mm/nommu.c

   1 /*
   2  *  linux/mm/nommu.c
   3  *
   4  *  Replacement code for mm functions to support CPU's that don't
   5  *  have any form of memory management unit (thus no virtual memory).
   6  *
   7  *  See Documentation/nommu-mmap.txt
   8  *
   9  *  Copyright (c) 2004-2008 David Howells <dhowells@redhat.com>
  10  *  Copyright (c) 2000-2003 David McCullough <davidm@snapgear.com>
  11  *  Copyright (c) 2000-2001 D Jeff Dionne <jeff@uClinux.org>
  12  *  Copyright (c) 2002      Greg Ungerer <gerg@snapgear.com>
  13  *  Copyright (c) 2007-2009 Paul Mundt <lethal@linux-sh.org>
  14  */
  15
  16 #include <linux/module.h>
  17 #include <linux/mm.h>
  18 #include <linux/mman.h>
  19 #include <linux/swap.h>
  20 #include <linux/file.h>
  21 #include <linux/highmem.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/slab.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/tracehook.h>
  26 #include <linux/blkdev.h>
  27 #include <linux/backing-dev.h>
  28 #include <linux/mount.h>
  29 #include <linux/personality.h>
  30 #include <linux/security.h>
  31 #include <linux/syscalls.h>
  32
  33 #include <asm/uaccess.h>
  34 #include <asm/tlb.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/mmu_context.h>
  37 #include "internal.h"
  38
  39 static inline __attribute__((format(printf, 1, 2)))
  40 void no_printk(const char *fmt, ...)
  41 {
  42 }
  43
  44 #if 0
  45 #define kenter(FMT, ...) \
  46         printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
  47 #define kleave(FMT, ...) \
  48         printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
  49 #define kdebug(FMT, ...) \
  50         printk(KERN_DEBUG "xxx" FMT"yyy\n", ##__VA_ARGS__)
  51 #else
  52 #define kenter(FMT, ...) \
  53         no_printk(KERN_DEBUG "==> %s("FMT")\n", __func__, ##__VA_ARGS__)
  54 #define kleave(FMT, ...) \
  55         no_printk(KERN_DEBUG "<== %s()"FMT"\n", __func__, ##__VA_ARGS__)
  56 #define kdebug(FMT, ...) \
  57         no_printk(KERN_DEBUG FMT"\n", ##__VA_ARGS__)
  58 #endif
  59
  60 void *high_memory;
  61 struct page *mem_map;
  62 unsigned long max_mapnr;
  63 unsigned long num_physpages;
  64 unsigned long highest_memmap_pfn;
  65 struct percpu_counter vm_committed_as;
  66 int sysctl_overcommit_memory = OVERCOMMIT_GUESS; /* heuristic overcommit */
  67 int sysctl_overcommit_ratio = 50; /* default is 50% */
  68 int sysctl_max_map_count = DEFAULT_MAX_MAP_COUNT;
  69 int sysctl_nr_trim_pages = CONFIG_NOMMU_INITIAL_TRIM_EXCESS;
  70 int heap_stack_gap = 0;
  71
  72 atomic_long_t mmap_pages_allocated;
  73
  74 EXPORT_SYMBOL(mem_map);
  75 EXPORT_SYMBOL(num_physpages);
  76
  77 /* list of mapped, potentially shareable regions */
  78 static struct kmem_cache *vm_region_jar;
  79 struct rb_root nommu_region_tree = RB_ROOT;
  80 DECLARE_RWSEM(nommu_region_sem);
  81
  82 const struct vm_operations_struct generic_file_vm_ops = {
  83 };
  84
  85 /*
  86  * Return the total memory allocated for this pointer, not
  87  * just what the caller asked for.
  88  *
  89  * Doesn't have to be accurate, i.e. may have races.
  90  */
  91 unsigned int kobjsize(const void *objp)
  92 {
  93         struct page *page;
  94
  95         /*
  96          * If the object we have should not have ksize performed on it,
  97          * return size of 0
  98          */
  99         if (!objp || !virt_addr_valid(objp))
 100                 return 0;
 101
 102         page = virt_to_head_page(objp);
 103
 104         /*
 105          * If the allocator sets PageSlab, we know the pointer came from
 106          * kmalloc().
 107          */
 108         if (PageSlab(page))
 109                 return ksize(objp);
 110
 111         /*
 112          * If it's not a compound page, see if we have a matching VMA
 113          * region. This test is intentionally done in reverse order,
 114          * so if there's no VMA, we still fall through and hand back
 115          * PAGE_SIZE for 0-order pages.
 116          */
 117         if (!PageCompound(page)) {
 118                 struct vm_area_struct *vma;
 119
 120                 vma = find_vma(current->mm, (unsigned long)objp);
 121                 if (vma)
 122                         return vma->vm_end - vma->vm_start;
 123         }
 124
 125         /*
 126          * The ksize() function is only guaranteed to work for pointers
 127          * returned by kmalloc(). So handle arbitrary pointers here.
 128          */
 129         return PAGE_SIZE << compound_order(page);
 130 }
 131
 132 int __get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 133                      unsigned long start, int nr_pages, unsigned int foll_flags,
 134                      struct page **pages, struct vm_area_struct **vmas)
 135 {
 136         struct vm_area_struct *vma;
 137         unsigned long vm_flags;
 138         int i;
 139
 140         /* calculate required read or write permissions.
 141          * If FOLL_FORCE is set, we only require the "MAY" flags.
 142          */
 143         vm_flags  = (foll_flags & FOLL_WRITE) ?
 144                         (VM_WRITE | VM_MAYWRITE) : (VM_READ | VM_MAYREAD);
 145         vm_flags &= (foll_flags & FOLL_FORCE) ?
 146                         (VM_MAYREAD | VM_MAYWRITE) : (VM_READ | VM_WRITE);
 147
 148         for (i = 0; i < nr_pages; i++) {
 149                 vma = find_vma(mm, start);
 150                 if (!vma)
 151                         goto finish_or_fault;
 152
 153                 /* protect what we can, including chardevs */
 154                 if ((vma->vm_flags & (VM_IO | VM_PFNMAP)) ||
 155                     !(vm_flags & vma->vm_flags))
 156                         goto finish_or_fault;
 157
 158                 if (pages) {
 159                         pages[i] = virt_to_page(start);
 160                         if (pages[i])
 161                                 page_cache_get(pages[i]);
 162                 }
 163                 if (vmas)
 164                         vmas[i] = vma;
 165                 start += PAGE_SIZE;
 166         }
 167
 168         return i;
 169
 170 finish_or_fault:
 171         return i ? : -EFAULT;
 172 }
 173
 174 /*
 175  * get a list of pages in an address range belonging to the specified process
 176  * and indicate the VMA that covers each page
 177  * - this is potentially dodgy as we may end incrementing the page count of a
 178  *   slab page or a secondary page from a compound page
 179  * - don't permit access to VMAs that don't support it, such as I/O mappings
 180  */
 181 int get_user_pages(struct task_struct *tsk, struct mm_struct *mm,
 182         unsigned long start, int nr_pages, int write, int force,
 183         struct page **pages, struct vm_area_struct **vmas)
 184 {
 185         int flags = 0;
 186
 187         if (write)
 188                 flags |= FOLL_WRITE;
 189         if (force)
 190                 flags |= FOLL_FORCE;
 191
 192         return __get_user_pages(tsk, mm, start, nr_pages, flags, pages, vmas);
 193 }
 194 EXPORT_SYMBOL(get_user_pages);
 195
 196 /**
 197  * follow_pfn - look up PFN at a user virtual address
 198  * @vma: memory mapping
 199  * @address: user virtual address
 200  * @pfn: location to store found PFN
 201  *
 202  * Only IO mappings and raw PFN mappings are allowed.
 203  *
 204  * Returns zero and the pfn at @pfn on success, -ve otherwise.
 205  */
 206 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
 207         unsigned long *pfn)
 208 {
 209         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
 210                 return -EINVAL;
 211
 212         *pfn = address >> PAGE_SHIFT;
 213         return 0;
 214 }
 215 EXPORT_SYMBOL(follow_pfn);
 216
 217 DEFINE_RWLOCK(vmlist_lock);
 218 struct vm_struct *vmlist;
 219
 220 void vfree(const void *addr)
 221 {
 222         kfree(addr);
 223 }
 224 EXPORT_SYMBOL(vfree);
 225
 226 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
 227 {
 228         /*
 229          *  You can't specify __GFP_HIGHMEM with kmalloc() since kmalloc()
 230          * returns only a logical address.
 231          */
 232         return kmalloc(size, (gfp_mask | __GFP_COMP) & ~__GFP_HIGHMEM);
 233 }
 234 EXPORT_SYMBOL(__vmalloc);
 235
 236 void *vmalloc_user(unsigned long size)
 237 {
 238         void *ret;
 239
 240         ret = __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO,
 241                         PAGE_KERNEL);
 242         if (ret) {
 243                 struct vm_area_struct *vma;
 244
 245                 down_write(&current->mm->mmap_sem);
 246                 vma = find_vma(current->mm, (unsigned long)ret);
 247                 if (vma)
 248                         vma->vm_flags |= VM_USERMAP;
 249                 up_write(&current->mm->mmap_sem);
 250         }
 251
 252         return ret;
 253 }
 254 EXPORT_SYMBOL(vmalloc_user);
 255
 256 struct page *vmalloc_to_page(const void *addr)
 257 {
 258         return virt_to_page(addr);
 259 }
 260 EXPORT_SYMBOL(vmalloc_to_page);
 261
 262 unsigned long vmalloc_to_pfn(const void *addr)
 263 {
 264         return page_to_pfn(virt_to_page(addr));
 265 }
 266 EXPORT_SYMBOL(vmalloc_to_pfn);
 267
 268 long vread(char *buf, char *addr, unsigned long count)
 269 {
 270         memcpy(buf, addr, count);
 271         return count;
 272 }
 273
 274 long vwrite(char *buf, char *addr, unsigned long count)
 275 {
 276         /* Don't allow overflow */
 277         if ((unsigned long) addr + count < count)
 278                 count = -(unsigned long) addr;
 279
 280         memcpy(addr, buf, count);
 281         return(count);
 282 }
 283
 284 /*
 285  *      vmalloc  -  allocate virtually continguos memory
 286  *
 287  *      @size:          allocation size
 288  *
 289  *      Allocate enough pages to cover @size from the page level
 290  *      allocator and map them into continguos kernel virtual space.
 291  *
 292  *      For tight control over page level allocator and protection flags
 293  *      use __vmalloc() instead.
 294  */
 295 void *vmalloc(unsigned long size)
 296 {
 297        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
 298 }
 299 EXPORT_SYMBOL(vmalloc);
 300
 301 void *vmalloc_node(unsigned long size, int node)
 302 {
 303         return vmalloc(size);
 304 }
 305 EXPORT_SYMBOL(vmalloc_node);
 306
 307 #ifndef PAGE_KERNEL_EXEC
 308 # define PAGE_KERNEL_EXEC PAGE_KERNEL
 309 #endif
 310
 311 /**
 312  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
 313  *      @size:          allocation size
 314  *
 315  *      Kernel-internal function to allocate enough pages to cover @size
 316  *      the page level allocator and map them into contiguous and
 317  *      executable kernel virtual space.
 318  *
 319  *      For tight control over page level allocator and protection flags
 320  *      use __vmalloc() instead.
 321  */
 322
 323 void *vmalloc_exec(unsigned long size)
 324 {
 325         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
 326 }
 327
 328 /**
 329  * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
 330  *      @size:          allocation size
 331  *
 332  *      Allocate enough 32bit PA addressable pages to cover @size from the
 333  *      page level allocator and map them into continguos kernel virtual space.
 334  */
 335 void *vmalloc_32(unsigned long size)
 336 {
 337         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
 338 }
 339 EXPORT_SYMBOL(vmalloc_32);
 340
 341 /**
 342  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
 343  *      @size:          allocation size
 344  *
 345  * The resulting memory area is 32bit addressable and zeroed so it can be
 346  * mapped to userspace without leaking data.
 347  *
 348  * VM_USERMAP is set on the corresponding VMA so that subsequent calls to
 349  * remap_vmalloc_range() are permissible.
 350  */
 351 void *vmalloc_32_user(unsigned long size)
 352 {
 353         /*
 354          * We'll have to sort out the ZONE_DMA bits for 64-bit,
 355          * but for now this can simply use vmalloc_user() directly.
 356          */
 357         return vmalloc_user(size);
 358 }
 359 EXPORT_SYMBOL(vmalloc_32_user);
 360
 361 void *vmap(struct page **pages, unsigned int count, unsigned long flags, pgprot_t prot)
 362 {
 363         BUG();
 364         return NULL;
 365 }
 366 EXPORT_SYMBOL(vmap);
 367
 368 void vunmap(const void *addr)
 369 {
 370         BUG();
 371 }
 372 EXPORT_SYMBOL(vunmap);
 373
 374 void *vm_map_ram(struct page **pages, unsigned int count, int node, pgprot_t prot)
 375 {
 376         BUG();
 377         return NULL;
 378 }
 379 EXPORT_SYMBOL(vm_map_ram);
 380
 381 void vm_unmap_ram(const void *mem, unsigned int count)
 382 {
 383         BUG();
 384 }
 385 EXPORT_SYMBOL(vm_unmap_ram);
 386
 387 void vm_unmap_aliases(void)
 388 {
 389 }
 390 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
 391
 392 /*
 393  * Implement a stub for vmalloc_sync_all() if the architecture chose not to
 394  * have one.
 395  */
 396 void  __attribute__((weak)) vmalloc_sync_all(void)
 397 {
 398 }
 399
 400 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
 401                    struct page *page)
 402 {
 403         return -EINVAL;
 404 }
 405 EXPORT_SYMBOL(vm_insert_page);
 406
 407 /*
 408  *  sys_brk() for the most part doesn't need the global kernel
 409  *  lock, except when an application is doing something nasty
 410  *  like trying to un-brk an area that has already been mapped
 411  *  to a regular file.  in this case, the unmapping will need
 412  *  to invoke file system routines that need the global lock.
 413  */
 414 SYSCALL_DEFINE1(brk, unsigned long, brk)
 415 {
 416         struct mm_struct *mm = current->mm;
 417
 418         if (brk < mm->start_brk || brk > mm->context.end_brk)
 419                 return mm->brk;
 420
 421         if (mm->brk == brk)
 422                 return mm->brk;
 423
 424         /*
 425          * Always allow shrinking brk
 426          */
 427         if (brk <= mm->brk) {
 428                 mm->brk = brk;
 429                 return brk;
 430         }
 431
 432         /*
 433          * Ok, looks good - let it rip.
 434          */
 435         flush_icache_range(mm->brk, brk);
 436         return mm->brk = brk;
 437 }
 438
 439 /*
 440  * initialise the VMA and region record slabs
 441  */
 442 void __init mmap_init(void)
 443 {
 444         int ret;
 445
 446         ret = percpu_counter_init(&vm_committed_as, 0);
 447         VM_BUG_ON(ret);
 448         vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC);
 449 }
 450
 451 /*
 452  * validate the region tree
 453  * - the caller must hold the region lock
 454  */
 455 #ifdef CONFIG_DEBUG_NOMMU_REGIONS
 456 static noinline void validate_nommu_regions(void)
 457 {
 458         struct vm_region *region, *last;
 459         struct rb_node *p, *lastp;
 460
 461         lastp = rb_first(&nommu_region_tree);
 462         if (!lastp)
 463                 return;
 464
 465         last = rb_entry(lastp, struct vm_region, vm_rb);
 466         BUG_ON(unlikely(last->vm_end <= last->vm_start));
 467         BUG_ON(unlikely(last->vm_top < last->vm_end));
 468
 469         while ((p = rb_next(lastp))) {
 470                 region = rb_entry(p, struct vm_region, vm_rb);
 471                 last = rb_entry(lastp, struct vm_region, vm_rb);
 472
 473                 BUG_ON(unlikely(region->vm_end <= region->vm_start));
 474                 BUG_ON(unlikely(region->vm_top < region->vm_end));
 475                 BUG_ON(unlikely(region->vm_start < last->vm_top));
 476
 477                 lastp = p;
 478         }
 479 }
 480 #else
 481 static void validate_nommu_regions(void)
 482 {
 483 }
 484 #endif
 485
 486 /*
 487  * add a region into the global tree
 488  */
 489 static void add_nommu_region(struct vm_region *region)
 490 {
 491         struct vm_region *pregion;
 492         struct rb_node **p, *parent;
 493
 494         validate_nommu_regions();
 495
 496         parent = NULL;
 497         p = &nommu_region_tree.rb_node;
 498         while (*p) {
 499                 parent = *p;
 500                 pregion = rb_entry(parent, struct vm_region, vm_rb);
 501                 if (region->vm_start < pregion->vm_start)
 502                         p = &(*p)->rb_left;
 503                 else if (region->vm_start > pregion->vm_start)
 504                         p = &(*p)->rb_right;
 505                 else if (pregion == region)
 506                         return;
 507                 else
 508                         BUG();
 509         }
 510
 511         rb_link_node(&region->vm_rb, parent, p);
 512         rb_insert_color(&region->vm_rb, &nommu_region_tree);
 513
 514         validate_nommu_regions();
 515 }
 516
 517 /*
 518  * delete a region from the global tree
 519  */
 520 static void delete_nommu_region(struct vm_region *region)
 521 {
 522         BUG_ON(!nommu_region_tree.rb_node);
 523
 524         validate_nommu_regions();
 525         rb_erase(&region->vm_rb, &nommu_region_tree);
 526         validate_nommu_regions();
 527 }
 528
 529 /*
 530  * free a contiguous series of pages
 531  */
 532 static void free_page_series(unsigned long from, unsigned long to)
 533 {
 534         for (; from < to; from += PAGE_SIZE) {
 535                 struct page *page = virt_to_page(from);
 536
 537                 kdebug("- free %lx", from);
 538                 atomic_long_dec(&mmap_pages_allocated);
 539                 if (page_count(page) != 1)
 540                         kdebug("free page %p: refcount not one: %d",
 541                                page, page_count(page));
 542                 put_page(page);
 543         }
 544 }
 545
 546 /*
 547  * release a reference to a region
 548  * - the caller must hold the region semaphore for writing, which this releases
 549  * - the region may not have been added to the tree yet, in which case vm_top
 550  *   will equal vm_start
 551  */
 552 static void __put_nommu_region(struct vm_region *region)
 553         __releases(nommu_region_sem)
 554 {
 555         kenter("%p{%d}", region, region->vm_usage);
 556
 557         BUG_ON(!nommu_region_tree.rb_node);
 558
 559         if (--region->vm_usage == 0) {
 560                 if (region->vm_top > region->vm_start)
 561                         delete_nommu_region(region);
 562                 up_write(&nommu_region_sem);
 563
 564                 if (region->vm_file)
 565                         fput(region->vm_file);
 566
 567                 /* IO memory and memory shared directly out of the pagecache
 568                  * from ramfs/tmpfs mustn't be released here */
 569                 if (region->vm_flags & VM_MAPPED_COPY) {
 570                         kdebug("free series");
 571                         free_page_series(region->vm_start, region->vm_top);
 572                 }
 573                 kmem_cache_free(vm_region_jar, region);
 574         } else {
 575                 up_write(&nommu_region_sem);
 576         }
 577 }
 578
 579 /*
 580  * release a reference to a region
 581  */
 582 static void put_nommu_region(struct vm_region *region)
 583 {
 584         down_write(&nommu_region_sem);
 585         __put_nommu_region(region);
 586 }
 587
 588 /*
 589  * update protection on a vma
 590  */
 591 static void protect_vma(struct vm_area_struct *vma, unsigned long flags)
 592 {
 593 #ifdef CONFIG_MPU
 594         struct mm_struct *mm = vma->vm_mm;
 595         long start = vma->vm_start & PAGE_MASK;
 596         while (start < vma->vm_end) {
 597                 protect_page(mm, start, flags);
 598                 start += PAGE_SIZE;
 599         }
 600         update_protections(mm);
 601 #endif
 602 }
 603
 604 /*
 605  * add a VMA into a process's mm_struct in the appropriate place in the list
 606  * and tree and add to the address space's page tree also if not an anonymous
 607  * page
 608  * - should be called with mm->mmap_sem held writelocked
 609  */
 610 static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
 611 {
 612         struct vm_area_struct *pvma, **pp;
 613         struct address_space *mapping;
 614         struct rb_node **p, *parent;
 615
 616         kenter(",%p", vma);
 617
 618         BUG_ON(!vma->vm_region);
 619
 620         mm->map_count++;
 621         vma->vm_mm = mm;
 622
 623         protect_vma(vma, vma->vm_flags);
 624
 625         /* add the VMA to the mapping */
 626         if (vma->vm_file) {
 627                 mapping = vma->vm_file->f_mapping;
 628
 629                 flush_dcache_mmap_lock(mapping);
 630                 vma_prio_tree_insert(vma, &mapping->i_mmap);
 631                 flush_dcache_mmap_unlock(mapping);
 632         }
 633
 634         /* add the VMA to the tree */
 635         parent = NULL;
 636         p = &mm->mm_rb.rb_node;
 637         while (*p) {
 638                 parent = *p;
 639                 pvma = rb_entry(parent, struct vm_area_struct, vm_rb);
 640
 641                 /* sort by: start addr, end addr, VMA struct addr in that order
 642                  * (the latter is necessary as we may get identical VMAs) */
 643                 if (vma->vm_start < pvma->vm_start)
 644                         p = &(*p)->rb_left;
 645                 else if (vma->vm_start > pvma->vm_start)
 646                         p = &(*p)->rb_right;
 647                 else if (vma->vm_end < pvma->vm_end)
 648                         p = &(*p)->rb_left;
 649                 else if (vma->vm_end > pvma->vm_end)
 650                         p = &(*p)->rb_right;
 651                 else if (vma < pvma)
 652                         p = &(*p)->rb_left;
 653                 else if (vma > pvma)
 654                         p = &(*p)->rb_right;
 655                 else
 656                         BUG();
 657         }
 658
 659         rb_link_node(&vma->vm_rb, parent, p);
 660         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
 661
 662         /* add VMA to the VMA list also */
 663         for (pp = &mm->mmap; (pvma = *pp); pp = &(*pp)->vm_next) {
 664                 if (pvma->vm_start > vma->vm_start)
 665                         break;
 666                 if (pvma->vm_start < vma->vm_start)
 667                         continue;
 668                 if (pvma->vm_end < vma->vm_end)
 669                         break;
 670         }
 671
 672         vma->vm_next = *pp;
 673         *pp = vma;
 674 }
 675
 676 /*
 677  * delete a VMA from its owning mm_struct and address space
 678  */
 679 static void delete_vma_from_mm(struct vm_area_struct *vma)
 680 {
 681         struct vm_area_struct **pp;
 682         struct address_space *mapping;
 683         struct mm_struct *mm = vma->vm_mm;
 684
 685         kenter("%p", vma);
 686
 687         protect_vma(vma, 0);
 688
 689         mm->map_count--;
 690         if (mm->mmap_cache == vma)
 691                 mm->mmap_cache = NULL;
 692
 693         /* remove the VMA from the mapping */
 694         if (vma->vm_file) {
 695                 mapping = vma->vm_file->f_mapping;
 696
 697                 flush_dcache_mmap_lock(mapping);
 698                 vma_prio_tree_remove(vma, &mapping->i_mmap);
 699                 flush_dcache_mmap_unlock(mapping);
 700         }
 701
 702         /* remove from the MM's tree and list */
 703         rb_erase(&vma->vm_rb, &mm->mm_rb);
 704         for (pp = &mm->mmap; *pp; pp = &(*pp)->vm_next) {
 705                 if (*pp == vma) {
 706                         *pp = vma->vm_next;
 707                         break;
 708                 }
 709         }
 710
 711         vma->vm_mm = NULL;
 712 }
 713
 714 /*
 715  * destroy a VMA record
 716  */
 717 static void delete_vma(struct mm_struct *mm, struct vm_area_struct *vma)
 718 {
 719         kenter("%p", vma);
 720         if (vma->vm_ops && vma->vm_ops->close)
 721                 vma->vm_ops->close(vma);
 722         if (vma->vm_file) {
 723                 fput(vma->vm_file);
 724                 if (vma->vm_flags & VM_EXECUTABLE)
 725                         removed_exe_file_vma(mm);
 726         }
 727         put_nommu_region(vma->vm_region);
 728         kmem_cache_free(vm_area_cachep, vma);
 729 }
 730
 731 /*
 732  * look up the first VMA in which addr resides, NULL if none
 733  * - should be called with mm->mmap_sem at least held readlocked
 734  */
 735 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 736 {
 737         struct vm_area_struct *vma;
 738         struct rb_node *n = mm->mm_rb.rb_node;
 739
 740         /* check the cache first */
 741         vma = mm->mmap_cache;
 742         if (vma && vma->vm_start <= addr && vma->vm_end > addr)
 743                 return vma;
 744
 745         /* trawl the tree (there may be multiple mappings in which addr
 746          * resides) */
 747         for (n = rb_first(&mm->mm_rb); n; n = rb_next(n)) {
 748                 vma = rb_entry(n, struct vm_area_struct, vm_rb);
 749                 if (vma->vm_start > addr)
 750                         return NULL;
 751                 if (vma->vm_end > addr) {
 752                         mm->mmap_cache = vma;
 753                         return vma;
 754                 }
 755         }
 756
 757         return NULL;
 758 }
 759 EXPORT_SYMBOL(find_vma);
 760
 761 /*
 762  * find a VMA
 763  * - we don't extend stack VMAs under NOMMU conditions
 764  */
 765 struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr)
 766 {
 767         return find_vma(mm, addr);
 768 }
 769
 770 /*
 771  * expand a stack to a given address
 772  * - not supported under NOMMU conditions
 773  */
 774 int expand_stack(struct vm_area_struct *vma, unsigned long address)
 775 {
 776         return -ENOMEM;
 777 }
 778
 779 /*
 780  * look up the first VMA exactly that exactly matches addr
 781  * - should be called with mm->mmap_sem at least held readlocked
 782  */
 783 static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
 784                                              unsigned long addr,
 785                                              unsigned long len)
 786 {
 787         struct vm_area_struct *vma;
 788         struct rb_node *n = mm->mm_rb.rb_node;
 789         unsigned long end = addr + len;
 790
 791         /* check the cache first */
 792         vma = mm->mmap_cache;
 793         if (vma && vma->vm_start == addr && vma->vm_end == end)
 794                 return vma;
 795
 796         /* trawl the tree (there may be multiple mappings in which addr
 797          * resides) */
 798         for (n = rb_first(&mm->mm_rb); n; n = rb_next(n)) {
 799                 vma = rb_entry(n, struct vm_area_struct, vm_rb);
 800                 if (vma->vm_start < addr)
 801                         continue;
 802                 if (vma->vm_start > addr)
 803                         return NULL;
 804                 if (vma->vm_end == end) {
 805                         mm->mmap_cache = vma;
 806                         return vma;
 807                 }
 808         }
 809
 810         return NULL;
 811 }
 812
 813 /*
 814  * determine whether a mapping should be permitted and, if so, what sort of
 815  * mapping we're capable of supporting
 816  */
 817 static int validate_mmap_request(struct file *file,
 818                                  unsigned long addr,
 819                                  unsigned long len,
 820                                  unsigned long prot,
 821                                  unsigned long flags,
 822                                  unsigned long pgoff,
 823                                  unsigned long *_capabilities)
 824 {
 825         unsigned long capabilities, rlen;
 826         unsigned long reqprot = prot;
 827         int ret;
 828
 829         /* do the simple checks first */
 830         if (flags & MAP_FIXED) {
 831                 printk(KERN_DEBUG
 832                        "%d: Can't do fixed-address/overlay mmap of RAM\n",
 833                        current->pid);
 834                 return -EINVAL;
 835         }
 836
 837         if ((flags & MAP_TYPE) != MAP_PRIVATE &&
 838             (flags & MAP_TYPE) != MAP_SHARED)
 839                 return -EINVAL;
 840
 841         if (!len)
 842                 return -EINVAL;
 843
 844         /* Careful about overflows.. */
 845         rlen = PAGE_ALIGN(len);
 846         if (!rlen || rlen > TASK_SIZE)
 847                 return -ENOMEM;
 848
 849         /* offset overflow? */
 850         if ((pgoff + (rlen >> PAGE_SHIFT)) < pgoff)
 851                 return -EOVERFLOW;
 852
 853         if (file) {
 854                 /* validate file mapping requests */
 855                 struct address_space *mapping;
 856
 857                 /* files must support mmap */
 858                 if (!file->f_op || !file->f_op->mmap)
 859                         return -ENODEV;
 860
 861                 /* work out if what we've got could possibly be shared
 862                  * - we support chardevs that provide their own "memory"
 863                  * - we support files/blockdevs that are memory backed
 864                  */
 865                 mapping = file->f_mapping;
 866                 if (!mapping)
 867                         mapping = file->f_path.dentry->d_inode->i_mapping;
 868
 869                 capabilities = 0;
 870                 if (mapping && mapping->backing_dev_info)
 871                         capabilities = mapping->backing_dev_info->capabilities;
 872
 873                 if (!capabilities) {
 874                         /* no explicit capabilities set, so assume some
 875                          * defaults */
 876                         switch (file->f_path.dentry->d_inode->i_mode & S_IFMT) {
 877                         case S_IFREG:
 878                         case S_IFBLK:
 879                                 capabilities = BDI_CAP_MAP_COPY;
 880                                 break;
 881
 882                         case S_IFCHR:
 883                                 capabilities =
 884                                         BDI_CAP_MAP_DIRECT |
 885                                         BDI_CAP_READ_MAP |
 886                                         BDI_CAP_WRITE_MAP;
 887                                 break;
 888
 889                         default:
 890                                 return -EINVAL;
 891                         }
 892                 }
 893
 894                 /* eliminate any capabilities that we can't support on this
 895                  * device */
 896                 if (!file->f_op->get_unmapped_area)
 897                         capabilities &= ~BDI_CAP_MAP_DIRECT;
 898                 if (!file->f_op->read)
 899                         capabilities &= ~BDI_CAP_MAP_COPY;
 900
 901                 /* The file shall have been opened with read permission. */
 902                 if (!(file->f_mode & FMODE_READ))
 903                         return -EACCES;
 904
 905                 if (flags & MAP_SHARED) {
 906                         /* do checks for writing, appending and locking */
 907                         if ((prot & PROT_WRITE) &&
 908                             !(file->f_mode & FMODE_WRITE))
 909                                 return -EACCES;
 910
 911                         if (IS_APPEND(file->f_path.dentry->d_inode) &&
 912                             (file->f_mode & FMODE_WRITE))
 913                                 return -EACCES;
 914
 915                         if (locks_verify_locked(file->f_path.dentry->d_inode))
 916                                 return -EAGAIN;
 917
 918                         if (!(capabilities & BDI_CAP_MAP_DIRECT))
 919                                 return -ENODEV;
 920
 921                         if (((prot & PROT_READ)  && !(capabilities & BDI_CAP_READ_MAP))  ||
 922                             ((prot & PROT_WRITE) && !(capabilities & BDI_CAP_WRITE_MAP)) ||
 923                             ((prot & PROT_EXEC)  && !(capabilities & BDI_CAP_EXEC_MAP))
 924                             ) {
 925                                 printk("MAP_SHARED not completely supported on !MMU\n");
 926                                 return -EINVAL;
 927                         }
 928
 929                         /* we mustn't privatise shared mappings */
 930                         capabilities &= ~BDI_CAP_MAP_COPY;
 931                 }
 932                 else {
 933                         /* we're going to read the file into private memory we
 934                          * allocate */
 935                         if (!(capabilities & BDI_CAP_MAP_COPY))
 936                                 return -ENODEV;
 937
 938                         /* we don't permit a private writable mapping to be
 939                          * shared with the backing device */
 940                         if (prot & PROT_WRITE)
 941                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
 942                 }
 943
 944                 /* handle executable mappings and implied executable
 945                  * mappings */
 946                 if (file->f_path.mnt->mnt_flags & MNT_NOEXEC) {
 947                         if (prot & PROT_EXEC)
 948                                 return -EPERM;
 949                 }
 950                 else if ((prot & PROT_READ) && !(prot & PROT_EXEC)) {
 951                         /* handle implication of PROT_EXEC by PROT_READ */
 952                         if (current->personality & READ_IMPLIES_EXEC) {
 953                                 if (capabilities & BDI_CAP_EXEC_MAP)
 954                                         prot |= PROT_EXEC;
 955                         }
 956                 }
 957                 else if ((prot & PROT_READ) &&
 958                          (prot & PROT_EXEC) &&
 959                          !(capabilities & BDI_CAP_EXEC_MAP)
 960                          ) {
 961                         /* backing file is not executable, try to copy */
 962                         capabilities &= ~BDI_CAP_MAP_DIRECT;
 963                 }
 964         }
 965         else {
 966                 /* anonymous mappings are always memory backed and can be
 967                  * privately mapped
 968                  */
 969                 capabilities = BDI_CAP_MAP_COPY;
 970
 971                 /* handle PROT_EXEC implication by PROT_READ */
 972                 if ((prot & PROT_READ) &&
 973                     (current->personality & READ_IMPLIES_EXEC))
 974                         prot |= PROT_EXEC;
 975         }
 976
 977         /* allow the security API to have its say */
 978         ret = security_file_mmap(file, reqprot, prot, flags, addr, 0);
 979         if (ret < 0)
 980                 return ret;
 981
 982         /* looks okay */
 983         *_capabilities = capabilities;
 984         return 0;
 985 }
 986
 987 /*
 988  * we've determined that we can make the mapping, now translate what we
 989  * now know into VMA flags
 990  */
 991 static unsigned long determine_vm_flags(struct file *file,
 992                                         unsigned long prot,
 993                                         unsigned long flags,
 994                                         unsigned long capabilities)
 995 {
 996         unsigned long vm_flags;
 997
 998         vm_flags = calc_vm_prot_bits(prot) | calc_vm_flag_bits(flags);
 999         vm_flags |= VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
1000         /* vm_flags |= mm->def_flags; */
1001
1002         if (!(capabilities & BDI_CAP_MAP_DIRECT)) {
1003                 /* attempt to share read-only copies of mapped file chunks */
1004                 if (file && !(prot & PROT_WRITE))
1005                         vm_flags |= VM_MAYSHARE;
1006         }
1007         else {
1008                 /* overlay a shareable mapping on the backing device or inode
1009                  * if possible - used for chardevs, ramfs/tmpfs/shmfs and
1010                  * romfs/cramfs */
1011                 if (flags & MAP_SHARED)
1012                         vm_flags |= VM_MAYSHARE | VM_SHARED;
1013                 else if ((((vm_flags & capabilities) ^ vm_flags) & BDI_CAP_VMFLAGS) == 0)
1014                         vm_flags |= VM_MAYSHARE;
1015         }
1016
1017         /* refuse to let anyone share private mappings with this process if
1018          * it's being traced - otherwise breakpoints set in it may interfere
1019          * with another untraced process
1020          */
1021         if ((flags & MAP_PRIVATE) && tracehook_expect_breakpoints(current))
1022                 vm_flags &= ~VM_MAYSHARE;
1023
1024         return vm_flags;
1025 }
1026
1027 /*
1028  * set up a shared mapping on a file (the driver or filesystem provides and
1029  * pins the storage)
1030  */
1031 static int do_mmap_shared_file(struct vm_area_struct *vma)
1032 {
1033         int ret;
1034
1035         ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1036         if (ret == 0) {
1037                 vma->vm_region->vm_top = vma->vm_region->vm_end;
1038                 return 0;
1039         }
1040         if (ret != -ENOSYS)
1041                 return ret;
1042
1043         /* getting an ENOSYS error indicates that direct mmap isn't
1044          * possible (as opposed to tried but failed) so we'll fall
1045          * through to making a private copy of the data and mapping
1046          * that if we can */
1047         return -ENODEV;
1048 }
1049
1050 /*
1051  * set up a private mapping or an anonymous shared mapping
1052  */
1053 static int do_mmap_private(struct vm_area_struct *vma,
1054                            struct vm_region *region,
1055                            unsigned long len,
1056                            unsigned long capabilities)
1057 {
1058         struct page *pages;
1059         unsigned long total, point, n, rlen;
1060         void *base;
1061         int ret, order;
1062
1063         /* invoke the file's mapping function so that it can keep track of
1064          * shared mappings on devices or memory
1065          * - VM_MAYSHARE will be set if it may attempt to share
1066          */
1067         if (capabilities & BDI_CAP_MAP_DIRECT) {
1068                 ret = vma->vm_file->f_op->mmap(vma->vm_file, vma);
1069                 if (ret == 0) {
1070                         /* shouldn't return success if we're not sharing */
1071                         BUG_ON(!(vma->vm_flags & VM_MAYSHARE));
1072                         vma->vm_region->vm_top = vma->vm_region->vm_end;
1073                         return 0;
1074                 }
1075                 if (ret != -ENOSYS)
1076                         return ret;
1077
1078                 /* getting an ENOSYS error indicates that direct mmap isn't
1079                  * possible (as opposed to tried but failed) so we'll try to
1080                  * make a private copy of the data and map that instead */
1081         }
1082
1083         rlen = PAGE_ALIGN(len);
1084
1085         /* allocate some memory to hold the mapping
1086          * - note that this may not return a page-aligned address if the object
1087          *   we're allocating is smaller than a page
1088          */
1089         order = get_order(rlen);
1090         kdebug("alloc order %d for %lx", order, len);
1091
1092         pages = alloc_pages(GFP_KERNEL, order);
1093         if (!pages)
1094                 goto enomem;
1095
1096         total = 1 << order;
1097         atomic_long_add(total, &mmap_pages_allocated);
1098
1099         point = rlen >> PAGE_SHIFT;
1100
1101         /* we allocated a power-of-2 sized page set, so we may want to trim off
1102          * the excess */
1103         if (sysctl_nr_trim_pages && total - point >= sysctl_nr_trim_pages) {
1104                 while (total > point) {
1105                         order = ilog2(total - point);
1106                         n = 1 << order;
1107                         kdebug("shave %lu/%lu @%lu", n, total - point, total);
1108                         atomic_long_sub(n, &mmap_pages_allocated);
1109                         total -= n;
1110                         set_page_refcounted(pages + total);
1111                         __free_pages(pages + total, order);
1112                 }
1113         }
1114
1115         for (point = 1; point < total; point++)
1116                 set_page_refcounted(&pages[point]);
1117
1118         base = page_address(pages);
1119         region->vm_flags = vma->vm_flags |= VM_MAPPED_COPY;
1120         region->vm_start = (unsigned long) base;
1121         region->vm_end   = region->vm_start + rlen;
1122         region->vm_top   = region->vm_start + (total << PAGE_SHIFT);
1123
1124         vma->vm_start = region->vm_start;
1125         vma->vm_end   = region->vm_start + len;
1126
1127         if (vma->vm_file) {
1128                 /* read the contents of a file into the copy */
1129                 mm_segment_t old_fs;
1130                 loff_t fpos;
1131
1132                 fpos = vma->vm_pgoff;
1133                 fpos <<= PAGE_SHIFT;
1134
1135                 old_fs = get_fs();
1136                 set_fs(KERNEL_DS);
1137                 ret = vma->vm_file->f_op->read(vma->vm_file, base, rlen, &fpos);
1138                 set_fs(old_fs);
1139
1140                 if (ret < 0)
1141                         goto error_free;
1142
1143                 /* clear the last little bit */
1144                 if (ret < rlen)
1145                         memset(base + ret, 0, rlen - ret);
1146
1147         }
1148
1149         return 0;
1150
1151 error_free:
1152         free_page_series(region->vm_start, region->vm_end);
1153         region->vm_start = vma->vm_start = 0;
1154         region->vm_end   = vma->vm_end = 0;
1155         region->vm_top   = 0;
1156         return ret;
1157
1158 enomem:
1159         printk("Allocation of length %lu from process %d (%s) failed\n",
1160                len, current->pid, current->comm);
1161         show_free_areas();
1162         return -ENOMEM;
1163 }
1164
1165 /*
1166  * handle mapping creation for uClinux
1167  */
1168 unsigned long do_mmap_pgoff(struct file *file,
1169                             unsigned long addr,
1170                             unsigned long len,
1171                             unsigned long prot,
1172                             unsigned long flags,
1173                             unsigned long pgoff)
1174 {
1175         struct vm_area_struct *vma;
1176         struct vm_region *region;
1177         struct rb_node *rb;
1178         unsigned long capabilities, vm_flags, result;
1179         int ret;
1180
1181         kenter(",%lx,%lx,%lx,%lx,%lx", addr, len, prot, flags, pgoff);
1182
1183         /* decide whether we should attempt the mapping, and if so what sort of
1184          * mapping */
1185         ret = validate_mmap_request(file, addr, len, prot, flags, pgoff,
1186                                     &capabilities);
1187         if (ret < 0) {
1188                 kleave(" = %d [val]", ret);
1189                 return ret;
1190         }
1191
1192         /* we ignore the address hint */
1193         addr = 0;
1194
1195         /* we've determined that we can make the mapping, now translate what we
1196          * now know into VMA flags */
1197         vm_flags = determine_vm_flags(file, prot, flags, capabilities);
1198
1199         /* we're going to need to record the mapping */
1200         region = kmem_cache_zalloc(vm_region_jar, GFP_KERNEL);
1201         if (!region)
1202                 goto error_getting_region;
1203
1204         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
1205         if (!vma)
1206                 goto error_getting_vma;
1207
1208         region->vm_usage = 1;
1209         region->vm_flags = vm_flags;
1210         region->vm_pgoff = pgoff;
1211
1212         INIT_LIST_HEAD(&vma->anon_vma_chain);
1213         vma->vm_flags = vm_flags;
1214         vma->vm_pgoff = pgoff;
1215
1216         if (file) {
1217                 region->vm_file = file;
1218                 get_file(file);
1219                 vma->vm_file = file;
1220                 get_file(file);
1221                 if (vm_flags & VM_EXECUTABLE) {
1222                         added_exe_file_vma(current->mm);
1223                         vma->vm_mm = current->mm;
1224                 }
1225         }
1226
1227         down_write(&nommu_region_sem);
1228
1229         /* if we want to share, we need to check for regions created by other
1230          * mmap() calls that overlap with our proposed mapping
1231          * - we can only share with a superset match on most regular files
1232          * - shared mappings on character devices and memory backed files are
1233          *   permitted to overlap inexactly as far as we are concerned for in
1234          *   these cases, sharing is handled in the driver or filesystem rather
1235          *   than here
1236          */
1237         if (vm_flags & VM_MAYSHARE) {
1238                 struct vm_region *pregion;
1239                 unsigned long pglen, rpglen, pgend, rpgend, start;
1240
1241                 pglen = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1242                 pgend = pgoff + pglen;
1243
1244                 for (rb = rb_first(&nommu_region_tree); rb; rb = rb_next(rb)) {
1245                         pregion = rb_entry(rb, struct vm_region, vm_rb);
1246
1247                         if (!(pregion->vm_flags & VM_MAYSHARE))
1248                                 continue;
1249
1250                         /* search for overlapping mappings on the same file */
1251                         if (pregion->vm_file->f_path.dentry->d_inode !=
1252                             file->f_path.dentry->d_inode)
1253                                 continue;
1254
1255                         if (pregion->vm_pgoff >= pgend)
1256                                 continue;
1257
1258                         rpglen = pregion->vm_end - pregion->vm_start;
1259                         rpglen = (rpglen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1260                         rpgend = pregion->vm_pgoff + rpglen;
1261                         if (pgoff >= rpgend)
1262                                 continue;
1263
1264                         /* handle inexactly overlapping matches between
1265                          * mappings */
1266                         if ((pregion->vm_pgoff != pgoff || rpglen != pglen) &&
1267                             !(pgoff >= pregion->vm_pgoff && pgend <= rpgend)) {
1268                                 /* new mapping is not a subset of the region */
1269                                 if (!(capabilities & BDI_CAP_MAP_DIRECT))
1270                                         goto sharing_violation;
1271                                 continue;
1272                         }
1273
1274                         /* we've found a region we can share */
1275                         pregion->vm_usage++;
1276                         vma->vm_region = pregion;
1277                         start = pregion->vm_start;
1278                         start += (pgoff - pregion->vm_pgoff) << PAGE_SHIFT;
1279                         vma->vm_start = start;
1280                         vma->vm_end = start + len;
1281
1282                         if (pregion->vm_flags & VM_MAPPED_COPY) {
1283                                 kdebug("share copy");
1284                                 vma->vm_flags |= VM_MAPPED_COPY;
1285                         } else {
1286                                 kdebug("share mmap");
1287                                 ret = do_mmap_shared_file(vma);
1288                                 if (ret < 0) {
1289                                         vma->vm_region = NULL;
1290                                         vma->vm_start = 0;
1291                                         vma->vm_end = 0;
1292                                         pregion->vm_usage--;
1293                                         pregion = NULL;
1294                                         goto error_just_free;
1295                                 }
1296                         }
1297                         fput(region->vm_file);
1298                         kmem_cache_free(vm_region_jar, region);
1299                         region = pregion;
1300                         result = start;
1301                         goto share;
1302                 }
1303
1304                 /* obtain the address at which to make a shared mapping
1305                  * - this is the hook for quasi-memory character devices to
1306                  *   tell us the location of a shared mapping
1307                  */
1308                 if (capabilities & BDI_CAP_MAP_DIRECT) {
1309                         addr = file->f_op->get_unmapped_area(file, addr, len,
1310                                                              pgoff, flags);
1311                         if (IS_ERR((void *) addr)) {
1312                                 ret = addr;
1313                                 if (ret != (unsigned long) -ENOSYS)
1314                                         goto error_just_free;
1315
1316                                 /* the driver refused to tell us where to site
1317                                  * the mapping so we'll have to attempt to copy
1318                                  * it */
1319                                 ret = (unsigned long) -ENODEV;
1320                                 if (!(capabilities & BDI_CAP_MAP_COPY))
1321                                         goto error_just_free;
1322
1323                                 capabilities &= ~BDI_CAP_MAP_DIRECT;
1324                         } else {
1325                                 vma->vm_start = region->vm_start = addr;
1326                                 vma->vm_end = region->vm_end = addr + len;
1327                         }
1328                 }
1329         }
1330
1331         vma->vm_region = region;
1332
1333         /* set up the mapping
1334          * - the region is filled in if BDI_CAP_MAP_DIRECT is still set
1335          */
1336         if (file && vma->vm_flags & VM_SHARED)
1337                 ret = do_mmap_shared_file(vma);
1338         else
1339                 ret = do_mmap_private(vma, region, len, capabilities);
1340         if (ret < 0)
1341                 goto error_just_free;
1342         add_nommu_region(region);
1343
1344         /* clear anonymous mappings that don't ask for uninitialized data */
1345         if (!vma->vm_file && !(flags & MAP_UNINITIALIZED))
1346                 memset((void *)region->vm_start, 0,
1347                        region->vm_end - region->vm_start);
1348
1349         /* okay... we have a mapping; now we have to register it */
1350         result = vma->vm_start;
1351
1352         current->mm->total_vm += len >> PAGE_SHIFT;
1353
1354 share:
1355         add_vma_to_mm(current->mm, vma);
1356
1357         /* we flush the region from the icache only when the first executable
1358          * mapping of it is made  */
1359         if (vma->vm_flags & VM_EXEC && !region->vm_icache_flushed) {
1360                 flush_icache_range(region->vm_start, region->vm_end);
1361                 region->vm_icache_flushed = true;
1362         }
1363
1364         up_write(&nommu_region_sem);
1365
1366         kleave(" = %lx", result);
1367         return result;
1368
1369 error_just_free:
1370         up_write(&nommu_region_sem);
1371 error:
1372         if (region->vm_file)
1373                 fput(region->vm_file);
1374         kmem_cache_free(vm_region_jar, region);
1375         if (vma->vm_file)
1376                 fput(vma->vm_file);
1377         if (vma->vm_flags & VM_EXECUTABLE)
1378                 removed_exe_file_vma(vma->vm_mm);
1379         kmem_cache_free(vm_area_cachep, vma);
1380         kleave(" = %d", ret);
1381         return ret;
1382
1383 sharing_violation:
1384         up_write(&nommu_region_sem);
1385         printk(KERN_WARNING "Attempt to share mismatched mappings\n");
1386         ret = -EINVAL;
1387         goto error;
1388
1389 error_getting_vma:
1390         kmem_cache_free(vm_region_jar, region);
1391         printk(KERN_WARNING "Allocation of vma for %lu byte allocation"
1392                " from process %d failed\n",
1393                len, current->pid);
1394         show_free_areas();
1395         return -ENOMEM;
1396
1397 error_getting_region:
1398         printk(KERN_WARNING "Allocation of vm region for %lu byte allocation"
1399                " from process %d failed\n",
1400                len, current->pid);
1401         show_free_areas();
1402         return -ENOMEM;
1403 }
1404 EXPORT_SYMBOL(do_mmap_pgoff);
1405
1406 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1407                 unsigned long, prot, unsigned long, flags,
1408                 unsigned long, fd, unsigned long, pgoff)
1409 {
1410         struct file *file = NULL;
1411         unsigned long retval = -EBADF;
1412
1413         if (!(flags & MAP_ANONYMOUS)) {
1414                 file = fget(fd);
1415                 if (!file)
1416                         goto out;
1417         }
1418
1419         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1420
1421         down_write(&current->mm->mmap_sem);
1422         retval = do_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1423         up_write(&current->mm->mmap_sem);
1424
1425         if (file)
1426                 fput(file);
1427 out:
1428         return retval;
1429 }
1430
1431 /*
1432  * split a vma into two pieces at address 'addr', a new vma is allocated either
1433  * for the first part or the tail.
1434  */
1435 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1436               unsigned long addr, int new_below)
1437 {
1438         struct vm_area_struct *new;
1439         struct vm_region *region;
1440         unsigned long npages;
1441
1442         kenter("");
1443
1444         /* we're only permitted to split anonymous regions (these should have
1445          * only a single usage on the region) */
1446         if (vma->vm_file)
1447                 return -ENOMEM;
1448
1449         if (mm->map_count >= sysctl_max_map_count)
1450                 return -ENOMEM;
1451
1452         region = kmem_cache_alloc(vm_region_jar, GFP_KERNEL);
1453         if (!region)
1454                 return -ENOMEM;
1455
1456         new = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
1457         if (!new) {
1458                 kmem_cache_free(vm_region_jar, region);
1459                 return -ENOMEM;
1460         }
1461
1462         /* most fields are the same, copy all, and then fixup */
1463         *new = *vma;
1464         *region = *vma->vm_region;
1465         new->vm_region = region;
1466
1467         npages = (addr - vma->vm_start) >> PAGE_SHIFT;
1468
1469         if (new_below) {
1470                 region->vm_top = region->vm_end = new->vm_end = addr;
1471         } else {
1472                 region->vm_start = new->vm_start = addr;
1473                 region->vm_pgoff = new->vm_pgoff += npages;
1474         }
1475
1476         if (new->vm_ops && new->vm_ops->open)
1477                 new->vm_ops->open(new);
1478
1479         delete_vma_from_mm(vma);
1480         down_write(&nommu_region_sem);
1481         delete_nommu_region(vma->vm_region);
1482         if (new_below) {
1483                 vma->vm_region->vm_start = vma->vm_start = addr;
1484                 vma->vm_region->vm_pgoff = vma->vm_pgoff += npages;
1485         } else {
1486                 vma->vm_region->vm_end = vma->vm_end = addr;
1487                 vma->vm_region->vm_top = addr;
1488         }
1489         add_nommu_region(vma->vm_region);
1490         add_nommu_region(new->vm_region);
1491         up_write(&nommu_region_sem);
1492         add_vma_to_mm(mm, vma);
1493         add_vma_to_mm(mm, new);
1494         return 0;
1495 }
1496
1497 /*
1498  * shrink a VMA by removing the specified chunk from either the beginning or
1499  * the end
1500  */
1501 static int shrink_vma(struct mm_struct *mm,
1502                       struct vm_area_struct *vma,
1503                       unsigned long from, unsigned long to)
1504 {
1505         struct vm_region *region;
1506
1507         kenter("");
1508
1509         /* adjust the VMA's pointers, which may reposition it in the MM's tree
1510          * and list */
1511         delete_vma_from_mm(vma);
1512         if (from > vma->vm_start)
1513                 vma->vm_end = from;
1514         else
1515                 vma->vm_start = to;
1516         add_vma_to_mm(mm, vma);
1517
1518         /* cut the backing region down to size */
1519         region = vma->vm_region;
1520         BUG_ON(region->vm_usage != 1);
1521
1522         down_write(&nommu_region_sem);
1523         delete_nommu_region(region);
1524         if (from > region->vm_start) {
1525                 to = region->vm_top;
1526                 region->vm_top = region->vm_end = from;
1527         } else {
1528                 region->vm_start = to;
1529         }
1530         add_nommu_region(region);
1531         up_write(&nommu_region_sem);
1532
1533         free_page_series(from, to);
1534         return 0;
1535 }
1536
1537 /*
1538  * release a mapping
1539  * - under NOMMU conditions the chunk to be unmapped must be backed by a single
1540  *   VMA, though it need not cover the whole VMA
1541  */
1542 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len)
1543 {
1544         struct vm_area_struct *vma;
1545         struct rb_node *rb;
1546         unsigned long end = start + len;
1547         int ret;
1548
1549         kenter(",%lx,%zx", start, len);
1550
1551         if (len == 0)
1552                 return -EINVAL;
1553
1554         /* find the first potentially overlapping VMA */
1555         vma = find_vma(mm, start);
1556         if (!vma) {
1557                 static int limit = 0;
1558                 if (limit < 5) {
1559                         printk(KERN_WARNING
1560                                "munmap of memory not mmapped by process %d"
1561                                " (%s): 0x%lx-0x%lx\n",
1562                                current->pid, current->comm,
1563                                start, start + len - 1);
1564                         limit++;
1565                 }
1566                 return -EINVAL;
1567         }
1568
1569         /* we're allowed to split an anonymous VMA but not a file-backed one */
1570         if (vma->vm_file) {
1571                 do {
1572                         if (start > vma->vm_start) {
1573                                 kleave(" = -EINVAL [miss]");
1574                                 return -EINVAL;
1575                         }
1576                         if (end == vma->vm_end)
1577                                 goto erase_whole_vma;
1578                         rb = rb_next(&vma->vm_rb);
1579                         vma = rb_entry(rb, struct vm_area_struct, vm_rb);
1580                 } while (rb);
1581                 kleave(" = -EINVAL [split file]");
1582                 return -EINVAL;
1583         } else {
1584                 /* the chunk must be a subset of the VMA found */
1585                 if (start == vma->vm_start && end == vma->vm_end)
1586                         goto erase_whole_vma;
1587                 if (start < vma->vm_start || end > vma->vm_end) {
1588                         kleave(" = -EINVAL [superset]");
1589                         return -EINVAL;
1590                 }
1591                 if (start & ~PAGE_MASK) {
1592                         kleave(" = -EINVAL [unaligned start]");
1593                         return -EINVAL;
1594                 }
1595                 if (end != vma->vm_end && end & ~PAGE_MASK) {
1596                         kleave(" = -EINVAL [unaligned split]");
1597                         return -EINVAL;
1598                 }
1599                 if (start != vma->vm_start && end != vma->vm_end) {
1600                         ret = split_vma(mm, vma, start, 1);
1601                         if (ret < 0) {
1602                                 kleave(" = %d [split]", ret);
1603                                 return ret;
1604                         }
1605                 }
1606                 return shrink_vma(mm, vma, start, end);
1607         }
1608
1609 erase_whole_vma:
1610         delete_vma_from_mm(vma);
1611         delete_vma(mm, vma);
1612         kleave(" = 0");
1613         return 0;
1614 }
1615 EXPORT_SYMBOL(do_munmap);
1616
1617 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1618 {
1619         int ret;
1620         struct mm_struct *mm = current->mm;
1621
1622         down_write(&mm->mmap_sem);
1623         ret = do_munmap(mm, addr, len);
1624         up_write(&mm->mmap_sem);
1625         return ret;
1626 }
1627
1628 /*
1629  * release all the mappings made in a process's VM space
1630  */
1631 void exit_mmap(struct mm_struct *mm)
1632 {
1633         struct vm_area_struct *vma;
1634
1635         if (!mm)
1636                 return;
1637
1638         kenter("");
1639
1640         mm->total_vm = 0;
1641
1642         while ((vma = mm->mmap)) {
1643                 mm->mmap = vma->vm_next;
1644                 delete_vma_from_mm(vma);
1645                 delete_vma(mm, vma);
1646         }
1647
1648         kleave("");
1649 }
1650
1651 unsigned long do_brk(unsigned long addr, unsigned long len)
1652 {
1653         return -ENOMEM;
1654 }
1655
1656 /*
1657  * expand (or shrink) an existing mapping, potentially moving it at the same
1658  * time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1659  *
1660  * under NOMMU conditions, we only permit changing a mapping's size, and only
1661  * as long as it stays within the region allocated by do_mmap_private() and the
1662  * block is not shareable
1663  *
1664  * MREMAP_FIXED is not supported under NOMMU conditions
1665  */
1666 unsigned long do_mremap(unsigned long addr,
1667                         unsigned long old_len, unsigned long new_len,
1668                         unsigned long flags, unsigned long new_addr)
1669 {
1670         struct vm_area_struct *vma;
1671
1672         /* insanity checks first */
1673         if (old_len == 0 || new_len == 0)
1674                 return (unsigned long) -EINVAL;
1675
1676         if (addr & ~PAGE_MASK)
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 EXPORT_SYMBOL(do_mremap);
1700
1701 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1702                 unsigned long, new_len, unsigned long, flags,
1703                 unsigned long, new_addr)
1704 {
1705         unsigned long ret;
1706
1707         down_write(&current->mm->mmap_sem);
1708         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
1709         up_write(&current->mm->mmap_sem);
1710         return ret;
1711 }
1712
1713 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1714                         unsigned int foll_flags)
1715 {
1716         return NULL;
1717 }
1718
1719 int remap_pfn_range(struct vm_area_struct *vma, unsigned long from,
1720                 unsigned long to, unsigned long size, pgprot_t prot)
1721 {
1722         vma->vm_start = vma->vm_pgoff << PAGE_SHIFT;
1723         return 0;
1724 }
1725 EXPORT_SYMBOL(remap_pfn_range);
1726
1727 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1728                         unsigned long pgoff)
1729 {
1730         unsigned int size = vma->vm_end - vma->vm_start;
1731
1732         if (!(vma->vm_flags & VM_USERMAP))
1733                 return -EINVAL;
1734
1735         vma->vm_start = (unsigned long)(addr + (pgoff << PAGE_SHIFT));
1736         vma->vm_end = vma->vm_start + size;
1737
1738         return 0;
1739 }
1740 EXPORT_SYMBOL(remap_vmalloc_range);
1741
1742 void swap_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
1743 {
1744 }
1745
1746 unsigned long arch_get_unmapped_area(struct file *file, unsigned long addr,
1747         unsigned long len, unsigned long pgoff, unsigned long flags)
1748 {
1749         return -ENOMEM;
1750 }
1751
1752 void arch_unmap_area(struct mm_struct *mm, unsigned long addr)
1753 {
1754 }
1755
1756 void unmap_mapping_range(struct address_space *mapping,
1757                          loff_t const holebegin, loff_t const holelen,
1758                          int even_cows)
1759 {
1760 }
1761 EXPORT_SYMBOL(unmap_mapping_range);
1762
1763 /*
1764  * Check that a process has enough memory to allocate a new virtual
1765  * mapping. 0 means there is enough memory for the allocation to
1766  * succeed and -ENOMEM implies there is not.
1767  *
1768  * We currently support three overcommit policies, which are set via the
1769  * vm.overcommit_memory sysctl.  See Documentation/vm/overcommit-accounting
1770  *
1771  * Strict overcommit modes added 2002 Feb 26 by Alan Cox.
1772  * Additional code 2002 Jul 20 by Robert Love.
1773  *
1774  * cap_sys_admin is 1 if the process has admin privileges, 0 otherwise.
1775  *
1776  * Note this is a helper function intended to be used by LSMs which
1777  * wish to use this logic.
1778  */
1779 int __vm_enough_memory(struct mm_struct *mm, long pages, int cap_sys_admin)
1780 {
1781         unsigned long free, allowed;
1782
1783         vm_acct_memory(pages);
1784
1785         /*
1786          * Sometimes we want to use more memory than we have
1787          */
1788         if (sysctl_overcommit_memory == OVERCOMMIT_ALWAYS)
1789                 return 0;
1790
1791         if (sysctl_overcommit_memory == OVERCOMMIT_GUESS) {
1792                 unsigned long n;
1793
1794                 free = global_page_state(NR_FILE_PAGES);
1795                 free += nr_swap_pages;
1796
1797                 /*
1798                  * Any slabs which are created with the
1799                  * SLAB_RECLAIM_ACCOUNT flag claim to have contents
1800                  * which are reclaimable, under pressure.  The dentry
1801                  * cache and most inode caches should fall into this
1802                  */
1803                 free += global_page_state(NR_SLAB_RECLAIMABLE);
1804
1805                 /*
1806                  * Leave the last 3% for root
1807                  */
1808                 if (!cap_sys_admin)
1809                         free -= free / 32;
1810
1811                 if (free > pages)
1812                         return 0;
1813
1814                 /*
1815                  * nr_free_pages() is very expensive on large systems,
1816                  * only call if we're about to fail.
1817                  */
1818                 n = nr_free_pages();
1819
1820                 /*
1821                  * Leave reserved pages. The pages are not for anonymous pages.
1822                  */
1823                 if (n <= totalreserve_pages)
1824                         goto error;
1825                 else
1826                         n -= totalreserve_pages;
1827
1828                 /*
1829                  * Leave the last 3% for root
1830                  */
1831                 if (!cap_sys_admin)
1832                         n -= n / 32;
1833                 free += n;
1834
1835                 if (free > pages)
1836                         return 0;
1837
1838                 goto error;
1839         }
1840
1841         allowed = totalram_pages * sysctl_overcommit_ratio / 100;
1842         /*
1843          * Leave the last 3% for root
1844          */
1845         if (!cap_sys_admin)
1846                 allowed -= allowed / 32;
1847         allowed += total_swap_pages;
1848
1849         /* Don't let a single process grow too big:
1850            leave 3% of the size of this process for other processes */
1851         if (mm)
1852                 allowed -= mm->total_vm / 32;
1853
1854         if (percpu_counter_read_positive(&vm_committed_as) < allowed)
1855                 return 0;
1856
1857 error:
1858         vm_unacct_memory(pages);
1859
1860         return -ENOMEM;
1861 }
1862
1863 int in_gate_area_no_task(unsigned long addr)
1864 {
1865         return 0;
1866 }
1867
1868 int filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1869 {
1870         BUG();
1871         return 0;
1872 }
1873 EXPORT_SYMBOL(filemap_fault);
1874
1875 /*
1876  * Access another process' address space.
1877  * - source/target buffer must be kernel space
1878  */
1879 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len, int write)
1880 {
1881         struct vm_area_struct *vma;
1882         struct mm_struct *mm;
1883
1884         if (addr + len < addr)
1885                 return 0;
1886
1887         mm = get_task_mm(tsk);
1888         if (!mm)
1889                 return 0;
1890
1891         down_read(&mm->mmap_sem);
1892
1893         /* the access must start within one of the target process's mappings */
1894         vma = find_vma(mm, addr);
1895         if (vma) {
1896                 /* don't overrun this mapping */
1897                 if (addr + len >= vma->vm_end)
1898                         len = vma->vm_end - addr;
1899
1900                 /* only read or write mappings where it is permitted */
1901                 if (write && vma->vm_flags & VM_MAYWRITE)
1902                         copy_to_user_page(vma, NULL, addr,
1903                                          (void *) addr, buf, len);
1904                 else if (!write && vma->vm_flags & VM_MAYREAD)
1905                         copy_from_user_page(vma, NULL, addr,
1906                                             buf, (void *) addr, len);
1907                 else
1908                         len = 0;
1909         } else {
1910                 len = 0;
1911         }
1912
1913         up_read(&mm->mmap_sem);
1914         mmput(mm);
1915         return len;
1916 }
1917
1918 /**
1919  * nommu_shrink_inode_mappings - Shrink the shared mappings on an inode
1920  * @inode: The inode to check
1921  * @size: The current filesize of the inode
1922  * @newsize: The proposed filesize of the inode
1923  *
1924  * Check the shared mappings on an inode on behalf of a shrinking truncate to
1925  * make sure that that any outstanding VMAs aren't broken and then shrink the
1926  * vm_regions that extend that beyond so that do_mmap_pgoff() doesn't
1927  * automatically grant mappings that are too large.
1928  */
1929 int nommu_shrink_inode_mappings(struct inode *inode, size_t size,
1930                                 size_t newsize)
1931 {
1932         struct vm_area_struct *vma;
1933         struct prio_tree_iter iter;
1934         struct vm_region *region;
1935         pgoff_t low, high;
1936         size_t r_size, r_top;
1937
1938         low = newsize >> PAGE_SHIFT;
1939         high = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1940
1941         down_write(&nommu_region_sem);
1942
1943         /* search for VMAs that fall within the dead zone */
1944         vma_prio_tree_foreach(vma, &iter, &inode->i_mapping->i_mmap,
1945                               low, high) {
1946                 /* found one - only interested if it's shared out of the page
1947                  * cache */
1948                 if (vma->vm_flags & VM_SHARED) {
1949                         up_write(&nommu_region_sem);
1950                         return -ETXTBSY; /* not quite true, but near enough */
1951                 }
1952         }
1953
1954         /* reduce any regions that overlap the dead zone - if in existence,
1955          * these will be pointed to by VMAs that don't overlap the dead zone
1956          *
1957          * we don't check for any regions that start beyond the EOF as there
1958          * shouldn't be any
1959          */
1960         vma_prio_tree_foreach(vma, &iter, &inode->i_mapping->i_mmap,
1961                               0, ULONG_MAX) {
1962                 if (!(vma->vm_flags & VM_SHARED))
1963                         continue;
1964
1965                 region = vma->vm_region;
1966                 r_size = region->vm_top - region->vm_start;
1967                 r_top = (region->vm_pgoff << PAGE_SHIFT) + r_size;
1968
1969                 if (r_top > newsize) {
1970                         region->vm_top -= r_top - newsize;
1971                         if (region->vm_end > region->vm_top)
1972                                 region->vm_end = region->vm_top;
1973                 }
1974         }
1975
1976         up_write(&nommu_region_sem);
1977         return 0;
1978 }