do_mremap() untangling, part 2
[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/slab.h>
13 #include <linux/shm.h>
14 #include <linux/ksm.h>
15 #include <linux/mman.h>
16 #include <linux/swap.h>
17 #include <linux/capability.h>
18 #include <linux/fs.h>
19 #include <linux/highmem.h>
20 #include <linux/security.h>
21 #include <linux/syscalls.h>
22 #include <linux/mmu_notifier.h>
23
24 #include <asm/uaccess.h>
25 #include <asm/cacheflush.h>
26 #include <asm/tlbflush.h>
27
28 #include "internal.h"
29
30 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
31 {
32         pgd_t *pgd;
33         pud_t *pud;
34         pmd_t *pmd;
35
36         pgd = pgd_offset(mm, addr);
37         if (pgd_none_or_clear_bad(pgd))
38                 return NULL;
39
40         pud = pud_offset(pgd, addr);
41         if (pud_none_or_clear_bad(pud))
42                 return NULL;
43
44         pmd = pmd_offset(pud, addr);
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, 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         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
67                 return NULL;
68
69         return pmd;
70 }
71
72 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
73                 unsigned long old_addr, unsigned long old_end,
74                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
75                 unsigned long new_addr)
76 {
77         struct address_space *mapping = NULL;
78         struct mm_struct *mm = vma->vm_mm;
79         pte_t *old_pte, *new_pte, pte;
80         spinlock_t *old_ptl, *new_ptl;
81         unsigned long old_start;
82
83         old_start = old_addr;
84         mmu_notifier_invalidate_range_start(vma->vm_mm,
85                                             old_start, old_end);
86         if (vma->vm_file) {
87                 /*
88                  * Subtle point from Rajesh Venkatasubramanian: before
89                  * moving file-based ptes, we must lock truncate_pagecache
90                  * out, since it might clean the dst vma before the src vma,
91                  * and we propagate stale pages into the dst afterward.
92                  */
93                 mapping = vma->vm_file->f_mapping;
94                 spin_lock(&mapping->i_mmap_lock);
95                 if (new_vma->vm_truncate_count &&
96                     new_vma->vm_truncate_count != vma->vm_truncate_count)
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_nested(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_nested(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, 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         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
281                 if (new_len > old_len)
282                         goto Efault;
283         }
284
285         if (vma->vm_flags & VM_LOCKED) {
286                 unsigned long locked, lock_limit;
287                 locked = mm->locked_vm << PAGE_SHIFT;
288                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
289                 locked += new_len - old_len;
290                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
291                         goto Eagain;
292         }
293
294         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
295                 goto Enomem;
296
297         if (vma->vm_flags & VM_ACCOUNT) {
298                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
299                 if (security_vm_enough_memory(charged))
300                         goto Efault;
301                 *p = charged;
302         }
303
304         return vma;
305
306 Efault: /* very odd choice for most of the cases, but... */
307         return ERR_PTR(-EFAULT);
308 Einval:
309         return ERR_PTR(-EINVAL);
310 Enomem:
311         return ERR_PTR(-ENOMEM);
312 Eagain:
313         return ERR_PTR(-EAGAIN);
314 }
315
316 static unsigned long mremap_to(unsigned long addr,
317         unsigned long old_len, unsigned long new_addr,
318         unsigned long new_len)
319 {
320         struct mm_struct *mm = current->mm;
321         struct vm_area_struct *vma;
322         unsigned long ret = -EINVAL;
323         unsigned long charged = 0;
324
325         if (new_addr & ~PAGE_MASK)
326                 goto out;
327
328         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
329                 goto out;
330
331         /* Check if the location we're moving into overlaps the
332          * old location at all, and fail if it does.
333          */
334         if ((new_addr <= addr) && (new_addr+new_len) > addr)
335                 goto out;
336
337         if ((addr <= new_addr) && (addr+old_len) > new_addr)
338                 goto out;
339
340         ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
341         if (ret)
342                 goto out;
343
344         ret = do_munmap(mm, new_addr, new_len);
345         if (ret)
346                 goto out;
347
348         if (old_len >= new_len) {
349                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
350                 if (ret && old_len != new_len)
351                         goto out;
352                 old_len = new_len;
353         }
354
355         vma = vma_to_resize(addr, old_len, new_len, &charged);
356         if (IS_ERR(vma)) {
357                 ret = PTR_ERR(vma);
358                 goto out;
359         }
360
361         ret = move_vma(vma, addr, old_len, new_len, new_addr);
362         if (ret & ~PAGE_MASK)
363                 vm_unacct_memory(charged);
364
365 out:
366         return ret;
367 }
368
369 /*
370  * Expand (or shrink) an existing mapping, potentially moving it at the
371  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
372  *
373  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
374  * This option implies MREMAP_MAYMOVE.
375  */
376 unsigned long do_mremap(unsigned long addr,
377         unsigned long old_len, unsigned long new_len,
378         unsigned long flags, unsigned long new_addr)
379 {
380         struct mm_struct *mm = current->mm;
381         struct vm_area_struct *vma;
382         unsigned long ret = -EINVAL;
383         unsigned long charged = 0;
384
385         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
386                 goto out;
387
388         if (addr & ~PAGE_MASK)
389                 goto out;
390
391         old_len = PAGE_ALIGN(old_len);
392         new_len = PAGE_ALIGN(new_len);
393
394         /*
395          * We allow a zero old-len as a special case
396          * for DOS-emu "duplicate shm area" thing. But
397          * a zero new-len is nonsensical.
398          */
399         if (!new_len)
400                 goto out;
401
402         if (flags & MREMAP_FIXED) {
403                 if (flags & MREMAP_MAYMOVE)
404                         ret = mremap_to(addr, old_len, new_addr, new_len);
405                 goto out;
406         }
407
408         /*
409          * Always allow a shrinking remap: that just unmaps
410          * the unnecessary pages..
411          * do_munmap does all the needed commit accounting
412          */
413         if (old_len >= new_len) {
414                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
415                 if (ret && old_len != new_len)
416                         goto out;
417                 ret = addr;
418                 goto out;
419         }
420
421         /*
422          * Ok, we need to grow..
423          */
424         vma = vma_to_resize(addr, old_len, new_len, &charged);
425         if (IS_ERR(vma)) {
426                 ret = PTR_ERR(vma);
427                 goto out;
428         }
429
430         /* old_len exactly to the end of the area..
431          */
432         if (old_len == vma->vm_end - addr) {
433                 unsigned long max_addr = TASK_SIZE;
434                 if (vma->vm_next)
435                         max_addr = vma->vm_next->vm_start;
436                 /* can we just expand the current mapping? */
437                 if (max_addr - addr >= new_len) {
438                         int pages = (new_len - old_len) >> PAGE_SHIFT;
439
440                         vma_adjust(vma, vma->vm_start,
441                                 addr + new_len, vma->vm_pgoff, NULL);
442
443                         mm->total_vm += pages;
444                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
445                         if (vma->vm_flags & VM_LOCKED) {
446                                 mm->locked_vm += pages;
447                                 mlock_vma_pages_range(vma, addr + old_len,
448                                                    addr + new_len);
449                         }
450                         ret = addr;
451                         goto out;
452                 }
453         }
454
455         /*
456          * We weren't able to just expand or shrink the area,
457          * we need to create a new one and move it..
458          */
459         ret = -ENOMEM;
460         if (flags & MREMAP_MAYMOVE) {
461                 unsigned long map_flags = 0;
462                 if (vma->vm_flags & VM_MAYSHARE)
463                         map_flags |= MAP_SHARED;
464
465                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
466                                         vma->vm_pgoff, map_flags);
467                 if (new_addr & ~PAGE_MASK) {
468                         ret = new_addr;
469                         goto out;
470                 }
471
472                 ret = security_file_mmap(NULL, 0, 0, 0, new_addr, 1);
473                 if (ret)
474                         goto out;
475                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
476         }
477 out:
478         if (ret & ~PAGE_MASK)
479                 vm_unacct_memory(charged);
480         return ret;
481 }
482
483 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
484                 unsigned long, new_len, unsigned long, flags,
485                 unsigned long, new_addr)
486 {
487         unsigned long ret;
488
489         down_write(&current->mm->mmap_sem);
490         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
491         up_write(&current->mm->mmap_sem);
492         return ret;
493 }