mm/nommu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  linux/mm/nommu.c
   4  *
   5  *  Replacement code for mm functions to support CPU's that don't
   6  *  have any form of memory management unit (thus no virtual memory).
   7  *
   8  *  See Documentation/admin-guide/mm/nommu-mmap.rst
   9  *
  10  *  Copyright (c) 2004-2008 David Howells <dhowells@redhat.com>
  11  *  Copyright (c) 2000-2003 David McCullough <davidm@snapgear.com>
  12  *  Copyright (c) 2000-2001 D Jeff Dionne <jeff@uClinux.org>
  13  *  Copyright (c) 2002      Greg Ungerer <gerg@snapgear.com>
  14  *  Copyright (c) 2007-2010 Paul Mundt <lethal@linux-sh.org>
  15  */
  16
  17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  18
  19 #include <linux/export.h>
  20 #include <linux/mm.h>
  21 #include <linux/sched/mm.h>
  22 #include <linux/vmacache.h>
  23 #include <linux/mman.h>
  24 #include <linux/swap.h>
  25 #include <linux/file.h>
  26 #include <linux/highmem.h>
  27 #include <linux/pagemap.h>
  28 #include <linux/slab.h>
  29 #include <linux/vmalloc.h>
  30 #include <linux/blkdev.h>
  31 #include <linux/backing-dev.h>
  32 #include <linux/compiler.h>
  33 #include <linux/mount.h>
  34 #include <linux/personality.h>
  35 #include <linux/security.h>
  36 #include <linux/syscalls.h>
  37 #include <linux/audit.h>
  38 #include <linux/printk.h>
  39
  40 #include <linux/uaccess.h>
  41 #include <asm/tlb.h>
  42 #include <asm/tlbflush.h>
  43 #include <asm/mmu_context.h>
  44 #include "internal.h"
  45
  46 void *high_memory;
  47 EXPORT_SYMBOL(high_memory);
  48 struct page *mem_map;
  49 unsigned long max_mapnr;
  50 EXPORT_SYMBOL(max_mapnr);
  51 unsigned long highest_memmap_pfn;
  52 int sysctl_nr_trim_pages = CONFIG_NOMMU_INITIAL_TRIM_EXCESS;
  53 int heap_stack_gap = 0;
  54
  55 atomic_long_t mmap_pages_allocated;
  56
  57 EXPORT_SYMBOL(mem_map);
  58
  59 /* list of mapped, potentially shareable regions */
  60 static struct kmem_cache *vm_region_jar;
  61 struct rb_root nommu_region_tree = RB_ROOT;
  62 DECLARE_RWSEM(nommu_region_sem);
  63
  64 const struct vm_operations_struct generic_file_vm_ops = {
  65 };
  66
  67 /*
  68  * Return the total memory allocated for this pointer, not
  69  * just what the caller asked for.
  70  *
  71  * Doesn't have to be accurate, i.e. may have races.
  72  */
  73 unsigned int kobjsize(const void *objp)
  74 {
  75         struct page *page;
  76
  77         /*
  78          * If the object we have should not have ksize performed on it,
  79          * return size of 0
  80          */
  81         if (!objp || !virt_addr_valid(objp))
  82                 return 0;
  83
  84         page = virt_to_head_page(objp);
  85
  86         /*
  87          * If the allocator sets PageSlab, we know the pointer came from
  88          * kmalloc().
  89          */
  90         if (PageSlab(page))
  91                 return ksize(objp);
  92
  93         /*
  94          * If it's not a compound page, see if we have a matching VMA
  95          * region. This test is intentionally done in reverse order,
  96          * so if there's no VMA, we still fall through and hand back
  97          * PAGE_SIZE for 0-order pages.
  98          */
  99         if (!PageCompound(page)) {
 100                 struct vm_area_struct *vma;
 101
 102                 vma = find_vma(current->mm, (unsigned long)objp);
 103                 if (vma)
 104                         return vma->vm_end - vma->vm_start;
 105         }
 106
 107         /*
 108          * The ksize() function is only guaranteed to work for pointers
 109          * returned by kmalloc(). So handle arbitrary pointers here.
 110          */
 111         return page_size(page);
 112 }
 113
 114 /**
 115  * follow_pfn - look up PFN at a user virtual address
 116  * @vma: memory mapping
 117  * @address: user virtual address
 118  * @pfn: location to store found PFN
 119  *
 120  * Only IO mappings and raw PFN mappings are allowed.
 121  *
 122  * Returns zero and the pfn at @pfn on success, -ve otherwise.
 123  */
 124 int follow_pfn(struct vm_area_struct *vma, unsigned long address,
 125         unsigned long *pfn)
 126 {
 127         if (!(vma->vm_flags & (VM_IO | VM_PFNMAP)))
 128                 return -EINVAL;
 129
 130         *pfn = address >> PAGE_SHIFT;
 131         return 0;
 132 }
 133 EXPORT_SYMBOL(follow_pfn);
 134
 135 LIST_HEAD(vmap_area_list);
 136
 137 void vfree(const void *addr)
 138 {
 139         kfree(addr);
 140 }
 141 EXPORT_SYMBOL(vfree);
 142
 143 void *__vmalloc(unsigned long size, gfp_t gfp_mask)
 144 {
 145         /*
 146          *  You can't specify __GFP_HIGHMEM with kmalloc() since kmalloc()
 147          * returns only a logical address.
 148          */
 149         return kmalloc(size, (gfp_mask | __GFP_COMP) & ~__GFP_HIGHMEM);
 150 }
 151 EXPORT_SYMBOL(__vmalloc);
 152
 153 void *__vmalloc_node_range(unsigned long size, unsigned long align,
 154                 unsigned long start, unsigned long end, gfp_t gfp_mask,
 155                 pgprot_t prot, unsigned long vm_flags, int node,
 156                 const void *caller)
 157 {
 158         return __vmalloc(size, gfp_mask);
 159 }
 160
 161 void *__vmalloc_node(unsigned long size, unsigned long align, gfp_t gfp_mask,
 162                 int node, const void *caller)
 163 {
 164         return __vmalloc(size, gfp_mask);
 165 }
 166
 167 static void *__vmalloc_user_flags(unsigned long size, gfp_t flags)
 168 {
 169         void *ret;
 170
 171         ret = __vmalloc(size, flags);
 172         if (ret) {
 173                 struct vm_area_struct *vma;
 174
 175                 mmap_write_lock(current->mm);
 176                 vma = find_vma(current->mm, (unsigned long)ret);
 177                 if (vma)
 178                         vma->vm_flags |= VM_USERMAP;
 179                 mmap_write_unlock(current->mm);
 180         }
 181
 182         return ret;
 183 }
 184
 185 void *vmalloc_user(unsigned long size)
 186 {
 187         return __vmalloc_user_flags(size, GFP_KERNEL | __GFP_ZERO);
 188 }
 189 EXPORT_SYMBOL(vmalloc_user);
 190
 191 struct page *vmalloc_to_page(const void *addr)
 192 {
 193         return virt_to_page(addr);
 194 }
 195 EXPORT_SYMBOL(vmalloc_to_page);
 196
 197 unsigned long vmalloc_to_pfn(const void *addr)
 198 {
 199         return page_to_pfn(virt_to_page(addr));
 200 }
 201 EXPORT_SYMBOL(vmalloc_to_pfn);
 202
 203 long vread(char *buf, char *addr, unsigned long count)
 204 {
 205         /* Don't allow overflow */
 206         if ((unsigned long) buf + count < count)
 207                 count = -(unsigned long) buf;
 208
 209         memcpy(buf, addr, count);
 210         return count;
 211 }
 212
 213 long vwrite(char *buf, char *addr, unsigned long count)
 214 {
 215         /* Don't allow overflow */
 216         if ((unsigned long) addr + count < count)
 217                 count = -(unsigned long) addr;
 218
 219         memcpy(addr, buf, count);
 220         return count;
 221 }
 222
 223 /*
 224  *      vmalloc  -  allocate virtually contiguous memory
 225  *
 226  *      @size:          allocation size
 227  *
 228  *      Allocate enough pages to cover @size from the page level
 229  *      allocator and map them into contiguous kernel virtual space.
 230  *
 231  *      For tight control over page level allocator and protection flags
 232  *      use __vmalloc() instead.
 233  */
 234 void *vmalloc(unsigned long size)
 235 {
 236        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM);
 237 }
 238 EXPORT_SYMBOL(vmalloc);
 239
 240 /*
 241  *      vzalloc - allocate virtually contiguous memory with zero fill
 242  *
 243  *      @size:          allocation size
 244  *
 245  *      Allocate enough pages to cover @size from the page level
 246  *      allocator and map them into contiguous kernel virtual space.
 247  *      The memory allocated is set to zero.
 248  *
 249  *      For tight control over page level allocator and protection flags
 250  *      use __vmalloc() instead.
 251  */
 252 void *vzalloc(unsigned long size)
 253 {
 254         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM | __GFP_ZERO);
 255 }
 256 EXPORT_SYMBOL(vzalloc);
 257
 258 /**
 259  * vmalloc_node - allocate memory on a specific node
 260  * @size:       allocation size
 261  * @node:       numa node
 262  *
 263  * Allocate enough pages to cover @size from the page level
 264  * allocator and map them into contiguous kernel virtual space.
 265  *
 266  * For tight control over page level allocator and protection flags
 267  * use __vmalloc() instead.
 268  */
 269 void *vmalloc_node(unsigned long size, int node)
 270 {
 271         return vmalloc(size);
 272 }
 273 EXPORT_SYMBOL(vmalloc_node);
 274
 275 /**
 276  * vzalloc_node - allocate memory on a specific node with zero fill
 277  * @size:       allocation size
 278  * @node:       numa node
 279  *
 280  * Allocate enough pages to cover @size from the page level
 281  * allocator and map them into contiguous kernel virtual space.
 282  * The memory allocated is set to zero.
 283  *
 284  * For tight control over page level allocator and protection flags
 285  * use __vmalloc() instead.
 286  */
 287 void *vzalloc_node(unsigned long size, int node)
 288 {
 289         return vzalloc(size);
 290 }
 291 EXPORT_SYMBOL(vzalloc_node);
 292
 293 /**
 294  * vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
 295  *      @size:          allocation size
 296  *
 297  *      Allocate enough 32bit PA addressable pages to cover @size from the
 298  *      page level allocator and map them into contiguous kernel virtual space.
 299  */
 300 void *vmalloc_32(unsigned long size)
 301 {
 302         return __vmalloc(size, GFP_KERNEL);
 303 }
 304 EXPORT_SYMBOL(vmalloc_32);
 305
 306 /**
 307  * vmalloc_32_user - allocate zeroed virtually contiguous 32bit memory
 308  *      @size:          allocation size
 309  *
 310  * The resulting memory area is 32bit addressable and zeroed so it can be
 311  * mapped to userspace without leaking data.
 312  *
 313  * VM_USERMAP is set on the corresponding VMA so that subsequent calls to
 314  * remap_vmalloc_range() are permissible.
 315  */
 316 void *vmalloc_32_user(unsigned long size)
 317 {
 318         /*
 319          * We'll have to sort out the ZONE_DMA bits for 64-bit,
 320          * but for now this can simply use vmalloc_user() directly.
 321          */
 322         return vmalloc_user(size);
 323 }
 324 EXPORT_SYMBOL(vmalloc_32_user);
 325
 326 void *vmap(struct page **pages, unsigned int count, unsigned long flags, pgprot_t prot)
 327 {
 328         BUG();
 329         return NULL;
 330 }
 331 EXPORT_SYMBOL(vmap);
 332
 333 void vunmap(const void *addr)
 334 {
 335         BUG();
 336 }
 337 EXPORT_SYMBOL(vunmap);
 338
 339 void *vm_map_ram(struct page **pages, unsigned int count, int node)
 340 {
 341         BUG();
 342         return NULL;
 343 }
 344 EXPORT_SYMBOL(vm_map_ram);
 345
 346 void vm_unmap_ram(const void *mem, unsigned int count)
 347 {
 348         BUG();
 349 }
 350 EXPORT_SYMBOL(vm_unmap_ram);
 351
 352 void vm_unmap_aliases(void)
 353 {
 354 }
 355 EXPORT_SYMBOL_GPL(vm_unmap_aliases);
 356
 357 void free_vm_area(struct vm_struct *area)
 358 {
 359         BUG();
 360 }
 361 EXPORT_SYMBOL_GPL(free_vm_area);
 362
 363 int vm_insert_page(struct vm_area_struct *vma, unsigned long addr,
 364                    struct page *page)
 365 {
 366         return -EINVAL;
 367 }
 368 EXPORT_SYMBOL(vm_insert_page);
 369
 370 int vm_map_pages(struct vm_area_struct *vma, struct page **pages,
 371                         unsigned long num)
 372 {
 373         return -EINVAL;
 374 }
 375 EXPORT_SYMBOL(vm_map_pages);
 376
 377 int vm_map_pages_zero(struct vm_area_struct *vma, struct page **pages,
 378                                 unsigned long num)
 379 {
 380         return -EINVAL;
 381 }
 382 EXPORT_SYMBOL(vm_map_pages_zero);
 383
 384 /*
 385  *  sys_brk() for the most part doesn't need the global kernel
 386  *  lock, except when an application is doing something nasty
 387  *  like trying to un-brk an area that has already been mapped
 388  *  to a regular file.  in this case, the unmapping will need
 389  *  to invoke file system routines that need the global lock.
 390  */
 391 SYSCALL_DEFINE1(brk, unsigned long, brk)
 392 {
 393         struct mm_struct *mm = current->mm;
 394
 395         if (brk < mm->start_brk || brk > mm->context.end_brk)
 396                 return mm->brk;
 397
 398         if (mm->brk == brk)
 399                 return mm->brk;
 400
 401         /*
 402          * Always allow shrinking brk
 403          */
 404         if (brk <= mm->brk) {
 405                 mm->brk = brk;
 406                 return brk;
 407         }
 408
 409         /*
 410          * Ok, looks good - let it rip.
 411          */
 412         flush_icache_user_range(mm->brk, brk);
 413         return mm->brk = brk;
 414 }
 415
 416 /*
 417  * initialise the percpu counter for VM and region record slabs
 418  */
 419 void __init mmap_init(void)
 420 {
 421         int ret;
 422
 423         ret = percpu_counter_init(&vm_committed_as, 0, GFP_KERNEL);
 424         VM_BUG_ON(ret);
 425         vm_region_jar = KMEM_CACHE(vm_region, SLAB_PANIC|SLAB_ACCOUNT);
 426 }
 427
 428 /*
 429  * validate the region tree
 430  * - the caller must hold the region lock
 431  */
 432 #ifdef CONFIG_DEBUG_NOMMU_REGIONS
 433 static noinline void validate_nommu_regions(void)
 434 {
 435         struct vm_region *region, *last;
 436         struct rb_node *p, *lastp;
 437
 438         lastp = rb_first(&nommu_region_tree);
 439         if (!lastp)
 440                 return;
 441
 442         last = rb_entry(lastp, struct vm_region, vm_rb);
 443         BUG_ON(last->vm_end <= last->vm_start);
 444         BUG_ON(last->vm_top < last->vm_end);
 445
 446         while ((p = rb_next(lastp))) {
 447                 region = rb_entry(p, struct vm_region, vm_rb);
 448                 last = rb_entry(lastp, struct vm_region, vm_rb);
 449
 450                 BUG_ON(region->vm_end <= region->vm_start);
 451                 BUG_ON(region->vm_top < region->vm_end);
 452                 BUG_ON(region->vm_start < last->vm_top);
 453
 454                 lastp = p;
 455         }
 456 }
 457 #else
 458 static void validate_nommu_regions(void)
 459 {
 460 }
 461 #endif
 462
 463 /*
 464  * add a region into the global tree
 465  */
 466 static void add_nommu_region(struct vm_region *region)
 467 {
 468         struct vm_region *pregion;
 469         struct rb_node **p, *parent;
 470
 471         validate_nommu_regions();
 472
 473         parent = NULL;
 474         p = &nommu_region_tree.rb_node;
 475         while (*p) {
 476                 parent = *p;
 477                 pregion = rb_entry(parent, struct vm_region, vm_rb);
 478                 if (region->vm_start < pregion->vm_start)
 479                         p = &(*p)->rb_left;
 480                 else if (region->vm_start > pregion->vm_start)
 481                         p = &(*p)->rb_right;
 482                 else if (pregion == region)
 483                         return;
 484                 else
 485                         BUG();
 486         }
 487
 488         rb_link_node(&region->vm_rb, parent, p);
 489         rb_insert_color(&region->vm_rb, &nommu_region_tree);
 490
 491         validate_nommu_regions();
 492 }
 493
 494 /*
 495  * delete a region from the global tree
 496  */
 497 static void delete_nommu_region(struct vm_region *region)
 498 {
 499         BUG_ON(!nommu_region_tree.rb_node);
 500
 501         validate_nommu_regions();
 502         rb_erase(&region->vm_rb, &nommu_region_tree);
 503         validate_nommu_regions();
 504 }
 505
 506 /*
 507  * free a contiguous series of pages
 508  */
 509 static void free_page_series(unsigned long from, unsigned long to)
 510 {
 511         for (; from < to; from += PAGE_SIZE) {
 512                 struct page *page = virt_to_page(from);
 513
 514                 atomic_long_dec(&mmap_pages_allocated);
 515                 put_page(page);
 516         }
 517 }
 518
 519 /*
 520  * release a reference to a region
 521  * - the caller must hold the region semaphore for writing, which this releases
 522  * - the region may not have been added to the tree yet, in which case vm_top
 523  *   will equal vm_start
 524  */
 525 static void __put_nommu_region(struct vm_region *region)
 526         __releases(nommu_region_sem)
 527 {
 528         BUG_ON(!nommu_region_tree.rb_node);
 529
 530         if (--region->vm_usage == 0) {
 531                 if (region->vm_top > region->vm_start)
 532                         delete_nommu_region(region);
 533                 up_write(&nommu_region_sem);
 534
 535                 if (region->vm_file)
 536                         fput(region->vm_file);
 537
 538                 /* IO memory and memory shared directly out of the pagecache
 539                  * from ramfs/tmpfs mustn't be released here */
 540                 if (region->vm_flags & VM_MAPPED_COPY)
 541                         free_page_series(region->vm_start, region->vm_top);
 542                 kmem_cache_free(vm_region_jar, region);
 543         } else {
 544                 up_write(&nommu_region_sem);
 545         }
 546 }
 547
 548 /*
 549  * release a reference to a region
 550  */
 551 static void put_nommu_region(struct vm_region *region)
 552 {
 553         down_write(&nommu_region_sem);
 554         __put_nommu_region(region);
 555 }
 556
 557 /*
 558  * add a VMA into a process's mm_struct in the appropriate place in the list
 559  * and tree and add to the address space's page tree also if not an anonymous
 560  * page
 561  * - should be called with mm->mmap_lock held writelocked
 562  */
 563 static void add_vma_to_mm(struct mm_struct *mm, struct vm_area_struct *vma)
 564 {
 565         struct vm_area_struct *pvma, *prev;
 566         struct address_space *mapping;
 567         struct rb_node **p, *parent, *rb_prev;
 568
 569         BUG_ON(!vma->vm_region);
 570
 571         mm->map_count++;
 572         vma->vm_mm = mm;
 573
 574         /* add the VMA to the mapping */
 575         if (vma->vm_file) {
 576                 mapping = vma->vm_file->f_mapping;
 577
 578                 i_mmap_lock_write(mapping);
 579                 flush_dcache_mmap_lock(mapping);
 580                 vma_interval_tree_insert(vma, &mapping->i_mmap);
 581                 flush_dcache_mmap_unlock(mapping);
 582                 i_mmap_unlock_write(mapping);
 583         }
 584
 585         /* add the VMA to the tree */
 586         parent = rb_prev = NULL;
 587         p = &mm->mm_rb.rb_node;
 588         while (*p) {
 589                 parent = *p;
 590                 pvma = rb_entry(parent, struct vm_area_struct, vm_rb);
 591
 592                 /* sort by: start addr, end addr, VMA struct addr in that order
 593                  * (the latter is necessary as we may get identical VMAs) */
 594                 if (vma->vm_start < pvma->vm_start)
 595                         p = &(*p)->rb_left;
 596                 else if (vma->vm_start > pvma->vm_start) {
 597                         rb_prev = parent;
 598                         p = &(*p)->rb_right;
 599                 } else if (vma->vm_end < pvma->vm_end)
 600                         p = &(*p)->rb_left;
 601                 else if (vma->vm_end > pvma->vm_end) {
 602                         rb_prev = parent;
 603                         p = &(*p)->rb_right;
 604                 } else if (vma < pvma)
 605                         p = &(*p)->rb_left;
 606                 else if (vma > pvma) {
 607                         rb_prev = parent;
 608                         p = &(*p)->rb_right;
 609                 } else
 610                         BUG();
 611         }
 612
 613         rb_link_node(&vma->vm_rb, parent, p);
 614         rb_insert_color(&vma->vm_rb, &mm->mm_rb);
 615
 616         /* add VMA to the VMA list also */
 617         prev = NULL;
 618         if (rb_prev)
 619                 prev = rb_entry(rb_prev, struct vm_area_struct, vm_rb);
 620
 621         __vma_link_list(mm, vma, prev);
 622 }
 623
 624 /*
 625  * delete a VMA from its owning mm_struct and address space
 626  */
 627 static void delete_vma_from_mm(struct vm_area_struct *vma)
 628 {
 629         int i;
 630         struct address_space *mapping;
 631         struct mm_struct *mm = vma->vm_mm;
 632         struct task_struct *curr = current;
 633
 634         mm->map_count--;
 635         for (i = 0; i < VMACACHE_SIZE; i++) {
 636                 /* if the vma is cached, invalidate the entire cache */
 637                 if (curr->vmacache.vmas[i] == vma) {
 638                         vmacache_invalidate(mm);
 639                         break;
 640                 }
 641         }
 642
 643         /* remove the VMA from the mapping */
 644         if (vma->vm_file) {
 645                 mapping = vma->vm_file->f_mapping;
 646
 647                 i_mmap_lock_write(mapping);
 648                 flush_dcache_mmap_lock(mapping);
 649                 vma_interval_tree_remove(vma, &mapping->i_mmap);
 650                 flush_dcache_mmap_unlock(mapping);
 651                 i_mmap_unlock_write(mapping);
 652         }
 653
 654         /* remove from the MM's tree and list */
 655         rb_erase(&vma->vm_rb, &mm->mm_rb);
 656
 657         __vma_unlink_list(mm, vma);
 658 }
 659
 660 /*
 661  * destroy a VMA record
 662  */
 663 static void delete_vma(struct mm_struct *mm, struct vm_area_struct *vma)
 664 {
 665         if (vma->vm_ops && vma->vm_ops->close)
 666                 vma->vm_ops->close(vma);
 667         if (vma->vm_file)
 668                 fput(vma->vm_file);
 669         put_nommu_region(vma->vm_region);
 670         vm_area_free(vma);
 671 }
 672
 673 /*
 674  * look up the first VMA in which addr resides, NULL if none
 675  * - should be called with mm->mmap_lock at least held readlocked
 676  */
 677 struct vm_area_struct *find_vma(struct mm_struct *mm, unsigned long addr)
 678 {
 679         struct vm_area_struct *vma;
 680
 681         /* check the cache first */
 682         vma = vmacache_find(mm, addr);
 683         if (likely(vma))
 684                 return vma;
 685
 686         /* trawl the list (there may be multiple mappings in which addr
 687          * resides) */
 688         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 689                 if (vma->vm_start > addr)
 690                         return NULL;
 691                 if (vma->vm_end > addr) {
 692                         vmacache_update(addr, vma);
 693                         return vma;
 694                 }
 695         }
 696
 697         return NULL;
 698 }
 699 EXPORT_SYMBOL(find_vma);
 700
 701 /*
 702  * find a VMA
 703  * - we don't extend stack VMAs under NOMMU conditions
 704  */
 705 struct vm_area_struct *find_extend_vma(struct mm_struct *mm, unsigned long addr)
 706 {
 707         return find_vma(mm, addr);
 708 }
 709
 710 /*
 711  * expand a stack to a given address
 712  * - not supported under NOMMU conditions
 713  */
 714 int expand_stack(struct vm_area_struct *vma, unsigned long address)
 715 {
 716         return -ENOMEM;
 717 }
 718
 719 /*
 720  * look up the first VMA exactly that exactly matches addr
 721  * - should be called with mm->mmap_lock at least held readlocked
 722  */
 723 static struct vm_area_struct *find_vma_exact(struct mm_struct *mm,
 724                                              unsigned long addr,
 725                                              unsigned long len)
 726 {
 727         struct vm_area_struct *vma;
 728         unsigned long end = addr + len;
 729
 730         /* check the cache first */
 731         vma = vmacache_find_exact(mm, addr, end);
 732         if (vma)
 733                 return vma;
 734
 735         /* trawl the list (there may be multiple mappings in which addr
 736          * resides) */
 737         for (vma = mm->mmap; vma; vma = vma->vm_next) {
 738                 if (vma->vm_start < addr)
 739                         continue;
 740                 if (vma->vm_start > addr)
 741                         return NULL;
 742                 if (vma->vm_end == end) {
 743                         vmacache_update(addr, vma);
 744                         return vma;
 745                 }
 746         }
 747
 748         return NULL;
 749 }
 750
 751 /*
 752  * determine whether a mapping should be permitted and, if so, what sort of
 753  * mapping we're capable of supporting
 754  */
 755 static int validate_mmap_request(struct file *file,
 756                                  unsigned long addr,
 757                                  unsigned long len,
 758                                  unsigned long prot,
 759                                  unsigned long flags,
 760                                  unsigned long pgoff,
 761                                  unsigned long *_capabilities)
 762 {
 763         unsigned long capabilities, rlen;
 764         int ret;
 765
 766         /* do the simple checks first */
 767         if (flags & MAP_FIXED)
 768                 return -EINVAL;
 769
 770         if ((flags & MAP_TYPE) != MAP_PRIVATE &&
 771             (flags & MAP_TYPE) != MAP_SHARED)
 772                 return -EINVAL;
 773
 774         if (!len)
 775                 return -EINVAL;
 776
 777         /* Careful about overflows.. */
 778         rlen = PAGE_ALIGN(len);
 779         if (!rlen || rlen > TASK_SIZE)
 780                 return -ENOMEM;
 781
 782         /* offset overflow? */
 783         if ((pgoff + (rlen >> PAGE_SHIFT)) < pgoff)
 784                 return -EOVERFLOW;
 785
 786         if (file) {
 787                 /* files must support mmap */
 788                 if (!file->f_op->mmap)
 789                         return -ENODEV;
 790
 791                 /* work out if what we've got could possibly be shared
 792                  * - we support chardevs that provide their own "memory"
 793                  * - we support files/blockdevs that are memory backed
 794                  */
 795                 if (file->f_op->mmap_capabilities) {
 796                         capabilities = file->f_op->mmap_capabilities(file);
 797                 } else {
 798                         /* no explicit capabilities set, so assume some
 799                          * defaults */
 800                         switch (file_inode(file)->i_mode & S_IFMT) {
 801                         case S_IFREG:
 802                         case S_IFBLK:
 803                                 capabilities = NOMMU_MAP_COPY;
 804                                 break;
 805
 806                         case S_IFCHR:
 807                                 capabilities =
 808                                         NOMMU_MAP_DIRECT |
 809                                         NOMMU_MAP_READ |
 810                                         NOMMU_MAP_WRITE;
 811                                 break;
 812
 813                         default:
 814                                 return -EINVAL;
 815                         }
 816                 }
 817
 818                 /* eliminate any capabilities that we can't support on this
 819                  * device */
 820                 if (!file->f_op->get_unmapped_area)
 821                         capabilities &= ~NOMMU_MAP_DIRECT;
 822                 if (!(file->f_mode & FMODE_CAN_READ))
 823                         capabilities &= ~NOMMU_MAP_COPY;
 824
 825                 /* The file shall have been opened with read permission. */
 826                 if (!(file->f_mode & FMODE_READ))
 827                         return -EACCES;
 828
 829                 if (flags & MAP_SHARED) {
 830                         /* do checks for writing, appending and locking */
 831                         if ((prot & PROT_WRITE) &&
 832                             !(file->f_mode & FMODE_WRITE))
 833                                 return -EACCES;
 834
 835                         if (IS_APPEND(file_inode(file)) &&
 836                             (file->f_mode & FMODE_WRITE))
 837                                 return -EACCES;
 838
 839                         if (locks_verify_locked(file))
 840                                 return -EAGAIN;
 841
 842                         if (!(capabilities & NOMMU_MAP_DIRECT))
 843                                 return -ENODEV;
 844
 845                         /* we mustn't privatise shared mappings */
 846                         capabilities &= ~NOMMU_MAP_COPY;
 847                 } else {
 848                         /* we're going to read the file into private memory we
 849                          * allocate */
 850                         if (!(capabilities & NOMMU_MAP_COPY))
 851                                 return -ENODEV;
 852
 853                         /* we don't permit a private writable mapping to be
 854                          * shared with the backing device */
 855                         if (prot & PROT_WRITE)
 856                                 capabilities &= ~NOMMU_MAP_DIRECT;
 857                 }
 858
 859                 if (capabilities & NOMMU_MAP_DIRECT) {
 860                         if (((prot & PROT_READ)  && !(capabilities & NOMMU_MAP_READ))  ||
 861                             ((prot & PROT_WRITE) && !(capabilities & NOMMU_MAP_WRITE)) ||
 862                             ((prot & PROT_EXEC)  && !(capabilities & NOMMU_MAP_EXEC))
 863                             ) {
 864                                 capabilities &= ~NOMMU_MAP_DIRECT;
 865                                 if (flags & MAP_SHARED) {
 866                                         pr_warn("MAP_SHARED not completely supported on !MMU\n");
 867                                         return -EINVAL;
 868                                 }
 869                         }
 870                 }
 871
 872                 /* handle executable mappings and implied executable
 873                  * mappings */
 874                 if (path_noexec(&file->f_path)) {
 875                         if (prot & PROT_EXEC)
 876                                 return -EPERM;
 877                 } else if ((prot & PROT_READ) && !(prot & PROT_EXEC)) {
 878                         /* handle implication of PROT_EXEC by PROT_READ */
 879                         if (current->personality & READ_IMPLIES_EXEC) {
 880                                 if (capabilities & NOMMU_MAP_EXEC)
 881                                         prot |= PROT_EXEC;
 882                         }
 883                 } else if ((prot & PROT_READ) &&
 884                          (prot & PROT_EXEC) &&
 885                          !(capabilities & NOMMU_MAP_EXEC)
 886                          ) {
 887                         /* backing file is not executable, try to copy */
 888                         capabilities &= ~NOMMU_MAP_DIRECT;
 889                 }
 890         } else {
 891                 /* anonymous mappings are always memory backed and can be
 892                  * privately mapped
 893                  */
 894                 capabilities = NOMMU_MAP_COPY;
 895
 896                 /* handle PROT_EXEC implication by PROT_READ */
 897                 if ((prot & PROT_READ) &&
 898                     (current->personality & READ_IMPLIES_EXEC))
 899                         prot |= PROT_EXEC;
 900         }
 901
 902         /* allow the security API to have its say */
 903         ret = security_mmap_addr(addr);
 904         if (ret < 0)
 905                 return ret;
 906
 907         /* looks okay */
 908         *_capabilities = capabilities;
 909         return 0;
 910 }
 911
 912 /*
 913  * we've determined that we can make the mapping, now translate what we
 914  * now know into VMA flags
 915  */
 916 static unsigned long determine_vm_flags(struct file *file,
 917                                         unsigned long prot,
 918                                         unsigned long flags,
 919                                         unsigned long capabilities)
 920 {
 921         unsigned long vm_flags;
 922
 923         vm_flags = calc_vm_prot_bits(prot, 0) | calc_vm_flag_bits(flags);
 924         /* vm_flags |= mm->def_flags; */
 925
 926         if (!(capabilities & NOMMU_MAP_DIRECT)) {
 927                 /* attempt to share read-only copies of mapped file chunks */
 928                 vm_flags |= VM_MAYREAD | VM_MAYWRITE | VM_MAYEXEC;
 929                 if (file && !(prot & PROT_WRITE))
 930                         vm_flags |= VM_MAYSHARE;
 931         } else {
 932                 /* overlay a shareable mapping on the backing device or inode
 933                  * if possible - used for chardevs, ramfs/tmpfs/shmfs and
 934                  * romfs/cramfs */
 935                 vm_flags |= VM_MAYSHARE | (capabilities & NOMMU_VMFLAGS);
 936                 if (flags & MAP_SHARED)
 937                         vm_flags |= VM_SHARED;
 938         }
 939
 940         /* refuse to let anyone share private mappings with this process if
 941          * it's being traced - otherwise breakpoints set in it may interfere
 942          * with another untraced process
 943          */
 944         if ((flags & MAP_PRIVATE) && current->ptrace)
 945                 vm_flags &= ~VM_MAYSHARE;
 946
 947         return vm_flags;
 948 }
 949
 950 /*
 951  * set up a shared mapping on a file (the driver or filesystem provides and
 952  * pins the storage)
 953  */
 954 static int do_mmap_shared_file(struct vm_area_struct *vma)
 955 {
 956         int ret;
 957
 958         ret = call_mmap(vma->vm_file, vma);
 959         if (ret == 0) {
 960                 vma->vm_region->vm_top = vma->vm_region->vm_end;
 961                 return 0;
 962         }
 963         if (ret != -ENOSYS)
 964                 return ret;
 965
 966         /* getting -ENOSYS indicates that direct mmap isn't possible (as
 967          * opposed to tried but failed) so we can only give a suitable error as
 968          * it's not possible to make a private copy if MAP_SHARED was given */
 969         return -ENODEV;
 970 }
 971
 972 /*
 973  * set up a private mapping or an anonymous shared mapping
 974  */
 975 static int do_mmap_private(struct vm_area_struct *vma,
 976                            struct vm_region *region,
 977                            unsigned long len,
 978                            unsigned long capabilities)
 979 {
 980         unsigned long total, point;
 981         void *base;
 982         int ret, order;
 983
 984         /* invoke the file's mapping function so that it can keep track of
 985          * shared mappings on devices or memory
 986          * - VM_MAYSHARE will be set if it may attempt to share
 987          */
 988         if (capabilities & NOMMU_MAP_DIRECT) {
 989                 ret = call_mmap(vma->vm_file, vma);
 990                 if (ret == 0) {
 991                         /* shouldn't return success if we're not sharing */
 992                         BUG_ON(!(vma->vm_flags & VM_MAYSHARE));
 993                         vma->vm_region->vm_top = vma->vm_region->vm_end;
 994                         return 0;
 995                 }
 996                 if (ret != -ENOSYS)
 997                         return ret;
 998
 999                 /* getting an ENOSYS error indicates that direct mmap isn't
1000                  * possible (as opposed to tried but failed) so we'll try to
1001                  * make a private copy of the data and map that instead */
1002         }
1003
1004
1005         /* allocate some memory to hold the mapping
1006          * - note that this may not return a page-aligned address if the object
1007          *   we're allocating is smaller than a page
1008          */
1009         order = get_order(len);
1010         total = 1 << order;
1011         point = len >> PAGE_SHIFT;
1012
1013         /* we don't want to allocate a power-of-2 sized page set */
1014         if (sysctl_nr_trim_pages && total - point >= sysctl_nr_trim_pages)
1015                 total = point;
1016
1017         base = alloc_pages_exact(total << PAGE_SHIFT, GFP_KERNEL);
1018         if (!base)
1019                 goto enomem;
1020
1021         atomic_long_add(total, &mmap_pages_allocated);
1022
1023         region->vm_flags = vma->vm_flags |= VM_MAPPED_COPY;
1024         region->vm_start = (unsigned long) base;
1025         region->vm_end   = region->vm_start + len;
1026         region->vm_top   = region->vm_start + (total << PAGE_SHIFT);
1027
1028         vma->vm_start = region->vm_start;
1029         vma->vm_end   = region->vm_start + len;
1030
1031         if (vma->vm_file) {
1032                 /* read the contents of a file into the copy */
1033                 loff_t fpos;
1034
1035                 fpos = vma->vm_pgoff;
1036                 fpos <<= PAGE_SHIFT;
1037
1038                 ret = kernel_read(vma->vm_file, base, len, &fpos);
1039                 if (ret < 0)
1040                         goto error_free;
1041
1042                 /* clear the last little bit */
1043                 if (ret < len)
1044                         memset(base + ret, 0, len - ret);
1045
1046         } else {
1047                 vma_set_anonymous(vma);
1048         }
1049
1050         return 0;
1051
1052 error_free:
1053         free_page_series(region->vm_start, region->vm_top);
1054         region->vm_start = vma->vm_start = 0;
1055         region->vm_end   = vma->vm_end = 0;
1056         region->vm_top   = 0;
1057         return ret;
1058
1059 enomem:
1060         pr_err("Allocation of length %lu from process %d (%s) failed\n",
1061                len, current->pid, current->comm);
1062         show_free_areas(0, NULL);
1063         return -ENOMEM;
1064 }
1065
1066 /*
1067  * handle mapping creation for uClinux
1068  */
1069 unsigned long do_mmap(struct file *file,
1070                         unsigned long addr,
1071                         unsigned long len,
1072                         unsigned long prot,
1073                         unsigned long flags,
1074                         unsigned long pgoff,
1075                         unsigned long *populate,
1076                         struct list_head *uf)
1077 {
1078         struct vm_area_struct *vma;
1079         struct vm_region *region;
1080         struct rb_node *rb;
1081         vm_flags_t vm_flags;
1082         unsigned long capabilities, result;
1083         int ret;
1084
1085         *populate = 0;
1086
1087         /* decide whether we should attempt the mapping, and if so what sort of
1088          * mapping */
1089         ret = validate_mmap_request(file, addr, len, prot, flags, pgoff,
1090                                     &capabilities);
1091         if (ret < 0)
1092                 return ret;
1093
1094         /* we ignore the address hint */
1095         addr = 0;
1096         len = PAGE_ALIGN(len);
1097
1098         /* we've determined that we can make the mapping, now translate what we
1099          * now know into VMA flags */
1100         vm_flags = determine_vm_flags(file, prot, flags, capabilities);
1101
1102         /* we're going to need to record the mapping */
1103         region = kmem_cache_zalloc(vm_region_jar, GFP_KERNEL);
1104         if (!region)
1105                 goto error_getting_region;
1106
1107         vma = vm_area_alloc(current->mm);
1108         if (!vma)
1109                 goto error_getting_vma;
1110
1111         region->vm_usage = 1;
1112         region->vm_flags = vm_flags;
1113         region->vm_pgoff = pgoff;
1114
1115         vma->vm_flags = vm_flags;
1116         vma->vm_pgoff = pgoff;
1117
1118         if (file) {
1119                 region->vm_file = get_file(file);
1120                 vma->vm_file = get_file(file);
1121         }
1122
1123         down_write(&nommu_region_sem);
1124
1125         /* if we want to share, we need to check for regions created by other
1126          * mmap() calls that overlap with our proposed mapping
1127          * - we can only share with a superset match on most regular files
1128          * - shared mappings on character devices and memory backed files are
1129          *   permitted to overlap inexactly as far as we are concerned for in
1130          *   these cases, sharing is handled in the driver or filesystem rather
1131          *   than here
1132          */
1133         if (vm_flags & VM_MAYSHARE) {
1134                 struct vm_region *pregion;
1135                 unsigned long pglen, rpglen, pgend, rpgend, start;
1136
1137                 pglen = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1138                 pgend = pgoff + pglen;
1139
1140                 for (rb = rb_first(&nommu_region_tree); rb; rb = rb_next(rb)) {
1141                         pregion = rb_entry(rb, struct vm_region, vm_rb);
1142
1143                         if (!(pregion->vm_flags & VM_MAYSHARE))
1144                                 continue;
1145
1146                         /* search for overlapping mappings on the same file */
1147                         if (file_inode(pregion->vm_file) !=
1148                             file_inode(file))
1149                                 continue;
1150
1151                         if (pregion->vm_pgoff >= pgend)
1152                                 continue;
1153
1154                         rpglen = pregion->vm_end - pregion->vm_start;
1155                         rpglen = (rpglen + PAGE_SIZE - 1) >> PAGE_SHIFT;
1156                         rpgend = pregion->vm_pgoff + rpglen;
1157                         if (pgoff >= rpgend)
1158                                 continue;
1159
1160                         /* handle inexactly overlapping matches between
1161                          * mappings */
1162                         if ((pregion->vm_pgoff != pgoff || rpglen != pglen) &&
1163                             !(pgoff >= pregion->vm_pgoff && pgend <= rpgend)) {
1164                                 /* new mapping is not a subset of the region */
1165                                 if (!(capabilities & NOMMU_MAP_DIRECT))
1166                                         goto sharing_violation;
1167                                 continue;
1168                         }
1169
1170                         /* we've found a region we can share */
1171                         pregion->vm_usage++;
1172                         vma->vm_region = pregion;
1173                         start = pregion->vm_start;
1174                         start += (pgoff - pregion->vm_pgoff) << PAGE_SHIFT;
1175                         vma->vm_start = start;
1176                         vma->vm_end = start + len;
1177
1178                         if (pregion->vm_flags & VM_MAPPED_COPY)
1179                                 vma->vm_flags |= VM_MAPPED_COPY;
1180                         else {
1181                                 ret = do_mmap_shared_file(vma);
1182                                 if (ret < 0) {
1183                                         vma->vm_region = NULL;
1184                                         vma->vm_start = 0;
1185                                         vma->vm_end = 0;
1186                                         pregion->vm_usage--;
1187                                         pregion = NULL;
1188                                         goto error_just_free;
1189                                 }
1190                         }
1191                         fput(region->vm_file);
1192                         kmem_cache_free(vm_region_jar, region);
1193                         region = pregion;
1194                         result = start;
1195                         goto share;
1196                 }
1197
1198                 /* obtain the address at which to make a shared mapping
1199                  * - this is the hook for quasi-memory character devices to
1200                  *   tell us the location of a shared mapping
1201                  */
1202                 if (capabilities & NOMMU_MAP_DIRECT) {
1203                         addr = file->f_op->get_unmapped_area(file, addr, len,
1204                                                              pgoff, flags);
1205                         if (IS_ERR_VALUE(addr)) {
1206                                 ret = addr;
1207                                 if (ret != -ENOSYS)
1208                                         goto error_just_free;
1209
1210                                 /* the driver refused to tell us where to site
1211                                  * the mapping so we'll have to attempt to copy
1212                                  * it */
1213                                 ret = -ENODEV;
1214                                 if (!(capabilities & NOMMU_MAP_COPY))
1215                                         goto error_just_free;
1216
1217                                 capabilities &= ~NOMMU_MAP_DIRECT;
1218                         } else {
1219                                 vma->vm_start = region->vm_start = addr;
1220                                 vma->vm_end = region->vm_end = addr + len;
1221                         }
1222                 }
1223         }
1224
1225         vma->vm_region = region;
1226
1227         /* set up the mapping
1228          * - the region is filled in if NOMMU_MAP_DIRECT is still set
1229          */
1230         if (file && vma->vm_flags & VM_SHARED)
1231                 ret = do_mmap_shared_file(vma);
1232         else
1233                 ret = do_mmap_private(vma, region, len, capabilities);
1234         if (ret < 0)
1235                 goto error_just_free;
1236         add_nommu_region(region);
1237
1238         /* clear anonymous mappings that don't ask for uninitialized data */
1239         if (!vma->vm_file &&
1240             (!IS_ENABLED(CONFIG_MMAP_ALLOW_UNINITIALIZED) ||
1241              !(flags & MAP_UNINITIALIZED)))
1242                 memset((void *)region->vm_start, 0,
1243                        region->vm_end - region->vm_start);
1244
1245         /* okay... we have a mapping; now we have to register it */
1246         result = vma->vm_start;
1247
1248         current->mm->total_vm += len >> PAGE_SHIFT;
1249
1250 share:
1251         add_vma_to_mm(current->mm, vma);
1252
1253         /* we flush the region from the icache only when the first executable
1254          * mapping of it is made  */
1255         if (vma->vm_flags & VM_EXEC && !region->vm_icache_flushed) {
1256                 flush_icache_user_range(region->vm_start, region->vm_end);
1257                 region->vm_icache_flushed = true;
1258         }
1259
1260         up_write(&nommu_region_sem);
1261
1262         return result;
1263
1264 error_just_free:
1265         up_write(&nommu_region_sem);
1266 error:
1267         if (region->vm_file)
1268                 fput(region->vm_file);
1269         kmem_cache_free(vm_region_jar, region);
1270         if (vma->vm_file)
1271                 fput(vma->vm_file);
1272         vm_area_free(vma);
1273         return ret;
1274
1275 sharing_violation:
1276         up_write(&nommu_region_sem);
1277         pr_warn("Attempt to share mismatched mappings\n");
1278         ret = -EINVAL;
1279         goto error;
1280
1281 error_getting_vma:
1282         kmem_cache_free(vm_region_jar, region);
1283         pr_warn("Allocation of vma for %lu byte allocation from process %d failed\n",
1284                         len, current->pid);
1285         show_free_areas(0, NULL);
1286         return -ENOMEM;
1287
1288 error_getting_region:
1289         pr_warn("Allocation of vm region for %lu byte allocation from process %d failed\n",
1290                         len, current->pid);
1291         show_free_areas(0, NULL);
1292         return -ENOMEM;
1293 }
1294
1295 unsigned long ksys_mmap_pgoff(unsigned long addr, unsigned long len,
1296                               unsigned long prot, unsigned long flags,
1297                               unsigned long fd, unsigned long pgoff)
1298 {
1299         struct file *file = NULL;
1300         unsigned long retval = -EBADF;
1301
1302         audit_mmap_fd(fd, flags);
1303         if (!(flags & MAP_ANONYMOUS)) {
1304                 file = fget(fd);
1305                 if (!file)
1306                         goto out;
1307         }
1308
1309         flags &= ~(MAP_EXECUTABLE | MAP_DENYWRITE);
1310
1311         retval = vm_mmap_pgoff(file, addr, len, prot, flags, pgoff);
1312
1313         if (file)
1314                 fput(file);
1315 out:
1316         return retval;
1317 }
1318
1319 SYSCALL_DEFINE6(mmap_pgoff, unsigned long, addr, unsigned long, len,
1320                 unsigned long, prot, unsigned long, flags,
1321                 unsigned long, fd, unsigned long, pgoff)
1322 {
1323         return ksys_mmap_pgoff(addr, len, prot, flags, fd, pgoff);
1324 }
1325
1326 #ifdef __ARCH_WANT_SYS_OLD_MMAP
1327 struct mmap_arg_struct {
1328         unsigned long addr;
1329         unsigned long len;
1330         unsigned long prot;
1331         unsigned long flags;
1332         unsigned long fd;
1333         unsigned long offset;
1334 };
1335
1336 SYSCALL_DEFINE1(old_mmap, struct mmap_arg_struct __user *, arg)
1337 {
1338         struct mmap_arg_struct a;
1339
1340         if (copy_from_user(&a, arg, sizeof(a)))
1341                 return -EFAULT;
1342         if (offset_in_page(a.offset))
1343                 return -EINVAL;
1344
1345         return ksys_mmap_pgoff(a.addr, a.len, a.prot, a.flags, a.fd,
1346                                a.offset >> PAGE_SHIFT);
1347 }
1348 #endif /* __ARCH_WANT_SYS_OLD_MMAP */
1349
1350 /*
1351  * split a vma into two pieces at address 'addr', a new vma is allocated either
1352  * for the first part or the tail.
1353  */
1354 int split_vma(struct mm_struct *mm, struct vm_area_struct *vma,
1355               unsigned long addr, int new_below)
1356 {
1357         struct vm_area_struct *new;
1358         struct vm_region *region;
1359         unsigned long npages;
1360
1361         /* we're only permitted to split anonymous regions (these should have
1362          * only a single usage on the region) */
1363         if (vma->vm_file)
1364                 return -ENOMEM;
1365
1366         if (mm->map_count >= sysctl_max_map_count)
1367                 return -ENOMEM;
1368
1369         region = kmem_cache_alloc(vm_region_jar, GFP_KERNEL);
1370         if (!region)
1371                 return -ENOMEM;
1372
1373         new = vm_area_dup(vma);
1374         if (!new) {
1375                 kmem_cache_free(vm_region_jar, region);
1376                 return -ENOMEM;
1377         }
1378
1379         /* most fields are the same, copy all, and then fixup */
1380         *region = *vma->vm_region;
1381         new->vm_region = region;
1382
1383         npages = (addr - vma->vm_start) >> PAGE_SHIFT;
1384
1385         if (new_below) {
1386                 region->vm_top = region->vm_end = new->vm_end = addr;
1387         } else {
1388                 region->vm_start = new->vm_start = addr;
1389                 region->vm_pgoff = new->vm_pgoff += npages;
1390         }
1391
1392         if (new->vm_ops && new->vm_ops->open)
1393                 new->vm_ops->open(new);
1394
1395         delete_vma_from_mm(vma);
1396         down_write(&nommu_region_sem);
1397         delete_nommu_region(vma->vm_region);
1398         if (new_below) {
1399                 vma->vm_region->vm_start = vma->vm_start = addr;
1400                 vma->vm_region->vm_pgoff = vma->vm_pgoff += npages;
1401         } else {
1402                 vma->vm_region->vm_end = vma->vm_end = addr;
1403                 vma->vm_region->vm_top = addr;
1404         }
1405         add_nommu_region(vma->vm_region);
1406         add_nommu_region(new->vm_region);
1407         up_write(&nommu_region_sem);
1408         add_vma_to_mm(mm, vma);
1409         add_vma_to_mm(mm, new);
1410         return 0;
1411 }
1412
1413 /*
1414  * shrink a VMA by removing the specified chunk from either the beginning or
1415  * the end
1416  */
1417 static int shrink_vma(struct mm_struct *mm,
1418                       struct vm_area_struct *vma,
1419                       unsigned long from, unsigned long to)
1420 {
1421         struct vm_region *region;
1422
1423         /* adjust the VMA's pointers, which may reposition it in the MM's tree
1424          * and list */
1425         delete_vma_from_mm(vma);
1426         if (from > vma->vm_start)
1427                 vma->vm_end = from;
1428         else
1429                 vma->vm_start = to;
1430         add_vma_to_mm(mm, vma);
1431
1432         /* cut the backing region down to size */
1433         region = vma->vm_region;
1434         BUG_ON(region->vm_usage != 1);
1435
1436         down_write(&nommu_region_sem);
1437         delete_nommu_region(region);
1438         if (from > region->vm_start) {
1439                 to = region->vm_top;
1440                 region->vm_top = region->vm_end = from;
1441         } else {
1442                 region->vm_start = to;
1443         }
1444         add_nommu_region(region);
1445         up_write(&nommu_region_sem);
1446
1447         free_page_series(from, to);
1448         return 0;
1449 }
1450
1451 /*
1452  * release a mapping
1453  * - under NOMMU conditions the chunk to be unmapped must be backed by a single
1454  *   VMA, though it need not cover the whole VMA
1455  */
1456 int do_munmap(struct mm_struct *mm, unsigned long start, size_t len, struct list_head *uf)
1457 {
1458         struct vm_area_struct *vma;
1459         unsigned long end;
1460         int ret;
1461
1462         len = PAGE_ALIGN(len);
1463         if (len == 0)
1464                 return -EINVAL;
1465
1466         end = start + len;
1467
1468         /* find the first potentially overlapping VMA */
1469         vma = find_vma(mm, start);
1470         if (!vma) {
1471                 static int limit;
1472                 if (limit < 5) {
1473                         pr_warn("munmap of memory not mmapped by process %d (%s): 0x%lx-0x%lx\n",
1474                                         current->pid, current->comm,
1475                                         start, start + len - 1);
1476                         limit++;
1477                 }
1478                 return -EINVAL;
1479         }
1480
1481         /* we're allowed to split an anonymous VMA but not a file-backed one */
1482         if (vma->vm_file) {
1483                 do {
1484                         if (start > vma->vm_start)
1485                                 return -EINVAL;
1486                         if (end == vma->vm_end)
1487                                 goto erase_whole_vma;
1488                         vma = vma->vm_next;
1489                 } while (vma);
1490                 return -EINVAL;
1491         } else {
1492                 /* the chunk must be a subset of the VMA found */
1493                 if (start == vma->vm_start && end == vma->vm_end)
1494                         goto erase_whole_vma;
1495                 if (start < vma->vm_start || end > vma->vm_end)
1496                         return -EINVAL;
1497                 if (offset_in_page(start))
1498                         return -EINVAL;
1499                 if (end != vma->vm_end && offset_in_page(end))
1500                         return -EINVAL;
1501                 if (start != vma->vm_start && end != vma->vm_end) {
1502                         ret = split_vma(mm, vma, start, 1);
1503                         if (ret < 0)
1504                                 return ret;
1505                 }
1506                 return shrink_vma(mm, vma, start, end);
1507         }
1508
1509 erase_whole_vma:
1510         delete_vma_from_mm(vma);
1511         delete_vma(mm, vma);
1512         return 0;
1513 }
1514 EXPORT_SYMBOL(do_munmap);
1515
1516 int vm_munmap(unsigned long addr, size_t len)
1517 {
1518         struct mm_struct *mm = current->mm;
1519         int ret;
1520
1521         mmap_write_lock(mm);
1522         ret = do_munmap(mm, addr, len, NULL);
1523         mmap_write_unlock(mm);
1524         return ret;
1525 }
1526 EXPORT_SYMBOL(vm_munmap);
1527
1528 SYSCALL_DEFINE2(munmap, unsigned long, addr, size_t, len)
1529 {
1530         return vm_munmap(addr, len);
1531 }
1532
1533 /*
1534  * release all the mappings made in a process's VM space
1535  */
1536 void exit_mmap(struct mm_struct *mm)
1537 {
1538         struct vm_area_struct *vma;
1539
1540         if (!mm)
1541                 return;
1542
1543         mm->total_vm = 0;
1544
1545         while ((vma = mm->mmap)) {
1546                 mm->mmap = vma->vm_next;
1547                 delete_vma_from_mm(vma);
1548                 delete_vma(mm, vma);
1549                 cond_resched();
1550         }
1551 }
1552
1553 int vm_brk(unsigned long addr, unsigned long len)
1554 {
1555         return -ENOMEM;
1556 }
1557
1558 /*
1559  * expand (or shrink) an existing mapping, potentially moving it at the same
1560  * time (controlled by the MREMAP_MAYMOVE flag and available VM space)
1561  *
1562  * under NOMMU conditions, we only permit changing a mapping's size, and only
1563  * as long as it stays within the region allocated by do_mmap_private() and the
1564  * block is not shareable
1565  *
1566  * MREMAP_FIXED is not supported under NOMMU conditions
1567  */
1568 static unsigned long do_mremap(unsigned long addr,
1569                         unsigned long old_len, unsigned long new_len,
1570                         unsigned long flags, unsigned long new_addr)
1571 {
1572         struct vm_area_struct *vma;
1573
1574         /* insanity checks first */
1575         old_len = PAGE_ALIGN(old_len);
1576         new_len = PAGE_ALIGN(new_len);
1577         if (old_len == 0 || new_len == 0)
1578                 return (unsigned long) -EINVAL;
1579
1580         if (offset_in_page(addr))
1581                 return -EINVAL;
1582
1583         if (flags & MREMAP_FIXED && new_addr != addr)
1584                 return (unsigned long) -EINVAL;
1585
1586         vma = find_vma_exact(current->mm, addr, old_len);
1587         if (!vma)
1588                 return (unsigned long) -EINVAL;
1589
1590         if (vma->vm_end != vma->vm_start + old_len)
1591                 return (unsigned long) -EFAULT;
1592
1593         if (vma->vm_flags & VM_MAYSHARE)
1594                 return (unsigned long) -EPERM;
1595
1596         if (new_len > vma->vm_region->vm_end - vma->vm_region->vm_start)
1597                 return (unsigned long) -ENOMEM;
1598
1599         /* all checks complete - do it */
1600         vma->vm_end = vma->vm_start + new_len;
1601         return vma->vm_start;
1602 }
1603
1604 SYSCALL_DEFINE5(mremap, unsigned long, addr, unsigned long, old_len,
1605                 unsigned long, new_len, unsigned long, flags,
1606                 unsigned long, new_addr)
1607 {
1608         unsigned long ret;
1609
1610         mmap_write_lock(current->mm);
1611         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
1612         mmap_write_unlock(current->mm);
1613         return ret;
1614 }
1615
1616 struct page *follow_page(struct vm_area_struct *vma, unsigned long address,
1617                          unsigned int foll_flags)
1618 {
1619         return NULL;
1620 }
1621
1622 int remap_pfn_range(struct vm_area_struct *vma, unsigned long addr,
1623                 unsigned long pfn, unsigned long size, pgprot_t prot)
1624 {
1625         if (addr != (pfn << PAGE_SHIFT))
1626                 return -EINVAL;
1627
1628         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1629         return 0;
1630 }
1631 EXPORT_SYMBOL(remap_pfn_range);
1632
1633 int vm_iomap_memory(struct vm_area_struct *vma, phys_addr_t start, unsigned long len)
1634 {
1635         unsigned long pfn = start >> PAGE_SHIFT;
1636         unsigned long vm_len = vma->vm_end - vma->vm_start;
1637
1638         pfn += vma->vm_pgoff;
1639         return io_remap_pfn_range(vma, vma->vm_start, pfn, vm_len, vma->vm_page_prot);
1640 }
1641 EXPORT_SYMBOL(vm_iomap_memory);
1642
1643 int remap_vmalloc_range(struct vm_area_struct *vma, void *addr,
1644                         unsigned long pgoff)
1645 {
1646         unsigned int size = vma->vm_end - vma->vm_start;
1647
1648         if (!(vma->vm_flags & VM_USERMAP))
1649                 return -EINVAL;
1650
1651         vma->vm_start = (unsigned long)(addr + (pgoff << PAGE_SHIFT));
1652         vma->vm_end = vma->vm_start + size;
1653
1654         return 0;
1655 }
1656 EXPORT_SYMBOL(remap_vmalloc_range);
1657
1658 unsigned long arch_get_unmapped_area(struct file *file, unsigned long addr,
1659         unsigned long len, unsigned long pgoff, unsigned long flags)
1660 {
1661         return -ENOMEM;
1662 }
1663
1664 vm_fault_t filemap_fault(struct vm_fault *vmf)
1665 {
1666         BUG();
1667         return 0;
1668 }
1669 EXPORT_SYMBOL(filemap_fault);
1670
1671 void filemap_map_pages(struct vm_fault *vmf,
1672                 pgoff_t start_pgoff, pgoff_t end_pgoff)
1673 {
1674         BUG();
1675 }
1676 EXPORT_SYMBOL(filemap_map_pages);
1677
1678 int __access_remote_vm(struct task_struct *tsk, struct mm_struct *mm,
1679                 unsigned long addr, void *buf, int len, unsigned int gup_flags)
1680 {
1681         struct vm_area_struct *vma;
1682         int write = gup_flags & FOLL_WRITE;
1683
1684         if (mmap_read_lock_killable(mm))
1685                 return 0;
1686
1687         /* the access must start within one of the target process's mappings */
1688         vma = find_vma(mm, addr);
1689         if (vma) {
1690                 /* don't overrun this mapping */
1691                 if (addr + len >= vma->vm_end)
1692                         len = vma->vm_end - addr;
1693
1694                 /* only read or write mappings where it is permitted */
1695                 if (write && vma->vm_flags & VM_MAYWRITE)
1696                         copy_to_user_page(vma, NULL, addr,
1697                                          (void *) addr, buf, len);
1698                 else if (!write && vma->vm_flags & VM_MAYREAD)
1699                         copy_from_user_page(vma, NULL, addr,
1700                                             buf, (void *) addr, len);
1701                 else
1702                         len = 0;
1703         } else {
1704                 len = 0;
1705         }
1706
1707         mmap_read_unlock(mm);
1708
1709         return len;
1710 }
1711
1712 /**
1713  * access_remote_vm - access another process' address space
1714  * @mm:         the mm_struct of the target address space
1715  * @addr:       start address to access
1716  * @buf:        source or destination buffer
1717  * @len:        number of bytes to transfer
1718  * @gup_flags:  flags modifying lookup behaviour
1719  *
1720  * The caller must hold a reference on @mm.
1721  */
1722 int access_remote_vm(struct mm_struct *mm, unsigned long addr,
1723                 void *buf, int len, unsigned int gup_flags)
1724 {
1725         return __access_remote_vm(NULL, mm, addr, buf, len, gup_flags);
1726 }
1727
1728 /*
1729  * Access another process' address space.
1730  * - source/target buffer must be kernel space
1731  */
1732 int access_process_vm(struct task_struct *tsk, unsigned long addr, void *buf, int len,
1733                 unsigned int gup_flags)
1734 {
1735         struct mm_struct *mm;
1736
1737         if (addr + len < addr)
1738                 return 0;
1739
1740         mm = get_task_mm(tsk);
1741         if (!mm)
1742                 return 0;
1743
1744         len = __access_remote_vm(tsk, mm, addr, buf, len, gup_flags);
1745
1746         mmput(mm);
1747         return len;
1748 }
1749 EXPORT_SYMBOL_GPL(access_process_vm);
1750
1751 /**
1752  * nommu_shrink_inode_mappings - Shrink the shared mappings on an inode
1753  * @inode: The inode to check
1754  * @size: The current filesize of the inode
1755  * @newsize: The proposed filesize of the inode
1756  *
1757  * Check the shared mappings on an inode on behalf of a shrinking truncate to
1758  * make sure that any outstanding VMAs aren't broken and then shrink the
1759  * vm_regions that extend beyond so that do_mmap() doesn't
1760  * automatically grant mappings that are too large.
1761  */
1762 int nommu_shrink_inode_mappings(struct inode *inode, size_t size,
1763                                 size_t newsize)
1764 {
1765         struct vm_area_struct *vma;
1766         struct vm_region *region;
1767         pgoff_t low, high;
1768         size_t r_size, r_top;
1769
1770         low = newsize >> PAGE_SHIFT;
1771         high = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
1772
1773         down_write(&nommu_region_sem);
1774         i_mmap_lock_read(inode->i_mapping);
1775
1776         /* search for VMAs that fall within the dead zone */
1777         vma_interval_tree_foreach(vma, &inode->i_mapping->i_mmap, low, high) {
1778                 /* found one - only interested if it's shared out of the page
1779                  * cache */
1780                 if (vma->vm_flags & VM_SHARED) {
1781                         i_mmap_unlock_read(inode->i_mapping);
1782                         up_write(&nommu_region_sem);
1783                         return -ETXTBSY; /* not quite true, but near enough */
1784                 }
1785         }
1786
1787         /* reduce any regions that overlap the dead zone - if in existence,
1788          * these will be pointed to by VMAs that don't overlap the dead zone
1789          *
1790          * we don't check for any regions that start beyond the EOF as there
1791          * shouldn't be any
1792          */
1793         vma_interval_tree_foreach(vma, &inode->i_mapping->i_mmap, 0, ULONG_MAX) {
1794                 if (!(vma->vm_flags & VM_SHARED))
1795                         continue;
1796
1797                 region = vma->vm_region;
1798                 r_size = region->vm_top - region->vm_start;
1799                 r_top = (region->vm_pgoff << PAGE_SHIFT) + r_size;
1800
1801                 if (r_top > newsize) {
1802                         region->vm_top -= r_top - newsize;
1803                         if (region->vm_end > region->vm_top)
1804                                 region->vm_end = region->vm_top;
1805                 }
1806         }
1807
1808         i_mmap_unlock_read(inode->i_mapping);
1809         up_write(&nommu_region_sem);
1810         return 0;
1811 }
1812
1813 /*
1814  * Initialise sysctl_user_reserve_kbytes.
1815  *
1816  * This is intended to prevent a user from starting a single memory hogging
1817  * process, such that they cannot recover (kill the hog) in OVERCOMMIT_NEVER
1818  * mode.
1819  *
1820  * The default value is min(3% of free memory, 128MB)
1821  * 128MB is enough to recover with sshd/login, bash, and top/kill.
1822  */
1823 static int __meminit init_user_reserve(void)
1824 {
1825         unsigned long free_kbytes;
1826
1827         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
1828
1829         sysctl_user_reserve_kbytes = min(free_kbytes / 32, 1UL << 17);
1830         return 0;
1831 }
1832 subsys_initcall(init_user_reserve);
1833
1834 /*
1835  * Initialise sysctl_admin_reserve_kbytes.
1836  *
1837  * The purpose of sysctl_admin_reserve_kbytes is to allow the sys admin
1838  * to log in and kill a memory hogging process.
1839  *
1840  * Systems with more than 256MB will reserve 8MB, enough to recover
1841  * with sshd, bash, and top in OVERCOMMIT_GUESS. Smaller systems will
1842  * only reserve 3% of free pages by default.
1843  */
1844 static int __meminit init_admin_reserve(void)
1845 {
1846         unsigned long free_kbytes;
1847
1848         free_kbytes = global_zone_page_state(NR_FREE_PAGES) << (PAGE_SHIFT - 10);
1849
1850         sysctl_admin_reserve_kbytes = min(free_kbytes / 32, 1UL << 13);
1851         return 0;
1852 }
1853 subsys_initcall(init_admin_reserve);