[PATCH] mm: ptd_alloc take ptlock
[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@redhat.com>
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/mman.h>
15 #include <linux/swap.h>
16 #include <linux/fs.h>
17 #include <linux/highmem.h>
18 #include <linux/security.h>
19 #include <linux/syscalls.h>
20
21 #include <asm/uaccess.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
24
25 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
26 {
27         pgd_t *pgd;
28         pud_t *pud;
29         pmd_t *pmd;
30
31         pgd = pgd_offset(mm, addr);
32         if (pgd_none_or_clear_bad(pgd))
33                 return NULL;
34
35         pud = pud_offset(pgd, addr);
36         if (pud_none_or_clear_bad(pud))
37                 return NULL;
38
39         pmd = pmd_offset(pud, addr);
40         if (pmd_none_or_clear_bad(pmd))
41                 return NULL;
42
43         return pmd;
44 }
45
46 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
47 {
48         pgd_t *pgd;
49         pud_t *pud;
50         pmd_t *pmd;
51
52         pgd = pgd_offset(mm, addr);
53         pud = pud_alloc(mm, pgd, addr);
54         if (!pud)
55                 return NULL;
56
57         pmd = pmd_alloc(mm, pud, addr);
58         if (!pmd)
59                 return NULL;
60
61         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
62                 return NULL;
63
64         return pmd;
65 }
66
67 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
68                 unsigned long old_addr, unsigned long old_end,
69                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
70                 unsigned long new_addr)
71 {
72         struct address_space *mapping = NULL;
73         struct mm_struct *mm = vma->vm_mm;
74         pte_t *old_pte, *new_pte, pte;
75         spinlock_t *old_ptl;
76
77         if (vma->vm_file) {
78                 /*
79                  * Subtle point from Rajesh Venkatasubramanian: before
80                  * moving file-based ptes, we must lock vmtruncate out,
81                  * since it might clean the dst vma before the src vma,
82                  * and we propagate stale pages into the dst afterward.
83                  */
84                 mapping = vma->vm_file->f_mapping;
85                 spin_lock(&mapping->i_mmap_lock);
86                 if (new_vma->vm_truncate_count &&
87                     new_vma->vm_truncate_count != vma->vm_truncate_count)
88                         new_vma->vm_truncate_count = 0;
89         }
90
91         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
92         new_pte = pte_offset_map_nested(new_pmd, new_addr);
93
94         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
95                                    new_pte++, new_addr += PAGE_SIZE) {
96                 if (pte_none(*old_pte))
97                         continue;
98                 pte = ptep_clear_flush(vma, old_addr, old_pte);
99                 /* ZERO_PAGE can be dependant on virtual addr */
100                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
101                 set_pte_at(mm, new_addr, new_pte, pte);
102         }
103
104         pte_unmap_nested(new_pte - 1);
105         pte_unmap_unlock(old_pte - 1, old_ptl);
106         if (mapping)
107                 spin_unlock(&mapping->i_mmap_lock);
108 }
109
110 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
111
112 static unsigned long move_page_tables(struct vm_area_struct *vma,
113                 unsigned long old_addr, struct vm_area_struct *new_vma,
114                 unsigned long new_addr, unsigned long len)
115 {
116         unsigned long extent, next, old_end;
117         pmd_t *old_pmd, *new_pmd;
118
119         old_end = old_addr + len;
120         flush_cache_range(vma, old_addr, old_end);
121
122         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
123                 cond_resched();
124                 next = (old_addr + PMD_SIZE) & PMD_MASK;
125                 if (next - 1 > old_end)
126                         next = old_end;
127                 extent = next - old_addr;
128                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
129                 if (!old_pmd)
130                         continue;
131                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
132                 if (!new_pmd)
133                         break;
134                 next = (new_addr + PMD_SIZE) & PMD_MASK;
135                 if (extent > next - new_addr)
136                         extent = next - new_addr;
137                 if (extent > LATENCY_LIMIT)
138                         extent = LATENCY_LIMIT;
139                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
140                                 new_vma, new_pmd, new_addr);
141         }
142
143         return len + old_addr - old_end;        /* how much done */
144 }
145
146 static unsigned long move_vma(struct vm_area_struct *vma,
147                 unsigned long old_addr, unsigned long old_len,
148                 unsigned long new_len, unsigned long new_addr)
149 {
150         struct mm_struct *mm = vma->vm_mm;
151         struct vm_area_struct *new_vma;
152         unsigned long vm_flags = vma->vm_flags;
153         unsigned long new_pgoff;
154         unsigned long moved_len;
155         unsigned long excess = 0;
156         unsigned long hiwater_vm;
157         int split = 0;
158
159         /*
160          * We'd prefer to avoid failure later on in do_munmap:
161          * which may split one vma into three before unmapping.
162          */
163         if (mm->map_count >= sysctl_max_map_count - 3)
164                 return -ENOMEM;
165
166         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
167         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
168         if (!new_vma)
169                 return -ENOMEM;
170
171         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
172         if (moved_len < old_len) {
173                 /*
174                  * On error, move entries back from new area to old,
175                  * which will succeed since page tables still there,
176                  * and then proceed to unmap new area instead of old.
177                  */
178                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
179                 vma = new_vma;
180                 old_len = new_len;
181                 old_addr = new_addr;
182                 new_addr = -ENOMEM;
183         }
184
185         /* Conceal VM_ACCOUNT so old reservation is not undone */
186         if (vm_flags & VM_ACCOUNT) {
187                 vma->vm_flags &= ~VM_ACCOUNT;
188                 excess = vma->vm_end - vma->vm_start - old_len;
189                 if (old_addr > vma->vm_start &&
190                     old_addr + old_len < vma->vm_end)
191                         split = 1;
192         }
193
194         /*
195          * If we failed to move page tables we still do total_vm increment
196          * since do_munmap() will decrement it by old_len == new_len.
197          *
198          * Since total_vm is about to be raised artificially high for a
199          * moment, we need to restore high watermark afterwards: if stats
200          * are taken meanwhile, total_vm and hiwater_vm appear too high.
201          * If this were a serious issue, we'd add a flag to do_munmap().
202          */
203         hiwater_vm = mm->hiwater_vm;
204         mm->total_vm += new_len >> PAGE_SHIFT;
205         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
206
207         if (do_munmap(mm, old_addr, old_len) < 0) {
208                 /* OOM: unable to split vma, just get accounts right */
209                 vm_unacct_memory(excess >> PAGE_SHIFT);
210                 excess = 0;
211         }
212         mm->hiwater_vm = hiwater_vm;
213
214         /* Restore VM_ACCOUNT if one or two pieces of vma left */
215         if (excess) {
216                 vma->vm_flags |= VM_ACCOUNT;
217                 if (split)
218                         vma->vm_next->vm_flags |= VM_ACCOUNT;
219         }
220
221         if (vm_flags & VM_LOCKED) {
222                 mm->locked_vm += new_len >> PAGE_SHIFT;
223                 if (new_len > old_len)
224                         make_pages_present(new_addr + old_len,
225                                            new_addr + new_len);
226         }
227
228         return new_addr;
229 }
230
231 /*
232  * Expand (or shrink) an existing mapping, potentially moving it at the
233  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
234  *
235  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
236  * This option implies MREMAP_MAYMOVE.
237  */
238 unsigned long do_mremap(unsigned long addr,
239         unsigned long old_len, unsigned long new_len,
240         unsigned long flags, unsigned long new_addr)
241 {
242         struct mm_struct *mm = current->mm;
243         struct vm_area_struct *vma;
244         unsigned long ret = -EINVAL;
245         unsigned long charged = 0;
246
247         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
248                 goto out;
249
250         if (addr & ~PAGE_MASK)
251                 goto out;
252
253         old_len = PAGE_ALIGN(old_len);
254         new_len = PAGE_ALIGN(new_len);
255
256         /*
257          * We allow a zero old-len as a special case
258          * for DOS-emu "duplicate shm area" thing. But
259          * a zero new-len is nonsensical.
260          */
261         if (!new_len)
262                 goto out;
263
264         /* new_addr is only valid if MREMAP_FIXED is specified */
265         if (flags & MREMAP_FIXED) {
266                 if (new_addr & ~PAGE_MASK)
267                         goto out;
268                 if (!(flags & MREMAP_MAYMOVE))
269                         goto out;
270
271                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
272                         goto out;
273
274                 /* Check if the location we're moving into overlaps the
275                  * old location at all, and fail if it does.
276                  */
277                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
278                         goto out;
279
280                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
281                         goto out;
282
283                 ret = do_munmap(mm, new_addr, new_len);
284                 if (ret)
285                         goto out;
286         }
287
288         /*
289          * Always allow a shrinking remap: that just unmaps
290          * the unnecessary pages..
291          * do_munmap does all the needed commit accounting
292          */
293         if (old_len >= new_len) {
294                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
295                 if (ret && old_len != new_len)
296                         goto out;
297                 ret = addr;
298                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
299                         goto out;
300                 old_len = new_len;
301         }
302
303         /*
304          * Ok, we need to grow..  or relocate.
305          */
306         ret = -EFAULT;
307         vma = find_vma(mm, addr);
308         if (!vma || vma->vm_start > addr)
309                 goto out;
310         if (is_vm_hugetlb_page(vma)) {
311                 ret = -EINVAL;
312                 goto out;
313         }
314         /* We can't remap across vm area boundaries */
315         if (old_len > vma->vm_end - addr)
316                 goto out;
317         if (vma->vm_flags & VM_DONTEXPAND) {
318                 if (new_len > old_len)
319                         goto out;
320         }
321         if (vma->vm_flags & VM_LOCKED) {
322                 unsigned long locked, lock_limit;
323                 locked = mm->locked_vm << PAGE_SHIFT;
324                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
325                 locked += new_len - old_len;
326                 ret = -EAGAIN;
327                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
328                         goto out;
329         }
330         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
331                 ret = -ENOMEM;
332                 goto out;
333         }
334
335         if (vma->vm_flags & VM_ACCOUNT) {
336                 charged = (new_len - old_len) >> PAGE_SHIFT;
337                 if (security_vm_enough_memory(charged))
338                         goto out_nc;
339         }
340
341         /* old_len exactly to the end of the area..
342          * And we're not relocating the area.
343          */
344         if (old_len == vma->vm_end - addr &&
345             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
346             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
347                 unsigned long max_addr = TASK_SIZE;
348                 if (vma->vm_next)
349                         max_addr = vma->vm_next->vm_start;
350                 /* can we just expand the current mapping? */
351                 if (max_addr - addr >= new_len) {
352                         int pages = (new_len - old_len) >> PAGE_SHIFT;
353
354                         vma_adjust(vma, vma->vm_start,
355                                 addr + new_len, vma->vm_pgoff, NULL);
356
357                         mm->total_vm += pages;
358                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
359                         if (vma->vm_flags & VM_LOCKED) {
360                                 mm->locked_vm += pages;
361                                 make_pages_present(addr + old_len,
362                                                    addr + new_len);
363                         }
364                         ret = addr;
365                         goto out;
366                 }
367         }
368
369         /*
370          * We weren't able to just expand or shrink the area,
371          * we need to create a new one and move it..
372          */
373         ret = -ENOMEM;
374         if (flags & MREMAP_MAYMOVE) {
375                 if (!(flags & MREMAP_FIXED)) {
376                         unsigned long map_flags = 0;
377                         if (vma->vm_flags & VM_MAYSHARE)
378                                 map_flags |= MAP_SHARED;
379
380                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
381                                                 vma->vm_pgoff, map_flags);
382                         ret = new_addr;
383                         if (new_addr & ~PAGE_MASK)
384                                 goto out;
385                 }
386                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
387         }
388 out:
389         if (ret & ~PAGE_MASK)
390                 vm_unacct_memory(charged);
391 out_nc:
392         return ret;
393 }
394
395 asmlinkage unsigned long sys_mremap(unsigned long addr,
396         unsigned long old_len, unsigned long new_len,
397         unsigned long flags, unsigned long new_addr)
398 {
399         unsigned long ret;
400
401         down_write(&current->mm->mmap_sem);
402         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
403         up_write(&current->mm->mmap_sem);
404         return ret;
405 }