[PATCH] move capable() to capability.h
[linux-2.6.git] / mm / mremap.c
1 /*
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>
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
22 #include <asm/uaccess.h>
23 #include <asm/cacheflush.h>
24 #include <asm/tlbflush.h>
25
26 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
27 {
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
44         return pmd;
45 }
46
47 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
48 {
49         pgd_t *pgd;
50         pud_t *pud;
51         pmd_t *pmd;
52
53         pgd = pgd_offset(mm, addr);
54         pud = pud_alloc(mm, pgd, addr);
55         if (!pud)
56                 return NULL;
57
58         pmd = pmd_alloc(mm, pud, addr);
59         if (!pmd)
60                 return NULL;
61
62         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
63                 return NULL;
64
65         return pmd;
66 }
67
68 static 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)
72 {
73         struct address_space *mapping = NULL;
74         struct mm_struct *mm = vma->vm_mm;
75         pte_t *old_pte, *new_pte, pte;
76         spinlock_t *old_ptl, *new_ptl;
77
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         }
91
92         /*
93          * We don't have to worry about the ordering of src and dst
94          * pte locks because exclusive mmap_sem prevents deadlock.
95          */
96         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
97         new_pte = pte_offset_map_nested(new_pmd, new_addr);
98         new_ptl = pte_lockptr(mm, new_pmd);
99         if (new_ptl != old_ptl)
100                 spin_lock(new_ptl);
101
102         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
103                                    new_pte++, new_addr += PAGE_SIZE) {
104                 if (pte_none(*old_pte))
105                         continue;
106                 pte = ptep_clear_flush(vma, old_addr, old_pte);
107                 /* ZERO_PAGE can be dependant on virtual addr */
108                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
109                 set_pte_at(mm, new_addr, new_pte, pte);
110         }
111
112         if (new_ptl != old_ptl)
113                 spin_unlock(new_ptl);
114         pte_unmap_nested(new_pte - 1);
115         pte_unmap_unlock(old_pte - 1, old_ptl);
116         if (mapping)
117                 spin_unlock(&mapping->i_mmap_lock);
118 }
119
120 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
121
122 static unsigned long move_page_tables(struct vm_area_struct *vma,
123                 unsigned long old_addr, struct vm_area_struct *new_vma,
124                 unsigned long new_addr, unsigned long len)
125 {
126         unsigned long extent, next, old_end;
127         pmd_t *old_pmd, *new_pmd;
128
129         old_end = old_addr + len;
130         flush_cache_range(vma, old_addr, old_end);
131
132         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
133                 cond_resched();
134                 next = (old_addr + PMD_SIZE) & PMD_MASK;
135                 if (next - 1 > old_end)
136                         next = old_end;
137                 extent = next - old_addr;
138                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
139                 if (!old_pmd)
140                         continue;
141                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
142                 if (!new_pmd)
143                         break;
144                 next = (new_addr + PMD_SIZE) & PMD_MASK;
145                 if (extent > next - new_addr)
146                         extent = next - new_addr;
147                 if (extent > LATENCY_LIMIT)
148                         extent = LATENCY_LIMIT;
149                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
150                                 new_vma, new_pmd, new_addr);
151         }
152
153         return len + old_addr - old_end;        /* how much done */
154 }
155
156 static unsigned long move_vma(struct vm_area_struct *vma,
157                 unsigned long old_addr, unsigned long old_len,
158                 unsigned long new_len, unsigned long new_addr)
159 {
160         struct mm_struct *mm = vma->vm_mm;
161         struct vm_area_struct *new_vma;
162         unsigned long vm_flags = vma->vm_flags;
163         unsigned long new_pgoff;
164         unsigned long moved_len;
165         unsigned long excess = 0;
166         unsigned long hiwater_vm;
167         int split = 0;
168
169         /*
170          * We'd prefer to avoid failure later on in do_munmap:
171          * which may split one vma into three before unmapping.
172          */
173         if (mm->map_count >= sysctl_max_map_count - 3)
174                 return -ENOMEM;
175
176         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
177         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
178         if (!new_vma)
179                 return -ENOMEM;
180
181         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
182         if (moved_len < old_len) {
183                 /*
184                  * On error, move entries back from new area to old,
185                  * which will succeed since page tables still there,
186                  * and then proceed to unmap new area instead of old.
187                  */
188                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
189                 vma = new_vma;
190                 old_len = new_len;
191                 old_addr = new_addr;
192                 new_addr = -ENOMEM;
193         }
194
195         /* Conceal VM_ACCOUNT so old reservation is not undone */
196         if (vm_flags & VM_ACCOUNT) {
197                 vma->vm_flags &= ~VM_ACCOUNT;
198                 excess = vma->vm_end - vma->vm_start - old_len;
199                 if (old_addr > vma->vm_start &&
200                     old_addr + old_len < vma->vm_end)
201                         split = 1;
202         }
203
204         /*
205          * If we failed to move page tables we still do total_vm increment
206          * since do_munmap() will decrement it by old_len == new_len.
207          *
208          * Since total_vm is about to be raised artificially high for a
209          * moment, we need to restore high watermark afterwards: if stats
210          * are taken meanwhile, total_vm and hiwater_vm appear too high.
211          * If this were a serious issue, we'd add a flag to do_munmap().
212          */
213         hiwater_vm = mm->hiwater_vm;
214         mm->total_vm += new_len >> PAGE_SHIFT;
215         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
216
217         if (do_munmap(mm, old_addr, old_len) < 0) {
218                 /* OOM: unable to split vma, just get accounts right */
219                 vm_unacct_memory(excess >> PAGE_SHIFT);
220                 excess = 0;
221         }
222         mm->hiwater_vm = hiwater_vm;
223
224         /* Restore VM_ACCOUNT if one or two pieces of vma left */
225         if (excess) {
226                 vma->vm_flags |= VM_ACCOUNT;
227                 if (split)
228                         vma->vm_next->vm_flags |= VM_ACCOUNT;
229         }
230
231         if (vm_flags & VM_LOCKED) {
232                 mm->locked_vm += new_len >> PAGE_SHIFT;
233                 if (new_len > old_len)
234                         make_pages_present(new_addr + old_len,
235                                            new_addr + new_len);
236         }
237
238         return new_addr;
239 }
240
241 /*
242  * Expand (or shrink) an existing mapping, potentially moving it at the
243  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
244  *
245  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
246  * This option implies MREMAP_MAYMOVE.
247  */
248 unsigned long do_mremap(unsigned long addr,
249         unsigned long old_len, unsigned long new_len,
250         unsigned long flags, unsigned long new_addr)
251 {
252         struct mm_struct *mm = current->mm;
253         struct vm_area_struct *vma;
254         unsigned long ret = -EINVAL;
255         unsigned long charged = 0;
256
257         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
258                 goto out;
259
260         if (addr & ~PAGE_MASK)
261                 goto out;
262
263         old_len = PAGE_ALIGN(old_len);
264         new_len = PAGE_ALIGN(new_len);
265
266         /*
267          * We allow a zero old-len as a special case
268          * for DOS-emu "duplicate shm area" thing. But
269          * a zero new-len is nonsensical.
270          */
271         if (!new_len)
272                 goto out;
273
274         /* new_addr is only valid if MREMAP_FIXED is specified */
275         if (flags & MREMAP_FIXED) {
276                 if (new_addr & ~PAGE_MASK)
277                         goto out;
278                 if (!(flags & MREMAP_MAYMOVE))
279                         goto out;
280
281                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
282                         goto out;
283
284                 /* Check if the location we're moving into overlaps the
285                  * old location at all, and fail if it does.
286                  */
287                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
288                         goto out;
289
290                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
291                         goto out;
292
293                 ret = do_munmap(mm, new_addr, new_len);
294                 if (ret)
295                         goto out;
296         }
297
298         /*
299          * Always allow a shrinking remap: that just unmaps
300          * the unnecessary pages..
301          * do_munmap does all the needed commit accounting
302          */
303         if (old_len >= new_len) {
304                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
305                 if (ret && old_len != new_len)
306                         goto out;
307                 ret = addr;
308                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
309                         goto out;
310                 old_len = new_len;
311         }
312
313         /*
314          * Ok, we need to grow..  or relocate.
315          */
316         ret = -EFAULT;
317         vma = find_vma(mm, addr);
318         if (!vma || vma->vm_start > addr)
319                 goto out;
320         if (is_vm_hugetlb_page(vma)) {
321                 ret = -EINVAL;
322                 goto out;
323         }
324         /* We can't remap across vm area boundaries */
325         if (old_len > vma->vm_end - addr)
326                 goto out;
327         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
328                 if (new_len > old_len)
329                         goto out;
330         }
331         if (vma->vm_flags & VM_LOCKED) {
332                 unsigned long locked, lock_limit;
333                 locked = mm->locked_vm << PAGE_SHIFT;
334                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
335                 locked += new_len - old_len;
336                 ret = -EAGAIN;
337                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
338                         goto out;
339         }
340         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
341                 ret = -ENOMEM;
342                 goto out;
343         }
344
345         if (vma->vm_flags & VM_ACCOUNT) {
346                 charged = (new_len - old_len) >> PAGE_SHIFT;
347                 if (security_vm_enough_memory(charged))
348                         goto out_nc;
349         }
350
351         /* old_len exactly to the end of the area..
352          * And we're not relocating the area.
353          */
354         if (old_len == vma->vm_end - addr &&
355             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
356             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
357                 unsigned long max_addr = TASK_SIZE;
358                 if (vma->vm_next)
359                         max_addr = vma->vm_next->vm_start;
360                 /* can we just expand the current mapping? */
361                 if (max_addr - addr >= new_len) {
362                         int pages = (new_len - old_len) >> PAGE_SHIFT;
363
364                         vma_adjust(vma, vma->vm_start,
365                                 addr + new_len, vma->vm_pgoff, NULL);
366
367                         mm->total_vm += pages;
368                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
369                         if (vma->vm_flags & VM_LOCKED) {
370                                 mm->locked_vm += pages;
371                                 make_pages_present(addr + old_len,
372                                                    addr + new_len);
373                         }
374                         ret = addr;
375                         goto out;
376                 }
377         }
378
379         /*
380          * We weren't able to just expand or shrink the area,
381          * we need to create a new one and move it..
382          */
383         ret = -ENOMEM;
384         if (flags & MREMAP_MAYMOVE) {
385                 if (!(flags & MREMAP_FIXED)) {
386                         unsigned long map_flags = 0;
387                         if (vma->vm_flags & VM_MAYSHARE)
388                                 map_flags |= MAP_SHARED;
389
390                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
391                                                 vma->vm_pgoff, map_flags);
392                         ret = new_addr;
393                         if (new_addr & ~PAGE_MASK)
394                                 goto out;
395                 }
396                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
397         }
398 out:
399         if (ret & ~PAGE_MASK)
400                 vm_unacct_memory(charged);
401 out_nc:
402         return ret;
403 }
404
405 asmlinkage unsigned long sys_mremap(unsigned long addr,
406         unsigned long old_len, unsigned long new_len,
407         unsigned long flags, unsigned long new_addr)
408 {
409         unsigned long ret;
410
411         down_write(&current->mm->mmap_sem);
412         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
413         up_write(&current->mm->mmap_sem);
414         return ret;
415 }