nommu: fix a number of issues with the per-MM VMA patch
[linux-2.6.git] / mm / nommu.c
1 /*
2  *  linux/mm/nommu.c
3  *
4  *  Replacement code for mm functions to support CPU's that don't
5  *  have any form of memory management unit (thus no virtual memory).
6  *
7  *  See Documentation/nommu-mmap.txt
8  *
9  *  Copyright (c) 2004-2008 David Howells <dhowells@redhat.com>
10  *  Copyright (c) 2000-2003 David McCullough <davidm@snapgear.com>
11  *  Copyright (c) 2000-2001 D Jeff Dionne <jeff@uClinux.org>
12  *  Copyright (c) 2002      Greg Ungerer <gerg@snapgear.com>
13  *  Copyright (c) 2007-2009 Paul Mundt <lethal@linux-sh.org>
14  */
15
16 #include <linux/module.h>
17 #include <linux/mm.h>
18 #include <linux/mman.h>
19 #include <linux/swap.h>
20 #include <linux/file.h>
21 #include <linux/highmem.h>
22 #include <linux/pagemap.h>
23 #include <linux/slab.h>
24 #include <linux/vmalloc.h>
25 #include <linux/tracehook.h>
26 #include <linux/blkdev.h>
27 #include <linux/backing-dev.h>
28 #include <linux/mount.h>
29 #include <linux/personality.h>
30 #include <linux/security.h>
31 #include <linux/syscalls.h>
32
33 #include <asm/uaccess.h>
34 #include <asm/tlb.h>
35 #include <asm/tlbflush.h>
36 #include "internal.h"
37
38 static inline __attribute__((format(printf, 1, 2)))
39 void no_printk(const char *fmt, ...)
40 {
41 }
42
43 #if 0
44 #define kenter(FMT, ...) \
45         printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
46 #define kleave(FMT, ...) \
47         printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
48 #define kdebug(FMT, ...) \
49         printk(KERN_DEBUG "xxx" FMT"yyy\n", ##__VA_ARGS__)
50 #else
51 #define kenter(FMT, ...) \
52         no_printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
53 #define kleave(FMT, ...) \
54         no_printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
55 #define kdebug(FMT, ...) \
56         no_printk(KERN_DEBUG FMT"\n", ##__VA_ARGS__)
57 #endif
58
59 #include "internal.h"
60
61 void *high_memory;
62 struct page *mem_map;
63 unsigned long max_mapnr;
64 unsigned long num_physpages;
65 atomic_long_t vm_committed_space = ATOMIC_LONG_INIT(0);
66 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
67 int sysctl_overcommit_ratio = 50; /* default is 50% */
68 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT;
69 int sysctl_nr_trim_pages = 1; /* page trimming behaviour */
70 int heap_stack_gap = 0;
71
72 atomic_long_t mmap_pages_allocated;
73
74 EXPORT_SYMBOL(mem_map);
75 EXPORT_SYMBOL(num_physpages);
76
77 /* list of mapped, potentially shareable regions */
78 static struct kmem_cache *vm_region_jar;
79 struct rb_root nommu_region_tree = RB_ROOT;
80 DECLARE_RWSEM(nommu_region_sem);
81
82 struct vm_operations_struct generic_file_vm_ops = {
83 };
84
85 /*
86  * Handle all mappings that got truncated by a "truncate()"
87  * system call.
88  *
89  * NOTE! We have to be ready to update the memory sharing
90  * between the file and the memory map for a potential last
91  * incomplete page.  Ugly, but necessary.
92  */
93 int vmtruncate(struct inode *inode, loff_t offset)
94 {
95         struct address_space *mapping = inode->i_mapping;
96         unsigned long limit;
97
98         if (inode->i_size < offset)
99                 goto do_expand;
100         i_size_write(inode, offset);
101
102         truncate_inode_pages(mapping, offset);
103         goto out_truncate;
104
105 do_expand:
106         limit = current->signal->rlim[RLIMIT_FSIZE].rlim_cur;
107         if (limit != RLIM_INFINITY && offset > limit)
108                 goto out_sig;
109         if (offset > inode->i_sb->s_maxbytes)
110                 goto out;
111         i_size_write(inode, offset);
112
113 out_truncate:
114         if (inode->i_op->truncate)
115                 inode->i_op->truncate(inode);
116         return 0;
117 out_sig:
118         send_sig(SIGXFSZ, current, 0);
119 out:
120         return -EFBIG;
121 }
122
123 EXPORT_SYMBOL(vmtruncate);
124
125 /*
126  * Return the total memory allocated for this pointer, not
127  * just what the caller asked for.
128  *
129  * Doesn't have to be accurate, i.e. may have races.
130  */
131 unsigned int kobjsize(const void *objp)
132 {
133         struct page *page;
134
135         /*
136          * If the object we have should not have ksize performed on it,
137          * return size of 0
138          */
139         if (!objp || !virt_addr_valid(objp))
140                 return 0;
141
142         page = virt_to_head_page(objp);
143
144         /*
145          * If the allocator sets PageSlab, we know the pointer came from
146          * kmalloc().
147          */
148         if (PageSlab(page))
149                 return ksize(objp);
150
151         /*
152          * If it's not a compound page, see if we have a matching VMA
153          * region. This test is intentionally done in reverse order,
154          * so if there's no VMA, we still fall through and hand back
155          * PAGE_SIZE for 0-order pages.
156          */
157         if (!PageCompound(page)) {
158                 struct vm_area_struct *vma;
159
160                 vma = find_vma(current->mm, (unsigned long)objp);
161                 if (vma)
162                         return vma->vm_end - vma->vm_start;
163         }
164
165         /*
166          * The ksize() function is only guaranteed to work for pointers
167          * returned by kmalloc(). So handle arbitrary pointers here.
168          */
169         return PAGE_SIZE << compound_order(page);
170 }
171
172 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
173                      unsigned long start, int len, int flags,
174                 struct page **pages, struct vm_area_struct **vmas)
175 {
176         struct vm_area_struct *vma;
177         unsigned long vm_flags;
178         int i;
179         int write = !!(flags & GUP_FLAGS_WRITE);
180         int force = !!(flags & GUP_FLAGS_FORCE);
181         int ignore = !!(flags & GUP_FLAGS_IGNORE_VMA_PERMISSIONS);
182
183         /* calculate required read or write permissions.
184          * - if 'force' is set, we only require the "MAY" flags.
185          */
186         vm_flags  = write ? (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
187         vm_flags &= force ? (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
188
189         for (i = 0; i < len; i++) {
190                 vma = find_vma(mm, start);
191                 if (!vma)
192                         goto finish_or_fault;
193
194                 /* protect what we can, including chardevs */
195                 if (vma->vm_flags & (VM_IO | VM_PFNMAP) ||
196                     (!ignore && !(vm_flags & vma->vm_flags)))
197                         goto finish_or_fault;
198
199                 if (pages) {
200                         pages[i] = virt_to_page(start);
201                         if (pages[i])
202                                 page_cache_get(pages[i]);
203                 }
204                 if (vmas)
205                         vmas[i] = vma;
206                 start += PAGE_SIZE;
207         }
208
209         return i;
210
211 finish_or_fault:
212         return i ? : -EFAULT;
213 }
214
215
216 /*
217  * get a list of pages in an address range belonging to the specified process
218  * and indicate the VMA that covers each page
219  * - this is potentially dodgy as we may end incrementing the page count of a
220  *   slab page or a secondary page from a compound page
221  * - don't permit access to VMAs that don't support it, such as I/O mappings
222  */
223 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
224         unsigned long start, int len, int write, int force,
225         struct page **pages, struct vm_area_struct **vmas)
226 {
227         int flags = 0;
228
229         if (write)
230                 flags |= GUP_FLAGS_WRITE;
231         if (force)
232                 flags |= GUP_FLAGS_FORCE;
233
234         return __get_user_pages(tsk, mm,
235                                 start, len, flags,
236                                 pages, vmas);
237 }
238 EXPORT_SYMBOL(get_user_pages);
239
240 DEFINE_RWLOCK(vmlist_lock);
241 struct vm_struct *vmlist;
242
243 void vfree(const void *addr)
244 {
245         kfree(addr);
246 }
247 EXPORT_SYMBOL(vfree);
248
249 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
250 {
251         /*
252          *  You can't specify __GFP_HIGHMEM with kmalloc() since kmalloc()
253          * returns only a logical address.
254          */
255         return kmalloc(size, (gfp_mask | __GFP_COMP) & ~__GFP_HIGHMEM);
256 }
257 EXPORT_SYMBOL(__vmalloc);
258
259 void *vmalloc_user(unsigned long size)
260 {
261         void *ret;
262
263         ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
264                         PAGE_KERNEL);
265         if (ret) {
266                 struct vm_area_struct *vma;
267
268                 down_write(&current->mm->mmap_sem);
269                 vma = find_vma(current->mm, (unsigned long)ret);
270                 if (vma)
271                         vma->vm_flags |= VM_USERMAP;
272                 up_write(&current->mm->mmap_sem);
273         }
274
275         return ret;
276 }
277 EXPORT_SYMBOL(vmalloc_user);
278
279 struct page *vmalloc_to_page(const void *addr)
280 {
281         return virt_to_page(addr);
282 }
283 EXPORT_SYMBOL(vmalloc_to_page);
284
285 unsigned long vmalloc_to_pfn(const void *addr)
286 {
287         return page_to_pfn(virt_to_page(addr));
288 }
289 EXPORT_SYMBOL(vmalloc_to_pfn);
290
291 long vread(char *buf, char *addr, unsigned long count)
292 {
293         memcpy(buf, addr, count);
294         return count;
295 }
296
297 long vwrite(char *buf, char *addr, unsigned long count)
298 {
299         /* Don't allow overflow */
300         if ((unsigned long) addr + count < count)
301                 count = -(unsigned long) addr;
302
303         memcpy(addr, buf, count);
304         return(count);
305 }
306
307 /*
308  *      vmalloc  -  allocate virtually continguos memory
309  *
310  *      @size:          allocation size
311  *
312  *      Allocate enough pages to cover @size from the page level
313  *      allocator and map them into continguos kernel virtual space.
314  *
315  *      For tight control over page level allocator and protection flags
316  *      use __vmalloc() instead.
317  */
318 void *vmalloc(unsigned long size)
319 {
320        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
321 }
322 EXPORT_SYMBOL(vmalloc);
323
324 void *vmalloc_node(unsigned long size, int node)
325 {
326         return vmalloc(size);
327 }
328 EXPORT_SYMBOL(vmalloc_node);
329
330 #ifndef PAGE_KERNEL_EXEC
331 # define PAGE_KERNEL_EXEC PAGE_KERNEL
332 #endif
333
334 /**
335  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
336  *      @size:          allocation size
337  *
338  *      Kernel-internal function to allocate enough pages to cover @size
339  *      the page level allocator and map them into contiguous and
340  *      executable kernel virtual space.
341  *
342  *      For tight control over page level allocator and protection flags
343  *      use __vmalloc() instead.
344  */
345
346 void *vmalloc_exec(unsigned long size)
347 {
348         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
349 }
350
351 /**
352  * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
353  *      @size:          allocation size
354  *
355  *      Allocate enough 32bit PA addressable pages to cover @size from the
356  *      page level allocator and map them into continguos kernel virtual space.
357  */
358 void *vmalloc_32(unsigned long size)
359 {
360         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
361 }
362 EXPORT_SYMBOL(vmalloc_32);
363
364 /**
365  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
366  *      @size:          allocation size
367  *
368  * The resulting memory area is 32bit addressable and zeroed so it can be
369  * mapped to userspace without leaking data.
370  *
371  * VM_USERMAP is set on the corresponding VMA so that subsequent calls to
372  * remap_vmalloc_range() are permissible.
373  */
374 void *vmalloc_32_user(unsigned long size)
375 {
376         /*
377          * We'll have to sort out the ZONE_DMA bits for 64-bit,
378          * but for now this can simply use vmalloc_user() directly.
379          */
380         return vmalloc_user(size);
381 }
382 EXPORT_SYMBOL(vmalloc_32_user);
383
384 void *vmap(struct page **pages, unsigned int count, unsigned long flags, pgprot_t prot)
385 {
386         BUG();
387         return NULL;
388 }
389 EXPORT_SYMBOL(vmap);
390
391 void vunmap(const void *addr)
392 {
393         BUG();
394 }
395 EXPORT_SYMBOL(vunmap);
396
397 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
398 {
399         BUG();
400         return NULL;
401 }
402 EXPORT_SYMBOL(vm_map_ram);
403
404 void vm_unmap_ram(const void *mem, unsigned int count)
405 {
406         BUG();
407 }
408 EXPORT_SYMBOL(vm_unmap_ram);
409
410 void vm_unmap_aliases(void)
411 {
412 }
413 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
414
415 /*
416  * Implement a stub for vmalloc_sync_all() if the architecture chose not to
417  * have one.
418  */
419 void  __attribute__((weak)) vmalloc_sync_all(void)
420 {
421 }
422
423 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
424                    struct page *page)
425 {
426         return -EINVAL;
427 }
428 EXPORT_SYMBOL(vm_insert_page);
429
430 /*
431  *  sys_brk() for the most part doesn't need the global kernel
432  *  lock, except when an application is doing something nasty
433  *  like trying to un-brk an area that has already been mapped
434  *  to a regular file.  in this case, the unmapping will need
435  *  to invoke file system routines that need the global lock.
436  */
437 SYSCALL_DEFINE1(brk, unsigned long, brk)
438 {
439         struct mm_struct *mm = current->mm;
440
441         if (brk < mm->start_brk || brk > mm->context.end_brk)
442                 return mm->brk;
443
444         if (mm->brk == brk)
445                 return mm->brk;
446
447         /*
448          * Always allow shrinking brk
449          */
450         if (brk <= mm->brk) {
451                 mm->brk = brk;
452                 return brk;
453         }
454
455         /*
456          * Ok, looks good - let it rip.
457          */
458         return mm->brk = brk;
459 }
460
461 /*
462  * initialise the VMA and region record slabs
463  */
464 void __init mmap_init(void)
465 {
466         vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);
467 }
468
469 /*
470  * validate the region tree
471  * - the caller must hold the region lock
472  */
473 #ifdef CONFIG_DEBUG_NOMMU_REGIONS
474 static noinline void validate_nommu_regions(void)
475 {
476         struct vm_region *region, *last;
477         struct rb_node *p, *lastp;
478
479         lastp = rb_first(&nommu_region_tree);
480         if (!lastp)
481                 return;
482
483         last = rb_entry(lastp, struct vm_region, vm_rb);
484         BUG_ON(unlikely(last->vm_end <= last->vm_start));
485         BUG_ON(unlikely(last->vm_top < last->vm_end));
486
487         while ((p = rb_next(lastp))) {
488                 region = rb_entry(p, struct vm_region, vm_rb);
489                 last = rb_entry(lastp, struct vm_region, vm_rb);
490
491                 BUG_ON(unlikely(region->vm_end <= region->vm_start));
492                 BUG_ON(unlikely(region->vm_top < region->vm_end));
493                 BUG_ON(unlikely(region->vm_start < last->vm_top));
494
495                 lastp = p;
496         }
497 }
498 #else
499 static void validate_nommu_regions(void)
500 {
501 }
502 #endif
503
504 /*
505  * add a region into the global tree
506  */
507 static void add_nommu_region(struct vm_region *region)
508 {
509         struct vm_region *pregion;
510         struct rb_node **p, *parent;
511
512         validate_nommu_regions();
513
514         BUG_ON(region->vm_start & ~PAGE_MASK);
515
516         parent = NULL;
517         p = &nommu_region_tree.rb_node;
518         while (*p) {
519                 parent = *p;
520                 pregion = rb_entry(parent, struct vm_region, vm_rb);
521                 if (region->vm_start < pregion->vm_start)
522                         p = &(*p)->rb_left;
523                 else if (region->vm_start > pregion->vm_start)
524                         p = &(*p)->rb_right;
525                 else if (pregion == region)
526                         return;
527                 else
528                         BUG();
529         }
530
531         rb_link_node(&region->vm_rb, parent, p);
532         rb_insert_color(&region->vm_rb, &nommu_region_tree);
533
534         validate_nommu_regions();
535 }
536
537 /*
538  * delete a region from the global tree
539  */
540 static void delete_nommu_region(struct vm_region *region)
541 {
542         BUG_ON(!nommu_region_tree.rb_node);
543
544         validate_nommu_regions();
545         rb_erase(&region->vm_rb, &nommu_region_tree);
546         validate_nommu_regions();
547 }
548
549 /*
550  * free a contiguous series of pages
551  */
552 static void free_page_series(unsigned long from, unsigned long to)
553 {
554         for (; from < to; from += PAGE_SIZE) {
555                 struct page *page = virt_to_page(from);
556
557                 kdebug("- free %lx", from);
558                 atomic_long_dec(&mmap_pages_allocated);
559                 if (page_count(page) != 1)
560                         kdebug("free page %p: refcount not one: %d",
561                                page, page_count(page));
562                 put_page(page);
563         }
564 }
565
566 /*
567  * release a reference to a region
568  * - the caller must hold the region semaphore for writing, which this releases
569  * - the region may not have been added to the tree yet, in which case vm_top
570  *   will equal vm_start
571  */
572 static void __put_nommu_region(struct vm_region *region)
573         __releases(nommu_region_sem)
574 {
575         kenter("%p{%d}", region, atomic_read(&region->vm_usage));
576
577         BUG_ON(!nommu_region_tree.rb_node);
578
579         if (atomic_dec_and_test(&region->vm_usage)) {
580                 if (region->vm_top > region->vm_start)
581                         delete_nommu_region(region);
582                 up_write(&nommu_region_sem);
583
584                 if (region->vm_file)
585                         fput(region->vm_file);
586
587                 /* IO memory and memory shared directly out of the pagecache
588                  * from ramfs/tmpfs mustn't be released here */
589                 if (region->vm_flags & VM_MAPPED_COPY) {
590                         kdebug("free series");
591                         free_page_series(region->vm_start, region->vm_top);
592                 }
593                 kmem_cache_free(vm_region_jar, region);
594         } else {
595                 up_write(&nommu_region_sem);
596         }
597 }
598
599 /*
600  * release a reference to a region
601  */
602 static void put_nommu_region(struct vm_region *region)
603 {
604         down_write(&nommu_region_sem);
605         __put_nommu_region(region);
606 }
607
608 /*
609  * add a VMA into a process's mm_struct in the appropriate place in the list
610  * and tree and add to the address space's page tree also if not an anonymous
611  * page
612  * - should be called with mm->mmap_sem held writelocked
613  */
614 static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
615 {
616         struct vm_area_struct *pvma, **pp;
617         struct address_space *mapping;
618         struct rb_node **p, *parent;
619
620         kenter(",%p", vma);
621
622         BUG_ON(!vma->vm_region);
623
624         mm->map_count++;
625         vma->vm_mm = mm;
626
627         /* add the VMA to the mapping */
628         if (vma->vm_file) {
629                 mapping = vma->vm_file->f_mapping;
630
631                 flush_dcache_mmap_lock(mapping);
632                 vma_prio_tree_insert(vma, &mapping->i_mmap);
633                 flush_dcache_mmap_unlock(mapping);
634         }
635
636         /* add the VMA to the tree */
637         parent = NULL;
638         p = &mm->mm_rb.rb_node;
639         while (*p) {
640                 parent = *p;
641                 pvma = rb_entry(parent, struct vm_area_struct, vm_rb);
642
643                 /* sort by: start addr, end addr, VMA struct addr in that order
644                  * (the latter is necessary as we may get identical VMAs) */
645                 if (vma->vm_start < pvma->vm_start)
646                         p = &(*p)->rb_left;
647                 else if (vma->vm_start > pvma->vm_start)
648                         p = &(*p)->rb_right;
649                 else if (vma->vm_end < pvma->vm_end)
650                         p = &(*p)->rb_left;
651                 else if (vma->vm_end > pvma->vm_end)
652                         p = &(*p)->rb_right;
653                 else if (vma < pvma)
654                         p = &(*p)->rb_left;
655                 else if (vma > pvma)
656                         p = &(*p)->rb_right;
657                 else
658                         BUG();
659         }
660
661         rb_link_node(&vma->vm_rb, parent, p);
662         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
663
664         /* add VMA to the VMA list also */
665         for (pp = &mm->mmap; (pvma = *pp); pp = &(*pp)->vm_next) {
666                 if (pvma->vm_start > vma->vm_start)
667                         break;
668                 if (pvma->vm_start < vma->vm_start)
669                         continue;
670                 if (pvma->vm_end < vma->vm_end)
671                         break;
672         }
673
674         vma->vm_next = *pp;
675         *pp = vma;
676 }
677
678 /*
679  * delete a VMA from its owning mm_struct and address space
680  */
681 static void delete_vma_from_mm(struct vm_area_struct *vma)
682 {
683         struct vm_area_struct **pp;
684         struct address_space *mapping;
685         struct mm_struct *mm = vma->vm_mm;
686
687         kenter("%p", vma);
688
689         mm->map_count--;
690         if (mm->mmap_cache == vma)
691                 mm->mmap_cache = NULL;
692
693         /* remove the VMA from the mapping */
694         if (vma->vm_file) {
695                 mapping = vma->vm_file->f_mapping;
696
697                 flush_dcache_mmap_lock(mapping);
698                 vma_prio_tree_remove(vma, &mapping->i_mmap);
699                 flush_dcache_mmap_unlock(mapping);
700         }
701
702         /* remove from the MM's tree and list */
703         rb_erase(&vma->vm_rb, &mm->mm_rb);
704         for (pp = &mm->mmap; *pp; pp = &(*pp)->vm_next) {
705                 if (*pp == vma) {
706                         *pp = vma->vm_next;
707                         break;
708                 }
709         }
710
711         vma->vm_mm = NULL;
712 }
713
714 /*
715  * destroy a VMA record
716  */
717 static void delete_vma(struct mm_struct *mm, struct vm_area_struct *vma)
718 {
719         kenter("%p", vma);
720         if (vma->vm_ops && vma->vm_ops->close)
721                 vma->vm_ops->close(vma);
722         if (vma->vm_file) {
723                 fput(vma->vm_file);
724                 if (vma->vm_flags & VM_EXECUTABLE)
725                         removed_exe_file_vma(mm);
726         }
727         put_nommu_region(vma->vm_region);
728         kmem_cache_free(vm_area_cachep, vma);
729 }
730
731 /*
732  * look up the first VMA in which addr resides, NULL if none
733  * - should be called with mm->mmap_sem at least held readlocked
734  */
735 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
736 {
737         struct vm_area_struct *vma;
738         struct rb_node *n = mm->mm_rb.rb_node;
739
740         /* check the cache first */
741         vma = mm->mmap_cache;
742         if (vma && vma->vm_start <= addr && vma->vm_end > addr)
743                 return vma;
744
745         /* trawl the tree (there may be multiple mappings in which addr
746          * resides) */
747         for (n = rb_first(&mm->mm_rb); n; n = rb_next(n)) {
748                 vma = rb_entry(n, struct vm_area_struct, vm_rb);
749                 if (vma->vm_start > addr)
750                         return NULL;
751                 if (vma->vm_end > addr) {
752                         mm->mmap_cache = vma;
753                         return vma;
754                 }
755         }
756
757         return NULL;
758 }
759 EXPORT_SYMBOL(find_vma);
760
761 /*
762  * find a VMA
763  * - we don't extend stack VMAs under NOMMU conditions
764  */
765 struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr)
766 {
767         return find_vma(mm, addr);
768 }
769
770 /*
771  * expand a stack to a given address
772  * - not supported under NOMMU conditions
773  */
774 int expand_stack(struct vm_area_struct *vma, unsigned long address)
775 {
776         return -ENOMEM;
777 }
778
779 /*
780  * look up the first VMA exactly that exactly matches addr
781  * - should be called with mm->mmap_sem at least held readlocked
782  */
783 static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
784                                              unsigned long addr,
785                                              unsigned long len)
786 {
787         struct vm_area_struct *vma;
788         struct rb_node *n = mm->mm_rb.rb_node;
789         unsigned long end = addr + len;
790
791         /* check the cache first */
792         vma = mm->mmap_cache;
793         if (vma && vma->vm_start == addr && vma->vm_end == end)
794                 return vma;
795
796         /* trawl the tree (there may be multiple mappings in which addr
797          * resides) */
798         for (n = rb_first(&mm->mm_rb); n; n = rb_next(n)) {
799                 vma = rb_entry(n, struct vm_area_struct, vm_rb);
800                 if (vma->vm_start < addr)
801                         continue;
802                 if (vma->vm_start > addr)
803                         return NULL;
804                 if (vma->vm_end == end) {
805                         mm->mmap_cache = vma;
806                         return vma;
807                 }
808         }
809
810         return NULL;
811 }
812
813 /*
814  * determine whether a mapping should be permitted and, if so, what sort of
815  * mapping we're capable of supporting
816  */
817 static int validate_mmap_request(struct file *file,
818                                  unsigned long addr,
819                                  unsigned long len,
820                                  unsigned long prot,
821                                  unsigned long flags,
822                                  unsigned long pgoff,
823                                  unsigned long *_capabilities)
824 {
825         unsigned long capabilities, rlen;
826         unsigned long reqprot = prot;
827         int ret;
828
829         /* do the simple checks first */
830         if (flags & MAP_FIXED || addr) {
831                 printk(KERN_DEBUG
832                        "%d: Can't do fixed-address/overlay mmap of RAM\n",
833                        current->pid);
834                 return -EINVAL;
835         }
836
837         if ((flags & MAP_TYPE) != MAP_PRIVATE &&
838             (flags & MAP_TYPE) != MAP_SHARED)
839                 return -EINVAL;
840
841         if (!len)
842                 return -EINVAL;
843
844         /* Careful about overflows.. */
845         rlen = PAGE_ALIGN(len);
846         if (!rlen || rlen > TASK_SIZE)
847                 return -ENOMEM;
848
849         /* offset overflow? */
850         if ((pgoff + (rlen >> PAGE_SHIFT)) < pgoff)
851                 return -EOVERFLOW;
852
853         if (file) {
854                 /* validate file mapping requests */
855                 struct address_space *mapping;
856
857                 /* files must support mmap */
858                 if (!file->f_op || !file->f_op->mmap)
859                         return -ENODEV;
860
861                 /* work out if what we've got could possibly be shared
862                  * - we support chardevs that provide their own "memory"
863                  * - we support files/blockdevs that are memory backed
864                  */
865                 mapping = file->f_mapping;
866                 if (!mapping)
867                         mapping = file->f_path.dentry->d_inode->i_mapping;
868
869                 capabilities = 0;
870                 if (mapping && mapping->backing_dev_info)
871                         capabilities = mapping->backing_dev_info->capabilities;
872
873                 if (!capabilities) {
874                         /* no explicit capabilities set, so assume some
875                          * defaults */
876                         switch (file->f_path.dentry->d_inode->i_mode & S_IFMT) {
877                         case S_IFREG:
878                         case S_IFBLK:
879                                 capabilities = BDI_CAP_MAP_COPY;
880                                 break;
881
882                         case S_IFCHR:
883                                 capabilities =
884                                         BDI_CAP_MAP_DIRECT |
885                                         BDI_CAP_READ_MAP |
886                                         BDI_CAP_WRITE_MAP;
887                                 break;
888
889                         default:
890                                 return -EINVAL;
891                         }
892                 }
893
894                 /* eliminate any capabilities that we can't support on this
895                  * device */
896                 if (!file->f_op->get_unmapped_area)
897                         capabilities &= ~BDI_CAP_MAP_DIRECT;
898                 if (!file->f_op->read)
899                         capabilities &= ~BDI_CAP_MAP_COPY;
900
901                 if (flags & MAP_SHARED) {
902                         /* do checks for writing, appending and locking */
903                         if ((prot & PROT_WRITE) &&
904                             !(file->f_mode & FMODE_WRITE))
905                                 return -EACCES;
906
907                         if (IS_APPEND(file->f_path.dentry->d_inode) &&
908                             (file->f_mode & FMODE_WRITE))
909                                 return -EACCES;
910
911                         if (locks_verify_locked(file->f_path.dentry->d_inode))
912                                 return -EAGAIN;
913
914                         if (!(capabilities & BDI_CAP_MAP_DIRECT))
915                                 return -ENODEV;
916
917                         if (((prot & PROT_READ)  && !(capabilities & BDI_CAP_READ_MAP))  ||
918                             ((prot & PROT_WRITE) && !(capabilities & BDI_CAP_WRITE_MAP)) ||
919                             ((prot & PROT_EXEC)  && !(capabilities & BDI_CAP_EXEC_MAP))
920                             ) {
921                                 printk("MAP_SHARED not completely supported on !MMU\n");
922                                 return -EINVAL;
923                         }
924
925                         /* we mustn't privatise shared mappings */
926                         capabilities &= ~BDI_CAP_MAP_COPY;
927                 }
928                 else {
929                         /* we're going to read the file into private memory we
930                          * allocate */
931                         if (!(capabilities & BDI_CAP_MAP_COPY))
932                                 return -ENODEV;
933
934                         /* we don't permit a private writable mapping to be
935                          * shared with the backing device */
936                         if (prot & PROT_WRITE)
937                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
938                 }
939
940                 /* handle executable mappings and implied executable
941                  * mappings */
942                 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
943                         if (prot & PROT_EXEC)
944                                 return -EPERM;
945                 }
946                 else if ((prot & PROT_READ) && !(prot & PROT_EXEC)) {
947                         /* handle implication of PROT_EXEC by PROT_READ */
948                         if (current->personality & READ_IMPLIES_EXEC) {
949                                 if (capabilities & BDI_CAP_EXEC_MAP)
950                                         prot |= PROT_EXEC;
951                         }
952                 }
953                 else if ((prot & PROT_READ) &&
954                          (prot & PROT_EXEC) &&
955                          !(capabilities & BDI_CAP_EXEC_MAP)
956                          ) {
957                         /* backing file is not executable, try to copy */
958                         capabilities &= ~BDI_CAP_MAP_DIRECT;
959                 }
960         }
961         else {
962                 /* anonymous mappings are always memory backed and can be
963                  * privately mapped
964                  */
965                 capabilities = BDI_CAP_MAP_COPY;
966
967                 /* handle PROT_EXEC implication by PROT_READ */
968                 if ((prot & PROT_READ) &&
969                     (current->personality & READ_IMPLIES_EXEC))
970                         prot |= PROT_EXEC;
971         }
972
973         /* allow the security API to have its say */
974         ret = security_file_mmap(file, reqprot, prot, flags, addr, 0);
975         if (ret < 0)
976                 return ret;
977
978         /* looks okay */
979         *_capabilities = capabilities;
980         return 0;
981 }
982
983 /*
984  * we've determined that we can make the mapping, now translate what we
985  * now know into VMA flags
986  */
987 static unsigned long determine_vm_flags(struct file *file,
988                                         unsigned long prot,
989                                         unsigned long flags,
990                                         unsigned long capabilities)
991 {
992         unsigned long vm_flags;
993
994         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags);
995         vm_flags |= VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
996         /* vm_flags |= mm->def_flags; */
997
998         if (!(capabilities & BDI_CAP_MAP_DIRECT)) {
999                 /* attempt to share read-only copies of mapped file chunks */
1000                 if (file && !(prot & PROT_WRITE))
1001                         vm_flags |= VM_MAYSHARE;
1002         }
1003         else {
1004                 /* overlay a shareable mapping on the backing device or inode
1005                  * if possible - used for chardevs, ramfs/tmpfs/shmfs and
1006                  * romfs/cramfs */
1007                 if (flags & MAP_SHARED)
1008                         vm_flags |= VM_MAYSHARE | VM_SHARED;
1009                 else if ((((vm_flags & capabilities) ^ vm_flags) & BDI_CAP_VMFLAGS) == 0)
1010                         vm_flags |= VM_MAYSHARE;
1011         }
1012
1013         /* refuse to let anyone share private mappings with this process if
1014          * it's being traced - otherwise breakpoints set in it may interfere
1015          * with another untraced process
1016          */
1017         if ((flags & MAP_PRIVATE) && tracehook_expect_breakpoints(current))
1018                 vm_flags &= ~VM_MAYSHARE;
1019
1020         return vm_flags;
1021 }
1022
1023 /*
1024  * set up a shared mapping on a file (the driver or filesystem provides and
1025  * pins the storage)
1026  */
1027 static int do_mmap_shared_file(struct vm_area_struct *vma)
1028 {
1029         int ret;
1030
1031         ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1032         if (ret == 0) {
1033                 vma->vm_region->vm_top = vma->vm_region->vm_end;
1034                 return ret;
1035         }
1036         if (ret != -ENOSYS)
1037                 return ret;
1038
1039         /* getting an ENOSYS error indicates that direct mmap isn't
1040          * possible (as opposed to tried but failed) so we'll fall
1041          * through to making a private copy of the data and mapping
1042          * that if we can */
1043         return -ENODEV;
1044 }
1045
1046 /*
1047  * set up a private mapping or an anonymous shared mapping
1048  */
1049 static int do_mmap_private(struct vm_area_struct *vma,
1050                            struct vm_region *region,
1051                            unsigned long len)
1052 {
1053         struct page *pages;
1054         unsigned long total, point, n, rlen;
1055         void *base;
1056         int ret, order;
1057
1058         /* invoke the file's mapping function so that it can keep track of
1059          * shared mappings on devices or memory
1060          * - VM_MAYSHARE will be set if it may attempt to share
1061          */
1062         if (vma->vm_file) {
1063                 ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1064                 if (ret == 0) {
1065                         /* shouldn't return success if we're not sharing */
1066                         BUG_ON(!(vma->vm_flags & VM_MAYSHARE));
1067                         vma->vm_region->vm_top = vma->vm_region->vm_end;
1068                         return ret;
1069                 }
1070                 if (ret != -ENOSYS)
1071                         return ret;
1072
1073                 /* getting an ENOSYS error indicates that direct mmap isn't
1074                  * possible (as opposed to tried but failed) so we'll try to
1075                  * make a private copy of the data and map that instead */
1076         }
1077
1078         rlen = PAGE_ALIGN(len);
1079
1080         /* allocate some memory to hold the mapping
1081          * - note that this may not return a page-aligned address if the object
1082          *   we're allocating is smaller than a page
1083          */
1084         order = get_order(rlen);
1085         kdebug("alloc order %d for %lx", order, len);
1086
1087         pages = alloc_pages(GFP_KERNEL, order);
1088         if (!pages)
1089                 goto enomem;
1090
1091         total = 1 << order;
1092         atomic_long_add(total, &mmap_pages_allocated);
1093
1094         point = rlen >> PAGE_SHIFT;
1095
1096         /* we allocated a power-of-2 sized page set, so we may want to trim off
1097          * the excess */
1098         if (sysctl_nr_trim_pages && total - point >= sysctl_nr_trim_pages) {
1099                 while (total > point) {
1100                         order = ilog2(total - point);
1101                         n = 1 << order;
1102                         kdebug("shave %lu/%lu @%lu", n, total - point, total);
1103                         atomic_long_sub(n, &mmap_pages_allocated);
1104                         total -= n;
1105                         set_page_refcounted(pages + total);
1106                         __free_pages(pages + total, order);
1107                 }
1108         }
1109
1110         for (point = 1; point < total; point++)
1111                 set_page_refcounted(&pages[point]);
1112
1113         base = page_address(pages);
1114         region->vm_flags = vma->vm_flags |= VM_MAPPED_COPY;
1115         region->vm_start = (unsigned long) base;
1116         region->vm_end   = region->vm_start + rlen;
1117         region->vm_top   = region->vm_start + (total << PAGE_SHIFT);
1118
1119         vma->vm_start = region->vm_start;
1120         vma->vm_end   = region->vm_start + len;
1121
1122         if (vma->vm_file) {
1123                 /* read the contents of a file into the copy */
1124                 mm_segment_t old_fs;
1125                 loff_t fpos;
1126
1127                 fpos = vma->vm_pgoff;
1128                 fpos <<= PAGE_SHIFT;
1129
1130                 old_fs = get_fs();
1131                 set_fs(KERNEL_DS);
1132                 ret = vma->vm_file->f_op->read(vma->vm_file, base, rlen, &fpos);
1133                 set_fs(old_fs);
1134
1135                 if (ret < 0)
1136                         goto error_free;
1137
1138                 /* clear the last little bit */
1139                 if (ret < rlen)
1140                         memset(base + ret, 0, rlen - ret);
1141
1142         } else {
1143                 /* if it's an anonymous mapping, then just clear it */
1144                 memset(base, 0, rlen);
1145         }
1146
1147         return 0;
1148
1149 error_free:
1150         free_page_series(region->vm_start, region->vm_end);
1151         region->vm_start = vma->vm_start = 0;
1152         region->vm_end   = vma->vm_end = 0;
1153         region->vm_top   = 0;
1154         return ret;
1155
1156 enomem:
1157         printk("Allocation of length %lu from process %d (%s) failed\n",
1158                len, current->pid, current->comm);
1159         show_free_areas();
1160         return -ENOMEM;
1161 }
1162
1163 /*
1164  * handle mapping creation for uClinux
1165  */
1166 unsigned long do_mmap_pgoff(struct file *file,
1167                             unsigned long addr,
1168                             unsigned long len,
1169                             unsigned long prot,
1170                             unsigned long flags,
1171                             unsigned long pgoff)
1172 {
1173         struct vm_area_struct *vma;
1174         struct vm_region *region;
1175         struct rb_node *rb;
1176         unsigned long capabilities, vm_flags, result;
1177         int ret;
1178
1179         kenter(",%lx,%lx,%lx,%lx,%lx", addr, len, prot, flags, pgoff);
1180
1181         if (!(flags & MAP_FIXED))
1182                 addr = round_hint_to_min(addr);
1183
1184         /* decide whether we should attempt the mapping, and if so what sort of
1185          * mapping */
1186         ret = validate_mmap_request(file, addr, len, prot, flags, pgoff,
1187                                     &capabilities);
1188         if (ret < 0) {
1189                 kleave(" = %d [val]", ret);
1190                 return ret;
1191         }
1192
1193         /* we've determined that we can make the mapping, now translate what we
1194          * now know into VMA flags */
1195         vm_flags = determine_vm_flags(file, prot, flags, capabilities);
1196
1197         /* we're going to need to record the mapping */
1198         region = kmem_cache_zalloc(vm_region_jar, GFP_KERNEL);
1199         if (!region)
1200                 goto error_getting_region;
1201
1202         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1203         if (!vma)
1204                 goto error_getting_vma;
1205
1206         atomic_set(&region->vm_usage, 1);
1207         region->vm_flags = vm_flags;
1208         region->vm_pgoff = pgoff;
1209
1210         INIT_LIST_HEAD(&vma->anon_vma_node);
1211         vma->vm_flags = vm_flags;
1212         vma->vm_pgoff = pgoff;
1213
1214         if (file) {
1215                 region->vm_file = file;
1216                 get_file(file);
1217                 vma->vm_file = file;
1218                 get_file(file);
1219                 if (vm_flags & VM_EXECUTABLE) {
1220                         added_exe_file_vma(current->mm);
1221                         vma->vm_mm = current->mm;
1222                 }
1223         }
1224
1225         down_write(&nommu_region_sem);
1226
1227         /* if we want to share, we need to check for regions created by other
1228          * mmap() calls that overlap with our proposed mapping
1229          * - we can only share with a superset match on most regular files
1230          * - shared mappings on character devices and memory backed files are
1231          *   permitted to overlap inexactly as far as we are concerned for in
1232          *   these cases, sharing is handled in the driver or filesystem rather
1233          *   than here
1234          */
1235         if (vm_flags & VM_MAYSHARE) {
1236                 struct vm_region *pregion;
1237                 unsigned long pglen, rpglen, pgend, rpgend, start;
1238
1239                 pglen = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1240                 pgend = pgoff + pglen;
1241
1242                 for (rb = rb_first(&nommu_region_tree); rb; rb = rb_next(rb)) {
1243                         pregion = rb_entry(rb, struct vm_region, vm_rb);
1244
1245                         if (!(pregion->vm_flags & VM_MAYSHARE))
1246                                 continue;
1247
1248                         /* search for overlapping mappings on the same file */
1249                         if (pregion->vm_file->f_path.dentry->d_inode !=
1250                             file->f_path.dentry->d_inode)
1251                                 continue;
1252
1253                         if (pregion->vm_pgoff >= pgend)
1254                                 continue;
1255
1256                         rpglen = pregion->vm_end - pregion->vm_start;
1257                         rpglen = (rpglen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1258                         rpgend = pregion->vm_pgoff + rpglen;
1259                         if (pgoff >= rpgend)
1260                                 continue;
1261
1262                         /* handle inexactly overlapping matches between
1263                          * mappings */
1264                         if ((pregion->vm_pgoff != pgoff || rpglen != pglen) &&
1265                             !(pgoff >= pregion->vm_pgoff && pgend <= rpgend)) {
1266                                 /* new mapping is not a subset of the region */
1267                                 if (!(capabilities & BDI_CAP_MAP_DIRECT))
1268                                         goto sharing_violation;
1269                                 continue;
1270                         }
1271
1272                         /* we've found a region we can share */
1273                         atomic_inc(&pregion->vm_usage);
1274                         vma->vm_region = pregion;
1275                         start = pregion->vm_start;
1276                         start += (pgoff - pregion->vm_pgoff) << PAGE_SHIFT;
1277                         vma->vm_start = start;
1278                         vma->vm_end = start + len;
1279
1280                         if (pregion->vm_flags & VM_MAPPED_COPY) {
1281                                 kdebug("share copy");
1282                                 vma->vm_flags |= VM_MAPPED_COPY;
1283                         } else {
1284                                 kdebug("share mmap");
1285                                 ret = do_mmap_shared_file(vma);
1286                                 if (ret < 0) {
1287                                         vma->vm_region = NULL;
1288                                         vma->vm_start = 0;
1289                                         vma->vm_end = 0;
1290                                         atomic_dec(&pregion->vm_usage);
1291                                         pregion = NULL;
1292                                         goto error_just_free;
1293                                 }
1294                         }
1295                         fput(region->vm_file);
1296                         kmem_cache_free(vm_region_jar, region);
1297                         region = pregion;
1298                         result = start;
1299                         goto share;
1300                 }
1301
1302                 /* obtain the address at which to make a shared mapping
1303                  * - this is the hook for quasi-memory character devices to
1304                  *   tell us the location of a shared mapping
1305                  */
1306                 if (file && file->f_op->get_unmapped_area) {
1307                         addr = file->f_op->get_unmapped_area(file, addr, len,
1308                                                              pgoff, flags);
1309                         if (IS_ERR((void *) addr)) {
1310                                 ret = addr;
1311                                 if (ret != (unsigned long) -ENOSYS)
1312                                         goto error_just_free;
1313
1314                                 /* the driver refused to tell us where to site
1315                                  * the mapping so we'll have to attempt to copy
1316                                  * it */
1317                                 ret = (unsigned long) -ENODEV;
1318                                 if (!(capabilities & BDI_CAP_MAP_COPY))
1319                                         goto error_just_free;
1320
1321                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
1322                         } else {
1323                                 vma->vm_start = region->vm_start = addr;
1324                                 vma->vm_end = region->vm_end = addr + len;
1325                         }
1326                 }
1327         }
1328
1329         vma->vm_region = region;
1330
1331         /* set up the mapping */
1332         if (file && vma->vm_flags & VM_SHARED)
1333                 ret = do_mmap_shared_file(vma);
1334         else
1335                 ret = do_mmap_private(vma, region, len);
1336         if (ret < 0)
1337                 goto error_put_region;
1338
1339         add_nommu_region(region);
1340
1341         /* okay... we have a mapping; now we have to register it */
1342         result = vma->vm_start;
1343
1344         current->mm->total_vm += len >> PAGE_SHIFT;
1345
1346 share:
1347         add_vma_to_mm(current->mm, vma);
1348
1349         up_write(&nommu_region_sem);
1350
1351         if (prot & PROT_EXEC)
1352                 flush_icache_range(result, result + len);
1353
1354         kleave(" = %lx", result);
1355         return result;
1356
1357 error_put_region:
1358         __put_nommu_region(region);
1359         if (vma) {
1360                 if (vma->vm_file) {
1361                         fput(vma->vm_file);
1362                         if (vma->vm_flags & VM_EXECUTABLE)
1363                                 removed_exe_file_vma(vma->vm_mm);
1364                 }
1365                 kmem_cache_free(vm_area_cachep, vma);
1366         }
1367         kleave(" = %d [pr]", ret);
1368         return ret;
1369
1370 error_just_free:
1371         up_write(&nommu_region_sem);
1372 error:
1373         fput(region->vm_file);
1374         kmem_cache_free(vm_region_jar, region);
1375         fput(vma->vm_file);
1376         if (vma->vm_flags & VM_EXECUTABLE)
1377                 removed_exe_file_vma(vma->vm_mm);
1378         kmem_cache_free(vm_area_cachep, vma);
1379         kleave(" = %d", ret);
1380         return ret;
1381
1382 sharing_violation:
1383         up_write(&nommu_region_sem);
1384         printk(KERN_WARNING "Attempt to share mismatched mappings\n");
1385         ret = -EINVAL;
1386         goto error;
1387
1388 error_getting_vma:
1389         kmem_cache_free(vm_region_jar, region);
1390         printk(KERN_WARNING "Allocation of vma for %lu byte allocation"
1391                " from process %d failed\n",
1392                len, current->pid);
1393         show_free_areas();
1394         return -ENOMEM;
1395
1396 error_getting_region:
1397         printk(KERN_WARNING "Allocation of vm region for %lu byte allocation"
1398                " from process %d failed\n",
1399                len, current->pid);
1400         show_free_areas();
1401         return -ENOMEM;
1402 }
1403 EXPORT_SYMBOL(do_mmap_pgoff);
1404
1405 /*
1406  * split a vma into two pieces at address 'addr', a new vma is allocated either
1407  * for the first part or the tail.
1408  */
1409 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1410               unsigned long addr, int new_below)
1411 {
1412         struct vm_area_struct *new;
1413         struct vm_region *region;
1414         unsigned long npages;
1415
1416         kenter("");
1417
1418         /* we're only permitted to split anonymous regions that have a single
1419          * owner */
1420         if (vma->vm_file ||
1421             atomic_read(&vma->vm_region->vm_usage) != 1)
1422                 return -ENOMEM;
1423
1424         if (mm->map_count >= sysctl_max_map_count)
1425                 return -ENOMEM;
1426
1427         region = kmem_cache_alloc(vm_region_jar, GFP_KERNEL);
1428         if (!region)
1429                 return -ENOMEM;
1430
1431         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1432         if (!new) {
1433                 kmem_cache_free(vm_region_jar, region);
1434                 return -ENOMEM;
1435         }
1436
1437         /* most fields are the same, copy all, and then fixup */
1438         *new = *vma;
1439         *region = *vma->vm_region;
1440         new->vm_region = region;
1441
1442         npages = (addr - vma->vm_start) >> PAGE_SHIFT;
1443
1444         if (new_below) {
1445                 region->vm_top = region->vm_end = new->vm_end = addr;
1446         } else {
1447                 region->vm_start = new->vm_start = addr;
1448                 region->vm_pgoff = new->vm_pgoff += npages;
1449         }
1450
1451         if (new->vm_ops && new->vm_ops->open)
1452                 new->vm_ops->open(new);
1453
1454         delete_vma_from_mm(vma);
1455         down_write(&nommu_region_sem);
1456         delete_nommu_region(vma->vm_region);
1457         if (new_below) {
1458                 vma->vm_region->vm_start = vma->vm_start = addr;
1459                 vma->vm_region->vm_pgoff = vma->vm_pgoff += npages;
1460         } else {
1461                 vma->vm_region->vm_end = vma->vm_end = addr;
1462                 vma->vm_region->vm_top = addr;
1463         }
1464         add_nommu_region(vma->vm_region);
1465         add_nommu_region(new->vm_region);
1466         up_write(&nommu_region_sem);
1467         add_vma_to_mm(mm, vma);
1468         add_vma_to_mm(mm, new);
1469         return 0;
1470 }
1471
1472 /*
1473  * shrink a VMA by removing the specified chunk from either the beginning or
1474  * the end
1475  */
1476 static int shrink_vma(struct mm_struct *mm,
1477                       struct vm_area_struct *vma,
1478                       unsigned long from, unsigned long to)
1479 {
1480         struct vm_region *region;
1481
1482         kenter("");
1483
1484         /* adjust the VMA's pointers, which may reposition it in the MM's tree
1485          * and list */
1486         delete_vma_from_mm(vma);
1487         if (from > vma->vm_start)
1488                 vma->vm_end = from;
1489         else
1490                 vma->vm_start = to;
1491         add_vma_to_mm(mm, vma);
1492
1493         /* cut the backing region down to size */
1494         region = vma->vm_region;
1495         BUG_ON(atomic_read(&region->vm_usage) != 1);
1496
1497         down_write(&nommu_region_sem);
1498         delete_nommu_region(region);
1499         if (from > region->vm_start) {
1500                 to = region->vm_top;
1501                 region->vm_top = region->vm_end = from;
1502         } else {
1503                 region->vm_start = to;
1504         }
1505         add_nommu_region(region);
1506         up_write(&nommu_region_sem);
1507
1508         free_page_series(from, to);
1509         return 0;
1510 }
1511
1512 /*
1513  * release a mapping
1514  * - under NOMMU conditions the chunk to be unmapped must be backed by a single
1515  *   VMA, though it need not cover the whole VMA
1516  */
1517 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1518 {
1519         struct vm_area_struct *vma;
1520         struct rb_node *rb;
1521         unsigned long end = start + len;
1522         int ret;
1523
1524         kenter(",%lx,%zx", start, len);
1525
1526         if (len == 0)
1527                 return -EINVAL;
1528
1529         /* find the first potentially overlapping VMA */
1530         vma = find_vma(mm, start);
1531         if (!vma) {
1532                 static int limit = 0;
1533                 if (limit < 5) {
1534                         printk(KERN_WARNING
1535                                "munmap of memory not mmapped by process %d"
1536                                " (%s): 0x%lx-0x%lx\n",
1537                                current->pid, current->comm,
1538                                start, start + len - 1);
1539                         limit++;
1540                 }
1541                 return -EINVAL;
1542         }
1543
1544         /* we're allowed to split an anonymous VMA but not a file-backed one */
1545         if (vma->vm_file) {
1546                 do {
1547                         if (start > vma->vm_start) {
1548                                 kleave(" = -EINVAL [miss]");
1549                                 return -EINVAL;
1550                         }
1551                         if (end == vma->vm_end)
1552                                 goto erase_whole_vma;
1553                         rb = rb_next(&vma->vm_rb);
1554                         vma = rb_entry(rb, struct vm_area_struct, vm_rb);
1555                 } while (rb);
1556                 kleave(" = -EINVAL [split file]");
1557                 return -EINVAL;
1558         } else {
1559                 /* the chunk must be a subset of the VMA found */
1560                 if (start == vma->vm_start && end == vma->vm_end)
1561                         goto erase_whole_vma;
1562                 if (start < vma->vm_start || end > vma->vm_end) {
1563                         kleave(" = -EINVAL [superset]");
1564                         return -EINVAL;
1565                 }
1566                 if (start & ~PAGE_MASK) {
1567                         kleave(" = -EINVAL [unaligned start]");
1568                         return -EINVAL;
1569                 }
1570                 if (end != vma->vm_end && end & ~PAGE_MASK) {
1571                         kleave(" = -EINVAL [unaligned split]");
1572                         return -EINVAL;
1573                 }
1574                 if (start != vma->vm_start && end != vma->vm_end) {
1575                         ret = split_vma(mm, vma, start, 1);
1576                         if (ret < 0) {
1577                                 kleave(" = %d [split]", ret);
1578                                 return ret;
1579                         }
1580                 }
1581                 return shrink_vma(mm, vma, start, end);
1582         }
1583
1584 erase_whole_vma:
1585         delete_vma_from_mm(vma);
1586         delete_vma(mm, vma);
1587         kleave(" = 0");
1588         return 0;
1589 }
1590 EXPORT_SYMBOL(do_munmap);
1591
1592 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1593 {
1594         int ret;
1595         struct mm_struct *mm = current->mm;
1596
1597         down_write(&mm->mmap_sem);
1598         ret = do_munmap(mm, addr, len);
1599         up_write(&mm->mmap_sem);
1600         return ret;
1601 }
1602
1603 /*
1604  * release all the mappings made in a process's VM space
1605  */
1606 void exit_mmap(struct mm_struct *mm)
1607 {
1608         struct vm_area_struct *vma;
1609
1610         if (!mm)
1611                 return;
1612
1613         kenter("");
1614
1615         mm->total_vm = 0;
1616
1617         while ((vma = mm->mmap)) {
1618                 mm->mmap = vma->vm_next;
1619                 delete_vma_from_mm(vma);
1620                 delete_vma(mm, vma);
1621         }
1622
1623         kleave("");
1624 }
1625
1626 unsigned long do_brk(unsigned long addr, unsigned long len)
1627 {
1628         return -ENOMEM;
1629 }
1630
1631 /*
1632  * expand (or shrink) an existing mapping, potentially moving it at the same
1633  * time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1634  *
1635  * under NOMMU conditions, we only permit changing a mapping's size, and only
1636  * as long as it stays within the region allocated by do_mmap_private() and the
1637  * block is not shareable
1638  *
1639  * MREMAP_FIXED is not supported under NOMMU conditions
1640  */
1641 unsigned long do_mremap(unsigned long addr,
1642                         unsigned long old_len, unsigned long new_len,
1643                         unsigned long flags, unsigned long new_addr)
1644 {
1645         struct vm_area_struct *vma;
1646
1647         /* insanity checks first */
1648         if (old_len == 0 || new_len == 0)
1649                 return (unsigned long) -EINVAL;
1650
1651         if (addr & ~PAGE_MASK)
1652                 return -EINVAL;
1653
1654         if (flags & MREMAP_FIXED && new_addr != addr)
1655                 return (unsigned long) -EINVAL;
1656
1657         vma = find_vma_exact(current->mm, addr, old_len);
1658         if (!vma)
1659                 return (unsigned long) -EINVAL;
1660
1661         if (vma->vm_end != vma->vm_start + old_len)
1662                 return (unsigned long) -EFAULT;
1663
1664         if (vma->vm_flags & VM_MAYSHARE)
1665                 return (unsigned long) -EPERM;
1666
1667         if (new_len > vma->vm_region->vm_end - vma->vm_region->vm_start)
1668                 return (unsigned long) -ENOMEM;
1669
1670         /* all checks complete - do it */
1671         vma->vm_end = vma->vm_start + new_len;
1672         return vma->vm_start;
1673 }
1674 EXPORT_SYMBOL(do_mremap);
1675
1676 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1677                 unsigned long, new_len, unsigned long, flags,
1678                 unsigned long, new_addr)
1679 {
1680         unsigned long ret;
1681
1682         down_write(&current->mm->mmap_sem);
1683         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
1684         up_write(&current->mm->mmap_sem);
1685         return ret;
1686 }
1687
1688 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1689                         unsigned int foll_flags)
1690 {
1691         return NULL;
1692 }
1693
1694 int remap_pfn_range(struct vm_area_struct *vma, unsigned long from,
1695                 unsigned long to, unsigned long size, pgprot_t prot)
1696 {
1697         vma->vm_start = vma->vm_pgoff << PAGE_SHIFT;
1698         return 0;
1699 }
1700 EXPORT_SYMBOL(remap_pfn_range);
1701
1702 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1703                         unsigned long pgoff)
1704 {
1705         unsigned int size = vma->vm_end - vma->vm_start;
1706
1707         if (!(vma->vm_flags & VM_USERMAP))
1708                 return -EINVAL;
1709
1710         vma->vm_start = (unsigned long)(addr + (pgoff << PAGE_SHIFT));
1711         vma->vm_end = vma->vm_start + size;
1712
1713         return 0;
1714 }
1715 EXPORT_SYMBOL(remap_vmalloc_range);
1716
1717 void swap_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
1718 {
1719 }
1720
1721 unsigned long arch_get_unmapped_area(struct file *file, unsigned long addr,
1722         unsigned long len, unsigned long pgoff, unsigned long flags)
1723 {
1724         return -ENOMEM;
1725 }
1726
1727 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1728 {
1729 }
1730
1731 void unmap_mapping_range(struct address_space *mapping,
1732                          loff_t const holebegin, loff_t const holelen,
1733                          int even_cows)
1734 {
1735 }
1736 EXPORT_SYMBOL(unmap_mapping_range);
1737
1738 /*
1739  * ask for an unmapped area at which to create a mapping on a file
1740  */
1741 unsigned long get_unmapped_area(struct file *file, unsigned long addr,
1742                                 unsigned long len, unsigned long pgoff,
1743                                 unsigned long flags)
1744 {
1745         unsigned long (*get_area)(struct file *, unsigned long, unsigned long,
1746                                   unsigned long, unsigned long);
1747
1748         get_area = current->mm->get_unmapped_area;
1749         if (file && file->f_op && file->f_op->get_unmapped_area)
1750                 get_area = file->f_op->get_unmapped_area;
1751
1752         if (!get_area)
1753                 return -ENOSYS;
1754
1755         return get_area(file, addr, len, pgoff, flags);
1756 }
1757 EXPORT_SYMBOL(get_unmapped_area);
1758
1759 /*
1760  * Check that a process has enough memory to allocate a new virtual
1761  * mapping. 0 means there is enough memory for the allocation to
1762  * succeed and -ENOMEM implies there is not.
1763  *
1764  * We currently support three overcommit policies, which are set via the
1765  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
1766  *
1767  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
1768  * Additional code 2002 Jul 20 by Robert Love.
1769  *
1770  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
1771  *
1772  * Note this is a helper function intended to be used by LSMs which
1773  * wish to use this logic.
1774  */
1775 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
1776 {
1777         unsigned long free, allowed;
1778
1779         vm_acct_memory(pages);
1780
1781         /*
1782          * Sometimes we want to use more memory than we have
1783          */
1784         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
1785                 return 0;
1786
1787         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
1788                 unsigned long n;
1789
1790                 free = global_page_state(NR_FILE_PAGES);
1791                 free += nr_swap_pages;
1792
1793                 /*
1794                  * Any slabs which are created with the
1795                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
1796                  * which are reclaimable, under pressure.  The dentry
1797                  * cache and most inode caches should fall into this
1798                  */
1799                 free += global_page_state(NR_SLAB_RECLAIMABLE);
1800
1801                 /*
1802                  * Leave the last 3% for root
1803                  */
1804                 if (!cap_sys_admin)
1805                         free -= free / 32;
1806
1807                 if (free > pages)
1808                         return 0;
1809
1810                 /*
1811                  * nr_free_pages() is very expensive on large systems,
1812                  * only call if we're about to fail.
1813                  */
1814                 n = nr_free_pages();
1815
1816                 /*
1817                  * Leave reserved pages. The pages are not for anonymous pages.
1818                  */
1819                 if (n <= totalreserve_pages)
1820                         goto error;
1821                 else
1822                         n -= totalreserve_pages;
1823
1824                 /*
1825                  * Leave the last 3% for root
1826                  */
1827                 if (!cap_sys_admin)
1828                         n -= n / 32;
1829                 free += n;
1830
1831                 if (free > pages)
1832                         return 0;
1833
1834                 goto error;
1835         }
1836
1837         allowed = totalram_pages * sysctl_overcommit_ratio / 100;
1838         /*
1839          * Leave the last 3% for root
1840          */
1841         if (!cap_sys_admin)
1842                 allowed -= allowed / 32;
1843         allowed += total_swap_pages;
1844
1845         /* Don't let a single process grow too big:
1846            leave 3% of the size of this process for other processes */
1847         if (mm)
1848                 allowed -= mm->total_vm / 32;
1849
1850         /*
1851          * cast `allowed' as a signed long because vm_committed_space
1852          * sometimes has a negative value
1853          */
1854         if (atomic_long_read(&vm_committed_space) < (long)allowed)
1855                 return 0;
1856 error:
1857         vm_unacct_memory(pages);
1858
1859         return -ENOMEM;
1860 }
1861
1862 int in_gate_area_no_task(unsigned long addr)
1863 {
1864         return 0;
1865 }
1866
1867 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1868 {
1869         BUG();
1870         return 0;
1871 }
1872 EXPORT_SYMBOL(filemap_fault);
1873
1874 /*
1875  * Access another process' address space.
1876  * - source/target buffer must be kernel space
1877  */
1878 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
1879 {
1880         struct vm_area_struct *vma;
1881         struct mm_struct *mm;
1882
1883         if (addr + len < addr)
1884                 return 0;
1885
1886         mm = get_task_mm(tsk);
1887         if (!mm)
1888                 return 0;
1889
1890         down_read(&mm->mmap_sem);
1891
1892         /* the access must start within one of the target process's mappings */
1893         vma = find_vma(mm, addr);
1894         if (vma) {
1895                 /* don't overrun this mapping */
1896                 if (addr + len >= vma->vm_end)
1897                         len = vma->vm_end - addr;
1898
1899                 /* only read or write mappings where it is permitted */
1900                 if (write && vma->vm_flags & VM_MAYWRITE)
1901                         len -= copy_to_user((void *) addr, buf, len);
1902                 else if (!write && vma->vm_flags & VM_MAYREAD)
1903                         len -= copy_from_user(buf, (void *) addr, len);
1904                 else
1905                         len = 0;
1906         } else {
1907                 len = 0;
1908         }
1909
1910         up_read(&mm->mmap_sem);
1911         mmput(mm);
1912         return len;
1913 }