blob: 08e3c7f2bd15a1b2e502aeb3e9f509148cb5dfa8 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * mm/mremap.c
3 *
4 * (C) Copyright 1996 Linus Torvalds
5 *
6 * Address space accounting code <alan@redhat.com>
7 * (C) Copyright 2002 Red Hat Inc, All Rights Reserved
8 */
9
10#include <linux/mm.h>
11#include <linux/hugetlb.h>
12#include <linux/slab.h>
13#include <linux/shm.h>
14#include <linux/mman.h>
15#include <linux/swap.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080016#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/fs.h>
18#include <linux/highmem.h>
19#include <linux/security.h>
20#include <linux/syscalls.h>
21
22#include <asm/uaccess.h>
23#include <asm/cacheflush.h>
24#include <asm/tlbflush.h>
25
Hugh Dickins7be7a542005-10-29 18:16:00 -070026static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070027{
28 pgd_t *pgd;
29 pud_t *pud;
30 pmd_t *pmd;
31
32 pgd = pgd_offset(mm, addr);
33 if (pgd_none_or_clear_bad(pgd))
34 return NULL;
35
36 pud = pud_offset(pgd, addr);
37 if (pud_none_or_clear_bad(pud))
38 return NULL;
39
40 pmd = pmd_offset(pud, addr);
41 if (pmd_none_or_clear_bad(pmd))
42 return NULL;
43
Hugh Dickins7be7a542005-10-29 18:16:00 -070044 return pmd;
Linus Torvalds1da177e2005-04-16 15:20:36 -070045}
46
Hugh Dickins7be7a542005-10-29 18:16:00 -070047static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070048{
49 pgd_t *pgd;
50 pud_t *pud;
Hugh Dickinsc74df322005-10-29 18:16:23 -070051 pmd_t *pmd;
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
53 pgd = pgd_offset(mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 pud = pud_alloc(mm, pgd, addr);
55 if (!pud)
Hugh Dickinsc74df322005-10-29 18:16:23 -070056 return NULL;
Hugh Dickins7be7a542005-10-29 18:16:00 -070057
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 pmd = pmd_alloc(mm, pud, addr);
Hugh Dickins7be7a542005-10-29 18:16:00 -070059 if (!pmd)
Hugh Dickinsc74df322005-10-29 18:16:23 -070060 return NULL;
Hugh Dickins7be7a542005-10-29 18:16:00 -070061
Hugh Dickins1bb36302005-10-29 18:16:22 -070062 if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
Hugh Dickinsc74df322005-10-29 18:16:23 -070063 return NULL;
64
Hugh Dickins7be7a542005-10-29 18:16:00 -070065 return pmd;
Linus Torvalds1da177e2005-04-16 15:20:36 -070066}
67
Hugh Dickins7be7a542005-10-29 18:16:00 -070068static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
69 unsigned long old_addr, unsigned long old_end,
70 struct vm_area_struct *new_vma, pmd_t *new_pmd,
71 unsigned long new_addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -070072{
73 struct address_space *mapping = NULL;
74 struct mm_struct *mm = vma->vm_mm;
Hugh Dickins7be7a542005-10-29 18:16:00 -070075 pte_t *old_pte, *new_pte, pte;
Hugh Dickins4c21e2f2005-10-29 18:16:40 -070076 spinlock_t *old_ptl, *new_ptl;
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
78 if (vma->vm_file) {
79 /*
80 * Subtle point from Rajesh Venkatasubramanian: before
81 * moving file-based ptes, we must lock vmtruncate out,
82 * since it might clean the dst vma before the src vma,
83 * and we propagate stale pages into the dst afterward.
84 */
85 mapping = vma->vm_file->f_mapping;
86 spin_lock(&mapping->i_mmap_lock);
87 if (new_vma->vm_truncate_count &&
88 new_vma->vm_truncate_count != vma->vm_truncate_count)
89 new_vma->vm_truncate_count = 0;
90 }
Hugh Dickins7be7a542005-10-29 18:16:00 -070091
Hugh Dickins4c21e2f2005-10-29 18:16:40 -070092 /*
93 * We don't have to worry about the ordering of src and dst
94 * pte locks because exclusive mmap_sem prevents deadlock.
95 */
Hugh Dickinsc74df322005-10-29 18:16:23 -070096 old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
97 new_pte = pte_offset_map_nested(new_pmd, new_addr);
Hugh Dickins4c21e2f2005-10-29 18:16:40 -070098 new_ptl = pte_lockptr(mm, new_pmd);
99 if (new_ptl != old_ptl)
Ingo Molnarf20dc5f2006-07-03 00:25:08 -0700100 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700101 arch_enter_lazy_mmu_mode();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102
Hugh Dickins7be7a542005-10-29 18:16:00 -0700103 for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
104 new_pte++, new_addr += PAGE_SIZE) {
105 if (pte_none(*old_pte))
106 continue;
107 pte = ptep_clear_flush(vma, old_addr, old_pte);
Hugh Dickins7be7a542005-10-29 18:16:00 -0700108 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
109 set_pte_at(mm, new_addr, new_pte, pte);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110 }
Hugh Dickins7be7a542005-10-29 18:16:00 -0700111
Zachary Amsden6606c3e2006-09-30 23:29:33 -0700112 arch_leave_lazy_mmu_mode();
Hugh Dickins4c21e2f2005-10-29 18:16:40 -0700113 if (new_ptl != old_ptl)
114 spin_unlock(new_ptl);
Hugh Dickins7be7a542005-10-29 18:16:00 -0700115 pte_unmap_nested(new_pte - 1);
Hugh Dickinsc74df322005-10-29 18:16:23 -0700116 pte_unmap_unlock(old_pte - 1, old_ptl);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117 if (mapping)
118 spin_unlock(&mapping->i_mmap_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119}
120
Hugh Dickins7be7a542005-10-29 18:16:00 -0700121#define LATENCY_LIMIT (64 * PAGE_SIZE)
122
Ollie Wildb6a2fea2007-07-19 01:48:16 -0700123unsigned long move_page_tables(struct vm_area_struct *vma,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 unsigned long old_addr, struct vm_area_struct *new_vma,
125 unsigned long new_addr, unsigned long len)
126{
Hugh Dickins7be7a542005-10-29 18:16:00 -0700127 unsigned long extent, next, old_end;
128 pmd_t *old_pmd, *new_pmd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
Hugh Dickins7be7a542005-10-29 18:16:00 -0700130 old_end = old_addr + len;
131 flush_cache_range(vma, old_addr, old_end);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
Hugh Dickins7be7a542005-10-29 18:16:00 -0700133 for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134 cond_resched();
Hugh Dickins7be7a542005-10-29 18:16:00 -0700135 next = (old_addr + PMD_SIZE) & PMD_MASK;
136 if (next - 1 > old_end)
137 next = old_end;
138 extent = next - old_addr;
139 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
140 if (!old_pmd)
141 continue;
142 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
143 if (!new_pmd)
144 break;
145 next = (new_addr + PMD_SIZE) & PMD_MASK;
146 if (extent > next - new_addr)
147 extent = next - new_addr;
148 if (extent > LATENCY_LIMIT)
149 extent = LATENCY_LIMIT;
150 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
151 new_vma, new_pmd, new_addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700152 }
Hugh Dickins7be7a542005-10-29 18:16:00 -0700153
154 return len + old_addr - old_end; /* how much done */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155}
156
157static unsigned long move_vma(struct vm_area_struct *vma,
158 unsigned long old_addr, unsigned long old_len,
159 unsigned long new_len, unsigned long new_addr)
160{
161 struct mm_struct *mm = vma->vm_mm;
162 struct vm_area_struct *new_vma;
163 unsigned long vm_flags = vma->vm_flags;
164 unsigned long new_pgoff;
165 unsigned long moved_len;
166 unsigned long excess = 0;
Hugh Dickins365e9c872005-10-29 18:16:18 -0700167 unsigned long hiwater_vm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700168 int split = 0;
169
170 /*
171 * We'd prefer to avoid failure later on in do_munmap:
172 * which may split one vma into three before unmapping.
173 */
174 if (mm->map_count >= sysctl_max_map_count - 3)
175 return -ENOMEM;
176
177 new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
178 new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
179 if (!new_vma)
180 return -ENOMEM;
181
182 moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
183 if (moved_len < old_len) {
184 /*
185 * On error, move entries back from new area to old,
186 * which will succeed since page tables still there,
187 * and then proceed to unmap new area instead of old.
188 */
189 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
190 vma = new_vma;
191 old_len = new_len;
192 old_addr = new_addr;
193 new_addr = -ENOMEM;
194 }
195
196 /* Conceal VM_ACCOUNT so old reservation is not undone */
197 if (vm_flags & VM_ACCOUNT) {
198 vma->vm_flags &= ~VM_ACCOUNT;
199 excess = vma->vm_end - vma->vm_start - old_len;
200 if (old_addr > vma->vm_start &&
201 old_addr + old_len < vma->vm_end)
202 split = 1;
203 }
204
Kirill Korotaev71799062005-05-16 21:53:18 -0700205 /*
Hugh Dickins365e9c872005-10-29 18:16:18 -0700206 * If we failed to move page tables we still do total_vm increment
207 * since do_munmap() will decrement it by old_len == new_len.
208 *
209 * Since total_vm is about to be raised artificially high for a
210 * moment, we need to restore high watermark afterwards: if stats
211 * are taken meanwhile, total_vm and hiwater_vm appear too high.
212 * If this were a serious issue, we'd add a flag to do_munmap().
Kirill Korotaev71799062005-05-16 21:53:18 -0700213 */
Hugh Dickins365e9c872005-10-29 18:16:18 -0700214 hiwater_vm = mm->hiwater_vm;
Kirill Korotaev71799062005-05-16 21:53:18 -0700215 mm->total_vm += new_len >> PAGE_SHIFT;
Hugh Dickinsab50b8e2005-10-29 18:15:56 -0700216 vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
Kirill Korotaev71799062005-05-16 21:53:18 -0700217
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218 if (do_munmap(mm, old_addr, old_len) < 0) {
219 /* OOM: unable to split vma, just get accounts right */
220 vm_unacct_memory(excess >> PAGE_SHIFT);
221 excess = 0;
222 }
Hugh Dickins365e9c872005-10-29 18:16:18 -0700223 mm->hiwater_vm = hiwater_vm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700224
225 /* Restore VM_ACCOUNT if one or two pieces of vma left */
226 if (excess) {
227 vma->vm_flags |= VM_ACCOUNT;
228 if (split)
229 vma->vm_next->vm_flags |= VM_ACCOUNT;
230 }
231
Linus Torvalds1da177e2005-04-16 15:20:36 -0700232 if (vm_flags & VM_LOCKED) {
233 mm->locked_vm += new_len >> PAGE_SHIFT;
234 if (new_len > old_len)
235 make_pages_present(new_addr + old_len,
236 new_addr + new_len);
237 }
238
239 return new_addr;
240}
241
242/*
243 * Expand (or shrink) an existing mapping, potentially moving it at the
244 * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
245 *
246 * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
247 * This option implies MREMAP_MAYMOVE.
248 */
249unsigned long do_mremap(unsigned long addr,
250 unsigned long old_len, unsigned long new_len,
251 unsigned long flags, unsigned long new_addr)
252{
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700253 struct mm_struct *mm = current->mm;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 struct vm_area_struct *vma;
255 unsigned long ret = -EINVAL;
256 unsigned long charged = 0;
257
258 if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
259 goto out;
260
261 if (addr & ~PAGE_MASK)
262 goto out;
263
264 old_len = PAGE_ALIGN(old_len);
265 new_len = PAGE_ALIGN(new_len);
266
267 /*
268 * We allow a zero old-len as a special case
269 * for DOS-emu "duplicate shm area" thing. But
270 * a zero new-len is nonsensical.
271 */
272 if (!new_len)
273 goto out;
274
275 /* new_addr is only valid if MREMAP_FIXED is specified */
276 if (flags & MREMAP_FIXED) {
277 if (new_addr & ~PAGE_MASK)
278 goto out;
279 if (!(flags & MREMAP_MAYMOVE))
280 goto out;
281
282 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
283 goto out;
284
285 /* Check if the location we're moving into overlaps the
286 * old location at all, and fail if it does.
287 */
288 if ((new_addr <= addr) && (new_addr+new_len) > addr)
289 goto out;
290
291 if ((addr <= new_addr) && (addr+old_len) > new_addr)
292 goto out;
293
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700294 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
Eric Parised032182007-06-28 15:55:21 -0400295 if (ret)
296 goto out;
297
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700298 ret = do_munmap(mm, new_addr, new_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 if (ret)
300 goto out;
301 }
302
303 /*
304 * Always allow a shrinking remap: that just unmaps
305 * the unnecessary pages..
306 * do_munmap does all the needed commit accounting
307 */
308 if (old_len >= new_len) {
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700309 ret = do_munmap(mm, addr+new_len, old_len - new_len);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310 if (ret && old_len != new_len)
311 goto out;
312 ret = addr;
313 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
314 goto out;
315 old_len = new_len;
316 }
317
318 /*
319 * Ok, we need to grow.. or relocate.
320 */
321 ret = -EFAULT;
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700322 vma = find_vma(mm, addr);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 if (!vma || vma->vm_start > addr)
324 goto out;
325 if (is_vm_hugetlb_page(vma)) {
326 ret = -EINVAL;
327 goto out;
328 }
329 /* We can't remap across vm area boundaries */
330 if (old_len > vma->vm_end - addr)
331 goto out;
Linus Torvalds4d7672b2005-12-16 10:21:23 -0800332 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700333 if (new_len > old_len)
334 goto out;
335 }
336 if (vma->vm_flags & VM_LOCKED) {
337 unsigned long locked, lock_limit;
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700338 locked = mm->locked_vm << PAGE_SHIFT;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
340 locked += new_len - old_len;
341 ret = -EAGAIN;
342 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
343 goto out;
344 }
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700345 if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
akpm@osdl.org119f6572005-05-01 08:58:35 -0700346 ret = -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 goto out;
akpm@osdl.org119f6572005-05-01 08:58:35 -0700348 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349
350 if (vma->vm_flags & VM_ACCOUNT) {
351 charged = (new_len - old_len) >> PAGE_SHIFT;
352 if (security_vm_enough_memory(charged))
353 goto out_nc;
354 }
355
356 /* old_len exactly to the end of the area..
357 * And we're not relocating the area.
358 */
359 if (old_len == vma->vm_end - addr &&
360 !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
361 (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
362 unsigned long max_addr = TASK_SIZE;
363 if (vma->vm_next)
364 max_addr = vma->vm_next->vm_start;
365 /* can we just expand the current mapping? */
366 if (max_addr - addr >= new_len) {
367 int pages = (new_len - old_len) >> PAGE_SHIFT;
368
369 vma_adjust(vma, vma->vm_start,
370 addr + new_len, vma->vm_pgoff, NULL);
371
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700372 mm->total_vm += pages;
373 vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 if (vma->vm_flags & VM_LOCKED) {
Hugh Dickinsd0de32d2005-10-29 18:16:16 -0700375 mm->locked_vm += pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 make_pages_present(addr + old_len,
377 addr + new_len);
378 }
379 ret = addr;
380 goto out;
381 }
382 }
383
384 /*
385 * We weren't able to just expand or shrink the area,
386 * we need to create a new one and move it..
387 */
388 ret = -ENOMEM;
389 if (flags & MREMAP_MAYMOVE) {
390 if (!(flags & MREMAP_FIXED)) {
391 unsigned long map_flags = 0;
392 if (vma->vm_flags & VM_MAYSHARE)
393 map_flags |= MAP_SHARED;
394
395 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
396 vma->vm_pgoff, map_flags);
Eric Parised032182007-06-28 15:55:21 -0400397 if (new_addr & ~PAGE_MASK) {
398 ret = new_addr;
399 goto out;
400 }
401
Stephen Hemmingerc80544d2007-10-18 03:07:05 -0700402 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
Eric Parised032182007-06-28 15:55:21 -0400403 if (ret)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404 goto out;
405 }
406 ret = move_vma(vma, addr, old_len, new_len, new_addr);
407 }
408out:
409 if (ret & ~PAGE_MASK)
410 vm_unacct_memory(charged);
411out_nc:
412 return ret;
413}
414
415asmlinkage unsigned long sys_mremap(unsigned long addr,
416 unsigned long old_len, unsigned long new_len,
417 unsigned long flags, unsigned long new_addr)
418{
419 unsigned long ret;
420
421 down_write(&current->mm->mmap_sem);
422 ret = do_mremap(addr, old_len, new_len, flags, new_addr);
423 up_write(&current->mm->mmap_sem);
424 return ret;
425}