mremap: avoid sending one IPI per page
[pandora-kernel.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
84         if (vma->vm_file) {
85                 /*
86                  * Subtle point from Rajesh Venkatasubramanian: before
87                  * moving file-based ptes, we must lock truncate_pagecache
88                  * out, since it might clean the dst vma before the src vma,
89                  * and we propagate stale pages into the dst afterward.
90                  */
91                 mapping = vma->vm_file->f_mapping;
92                 mutex_lock(&mapping->i_mmap_mutex);
93         }
94
95         /*
96          * We don't have to worry about the ordering of src and dst
97          * pte locks because exclusive mmap_sem prevents deadlock.
98          */
99         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
100         new_pte = pte_offset_map(new_pmd, new_addr);
101         new_ptl = pte_lockptr(mm, new_pmd);
102         if (new_ptl != old_ptl)
103                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
104         arch_enter_lazy_mmu_mode();
105
106         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
107                                    new_pte++, new_addr += PAGE_SIZE) {
108                 if (pte_none(*old_pte))
109                         continue;
110                 pte = ptep_get_and_clear(mm, old_addr, old_pte);
111                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
112                 set_pte_at(mm, new_addr, new_pte, pte);
113         }
114
115         arch_leave_lazy_mmu_mode();
116         if (new_ptl != old_ptl)
117                 spin_unlock(new_ptl);
118         pte_unmap(new_pte - 1);
119         pte_unmap_unlock(old_pte - 1, old_ptl);
120         if (mapping)
121                 mutex_unlock(&mapping->i_mmap_mutex);
122 }
123
124 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
125
126 unsigned long move_page_tables(struct vm_area_struct *vma,
127                 unsigned long old_addr, struct vm_area_struct *new_vma,
128                 unsigned long new_addr, unsigned long len)
129 {
130         unsigned long extent, next, old_end;
131         pmd_t *old_pmd, *new_pmd;
132         bool need_flush = false;
133
134         old_end = old_addr + len;
135         flush_cache_range(vma, old_addr, old_end);
136
137         mmu_notifier_invalidate_range_start(vma->vm_mm, old_addr, old_end);
138
139         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
140                 cond_resched();
141                 next = (old_addr + PMD_SIZE) & PMD_MASK;
142                 /* even if next overflowed, extent below will be ok */
143                 extent = next - old_addr;
144                 if (extent > old_end - old_addr)
145                         extent = old_end - old_addr;
146                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
147                 if (!old_pmd)
148                         continue;
149                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
150                 if (!new_pmd)
151                         break;
152                 next = (new_addr + PMD_SIZE) & PMD_MASK;
153                 if (extent > next - new_addr)
154                         extent = next - new_addr;
155                 if (extent > LATENCY_LIMIT)
156                         extent = LATENCY_LIMIT;
157                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
158                                 new_vma, new_pmd, new_addr);
159                 need_flush = true;
160         }
161         if (likely(need_flush))
162                 flush_tlb_range(vma, old_end-len, old_addr);
163
164         mmu_notifier_invalidate_range_end(vma->vm_mm, old_end-len, old_end);
165
166         return len + old_addr - old_end;        /* how much done */
167 }
168
169 static unsigned long move_vma(struct vm_area_struct *vma,
170                 unsigned long old_addr, unsigned long old_len,
171                 unsigned long new_len, unsigned long new_addr)
172 {
173         struct mm_struct *mm = vma->vm_mm;
174         struct vm_area_struct *new_vma;
175         unsigned long vm_flags = vma->vm_flags;
176         unsigned long new_pgoff;
177         unsigned long moved_len;
178         unsigned long excess = 0;
179         unsigned long hiwater_vm;
180         int split = 0;
181         int err;
182
183         /*
184          * We'd prefer to avoid failure later on in do_munmap:
185          * which may split one vma into three before unmapping.
186          */
187         if (mm->map_count >= sysctl_max_map_count - 3)
188                 return -ENOMEM;
189
190         /*
191          * Advise KSM to break any KSM pages in the area to be moved:
192          * it would be confusing if they were to turn up at the new
193          * location, where they happen to coincide with different KSM
194          * pages recently unmapped.  But leave vma->vm_flags as it was,
195          * so KSM can come around to merge on vma and new_vma afterwards.
196          */
197         err = ksm_madvise(vma, old_addr, old_addr + old_len,
198                                                 MADV_UNMERGEABLE, &vm_flags);
199         if (err)
200                 return err;
201
202         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
203         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
204         if (!new_vma)
205                 return -ENOMEM;
206
207         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
208         if (moved_len < old_len) {
209                 /*
210                  * On error, move entries back from new area to old,
211                  * which will succeed since page tables still there,
212                  * and then proceed to unmap new area instead of old.
213                  */
214                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
215                 vma = new_vma;
216                 old_len = new_len;
217                 old_addr = new_addr;
218                 new_addr = -ENOMEM;
219         }
220
221         /* Conceal VM_ACCOUNT so old reservation is not undone */
222         if (vm_flags & VM_ACCOUNT) {
223                 vma->vm_flags &= ~VM_ACCOUNT;
224                 excess = vma->vm_end - vma->vm_start - old_len;
225                 if (old_addr > vma->vm_start &&
226                     old_addr + old_len < vma->vm_end)
227                         split = 1;
228         }
229
230         /*
231          * If we failed to move page tables we still do total_vm increment
232          * since do_munmap() will decrement it by old_len == new_len.
233          *
234          * Since total_vm is about to be raised artificially high for a
235          * moment, we need to restore high watermark afterwards: if stats
236          * are taken meanwhile, total_vm and hiwater_vm appear too high.
237          * If this were a serious issue, we'd add a flag to do_munmap().
238          */
239         hiwater_vm = mm->hiwater_vm;
240         mm->total_vm += new_len >> PAGE_SHIFT;
241         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
242
243         if (do_munmap(mm, old_addr, old_len) < 0) {
244                 /* OOM: unable to split vma, just get accounts right */
245                 vm_unacct_memory(excess >> PAGE_SHIFT);
246                 excess = 0;
247         }
248         mm->hiwater_vm = hiwater_vm;
249
250         /* Restore VM_ACCOUNT if one or two pieces of vma left */
251         if (excess) {
252                 vma->vm_flags |= VM_ACCOUNT;
253                 if (split)
254                         vma->vm_next->vm_flags |= VM_ACCOUNT;
255         }
256
257         if (vm_flags & VM_LOCKED) {
258                 mm->locked_vm += new_len >> PAGE_SHIFT;
259                 if (new_len > old_len)
260                         mlock_vma_pages_range(new_vma, new_addr + old_len,
261                                                        new_addr + new_len);
262         }
263
264         return new_addr;
265 }
266
267 static struct vm_area_struct *vma_to_resize(unsigned long addr,
268         unsigned long old_len, unsigned long new_len, unsigned long *p)
269 {
270         struct mm_struct *mm = current->mm;
271         struct vm_area_struct *vma = find_vma(mm, addr);
272
273         if (!vma || vma->vm_start > addr)
274                 goto Efault;
275
276         if (is_vm_hugetlb_page(vma))
277                 goto Einval;
278
279         /* We can't remap across vm area boundaries */
280         if (old_len > vma->vm_end - addr)
281                 goto Efault;
282
283         /* Need to be careful about a growing mapping */
284         if (new_len > old_len) {
285                 unsigned long pgoff;
286
287                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
288                         goto Efault;
289                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
290                 pgoff += vma->vm_pgoff;
291                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
292                         goto Einval;
293         }
294
295         if (vma->vm_flags & VM_LOCKED) {
296                 unsigned long locked, lock_limit;
297                 locked = mm->locked_vm << PAGE_SHIFT;
298                 lock_limit = rlimit(RLIMIT_MEMLOCK);
299                 locked += new_len - old_len;
300                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
301                         goto Eagain;
302         }
303
304         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
305                 goto Enomem;
306
307         if (vma->vm_flags & VM_ACCOUNT) {
308                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
309                 if (security_vm_enough_memory(charged))
310                         goto Efault;
311                 *p = charged;
312         }
313
314         return vma;
315
316 Efault: /* very odd choice for most of the cases, but... */
317         return ERR_PTR(-EFAULT);
318 Einval:
319         return ERR_PTR(-EINVAL);
320 Enomem:
321         return ERR_PTR(-ENOMEM);
322 Eagain:
323         return ERR_PTR(-EAGAIN);
324 }
325
326 static unsigned long mremap_to(unsigned long addr,
327         unsigned long old_len, unsigned long new_addr,
328         unsigned long new_len)
329 {
330         struct mm_struct *mm = current->mm;
331         struct vm_area_struct *vma;
332         unsigned long ret = -EINVAL;
333         unsigned long charged = 0;
334         unsigned long map_flags;
335
336         if (new_addr & ~PAGE_MASK)
337                 goto out;
338
339         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
340                 goto out;
341
342         /* Check if the location we're moving into overlaps the
343          * old location at all, and fail if it does.
344          */
345         if ((new_addr <= addr) && (new_addr+new_len) > addr)
346                 goto out;
347
348         if ((addr <= new_addr) && (addr+old_len) > new_addr)
349                 goto out;
350
351         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
352         if (ret)
353                 goto out;
354
355         ret = do_munmap(mm, new_addr, new_len);
356         if (ret)
357                 goto out;
358
359         if (old_len >= new_len) {
360                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
361                 if (ret && old_len != new_len)
362                         goto out;
363                 old_len = new_len;
364         }
365
366         vma = vma_to_resize(addr, old_len, new_len, &charged);
367         if (IS_ERR(vma)) {
368                 ret = PTR_ERR(vma);
369                 goto out;
370         }
371
372         map_flags = MAP_FIXED;
373         if (vma->vm_flags & VM_MAYSHARE)
374                 map_flags |= MAP_SHARED;
375
376         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
377                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
378                                 map_flags);
379         if (ret & ~PAGE_MASK)
380                 goto out1;
381
382         ret = move_vma(vma, addr, old_len, new_len, new_addr);
383         if (!(ret & ~PAGE_MASK))
384                 goto out;
385 out1:
386         vm_unacct_memory(charged);
387
388 out:
389         return ret;
390 }
391
392 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
393 {
394         unsigned long end = vma->vm_end + delta;
395         if (end < vma->vm_end) /* overflow */
396                 return 0;
397         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
398                 return 0;
399         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
400                               0, MAP_FIXED) & ~PAGE_MASK)
401                 return 0;
402         return 1;
403 }
404
405 /*
406  * Expand (or shrink) an existing mapping, potentially moving it at the
407  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
408  *
409  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
410  * This option implies MREMAP_MAYMOVE.
411  */
412 unsigned long do_mremap(unsigned long addr,
413         unsigned long old_len, unsigned long new_len,
414         unsigned long flags, unsigned long new_addr)
415 {
416         struct mm_struct *mm = current->mm;
417         struct vm_area_struct *vma;
418         unsigned long ret = -EINVAL;
419         unsigned long charged = 0;
420
421         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
422                 goto out;
423
424         if (addr & ~PAGE_MASK)
425                 goto out;
426
427         old_len = PAGE_ALIGN(old_len);
428         new_len = PAGE_ALIGN(new_len);
429
430         /*
431          * We allow a zero old-len as a special case
432          * for DOS-emu "duplicate shm area" thing. But
433          * a zero new-len is nonsensical.
434          */
435         if (!new_len)
436                 goto out;
437
438         if (flags & MREMAP_FIXED) {
439                 if (flags & MREMAP_MAYMOVE)
440                         ret = mremap_to(addr, old_len, new_addr, new_len);
441                 goto out;
442         }
443
444         /*
445          * Always allow a shrinking remap: that just unmaps
446          * the unnecessary pages..
447          * do_munmap does all the needed commit accounting
448          */
449         if (old_len >= new_len) {
450                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
451                 if (ret && old_len != new_len)
452                         goto out;
453                 ret = addr;
454                 goto out;
455         }
456
457         /*
458          * Ok, we need to grow..
459          */
460         vma = vma_to_resize(addr, old_len, new_len, &charged);
461         if (IS_ERR(vma)) {
462                 ret = PTR_ERR(vma);
463                 goto out;
464         }
465
466         /* old_len exactly to the end of the area..
467          */
468         if (old_len == vma->vm_end - addr) {
469                 /* can we just expand the current mapping? */
470                 if (vma_expandable(vma, new_len - old_len)) {
471                         int pages = (new_len - old_len) >> PAGE_SHIFT;
472
473                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
474                                        vma->vm_pgoff, NULL)) {
475                                 ret = -ENOMEM;
476                                 goto out;
477                         }
478
479                         mm->total_vm += pages;
480                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
481                         if (vma->vm_flags & VM_LOCKED) {
482                                 mm->locked_vm += pages;
483                                 mlock_vma_pages_range(vma, addr + old_len,
484                                                    addr + new_len);
485                         }
486                         ret = addr;
487                         goto out;
488                 }
489         }
490
491         /*
492          * We weren't able to just expand or shrink the area,
493          * we need to create a new one and move it..
494          */
495         ret = -ENOMEM;
496         if (flags & MREMAP_MAYMOVE) {
497                 unsigned long map_flags = 0;
498                 if (vma->vm_flags & VM_MAYSHARE)
499                         map_flags |= MAP_SHARED;
500
501                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
502                                         vma->vm_pgoff +
503                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
504                                         map_flags);
505                 if (new_addr & ~PAGE_MASK) {
506                         ret = new_addr;
507                         goto out;
508                 }
509
510                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
511                 if (ret)
512                         goto out;
513                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
514         }
515 out:
516         if (ret & ~PAGE_MASK)
517                 vm_unacct_memory(charged);
518         return ret;
519 }
520
521 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
522                 unsigned long, new_len, unsigned long, flags,
523                 unsigned long, new_addr)
524 {
525         unsigned long ret;
526
527         down_write(&current->mm->mmap_sem);
528         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
529         up_write(&current->mm->mmap_sem);
530         return ret;
531 }