[PATCH] mm: move_page_tables by extents
[linux-2.6] / 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         /*
32          * We don't need page_table_lock: we have mmap_sem exclusively.
33          */
34         pgd = pgd_offset(mm, addr);
35         if (pgd_none_or_clear_bad(pgd))
36                 return NULL;
37
38         pud = pud_offset(pgd, addr);
39         if (pud_none_or_clear_bad(pud))
40                 return NULL;
41
42         pmd = pmd_offset(pud, addr);
43         if (pmd_none_or_clear_bad(pmd))
44                 return NULL;
45
46         return pmd;
47 }
48
49 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
50 {
51         pgd_t *pgd;
52         pud_t *pud;
53         pmd_t *pmd = NULL;
54         pte_t *pte;
55
56         /*
57          * We do need page_table_lock: because allocators expect that.
58          */
59         spin_lock(&mm->page_table_lock);
60         pgd = pgd_offset(mm, addr);
61         pud = pud_alloc(mm, pgd, addr);
62         if (!pud)
63                 goto out;
64
65         pmd = pmd_alloc(mm, pud, addr);
66         if (!pmd)
67                 goto out;
68
69         pte = pte_alloc_map(mm, pmd, addr);
70         if (!pte) {
71                 pmd = NULL;
72                 goto out;
73         }
74         pte_unmap(pte);
75 out:
76         spin_unlock(&mm->page_table_lock);
77         return pmd;
78 }
79
80 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
81                 unsigned long old_addr, unsigned long old_end,
82                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
83                 unsigned long new_addr)
84 {
85         struct address_space *mapping = NULL;
86         struct mm_struct *mm = vma->vm_mm;
87         pte_t *old_pte, *new_pte, pte;
88
89         if (vma->vm_file) {
90                 /*
91                  * Subtle point from Rajesh Venkatasubramanian: before
92                  * moving file-based ptes, we must lock vmtruncate out,
93                  * since it might clean the dst vma before the src vma,
94                  * and we propagate stale pages into the dst afterward.
95                  */
96                 mapping = vma->vm_file->f_mapping;
97                 spin_lock(&mapping->i_mmap_lock);
98                 if (new_vma->vm_truncate_count &&
99                     new_vma->vm_truncate_count != vma->vm_truncate_count)
100                         new_vma->vm_truncate_count = 0;
101         }
102
103         spin_lock(&mm->page_table_lock);
104         old_pte = pte_offset_map(old_pmd, old_addr);
105         new_pte = pte_offset_map_nested(new_pmd, new_addr);
106
107         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
108                                    new_pte++, new_addr += PAGE_SIZE) {
109                 if (pte_none(*old_pte))
110                         continue;
111                 pte = ptep_clear_flush(vma, old_addr, old_pte);
112                 /* ZERO_PAGE can be dependant on virtual addr */
113                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
114                 set_pte_at(mm, new_addr, new_pte, pte);
115         }
116
117         pte_unmap_nested(new_pte - 1);
118         pte_unmap(old_pte - 1);
119         spin_unlock(&mm->page_table_lock);
120         if (mapping)
121                 spin_unlock(&mapping->i_mmap_lock);
122 }
123
124 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
125
126 static 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
133         old_end = old_addr + len;
134         flush_cache_range(vma, 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                 if (next - 1 > old_end)
140                         next = old_end;
141                 extent = next - old_addr;
142                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
143                 if (!old_pmd)
144                         continue;
145                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
146                 if (!new_pmd)
147                         break;
148                 next = (new_addr + PMD_SIZE) & PMD_MASK;
149                 if (extent > next - new_addr)
150                         extent = next - new_addr;
151                 if (extent > LATENCY_LIMIT)
152                         extent = LATENCY_LIMIT;
153                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
154                                 new_vma, new_pmd, new_addr);
155         }
156
157         return len + old_addr - old_end;        /* how much done */
158 }
159
160 static unsigned long move_vma(struct vm_area_struct *vma,
161                 unsigned long old_addr, unsigned long old_len,
162                 unsigned long new_len, unsigned long new_addr)
163 {
164         struct mm_struct *mm = vma->vm_mm;
165         struct vm_area_struct *new_vma;
166         unsigned long vm_flags = vma->vm_flags;
167         unsigned long new_pgoff;
168         unsigned long moved_len;
169         unsigned long excess = 0;
170         int split = 0;
171
172         /*
173          * We'd prefer to avoid failure later on in do_munmap:
174          * which may split one vma into three before unmapping.
175          */
176         if (mm->map_count >= sysctl_max_map_count - 3)
177                 return -ENOMEM;
178
179         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
180         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
181         if (!new_vma)
182                 return -ENOMEM;
183
184         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
185         if (moved_len < old_len) {
186                 /*
187                  * On error, move entries back from new area to old,
188                  * which will succeed since page tables still there,
189                  * and then proceed to unmap new area instead of old.
190                  */
191                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
192                 vma = new_vma;
193                 old_len = new_len;
194                 old_addr = new_addr;
195                 new_addr = -ENOMEM;
196         }
197
198         /* Conceal VM_ACCOUNT so old reservation is not undone */
199         if (vm_flags & VM_ACCOUNT) {
200                 vma->vm_flags &= ~VM_ACCOUNT;
201                 excess = vma->vm_end - vma->vm_start - old_len;
202                 if (old_addr > vma->vm_start &&
203                     old_addr + old_len < vma->vm_end)
204                         split = 1;
205         }
206
207         /*
208          * if we failed to move page tables we still do total_vm increment
209          * since do_munmap() will decrement it by old_len == new_len
210          */
211         mm->total_vm += new_len >> PAGE_SHIFT;
212         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
213
214         if (do_munmap(mm, old_addr, old_len) < 0) {
215                 /* OOM: unable to split vma, just get accounts right */
216                 vm_unacct_memory(excess >> PAGE_SHIFT);
217                 excess = 0;
218         }
219
220         /* Restore VM_ACCOUNT if one or two pieces of vma left */
221         if (excess) {
222                 vma->vm_flags |= VM_ACCOUNT;
223                 if (split)
224                         vma->vm_next->vm_flags |= VM_ACCOUNT;
225         }
226
227         if (vm_flags & VM_LOCKED) {
228                 mm->locked_vm += new_len >> PAGE_SHIFT;
229                 if (new_len > old_len)
230                         make_pages_present(new_addr + old_len,
231                                            new_addr + new_len);
232         }
233
234         return new_addr;
235 }
236
237 /*
238  * Expand (or shrink) an existing mapping, potentially moving it at the
239  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
240  *
241  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
242  * This option implies MREMAP_MAYMOVE.
243  */
244 unsigned long do_mremap(unsigned long addr,
245         unsigned long old_len, unsigned long new_len,
246         unsigned long flags, unsigned long new_addr)
247 {
248         struct vm_area_struct *vma;
249         unsigned long ret = -EINVAL;
250         unsigned long charged = 0;
251
252         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
253                 goto out;
254
255         if (addr & ~PAGE_MASK)
256                 goto out;
257
258         old_len = PAGE_ALIGN(old_len);
259         new_len = PAGE_ALIGN(new_len);
260
261         /*
262          * We allow a zero old-len as a special case
263          * for DOS-emu "duplicate shm area" thing. But
264          * a zero new-len is nonsensical.
265          */
266         if (!new_len)
267                 goto out;
268
269         /* new_addr is only valid if MREMAP_FIXED is specified */
270         if (flags & MREMAP_FIXED) {
271                 if (new_addr & ~PAGE_MASK)
272                         goto out;
273                 if (!(flags & MREMAP_MAYMOVE))
274                         goto out;
275
276                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
277                         goto out;
278
279                 /* Check if the location we're moving into overlaps the
280                  * old location at all, and fail if it does.
281                  */
282                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
283                         goto out;
284
285                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
286                         goto out;
287
288                 ret = do_munmap(current->mm, new_addr, new_len);
289                 if (ret)
290                         goto out;
291         }
292
293         /*
294          * Always allow a shrinking remap: that just unmaps
295          * the unnecessary pages..
296          * do_munmap does all the needed commit accounting
297          */
298         if (old_len >= new_len) {
299                 ret = do_munmap(current->mm, addr+new_len, old_len - new_len);
300                 if (ret && old_len != new_len)
301                         goto out;
302                 ret = addr;
303                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
304                         goto out;
305                 old_len = new_len;
306         }
307
308         /*
309          * Ok, we need to grow..  or relocate.
310          */
311         ret = -EFAULT;
312         vma = find_vma(current->mm, addr);
313         if (!vma || vma->vm_start > addr)
314                 goto out;
315         if (is_vm_hugetlb_page(vma)) {
316                 ret = -EINVAL;
317                 goto out;
318         }
319         /* We can't remap across vm area boundaries */
320         if (old_len > vma->vm_end - addr)
321                 goto out;
322         if (vma->vm_flags & VM_DONTEXPAND) {
323                 if (new_len > old_len)
324                         goto out;
325         }
326         if (vma->vm_flags & VM_LOCKED) {
327                 unsigned long locked, lock_limit;
328                 locked = current->mm->locked_vm << PAGE_SHIFT;
329                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
330                 locked += new_len - old_len;
331                 ret = -EAGAIN;
332                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
333                         goto out;
334         }
335         if (!may_expand_vm(current->mm, (new_len - old_len) >> PAGE_SHIFT)) {
336                 ret = -ENOMEM;
337                 goto out;
338         }
339
340         if (vma->vm_flags & VM_ACCOUNT) {
341                 charged = (new_len - old_len) >> PAGE_SHIFT;
342                 if (security_vm_enough_memory(charged))
343                         goto out_nc;
344         }
345
346         /* old_len exactly to the end of the area..
347          * And we're not relocating the area.
348          */
349         if (old_len == vma->vm_end - addr &&
350             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
351             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
352                 unsigned long max_addr = TASK_SIZE;
353                 if (vma->vm_next)
354                         max_addr = vma->vm_next->vm_start;
355                 /* can we just expand the current mapping? */
356                 if (max_addr - addr >= new_len) {
357                         int pages = (new_len - old_len) >> PAGE_SHIFT;
358
359                         vma_adjust(vma, vma->vm_start,
360                                 addr + new_len, vma->vm_pgoff, NULL);
361
362                         current->mm->total_vm += pages;
363                         vm_stat_account(vma->vm_mm, vma->vm_flags,
364                                                         vma->vm_file, pages);
365                         if (vma->vm_flags & VM_LOCKED) {
366                                 current->mm->locked_vm += pages;
367                                 make_pages_present(addr + old_len,
368                                                    addr + new_len);
369                         }
370                         ret = addr;
371                         goto out;
372                 }
373         }
374
375         /*
376          * We weren't able to just expand or shrink the area,
377          * we need to create a new one and move it..
378          */
379         ret = -ENOMEM;
380         if (flags & MREMAP_MAYMOVE) {
381                 if (!(flags & MREMAP_FIXED)) {
382                         unsigned long map_flags = 0;
383                         if (vma->vm_flags & VM_MAYSHARE)
384                                 map_flags |= MAP_SHARED;
385
386                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
387                                                 vma->vm_pgoff, map_flags);
388                         ret = new_addr;
389                         if (new_addr & ~PAGE_MASK)
390                                 goto out;
391                 }
392                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
393         }
394 out:
395         if (ret & ~PAGE_MASK)
396                 vm_unacct_memory(charged);
397 out_nc:
398         return ret;
399 }
400
401 asmlinkage unsigned long sys_mremap(unsigned long addr,
402         unsigned long old_len, unsigned long new_len,
403         unsigned long flags, unsigned long new_addr)
404 {
405         unsigned long ret;
406
407         down_write(&current->mm->mmap_sem);
408         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
409         up_write(&current->mm->mmap_sem);
410         return ret;
411 }