mm: have order > 0 compaction start near a pageblock with free pages
[linux-3.10.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/shm.h>
13 #include <linux/ksm.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 #include <linux/mmu_notifier.h>
22
23 #include <asm/uaccess.h>
24 #include <asm/cacheflush.h>
25 #include <asm/tlbflush.h>
26
27 #include "internal.h"
28
29 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
30 {
31         pgd_t *pgd;
32         pud_t *pud;
33         pmd_t *pmd;
34
35         pgd = pgd_offset(mm, addr);
36         if (pgd_none_or_clear_bad(pgd))
37                 return NULL;
38
39         pud = pud_offset(pgd, addr);
40         if (pud_none_or_clear_bad(pud))
41                 return NULL;
42
43         pmd = pmd_offset(pud, addr);
44         if (pmd_none(*pmd))
45                 return NULL;
46
47         return pmd;
48 }
49
50 static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
51                             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         VM_BUG_ON(pmd_trans_huge(*pmd));
67
68         return pmd;
69 }
70
71 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
72                 unsigned long old_addr, unsigned long old_end,
73                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
74                 unsigned long new_addr)
75 {
76         struct address_space *mapping = NULL;
77         struct mm_struct *mm = vma->vm_mm;
78         pte_t *old_pte, *new_pte, pte;
79         spinlock_t *old_ptl, *new_ptl;
80
81         if (vma->vm_file) {
82                 /*
83                  * Subtle point from Rajesh Venkatasubramanian: before
84                  * moving file-based ptes, we must lock truncate_pagecache
85                  * out, since it might clean the dst vma before the src vma,
86                  * and we propagate stale pages into the dst afterward.
87                  */
88                 mapping = vma->vm_file->f_mapping;
89                 mutex_lock(&mapping->i_mmap_mutex);
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(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_get_and_clear(mm, old_addr, old_pte);
108                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
109                 set_pte_at(mm, new_addr, new_pte, pte);
110         }
111
112         arch_leave_lazy_mmu_mode();
113         if (new_ptl != old_ptl)
114                 spin_unlock(new_ptl);
115         pte_unmap(new_pte - 1);
116         pte_unmap_unlock(old_pte - 1, old_ptl);
117         if (mapping)
118                 mutex_unlock(&mapping->i_mmap_mutex);
119 }
120
121 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
122
123 unsigned long move_page_tables(struct vm_area_struct *vma,
124                 unsigned long old_addr, struct vm_area_struct *new_vma,
125                 unsigned long new_addr, unsigned long len)
126 {
127         unsigned long extent, next, old_end;
128         pmd_t *old_pmd, *new_pmd;
129         bool need_flush = false;
130
131         old_end = old_addr + len;
132         flush_cache_range(vma, old_addr, old_end);
133
134         mmu_notifier_invalidate_range_start(vma->vm_mm, 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                 /* even if next overflowed, extent below will be ok */
140                 extent = next - old_addr;
141                 if (extent > old_end - old_addr)
142                         extent = old_end - old_addr;
143                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
144                 if (!old_pmd)
145                         continue;
146                 new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
147                 if (!new_pmd)
148                         break;
149                 if (pmd_trans_huge(*old_pmd)) {
150                         int err = 0;
151                         if (extent == HPAGE_PMD_SIZE)
152                                 err = move_huge_pmd(vma, new_vma, old_addr,
153                                                     new_addr, old_end,
154                                                     old_pmd, new_pmd);
155                         if (err > 0) {
156                                 need_flush = true;
157                                 continue;
158                         } else if (!err) {
159                                 split_huge_page_pmd(vma->vm_mm, old_pmd);
160                         }
161                         VM_BUG_ON(pmd_trans_huge(*old_pmd));
162                 }
163                 if (pmd_none(*new_pmd) && __pte_alloc(new_vma->vm_mm, new_vma,
164                                                       new_pmd, new_addr))
165                         break;
166                 next = (new_addr + PMD_SIZE) & PMD_MASK;
167                 if (extent > next - new_addr)
168                         extent = next - new_addr;
169                 if (extent > LATENCY_LIMIT)
170                         extent = LATENCY_LIMIT;
171                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
172                                 new_vma, new_pmd, new_addr);
173                 need_flush = true;
174         }
175         if (likely(need_flush))
176                 flush_tlb_range(vma, old_end-len, old_addr);
177
178         mmu_notifier_invalidate_range_end(vma->vm_mm, old_end-len, old_end);
179
180         return len + old_addr - old_end;        /* how much done */
181 }
182
183 static unsigned long move_vma(struct vm_area_struct *vma,
184                 unsigned long old_addr, unsigned long old_len,
185                 unsigned long new_len, unsigned long new_addr)
186 {
187         struct mm_struct *mm = vma->vm_mm;
188         struct vm_area_struct *new_vma;
189         unsigned long vm_flags = vma->vm_flags;
190         unsigned long new_pgoff;
191         unsigned long moved_len;
192         unsigned long excess = 0;
193         unsigned long hiwater_vm;
194         int split = 0;
195         int err;
196
197         /*
198          * We'd prefer to avoid failure later on in do_munmap:
199          * which may split one vma into three before unmapping.
200          */
201         if (mm->map_count >= sysctl_max_map_count - 3)
202                 return -ENOMEM;
203
204         /*
205          * Advise KSM to break any KSM pages in the area to be moved:
206          * it would be confusing if they were to turn up at the new
207          * location, where they happen to coincide with different KSM
208          * pages recently unmapped.  But leave vma->vm_flags as it was,
209          * so KSM can come around to merge on vma and new_vma afterwards.
210          */
211         err = ksm_madvise(vma, old_addr, old_addr + old_len,
212                                                 MADV_UNMERGEABLE, &vm_flags);
213         if (err)
214                 return err;
215
216         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
217         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
218         if (!new_vma)
219                 return -ENOMEM;
220
221         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
222         if (moved_len < old_len) {
223                 /*
224                  * Before moving the page tables from the new vma to
225                  * the old vma, we need to be sure the old vma is
226                  * queued after new vma in the same_anon_vma list to
227                  * prevent SMP races with rmap_walk (that could lead
228                  * rmap_walk to miss some page table).
229                  */
230                 anon_vma_moveto_tail(vma);
231
232                 /*
233                  * On error, move entries back from new area to old,
234                  * which will succeed since page tables still there,
235                  * and then proceed to unmap new area instead of old.
236                  */
237                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
238                 vma = new_vma;
239                 old_len = new_len;
240                 old_addr = new_addr;
241                 new_addr = -ENOMEM;
242         }
243
244         /* Conceal VM_ACCOUNT so old reservation is not undone */
245         if (vm_flags & VM_ACCOUNT) {
246                 vma->vm_flags &= ~VM_ACCOUNT;
247                 excess = vma->vm_end - vma->vm_start - old_len;
248                 if (old_addr > vma->vm_start &&
249                     old_addr + old_len < vma->vm_end)
250                         split = 1;
251         }
252
253         /*
254          * If we failed to move page tables we still do total_vm increment
255          * since do_munmap() will decrement it by old_len == new_len.
256          *
257          * Since total_vm is about to be raised artificially high for a
258          * moment, we need to restore high watermark afterwards: if stats
259          * are taken meanwhile, total_vm and hiwater_vm appear too high.
260          * If this were a serious issue, we'd add a flag to do_munmap().
261          */
262         hiwater_vm = mm->hiwater_vm;
263         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
264
265         if (do_munmap(mm, old_addr, old_len) < 0) {
266                 /* OOM: unable to split vma, just get accounts right */
267                 vm_unacct_memory(excess >> PAGE_SHIFT);
268                 excess = 0;
269         }
270         mm->hiwater_vm = hiwater_vm;
271
272         /* Restore VM_ACCOUNT if one or two pieces of vma left */
273         if (excess) {
274                 vma->vm_flags |= VM_ACCOUNT;
275                 if (split)
276                         vma->vm_next->vm_flags |= VM_ACCOUNT;
277         }
278
279         if (vm_flags & VM_LOCKED) {
280                 mm->locked_vm += new_len >> PAGE_SHIFT;
281                 if (new_len > old_len)
282                         mlock_vma_pages_range(new_vma, new_addr + old_len,
283                                                        new_addr + new_len);
284         }
285
286         return new_addr;
287 }
288
289 static struct vm_area_struct *vma_to_resize(unsigned long addr,
290         unsigned long old_len, unsigned long new_len, unsigned long *p)
291 {
292         struct mm_struct *mm = current->mm;
293         struct vm_area_struct *vma = find_vma(mm, addr);
294
295         if (!vma || vma->vm_start > addr)
296                 goto Efault;
297
298         if (is_vm_hugetlb_page(vma))
299                 goto Einval;
300
301         /* We can't remap across vm area boundaries */
302         if (old_len > vma->vm_end - addr)
303                 goto Efault;
304
305         /* Need to be careful about a growing mapping */
306         if (new_len > old_len) {
307                 unsigned long pgoff;
308
309                 if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP))
310                         goto Efault;
311                 pgoff = (addr - vma->vm_start) >> PAGE_SHIFT;
312                 pgoff += vma->vm_pgoff;
313                 if (pgoff + (new_len >> PAGE_SHIFT) < pgoff)
314                         goto Einval;
315         }
316
317         if (vma->vm_flags & VM_LOCKED) {
318                 unsigned long locked, lock_limit;
319                 locked = mm->locked_vm << PAGE_SHIFT;
320                 lock_limit = rlimit(RLIMIT_MEMLOCK);
321                 locked += new_len - old_len;
322                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
323                         goto Eagain;
324         }
325
326         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT))
327                 goto Enomem;
328
329         if (vma->vm_flags & VM_ACCOUNT) {
330                 unsigned long charged = (new_len - old_len) >> PAGE_SHIFT;
331                 if (security_vm_enough_memory_mm(mm, charged))
332                         goto Efault;
333                 *p = charged;
334         }
335
336         return vma;
337
338 Efault: /* very odd choice for most of the cases, but... */
339         return ERR_PTR(-EFAULT);
340 Einval:
341         return ERR_PTR(-EINVAL);
342 Enomem:
343         return ERR_PTR(-ENOMEM);
344 Eagain:
345         return ERR_PTR(-EAGAIN);
346 }
347
348 static unsigned long mremap_to(unsigned long addr,
349         unsigned long old_len, unsigned long new_addr,
350         unsigned long new_len)
351 {
352         struct mm_struct *mm = current->mm;
353         struct vm_area_struct *vma;
354         unsigned long ret = -EINVAL;
355         unsigned long charged = 0;
356         unsigned long map_flags;
357
358         if (new_addr & ~PAGE_MASK)
359                 goto out;
360
361         if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
362                 goto out;
363
364         /* Check if the location we're moving into overlaps the
365          * old location at all, and fail if it does.
366          */
367         if ((new_addr <= addr) && (new_addr+new_len) > addr)
368                 goto out;
369
370         if ((addr <= new_addr) && (addr+old_len) > new_addr)
371                 goto out;
372
373         ret = do_munmap(mm, new_addr, new_len);
374         if (ret)
375                 goto out;
376
377         if (old_len >= new_len) {
378                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
379                 if (ret && old_len != new_len)
380                         goto out;
381                 old_len = new_len;
382         }
383
384         vma = vma_to_resize(addr, old_len, new_len, &charged);
385         if (IS_ERR(vma)) {
386                 ret = PTR_ERR(vma);
387                 goto out;
388         }
389
390         map_flags = MAP_FIXED;
391         if (vma->vm_flags & VM_MAYSHARE)
392                 map_flags |= MAP_SHARED;
393
394         ret = get_unmapped_area(vma->vm_file, new_addr, new_len, vma->vm_pgoff +
395                                 ((addr - vma->vm_start) >> PAGE_SHIFT),
396                                 map_flags);
397         if (ret & ~PAGE_MASK)
398                 goto out1;
399
400         ret = move_vma(vma, addr, old_len, new_len, new_addr);
401         if (!(ret & ~PAGE_MASK))
402                 goto out;
403 out1:
404         vm_unacct_memory(charged);
405
406 out:
407         return ret;
408 }
409
410 static int vma_expandable(struct vm_area_struct *vma, unsigned long delta)
411 {
412         unsigned long end = vma->vm_end + delta;
413         if (end < vma->vm_end) /* overflow */
414                 return 0;
415         if (vma->vm_next && vma->vm_next->vm_start < end) /* intersection */
416                 return 0;
417         if (get_unmapped_area(NULL, vma->vm_start, end - vma->vm_start,
418                               0, MAP_FIXED) & ~PAGE_MASK)
419                 return 0;
420         return 1;
421 }
422
423 /*
424  * Expand (or shrink) an existing mapping, potentially moving it at the
425  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
426  *
427  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
428  * This option implies MREMAP_MAYMOVE.
429  */
430 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
431                 unsigned long, new_len, unsigned long, flags,
432                 unsigned long, new_addr)
433 {
434         struct mm_struct *mm = current->mm;
435         struct vm_area_struct *vma;
436         unsigned long ret = -EINVAL;
437         unsigned long charged = 0;
438
439         down_write(&current->mm->mmap_sem);
440
441         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
442                 goto out;
443
444         if (addr & ~PAGE_MASK)
445                 goto out;
446
447         old_len = PAGE_ALIGN(old_len);
448         new_len = PAGE_ALIGN(new_len);
449
450         /*
451          * We allow a zero old-len as a special case
452          * for DOS-emu "duplicate shm area" thing. But
453          * a zero new-len is nonsensical.
454          */
455         if (!new_len)
456                 goto out;
457
458         if (flags & MREMAP_FIXED) {
459                 if (flags & MREMAP_MAYMOVE)
460                         ret = mremap_to(addr, old_len, new_addr, new_len);
461                 goto out;
462         }
463
464         /*
465          * Always allow a shrinking remap: that just unmaps
466          * the unnecessary pages..
467          * do_munmap does all the needed commit accounting
468          */
469         if (old_len >= new_len) {
470                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
471                 if (ret && old_len != new_len)
472                         goto out;
473                 ret = addr;
474                 goto out;
475         }
476
477         /*
478          * Ok, we need to grow..
479          */
480         vma = vma_to_resize(addr, old_len, new_len, &charged);
481         if (IS_ERR(vma)) {
482                 ret = PTR_ERR(vma);
483                 goto out;
484         }
485
486         /* old_len exactly to the end of the area..
487          */
488         if (old_len == vma->vm_end - addr) {
489                 /* can we just expand the current mapping? */
490                 if (vma_expandable(vma, new_len - old_len)) {
491                         int pages = (new_len - old_len) >> PAGE_SHIFT;
492
493                         if (vma_adjust(vma, vma->vm_start, addr + new_len,
494                                        vma->vm_pgoff, NULL)) {
495                                 ret = -ENOMEM;
496                                 goto out;
497                         }
498
499                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
500                         if (vma->vm_flags & VM_LOCKED) {
501                                 mm->locked_vm += pages;
502                                 mlock_vma_pages_range(vma, addr + old_len,
503                                                    addr + new_len);
504                         }
505                         ret = addr;
506                         goto out;
507                 }
508         }
509
510         /*
511          * We weren't able to just expand or shrink the area,
512          * we need to create a new one and move it..
513          */
514         ret = -ENOMEM;
515         if (flags & MREMAP_MAYMOVE) {
516                 unsigned long map_flags = 0;
517                 if (vma->vm_flags & VM_MAYSHARE)
518                         map_flags |= MAP_SHARED;
519
520                 new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
521                                         vma->vm_pgoff +
522                                         ((addr - vma->vm_start) >> PAGE_SHIFT),
523                                         map_flags);
524                 if (new_addr & ~PAGE_MASK) {
525                         ret = new_addr;
526                         goto out;
527                 }
528
529                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
530         }
531 out:
532         if (ret & ~PAGE_MASK)
533                 vm_unacct_memory(charged);
534         up_write(&current->mm->mmap_sem);
535         return ret;
536 }