[PATCH] fix VmSize and VmData after mremap
[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/fs.h>
17 #include <linux/highmem.h>
18 #include <linux/security.h>
19 #include <linux/syscalls.h>
20
21 #include <asm/uaccess.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
24
25 static pte_t *get_one_pte_map_nested(struct mm_struct *mm, unsigned long addr)
26 {
27         pgd_t *pgd;
28         pud_t *pud;
29         pmd_t *pmd;
30         pte_t *pte = NULL;
31
32         pgd = pgd_offset(mm, addr);
33         if (pgd_none_or_clear_bad(pgd))
34                 goto end;
35
36         pud = pud_offset(pgd, addr);
37         if (pud_none_or_clear_bad(pud))
38                 goto end;
39
40         pmd = pmd_offset(pud, addr);
41         if (pmd_none_or_clear_bad(pmd))
42                 goto end;
43
44         pte = pte_offset_map_nested(pmd, addr);
45         if (pte_none(*pte)) {
46                 pte_unmap_nested(pte);
47                 pte = NULL;
48         }
49 end:
50         return pte;
51 }
52
53 static pte_t *get_one_pte_map(struct mm_struct *mm, unsigned long addr)
54 {
55         pgd_t *pgd;
56         pud_t *pud;
57         pmd_t *pmd;
58
59         pgd = pgd_offset(mm, addr);
60         if (pgd_none_or_clear_bad(pgd))
61                 return NULL;
62
63         pud = pud_offset(pgd, addr);
64         if (pud_none_or_clear_bad(pud))
65                 return NULL;
66
67         pmd = pmd_offset(pud, addr);
68         if (pmd_none_or_clear_bad(pmd))
69                 return NULL;
70
71         return pte_offset_map(pmd, addr);
72 }
73
74 static inline pte_t *alloc_one_pte_map(struct mm_struct *mm, unsigned long addr)
75 {
76         pgd_t *pgd;
77         pud_t *pud;
78         pmd_t *pmd;
79         pte_t *pte = NULL;
80
81         pgd = pgd_offset(mm, addr);
82
83         pud = pud_alloc(mm, pgd, addr);
84         if (!pud)
85                 return NULL;
86         pmd = pmd_alloc(mm, pud, addr);
87         if (pmd)
88                 pte = pte_alloc_map(mm, pmd, addr);
89         return pte;
90 }
91
92 static int
93 move_one_page(struct vm_area_struct *vma, unsigned long old_addr,
94                 struct vm_area_struct *new_vma, unsigned long new_addr)
95 {
96         struct address_space *mapping = NULL;
97         struct mm_struct *mm = vma->vm_mm;
98         int error = 0;
99         pte_t *src, *dst;
100
101         if (vma->vm_file) {
102                 /*
103                  * Subtle point from Rajesh Venkatasubramanian: before
104                  * moving file-based ptes, we must lock vmtruncate out,
105                  * since it might clean the dst vma before the src vma,
106                  * and we propagate stale pages into the dst afterward.
107                  */
108                 mapping = vma->vm_file->f_mapping;
109                 spin_lock(&mapping->i_mmap_lock);
110                 if (new_vma->vm_truncate_count &&
111                     new_vma->vm_truncate_count != vma->vm_truncate_count)
112                         new_vma->vm_truncate_count = 0;
113         }
114         spin_lock(&mm->page_table_lock);
115
116         src = get_one_pte_map_nested(mm, old_addr);
117         if (src) {
118                 /*
119                  * Look to see whether alloc_one_pte_map needs to perform a
120                  * memory allocation.  If it does then we need to drop the
121                  * atomic kmap
122                  */
123                 dst = get_one_pte_map(mm, new_addr);
124                 if (unlikely(!dst)) {
125                         pte_unmap_nested(src);
126                         if (mapping)
127                                 spin_unlock(&mapping->i_mmap_lock);
128                         dst = alloc_one_pte_map(mm, new_addr);
129                         if (mapping && !spin_trylock(&mapping->i_mmap_lock)) {
130                                 spin_unlock(&mm->page_table_lock);
131                                 spin_lock(&mapping->i_mmap_lock);
132                                 spin_lock(&mm->page_table_lock);
133                         }
134                         src = get_one_pte_map_nested(mm, old_addr);
135                 }
136                 /*
137                  * Since alloc_one_pte_map can drop and re-acquire
138                  * page_table_lock, we should re-check the src entry...
139                  */
140                 if (src) {
141                         if (dst) {
142                                 pte_t pte;
143                                 pte = ptep_clear_flush(vma, old_addr, src);
144                                 set_pte_at(mm, new_addr, dst, pte);
145                         } else
146                                 error = -ENOMEM;
147                         pte_unmap_nested(src);
148                 }
149                 if (dst)
150                         pte_unmap(dst);
151         }
152         spin_unlock(&mm->page_table_lock);
153         if (mapping)
154                 spin_unlock(&mapping->i_mmap_lock);
155         return error;
156 }
157
158 static unsigned long move_page_tables(struct vm_area_struct *vma,
159                 unsigned long old_addr, struct vm_area_struct *new_vma,
160                 unsigned long new_addr, unsigned long len)
161 {
162         unsigned long offset;
163
164         flush_cache_range(vma, old_addr, old_addr + len);
165
166         /*
167          * This is not the clever way to do this, but we're taking the
168          * easy way out on the assumption that most remappings will be
169          * only a few pages.. This also makes error recovery easier.
170          */
171         for (offset = 0; offset < len; offset += PAGE_SIZE) {
172                 if (move_one_page(vma, old_addr + offset,
173                                 new_vma, new_addr + offset) < 0)
174                         break;
175                 cond_resched();
176         }
177         return offset;
178 }
179
180 static unsigned long move_vma(struct vm_area_struct *vma,
181                 unsigned long old_addr, unsigned long old_len,
182                 unsigned long new_len, unsigned long new_addr)
183 {
184         struct mm_struct *mm = vma->vm_mm;
185         struct vm_area_struct *new_vma;
186         unsigned long vm_flags = vma->vm_flags;
187         unsigned long new_pgoff;
188         unsigned long moved_len;
189         unsigned long excess = 0;
190         int split = 0;
191
192         /*
193          * We'd prefer to avoid failure later on in do_munmap:
194          * which may split one vma into three before unmapping.
195          */
196         if (mm->map_count >= sysctl_max_map_count - 3)
197                 return -ENOMEM;
198
199         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
200         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
201         if (!new_vma)
202                 return -ENOMEM;
203
204         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
205         if (moved_len < old_len) {
206                 /*
207                  * On error, move entries back from new area to old,
208                  * which will succeed since page tables still there,
209                  * and then proceed to unmap new area instead of old.
210                  */
211                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
212                 vma = new_vma;
213                 old_len = new_len;
214                 old_addr = new_addr;
215                 new_addr = -ENOMEM;
216         }
217
218         /* Conceal VM_ACCOUNT so old reservation is not undone */
219         if (vm_flags & VM_ACCOUNT) {
220                 vma->vm_flags &= ~VM_ACCOUNT;
221                 excess = vma->vm_end - vma->vm_start - old_len;
222                 if (old_addr > vma->vm_start &&
223                     old_addr + old_len < vma->vm_end)
224                         split = 1;
225         }
226
227         /*
228          * if we failed to move page tables we still do total_vm increment
229          * since do_munmap() will decrement it by old_len == new_len
230          */
231         mm->total_vm += new_len >> PAGE_SHIFT;
232         __vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
233
234         if (do_munmap(mm, old_addr, old_len) < 0) {
235                 /* OOM: unable to split vma, just get accounts right */
236                 vm_unacct_memory(excess >> PAGE_SHIFT);
237                 excess = 0;
238         }
239
240         /* Restore VM_ACCOUNT if one or two pieces of vma left */
241         if (excess) {
242                 vma->vm_flags |= VM_ACCOUNT;
243                 if (split)
244                         vma->vm_next->vm_flags |= VM_ACCOUNT;
245         }
246
247         if (vm_flags & VM_LOCKED) {
248                 mm->locked_vm += new_len >> PAGE_SHIFT;
249                 if (new_len > old_len)
250                         make_pages_present(new_addr + old_len,
251                                            new_addr + new_len);
252         }
253
254         return new_addr;
255 }
256
257 /*
258  * Expand (or shrink) an existing mapping, potentially moving it at the
259  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
260  *
261  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
262  * This option implies MREMAP_MAYMOVE.
263  */
264 unsigned long do_mremap(unsigned long addr,
265         unsigned long old_len, unsigned long new_len,
266         unsigned long flags, unsigned long new_addr)
267 {
268         struct vm_area_struct *vma;
269         unsigned long ret = -EINVAL;
270         unsigned long charged = 0;
271
272         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
273                 goto out;
274
275         if (addr & ~PAGE_MASK)
276                 goto out;
277
278         old_len = PAGE_ALIGN(old_len);
279         new_len = PAGE_ALIGN(new_len);
280
281         /*
282          * We allow a zero old-len as a special case
283          * for DOS-emu "duplicate shm area" thing. But
284          * a zero new-len is nonsensical.
285          */
286         if (!new_len)
287                 goto out;
288
289         /* new_addr is only valid if MREMAP_FIXED is specified */
290         if (flags & MREMAP_FIXED) {
291                 if (new_addr & ~PAGE_MASK)
292                         goto out;
293                 if (!(flags & MREMAP_MAYMOVE))
294                         goto out;
295
296                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
297                         goto out;
298
299                 /* Check if the location we're moving into overlaps the
300                  * old location at all, and fail if it does.
301                  */
302                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
303                         goto out;
304
305                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
306                         goto out;
307
308                 ret = do_munmap(current->mm, new_addr, new_len);
309                 if (ret)
310                         goto out;
311         }
312
313         /*
314          * Always allow a shrinking remap: that just unmaps
315          * the unnecessary pages..
316          * do_munmap does all the needed commit accounting
317          */
318         if (old_len >= new_len) {
319                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
320                 if (ret && old_len != new_len)
321                         goto out;
322                 ret = addr;
323                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
324                         goto out;
325                 old_len = new_len;
326         }
327
328         /*
329          * Ok, we need to grow..  or relocate.
330          */
331         ret = -EFAULT;
332         vma = find_vma(current->mm, addr);
333         if (!vma || vma->vm_start > addr)
334                 goto out;
335         if (is_vm_hugetlb_page(vma)) {
336                 ret = -EINVAL;
337                 goto out;
338         }
339         /* We can't remap across vm area boundaries */
340         if (old_len > vma->vm_end - addr)
341                 goto out;
342         if (vma->vm_flags & VM_DONTEXPAND) {
343                 if (new_len > old_len)
344                         goto out;
345         }
346         if (vma->vm_flags & VM_LOCKED) {
347                 unsigned long locked, lock_limit;
348                 locked = current->mm->locked_vm << PAGE_SHIFT;
349                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
350                 locked += new_len - old_len;
351                 ret = -EAGAIN;
352                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
353                         goto out;
354         }
355         if (!may_expand_vm(current->mm, (new_len - old_len) >> PAGE_SHIFT)) {
356                 ret = -ENOMEM;
357                 goto out;
358         }
359
360         if (vma->vm_flags & VM_ACCOUNT) {
361                 charged = (new_len - old_len) >> PAGE_SHIFT;
362                 if (security_vm_enough_memory(charged))
363                         goto out_nc;
364         }
365
366         /* old_len exactly to the end of the area..
367          * And we're not relocating the area.
368          */
369         if (old_len == vma->vm_end - addr &&
370             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
371             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
372                 unsigned long max_addr = TASK_SIZE;
373                 if (vma->vm_next)
374                         max_addr = vma->vm_next->vm_start;
375                 /* can we just expand the current mapping? */
376                 if (max_addr - addr >= new_len) {
377                         int pages = (new_len - old_len) >> PAGE_SHIFT;
378
379                         vma_adjust(vma, vma->vm_start,
380                                 addr + new_len, vma->vm_pgoff, NULL);
381
382                         current->mm->total_vm += pages;
383                         __vm_stat_account(vma->vm_mm, vma->vm_flags,
384                                                         vma->vm_file, pages);
385                         if (vma->vm_flags & VM_LOCKED) {
386                                 current->mm->locked_vm += pages;
387                                 make_pages_present(addr + old_len,
388                                                    addr + new_len);
389                         }
390                         ret = addr;
391                         goto out;
392                 }
393         }
394
395         /*
396          * We weren't able to just expand or shrink the area,
397          * we need to create a new one and move it..
398          */
399         ret = -ENOMEM;
400         if (flags & MREMAP_MAYMOVE) {
401                 if (!(flags & MREMAP_FIXED)) {
402                         unsigned long map_flags = 0;
403                         if (vma->vm_flags & VM_MAYSHARE)
404                                 map_flags |= MAP_SHARED;
405
406                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
407                                                 vma->vm_pgoff, map_flags);
408                         ret = new_addr;
409                         if (new_addr & ~PAGE_MASK)
410                                 goto out;
411                 }
412                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
413         }
414 out:
415         if (ret & ~PAGE_MASK)
416                 vm_unacct_memory(charged);
417 out_nc:
418         return ret;
419 }
420
421 asmlinkage unsigned long sys_mremap(unsigned long addr,
422         unsigned long old_len, unsigned long new_len,
423         unsigned long flags, unsigned long new_addr)
424 {
425         unsigned long ret;
426
427         down_write(&current->mm->mmap_sem);
428         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
429         up_write(&current->mm->mmap_sem);
430         return ret;
431 }