mm: nommu: find vma using the sorted vma list
[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) 2007-2010 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 #include <linux/audit.h>
33
34 #include <asm/uaccess.h>
35 #include <asm/tlb.h>
36 #include <asm/tlbflush.h>
37 #include <asm/mmu_context.h>
38 #include "internal.h"
39
40 #if 0
41 #define kenter(FMT, ...) \
42         printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
43 #define kleave(FMT, ...) \
44         printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
45 #define kdebug(FMT, ...) \
46         printk(KERN_DEBUG "xxx" FMT"yyy\n", ##__VA_ARGS__)
47 #else
48 #define kenter(FMT, ...) \
49         no_printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
50 #define kleave(FMT, ...) \
51         no_printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
52 #define kdebug(FMT, ...) \
53         no_printk(KERN_DEBUG FMT"\n", ##__VA_ARGS__)
54 #endif
55
56 void *high_memory;
57 struct page *mem_map;
58 unsigned long max_mapnr;
59 unsigned long num_physpages;
60 unsigned long highest_memmap_pfn;
61 struct percpu_counter vm_committed_as;
62 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
63 int sysctl_overcommit_ratio = 50; /* default is 50% */
64 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT;
65 int sysctl_nr_trim_pages = CONFIG_NOMMU_INITIAL_TRIM_EXCESS;
66 int heap_stack_gap = 0;
67
68 atomic_long_t mmap_pages_allocated;
69
70 EXPORT_SYMBOL(mem_map);
71 EXPORT_SYMBOL(num_physpages);
72
73 /* list of mapped, potentially shareable regions */
74 static struct kmem_cache *vm_region_jar;
75 struct rb_root nommu_region_tree = RB_ROOT;
76 DECLARE_RWSEM(nommu_region_sem);
77
78 const struct vm_operations_struct generic_file_vm_ops = {
79 };
80
81 /*
82  * Return the total memory allocated for this pointer, not
83  * just what the caller asked for.
84  *
85  * Doesn't have to be accurate, i.e. may have races.
86  */
87 unsigned int kobjsize(const void *objp)
88 {
89         struct page *page;
90
91         /*
92          * If the object we have should not have ksize performed on it,
93          * return size of 0
94          */
95         if (!objp || !virt_addr_valid(objp))
96                 return 0;
97
98         page = virt_to_head_page(objp);
99
100         /*
101          * If the allocator sets PageSlab, we know the pointer came from
102          * kmalloc().
103          */
104         if (PageSlab(page))
105                 return ksize(objp);
106
107         /*
108          * If it's not a compound page, see if we have a matching VMA
109          * region. This test is intentionally done in reverse order,
110          * so if there's no VMA, we still fall through and hand back
111          * PAGE_SIZE for 0-order pages.
112          */
113         if (!PageCompound(page)) {
114                 struct vm_area_struct *vma;
115
116                 vma = find_vma(current->mm, (unsigned long)objp);
117                 if (vma)
118                         return vma->vm_end - vma->vm_start;
119         }
120
121         /*
122          * The ksize() function is only guaranteed to work for pointers
123          * returned by kmalloc(). So handle arbitrary pointers here.
124          */
125         return PAGE_SIZE << compound_order(page);
126 }
127
128 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
129                      unsigned long start, int nr_pages, unsigned int foll_flags,
130                      struct page **pages, struct vm_area_struct **vmas,
131                      int *retry)
132 {
133         struct vm_area_struct *vma;
134         unsigned long vm_flags;
135         int i;
136
137         /* calculate required read or write permissions.
138          * If FOLL_FORCE is set, we only require the "MAY" flags.
139          */
140         vm_flags  = (foll_flags & FOLL_WRITE) ?
141                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
142         vm_flags &= (foll_flags & FOLL_FORCE) ?
143                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
144
145         for (i = 0; i < nr_pages; i++) {
146                 vma = find_vma(mm, start);
147                 if (!vma)
148                         goto finish_or_fault;
149
150                 /* protect what we can, including chardevs */
151                 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
152                     !(vm_flags & vma->vm_flags))
153                         goto finish_or_fault;
154
155                 if (pages) {
156                         pages[i] = virt_to_page(start);
157                         if (pages[i])
158                                 page_cache_get(pages[i]);
159                 }
160                 if (vmas)
161                         vmas[i] = vma;
162                 start = (start + PAGE_SIZE) & PAGE_MASK;
163         }
164
165         return i;
166
167 finish_or_fault:
168         return i ? : -EFAULT;
169 }
170
171 /*
172  * get a list of pages in an address range belonging to the specified process
173  * and indicate the VMA that covers each page
174  * - this is potentially dodgy as we may end incrementing the page count of a
175  *   slab page or a secondary page from a compound page
176  * - don't permit access to VMAs that don't support it, such as I/O mappings
177  */
178 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
179         unsigned long start, int nr_pages, int write, int force,
180         struct page **pages, struct vm_area_struct **vmas)
181 {
182         int flags = 0;
183
184         if (write)
185                 flags |= FOLL_WRITE;
186         if (force)
187                 flags |= FOLL_FORCE;
188
189         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas,
190                                 NULL);
191 }
192 EXPORT_SYMBOL(get_user_pages);
193
194 /**
195  * follow_pfn - look up PFN at a user virtual address
196  * @vma: memory mapping
197  * @address: user virtual address
198  * @pfn: location to store found PFN
199  *
200  * Only IO mappings and raw PFN mappings are allowed.
201  *
202  * Returns zero and the pfn at @pfn on success, -ve otherwise.
203  */
204 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
205         unsigned long *pfn)
206 {
207         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
208                 return -EINVAL;
209
210         *pfn = address >> PAGE_SHIFT;
211         return 0;
212 }
213 EXPORT_SYMBOL(follow_pfn);
214
215 DEFINE_RWLOCK(vmlist_lock);
216 struct vm_struct *vmlist;
217
218 void vfree(const void *addr)
219 {
220         kfree(addr);
221 }
222 EXPORT_SYMBOL(vfree);
223
224 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
225 {
226         /*
227          *  You can't specify __GFP_HIGHMEM with kmalloc() since kmalloc()
228          * returns only a logical address.
229          */
230         return kmalloc(size, (gfp_mask | __GFP_COMP) & ~__GFP_HIGHMEM);
231 }
232 EXPORT_SYMBOL(__vmalloc);
233
234 void *vmalloc_user(unsigned long size)
235 {
236         void *ret;
237
238         ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
239                         PAGE_KERNEL);
240         if (ret) {
241                 struct vm_area_struct *vma;
242
243                 down_write(&current->mm->mmap_sem);
244                 vma = find_vma(current->mm, (unsigned long)ret);
245                 if (vma)
246                         vma->vm_flags |= VM_USERMAP;
247                 up_write(&current->mm->mmap_sem);
248         }
249
250         return ret;
251 }
252 EXPORT_SYMBOL(vmalloc_user);
253
254 struct page *vmalloc_to_page(const void *addr)
255 {
256         return virt_to_page(addr);
257 }
258 EXPORT_SYMBOL(vmalloc_to_page);
259
260 unsigned long vmalloc_to_pfn(const void *addr)
261 {
262         return page_to_pfn(virt_to_page(addr));
263 }
264 EXPORT_SYMBOL(vmalloc_to_pfn);
265
266 long vread(char *buf, char *addr, unsigned long count)
267 {
268         memcpy(buf, addr, count);
269         return count;
270 }
271
272 long vwrite(char *buf, char *addr, unsigned long count)
273 {
274         /* Don't allow overflow */
275         if ((unsigned long) addr + count < count)
276                 count = -(unsigned long) addr;
277
278         memcpy(addr, buf, count);
279         return(count);
280 }
281
282 /*
283  *      vmalloc  -  allocate virtually continguos memory
284  *
285  *      @size:          allocation size
286  *
287  *      Allocate enough pages to cover @size from the page level
288  *      allocator and map them into continguos kernel virtual space.
289  *
290  *      For tight control over page level allocator and protection flags
291  *      use __vmalloc() instead.
292  */
293 void *vmalloc(unsigned long size)
294 {
295        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
296 }
297 EXPORT_SYMBOL(vmalloc);
298
299 /*
300  *      vzalloc - allocate virtually continguos memory with zero fill
301  *
302  *      @size:          allocation size
303  *
304  *      Allocate enough pages to cover @size from the page level
305  *      allocator and map them into continguos kernel virtual space.
306  *      The memory allocated is set to zero.
307  *
308  *      For tight control over page level allocator and protection flags
309  *      use __vmalloc() instead.
310  */
311 void *vzalloc(unsigned long size)
312 {
313         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
314                         PAGE_KERNEL);
315 }
316 EXPORT_SYMBOL(vzalloc);
317
318 /**
319  * vmalloc_node - allocate memory on a specific node
320  * @size:       allocation size
321  * @node:       numa node
322  *
323  * Allocate enough pages to cover @size from the page level
324  * allocator and map them into contiguous kernel virtual space.
325  *
326  * For tight control over page level allocator and protection flags
327  * use __vmalloc() instead.
328  */
329 void *vmalloc_node(unsigned long size, int node)
330 {
331         return vmalloc(size);
332 }
333 EXPORT_SYMBOL(vmalloc_node);
334
335 /**
336  * vzalloc_node - allocate memory on a specific node with zero fill
337  * @size:       allocation size
338  * @node:       numa node
339  *
340  * Allocate enough pages to cover @size from the page level
341  * allocator and map them into contiguous kernel virtual space.
342  * The memory allocated is set to zero.
343  *
344  * For tight control over page level allocator and protection flags
345  * use __vmalloc() instead.
346  */
347 void *vzalloc_node(unsigned long size, int node)
348 {
349         return vzalloc(size);
350 }
351 EXPORT_SYMBOL(vzalloc_node);
352
353 #ifndef PAGE_KERNEL_EXEC
354 # define PAGE_KERNEL_EXEC PAGE_KERNEL
355 #endif
356
357 /**
358  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
359  *      @size:          allocation size
360  *
361  *      Kernel-internal function to allocate enough pages to cover @size
362  *      the page level allocator and map them into contiguous and
363  *      executable kernel virtual space.
364  *
365  *      For tight control over page level allocator and protection flags
366  *      use __vmalloc() instead.
367  */
368
369 void *vmalloc_exec(unsigned long size)
370 {
371         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
372 }
373
374 /**
375  * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
376  *      @size:          allocation size
377  *
378  *      Allocate enough 32bit PA addressable pages to cover @size from the
379  *      page level allocator and map them into continguos kernel virtual space.
380  */
381 void *vmalloc_32(unsigned long size)
382 {
383         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
384 }
385 EXPORT_SYMBOL(vmalloc_32);
386
387 /**
388  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
389  *      @size:          allocation size
390  *
391  * The resulting memory area is 32bit addressable and zeroed so it can be
392  * mapped to userspace without leaking data.
393  *
394  * VM_USERMAP is set on the corresponding VMA so that subsequent calls to
395  * remap_vmalloc_range() are permissible.
396  */
397 void *vmalloc_32_user(unsigned long size)
398 {
399         /*
400          * We'll have to sort out the ZONE_DMA bits for 64-bit,
401          * but for now this can simply use vmalloc_user() directly.
402          */
403         return vmalloc_user(size);
404 }
405 EXPORT_SYMBOL(vmalloc_32_user);
406
407 void *vmap(struct page **pages, unsigned int count, unsigned long flags, pgprot_t prot)
408 {
409         BUG();
410         return NULL;
411 }
412 EXPORT_SYMBOL(vmap);
413
414 void vunmap(const void *addr)
415 {
416         BUG();
417 }
418 EXPORT_SYMBOL(vunmap);
419
420 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
421 {
422         BUG();
423         return NULL;
424 }
425 EXPORT_SYMBOL(vm_map_ram);
426
427 void vm_unmap_ram(const void *mem, unsigned int count)
428 {
429         BUG();
430 }
431 EXPORT_SYMBOL(vm_unmap_ram);
432
433 void vm_unmap_aliases(void)
434 {
435 }
436 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
437
438 /*
439  * Implement a stub for vmalloc_sync_all() if the architecture chose not to
440  * have one.
441  */
442 void  __attribute__((weak)) vmalloc_sync_all(void)
443 {
444 }
445
446 /**
447  *      alloc_vm_area - allocate a range of kernel address space
448  *      @size:          size of the area
449  *
450  *      Returns:        NULL on failure, vm_struct on success
451  *
452  *      This function reserves a range of kernel address space, and
453  *      allocates pagetables to map that range.  No actual mappings
454  *      are created.  If the kernel address space is not shared
455  *      between processes, it syncs the pagetable across all
456  *      processes.
457  */
458 struct vm_struct *alloc_vm_area(size_t size)
459 {
460         BUG();
461         return NULL;
462 }
463 EXPORT_SYMBOL_GPL(alloc_vm_area);
464
465 void free_vm_area(struct vm_struct *area)
466 {
467         BUG();
468 }
469 EXPORT_SYMBOL_GPL(free_vm_area);
470
471 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
472                    struct page *page)
473 {
474         return -EINVAL;
475 }
476 EXPORT_SYMBOL(vm_insert_page);
477
478 /*
479  *  sys_brk() for the most part doesn't need the global kernel
480  *  lock, except when an application is doing something nasty
481  *  like trying to un-brk an area that has already been mapped
482  *  to a regular file.  in this case, the unmapping will need
483  *  to invoke file system routines that need the global lock.
484  */
485 SYSCALL_DEFINE1(brk, unsigned long, brk)
486 {
487         struct mm_struct *mm = current->mm;
488
489         if (brk < mm->start_brk || brk > mm->context.end_brk)
490                 return mm->brk;
491
492         if (mm->brk == brk)
493                 return mm->brk;
494
495         /*
496          * Always allow shrinking brk
497          */
498         if (brk <= mm->brk) {
499                 mm->brk = brk;
500                 return brk;
501         }
502
503         /*
504          * Ok, looks good - let it rip.
505          */
506         flush_icache_range(mm->brk, brk);
507         return mm->brk = brk;
508 }
509
510 /*
511  * initialise the VMA and region record slabs
512  */
513 void __init mmap_init(void)
514 {
515         int ret;
516
517         ret = percpu_counter_init(&vm_committed_as, 0);
518         VM_BUG_ON(ret);
519         vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);
520 }
521
522 /*
523  * validate the region tree
524  * - the caller must hold the region lock
525  */
526 #ifdef CONFIG_DEBUG_NOMMU_REGIONS
527 static noinline void validate_nommu_regions(void)
528 {
529         struct vm_region *region, *last;
530         struct rb_node *p, *lastp;
531
532         lastp = rb_first(&nommu_region_tree);
533         if (!lastp)
534                 return;
535
536         last = rb_entry(lastp, struct vm_region, vm_rb);
537         BUG_ON(unlikely(last->vm_end <= last->vm_start));
538         BUG_ON(unlikely(last->vm_top < last->vm_end));
539
540         while ((p = rb_next(lastp))) {
541                 region = rb_entry(p, struct vm_region, vm_rb);
542                 last = rb_entry(lastp, struct vm_region, vm_rb);
543
544                 BUG_ON(unlikely(region->vm_end <= region->vm_start));
545                 BUG_ON(unlikely(region->vm_top < region->vm_end));
546                 BUG_ON(unlikely(region->vm_start < last->vm_top));
547
548                 lastp = p;
549         }
550 }
551 #else
552 static void validate_nommu_regions(void)
553 {
554 }
555 #endif
556
557 /*
558  * add a region into the global tree
559  */
560 static void add_nommu_region(struct vm_region *region)
561 {
562         struct vm_region *pregion;
563         struct rb_node **p, *parent;
564
565         validate_nommu_regions();
566
567         parent = NULL;
568         p = &nommu_region_tree.rb_node;
569         while (*p) {
570                 parent = *p;
571                 pregion = rb_entry(parent, struct vm_region, vm_rb);
572                 if (region->vm_start < pregion->vm_start)
573                         p = &(*p)->rb_left;
574                 else if (region->vm_start > pregion->vm_start)
575                         p = &(*p)->rb_right;
576                 else if (pregion == region)
577                         return;
578                 else
579                         BUG();
580         }
581
582         rb_link_node(&region->vm_rb, parent, p);
583         rb_insert_color(&region->vm_rb, &nommu_region_tree);
584
585         validate_nommu_regions();
586 }
587
588 /*
589  * delete a region from the global tree
590  */
591 static void delete_nommu_region(struct vm_region *region)
592 {
593         BUG_ON(!nommu_region_tree.rb_node);
594
595         validate_nommu_regions();
596         rb_erase(&region->vm_rb, &nommu_region_tree);
597         validate_nommu_regions();
598 }
599
600 /*
601  * free a contiguous series of pages
602  */
603 static void free_page_series(unsigned long from, unsigned long to)
604 {
605         for (; from < to; from += PAGE_SIZE) {
606                 struct page *page = virt_to_page(from);
607
608                 kdebug("- free %lx", from);
609                 atomic_long_dec(&mmap_pages_allocated);
610                 if (page_count(page) != 1)
611                         kdebug("free page %p: refcount not one: %d",
612                                page, page_count(page));
613                 put_page(page);
614         }
615 }
616
617 /*
618  * release a reference to a region
619  * - the caller must hold the region semaphore for writing, which this releases
620  * - the region may not have been added to the tree yet, in which case vm_top
621  *   will equal vm_start
622  */
623 static void __put_nommu_region(struct vm_region *region)
624         __releases(nommu_region_sem)
625 {
626         kenter("%p{%d}", region, region->vm_usage);
627
628         BUG_ON(!nommu_region_tree.rb_node);
629
630         if (--region->vm_usage == 0) {
631                 if (region->vm_top > region->vm_start)
632                         delete_nommu_region(region);
633                 up_write(&nommu_region_sem);
634
635                 if (region->vm_file)
636                         fput(region->vm_file);
637
638                 /* IO memory and memory shared directly out of the pagecache
639                  * from ramfs/tmpfs mustn't be released here */
640                 if (region->vm_flags & VM_MAPPED_COPY) {
641                         kdebug("free series");
642                         free_page_series(region->vm_start, region->vm_top);
643                 }
644                 kmem_cache_free(vm_region_jar, region);
645         } else {
646                 up_write(&nommu_region_sem);
647         }
648 }
649
650 /*
651  * release a reference to a region
652  */
653 static void put_nommu_region(struct vm_region *region)
654 {
655         down_write(&nommu_region_sem);
656         __put_nommu_region(region);
657 }
658
659 /*
660  * update protection on a vma
661  */
662 static void protect_vma(struct vm_area_struct *vma, unsigned long flags)
663 {
664 #ifdef CONFIG_MPU
665         struct mm_struct *mm = vma->vm_mm;
666         long start = vma->vm_start & PAGE_MASK;
667         while (start < vma->vm_end) {
668                 protect_page(mm, start, flags);
669                 start += PAGE_SIZE;
670         }
671         update_protections(mm);
672 #endif
673 }
674
675 /*
676  * add a VMA into a process's mm_struct in the appropriate place in the list
677  * and tree and add to the address space's page tree also if not an anonymous
678  * page
679  * - should be called with mm->mmap_sem held writelocked
680  */
681 static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
682 {
683         struct vm_area_struct *pvma, *prev;
684         struct address_space *mapping;
685         struct rb_node **p, *parent, *rb_prev;
686
687         kenter(",%p", vma);
688
689         BUG_ON(!vma->vm_region);
690
691         mm->map_count++;
692         vma->vm_mm = mm;
693
694         protect_vma(vma, vma->vm_flags);
695
696         /* add the VMA to the mapping */
697         if (vma->vm_file) {
698                 mapping = vma->vm_file->f_mapping;
699
700                 flush_dcache_mmap_lock(mapping);
701                 vma_prio_tree_insert(vma, &mapping->i_mmap);
702                 flush_dcache_mmap_unlock(mapping);
703         }
704
705         /* add the VMA to the tree */
706         parent = rb_prev = NULL;
707         p = &mm->mm_rb.rb_node;
708         while (*p) {
709                 parent = *p;
710                 pvma = rb_entry(parent, struct vm_area_struct, vm_rb);
711
712                 /* sort by: start addr, end addr, VMA struct addr in that order
713                  * (the latter is necessary as we may get identical VMAs) */
714                 if (vma->vm_start < pvma->vm_start)
715                         p = &(*p)->rb_left;
716                 else if (vma->vm_start > pvma->vm_start) {
717                         rb_prev = parent;
718                         p = &(*p)->rb_right;
719                 } else if (vma->vm_end < pvma->vm_end)
720                         p = &(*p)->rb_left;
721                 else if (vma->vm_end > pvma->vm_end) {
722                         rb_prev = parent;
723                         p = &(*p)->rb_right;
724                 } else if (vma < pvma)
725                         p = &(*p)->rb_left;
726                 else if (vma > pvma) {
727                         rb_prev = parent;
728                         p = &(*p)->rb_right;
729                 } else
730                         BUG();
731         }
732
733         rb_link_node(&vma->vm_rb, parent, p);
734         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
735
736         /* add VMA to the VMA list also */
737         prev = NULL;
738         if (rb_prev)
739                 prev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
740
741         __vma_link_list(mm, vma, prev, parent);
742 }
743
744 /*
745  * delete a VMA from its owning mm_struct and address space
746  */
747 static void delete_vma_from_mm(struct vm_area_struct *vma)
748 {
749         struct address_space *mapping;
750         struct mm_struct *mm = vma->vm_mm;
751
752         kenter("%p", vma);
753
754         protect_vma(vma, 0);
755
756         mm->map_count--;
757         if (mm->mmap_cache == vma)
758                 mm->mmap_cache = NULL;
759
760         /* remove the VMA from the mapping */
761         if (vma->vm_file) {
762                 mapping = vma->vm_file->f_mapping;
763
764                 flush_dcache_mmap_lock(mapping);
765                 vma_prio_tree_remove(vma, &mapping->i_mmap);
766                 flush_dcache_mmap_unlock(mapping);
767         }
768
769         /* remove from the MM's tree and list */
770         rb_erase(&vma->vm_rb, &mm->mm_rb);
771
772         if (vma->vm_prev)
773                 vma->vm_prev->vm_next = vma->vm_next;
774         else
775                 mm->mmap = vma->vm_next;
776
777         if (vma->vm_next)
778                 vma->vm_next->vm_prev = vma->vm_prev;
779
780         vma->vm_mm = NULL;
781 }
782
783 /*
784  * destroy a VMA record
785  */
786 static void delete_vma(struct mm_struct *mm, struct vm_area_struct *vma)
787 {
788         kenter("%p", vma);
789         if (vma->vm_ops && vma->vm_ops->close)
790                 vma->vm_ops->close(vma);
791         if (vma->vm_file) {
792                 fput(vma->vm_file);
793                 if (vma->vm_flags & VM_EXECUTABLE)
794                         removed_exe_file_vma(mm);
795         }
796         put_nommu_region(vma->vm_region);
797         kmem_cache_free(vm_area_cachep, vma);
798 }
799
800 /*
801  * look up the first VMA in which addr resides, NULL if none
802  * - should be called with mm->mmap_sem at least held readlocked
803  */
804 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
805 {
806         struct vm_area_struct *vma;
807
808         /* check the cache first */
809         vma = mm->mmap_cache;
810         if (vma && vma->vm_start <= addr && vma->vm_end > addr)
811                 return vma;
812
813         /* trawl the list (there may be multiple mappings in which addr
814          * resides) */
815         for (vma = mm->mmap; vma; vma = vma->vm_next) {
816                 if (vma->vm_start > addr)
817                         return NULL;
818                 if (vma->vm_end > addr) {
819                         mm->mmap_cache = vma;
820                         return vma;
821                 }
822         }
823
824         return NULL;
825 }
826 EXPORT_SYMBOL(find_vma);
827
828 /*
829  * find a VMA
830  * - we don't extend stack VMAs under NOMMU conditions
831  */
832 struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr)
833 {
834         return find_vma(mm, addr);
835 }
836
837 /*
838  * expand a stack to a given address
839  * - not supported under NOMMU conditions
840  */
841 int expand_stack(struct vm_area_struct *vma, unsigned long address)
842 {
843         return -ENOMEM;
844 }
845
846 /*
847  * look up the first VMA exactly that exactly matches addr
848  * - should be called with mm->mmap_sem at least held readlocked
849  */
850 static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
851                                              unsigned long addr,
852                                              unsigned long len)
853 {
854         struct vm_area_struct *vma;
855         unsigned long end = addr + len;
856
857         /* check the cache first */
858         vma = mm->mmap_cache;
859         if (vma && vma->vm_start == addr && vma->vm_end == end)
860                 return vma;
861
862         /* trawl the list (there may be multiple mappings in which addr
863          * resides) */
864         for (vma = mm->mmap; vma; vma = vma->vm_next) {
865                 if (vma->vm_start < addr)
866                         continue;
867                 if (vma->vm_start > addr)
868                         return NULL;
869                 if (vma->vm_end == end) {
870                         mm->mmap_cache = vma;
871                         return vma;
872                 }
873         }
874
875         return NULL;
876 }
877
878 /*
879  * determine whether a mapping should be permitted and, if so, what sort of
880  * mapping we're capable of supporting
881  */
882 static int validate_mmap_request(struct file *file,
883                                  unsigned long addr,
884                                  unsigned long len,
885                                  unsigned long prot,
886                                  unsigned long flags,
887                                  unsigned long pgoff,
888                                  unsigned long *_capabilities)
889 {
890         unsigned long capabilities, rlen;
891         unsigned long reqprot = prot;
892         int ret;
893
894         /* do the simple checks first */
895         if (flags & MAP_FIXED) {
896                 printk(KERN_DEBUG
897                        "%d: Can't do fixed-address/overlay mmap of RAM\n",
898                        current->pid);
899                 return -EINVAL;
900         }
901
902         if ((flags & MAP_TYPE) != MAP_PRIVATE &&
903             (flags & MAP_TYPE) != MAP_SHARED)
904                 return -EINVAL;
905
906         if (!len)
907                 return -EINVAL;
908
909         /* Careful about overflows.. */
910         rlen = PAGE_ALIGN(len);
911         if (!rlen || rlen > TASK_SIZE)
912                 return -ENOMEM;
913
914         /* offset overflow? */
915         if ((pgoff + (rlen >> PAGE_SHIFT)) < pgoff)
916                 return -EOVERFLOW;
917
918         if (file) {
919                 /* validate file mapping requests */
920                 struct address_space *mapping;
921
922                 /* files must support mmap */
923                 if (!file->f_op || !file->f_op->mmap)
924                         return -ENODEV;
925
926                 /* work out if what we've got could possibly be shared
927                  * - we support chardevs that provide their own "memory"
928                  * - we support files/blockdevs that are memory backed
929                  */
930                 mapping = file->f_mapping;
931                 if (!mapping)
932                         mapping = file->f_path.dentry->d_inode->i_mapping;
933
934                 capabilities = 0;
935                 if (mapping && mapping->backing_dev_info)
936                         capabilities = mapping->backing_dev_info->capabilities;
937
938                 if (!capabilities) {
939                         /* no explicit capabilities set, so assume some
940                          * defaults */
941                         switch (file->f_path.dentry->d_inode->i_mode & S_IFMT) {
942                         case S_IFREG:
943                         case S_IFBLK:
944                                 capabilities = BDI_CAP_MAP_COPY;
945                                 break;
946
947                         case S_IFCHR:
948                                 capabilities =
949                                         BDI_CAP_MAP_DIRECT |
950                                         BDI_CAP_READ_MAP |
951                                         BDI_CAP_WRITE_MAP;
952                                 break;
953
954                         default:
955                                 return -EINVAL;
956                         }
957                 }
958
959                 /* eliminate any capabilities that we can't support on this
960                  * device */
961                 if (!file->f_op->get_unmapped_area)
962                         capabilities &= ~BDI_CAP_MAP_DIRECT;
963                 if (!file->f_op->read)
964                         capabilities &= ~BDI_CAP_MAP_COPY;
965
966                 /* The file shall have been opened with read permission. */
967                 if (!(file->f_mode & FMODE_READ))
968                         return -EACCES;
969
970                 if (flags & MAP_SHARED) {
971                         /* do checks for writing, appending and locking */
972                         if ((prot & PROT_WRITE) &&
973                             !(file->f_mode & FMODE_WRITE))
974                                 return -EACCES;
975
976                         if (IS_APPEND(file->f_path.dentry->d_inode) &&
977                             (file->f_mode & FMODE_WRITE))
978                                 return -EACCES;
979
980                         if (locks_verify_locked(file->f_path.dentry->d_inode))
981                                 return -EAGAIN;
982
983                         if (!(capabilities & BDI_CAP_MAP_DIRECT))
984                                 return -ENODEV;
985
986                         /* we mustn't privatise shared mappings */
987                         capabilities &= ~BDI_CAP_MAP_COPY;
988                 }
989                 else {
990                         /* we're going to read the file into private memory we
991                          * allocate */
992                         if (!(capabilities & BDI_CAP_MAP_COPY))
993                                 return -ENODEV;
994
995                         /* we don't permit a private writable mapping to be
996                          * shared with the backing device */
997                         if (prot & PROT_WRITE)
998                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
999                 }
1000
1001                 if (capabilities & BDI_CAP_MAP_DIRECT) {
1002                         if (((prot & PROT_READ)  && !(capabilities & BDI_CAP_READ_MAP))  ||
1003                             ((prot & PROT_WRITE) && !(capabilities & BDI_CAP_WRITE_MAP)) ||
1004                             ((prot & PROT_EXEC)  && !(capabilities & BDI_CAP_EXEC_MAP))
1005                             ) {
1006                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
1007                                 if (flags & MAP_SHARED) {
1008                                         printk(KERN_WARNING
1009                                                "MAP_SHARED not completely supported on !MMU\n");
1010                                         return -EINVAL;
1011                                 }
1012                         }
1013                 }
1014
1015                 /* handle executable mappings and implied executable
1016                  * mappings */
1017                 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
1018                         if (prot & PROT_EXEC)
1019                                 return -EPERM;
1020                 }
1021                 else if ((prot & PROT_READ) && !(prot & PROT_EXEC)) {
1022                         /* handle implication of PROT_EXEC by PROT_READ */
1023                         if (current->personality & READ_IMPLIES_EXEC) {
1024                                 if (capabilities & BDI_CAP_EXEC_MAP)
1025                                         prot |= PROT_EXEC;
1026                         }
1027                 }
1028                 else if ((prot & PROT_READ) &&
1029                          (prot & PROT_EXEC) &&
1030                          !(capabilities & BDI_CAP_EXEC_MAP)
1031                          ) {
1032                         /* backing file is not executable, try to copy */
1033                         capabilities &= ~BDI_CAP_MAP_DIRECT;
1034                 }
1035         }
1036         else {
1037                 /* anonymous mappings are always memory backed and can be
1038                  * privately mapped
1039                  */
1040                 capabilities = BDI_CAP_MAP_COPY;
1041
1042                 /* handle PROT_EXEC implication by PROT_READ */
1043                 if ((prot & PROT_READ) &&
1044                     (current->personality & READ_IMPLIES_EXEC))
1045                         prot |= PROT_EXEC;
1046         }
1047
1048         /* allow the security API to have its say */
1049         ret = security_file_mmap(file, reqprot, prot, flags, addr, 0);
1050         if (ret < 0)
1051                 return ret;
1052
1053         /* looks okay */
1054         *_capabilities = capabilities;
1055         return 0;
1056 }
1057
1058 /*
1059  * we've determined that we can make the mapping, now translate what we
1060  * now know into VMA flags
1061  */
1062 static unsigned long determine_vm_flags(struct file *file,
1063                                         unsigned long prot,
1064                                         unsigned long flags,
1065                                         unsigned long capabilities)
1066 {
1067         unsigned long vm_flags;
1068
1069         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags);
1070         /* vm_flags |= mm->def_flags; */
1071
1072         if (!(capabilities & BDI_CAP_MAP_DIRECT)) {
1073                 /* attempt to share read-only copies of mapped file chunks */
1074                 vm_flags |= VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1075                 if (file && !(prot & PROT_WRITE))
1076                         vm_flags |= VM_MAYSHARE;
1077         } else {
1078                 /* overlay a shareable mapping on the backing device or inode
1079                  * if possible - used for chardevs, ramfs/tmpfs/shmfs and
1080                  * romfs/cramfs */
1081                 vm_flags |= VM_MAYSHARE | (capabilities & BDI_CAP_VMFLAGS);
1082                 if (flags & MAP_SHARED)
1083                         vm_flags |= VM_SHARED;
1084         }
1085
1086         /* refuse to let anyone share private mappings with this process if
1087          * it's being traced - otherwise breakpoints set in it may interfere
1088          * with another untraced process
1089          */
1090         if ((flags & MAP_PRIVATE) && tracehook_expect_breakpoints(current))
1091                 vm_flags &= ~VM_MAYSHARE;
1092
1093         return vm_flags;
1094 }
1095
1096 /*
1097  * set up a shared mapping on a file (the driver or filesystem provides and
1098  * pins the storage)
1099  */
1100 static int do_mmap_shared_file(struct vm_area_struct *vma)
1101 {
1102         int ret;
1103
1104         ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1105         if (ret == 0) {
1106                 vma->vm_region->vm_top = vma->vm_region->vm_end;
1107                 return 0;
1108         }
1109         if (ret != -ENOSYS)
1110                 return ret;
1111
1112         /* getting -ENOSYS indicates that direct mmap isn't possible (as
1113          * opposed to tried but failed) so we can only give a suitable error as
1114          * it's not possible to make a private copy if MAP_SHARED was given */
1115         return -ENODEV;
1116 }
1117
1118 /*
1119  * set up a private mapping or an anonymous shared mapping
1120  */
1121 static int do_mmap_private(struct vm_area_struct *vma,
1122                            struct vm_region *region,
1123                            unsigned long len,
1124                            unsigned long capabilities)
1125 {
1126         struct page *pages;
1127         unsigned long total, point, n, rlen;
1128         void *base;
1129         int ret, order;
1130
1131         /* invoke the file's mapping function so that it can keep track of
1132          * shared mappings on devices or memory
1133          * - VM_MAYSHARE will be set if it may attempt to share
1134          */
1135         if (capabilities & BDI_CAP_MAP_DIRECT) {
1136                 ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1137                 if (ret == 0) {
1138                         /* shouldn't return success if we're not sharing */
1139                         BUG_ON(!(vma->vm_flags & VM_MAYSHARE));
1140                         vma->vm_region->vm_top = vma->vm_region->vm_end;
1141                         return 0;
1142                 }
1143                 if (ret != -ENOSYS)
1144                         return ret;
1145
1146                 /* getting an ENOSYS error indicates that direct mmap isn't
1147                  * possible (as opposed to tried but failed) so we'll try to
1148                  * make a private copy of the data and map that instead */
1149         }
1150
1151         rlen = PAGE_ALIGN(len);
1152
1153         /* allocate some memory to hold the mapping
1154          * - note that this may not return a page-aligned address if the object
1155          *   we're allocating is smaller than a page
1156          */
1157         order = get_order(rlen);
1158         kdebug("alloc order %d for %lx", order, len);
1159
1160         pages = alloc_pages(GFP_KERNEL, order);
1161         if (!pages)
1162                 goto enomem;
1163
1164         total = 1 << order;
1165         atomic_long_add(total, &mmap_pages_allocated);
1166
1167         point = rlen >> PAGE_SHIFT;
1168
1169         /* we allocated a power-of-2 sized page set, so we may want to trim off
1170          * the excess */
1171         if (sysctl_nr_trim_pages && total - point >= sysctl_nr_trim_pages) {
1172                 while (total > point) {
1173                         order = ilog2(total - point);
1174                         n = 1 << order;
1175                         kdebug("shave %lu/%lu @%lu", n, total - point, total);
1176                         atomic_long_sub(n, &mmap_pages_allocated);
1177                         total -= n;
1178                         set_page_refcounted(pages + total);
1179                         __free_pages(pages + total, order);
1180                 }
1181         }
1182
1183         for (point = 1; point < total; point++)
1184                 set_page_refcounted(&pages[point]);
1185
1186         base = page_address(pages);
1187         region->vm_flags = vma->vm_flags |= VM_MAPPED_COPY;
1188         region->vm_start = (unsigned long) base;
1189         region->vm_end   = region->vm_start + rlen;
1190         region->vm_top   = region->vm_start + (total << PAGE_SHIFT);
1191
1192         vma->vm_start = region->vm_start;
1193         vma->vm_end   = region->vm_start + len;
1194
1195         if (vma->vm_file) {
1196                 /* read the contents of a file into the copy */
1197                 mm_segment_t old_fs;
1198                 loff_t fpos;
1199
1200                 fpos = vma->vm_pgoff;
1201                 fpos <<= PAGE_SHIFT;
1202
1203                 old_fs = get_fs();
1204                 set_fs(KERNEL_DS);
1205                 ret = vma->vm_file->f_op->read(vma->vm_file, base, rlen, &fpos);
1206                 set_fs(old_fs);
1207
1208                 if (ret < 0)
1209                         goto error_free;
1210
1211                 /* clear the last little bit */
1212                 if (ret < rlen)
1213                         memset(base + ret, 0, rlen - ret);
1214
1215         }
1216
1217         return 0;
1218
1219 error_free:
1220         free_page_series(region->vm_start, region->vm_end);
1221         region->vm_start = vma->vm_start = 0;
1222         region->vm_end   = vma->vm_end = 0;
1223         region->vm_top   = 0;
1224         return ret;
1225
1226 enomem:
1227         printk("Allocation of length %lu from process %d (%s) failed\n",
1228                len, current->pid, current->comm);
1229         show_free_areas(0);
1230         return -ENOMEM;
1231 }
1232
1233 /*
1234  * handle mapping creation for uClinux
1235  */
1236 unsigned long do_mmap_pgoff(struct file *file,
1237                             unsigned long addr,
1238                             unsigned long len,
1239                             unsigned long prot,
1240                             unsigned long flags,
1241                             unsigned long pgoff)
1242 {
1243         struct vm_area_struct *vma;
1244         struct vm_region *region;
1245         struct rb_node *rb;
1246         unsigned long capabilities, vm_flags, result;
1247         int ret;
1248
1249         kenter(",%lx,%lx,%lx,%lx,%lx", addr, len, prot, flags, pgoff);
1250
1251         /* decide whether we should attempt the mapping, and if so what sort of
1252          * mapping */
1253         ret = validate_mmap_request(file, addr, len, prot, flags, pgoff,
1254                                     &capabilities);
1255         if (ret < 0) {
1256                 kleave(" = %d [val]", ret);
1257                 return ret;
1258         }
1259
1260         /* we ignore the address hint */
1261         addr = 0;
1262
1263         /* we've determined that we can make the mapping, now translate what we
1264          * now know into VMA flags */
1265         vm_flags = determine_vm_flags(file, prot, flags, capabilities);
1266
1267         /* we're going to need to record the mapping */
1268         region = kmem_cache_zalloc(vm_region_jar, GFP_KERNEL);
1269         if (!region)
1270                 goto error_getting_region;
1271
1272         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1273         if (!vma)
1274                 goto error_getting_vma;
1275
1276         region->vm_usage = 1;
1277         region->vm_flags = vm_flags;
1278         region->vm_pgoff = pgoff;
1279
1280         INIT_LIST_HEAD(&vma->anon_vma_chain);
1281         vma->vm_flags = vm_flags;
1282         vma->vm_pgoff = pgoff;
1283
1284         if (file) {
1285                 region->vm_file = file;
1286                 get_file(file);
1287                 vma->vm_file = file;
1288                 get_file(file);
1289                 if (vm_flags & VM_EXECUTABLE) {
1290                         added_exe_file_vma(current->mm);
1291                         vma->vm_mm = current->mm;
1292                 }
1293         }
1294
1295         down_write(&nommu_region_sem);
1296
1297         /* if we want to share, we need to check for regions created by other
1298          * mmap() calls that overlap with our proposed mapping
1299          * - we can only share with a superset match on most regular files
1300          * - shared mappings on character devices and memory backed files are
1301          *   permitted to overlap inexactly as far as we are concerned for in
1302          *   these cases, sharing is handled in the driver or filesystem rather
1303          *   than here
1304          */
1305         if (vm_flags & VM_MAYSHARE) {
1306                 struct vm_region *pregion;
1307                 unsigned long pglen, rpglen, pgend, rpgend, start;
1308
1309                 pglen = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1310                 pgend = pgoff + pglen;
1311
1312                 for (rb = rb_first(&nommu_region_tree); rb; rb = rb_next(rb)) {
1313                         pregion = rb_entry(rb, struct vm_region, vm_rb);
1314
1315                         if (!(pregion->vm_flags & VM_MAYSHARE))
1316                                 continue;
1317
1318                         /* search for overlapping mappings on the same file */
1319                         if (pregion->vm_file->f_path.dentry->d_inode !=
1320                             file->f_path.dentry->d_inode)
1321                                 continue;
1322
1323                         if (pregion->vm_pgoff >= pgend)
1324                                 continue;
1325
1326                         rpglen = pregion->vm_end - pregion->vm_start;
1327                         rpglen = (rpglen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1328                         rpgend = pregion->vm_pgoff + rpglen;
1329                         if (pgoff >= rpgend)
1330                                 continue;
1331
1332                         /* handle inexactly overlapping matches between
1333                          * mappings */
1334                         if ((pregion->vm_pgoff != pgoff || rpglen != pglen) &&
1335                             !(pgoff >= pregion->vm_pgoff && pgend <= rpgend)) {
1336                                 /* new mapping is not a subset of the region */
1337                                 if (!(capabilities & BDI_CAP_MAP_DIRECT))
1338                                         goto sharing_violation;
1339                                 continue;
1340                         }
1341
1342                         /* we've found a region we can share */
1343                         pregion->vm_usage++;
1344                         vma->vm_region = pregion;
1345                         start = pregion->vm_start;
1346                         start += (pgoff - pregion->vm_pgoff) << PAGE_SHIFT;
1347                         vma->vm_start = start;
1348                         vma->vm_end = start + len;
1349
1350                         if (pregion->vm_flags & VM_MAPPED_COPY) {
1351                                 kdebug("share copy");
1352                                 vma->vm_flags |= VM_MAPPED_COPY;
1353                         } else {
1354                                 kdebug("share mmap");
1355                                 ret = do_mmap_shared_file(vma);
1356                                 if (ret < 0) {
1357                                         vma->vm_region = NULL;
1358                                         vma->vm_start = 0;
1359                                         vma->vm_end = 0;
1360                                         pregion->vm_usage--;
1361                                         pregion = NULL;
1362                                         goto error_just_free;
1363                                 }
1364                         }
1365                         fput(region->vm_file);
1366                         kmem_cache_free(vm_region_jar, region);
1367                         region = pregion;
1368                         result = start;
1369                         goto share;
1370                 }
1371
1372                 /* obtain the address at which to make a shared mapping
1373                  * - this is the hook for quasi-memory character devices to
1374                  *   tell us the location of a shared mapping
1375                  */
1376                 if (capabilities & BDI_CAP_MAP_DIRECT) {
1377                         addr = file->f_op->get_unmapped_area(file, addr, len,
1378                                                              pgoff, flags);
1379                         if (IS_ERR((void *) addr)) {
1380                                 ret = addr;
1381                                 if (ret != (unsigned long) -ENOSYS)
1382                                         goto error_just_free;
1383
1384                                 /* the driver refused to tell us where to site
1385                                  * the mapping so we'll have to attempt to copy
1386                                  * it */
1387                                 ret = (unsigned long) -ENODEV;
1388                                 if (!(capabilities & BDI_CAP_MAP_COPY))
1389                                         goto error_just_free;
1390
1391                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
1392                         } else {
1393                                 vma->vm_start = region->vm_start = addr;
1394                                 vma->vm_end = region->vm_end = addr + len;
1395                         }
1396                 }
1397         }
1398
1399         vma->vm_region = region;
1400
1401         /* set up the mapping
1402          * - the region is filled in if BDI_CAP_MAP_DIRECT is still set
1403          */
1404         if (file && vma->vm_flags & VM_SHARED)
1405                 ret = do_mmap_shared_file(vma);
1406         else
1407                 ret = do_mmap_private(vma, region, len, capabilities);
1408         if (ret < 0)
1409                 goto error_just_free;
1410         add_nommu_region(region);
1411
1412         /* clear anonymous mappings that don't ask for uninitialized data */
1413         if (!vma->vm_file && !(flags & MAP_UNINITIALIZED))
1414                 memset((void *)region->vm_start, 0,
1415                        region->vm_end - region->vm_start);
1416
1417         /* okay... we have a mapping; now we have to register it */
1418         result = vma->vm_start;
1419
1420         current->mm->total_vm += len >> PAGE_SHIFT;
1421
1422 share:
1423         add_vma_to_mm(current->mm, vma);
1424
1425         /* we flush the region from the icache only when the first executable
1426          * mapping of it is made  */
1427         if (vma->vm_flags & VM_EXEC && !region->vm_icache_flushed) {
1428                 flush_icache_range(region->vm_start, region->vm_end);
1429                 region->vm_icache_flushed = true;
1430         }
1431
1432         up_write(&nommu_region_sem);
1433
1434         kleave(" = %lx", result);
1435         return result;
1436
1437 error_just_free:
1438         up_write(&nommu_region_sem);
1439 error:
1440         if (region->vm_file)
1441                 fput(region->vm_file);
1442         kmem_cache_free(vm_region_jar, region);
1443         if (vma->vm_file)
1444                 fput(vma->vm_file);
1445         if (vma->vm_flags & VM_EXECUTABLE)
1446                 removed_exe_file_vma(vma->vm_mm);
1447         kmem_cache_free(vm_area_cachep, vma);
1448         kleave(" = %d", ret);
1449         return ret;
1450
1451 sharing_violation:
1452         up_write(&nommu_region_sem);
1453         printk(KERN_WARNING "Attempt to share mismatched mappings\n");
1454         ret = -EINVAL;
1455         goto error;
1456
1457 error_getting_vma:
1458         kmem_cache_free(vm_region_jar, region);
1459         printk(KERN_WARNING "Allocation of vma for %lu byte allocation"
1460                " from process %d failed\n",
1461                len, current->pid);
1462         show_free_areas(0);
1463         return -ENOMEM;
1464
1465 error_getting_region:
1466         printk(KERN_WARNING "Allocation of vm region for %lu byte allocation"
1467                " from process %d failed\n",
1468                len, current->pid);
1469         show_free_areas(0);
1470         return -ENOMEM;
1471 }
1472 EXPORT_SYMBOL(do_mmap_pgoff);
1473
1474 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1475                 unsigned long, prot, unsigned long, flags,
1476                 unsigned long, fd, unsigned long, pgoff)
1477 {
1478         struct file *file = NULL;
1479         unsigned long retval = -EBADF;
1480
1481         audit_mmap_fd(fd, flags);
1482         if (!(flags & MAP_ANONYMOUS)) {
1483                 file = fget(fd);
1484                 if (!file)
1485                         goto out;
1486         }
1487
1488         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1489
1490         down_write(&current->mm->mmap_sem);
1491         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1492         up_write(&current->mm->mmap_sem);
1493
1494         if (file)
1495                 fput(file);
1496 out:
1497         return retval;
1498 }
1499
1500 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1501 struct mmap_arg_struct {
1502         unsigned long addr;
1503         unsigned long len;
1504         unsigned long prot;
1505         unsigned long flags;
1506         unsigned long fd;
1507         unsigned long offset;
1508 };
1509
1510 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1511 {
1512         struct mmap_arg_struct a;
1513
1514         if (copy_from_user(&a, arg, sizeof(a)))
1515                 return -EFAULT;
1516         if (a.offset & ~PAGE_MASK)
1517                 return -EINVAL;
1518
1519         return sys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1520                               a.offset >> PAGE_SHIFT);
1521 }
1522 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1523
1524 /*
1525  * split a vma into two pieces at address 'addr', a new vma is allocated either
1526  * for the first part or the tail.
1527  */
1528 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1529               unsigned long addr, int new_below)
1530 {
1531         struct vm_area_struct *new;
1532         struct vm_region *region;
1533         unsigned long npages;
1534
1535         kenter("");
1536
1537         /* we're only permitted to split anonymous regions (these should have
1538          * only a single usage on the region) */
1539         if (vma->vm_file)
1540                 return -ENOMEM;
1541
1542         if (mm->map_count >= sysctl_max_map_count)
1543                 return -ENOMEM;
1544
1545         region = kmem_cache_alloc(vm_region_jar, GFP_KERNEL);
1546         if (!region)
1547                 return -ENOMEM;
1548
1549         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1550         if (!new) {
1551                 kmem_cache_free(vm_region_jar, region);
1552                 return -ENOMEM;
1553         }
1554
1555         /* most fields are the same, copy all, and then fixup */
1556         *new = *vma;
1557         *region = *vma->vm_region;
1558         new->vm_region = region;
1559
1560         npages = (addr - vma->vm_start) >> PAGE_SHIFT;
1561
1562         if (new_below) {
1563                 region->vm_top = region->vm_end = new->vm_end = addr;
1564         } else {
1565                 region->vm_start = new->vm_start = addr;
1566                 region->vm_pgoff = new->vm_pgoff += npages;
1567         }
1568
1569         if (new->vm_ops && new->vm_ops->open)
1570                 new->vm_ops->open(new);
1571
1572         delete_vma_from_mm(vma);
1573         down_write(&nommu_region_sem);
1574         delete_nommu_region(vma->vm_region);
1575         if (new_below) {
1576                 vma->vm_region->vm_start = vma->vm_start = addr;
1577                 vma->vm_region->vm_pgoff = vma->vm_pgoff += npages;
1578         } else {
1579                 vma->vm_region->vm_end = vma->vm_end = addr;
1580                 vma->vm_region->vm_top = addr;
1581         }
1582         add_nommu_region(vma->vm_region);
1583         add_nommu_region(new->vm_region);
1584         up_write(&nommu_region_sem);
1585         add_vma_to_mm(mm, vma);
1586         add_vma_to_mm(mm, new);
1587         return 0;
1588 }
1589
1590 /*
1591  * shrink a VMA by removing the specified chunk from either the beginning or
1592  * the end
1593  */
1594 static int shrink_vma(struct mm_struct *mm,
1595                       struct vm_area_struct *vma,
1596                       unsigned long from, unsigned long to)
1597 {
1598         struct vm_region *region;
1599
1600         kenter("");
1601
1602         /* adjust the VMA's pointers, which may reposition it in the MM's tree
1603          * and list */
1604         delete_vma_from_mm(vma);
1605         if (from > vma->vm_start)
1606                 vma->vm_end = from;
1607         else
1608                 vma->vm_start = to;
1609         add_vma_to_mm(mm, vma);
1610
1611         /* cut the backing region down to size */
1612         region = vma->vm_region;
1613         BUG_ON(region->vm_usage != 1);
1614
1615         down_write(&nommu_region_sem);
1616         delete_nommu_region(region);
1617         if (from > region->vm_start) {
1618                 to = region->vm_top;
1619                 region->vm_top = region->vm_end = from;
1620         } else {
1621                 region->vm_start = to;
1622         }
1623         add_nommu_region(region);
1624         up_write(&nommu_region_sem);
1625
1626         free_page_series(from, to);
1627         return 0;
1628 }
1629
1630 /*
1631  * release a mapping
1632  * - under NOMMU conditions the chunk to be unmapped must be backed by a single
1633  *   VMA, though it need not cover the whole VMA
1634  */
1635 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1636 {
1637         struct vm_area_struct *vma;
1638         struct rb_node *rb;
1639         unsigned long end = start + len;
1640         int ret;
1641
1642         kenter(",%lx,%zx", start, len);
1643
1644         if (len == 0)
1645                 return -EINVAL;
1646
1647         /* find the first potentially overlapping VMA */
1648         vma = find_vma(mm, start);
1649         if (!vma) {
1650                 static int limit = 0;
1651                 if (limit < 5) {
1652                         printk(KERN_WARNING
1653                                "munmap of memory not mmapped by process %d"
1654                                " (%s): 0x%lx-0x%lx\n",
1655                                current->pid, current->comm,
1656                                start, start + len - 1);
1657                         limit++;
1658                 }
1659                 return -EINVAL;
1660         }
1661
1662         /* we're allowed to split an anonymous VMA but not a file-backed one */
1663         if (vma->vm_file) {
1664                 do {
1665                         if (start > vma->vm_start) {
1666                                 kleave(" = -EINVAL [miss]");
1667                                 return -EINVAL;
1668                         }
1669                         if (end == vma->vm_end)
1670                                 goto erase_whole_vma;
1671                         rb = rb_next(&vma->vm_rb);
1672                         vma = rb_entry(rb, struct vm_area_struct, vm_rb);
1673                 } while (rb);
1674                 kleave(" = -EINVAL [split file]");
1675                 return -EINVAL;
1676         } else {
1677                 /* the chunk must be a subset of the VMA found */
1678                 if (start == vma->vm_start && end == vma->vm_end)
1679                         goto erase_whole_vma;
1680                 if (start < vma->vm_start || end > vma->vm_end) {
1681                         kleave(" = -EINVAL [superset]");
1682                         return -EINVAL;
1683                 }
1684                 if (start & ~PAGE_MASK) {
1685                         kleave(" = -EINVAL [unaligned start]");
1686                         return -EINVAL;
1687                 }
1688                 if (end != vma->vm_end && end & ~PAGE_MASK) {
1689                         kleave(" = -EINVAL [unaligned split]");
1690                         return -EINVAL;
1691                 }
1692                 if (start != vma->vm_start && end != vma->vm_end) {
1693                         ret = split_vma(mm, vma, start, 1);
1694                         if (ret < 0) {
1695                                 kleave(" = %d [split]", ret);
1696                                 return ret;
1697                         }
1698                 }
1699                 return shrink_vma(mm, vma, start, end);
1700         }
1701
1702 erase_whole_vma:
1703         delete_vma_from_mm(vma);
1704         delete_vma(mm, vma);
1705         kleave(" = 0");
1706         return 0;
1707 }
1708 EXPORT_SYMBOL(do_munmap);
1709
1710 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1711 {
1712         int ret;
1713         struct mm_struct *mm = current->mm;
1714
1715         down_write(&mm->mmap_sem);
1716         ret = do_munmap(mm, addr, len);
1717         up_write(&mm->mmap_sem);
1718         return ret;
1719 }
1720
1721 /*
1722  * release all the mappings made in a process's VM space
1723  */
1724 void exit_mmap(struct mm_struct *mm)
1725 {
1726         struct vm_area_struct *vma;
1727
1728         if (!mm)
1729                 return;
1730
1731         kenter("");
1732
1733         mm->total_vm = 0;
1734
1735         while ((vma = mm->mmap)) {
1736                 mm->mmap = vma->vm_next;
1737                 delete_vma_from_mm(vma);
1738                 delete_vma(mm, vma);
1739                 cond_resched();
1740         }
1741
1742         kleave("");
1743 }
1744
1745 unsigned long do_brk(unsigned long addr, unsigned long len)
1746 {
1747         return -ENOMEM;
1748 }
1749
1750 /*
1751  * expand (or shrink) an existing mapping, potentially moving it at the same
1752  * time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1753  *
1754  * under NOMMU conditions, we only permit changing a mapping's size, and only
1755  * as long as it stays within the region allocated by do_mmap_private() and the
1756  * block is not shareable
1757  *
1758  * MREMAP_FIXED is not supported under NOMMU conditions
1759  */
1760 unsigned long do_mremap(unsigned long addr,
1761                         unsigned long old_len, unsigned long new_len,
1762                         unsigned long flags, unsigned long new_addr)
1763 {
1764         struct vm_area_struct *vma;
1765
1766         /* insanity checks first */
1767         if (old_len == 0 || new_len == 0)
1768                 return (unsigned long) -EINVAL;
1769
1770         if (addr & ~PAGE_MASK)
1771                 return -EINVAL;
1772
1773         if (flags & MREMAP_FIXED && new_addr != addr)
1774                 return (unsigned long) -EINVAL;
1775
1776         vma = find_vma_exact(current->mm, addr, old_len);
1777         if (!vma)
1778                 return (unsigned long) -EINVAL;
1779
1780         if (vma->vm_end != vma->vm_start + old_len)
1781                 return (unsigned long) -EFAULT;
1782
1783         if (vma->vm_flags & VM_MAYSHARE)
1784                 return (unsigned long) -EPERM;
1785
1786         if (new_len > vma->vm_region->vm_end - vma->vm_region->vm_start)
1787                 return (unsigned long) -ENOMEM;
1788
1789         /* all checks complete - do it */
1790         vma->vm_end = vma->vm_start + new_len;
1791         return vma->vm_start;
1792 }
1793 EXPORT_SYMBOL(do_mremap);
1794
1795 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1796                 unsigned long, new_len, unsigned long, flags,
1797                 unsigned long, new_addr)
1798 {
1799         unsigned long ret;
1800
1801         down_write(&current->mm->mmap_sem);
1802         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
1803         up_write(&current->mm->mmap_sem);
1804         return ret;
1805 }
1806
1807 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1808                         unsigned int foll_flags)
1809 {
1810         return NULL;
1811 }
1812
1813 int remap_pfn_range(struct vm_area_struct *vma, unsigned long from,
1814                 unsigned long to, unsigned long size, pgprot_t prot)
1815 {
1816         vma->vm_start = vma->vm_pgoff << PAGE_SHIFT;
1817         return 0;
1818 }
1819 EXPORT_SYMBOL(remap_pfn_range);
1820
1821 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1822                         unsigned long pgoff)
1823 {
1824         unsigned int size = vma->vm_end - vma->vm_start;
1825
1826         if (!(vma->vm_flags & VM_USERMAP))
1827                 return -EINVAL;
1828
1829         vma->vm_start = (unsigned long)(addr + (pgoff << PAGE_SHIFT));
1830         vma->vm_end = vma->vm_start + size;
1831
1832         return 0;
1833 }
1834 EXPORT_SYMBOL(remap_vmalloc_range);
1835
1836 unsigned long arch_get_unmapped_area(struct file *file, unsigned long addr,
1837         unsigned long len, unsigned long pgoff, unsigned long flags)
1838 {
1839         return -ENOMEM;
1840 }
1841
1842 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1843 {
1844 }
1845
1846 void unmap_mapping_range(struct address_space *mapping,
1847                          loff_t const holebegin, loff_t const holelen,
1848                          int even_cows)
1849 {
1850 }
1851 EXPORT_SYMBOL(unmap_mapping_range);
1852
1853 /*
1854  * Check that a process has enough memory to allocate a new virtual
1855  * mapping. 0 means there is enough memory for the allocation to
1856  * succeed and -ENOMEM implies there is not.
1857  *
1858  * We currently support three overcommit policies, which are set via the
1859  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
1860  *
1861  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
1862  * Additional code 2002 Jul 20 by Robert Love.
1863  *
1864  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
1865  *
1866  * Note this is a helper function intended to be used by LSMs which
1867  * wish to use this logic.
1868  */
1869 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
1870 {
1871         unsigned long free, allowed;
1872
1873         vm_acct_memory(pages);
1874
1875         /*
1876          * Sometimes we want to use more memory than we have
1877          */
1878         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
1879                 return 0;
1880
1881         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
1882                 unsigned long n;
1883
1884                 free = global_page_state(NR_FILE_PAGES);
1885                 free += nr_swap_pages;
1886
1887                 /*
1888                  * Any slabs which are created with the
1889                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
1890                  * which are reclaimable, under pressure.  The dentry
1891                  * cache and most inode caches should fall into this
1892                  */
1893                 free += global_page_state(NR_SLAB_RECLAIMABLE);
1894
1895                 /*
1896                  * Leave the last 3% for root
1897                  */
1898                 if (!cap_sys_admin)
1899                         free -= free / 32;
1900
1901                 if (free > pages)
1902                         return 0;
1903
1904                 /*
1905                  * nr_free_pages() is very expensive on large systems,
1906                  * only call if we're about to fail.
1907                  */
1908                 n = nr_free_pages();
1909
1910                 /*
1911                  * Leave reserved pages. The pages are not for anonymous pages.
1912                  */
1913                 if (n <= totalreserve_pages)
1914                         goto error;
1915                 else
1916                         n -= totalreserve_pages;
1917
1918                 /*
1919                  * Leave the last 3% for root
1920                  */
1921                 if (!cap_sys_admin)
1922                         n -= n / 32;
1923                 free += n;
1924
1925                 if (free > pages)
1926                         return 0;
1927
1928                 goto error;
1929         }
1930
1931         allowed = totalram_pages * sysctl_overcommit_ratio / 100;
1932         /*
1933          * Leave the last 3% for root
1934          */
1935         if (!cap_sys_admin)
1936                 allowed -= allowed / 32;
1937         allowed += total_swap_pages;
1938
1939         /* Don't let a single process grow too big:
1940            leave 3% of the size of this process for other processes */
1941         if (mm)
1942                 allowed -= mm->total_vm / 32;
1943
1944         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
1945                 return 0;
1946
1947 error:
1948         vm_unacct_memory(pages);
1949
1950         return -ENOMEM;
1951 }
1952
1953 int in_gate_area_no_mm(unsigned long addr)
1954 {
1955         return 0;
1956 }
1957
1958 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1959 {
1960         BUG();
1961         return 0;
1962 }
1963 EXPORT_SYMBOL(filemap_fault);
1964
1965 static int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
1966                 unsigned long addr, void *buf, int len, int write)
1967 {
1968         struct vm_area_struct *vma;
1969
1970         down_read(&mm->mmap_sem);
1971
1972         /* the access must start within one of the target process's mappings */
1973         vma = find_vma(mm, addr);
1974         if (vma) {
1975                 /* don't overrun this mapping */
1976                 if (addr + len >= vma->vm_end)
1977                         len = vma->vm_end - addr;
1978
1979                 /* only read or write mappings where it is permitted */
1980                 if (write && vma->vm_flags & VM_MAYWRITE)
1981                         copy_to_user_page(vma, NULL, addr,
1982                                          (void *) addr, buf, len);
1983                 else if (!write && vma->vm_flags & VM_MAYREAD)
1984                         copy_from_user_page(vma, NULL, addr,
1985                                             buf, (void *) addr, len);
1986                 else
1987                         len = 0;
1988         } else {
1989                 len = 0;
1990         }
1991
1992         up_read(&mm->mmap_sem);
1993
1994         return len;
1995 }
1996
1997 /**
1998  * @access_remote_vm - access another process' address space
1999  * @mm:         the mm_struct of the target address space
2000  * @addr:       start address to access
2001  * @buf:        source or destination buffer
2002  * @len:        number of bytes to transfer
2003  * @write:      whether the access is a write
2004  *
2005  * The caller must hold a reference on @mm.
2006  */
2007 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
2008                 void *buf, int len, int write)
2009 {
2010         return __access_remote_vm(NULL, mm, addr, buf, len, write);
2011 }
2012
2013 /*
2014  * Access another process' address space.
2015  * - source/target buffer must be kernel space
2016  */
2017 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
2018 {
2019         struct mm_struct *mm;
2020
2021         if (addr + len < addr)
2022                 return 0;
2023
2024         mm = get_task_mm(tsk);
2025         if (!mm)
2026                 return 0;
2027
2028         len = __access_remote_vm(tsk, mm, addr, buf, len, write);
2029
2030         mmput(mm);
2031         return len;
2032 }
2033
2034 /**
2035  * nommu_shrink_inode_mappings - Shrink the shared mappings on an inode
2036  * @inode: The inode to check
2037  * @size: The current filesize of the inode
2038  * @newsize: The proposed filesize of the inode
2039  *
2040  * Check the shared mappings on an inode on behalf of a shrinking truncate to
2041  * make sure that that any outstanding VMAs aren't broken and then shrink the
2042  * vm_regions that extend that beyond so that do_mmap_pgoff() doesn't
2043  * automatically grant mappings that are too large.
2044  */
2045 int nommu_shrink_inode_mappings(struct inode *inode, size_t size,
2046                                 size_t newsize)
2047 {
2048         struct vm_area_struct *vma;
2049         struct prio_tree_iter iter;
2050         struct vm_region *region;
2051         pgoff_t low, high;
2052         size_t r_size, r_top;
2053
2054         low = newsize >> PAGE_SHIFT;
2055         high = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
2056
2057         down_write(&nommu_region_sem);
2058
2059         /* search for VMAs that fall within the dead zone */
2060         vma_prio_tree_foreach(vma, &iter, &inode->i_mapping->i_mmap,
2061                               low, high) {
2062                 /* found one - only interested if it's shared out of the page
2063                  * cache */
2064                 if (vma->vm_flags & VM_SHARED) {
2065                         up_write(&nommu_region_sem);
2066                         return -ETXTBSY; /* not quite true, but near enough */
2067                 }
2068         }
2069
2070         /* reduce any regions that overlap the dead zone - if in existence,
2071          * these will be pointed to by VMAs that don't overlap the dead zone
2072          *
2073          * we don't check for any regions that start beyond the EOF as there
2074          * shouldn't be any
2075          */
2076         vma_prio_tree_foreach(vma, &iter, &inode->i_mapping->i_mmap,
2077                               0, ULONG_MAX) {
2078                 if (!(vma->vm_flags & VM_SHARED))
2079                         continue;
2080
2081                 region = vma->vm_region;
2082                 r_size = region->vm_top - region->vm_start;
2083                 r_top = (region->vm_pgoff << PAGE_SHIFT) + r_size;
2084
2085                 if (r_top > newsize) {
2086                         region->vm_top -= r_top - newsize;
2087                         if (region->vm_end > region->vm_top)
2088                                 region->vm_end = region->vm_top;
2089                 }
2090         }
2091
2092         up_write(&nommu_region_sem);
2093         return 0;
2094 }