net: ipv6 bind to device issue
[linux-2.6-block.git] / mm / mremap.c
1 /*
2  *      mm/mremap.c
3  *
4  *      (C) Copyright 1996 Linus Torvalds
5  *
6  *      Address space accounting code   <alan@lxorguk.ukuu.org.uk>
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/shm.h>
13 #include <linux/ksm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
16 #include <linux/capability.h>
17 #include <linux/fs.h>
18 #include <linux/highmem.h>
19 #include <linux/security.h>
20 #include <linux/syscalls.h>
21 #include <linux/mmu_notifier.h>
22
23 #include <asm/uaccess.h>
24 #include <asm/cacheflush.h>
25 #include <asm/tlbflush.h>
26
27 #include "internal.h"
28
29 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
30 {
31         pgd_t *pgd;
32         pud_t *pud;
33         pmd_t *pmd;
34
35         pgd = pgd_offset(mm, addr);
36         if (pgd_none_or_clear_bad(pgd))
37                 return NULL;
38
39         pud = pud_offset(pgd, addr);
40         if (pud_none_or_clear_bad(pud))
41                 return NULL;
42
43         pmd = pmd_offset(pud, addr);
44         if (pmd_none_or_clear_bad(pmd))
45                 return NULL;
46
47         return pmd;
48 }
49
50 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
51 {
52         pgd_t *pgd;
53         pud_t *pud;
54         pmd_t *pmd;
55
56         pgd = pgd_offset(mm, addr);
57         pud = pud_alloc(mm, pgd, addr);
58         if (!pud)
59                 return NULL;
60
61         pmd = pmd_alloc(mm, pud, addr);
62         if (!pmd)
63                 return NULL;
64
65         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
66                 return NULL;
67
68         return pmd;
69 }
70
71 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
72                 unsigned long old_addr, unsigned long old_end,
73                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
74                 unsigned long new_addr)
75 {
76         struct address_space *mapping = NULL;
77         struct mm_struct *mm = vma->vm_mm;
78         pte_t *old_pte, *new_pte, pte;
79         spinlock_t *old_ptl, *new_ptl;
80         unsigned long old_start;
81
82         old_start = old_addr;
83         mmu_notifier_invalidate_range_start(vma->vm_mm,
84                                             old_start, old_end);
85         if (vma->vm_file) {
86                 /*
87                  * Subtle point from Rajesh Venkatasubramanian: before
88                  * moving file-based ptes, we must lock truncate_pagecache
89                  * out, since it might clean the dst vma before the src vma,
90                  * and we propagate stale pages into the dst afterward.
91                  */
92                 mapping = vma->vm_file->f_mapping;
93                 spin_lock(&mapping->i_mmap_lock);
94                 if (new_vma->vm_truncate_count &&
95                     new_vma->vm_truncate_count != vma->vm_truncate_count)
96                         new_vma->vm_truncate_count = 0;
97         }
98
99         /*
100          * We don't have to worry about the ordering of src and dst
101          * pte locks because exclusive mmap_sem prevents deadlock.
102          */
103         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
104         new_pte = pte_offset_map_nested(new_pmd, new_addr);
105         new_ptl = pte_lockptr(mm, new_pmd);
106         if (new_ptl != old_ptl)
107                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
108         arch_enter_lazy_mmu_mode();
109
110         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
111                                    new_pte++, new_addr += PAGE_SIZE) {
112                 if (pte_none(*old_pte))
113                         continue;
114                 pte = ptep_clear_flush(vma, old_addr, old_pte);
115                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
116                 set_pte_at(mm, new_addr, new_pte, pte);
117         }
118
119         arch_leave_lazy_mmu_mode();
120         if (new_ptl != old_ptl)
121                 spin_unlock(new_ptl);
122         pte_unmap_nested(new_pte - 1);
123         pte_unmap_unlock(old_pte - 1, old_ptl);
124         if (mapping)
125                 spin_unlock(&mapping->i_mmap_lock);
126         mmu_notifier_invalidate_range_end(vma->vm_mm, old_start, old_end);
127 }
128
129 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
130
131 unsigned long move_page_tables(struct vm_area_struct *vma,
132                 unsigned long old_addr, struct vm_area_struct *new_vma,
133                 unsigned long new_addr, unsigned long len)
134 {
135         unsigned long extent, next, old_end;
136         pmd_t *old_pmd, *new_pmd;
137
138         old_end = old_addr + len;
139         flush_cache_range(vma, old_addr, old_end);
140
141         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
142                 cond_resched();
143                 next = (old_addr + PMD_SIZE) & PMD_MASK;
144                 if (next - 1 > old_end)
145                         next = old_end;
146                 extent = next - old_addr;
147                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
148                 if (!old_pmd)
149                         continue;
150                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
151                 if (!new_pmd)
152                         break;
153                 next = (new_addr + PMD_SIZE) & PMD_MASK;
154                 if (extent > next - new_addr)
155                         extent = next - new_addr;
156                 if (extent > LATENCY_LIMIT)
157                         extent = LATENCY_LIMIT;
158                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
159                                 new_vma, new_pmd, new_addr);
160         }
161
162         return len + old_addr - old_end;        /* how much done */
163 }
164
165 static unsigned long move_vma(struct vm_area_struct *vma,
166                 unsigned long old_addr, unsigned long old_len,
167                 unsigned long new_len, unsigned long new_addr)
168 {
169         struct mm_struct *mm = vma->vm_mm;
170         struct vm_area_struct *new_vma;
171         unsigned long vm_flags = vma->vm_flags;
172         unsigned long new_pgoff;
173         unsigned long moved_len;
174         unsigned long excess = 0;
175         unsigned long hiwater_vm;
176         int split = 0;
177         int err;
178
179         /*
180          * We'd prefer to avoid failure later on in do_munmap:
181          * which may split one vma into three before unmapping.
182          */
183         if (mm->map_count >= sysctl_max_map_count - 3)
184                 return -ENOMEM;
185
186         /*
187          * Advise KSM to break any KSM pages in the area to be moved:
188          * it would be confusing if they were to turn up at the new
189          * location, where they happen to coincide with different KSM
190          * pages recently unmapped.  But leave vma->vm_flags as it was,
191          * so KSM can come around to merge on vma and new_vma afterwards.
192          */
193         err = ksm_madvise(vma, old_addr, old_addr + old_len,
194                                                 MADV_UNMERGEABLE, &vm_flags);
195         if (err)
196                 return err;
197
198         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
199         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
200         if (!new_vma)
201                 return -ENOMEM;
202
203         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
204         if (moved_len < old_len) {
205                 /*
206                  * On error, move entries back from new area to old,
207                  * which will succeed since page tables still there,
208                  * and then proceed to unmap new area instead of old.
209                  */
210                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
211                 vma = new_vma;
212                 old_len = new_len;
213                 old_addr = new_addr;
214                 new_addr = -ENOMEM;
215         }
216
217         /* Conceal VM_ACCOUNT so old reservation is not undone */
218         if (vm_flags & VM_ACCOUNT) {
219                 vma->vm_flags &= ~VM_ACCOUNT;
220                 excess = vma->vm_end - vma->vm_start - old_len;
221                 if (old_addr > vma->vm_start &&
222                     old_addr + old_len < vma->vm_end)
223                         split = 1;
224         }
225
226         /*
227          * If we failed to move page tables we still do total_vm increment
228          * since do_munmap() will decrement it by old_len == new_len.
229          *
230          * Since total_vm is about to be raised artificially high for a
231          * moment, we need to restore high watermark afterwards: if stats
232          * are taken meanwhile, total_vm and hiwater_vm appear too high.
233          * If this were a serious issue, we'd add a flag to do_munmap().
234          */
235         hiwater_vm = mm->hiwater_vm;
236         mm->total_vm += new_len >> PAGE_SHIFT;
237         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
238
239         if (do_munmap(mm, old_addr, old_len) < 0) {
240                 /* OOM: unable to split vma, just get accounts right */
241                 vm_unacct_memory(excess >> PAGE_SHIFT);
242                 excess = 0;
243         }
244         mm->hiwater_vm = hiwater_vm;
245
246         /* Restore VM_ACCOUNT if one or two pieces of vma left */
247         if (excess) {
248                 vma->vm_flags |= VM_ACCOUNT;
249                 if (split)
250                         vma->vm_next->vm_flags |= VM_ACCOUNT;
251         }
252
253         if (vm_flags & VM_LOCKED) {
254                 mm->locked_vm += new_len >> PAGE_SHIFT;
255                 if (new_len > old_len)
256                         mlock_vma_pages_range(new_vma, new_addr + old_len,
257                                                        new_addr + new_len);
258         }
259
260         return new_addr;
261 }
262
263 static struct vm_area_struct *vma_to_resize(unsigned long addr,
264         unsigned long old_len, unsigned long new_len, unsigned long *p)
265 {
266         struct mm_struct *mm = current->mm;
267         struct vm_area_struct *vma = find_vma(mm, addr);
268
269         if (!vma || vma->vm_start > addr)
270                 goto Efault;
271
272         if (is_vm_hugetlb_page(vma))
273                 goto Einval;
274
275         /* We can't remap across vm area boundaries */
276         if (old_len > vma->vm_end - addr)
277                 goto Efault;
278
279         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
280                 if (new_len > old_len)
281                         goto Efault;
282         }
283
284         if (vma->vm_flags & VM_LOCKED) {
285                 unsigned long locked, lock_limit;
286                 locked = mm->locked_vm << PAGE_SHIFT;
287                 lock_limit = rlimit(RLIMIT_MEMLOCK);
288                 locked += new_len - old_len;
289                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
290                         goto Eagain;
291         }
292
293         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
294                 goto Enomem;
295
296         if (vma->vm_flags & VM_ACCOUNT) {
297                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
298                 if (security_vm_enough_memory(charged))
299                         goto Efault;
300                 *p = charged;
301         }
302
303         return vma;
304
305 Efault: /* very odd choice for most of the cases, but... */
306         return ERR_PTR(-EFAULT);
307 Einval:
308         return ERR_PTR(-EINVAL);
309 Enomem:
310         return ERR_PTR(-ENOMEM);
311 Eagain:
312         return ERR_PTR(-EAGAIN);
313 }
314
315 static unsigned long mremap_to(unsigned long addr,
316         unsigned long old_len, unsigned long new_addr,
317         unsigned long new_len)
318 {
319         struct mm_struct *mm = current->mm;
320         struct vm_area_struct *vma;
321         unsigned long ret = -EINVAL;
322         unsigned long charged = 0;
323         unsigned long map_flags;
324
325         if (new_addr & ~PAGE_MASK)
326                 goto out;
327
328         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
329                 goto out;
330
331         /* Check if the location we're moving into overlaps the
332          * old location at all, and fail if it does.
333          */
334         if ((new_addr <= addr) && (new_addr+new_len) > addr)
335                 goto out;
336
337         if ((addr <= new_addr) && (addr+old_len) > new_addr)
338                 goto out;
339
340         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
341         if (ret)
342                 goto out;
343
344         ret = do_munmap(mm, new_addr, new_len);
345         if (ret)
346                 goto out;
347
348         if (old_len >= new_len) {
349                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
350                 if (ret && old_len != new_len)
351                         goto out;
352                 old_len = new_len;
353         }
354
355         vma = vma_to_resize(addr, old_len, new_len, &charged);
356         if (IS_ERR(vma)) {
357                 ret = PTR_ERR(vma);
358                 goto out;
359         }
360
361         map_flags = MAP_FIXED;
362         if (vma->vm_flags & VM_MAYSHARE)
363                 map_flags |= MAP_SHARED;
364
365         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
366                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
367                                 map_flags);
368         if (ret & ~PAGE_MASK)
369                 goto out1;
370
371         ret = move_vma(vma, addr, old_len, new_len, new_addr);
372         if (!(ret & ~PAGE_MASK))
373                 goto out;
374 out1:
375         vm_unacct_memory(charged);
376
377 out:
378         return ret;
379 }
380
381 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
382 {
383         unsigned long end = vma->vm_end + delta;
384         if (end < vma->vm_end) /* overflow */
385                 return 0;
386         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
387                 return 0;
388         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
389                               0, MAP_FIXED) & ~PAGE_MASK)
390                 return 0;
391         return 1;
392 }
393
394 /*
395  * Expand (or shrink) an existing mapping, potentially moving it at the
396  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
397  *
398  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
399  * This option implies MREMAP_MAYMOVE.
400  */
401 unsigned long do_mremap(unsigned long addr,
402         unsigned long old_len, unsigned long new_len,
403         unsigned long flags, unsigned long new_addr)
404 {
405         struct mm_struct *mm = current->mm;
406         struct vm_area_struct *vma;
407         unsigned long ret = -EINVAL;
408         unsigned long charged = 0;
409
410         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
411                 goto out;
412
413         if (addr & ~PAGE_MASK)
414                 goto out;
415
416         old_len = PAGE_ALIGN(old_len);
417         new_len = PAGE_ALIGN(new_len);
418
419         /*
420          * We allow a zero old-len as a special case
421          * for DOS-emu "duplicate shm area" thing. But
422          * a zero new-len is nonsensical.
423          */
424         if (!new_len)
425                 goto out;
426
427         if (flags & MREMAP_FIXED) {
428                 if (flags & MREMAP_MAYMOVE)
429                         ret = mremap_to(addr, old_len, new_addr, new_len);
430                 goto out;
431         }
432
433         /*
434          * Always allow a shrinking remap: that just unmaps
435          * the unnecessary pages..
436          * do_munmap does all the needed commit accounting
437          */
438         if (old_len >= new_len) {
439                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
440                 if (ret && old_len != new_len)
441                         goto out;
442                 ret = addr;
443                 goto out;
444         }
445
446         /*
447          * Ok, we need to grow..
448          */
449         vma = vma_to_resize(addr, old_len, new_len, &charged);
450         if (IS_ERR(vma)) {
451                 ret = PTR_ERR(vma);
452                 goto out;
453         }
454
455         /* old_len exactly to the end of the area..
456          */
457         if (old_len == vma->vm_end - addr) {
458                 /* can we just expand the current mapping? */
459                 if (vma_expandable(vma, new_len - old_len)) {
460                         int pages = (new_len - old_len) >> PAGE_SHIFT;
461
462                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
463                                        vma->vm_pgoff, NULL)) {
464                                 ret = -ENOMEM;
465                                 goto out;
466                         }
467
468                         mm->total_vm += pages;
469                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
470                         if (vma->vm_flags & VM_LOCKED) {
471                                 mm->locked_vm += pages;
472                                 mlock_vma_pages_range(vma, addr + old_len,
473                                                    addr + new_len);
474                         }
475                         ret = addr;
476                         goto out;
477                 }
478         }
479
480         /*
481          * We weren't able to just expand or shrink the area,
482          * we need to create a new one and move it..
483          */
484         ret = -ENOMEM;
485         if (flags & MREMAP_MAYMOVE) {
486                 unsigned long map_flags = 0;
487                 if (vma->vm_flags & VM_MAYSHARE)
488                         map_flags |= MAP_SHARED;
489
490                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
491                                         vma->vm_pgoff +
492                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
493                                         map_flags);
494                 if (new_addr & ~PAGE_MASK) {
495                         ret = new_addr;
496                         goto out;
497                 }
498
499                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
500                 if (ret)
501                         goto out;
502                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
503         }
504 out:
505         if (ret & ~PAGE_MASK)
506                 vm_unacct_memory(charged);
507         return ret;
508 }
509
510 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
511                 unsigned long, new_len, unsigned long, flags,
512                 unsigned long, new_addr)
513 {
514         unsigned long ret;
515
516         down_write(&current->mm->mmap_sem);
517         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
518         up_write(&current->mm->mmap_sem);
519         return ret;
520 }