blob: d1f46f4e4c8a8594ffa649f15dd03d6cd6eda944 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
Andy Whitcroftd41dee32005-06-23 00:07:54 -070061#ifndef CONFIG_NEED_MULTIPLE_NODES
Linus Torvalds1da177e2005-04-16 15:20:36 -070062/* 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 */
Hugh Dickinse0da3822005-04-19 13:29:15 -0700113static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114{
Hugh Dickinse0da3822005-04-19 13:29:15 -0700115 struct page *page = pmd_page(*pmd);
116 pmd_clear(pmd);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700117 pte_lock_deinit(page);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700118 pte_free_tlb(tlb, page);
119 dec_page_state(nr_page_table_pages);
120 tlb->mm->nr_ptes--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700121}
122
Hugh Dickinse0da3822005-04-19 13:29:15 -0700123static 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)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700126{
127 pmd_t *pmd;
128 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700129 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700130
Hugh Dickinse0da3822005-04-19 13:29:15 -0700131 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 pmd = pmd_offset(pud, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133 do {
134 next = pmd_addr_end(addr, end);
135 if (pmd_none_or_clear_bad(pmd))
136 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700137 free_pte_range(tlb, pmd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138 } while (pmd++, addr = next, addr != end);
139
Hugh Dickinse0da3822005-04-19 13:29:15 -0700140 start &= PUD_MASK;
141 if (start < floor)
142 return;
143 if (ceiling) {
144 ceiling &= PUD_MASK;
145 if (!ceiling)
146 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700148 if (end - 1 > ceiling - 1)
149 return;
150
151 pmd = pmd_offset(pud, start);
152 pud_clear(pud);
153 pmd_free_tlb(tlb, pmd);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154}
155
Hugh Dickinse0da3822005-04-19 13:29:15 -0700156static 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)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159{
160 pud_t *pud;
161 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700162 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700163
Hugh Dickinse0da3822005-04-19 13:29:15 -0700164 start = addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165 pud = pud_offset(pgd, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 do {
167 next = pud_addr_end(addr, end);
168 if (pud_none_or_clear_bad(pud))
169 continue;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700170 free_pmd_range(tlb, pud, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 } while (pud++, addr = next, addr != end);
172
Hugh Dickinse0da3822005-04-19 13:29:15 -0700173 start &= PGDIR_MASK;
174 if (start < floor)
175 return;
176 if (ceiling) {
177 ceiling &= PGDIR_MASK;
178 if (!ceiling)
179 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700181 if (end - 1 > ceiling - 1)
182 return;
183
184 pud = pud_offset(pgd, start);
185 pgd_clear(pgd);
186 pud_free_tlb(tlb, pud);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
188
189/*
Hugh Dickinse0da3822005-04-19 13:29:15 -0700190 * This function frees user-level page tables of a process.
191 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192 * Must be called with pagetable lock held.
193 */
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700194void free_pgd_range(struct mmu_gather **tlb,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700195 unsigned long addr, unsigned long end,
196 unsigned long floor, unsigned long ceiling)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700197{
198 pgd_t *pgd;
199 unsigned long next;
Hugh Dickinse0da3822005-04-19 13:29:15 -0700200 unsigned long start;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700201
Hugh Dickinse0da3822005-04-19 13:29:15 -0700202 /*
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 */
227
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;
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700245 pgd = pgd_offset((*tlb)->mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246 do {
247 next = pgd_addr_end(addr, end);
248 if (pgd_none_or_clear_bad(pgd))
249 continue;
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700250 free_pud_range(*tlb, pgd, addr, next, floor, ceiling);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251 } while (pgd++, addr = next, addr != end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700252
Hugh Dickins4d6ddfa2005-10-29 18:16:02 -0700253 if (!(*tlb)->fullmm)
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700254 flush_tlb_pgtables((*tlb)->mm, start, end);
Hugh Dickinse0da3822005-04-19 13:29:15 -0700255}
256
257void free_pgtables(struct mmu_gather **tlb, struct vm_area_struct *vma,
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700258 unsigned long floor, unsigned long ceiling)
Hugh Dickinse0da3822005-04-19 13:29:15 -0700259{
260 while (vma) {
261 struct vm_area_struct *next = vma->vm_next;
262 unsigned long addr = vma->vm_start;
263
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700264 /*
265 * Hide vma from rmap and vmtruncate before freeing pgtables
266 */
267 anon_vma_unlink(vma);
268 unlink_file_vma(vma);
269
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700270 if (is_hugepage_only_range(vma->vm_mm, addr, HPAGE_SIZE)) {
271 hugetlb_free_pgd_range(tlb, addr, vma->vm_end,
Hugh Dickinse0da3822005-04-19 13:29:15 -0700272 floor, next? next->vm_start: ceiling);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700273 } 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;
Hugh Dickins8f4f8c12005-10-29 18:16:29 -0700282 anon_vma_unlink(vma);
283 unlink_file_vma(vma);
Hugh Dickins3bf5ee92005-04-19 13:29:16 -0700284 }
285 free_pgd_range(tlb, addr, vma->vm_end,
286 floor, next? next->vm_start: ceiling);
287 }
Hugh Dickinse0da3822005-04-19 13:29:15 -0700288 vma = next;
289 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290}
291
Hugh Dickins1bb36302005-10-29 18:16:22 -0700292int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293{
Hugh Dickinsc74df322005-10-29 18:16:23 -0700294 struct page *new = pte_alloc_one(mm, address);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700295 if (!new)
296 return -ENOMEM;
297
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700298 pte_lock_init(new);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700299 spin_lock(&mm->page_table_lock);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700300 if (pmd_present(*pmd)) { /* Another has populated it */
301 pte_lock_deinit(new);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700302 pte_free(new);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700303 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 mm->nr_ptes++;
305 inc_page_state(nr_page_table_pages);
306 pmd_populate(mm, pmd, new);
307 }
Hugh Dickinsc74df322005-10-29 18:16:23 -0700308 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -0700309 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310}
311
Hugh Dickins1bb36302005-10-29 18:16:22 -0700312int __pte_alloc_kernel(pmd_t *pmd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313{
Hugh Dickins1bb36302005-10-29 18:16:22 -0700314 pte_t *new = pte_alloc_one_kernel(&init_mm, address);
315 if (!new)
316 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700317
Hugh Dickins1bb36302005-10-29 18:16:22 -0700318 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;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325}
326
Hugh Dickinsae859762005-10-29 18:16:05 -0700327static 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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700335/*
Nick Pigginb5810032005-10-29 18:16:12 -0700336 * This function is called to print an error when a pte in a
Hugh Dickins0b14c172005-11-21 21:32:15 -0800337 * !VM_UNPAGED region is found pointing to an invalid pfn (which
Nick Pigginb5810032005-10-29 18:16:12 -0700338 * 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
352/*
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800353 * page_is_anon applies strict checks for an anonymous page belonging to
354 * this vma at this address. It is used on VM_UNPAGED vmas, which are
355 * usually populated with shared originals (which must not be counted),
356 * but occasionally contain private COWed copies (when !VM_SHARED, or
357 * perhaps via ptrace when VM_SHARED). An mmap of /dev/mem might window
358 * free pages, pages from other processes, or from other parts of this:
359 * it's tricky, but try not to be deceived by foreign anonymous pages.
360 */
361static inline int page_is_anon(struct page *page,
362 struct vm_area_struct *vma, unsigned long addr)
363{
364 return page && PageAnon(page) && page_mapped(page) &&
365 page_address_in_vma(page, vma) == addr;
366}
367
368/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 * copy one vm_area from one task to the other. Assumes the page tables
370 * already present in the new task to be cleared in the whole range
371 * covered by this vma.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372 */
373
Hugh Dickins8c103762005-10-29 18:16:13 -0700374static inline void
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375copy_one_pte(struct mm_struct *dst_mm, struct mm_struct *src_mm,
Nick Pigginb5810032005-10-29 18:16:12 -0700376 pte_t *dst_pte, pte_t *src_pte, struct vm_area_struct *vma,
Hugh Dickins8c103762005-10-29 18:16:13 -0700377 unsigned long addr, int *rss)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378{
Nick Pigginb5810032005-10-29 18:16:12 -0700379 unsigned long vm_flags = vma->vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 pte_t pte = *src_pte;
381 struct page *page;
382 unsigned long pfn;
383
384 /* pte contains position in swap or file, so copy. */
385 if (unlikely(!pte_present(pte))) {
386 if (!pte_file(pte)) {
387 swap_duplicate(pte_to_swp_entry(pte));
388 /* make sure dst_mm is on swapoff's mmlist. */
389 if (unlikely(list_empty(&dst_mm->mmlist))) {
390 spin_lock(&mmlist_lock);
Hugh Dickinsf412ac02005-10-29 18:16:41 -0700391 if (list_empty(&dst_mm->mmlist))
392 list_add(&dst_mm->mmlist,
393 &src_mm->mmlist);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 spin_unlock(&mmlist_lock);
395 }
396 }
Hugh Dickinsae859762005-10-29 18:16:05 -0700397 goto out_set_pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 }
399
Nick Pigginb5810032005-10-29 18:16:12 -0700400 pfn = pte_pfn(pte);
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800401 page = pfn_valid(pfn)? pfn_to_page(pfn): NULL;
402
403 if (unlikely(vm_flags & VM_UNPAGED))
404 if (!page_is_anon(page, vma, addr))
405 goto out_set_pte;
406
407 /*
408 * If the pte points outside of valid memory but
Hugh Dickins0b14c172005-11-21 21:32:15 -0800409 * the region is not VM_UNPAGED, we have a problem.
Nick Pigginb5810032005-10-29 18:16:12 -0700410 */
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800411 if (unlikely(!page)) {
Nick Pigginb5810032005-10-29 18:16:12 -0700412 print_bad_pte(vma, pte, addr);
413 goto out_set_pte; /* try to do something sane */
414 }
415
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416 /*
417 * If it's a COW mapping, write protect it both
418 * in the parent and the child
419 */
420 if ((vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE) {
421 ptep_set_wrprotect(src_mm, addr, src_pte);
422 pte = *src_pte;
423 }
424
425 /*
426 * If it's a shared mapping, mark it clean in
427 * the child
428 */
429 if (vm_flags & VM_SHARED)
430 pte = pte_mkclean(pte);
431 pte = pte_mkold(pte);
432 get_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433 page_dup_rmap(page);
Hugh Dickins8c103762005-10-29 18:16:13 -0700434 rss[!!PageAnon(page)]++;
Hugh Dickinsae859762005-10-29 18:16:05 -0700435
436out_set_pte:
437 set_pte_at(dst_mm, addr, dst_pte, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438}
439
440static int copy_pte_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
441 pmd_t *dst_pmd, pmd_t *src_pmd, struct vm_area_struct *vma,
442 unsigned long addr, unsigned long end)
443{
444 pte_t *src_pte, *dst_pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700445 spinlock_t *src_ptl, *dst_ptl;
Hugh Dickinse040f212005-10-29 18:15:53 -0700446 int progress = 0;
Hugh Dickins8c103762005-10-29 18:16:13 -0700447 int rss[2];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448
449again:
Hugh Dickinsae859762005-10-29 18:16:05 -0700450 rss[1] = rss[0] = 0;
Hugh Dickinsc74df322005-10-29 18:16:23 -0700451 dst_pte = pte_alloc_map_lock(dst_mm, dst_pmd, addr, &dst_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452 if (!dst_pte)
453 return -ENOMEM;
454 src_pte = pte_offset_map_nested(src_pmd, addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700455 src_ptl = pte_lockptr(src_mm, src_pmd);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700456 spin_lock(src_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 do {
459 /*
460 * We are holding two locks at this point - either of them
461 * could generate latencies in another task on another CPU.
462 */
Hugh Dickinse040f212005-10-29 18:15:53 -0700463 if (progress >= 32) {
464 progress = 0;
465 if (need_resched() ||
Hugh Dickinsc74df322005-10-29 18:16:23 -0700466 need_lockbreak(src_ptl) ||
467 need_lockbreak(dst_ptl))
Hugh Dickinse040f212005-10-29 18:15:53 -0700468 break;
469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700470 if (pte_none(*src_pte)) {
471 progress++;
472 continue;
473 }
Hugh Dickins8c103762005-10-29 18:16:13 -0700474 copy_one_pte(dst_mm, src_mm, dst_pte, src_pte, vma, addr, rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 progress += 8;
476 } while (dst_pte++, src_pte++, addr += PAGE_SIZE, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700477
Hugh Dickinsc74df322005-10-29 18:16:23 -0700478 spin_unlock(src_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479 pte_unmap_nested(src_pte - 1);
Hugh Dickinsae859762005-10-29 18:16:05 -0700480 add_mm_rss(dst_mm, rss[0], rss[1]);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700481 pte_unmap_unlock(dst_pte - 1, dst_ptl);
482 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483 if (addr != end)
484 goto again;
485 return 0;
486}
487
488static inline int copy_pmd_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
489 pud_t *dst_pud, pud_t *src_pud, struct vm_area_struct *vma,
490 unsigned long addr, unsigned long end)
491{
492 pmd_t *src_pmd, *dst_pmd;
493 unsigned long next;
494
495 dst_pmd = pmd_alloc(dst_mm, dst_pud, addr);
496 if (!dst_pmd)
497 return -ENOMEM;
498 src_pmd = pmd_offset(src_pud, addr);
499 do {
500 next = pmd_addr_end(addr, end);
501 if (pmd_none_or_clear_bad(src_pmd))
502 continue;
503 if (copy_pte_range(dst_mm, src_mm, dst_pmd, src_pmd,
504 vma, addr, next))
505 return -ENOMEM;
506 } while (dst_pmd++, src_pmd++, addr = next, addr != end);
507 return 0;
508}
509
510static inline int copy_pud_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
511 pgd_t *dst_pgd, pgd_t *src_pgd, struct vm_area_struct *vma,
512 unsigned long addr, unsigned long end)
513{
514 pud_t *src_pud, *dst_pud;
515 unsigned long next;
516
517 dst_pud = pud_alloc(dst_mm, dst_pgd, addr);
518 if (!dst_pud)
519 return -ENOMEM;
520 src_pud = pud_offset(src_pgd, addr);
521 do {
522 next = pud_addr_end(addr, end);
523 if (pud_none_or_clear_bad(src_pud))
524 continue;
525 if (copy_pmd_range(dst_mm, src_mm, dst_pud, src_pud,
526 vma, addr, next))
527 return -ENOMEM;
528 } while (dst_pud++, src_pud++, addr = next, addr != end);
529 return 0;
530}
531
532int copy_page_range(struct mm_struct *dst_mm, struct mm_struct *src_mm,
533 struct vm_area_struct *vma)
534{
535 pgd_t *src_pgd, *dst_pgd;
536 unsigned long next;
537 unsigned long addr = vma->vm_start;
538 unsigned long end = vma->vm_end;
539
Nick Piggind9928952005-08-28 16:49:11 +1000540 /*
541 * Don't copy ptes where a page fault will fill them correctly.
542 * Fork becomes much lighter when there are big shared or private
543 * readonly mappings. The tradeoff is that copy_page_range is more
544 * efficient than faulting.
545 */
Hugh Dickins0b14c172005-11-21 21:32:15 -0800546 if (!(vma->vm_flags & (VM_HUGETLB|VM_NONLINEAR|VM_UNPAGED))) {
Nick Piggind9928952005-08-28 16:49:11 +1000547 if (!vma->anon_vma)
548 return 0;
549 }
550
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 if (is_vm_hugetlb_page(vma))
552 return copy_hugetlb_page_range(dst_mm, src_mm, vma);
553
554 dst_pgd = pgd_offset(dst_mm, addr);
555 src_pgd = pgd_offset(src_mm, addr);
556 do {
557 next = pgd_addr_end(addr, end);
558 if (pgd_none_or_clear_bad(src_pgd))
559 continue;
560 if (copy_pud_range(dst_mm, src_mm, dst_pgd, src_pgd,
561 vma, addr, next))
562 return -ENOMEM;
563 } while (dst_pgd++, src_pgd++, addr = next, addr != end);
564 return 0;
565}
566
Robin Holt51c6f662005-11-13 16:06:42 -0800567static unsigned long zap_pte_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700568 struct vm_area_struct *vma, pmd_t *pmd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800570 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571{
Nick Pigginb5810032005-10-29 18:16:12 -0700572 struct mm_struct *mm = tlb->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 pte_t *pte;
Hugh Dickins508034a2005-10-29 18:16:30 -0700574 spinlock_t *ptl;
Hugh Dickinsae859762005-10-29 18:16:05 -0700575 int file_rss = 0;
576 int anon_rss = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577
Hugh Dickins508034a2005-10-29 18:16:30 -0700578 pte = pte_offset_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 do {
580 pte_t ptent = *pte;
Robin Holt51c6f662005-11-13 16:06:42 -0800581 if (pte_none(ptent)) {
582 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800584 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 if (pte_present(ptent)) {
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800586 struct page *page;
587 unsigned long pfn;
Robin Holt51c6f662005-11-13 16:06:42 -0800588
589 (*zap_work) -= PAGE_SIZE;
590
Hugh Dickinsee498ed2005-11-21 21:32:18 -0800591 pfn = pte_pfn(ptent);
592 page = pfn_valid(pfn)? pfn_to_page(pfn): NULL;
593
594 if (unlikely(vma->vm_flags & VM_UNPAGED)) {
595 if (!page_is_anon(page, vma, addr))
596 page = NULL;
597 } else if (unlikely(!page))
598 print_bad_pte(vma, ptent, addr);
599
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 if (unlikely(details) && page) {
601 /*
602 * unmap_shared_mapping_pages() wants to
603 * invalidate cache without truncating:
604 * unmap shared but keep private pages.
605 */
606 if (details->check_mapping &&
607 details->check_mapping != page->mapping)
608 continue;
609 /*
610 * Each page->index must be checked when
611 * invalidating or truncating nonlinear.
612 */
613 if (details->nonlinear_vma &&
614 (page->index < details->first_index ||
615 page->index > details->last_index))
616 continue;
617 }
Nick Pigginb5810032005-10-29 18:16:12 -0700618 ptent = ptep_get_and_clear_full(mm, addr, pte,
Zachary Amsdena6003882005-09-03 15:55:04 -0700619 tlb->fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 tlb_remove_tlb_entry(tlb, pte, addr);
621 if (unlikely(!page))
622 continue;
623 if (unlikely(details) && details->nonlinear_vma
624 && linear_page_index(details->nonlinear_vma,
625 addr) != page->index)
Nick Pigginb5810032005-10-29 18:16:12 -0700626 set_pte_at(mm, addr, pte,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 pgoff_to_pte(page->index));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628 if (PageAnon(page))
Hugh Dickins86d912f2005-10-29 18:16:14 -0700629 anon_rss--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700630 else {
631 if (pte_dirty(ptent))
632 set_page_dirty(page);
633 if (pte_young(ptent))
634 mark_page_accessed(page);
Hugh Dickins86d912f2005-10-29 18:16:14 -0700635 file_rss--;
Hugh Dickins6237bcd2005-10-29 18:15:54 -0700636 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 page_remove_rmap(page);
638 tlb_remove_page(tlb, page);
639 continue;
640 }
641 /*
642 * If details->check_mapping, we leave swap entries;
643 * if details->nonlinear_vma, we leave file entries.
644 */
645 if (unlikely(details))
646 continue;
647 if (!pte_file(ptent))
648 free_swap_and_cache(pte_to_swp_entry(ptent));
Nick Pigginb5810032005-10-29 18:16:12 -0700649 pte_clear_full(mm, addr, pte, tlb->fullmm);
Robin Holt51c6f662005-11-13 16:06:42 -0800650 } while (pte++, addr += PAGE_SIZE, (addr != end && *zap_work > 0));
Hugh Dickinsae859762005-10-29 18:16:05 -0700651
Hugh Dickins86d912f2005-10-29 18:16:14 -0700652 add_mm_rss(mm, file_rss, anon_rss);
Hugh Dickins508034a2005-10-29 18:16:30 -0700653 pte_unmap_unlock(pte - 1, ptl);
Robin Holt51c6f662005-11-13 16:06:42 -0800654
655 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656}
657
Robin Holt51c6f662005-11-13 16:06:42 -0800658static inline unsigned long zap_pmd_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700659 struct vm_area_struct *vma, pud_t *pud,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800661 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662{
663 pmd_t *pmd;
664 unsigned long next;
665
666 pmd = pmd_offset(pud, addr);
667 do {
668 next = pmd_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -0800669 if (pmd_none_or_clear_bad(pmd)) {
670 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800672 }
673 next = zap_pte_range(tlb, vma, pmd, addr, next,
674 zap_work, details);
675 } while (pmd++, addr = next, (addr != end && *zap_work > 0));
676
677 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678}
679
Robin Holt51c6f662005-11-13 16:06:42 -0800680static inline unsigned long zap_pud_range(struct mmu_gather *tlb,
Nick Pigginb5810032005-10-29 18:16:12 -0700681 struct vm_area_struct *vma, pgd_t *pgd,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800683 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
685 pud_t *pud;
686 unsigned long next;
687
688 pud = pud_offset(pgd, addr);
689 do {
690 next = pud_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -0800691 if (pud_none_or_clear_bad(pud)) {
692 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800694 }
695 next = zap_pmd_range(tlb, vma, pud, addr, next,
696 zap_work, details);
697 } while (pud++, addr = next, (addr != end && *zap_work > 0));
698
699 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700}
701
Robin Holt51c6f662005-11-13 16:06:42 -0800702static unsigned long unmap_page_range(struct mmu_gather *tlb,
703 struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 unsigned long addr, unsigned long end,
Robin Holt51c6f662005-11-13 16:06:42 -0800705 long *zap_work, struct zap_details *details)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706{
707 pgd_t *pgd;
708 unsigned long next;
709
710 if (details && !details->check_mapping && !details->nonlinear_vma)
711 details = NULL;
712
713 BUG_ON(addr >= end);
714 tlb_start_vma(tlb, vma);
715 pgd = pgd_offset(vma->vm_mm, addr);
716 do {
717 next = pgd_addr_end(addr, end);
Robin Holt51c6f662005-11-13 16:06:42 -0800718 if (pgd_none_or_clear_bad(pgd)) {
719 (*zap_work)--;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 continue;
Robin Holt51c6f662005-11-13 16:06:42 -0800721 }
722 next = zap_pud_range(tlb, vma, pgd, addr, next,
723 zap_work, details);
724 } while (pgd++, addr = next, (addr != end && *zap_work > 0));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 tlb_end_vma(tlb, vma);
Robin Holt51c6f662005-11-13 16:06:42 -0800726
727 return addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728}
729
730#ifdef CONFIG_PREEMPT
731# define ZAP_BLOCK_SIZE (8 * PAGE_SIZE)
732#else
733/* No preempt: go for improved straight-line efficiency */
734# define ZAP_BLOCK_SIZE (1024 * PAGE_SIZE)
735#endif
736
737/**
738 * unmap_vmas - unmap a range of memory covered by a list of vma's
739 * @tlbp: address of the caller's struct mmu_gather
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 * @vma: the starting vma
741 * @start_addr: virtual address at which to start unmapping
742 * @end_addr: virtual address at which to end unmapping
743 * @nr_accounted: Place number of unmapped pages in vm-accountable vma's here
744 * @details: details of nonlinear truncation or shared cache invalidation
745 *
Hugh Dickinsee39b372005-04-19 13:29:15 -0700746 * Returns the end address of the unmapping (restart addr if interrupted).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 *
Hugh Dickins508034a2005-10-29 18:16:30 -0700748 * Unmap all pages in the vma list.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 *
Hugh Dickins508034a2005-10-29 18:16:30 -0700750 * We aim to not hold locks for too long (for scheduling latency reasons).
751 * So zap pages in ZAP_BLOCK_SIZE bytecounts. This means we need to
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752 * return the ending mmu_gather to the caller.
753 *
754 * Only addresses between `start' and `end' will be unmapped.
755 *
756 * The VMA list must be sorted in ascending virtual address order.
757 *
758 * unmap_vmas() assumes that the caller will flush the whole unmapped address
759 * range after unmap_vmas() returns. So the only responsibility here is to
760 * ensure that any thus-far unmapped pages are flushed before unmap_vmas()
761 * drops the lock and schedules.
762 */
Hugh Dickins508034a2005-10-29 18:16:30 -0700763unsigned long unmap_vmas(struct mmu_gather **tlbp,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 struct vm_area_struct *vma, unsigned long start_addr,
765 unsigned long end_addr, unsigned long *nr_accounted,
766 struct zap_details *details)
767{
Robin Holt51c6f662005-11-13 16:06:42 -0800768 long zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769 unsigned long tlb_start = 0; /* For tlb_finish_mmu */
770 int tlb_start_valid = 0;
Hugh Dickinsee39b372005-04-19 13:29:15 -0700771 unsigned long start = start_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772 spinlock_t *i_mmap_lock = details? details->i_mmap_lock: NULL;
Hugh Dickins4d6ddfa2005-10-29 18:16:02 -0700773 int fullmm = (*tlbp)->fullmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774
775 for ( ; vma && vma->vm_start < end_addr; vma = vma->vm_next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 unsigned long end;
777
778 start = max(vma->vm_start, start_addr);
779 if (start >= vma->vm_end)
780 continue;
781 end = min(vma->vm_end, end_addr);
782 if (end <= vma->vm_start)
783 continue;
784
785 if (vma->vm_flags & VM_ACCOUNT)
786 *nr_accounted += (end - start) >> PAGE_SHIFT;
787
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 while (start != end) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 if (!tlb_start_valid) {
790 tlb_start = start;
791 tlb_start_valid = 1;
792 }
793
Robin Holt51c6f662005-11-13 16:06:42 -0800794 if (unlikely(is_vm_hugetlb_page(vma))) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 unmap_hugepage_range(vma, start, end);
Robin Holt51c6f662005-11-13 16:06:42 -0800796 zap_work -= (end - start) /
797 (HPAGE_SIZE / PAGE_SIZE);
798 start = end;
799 } else
800 start = unmap_page_range(*tlbp, vma,
801 start, end, &zap_work, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802
Robin Holt51c6f662005-11-13 16:06:42 -0800803 if (zap_work > 0) {
804 BUG_ON(start != end);
805 break;
806 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807
808 tlb_finish_mmu(*tlbp, tlb_start, start);
809
810 if (need_resched() ||
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 (i_mmap_lock && need_lockbreak(i_mmap_lock))) {
812 if (i_mmap_lock) {
Hugh Dickins508034a2005-10-29 18:16:30 -0700813 *tlbp = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 goto out;
815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 cond_resched();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 }
818
Hugh Dickins508034a2005-10-29 18:16:30 -0700819 *tlbp = tlb_gather_mmu(vma->vm_mm, fullmm);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820 tlb_start_valid = 0;
Robin Holt51c6f662005-11-13 16:06:42 -0800821 zap_work = ZAP_BLOCK_SIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822 }
823 }
824out:
Hugh Dickinsee39b372005-04-19 13:29:15 -0700825 return start; /* which is now the end (or restart) address */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826}
827
828/**
829 * zap_page_range - remove user pages in a given range
830 * @vma: vm_area_struct holding the applicable pages
831 * @address: starting address of pages to zap
832 * @size: number of bytes to zap
833 * @details: details of nonlinear truncation or shared cache invalidation
834 */
Hugh Dickinsee39b372005-04-19 13:29:15 -0700835unsigned long zap_page_range(struct vm_area_struct *vma, unsigned long address,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700836 unsigned long size, struct zap_details *details)
837{
838 struct mm_struct *mm = vma->vm_mm;
839 struct mmu_gather *tlb;
840 unsigned long end = address + size;
841 unsigned long nr_accounted = 0;
842
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843 lru_add_drain();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700844 tlb = tlb_gather_mmu(mm, 0);
Hugh Dickins365e9c872005-10-29 18:16:18 -0700845 update_hiwater_rss(mm);
Hugh Dickins508034a2005-10-29 18:16:30 -0700846 end = unmap_vmas(&tlb, vma, address, end, &nr_accounted, details);
847 if (tlb)
848 tlb_finish_mmu(tlb, address, end);
Hugh Dickinsee39b372005-04-19 13:29:15 -0700849 return end;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700850}
851
852/*
853 * Do a quick page-table lookup for a single page.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700855struct page *follow_page(struct mm_struct *mm, unsigned long address,
856 unsigned int flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857{
858 pgd_t *pgd;
859 pud_t *pud;
860 pmd_t *pmd;
861 pte_t *ptep, pte;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700862 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 unsigned long pfn;
864 struct page *page;
865
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700866 page = follow_huge_addr(mm, address, flags & FOLL_WRITE);
867 if (!IS_ERR(page)) {
868 BUG_ON(flags & FOLL_GET);
869 goto out;
870 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700872 page = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873 pgd = pgd_offset(mm, address);
874 if (pgd_none(*pgd) || unlikely(pgd_bad(*pgd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700875 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
877 pud = pud_offset(pgd, address);
878 if (pud_none(*pud) || unlikely(pud_bad(*pud)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700879 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880
881 pmd = pmd_offset(pud, address);
882 if (pmd_none(*pmd) || unlikely(pmd_bad(*pmd)))
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700883 goto no_page_table;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700885 if (pmd_huge(*pmd)) {
886 BUG_ON(flags & FOLL_GET);
887 page = follow_huge_pmd(mm, address, pmd, flags & FOLL_WRITE);
888 goto out;
889 }
890
891 ptep = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700892 if (!ptep)
893 goto out;
894
895 pte = *ptep;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700896 if (!pte_present(pte))
897 goto unlock;
898 if ((flags & FOLL_WRITE) && !pte_write(pte))
899 goto unlock;
900 pfn = pte_pfn(pte);
901 if (!pfn_valid(pfn))
902 goto unlock;
903
904 page = pfn_to_page(pfn);
905 if (flags & FOLL_GET)
906 get_page(page);
907 if (flags & FOLL_TOUCH) {
908 if ((flags & FOLL_WRITE) &&
909 !pte_dirty(pte) && !PageDirty(page))
910 set_page_dirty(page);
911 mark_page_accessed(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700913unlock:
914 pte_unmap_unlock(ptep, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915out:
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700916 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700918no_page_table:
919 /*
920 * When core dumping an enormous anonymous area that nobody
921 * has touched so far, we don't want to allocate page tables.
922 */
923 if (flags & FOLL_ANON) {
924 page = ZERO_PAGE(address);
925 if (flags & FOLL_GET)
926 get_page(page);
927 BUG_ON(flags & FOLL_WRITE);
928 }
929 return page;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930}
931
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
933 unsigned long start, int len, int write, int force,
934 struct page **pages, struct vm_area_struct **vmas)
935{
936 int i;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700937 unsigned int vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
939 /*
940 * Require read or write permissions.
941 * If 'force' is set, we only require the "MAY" flags.
942 */
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700943 vm_flags = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
944 vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 i = 0;
946
947 do {
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700948 struct vm_area_struct *vma;
949 unsigned int foll_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
951 vma = find_extend_vma(mm, start);
952 if (!vma && in_gate_area(tsk, start)) {
953 unsigned long pg = start & PAGE_MASK;
954 struct vm_area_struct *gate_vma = get_gate_vma(tsk);
955 pgd_t *pgd;
956 pud_t *pud;
957 pmd_t *pmd;
958 pte_t *pte;
959 if (write) /* user gate pages are read-only */
960 return i ? : -EFAULT;
961 if (pg > TASK_SIZE)
962 pgd = pgd_offset_k(pg);
963 else
964 pgd = pgd_offset_gate(mm, pg);
965 BUG_ON(pgd_none(*pgd));
966 pud = pud_offset(pgd, pg);
967 BUG_ON(pud_none(*pud));
968 pmd = pmd_offset(pud, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -0700969 if (pmd_none(*pmd))
970 return i ? : -EFAULT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971 pte = pte_offset_map(pmd, pg);
Hugh Dickins690dbe12005-08-01 21:11:42 -0700972 if (pte_none(*pte)) {
973 pte_unmap(pte);
974 return i ? : -EFAULT;
975 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 if (pages) {
977 pages[i] = pte_page(*pte);
978 get_page(pages[i]);
979 }
980 pte_unmap(pte);
981 if (vmas)
982 vmas[i] = gate_vma;
983 i++;
984 start += PAGE_SIZE;
985 len--;
986 continue;
987 }
988
Hugh Dickinsed5297a2005-11-21 21:32:11 -0800989 if (!vma || (vma->vm_flags & VM_IO)
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700990 || !(vm_flags & vma->vm_flags))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991 return i ? : -EFAULT;
992
993 if (is_vm_hugetlb_page(vma)) {
994 i = follow_hugetlb_page(mm, vma, pages, vmas,
995 &start, &len, i);
996 continue;
997 }
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -0700998
999 foll_flags = FOLL_TOUCH;
1000 if (pages)
1001 foll_flags |= FOLL_GET;
1002 if (!write && !(vma->vm_flags & VM_LOCKED) &&
1003 (!vma->vm_ops || !vma->vm_ops->nopage))
1004 foll_flags |= FOLL_ANON;
1005
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006 do {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001007 struct page *page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001009 if (write)
1010 foll_flags |= FOLL_WRITE;
1011
1012 cond_resched();
1013 while (!(page = follow_page(mm, start, foll_flags))) {
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001014 int ret;
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001015 ret = __handle_mm_fault(mm, vma, start,
1016 foll_flags & FOLL_WRITE);
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001017 /*
1018 * The VM_FAULT_WRITE bit tells us that do_wp_page has
1019 * broken COW when necessary, even if maybe_mkwrite
1020 * decided not to set pte_write. We can thus safely do
1021 * subsequent page lookups as if they were reads.
1022 */
1023 if (ret & VM_FAULT_WRITE)
Hugh Dickinsdeceb6c2005-10-29 18:16:33 -07001024 foll_flags &= ~FOLL_WRITE;
Linus Torvaldsa68d2eb2005-08-03 10:07:09 -07001025
1026 switch (ret & ~VM_FAULT_WRITE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 case VM_FAULT_MINOR:
1028 tsk->min_flt++;
1029 break;
1030 case VM_FAULT_MAJOR:
1031 tsk->maj_flt++;
1032 break;
1033 case VM_FAULT_SIGBUS:
1034 return i ? i : -EFAULT;
1035 case VM_FAULT_OOM:
1036 return i ? i : -ENOMEM;
1037 default:
1038 BUG();
1039 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 }
1041 if (pages) {
Hugh Dickins08ef4722005-06-21 17:15:10 -07001042 pages[i] = page;
1043 flush_dcache_page(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044 }
1045 if (vmas)
1046 vmas[i] = vma;
1047 i++;
1048 start += PAGE_SIZE;
1049 len--;
Hugh Dickins08ef4722005-06-21 17:15:10 -07001050 } while (len && start < vma->vm_end);
Hugh Dickins08ef4722005-06-21 17:15:10 -07001051 } while (len);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052 return i;
1053}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054EXPORT_SYMBOL(get_user_pages);
1055
1056static int zeromap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1057 unsigned long addr, unsigned long end, pgprot_t prot)
1058{
1059 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001060 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061
Hugh Dickinsc74df322005-10-29 18:16:23 -07001062 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001063 if (!pte)
1064 return -ENOMEM;
1065 do {
Nick Pigginb5810032005-10-29 18:16:12 -07001066 struct page *page = ZERO_PAGE(addr);
1067 pte_t zero_pte = pte_wrprotect(mk_pte(page, prot));
1068 page_cache_get(page);
1069 page_add_file_rmap(page);
1070 inc_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001071 BUG_ON(!pte_none(*pte));
1072 set_pte_at(mm, addr, pte, zero_pte);
1073 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001074 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 return 0;
1076}
1077
1078static inline int zeromap_pmd_range(struct mm_struct *mm, pud_t *pud,
1079 unsigned long addr, unsigned long end, pgprot_t prot)
1080{
1081 pmd_t *pmd;
1082 unsigned long next;
1083
1084 pmd = pmd_alloc(mm, pud, addr);
1085 if (!pmd)
1086 return -ENOMEM;
1087 do {
1088 next = pmd_addr_end(addr, end);
1089 if (zeromap_pte_range(mm, pmd, addr, next, prot))
1090 return -ENOMEM;
1091 } while (pmd++, addr = next, addr != end);
1092 return 0;
1093}
1094
1095static inline int zeromap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1096 unsigned long addr, unsigned long end, pgprot_t prot)
1097{
1098 pud_t *pud;
1099 unsigned long next;
1100
1101 pud = pud_alloc(mm, pgd, addr);
1102 if (!pud)
1103 return -ENOMEM;
1104 do {
1105 next = pud_addr_end(addr, end);
1106 if (zeromap_pmd_range(mm, pud, addr, next, prot))
1107 return -ENOMEM;
1108 } while (pud++, addr = next, addr != end);
1109 return 0;
1110}
1111
1112int zeromap_page_range(struct vm_area_struct *vma,
1113 unsigned long addr, unsigned long size, pgprot_t prot)
1114{
1115 pgd_t *pgd;
1116 unsigned long next;
1117 unsigned long end = addr + size;
1118 struct mm_struct *mm = vma->vm_mm;
1119 int err;
1120
1121 BUG_ON(addr >= end);
1122 pgd = pgd_offset(mm, addr);
1123 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 do {
1125 next = pgd_addr_end(addr, end);
1126 err = zeromap_pud_range(mm, pgd, addr, next, prot);
1127 if (err)
1128 break;
1129 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 return err;
1131}
1132
1133/*
1134 * maps a range of physical memory into the requested pages. the old
1135 * mappings are removed. any references to nonexistent pages results
1136 * in null mappings (currently treated as "copy-on-access")
1137 */
1138static int remap_pte_range(struct mm_struct *mm, pmd_t *pmd,
1139 unsigned long addr, unsigned long end,
1140 unsigned long pfn, pgprot_t prot)
1141{
1142 pte_t *pte;
Hugh Dickinsc74df322005-10-29 18:16:23 -07001143 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144
Hugh Dickinsc74df322005-10-29 18:16:23 -07001145 pte = pte_alloc_map_lock(mm, pmd, addr, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 if (!pte)
1147 return -ENOMEM;
1148 do {
1149 BUG_ON(!pte_none(*pte));
Nick Pigginb5810032005-10-29 18:16:12 -07001150 set_pte_at(mm, addr, pte, pfn_pte(pfn, prot));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151 pfn++;
1152 } while (pte++, addr += PAGE_SIZE, addr != end);
Hugh Dickinsc74df322005-10-29 18:16:23 -07001153 pte_unmap_unlock(pte - 1, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 return 0;
1155}
1156
1157static inline int remap_pmd_range(struct mm_struct *mm, pud_t *pud,
1158 unsigned long addr, unsigned long end,
1159 unsigned long pfn, pgprot_t prot)
1160{
1161 pmd_t *pmd;
1162 unsigned long next;
1163
1164 pfn -= addr >> PAGE_SHIFT;
1165 pmd = pmd_alloc(mm, pud, addr);
1166 if (!pmd)
1167 return -ENOMEM;
1168 do {
1169 next = pmd_addr_end(addr, end);
1170 if (remap_pte_range(mm, pmd, addr, next,
1171 pfn + (addr >> PAGE_SHIFT), prot))
1172 return -ENOMEM;
1173 } while (pmd++, addr = next, addr != end);
1174 return 0;
1175}
1176
1177static inline int remap_pud_range(struct mm_struct *mm, pgd_t *pgd,
1178 unsigned long addr, unsigned long end,
1179 unsigned long pfn, pgprot_t prot)
1180{
1181 pud_t *pud;
1182 unsigned long next;
1183
1184 pfn -= addr >> PAGE_SHIFT;
1185 pud = pud_alloc(mm, pgd, addr);
1186 if (!pud)
1187 return -ENOMEM;
1188 do {
1189 next = pud_addr_end(addr, end);
1190 if (remap_pmd_range(mm, pud, addr, next,
1191 pfn + (addr >> PAGE_SHIFT), prot))
1192 return -ENOMEM;
1193 } while (pud++, addr = next, addr != end);
1194 return 0;
1195}
1196
1197/* Note: this is only safe if the mm semaphore is held when called. */
1198int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1199 unsigned long pfn, unsigned long size, pgprot_t prot)
1200{
1201 pgd_t *pgd;
1202 unsigned long next;
Hugh Dickins2d15cab2005-06-25 14:54:33 -07001203 unsigned long end = addr + PAGE_ALIGN(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 struct mm_struct *mm = vma->vm_mm;
1205 int err;
1206
1207 /*
1208 * Physically remapped pages are special. Tell the
1209 * rest of the world about it:
1210 * VM_IO tells people not to look at these pages
1211 * (accesses can have side effects).
Hugh Dickins0b14c172005-11-21 21:32:15 -08001212 * VM_RESERVED is specified all over the place, because
1213 * in 2.4 it kept swapout's vma scan off this vma; but
1214 * in 2.6 the LRU scan won't even find its pages, so this
1215 * flag means no more than count its pages in reserved_vm,
1216 * and omit it from core dump, even when VM_IO turned off.
1217 * VM_UNPAGED tells the core MM not to "manage" these pages
1218 * (e.g. refcount, mapcount, try to swap them out): in
1219 * particular, zap_pte_range does not try to free them.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 */
Hugh Dickins0b14c172005-11-21 21:32:15 -08001221 vma->vm_flags |= VM_IO | VM_RESERVED | VM_UNPAGED;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222
1223 BUG_ON(addr >= end);
1224 pfn -= addr >> PAGE_SHIFT;
1225 pgd = pgd_offset(mm, addr);
1226 flush_cache_range(vma, addr, end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 do {
1228 next = pgd_addr_end(addr, end);
1229 err = remap_pud_range(mm, pgd, addr, next,
1230 pfn + (addr >> PAGE_SHIFT), prot);
1231 if (err)
1232 break;
1233 } while (pgd++, addr = next, addr != end);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234 return err;
1235}
1236EXPORT_SYMBOL(remap_pfn_range);
1237
1238/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001239 * handle_pte_fault chooses page fault handler according to an entry
1240 * which was read non-atomically. Before making any commitment, on
1241 * those architectures or configurations (e.g. i386 with PAE) which
1242 * might give a mix of unmatched parts, do_swap_page and do_file_page
1243 * must check under lock before unmapping the pte and proceeding
1244 * (but do_wp_page is only called after already making such a check;
1245 * and do_anonymous_page and do_no_page can safely check later on).
1246 */
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001247static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001248 pte_t *page_table, pte_t orig_pte)
1249{
1250 int same = 1;
1251#if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
1252 if (sizeof(pte_t) > sizeof(unsigned long)) {
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001253 spinlock_t *ptl = pte_lockptr(mm, pmd);
1254 spin_lock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001255 same = pte_same(*page_table, orig_pte);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001256 spin_unlock(ptl);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001257 }
1258#endif
1259 pte_unmap(page_table);
1260 return same;
1261}
1262
1263/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 * Do pte_mkwrite, but only if the vma says VM_WRITE. We do this when
1265 * servicing faults for write access. In the normal case, do always want
1266 * pte_mkwrite. But get_user_pages can cause write faults for mappings
1267 * that do not have writing enabled, when used by access_process_vm.
1268 */
1269static inline pte_t maybe_mkwrite(pte_t pte, struct vm_area_struct *vma)
1270{
1271 if (likely(vma->vm_flags & VM_WRITE))
1272 pte = pte_mkwrite(pte);
1273 return pte;
1274}
1275
1276/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 * This routine handles present pages, when users try to write
1278 * to a shared page. It is done by copying the page to a new address
1279 * and decrementing the shared-page counter for the old page.
1280 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001281 * Note that this routine assumes that the protection checks have been
1282 * done by the caller (the low-level page fault routine in most cases).
1283 * Thus we can safely just mark it writable once we've done any necessary
1284 * COW.
1285 *
1286 * We also mark the page dirty at this point even though the page will
1287 * change only once the write actually happens. This avoids a few races,
1288 * and potentially makes it more efficient.
1289 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001290 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1291 * but allow concurrent faults), with pte both mapped and locked.
1292 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 */
Hugh Dickins65500d22005-10-29 18:15:59 -07001294static int do_wp_page(struct mm_struct *mm, struct vm_area_struct *vma,
1295 unsigned long address, pte_t *page_table, pmd_t *pmd,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001296 spinlock_t *ptl, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297{
Hugh Dickins920fc352005-11-21 21:32:17 -08001298 struct page *old_page, *src_page, *new_page;
Hugh Dickins65500d22005-10-29 18:15:59 -07001299 unsigned long pfn = pte_pfn(orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 pte_t entry;
Hugh Dickins65500d22005-10-29 18:15:59 -07001301 int ret = VM_FAULT_MINOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302
1303 if (unlikely(!pfn_valid(pfn))) {
1304 /*
Hugh Dickins65500d22005-10-29 18:15:59 -07001305 * Page table corrupted: show pte and kill process.
Hugh Dickins920fc352005-11-21 21:32:17 -08001306 * Or it's an attempt to COW an out-of-map VM_UNPAGED
1307 * entry, which copy_user_highpage does not support.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 */
Nick Pigginb5810032005-10-29 18:16:12 -07001309 print_bad_pte(vma, orig_pte, address);
Hugh Dickins65500d22005-10-29 18:15:59 -07001310 ret = VM_FAULT_OOM;
1311 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 }
1313 old_page = pfn_to_page(pfn);
Hugh Dickins920fc352005-11-21 21:32:17 -08001314 src_page = old_page;
1315
Hugh Dickinsee498ed2005-11-21 21:32:18 -08001316 if (unlikely(vma->vm_flags & VM_UNPAGED))
1317 if (!page_is_anon(old_page, vma, address)) {
1318 old_page = NULL;
1319 goto gotten;
1320 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001321
Hugh Dickinsd296e9c2005-06-21 17:15:11 -07001322 if (PageAnon(old_page) && !TestSetPageLocked(old_page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323 int reuse = can_share_swap_page(old_page);
1324 unlock_page(old_page);
1325 if (reuse) {
1326 flush_cache_page(vma, address, pfn);
Hugh Dickins65500d22005-10-29 18:15:59 -07001327 entry = pte_mkyoung(orig_pte);
1328 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 ptep_set_access_flags(vma, address, page_table, entry, 1);
1330 update_mmu_cache(vma, address, entry);
1331 lazy_mmu_prot_update(entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07001332 ret |= VM_FAULT_WRITE;
1333 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 }
1335 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336
1337 /*
1338 * Ok, we need to copy. Oh, well..
1339 */
Nick Pigginb5810032005-10-29 18:16:12 -07001340 page_cache_get(old_page);
Hugh Dickins920fc352005-11-21 21:32:17 -08001341gotten:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001342 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343
1344 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07001345 goto oom;
Hugh Dickins920fc352005-11-21 21:32:17 -08001346 if (src_page == ZERO_PAGE(address)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 new_page = alloc_zeroed_user_highpage(vma, address);
1348 if (!new_page)
Hugh Dickins65500d22005-10-29 18:15:59 -07001349 goto oom;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 } else {
1351 new_page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1352 if (!new_page)
Hugh Dickins65500d22005-10-29 18:15:59 -07001353 goto oom;
Hugh Dickins920fc352005-11-21 21:32:17 -08001354 copy_user_highpage(new_page, src_page, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 }
Hugh Dickins65500d22005-10-29 18:15:59 -07001356
Linus Torvalds1da177e2005-04-16 15:20:36 -07001357 /*
1358 * Re-check the pte - we dropped the lock
1359 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001360 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins65500d22005-10-29 18:15:59 -07001361 if (likely(pte_same(*page_table, orig_pte))) {
Hugh Dickins920fc352005-11-21 21:32:17 -08001362 if (old_page) {
1363 page_remove_rmap(old_page);
1364 if (!PageAnon(old_page)) {
1365 dec_mm_counter(mm, file_rss);
1366 inc_mm_counter(mm, anon_rss);
1367 }
1368 } else
Hugh Dickins42946212005-10-29 18:16:05 -07001369 inc_mm_counter(mm, anon_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 flush_cache_page(vma, address, pfn);
Hugh Dickins65500d22005-10-29 18:15:59 -07001371 entry = mk_pte(new_page, vma->vm_page_prot);
1372 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1373 ptep_establish(vma, address, page_table, entry);
1374 update_mmu_cache(vma, address, entry);
1375 lazy_mmu_prot_update(entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 lru_cache_add_active(new_page);
1377 page_add_anon_rmap(new_page, vma, address);
1378
1379 /* Free the old page.. */
1380 new_page = old_page;
Nick Pigginf33ea7f2005-08-03 20:24:01 +10001381 ret |= VM_FAULT_WRITE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 }
Hugh Dickins920fc352005-11-21 21:32:17 -08001383 if (new_page)
1384 page_cache_release(new_page);
1385 if (old_page)
1386 page_cache_release(old_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001387unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001388 pte_unmap_unlock(page_table, ptl);
Nick Pigginf33ea7f2005-08-03 20:24:01 +10001389 return ret;
Hugh Dickins65500d22005-10-29 18:15:59 -07001390oom:
Hugh Dickins920fc352005-11-21 21:32:17 -08001391 if (old_page)
1392 page_cache_release(old_page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393 return VM_FAULT_OOM;
1394}
1395
1396/*
1397 * Helper functions for unmap_mapping_range().
1398 *
1399 * __ Notes on dropping i_mmap_lock to reduce latency while unmapping __
1400 *
1401 * We have to restart searching the prio_tree whenever we drop the lock,
1402 * since the iterator is only valid while the lock is held, and anyway
1403 * a later vma might be split and reinserted earlier while lock dropped.
1404 *
1405 * The list of nonlinear vmas could be handled more efficiently, using
1406 * a placeholder, but handle it in the same way until a need is shown.
1407 * It is important to search the prio_tree before nonlinear list: a vma
1408 * may become nonlinear and be shifted from prio_tree to nonlinear list
1409 * while the lock is dropped; but never shifted from list to prio_tree.
1410 *
1411 * In order to make forward progress despite restarting the search,
1412 * vm_truncate_count is used to mark a vma as now dealt with, so we can
1413 * quickly skip it next time around. Since the prio_tree search only
1414 * shows us those vmas affected by unmapping the range in question, we
1415 * can't efficiently keep all vmas in step with mapping->truncate_count:
1416 * so instead reset them all whenever it wraps back to 0 (then go to 1).
1417 * mapping->truncate_count and vma->vm_truncate_count are protected by
1418 * i_mmap_lock.
1419 *
1420 * In order to make forward progress despite repeatedly restarting some
Hugh Dickinsee39b372005-04-19 13:29:15 -07001421 * large vma, note the restart_addr from unmap_vmas when it breaks out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422 * and restart from that address when we reach that vma again. It might
1423 * have been split or merged, shrunk or extended, but never shifted: so
1424 * restart_addr remains valid so long as it remains in the vma's range.
1425 * unmap_mapping_range forces truncate_count to leap over page-aligned
1426 * values so we can save vma's restart_addr in its truncate_count field.
1427 */
1428#define is_restart_addr(truncate_count) (!((truncate_count) & ~PAGE_MASK))
1429
1430static void reset_vma_truncate_counts(struct address_space *mapping)
1431{
1432 struct vm_area_struct *vma;
1433 struct prio_tree_iter iter;
1434
1435 vma_prio_tree_foreach(vma, &iter, &mapping->i_mmap, 0, ULONG_MAX)
1436 vma->vm_truncate_count = 0;
1437 list_for_each_entry(vma, &mapping->i_mmap_nonlinear, shared.vm_set.list)
1438 vma->vm_truncate_count = 0;
1439}
1440
1441static int unmap_mapping_range_vma(struct vm_area_struct *vma,
1442 unsigned long start_addr, unsigned long end_addr,
1443 struct zap_details *details)
1444{
1445 unsigned long restart_addr;
1446 int need_break;
1447
1448again:
1449 restart_addr = vma->vm_truncate_count;
1450 if (is_restart_addr(restart_addr) && start_addr < restart_addr) {
1451 start_addr = restart_addr;
1452 if (start_addr >= end_addr) {
1453 /* Top of vma has been split off since last time */
1454 vma->vm_truncate_count = details->truncate_count;
1455 return 0;
1456 }
1457 }
1458
Hugh Dickinsee39b372005-04-19 13:29:15 -07001459 restart_addr = zap_page_range(vma, start_addr,
1460 end_addr - start_addr, details);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 need_break = need_resched() ||
1462 need_lockbreak(details->i_mmap_lock);
1463
Hugh Dickinsee39b372005-04-19 13:29:15 -07001464 if (restart_addr >= end_addr) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 /* We have now completed this vma: mark it so */
1466 vma->vm_truncate_count = details->truncate_count;
1467 if (!need_break)
1468 return 0;
1469 } else {
1470 /* Note restart_addr in vma's truncate_count field */
Hugh Dickinsee39b372005-04-19 13:29:15 -07001471 vma->vm_truncate_count = restart_addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 if (!need_break)
1473 goto again;
1474 }
1475
1476 spin_unlock(details->i_mmap_lock);
1477 cond_resched();
1478 spin_lock(details->i_mmap_lock);
1479 return -EINTR;
1480}
1481
1482static inline void unmap_mapping_range_tree(struct prio_tree_root *root,
1483 struct zap_details *details)
1484{
1485 struct vm_area_struct *vma;
1486 struct prio_tree_iter iter;
1487 pgoff_t vba, vea, zba, zea;
1488
1489restart:
1490 vma_prio_tree_foreach(vma, &iter, root,
1491 details->first_index, details->last_index) {
1492 /* Skip quickly over those we have already dealt with */
1493 if (vma->vm_truncate_count == details->truncate_count)
1494 continue;
1495
1496 vba = vma->vm_pgoff;
1497 vea = vba + ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT) - 1;
1498 /* Assume for now that PAGE_CACHE_SHIFT == PAGE_SHIFT */
1499 zba = details->first_index;
1500 if (zba < vba)
1501 zba = vba;
1502 zea = details->last_index;
1503 if (zea > vea)
1504 zea = vea;
1505
1506 if (unmap_mapping_range_vma(vma,
1507 ((zba - vba) << PAGE_SHIFT) + vma->vm_start,
1508 ((zea - vba + 1) << PAGE_SHIFT) + vma->vm_start,
1509 details) < 0)
1510 goto restart;
1511 }
1512}
1513
1514static inline void unmap_mapping_range_list(struct list_head *head,
1515 struct zap_details *details)
1516{
1517 struct vm_area_struct *vma;
1518
1519 /*
1520 * In nonlinear VMAs there is no correspondence between virtual address
1521 * offset and file offset. So we must perform an exhaustive search
1522 * across *all* the pages in each nonlinear VMA, not just the pages
1523 * whose virtual address lies outside the file truncation point.
1524 */
1525restart:
1526 list_for_each_entry(vma, head, shared.vm_set.list) {
1527 /* Skip quickly over those we have already dealt with */
1528 if (vma->vm_truncate_count == details->truncate_count)
1529 continue;
1530 details->nonlinear_vma = vma;
1531 if (unmap_mapping_range_vma(vma, vma->vm_start,
1532 vma->vm_end, details) < 0)
1533 goto restart;
1534 }
1535}
1536
1537/**
1538 * unmap_mapping_range - unmap the portion of all mmaps
1539 * in the specified address_space corresponding to the specified
1540 * page range in the underlying file.
Martin Waitz3d410882005-06-23 22:05:21 -07001541 * @mapping: the address space containing mmaps to be unmapped.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 * @holebegin: byte in first page to unmap, relative to the start of
1543 * the underlying file. This will be rounded down to a PAGE_SIZE
1544 * boundary. Note that this is different from vmtruncate(), which
1545 * must keep the partial page. In contrast, we must get rid of
1546 * partial pages.
1547 * @holelen: size of prospective hole in bytes. This will be rounded
1548 * up to a PAGE_SIZE boundary. A holelen of zero truncates to the
1549 * end of the file.
1550 * @even_cows: 1 when truncating a file, unmap even private COWed pages;
1551 * but 0 when invalidating pagecache, don't throw away private data.
1552 */
1553void unmap_mapping_range(struct address_space *mapping,
1554 loff_t const holebegin, loff_t const holelen, int even_cows)
1555{
1556 struct zap_details details;
1557 pgoff_t hba = holebegin >> PAGE_SHIFT;
1558 pgoff_t hlen = (holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1559
1560 /* Check for overflow. */
1561 if (sizeof(holelen) > sizeof(hlen)) {
1562 long long holeend =
1563 (holebegin + holelen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1564 if (holeend & ~(long long)ULONG_MAX)
1565 hlen = ULONG_MAX - hba + 1;
1566 }
1567
1568 details.check_mapping = even_cows? NULL: mapping;
1569 details.nonlinear_vma = NULL;
1570 details.first_index = hba;
1571 details.last_index = hba + hlen - 1;
1572 if (details.last_index < details.first_index)
1573 details.last_index = ULONG_MAX;
1574 details.i_mmap_lock = &mapping->i_mmap_lock;
1575
1576 spin_lock(&mapping->i_mmap_lock);
1577
1578 /* serialize i_size write against truncate_count write */
1579 smp_wmb();
1580 /* Protect against page faults, and endless unmapping loops */
1581 mapping->truncate_count++;
1582 /*
1583 * For archs where spin_lock has inclusive semantics like ia64
1584 * this smp_mb() will prevent to read pagetable contents
1585 * before the truncate_count increment is visible to
1586 * other cpus.
1587 */
1588 smp_mb();
1589 if (unlikely(is_restart_addr(mapping->truncate_count))) {
1590 if (mapping->truncate_count == 0)
1591 reset_vma_truncate_counts(mapping);
1592 mapping->truncate_count++;
1593 }
1594 details.truncate_count = mapping->truncate_count;
1595
1596 if (unlikely(!prio_tree_empty(&mapping->i_mmap)))
1597 unmap_mapping_range_tree(&mapping->i_mmap, &details);
1598 if (unlikely(!list_empty(&mapping->i_mmap_nonlinear)))
1599 unmap_mapping_range_list(&mapping->i_mmap_nonlinear, &details);
1600 spin_unlock(&mapping->i_mmap_lock);
1601}
1602EXPORT_SYMBOL(unmap_mapping_range);
1603
1604/*
1605 * Handle all mappings that got truncated by a "truncate()"
1606 * system call.
1607 *
1608 * NOTE! We have to be ready to update the memory sharing
1609 * between the file and the memory map for a potential last
1610 * incomplete page. Ugly, but necessary.
1611 */
1612int vmtruncate(struct inode * inode, loff_t offset)
1613{
1614 struct address_space *mapping = inode->i_mapping;
1615 unsigned long limit;
1616
1617 if (inode->i_size < offset)
1618 goto do_expand;
1619 /*
1620 * truncation of in-use swapfiles is disallowed - it would cause
1621 * subsequent swapout to scribble on the now-freed blocks.
1622 */
1623 if (IS_SWAPFILE(inode))
1624 goto out_busy;
1625 i_size_write(inode, offset);
1626 unmap_mapping_range(mapping, offset + PAGE_SIZE - 1, 0, 1);
1627 truncate_inode_pages(mapping, offset);
1628 goto out_truncate;
1629
1630do_expand:
1631 limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
1632 if (limit != RLIM_INFINITY && offset > limit)
1633 goto out_sig;
1634 if (offset > inode->i_sb->s_maxbytes)
1635 goto out_big;
1636 i_size_write(inode, offset);
1637
1638out_truncate:
1639 if (inode->i_op && inode->i_op->truncate)
1640 inode->i_op->truncate(inode);
1641 return 0;
1642out_sig:
1643 send_sig(SIGXFSZ, current, 0);
1644out_big:
1645 return -EFBIG;
1646out_busy:
1647 return -ETXTBSY;
1648}
1649
1650EXPORT_SYMBOL(vmtruncate);
1651
1652/*
1653 * Primitive swap readahead code. We simply read an aligned block of
1654 * (1 << page_cluster) entries in the swap area. This method is chosen
1655 * because it doesn't cost us any seek time. We also make sure to queue
1656 * the 'original' request together with the readahead ones...
1657 *
1658 * This has been extended to use the NUMA policies from the mm triggering
1659 * the readahead.
1660 *
1661 * Caller must hold down_read on the vma->vm_mm if vma is not NULL.
1662 */
1663void swapin_readahead(swp_entry_t entry, unsigned long addr,struct vm_area_struct *vma)
1664{
1665#ifdef CONFIG_NUMA
1666 struct vm_area_struct *next_vma = vma ? vma->vm_next : NULL;
1667#endif
1668 int i, num;
1669 struct page *new_page;
1670 unsigned long offset;
1671
1672 /*
1673 * Get the number of handles we should do readahead io to.
1674 */
1675 num = valid_swaphandles(entry, &offset);
1676 for (i = 0; i < num; offset++, i++) {
1677 /* Ok, do the async read-ahead now */
1678 new_page = read_swap_cache_async(swp_entry(swp_type(entry),
1679 offset), vma, addr);
1680 if (!new_page)
1681 break;
1682 page_cache_release(new_page);
1683#ifdef CONFIG_NUMA
1684 /*
1685 * Find the next applicable VMA for the NUMA policy.
1686 */
1687 addr += PAGE_SIZE;
1688 if (addr == 0)
1689 vma = NULL;
1690 if (vma) {
1691 if (addr >= vma->vm_end) {
1692 vma = next_vma;
1693 next_vma = vma ? vma->vm_next : NULL;
1694 }
1695 if (vma && addr < vma->vm_start)
1696 vma = NULL;
1697 } else {
1698 if (next_vma && addr >= next_vma->vm_start) {
1699 vma = next_vma;
1700 next_vma = vma->vm_next;
1701 }
1702 }
1703#endif
1704 }
1705 lru_add_drain(); /* Push any new pages onto the LRU now */
1706}
1707
1708/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001709 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1710 * but allow concurrent faults), and pte mapped but not yet locked.
1711 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 */
Hugh Dickins65500d22005-10-29 18:15:59 -07001713static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
1714 unsigned long address, pte_t *page_table, pmd_t *pmd,
1715 int write_access, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001717 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 struct page *page;
Hugh Dickins65500d22005-10-29 18:15:59 -07001719 swp_entry_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 pte_t pte;
1721 int ret = VM_FAULT_MINOR;
1722
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001723 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001724 goto out;
Hugh Dickins65500d22005-10-29 18:15:59 -07001725
1726 entry = pte_to_swp_entry(orig_pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 page = lookup_swap_cache(entry);
1728 if (!page) {
1729 swapin_readahead(entry, address, vma);
1730 page = read_swap_cache_async(entry, vma, address);
1731 if (!page) {
1732 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001733 * Back out if somebody else faulted in this pte
1734 * while we released the pte lock.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001736 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 if (likely(pte_same(*page_table, orig_pte)))
1738 ret = VM_FAULT_OOM;
Hugh Dickins65500d22005-10-29 18:15:59 -07001739 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 }
1741
1742 /* Had to read the page from swap area: Major fault */
1743 ret = VM_FAULT_MAJOR;
1744 inc_page_state(pgmajfault);
1745 grab_swap_token();
1746 }
1747
1748 mark_page_accessed(page);
1749 lock_page(page);
1750
1751 /*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001752 * Back out if somebody else already faulted in this pte.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753 */
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001754 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Hugh Dickins9e9bef02005-10-29 18:16:15 -07001755 if (unlikely(!pte_same(*page_table, orig_pte)))
Kirill Korotaevb8107482005-05-16 21:53:50 -07001756 goto out_nomap;
Kirill Korotaevb8107482005-05-16 21:53:50 -07001757
1758 if (unlikely(!PageUptodate(page))) {
1759 ret = VM_FAULT_SIGBUS;
1760 goto out_nomap;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 }
1762
1763 /* The page isn't present yet, go ahead with the fault. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764
Hugh Dickins42946212005-10-29 18:16:05 -07001765 inc_mm_counter(mm, anon_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 pte = mk_pte(page, vma->vm_page_prot);
1767 if (write_access && can_share_swap_page(page)) {
1768 pte = maybe_mkwrite(pte_mkdirty(pte), vma);
1769 write_access = 0;
1770 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771
1772 flush_icache_page(vma, page);
1773 set_pte_at(mm, address, page_table, pte);
1774 page_add_anon_rmap(page, vma, address);
1775
Hugh Dickinsc475a8a2005-06-21 17:15:12 -07001776 swap_free(entry);
1777 if (vm_swap_full())
1778 remove_exclusive_swap_page(page);
1779 unlock_page(page);
1780
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 if (write_access) {
1782 if (do_wp_page(mm, vma, address,
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001783 page_table, pmd, ptl, pte) == VM_FAULT_OOM)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 ret = VM_FAULT_OOM;
1785 goto out;
1786 }
1787
1788 /* No need to invalidate - it was non-present before */
1789 update_mmu_cache(vma, address, pte);
1790 lazy_mmu_prot_update(pte);
Hugh Dickins65500d22005-10-29 18:15:59 -07001791unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001792 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793out:
1794 return ret;
Kirill Korotaevb8107482005-05-16 21:53:50 -07001795out_nomap:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001796 pte_unmap_unlock(page_table, ptl);
Kirill Korotaevb8107482005-05-16 21:53:50 -07001797 unlock_page(page);
1798 page_cache_release(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001799 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800}
1801
1802/*
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001803 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1804 * but allow concurrent faults), and pte mapped but not yet locked.
1805 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 */
Hugh Dickins65500d22005-10-29 18:15:59 -07001807static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
1808 unsigned long address, pte_t *page_table, pmd_t *pmd,
1809 int write_access)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001811 struct page *page;
1812 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 pte_t entry;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08001815 /*
1816 * A VM_UNPAGED vma will normally be filled with present ptes
1817 * by remap_pfn_range, and never arrive here; but it might have
1818 * holes, or if !VM_DONTEXPAND, mremap might have expanded it.
1819 * It's weird enough handling anon pages in unpaged vmas, we do
1820 * not want to worry about ZERO_PAGEs too (it may or may not
1821 * matter if their counts wrap): just give them anon pages.
1822 */
1823
1824 if (write_access || (vma->vm_flags & VM_UNPAGED)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 /* Allocate our own private page. */
1826 pte_unmap(page_table);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827
1828 if (unlikely(anon_vma_prepare(vma)))
Hugh Dickins65500d22005-10-29 18:15:59 -07001829 goto oom;
1830 page = alloc_zeroed_user_highpage(vma, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 if (!page)
Hugh Dickins65500d22005-10-29 18:15:59 -07001832 goto oom;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
Hugh Dickins65500d22005-10-29 18:15:59 -07001834 entry = mk_pte(page, vma->vm_page_prot);
1835 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001836
1837 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
1838 if (!pte_none(*page_table))
1839 goto release;
1840 inc_mm_counter(mm, anon_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 lru_cache_add_active(page);
1842 SetPageReferenced(page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001843 page_add_anon_rmap(page, vma, address);
Nick Pigginb5810032005-10-29 18:16:12 -07001844 } else {
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001845 /* Map the ZERO_PAGE - vm_page_prot is readonly */
1846 page = ZERO_PAGE(address);
1847 page_cache_get(page);
1848 entry = mk_pte(page, vma->vm_page_prot);
1849
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07001850 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001851 spin_lock(ptl);
1852 if (!pte_none(*page_table))
1853 goto release;
Nick Pigginb5810032005-10-29 18:16:12 -07001854 inc_mm_counter(mm, file_rss);
1855 page_add_file_rmap(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 }
1857
Hugh Dickins65500d22005-10-29 18:15:59 -07001858 set_pte_at(mm, address, page_table, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859
1860 /* No need to invalidate - it was non-present before */
Hugh Dickins65500d22005-10-29 18:15:59 -07001861 update_mmu_cache(vma, address, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 lazy_mmu_prot_update(entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07001863unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001864 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 return VM_FAULT_MINOR;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001866release:
1867 page_cache_release(page);
1868 goto unlock;
Hugh Dickins65500d22005-10-29 18:15:59 -07001869oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 return VM_FAULT_OOM;
1871}
1872
1873/*
1874 * do_no_page() tries to create a new page mapping. It aggressively
1875 * tries to share with existing pages, but makes a separate copy if
1876 * the "write_access" parameter is true in order to avoid the next
1877 * page fault.
1878 *
1879 * As this is called only for pages that do not currently exist, we
1880 * do not need to flush old virtual caches or the TLB.
1881 *
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001882 * We enter with non-exclusive mmap_sem (to exclude vma changes,
1883 * but allow concurrent faults), and pte mapped but not yet locked.
1884 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885 */
Hugh Dickins65500d22005-10-29 18:15:59 -07001886static int do_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1887 unsigned long address, pte_t *page_table, pmd_t *pmd,
1888 int write_access)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889{
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001890 spinlock_t *ptl;
Hugh Dickins65500d22005-10-29 18:15:59 -07001891 struct page *new_page;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 struct address_space *mapping = NULL;
1893 pte_t entry;
1894 unsigned int sequence = 0;
1895 int ret = VM_FAULT_MINOR;
1896 int anon = 0;
1897
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 pte_unmap(page_table);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08001899 BUG_ON(vma->vm_flags & VM_UNPAGED);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900
1901 if (vma->vm_file) {
1902 mapping = vma->vm_file->f_mapping;
1903 sequence = mapping->truncate_count;
1904 smp_rmb(); /* serializes i_size against truncate_count */
1905 }
1906retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 new_page = vma->vm_ops->nopage(vma, address & PAGE_MASK, &ret);
1908 /*
1909 * No smp_rmb is needed here as long as there's a full
1910 * spin_lock/unlock sequence inside the ->nopage callback
1911 * (for the pagecache lookup) that acts as an implicit
1912 * smp_mb() and prevents the i_size read to happen
1913 * after the next truncate_count read.
1914 */
1915
1916 /* no page was available -- either SIGBUS or OOM */
1917 if (new_page == NOPAGE_SIGBUS)
1918 return VM_FAULT_SIGBUS;
1919 if (new_page == NOPAGE_OOM)
1920 return VM_FAULT_OOM;
1921
1922 /*
1923 * Should we do an early C-O-W break?
1924 */
1925 if (write_access && !(vma->vm_flags & VM_SHARED)) {
1926 struct page *page;
1927
1928 if (unlikely(anon_vma_prepare(vma)))
1929 goto oom;
1930 page = alloc_page_vma(GFP_HIGHUSER, vma, address);
1931 if (!page)
1932 goto oom;
1933 copy_user_highpage(page, new_page, address);
1934 page_cache_release(new_page);
1935 new_page = page;
1936 anon = 1;
1937 }
1938
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001939 page_table = pte_offset_map_lock(mm, pmd, address, &ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 /*
1941 * For a file-backed vma, someone could have truncated or otherwise
1942 * invalidated this page. If unmap_mapping_range got called,
1943 * retry getting the page.
1944 */
1945 if (mapping && unlikely(sequence != mapping->truncate_count)) {
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001946 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001948 cond_resched();
1949 sequence = mapping->truncate_count;
1950 smp_rmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 goto retry;
1952 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953
1954 /*
1955 * This silly early PAGE_DIRTY setting removes a race
1956 * due to the bad i386 page protection. But it's valid
1957 * for other architectures too.
1958 *
1959 * Note that if write_access is true, we either now have
1960 * an exclusive copy of the page, or this is a shared mapping,
1961 * so we can make it writable and dirty to avoid having to
1962 * handle that later.
1963 */
1964 /* Only go through if we didn't race with anybody else... */
1965 if (pte_none(*page_table)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 flush_icache_page(vma, new_page);
1967 entry = mk_pte(new_page, vma->vm_page_prot);
1968 if (write_access)
1969 entry = maybe_mkwrite(pte_mkdirty(entry), vma);
1970 set_pte_at(mm, address, page_table, entry);
1971 if (anon) {
Hugh Dickins42946212005-10-29 18:16:05 -07001972 inc_mm_counter(mm, anon_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 lru_cache_add_active(new_page);
1974 page_add_anon_rmap(new_page, vma, address);
Hugh Dickinsf57e88a2005-11-21 21:32:19 -08001975 } else {
Hugh Dickins42946212005-10-29 18:16:05 -07001976 inc_mm_counter(mm, file_rss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 page_add_file_rmap(new_page);
Hugh Dickins42946212005-10-29 18:16:05 -07001978 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 } else {
1980 /* One of our sibling threads was faster, back out. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001982 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 }
1984
1985 /* no need to invalidate: a not-present page shouldn't be cached */
1986 update_mmu_cache(vma, address, entry);
1987 lazy_mmu_prot_update(entry);
Hugh Dickins65500d22005-10-29 18:15:59 -07001988unlock:
Hugh Dickins8f4e2102005-10-29 18:16:26 -07001989 pte_unmap_unlock(page_table, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990 return ret;
1991oom:
1992 page_cache_release(new_page);
Hugh Dickins65500d22005-10-29 18:15:59 -07001993 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994}
1995
1996/*
1997 * Fault of a previously existing named mapping. Repopulate the pte
1998 * from the encoded file_pte if possible. This enables swappable
1999 * nonlinear vmas.
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002000 *
2001 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2002 * but allow concurrent faults), and pte mapped but not yet locked.
2003 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002005static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
2006 unsigned long address, pte_t *page_table, pmd_t *pmd,
2007 int write_access, pte_t orig_pte)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008{
Hugh Dickins65500d22005-10-29 18:15:59 -07002009 pgoff_t pgoff;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 int err;
2011
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002012 if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002013 return VM_FAULT_MINOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014
Hugh Dickins65500d22005-10-29 18:15:59 -07002015 if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
2016 /*
2017 * Page table corrupted: show pte and kill process.
2018 */
Nick Pigginb5810032005-10-29 18:16:12 -07002019 print_bad_pte(vma, orig_pte, address);
Hugh Dickins65500d22005-10-29 18:15:59 -07002020 return VM_FAULT_OOM;
2021 }
2022 /* We can then assume vm->vm_ops && vma->vm_ops->populate */
2023
2024 pgoff = pte_to_pgoff(orig_pte);
2025 err = vma->vm_ops->populate(vma, address & PAGE_MASK, PAGE_SIZE,
2026 vma->vm_page_prot, pgoff, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 if (err == -ENOMEM)
2028 return VM_FAULT_OOM;
2029 if (err)
2030 return VM_FAULT_SIGBUS;
2031 return VM_FAULT_MAJOR;
2032}
2033
2034/*
2035 * These routines also need to handle stuff like marking pages dirty
2036 * and/or accessed for architectures that don't do it in hardware (most
2037 * RISC architectures). The early dirtying is also good on the i386.
2038 *
2039 * There is also a hook called "update_mmu_cache()" that architectures
2040 * with external mmu caches can use to update those (ie the Sparc or
2041 * PowerPC hashed page tables that act as extended TLBs).
2042 *
Hugh Dickinsc74df322005-10-29 18:16:23 -07002043 * We enter with non-exclusive mmap_sem (to exclude vma changes,
2044 * but allow concurrent faults), and pte mapped but not yet locked.
2045 * We return with mmap_sem still held, but pte unmapped and unlocked.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046 */
2047static inline int handle_pte_fault(struct mm_struct *mm,
Hugh Dickins65500d22005-10-29 18:15:59 -07002048 struct vm_area_struct *vma, unsigned long address,
2049 pte_t *pte, pmd_t *pmd, int write_access)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050{
2051 pte_t entry;
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07002052 pte_t old_entry;
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002053 spinlock_t *ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07002055 old_entry = entry = *pte;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 if (!pte_present(entry)) {
Hugh Dickins65500d22005-10-29 18:15:59 -07002057 if (pte_none(entry)) {
2058 if (!vma->vm_ops || !vma->vm_ops->nopage)
2059 return do_anonymous_page(mm, vma, address,
2060 pte, pmd, write_access);
2061 return do_no_page(mm, vma, address,
2062 pte, pmd, write_access);
2063 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 if (pte_file(entry))
Hugh Dickins65500d22005-10-29 18:15:59 -07002065 return do_file_page(mm, vma, address,
2066 pte, pmd, write_access, entry);
2067 return do_swap_page(mm, vma, address,
2068 pte, pmd, write_access, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 }
2070
Hugh Dickins4c21e2f2005-10-29 18:16:40 -07002071 ptl = pte_lockptr(mm, pmd);
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002072 spin_lock(ptl);
2073 if (unlikely(!pte_same(*pte, entry)))
2074 goto unlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 if (write_access) {
2076 if (!pte_write(entry))
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002077 return do_wp_page(mm, vma, address,
2078 pte, pmd, ptl, entry);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079 entry = pte_mkdirty(entry);
2080 }
2081 entry = pte_mkyoung(entry);
Andrea Arcangeli1a44e142005-10-29 18:16:48 -07002082 if (!pte_same(old_entry, entry)) {
2083 ptep_set_access_flags(vma, address, pte, entry, write_access);
2084 update_mmu_cache(vma, address, entry);
2085 lazy_mmu_prot_update(entry);
2086 } else {
2087 /*
2088 * This is needed only for protection faults but the arch code
2089 * is not yet telling us if this is a protection fault or not.
2090 * This still avoids useless tlb flushes for .text page faults
2091 * with threads.
2092 */
2093 if (write_access)
2094 flush_tlb_page(vma, address);
2095 }
Hugh Dickins8f4e2102005-10-29 18:16:26 -07002096unlock:
2097 pte_unmap_unlock(pte, ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 return VM_FAULT_MINOR;
2099}
2100
2101/*
2102 * By the time we get here, we already hold the mm semaphore
2103 */
Hugh Dickins65500d22005-10-29 18:15:59 -07002104int __handle_mm_fault(struct mm_struct *mm, struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 unsigned long address, int write_access)
2106{
2107 pgd_t *pgd;
2108 pud_t *pud;
2109 pmd_t *pmd;
2110 pte_t *pte;
2111
2112 __set_current_state(TASK_RUNNING);
2113
2114 inc_page_state(pgfault);
2115
Hugh Dickinsac9b9c62005-10-20 16:24:28 +01002116 if (unlikely(is_vm_hugetlb_page(vma)))
2117 return hugetlb_fault(mm, vma, address, write_access);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 pgd = pgd_offset(mm, address);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 pud = pud_alloc(mm, pgd, address);
2121 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -07002122 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 pmd = pmd_alloc(mm, pud, address);
2124 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -07002125 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 pte = pte_alloc_map(mm, pmd, address);
2127 if (!pte)
Hugh Dickinsc74df322005-10-29 18:16:23 -07002128 return VM_FAULT_OOM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Hugh Dickinsc74df322005-10-29 18:16:23 -07002130 return handle_pte_fault(mm, vma, address, pte, pmd, write_access);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131}
2132
2133#ifndef __PAGETABLE_PUD_FOLDED
2134/*
2135 * Allocate page upper directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07002136 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07002138int __pud_alloc(struct mm_struct *mm, pgd_t *pgd, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139{
Hugh Dickinsc74df322005-10-29 18:16:23 -07002140 pud_t *new = pud_alloc_one(mm, address);
2141 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07002142 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143
Hugh Dickins872fec12005-10-29 18:16:21 -07002144 spin_lock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002145 if (pgd_present(*pgd)) /* Another has populated it */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 pud_free(new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002147 else
2148 pgd_populate(mm, pgd, new);
Hugh Dickinsc74df322005-10-29 18:16:23 -07002149 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002150 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151}
2152#endif /* __PAGETABLE_PUD_FOLDED */
2153
2154#ifndef __PAGETABLE_PMD_FOLDED
2155/*
2156 * Allocate page middle directory.
Hugh Dickins872fec12005-10-29 18:16:21 -07002157 * We've already handled the fast-path in-line.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 */
Hugh Dickins1bb36302005-10-29 18:16:22 -07002159int __pmd_alloc(struct mm_struct *mm, pud_t *pud, unsigned long address)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160{
Hugh Dickinsc74df322005-10-29 18:16:23 -07002161 pmd_t *new = pmd_alloc_one(mm, address);
2162 if (!new)
Hugh Dickins1bb36302005-10-29 18:16:22 -07002163 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164
Hugh Dickins872fec12005-10-29 18:16:21 -07002165 spin_lock(&mm->page_table_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166#ifndef __ARCH_HAS_4LEVEL_HACK
Hugh Dickins1bb36302005-10-29 18:16:22 -07002167 if (pud_present(*pud)) /* Another has populated it */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 pmd_free(new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002169 else
2170 pud_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171#else
Hugh Dickins1bb36302005-10-29 18:16:22 -07002172 if (pgd_present(*pud)) /* Another has populated it */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173 pmd_free(new);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002174 else
2175 pgd_populate(mm, pud, new);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176#endif /* __ARCH_HAS_4LEVEL_HACK */
Hugh Dickinsc74df322005-10-29 18:16:23 -07002177 spin_unlock(&mm->page_table_lock);
Hugh Dickins1bb36302005-10-29 18:16:22 -07002178 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179}
2180#endif /* __PAGETABLE_PMD_FOLDED */
2181
2182int make_pages_present(unsigned long addr, unsigned long end)
2183{
2184 int ret, len, write;
2185 struct vm_area_struct * vma;
2186
2187 vma = find_vma(current->mm, addr);
2188 if (!vma)
2189 return -1;
2190 write = (vma->vm_flags & VM_WRITE) != 0;
2191 if (addr >= end)
2192 BUG();
2193 if (end > vma->vm_end)
2194 BUG();
2195 len = (end+PAGE_SIZE-1)/PAGE_SIZE-addr/PAGE_SIZE;
2196 ret = get_user_pages(current, current->mm, addr,
2197 len, write, 0, NULL, NULL);
2198 if (ret < 0)
2199 return ret;
2200 return ret == len ? 0 : -1;
2201}
2202
2203/*
2204 * Map a vmalloc()-space virtual address to the physical page.
2205 */
2206struct page * vmalloc_to_page(void * vmalloc_addr)
2207{
2208 unsigned long addr = (unsigned long) vmalloc_addr;
2209 struct page *page = NULL;
2210 pgd_t *pgd = pgd_offset_k(addr);
2211 pud_t *pud;
2212 pmd_t *pmd;
2213 pte_t *ptep, pte;
2214
2215 if (!pgd_none(*pgd)) {
2216 pud = pud_offset(pgd, addr);
2217 if (!pud_none(*pud)) {
2218 pmd = pmd_offset(pud, addr);
2219 if (!pmd_none(*pmd)) {
2220 ptep = pte_offset_map(pmd, addr);
2221 pte = *ptep;
2222 if (pte_present(pte))
2223 page = pte_page(pte);
2224 pte_unmap(ptep);
2225 }
2226 }
2227 }
2228 return page;
2229}
2230
2231EXPORT_SYMBOL(vmalloc_to_page);
2232
2233/*
2234 * Map a vmalloc()-space virtual address to the physical page frame number.
2235 */
2236unsigned long vmalloc_to_pfn(void * vmalloc_addr)
2237{
2238 return page_to_pfn(vmalloc_to_page(vmalloc_addr));
2239}
2240
2241EXPORT_SYMBOL(vmalloc_to_pfn);
2242
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243#if !defined(__HAVE_ARCH_GATE_AREA)
2244
2245#if defined(AT_SYSINFO_EHDR)
Adrian Bunk5ce78522005-09-10 00:26:28 -07002246static struct vm_area_struct gate_vma;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247
2248static int __init gate_vma_init(void)
2249{
2250 gate_vma.vm_mm = NULL;
2251 gate_vma.vm_start = FIXADDR_USER_START;
2252 gate_vma.vm_end = FIXADDR_USER_END;
2253 gate_vma.vm_page_prot = PAGE_READONLY;
Hugh Dickins0b14c172005-11-21 21:32:15 -08002254 gate_vma.vm_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 return 0;
2256}
2257__initcall(gate_vma_init);
2258#endif
2259
2260struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
2261{
2262#ifdef AT_SYSINFO_EHDR
2263 return &gate_vma;
2264#else
2265 return NULL;
2266#endif
2267}
2268
2269int in_gate_area_no_task(unsigned long addr)
2270{
2271#ifdef AT_SYSINFO_EHDR
2272 if ((addr >= FIXADDR_USER_START) && (addr < FIXADDR_USER_END))
2273 return 1;
2274#endif
2275 return 0;
2276}
2277
2278#endif /* __HAVE_ARCH_GATE_AREA */