mm/gup: cache dev_pagemap while pinning pages
[muen/linux.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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17
18 #include <linux/export.h>
19 #include <linux/mm.h>
20 #include <linux/sched/mm.h>
21 #include <linux/vmacache.h>
22 #include <linux/mman.h>
23 #include <linux/swap.h>
24 #include <linux/file.h>
25 #include <linux/highmem.h>
26 #include <linux/pagemap.h>
27 #include <linux/slab.h>
28 #include <linux/vmalloc.h>
29 #include <linux/blkdev.h>
30 #include <linux/backing-dev.h>
31 #include <linux/compiler.h>
32 #include <linux/mount.h>
33 #include <linux/personality.h>
34 #include <linux/security.h>
35 #include <linux/syscalls.h>
36 #include <linux/audit.h>
37 #include <linux/printk.h>
38
39 #include <linux/uaccess.h>
40 #include <asm/tlb.h>
41 #include <asm/tlbflush.h>
42 #include <asm/mmu_context.h>
43 #include "internal.h"
44
45 void *high_memory;
46 EXPORT_SYMBOL(high_memory);
47 struct page *mem_map;
48 unsigned long max_mapnr;
49 EXPORT_SYMBOL(max_mapnr);
50 unsigned long highest_memmap_pfn;
51 int sysctl_nr_trim_pages = CONFIG_NOMMU_INITIAL_TRIM_EXCESS;
52 int heap_stack_gap = 0;
53
54 atomic_long_t mmap_pages_allocated;
55
56 EXPORT_SYMBOL(mem_map);
57
58 /* list of mapped, potentially shareable regions */
59 static struct kmem_cache *vm_region_jar;
60 struct rb_root nommu_region_tree = RB_ROOT;
61 DECLARE_RWSEM(nommu_region_sem);
62
63 const struct vm_operations_struct generic_file_vm_ops = {
64 };
65
66 /*
67  * Return the total memory allocated for this pointer, not
68  * just what the caller asked for.
69  *
70  * Doesn't have to be accurate, i.e. may have races.
71  */
72 unsigned int kobjsize(const void *objp)
73 {
74         struct page *page;
75
76         /*
77          * If the object we have should not have ksize performed on it,
78          * return size of 0
79          */
80         if (!objp || !virt_addr_valid(objp))
81                 return 0;
82
83         page = virt_to_head_page(objp);
84
85         /*
86          * If the allocator sets PageSlab, we know the pointer came from
87          * kmalloc().
88          */
89         if (PageSlab(page))
90                 return ksize(objp);
91
92         /*
93          * If it's not a compound page, see if we have a matching VMA
94          * region. This test is intentionally done in reverse order,
95          * so if there's no VMA, we still fall through and hand back
96          * PAGE_SIZE for 0-order pages.
97          */
98         if (!PageCompound(page)) {
99                 struct vm_area_struct *vma;
100
101                 vma = find_vma(current->mm, (unsigned long)objp);
102                 if (vma)
103                         return vma->vm_end - vma->vm_start;
104         }
105
106         /*
107          * The ksize() function is only guaranteed to work for pointers
108          * returned by kmalloc(). So handle arbitrary pointers here.
109          */
110         return PAGE_SIZE << compound_order(page);
111 }
112
113 static long __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
114                       unsigned long start, unsigned long nr_pages,
115                       unsigned int foll_flags, struct page **pages,
116                       struct vm_area_struct **vmas, int *nonblocking)
117 {
118         struct vm_area_struct *vma;
119         unsigned long vm_flags;
120         int i;
121
122         /* calculate required read or write permissions.
123          * If FOLL_FORCE is set, we only require the "MAY" flags.
124          */
125         vm_flags  = (foll_flags & FOLL_WRITE) ?
126                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
127         vm_flags &= (foll_flags & FOLL_FORCE) ?
128                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
129
130         for (i = 0; i < nr_pages; i++) {
131                 vma = find_vma(mm, start);
132                 if (!vma)
133                         goto finish_or_fault;
134
135                 /* protect what we can, including chardevs */
136                 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
137                     !(vm_flags & vma->vm_flags))
138                         goto finish_or_fault;
139
140                 if (pages) {
141                         pages[i] = virt_to_page(start);
142                         if (pages[i])
143                                 get_page(pages[i]);
144                 }
145                 if (vmas)
146                         vmas[i] = vma;
147                 start = (start + PAGE_SIZE) & PAGE_MASK;
148         }
149
150         return i;
151
152 finish_or_fault:
153         return i ? : -EFAULT;
154 }
155
156 /*
157  * get a list of pages in an address range belonging to the specified process
158  * and indicate the VMA that covers each page
159  * - this is potentially dodgy as we may end incrementing the page count of a
160  *   slab page or a secondary page from a compound page
161  * - don't permit access to VMAs that don't support it, such as I/O mappings
162  */
163 long get_user_pages(unsigned long start, unsigned long nr_pages,
164                     unsigned int gup_flags, struct page **pages,
165                     struct vm_area_struct **vmas)
166 {
167         return __get_user_pages(current, current->mm, start, nr_pages,
168                                 gup_flags, pages, vmas, NULL);
169 }
170 EXPORT_SYMBOL(get_user_pages);
171
172 long get_user_pages_locked(unsigned long start, unsigned long nr_pages,
173                             unsigned int gup_flags, struct page **pages,
174                             int *locked)
175 {
176         return get_user_pages(start, nr_pages, gup_flags, pages, NULL);
177 }
178 EXPORT_SYMBOL(get_user_pages_locked);
179
180 static long __get_user_pages_unlocked(struct task_struct *tsk,
181                         struct mm_struct *mm, unsigned long start,
182                         unsigned long nr_pages, struct page **pages,
183                         unsigned int gup_flags)
184 {
185         long ret;
186         down_read(&mm->mmap_sem);
187         ret = __get_user_pages(tsk, mm, start, nr_pages, gup_flags, pages,
188                                 NULL, NULL);
189         up_read(&mm->mmap_sem);
190         return ret;
191 }
192
193 long get_user_pages_unlocked(unsigned long start, unsigned long nr_pages,
194                              struct page **pages, unsigned int gup_flags)
195 {
196         return __get_user_pages_unlocked(current, current->mm, start, nr_pages,
197                                          pages, gup_flags);
198 }
199 EXPORT_SYMBOL(get_user_pages_unlocked);
200
201 /**
202  * follow_pfn - look up PFN at a user virtual address
203  * @vma: memory mapping
204  * @address: user virtual address
205  * @pfn: location to store found PFN
206  *
207  * Only IO mappings and raw PFN mappings are allowed.
208  *
209  * Returns zero and the pfn at @pfn on success, -ve otherwise.
210  */
211 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
212         unsigned long *pfn)
213 {
214         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
215                 return -EINVAL;
216
217         *pfn = address >> PAGE_SHIFT;
218         return 0;
219 }
220 EXPORT_SYMBOL(follow_pfn);
221
222 LIST_HEAD(vmap_area_list);
223
224 void vfree(const void *addr)
225 {
226         kfree(addr);
227 }
228 EXPORT_SYMBOL(vfree);
229
230 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
231 {
232         /*
233          *  You can't specify __GFP_HIGHMEM with kmalloc() since kmalloc()
234          * returns only a logical address.
235          */
236         return kmalloc(size, (gfp_mask | __GFP_COMP) & ~__GFP_HIGHMEM);
237 }
238 EXPORT_SYMBOL(__vmalloc);
239
240 void *__vmalloc_node_flags(unsigned long size, int node, gfp_t flags)
241 {
242         return __vmalloc(size, flags, PAGE_KERNEL);
243 }
244
245 void *vmalloc_user(unsigned long size)
246 {
247         void *ret;
248
249         ret = __vmalloc(size, GFP_KERNEL | __GFP_ZERO, PAGE_KERNEL);
250         if (ret) {
251                 struct vm_area_struct *vma;
252
253                 down_write(&current->mm->mmap_sem);
254                 vma = find_vma(current->mm, (unsigned long)ret);
255                 if (vma)
256                         vma->vm_flags |= VM_USERMAP;
257                 up_write(&current->mm->mmap_sem);
258         }
259
260         return ret;
261 }
262 EXPORT_SYMBOL(vmalloc_user);
263
264 struct page *vmalloc_to_page(const void *addr)
265 {
266         return virt_to_page(addr);
267 }
268 EXPORT_SYMBOL(vmalloc_to_page);
269
270 unsigned long vmalloc_to_pfn(const void *addr)
271 {
272         return page_to_pfn(virt_to_page(addr));
273 }
274 EXPORT_SYMBOL(vmalloc_to_pfn);
275
276 long vread(char *buf, char *addr, unsigned long count)
277 {
278         /* Don't allow overflow */
279         if ((unsigned long) buf + count < count)
280                 count = -(unsigned long) buf;
281
282         memcpy(buf, addr, count);
283         return count;
284 }
285
286 long vwrite(char *buf, char *addr, unsigned long count)
287 {
288         /* Don't allow overflow */
289         if ((unsigned long) addr + count < count)
290                 count = -(unsigned long) addr;
291
292         memcpy(addr, buf, count);
293         return count;
294 }
295
296 /*
297  *      vmalloc  -  allocate virtually contiguous memory
298  *
299  *      @size:          allocation size
300  *
301  *      Allocate enough pages to cover @size from the page level
302  *      allocator and map them into contiguous kernel virtual space.
303  *
304  *      For tight control over page level allocator and protection flags
305  *      use __vmalloc() instead.
306  */
307 void *vmalloc(unsigned long size)
308 {
309        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
310 }
311 EXPORT_SYMBOL(vmalloc);
312
313 /*
314  *      vzalloc - allocate virtually contiguous memory with zero fill
315  *
316  *      @size:          allocation size
317  *
318  *      Allocate enough pages to cover @size from the page level
319  *      allocator and map them into contiguous kernel virtual space.
320  *      The memory allocated is set to zero.
321  *
322  *      For tight control over page level allocator and protection flags
323  *      use __vmalloc() instead.
324  */
325 void *vzalloc(unsigned long size)
326 {
327         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
328                         PAGE_KERNEL);
329 }
330 EXPORT_SYMBOL(vzalloc);
331
332 /**
333  * vmalloc_node - allocate memory on a specific node
334  * @size:       allocation size
335  * @node:       numa node
336  *
337  * Allocate enough pages to cover @size from the page level
338  * allocator and map them into contiguous kernel virtual space.
339  *
340  * For tight control over page level allocator and protection flags
341  * use __vmalloc() instead.
342  */
343 void *vmalloc_node(unsigned long size, int node)
344 {
345         return vmalloc(size);
346 }
347 EXPORT_SYMBOL(vmalloc_node);
348
349 /**
350  * vzalloc_node - allocate memory on a specific node with zero fill
351  * @size:       allocation size
352  * @node:       numa node
353  *
354  * Allocate enough pages to cover @size from the page level
355  * allocator and map them into contiguous kernel virtual space.
356  * The memory allocated is set to zero.
357  *
358  * For tight control over page level allocator and protection flags
359  * use __vmalloc() instead.
360  */
361 void *vzalloc_node(unsigned long size, int node)
362 {
363         return vzalloc(size);
364 }
365 EXPORT_SYMBOL(vzalloc_node);
366
367 /**
368  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
369  *      @size:          allocation size
370  *
371  *      Kernel-internal function to allocate enough pages to cover @size
372  *      the page level allocator and map them into contiguous and
373  *      executable kernel virtual space.
374  *
375  *      For tight control over page level allocator and protection flags
376  *      use __vmalloc() instead.
377  */
378
379 void *vmalloc_exec(unsigned long size)
380 {
381         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
382 }
383
384 /**
385  * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
386  *      @size:          allocation size
387  *
388  *      Allocate enough 32bit PA addressable pages to cover @size from the
389  *      page level allocator and map them into contiguous kernel virtual space.
390  */
391 void *vmalloc_32(unsigned long size)
392 {
393         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
394 }
395 EXPORT_SYMBOL(vmalloc_32);
396
397 /**
398  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
399  *      @size:          allocation size
400  *
401  * The resulting memory area is 32bit addressable and zeroed so it can be
402  * mapped to userspace without leaking data.
403  *
404  * VM_USERMAP is set on the corresponding VMA so that subsequent calls to
405  * remap_vmalloc_range() are permissible.
406  */
407 void *vmalloc_32_user(unsigned long size)
408 {
409         /*
410          * We'll have to sort out the ZONE_DMA bits for 64-bit,
411          * but for now this can simply use vmalloc_user() directly.
412          */
413         return vmalloc_user(size);
414 }
415 EXPORT_SYMBOL(vmalloc_32_user);
416
417 void *vmap(struct page **pages, unsigned int count, unsigned long flags, pgprot_t prot)
418 {
419         BUG();
420         return NULL;
421 }
422 EXPORT_SYMBOL(vmap);
423
424 void vunmap(const void *addr)
425 {
426         BUG();
427 }
428 EXPORT_SYMBOL(vunmap);
429
430 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
431 {
432         BUG();
433         return NULL;
434 }
435 EXPORT_SYMBOL(vm_map_ram);
436
437 void vm_unmap_ram(const void *mem, unsigned int count)
438 {
439         BUG();
440 }
441 EXPORT_SYMBOL(vm_unmap_ram);
442
443 void vm_unmap_aliases(void)
444 {
445 }
446 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
447
448 /*
449  * Implement a stub for vmalloc_sync_all() if the architecture chose not to
450  * have one.
451  */
452 void __weak vmalloc_sync_all(void)
453 {
454 }
455
456 struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
457 {
458         BUG();
459         return NULL;
460 }
461 EXPORT_SYMBOL_GPL(alloc_vm_area);
462
463 void free_vm_area(struct vm_struct *area)
464 {
465         BUG();
466 }
467 EXPORT_SYMBOL_GPL(free_vm_area);
468
469 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
470                    struct page *page)
471 {
472         return -EINVAL;
473 }
474 EXPORT_SYMBOL(vm_insert_page);
475
476 /*
477  *  sys_brk() for the most part doesn't need the global kernel
478  *  lock, except when an application is doing something nasty
479  *  like trying to un-brk an area that has already been mapped
480  *  to a regular file.  in this case, the unmapping will need
481  *  to invoke file system routines that need the global lock.
482  */
483 SYSCALL_DEFINE1(brk, unsigned long, brk)
484 {
485         struct mm_struct *mm = current->mm;
486
487         if (brk < mm->start_brk || brk > mm->context.end_brk)
488                 return mm->brk;
489
490         if (mm->brk == brk)
491                 return mm->brk;
492
493         /*
494          * Always allow shrinking brk
495          */
496         if (brk <= mm->brk) {
497                 mm->brk = brk;
498                 return brk;
499         }
500
501         /*
502          * Ok, looks good - let it rip.
503          */
504         flush_icache_range(mm->brk, brk);
505         return mm->brk = brk;
506 }
507
508 /*
509  * initialise the percpu counter for VM and region record slabs
510  */
511 void __init mmap_init(void)
512 {
513         int ret;
514
515         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
516         VM_BUG_ON(ret);
517         vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC|SLAB_ACCOUNT);
518 }
519
520 /*
521  * validate the region tree
522  * - the caller must hold the region lock
523  */
524 #ifdef CONFIG_DEBUG_NOMMU_REGIONS
525 static noinline void validate_nommu_regions(void)
526 {
527         struct vm_region *region, *last;
528         struct rb_node *p, *lastp;
529
530         lastp = rb_first(&nommu_region_tree);
531         if (!lastp)
532                 return;
533
534         last = rb_entry(lastp, struct vm_region, vm_rb);
535         BUG_ON(last->vm_end <= last->vm_start);
536         BUG_ON(last->vm_top < last->vm_end);
537
538         while ((p = rb_next(lastp))) {
539                 region = rb_entry(p, struct vm_region, vm_rb);
540                 last = rb_entry(lastp, struct vm_region, vm_rb);
541
542                 BUG_ON(region->vm_end <= region->vm_start);
543                 BUG_ON(region->vm_top < region->vm_end);
544                 BUG_ON(region->vm_start < last->vm_top);
545
546                 lastp = p;
547         }
548 }
549 #else
550 static void validate_nommu_regions(void)
551 {
552 }
553 #endif
554
555 /*
556  * add a region into the global tree
557  */
558 static void add_nommu_region(struct vm_region *region)
559 {
560         struct vm_region *pregion;
561         struct rb_node **p, *parent;
562
563         validate_nommu_regions();
564
565         parent = NULL;
566         p = &nommu_region_tree.rb_node;
567         while (*p) {
568                 parent = *p;
569                 pregion = rb_entry(parent, struct vm_region, vm_rb);
570                 if (region->vm_start < pregion->vm_start)
571                         p = &(*p)->rb_left;
572                 else if (region->vm_start > pregion->vm_start)
573                         p = &(*p)->rb_right;
574                 else if (pregion == region)
575                         return;
576                 else
577                         BUG();
578         }
579
580         rb_link_node(&region->vm_rb, parent, p);
581         rb_insert_color(&region->vm_rb, &nommu_region_tree);
582
583         validate_nommu_regions();
584 }
585
586 /*
587  * delete a region from the global tree
588  */
589 static void delete_nommu_region(struct vm_region *region)
590 {
591         BUG_ON(!nommu_region_tree.rb_node);
592
593         validate_nommu_regions();
594         rb_erase(&region->vm_rb, &nommu_region_tree);
595         validate_nommu_regions();
596 }
597
598 /*
599  * free a contiguous series of pages
600  */
601 static void free_page_series(unsigned long from, unsigned long to)
602 {
603         for (; from < to; from += PAGE_SIZE) {
604                 struct page *page = virt_to_page(from);
605
606                 atomic_long_dec(&mmap_pages_allocated);
607                 put_page(page);
608         }
609 }
610
611 /*
612  * release a reference to a region
613  * - the caller must hold the region semaphore for writing, which this releases
614  * - the region may not have been added to the tree yet, in which case vm_top
615  *   will equal vm_start
616  */
617 static void __put_nommu_region(struct vm_region *region)
618         __releases(nommu_region_sem)
619 {
620         BUG_ON(!nommu_region_tree.rb_node);
621
622         if (--region->vm_usage == 0) {
623                 if (region->vm_top > region->vm_start)
624                         delete_nommu_region(region);
625                 up_write(&nommu_region_sem);
626
627                 if (region->vm_file)
628                         fput(region->vm_file);
629
630                 /* IO memory and memory shared directly out of the pagecache
631                  * from ramfs/tmpfs mustn't be released here */
632                 if (region->vm_flags & VM_MAPPED_COPY)
633                         free_page_series(region->vm_start, region->vm_top);
634                 kmem_cache_free(vm_region_jar, region);
635         } else {
636                 up_write(&nommu_region_sem);
637         }
638 }
639
640 /*
641  * release a reference to a region
642  */
643 static void put_nommu_region(struct vm_region *region)
644 {
645         down_write(&nommu_region_sem);
646         __put_nommu_region(region);
647 }
648
649 /*
650  * add a VMA into a process's mm_struct in the appropriate place in the list
651  * and tree and add to the address space's page tree also if not an anonymous
652  * page
653  * - should be called with mm->mmap_sem held writelocked
654  */
655 static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
656 {
657         struct vm_area_struct *pvma, *prev;
658         struct address_space *mapping;
659         struct rb_node **p, *parent, *rb_prev;
660
661         BUG_ON(!vma->vm_region);
662
663         mm->map_count++;
664         vma->vm_mm = mm;
665
666         /* add the VMA to the mapping */
667         if (vma->vm_file) {
668                 mapping = vma->vm_file->f_mapping;
669
670                 i_mmap_lock_write(mapping);
671                 flush_dcache_mmap_lock(mapping);
672                 vma_interval_tree_insert(vma, &mapping->i_mmap);
673                 flush_dcache_mmap_unlock(mapping);
674                 i_mmap_unlock_write(mapping);
675         }
676
677         /* add the VMA to the tree */
678         parent = rb_prev = NULL;
679         p = &mm->mm_rb.rb_node;
680         while (*p) {
681                 parent = *p;
682                 pvma = rb_entry(parent, struct vm_area_struct, vm_rb);
683
684                 /* sort by: start addr, end addr, VMA struct addr in that order
685                  * (the latter is necessary as we may get identical VMAs) */
686                 if (vma->vm_start < pvma->vm_start)
687                         p = &(*p)->rb_left;
688                 else if (vma->vm_start > pvma->vm_start) {
689                         rb_prev = parent;
690                         p = &(*p)->rb_right;
691                 } else if (vma->vm_end < pvma->vm_end)
692                         p = &(*p)->rb_left;
693                 else if (vma->vm_end > pvma->vm_end) {
694                         rb_prev = parent;
695                         p = &(*p)->rb_right;
696                 } else if (vma < pvma)
697                         p = &(*p)->rb_left;
698                 else if (vma > pvma) {
699                         rb_prev = parent;
700                         p = &(*p)->rb_right;
701                 } else
702                         BUG();
703         }
704
705         rb_link_node(&vma->vm_rb, parent, p);
706         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
707
708         /* add VMA to the VMA list also */
709         prev = NULL;
710         if (rb_prev)
711                 prev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
712
713         __vma_link_list(mm, vma, prev, parent);
714 }
715
716 /*
717  * delete a VMA from its owning mm_struct and address space
718  */
719 static void delete_vma_from_mm(struct vm_area_struct *vma)
720 {
721         int i;
722         struct address_space *mapping;
723         struct mm_struct *mm = vma->vm_mm;
724         struct task_struct *curr = current;
725
726         mm->map_count--;
727         for (i = 0; i < VMACACHE_SIZE; i++) {
728                 /* if the vma is cached, invalidate the entire cache */
729                 if (curr->vmacache.vmas[i] == vma) {
730                         vmacache_invalidate(mm);
731                         break;
732                 }
733         }
734
735         /* remove the VMA from the mapping */
736         if (vma->vm_file) {
737                 mapping = vma->vm_file->f_mapping;
738
739                 i_mmap_lock_write(mapping);
740                 flush_dcache_mmap_lock(mapping);
741                 vma_interval_tree_remove(vma, &mapping->i_mmap);
742                 flush_dcache_mmap_unlock(mapping);
743                 i_mmap_unlock_write(mapping);
744         }
745
746         /* remove from the MM's tree and list */
747         rb_erase(&vma->vm_rb, &mm->mm_rb);
748
749         if (vma->vm_prev)
750                 vma->vm_prev->vm_next = vma->vm_next;
751         else
752                 mm->mmap = vma->vm_next;
753
754         if (vma->vm_next)
755                 vma->vm_next->vm_prev = vma->vm_prev;
756 }
757
758 /*
759  * destroy a VMA record
760  */
761 static void delete_vma(struct mm_struct *mm, struct vm_area_struct *vma)
762 {
763         if (vma->vm_ops && vma->vm_ops->close)
764                 vma->vm_ops->close(vma);
765         if (vma->vm_file)
766                 fput(vma->vm_file);
767         put_nommu_region(vma->vm_region);
768         vm_area_free(vma);
769 }
770
771 /*
772  * look up the first VMA in which addr resides, NULL if none
773  * - should be called with mm->mmap_sem at least held readlocked
774  */
775 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
776 {
777         struct vm_area_struct *vma;
778
779         /* check the cache first */
780         vma = vmacache_find(mm, addr);
781         if (likely(vma))
782                 return vma;
783
784         /* trawl the list (there may be multiple mappings in which addr
785          * resides) */
786         for (vma = mm->mmap; vma; vma = vma->vm_next) {
787                 if (vma->vm_start > addr)
788                         return NULL;
789                 if (vma->vm_end > addr) {
790                         vmacache_update(addr, vma);
791                         return vma;
792                 }
793         }
794
795         return NULL;
796 }
797 EXPORT_SYMBOL(find_vma);
798
799 /*
800  * find a VMA
801  * - we don't extend stack VMAs under NOMMU conditions
802  */
803 struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr)
804 {
805         return find_vma(mm, addr);
806 }
807
808 /*
809  * expand a stack to a given address
810  * - not supported under NOMMU conditions
811  */
812 int expand_stack(struct vm_area_struct *vma, unsigned long address)
813 {
814         return -ENOMEM;
815 }
816
817 /*
818  * look up the first VMA exactly that exactly matches addr
819  * - should be called with mm->mmap_sem at least held readlocked
820  */
821 static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
822                                              unsigned long addr,
823                                              unsigned long len)
824 {
825         struct vm_area_struct *vma;
826         unsigned long end = addr + len;
827
828         /* check the cache first */
829         vma = vmacache_find_exact(mm, addr, end);
830         if (vma)
831                 return vma;
832
833         /* trawl the list (there may be multiple mappings in which addr
834          * resides) */
835         for (vma = mm->mmap; vma; vma = vma->vm_next) {
836                 if (vma->vm_start < addr)
837                         continue;
838                 if (vma->vm_start > addr)
839                         return NULL;
840                 if (vma->vm_end == end) {
841                         vmacache_update(addr, vma);
842                         return vma;
843                 }
844         }
845
846         return NULL;
847 }
848
849 /*
850  * determine whether a mapping should be permitted and, if so, what sort of
851  * mapping we're capable of supporting
852  */
853 static int validate_mmap_request(struct file *file,
854                                  unsigned long addr,
855                                  unsigned long len,
856                                  unsigned long prot,
857                                  unsigned long flags,
858                                  unsigned long pgoff,
859                                  unsigned long *_capabilities)
860 {
861         unsigned long capabilities, rlen;
862         int ret;
863
864         /* do the simple checks first */
865         if (flags & MAP_FIXED)
866                 return -EINVAL;
867
868         if ((flags & MAP_TYPE) != MAP_PRIVATE &&
869             (flags & MAP_TYPE) != MAP_SHARED)
870                 return -EINVAL;
871
872         if (!len)
873                 return -EINVAL;
874
875         /* Careful about overflows.. */
876         rlen = PAGE_ALIGN(len);
877         if (!rlen || rlen > TASK_SIZE)
878                 return -ENOMEM;
879
880         /* offset overflow? */
881         if ((pgoff + (rlen >> PAGE_SHIFT)) < pgoff)
882                 return -EOVERFLOW;
883
884         if (file) {
885                 /* files must support mmap */
886                 if (!file->f_op->mmap)
887                         return -ENODEV;
888
889                 /* work out if what we've got could possibly be shared
890                  * - we support chardevs that provide their own "memory"
891                  * - we support files/blockdevs that are memory backed
892                  */
893                 if (file->f_op->mmap_capabilities) {
894                         capabilities = file->f_op->mmap_capabilities(file);
895                 } else {
896                         /* no explicit capabilities set, so assume some
897                          * defaults */
898                         switch (file_inode(file)->i_mode & S_IFMT) {
899                         case S_IFREG:
900                         case S_IFBLK:
901                                 capabilities = NOMMU_MAP_COPY;
902                                 break;
903
904                         case S_IFCHR:
905                                 capabilities =
906                                         NOMMU_MAP_DIRECT |
907                                         NOMMU_MAP_READ |
908                                         NOMMU_MAP_WRITE;
909                                 break;
910
911                         default:
912                                 return -EINVAL;
913                         }
914                 }
915
916                 /* eliminate any capabilities that we can't support on this
917                  * device */
918                 if (!file->f_op->get_unmapped_area)
919                         capabilities &= ~NOMMU_MAP_DIRECT;
920                 if (!(file->f_mode & FMODE_CAN_READ))
921                         capabilities &= ~NOMMU_MAP_COPY;
922
923                 /* The file shall have been opened with read permission. */
924                 if (!(file->f_mode & FMODE_READ))
925                         return -EACCES;
926
927                 if (flags & MAP_SHARED) {
928                         /* do checks for writing, appending and locking */
929                         if ((prot & PROT_WRITE) &&
930                             !(file->f_mode & FMODE_WRITE))
931                                 return -EACCES;
932
933                         if (IS_APPEND(file_inode(file)) &&
934                             (file->f_mode & FMODE_WRITE))
935                                 return -EACCES;
936
937                         if (locks_verify_locked(file))
938                                 return -EAGAIN;
939
940                         if (!(capabilities & NOMMU_MAP_DIRECT))
941                                 return -ENODEV;
942
943                         /* we mustn't privatise shared mappings */
944                         capabilities &= ~NOMMU_MAP_COPY;
945                 } else {
946                         /* we're going to read the file into private memory we
947                          * allocate */
948                         if (!(capabilities & NOMMU_MAP_COPY))
949                                 return -ENODEV;
950
951                         /* we don't permit a private writable mapping to be
952                          * shared with the backing device */
953                         if (prot & PROT_WRITE)
954                                 capabilities &= ~NOMMU_MAP_DIRECT;
955                 }
956
957                 if (capabilities & NOMMU_MAP_DIRECT) {
958                         if (((prot & PROT_READ)  && !(capabilities & NOMMU_MAP_READ))  ||
959                             ((prot & PROT_WRITE) && !(capabilities & NOMMU_MAP_WRITE)) ||
960                             ((prot & PROT_EXEC)  && !(capabilities & NOMMU_MAP_EXEC))
961                             ) {
962                                 capabilities &= ~NOMMU_MAP_DIRECT;
963                                 if (flags & MAP_SHARED) {
964                                         pr_warn("MAP_SHARED not completely supported on !MMU\n");
965                                         return -EINVAL;
966                                 }
967                         }
968                 }
969
970                 /* handle executable mappings and implied executable
971                  * mappings */
972                 if (path_noexec(&file->f_path)) {
973                         if (prot & PROT_EXEC)
974                                 return -EPERM;
975                 } else if ((prot & PROT_READ) && !(prot & PROT_EXEC)) {
976                         /* handle implication of PROT_EXEC by PROT_READ */
977                         if (current->personality & READ_IMPLIES_EXEC) {
978                                 if (capabilities & NOMMU_MAP_EXEC)
979                                         prot |= PROT_EXEC;
980                         }
981                 } else if ((prot & PROT_READ) &&
982                          (prot & PROT_EXEC) &&
983                          !(capabilities & NOMMU_MAP_EXEC)
984                          ) {
985                         /* backing file is not executable, try to copy */
986                         capabilities &= ~NOMMU_MAP_DIRECT;
987                 }
988         } else {
989                 /* anonymous mappings are always memory backed and can be
990                  * privately mapped
991                  */
992                 capabilities = NOMMU_MAP_COPY;
993
994                 /* handle PROT_EXEC implication by PROT_READ */
995                 if ((prot & PROT_READ) &&
996                     (current->personality & READ_IMPLIES_EXEC))
997                         prot |= PROT_EXEC;
998         }
999
1000         /* allow the security API to have its say */
1001         ret = security_mmap_addr(addr);
1002         if (ret < 0)
1003                 return ret;
1004
1005         /* looks okay */
1006         *_capabilities = capabilities;
1007         return 0;
1008 }
1009
1010 /*
1011  * we've determined that we can make the mapping, now translate what we
1012  * now know into VMA flags
1013  */
1014 static unsigned long determine_vm_flags(struct file *file,
1015                                         unsigned long prot,
1016                                         unsigned long flags,
1017                                         unsigned long capabilities)
1018 {
1019         unsigned long vm_flags;
1020
1021         vm_flags = calc_vm_prot_bits(prot, 0) | calc_vm_flag_bits(flags);
1022         /* vm_flags |= mm->def_flags; */
1023
1024         if (!(capabilities & NOMMU_MAP_DIRECT)) {
1025                 /* attempt to share read-only copies of mapped file chunks */
1026                 vm_flags |= VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1027                 if (file && !(prot & PROT_WRITE))
1028                         vm_flags |= VM_MAYSHARE;
1029         } else {
1030                 /* overlay a shareable mapping on the backing device or inode
1031                  * if possible - used for chardevs, ramfs/tmpfs/shmfs and
1032                  * romfs/cramfs */
1033                 vm_flags |= VM_MAYSHARE | (capabilities & NOMMU_VMFLAGS);
1034                 if (flags & MAP_SHARED)
1035                         vm_flags |= VM_SHARED;
1036         }
1037
1038         /* refuse to let anyone share private mappings with this process if
1039          * it's being traced - otherwise breakpoints set in it may interfere
1040          * with another untraced process
1041          */
1042         if ((flags & MAP_PRIVATE) && current->ptrace)
1043                 vm_flags &= ~VM_MAYSHARE;
1044
1045         return vm_flags;
1046 }
1047
1048 /*
1049  * set up a shared mapping on a file (the driver or filesystem provides and
1050  * pins the storage)
1051  */
1052 static int do_mmap_shared_file(struct vm_area_struct *vma)
1053 {
1054         int ret;
1055
1056         ret = call_mmap(vma->vm_file, vma);
1057         if (ret == 0) {
1058                 vma->vm_region->vm_top = vma->vm_region->vm_end;
1059                 return 0;
1060         }
1061         if (ret != -ENOSYS)
1062                 return ret;
1063
1064         /* getting -ENOSYS indicates that direct mmap isn't possible (as
1065          * opposed to tried but failed) so we can only give a suitable error as
1066          * it's not possible to make a private copy if MAP_SHARED was given */
1067         return -ENODEV;
1068 }
1069
1070 /*
1071  * set up a private mapping or an anonymous shared mapping
1072  */
1073 static int do_mmap_private(struct vm_area_struct *vma,
1074                            struct vm_region *region,
1075                            unsigned long len,
1076                            unsigned long capabilities)
1077 {
1078         unsigned long total, point;
1079         void *base;
1080         int ret, order;
1081
1082         /* invoke the file's mapping function so that it can keep track of
1083          * shared mappings on devices or memory
1084          * - VM_MAYSHARE will be set if it may attempt to share
1085          */
1086         if (capabilities & NOMMU_MAP_DIRECT) {
1087                 ret = call_mmap(vma->vm_file, vma);
1088                 if (ret == 0) {
1089                         /* shouldn't return success if we're not sharing */
1090                         BUG_ON(!(vma->vm_flags & VM_MAYSHARE));
1091                         vma->vm_region->vm_top = vma->vm_region->vm_end;
1092                         return 0;
1093                 }
1094                 if (ret != -ENOSYS)
1095                         return ret;
1096
1097                 /* getting an ENOSYS error indicates that direct mmap isn't
1098                  * possible (as opposed to tried but failed) so we'll try to
1099                  * make a private copy of the data and map that instead */
1100         }
1101
1102
1103         /* allocate some memory to hold the mapping
1104          * - note that this may not return a page-aligned address if the object
1105          *   we're allocating is smaller than a page
1106          */
1107         order = get_order(len);
1108         total = 1 << order;
1109         point = len >> PAGE_SHIFT;
1110
1111         /* we don't want to allocate a power-of-2 sized page set */
1112         if (sysctl_nr_trim_pages && total - point >= sysctl_nr_trim_pages)
1113                 total = point;
1114
1115         base = alloc_pages_exact(total << PAGE_SHIFT, GFP_KERNEL);
1116         if (!base)
1117                 goto enomem;
1118
1119         atomic_long_add(total, &mmap_pages_allocated);
1120
1121         region->vm_flags = vma->vm_flags |= VM_MAPPED_COPY;
1122         region->vm_start = (unsigned long) base;
1123         region->vm_end   = region->vm_start + len;
1124         region->vm_top   = region->vm_start + (total << PAGE_SHIFT);
1125
1126         vma->vm_start = region->vm_start;
1127         vma->vm_end   = region->vm_start + len;
1128
1129         if (vma->vm_file) {
1130                 /* read the contents of a file into the copy */
1131                 loff_t fpos;
1132
1133                 fpos = vma->vm_pgoff;
1134                 fpos <<= PAGE_SHIFT;
1135
1136                 ret = kernel_read(vma->vm_file, base, len, &fpos);
1137                 if (ret < 0)
1138                         goto error_free;
1139
1140                 /* clear the last little bit */
1141                 if (ret < len)
1142                         memset(base + ret, 0, len - ret);
1143
1144         } else {
1145                 vma_set_anonymous(vma);
1146         }
1147
1148         return 0;
1149
1150 error_free:
1151         free_page_series(region->vm_start, region->vm_top);
1152         region->vm_start = vma->vm_start = 0;
1153         region->vm_end   = vma->vm_end = 0;
1154         region->vm_top   = 0;
1155         return ret;
1156
1157 enomem:
1158         pr_err("Allocation of length %lu from process %d (%s) failed\n",
1159                len, current->pid, current->comm);
1160         show_free_areas(0, NULL);
1161         return -ENOMEM;
1162 }
1163
1164 /*
1165  * handle mapping creation for uClinux
1166  */
1167 unsigned long do_mmap(struct file *file,
1168                         unsigned long addr,
1169                         unsigned long len,
1170                         unsigned long prot,
1171                         unsigned long flags,
1172                         vm_flags_t vm_flags,
1173                         unsigned long pgoff,
1174                         unsigned long *populate,
1175                         struct list_head *uf)
1176 {
1177         struct vm_area_struct *vma;
1178         struct vm_region *region;
1179         struct rb_node *rb;
1180         unsigned long capabilities, result;
1181         int ret;
1182
1183         *populate = 0;
1184
1185         /* decide whether we should attempt the mapping, and if so what sort of
1186          * mapping */
1187         ret = validate_mmap_request(file, addr, len, prot, flags, pgoff,
1188                                     &capabilities);
1189         if (ret < 0)
1190                 return ret;
1191
1192         /* we ignore the address hint */
1193         addr = 0;
1194         len = PAGE_ALIGN(len);
1195
1196         /* we've determined that we can make the mapping, now translate what we
1197          * now know into VMA flags */
1198         vm_flags |= determine_vm_flags(file, prot, flags, capabilities);
1199
1200         /* we're going to need to record the mapping */
1201         region = kmem_cache_zalloc(vm_region_jar, GFP_KERNEL);
1202         if (!region)
1203                 goto error_getting_region;
1204
1205         vma = vm_area_alloc(current->mm);
1206         if (!vma)
1207                 goto error_getting_vma;
1208
1209         region->vm_usage = 1;
1210         region->vm_flags = vm_flags;
1211         region->vm_pgoff = pgoff;
1212
1213         vma->vm_flags = vm_flags;
1214         vma->vm_pgoff = pgoff;
1215
1216         if (file) {
1217                 region->vm_file = get_file(file);
1218                 vma->vm_file = get_file(file);
1219         }
1220
1221         down_write(&nommu_region_sem);
1222
1223         /* if we want to share, we need to check for regions created by other
1224          * mmap() calls that overlap with our proposed mapping
1225          * - we can only share with a superset match on most regular files
1226          * - shared mappings on character devices and memory backed files are
1227          *   permitted to overlap inexactly as far as we are concerned for in
1228          *   these cases, sharing is handled in the driver or filesystem rather
1229          *   than here
1230          */
1231         if (vm_flags & VM_MAYSHARE) {
1232                 struct vm_region *pregion;
1233                 unsigned long pglen, rpglen, pgend, rpgend, start;
1234
1235                 pglen = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1236                 pgend = pgoff + pglen;
1237
1238                 for (rb = rb_first(&nommu_region_tree); rb; rb = rb_next(rb)) {
1239                         pregion = rb_entry(rb, struct vm_region, vm_rb);
1240
1241                         if (!(pregion->vm_flags & VM_MAYSHARE))
1242                                 continue;
1243
1244                         /* search for overlapping mappings on the same file */
1245                         if (file_inode(pregion->vm_file) !=
1246                             file_inode(file))
1247                                 continue;
1248
1249                         if (pregion->vm_pgoff >= pgend)
1250                                 continue;
1251
1252                         rpglen = pregion->vm_end - pregion->vm_start;
1253                         rpglen = (rpglen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1254                         rpgend = pregion->vm_pgoff + rpglen;
1255                         if (pgoff >= rpgend)
1256                                 continue;
1257
1258                         /* handle inexactly overlapping matches between
1259                          * mappings */
1260                         if ((pregion->vm_pgoff != pgoff || rpglen != pglen) &&
1261                             !(pgoff >= pregion->vm_pgoff && pgend <= rpgend)) {
1262                                 /* new mapping is not a subset of the region */
1263                                 if (!(capabilities & NOMMU_MAP_DIRECT))
1264                                         goto sharing_violation;
1265                                 continue;
1266                         }
1267
1268                         /* we've found a region we can share */
1269                         pregion->vm_usage++;
1270                         vma->vm_region = pregion;
1271                         start = pregion->vm_start;
1272                         start += (pgoff - pregion->vm_pgoff) << PAGE_SHIFT;
1273                         vma->vm_start = start;
1274                         vma->vm_end = start + len;
1275
1276                         if (pregion->vm_flags & VM_MAPPED_COPY)
1277                                 vma->vm_flags |= VM_MAPPED_COPY;
1278                         else {
1279                                 ret = do_mmap_shared_file(vma);
1280                                 if (ret < 0) {
1281                                         vma->vm_region = NULL;
1282                                         vma->vm_start = 0;
1283                                         vma->vm_end = 0;
1284                                         pregion->vm_usage--;
1285                                         pregion = NULL;
1286                                         goto error_just_free;
1287                                 }
1288                         }
1289                         fput(region->vm_file);
1290                         kmem_cache_free(vm_region_jar, region);
1291                         region = pregion;
1292                         result = start;
1293                         goto share;
1294                 }
1295
1296                 /* obtain the address at which to make a shared mapping
1297                  * - this is the hook for quasi-memory character devices to
1298                  *   tell us the location of a shared mapping
1299                  */
1300                 if (capabilities & NOMMU_MAP_DIRECT) {
1301                         addr = file->f_op->get_unmapped_area(file, addr, len,
1302                                                              pgoff, flags);
1303                         if (IS_ERR_VALUE(addr)) {
1304                                 ret = addr;
1305                                 if (ret != -ENOSYS)
1306                                         goto error_just_free;
1307
1308                                 /* the driver refused to tell us where to site
1309                                  * the mapping so we'll have to attempt to copy
1310                                  * it */
1311                                 ret = -ENODEV;
1312                                 if (!(capabilities & NOMMU_MAP_COPY))
1313                                         goto error_just_free;
1314
1315                                 capabilities &= ~NOMMU_MAP_DIRECT;
1316                         } else {
1317                                 vma->vm_start = region->vm_start = addr;
1318                                 vma->vm_end = region->vm_end = addr + len;
1319                         }
1320                 }
1321         }
1322
1323         vma->vm_region = region;
1324
1325         /* set up the mapping
1326          * - the region is filled in if NOMMU_MAP_DIRECT is still set
1327          */
1328         if (file && vma->vm_flags & VM_SHARED)
1329                 ret = do_mmap_shared_file(vma);
1330         else
1331                 ret = do_mmap_private(vma, region, len, capabilities);
1332         if (ret < 0)
1333                 goto error_just_free;
1334         add_nommu_region(region);
1335
1336         /* clear anonymous mappings that don't ask for uninitialized data */
1337         if (!vma->vm_file && !(flags & MAP_UNINITIALIZED))
1338                 memset((void *)region->vm_start, 0,
1339                        region->vm_end - region->vm_start);
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         /* we flush the region from the icache only when the first executable
1350          * mapping of it is made  */
1351         if (vma->vm_flags & VM_EXEC && !region->vm_icache_flushed) {
1352                 flush_icache_range(region->vm_start, region->vm_end);
1353                 region->vm_icache_flushed = true;
1354         }
1355
1356         up_write(&nommu_region_sem);
1357
1358         return result;
1359
1360 error_just_free:
1361         up_write(&nommu_region_sem);
1362 error:
1363         if (region->vm_file)
1364                 fput(region->vm_file);
1365         kmem_cache_free(vm_region_jar, region);
1366         if (vma->vm_file)
1367                 fput(vma->vm_file);
1368         vm_area_free(vma);
1369         return ret;
1370
1371 sharing_violation:
1372         up_write(&nommu_region_sem);
1373         pr_warn("Attempt to share mismatched mappings\n");
1374         ret = -EINVAL;
1375         goto error;
1376
1377 error_getting_vma:
1378         kmem_cache_free(vm_region_jar, region);
1379         pr_warn("Allocation of vma for %lu byte allocation from process %d failed\n",
1380                         len, current->pid);
1381         show_free_areas(0, NULL);
1382         return -ENOMEM;
1383
1384 error_getting_region:
1385         pr_warn("Allocation of vm region for %lu byte allocation from process %d failed\n",
1386                         len, current->pid);
1387         show_free_areas(0, NULL);
1388         return -ENOMEM;
1389 }
1390
1391 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1392                               unsigned long prot, unsigned long flags,
1393                               unsigned long fd, unsigned long pgoff)
1394 {
1395         struct file *file = NULL;
1396         unsigned long retval = -EBADF;
1397
1398         audit_mmap_fd(fd, flags);
1399         if (!(flags & MAP_ANONYMOUS)) {
1400                 file = fget(fd);
1401                 if (!file)
1402                         goto out;
1403         }
1404
1405         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1406
1407         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1408
1409         if (file)
1410                 fput(file);
1411 out:
1412         return retval;
1413 }
1414
1415 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1416                 unsigned long, prot, unsigned long, flags,
1417                 unsigned long, fd, unsigned long, pgoff)
1418 {
1419         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1420 }
1421
1422 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1423 struct mmap_arg_struct {
1424         unsigned long addr;
1425         unsigned long len;
1426         unsigned long prot;
1427         unsigned long flags;
1428         unsigned long fd;
1429         unsigned long offset;
1430 };
1431
1432 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1433 {
1434         struct mmap_arg_struct a;
1435
1436         if (copy_from_user(&a, arg, sizeof(a)))
1437                 return -EFAULT;
1438         if (offset_in_page(a.offset))
1439                 return -EINVAL;
1440
1441         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1442                                a.offset >> PAGE_SHIFT);
1443 }
1444 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1445
1446 /*
1447  * split a vma into two pieces at address 'addr', a new vma is allocated either
1448  * for the first part or the tail.
1449  */
1450 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1451               unsigned long addr, int new_below)
1452 {
1453         struct vm_area_struct *new;
1454         struct vm_region *region;
1455         unsigned long npages;
1456
1457         /* we're only permitted to split anonymous regions (these should have
1458          * only a single usage on the region) */
1459         if (vma->vm_file)
1460                 return -ENOMEM;
1461
1462         if (mm->map_count >= sysctl_max_map_count)
1463                 return -ENOMEM;
1464
1465         region = kmem_cache_alloc(vm_region_jar, GFP_KERNEL);
1466         if (!region)
1467                 return -ENOMEM;
1468
1469         new = vm_area_dup(vma);
1470         if (!new) {
1471                 kmem_cache_free(vm_region_jar, region);
1472                 return -ENOMEM;
1473         }
1474
1475         /* most fields are the same, copy all, and then fixup */
1476         *region = *vma->vm_region;
1477         new->vm_region = region;
1478
1479         npages = (addr - vma->vm_start) >> PAGE_SHIFT;
1480
1481         if (new_below) {
1482                 region->vm_top = region->vm_end = new->vm_end = addr;
1483         } else {
1484                 region->vm_start = new->vm_start = addr;
1485                 region->vm_pgoff = new->vm_pgoff += npages;
1486         }
1487
1488         if (new->vm_ops && new->vm_ops->open)
1489                 new->vm_ops->open(new);
1490
1491         delete_vma_from_mm(vma);
1492         down_write(&nommu_region_sem);
1493         delete_nommu_region(vma->vm_region);
1494         if (new_below) {
1495                 vma->vm_region->vm_start = vma->vm_start = addr;
1496                 vma->vm_region->vm_pgoff = vma->vm_pgoff += npages;
1497         } else {
1498                 vma->vm_region->vm_end = vma->vm_end = addr;
1499                 vma->vm_region->vm_top = addr;
1500         }
1501         add_nommu_region(vma->vm_region);
1502         add_nommu_region(new->vm_region);
1503         up_write(&nommu_region_sem);
1504         add_vma_to_mm(mm, vma);
1505         add_vma_to_mm(mm, new);
1506         return 0;
1507 }
1508
1509 /*
1510  * shrink a VMA by removing the specified chunk from either the beginning or
1511  * the end
1512  */
1513 static int shrink_vma(struct mm_struct *mm,
1514                       struct vm_area_struct *vma,
1515                       unsigned long from, unsigned long to)
1516 {
1517         struct vm_region *region;
1518
1519         /* adjust the VMA's pointers, which may reposition it in the MM's tree
1520          * and list */
1521         delete_vma_from_mm(vma);
1522         if (from > vma->vm_start)
1523                 vma->vm_end = from;
1524         else
1525                 vma->vm_start = to;
1526         add_vma_to_mm(mm, vma);
1527
1528         /* cut the backing region down to size */
1529         region = vma->vm_region;
1530         BUG_ON(region->vm_usage != 1);
1531
1532         down_write(&nommu_region_sem);
1533         delete_nommu_region(region);
1534         if (from > region->vm_start) {
1535                 to = region->vm_top;
1536                 region->vm_top = region->vm_end = from;
1537         } else {
1538                 region->vm_start = to;
1539         }
1540         add_nommu_region(region);
1541         up_write(&nommu_region_sem);
1542
1543         free_page_series(from, to);
1544         return 0;
1545 }
1546
1547 /*
1548  * release a mapping
1549  * - under NOMMU conditions the chunk to be unmapped must be backed by a single
1550  *   VMA, though it need not cover the whole VMA
1551  */
1552 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, struct list_head *uf)
1553 {
1554         struct vm_area_struct *vma;
1555         unsigned long end;
1556         int ret;
1557
1558         len = PAGE_ALIGN(len);
1559         if (len == 0)
1560                 return -EINVAL;
1561
1562         end = start + len;
1563
1564         /* find the first potentially overlapping VMA */
1565         vma = find_vma(mm, start);
1566         if (!vma) {
1567                 static int limit;
1568                 if (limit < 5) {
1569                         pr_warn("munmap of memory not mmapped by process %d (%s): 0x%lx-0x%lx\n",
1570                                         current->pid, current->comm,
1571                                         start, start + len - 1);
1572                         limit++;
1573                 }
1574                 return -EINVAL;
1575         }
1576
1577         /* we're allowed to split an anonymous VMA but not a file-backed one */
1578         if (vma->vm_file) {
1579                 do {
1580                         if (start > vma->vm_start)
1581                                 return -EINVAL;
1582                         if (end == vma->vm_end)
1583                                 goto erase_whole_vma;
1584                         vma = vma->vm_next;
1585                 } while (vma);
1586                 return -EINVAL;
1587         } else {
1588                 /* the chunk must be a subset of the VMA found */
1589                 if (start == vma->vm_start && end == vma->vm_end)
1590                         goto erase_whole_vma;
1591                 if (start < vma->vm_start || end > vma->vm_end)
1592                         return -EINVAL;
1593                 if (offset_in_page(start))
1594                         return -EINVAL;
1595                 if (end != vma->vm_end && offset_in_page(end))
1596                         return -EINVAL;
1597                 if (start != vma->vm_start && end != vma->vm_end) {
1598                         ret = split_vma(mm, vma, start, 1);
1599                         if (ret < 0)
1600                                 return ret;
1601                 }
1602                 return shrink_vma(mm, vma, start, end);
1603         }
1604
1605 erase_whole_vma:
1606         delete_vma_from_mm(vma);
1607         delete_vma(mm, vma);
1608         return 0;
1609 }
1610 EXPORT_SYMBOL(do_munmap);
1611
1612 int vm_munmap(unsigned long addr, size_t len)
1613 {
1614         struct mm_struct *mm = current->mm;
1615         int ret;
1616
1617         down_write(&mm->mmap_sem);
1618         ret = do_munmap(mm, addr, len, NULL);
1619         up_write(&mm->mmap_sem);
1620         return ret;
1621 }
1622 EXPORT_SYMBOL(vm_munmap);
1623
1624 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1625 {
1626         return vm_munmap(addr, len);
1627 }
1628
1629 /*
1630  * release all the mappings made in a process's VM space
1631  */
1632 void exit_mmap(struct mm_struct *mm)
1633 {
1634         struct vm_area_struct *vma;
1635
1636         if (!mm)
1637                 return;
1638
1639         mm->total_vm = 0;
1640
1641         while ((vma = mm->mmap)) {
1642                 mm->mmap = vma->vm_next;
1643                 delete_vma_from_mm(vma);
1644                 delete_vma(mm, vma);
1645                 cond_resched();
1646         }
1647 }
1648
1649 int vm_brk(unsigned long addr, unsigned long len)
1650 {
1651         return -ENOMEM;
1652 }
1653
1654 /*
1655  * expand (or shrink) an existing mapping, potentially moving it at the same
1656  * time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1657  *
1658  * under NOMMU conditions, we only permit changing a mapping's size, and only
1659  * as long as it stays within the region allocated by do_mmap_private() and the
1660  * block is not shareable
1661  *
1662  * MREMAP_FIXED is not supported under NOMMU conditions
1663  */
1664 static unsigned long do_mremap(unsigned long addr,
1665                         unsigned long old_len, unsigned long new_len,
1666                         unsigned long flags, unsigned long new_addr)
1667 {
1668         struct vm_area_struct *vma;
1669
1670         /* insanity checks first */
1671         old_len = PAGE_ALIGN(old_len);
1672         new_len = PAGE_ALIGN(new_len);
1673         if (old_len == 0 || new_len == 0)
1674                 return (unsigned long) -EINVAL;
1675
1676         if (offset_in_page(addr))
1677                 return -EINVAL;
1678
1679         if (flags & MREMAP_FIXED && new_addr != addr)
1680                 return (unsigned long) -EINVAL;
1681
1682         vma = find_vma_exact(current->mm, addr, old_len);
1683         if (!vma)
1684                 return (unsigned long) -EINVAL;
1685
1686         if (vma->vm_end != vma->vm_start + old_len)
1687                 return (unsigned long) -EFAULT;
1688
1689         if (vma->vm_flags & VM_MAYSHARE)
1690                 return (unsigned long) -EPERM;
1691
1692         if (new_len > vma->vm_region->vm_end - vma->vm_region->vm_start)
1693                 return (unsigned long) -ENOMEM;
1694
1695         /* all checks complete - do it */
1696         vma->vm_end = vma->vm_start + new_len;
1697         return vma->vm_start;
1698 }
1699
1700 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1701                 unsigned long, new_len, unsigned long, flags,
1702                 unsigned long, new_addr)
1703 {
1704         unsigned long ret;
1705
1706         down_write(&current->mm->mmap_sem);
1707         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
1708         up_write(&current->mm->mmap_sem);
1709         return ret;
1710 }
1711
1712 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1713                          unsigned int foll_flags)
1714 {
1715         return NULL;
1716 }
1717
1718 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1719                 unsigned long pfn, unsigned long size, pgprot_t prot)
1720 {
1721         if (addr != (pfn << PAGE_SHIFT))
1722                 return -EINVAL;
1723
1724         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1725         return 0;
1726 }
1727 EXPORT_SYMBOL(remap_pfn_range);
1728
1729 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1730 {
1731         unsigned long pfn = start >> PAGE_SHIFT;
1732         unsigned long vm_len = vma->vm_end - vma->vm_start;
1733
1734         pfn += vma->vm_pgoff;
1735         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1736 }
1737 EXPORT_SYMBOL(vm_iomap_memory);
1738
1739 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1740                         unsigned long pgoff)
1741 {
1742         unsigned int size = vma->vm_end - vma->vm_start;
1743
1744         if (!(vma->vm_flags & VM_USERMAP))
1745                 return -EINVAL;
1746
1747         vma->vm_start = (unsigned long)(addr + (pgoff << PAGE_SHIFT));
1748         vma->vm_end = vma->vm_start + size;
1749
1750         return 0;
1751 }
1752 EXPORT_SYMBOL(remap_vmalloc_range);
1753
1754 unsigned long arch_get_unmapped_area(struct file *file, unsigned long addr,
1755         unsigned long len, unsigned long pgoff, unsigned long flags)
1756 {
1757         return -ENOMEM;
1758 }
1759
1760 vm_fault_t filemap_fault(struct vm_fault *vmf)
1761 {
1762         BUG();
1763         return 0;
1764 }
1765 EXPORT_SYMBOL(filemap_fault);
1766
1767 void filemap_map_pages(struct vm_fault *vmf,
1768                 pgoff_t start_pgoff, pgoff_t end_pgoff)
1769 {
1770         BUG();
1771 }
1772 EXPORT_SYMBOL(filemap_map_pages);
1773
1774 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
1775                 unsigned long addr, void *buf, int len, unsigned int gup_flags)
1776 {
1777         struct vm_area_struct *vma;
1778         int write = gup_flags & FOLL_WRITE;
1779
1780         down_read(&mm->mmap_sem);
1781
1782         /* the access must start within one of the target process's mappings */
1783         vma = find_vma(mm, addr);
1784         if (vma) {
1785                 /* don't overrun this mapping */
1786                 if (addr + len >= vma->vm_end)
1787                         len = vma->vm_end - addr;
1788
1789                 /* only read or write mappings where it is permitted */
1790                 if (write && vma->vm_flags & VM_MAYWRITE)
1791                         copy_to_user_page(vma, NULL, addr,
1792                                          (void *) addr, buf, len);
1793                 else if (!write && vma->vm_flags & VM_MAYREAD)
1794                         copy_from_user_page(vma, NULL, addr,
1795                                             buf, (void *) addr, len);
1796                 else
1797                         len = 0;
1798         } else {
1799                 len = 0;
1800         }
1801
1802         up_read(&mm->mmap_sem);
1803
1804         return len;
1805 }
1806
1807 /**
1808  * access_remote_vm - access another process' address space
1809  * @mm:         the mm_struct of the target address space
1810  * @addr:       start address to access
1811  * @buf:        source or destination buffer
1812  * @len:        number of bytes to transfer
1813  * @gup_flags:  flags modifying lookup behaviour
1814  *
1815  * The caller must hold a reference on @mm.
1816  */
1817 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
1818                 void *buf, int len, unsigned int gup_flags)
1819 {
1820         return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
1821 }
1822
1823 /*
1824  * Access another process' address space.
1825  * - source/target buffer must be kernel space
1826  */
1827 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len,
1828                 unsigned int gup_flags)
1829 {
1830         struct mm_struct *mm;
1831
1832         if (addr + len < addr)
1833                 return 0;
1834
1835         mm = get_task_mm(tsk);
1836         if (!mm)
1837                 return 0;
1838
1839         len = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
1840
1841         mmput(mm);
1842         return len;
1843 }
1844 EXPORT_SYMBOL_GPL(access_process_vm);
1845
1846 /**
1847  * nommu_shrink_inode_mappings - Shrink the shared mappings on an inode
1848  * @inode: The inode to check
1849  * @size: The current filesize of the inode
1850  * @newsize: The proposed filesize of the inode
1851  *
1852  * Check the shared mappings on an inode on behalf of a shrinking truncate to
1853  * make sure that that any outstanding VMAs aren't broken and then shrink the
1854  * vm_regions that extend that beyond so that do_mmap_pgoff() doesn't
1855  * automatically grant mappings that are too large.
1856  */
1857 int nommu_shrink_inode_mappings(struct inode *inode, size_t size,
1858                                 size_t newsize)
1859 {
1860         struct vm_area_struct *vma;
1861         struct vm_region *region;
1862         pgoff_t low, high;
1863         size_t r_size, r_top;
1864
1865         low = newsize >> PAGE_SHIFT;
1866         high = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1867
1868         down_write(&nommu_region_sem);
1869         i_mmap_lock_read(inode->i_mapping);
1870
1871         /* search for VMAs that fall within the dead zone */
1872         vma_interval_tree_foreach(vma, &inode->i_mapping->i_mmap, low, high) {
1873                 /* found one - only interested if it's shared out of the page
1874                  * cache */
1875                 if (vma->vm_flags & VM_SHARED) {
1876                         i_mmap_unlock_read(inode->i_mapping);
1877                         up_write(&nommu_region_sem);
1878                         return -ETXTBSY; /* not quite true, but near enough */
1879                 }
1880         }
1881
1882         /* reduce any regions that overlap the dead zone - if in existence,
1883          * these will be pointed to by VMAs that don't overlap the dead zone
1884          *
1885          * we don't check for any regions that start beyond the EOF as there
1886          * shouldn't be any
1887          */
1888         vma_interval_tree_foreach(vma, &inode->i_mapping->i_mmap, 0, ULONG_MAX) {
1889                 if (!(vma->vm_flags & VM_SHARED))
1890                         continue;
1891
1892                 region = vma->vm_region;
1893                 r_size = region->vm_top - region->vm_start;
1894                 r_top = (region->vm_pgoff << PAGE_SHIFT) + r_size;
1895
1896                 if (r_top > newsize) {
1897                         region->vm_top -= r_top - newsize;
1898                         if (region->vm_end > region->vm_top)
1899                                 region->vm_end = region->vm_top;
1900                 }
1901         }
1902
1903         i_mmap_unlock_read(inode->i_mapping);
1904         up_write(&nommu_region_sem);
1905         return 0;
1906 }
1907
1908 /*
1909  * Initialise sysctl_user_reserve_kbytes.
1910  *
1911  * This is intended to prevent a user from starting a single memory hogging
1912  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
1913  * mode.
1914  *
1915  * The default value is min(3% of free memory, 128MB)
1916  * 128MB is enough to recover with sshd/login, bash, and top/kill.
1917  */
1918 static int __meminit init_user_reserve(void)
1919 {
1920         unsigned long free_kbytes;
1921
1922         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
1923
1924         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
1925         return 0;
1926 }
1927 subsys_initcall(init_user_reserve);
1928
1929 /*
1930  * Initialise sysctl_admin_reserve_kbytes.
1931  *
1932  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
1933  * to log in and kill a memory hogging process.
1934  *
1935  * Systems with more than 256MB will reserve 8MB, enough to recover
1936  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
1937  * only reserve 3% of free pages by default.
1938  */
1939 static int __meminit init_admin_reserve(void)
1940 {
1941         unsigned long free_kbytes;
1942
1943         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
1944
1945         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
1946         return 0;
1947 }
1948 subsys_initcall(init_admin_reserve);