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