[PATCH] mm: update_hiwaters just in time
[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         /*
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         unsigned long hiwater_vm;
171         int split = 0;
172
173         /*
174          * We'd prefer to avoid failure later on in do_munmap:
175          * which may split one vma into three before unmapping.
176          */
177         if (mm->map_count >= sysctl_max_map_count - 3)
178                 return -ENOMEM;
179
180         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
181         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
182         if (!new_vma)
183                 return -ENOMEM;
184
185         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
186         if (moved_len < old_len) {
187                 /*
188                  * On error, move entries back from new area to old,
189                  * which will succeed since page tables still there,
190                  * and then proceed to unmap new area instead of old.
191                  */
192                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
193                 vma = new_vma;
194                 old_len = new_len;
195                 old_addr = new_addr;
196                 new_addr = -ENOMEM;
197         }
198
199         /* Conceal VM_ACCOUNT so old reservation is not undone */
200         if (vm_flags & VM_ACCOUNT) {
201                 vma->vm_flags &= ~VM_ACCOUNT;
202                 excess = vma->vm_end - vma->vm_start - old_len;
203                 if (old_addr > vma->vm_start &&
204                     old_addr + old_len < vma->vm_end)
205                         split = 1;
206         }
207
208         /*
209          * If we failed to move page tables we still do total_vm increment
210          * since do_munmap() will decrement it by old_len == new_len.
211          *
212          * Since total_vm is about to be raised artificially high for a
213          * moment, we need to restore high watermark afterwards: if stats
214          * are taken meanwhile, total_vm and hiwater_vm appear too high.
215          * If this were a serious issue, we'd add a flag to do_munmap().
216          */
217         hiwater_vm = mm->hiwater_vm;
218         mm->total_vm += new_len >> PAGE_SHIFT;
219         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
220
221         if (do_munmap(mm, old_addr, old_len) < 0) {
222                 /* OOM: unable to split vma, just get accounts right */
223                 vm_unacct_memory(excess >> PAGE_SHIFT);
224                 excess = 0;
225         }
226         mm->hiwater_vm = hiwater_vm;
227
228         /* Restore VM_ACCOUNT if one or two pieces of vma left */
229         if (excess) {
230                 vma->vm_flags |= VM_ACCOUNT;
231                 if (split)
232                         vma->vm_next->vm_flags |= VM_ACCOUNT;
233         }
234
235         if (vm_flags & VM_LOCKED) {
236                 mm->locked_vm += new_len >> PAGE_SHIFT;
237                 if (new_len > old_len)
238                         make_pages_present(new_addr + old_len,
239                                            new_addr + new_len);
240         }
241
242         return new_addr;
243 }
244
245 /*
246  * Expand (or shrink) an existing mapping, potentially moving it at the
247  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
248  *
249  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
250  * This option implies MREMAP_MAYMOVE.
251  */
252 unsigned long do_mremap(unsigned long addr,
253         unsigned long old_len, unsigned long new_len,
254         unsigned long flags, unsigned long new_addr)
255 {
256         struct mm_struct *mm = current->mm;
257         struct vm_area_struct *vma;
258         unsigned long ret = -EINVAL;
259         unsigned long charged = 0;
260
261         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
262                 goto out;
263
264         if (addr & ~PAGE_MASK)
265                 goto out;
266
267         old_len = PAGE_ALIGN(old_len);
268         new_len = PAGE_ALIGN(new_len);
269
270         /*
271          * We allow a zero old-len as a special case
272          * for DOS-emu "duplicate shm area" thing. But
273          * a zero new-len is nonsensical.
274          */
275         if (!new_len)
276                 goto out;
277
278         /* new_addr is only valid if MREMAP_FIXED is specified */
279         if (flags & MREMAP_FIXED) {
280                 if (new_addr & ~PAGE_MASK)
281                         goto out;
282                 if (!(flags & MREMAP_MAYMOVE))
283                         goto out;
284
285                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
286                         goto out;
287
288                 /* Check if the location we're moving into overlaps the
289                  * old location at all, and fail if it does.
290                  */
291                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
292                         goto out;
293
294                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
295                         goto out;
296
297                 ret = do_munmap(mm, new_addr, new_len);
298                 if (ret)
299                         goto out;
300         }
301
302         /*
303          * Always allow a shrinking remap: that just unmaps
304          * the unnecessary pages..
305          * do_munmap does all the needed commit accounting
306          */
307         if (old_len >= new_len) {
308                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
309                 if (ret && old_len != new_len)
310                         goto out;
311                 ret = addr;
312                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
313                         goto out;
314                 old_len = new_len;
315         }
316
317         /*
318          * Ok, we need to grow..  or relocate.
319          */
320         ret = -EFAULT;
321         vma = find_vma(mm, addr);
322         if (!vma || vma->vm_start > addr)
323                 goto out;
324         if (is_vm_hugetlb_page(vma)) {
325                 ret = -EINVAL;
326                 goto out;
327         }
328         /* We can't remap across vm area boundaries */
329         if (old_len > vma->vm_end - addr)
330                 goto out;
331         if (vma->vm_flags & VM_DONTEXPAND) {
332                 if (new_len > old_len)
333                         goto out;
334         }
335         if (vma->vm_flags & VM_LOCKED) {
336                 unsigned long locked, lock_limit;
337                 locked = mm->locked_vm << PAGE_SHIFT;
338                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
339                 locked += new_len - old_len;
340                 ret = -EAGAIN;
341                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
342                         goto out;
343         }
344         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
345                 ret = -ENOMEM;
346                 goto out;
347         }
348
349         if (vma->vm_flags & VM_ACCOUNT) {
350                 charged = (new_len - old_len) >> PAGE_SHIFT;
351                 if (security_vm_enough_memory(charged))
352                         goto out_nc;
353         }
354
355         /* old_len exactly to the end of the area..
356          * And we're not relocating the area.
357          */
358         if (old_len == vma->vm_end - addr &&
359             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
360             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
361                 unsigned long max_addr = TASK_SIZE;
362                 if (vma->vm_next)
363                         max_addr = vma->vm_next->vm_start;
364                 /* can we just expand the current mapping? */
365                 if (max_addr - addr >= new_len) {
366                         int pages = (new_len - old_len) >> PAGE_SHIFT;
367
368                         vma_adjust(vma, vma->vm_start,
369                                 addr + new_len, vma->vm_pgoff, NULL);
370
371                         mm->total_vm += pages;
372                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
373                         if (vma->vm_flags & VM_LOCKED) {
374                                 mm->locked_vm += pages;
375                                 make_pages_present(addr + old_len,
376                                                    addr + new_len);
377                         }
378                         ret = addr;
379                         goto out;
380                 }
381         }
382
383         /*
384          * We weren't able to just expand or shrink the area,
385          * we need to create a new one and move it..
386          */
387         ret = -ENOMEM;
388         if (flags & MREMAP_MAYMOVE) {
389                 if (!(flags & MREMAP_FIXED)) {
390                         unsigned long map_flags = 0;
391                         if (vma->vm_flags & VM_MAYSHARE)
392                                 map_flags |= MAP_SHARED;
393
394                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
395                                                 vma->vm_pgoff, map_flags);
396                         ret = new_addr;
397                         if (new_addr & ~PAGE_MASK)
398                                 goto out;
399                 }
400                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
401         }
402 out:
403         if (ret & ~PAGE_MASK)
404                 vm_unacct_memory(charged);
405 out_nc:
406         return ret;
407 }
408
409 asmlinkage unsigned long sys_mremap(unsigned long addr,
410         unsigned long old_len, unsigned long new_len,
411         unsigned long flags, unsigned long new_addr)
412 {
413         unsigned long ret;
414
415         down_write(&current->mm->mmap_sem);
416         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
417         up_write(&current->mm->mmap_sem);
418         return ret;
419 }