ia64/xen-unstable

view linux-2.6-xen-sparse/mm/memory.c @ 11624:f5fd563bcc84

[XEN] Small clean up.
Signed-off-by: Keir Fraser <keir@xensource.com>
author kfraser@localhost.localdomain
date Mon Sep 25 17:45:28 2006 +0100 (2006-09-25)
parents eb7e5d95e7ea
children e1f3af226a8e
line source
1 /*
2 * linux/mm/memory.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 */
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 */
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 */
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 */
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 */
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>
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>
58 #include <linux/swapops.h>
59 #include <linux/elf.h>
61 #ifndef CONFIG_NEED_MULTIPLE_NODES
62 /* use the per-pgdat data instead for discontigmem - mbligh */
63 unsigned long max_mapnr;
64 struct page *mem_map;
66 EXPORT_SYMBOL(max_mapnr);
67 EXPORT_SYMBOL(mem_map);
68 #endif
70 unsigned 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 */
78 void * high_memory;
79 unsigned long vmalloc_earlyreserve;
81 EXPORT_SYMBOL(num_physpages);
82 EXPORT_SYMBOL(high_memory);
83 EXPORT_SYMBOL(vmalloc_earlyreserve);
85 int randomize_va_space __read_mostly = 1;
87 static int __init disable_randmaps(char *s)
88 {
89 randomize_va_space = 0;
90 return 0;
91 }
92 __setup("norandmaps", disable_randmaps);
95 /*
96 * If a p?d_bad entry is found while walking page tables, report
97 * the error, before resetting entry to p?d_none. Usually (but
98 * very seldom) called out from the p?d_none_or_clear_bad macros.
99 */
101 void pgd_clear_bad(pgd_t *pgd)
102 {
103 pgd_ERROR(*pgd);
104 pgd_clear(pgd);
105 }
107 void pud_clear_bad(pud_t *pud)
108 {
109 pud_ERROR(*pud);
110 pud_clear(pud);
111 }
113 void pmd_clear_bad(pmd_t *pmd)
114 {
115 pmd_ERROR(*pmd);
116 pmd_clear(pmd);
117 }
119 /*
120 * Note: this doesn't free the actual pages themselves. That
121 * has been handled earlier when unmapping all the memory regions.
122 */
123 static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
124 {
125 struct page *page = pmd_page(*pmd);
126 pmd_clear(pmd);
127 pte_lock_deinit(page);
128 pte_free_tlb(tlb, page);
129 dec_page_state(nr_page_table_pages);
130 tlb->mm->nr_ptes--;
131 }
133 static inline void free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
134 unsigned long addr, unsigned long end,
135 unsigned long floor, unsigned long ceiling)
136 {
137 pmd_t *pmd;
138 unsigned long next;
139 unsigned long start;
141 start = addr;
142 pmd = pmd_offset(pud, addr);
143 do {
144 next = pmd_addr_end(addr, end);
145 if (pmd_none_or_clear_bad(pmd))
146 continue;
147 free_pte_range(tlb, pmd);
148 } while (pmd++, addr = next, addr != end);
150 start &= PUD_MASK;
151 if (start < floor)
152 return;
153 if (ceiling) {
154 ceiling &= PUD_MASK;
155 if (!ceiling)
156 return;
157 }
158 if (end - 1 > ceiling - 1)
159 return;
161 pmd = pmd_offset(pud, start);
162 pud_clear(pud);
163 pmd_free_tlb(tlb, pmd);
164 }
166 static inline void free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
167 unsigned long addr, unsigned long end,
168 unsigned long floor, unsigned long ceiling)
169 {
170 pud_t *pud;
171 unsigned long next;
172 unsigned long start;
174 start = addr;
175 pud = pud_offset(pgd, addr);
176 do {
177 next = pud_addr_end(addr, end);
178 if (pud_none_or_clear_bad(pud))
179 continue;
180 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
181 } while (pud++, addr = next, addr != end);
183 start &= PGDIR_MASK;
184 if (start < floor)
185 return;
186 if (ceiling) {
187 ceiling &= PGDIR_MASK;
188 if (!ceiling)
189 return;
190 }
191 if (end - 1 > ceiling - 1)
192 return;
194 pud = pud_offset(pgd, start);
195 pgd_clear(pgd);
196 pud_free_tlb(tlb, pud);
197 }
199 /*
200 * This function frees user-level page tables of a process.
201 *
202 * Must be called with pagetable lock held.
203 */
204 void free_pgd_range(struct mmu_gather **tlb,
205 unsigned long addr, unsigned long end,
206 unsigned long floor, unsigned long ceiling)
207 {
208 pgd_t *pgd;
209 unsigned long next;
210 unsigned long start;
212 /*
213 * The next few lines have given us lots of grief...
214 *
215 * Why are we testing PMD* at this top level? Because often
216 * there will be no work to do at all, and we'd prefer not to
217 * go all the way down to the bottom just to discover that.
218 *
219 * Why all these "- 1"s? Because 0 represents both the bottom
220 * of the address space and the top of it (using -1 for the
221 * top wouldn't help much: the masks would do the wrong thing).
222 * The rule is that addr 0 and floor 0 refer to the bottom of
223 * the address space, but end 0 and ceiling 0 refer to the top
224 * Comparisons need to use "end - 1" and "ceiling - 1" (though
225 * that end 0 case should be mythical).
226 *
227 * Wherever addr is brought up or ceiling brought down, we must
228 * be careful to reject "the opposite 0" before it confuses the
229 * subsequent tests. But what about where end is brought down
230 * by PMD_SIZE below? no, end can't go down to 0 there.
231 *
232 * Whereas we round start (addr) and ceiling down, by different
233 * masks at different levels, in order to test whether a table
234 * now has no other vmas using it, so can be freed, we don't
235 * bother to round floor or end up - the tests don't need that.
236 */
238 addr &= PMD_MASK;
239 if (addr < floor) {
240 addr += PMD_SIZE;
241 if (!addr)
242 return;
243 }
244 if (ceiling) {
245 ceiling &= PMD_MASK;
246 if (!ceiling)
247 return;
248 }
249 if (end - 1 > ceiling - 1)
250 end -= PMD_SIZE;
251 if (addr > end - 1)
252 return;
254 start = addr;
255 pgd = pgd_offset((*tlb)->mm, addr);
256 do {
257 next = pgd_addr_end(addr, end);
258 if (pgd_none_or_clear_bad(pgd))
259 continue;
260 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
261 } while (pgd++, addr = next, addr != end);
263 if (!(*tlb)->fullmm)
264 flush_tlb_pgtables((*tlb)->mm, start, end);
265 }
267 void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
268 unsigned long floor, unsigned long ceiling)
269 {
270 while (vma) {
271 struct vm_area_struct *next = vma->vm_next;
272 unsigned long addr = vma->vm_start;
274 /*
275 * Hide vma from rmap and vmtruncate before freeing pgtables
276 */
277 anon_vma_unlink(vma);
278 unlink_file_vma(vma);
280 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
281 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
282 floor, next? next->vm_start: ceiling);
283 } else {
284 /*
285 * Optimization: gather nearby vmas into one call down
286 */
287 while (next && next->vm_start <= vma->vm_end + PMD_SIZE
288 && !is_hugepage_only_range(vma->vm_mm, next->vm_start,
289 HPAGE_SIZE)) {
290 vma = next;
291 next = vma->vm_next;
292 anon_vma_unlink(vma);
293 unlink_file_vma(vma);
294 }
295 free_pgd_range(tlb, addr, vma->vm_end,
296 floor, next? next->vm_start: ceiling);
297 }
298 vma = next;
299 }
300 }
302 int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
303 {
304 struct page *new = pte_alloc_one(mm, address);
305 if (!new)
306 return -ENOMEM;
308 pte_lock_init(new);
309 spin_lock(&mm->page_table_lock);
310 if (pmd_present(*pmd)) { /* Another has populated it */
311 pte_lock_deinit(new);
312 pte_free(new);
313 } else {
314 mm->nr_ptes++;
315 inc_page_state(nr_page_table_pages);
316 pmd_populate(mm, pmd, new);
317 }
318 spin_unlock(&mm->page_table_lock);
319 return 0;
320 }
322 int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
323 {
324 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
325 if (!new)
326 return -ENOMEM;
328 spin_lock(&init_mm.page_table_lock);
329 if (pmd_present(*pmd)) /* Another has populated it */
330 pte_free_kernel(new);
331 else
332 pmd_populate_kernel(&init_mm, pmd, new);
333 spin_unlock(&init_mm.page_table_lock);
334 return 0;
335 }
337 static inline void add_mm_rss(struct mm_struct *mm, int file_rss, int anon_rss)
338 {
339 if (file_rss)
340 add_mm_counter(mm, file_rss, file_rss);
341 if (anon_rss)
342 add_mm_counter(mm, anon_rss, anon_rss);
343 }
345 /*
346 * This function is called to print an error when a bad pte
347 * is found. For example, we might have a PFN-mapped pte in
348 * a region that doesn't allow it.
349 *
350 * The calling function must still handle the error.
351 */
352 void print_bad_pte(struct vm_area_struct *vma, pte_t pte, unsigned long vaddr)
353 {
354 printk(KERN_ERR "Bad pte = %08llx, process = %s, "
355 "vm_flags = %lx, vaddr = %lx\n",
356 (long long)pte_val(pte),
357 (vma->vm_mm == current->mm ? current->comm : "???"),
358 vma->vm_flags, vaddr);
359 dump_stack();
360 }
362 static inline int is_cow_mapping(unsigned int flags)
363 {
364 return (flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
365 }
367 /*
368 * This function gets the "struct page" associated with a pte.
369 *
370 * NOTE! Some mappings do not have "struct pages". A raw PFN mapping
371 * will have each page table entry just pointing to a raw page frame
372 * number, and as far as the VM layer is concerned, those do not have
373 * pages associated with them - even if the PFN might point to memory
374 * that otherwise is perfectly fine and has a "struct page".
375 *
376 * The way we recognize those mappings is through the rules set up
377 * by "remap_pfn_range()": the vma will have the VM_PFNMAP bit set,
378 * and the vm_pgoff will point to the first PFN mapped: thus every
379 * page that is a raw mapping will always honor the rule
380 *
381 * pfn_of_page == vma->vm_pgoff + ((addr - vma->vm_start) >> PAGE_SHIFT)
382 *
383 * and if that isn't true, the page has been COW'ed (in which case it
384 * _does_ have a "struct page" associated with it even if it is in a
385 * VM_PFNMAP range).
386 */
387 struct page *vm_normal_page(struct vm_area_struct *vma, unsigned long addr, pte_t pte)
388 {
389 unsigned long pfn = pte_pfn(pte);
391 if (vma->vm_flags & VM_PFNMAP) {
392 unsigned long off = (addr - vma->vm_start) >> PAGE_SHIFT;
393 if ((pfn == vma->vm_pgoff + off) || !pfn_valid(pfn))
394 return NULL;
395 if (!is_cow_mapping(vma->vm_flags))
396 return NULL;
397 }
399 /*
400 * Add some anal sanity checks for now. Eventually,
401 * we should just do "return pfn_to_page(pfn)", but
402 * in the meantime we check that we get a valid pfn,
403 * and that the resulting page looks ok.
404 *
405 * Remove this test eventually!
406 */
407 if (unlikely(!pfn_valid(pfn))) {
408 print_bad_pte(vma, pte, addr);
409 return NULL;
410 }
412 /*
413 * NOTE! We still have PageReserved() pages in the page
414 * tables.
415 *
416 * The PAGE_ZERO() pages and various VDSO mappings can
417 * cause them to exist.
418 */
419 return pfn_to_page(pfn);
420 }
422 /*
423 * copy one vm_area from one task to the other. Assumes the page tables
424 * already present in the new task to be cleared in the whole range
425 * covered by this vma.
426 */
428 static inline void
429 copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
430 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
431 unsigned long addr, int *rss)
432 {
433 unsigned long vm_flags = vma->vm_flags;
434 pte_t pte = *src_pte;
435 struct page *page;
437 /* pte contains position in swap or file, so copy. */
438 if (unlikely(!pte_present(pte))) {
439 if (!pte_file(pte)) {
440 swap_duplicate(pte_to_swp_entry(pte));
441 /* make sure dst_mm is on swapoff's mmlist. */
442 if (unlikely(list_empty(&dst_mm->mmlist))) {
443 spin_lock(&mmlist_lock);
444 if (list_empty(&dst_mm->mmlist))
445 list_add(&dst_mm->mmlist,
446 &src_mm->mmlist);
447 spin_unlock(&mmlist_lock);
448 }
449 }
450 goto out_set_pte;
451 }
453 /*
454 * If it's a COW mapping, write protect it both
455 * in the parent and the child
456 */
457 if (is_cow_mapping(vm_flags)) {
458 ptep_set_wrprotect(src_mm, addr, src_pte);
459 pte = *src_pte;
460 }
462 /*
463 * If it's a shared mapping, mark it clean in
464 * the child
465 */
466 if (vm_flags & VM_SHARED)
467 pte = pte_mkclean(pte);
468 pte = pte_mkold(pte);
470 page = vm_normal_page(vma, addr, pte);
471 if (page) {
472 get_page(page);
473 page_dup_rmap(page);
474 rss[!!PageAnon(page)]++;
475 }
477 out_set_pte:
478 set_pte_at(dst_mm, addr, dst_pte, pte);
479 }
481 static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
482 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
483 unsigned long addr, unsigned long end)
484 {
485 pte_t *src_pte, *dst_pte;
486 spinlock_t *src_ptl, *dst_ptl;
487 int progress = 0;
488 int rss[2];
490 again:
491 rss[1] = rss[0] = 0;
492 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
493 if (!dst_pte)
494 return -ENOMEM;
495 src_pte = pte_offset_map_nested(src_pmd, addr);
496 src_ptl = pte_lockptr(src_mm, src_pmd);
497 spin_lock(src_ptl);
499 do {
500 /*
501 * We are holding two locks at this point - either of them
502 * could generate latencies in another task on another CPU.
503 */
504 if (progress >= 32) {
505 progress = 0;
506 if (need_resched() ||
507 need_lockbreak(src_ptl) ||
508 need_lockbreak(dst_ptl))
509 break;
510 }
511 if (pte_none(*src_pte)) {
512 progress++;
513 continue;
514 }
515 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
516 progress += 8;
517 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
519 spin_unlock(src_ptl);
520 pte_unmap_nested(src_pte - 1);
521 add_mm_rss(dst_mm, rss[0], rss[1]);
522 pte_unmap_unlock(dst_pte - 1, dst_ptl);
523 cond_resched();
524 if (addr != end)
525 goto again;
526 return 0;
527 }
529 static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
530 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
531 unsigned long addr, unsigned long end)
532 {
533 pmd_t *src_pmd, *dst_pmd;
534 unsigned long next;
536 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
537 if (!dst_pmd)
538 return -ENOMEM;
539 src_pmd = pmd_offset(src_pud, addr);
540 do {
541 next = pmd_addr_end(addr, end);
542 if (pmd_none_or_clear_bad(src_pmd))
543 continue;
544 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
545 vma, addr, next))
546 return -ENOMEM;
547 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
548 return 0;
549 }
551 static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
552 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
553 unsigned long addr, unsigned long end)
554 {
555 pud_t *src_pud, *dst_pud;
556 unsigned long next;
558 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
559 if (!dst_pud)
560 return -ENOMEM;
561 src_pud = pud_offset(src_pgd, addr);
562 do {
563 next = pud_addr_end(addr, end);
564 if (pud_none_or_clear_bad(src_pud))
565 continue;
566 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
567 vma, addr, next))
568 return -ENOMEM;
569 } while (dst_pud++, src_pud++, addr = next, addr != end);
570 return 0;
571 }
573 int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
574 struct vm_area_struct *vma)
575 {
576 pgd_t *src_pgd, *dst_pgd;
577 unsigned long next;
578 unsigned long addr = vma->vm_start;
579 unsigned long end = vma->vm_end;
581 /*
582 * Don't copy ptes where a page fault will fill them correctly.
583 * Fork becomes much lighter when there are big shared or private
584 * readonly mappings. The tradeoff is that copy_page_range is more
585 * efficient than faulting.
586 */
587 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_PFNMAP|VM_INSERTPAGE))) {
588 if (!vma->anon_vma)
589 return 0;
590 }
592 if (is_vm_hugetlb_page(vma))
593 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
595 dst_pgd = pgd_offset(dst_mm, addr);
596 src_pgd = pgd_offset(src_mm, addr);
597 do {
598 next = pgd_addr_end(addr, end);
599 if (pgd_none_or_clear_bad(src_pgd))
600 continue;
601 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
602 vma, addr, next))
603 return -ENOMEM;
604 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
605 return 0;
606 }
608 static unsigned long zap_pte_range(struct mmu_gather *tlb,
609 struct vm_area_struct *vma, pmd_t *pmd,
610 unsigned long addr, unsigned long end,
611 long *zap_work, struct zap_details *details)
612 {
613 struct mm_struct *mm = tlb->mm;
614 pte_t *pte;
615 spinlock_t *ptl;
616 int file_rss = 0;
617 int anon_rss = 0;
619 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
620 do {
621 pte_t ptent = *pte;
622 if (pte_none(ptent)) {
623 (*zap_work)--;
624 continue;
625 }
627 (*zap_work) -= PAGE_SIZE;
629 if (pte_present(ptent)) {
630 struct page *page;
632 page = vm_normal_page(vma, addr, ptent);
633 if (unlikely(details) && page) {
634 /*
635 * unmap_shared_mapping_pages() wants to
636 * invalidate cache without truncating:
637 * unmap shared but keep private pages.
638 */
639 if (details->check_mapping &&
640 details->check_mapping != page->mapping)
641 continue;
642 /*
643 * Each page->index must be checked when
644 * invalidating or truncating nonlinear.
645 */
646 if (details->nonlinear_vma &&
647 (page->index < details->first_index ||
648 page->index > details->last_index))
649 continue;
650 }
651 ptent = ptep_get_and_clear_full(mm, addr, pte,
652 tlb->fullmm);
653 tlb_remove_tlb_entry(tlb, pte, addr);
654 if (unlikely(!page))
655 continue;
656 if (unlikely(details) && details->nonlinear_vma
657 && linear_page_index(details->nonlinear_vma,
658 addr) != page->index)
659 set_pte_at(mm, addr, pte,
660 pgoff_to_pte(page->index));
661 if (PageAnon(page))
662 anon_rss--;
663 else {
664 if (pte_dirty(ptent))
665 set_page_dirty(page);
666 if (pte_young(ptent))
667 mark_page_accessed(page);
668 file_rss--;
669 }
670 page_remove_rmap(page);
671 tlb_remove_page(tlb, page);
672 continue;
673 }
674 /*
675 * If details->check_mapping, we leave swap entries;
676 * if details->nonlinear_vma, we leave file entries.
677 */
678 if (unlikely(details))
679 continue;
680 if (!pte_file(ptent))
681 free_swap_and_cache(pte_to_swp_entry(ptent));
682 pte_clear_full(mm, addr, pte, tlb->fullmm);
683 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
685 add_mm_rss(mm, file_rss, anon_rss);
686 pte_unmap_unlock(pte - 1, ptl);
688 return addr;
689 }
691 static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
692 struct vm_area_struct *vma, pud_t *pud,
693 unsigned long addr, unsigned long end,
694 long *zap_work, struct zap_details *details)
695 {
696 pmd_t *pmd;
697 unsigned long next;
699 pmd = pmd_offset(pud, addr);
700 do {
701 next = pmd_addr_end(addr, end);
702 if (pmd_none_or_clear_bad(pmd)) {
703 (*zap_work)--;
704 continue;
705 }
706 next = zap_pte_range(tlb, vma, pmd, addr, next,
707 zap_work, details);
708 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
710 return addr;
711 }
713 static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
714 struct vm_area_struct *vma, pgd_t *pgd,
715 unsigned long addr, unsigned long end,
716 long *zap_work, struct zap_details *details)
717 {
718 pud_t *pud;
719 unsigned long next;
721 pud = pud_offset(pgd, addr);
722 do {
723 next = pud_addr_end(addr, end);
724 if (pud_none_or_clear_bad(pud)) {
725 (*zap_work)--;
726 continue;
727 }
728 next = zap_pmd_range(tlb, vma, pud, addr, next,
729 zap_work, details);
730 } while (pud++, addr = next, (addr != end && *zap_work > 0));
732 return addr;
733 }
735 static unsigned long unmap_page_range(struct mmu_gather *tlb,
736 struct vm_area_struct *vma,
737 unsigned long addr, unsigned long end,
738 long *zap_work, struct zap_details *details)
739 {
740 pgd_t *pgd;
741 unsigned long next;
743 if (details && !details->check_mapping && !details->nonlinear_vma)
744 details = NULL;
746 BUG_ON(addr >= end);
747 tlb_start_vma(tlb, vma);
748 pgd = pgd_offset(vma->vm_mm, addr);
749 do {
750 next = pgd_addr_end(addr, end);
751 if (pgd_none_or_clear_bad(pgd)) {
752 (*zap_work)--;
753 continue;
754 }
755 next = zap_pud_range(tlb, vma, pgd, addr, next,
756 zap_work, details);
757 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
758 tlb_end_vma(tlb, vma);
760 return addr;
761 }
763 #ifdef CONFIG_PREEMPT
764 # define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
765 #else
766 /* No preempt: go for improved straight-line efficiency */
767 # define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
768 #endif
770 /**
771 * unmap_vmas - unmap a range of memory covered by a list of vma's
772 * @tlbp: address of the caller's struct mmu_gather
773 * @vma: the starting vma
774 * @start_addr: virtual address at which to start unmapping
775 * @end_addr: virtual address at which to end unmapping
776 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
777 * @details: details of nonlinear truncation or shared cache invalidation
778 *
779 * Returns the end address of the unmapping (restart addr if interrupted).
780 *
781 * Unmap all pages in the vma list.
782 *
783 * We aim to not hold locks for too long (for scheduling latency reasons).
784 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
785 * return the ending mmu_gather to the caller.
786 *
787 * Only addresses between `start' and `end' will be unmapped.
788 *
789 * The VMA list must be sorted in ascending virtual address order.
790 *
791 * unmap_vmas() assumes that the caller will flush the whole unmapped address
792 * range after unmap_vmas() returns. So the only responsibility here is to
793 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
794 * drops the lock and schedules.
795 */
796 unsigned long unmap_vmas(struct mmu_gather **tlbp,
797 struct vm_area_struct *vma, unsigned long start_addr,
798 unsigned long end_addr, unsigned long *nr_accounted,
799 struct zap_details *details)
800 {
801 long zap_work = ZAP_BLOCK_SIZE;
802 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
803 int tlb_start_valid = 0;
804 unsigned long start = start_addr;
805 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
806 int fullmm = (*tlbp)->fullmm;
808 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
809 unsigned long end;
811 start = max(vma->vm_start, start_addr);
812 if (start >= vma->vm_end)
813 continue;
814 end = min(vma->vm_end, end_addr);
815 if (end <= vma->vm_start)
816 continue;
818 if (vma->vm_flags & VM_ACCOUNT)
819 *nr_accounted += (end - start) >> PAGE_SHIFT;
821 while (start != end) {
822 if (!tlb_start_valid) {
823 tlb_start = start;
824 tlb_start_valid = 1;
825 }
827 if (unlikely(is_vm_hugetlb_page(vma))) {
828 unmap_hugepage_range(vma, start, end);
829 zap_work -= (end - start) /
830 (HPAGE_SIZE / PAGE_SIZE);
831 start = end;
832 } else
833 start = unmap_page_range(*tlbp, vma,
834 start, end, &zap_work, details);
836 if (zap_work > 0) {
837 BUG_ON(start != end);
838 break;
839 }
841 tlb_finish_mmu(*tlbp, tlb_start, start);
843 if (need_resched() ||
844 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
845 if (i_mmap_lock) {
846 *tlbp = NULL;
847 goto out;
848 }
849 cond_resched();
850 }
852 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
853 tlb_start_valid = 0;
854 zap_work = ZAP_BLOCK_SIZE;
855 }
856 }
857 out:
858 return start; /* which is now the end (or restart) address */
859 }
861 /**
862 * zap_page_range - remove user pages in a given range
863 * @vma: vm_area_struct holding the applicable pages
864 * @address: starting address of pages to zap
865 * @size: number of bytes to zap
866 * @details: details of nonlinear truncation or shared cache invalidation
867 */
868 unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
869 unsigned long size, struct zap_details *details)
870 {
871 struct mm_struct *mm = vma->vm_mm;
872 struct mmu_gather *tlb;
873 unsigned long end = address + size;
874 unsigned long nr_accounted = 0;
876 lru_add_drain();
877 tlb = tlb_gather_mmu(mm, 0);
878 update_hiwater_rss(mm);
879 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
880 if (tlb)
881 tlb_finish_mmu(tlb, address, end);
882 return end;
883 }
885 /*
886 * Do a quick page-table lookup for a single page.
887 */
888 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
889 unsigned int flags)
890 {
891 pgd_t *pgd;
892 pud_t *pud;
893 pmd_t *pmd;
894 pte_t *ptep, pte;
895 spinlock_t *ptl;
896 struct page *page;
897 struct mm_struct *mm = vma->vm_mm;
899 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
900 if (!IS_ERR(page)) {
901 BUG_ON(flags & FOLL_GET);
902 goto out;
903 }
905 page = NULL;
906 pgd = pgd_offset(mm, address);
907 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
908 goto no_page_table;
910 pud = pud_offset(pgd, address);
911 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
912 goto no_page_table;
914 pmd = pmd_offset(pud, address);
915 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
916 goto no_page_table;
918 if (pmd_huge(*pmd)) {
919 BUG_ON(flags & FOLL_GET);
920 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
921 goto out;
922 }
924 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
925 if (!ptep)
926 goto out;
928 pte = *ptep;
929 if (!pte_present(pte))
930 goto unlock;
931 if ((flags & FOLL_WRITE) && !pte_write(pte))
932 goto unlock;
933 page = vm_normal_page(vma, address, pte);
934 if (unlikely(!page))
935 goto unlock;
937 if (flags & FOLL_GET)
938 get_page(page);
939 if (flags & FOLL_TOUCH) {
940 if ((flags & FOLL_WRITE) &&
941 !pte_dirty(pte) && !PageDirty(page))
942 set_page_dirty(page);
943 mark_page_accessed(page);
944 }
945 unlock:
946 pte_unmap_unlock(ptep, ptl);
947 out:
948 return page;
950 no_page_table:
951 /*
952 * When core dumping an enormous anonymous area that nobody
953 * has touched so far, we don't want to allocate page tables.
954 */
955 if (flags & FOLL_ANON) {
956 page = ZERO_PAGE(address);
957 if (flags & FOLL_GET)
958 get_page(page);
959 BUG_ON(flags & FOLL_WRITE);
960 }
961 return page;
962 }
964 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
965 unsigned long start, int len, int write, int force,
966 struct page **pages, struct vm_area_struct **vmas)
967 {
968 int i;
969 unsigned int vm_flags;
971 /*
972 * Require read or write permissions.
973 * If 'force' is set, we only require the "MAY" flags.
974 */
975 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
976 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
977 i = 0;
979 do {
980 struct vm_area_struct *vma;
981 unsigned int foll_flags;
983 vma = find_extend_vma(mm, start);
984 if (!vma && in_gate_area(tsk, start)) {
985 unsigned long pg = start & PAGE_MASK;
986 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
987 pgd_t *pgd;
988 pud_t *pud;
989 pmd_t *pmd;
990 pte_t *pte;
991 if (write) /* user gate pages are read-only */
992 return i ? : -EFAULT;
993 if (pg > TASK_SIZE)
994 pgd = pgd_offset_k(pg);
995 else
996 pgd = pgd_offset_gate(mm, pg);
997 BUG_ON(pgd_none(*pgd));
998 pud = pud_offset(pgd, pg);
999 BUG_ON(pud_none(*pud));
1000 pmd = pmd_offset(pud, pg);
1001 if (pmd_none(*pmd))
1002 return i ? : -EFAULT;
1003 pte = pte_offset_map(pmd, pg);
1004 if (pte_none(*pte)) {
1005 pte_unmap(pte);
1006 return i ? : -EFAULT;
1008 if (pages) {
1009 struct page *page = vm_normal_page(gate_vma, start, *pte);
1010 pages[i] = page;
1011 if (page)
1012 get_page(page);
1014 pte_unmap(pte);
1015 if (vmas)
1016 vmas[i] = gate_vma;
1017 i++;
1018 start += PAGE_SIZE;
1019 len--;
1020 continue;
1023 #ifdef CONFIG_XEN
1024 if (vma && (vma->vm_flags & VM_FOREIGN)) {
1025 struct page **map = vma->vm_private_data;
1026 int offset = (start - vma->vm_start) >> PAGE_SHIFT;
1027 if (map[offset] != NULL) {
1028 if (pages) {
1029 struct page *page = map[offset];
1031 pages[i] = page;
1032 get_page(page);
1034 if (vmas)
1035 vmas[i] = vma;
1036 i++;
1037 start += PAGE_SIZE;
1038 len--;
1039 continue;
1042 #endif
1043 if (!vma || (vma->vm_flags & (VM_IO | VM_PFNMAP))
1044 || !(vm_flags & vma->vm_flags))
1045 return i ? : -EFAULT;
1047 if (is_vm_hugetlb_page(vma)) {
1048 i = follow_hugetlb_page(mm, vma, pages, vmas,
1049 &start, &len, i);
1050 continue;
1053 foll_flags = FOLL_TOUCH;
1054 if (pages)
1055 foll_flags |= FOLL_GET;
1056 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1057 (!vma->vm_ops || !vma->vm_ops->nopage))
1058 foll_flags |= FOLL_ANON;
1060 do {
1061 struct page *page;
1063 if (write)
1064 foll_flags |= FOLL_WRITE;
1066 cond_resched();
1067 while (!(page = follow_page(vma, start, foll_flags))) {
1068 int ret;
1069 ret = __handle_mm_fault(mm, vma, start,
1070 foll_flags & FOLL_WRITE);
1071 /*
1072 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1073 * broken COW when necessary, even if maybe_mkwrite
1074 * decided not to set pte_write. We can thus safely do
1075 * subsequent page lookups as if they were reads.
1076 */
1077 if (ret & VM_FAULT_WRITE)
1078 foll_flags &= ~FOLL_WRITE;
1080 switch (ret & ~VM_FAULT_WRITE) {
1081 case VM_FAULT_MINOR:
1082 tsk->min_flt++;
1083 break;
1084 case VM_FAULT_MAJOR:
1085 tsk->maj_flt++;
1086 break;
1087 case VM_FAULT_SIGBUS:
1088 return i ? i : -EFAULT;
1089 case VM_FAULT_OOM:
1090 return i ? i : -ENOMEM;
1091 default:
1092 BUG();
1095 if (pages) {
1096 pages[i] = page;
1097 flush_dcache_page(page);
1099 if (vmas)
1100 vmas[i] = vma;
1101 i++;
1102 start += PAGE_SIZE;
1103 len--;
1104 } while (len && start < vma->vm_end);
1105 } while (len);
1106 return i;
1108 EXPORT_SYMBOL(get_user_pages);
1110 static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1111 unsigned long addr, unsigned long end, pgprot_t prot)
1113 pte_t *pte;
1114 spinlock_t *ptl;
1116 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1117 if (!pte)
1118 return -ENOMEM;
1119 do {
1120 struct page *page = ZERO_PAGE(addr);
1121 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1122 page_cache_get(page);
1123 page_add_file_rmap(page);
1124 inc_mm_counter(mm, file_rss);
1125 BUG_ON(!pte_none(*pte));
1126 set_pte_at(mm, addr, pte, zero_pte);
1127 } while (pte++, addr += PAGE_SIZE, addr != end);
1128 pte_unmap_unlock(pte - 1, ptl);
1129 return 0;
1132 static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1133 unsigned long addr, unsigned long end, pgprot_t prot)
1135 pmd_t *pmd;
1136 unsigned long next;
1138 pmd = pmd_alloc(mm, pud, addr);
1139 if (!pmd)
1140 return -ENOMEM;
1141 do {
1142 next = pmd_addr_end(addr, end);
1143 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1144 return -ENOMEM;
1145 } while (pmd++, addr = next, addr != end);
1146 return 0;
1149 static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1150 unsigned long addr, unsigned long end, pgprot_t prot)
1152 pud_t *pud;
1153 unsigned long next;
1155 pud = pud_alloc(mm, pgd, addr);
1156 if (!pud)
1157 return -ENOMEM;
1158 do {
1159 next = pud_addr_end(addr, end);
1160 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1161 return -ENOMEM;
1162 } while (pud++, addr = next, addr != end);
1163 return 0;
1166 int zeromap_page_range(struct vm_area_struct *vma,
1167 unsigned long addr, unsigned long size, pgprot_t prot)
1169 pgd_t *pgd;
1170 unsigned long next;
1171 unsigned long end = addr + size;
1172 struct mm_struct *mm = vma->vm_mm;
1173 int err;
1175 BUG_ON(addr >= end);
1176 pgd = pgd_offset(mm, addr);
1177 flush_cache_range(vma, addr, end);
1178 do {
1179 next = pgd_addr_end(addr, end);
1180 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1181 if (err)
1182 break;
1183 } while (pgd++, addr = next, addr != end);
1184 return err;
1187 pte_t * fastcall get_locked_pte(struct mm_struct *mm, unsigned long addr, spinlock_t **ptl)
1189 pgd_t * pgd = pgd_offset(mm, addr);
1190 pud_t * pud = pud_alloc(mm, pgd, addr);
1191 if (pud) {
1192 pmd_t * pmd = pmd_alloc(mm, pud, addr);
1193 if (pmd)
1194 return pte_alloc_map_lock(mm, pmd, addr, ptl);
1196 return NULL;
1199 /*
1200 * This is the old fallback for page remapping.
1202 * For historical reasons, it only allows reserved pages. Only
1203 * old drivers should use this, and they needed to mark their
1204 * pages reserved for the old functions anyway.
1205 */
1206 static int insert_page(struct mm_struct *mm, unsigned long addr, struct page *page, pgprot_t prot)
1208 int retval;
1209 pte_t *pte;
1210 spinlock_t *ptl;
1212 retval = -EINVAL;
1213 if (PageAnon(page))
1214 goto out;
1215 retval = -ENOMEM;
1216 flush_dcache_page(page);
1217 pte = get_locked_pte(mm, addr, &ptl);
1218 if (!pte)
1219 goto out;
1220 retval = -EBUSY;
1221 if (!pte_none(*pte))
1222 goto out_unlock;
1224 /* Ok, finally just insert the thing.. */
1225 get_page(page);
1226 inc_mm_counter(mm, file_rss);
1227 page_add_file_rmap(page);
1228 set_pte_at(mm, addr, pte, mk_pte(page, prot));
1230 retval = 0;
1231 out_unlock:
1232 pte_unmap_unlock(pte, ptl);
1233 out:
1234 return retval;
1237 /*
1238 * This allows drivers to insert individual pages they've allocated
1239 * into a user vma.
1241 * The page has to be a nice clean _individual_ kernel allocation.
1242 * If you allocate a compound page, you need to have marked it as
1243 * such (__GFP_COMP), or manually just split the page up yourself
1244 * (which is mainly an issue of doing "set_page_count(page, 1)" for
1245 * each sub-page, and then freeing them one by one when you free
1246 * them rather than freeing it as a compound page).
1248 * NOTE! Traditionally this was done with "remap_pfn_range()" which
1249 * took an arbitrary page protection parameter. This doesn't allow
1250 * that. Your vma protection will have to be set up correctly, which
1251 * means that if you want a shared writable mapping, you'd better
1252 * ask for a shared writable mapping!
1254 * The page does not need to be reserved.
1255 */
1256 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr, struct page *page)
1258 if (addr < vma->vm_start || addr >= vma->vm_end)
1259 return -EFAULT;
1260 if (!page_count(page))
1261 return -EINVAL;
1262 vma->vm_flags |= VM_INSERTPAGE;
1263 return insert_page(vma->vm_mm, addr, page, vma->vm_page_prot);
1265 EXPORT_SYMBOL(vm_insert_page);
1267 /*
1268 * maps a range of physical memory into the requested pages. the old
1269 * mappings are removed. any references to nonexistent pages results
1270 * in null mappings (currently treated as "copy-on-access")
1271 */
1272 static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1273 unsigned long addr, unsigned long end,
1274 unsigned long pfn, pgprot_t prot)
1276 pte_t *pte;
1277 spinlock_t *ptl;
1279 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
1280 if (!pte)
1281 return -ENOMEM;
1282 do {
1283 BUG_ON(!pte_none(*pte));
1284 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
1285 pfn++;
1286 } while (pte++, addr += PAGE_SIZE, addr != end);
1287 pte_unmap_unlock(pte - 1, ptl);
1288 return 0;
1291 static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1292 unsigned long addr, unsigned long end,
1293 unsigned long pfn, pgprot_t prot)
1295 pmd_t *pmd;
1296 unsigned long next;
1298 pfn -= addr >> PAGE_SHIFT;
1299 pmd = pmd_alloc(mm, pud, addr);
1300 if (!pmd)
1301 return -ENOMEM;
1302 do {
1303 next = pmd_addr_end(addr, end);
1304 if (remap_pte_range(mm, pmd, addr, next,
1305 pfn + (addr >> PAGE_SHIFT), prot))
1306 return -ENOMEM;
1307 } while (pmd++, addr = next, addr != end);
1308 return 0;
1311 static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1312 unsigned long addr, unsigned long end,
1313 unsigned long pfn, pgprot_t prot)
1315 pud_t *pud;
1316 unsigned long next;
1318 pfn -= addr >> PAGE_SHIFT;
1319 pud = pud_alloc(mm, pgd, addr);
1320 if (!pud)
1321 return -ENOMEM;
1322 do {
1323 next = pud_addr_end(addr, end);
1324 if (remap_pmd_range(mm, pud, addr, next,
1325 pfn + (addr >> PAGE_SHIFT), prot))
1326 return -ENOMEM;
1327 } while (pud++, addr = next, addr != end);
1328 return 0;
1331 /* Note: this is only safe if the mm semaphore is held when called. */
1332 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1333 unsigned long pfn, unsigned long size, pgprot_t prot)
1335 pgd_t *pgd;
1336 unsigned long next;
1337 unsigned long end = addr + PAGE_ALIGN(size);
1338 struct mm_struct *mm = vma->vm_mm;
1339 int err;
1341 /*
1342 * Physically remapped pages are special. Tell the
1343 * rest of the world about it:
1344 * VM_IO tells people not to look at these pages
1345 * (accesses can have side effects).
1346 * VM_RESERVED is specified all over the place, because
1347 * in 2.4 it kept swapout's vma scan off this vma; but
1348 * in 2.6 the LRU scan won't even find its pages, so this
1349 * flag means no more than count its pages in reserved_vm,
1350 * and omit it from core dump, even when VM_IO turned off.
1351 * VM_PFNMAP tells the core MM that the base pages are just
1352 * raw PFN mappings, and do not have a "struct page" associated
1353 * with them.
1355 * There's a horrible special case to handle copy-on-write
1356 * behaviour that some programs depend on. We mark the "original"
1357 * un-COW'ed pages by matching them up with "vma->vm_pgoff".
1358 */
1359 if (is_cow_mapping(vma->vm_flags)) {
1360 if (addr != vma->vm_start || end != vma->vm_end)
1361 return -EINVAL;
1362 vma->vm_pgoff = pfn;
1365 vma->vm_flags |= VM_IO | VM_RESERVED | VM_PFNMAP;
1367 BUG_ON(addr >= end);
1368 pfn -= addr >> PAGE_SHIFT;
1369 pgd = pgd_offset(mm, addr);
1370 flush_cache_range(vma, addr, end);
1371 do {
1372 next = pgd_addr_end(addr, end);
1373 err = remap_pud_range(mm, pgd, addr, next,
1374 pfn + (addr >> PAGE_SHIFT), prot);
1375 if (err)
1376 break;
1377 } while (pgd++, addr = next, addr != end);
1378 return err;
1380 EXPORT_SYMBOL(remap_pfn_range);
1382 #ifdef CONFIG_XEN
1383 static inline int apply_to_pte_range(struct mm_struct *mm, pmd_t *pmd,
1384 unsigned long addr, unsigned long end,
1385 pte_fn_t fn, void *data)
1387 pte_t *pte;
1388 int err;
1389 struct page *pmd_page;
1390 spinlock_t *ptl;
1392 pte = (mm == &init_mm) ?
1393 pte_alloc_kernel(pmd, addr) :
1394 pte_alloc_map_lock(mm, pmd, addr, &ptl);
1395 if (!pte)
1396 return -ENOMEM;
1398 BUG_ON(pmd_huge(*pmd));
1400 pmd_page = pmd_page(*pmd);
1402 do {
1403 err = fn(pte, pmd_page, addr, data);
1404 if (err)
1405 break;
1406 } while (pte++, addr += PAGE_SIZE, addr != end);
1408 if (mm != &init_mm)
1409 pte_unmap_unlock(pte-1, ptl);
1410 return err;
1413 static inline int apply_to_pmd_range(struct mm_struct *mm, pud_t *pud,
1414 unsigned long addr, unsigned long end,
1415 pte_fn_t fn, void *data)
1417 pmd_t *pmd;
1418 unsigned long next;
1419 int err;
1421 pmd = pmd_alloc(mm, pud, addr);
1422 if (!pmd)
1423 return -ENOMEM;
1424 do {
1425 next = pmd_addr_end(addr, end);
1426 err = apply_to_pte_range(mm, pmd, addr, next, fn, data);
1427 if (err)
1428 break;
1429 } while (pmd++, addr = next, addr != end);
1430 return err;
1433 static inline int apply_to_pud_range(struct mm_struct *mm, pgd_t *pgd,
1434 unsigned long addr, unsigned long end,
1435 pte_fn_t fn, void *data)
1437 pud_t *pud;
1438 unsigned long next;
1439 int err;
1441 pud = pud_alloc(mm, pgd, addr);
1442 if (!pud)
1443 return -ENOMEM;
1444 do {
1445 next = pud_addr_end(addr, end);
1446 err = apply_to_pmd_range(mm, pud, addr, next, fn, data);
1447 if (err)
1448 break;
1449 } while (pud++, addr = next, addr != end);
1450 return err;
1453 /*
1454 * Scan a region of virtual memory, filling in page tables as necessary
1455 * and calling a provided function on each leaf page table.
1456 */
1457 int apply_to_page_range(struct mm_struct *mm, unsigned long addr,
1458 unsigned long size, pte_fn_t fn, void *data)
1460 pgd_t *pgd;
1461 unsigned long next;
1462 unsigned long end = addr + size;
1463 int err;
1465 BUG_ON(addr >= end);
1466 pgd = pgd_offset(mm, addr);
1467 do {
1468 next = pgd_addr_end(addr, end);
1469 err = apply_to_pud_range(mm, pgd, addr, next, fn, data);
1470 if (err)
1471 break;
1472 } while (pgd++, addr = next, addr != end);
1473 return err;
1475 EXPORT_SYMBOL_GPL(apply_to_page_range);
1476 #endif
1478 /*
1479 * handle_pte_fault chooses page fault handler according to an entry
1480 * which was read non-atomically. Before making any commitment, on
1481 * those architectures or configurations (e.g. i386 with PAE) which
1482 * might give a mix of unmatched parts, do_swap_page and do_file_page
1483 * must check under lock before unmapping the pte and proceeding
1484 * (but do_wp_page is only called after already making such a check;
1485 * and do_anonymous_page and do_no_page can safely check later on).
1486 */
1487 static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
1488 pte_t *page_table, pte_t orig_pte)
1490 int same = 1;
1491 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1492 if (sizeof(pte_t) > sizeof(unsigned long)) {
1493 spinlock_t *ptl = pte_lockptr(mm, pmd);
1494 spin_lock(ptl);
1495 same = pte_same(*page_table, orig_pte);
1496 spin_unlock(ptl);
1498 #endif
1499 pte_unmap(page_table);
1500 return same;
1503 /*
1504 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1505 * servicing faults for write access. In the normal case, do always want
1506 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1507 * that do not have writing enabled, when used by access_process_vm.
1508 */
1509 static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1511 if (likely(vma->vm_flags & VM_WRITE))
1512 pte = pte_mkwrite(pte);
1513 return pte;
1516 static inline void cow_user_page(struct page *dst, struct page *src, unsigned long va)
1518 /*
1519 * If the source page was a PFN mapping, we don't have
1520 * a "struct page" for it. We do a best-effort copy by
1521 * just copying from the original user address. If that
1522 * fails, we just zero-fill it. Live with it.
1523 */
1524 if (unlikely(!src)) {
1525 void *kaddr = kmap_atomic(dst, KM_USER0);
1526 void __user *uaddr = (void __user *)(va & PAGE_MASK);
1528 /*
1529 * This really shouldn't fail, because the page is there
1530 * in the page tables. But it might just be unreadable,
1531 * in which case we just give up and fill the result with
1532 * zeroes.
1533 */
1534 if (__copy_from_user_inatomic(kaddr, uaddr, PAGE_SIZE))
1535 memset(kaddr, 0, PAGE_SIZE);
1536 kunmap_atomic(kaddr, KM_USER0);
1537 return;
1540 copy_user_highpage(dst, src, va);
1543 /*
1544 * This routine handles present pages, when users try to write
1545 * to a shared page. It is done by copying the page to a new address
1546 * and decrementing the shared-page counter for the old page.
1548 * Note that this routine assumes that the protection checks have been
1549 * done by the caller (the low-level page fault routine in most cases).
1550 * Thus we can safely just mark it writable once we've done any necessary
1551 * COW.
1553 * We also mark the page dirty at this point even though the page will
1554 * change only once the write actually happens. This avoids a few races,
1555 * and potentially makes it more efficient.
1557 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1558 * but allow concurrent faults), with pte both mapped and locked.
1559 * We return with mmap_sem still held, but pte unmapped and unlocked.
1560 */
1561 static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1562 unsigned long address, pte_t *page_table, pmd_t *pmd,
1563 spinlock_t *ptl, pte_t orig_pte)
1565 struct page *old_page, *new_page;
1566 pte_t entry;
1567 int ret = VM_FAULT_MINOR;
1569 old_page = vm_normal_page(vma, address, orig_pte);
1570 if (!old_page)
1571 goto gotten;
1573 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
1574 int reuse = can_share_swap_page(old_page);
1575 unlock_page(old_page);
1576 if (reuse) {
1577 flush_cache_page(vma, address, pte_pfn(orig_pte));
1578 entry = pte_mkyoung(orig_pte);
1579 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1580 ptep_set_access_flags(vma, address, page_table, entry, 1);
1581 update_mmu_cache(vma, address, entry);
1582 lazy_mmu_prot_update(entry);
1583 ret |= VM_FAULT_WRITE;
1584 goto unlock;
1588 /*
1589 * Ok, we need to copy. Oh, well..
1590 */
1591 page_cache_get(old_page);
1592 gotten:
1593 pte_unmap_unlock(page_table, ptl);
1595 if (unlikely(anon_vma_prepare(vma)))
1596 goto oom;
1597 if (old_page == ZERO_PAGE(address)) {
1598 new_page = alloc_zeroed_user_highpage(vma, address);
1599 if (!new_page)
1600 goto oom;
1601 } else {
1602 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1603 if (!new_page)
1604 goto oom;
1605 cow_user_page(new_page, old_page, address);
1608 /*
1609 * Re-check the pte - we dropped the lock
1610 */
1611 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1612 if (likely(pte_same(*page_table, orig_pte))) {
1613 if (old_page) {
1614 page_remove_rmap(old_page);
1615 if (!PageAnon(old_page)) {
1616 dec_mm_counter(mm, file_rss);
1617 inc_mm_counter(mm, anon_rss);
1619 } else
1620 inc_mm_counter(mm, anon_rss);
1621 flush_cache_page(vma, address, pte_pfn(orig_pte));
1622 entry = mk_pte(new_page, vma->vm_page_prot);
1623 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1624 ptep_establish(vma, address, page_table, entry);
1625 update_mmu_cache(vma, address, entry);
1626 lazy_mmu_prot_update(entry);
1627 lru_cache_add_active(new_page);
1628 page_add_new_anon_rmap(new_page, vma, address);
1630 /* Free the old page.. */
1631 new_page = old_page;
1632 ret |= VM_FAULT_WRITE;
1634 if (new_page)
1635 page_cache_release(new_page);
1636 if (old_page)
1637 page_cache_release(old_page);
1638 unlock:
1639 pte_unmap_unlock(page_table, ptl);
1640 return ret;
1641 oom:
1642 if (old_page)
1643 page_cache_release(old_page);
1644 return VM_FAULT_OOM;
1647 /*
1648 * Helper functions for unmap_mapping_range().
1650 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1652 * We have to restart searching the prio_tree whenever we drop the lock,
1653 * since the iterator is only valid while the lock is held, and anyway
1654 * a later vma might be split and reinserted earlier while lock dropped.
1656 * The list of nonlinear vmas could be handled more efficiently, using
1657 * a placeholder, but handle it in the same way until a need is shown.
1658 * It is important to search the prio_tree before nonlinear list: a vma
1659 * may become nonlinear and be shifted from prio_tree to nonlinear list
1660 * while the lock is dropped; but never shifted from list to prio_tree.
1662 * In order to make forward progress despite restarting the search,
1663 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1664 * quickly skip it next time around. Since the prio_tree search only
1665 * shows us those vmas affected by unmapping the range in question, we
1666 * can't efficiently keep all vmas in step with mapping->truncate_count:
1667 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1668 * mapping->truncate_count and vma->vm_truncate_count are protected by
1669 * i_mmap_lock.
1671 * In order to make forward progress despite repeatedly restarting some
1672 * large vma, note the restart_addr from unmap_vmas when it breaks out:
1673 * and restart from that address when we reach that vma again. It might
1674 * have been split or merged, shrunk or extended, but never shifted: so
1675 * restart_addr remains valid so long as it remains in the vma's range.
1676 * unmap_mapping_range forces truncate_count to leap over page-aligned
1677 * values so we can save vma's restart_addr in its truncate_count field.
1678 */
1679 #define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1681 static void reset_vma_truncate_counts(struct address_space *mapping)
1683 struct vm_area_struct *vma;
1684 struct prio_tree_iter iter;
1686 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1687 vma->vm_truncate_count = 0;
1688 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1689 vma->vm_truncate_count = 0;
1692 static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1693 unsigned long start_addr, unsigned long end_addr,
1694 struct zap_details *details)
1696 unsigned long restart_addr;
1697 int need_break;
1699 again:
1700 restart_addr = vma->vm_truncate_count;
1701 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1702 start_addr = restart_addr;
1703 if (start_addr >= end_addr) {
1704 /* Top of vma has been split off since last time */
1705 vma->vm_truncate_count = details->truncate_count;
1706 return 0;
1710 restart_addr = zap_page_range(vma, start_addr,
1711 end_addr - start_addr, details);
1712 need_break = need_resched() ||
1713 need_lockbreak(details->i_mmap_lock);
1715 if (restart_addr >= end_addr) {
1716 /* We have now completed this vma: mark it so */
1717 vma->vm_truncate_count = details->truncate_count;
1718 if (!need_break)
1719 return 0;
1720 } else {
1721 /* Note restart_addr in vma's truncate_count field */
1722 vma->vm_truncate_count = restart_addr;
1723 if (!need_break)
1724 goto again;
1727 spin_unlock(details->i_mmap_lock);
1728 cond_resched();
1729 spin_lock(details->i_mmap_lock);
1730 return -EINTR;
1733 static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1734 struct zap_details *details)
1736 struct vm_area_struct *vma;
1737 struct prio_tree_iter iter;
1738 pgoff_t vba, vea, zba, zea;
1740 restart:
1741 vma_prio_tree_foreach(vma, &iter, root,
1742 details->first_index, details->last_index) {
1743 /* Skip quickly over those we have already dealt with */
1744 if (vma->vm_truncate_count == details->truncate_count)
1745 continue;
1747 vba = vma->vm_pgoff;
1748 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1749 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1750 zba = details->first_index;
1751 if (zba < vba)
1752 zba = vba;
1753 zea = details->last_index;
1754 if (zea > vea)
1755 zea = vea;
1757 if (unmap_mapping_range_vma(vma,
1758 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1759 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1760 details) < 0)
1761 goto restart;
1765 static inline void unmap_mapping_range_list(struct list_head *head,
1766 struct zap_details *details)
1768 struct vm_area_struct *vma;
1770 /*
1771 * In nonlinear VMAs there is no correspondence between virtual address
1772 * offset and file offset. So we must perform an exhaustive search
1773 * across *all* the pages in each nonlinear VMA, not just the pages
1774 * whose virtual address lies outside the file truncation point.
1775 */
1776 restart:
1777 list_for_each_entry(vma, head, shared.vm_set.list) {
1778 /* Skip quickly over those we have already dealt with */
1779 if (vma->vm_truncate_count == details->truncate_count)
1780 continue;
1781 details->nonlinear_vma = vma;
1782 if (unmap_mapping_range_vma(vma, vma->vm_start,
1783 vma->vm_end, details) < 0)
1784 goto restart;
1788 /**
1789 * unmap_mapping_range - unmap the portion of all mmaps
1790 * in the specified address_space corresponding to the specified
1791 * page range in the underlying file.
1792 * @mapping: the address space containing mmaps to be unmapped.
1793 * @holebegin: byte in first page to unmap, relative to the start of
1794 * the underlying file. This will be rounded down to a PAGE_SIZE
1795 * boundary. Note that this is different from vmtruncate(), which
1796 * must keep the partial page. In contrast, we must get rid of
1797 * partial pages.
1798 * @holelen: size of prospective hole in bytes. This will be rounded
1799 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1800 * end of the file.
1801 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1802 * but 0 when invalidating pagecache, don't throw away private data.
1803 */
1804 void unmap_mapping_range(struct address_space *mapping,
1805 loff_t const holebegin, loff_t const holelen, int even_cows)
1807 struct zap_details details;
1808 pgoff_t hba = holebegin >> PAGE_SHIFT;
1809 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1811 /* Check for overflow. */
1812 if (sizeof(holelen) > sizeof(hlen)) {
1813 long long holeend =
1814 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1815 if (holeend & ~(long long)ULONG_MAX)
1816 hlen = ULONG_MAX - hba + 1;
1819 details.check_mapping = even_cows? NULL: mapping;
1820 details.nonlinear_vma = NULL;
1821 details.first_index = hba;
1822 details.last_index = hba + hlen - 1;
1823 if (details.last_index < details.first_index)
1824 details.last_index = ULONG_MAX;
1825 details.i_mmap_lock = &mapping->i_mmap_lock;
1827 spin_lock(&mapping->i_mmap_lock);
1829 /* serialize i_size write against truncate_count write */
1830 smp_wmb();
1831 /* Protect against page faults, and endless unmapping loops */
1832 mapping->truncate_count++;
1833 /*
1834 * For archs where spin_lock has inclusive semantics like ia64
1835 * this smp_mb() will prevent to read pagetable contents
1836 * before the truncate_count increment is visible to
1837 * other cpus.
1838 */
1839 smp_mb();
1840 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1841 if (mapping->truncate_count == 0)
1842 reset_vma_truncate_counts(mapping);
1843 mapping->truncate_count++;
1845 details.truncate_count = mapping->truncate_count;
1847 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1848 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1849 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1850 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1851 spin_unlock(&mapping->i_mmap_lock);
1853 EXPORT_SYMBOL(unmap_mapping_range);
1855 /*
1856 * Handle all mappings that got truncated by a "truncate()"
1857 * system call.
1859 * NOTE! We have to be ready to update the memory sharing
1860 * between the file and the memory map for a potential last
1861 * incomplete page. Ugly, but necessary.
1862 */
1863 int vmtruncate(struct inode * inode, loff_t offset)
1865 struct address_space *mapping = inode->i_mapping;
1866 unsigned long limit;
1868 if (inode->i_size < offset)
1869 goto do_expand;
1870 /*
1871 * truncation of in-use swapfiles is disallowed - it would cause
1872 * subsequent swapout to scribble on the now-freed blocks.
1873 */
1874 if (IS_SWAPFILE(inode))
1875 goto out_busy;
1876 i_size_write(inode, offset);
1877 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1878 truncate_inode_pages(mapping, offset);
1879 goto out_truncate;
1881 do_expand:
1882 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1883 if (limit != RLIM_INFINITY && offset > limit)
1884 goto out_sig;
1885 if (offset > inode->i_sb->s_maxbytes)
1886 goto out_big;
1887 i_size_write(inode, offset);
1889 out_truncate:
1890 if (inode->i_op && inode->i_op->truncate)
1891 inode->i_op->truncate(inode);
1892 return 0;
1893 out_sig:
1894 send_sig(SIGXFSZ, current, 0);
1895 out_big:
1896 return -EFBIG;
1897 out_busy:
1898 return -ETXTBSY;
1900 EXPORT_SYMBOL(vmtruncate);
1902 int vmtruncate_range(struct inode *inode, loff_t offset, loff_t end)
1904 struct address_space *mapping = inode->i_mapping;
1906 /*
1907 * If the underlying filesystem is not going to provide
1908 * a way to truncate a range of blocks (punch a hole) -
1909 * we should return failure right now.
1910 */
1911 if (!inode->i_op || !inode->i_op->truncate_range)
1912 return -ENOSYS;
1914 mutex_lock(&inode->i_mutex);
1915 down_write(&inode->i_alloc_sem);
1916 unmap_mapping_range(mapping, offset, (end - offset), 1);
1917 truncate_inode_pages_range(mapping, offset, end);
1918 inode->i_op->truncate_range(inode, offset, end);
1919 up_write(&inode->i_alloc_sem);
1920 mutex_unlock(&inode->i_mutex);
1922 return 0;
1924 EXPORT_SYMBOL(vmtruncate_range);
1926 /*
1927 * Primitive swap readahead code. We simply read an aligned block of
1928 * (1 << page_cluster) entries in the swap area. This method is chosen
1929 * because it doesn't cost us any seek time. We also make sure to queue
1930 * the 'original' request together with the readahead ones...
1932 * This has been extended to use the NUMA policies from the mm triggering
1933 * the readahead.
1935 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1936 */
1937 void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1939 #ifdef CONFIG_NUMA
1940 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1941 #endif
1942 int i, num;
1943 struct page *new_page;
1944 unsigned long offset;
1946 /*
1947 * Get the number of handles we should do readahead io to.
1948 */
1949 num = valid_swaphandles(entry, &offset);
1950 for (i = 0; i < num; offset++, i++) {
1951 /* Ok, do the async read-ahead now */
1952 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1953 offset), vma, addr);
1954 if (!new_page)
1955 break;
1956 page_cache_release(new_page);
1957 #ifdef CONFIG_NUMA
1958 /*
1959 * Find the next applicable VMA for the NUMA policy.
1960 */
1961 addr += PAGE_SIZE;
1962 if (addr == 0)
1963 vma = NULL;
1964 if (vma) {
1965 if (addr >= vma->vm_end) {
1966 vma = next_vma;
1967 next_vma = vma ? vma->vm_next : NULL;
1969 if (vma && addr < vma->vm_start)
1970 vma = NULL;
1971 } else {
1972 if (next_vma && addr >= next_vma->vm_start) {
1973 vma = next_vma;
1974 next_vma = vma->vm_next;
1977 #endif
1979 lru_add_drain(); /* Push any new pages onto the LRU now */
1982 /*
1983 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1984 * but allow concurrent faults), and pte mapped but not yet locked.
1985 * We return with mmap_sem still held, but pte unmapped and unlocked.
1986 */
1987 static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1988 unsigned long address, pte_t *page_table, pmd_t *pmd,
1989 int write_access, pte_t orig_pte)
1991 spinlock_t *ptl;
1992 struct page *page;
1993 swp_entry_t entry;
1994 pte_t pte;
1995 int ret = VM_FAULT_MINOR;
1997 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
1998 goto out;
2000 entry = pte_to_swp_entry(orig_pte);
2001 again:
2002 page = lookup_swap_cache(entry);
2003 if (!page) {
2004 swapin_readahead(entry, address, vma);
2005 page = read_swap_cache_async(entry, vma, address);
2006 if (!page) {
2007 /*
2008 * Back out if somebody else faulted in this pte
2009 * while we released the pte lock.
2010 */
2011 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2012 if (likely(pte_same(*page_table, orig_pte)))
2013 ret = VM_FAULT_OOM;
2014 goto unlock;
2017 /* Had to read the page from swap area: Major fault */
2018 ret = VM_FAULT_MAJOR;
2019 inc_page_state(pgmajfault);
2020 grab_swap_token();
2023 mark_page_accessed(page);
2024 lock_page(page);
2025 if (!PageSwapCache(page)) {
2026 /* Page migration has occured */
2027 unlock_page(page);
2028 page_cache_release(page);
2029 goto again;
2032 /*
2033 * Back out if somebody else already faulted in this pte.
2034 */
2035 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2036 if (unlikely(!pte_same(*page_table, orig_pte)))
2037 goto out_nomap;
2039 if (unlikely(!PageUptodate(page))) {
2040 ret = VM_FAULT_SIGBUS;
2041 goto out_nomap;
2044 /* The page isn't present yet, go ahead with the fault. */
2046 inc_mm_counter(mm, anon_rss);
2047 pte = mk_pte(page, vma->vm_page_prot);
2048 if (write_access && can_share_swap_page(page)) {
2049 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
2050 write_access = 0;
2053 flush_icache_page(vma, page);
2054 set_pte_at(mm, address, page_table, pte);
2055 page_add_anon_rmap(page, vma, address);
2057 swap_free(entry);
2058 if (vm_swap_full())
2059 remove_exclusive_swap_page(page);
2060 unlock_page(page);
2062 if (write_access) {
2063 if (do_wp_page(mm, vma, address,
2064 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
2065 ret = VM_FAULT_OOM;
2066 goto out;
2069 /* No need to invalidate - it was non-present before */
2070 update_mmu_cache(vma, address, pte);
2071 lazy_mmu_prot_update(pte);
2072 unlock:
2073 pte_unmap_unlock(page_table, ptl);
2074 out:
2075 return ret;
2076 out_nomap:
2077 pte_unmap_unlock(page_table, ptl);
2078 unlock_page(page);
2079 page_cache_release(page);
2080 return ret;
2083 /*
2084 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2085 * but allow concurrent faults), and pte mapped but not yet locked.
2086 * We return with mmap_sem still held, but pte unmapped and unlocked.
2087 */
2088 static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
2089 unsigned long address, pte_t *page_table, pmd_t *pmd,
2090 int write_access)
2092 struct page *page;
2093 spinlock_t *ptl;
2094 pte_t entry;
2096 if (write_access) {
2097 /* Allocate our own private page. */
2098 pte_unmap(page_table);
2100 if (unlikely(anon_vma_prepare(vma)))
2101 goto oom;
2102 page = alloc_zeroed_user_highpage(vma, address);
2103 if (!page)
2104 goto oom;
2106 entry = mk_pte(page, vma->vm_page_prot);
2107 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2109 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2110 if (!pte_none(*page_table))
2111 goto release;
2112 inc_mm_counter(mm, anon_rss);
2113 lru_cache_add_active(page);
2114 page_add_new_anon_rmap(page, vma, address);
2115 } else {
2116 /* Map the ZERO_PAGE - vm_page_prot is readonly */
2117 page = ZERO_PAGE(address);
2118 page_cache_get(page);
2119 entry = mk_pte(page, vma->vm_page_prot);
2121 ptl = pte_lockptr(mm, pmd);
2122 spin_lock(ptl);
2123 if (!pte_none(*page_table))
2124 goto release;
2125 inc_mm_counter(mm, file_rss);
2126 page_add_file_rmap(page);
2129 set_pte_at(mm, address, page_table, entry);
2131 /* No need to invalidate - it was non-present before */
2132 update_mmu_cache(vma, address, entry);
2133 lazy_mmu_prot_update(entry);
2134 unlock:
2135 pte_unmap_unlock(page_table, ptl);
2136 return VM_FAULT_MINOR;
2137 release:
2138 page_cache_release(page);
2139 goto unlock;
2140 oom:
2141 return VM_FAULT_OOM;
2144 /*
2145 * do_no_page() tries to create a new page mapping. It aggressively
2146 * tries to share with existing pages, but makes a separate copy if
2147 * the "write_access" parameter is true in order to avoid the next
2148 * page fault.
2150 * As this is called only for pages that do not currently exist, we
2151 * do not need to flush old virtual caches or the TLB.
2153 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2154 * but allow concurrent faults), and pte mapped but not yet locked.
2155 * We return with mmap_sem still held, but pte unmapped and unlocked.
2156 */
2157 static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
2158 unsigned long address, pte_t *page_table, pmd_t *pmd,
2159 int write_access)
2161 spinlock_t *ptl;
2162 struct page *new_page;
2163 struct address_space *mapping = NULL;
2164 pte_t entry;
2165 unsigned int sequence = 0;
2166 int ret = VM_FAULT_MINOR;
2167 int anon = 0;
2169 pte_unmap(page_table);
2170 BUG_ON(vma->vm_flags & VM_PFNMAP);
2172 if (vma->vm_file) {
2173 mapping = vma->vm_file->f_mapping;
2174 sequence = mapping->truncate_count;
2175 smp_rmb(); /* serializes i_size against truncate_count */
2177 retry:
2178 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
2179 /*
2180 * No smp_rmb is needed here as long as there's a full
2181 * spin_lock/unlock sequence inside the ->nopage callback
2182 * (for the pagecache lookup) that acts as an implicit
2183 * smp_mb() and prevents the i_size read to happen
2184 * after the next truncate_count read.
2185 */
2187 /* no page was available -- either SIGBUS or OOM */
2188 if (new_page == NOPAGE_SIGBUS)
2189 return VM_FAULT_SIGBUS;
2190 if (new_page == NOPAGE_OOM)
2191 return VM_FAULT_OOM;
2193 /*
2194 * Should we do an early C-O-W break?
2195 */
2196 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2197 struct page *page;
2199 if (unlikely(anon_vma_prepare(vma)))
2200 goto oom;
2201 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
2202 if (!page)
2203 goto oom;
2204 copy_user_highpage(page, new_page, address);
2205 page_cache_release(new_page);
2206 new_page = page;
2207 anon = 1;
2210 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
2211 /*
2212 * For a file-backed vma, someone could have truncated or otherwise
2213 * invalidated this page. If unmap_mapping_range got called,
2214 * retry getting the page.
2215 */
2216 if (mapping && unlikely(sequence != mapping->truncate_count)) {
2217 pte_unmap_unlock(page_table, ptl);
2218 page_cache_release(new_page);
2219 cond_resched();
2220 sequence = mapping->truncate_count;
2221 smp_rmb();
2222 goto retry;
2225 /*
2226 * This silly early PAGE_DIRTY setting removes a race
2227 * due to the bad i386 page protection. But it's valid
2228 * for other architectures too.
2230 * Note that if write_access is true, we either now have
2231 * an exclusive copy of the page, or this is a shared mapping,
2232 * so we can make it writable and dirty to avoid having to
2233 * handle that later.
2234 */
2235 /* Only go through if we didn't race with anybody else... */
2236 if (pte_none(*page_table)) {
2237 flush_icache_page(vma, new_page);
2238 entry = mk_pte(new_page, vma->vm_page_prot);
2239 if (write_access)
2240 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
2241 set_pte_at(mm, address, page_table, entry);
2242 if (anon) {
2243 inc_mm_counter(mm, anon_rss);
2244 lru_cache_add_active(new_page);
2245 page_add_new_anon_rmap(new_page, vma, address);
2246 } else {
2247 inc_mm_counter(mm, file_rss);
2248 page_add_file_rmap(new_page);
2250 } else {
2251 /* One of our sibling threads was faster, back out. */
2252 page_cache_release(new_page);
2253 goto unlock;
2256 /* no need to invalidate: a not-present page shouldn't be cached */
2257 update_mmu_cache(vma, address, entry);
2258 lazy_mmu_prot_update(entry);
2259 unlock:
2260 pte_unmap_unlock(page_table, ptl);
2261 return ret;
2262 oom:
2263 page_cache_release(new_page);
2264 return VM_FAULT_OOM;
2267 /*
2268 * Fault of a previously existing named mapping. Repopulate the pte
2269 * from the encoded file_pte if possible. This enables swappable
2270 * nonlinear vmas.
2272 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2273 * but allow concurrent faults), and pte mapped but not yet locked.
2274 * We return with mmap_sem still held, but pte unmapped and unlocked.
2275 */
2276 static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2277 unsigned long address, pte_t *page_table, pmd_t *pmd,
2278 int write_access, pte_t orig_pte)
2280 pgoff_t pgoff;
2281 int err;
2283 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
2284 return VM_FAULT_MINOR;
2286 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2287 /*
2288 * Page table corrupted: show pte and kill process.
2289 */
2290 print_bad_pte(vma, orig_pte, address);
2291 return VM_FAULT_OOM;
2293 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2295 pgoff = pte_to_pgoff(orig_pte);
2296 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2297 vma->vm_page_prot, pgoff, 0);
2298 if (err == -ENOMEM)
2299 return VM_FAULT_OOM;
2300 if (err)
2301 return VM_FAULT_SIGBUS;
2302 return VM_FAULT_MAJOR;
2305 /*
2306 * These routines also need to handle stuff like marking pages dirty
2307 * and/or accessed for architectures that don't do it in hardware (most
2308 * RISC architectures). The early dirtying is also good on the i386.
2310 * There is also a hook called "update_mmu_cache()" that architectures
2311 * with external mmu caches can use to update those (ie the Sparc or
2312 * PowerPC hashed page tables that act as extended TLBs).
2314 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2315 * but allow concurrent faults), and pte mapped but not yet locked.
2316 * We return with mmap_sem still held, but pte unmapped and unlocked.
2317 */
2318 static inline int handle_pte_fault(struct mm_struct *mm,
2319 struct vm_area_struct *vma, unsigned long address,
2320 pte_t *pte, pmd_t *pmd, int write_access)
2322 pte_t entry;
2323 pte_t old_entry;
2324 spinlock_t *ptl;
2326 old_entry = entry = *pte;
2327 if (!pte_present(entry)) {
2328 if (pte_none(entry)) {
2329 if (!vma->vm_ops || !vma->vm_ops->nopage)
2330 return do_anonymous_page(mm, vma, address,
2331 pte, pmd, write_access);
2332 return do_no_page(mm, vma, address,
2333 pte, pmd, write_access);
2335 if (pte_file(entry))
2336 return do_file_page(mm, vma, address,
2337 pte, pmd, write_access, entry);
2338 return do_swap_page(mm, vma, address,
2339 pte, pmd, write_access, entry);
2342 ptl = pte_lockptr(mm, pmd);
2343 spin_lock(ptl);
2344 if (unlikely(!pte_same(*pte, entry)))
2345 goto unlock;
2346 if (write_access) {
2347 if (!pte_write(entry))
2348 return do_wp_page(mm, vma, address,
2349 pte, pmd, ptl, entry);
2350 entry = pte_mkdirty(entry);
2352 entry = pte_mkyoung(entry);
2353 if (!pte_same(old_entry, entry)) {
2354 ptep_set_access_flags(vma, address, pte, entry, write_access);
2355 update_mmu_cache(vma, address, entry);
2356 lazy_mmu_prot_update(entry);
2357 } else {
2358 /*
2359 * This is needed only for protection faults but the arch code
2360 * is not yet telling us if this is a protection fault or not.
2361 * This still avoids useless tlb flushes for .text page faults
2362 * with threads.
2363 */
2364 if (write_access)
2365 flush_tlb_page(vma, address);
2367 unlock:
2368 pte_unmap_unlock(pte, ptl);
2369 return VM_FAULT_MINOR;
2372 /*
2373 * By the time we get here, we already hold the mm semaphore
2374 */
2375 int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2376 unsigned long address, int write_access)
2378 pgd_t *pgd;
2379 pud_t *pud;
2380 pmd_t *pmd;
2381 pte_t *pte;
2383 __set_current_state(TASK_RUNNING);
2385 inc_page_state(pgfault);
2387 if (unlikely(is_vm_hugetlb_page(vma)))
2388 return hugetlb_fault(mm, vma, address, write_access);
2390 pgd = pgd_offset(mm, address);
2391 pud = pud_alloc(mm, pgd, address);
2392 if (!pud)
2393 return VM_FAULT_OOM;
2394 pmd = pmd_alloc(mm, pud, address);
2395 if (!pmd)
2396 return VM_FAULT_OOM;
2397 pte = pte_alloc_map(mm, pmd, address);
2398 if (!pte)
2399 return VM_FAULT_OOM;
2401 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
2404 EXPORT_SYMBOL_GPL(__handle_mm_fault);
2406 #ifndef __PAGETABLE_PUD_FOLDED
2407 /*
2408 * Allocate page upper directory.
2409 * We've already handled the fast-path in-line.
2410 */
2411 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2413 pud_t *new = pud_alloc_one(mm, address);
2414 if (!new)
2415 return -ENOMEM;
2417 spin_lock(&mm->page_table_lock);
2418 if (pgd_present(*pgd)) /* Another has populated it */
2419 pud_free(new);
2420 else
2421 pgd_populate(mm, pgd, new);
2422 spin_unlock(&mm->page_table_lock);
2423 return 0;
2425 #else
2426 /* Workaround for gcc 2.96 */
2427 int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
2429 return 0;
2431 #endif /* __PAGETABLE_PUD_FOLDED */
2433 #ifndef __PAGETABLE_PMD_FOLDED
2434 /*
2435 * Allocate page middle directory.
2436 * We've already handled the fast-path in-line.
2437 */
2438 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2440 pmd_t *new = pmd_alloc_one(mm, address);
2441 if (!new)
2442 return -ENOMEM;
2444 spin_lock(&mm->page_table_lock);
2445 #ifndef __ARCH_HAS_4LEVEL_HACK
2446 if (pud_present(*pud)) /* Another has populated it */
2447 pmd_free(new);
2448 else
2449 pud_populate(mm, pud, new);
2450 #else
2451 if (pgd_present(*pud)) /* Another has populated it */
2452 pmd_free(new);
2453 else
2454 pgd_populate(mm, pud, new);
2455 #endif /* __ARCH_HAS_4LEVEL_HACK */
2456 spin_unlock(&mm->page_table_lock);
2457 return 0;
2459 #else
2460 /* Workaround for gcc 2.96 */
2461 int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
2463 return 0;
2465 #endif /* __PAGETABLE_PMD_FOLDED */
2467 int make_pages_present(unsigned long addr, unsigned long end)
2469 int ret, len, write;
2470 struct vm_area_struct * vma;
2472 vma = find_vma(current->mm, addr);
2473 if (!vma)
2474 return -1;
2475 write = (vma->vm_flags & VM_WRITE) != 0;
2476 if (addr >= end)
2477 BUG();
2478 if (end > vma->vm_end)
2479 BUG();
2480 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2481 ret = get_user_pages(current, current->mm, addr,
2482 len, write, 0, NULL, NULL);
2483 if (ret < 0)
2484 return ret;
2485 return ret == len ? 0 : -1;
2488 /*
2489 * Map a vmalloc()-space virtual address to the physical page.
2490 */
2491 struct page * vmalloc_to_page(void * vmalloc_addr)
2493 unsigned long addr = (unsigned long) vmalloc_addr;
2494 struct page *page = NULL;
2495 pgd_t *pgd = pgd_offset_k(addr);
2496 pud_t *pud;
2497 pmd_t *pmd;
2498 pte_t *ptep, pte;
2500 if (!pgd_none(*pgd)) {
2501 pud = pud_offset(pgd, addr);
2502 if (!pud_none(*pud)) {
2503 pmd = pmd_offset(pud, addr);
2504 if (!pmd_none(*pmd)) {
2505 ptep = pte_offset_map(pmd, addr);
2506 pte = *ptep;
2507 if (pte_present(pte))
2508 page = pte_page(pte);
2509 pte_unmap(ptep);
2513 return page;
2516 EXPORT_SYMBOL(vmalloc_to_page);
2518 /*
2519 * Map a vmalloc()-space virtual address to the physical page frame number.
2520 */
2521 unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2523 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2526 EXPORT_SYMBOL(vmalloc_to_pfn);
2528 #if !defined(__HAVE_ARCH_GATE_AREA)
2530 #if defined(AT_SYSINFO_EHDR)
2531 static struct vm_area_struct gate_vma;
2533 static int __init gate_vma_init(void)
2535 gate_vma.vm_mm = NULL;
2536 gate_vma.vm_start = FIXADDR_USER_START;
2537 gate_vma.vm_end = FIXADDR_USER_END;
2538 gate_vma.vm_page_prot = PAGE_READONLY;
2539 gate_vma.vm_flags = 0;
2540 return 0;
2542 __initcall(gate_vma_init);
2543 #endif
2545 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2547 #ifdef AT_SYSINFO_EHDR
2548 return &gate_vma;
2549 #else
2550 return NULL;
2551 #endif
2554 int in_gate_area_no_task(unsigned long addr)
2556 #ifdef AT_SYSINFO_EHDR
2557 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2558 return 1;
2559 #endif
2560 return 0;
2563 #endif /* __HAVE_ARCH_GATE_AREA */