mm: make swapin readahead skip over holes
[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         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                  * Before moving the page tables from the new vma to
225                  * the old vma, we need to be sure the old vma is
226                  * queued after new vma in the same_anon_vma list to
227                  * prevent SMP races with rmap_walk (that could lead
228                  * rmap_walk to miss some page table).
229                  */
230                 anon_vma_moveto_tail(vma);
231
232                 /*
233                  * On error, move entries back from new area to old,
234                  * which will succeed since page tables still there,
235                  * and then proceed to unmap new area instead of old.
236                  */
237                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
238                 vma = new_vma;
239                 old_len = new_len;
240                 old_addr = new_addr;
241                 new_addr = -ENOMEM;
242         }
243
244         /* Conceal VM_ACCOUNT so old reservation is not undone */
245         if (vm_flags & VM_ACCOUNT) {
246                 vma->vm_flags &= ~VM_ACCOUNT;
247                 excess = vma->vm_end - vma->vm_start - old_len;
248                 if (old_addr > vma->vm_start &&
249                     old_addr + old_len < vma->vm_end)
250                         split = 1;
251         }
252
253         /*
254          * If we failed to move page tables we still do total_vm increment
255          * since do_munmap() will decrement it by old_len == new_len.
256          *
257          * Since total_vm is about to be raised artificially high for a
258          * moment, we need to restore high watermark afterwards: if stats
259          * are taken meanwhile, total_vm and hiwater_vm appear too high.
260          * If this were a serious issue, we'd add a flag to do_munmap().
261          */
262         hiwater_vm = mm->hiwater_vm;
263         mm->total_vm += new_len >> PAGE_SHIFT;
264         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
265
266         if (do_munmap(mm, old_addr, old_len) < 0) {
267                 /* OOM: unable to split vma, just get accounts right */
268                 vm_unacct_memory(excess >> PAGE_SHIFT);
269                 excess = 0;
270         }
271         mm->hiwater_vm = hiwater_vm;
272
273         /* Restore VM_ACCOUNT if one or two pieces of vma left */
274         if (excess) {
275                 vma->vm_flags |= VM_ACCOUNT;
276                 if (split)
277                         vma->vm_next->vm_flags |= VM_ACCOUNT;
278         }
279
280         if (vm_flags & VM_LOCKED) {
281                 mm->locked_vm += new_len >> PAGE_SHIFT;
282                 if (new_len > old_len)
283                         mlock_vma_pages_range(new_vma, new_addr + old_len,
284                                                        new_addr + new_len);
285         }
286
287         return new_addr;
288 }
289
290 static struct vm_area_struct *vma_to_resize(unsigned long addr,
291         unsigned long old_len, unsigned long new_len, unsigned long *p)
292 {
293         struct mm_struct *mm = current->mm;
294         struct vm_area_struct *vma = find_vma(mm, addr);
295
296         if (!vma || vma->vm_start > addr)
297                 goto Efault;
298
299         if (is_vm_hugetlb_page(vma))
300                 goto Einval;
301
302         /* We can't remap across vm area boundaries */
303         if (old_len > vma->vm_end - addr)
304                 goto Efault;
305
306         /* Need to be careful about a growing mapping */
307         if (new_len > old_len) {
308                 unsigned long pgoff;
309
310                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
311                         goto Efault;
312                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
313                 pgoff += vma->vm_pgoff;
314                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
315                         goto Einval;
316         }
317
318         if (vma->vm_flags & VM_LOCKED) {
319                 unsigned long locked, lock_limit;
320                 locked = mm->locked_vm << PAGE_SHIFT;
321                 lock_limit = rlimit(RLIMIT_MEMLOCK);
322                 locked += new_len - old_len;
323                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
324                         goto Eagain;
325         }
326
327         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
328                 goto Enomem;
329
330         if (vma->vm_flags & VM_ACCOUNT) {
331                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
332                 if (security_vm_enough_memory(charged))
333                         goto Efault;
334                 *p = charged;
335         }
336
337         return vma;
338
339 Efault: /* very odd choice for most of the cases, but... */
340         return ERR_PTR(-EFAULT);
341 Einval:
342         return ERR_PTR(-EINVAL);
343 Enomem:
344         return ERR_PTR(-ENOMEM);
345 Eagain:
346         return ERR_PTR(-EAGAIN);
347 }
348
349 static unsigned long mremap_to(unsigned long addr,
350         unsigned long old_len, unsigned long new_addr,
351         unsigned long new_len)
352 {
353         struct mm_struct *mm = current->mm;
354         struct vm_area_struct *vma;
355         unsigned long ret = -EINVAL;
356         unsigned long charged = 0;
357         unsigned long map_flags;
358
359         if (new_addr & ~PAGE_MASK)
360                 goto out;
361
362         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
363                 goto out;
364
365         /* Check if the location we're moving into overlaps the
366          * old location at all, and fail if it does.
367          */
368         if ((new_addr <= addr) && (new_addr+new_len) > addr)
369                 goto out;
370
371         if ((addr <= new_addr) && (addr+old_len) > new_addr)
372                 goto out;
373
374         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
375         if (ret)
376                 goto out;
377
378         ret = do_munmap(mm, new_addr, new_len);
379         if (ret)
380                 goto out;
381
382         if (old_len >= new_len) {
383                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
384                 if (ret && old_len != new_len)
385                         goto out;
386                 old_len = new_len;
387         }
388
389         vma = vma_to_resize(addr, old_len, new_len, &charged);
390         if (IS_ERR(vma)) {
391                 ret = PTR_ERR(vma);
392                 goto out;
393         }
394
395         map_flags = MAP_FIXED;
396         if (vma->vm_flags & VM_MAYSHARE)
397                 map_flags |= MAP_SHARED;
398
399         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
400                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
401                                 map_flags);
402         if (ret & ~PAGE_MASK)
403                 goto out1;
404
405         ret = move_vma(vma, addr, old_len, new_len, new_addr);
406         if (!(ret & ~PAGE_MASK))
407                 goto out;
408 out1:
409         vm_unacct_memory(charged);
410
411 out:
412         return ret;
413 }
414
415 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
416 {
417         unsigned long end = vma->vm_end + delta;
418         if (end < vma->vm_end) /* overflow */
419                 return 0;
420         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
421                 return 0;
422         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
423                               0, MAP_FIXED) & ~PAGE_MASK)
424                 return 0;
425         return 1;
426 }
427
428 /*
429  * Expand (or shrink) an existing mapping, potentially moving it at the
430  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
431  *
432  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
433  * This option implies MREMAP_MAYMOVE.
434  */
435 unsigned long do_mremap(unsigned long addr,
436         unsigned long old_len, unsigned long new_len,
437         unsigned long flags, unsigned long new_addr)
438 {
439         struct mm_struct *mm = current->mm;
440         struct vm_area_struct *vma;
441         unsigned long ret = -EINVAL;
442         unsigned long charged = 0;
443
444         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
445                 goto out;
446
447         if (addr & ~PAGE_MASK)
448                 goto out;
449
450         old_len = PAGE_ALIGN(old_len);
451         new_len = PAGE_ALIGN(new_len);
452
453         /*
454          * We allow a zero old-len as a special case
455          * for DOS-emu "duplicate shm area" thing. But
456          * a zero new-len is nonsensical.
457          */
458         if (!new_len)
459                 goto out;
460
461         if (flags & MREMAP_FIXED) {
462                 if (flags & MREMAP_MAYMOVE)
463                         ret = mremap_to(addr, old_len, new_addr, new_len);
464                 goto out;
465         }
466
467         /*
468          * Always allow a shrinking remap: that just unmaps
469          * the unnecessary pages..
470          * do_munmap does all the needed commit accounting
471          */
472         if (old_len >= new_len) {
473                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
474                 if (ret && old_len != new_len)
475                         goto out;
476                 ret = addr;
477                 goto out;
478         }
479
480         /*
481          * Ok, we need to grow..
482          */
483         vma = vma_to_resize(addr, old_len, new_len, &charged);
484         if (IS_ERR(vma)) {
485                 ret = PTR_ERR(vma);
486                 goto out;
487         }
488
489         /* old_len exactly to the end of the area..
490          */
491         if (old_len == vma->vm_end - addr) {
492                 /* can we just expand the current mapping? */
493                 if (vma_expandable(vma, new_len - old_len)) {
494                         int pages = (new_len - old_len) >> PAGE_SHIFT;
495
496                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
497                                        vma->vm_pgoff, NULL)) {
498                                 ret = -ENOMEM;
499                                 goto out;
500                         }
501
502                         mm->total_vm += pages;
503                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
504                         if (vma->vm_flags & VM_LOCKED) {
505                                 mm->locked_vm += pages;
506                                 mlock_vma_pages_range(vma, addr + old_len,
507                                                    addr + new_len);
508                         }
509                         ret = addr;
510                         goto out;
511                 }
512         }
513
514         /*
515          * We weren't able to just expand or shrink the area,
516          * we need to create a new one and move it..
517          */
518         ret = -ENOMEM;
519         if (flags & MREMAP_MAYMOVE) {
520                 unsigned long map_flags = 0;
521                 if (vma->vm_flags & VM_MAYSHARE)
522                         map_flags |= MAP_SHARED;
523
524                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
525                                         vma->vm_pgoff +
526                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
527                                         map_flags);
528                 if (new_addr & ~PAGE_MASK) {
529                         ret = new_addr;
530                         goto out;
531                 }
532
533                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
534                 if (ret)
535                         goto out;
536                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
537         }
538 out:
539         if (ret & ~PAGE_MASK)
540                 vm_unacct_memory(charged);
541         return ret;
542 }
543
544 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
545                 unsigned long, new_len, unsigned long, flags,
546                 unsigned long, new_addr)
547 {
548         unsigned long ret;
549
550         down_write(&current->mm->mmap_sem);
551         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
552         up_write(&current->mm->mmap_sem);
553         return ret;
554 }