mm/vmalloc.c

   1 /*
   2  *  linux/mm/vmalloc.c
   3  *
   4  *  Copyright (C) 1993  Linus Torvalds
   5  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   6  *  SMP-safe vmalloc/vfree/ioremap, Tigran Aivazian <tigran@veritas.com>, May 2000
   7  *  Major rework to support vmap/vunmap, Christoph Hellwig, SGI, August 2002
   8  */
   9
  10 #include <linux/mm.h>
  11 #include <linux/module.h>
  12 #include <linux/highmem.h>
  13 #include <linux/slab.h>
  14 #include <linux/spinlock.h>
  15 #include <linux/interrupt.h>
  16
  17 #include <linux/vmalloc.h>
  18
  19 #include <asm/uaccess.h>
  20 #include <asm/tlbflush.h>
  21
  22
  23 rwlock_t vmlist_lock = RW_LOCK_UNLOCKED;
  24 struct vm_struct *vmlist;
  25
  26 static void unmap_area_pte(pmd_t *pmd, unsigned long address,
  27                                   unsigned long size)
  28 {
  29         unsigned long end;
  30         pte_t *pte;
  31
  32         if (pmd_none(*pmd))
  33                 return;
  34         if (pmd_bad(*pmd)) {
  35                 pmd_ERROR(*pmd);
  36                 pmd_clear(pmd);
  37                 return;
  38         }
  39
  40         pte = pte_offset_kernel(pmd, address);
  41         address &= ~PMD_MASK;
  42         end = address + size;
  43         if (end > PMD_SIZE)
  44                 end = PMD_SIZE;
  45
  46         do {
  47                 pte_t page;
  48                 page = ptep_get_and_clear(pte);
  49                 address += PAGE_SIZE;
  50                 pte++;
  51                 if (pte_none(page))
  52                         continue;
  53                 if (pte_present(page))
  54                         continue;
  55                 printk(KERN_CRIT "Whee.. Swapped out page in kernel page table\n");
  56         } while (address < end);
  57 }
  58
  59 static void unmap_area_pmd(pgd_t *dir, unsigned long address,
  60                                   unsigned long size)
  61 {
  62         unsigned long end;
  63         pmd_t *pmd;
  64
  65         if (pgd_none(*dir))
  66                 return;
  67         if (pgd_bad(*dir)) {
  68                 pgd_ERROR(*dir);
  69                 pgd_clear(dir);
  70                 return;
  71         }
  72
  73         pmd = pmd_offset(dir, address);
  74         address &= ~PGDIR_MASK;
  75         end = address + size;
  76         if (end > PGDIR_SIZE)
  77                 end = PGDIR_SIZE;
  78
  79         do {
  80                 unmap_area_pte(pmd, address, end - address);
  81                 address = (address + PMD_SIZE) & PMD_MASK;
  82                 pmd++;
  83         } while (address < end);
  84 }
  85
  86 static int map_area_pte(pte_t *pte, unsigned long address,
  87                                unsigned long size, pgprot_t prot,
  88                                struct page ***pages)
  89 {
  90         unsigned long end;
  91
  92         address &= ~PMD_MASK;
  93         end = address + size;
  94         if (end > PMD_SIZE)
  95                 end = PMD_SIZE;
  96
  97         do {
  98                 struct page *page = **pages;
  99
 100                 WARN_ON(!pte_none(*pte));
 101                 if (!page)
 102                         return -ENOMEM;
 103
 104                 set_pte(pte, mk_pte(page, prot));
 105                 address += PAGE_SIZE;
 106                 pte++;
 107                 (*pages)++;
 108         } while (address < end);
 109         return 0;
 110 }
 111
 112 static int map_area_pmd(pmd_t *pmd, unsigned long address,
 113                                unsigned long size, pgprot_t prot,
 114                                struct page ***pages)
 115 {
 116         unsigned long base, end;
 117
 118         base = address & PGDIR_MASK;
 119         address &= ~PGDIR_MASK;
 120         end = address + size;
 121         if (end > PGDIR_SIZE)
 122                 end = PGDIR_SIZE;
 123
 124         do {
 125                 pte_t * pte = pte_alloc_kernel(&init_mm, pmd, base + address);
 126                 if (!pte)
 127                         return -ENOMEM;
 128                 if (map_area_pte(pte, address, end - address, prot, pages))
 129                         return -ENOMEM;
 130                 address = (address + PMD_SIZE) & PMD_MASK;
 131                 pmd++;
 132         } while (address < end);
 133
 134         return 0;
 135 }
 136
 137 void unmap_vm_area(struct vm_struct *area)
 138 {
 139         unsigned long address = (unsigned long) area->addr;
 140         unsigned long end = (address + area->size);
 141         pgd_t *dir;
 142
 143         dir = pgd_offset_k(address);
 144         flush_cache_vunmap(address, end);
 145         do {
 146                 unmap_area_pmd(dir, address, end - address);
 147                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 148                 dir++;
 149         } while (address && (address < end));
 150         flush_tlb_kernel_range((unsigned long) area->addr, end);
 151 }
 152
 153 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
 154 {
 155         unsigned long address = (unsigned long) area->addr;
 156         unsigned long end = address + (area->size-PAGE_SIZE);
 157         pgd_t *dir;
 158         int err = 0;
 159
 160         dir = pgd_offset_k(address);
 161         spin_lock(&init_mm.page_table_lock);
 162         do {
 163                 pmd_t *pmd = pmd_alloc(&init_mm, dir, address);
 164                 if (!pmd) {
 165                         err = -ENOMEM;
 166                         break;
 167                 }
 168                 if (map_area_pmd(pmd, address, end - address, prot, pages)) {
 169                         err = -ENOMEM;
 170                         break;
 171                 }
 172
 173                 address = (address + PGDIR_SIZE) & PGDIR_MASK;
 174                 dir++;
 175         } while (address && (address < end));
 176
 177         spin_unlock(&init_mm.page_table_lock);
 178         flush_cache_vmap((unsigned long) area->addr, end);
 179         return err;
 180 }
 181
 182 #define IOREMAP_MAX_ORDER       (7 + PAGE_SHIFT)        /* 128 pages */
 183
 184 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
 185                                 unsigned long start, unsigned long end)
 186 {
 187         struct vm_struct **p, *tmp, *area;
 188         unsigned long align = 1;
 189         unsigned long addr;
 190
 191         if (flags & VM_IOREMAP) {
 192                 int bit = fls(size);
 193
 194                 if (bit > IOREMAP_MAX_ORDER)
 195                         bit = IOREMAP_MAX_ORDER;
 196                 else if (bit < PAGE_SHIFT)
 197                         bit = PAGE_SHIFT;
 198
 199                 align = 1ul << bit;
 200         }
 201         addr = ALIGN(start, align);
 202
 203         area = kmalloc(sizeof(*area), GFP_KERNEL);
 204         if (unlikely(!area))
 205                 return NULL;
 206
 207         /*
 208          * We always allocate a guard page.
 209          */
 210         size += PAGE_SIZE;
 211         if (unlikely(!size)) {
 212                 kfree (area);
 213                 return NULL;
 214         }
 215
 216         write_lock(&vmlist_lock);
 217         for (p = &vmlist; (tmp = *p) != NULL ;p = &tmp->next) {
 218                 if ((unsigned long)tmp->addr < addr) {
 219                         if((unsigned long)tmp->addr + tmp->size >= addr)
 220                                 addr = ALIGN(tmp->size +
 221                                              (unsigned long)tmp->addr, align);
 222                         continue;
 223                 }
 224                 if ((size + addr) < addr)
 225                         goto out;
 226                 if (size + addr <= (unsigned long)tmp->addr)
 227                         goto found;
 228                 addr = ALIGN(tmp->size + (unsigned long)tmp->addr, align);
 229                 if (addr > end - size)
 230                         goto out;
 231         }
 232
 233 found:
 234         area->next = *p;
 235         *p = area;
 236
 237         area->flags = flags;
 238         area->addr = (void *)addr;
 239         area->size = size;
 240         area->pages = NULL;
 241         area->nr_pages = 0;
 242         area->phys_addr = 0;
 243         write_unlock(&vmlist_lock);
 244
 245         return area;
 246
 247 out:
 248         write_unlock(&vmlist_lock);
 249         kfree(area);
 250         if (printk_ratelimit())
 251                 printk(KERN_WARNING "allocation failed: out of vmalloc space - use vmalloc=<size> to increase size.\n");
 252         return NULL;
 253 }
 254
 255 /**
 256  *      get_vm_area  -  reserve a contingous kernel virtual area
 257  *
 258  *      @size:          size of the area
 259  *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
 260  *
 261  *      Search an area of @size in the kernel virtual mapping area,
 262  *      and reserved it for out purposes.  Returns the area descriptor
 263  *      on success or %NULL on failure.
 264  */
 265 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
 266 {
 267         return __get_vm_area(size, flags, VMALLOC_START, VMALLOC_END);
 268 }
 269
 270 /**
 271  *      remove_vm_area  -  find and remove a contingous kernel virtual area
 272  *
 273  *      @addr:          base address
 274  *
 275  *      Search for the kernel VM area starting at @addr, and remove it.
 276  *      This function returns the found VM area, but using it is NOT safe
 277  *      on SMP machines.
 278  */
 279 struct vm_struct *remove_vm_area(void *addr)
 280 {
 281         struct vm_struct **p, *tmp;
 282
 283         write_lock(&vmlist_lock);
 284         for (p = &vmlist ; (tmp = *p) != NULL ;p = &tmp->next) {
 285                  if (tmp->addr == addr)
 286                          goto found;
 287         }
 288         write_unlock(&vmlist_lock);
 289         return NULL;
 290
 291 found:
 292         unmap_vm_area(tmp);
 293         *p = tmp->next;
 294         write_unlock(&vmlist_lock);
 295         return tmp;
 296 }
 297
 298 void __vunmap(void *addr, int deallocate_pages)
 299 {
 300         struct vm_struct *area;
 301
 302         if (!addr)
 303                 return;
 304
 305         if ((PAGE_SIZE-1) & (unsigned long)addr) {
 306                 printk(KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
 307                 WARN_ON(1);
 308                 return;
 309         }
 310
 311         area = remove_vm_area(addr);
 312         if (unlikely(!area)) {
 313                 printk(KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
 314                                 addr);
 315                 WARN_ON(1);
 316                 return;
 317         }
 318
 319         if (deallocate_pages) {
 320                 int i;
 321
 322                 for (i = 0; i < area->nr_pages; i++) {
 323                         if (unlikely(!area->pages[i]))
 324                                 BUG();
 325                         __free_page(area->pages[i]);
 326                 }
 327
 328                 kfree(area->pages);
 329         }
 330
 331         kfree(area);
 332         return;
 333 }
 334
 335 /**
 336  *      vfree  -  release memory allocated by vmalloc()
 337  *
 338  *      @addr:          memory base address
 339  *
 340  *      Free the virtually contiguous memory area starting at @addr, as
 341  *      obtained from vmalloc(), vmalloc_32() or __vmalloc().
 342  *
 343  *      May not be called in interrupt context.
 344  */
 345 void vfree(void *addr)
 346 {
 347         BUG_ON(in_interrupt());
 348         __vunmap(addr, 1);
 349 }
 350
 351 EXPORT_SYMBOL(vfree);
 352
 353 /**
 354  *      vunmap  -  release virtual mapping obtained by vmap()
 355  *
 356  *      @addr:          memory base address
 357  *
 358  *      Free the virtually contiguous memory area starting at @addr,
 359  *      which was created from the page array passed to vmap().
 360  *
 361  *      May not be called in interrupt context.
 362  */
 363 void vunmap(void *addr)
 364 {
 365         BUG_ON(in_interrupt());
 366         __vunmap(addr, 0);
 367 }
 368
 369 EXPORT_SYMBOL(vunmap);
 370
 371 /**
 372  *      vmap  -  map an array of pages into virtually contiguous space
 373  *
 374  *      @pages:         array of page pointers
 375  *      @count:         number of pages to map
 376  *      @flags:         vm_area->flags
 377  *      @prot:          page protection for the mapping
 378  *
 379  *      Maps @count pages from @pages into contiguous kernel virtual
 380  *      space.
 381  */
 382 void *vmap(struct page **pages, unsigned int count,
 383                 unsigned long flags, pgprot_t prot)
 384 {
 385         struct vm_struct *area;
 386
 387         if (count > num_physpages)
 388                 return NULL;
 389
 390         area = get_vm_area((count << PAGE_SHIFT), flags);
 391         if (!area)
 392                 return NULL;
 393         if (map_vm_area(area, prot, &pages)) {
 394                 vunmap(area->addr);
 395                 return NULL;
 396         }
 397
 398         return area->addr;
 399 }
 400
 401 EXPORT_SYMBOL(vmap);
 402
 403 /**
 404  *      __vmalloc  -  allocate virtually contiguous memory
 405  *
 406  *      @size:          allocation size
 407  *      @gfp_mask:      flags for the page level allocator
 408  *      @prot:          protection mask for the allocated pages
 409  *
 410  *      Allocate enough pages to cover @size from the page level
 411  *      allocator with @gfp_mask flags.  Map them into contiguous
 412  *      kernel virtual space, using a pagetable protection of @prot.
 413  */
 414 void *__vmalloc(unsigned long size, int gfp_mask, pgprot_t prot)
 415 {
 416         struct vm_struct *area;
 417         struct page **pages;
 418         unsigned int nr_pages, array_size, i;
 419
 420         size = PAGE_ALIGN(size);
 421         if (!size || (size >> PAGE_SHIFT) > num_physpages)
 422                 return NULL;
 423
 424         area = get_vm_area(size, VM_ALLOC);
 425         if (!area)
 426                 return NULL;
 427
 428         nr_pages = size >> PAGE_SHIFT;
 429         array_size = (nr_pages * sizeof(struct page *));
 430
 431         area->nr_pages = nr_pages;
 432         area->pages = pages = kmalloc(array_size, (gfp_mask & ~__GFP_HIGHMEM));
 433         if (!area->pages) {
 434                 remove_vm_area(area->addr);
 435                 kfree(area);
 436                 return NULL;
 437         }
 438         memset(area->pages, 0, array_size);
 439
 440         for (i = 0; i < area->nr_pages; i++) {
 441                 area->pages[i] = alloc_page(gfp_mask);
 442                 if (unlikely(!area->pages[i])) {
 443                         /* Successfully allocated i pages, free them in __vunmap() */
 444                         area->nr_pages = i;
 445                         goto fail;
 446                 }
 447         }
 448
 449         if (map_vm_area(area, prot, &pages))
 450                 goto fail;
 451         return area->addr;
 452
 453 fail:
 454         vfree(area->addr);
 455         return NULL;
 456 }
 457
 458 EXPORT_SYMBOL(__vmalloc);
 459
 460 /**
 461  *      vmalloc  -  allocate virtually contiguous memory
 462  *
 463  *      @size:          allocation size
 464  *
 465  *      Allocate enough pages to cover @size from the page level
 466  *      allocator and map them into contiguous kernel virtual space.
 467  *
 468  *      For tight cotrol over page level allocator and protection flags
 469  *      use __vmalloc() instead.
 470  */
 471 void *vmalloc(unsigned long size)
 472 {
 473        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
 474 }
 475
 476 EXPORT_SYMBOL(vmalloc);
 477
 478 /**
 479  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
 480  *
 481  *      @size:          allocation size
 482  *
 483  *      Kernel-internal function to allocate enough pages to cover @size
 484  *      the page level allocator and map them into contiguous and
 485  *      executable kernel virtual space.
 486  *
 487  *      For tight cotrol over page level allocator and protection flags
 488  *      use __vmalloc() instead.
 489  */
 490
 491 #ifndef PAGE_KERNEL_EXEC
 492 # define PAGE_KERNEL_EXEC PAGE_KERNEL
 493 #endif
 494
 495 void *vmalloc_exec(unsigned long size)
 496 {
 497         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
 498 }
 499
 500 /**
 501  *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
 502  *
 503  *      @size:          allocation size
 504  *
 505  *      Allocate enough 32bit PA addressable pages to cover @size from the
 506  *      page level allocator and map them into contiguous kernel virtual space.
 507  */
 508 void *vmalloc_32(unsigned long size)
 509 {
 510         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
 511 }
 512
 513 EXPORT_SYMBOL(vmalloc_32);
 514
 515 long vread(char *buf, char *addr, unsigned long count)
 516 {
 517         struct vm_struct *tmp;
 518         char *vaddr, *buf_start = buf;
 519         unsigned long n;
 520
 521         /* Don't allow overflow */
 522         if ((unsigned long) addr + count < count)
 523                 count = -(unsigned long) addr;
 524
 525         read_lock(&vmlist_lock);
 526         for (tmp = vmlist; tmp; tmp = tmp->next) {
 527                 vaddr = (char *) tmp->addr;
 528                 if (addr >= vaddr + tmp->size - PAGE_SIZE)
 529                         continue;
 530                 while (addr < vaddr) {
 531                         if (count == 0)
 532                                 goto finished;
 533                         *buf = '\0';
 534                         buf++;
 535                         addr++;
 536                         count--;
 537                 }
 538                 n = vaddr + tmp->size - PAGE_SIZE - addr;
 539                 do {
 540                         if (count == 0)
 541                                 goto finished;
 542                         *buf = *addr;
 543                         buf++;
 544                         addr++;
 545                         count--;
 546                 } while (--n > 0);
 547         }
 548 finished:
 549         read_unlock(&vmlist_lock);
 550         return buf - buf_start;
 551 }
 552
 553 long vwrite(char *buf, char *addr, unsigned long count)
 554 {
 555         struct vm_struct *tmp;
 556         char *vaddr, *buf_start = buf;
 557         unsigned long n;
 558
 559         /* Don't allow overflow */
 560         if ((unsigned long) addr + count < count)
 561                 count = -(unsigned long) addr;
 562
 563         read_lock(&vmlist_lock);
 564         for (tmp = vmlist; tmp; tmp = tmp->next) {
 565                 vaddr = (char *) tmp->addr;
 566                 if (addr >= vaddr + tmp->size - PAGE_SIZE)
 567                         continue;
 568                 while (addr < vaddr) {
 569                         if (count == 0)
 570                                 goto finished;
 571                         buf++;
 572                         addr++;
 573                         count--;
 574                 }
 575                 n = vaddr + tmp->size - PAGE_SIZE - addr;
 576                 do {
 577                         if (count == 0)
 578                                 goto finished;
 579                         *addr = *buf;
 580                         buf++;
 581                         addr++;
 582                         count--;
 583                 } while (--n > 0);
 584         }
 585 finished:
 586         read_unlock(&vmlist_lock);
 587         return buf - buf_start;
 588 }