[PATCH] paravirt: lazy mmu mode hooks.patch
[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/capability.h>
17 #include <linux/fs.h>
18 #include <linux/highmem.h>
19 #include <linux/security.h>
20 #include <linux/syscalls.h>
21
22 #include <asm/uaccess.h>
23 #include <asm/cacheflush.h>
24 #include <asm/tlbflush.h>
25
26 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
27 {
28         pgd_t *pgd;
29         pud_t *pud;
30         pmd_t *pmd;
31
32         pgd = pgd_offset(mm, addr);
33         if (pgd_none_or_clear_bad(pgd))
34                 return NULL;
35
36         pud = pud_offset(pgd, addr);
37         if (pud_none_or_clear_bad(pud))
38                 return NULL;
39
40         pmd = pmd_offset(pud, addr);
41         if (pmd_none_or_clear_bad(pmd))
42                 return NULL;
43
44         return pmd;
45 }
46
47 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
48 {
49         pgd_t *pgd;
50         pud_t *pud;
51         pmd_t *pmd;
52
53         pgd = pgd_offset(mm, addr);
54         pud = pud_alloc(mm, pgd, addr);
55         if (!pud)
56                 return NULL;
57
58         pmd = pmd_alloc(mm, pud, addr);
59         if (!pmd)
60                 return NULL;
61
62         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
63                 return NULL;
64
65         return pmd;
66 }
67
68 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
69                 unsigned long old_addr, unsigned long old_end,
70                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
71                 unsigned long new_addr)
72 {
73         struct address_space *mapping = NULL;
74         struct mm_struct *mm = vma->vm_mm;
75         pte_t *old_pte, *new_pte, pte;
76         spinlock_t *old_ptl, *new_ptl;
77
78         if (vma->vm_file) {
79                 /*
80                  * Subtle point from Rajesh Venkatasubramanian: before
81                  * moving file-based ptes, we must lock vmtruncate out,
82                  * since it might clean the dst vma before the src vma,
83                  * and we propagate stale pages into the dst afterward.
84                  */
85                 mapping = vma->vm_file->f_mapping;
86                 spin_lock(&mapping->i_mmap_lock);
87                 if (new_vma->vm_truncate_count &&
88                     new_vma->vm_truncate_count != vma->vm_truncate_count)
89                         new_vma->vm_truncate_count = 0;
90         }
91
92         /*
93          * We don't have to worry about the ordering of src and dst
94          * pte locks because exclusive mmap_sem prevents deadlock.
95          */
96         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
97         new_pte = pte_offset_map_nested(new_pmd, new_addr);
98         new_ptl = pte_lockptr(mm, new_pmd);
99         if (new_ptl != old_ptl)
100                 spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
101         arch_enter_lazy_mmu_mode();
102
103         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
104                                    new_pte++, new_addr += PAGE_SIZE) {
105                 if (pte_none(*old_pte))
106                         continue;
107                 pte = ptep_clear_flush(vma, old_addr, old_pte);
108                 /* ZERO_PAGE can be dependant on virtual addr */
109                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
110                 set_pte_at(mm, new_addr, new_pte, pte);
111         }
112
113         arch_leave_lazy_mmu_mode();
114         if (new_ptl != old_ptl)
115                 spin_unlock(new_ptl);
116         pte_unmap_nested(new_pte - 1);
117         pte_unmap_unlock(old_pte - 1, old_ptl);
118         if (mapping)
119                 spin_unlock(&mapping->i_mmap_lock);
120 }
121
122 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
123
124 static unsigned long move_page_tables(struct vm_area_struct *vma,
125                 unsigned long old_addr, struct vm_area_struct *new_vma,
126                 unsigned long new_addr, unsigned long len)
127 {
128         unsigned long extent, next, old_end;
129         pmd_t *old_pmd, *new_pmd;
130
131         old_end = old_addr + len;
132         flush_cache_range(vma, old_addr, old_end);
133
134         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
135                 cond_resched();
136                 next = (old_addr + PMD_SIZE) & PMD_MASK;
137                 if (next - 1 > old_end)
138                         next = old_end;
139                 extent = next - old_addr;
140                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
141                 if (!old_pmd)
142                         continue;
143                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
144                 if (!new_pmd)
145                         break;
146                 next = (new_addr + PMD_SIZE) & PMD_MASK;
147                 if (extent > next - new_addr)
148                         extent = next - new_addr;
149                 if (extent > LATENCY_LIMIT)
150                         extent = LATENCY_LIMIT;
151                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
152                                 new_vma, new_pmd, new_addr);
153         }
154
155         return len + old_addr - old_end;        /* how much done */
156 }
157
158 static unsigned long move_vma(struct vm_area_struct *vma,
159                 unsigned long old_addr, unsigned long old_len,
160                 unsigned long new_len, unsigned long new_addr)
161 {
162         struct mm_struct *mm = vma->vm_mm;
163         struct vm_area_struct *new_vma;
164         unsigned long vm_flags = vma->vm_flags;
165         unsigned long new_pgoff;
166         unsigned long moved_len;
167         unsigned long excess = 0;
168         unsigned long hiwater_vm;
169         int split = 0;
170
171         /*
172          * We'd prefer to avoid failure later on in do_munmap:
173          * which may split one vma into three before unmapping.
174          */
175         if (mm->map_count >= sysctl_max_map_count - 3)
176                 return -ENOMEM;
177
178         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
179         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
180         if (!new_vma)
181                 return -ENOMEM;
182
183         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
184         if (moved_len < old_len) {
185                 /*
186                  * On error, move entries back from new area to old,
187                  * which will succeed since page tables still there,
188                  * and then proceed to unmap new area instead of old.
189                  */
190                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
191                 vma = new_vma;
192                 old_len = new_len;
193                 old_addr = new_addr;
194                 new_addr = -ENOMEM;
195         }
196
197         /* Conceal VM_ACCOUNT so old reservation is not undone */
198         if (vm_flags & VM_ACCOUNT) {
199                 vma->vm_flags &= ~VM_ACCOUNT;
200                 excess = vma->vm_end - vma->vm_start - old_len;
201                 if (old_addr > vma->vm_start &&
202                     old_addr + old_len < vma->vm_end)
203                         split = 1;
204         }
205
206         /*
207          * If we failed to move page tables we still do total_vm increment
208          * since do_munmap() will decrement it by old_len == new_len.
209          *
210          * Since total_vm is about to be raised artificially high for a
211          * moment, we need to restore high watermark afterwards: if stats
212          * are taken meanwhile, total_vm and hiwater_vm appear too high.
213          * If this were a serious issue, we'd add a flag to do_munmap().
214          */
215         hiwater_vm = mm->hiwater_vm;
216         mm->total_vm += new_len >> PAGE_SHIFT;
217         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
218
219         if (do_munmap(mm, old_addr, old_len) < 0) {
220                 /* OOM: unable to split vma, just get accounts right */
221                 vm_unacct_memory(excess >> PAGE_SHIFT);
222                 excess = 0;
223         }
224         mm->hiwater_vm = hiwater_vm;
225
226         /* Restore VM_ACCOUNT if one or two pieces of vma left */
227         if (excess) {
228                 vma->vm_flags |= VM_ACCOUNT;
229                 if (split)
230                         vma->vm_next->vm_flags |= VM_ACCOUNT;
231         }
232
233         if (vm_flags & VM_LOCKED) {
234                 mm->locked_vm += new_len >> PAGE_SHIFT;
235                 if (new_len > old_len)
236                         make_pages_present(new_addr + old_len,
237                                            new_addr + new_len);
238         }
239
240         return new_addr;
241 }
242
243 /*
244  * Expand (or shrink) an existing mapping, potentially moving it at the
245  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
246  *
247  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
248  * This option implies MREMAP_MAYMOVE.
249  */
250 unsigned long do_mremap(unsigned long addr,
251         unsigned long old_len, unsigned long new_len,
252         unsigned long flags, unsigned long new_addr)
253 {
254         struct mm_struct *mm = current->mm;
255         struct vm_area_struct *vma;
256         unsigned long ret = -EINVAL;
257         unsigned long charged = 0;
258
259         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
260                 goto out;
261
262         if (addr & ~PAGE_MASK)
263                 goto out;
264
265         old_len = PAGE_ALIGN(old_len);
266         new_len = PAGE_ALIGN(new_len);
267
268         /*
269          * We allow a zero old-len as a special case
270          * for DOS-emu "duplicate shm area" thing. But
271          * a zero new-len is nonsensical.
272          */
273         if (!new_len)
274                 goto out;
275
276         /* new_addr is only valid if MREMAP_FIXED is specified */
277         if (flags & MREMAP_FIXED) {
278                 if (new_addr & ~PAGE_MASK)
279                         goto out;
280                 if (!(flags & MREMAP_MAYMOVE))
281                         goto out;
282
283                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
284                         goto out;
285
286                 /* Check if the location we're moving into overlaps the
287                  * old location at all, and fail if it does.
288                  */
289                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
290                         goto out;
291
292                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
293                         goto out;
294
295                 ret = do_munmap(mm, new_addr, new_len);
296                 if (ret)
297                         goto out;
298         }
299
300         /*
301          * Always allow a shrinking remap: that just unmaps
302          * the unnecessary pages..
303          * do_munmap does all the needed commit accounting
304          */
305         if (old_len >= new_len) {
306                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
307                 if (ret && old_len != new_len)
308                         goto out;
309                 ret = addr;
310                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
311                         goto out;
312                 old_len = new_len;
313         }
314
315         /*
316          * Ok, we need to grow..  or relocate.
317          */
318         ret = -EFAULT;
319         vma = find_vma(mm, addr);
320         if (!vma || vma->vm_start > addr)
321                 goto out;
322         if (is_vm_hugetlb_page(vma)) {
323                 ret = -EINVAL;
324                 goto out;
325         }
326         /* We can't remap across vm area boundaries */
327         if (old_len > vma->vm_end - addr)
328                 goto out;
329         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
330                 if (new_len > old_len)
331                         goto out;
332         }
333         if (vma->vm_flags & VM_LOCKED) {
334                 unsigned long locked, lock_limit;
335                 locked = mm->locked_vm << PAGE_SHIFT;
336                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
337                 locked += new_len - old_len;
338                 ret = -EAGAIN;
339                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
340                         goto out;
341         }
342         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
343                 ret = -ENOMEM;
344                 goto out;
345         }
346
347         if (vma->vm_flags & VM_ACCOUNT) {
348                 charged = (new_len - old_len) >> PAGE_SHIFT;
349                 if (security_vm_enough_memory(charged))
350                         goto out_nc;
351         }
352
353         /* old_len exactly to the end of the area..
354          * And we're not relocating the area.
355          */
356         if (old_len == vma->vm_end - addr &&
357             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
358             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
359                 unsigned long max_addr = TASK_SIZE;
360                 if (vma->vm_next)
361                         max_addr = vma->vm_next->vm_start;
362                 /* can we just expand the current mapping? */
363                 if (max_addr - addr >= new_len) {
364                         int pages = (new_len - old_len) >> PAGE_SHIFT;
365
366                         vma_adjust(vma, vma->vm_start,
367                                 addr + new_len, vma->vm_pgoff, NULL);
368
369                         mm->total_vm += pages;
370                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
371                         if (vma->vm_flags & VM_LOCKED) {
372                                 mm->locked_vm += pages;
373                                 make_pages_present(addr + old_len,
374                                                    addr + new_len);
375                         }
376                         ret = addr;
377                         goto out;
378                 }
379         }
380
381         /*
382          * We weren't able to just expand or shrink the area,
383          * we need to create a new one and move it..
384          */
385         ret = -ENOMEM;
386         if (flags & MREMAP_MAYMOVE) {
387                 if (!(flags & MREMAP_FIXED)) {
388                         unsigned long map_flags = 0;
389                         if (vma->vm_flags & VM_MAYSHARE)
390                                 map_flags |= MAP_SHARED;
391
392                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
393                                                 vma->vm_pgoff, map_flags);
394                         ret = new_addr;
395                         if (new_addr & ~PAGE_MASK)
396                                 goto out;
397                 }
398                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
399         }
400 out:
401         if (ret & ~PAGE_MASK)
402                 vm_unacct_memory(charged);
403 out_nc:
404         return ret;
405 }
406
407 asmlinkage unsigned long sys_mremap(unsigned long addr,
408         unsigned long old_len, unsigned long new_len,
409         unsigned long flags, unsigned long new_addr)
410 {
411         unsigned long ret;
412
413         down_write(&current->mm->mmap_sem);
414         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
415         up_write(&current->mm->mmap_sem);
416         return ret;
417 }