mm: nommu: find vma using the sorted vma list
[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/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         split_huge_page_pmd(mm, pmd);
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, struct vm_area_struct *vma,
52                             unsigned long addr)
53 {
54         pgd_t *pgd;
55         pud_t *pud;
56         pmd_t *pmd;
57
58         pgd = pgd_offset(mm, addr);
59         pud = pud_alloc(mm, pgd, addr);
60         if (!pud)
61                 return NULL;
62
63         pmd = pmd_alloc(mm, pud, addr);
64         if (!pmd)
65                 return NULL;
66
67         VM_BUG_ON(pmd_trans_huge(*pmd));
68         if (pmd_none(*pmd) && __pte_alloc(mm, vma, pmd, addr))
69                 return NULL;
70
71         return pmd;
72 }
73
74 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
75                 unsigned long old_addr, unsigned long old_end,
76                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
77                 unsigned long new_addr)
78 {
79         struct address_space *mapping = NULL;
80         struct mm_struct *mm = vma->vm_mm;
81         pte_t *old_pte, *new_pte, pte;
82         spinlock_t *old_ptl, *new_ptl;
83         unsigned long old_start;
84
85         old_start = old_addr;
86         mmu_notifier_invalidate_range_start(vma->vm_mm,
87                                             old_start, old_end);
88         if (vma->vm_file) {
89                 /*
90                  * Subtle point from Rajesh Venkatasubramanian: before
91                  * moving file-based ptes, we must lock truncate_pagecache
92                  * out, since it might clean the dst vma before the src vma,
93                  * and we propagate stale pages into the dst afterward.
94                  */
95                 mapping = vma->vm_file->f_mapping;
96                 spin_lock(&mapping->i_mmap_lock);
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(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(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, vma, 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         int err;
179
180         /*
181          * We'd prefer to avoid failure later on in do_munmap:
182          * which may split one vma into three before unmapping.
183          */
184         if (mm->map_count >= sysctl_max_map_count - 3)
185                 return -ENOMEM;
186
187         /*
188          * Advise KSM to break any KSM pages in the area to be moved:
189          * it would be confusing if they were to turn up at the new
190          * location, where they happen to coincide with different KSM
191          * pages recently unmapped.  But leave vma->vm_flags as it was,
192          * so KSM can come around to merge on vma and new_vma afterwards.
193          */
194         err = ksm_madvise(vma, old_addr, old_addr + old_len,
195                                                 MADV_UNMERGEABLE, &vm_flags);
196         if (err)
197                 return err;
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          * Since total_vm is about to be raised artificially high for a
232          * moment, we need to restore high watermark afterwards: if stats
233          * are taken meanwhile, total_vm and hiwater_vm appear too high.
234          * If this were a serious issue, we'd add a flag to do_munmap().
235          */
236         hiwater_vm = mm->hiwater_vm;
237         mm->total_vm += new_len >> PAGE_SHIFT;
238         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
239
240         if (do_munmap(mm, old_addr, old_len) < 0) {
241                 /* OOM: unable to split vma, just get accounts right */
242                 vm_unacct_memory(excess >> PAGE_SHIFT);
243                 excess = 0;
244         }
245         mm->hiwater_vm = hiwater_vm;
246
247         /* Restore VM_ACCOUNT if one or two pieces of vma left */
248         if (excess) {
249                 vma->vm_flags |= VM_ACCOUNT;
250                 if (split)
251                         vma->vm_next->vm_flags |= VM_ACCOUNT;
252         }
253
254         if (vm_flags & VM_LOCKED) {
255                 mm->locked_vm += new_len >> PAGE_SHIFT;
256                 if (new_len > old_len)
257                         mlock_vma_pages_range(new_vma, new_addr + old_len,
258                                                        new_addr + new_len);
259         }
260
261         return new_addr;
262 }
263
264 static struct vm_area_struct *vma_to_resize(unsigned long addr,
265         unsigned long old_len, unsigned long new_len, unsigned long *p)
266 {
267         struct mm_struct *mm = current->mm;
268         struct vm_area_struct *vma = find_vma(mm, addr);
269
270         if (!vma || vma->vm_start > addr)
271                 goto Efault;
272
273         if (is_vm_hugetlb_page(vma))
274                 goto Einval;
275
276         /* We can't remap across vm area boundaries */
277         if (old_len > vma->vm_end - addr)
278                 goto Efault;
279
280         /* Need to be careful about a growing mapping */
281         if (new_len > old_len) {
282                 unsigned long pgoff;
283
284                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
285                         goto Efault;
286                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
287                 pgoff += vma->vm_pgoff;
288                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
289                         goto Einval;
290         }
291
292         if (vma->vm_flags & VM_LOCKED) {
293                 unsigned long locked, lock_limit;
294                 locked = mm->locked_vm << PAGE_SHIFT;
295                 lock_limit = rlimit(RLIMIT_MEMLOCK);
296                 locked += new_len - old_len;
297                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
298                         goto Eagain;
299         }
300
301         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
302                 goto Enomem;
303
304         if (vma->vm_flags & VM_ACCOUNT) {
305                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
306                 if (security_vm_enough_memory(charged))
307                         goto Efault;
308                 *p = charged;
309         }
310
311         return vma;
312
313 Efault: /* very odd choice for most of the cases, but... */
314         return ERR_PTR(-EFAULT);
315 Einval:
316         return ERR_PTR(-EINVAL);
317 Enomem:
318         return ERR_PTR(-ENOMEM);
319 Eagain:
320         return ERR_PTR(-EAGAIN);
321 }
322
323 static unsigned long mremap_to(unsigned long addr,
324         unsigned long old_len, unsigned long new_addr,
325         unsigned long new_len)
326 {
327         struct mm_struct *mm = current->mm;
328         struct vm_area_struct *vma;
329         unsigned long ret = -EINVAL;
330         unsigned long charged = 0;
331         unsigned long map_flags;
332
333         if (new_addr & ~PAGE_MASK)
334                 goto out;
335
336         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
337                 goto out;
338
339         /* Check if the location we're moving into overlaps the
340          * old location at all, and fail if it does.
341          */
342         if ((new_addr <= addr) && (new_addr+new_len) > addr)
343                 goto out;
344
345         if ((addr <= new_addr) && (addr+old_len) > new_addr)
346                 goto out;
347
348         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
349         if (ret)
350                 goto out;
351
352         ret = do_munmap(mm, new_addr, new_len);
353         if (ret)
354                 goto out;
355
356         if (old_len >= new_len) {
357                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
358                 if (ret && old_len != new_len)
359                         goto out;
360                 old_len = new_len;
361         }
362
363         vma = vma_to_resize(addr, old_len, new_len, &charged);
364         if (IS_ERR(vma)) {
365                 ret = PTR_ERR(vma);
366                 goto out;
367         }
368
369         map_flags = MAP_FIXED;
370         if (vma->vm_flags & VM_MAYSHARE)
371                 map_flags |= MAP_SHARED;
372
373         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
374                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
375                                 map_flags);
376         if (ret & ~PAGE_MASK)
377                 goto out1;
378
379         ret = move_vma(vma, addr, old_len, new_len, new_addr);
380         if (!(ret & ~PAGE_MASK))
381                 goto out;
382 out1:
383         vm_unacct_memory(charged);
384
385 out:
386         return ret;
387 }
388
389 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
390 {
391         unsigned long end = vma->vm_end + delta;
392         if (end < vma->vm_end) /* overflow */
393                 return 0;
394         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
395                 return 0;
396         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
397                               0, MAP_FIXED) & ~PAGE_MASK)
398                 return 0;
399         return 1;
400 }
401
402 /*
403  * Expand (or shrink) an existing mapping, potentially moving it at the
404  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
405  *
406  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
407  * This option implies MREMAP_MAYMOVE.
408  */
409 unsigned long do_mremap(unsigned long addr,
410         unsigned long old_len, unsigned long new_len,
411         unsigned long flags, unsigned long new_addr)
412 {
413         struct mm_struct *mm = current->mm;
414         struct vm_area_struct *vma;
415         unsigned long ret = -EINVAL;
416         unsigned long charged = 0;
417
418         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
419                 goto out;
420
421         if (addr & ~PAGE_MASK)
422                 goto out;
423
424         old_len = PAGE_ALIGN(old_len);
425         new_len = PAGE_ALIGN(new_len);
426
427         /*
428          * We allow a zero old-len as a special case
429          * for DOS-emu "duplicate shm area" thing. But
430          * a zero new-len is nonsensical.
431          */
432         if (!new_len)
433                 goto out;
434
435         if (flags & MREMAP_FIXED) {
436                 if (flags & MREMAP_MAYMOVE)
437                         ret = mremap_to(addr, old_len, new_addr, new_len);
438                 goto out;
439         }
440
441         /*
442          * Always allow a shrinking remap: that just unmaps
443          * the unnecessary pages..
444          * do_munmap does all the needed commit accounting
445          */
446         if (old_len >= new_len) {
447                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
448                 if (ret && old_len != new_len)
449                         goto out;
450                 ret = addr;
451                 goto out;
452         }
453
454         /*
455          * Ok, we need to grow..
456          */
457         vma = vma_to_resize(addr, old_len, new_len, &charged);
458         if (IS_ERR(vma)) {
459                 ret = PTR_ERR(vma);
460                 goto out;
461         }
462
463         /* old_len exactly to the end of the area..
464          */
465         if (old_len == vma->vm_end - addr) {
466                 /* can we just expand the current mapping? */
467                 if (vma_expandable(vma, new_len - old_len)) {
468                         int pages = (new_len - old_len) >> PAGE_SHIFT;
469
470                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
471                                        vma->vm_pgoff, NULL)) {
472                                 ret = -ENOMEM;
473                                 goto out;
474                         }
475
476                         mm->total_vm += pages;
477                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
478                         if (vma->vm_flags & VM_LOCKED) {
479                                 mm->locked_vm += pages;
480                                 mlock_vma_pages_range(vma, addr + old_len,
481                                                    addr + new_len);
482                         }
483                         ret = addr;
484                         goto out;
485                 }
486         }
487
488         /*
489          * We weren't able to just expand or shrink the area,
490          * we need to create a new one and move it..
491          */
492         ret = -ENOMEM;
493         if (flags & MREMAP_MAYMOVE) {
494                 unsigned long map_flags = 0;
495                 if (vma->vm_flags & VM_MAYSHARE)
496                         map_flags |= MAP_SHARED;
497
498                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
499                                         vma->vm_pgoff +
500                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
501                                         map_flags);
502                 if (new_addr & ~PAGE_MASK) {
503                         ret = new_addr;
504                         goto out;
505                 }
506
507                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
508                 if (ret)
509                         goto out;
510                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
511         }
512 out:
513         if (ret & ~PAGE_MASK)
514                 vm_unacct_memory(charged);
515         return ret;
516 }
517
518 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
519                 unsigned long, new_len, unsigned long, flags,
520                 unsigned long, new_addr)
521 {
522         unsigned long ret;
523
524         down_write(&current->mm->mmap_sem);
525         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
526         up_write(&current->mm->mmap_sem);
527         return ret;
528 }