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