vfs: new helper - vfs_ustat()
[linux-3.10.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(*pmd))
45                 return NULL;
46
47         return pmd;
48 }
49
50 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
51                             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         VM_BUG_ON(pmd_trans_huge(*pmd));
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
81         if (vma->vm_file) {
82                 /*
83                  * Subtle point from Rajesh Venkatasubramanian: before
84                  * moving file-based ptes, we must lock truncate_pagecache
85                  * out, since it might clean the dst vma before the src vma,
86                  * and we propagate stale pages into the dst afterward.
87                  */
88                 mapping = vma->vm_file->f_mapping;
89                 mutex_lock(&mapping->i_mmap_mutex);
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(new_pmd, new_addr);
98         new_ptl = pte_lockptr(mm, new_pmd);
99         if (new_ptl != old_ptl)
100                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
101         arch_enter_lazy_mmu_mode();
102
103         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
104                                    new_pte++, new_addr += PAGE_SIZE) {
105                 if (pte_none(*old_pte))
106                         continue;
107                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
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         arch_leave_lazy_mmu_mode();
113         if (new_ptl != old_ptl)
114                 spin_unlock(new_ptl);
115         pte_unmap(new_pte - 1);
116         pte_unmap_unlock(old_pte - 1, old_ptl);
117         if (mapping)
118                 mutex_unlock(&mapping->i_mmap_mutex);
119 }
120
121 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
122
123 unsigned long move_page_tables(struct vm_area_struct *vma,
124                 unsigned long old_addr, struct vm_area_struct *new_vma,
125                 unsigned long new_addr, unsigned long len)
126 {
127         unsigned long extent, next, old_end;
128         pmd_t *old_pmd, *new_pmd;
129         bool need_flush = false;
130
131         old_end = old_addr + len;
132         flush_cache_range(vma, old_addr, old_end);
133
134         mmu_notifier_invalidate_range_start(vma->vm_mm, old_addr, old_end);
135
136         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
137                 cond_resched();
138                 next = (old_addr + PMD_SIZE) & PMD_MASK;
139                 /* even if next overflowed, extent below will be ok */
140                 extent = next - old_addr;
141                 if (extent > old_end - old_addr)
142                         extent = old_end - old_addr;
143                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
144                 if (!old_pmd)
145                         continue;
146                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
147                 if (!new_pmd)
148                         break;
149                 if (pmd_trans_huge(*old_pmd)) {
150                         int err = 0;
151                         if (extent == HPAGE_PMD_SIZE)
152                                 err = move_huge_pmd(vma, new_vma, old_addr,
153                                                     new_addr, old_end,
154                                                     old_pmd, new_pmd);
155                         if (err > 0) {
156                                 need_flush = true;
157                                 continue;
158                         } else if (!err) {
159                                 split_huge_page_pmd(vma->vm_mm, old_pmd);
160                         }
161                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
162                 }
163                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
164                                                       new_pmd, new_addr))
165                         break;
166                 next = (new_addr + PMD_SIZE) & PMD_MASK;
167                 if (extent > next - new_addr)
168                         extent = next - new_addr;
169                 if (extent > LATENCY_LIMIT)
170                         extent = LATENCY_LIMIT;
171                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
172                                 new_vma, new_pmd, new_addr);
173                 need_flush = true;
174         }
175         if (likely(need_flush))
176                 flush_tlb_range(vma, old_end-len, old_addr);
177
178         mmu_notifier_invalidate_range_end(vma->vm_mm, old_end-len, old_end);
179
180         return len + old_addr - old_end;        /* how much done */
181 }
182
183 static unsigned long move_vma(struct vm_area_struct *vma,
184                 unsigned long old_addr, unsigned long old_len,
185                 unsigned long new_len, unsigned long new_addr)
186 {
187         struct mm_struct *mm = vma->vm_mm;
188         struct vm_area_struct *new_vma;
189         unsigned long vm_flags = vma->vm_flags;
190         unsigned long new_pgoff;
191         unsigned long moved_len;
192         unsigned long excess = 0;
193         unsigned long hiwater_vm;
194         int split = 0;
195         int err;
196
197         /*
198          * We'd prefer to avoid failure later on in do_munmap:
199          * which may split one vma into three before unmapping.
200          */
201         if (mm->map_count >= sysctl_max_map_count - 3)
202                 return -ENOMEM;
203
204         /*
205          * Advise KSM to break any KSM pages in the area to be moved:
206          * it would be confusing if they were to turn up at the new
207          * location, where they happen to coincide with different KSM
208          * pages recently unmapped.  But leave vma->vm_flags as it was,
209          * so KSM can come around to merge on vma and new_vma afterwards.
210          */
211         err = ksm_madvise(vma, old_addr, old_addr + old_len,
212                                                 MADV_UNMERGEABLE, &vm_flags);
213         if (err)
214                 return err;
215
216         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
217         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
218         if (!new_vma)
219                 return -ENOMEM;
220
221         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
222         if (moved_len < old_len) {
223                 /*
224                  * On error, move entries back from new area to old,
225                  * which will succeed since page tables still there,
226                  * and then proceed to unmap new area instead of old.
227                  */
228                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
229                 vma = new_vma;
230                 old_len = new_len;
231                 old_addr = new_addr;
232                 new_addr = -ENOMEM;
233         }
234
235         /* Conceal VM_ACCOUNT so old reservation is not undone */
236         if (vm_flags & VM_ACCOUNT) {
237                 vma->vm_flags &= ~VM_ACCOUNT;
238                 excess = vma->vm_end - vma->vm_start - old_len;
239                 if (old_addr > vma->vm_start &&
240                     old_addr + old_len < vma->vm_end)
241                         split = 1;
242         }
243
244         /*
245          * If we failed to move page tables we still do total_vm increment
246          * since do_munmap() will decrement it by old_len == new_len.
247          *
248          * Since total_vm is about to be raised artificially high for a
249          * moment, we need to restore high watermark afterwards: if stats
250          * are taken meanwhile, total_vm and hiwater_vm appear too high.
251          * If this were a serious issue, we'd add a flag to do_munmap().
252          */
253         hiwater_vm = mm->hiwater_vm;
254         mm->total_vm += new_len >> PAGE_SHIFT;
255         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
256
257         if (do_munmap(mm, old_addr, old_len) < 0) {
258                 /* OOM: unable to split vma, just get accounts right */
259                 vm_unacct_memory(excess >> PAGE_SHIFT);
260                 excess = 0;
261         }
262         mm->hiwater_vm = hiwater_vm;
263
264         /* Restore VM_ACCOUNT if one or two pieces of vma left */
265         if (excess) {
266                 vma->vm_flags |= VM_ACCOUNT;
267                 if (split)
268                         vma->vm_next->vm_flags |= VM_ACCOUNT;
269         }
270
271         if (vm_flags & VM_LOCKED) {
272                 mm->locked_vm += new_len >> PAGE_SHIFT;
273                 if (new_len > old_len)
274                         mlock_vma_pages_range(new_vma, new_addr + old_len,
275                                                        new_addr + new_len);
276         }
277
278         return new_addr;
279 }
280
281 static struct vm_area_struct *vma_to_resize(unsigned long addr,
282         unsigned long old_len, unsigned long new_len, unsigned long *p)
283 {
284         struct mm_struct *mm = current->mm;
285         struct vm_area_struct *vma = find_vma(mm, addr);
286
287         if (!vma || vma->vm_start > addr)
288                 goto Efault;
289
290         if (is_vm_hugetlb_page(vma))
291                 goto Einval;
292
293         /* We can't remap across vm area boundaries */
294         if (old_len > vma->vm_end - addr)
295                 goto Efault;
296
297         /* Need to be careful about a growing mapping */
298         if (new_len > old_len) {
299                 unsigned long pgoff;
300
301                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
302                         goto Efault;
303                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
304                 pgoff += vma->vm_pgoff;
305                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
306                         goto Einval;
307         }
308
309         if (vma->vm_flags & VM_LOCKED) {
310                 unsigned long locked, lock_limit;
311                 locked = mm->locked_vm << PAGE_SHIFT;
312                 lock_limit = rlimit(RLIMIT_MEMLOCK);
313                 locked += new_len - old_len;
314                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
315                         goto Eagain;
316         }
317
318         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
319                 goto Enomem;
320
321         if (vma->vm_flags & VM_ACCOUNT) {
322                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
323                 if (security_vm_enough_memory(charged))
324                         goto Efault;
325                 *p = charged;
326         }
327
328         return vma;
329
330 Efault: /* very odd choice for most of the cases, but... */
331         return ERR_PTR(-EFAULT);
332 Einval:
333         return ERR_PTR(-EINVAL);
334 Enomem:
335         return ERR_PTR(-ENOMEM);
336 Eagain:
337         return ERR_PTR(-EAGAIN);
338 }
339
340 static unsigned long mremap_to(unsigned long addr,
341         unsigned long old_len, unsigned long new_addr,
342         unsigned long new_len)
343 {
344         struct mm_struct *mm = current->mm;
345         struct vm_area_struct *vma;
346         unsigned long ret = -EINVAL;
347         unsigned long charged = 0;
348         unsigned long map_flags;
349
350         if (new_addr & ~PAGE_MASK)
351                 goto out;
352
353         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
354                 goto out;
355
356         /* Check if the location we're moving into overlaps the
357          * old location at all, and fail if it does.
358          */
359         if ((new_addr <= addr) && (new_addr+new_len) > addr)
360                 goto out;
361
362         if ((addr <= new_addr) && (addr+old_len) > new_addr)
363                 goto out;
364
365         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
366         if (ret)
367                 goto out;
368
369         ret = do_munmap(mm, new_addr, new_len);
370         if (ret)
371                 goto out;
372
373         if (old_len >= new_len) {
374                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
375                 if (ret && old_len != new_len)
376                         goto out;
377                 old_len = new_len;
378         }
379
380         vma = vma_to_resize(addr, old_len, new_len, &charged);
381         if (IS_ERR(vma)) {
382                 ret = PTR_ERR(vma);
383                 goto out;
384         }
385
386         map_flags = MAP_FIXED;
387         if (vma->vm_flags & VM_MAYSHARE)
388                 map_flags |= MAP_SHARED;
389
390         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
391                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
392                                 map_flags);
393         if (ret & ~PAGE_MASK)
394                 goto out1;
395
396         ret = move_vma(vma, addr, old_len, new_len, new_addr);
397         if (!(ret & ~PAGE_MASK))
398                 goto out;
399 out1:
400         vm_unacct_memory(charged);
401
402 out:
403         return ret;
404 }
405
406 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
407 {
408         unsigned long end = vma->vm_end + delta;
409         if (end < vma->vm_end) /* overflow */
410                 return 0;
411         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
412                 return 0;
413         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
414                               0, MAP_FIXED) & ~PAGE_MASK)
415                 return 0;
416         return 1;
417 }
418
419 /*
420  * Expand (or shrink) an existing mapping, potentially moving it at the
421  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
422  *
423  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
424  * This option implies MREMAP_MAYMOVE.
425  */
426 unsigned long do_mremap(unsigned long addr,
427         unsigned long old_len, unsigned long new_len,
428         unsigned long flags, unsigned long new_addr)
429 {
430         struct mm_struct *mm = current->mm;
431         struct vm_area_struct *vma;
432         unsigned long ret = -EINVAL;
433         unsigned long charged = 0;
434
435         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
436                 goto out;
437
438         if (addr & ~PAGE_MASK)
439                 goto out;
440
441         old_len = PAGE_ALIGN(old_len);
442         new_len = PAGE_ALIGN(new_len);
443
444         /*
445          * We allow a zero old-len as a special case
446          * for DOS-emu "duplicate shm area" thing. But
447          * a zero new-len is nonsensical.
448          */
449         if (!new_len)
450                 goto out;
451
452         if (flags & MREMAP_FIXED) {
453                 if (flags & MREMAP_MAYMOVE)
454                         ret = mremap_to(addr, old_len, new_addr, new_len);
455                 goto out;
456         }
457
458         /*
459          * Always allow a shrinking remap: that just unmaps
460          * the unnecessary pages..
461          * do_munmap does all the needed commit accounting
462          */
463         if (old_len >= new_len) {
464                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
465                 if (ret && old_len != new_len)
466                         goto out;
467                 ret = addr;
468                 goto out;
469         }
470
471         /*
472          * Ok, we need to grow..
473          */
474         vma = vma_to_resize(addr, old_len, new_len, &charged);
475         if (IS_ERR(vma)) {
476                 ret = PTR_ERR(vma);
477                 goto out;
478         }
479
480         /* old_len exactly to the end of the area..
481          */
482         if (old_len == vma->vm_end - addr) {
483                 /* can we just expand the current mapping? */
484                 if (vma_expandable(vma, new_len - old_len)) {
485                         int pages = (new_len - old_len) >> PAGE_SHIFT;
486
487                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
488                                        vma->vm_pgoff, NULL)) {
489                                 ret = -ENOMEM;
490                                 goto out;
491                         }
492
493                         mm->total_vm += pages;
494                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
495                         if (vma->vm_flags & VM_LOCKED) {
496                                 mm->locked_vm += pages;
497                                 mlock_vma_pages_range(vma, addr + old_len,
498                                                    addr + new_len);
499                         }
500                         ret = addr;
501                         goto out;
502                 }
503         }
504
505         /*
506          * We weren't able to just expand or shrink the area,
507          * we need to create a new one and move it..
508          */
509         ret = -ENOMEM;
510         if (flags & MREMAP_MAYMOVE) {
511                 unsigned long map_flags = 0;
512                 if (vma->vm_flags & VM_MAYSHARE)
513                         map_flags |= MAP_SHARED;
514
515                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
516                                         vma->vm_pgoff +
517                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
518                                         map_flags);
519                 if (new_addr & ~PAGE_MASK) {
520                         ret = new_addr;
521                         goto out;
522                 }
523
524                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
525                 if (ret)
526                         goto out;
527                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
528         }
529 out:
530         if (ret & ~PAGE_MASK)
531                 vm_unacct_memory(charged);
532         return ret;
533 }
534
535 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
536                 unsigned long, new_len, unsigned long, flags,
537                 unsigned long, new_addr)
538 {
539         unsigned long ret;
540
541         down_write(&current->mm->mmap_sem);
542         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
543         up_write(&current->mm->mmap_sem);
544         return ret;
545 }