 358static inline void
1da177e4 359copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
b5810039 360 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
8c103762 361 unsigned long addr, int *rss)
1da177e4 362{
b5810039 363 unsigned long vm_flags = vma->vm_flags;
1da177e4
LT
364 pte_t pte = *src_pte;
365 struct page *page;
366 unsigned long pfn;
367
368 /* pte contains position in swap or file, so copy. */
369 if (unlikely(!pte_present(pte))) {
370 if (!pte_file(pte)) {
371 swap_duplicate(pte_to_swp_entry(pte));
372 /* make sure dst_mm is on swapoff's mmlist. */
373 if (unlikely(list_empty(&dst_mm->mmlist))) {
374 spin_lock(&mmlist_lock);
f412ac08
HD
375 if (list_empty(&dst_mm->mmlist))
376 list_add(&dst_mm->mmlist,
377 &src_mm->mmlist);
1da177e4
LT
378 spin_unlock(&mmlist_lock);
379 }
380 }
ae859762 381 goto out_set_pte;
1da177e4
LT
382 }
383
0b14c179 384 /* If the region is VM_UNPAGED, the mapping is not
b5810039
NP
385 * mapped via rmap - duplicate the pte as is.
386 */
0b14c179 387 if (vm_flags & VM_UNPAGED)
b5810039
NP
388 goto out_set_pte;
389
1da177e4 390 pfn = pte_pfn(pte);
b5810039 391 /* If the pte points outside of valid memory but
0b14c179 392 * the region is not VM_UNPAGED, we have a problem.
1da177e4 393 */
b5810039
NP
394 if (unlikely(!pfn_valid(pfn))) {
395 print_bad_pte(vma, pte, addr);
396 goto out_set_pte; /* try to do something sane */
397 }
1da177e4 398
b5810039 399 page = pfn_to_page(pfn);
1da177e4
LT
400
401 /*
402 * If it's a COW mapping, write protect it both
403 * in the parent and the child
404 */
405 if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
406 ptep_set_wrprotect(src_mm, addr, src_pte);
407 pte = *src_pte;
408 }
409
410 /*
411 * If it's a shared mapping, mark it clean in
412 * the child
413 */
414 if (vm_flags & VM_SHARED)
415 pte = pte_mkclean(pte);
416 pte = pte_mkold(pte);
417 get_page(page);
1da177e4 418 page_dup_rmap(page);
8c103762 419 rss[!!PageAnon(page)]++;
ae859762
HD
420
421out_set_pte:
422 set_pte_at(dst_mm, addr, dst_pte, pte);
1da177e4
LT
423}
424
425static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
426 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
427 unsigned long addr, unsigned long end)
428{
429 pte_t *src_pte, *dst_pte;
c74df32c 430 spinlock_t *src_ptl, *dst_ptl;
e040f218 431 int progress = 0;
8c103762 432 int rss[2];
1da177e4
LT
433
434again:
ae859762 435 rss[1] = rss[0] = 0;
c74df32c 436 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
1da177e4
LT
437 if (!dst_pte)
438 return -ENOMEM;
439 src_pte = pte_offset_map_nested(src_pmd, addr);
4c21e2f2 440 src_ptl = pte_lockptr(src_mm, src_pmd);
c74df32c 441 spin_lock(src_ptl);
1da177e4 442
1da177e4
LT
443 do {
444 /*
445 * We are holding two locks at this point - either of them
446 * could generate latencies in another task on another CPU.
447 */
e040f218
HD
448 if (progress >= 32) {
449 progress = 0;
450 if (need_resched() ||
c74df32c
HD
451 need_lockbreak(src_ptl) ||
452 need_lockbreak(dst_ptl))
e040f218
HD
453 break;
454 }
1da177e4
LT
455 if (pte_none(*src_pte)) {
456 progress++;
457 continue;
458 }
8c103762 459 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
1da177e4
LT
460 progress += 8;
461 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
1da177e4 462
c74df32c 463 spin_unlock(src_ptl);
1da177e4 464 pte_unmap_nested(src_pte - 1);
ae859762 465 add_mm_rss(dst_mm, rss[0], rss[1]);
c74df32c
HD
466 pte_unmap_unlock(dst_pte - 1, dst_ptl);
467 cond_resched();
1da177e4
LT
468 if (addr != end)
469 goto again;
470 return 0;
471}
472
473static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
474 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
475 unsigned long addr, unsigned long end)
476{
477 pmd_t *src_pmd, *dst_pmd;
478 unsigned long next;
479
480 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
481 if (!dst_pmd)
482 return -ENOMEM;
483 src_pmd = pmd_offset(src_pud, addr);
484 do {
485 next = pmd_addr_end(addr, end);
486 if (pmd_none_or_clear_bad(src_pmd))
487 continue;
488 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
489 vma, addr, next))
490 return -ENOMEM;
491 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
492 return 0;
493}
494
495static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
496 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
497 unsigned long addr, unsigned long end)
498{
499 pud_t *src_pud, *dst_pud;
500 unsigned long next;
501
502 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
503 if (!dst_pud)
504 return -ENOMEM;
505 src_pud = pud_offset(src_pgd, addr);
506 do {
507 next = pud_addr_end(addr, end);
508 if (pud_none_or_clear_bad(src_pud))
509 continue;
510 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
511 vma, addr, next))
512 return -ENOMEM;
513 } while (dst_pud++, src_pud++, addr = next, addr != end);
514 return 0;
515}
516
517int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
518 struct vm_area_struct *vma)
519{
520 pgd_t *src_pgd, *dst_pgd;
521 unsigned long next;
522 unsigned long addr = vma->vm_start;
523 unsigned long end = vma->vm_end;
524
d992895b
NP
525 /*
526 * Don't copy ptes where a page fault will fill them correctly.
527 * Fork becomes much lighter when there are big shared or private
528 * readonly mappings. The tradeoff is that copy_page_range is more
529 * efficient than faulting.
530 */
0b14c179 531 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_UNPAGED))) {
d992895b
NP
532 if (!vma->anon_vma)
533 return 0;
534 }
535
1da177e4
LT
536 if (is_vm_hugetlb_page(vma))
537 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
538
539 dst_pgd = pgd_offset(dst_mm, addr);
540 src_pgd = pgd_offset(src_mm, addr);
541 do {
542 next = pgd_addr_end(addr, end);
543 if (pgd_none_or_clear_bad(src_pgd))
544 continue;
545 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
546 vma, addr, next))
547 return -ENOMEM;
548 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
549 return 0;
550}
551
51c6f666 552static unsigned long zap_pte_range(struct mmu_gather *tlb,
b5810039 553 struct vm_area_struct *vma, pmd_t *pmd,
1da177e4 554 unsigned long addr, unsigned long end,
51c6f666 555 long *zap_work, struct zap_details *details)
1da177e4 556{
b5810039 557 struct mm_struct *mm = tlb->mm;
1da177e4 558 pte_t *pte;
508034a3 559 spinlock_t *ptl;
ae859762
HD
560 int file_rss = 0;
561 int anon_rss = 0;
1da177e4 562
508034a3 563 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
1da177e4
LT
564 do {
565 pte_t ptent = *pte;
51c6f666
RH
566 if (pte_none(ptent)) {
567 (*zap_work)--;
1da177e4 568 continue;
51c6f666 569 }
1da177e4
LT
570 if (pte_present(ptent)) {
571 struct page *page = NULL;
51c6f666
RH
572
573 (*zap_work) -= PAGE_SIZE;
574
0b14c179 575 if (!(vma->vm_flags & VM_UNPAGED)) {
b5810039
NP
576 unsigned long pfn = pte_pfn(ptent);
577 if (unlikely(!pfn_valid(pfn)))
578 print_bad_pte(vma, ptent, addr);
579 else
580 page = pfn_to_page(pfn);
1da177e4
LT
581 }
582 if (unlikely(details) && page) {
583 /*
584 * unmap_shared_mapping_pages() wants to
585 * invalidate cache without truncating:
586 * unmap shared but keep private pages.
587 */
588 if (details->check_mapping &&
589 details->check_mapping != page->mapping)
590 continue;
591 /*
592 * Each page->index must be checked when
593 * invalidating or truncating nonlinear.
594 */
595 if (details->nonlinear_vma &&
596 (page->index < details->first_index ||
597 page->index > details->last_index))
598 continue;
599 }
b5810039 600 ptent = ptep_get_and_clear_full(mm, addr, pte,
a600388d 601 tlb->fullmm);
1da177e4
LT
602 tlb_remove_tlb_entry(tlb, pte, addr);
603 if (unlikely(!page))
604 continue;
605 if (unlikely(details) && details->nonlinear_vma
606 && linear_page_index(details->nonlinear_vma,
607 addr) != page->index)
b5810039 608 set_pte_at(mm, addr, pte,
1da177e4 609 pgoff_to_pte(page->index));
1da177e4 610 if (PageAnon(page))
86d912f4 611 anon_rss--;
6237bcd9
HD
612 else {
613 if (pte_dirty(ptent))
614 set_page_dirty(page);
615 if (pte_young(ptent))
616 mark_page_accessed(page);
86d912f4 617 file_rss--;
6237bcd9 618 }
1da177e4
LT
619 page_remove_rmap(page);
620 tlb_remove_page(tlb, page);
621 continue;
622 }
623 /*
624 * If details->check_mapping, we leave swap entries;
625 * if details->nonlinear_vma, we leave file entries.
626 */
627 if (unlikely(details))
628 continue;
629 if (!pte_file(ptent))
630 free_swap_and_cache(pte_to_swp_entry(ptent));
b5810039 631 pte_clear_full(mm, addr, pte, tlb->fullmm);
51c6f666 632 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
ae859762 633
86d912f4 634 add_mm_rss(mm, file_rss, anon_rss);
508034a3 635 pte_unmap_unlock(pte - 1, ptl);
51c6f666
RH
636
637 return addr;
1da177e4
LT
638}
639
51c6f666 640static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
b5810039 641 struct vm_area_struct *vma, pud_t *pud,
1da177e4 642 unsigned long addr, unsigned long end,
51c6f666 643 long *zap_work, struct zap_details *details)
1da177e4
LT
644{
645 pmd_t *pmd;
646 unsigned long next;
647
648 pmd = pmd_offset(pud, addr);
649 do {
650 next = pmd_addr_end(addr, end);
51c6f666
RH
651 if (pmd_none_or_clear_bad(pmd)) {
652 (*zap_work)--;
1da177e4 653 continue;
51c6f666
RH
654 }
655 next = zap_pte_range(tlb, vma, pmd, addr, next,
656 zap_work, details);
657 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
658
659 return addr;
1da177e4
LT
660}
661
51c6f666 662static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
b5810039 663 struct vm_area_struct *vma, pgd_t *pgd,
1da177e4 664 unsigned long addr, unsigned long end,
51c6f666 665 long *zap_work, struct zap_details *details)
1da177e4
LT
666{
667 pud_t *pud;
668 unsigned long next;
669
670 pud = pud_offset(pgd, addr);
671 do {
672 next = pud_addr_end(addr, end);
51c6f666
RH
673 if (pud_none_or_clear_bad(pud)) {
674 (*zap_work)--;
1da177e4 675 continue;
51c6f666
RH
676 }
677 next = zap_pmd_range(tlb, vma, pud, addr, next,
678 zap_work, details);
679 } while (pud++, addr = next, (addr != end && *zap_work > 0));
680
681 return addr;
1da177e4
LT
682}
683
51c6f666
RH
684static unsigned long unmap_page_range(struct mmu_gather *tlb,
685 struct vm_area_struct *vma,
1da177e4 686 unsigned long addr, unsigned long end,
51c6f666 687 long *zap_work, struct zap_details *details)
1da177e4
LT
688{
689 pgd_t *pgd;
690 unsigned long next;
691
692 if (details && !details->check_mapping && !details->nonlinear_vma)
693 details = NULL;
694
695 BUG_ON(addr >= end);
696 tlb_start_vma(tlb, vma);
697 pgd = pgd_offset(vma->vm_mm, addr);
698 do {
699 next = pgd_addr_end(addr, end);
51c6f666
RH
700 if (pgd_none_or_clear_bad(pgd)) {
701 (*zap_work)--;
1da177e4 702 continue;
51c6f666
RH
703 }
704 next = zap_pud_range(tlb, vma, pgd, addr, next,
705 zap_work, details);
706 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
1da177e4 707 tlb_end_vma(tlb, vma);
51c6f666
RH
708
709 return addr;
1da177e4
LT
710}
711
712#ifdef CONFIG_PREEMPT
713# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
714#else
715/* No preempt: go for improved straight-line efficiency */
716# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
717#endif
718
719/**
720 * unmap_vmas - unmap a range of memory covered by a list of vma's
721 * @tlbp: address of the caller's struct mmu_gather
1da177e4
LT
722 * @vma: the starting vma
723 * @start_addr: virtual address at which to start unmapping
724 * @end_addr: virtual address at which to end unmapping
725 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
726 * @details: details of nonlinear truncation or shared cache invalidation
727 *
ee39b37b 728 * Returns the end address of the unmapping (restart addr if interrupted).
1da177e4 729 *
508034a3 730 * Unmap all pages in the vma list.
1da177e4 731 *
508034a3
HD
732 * We aim to not hold locks for too long (for scheduling latency reasons).
733 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
1da177e4
LT
734 * return the ending mmu_gather to the caller.
735 *
736 * Only addresses between `start' and `end' will be unmapped.
737 *
738 * The VMA list must be sorted in ascending virtual address order.
739 *
740 * unmap_vmas() assumes that the caller will flush the whole unmapped address
741 * range after unmap_vmas() returns. So the only responsibility here is to
742 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
743 * drops the lock and schedules.
744 */
508034a3 745unsigned long unmap_vmas(struct mmu_gather **tlbp,
1da177e4
LT
746 struct vm_area_struct *vma, unsigned long start_addr,
747 unsigned long end_addr, unsigned long *nr_accounted,
748 struct zap_details *details)
749{
51c6f666 750 long zap_work = ZAP_BLOCK_SIZE;
1da177e4
LT
751 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
752 int tlb_start_valid = 0;
ee39b37b 753 unsigned long start = start_addr;
1da177e4 754 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
4d6ddfa9 755 int fullmm = (*tlbp)->fullmm;
1da177e4
LT
756
757 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
1da177e4
LT
758 unsigned long end;
759
760 start = max(vma->vm_start, start_addr);
761 if (start >= vma->vm_end)
762 continue;
763 end = min(vma->vm_end, end_addr);
764 if (end <= vma->vm_start)
765 continue;
766
767 if (vma->vm_flags & VM_ACCOUNT)
768 *nr_accounted += (end - start) >> PAGE_SHIFT;
769
1da177e4 770 while (start != end) {
1da177e4
LT
771 if (!tlb_start_valid) {
772 tlb_start = start;
773 tlb_start_valid = 1;
774 }
775
51c6f666 776 if (unlikely(is_vm_hugetlb_page(vma))) {
1da177e4 777 unmap_hugepage_range(vma, start, end);
51c6f666
RH
778 zap_work -= (end - start) /
779 (HPAGE_SIZE / PAGE_SIZE);
780 start = end;
781 } else
782 start = unmap_page_range(*tlbp, vma,
783 start, end, &zap_work, details);
784
785 if (zap_work > 0) {
786 BUG_ON(start != end);
787 break;
1da177e4
LT
788 }
789
1da177e4
LT
790 tlb_finish_mmu(*tlbp, tlb_start, start);
791
792 if (need_resched() ||
1da177e4
LT
793 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
794 if (i_mmap_lock) {
508034a3 795 *tlbp = NULL;
1da177e4
LT
796 goto out;
797 }
1da177e4 798 cond_resched();
1da177e4
LT
799 }
800
508034a3 801 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
1da177e4 802 tlb_start_valid = 0;
51c6f666 803 zap_work = ZAP_BLOCK_SIZE;
1da177e4
LT
804 }
805 }
806out:
ee39b37b 807 return start; /* which is now the end (or restart) address */
1da177e4
LT
808}
809
810/**
811 * zap_page_range - remove user pages in a given range
812 * @vma: vm_area_struct holding the applicable pages
813 * @address: starting address of pages to zap
814 * @size: number of bytes to zap
815 * @details: details of nonlinear truncation or shared cache invalidation
816 */
ee39b37b 817unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
1da177e4
LT
818 unsigned long size, struct zap_details *details)
819{
820 struct mm_struct *mm = vma->vm_mm;
821 struct mmu_gather *tlb;
822 unsigned long end = address + size;
823 unsigned long nr_accounted = 0;
824
1da177e4 825 lru_add_drain();
1da177e4 826 tlb = tlb_gather_mmu(mm, 0);
365e9c87 827 update_hiwater_rss(mm);
508034a3
HD
828 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
829 if (tlb)
830 tlb_finish_mmu(tlb, address, end);
ee39b37b 831 return end;
1da177e4
LT
832}
833
834/*
835 * Do a quick page-table lookup for a single page.
1da177e4 836 */
deceb6cd
HD
837struct page *follow_page(struct mm_struct *mm, unsigned long address,
838 unsigned int flags)
1da177e4
LT
839{
840 pgd_t *pgd;
841 pud_t *pud;
842 pmd_t *pmd;
843 pte_t *ptep, pte;
deceb6cd 844 spinlock_t *ptl;
1da177e4
LT
845 unsigned long pfn;
846 struct page *page;
847
deceb6cd
HD
848 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
849 if (!IS_ERR(page)) {
850 BUG_ON(flags & FOLL_GET);
851 goto out;
852 }
1da177e4 853
deceb6cd 854 page = NULL;
1da177e4
LT
855 pgd = pgd_offset(mm, address);
856 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
deceb6cd 857 goto no_page_table;
1da177e4
LT
858
859 pud = pud_offset(pgd, address);
860 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
deceb6cd 861 goto no_page_table;
1da177e4
LT
862
863 pmd = pmd_offset(pud, address);
864 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
deceb6cd
HD
865 goto no_page_table;
866
867 if (pmd_huge(*pmd)) {
868 BUG_ON(flags & FOLL_GET);
869 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
1da177e4 870 goto out;
deceb6cd 871 }
1da177e4 872
deceb6cd 873 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
874 if (!ptep)
875 goto out;
876
877 pte = *ptep;
deceb6cd
HD
878 if (!pte_present(pte))
879 goto unlock;
880 if ((flags & FOLL_WRITE) && !pte_write(pte))
881 goto unlock;
882 pfn = pte_pfn(pte);
883 if (!pfn_valid(pfn))
884 goto unlock;
1da177e4 885
deceb6cd
HD
886 page = pfn_to_page(pfn);
887 if (flags & FOLL_GET)
888 get_page(page);
889 if (flags & FOLL_TOUCH) {
890 if ((flags & FOLL_WRITE) &&
891 !pte_dirty(pte) && !PageDirty(page))
892 set_page_dirty(page);
893 mark_page_accessed(page);
894 }
895unlock:
896 pte_unmap_unlock(ptep, ptl);
1da177e4 897out:
deceb6cd 898 return page;
1da177e4 899
deceb6cd
HD
900no_page_table:
901 /*
902 * When core dumping an enormous anonymous area that nobody
903 * has touched so far, we don't want to allocate page tables.
904 */
905 if (flags & FOLL_ANON) {
906 page = ZERO_PAGE(address);
907 if (flags & FOLL_GET)
908 get_page(page);
909 BUG_ON(flags & FOLL_WRITE);
910 }
911 return page;
1da177e4
LT
912}
913
1da177e4
LT
914int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
915 unsigned long start, int len, int write, int force,
916 struct page **pages, struct vm_area_struct **vmas)
917{
918 int i;
deceb6cd 919 unsigned int vm_flags;
1da177e4
LT
920
921 /*
922 * Require read or write permissions.
923 * If 'force' is set, we only require the "MAY" flags.
924 */
deceb6cd
HD
925 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
926 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
1da177e4
LT
927 i = 0;
928
929 do {
deceb6cd
HD
930 struct vm_area_struct *vma;
931 unsigned int foll_flags;
1da177e4
LT
932
933 vma = find_extend_vma(mm, start);
934 if (!vma && in_gate_area(tsk, start)) {
935 unsigned long pg = start & PAGE_MASK;
936 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
937 pgd_t *pgd;
938 pud_t *pud;
939 pmd_t *pmd;
940 pte_t *pte;
941 if (write) /* user gate pages are read-only */
942 return i ? : -EFAULT;
943 if (pg > TASK_SIZE)
944 pgd = pgd_offset_k(pg);
945 else
946 pgd = pgd_offset_gate(mm, pg);
947 BUG_ON(pgd_none(*pgd));
948 pud = pud_offset(pgd, pg);
949 BUG_ON(pud_none(*pud));
950 pmd = pmd_offset(pud, pg);
690dbe1c
HD
951 if (pmd_none(*pmd))
952 return i ? : -EFAULT;
1da177e4 953 pte = pte_offset_map(pmd, pg);
690dbe1c
HD
954 if (pte_none(*pte)) {
955 pte_unmap(pte);
956 return i ? : -EFAULT;
957 }
1da177e4
LT
958 if (pages) {
959 pages[i] = pte_page(*pte);
960 get_page(pages[i]);
961 }
962 pte_unmap(pte);
963 if (vmas)
964 vmas[i] = gate_vma;
965 i++;
966 start += PAGE_SIZE;
967 len--;
968 continue;
969 }
970
ed5297a9 971 if (!vma || (vma->vm_flags & VM_IO)
deceb6cd 972 || !(vm_flags & vma->vm_flags))
1da177e4
LT
973 return i ? : -EFAULT;
974
975 if (is_vm_hugetlb_page(vma)) {
976 i = follow_hugetlb_page(mm, vma, pages, vmas,
977 &start, &len, i);
978 continue;
979 }
deceb6cd
HD
980
981 foll_flags = FOLL_TOUCH;
982 if (pages)
983 foll_flags |= FOLL_GET;
984 if (!write && !(vma->vm_flags & VM_LOCKED) &&
985 (!vma->vm_ops || !vma->vm_ops->nopage))
986 foll_flags |= FOLL_ANON;
987
1da177e4 988 do {
08ef4729 989 struct page *page;
1da177e4 990
deceb6cd
HD
991 if (write)
992 foll_flags |= FOLL_WRITE;
a68d2ebc 993
deceb6cd
HD
994 cond_resched();
995 while (!(page = follow_page(mm, start, foll_flags))) {
996 int ret;
997 ret = __handle_mm_fault(mm, vma, start,
998 foll_flags & FOLL_WRITE);
a68d2ebc
LT
999 /*
1000 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1001 * broken COW when necessary, even if maybe_mkwrite
1002 * decided not to set pte_write. We can thus safely do
1003 * subsequent page lookups as if they were reads.
1004 */
1005 if (ret & VM_FAULT_WRITE)
deceb6cd 1006 foll_flags &= ~FOLL_WRITE;
a68d2ebc
LT
1007
1008 switch (ret & ~VM_FAULT_WRITE) {
1da177e4
LT
1009 case VM_FAULT_MINOR:
1010 tsk->min_flt++;
1011 break;
1012 case VM_FAULT_MAJOR:
1013 tsk->maj_flt++;
1014 break;
1015 case VM_FAULT_SIGBUS:
1016 return i ? i : -EFAULT;
1017 case VM_FAULT_OOM:
1018 return i ? i : -ENOMEM;
1019 default:
1020 BUG();
1021 }
1da177e4
LT
1022 }
1023 if (pages) {
08ef4729
HD
1024 pages[i] = page;
1025 flush_dcache_page(page);
1da177e4
LT
1026 }
1027 if (vmas)
1028 vmas[i] = vma;
1029 i++;
1030 start += PAGE_SIZE;
1031 len--;
08ef4729 1032 } while (len && start < vma->vm_end);
08ef4729 1033 } while (len);
1da177e4
LT
1034 return i;
1035}
1da177e4
LT
1036EXPORT_SYMBOL(get_user_pages);
1037
1038static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1039 unsigned long addr, unsigned long end, pgprot_t prot)
1040{
1041 pte_t *pte;
c74df32c 1042 spinlock_t *ptl;
1da177e4 1043
c74df32c 1044 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1da177e4
LT
1045 if (!pte)
1046 return -ENOMEM;
1047 do {
b5810039
NP
1048 struct page *page = ZERO_PAGE(addr);
1049 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1050 page_cache_get(page);
1051 page_add_file_rmap(page);
1052 inc_mm_counter(mm, file_rss);
1da177e4
LT
1053 BUG_ON(!pte_none(*pte));
1054 set_pte_at(mm, addr, pte, zero_pte);
1055 } while (pte++, addr += PAGE_SIZE, addr != end);
c74df32c 1056 pte_unmap_unlock(pte - 1, ptl);
1da177e4
LT
1057 return 0;
1058}
1059
1060static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1061 unsigned long addr, unsigned long end, pgprot_t prot)
1062{
1063 pmd_t *pmd;
1064 unsigned long next;
1065
1066 pmd = pmd_alloc(mm, pud, addr);
1067 if (!pmd)
1068 return -ENOMEM;
1069 do {
1070 next = pmd_addr_end(addr, end);
1071 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1072 return -ENOMEM;
1073 } while (pmd++, addr = next, addr != end);
1074 return 0;
1075}
1076
1077static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1078 unsigned long addr, unsigned long end, pgprot_t prot)
1079{
1080 pud_t *pud;
1081 unsigned long next;
1082
1083 pud = pud_alloc(mm, pgd, addr);
1084 if (!pud)
1085 return -ENOMEM;
1086 do {
1087 next = pud_addr_end(addr, end);
1088 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1089 return -ENOMEM;
1090 } while (pud++, addr = next, addr != end);
1091 return 0;
1092}
1093
1094int zeromap_page_range(struct vm_area_struct *vma,
1095 unsigned long addr, unsigned long size, pgprot_t prot)
1096{
1097 pgd_t *pgd;
1098 unsigned long next;
1099 unsigned long end = addr + size;
1100 struct mm_struct *mm = vma->vm_mm;
1101 int err;
1102
1103 BUG_ON(addr >= end);
1104 pgd = pgd_offset(mm, addr);
1105 flush_cache_range(vma, addr, end);
1da177e4
LT
1106 do {
1107 next = pgd_addr_end(addr, end);
1108 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1109 if (err)
1110 break;
1111 } while (pgd++, addr = next, addr != end);
1da177e4
LT
1112 return err;
1113}
1114
1115/*
1116 * maps a range of physical memory into the requested pages. the old
1117 * mappings are removed. any references to nonexistent pages results
1118 * in null mappings (currently treated as "copy-on-access")
1119 */
1120static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1121 unsigned long addr, unsigned long end,
1122 unsigned long pfn, pgprot_t prot)
1123{
1124 pte_t *pte;
c74df32c 1125 spinlock_t *ptl;
1da177e4 1126
c74df32c 1127 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1da177e4
LT
1128 if (!pte)
1129 return -ENOMEM;
1130 do {
1131 BUG_ON(!pte_none(*pte));
b5810039 1132 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1da177e4
LT
1133 pfn++;
1134 } while (pte++, addr += PAGE_SIZE, addr != end);
c74df32c 1135 pte_unmap_unlock(pte - 1, ptl);
1da177e4
LT
1136 return 0;
1137}
1138
1139static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1140 unsigned long addr, unsigned long end,
1141 unsigned long pfn, pgprot_t prot)
1142{
1143 pmd_t *pmd;
1144 unsigned long next;
1145
1146 pfn -= addr >> PAGE_SHIFT;
1147 pmd = pmd_alloc(mm, pud, addr);
1148 if (!pmd)
1149 return -ENOMEM;
1150 do {
1151 next = pmd_addr_end(addr, end);
1152 if (remap_pte_range(mm, pmd, addr, next,
1153 pfn + (addr >> PAGE_SHIFT), prot))
1154 return -ENOMEM;
1155 } while (pmd++, addr = next, addr != end);
1156 return 0;
1157}
1158
1159static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1160 unsigned long addr, unsigned long end,
1161 unsigned long pfn, pgprot_t prot)
1162{
1163 pud_t *pud;
1164 unsigned long next;
1165
1166 pfn -= addr >> PAGE_SHIFT;
1167 pud = pud_alloc(mm, pgd, addr);
1168 if (!pud)
1169 return -ENOMEM;
1170 do {
1171 next = pud_addr_end(addr, end);
1172 if (remap_pmd_range(mm, pud, addr, next,
1173 pfn + (addr >> PAGE_SHIFT), prot))
1174 return -ENOMEM;
1175 } while (pud++, addr = next, addr != end);
1176 return 0;
1177}
1178
1179/* Note: this is only safe if the mm semaphore is held when called. */
1180int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1181 unsigned long pfn, unsigned long size, pgprot_t prot)
1182{
1183 pgd_t *pgd;
1184 unsigned long next;
2d15cab8 1185 unsigned long end = addr + PAGE_ALIGN(size);
1da177e4
LT
1186 struct mm_struct *mm = vma->vm_mm;
1187 int err;
1188
1189 /*
1190 * Physically remapped pages are special. Tell the
1191 * rest of the world about it:
1192 * VM_IO tells people not to look at these pages
1193 * (accesses can have side effects).
0b14c179
HD
1194 * VM_RESERVED is specified all over the place, because
1195 * in 2.4 it kept swapout's vma scan off this vma; but
1196 * in 2.6 the LRU scan won't even find its pages, so this
1197 * flag means no more than count its pages in reserved_vm,
1198 * and omit it from core dump, even when VM_IO turned off.
1199 * VM_UNPAGED tells the core MM not to "manage" these pages
1200 * (e.g. refcount, mapcount, try to swap them out): in
1201 * particular, zap_pte_range does not try to free them.
1da177e4 1202 */
0b14c179 1203 vma->vm_flags |= VM_IO | VM_RESERVED | VM_UNPAGED;
1da177e4
LT
1204
1205 BUG_ON(addr >= end);
1206 pfn -= addr >> PAGE_SHIFT;
1207 pgd = pgd_offset(mm, addr);
1208 flush_cache_range(vma, addr, end);
1da177e4
LT
1209 do {
1210 next = pgd_addr_end(addr, end);
1211 err = remap_pud_range(mm, pgd, addr, next,
1212 pfn + (addr >> PAGE_SHIFT), prot);
1213 if (err)
1214 break;
1215 } while (pgd++, addr = next, addr != end);
1da177e4
LT
1216 return err;
1217}
1218EXPORT_SYMBOL(remap_pfn_range);
1219
8f4e2101
HD
1220/*
1221 * handle_pte_fault chooses page fault handler according to an entry
1222 * which was read non-atomically. Before making any commitment, on
1223 * those architectures or configurations (e.g. i386 with PAE) which
1224 * might give a mix of unmatched parts, do_swap_page and do_file_page
1225 * must check under lock before unmapping the pte and proceeding
1226 * (but do_wp_page is only called after already making such a check;
1227 * and do_anonymous_page and do_no_page can safely check later on).
1228 */
4c21e2f2 1229static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
8f4e2101
HD
1230 pte_t *page_table, pte_t orig_pte)
1231{
1232 int same = 1;
1233#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1234 if (sizeof(pte_t) > sizeof(unsigned long)) {
4c21e2f2
HD
1235 spinlock_t *ptl = pte_lockptr(mm, pmd);
1236 spin_lock(ptl);
8f4e2101 1237 same = pte_same(*page_table, orig_pte);
4c21e2f2 1238 spin_unlock(ptl);
8f4e2101
HD
1239 }
1240#endif
1241 pte_unmap(page_table);
1242 return same;
1243}
1244
1da177e4
LT
1245/*
1246 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1247 * servicing faults for write access. In the normal case, do always want
1248 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1249 * that do not have writing enabled, when used by access_process_vm.
1250 */
1251static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1252{
1253 if (likely(vma->vm_flags & VM_WRITE))
1254 pte = pte_mkwrite(pte);
1255 return pte;
1256}
1257
1da177e4
LT
1258/*
1259 * This routine handles present pages, when users try to write
1260 * to a shared page. It is done by copying the page to a new address
1261 * and decrementing the shared-page counter for the old page.
1262 *
1da177e4
LT
1263 * Note that this routine assumes that the protection checks have been
1264 * done by the caller (the low-level page fault routine in most cases).
1265 * Thus we can safely just mark it writable once we've done any necessary
1266 * COW.
1267 *
1268 * We also mark the page dirty at this point even though the page will
1269 * change only once the write actually happens. This avoids a few races,
1270 * and potentially makes it more efficient.
1271 *
8f4e2101
HD
1272 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1273 * but allow concurrent faults), with pte both mapped and locked.
1274 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 1275 */
65500d23
HD
1276static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1277 unsigned long address, pte_t *page_table, pmd_t *pmd,
8f4e2101 1278 spinlock_t *ptl, pte_t orig_pte)
1da177e4 1279{
920fc356 1280 struct page *old_page, *src_page, *new_page;
65500d23 1281 unsigned long pfn = pte_pfn(orig_pte);
1da177e4 1282 pte_t entry;
65500d23 1283 int ret = VM_FAULT_MINOR;
1da177e4
LT
1284
1285 if (unlikely(!pfn_valid(pfn))) {
1286 /*
65500d23 1287 * Page table corrupted: show pte and kill process.
920fc356
HD
1288 * Or it's an attempt to COW an out-of-map VM_UNPAGED
1289 * entry, which copy_user_highpage does not support.
1da177e4 1290 */
b5810039 1291 print_bad_pte(vma, orig_pte, address);
65500d23
HD
1292 ret = VM_FAULT_OOM;
1293 goto unlock;
1da177e4
LT
1294 }
1295 old_page = pfn_to_page(pfn);
920fc356
HD
1296 src_page = old_page;
1297
1298 if (unlikely(vma->vm_flags & VM_UNPAGED)) {
1299 old_page = NULL;
1300 goto gotten;
1301 }
1da177e4 1302
d296e9cd 1303 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1da177e4
LT
1304 int reuse = can_share_swap_page(old_page);
1305 unlock_page(old_page);
1306 if (reuse) {
1307 flush_cache_page(vma, address, pfn);
65500d23
HD
1308 entry = pte_mkyoung(orig_pte);
1309 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1da177e4
LT
1310 ptep_set_access_flags(vma, address, page_table, entry, 1);
1311 update_mmu_cache(vma, address, entry);
1312 lazy_mmu_prot_update(entry);
65500d23
HD
1313 ret |= VM_FAULT_WRITE;
1314 goto unlock;
1da177e4
LT
1315 }
1316 }
1da177e4
LT
1317
1318 /*
1319 * Ok, we need to copy. Oh, well..
1320 */
b5810039 1321 page_cache_get(old_page);
920fc356 1322gotten:
8f4e2101 1323 pte_unmap_unlock(page_table, ptl);
1da177e4
LT
1324
1325 if (unlikely(anon_vma_prepare(vma)))
65500d23 1326 goto oom;
920fc356 1327 if (src_page == ZERO_PAGE(address)) {
1da177e4
LT
1328 new_page = alloc_zeroed_user_highpage(vma, address);
1329 if (!new_page)
65500d23 1330 goto oom;
1da177e4
LT
1331 } else {
1332 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1333 if (!new_page)
65500d23 1334 goto oom;
920fc356 1335 copy_user_highpage(new_page, src_page, address);
1da177e4 1336 }
65500d23 1337
1da177e4
LT
1338 /*
1339 * Re-check the pte - we dropped the lock
1340 */
8f4e2101 1341 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
65500d23 1342 if (likely(pte_same(*page_table, orig_pte))) {
920fc356
HD
1343 if (old_page) {
1344 page_remove_rmap(old_page);
1345 if (!PageAnon(old_page)) {
1346 dec_mm_counter(mm, file_rss);
1347 inc_mm_counter(mm, anon_rss);
1348 }
1349 } else
4294621f 1350 inc_mm_counter(mm, anon_rss);
1da177e4 1351 flush_cache_page(vma, address, pfn);
65500d23
HD
1352 entry = mk_pte(new_page, vma->vm_page_prot);
1353 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1354 ptep_establish(vma, address, page_table, entry);
1355 update_mmu_cache(vma, address, entry);
1356 lazy_mmu_prot_update(entry);
1da177e4
LT
1357 lru_cache_add_active(new_page);
1358 page_add_anon_rmap(new_page, vma, address);
1359
1360 /* Free the old page.. */
1361 new_page = old_page;
f33ea7f4 1362 ret |= VM_FAULT_WRITE;
1da177e4 1363 }
920fc356
HD
1364 if (new_page)
1365 page_cache_release(new_page);
1366 if (old_page)
1367 page_cache_release(old_page);
65500d23 1368unlock:
8f4e2101 1369 pte_unmap_unlock(page_table, ptl);
f33ea7f4 1370 return ret;
65500d23 1371oom:
920fc356
HD
1372 if (old_page)
1373 page_cache_release(old_page);
1da177e4
LT
1374 return VM_FAULT_OOM;
1375}
1376
1377/*
1378 * Helper functions for unmap_mapping_range().
1379 *
1380 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1381 *
1382 * We have to restart searching the prio_tree whenever we drop the lock,
1383 * since the iterator is only valid while the lock is held, and anyway
1384 * a later vma might be split and reinserted earlier while lock dropped.
1385 *
1386 * The list of nonlinear vmas could be handled more efficiently, using
1387 * a placeholder, but handle it in the same way until a need is shown.
1388 * It is important to search the prio_tree before nonlinear list: a vma
1389 * may become nonlinear and be shifted from prio_tree to nonlinear list
1390 * while the lock is dropped; but never shifted from list to prio_tree.
1391 *
1392 * In order to make forward progress despite restarting the search,
1393 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1394 * quickly skip it next time around. Since the prio_tree search only
1395 * shows us those vmas affected by unmapping the range in question, we
1396 * can't efficiently keep all vmas in step with mapping->truncate_count:
1397 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1398 * mapping->truncate_count and vma->vm_truncate_count are protected by
1399 * i_mmap_lock.
1400 *
1401 * In order to make forward progress despite repeatedly restarting some
ee39b37b 1402 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1da177e4
LT
1403 * and restart from that address when we reach that vma again. It might
1404 * have been split or merged, shrunk or extended, but never shifted: so
1405 * restart_addr remains valid so long as it remains in the vma's range.
1406 * unmap_mapping_range forces truncate_count to leap over page-aligned
1407 * values so we can save vma's restart_addr in its truncate_count field.
1408 */
1409#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1410
1411static void reset_vma_truncate_counts(struct address_space *mapping)
1412{
1413 struct vm_area_struct *vma;
1414 struct prio_tree_iter iter;
1415
1416 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1417 vma->vm_truncate_count = 0;
1418 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1419 vma->vm_truncate_count = 0;
1420}
1421
1422static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1423 unsigned long start_addr, unsigned long end_addr,
1424 struct zap_details *details)
1425{
1426 unsigned long restart_addr;
1427 int need_break;
1428
1429again:
1430 restart_addr = vma->vm_truncate_count;
1431 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1432 start_addr = restart_addr;
1433 if (start_addr >= end_addr) {
1434 /* Top of vma has been split off since last time */
1435 vma->vm_truncate_count = details->truncate_count;
1436 return 0;
1437 }
1438 }
1439
ee39b37b
HD
1440 restart_addr = zap_page_range(vma, start_addr,
1441 end_addr - start_addr, details);
1da177e4
LT
1442 need_break = need_resched() ||
1443 need_lockbreak(details->i_mmap_lock);
1444
ee39b37b 1445 if (restart_addr >= end_addr) {
1da177e4
LT
1446 /* We have now completed this vma: mark it so */
1447 vma->vm_truncate_count = details->truncate_count;
1448 if (!need_break)
1449 return 0;
1450 } else {
1451 /* Note restart_addr in vma's truncate_count field */
ee39b37b 1452 vma->vm_truncate_count = restart_addr;
1da177e4
LT
1453 if (!need_break)
1454 goto again;
1455 }
1456
1457 spin_unlock(details->i_mmap_lock);
1458 cond_resched();
1459 spin_lock(details->i_mmap_lock);
1460 return -EINTR;
1461}
1462
1463static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1464 struct zap_details *details)
1465{
1466 struct vm_area_struct *vma;
1467 struct prio_tree_iter iter;
1468 pgoff_t vba, vea, zba, zea;
1469
1470restart:
1471 vma_prio_tree_foreach(vma, &iter, root,
1472 details->first_index, details->last_index) {
1473 /* Skip quickly over those we have already dealt with */
1474 if (vma->vm_truncate_count == details->truncate_count)
1475 continue;
1476
1477 vba = vma->vm_pgoff;
1478 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1479 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1480 zba = details->first_index;
1481 if (zba < vba)
1482 zba = vba;
1483 zea = details->last_index;
1484 if (zea > vea)
1485 zea = vea;
1486
1487 if (unmap_mapping_range_vma(vma,
1488 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1489 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1490 details) < 0)
1491 goto restart;
1492 }
1493}
1494
1495static inline void unmap_mapping_range_list(struct list_head *head,
1496 struct zap_details *details)
1497{
1498 struct vm_area_struct *vma;
1499
1500 /*
1501 * In nonlinear VMAs there is no correspondence between virtual address
1502 * offset and file offset. So we must perform an exhaustive search
1503 * across *all* the pages in each nonlinear VMA, not just the pages
1504 * whose virtual address lies outside the file truncation point.
1505 */
1506restart:
1507 list_for_each_entry(vma, head, shared.vm_set.list) {
1508 /* Skip quickly over those we have already dealt with */
1509 if (vma->vm_truncate_count == details->truncate_count)
1510 continue;
1511 details->nonlinear_vma = vma;
1512 if (unmap_mapping_range_vma(vma, vma->vm_start,
1513 vma->vm_end, details) < 0)
1514 goto restart;
1515 }
1516}
1517
1518/**
1519 * unmap_mapping_range - unmap the portion of all mmaps
1520 * in the specified address_space corresponding to the specified
1521 * page range in the underlying file.
3d41088f 1522 * @mapping: the address space containing mmaps to be unmapped.
1da177e4
LT
1523 * @holebegin: byte in first page to unmap, relative to the start of
1524 * the underlying file. This will be rounded down to a PAGE_SIZE
1525 * boundary. Note that this is different from vmtruncate(), which
1526 * must keep the partial page. In contrast, we must get rid of
1527 * partial pages.
1528 * @holelen: size of prospective hole in bytes. This will be rounded
1529 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1530 * end of the file.
1531 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1532 * but 0 when invalidating pagecache, don't throw away private data.
1533 */
1534void unmap_mapping_range(struct address_space *mapping,
1535 loff_t const holebegin, loff_t const holelen, int even_cows)
1536{
1537 struct zap_details details;
1538 pgoff_t hba = holebegin >> PAGE_SHIFT;
1539 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1540
1541 /* Check for overflow. */
1542 if (sizeof(holelen) > sizeof(hlen)) {
1543 long long holeend =
1544 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1545 if (holeend & ~(long long)ULONG_MAX)
1546 hlen = ULONG_MAX - hba + 1;
1547 }
1548
1549 details.check_mapping = even_cows? NULL: mapping;
1550 details.nonlinear_vma = NULL;
1551 details.first_index = hba;
1552 details.last_index = hba + hlen - 1;
1553 if (details.last_index < details.first_index)
1554 details.last_index = ULONG_MAX;
1555 details.i_mmap_lock = &mapping->i_mmap_lock;
1556
1557 spin_lock(&mapping->i_mmap_lock);
1558
1559 /* serialize i_size write against truncate_count write */
1560 smp_wmb();
1561 /* Protect against page faults, and endless unmapping loops */
1562 mapping->truncate_count++;
1563 /*
1564 * For archs where spin_lock has inclusive semantics like ia64
1565 * this smp_mb() will prevent to read pagetable contents
1566 * before the truncate_count increment is visible to
1567 * other cpus.
1568 */
1569 smp_mb();
1570 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1571 if (mapping->truncate_count == 0)
1572 reset_vma_truncate_counts(mapping);
1573 mapping->truncate_count++;
1574 }
1575 details.truncate_count = mapping->truncate_count;
1576
1577 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1578 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1579 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1580 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1581 spin_unlock(&mapping->i_mmap_lock);
1582}
1583EXPORT_SYMBOL(unmap_mapping_range);
1584
1585/*
1586 * Handle all mappings that got truncated by a "truncate()"
1587 * system call.
1588 *
1589 * NOTE! We have to be ready to update the memory sharing
1590 * between the file and the memory map for a potential last
1591 * incomplete page. Ugly, but necessary.
1592 */
1593int vmtruncate(struct inode * inode, loff_t offset)
1594{
1595 struct address_space *mapping = inode->i_mapping;
1596 unsigned long limit;
1597
1598 if (inode->i_size < offset)
1599 goto do_expand;
1600 /*
1601 * truncation of in-use swapfiles is disallowed - it would cause
1602 * subsequent swapout to scribble on the now-freed blocks.
1603 */
1604 if (IS_SWAPFILE(inode))
1605 goto out_busy;
1606 i_size_write(inode, offset);
1607 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1608 truncate_inode_pages(mapping, offset);
1609 goto out_truncate;
1610
1611do_expand:
1612 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1613 if (limit != RLIM_INFINITY && offset > limit)
1614 goto out_sig;
1615 if (offset > inode->i_sb->s_maxbytes)
1616 goto out_big;
1617 i_size_write(inode, offset);
1618
1619out_truncate:
1620 if (inode->i_op && inode->i_op->truncate)
1621 inode->i_op->truncate(inode);
1622 return 0;
1623out_sig:
1624 send_sig(SIGXFSZ, current, 0);
1625out_big:
1626 return -EFBIG;
1627out_busy:
1628 return -ETXTBSY;
1629}
1630
1631EXPORT_SYMBOL(vmtruncate);
1632
1633/*
1634 * Primitive swap readahead code. We simply read an aligned block of
1635 * (1 << page_cluster) entries in the swap area. This method is chosen
1636 * because it doesn't cost us any seek time. We also make sure to queue
1637 * the 'original' request together with the readahead ones...
1638 *
1639 * This has been extended to use the NUMA policies from the mm triggering
1640 * the readahead.
1641 *
1642 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1643 */
1644void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1645{
1646#ifdef CONFIG_NUMA
1647 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1648#endif
1649 int i, num;
1650 struct page *new_page;
1651 unsigned long offset;
1652
1653 /*
1654 * Get the number of handles we should do readahead io to.
1655 */
1656 num = valid_swaphandles(entry, &offset);
1657 for (i = 0; i < num; offset++, i++) {
1658 /* Ok, do the async read-ahead now */
1659 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1660 offset), vma, addr);
1661 if (!new_page)
1662 break;
1663 page_cache_release(new_page);
1664#ifdef CONFIG_NUMA
1665 /*
1666 * Find the next applicable VMA for the NUMA policy.
1667 */
1668 addr += PAGE_SIZE;
1669 if (addr == 0)
1670 vma = NULL;
1671 if (vma) {
1672 if (addr >= vma->vm_end) {
1673 vma = next_vma;
1674 next_vma = vma ? vma->vm_next : NULL;
1675 }
1676 if (vma && addr < vma->vm_start)
1677 vma = NULL;
1678 } else {
1679 if (next_vma && addr >= next_vma->vm_start) {
1680 vma = next_vma;
1681 next_vma = vma->vm_next;
1682 }
1683 }
1684#endif
1685 }
1686 lru_add_drain(); /* Push any new pages onto the LRU now */
1687}
1688
1689/*
8f4e2101
HD
1690 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1691 * but allow concurrent faults), and pte mapped but not yet locked.
1692 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 1693 */
65500d23
HD
1694static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1695 unsigned long address, pte_t *page_table, pmd_t *pmd,
1696 int write_access, pte_t orig_pte)
1da177e4 1697{
8f4e2101 1698 spinlock_t *ptl;
1da177e4 1699 struct page *page;
65500d23 1700 swp_entry_t entry;
1da177e4
LT
1701 pte_t pte;
1702 int ret = VM_FAULT_MINOR;
1703
4c21e2f2 1704 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
8f4e2101 1705 goto out;
65500d23
HD
1706
1707 entry = pte_to_swp_entry(orig_pte);
1da177e4
LT
1708 page = lookup_swap_cache(entry);
1709 if (!page) {
1710 swapin_readahead(entry, address, vma);
1711 page = read_swap_cache_async(entry, vma, address);
1712 if (!page) {
1713 /*
8f4e2101
HD
1714 * Back out if somebody else faulted in this pte
1715 * while we released the pte lock.
1da177e4 1716 */
8f4e2101 1717 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
1718 if (likely(pte_same(*page_table, orig_pte)))
1719 ret = VM_FAULT_OOM;
65500d23 1720 goto unlock;
1da177e4
LT
1721 }
1722
1723 /* Had to read the page from swap area: Major fault */
1724 ret = VM_FAULT_MAJOR;
1725 inc_page_state(pgmajfault);
1726 grab_swap_token();
1727 }
1728
1729 mark_page_accessed(page);
1730 lock_page(page);
1731
1732 /*
8f4e2101 1733 * Back out if somebody else already faulted in this pte.
1da177e4 1734 */
8f4e2101 1735 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
9e9bef07 1736 if (unlikely(!pte_same(*page_table, orig_pte)))
b8107480 1737 goto out_nomap;
b8107480
KK
1738
1739 if (unlikely(!PageUptodate(page))) {
1740 ret = VM_FAULT_SIGBUS;
1741 goto out_nomap;
1da177e4
LT
1742 }
1743
1744 /* The page isn't present yet, go ahead with the fault. */
1da177e4 1745
4294621f 1746 inc_mm_counter(mm, anon_rss);
1da177e4
LT
1747 pte = mk_pte(page, vma->vm_page_prot);
1748 if (write_access && can_share_swap_page(page)) {
1749 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1750 write_access = 0;
1751 }
1da177e4
LT
1752
1753 flush_icache_page(vma, page);
1754 set_pte_at(mm, address, page_table, pte);
1755 page_add_anon_rmap(page, vma, address);
1756
c475a8ab
HD
1757 swap_free(entry);
1758 if (vm_swap_full())
1759 remove_exclusive_swap_page(page);
1760 unlock_page(page);
1761
1da177e4
LT
1762 if (write_access) {
1763 if (do_wp_page(mm, vma, address,
8f4e2101 1764 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
1da177e4
LT
1765 ret = VM_FAULT_OOM;
1766 goto out;
1767 }
1768
1769 /* No need to invalidate - it was non-present before */
1770 update_mmu_cache(vma, address, pte);
1771 lazy_mmu_prot_update(pte);
65500d23 1772unlock:
8f4e2101 1773 pte_unmap_unlock(page_table, ptl);
1da177e4
LT
1774out:
1775 return ret;
b8107480 1776out_nomap:
8f4e2101 1777 pte_unmap_unlock(page_table, ptl);
b8107480
KK
1778 unlock_page(page);
1779 page_cache_release(page);
65500d23 1780 return ret;
1da177e4
LT
1781}
1782
1783/*
8f4e2101
HD
1784 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1785 * but allow concurrent faults), and pte mapped but not yet locked.
1786 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 1787 */
65500d23
HD
1788static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1789 unsigned long address, pte_t *page_table, pmd_t *pmd,
1790 int write_access)
1da177e4 1791{
8f4e2101
HD
1792 struct page *page;
1793 spinlock_t *ptl;
1da177e4 1794 pte_t entry;
1da177e4 1795
1da177e4
LT
1796 if (write_access) {
1797 /* Allocate our own private page. */
1798 pte_unmap(page_table);
1da177e4
LT
1799
1800 if (unlikely(anon_vma_prepare(vma)))
65500d23
HD
1801 goto oom;
1802 page = alloc_zeroed_user_highpage(vma, address);
1da177e4 1803 if (!page)
65500d23 1804 goto oom;
1da177e4 1805
65500d23
HD
1806 entry = mk_pte(page, vma->vm_page_prot);
1807 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
8f4e2101
HD
1808
1809 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1810 if (!pte_none(*page_table))
1811 goto release;
1812 inc_mm_counter(mm, anon_rss);
1da177e4
LT
1813 lru_cache_add_active(page);
1814 SetPageReferenced(page);
65500d23 1815 page_add_anon_rmap(page, vma, address);
b5810039 1816 } else {
8f4e2101
HD
1817 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1818 page = ZERO_PAGE(address);
1819 page_cache_get(page);
1820 entry = mk_pte(page, vma->vm_page_prot);
1821
4c21e2f2 1822 ptl = pte_lockptr(mm, pmd);
8f4e2101
HD
1823 spin_lock(ptl);
1824 if (!pte_none(*page_table))
1825 goto release;
b5810039
NP
1826 inc_mm_counter(mm, file_rss);
1827 page_add_file_rmap(page);
1da177e4
LT
1828 }
1829
65500d23 1830 set_pte_at(mm, address, page_table, entry);
1da177e4
LT
1831
1832 /* No need to invalidate - it was non-present before */
65500d23 1833 update_mmu_cache(vma, address, entry);
1da177e4 1834 lazy_mmu_prot_update(entry);
65500d23 1835unlock:
8f4e2101 1836 pte_unmap_unlock(page_table, ptl);
1da177e4 1837 return VM_FAULT_MINOR;
8f4e2101
HD
1838release:
1839 page_cache_release(page);
1840 goto unlock;
65500d23 1841oom:
1da177e4
LT
1842 return VM_FAULT_OOM;
1843}
1844
1845/*
1846 * do_no_page() tries to create a new page mapping. It aggressively
1847 * tries to share with existing pages, but makes a separate copy if
1848 * the "write_access" parameter is true in order to avoid the next
1849 * page fault.
1850 *
1851 * As this is called only for pages that do not currently exist, we
1852 * do not need to flush old virtual caches or the TLB.
1853 *
8f4e2101
HD
1854 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1855 * but allow concurrent faults), and pte mapped but not yet locked.
1856 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 1857 */
65500d23
HD
1858static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1859 unsigned long address, pte_t *page_table, pmd_t *pmd,
1860 int write_access)
1da177e4 1861{
8f4e2101 1862 spinlock_t *ptl;
65500d23 1863 struct page *new_page;
1da177e4
LT
1864 struct address_space *mapping = NULL;
1865 pte_t entry;
1866 unsigned int sequence = 0;
1867 int ret = VM_FAULT_MINOR;
1868 int anon = 0;
1869
1da177e4 1870 pte_unmap(page_table);
1da177e4
LT
1871
1872 if (vma->vm_file) {
1873 mapping = vma->vm_file->f_mapping;
1874 sequence = mapping->truncate_count;
1875 smp_rmb(); /* serializes i_size against truncate_count */
1876 }
1877retry:
1da177e4
LT
1878 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1879 /*
1880 * No smp_rmb is needed here as long as there's a full
1881 * spin_lock/unlock sequence inside the ->nopage callback
1882 * (for the pagecache lookup) that acts as an implicit
1883 * smp_mb() and prevents the i_size read to happen
1884 * after the next truncate_count read.
1885 */
1886
1887 /* no page was available -- either SIGBUS or OOM */
1888 if (new_page == NOPAGE_SIGBUS)
1889 return VM_FAULT_SIGBUS;
1890 if (new_page == NOPAGE_OOM)
1891 return VM_FAULT_OOM;
1892
1893 /*
1894 * Should we do an early C-O-W break?
1895 */
1896 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1897 struct page *page;
1898
1899 if (unlikely(anon_vma_prepare(vma)))
1900 goto oom;
1901 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1902 if (!page)
1903 goto oom;
1904 copy_user_highpage(page, new_page, address);
1905 page_cache_release(new_page);
1906 new_page = page;
1907 anon = 1;
1908 }
1909
8f4e2101 1910 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1da177e4
LT
1911 /*
1912 * For a file-backed vma, someone could have truncated or otherwise
1913 * invalidated this page. If unmap_mapping_range got called,
1914 * retry getting the page.
1915 */
1916 if (mapping && unlikely(sequence != mapping->truncate_count)) {
8f4e2101 1917 pte_unmap_unlock(page_table, ptl);
1da177e4 1918 page_cache_release(new_page);
65500d23
HD
1919 cond_resched();
1920 sequence = mapping->truncate_count;
1921 smp_rmb();
1da177e4
LT
1922 goto retry;
1923 }
1da177e4
LT
1924
1925 /*
1926 * This silly early PAGE_DIRTY setting removes a race
1927 * due to the bad i386 page protection. But it's valid
1928 * for other architectures too.
1929 *
1930 * Note that if write_access is true, we either now have
1931 * an exclusive copy of the page, or this is a shared mapping,
1932 * so we can make it writable and dirty to avoid having to
1933 * handle that later.
1934 */
1935 /* Only go through if we didn't race with anybody else... */
1936 if (pte_none(*page_table)) {
1da177e4
LT
1937 flush_icache_page(vma, new_page);
1938 entry = mk_pte(new_page, vma->vm_page_prot);
1939 if (write_access)
1940 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1941 set_pte_at(mm, address, page_table, entry);
1942 if (anon) {
4294621f 1943 inc_mm_counter(mm, anon_rss);
1da177e4
LT
1944 lru_cache_add_active(new_page);
1945 page_add_anon_rmap(new_page, vma, address);
0b14c179 1946 } else if (!(vma->vm_flags & VM_UNPAGED)) {
4294621f 1947 inc_mm_counter(mm, file_rss);
1da177e4 1948 page_add_file_rmap(new_page);
4294621f 1949 }
1da177e4
LT
1950 } else {
1951 /* One of our sibling threads was faster, back out. */
1da177e4 1952 page_cache_release(new_page);
65500d23 1953 goto unlock;
1da177e4
LT
1954 }
1955
1956 /* no need to invalidate: a not-present page shouldn't be cached */
1957 update_mmu_cache(vma, address, entry);
1958 lazy_mmu_prot_update(entry);
65500d23 1959unlock:
8f4e2101 1960 pte_unmap_unlock(page_table, ptl);
1da177e4
LT
1961 return ret;
1962oom:
1963 page_cache_release(new_page);
65500d23 1964 return VM_FAULT_OOM;
1da177e4
LT
1965}
1966
1967/*
1968 * Fault of a previously existing named mapping. Repopulate the pte
1969 * from the encoded file_pte if possible. This enables swappable
1970 * nonlinear vmas.
8f4e2101
HD
1971 *
1972 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1973 * but allow concurrent faults), and pte mapped but not yet locked.
1974 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4 1975 */
65500d23
HD
1976static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
1977 unsigned long address, pte_t *page_table, pmd_t *pmd,
1978 int write_access, pte_t orig_pte)
1da177e4 1979{
65500d23 1980 pgoff_t pgoff;
1da177e4
LT
1981 int err;
1982
4c21e2f2 1983 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
8f4e2101 1984 return VM_FAULT_MINOR;
1da177e4 1985
65500d23
HD
1986 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
1987 /*
1988 * Page table corrupted: show pte and kill process.
1989 */
b5810039 1990 print_bad_pte(vma, orig_pte, address);
65500d23
HD
1991 return VM_FAULT_OOM;
1992 }
1993 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
1994
1995 pgoff = pte_to_pgoff(orig_pte);
1996 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
1997 vma->vm_page_prot, pgoff, 0);
1da177e4
LT
1998 if (err == -ENOMEM)
1999 return VM_FAULT_OOM;
2000 if (err)
2001 return VM_FAULT_SIGBUS;
2002 return VM_FAULT_MAJOR;
2003}
2004
2005/*
2006 * These routines also need to handle stuff like marking pages dirty
2007 * and/or accessed for architectures that don't do it in hardware (most
2008 * RISC architectures). The early dirtying is also good on the i386.
2009 *
2010 * There is also a hook called "update_mmu_cache()" that architectures
2011 * with external mmu caches can use to update those (ie the Sparc or
2012 * PowerPC hashed page tables that act as extended TLBs).
2013 *
c74df32c
HD
2014 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2015 * but allow concurrent faults), and pte mapped but not yet locked.
2016 * We return with mmap_sem still held, but pte unmapped and unlocked.
1da177e4
LT
2017 */
2018static inline int handle_pte_fault(struct mm_struct *mm,
65500d23
HD
2019 struct vm_area_struct *vma, unsigned long address,
2020 pte_t *pte, pmd_t *pmd, int write_access)
1da177e4
LT
2021{
2022 pte_t entry;
1a44e149 2023 pte_t old_entry;
8f4e2101 2024 spinlock_t *ptl;
1da177e4 2025
1a44e149 2026 old_entry = entry = *pte;
1da177e4 2027 if (!pte_present(entry)) {
65500d23
HD
2028 if (pte_none(entry)) {
2029 if (!vma->vm_ops || !vma->vm_ops->nopage)
2030 return do_anonymous_page(mm, vma, address,
2031 pte, pmd, write_access);
2032 return do_no_page(mm, vma, address,
2033 pte, pmd, write_access);
2034 }
1da177e4 2035 if (pte_file(entry))
65500d23
HD
2036 return do_file_page(mm, vma, address,
2037 pte, pmd, write_access, entry);
2038 return do_swap_page(mm, vma, address,
2039 pte, pmd, write_access, entry);
1da177e4
LT
2040 }
2041
4c21e2f2 2042 ptl = pte_lockptr(mm, pmd);
8f4e2101
HD
2043 spin_lock(ptl);
2044 if (unlikely(!pte_same(*pte, entry)))
2045 goto unlock;
1da177e4
LT
2046 if (write_access) {
2047 if (!pte_write(entry))
8f4e2101
HD
2048 return do_wp_page(mm, vma, address,
2049 pte, pmd, ptl, entry);
1da177e4
LT
2050 entry = pte_mkdirty(entry);
2051 }
2052 entry = pte_mkyoung(entry);
1a44e149
AA
2053 if (!pte_same(old_entry, entry)) {
2054 ptep_set_access_flags(vma, address, pte, entry, write_access);
2055 update_mmu_cache(vma, address, entry);
2056 lazy_mmu_prot_update(entry);
2057 } else {
2058 /*
2059 * This is needed only for protection faults but the arch code
2060 * is not yet telling us if this is a protection fault or not.
2061 * This still avoids useless tlb flushes for .text page faults
2062 * with threads.
2063 */
2064 if (write_access)
2065 flush_tlb_page(vma, address);
2066 }
8f4e2101
HD
2067unlock:
2068 pte_unmap_unlock(pte, ptl);
1da177e4
LT
2069 return VM_FAULT_MINOR;
2070}
2071
2072/*
2073 * By the time we get here, we already hold the mm semaphore
2074 */
65500d23 2075int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
1da177e4
LT
2076 unsigned long address, int write_access)
2077{
2078 pgd_t *pgd;
2079 pud_t *pud;
2080 pmd_t *pmd;
2081 pte_t *pte;
2082
2083 __set_current_state(TASK_RUNNING);
2084
2085 inc_page_state(pgfault);
2086
ac9b9c66
HD
2087 if (unlikely(is_vm_hugetlb_page(vma)))
2088 return hugetlb_fault(mm, vma, address, write_access);
1da177e4 2089
1da177e4 2090 pgd = pgd_offset(mm, address);
1da177e4
LT
2091 pud = pud_alloc(mm, pgd, address);
2092 if (!pud)
c74df32c 2093 return VM_FAULT_OOM;
1da177e4
LT
2094 pmd = pmd_alloc(mm, pud, address);
2095 if (!pmd)
c74df32c 2096 return VM_FAULT_OOM;
1da177e4
LT
2097 pte = pte_alloc_map(mm, pmd, address);
2098 if (!pte)
c74df32c 2099 return VM_FAULT_OOM;
1da177e4 2100
c74df32c 2101 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
1da177e4
LT
2102}
2103
2104#ifndef __PAGETABLE_PUD_FOLDED
2105/*
2106 * Allocate page upper directory.
872fec16 2107 * We've already handled the fast-path in-line.
1da177e4 2108 */
1bb3630e 2109int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
1da177e4 2110{
c74df32c
HD
2111 pud_t *new = pud_alloc_one(mm, address);
2112 if (!new)
1bb3630e 2113 return -ENOMEM;
1da177e4 2114
872fec16 2115 spin_lock(&mm->page_table_lock);
1bb3630e 2116 if (pgd_present(*pgd)) /* Another has populated it */
1da177e4 2117 pud_free(new);
1bb3630e
HD
2118 else
2119 pgd_populate(mm, pgd, new);
c74df32c 2120 spin_unlock(&mm->page_table_lock);
1bb3630e 2121 return 0;
1da177e4
LT
2122}
2123#endif /* __PAGETABLE_PUD_FOLDED */
2124
2125#ifndef __PAGETABLE_PMD_FOLDED
2126/*
2127 * Allocate page middle directory.
872fec16 2128 * We've already handled the fast-path in-line.
1da177e4 2129 */
1bb3630e 2130int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
1da177e4 2131{
c74df32c
HD
2132 pmd_t *new = pmd_alloc_one(mm, address);
2133 if (!new)
1bb3630e 2134 return -ENOMEM;
1da177e4 2135
872fec16 2136 spin_lock(&mm->page_table_lock);
1da177e4 2137#ifndef __ARCH_HAS_4LEVEL_HACK
1bb3630e 2138 if (pud_present(*pud)) /* Another has populated it */
1da177e4 2139 pmd_free(new);
1bb3630e
HD
2140 else
2141 pud_populate(mm, pud, new);
1da177e4 2142#else
1bb3630e 2143 if (pgd_present(*pud)) /* Another has populated it */
1da177e4 2144 pmd_free(new);
1bb3630e
HD
2145 else
2146 pgd_populate(mm, pud, new);
1da177e4 2147#endif /* __ARCH_HAS_4LEVEL_HACK */
c74df32c 2148 spin_unlock(&mm->page_table_lock);
1bb3630e 2149 return 0;
1da177e4
LT
2150}
2151#endif /* __PAGETABLE_PMD_FOLDED */
2152
2153int make_pages_present(unsigned long addr, unsigned long end)
2154{
2155 int ret, len, write;
2156 struct vm_area_struct * vma;
2157
2158 vma = find_vma(current->mm, addr);
2159 if (!vma)
2160 return -1;
2161 write = (vma->vm_flags & VM_WRITE) != 0;
2162 if (addr >= end)
2163 BUG();
2164 if (end > vma->vm_end)
2165 BUG();
2166 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2167 ret = get_user_pages(current, current->mm, addr,
2168 len, write, 0, NULL, NULL);
2169 if (ret < 0)
2170 return ret;
2171 return ret == len ? 0 : -1;
2172}
2173
2174/*
2175 * Map a vmalloc()-space virtual address to the physical page.
2176 */
2177struct page * vmalloc_to_page(void * vmalloc_addr)
2178{
2179 unsigned long addr = (unsigned long) vmalloc_addr;
2180 struct page *page = NULL;
2181 pgd_t *pgd = pgd_offset_k(addr);
2182 pud_t *pud;
2183 pmd_t *pmd;
2184 pte_t *ptep, pte;
2185
2186 if (!pgd_none(*pgd)) {
2187 pud = pud_offset(pgd, addr);
2188 if (!pud_none(*pud)) {
2189 pmd = pmd_offset(pud, addr);
2190 if (!pmd_none(*pmd)) {
2191 ptep = pte_offset_map(pmd, addr);
2192 pte = *ptep;
2193 if (pte_present(pte))
2194 page = pte_page(pte);
2195 pte_unmap(ptep);
2196 }
2197 }
2198 }
2199 return page;
2200}
2201
2202EXPORT_SYMBOL(vmalloc_to_page);
2203
2204/*
2205 * Map a vmalloc()-space virtual address to the physical page frame number.
2206 */
2207unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2208{
2209 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2210}
2211
2212EXPORT_SYMBOL(vmalloc_to_pfn);
2213
1da177e4
LT
2214#if !defined(__HAVE_ARCH_GATE_AREA)
2215
2216#if defined(AT_SYSINFO_EHDR)
5ce7852c 2217static struct vm_area_struct gate_vma;
1da177e4
LT
2218
2219static int __init gate_vma_init(void)
2220{
2221 gate_vma.vm_mm = NULL;
2222 gate_vma.vm_start = FIXADDR_USER_START;
2223 gate_vma.vm_end = FIXADDR_USER_END;
2224 gate_vma.vm_page_prot = PAGE_READONLY;
0b14c179 2225 gate_vma.vm_flags = 0;
1da177e4
LT
2226 return 0;
2227}
2228__initcall(gate_vma_init);
2229#endif
2230
2231struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2232{
2233#ifdef AT_SYSINFO_EHDR
2234 return &gate_vma;
2235#else
2236 return NULL;
2237#endif
2238}
2239
2240int in_gate_area_no_task(unsigned long addr)
2241{
2242#ifdef AT_SYSINFO_EHDR
2243 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2244 return 1;
2245#endif
2246 return 0;
2247}
2248
2249#endif /* __HAVE_ARCH_GATE_AREA */