ia64/xen-unstable

view linux-2.6-xen-sparse/mm/mmap.c @ 13197:a9a43705f26b

Fix HVM booting through Xen-API when the kernel is unspecified.

Signed-off-by: Ewan Mellor <ewan@xensource.com>
author Ewan Mellor <ewan@xensource.com>
date Wed Dec 27 00:38:01 2006 +0000 (2006-12-27)
parents aaaa249e6f3b
children 4fad820a2233
line source
1 /*
2 * mm/mmap.c
3 *
4 * Written by obz.
5 *
6 * Address space accounting code <alan@redhat.com>
7 */
9 #include <linux/slab.h>
10 #include <linux/mm.h>
11 #include <linux/shm.h>
12 #include <linux/mman.h>
13 #include <linux/pagemap.h>
14 #include <linux/swap.h>
15 #include <linux/syscalls.h>
16 #include <linux/capability.h>
17 #include <linux/init.h>
18 #include <linux/file.h>
19 #include <linux/fs.h>
20 #include <linux/personality.h>
21 #include <linux/security.h>
22 #include <linux/hugetlb.h>
23 #include <linux/profile.h>
24 #include <linux/module.h>
25 #include <linux/mount.h>
26 #include <linux/mempolicy.h>
27 #include <linux/rmap.h>
29 #include <asm/uaccess.h>
30 #include <asm/cacheflush.h>
31 #include <asm/tlb.h>
33 #ifndef arch_mmap_check
34 #define arch_mmap_check(addr, len, flags) (0)
35 #endif
37 static void unmap_region(struct mm_struct *mm,
38 struct vm_area_struct *vma, struct vm_area_struct *prev,
39 unsigned long start, unsigned long end);
41 /*
42 * WARNING: the debugging will use recursive algorithms so never enable this
43 * unless you know what you are doing.
44 */
45 #undef DEBUG_MM_RB
47 /* description of effects of mapping type and prot in current implementation.
48 * this is due to the limited x86 page protection hardware. The expected
49 * behavior is in parens:
50 *
51 * map_type prot
52 * PROT_NONE PROT_READ PROT_WRITE PROT_EXEC
53 * MAP_SHARED r: (no) no r: (yes) yes r: (no) yes r: (no) yes
54 * w: (no) no w: (no) no w: (yes) yes w: (no) no
55 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
56 *
57 * MAP_PRIVATE r: (no) no r: (yes) yes r: (no) yes r: (no) yes
58 * w: (no) no w: (no) no w: (copy) copy w: (no) no
59 * x: (no) no x: (no) yes x: (no) yes x: (yes) yes
60 *
61 */
62 pgprot_t protection_map[16] = {
63 __P000, __P001, __P010, __P011, __P100, __P101, __P110, __P111,
64 __S000, __S001, __S010, __S011, __S100, __S101, __S110, __S111
65 };
67 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
68 int sysctl_overcommit_ratio = 50; /* default is 50% */
69 int sysctl_max_map_count __read_mostly = DEFAULT_MAX_MAP_COUNT;
70 atomic_t vm_committed_space = ATOMIC_INIT(0);
72 /*
73 * Check that a process has enough memory to allocate a new virtual
74 * mapping. 0 means there is enough memory for the allocation to
75 * succeed and -ENOMEM implies there is not.
76 *
77 * We currently support three overcommit policies, which are set via the
78 * vm.overcommit_memory sysctl. See Documentation/vm/overcommit-accounting
79 *
80 * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
81 * Additional code 2002 Jul 20 by Robert Love.
82 *
83 * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
84 *
85 * Note this is a helper function intended to be used by LSMs which
86 * wish to use this logic.
87 */
88 int __vm_enough_memory(long pages, int cap_sys_admin)
89 {
90 unsigned long free, allowed;
92 vm_acct_memory(pages);
94 /*
95 * Sometimes we want to use more memory than we have
96 */
97 if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
98 return 0;
100 if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
101 unsigned long n;
103 free = get_page_cache_size();
104 free += nr_swap_pages;
106 /*
107 * Any slabs which are created with the
108 * SLAB_RECLAIM_ACCOUNT flag claim to have contents
109 * which are reclaimable, under pressure. The dentry
110 * cache and most inode caches should fall into this
111 */
112 free += atomic_read(&slab_reclaim_pages);
114 /*
115 * Leave the last 3% for root
116 */
117 if (!cap_sys_admin)
118 free -= free / 32;
120 if (free > pages)
121 return 0;
123 /*
124 * nr_free_pages() is very expensive on large systems,
125 * only call if we're about to fail.
126 */
127 n = nr_free_pages();
128 if (!cap_sys_admin)
129 n -= n / 32;
130 free += n;
132 if (free > pages)
133 return 0;
134 vm_unacct_memory(pages);
135 return -ENOMEM;
136 }
138 allowed = (totalram_pages - hugetlb_total_pages())
139 * sysctl_overcommit_ratio / 100;
140 /*
141 * Leave the last 3% for root
142 */
143 if (!cap_sys_admin)
144 allowed -= allowed / 32;
145 allowed += total_swap_pages;
147 /* Don't let a single process grow too big:
148 leave 3% of the size of this process for other processes */
149 allowed -= current->mm->total_vm / 32;
151 /*
152 * cast `allowed' as a signed long because vm_committed_space
153 * sometimes has a negative value
154 */
155 if (atomic_read(&vm_committed_space) < (long)allowed)
156 return 0;
158 vm_unacct_memory(pages);
160 return -ENOMEM;
161 }
163 EXPORT_SYMBOL(__vm_enough_memory);
165 /*
166 * Requires inode->i_mapping->i_mmap_lock
167 */
168 static void __remove_shared_vm_struct(struct vm_area_struct *vma,
169 struct file *file, struct address_space *mapping)
170 {
171 if (vma->vm_flags & VM_DENYWRITE)
172 atomic_inc(&file->f_dentry->d_inode->i_writecount);
173 if (vma->vm_flags & VM_SHARED)
174 mapping->i_mmap_writable--;
176 flush_dcache_mmap_lock(mapping);
177 if (unlikely(vma->vm_flags & VM_NONLINEAR))
178 list_del_init(&vma->shared.vm_set.list);
179 else
180 vma_prio_tree_remove(vma, &mapping->i_mmap);
181 flush_dcache_mmap_unlock(mapping);
182 }
184 /*
185 * Unlink a file-based vm structure from its prio_tree, to hide
186 * vma from rmap and vmtruncate before freeing its page tables.
187 */
188 void unlink_file_vma(struct vm_area_struct *vma)
189 {
190 struct file *file = vma->vm_file;
192 if (file) {
193 struct address_space *mapping = file->f_mapping;
194 spin_lock(&mapping->i_mmap_lock);
195 __remove_shared_vm_struct(vma, file, mapping);
196 spin_unlock(&mapping->i_mmap_lock);
197 }
198 }
200 /*
201 * Close a vm structure and free it, returning the next.
202 */
203 static struct vm_area_struct *remove_vma(struct vm_area_struct *vma)
204 {
205 struct vm_area_struct *next = vma->vm_next;
207 might_sleep();
208 if (vma->vm_ops && vma->vm_ops->close)
209 vma->vm_ops->close(vma);
210 if (vma->vm_file)
211 fput(vma->vm_file);
212 mpol_free(vma_policy(vma));
213 kmem_cache_free(vm_area_cachep, vma);
214 return next;
215 }
217 asmlinkage unsigned long sys_brk(unsigned long brk)
218 {
219 unsigned long rlim, retval;
220 unsigned long newbrk, oldbrk;
221 struct mm_struct *mm = current->mm;
223 down_write(&mm->mmap_sem);
225 if (brk < mm->end_code)
226 goto out;
227 newbrk = PAGE_ALIGN(brk);
228 oldbrk = PAGE_ALIGN(mm->brk);
229 if (oldbrk == newbrk)
230 goto set_brk;
232 /* Always allow shrinking brk. */
233 if (brk <= mm->brk) {
234 if (!do_munmap(mm, newbrk, oldbrk-newbrk))
235 goto set_brk;
236 goto out;
237 }
239 /* Check against rlimit.. */
240 rlim = current->signal->rlim[RLIMIT_DATA].rlim_cur;
241 if (rlim < RLIM_INFINITY && brk - mm->start_data > rlim)
242 goto out;
244 /* Check against existing mmap mappings. */
245 if (find_vma_intersection(mm, oldbrk, newbrk+PAGE_SIZE))
246 goto out;
248 /* Ok, looks good - let it rip. */
249 if (do_brk(oldbrk, newbrk-oldbrk) != oldbrk)
250 goto out;
251 set_brk:
252 mm->brk = brk;
253 out:
254 retval = mm->brk;
255 up_write(&mm->mmap_sem);
256 return retval;
257 }
259 #ifdef DEBUG_MM_RB
260 static int browse_rb(struct rb_root *root)
261 {
262 int i = 0, j;
263 struct rb_node *nd, *pn = NULL;
264 unsigned long prev = 0, pend = 0;
266 for (nd = rb_first(root); nd; nd = rb_next(nd)) {
267 struct vm_area_struct *vma;
268 vma = rb_entry(nd, struct vm_area_struct, vm_rb);
269 if (vma->vm_start < prev)
270 printk("vm_start %lx prev %lx\n", vma->vm_start, prev), i = -1;
271 if (vma->vm_start < pend)
272 printk("vm_start %lx pend %lx\n", vma->vm_start, pend);
273 if (vma->vm_start > vma->vm_end)
274 printk("vm_end %lx < vm_start %lx\n", vma->vm_end, vma->vm_start);
275 i++;
276 pn = nd;
277 }
278 j = 0;
279 for (nd = pn; nd; nd = rb_prev(nd)) {
280 j++;
281 }
282 if (i != j)
283 printk("backwards %d, forwards %d\n", j, i), i = 0;
284 return i;
285 }
287 void validate_mm(struct mm_struct *mm)
288 {
289 int bug = 0;
290 int i = 0;
291 struct vm_area_struct *tmp = mm->mmap;
292 while (tmp) {
293 tmp = tmp->vm_next;
294 i++;
295 }
296 if (i != mm->map_count)
297 printk("map_count %d vm_next %d\n", mm->map_count, i), bug = 1;
298 i = browse_rb(&mm->mm_rb);
299 if (i != mm->map_count)
300 printk("map_count %d rb %d\n", mm->map_count, i), bug = 1;
301 if (bug)
302 BUG();
303 }
304 #else
305 #define validate_mm(mm) do { } while (0)
306 #endif
308 static struct vm_area_struct *
309 find_vma_prepare(struct mm_struct *mm, unsigned long addr,
310 struct vm_area_struct **pprev, struct rb_node ***rb_link,
311 struct rb_node ** rb_parent)
312 {
313 struct vm_area_struct * vma;
314 struct rb_node ** __rb_link, * __rb_parent, * rb_prev;
316 __rb_link = &mm->mm_rb.rb_node;
317 rb_prev = __rb_parent = NULL;
318 vma = NULL;
320 while (*__rb_link) {
321 struct vm_area_struct *vma_tmp;
323 __rb_parent = *__rb_link;
324 vma_tmp = rb_entry(__rb_parent, struct vm_area_struct, vm_rb);
326 if (vma_tmp->vm_end > addr) {
327 vma = vma_tmp;
328 if (vma_tmp->vm_start <= addr)
329 return vma;
330 __rb_link = &__rb_parent->rb_left;
331 } else {
332 rb_prev = __rb_parent;
333 __rb_link = &__rb_parent->rb_right;
334 }
335 }
337 *pprev = NULL;
338 if (rb_prev)
339 *pprev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
340 *rb_link = __rb_link;
341 *rb_parent = __rb_parent;
342 return vma;
343 }
345 static inline void
346 __vma_link_list(struct mm_struct *mm, struct vm_area_struct *vma,
347 struct vm_area_struct *prev, struct rb_node *rb_parent)
348 {
349 if (prev) {
350 vma->vm_next = prev->vm_next;
351 prev->vm_next = vma;
352 } else {
353 mm->mmap = vma;
354 if (rb_parent)
355 vma->vm_next = rb_entry(rb_parent,
356 struct vm_area_struct, vm_rb);
357 else
358 vma->vm_next = NULL;
359 }
360 }
362 void __vma_link_rb(struct mm_struct *mm, struct vm_area_struct *vma,
363 struct rb_node **rb_link, struct rb_node *rb_parent)
364 {
365 rb_link_node(&vma->vm_rb, rb_parent, rb_link);
366 rb_insert_color(&vma->vm_rb, &mm->mm_rb);
367 }
369 static inline void __vma_link_file(struct vm_area_struct *vma)
370 {
371 struct file * file;
373 file = vma->vm_file;
374 if (file) {
375 struct address_space *mapping = file->f_mapping;
377 if (vma->vm_flags & VM_DENYWRITE)
378 atomic_dec(&file->f_dentry->d_inode->i_writecount);
379 if (vma->vm_flags & VM_SHARED)
380 mapping->i_mmap_writable++;
382 flush_dcache_mmap_lock(mapping);
383 if (unlikely(vma->vm_flags & VM_NONLINEAR))
384 vma_nonlinear_insert(vma, &mapping->i_mmap_nonlinear);
385 else
386 vma_prio_tree_insert(vma, &mapping->i_mmap);
387 flush_dcache_mmap_unlock(mapping);
388 }
389 }
391 static void
392 __vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
393 struct vm_area_struct *prev, struct rb_node **rb_link,
394 struct rb_node *rb_parent)
395 {
396 __vma_link_list(mm, vma, prev, rb_parent);
397 __vma_link_rb(mm, vma, rb_link, rb_parent);
398 __anon_vma_link(vma);
399 }
401 static void vma_link(struct mm_struct *mm, struct vm_area_struct *vma,
402 struct vm_area_struct *prev, struct rb_node **rb_link,
403 struct rb_node *rb_parent)
404 {
405 struct address_space *mapping = NULL;
407 if (vma->vm_file)
408 mapping = vma->vm_file->f_mapping;
410 if (mapping) {
411 spin_lock(&mapping->i_mmap_lock);
412 vma->vm_truncate_count = mapping->truncate_count;
413 }
414 anon_vma_lock(vma);
416 __vma_link(mm, vma, prev, rb_link, rb_parent);
417 __vma_link_file(vma);
419 anon_vma_unlock(vma);
420 if (mapping)
421 spin_unlock(&mapping->i_mmap_lock);
423 mm->map_count++;
424 validate_mm(mm);
425 }
427 /*
428 * Helper for vma_adjust in the split_vma insert case:
429 * insert vm structure into list and rbtree and anon_vma,
430 * but it has already been inserted into prio_tree earlier.
431 */
432 static void
433 __insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
434 {
435 struct vm_area_struct * __vma, * prev;
436 struct rb_node ** rb_link, * rb_parent;
438 __vma = find_vma_prepare(mm, vma->vm_start,&prev, &rb_link, &rb_parent);
439 if (__vma && __vma->vm_start < vma->vm_end)
440 BUG();
441 __vma_link(mm, vma, prev, rb_link, rb_parent);
442 mm->map_count++;
443 }
445 static inline void
446 __vma_unlink(struct mm_struct *mm, struct vm_area_struct *vma,
447 struct vm_area_struct *prev)
448 {
449 prev->vm_next = vma->vm_next;
450 rb_erase(&vma->vm_rb, &mm->mm_rb);
451 if (mm->mmap_cache == vma)
452 mm->mmap_cache = prev;
453 }
455 /*
456 * We cannot adjust vm_start, vm_end, vm_pgoff fields of a vma that
457 * is already present in an i_mmap tree without adjusting the tree.
458 * The following helper function should be used when such adjustments
459 * are necessary. The "insert" vma (if any) is to be inserted
460 * before we drop the necessary locks.
461 */
462 void vma_adjust(struct vm_area_struct *vma, unsigned long start,
463 unsigned long end, pgoff_t pgoff, struct vm_area_struct *insert)
464 {
465 struct mm_struct *mm = vma->vm_mm;
466 struct vm_area_struct *next = vma->vm_next;
467 struct vm_area_struct *importer = NULL;
468 struct address_space *mapping = NULL;
469 struct prio_tree_root *root = NULL;
470 struct file *file = vma->vm_file;
471 struct anon_vma *anon_vma = NULL;
472 long adjust_next = 0;
473 int remove_next = 0;
475 if (next && !insert) {
476 if (end >= next->vm_end) {
477 /*
478 * vma expands, overlapping all the next, and
479 * perhaps the one after too (mprotect case 6).
480 */
481 again: remove_next = 1 + (end > next->vm_end);
482 end = next->vm_end;
483 anon_vma = next->anon_vma;
484 importer = vma;
485 } else if (end > next->vm_start) {
486 /*
487 * vma expands, overlapping part of the next:
488 * mprotect case 5 shifting the boundary up.
489 */
490 adjust_next = (end - next->vm_start) >> PAGE_SHIFT;
491 anon_vma = next->anon_vma;
492 importer = vma;
493 } else if (end < vma->vm_end) {
494 /*
495 * vma shrinks, and !insert tells it's not
496 * split_vma inserting another: so it must be
497 * mprotect case 4 shifting the boundary down.
498 */
499 adjust_next = - ((vma->vm_end - end) >> PAGE_SHIFT);
500 anon_vma = next->anon_vma;
501 importer = next;
502 }
503 }
505 if (file) {
506 mapping = file->f_mapping;
507 if (!(vma->vm_flags & VM_NONLINEAR))
508 root = &mapping->i_mmap;
509 spin_lock(&mapping->i_mmap_lock);
510 if (importer &&
511 vma->vm_truncate_count != next->vm_truncate_count) {
512 /*
513 * unmap_mapping_range might be in progress:
514 * ensure that the expanding vma is rescanned.
515 */
516 importer->vm_truncate_count = 0;
517 }
518 if (insert) {
519 insert->vm_truncate_count = vma->vm_truncate_count;
520 /*
521 * Put into prio_tree now, so instantiated pages
522 * are visible to arm/parisc __flush_dcache_page
523 * throughout; but we cannot insert into address
524 * space until vma start or end is updated.
525 */
526 __vma_link_file(insert);
527 }
528 }
530 /*
531 * When changing only vma->vm_end, we don't really need
532 * anon_vma lock: but is that case worth optimizing out?
533 */
534 if (vma->anon_vma)
535 anon_vma = vma->anon_vma;
536 if (anon_vma) {
537 spin_lock(&anon_vma->lock);
538 /*
539 * Easily overlooked: when mprotect shifts the boundary,
540 * make sure the expanding vma has anon_vma set if the
541 * shrinking vma had, to cover any anon pages imported.
542 */
543 if (importer && !importer->anon_vma) {
544 importer->anon_vma = anon_vma;
545 __anon_vma_link(importer);
546 }
547 }
549 if (root) {
550 flush_dcache_mmap_lock(mapping);
551 vma_prio_tree_remove(vma, root);
552 if (adjust_next)
553 vma_prio_tree_remove(next, root);
554 }
556 vma->vm_start = start;
557 vma->vm_end = end;
558 vma->vm_pgoff = pgoff;
559 if (adjust_next) {
560 next->vm_start += adjust_next << PAGE_SHIFT;
561 next->vm_pgoff += adjust_next;
562 }
564 if (root) {
565 if (adjust_next)
566 vma_prio_tree_insert(next, root);
567 vma_prio_tree_insert(vma, root);
568 flush_dcache_mmap_unlock(mapping);
569 }
571 if (remove_next) {
572 /*
573 * vma_merge has merged next into vma, and needs
574 * us to remove next before dropping the locks.
575 */
576 __vma_unlink(mm, next, vma);
577 if (file)
578 __remove_shared_vm_struct(next, file, mapping);
579 if (next->anon_vma)
580 __anon_vma_merge(vma, next);
581 } else if (insert) {
582 /*
583 * split_vma has split insert from vma, and needs
584 * us to insert it before dropping the locks
585 * (it may either follow vma or precede it).
586 */
587 __insert_vm_struct(mm, insert);
588 }
590 if (anon_vma)
591 spin_unlock(&anon_vma->lock);
592 if (mapping)
593 spin_unlock(&mapping->i_mmap_lock);
595 if (remove_next) {
596 if (file)
597 fput(file);
598 mm->map_count--;
599 mpol_free(vma_policy(next));
600 kmem_cache_free(vm_area_cachep, next);
601 /*
602 * In mprotect's case 6 (see comments on vma_merge),
603 * we must remove another next too. It would clutter
604 * up the code too much to do both in one go.
605 */
606 if (remove_next == 2) {
607 next = vma->vm_next;
608 goto again;
609 }
610 }
612 validate_mm(mm);
613 }
615 /*
616 * If the vma has a ->close operation then the driver probably needs to release
617 * per-vma resources, so we don't attempt to merge those.
618 */
619 #define VM_SPECIAL (VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_RESERVED | VM_PFNMAP)
621 static inline int is_mergeable_vma(struct vm_area_struct *vma,
622 struct file *file, unsigned long vm_flags)
623 {
624 if (vma->vm_flags != vm_flags)
625 return 0;
626 if (vma->vm_file != file)
627 return 0;
628 if (vma->vm_ops && vma->vm_ops->close)
629 return 0;
630 return 1;
631 }
633 static inline int is_mergeable_anon_vma(struct anon_vma *anon_vma1,
634 struct anon_vma *anon_vma2)
635 {
636 return !anon_vma1 || !anon_vma2 || (anon_vma1 == anon_vma2);
637 }
639 /*
640 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
641 * in front of (at a lower virtual address and file offset than) the vma.
642 *
643 * We cannot merge two vmas if they have differently assigned (non-NULL)
644 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
645 *
646 * We don't check here for the merged mmap wrapping around the end of pagecache
647 * indices (16TB on ia32) because do_mmap_pgoff() does not permit mmap's which
648 * wrap, nor mmaps which cover the final page at index -1UL.
649 */
650 static int
651 can_vma_merge_before(struct vm_area_struct *vma, unsigned long vm_flags,
652 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
653 {
654 if (is_mergeable_vma(vma, file, vm_flags) &&
655 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
656 if (vma->vm_pgoff == vm_pgoff)
657 return 1;
658 }
659 return 0;
660 }
662 /*
663 * Return true if we can merge this (vm_flags,anon_vma,file,vm_pgoff)
664 * beyond (at a higher virtual address and file offset than) the vma.
665 *
666 * We cannot merge two vmas if they have differently assigned (non-NULL)
667 * anon_vmas, nor if same anon_vma is assigned but offsets incompatible.
668 */
669 static int
670 can_vma_merge_after(struct vm_area_struct *vma, unsigned long vm_flags,
671 struct anon_vma *anon_vma, struct file *file, pgoff_t vm_pgoff)
672 {
673 if (is_mergeable_vma(vma, file, vm_flags) &&
674 is_mergeable_anon_vma(anon_vma, vma->anon_vma)) {
675 pgoff_t vm_pglen;
676 vm_pglen = (vma->vm_end - vma->vm_start) >> PAGE_SHIFT;
677 if (vma->vm_pgoff + vm_pglen == vm_pgoff)
678 return 1;
679 }
680 return 0;
681 }
683 /*
684 * Given a mapping request (addr,end,vm_flags,file,pgoff), figure out
685 * whether that can be merged with its predecessor or its successor.
686 * Or both (it neatly fills a hole).
687 *
688 * In most cases - when called for mmap, brk or mremap - [addr,end) is
689 * certain not to be mapped by the time vma_merge is called; but when
690 * called for mprotect, it is certain to be already mapped (either at
691 * an offset within prev, or at the start of next), and the flags of
692 * this area are about to be changed to vm_flags - and the no-change
693 * case has already been eliminated.
694 *
695 * The following mprotect cases have to be considered, where AAAA is
696 * the area passed down from mprotect_fixup, never extending beyond one
697 * vma, PPPPPP is the prev vma specified, and NNNNNN the next vma after:
698 *
699 * AAAA AAAA AAAA AAAA
700 * PPPPPPNNNNNN PPPPPPNNNNNN PPPPPPNNNNNN PPPPNNNNXXXX
701 * cannot merge might become might become might become
702 * PPNNNNNNNNNN PPPPPPPPPPNN PPPPPPPPPPPP 6 or
703 * mmap, brk or case 4 below case 5 below PPPPPPPPXXXX 7 or
704 * mremap move: PPPPNNNNNNNN 8
705 * AAAA
706 * PPPP NNNN PPPPPPPPPPPP PPPPPPPPNNNN PPPPNNNNNNNN
707 * might become case 1 below case 2 below case 3 below
708 *
709 * Odd one out? Case 8, because it extends NNNN but needs flags of XXXX:
710 * mprotect_fixup updates vm_flags & vm_page_prot on successful return.
711 */
712 struct vm_area_struct *vma_merge(struct mm_struct *mm,
713 struct vm_area_struct *prev, unsigned long addr,
714 unsigned long end, unsigned long vm_flags,
715 struct anon_vma *anon_vma, struct file *file,
716 pgoff_t pgoff, struct mempolicy *policy)
717 {
718 pgoff_t pglen = (end - addr) >> PAGE_SHIFT;
719 struct vm_area_struct *area, *next;
721 /*
722 * We later require that vma->vm_flags == vm_flags,
723 * so this tests vma->vm_flags & VM_SPECIAL, too.
724 */
725 if (vm_flags & VM_SPECIAL)
726 return NULL;
728 if (prev)
729 next = prev->vm_next;
730 else
731 next = mm->mmap;
732 area = next;
733 if (next && next->vm_end == end) /* cases 6, 7, 8 */
734 next = next->vm_next;
736 /*
737 * Can it merge with the predecessor?
738 */
739 if (prev && prev->vm_end == addr &&
740 mpol_equal(vma_policy(prev), policy) &&
741 can_vma_merge_after(prev, vm_flags,
742 anon_vma, file, pgoff)) {
743 /*
744 * OK, it can. Can we now merge in the successor as well?
745 */
746 if (next && end == next->vm_start &&
747 mpol_equal(policy, vma_policy(next)) &&
748 can_vma_merge_before(next, vm_flags,
749 anon_vma, file, pgoff+pglen) &&
750 is_mergeable_anon_vma(prev->anon_vma,
751 next->anon_vma)) {
752 /* cases 1, 6 */
753 vma_adjust(prev, prev->vm_start,
754 next->vm_end, prev->vm_pgoff, NULL);
755 } else /* cases 2, 5, 7 */
756 vma_adjust(prev, prev->vm_start,
757 end, prev->vm_pgoff, NULL);
758 return prev;
759 }
761 /*
762 * Can this new request be merged in front of next?
763 */
764 if (next && end == next->vm_start &&
765 mpol_equal(policy, vma_policy(next)) &&
766 can_vma_merge_before(next, vm_flags,
767 anon_vma, file, pgoff+pglen)) {
768 if (prev && addr < prev->vm_end) /* case 4 */
769 vma_adjust(prev, prev->vm_start,
770 addr, prev->vm_pgoff, NULL);
771 else /* cases 3, 8 */
772 vma_adjust(area, addr, next->vm_end,
773 next->vm_pgoff - pglen, NULL);
774 return area;
775 }
777 return NULL;
778 }
780 /*
781 * find_mergeable_anon_vma is used by anon_vma_prepare, to check
782 * neighbouring vmas for a suitable anon_vma, before it goes off
783 * to allocate a new anon_vma. It checks because a repetitive
784 * sequence of mprotects and faults may otherwise lead to distinct
785 * anon_vmas being allocated, preventing vma merge in subsequent
786 * mprotect.
787 */
788 struct anon_vma *find_mergeable_anon_vma(struct vm_area_struct *vma)
789 {
790 struct vm_area_struct *near;
791 unsigned long vm_flags;
793 near = vma->vm_next;
794 if (!near)
795 goto try_prev;
797 /*
798 * Since only mprotect tries to remerge vmas, match flags
799 * which might be mprotected into each other later on.
800 * Neither mlock nor madvise tries to remerge at present,
801 * so leave their flags as obstructing a merge.
802 */
803 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
804 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
806 if (near->anon_vma && vma->vm_end == near->vm_start &&
807 mpol_equal(vma_policy(vma), vma_policy(near)) &&
808 can_vma_merge_before(near, vm_flags,
809 NULL, vma->vm_file, vma->vm_pgoff +
810 ((vma->vm_end - vma->vm_start) >> PAGE_SHIFT)))
811 return near->anon_vma;
812 try_prev:
813 /*
814 * It is potentially slow to have to call find_vma_prev here.
815 * But it's only on the first write fault on the vma, not
816 * every time, and we could devise a way to avoid it later
817 * (e.g. stash info in next's anon_vma_node when assigning
818 * an anon_vma, or when trying vma_merge). Another time.
819 */
820 if (find_vma_prev(vma->vm_mm, vma->vm_start, &near) != vma)
821 BUG();
822 if (!near)
823 goto none;
825 vm_flags = vma->vm_flags & ~(VM_READ|VM_WRITE|VM_EXEC);
826 vm_flags |= near->vm_flags & (VM_READ|VM_WRITE|VM_EXEC);
828 if (near->anon_vma && near->vm_end == vma->vm_start &&
829 mpol_equal(vma_policy(near), vma_policy(vma)) &&
830 can_vma_merge_after(near, vm_flags,
831 NULL, vma->vm_file, vma->vm_pgoff))
832 return near->anon_vma;
833 none:
834 /*
835 * There's no absolute need to look only at touching neighbours:
836 * we could search further afield for "compatible" anon_vmas.
837 * But it would probably just be a waste of time searching,
838 * or lead to too many vmas hanging off the same anon_vma.
839 * We're trying to allow mprotect remerging later on,
840 * not trying to minimize memory used for anon_vmas.
841 */
842 return NULL;
843 }
845 #ifdef CONFIG_PROC_FS
846 void vm_stat_account(struct mm_struct *mm, unsigned long flags,
847 struct file *file, long pages)
848 {
849 const unsigned long stack_flags
850 = VM_STACK_FLAGS & (VM_GROWSUP|VM_GROWSDOWN);
852 #ifdef CONFIG_HUGETLB
853 if (flags & VM_HUGETLB) {
854 if (!(flags & VM_DONTCOPY))
855 mm->shared_vm += pages;
856 return;
857 }
858 #endif /* CONFIG_HUGETLB */
860 if (file) {
861 mm->shared_vm += pages;
862 if ((flags & (VM_EXEC|VM_WRITE)) == VM_EXEC)
863 mm->exec_vm += pages;
864 } else if (flags & stack_flags)
865 mm->stack_vm += pages;
866 if (flags & (VM_RESERVED|VM_IO))
867 mm->reserved_vm += pages;
868 }
869 #endif /* CONFIG_PROC_FS */
871 /*
872 * The caller must hold down_write(current->mm->mmap_sem).
873 */
875 unsigned long do_mmap_pgoff(struct file * file, unsigned long addr,
876 unsigned long len, unsigned long prot,
877 unsigned long flags, unsigned long pgoff)
878 {
879 struct mm_struct * mm = current->mm;
880 struct vm_area_struct * vma, * prev;
881 struct inode *inode;
882 unsigned int vm_flags;
883 int correct_wcount = 0;
884 int error;
885 struct rb_node ** rb_link, * rb_parent;
886 int accountable = 1;
887 unsigned long charged = 0, reqprot = prot;
889 if (file) {
890 if (is_file_hugepages(file))
891 accountable = 0;
893 if (!file->f_op || !file->f_op->mmap)
894 return -ENODEV;
896 if ((prot & PROT_EXEC) &&
897 (file->f_vfsmnt->mnt_flags & MNT_NOEXEC))
898 return -EPERM;
899 }
900 /*
901 * Does the application expect PROT_READ to imply PROT_EXEC?
902 *
903 * (the exception is when the underlying filesystem is noexec
904 * mounted, in which case we dont add PROT_EXEC.)
905 */
906 if ((prot & PROT_READ) && (current->personality & READ_IMPLIES_EXEC))
907 if (!(file && (file->f_vfsmnt->mnt_flags & MNT_NOEXEC)))
908 prot |= PROT_EXEC;
910 if (!len)
911 return -EINVAL;
913 error = arch_mmap_check(addr, len, flags);
914 if (error)
915 return error;
917 /* Careful about overflows.. */
918 len = PAGE_ALIGN(len);
919 if (!len || len > TASK_SIZE)
920 return -ENOMEM;
922 /* offset overflow? */
923 if ((pgoff + (len >> PAGE_SHIFT)) < pgoff)
924 return -EOVERFLOW;
926 /* Too many mappings? */
927 if (mm->map_count > sysctl_max_map_count)
928 return -ENOMEM;
930 /* Obtain the address to map to. we verify (or select) it and ensure
931 * that it represents a valid section of the address space.
932 */
933 addr = get_unmapped_area(file, addr, len, pgoff, flags);
934 if (addr & ~PAGE_MASK)
935 return addr;
937 /* Do simple checking here so the lower-level routines won't have
938 * to. we assume access permissions have been handled by the open
939 * of the memory object, so we don't do any here.
940 */
941 vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags) |
942 mm->def_flags | VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
944 if (flags & MAP_LOCKED) {
945 if (!can_do_mlock())
946 return -EPERM;
947 vm_flags |= VM_LOCKED;
948 }
949 /* mlock MCL_FUTURE? */
950 if (vm_flags & VM_LOCKED) {
951 unsigned long locked, lock_limit;
952 locked = len >> PAGE_SHIFT;
953 locked += mm->locked_vm;
954 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
955 lock_limit >>= PAGE_SHIFT;
956 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
957 return -EAGAIN;
958 }
960 inode = file ? file->f_dentry->d_inode : NULL;
962 if (file) {
963 switch (flags & MAP_TYPE) {
964 case MAP_SHARED:
965 if ((prot&PROT_WRITE) && !(file->f_mode&FMODE_WRITE))
966 return -EACCES;
968 /*
969 * Make sure we don't allow writing to an append-only
970 * file..
971 */
972 if (IS_APPEND(inode) && (file->f_mode & FMODE_WRITE))
973 return -EACCES;
975 /*
976 * Make sure there are no mandatory locks on the file.
977 */
978 if (locks_verify_locked(inode))
979 return -EAGAIN;
981 vm_flags |= VM_SHARED | VM_MAYSHARE;
982 if (!(file->f_mode & FMODE_WRITE))
983 vm_flags &= ~(VM_MAYWRITE | VM_SHARED);
985 /* fall through */
986 case MAP_PRIVATE:
987 if (!(file->f_mode & FMODE_READ))
988 return -EACCES;
989 break;
991 default:
992 return -EINVAL;
993 }
994 } else {
995 switch (flags & MAP_TYPE) {
996 case MAP_SHARED:
997 vm_flags |= VM_SHARED | VM_MAYSHARE;
998 break;
999 case MAP_PRIVATE:
1000 /*
1001 * Set pgoff according to addr for anon_vma.
1002 */
1003 pgoff = addr >> PAGE_SHIFT;
1004 break;
1005 default:
1006 return -EINVAL;
1010 error = security_file_mmap(file, reqprot, prot, flags);
1011 if (error)
1012 return error;
1014 /* Clear old maps */
1015 error = -ENOMEM;
1016 munmap_back:
1017 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1018 if (vma && vma->vm_start < addr + len) {
1019 if (do_munmap(mm, addr, len))
1020 return -ENOMEM;
1021 goto munmap_back;
1024 /* Check against address space limit. */
1025 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1026 return -ENOMEM;
1028 if (accountable && (!(flags & MAP_NORESERVE) ||
1029 sysctl_overcommit_memory == OVERCOMMIT_NEVER)) {
1030 if (vm_flags & VM_SHARED) {
1031 /* Check memory availability in shmem_file_setup? */
1032 vm_flags |= VM_ACCOUNT;
1033 } else if (vm_flags & VM_WRITE) {
1034 /*
1035 * Private writable mapping: check memory availability
1036 */
1037 charged = len >> PAGE_SHIFT;
1038 if (security_vm_enough_memory(charged))
1039 return -ENOMEM;
1040 vm_flags |= VM_ACCOUNT;
1044 /*
1045 * Can we just expand an old private anonymous mapping?
1046 * The VM_SHARED test is necessary because shmem_zero_setup
1047 * will create the file object for a shared anonymous map below.
1048 */
1049 if (!file && !(vm_flags & VM_SHARED) &&
1050 vma_merge(mm, prev, addr, addr + len, vm_flags,
1051 NULL, NULL, pgoff, NULL))
1052 goto out;
1054 /*
1055 * Determine the object being mapped and call the appropriate
1056 * specific mapper. the address has already been validated, but
1057 * not unmapped, but the maps are removed from the list.
1058 */
1059 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1060 if (!vma) {
1061 error = -ENOMEM;
1062 goto unacct_error;
1064 memset(vma, 0, sizeof(*vma));
1066 vma->vm_mm = mm;
1067 vma->vm_start = addr;
1068 vma->vm_end = addr + len;
1069 vma->vm_flags = vm_flags;
1070 vma->vm_page_prot = protection_map[vm_flags & 0x0f];
1071 vma->vm_pgoff = pgoff;
1073 if (file) {
1074 error = -EINVAL;
1075 if (vm_flags & (VM_GROWSDOWN|VM_GROWSUP))
1076 goto free_vma;
1077 if (vm_flags & VM_DENYWRITE) {
1078 error = deny_write_access(file);
1079 if (error)
1080 goto free_vma;
1081 correct_wcount = 1;
1083 vma->vm_file = file;
1084 get_file(file);
1085 error = file->f_op->mmap(file, vma);
1086 if (error)
1087 goto unmap_and_free_vma;
1088 } else if (vm_flags & VM_SHARED) {
1089 error = shmem_zero_setup(vma);
1090 if (error)
1091 goto free_vma;
1094 /* We set VM_ACCOUNT in a shared mapping's vm_flags, to inform
1095 * shmem_zero_setup (perhaps called through /dev/zero's ->mmap)
1096 * that memory reservation must be checked; but that reservation
1097 * belongs to shared memory object, not to vma: so now clear it.
1098 */
1099 if ((vm_flags & (VM_SHARED|VM_ACCOUNT)) == (VM_SHARED|VM_ACCOUNT))
1100 vma->vm_flags &= ~VM_ACCOUNT;
1102 /* Can addr have changed??
1104 * Answer: Yes, several device drivers can do it in their
1105 * f_op->mmap method. -DaveM
1106 */
1107 addr = vma->vm_start;
1108 pgoff = vma->vm_pgoff;
1109 vm_flags = vma->vm_flags;
1111 if (!file || !vma_merge(mm, prev, addr, vma->vm_end,
1112 vma->vm_flags, NULL, file, pgoff, vma_policy(vma))) {
1113 file = vma->vm_file;
1114 vma_link(mm, vma, prev, rb_link, rb_parent);
1115 if (correct_wcount)
1116 atomic_inc(&inode->i_writecount);
1117 } else {
1118 if (file) {
1119 if (correct_wcount)
1120 atomic_inc(&inode->i_writecount);
1121 fput(file);
1123 mpol_free(vma_policy(vma));
1124 kmem_cache_free(vm_area_cachep, vma);
1126 out:
1127 mm->total_vm += len >> PAGE_SHIFT;
1128 vm_stat_account(mm, vm_flags, file, len >> PAGE_SHIFT);
1129 if (vm_flags & VM_LOCKED) {
1130 mm->locked_vm += len >> PAGE_SHIFT;
1131 make_pages_present(addr, addr + len);
1133 if (flags & MAP_POPULATE) {
1134 up_write(&mm->mmap_sem);
1135 sys_remap_file_pages(addr, len, 0,
1136 pgoff, flags & MAP_NONBLOCK);
1137 down_write(&mm->mmap_sem);
1139 return addr;
1141 unmap_and_free_vma:
1142 if (correct_wcount)
1143 atomic_inc(&inode->i_writecount);
1144 vma->vm_file = NULL;
1145 fput(file);
1147 /* Undo any partial mapping done by a device driver. */
1148 unmap_region(mm, vma, prev, vma->vm_start, vma->vm_end);
1149 charged = 0;
1150 free_vma:
1151 kmem_cache_free(vm_area_cachep, vma);
1152 unacct_error:
1153 if (charged)
1154 vm_unacct_memory(charged);
1155 return error;
1158 EXPORT_SYMBOL(do_mmap_pgoff);
1160 /* Get an address range which is currently unmapped.
1161 * For shmat() with addr=0.
1163 * Ugly calling convention alert:
1164 * Return value with the low bits set means error value,
1165 * ie
1166 * if (ret & ~PAGE_MASK)
1167 * error = ret;
1169 * This function "knows" that -ENOMEM has the bits set.
1170 */
1171 #ifndef HAVE_ARCH_UNMAPPED_AREA
1172 unsigned long
1173 arch_get_unmapped_area(struct file *filp, unsigned long addr,
1174 unsigned long len, unsigned long pgoff, unsigned long flags)
1176 struct mm_struct *mm = current->mm;
1177 struct vm_area_struct *vma;
1178 unsigned long start_addr;
1180 if (len > TASK_SIZE)
1181 return -ENOMEM;
1183 if (addr) {
1184 addr = PAGE_ALIGN(addr);
1185 vma = find_vma(mm, addr);
1186 if (TASK_SIZE - len >= addr &&
1187 (!vma || addr + len <= vma->vm_start))
1188 return addr;
1190 if (len > mm->cached_hole_size) {
1191 start_addr = addr = mm->free_area_cache;
1192 } else {
1193 start_addr = addr = TASK_UNMAPPED_BASE;
1194 mm->cached_hole_size = 0;
1197 full_search:
1198 for (vma = find_vma(mm, addr); ; vma = vma->vm_next) {
1199 /* At this point: (!vma || addr < vma->vm_end). */
1200 if (TASK_SIZE - len < addr) {
1201 /*
1202 * Start a new search - just in case we missed
1203 * some holes.
1204 */
1205 if (start_addr != TASK_UNMAPPED_BASE) {
1206 addr = TASK_UNMAPPED_BASE;
1207 start_addr = addr;
1208 mm->cached_hole_size = 0;
1209 goto full_search;
1211 return -ENOMEM;
1213 if (!vma || addr + len <= vma->vm_start) {
1214 /*
1215 * Remember the place where we stopped the search:
1216 */
1217 mm->free_area_cache = addr + len;
1218 return addr;
1220 if (addr + mm->cached_hole_size < vma->vm_start)
1221 mm->cached_hole_size = vma->vm_start - addr;
1222 addr = vma->vm_end;
1225 #endif
1227 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1229 /*
1230 * Is this a new hole at the lowest possible address?
1231 */
1232 if (addr >= TASK_UNMAPPED_BASE && addr < mm->free_area_cache) {
1233 mm->free_area_cache = addr;
1234 mm->cached_hole_size = ~0UL;
1238 /*
1239 * This mmap-allocator allocates new areas top-down from below the
1240 * stack's low limit (the base):
1241 */
1242 #ifndef HAVE_ARCH_UNMAPPED_AREA_TOPDOWN
1243 unsigned long
1244 arch_get_unmapped_area_topdown(struct file *filp, const unsigned long addr0,
1245 const unsigned long len, const unsigned long pgoff,
1246 const unsigned long flags)
1248 struct vm_area_struct *vma;
1249 struct mm_struct *mm = current->mm;
1250 unsigned long addr = addr0;
1252 /* requested length too big for entire address space */
1253 if (len > TASK_SIZE)
1254 return -ENOMEM;
1256 /* requesting a specific address */
1257 if (addr) {
1258 addr = PAGE_ALIGN(addr);
1259 vma = find_vma(mm, addr);
1260 if (TASK_SIZE - len >= addr &&
1261 (!vma || addr + len <= vma->vm_start))
1262 return addr;
1265 /* check if free_area_cache is useful for us */
1266 if (len <= mm->cached_hole_size) {
1267 mm->cached_hole_size = 0;
1268 mm->free_area_cache = mm->mmap_base;
1271 /* either no address requested or can't fit in requested address hole */
1272 addr = mm->free_area_cache;
1274 /* make sure it can fit in the remaining address space */
1275 if (addr > len) {
1276 vma = find_vma(mm, addr-len);
1277 if (!vma || addr <= vma->vm_start)
1278 /* remember the address as a hint for next time */
1279 return (mm->free_area_cache = addr-len);
1282 if (mm->mmap_base < len)
1283 goto bottomup;
1285 addr = mm->mmap_base-len;
1287 do {
1288 /*
1289 * Lookup failure means no vma is above this address,
1290 * else if new region fits below vma->vm_start,
1291 * return with success:
1292 */
1293 vma = find_vma(mm, addr);
1294 if (!vma || addr+len <= vma->vm_start)
1295 /* remember the address as a hint for next time */
1296 return (mm->free_area_cache = addr);
1298 /* remember the largest hole we saw so far */
1299 if (addr + mm->cached_hole_size < vma->vm_start)
1300 mm->cached_hole_size = vma->vm_start - addr;
1302 /* try just below the current vma->vm_start */
1303 addr = vma->vm_start-len;
1304 } while (len < vma->vm_start);
1306 bottomup:
1307 /*
1308 * A failed mmap() very likely causes application failure,
1309 * so fall back to the bottom-up function here. This scenario
1310 * can happen with large stack limits and large mmap()
1311 * allocations.
1312 */
1313 mm->cached_hole_size = ~0UL;
1314 mm->free_area_cache = TASK_UNMAPPED_BASE;
1315 addr = arch_get_unmapped_area(filp, addr0, len, pgoff, flags);
1316 /*
1317 * Restore the topdown base:
1318 */
1319 mm->free_area_cache = mm->mmap_base;
1320 mm->cached_hole_size = ~0UL;
1322 return addr;
1324 #endif
1326 void arch_unmap_area_topdown(struct mm_struct *mm, unsigned long addr)
1328 /*
1329 * Is this a new hole at the highest possible address?
1330 */
1331 if (addr > mm->free_area_cache)
1332 mm->free_area_cache = addr;
1334 /* dont allow allocations above current base */
1335 if (mm->free_area_cache > mm->mmap_base)
1336 mm->free_area_cache = mm->mmap_base;
1339 unsigned long
1340 get_unmapped_area(struct file *file, unsigned long addr, unsigned long len,
1341 unsigned long pgoff, unsigned long flags)
1343 unsigned long ret;
1345 if (!(flags & MAP_FIXED)) {
1346 unsigned long (*get_area)(struct file *, unsigned long, unsigned long, unsigned long, unsigned long);
1348 get_area = current->mm->get_unmapped_area;
1349 if (file && file->f_op && file->f_op->get_unmapped_area)
1350 get_area = file->f_op->get_unmapped_area;
1351 addr = get_area(file, addr, len, pgoff, flags);
1352 if (IS_ERR_VALUE(addr))
1353 return addr;
1356 if (addr > TASK_SIZE - len)
1357 return -ENOMEM;
1358 if (addr & ~PAGE_MASK)
1359 return -EINVAL;
1360 if (file && is_file_hugepages(file)) {
1361 /*
1362 * Check if the given range is hugepage aligned, and
1363 * can be made suitable for hugepages.
1364 */
1365 ret = prepare_hugepage_range(addr, len);
1366 } else {
1367 /*
1368 * Ensure that a normal request is not falling in a
1369 * reserved hugepage range. For some archs like IA-64,
1370 * there is a separate region for hugepages.
1371 */
1372 ret = is_hugepage_only_range(current->mm, addr, len);
1374 if (ret)
1375 return -EINVAL;
1376 return addr;
1379 EXPORT_SYMBOL(get_unmapped_area);
1381 /* Look up the first VMA which satisfies addr < vm_end, NULL if none. */
1382 struct vm_area_struct * find_vma(struct mm_struct * mm, unsigned long addr)
1384 struct vm_area_struct *vma = NULL;
1386 if (mm) {
1387 /* Check the cache first. */
1388 /* (Cache hit rate is typically around 35%.) */
1389 vma = mm->mmap_cache;
1390 if (!(vma && vma->vm_end > addr && vma->vm_start <= addr)) {
1391 struct rb_node * rb_node;
1393 rb_node = mm->mm_rb.rb_node;
1394 vma = NULL;
1396 while (rb_node) {
1397 struct vm_area_struct * vma_tmp;
1399 vma_tmp = rb_entry(rb_node,
1400 struct vm_area_struct, vm_rb);
1402 if (vma_tmp->vm_end > addr) {
1403 vma = vma_tmp;
1404 if (vma_tmp->vm_start <= addr)
1405 break;
1406 rb_node = rb_node->rb_left;
1407 } else
1408 rb_node = rb_node->rb_right;
1410 if (vma)
1411 mm->mmap_cache = vma;
1414 return vma;
1417 EXPORT_SYMBOL(find_vma);
1419 /* Same as find_vma, but also return a pointer to the previous VMA in *pprev. */
1420 struct vm_area_struct *
1421 find_vma_prev(struct mm_struct *mm, unsigned long addr,
1422 struct vm_area_struct **pprev)
1424 struct vm_area_struct *vma = NULL, *prev = NULL;
1425 struct rb_node * rb_node;
1426 if (!mm)
1427 goto out;
1429 /* Guard against addr being lower than the first VMA */
1430 vma = mm->mmap;
1432 /* Go through the RB tree quickly. */
1433 rb_node = mm->mm_rb.rb_node;
1435 while (rb_node) {
1436 struct vm_area_struct *vma_tmp;
1437 vma_tmp = rb_entry(rb_node, struct vm_area_struct, vm_rb);
1439 if (addr < vma_tmp->vm_end) {
1440 rb_node = rb_node->rb_left;
1441 } else {
1442 prev = vma_tmp;
1443 if (!prev->vm_next || (addr < prev->vm_next->vm_end))
1444 break;
1445 rb_node = rb_node->rb_right;
1449 out:
1450 *pprev = prev;
1451 return prev ? prev->vm_next : vma;
1454 /*
1455 * Verify that the stack growth is acceptable and
1456 * update accounting. This is shared with both the
1457 * grow-up and grow-down cases.
1458 */
1459 static int acct_stack_growth(struct vm_area_struct * vma, unsigned long size, unsigned long grow)
1461 struct mm_struct *mm = vma->vm_mm;
1462 struct rlimit *rlim = current->signal->rlim;
1464 /* address space limit tests */
1465 if (!may_expand_vm(mm, grow))
1466 return -ENOMEM;
1468 /* Stack limit test */
1469 if (size > rlim[RLIMIT_STACK].rlim_cur)
1470 return -ENOMEM;
1472 /* mlock limit tests */
1473 if (vma->vm_flags & VM_LOCKED) {
1474 unsigned long locked;
1475 unsigned long limit;
1476 locked = mm->locked_vm + grow;
1477 limit = rlim[RLIMIT_MEMLOCK].rlim_cur >> PAGE_SHIFT;
1478 if (locked > limit && !capable(CAP_IPC_LOCK))
1479 return -ENOMEM;
1482 /*
1483 * Overcommit.. This must be the final test, as it will
1484 * update security statistics.
1485 */
1486 if (security_vm_enough_memory(grow))
1487 return -ENOMEM;
1489 /* Ok, everything looks good - let it rip */
1490 mm->total_vm += grow;
1491 if (vma->vm_flags & VM_LOCKED)
1492 mm->locked_vm += grow;
1493 vm_stat_account(mm, vma->vm_flags, vma->vm_file, grow);
1494 return 0;
1497 #if defined(CONFIG_STACK_GROWSUP) || defined(CONFIG_IA64)
1498 /*
1499 * PA-RISC uses this for its stack; IA64 for its Register Backing Store.
1500 * vma is the last one with address > vma->vm_end. Have to extend vma.
1501 */
1502 #ifndef CONFIG_IA64
1503 static inline
1504 #endif
1505 int expand_upwards(struct vm_area_struct *vma, unsigned long address)
1507 int error;
1509 if (!(vma->vm_flags & VM_GROWSUP))
1510 return -EFAULT;
1512 /*
1513 * We must make sure the anon_vma is allocated
1514 * so that the anon_vma locking is not a noop.
1515 */
1516 if (unlikely(anon_vma_prepare(vma)))
1517 return -ENOMEM;
1518 anon_vma_lock(vma);
1520 /*
1521 * vma->vm_start/vm_end cannot change under us because the caller
1522 * is required to hold the mmap_sem in read mode. We need the
1523 * anon_vma lock to serialize against concurrent expand_stacks.
1524 */
1525 address += 4 + PAGE_SIZE - 1;
1526 address &= PAGE_MASK;
1527 error = 0;
1529 /* Somebody else might have raced and expanded it already */
1530 if (address > vma->vm_end) {
1531 unsigned long size, grow;
1533 size = address - vma->vm_start;
1534 grow = (address - vma->vm_end) >> PAGE_SHIFT;
1536 error = acct_stack_growth(vma, size, grow);
1537 if (!error)
1538 vma->vm_end = address;
1540 anon_vma_unlock(vma);
1541 return error;
1543 #endif /* CONFIG_STACK_GROWSUP || CONFIG_IA64 */
1545 #ifdef CONFIG_STACK_GROWSUP
1546 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1548 return expand_upwards(vma, address);
1551 struct vm_area_struct *
1552 find_extend_vma(struct mm_struct *mm, unsigned long addr)
1554 struct vm_area_struct *vma, *prev;
1556 addr &= PAGE_MASK;
1557 vma = find_vma_prev(mm, addr, &prev);
1558 if (vma && (vma->vm_start <= addr))
1559 return vma;
1560 if (!prev || expand_stack(prev, addr))
1561 return NULL;
1562 if (prev->vm_flags & VM_LOCKED) {
1563 make_pages_present(addr, prev->vm_end);
1565 return prev;
1567 #else
1568 /*
1569 * vma is the first one with address < vma->vm_start. Have to extend vma.
1570 */
1571 int expand_stack(struct vm_area_struct *vma, unsigned long address)
1573 int error;
1575 /*
1576 * We must make sure the anon_vma is allocated
1577 * so that the anon_vma locking is not a noop.
1578 */
1579 if (unlikely(anon_vma_prepare(vma)))
1580 return -ENOMEM;
1581 anon_vma_lock(vma);
1583 /*
1584 * vma->vm_start/vm_end cannot change under us because the caller
1585 * is required to hold the mmap_sem in read mode. We need the
1586 * anon_vma lock to serialize against concurrent expand_stacks.
1587 */
1588 address &= PAGE_MASK;
1589 error = 0;
1591 /* Somebody else might have raced and expanded it already */
1592 if (address < vma->vm_start) {
1593 unsigned long size, grow;
1595 size = vma->vm_end - address;
1596 grow = (vma->vm_start - address) >> PAGE_SHIFT;
1598 error = acct_stack_growth(vma, size, grow);
1599 if (!error) {
1600 vma->vm_start = address;
1601 vma->vm_pgoff -= grow;
1604 anon_vma_unlock(vma);
1605 return error;
1608 struct vm_area_struct *
1609 find_extend_vma(struct mm_struct * mm, unsigned long addr)
1611 struct vm_area_struct * vma;
1612 unsigned long start;
1614 addr &= PAGE_MASK;
1615 vma = find_vma(mm,addr);
1616 if (!vma)
1617 return NULL;
1618 if (vma->vm_start <= addr)
1619 return vma;
1620 if (!(vma->vm_flags & VM_GROWSDOWN))
1621 return NULL;
1622 start = vma->vm_start;
1623 if (expand_stack(vma, addr))
1624 return NULL;
1625 if (vma->vm_flags & VM_LOCKED) {
1626 make_pages_present(addr, start);
1628 return vma;
1630 #endif
1632 /*
1633 * Ok - we have the memory areas we should free on the vma list,
1634 * so release them, and do the vma updates.
1636 * Called with the mm semaphore held.
1637 */
1638 static void remove_vma_list(struct mm_struct *mm, struct vm_area_struct *vma)
1640 /* Update high watermark before we lower total_vm */
1641 update_hiwater_vm(mm);
1642 do {
1643 long nrpages = vma_pages(vma);
1645 mm->total_vm -= nrpages;
1646 if (vma->vm_flags & VM_LOCKED)
1647 mm->locked_vm -= nrpages;
1648 vm_stat_account(mm, vma->vm_flags, vma->vm_file, -nrpages);
1649 vma = remove_vma(vma);
1650 } while (vma);
1651 validate_mm(mm);
1654 /*
1655 * Get rid of page table information in the indicated region.
1657 * Called with the mm semaphore held.
1658 */
1659 static void unmap_region(struct mm_struct *mm,
1660 struct vm_area_struct *vma, struct vm_area_struct *prev,
1661 unsigned long start, unsigned long end)
1663 struct vm_area_struct *next = prev? prev->vm_next: mm->mmap;
1664 struct mmu_gather *tlb;
1665 unsigned long nr_accounted = 0;
1667 lru_add_drain();
1668 tlb = tlb_gather_mmu(mm, 0);
1669 update_hiwater_rss(mm);
1670 unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
1671 vm_unacct_memory(nr_accounted);
1672 free_pgtables(&tlb, vma, prev? prev->vm_end: FIRST_USER_ADDRESS,
1673 next? next->vm_start: 0);
1674 tlb_finish_mmu(tlb, start, end);
1677 /*
1678 * Create a list of vma's touched by the unmap, removing them from the mm's
1679 * vma list as we go..
1680 */
1681 static void
1682 detach_vmas_to_be_unmapped(struct mm_struct *mm, struct vm_area_struct *vma,
1683 struct vm_area_struct *prev, unsigned long end)
1685 struct vm_area_struct **insertion_point;
1686 struct vm_area_struct *tail_vma = NULL;
1687 unsigned long addr;
1689 insertion_point = (prev ? &prev->vm_next : &mm->mmap);
1690 do {
1691 rb_erase(&vma->vm_rb, &mm->mm_rb);
1692 mm->map_count--;
1693 tail_vma = vma;
1694 vma = vma->vm_next;
1695 } while (vma && vma->vm_start < end);
1696 *insertion_point = vma;
1697 tail_vma->vm_next = NULL;
1698 if (mm->unmap_area == arch_unmap_area)
1699 addr = prev ? prev->vm_end : mm->mmap_base;
1700 else
1701 addr = vma ? vma->vm_start : mm->mmap_base;
1702 mm->unmap_area(mm, addr);
1703 mm->mmap_cache = NULL; /* Kill the cache. */
1706 /*
1707 * Split a vma into two pieces at address 'addr', a new vma is allocated
1708 * either for the first part or the the tail.
1709 */
1710 int split_vma(struct mm_struct * mm, struct vm_area_struct * vma,
1711 unsigned long addr, int new_below)
1713 struct mempolicy *pol;
1714 struct vm_area_struct *new;
1716 if (is_vm_hugetlb_page(vma) && (addr & ~HPAGE_MASK))
1717 return -EINVAL;
1719 if (mm->map_count >= sysctl_max_map_count)
1720 return -ENOMEM;
1722 new = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1723 if (!new)
1724 return -ENOMEM;
1726 /* most fields are the same, copy all, and then fixup */
1727 *new = *vma;
1729 if (new_below)
1730 new->vm_end = addr;
1731 else {
1732 new->vm_start = addr;
1733 new->vm_pgoff += ((addr - vma->vm_start) >> PAGE_SHIFT);
1736 pol = mpol_copy(vma_policy(vma));
1737 if (IS_ERR(pol)) {
1738 kmem_cache_free(vm_area_cachep, new);
1739 return PTR_ERR(pol);
1741 vma_set_policy(new, pol);
1743 if (new->vm_file)
1744 get_file(new->vm_file);
1746 if (new->vm_ops && new->vm_ops->open)
1747 new->vm_ops->open(new);
1749 if (new_below)
1750 vma_adjust(vma, addr, vma->vm_end, vma->vm_pgoff +
1751 ((addr - new->vm_start) >> PAGE_SHIFT), new);
1752 else
1753 vma_adjust(vma, vma->vm_start, addr, vma->vm_pgoff, new);
1755 return 0;
1758 /* Munmap is split into 2 main parts -- this part which finds
1759 * what needs doing, and the areas themselves, which do the
1760 * work. This now handles partial unmappings.
1761 * Jeremy Fitzhardinge <jeremy@goop.org>
1762 */
1763 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1765 unsigned long end;
1766 struct vm_area_struct *vma, *prev, *last;
1768 if ((start & ~PAGE_MASK) || start > TASK_SIZE || len > TASK_SIZE-start)
1769 return -EINVAL;
1771 if ((len = PAGE_ALIGN(len)) == 0)
1772 return -EINVAL;
1774 /* Find the first overlapping VMA */
1775 vma = find_vma_prev(mm, start, &prev);
1776 if (!vma)
1777 return 0;
1778 /* we have start < vma->vm_end */
1780 /* if it doesn't overlap, we have nothing.. */
1781 end = start + len;
1782 if (vma->vm_start >= end)
1783 return 0;
1785 /*
1786 * If we need to split any vma, do it now to save pain later.
1788 * Note: mremap's move_vma VM_ACCOUNT handling assumes a partially
1789 * unmapped vm_area_struct will remain in use: so lower split_vma
1790 * places tmp vma above, and higher split_vma places tmp vma below.
1791 */
1792 if (start > vma->vm_start) {
1793 int error = split_vma(mm, vma, start, 0);
1794 if (error)
1795 return error;
1796 prev = vma;
1799 /* Does it split the last one? */
1800 last = find_vma(mm, end);
1801 if (last && end > last->vm_start) {
1802 int error = split_vma(mm, last, end, 1);
1803 if (error)
1804 return error;
1806 vma = prev? prev->vm_next: mm->mmap;
1808 /*
1809 * Remove the vma's, and unmap the actual pages
1810 */
1811 detach_vmas_to_be_unmapped(mm, vma, prev, end);
1812 unmap_region(mm, vma, prev, start, end);
1814 /* Fix up all other VM information */
1815 remove_vma_list(mm, vma);
1817 return 0;
1820 EXPORT_SYMBOL(do_munmap);
1822 asmlinkage long sys_munmap(unsigned long addr, size_t len)
1824 int ret;
1825 struct mm_struct *mm = current->mm;
1827 profile_munmap(addr);
1829 down_write(&mm->mmap_sem);
1830 ret = do_munmap(mm, addr, len);
1831 up_write(&mm->mmap_sem);
1832 return ret;
1835 static inline void verify_mm_writelocked(struct mm_struct *mm)
1837 #ifdef CONFIG_DEBUG_VM
1838 if (unlikely(down_read_trylock(&mm->mmap_sem))) {
1839 WARN_ON(1);
1840 up_read(&mm->mmap_sem);
1842 #endif
1845 /*
1846 * this is really a simplified "do_mmap". it only handles
1847 * anonymous maps. eventually we may be able to do some
1848 * brk-specific accounting here.
1849 */
1850 unsigned long do_brk(unsigned long addr, unsigned long len)
1852 struct mm_struct * mm = current->mm;
1853 struct vm_area_struct * vma, * prev;
1854 unsigned long flags;
1855 struct rb_node ** rb_link, * rb_parent;
1856 pgoff_t pgoff = addr >> PAGE_SHIFT;
1857 int error;
1859 len = PAGE_ALIGN(len);
1860 if (!len)
1861 return addr;
1863 if ((addr + len) > TASK_SIZE || (addr + len) < addr)
1864 return -EINVAL;
1866 flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
1868 error = arch_mmap_check(addr, len, flags);
1869 if (error)
1870 return error;
1872 /*
1873 * mlock MCL_FUTURE?
1874 */
1875 if (mm->def_flags & VM_LOCKED) {
1876 unsigned long locked, lock_limit;
1877 locked = len >> PAGE_SHIFT;
1878 locked += mm->locked_vm;
1879 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
1880 lock_limit >>= PAGE_SHIFT;
1881 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
1882 return -EAGAIN;
1885 /*
1886 * mm->mmap_sem is required to protect against another thread
1887 * changing the mappings in case we sleep.
1888 */
1889 verify_mm_writelocked(mm);
1891 /*
1892 * Clear old maps. this also does some error checking for us
1893 */
1894 munmap_back:
1895 vma = find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
1896 if (vma && vma->vm_start < addr + len) {
1897 if (do_munmap(mm, addr, len))
1898 return -ENOMEM;
1899 goto munmap_back;
1902 /* Check against address space limits *after* clearing old maps... */
1903 if (!may_expand_vm(mm, len >> PAGE_SHIFT))
1904 return -ENOMEM;
1906 if (mm->map_count > sysctl_max_map_count)
1907 return -ENOMEM;
1909 if (security_vm_enough_memory(len >> PAGE_SHIFT))
1910 return -ENOMEM;
1912 /* Can we just expand an old private anonymous mapping? */
1913 if (vma_merge(mm, prev, addr, addr + len, flags,
1914 NULL, NULL, pgoff, NULL))
1915 goto out;
1917 /*
1918 * create a vma struct for an anonymous mapping
1919 */
1920 vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
1921 if (!vma) {
1922 vm_unacct_memory(len >> PAGE_SHIFT);
1923 return -ENOMEM;
1925 memset(vma, 0, sizeof(*vma));
1927 vma->vm_mm = mm;
1928 vma->vm_start = addr;
1929 vma->vm_end = addr + len;
1930 vma->vm_pgoff = pgoff;
1931 vma->vm_flags = flags;
1932 vma->vm_page_prot = protection_map[flags & 0x0f];
1933 vma_link(mm, vma, prev, rb_link, rb_parent);
1934 out:
1935 mm->total_vm += len >> PAGE_SHIFT;
1936 if (flags & VM_LOCKED) {
1937 mm->locked_vm += len >> PAGE_SHIFT;
1938 make_pages_present(addr, addr + len);
1940 return addr;
1943 EXPORT_SYMBOL(do_brk);
1945 /* Release all mmaps. */
1946 void exit_mmap(struct mm_struct *mm)
1948 struct mmu_gather *tlb;
1949 struct vm_area_struct *vma = mm->mmap;
1950 unsigned long nr_accounted = 0;
1951 unsigned long end;
1953 #ifdef arch_exit_mmap
1954 arch_exit_mmap(mm);
1955 #endif
1957 lru_add_drain();
1958 flush_cache_mm(mm);
1959 tlb = tlb_gather_mmu(mm, 1);
1960 /* Don't update_hiwater_rss(mm) here, do_exit already did */
1961 /* Use -1 here to ensure all VMAs in the mm are unmapped */
1962 end = unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
1963 vm_unacct_memory(nr_accounted);
1964 free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
1965 tlb_finish_mmu(tlb, 0, end);
1967 /*
1968 * Walk the list again, actually closing and freeing it,
1969 * with preemption enabled, without holding any MM locks.
1970 */
1971 while (vma)
1972 vma = remove_vma(vma);
1974 BUG_ON(mm->nr_ptes > (FIRST_USER_ADDRESS+PMD_SIZE-1)>>PMD_SHIFT);
1977 /* Insert vm structure into process list sorted by address
1978 * and into the inode's i_mmap tree. If vm_file is non-NULL
1979 * then i_mmap_lock is taken here.
1980 */
1981 int insert_vm_struct(struct mm_struct * mm, struct vm_area_struct * vma)
1983 struct vm_area_struct * __vma, * prev;
1984 struct rb_node ** rb_link, * rb_parent;
1986 /*
1987 * The vm_pgoff of a purely anonymous vma should be irrelevant
1988 * until its first write fault, when page's anon_vma and index
1989 * are set. But now set the vm_pgoff it will almost certainly
1990 * end up with (unless mremap moves it elsewhere before that
1991 * first wfault), so /proc/pid/maps tells a consistent story.
1993 * By setting it to reflect the virtual start address of the
1994 * vma, merges and splits can happen in a seamless way, just
1995 * using the existing file pgoff checks and manipulations.
1996 * Similarly in do_mmap_pgoff and in do_brk.
1997 */
1998 if (!vma->vm_file) {
1999 BUG_ON(vma->anon_vma);
2000 vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2002 __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
2003 if (__vma && __vma->vm_start < vma->vm_end)
2004 return -ENOMEM;
2005 if ((vma->vm_flags & VM_ACCOUNT) &&
2006 security_vm_enough_memory(vma_pages(vma)))
2007 return -ENOMEM;
2008 vma_link(mm, vma, prev, rb_link, rb_parent);
2009 return 0;
2012 /*
2013 * Copy the vma structure to a new location in the same mm,
2014 * prior to moving page table entries, to effect an mremap move.
2015 */
2016 struct vm_area_struct *copy_vma(struct vm_area_struct **vmap,
2017 unsigned long addr, unsigned long len, pgoff_t pgoff)
2019 struct vm_area_struct *vma = *vmap;
2020 unsigned long vma_start = vma->vm_start;
2021 struct mm_struct *mm = vma->vm_mm;
2022 struct vm_area_struct *new_vma, *prev;
2023 struct rb_node **rb_link, *rb_parent;
2024 struct mempolicy *pol;
2026 /*
2027 * If anonymous vma has not yet been faulted, update new pgoff
2028 * to match new location, to increase its chance of merging.
2029 */
2030 if (!vma->vm_file && !vma->anon_vma)
2031 pgoff = addr >> PAGE_SHIFT;
2033 find_vma_prepare(mm, addr, &prev, &rb_link, &rb_parent);
2034 new_vma = vma_merge(mm, prev, addr, addr + len, vma->vm_flags,
2035 vma->anon_vma, vma->vm_file, pgoff, vma_policy(vma));
2036 if (new_vma) {
2037 /*
2038 * Source vma may have been merged into new_vma
2039 */
2040 if (vma_start >= new_vma->vm_start &&
2041 vma_start < new_vma->vm_end)
2042 *vmap = new_vma;
2043 } else {
2044 new_vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
2045 if (new_vma) {
2046 *new_vma = *vma;
2047 pol = mpol_copy(vma_policy(vma));
2048 if (IS_ERR(pol)) {
2049 kmem_cache_free(vm_area_cachep, new_vma);
2050 return NULL;
2052 vma_set_policy(new_vma, pol);
2053 new_vma->vm_start = addr;
2054 new_vma->vm_end = addr + len;
2055 new_vma->vm_pgoff = pgoff;
2056 if (new_vma->vm_file)
2057 get_file(new_vma->vm_file);
2058 if (new_vma->vm_ops && new_vma->vm_ops->open)
2059 new_vma->vm_ops->open(new_vma);
2060 vma_link(mm, new_vma, prev, rb_link, rb_parent);
2063 return new_vma;
2066 /*
2067 * Return true if the calling process may expand its vm space by the passed
2068 * number of pages
2069 */
2070 int may_expand_vm(struct mm_struct *mm, unsigned long npages)
2072 unsigned long cur = mm->total_vm; /* pages */
2073 unsigned long lim;
2075 lim = current->signal->rlim[RLIMIT_AS].rlim_cur >> PAGE_SHIFT;
2077 if (cur + npages > lim)
2078 return 0;
2079 return 1;