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  *  Numa awareness, Christoph Lameter, SGI, June 2005
   9  */
  10
  11 #include <linux/mm.h>
  12 #include <linux/module.h>
  13 #include <linux/highmem.h>
  14 #include <linux/slab.h>
  15 #include <linux/spinlock.h>
  16 #include <linux/interrupt.h>
  17
  18 #include <linux/vmalloc.h>
  19
  20 #include <asm/uaccess.h>
  21 #include <asm/tlbflush.h>
  22
  23
  24 DEFINE_RWLOCK(vmlist_lock);
  25 struct vm_struct *vmlist;
  26
  27 static void vunmap_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end)
  28 {
  29         pte_t *pte;
  30
  31         pte = pte_offset_kernel(pmd, addr);
  32         do {
  33                 pte_t ptent = ptep_get_and_clear(&init_mm, addr, pte);
  34                 WARN_ON(!pte_none(ptent) && !pte_present(ptent));
  35         } while (pte++, addr += PAGE_SIZE, addr != end);
  36 }
  37
  38 static inline void vunmap_pmd_range(pud_t *pud, unsigned long addr,
  39                                                 unsigned long end)
  40 {
  41         pmd_t *pmd;
  42         unsigned long next;
  43
  44         pmd = pmd_offset(pud, addr);
  45         do {
  46                 next = pmd_addr_end(addr, end);
  47                 if (pmd_none_or_clear_bad(pmd))
  48                         continue;
  49                 vunmap_pte_range(pmd, addr, next);
  50         } while (pmd++, addr = next, addr != end);
  51 }
  52
  53 static inline void vunmap_pud_range(pgd_t *pgd, unsigned long addr,
  54                                                 unsigned long end)
  55 {
  56         pud_t *pud;
  57         unsigned long next;
  58
  59         pud = pud_offset(pgd, addr);
  60         do {
  61                 next = pud_addr_end(addr, end);
  62                 if (pud_none_or_clear_bad(pud))
  63                         continue;
  64                 vunmap_pmd_range(pud, addr, next);
  65         } while (pud++, addr = next, addr != end);
  66 }
  67
  68 void unmap_vm_area(struct vm_struct *area)
  69 {
  70         pgd_t *pgd;
  71         unsigned long next;
  72         unsigned long addr = (unsigned long) area->addr;
  73         unsigned long end = addr + area->size;
  74
  75         BUG_ON(addr >= end);
  76         pgd = pgd_offset_k(addr);
  77         flush_cache_vunmap(addr, end);
  78         do {
  79                 next = pgd_addr_end(addr, end);
  80                 if (pgd_none_or_clear_bad(pgd))
  81                         continue;
  82                 vunmap_pud_range(pgd, addr, next);
  83         } while (pgd++, addr = next, addr != end);
  84         flush_tlb_kernel_range((unsigned long) area->addr, end);
  85 }
  86
  87 static int vmap_pte_range(pmd_t *pmd, unsigned long addr,
  88                         unsigned long end, pgprot_t prot, struct page ***pages)
  89 {
  90         pte_t *pte;
  91
  92         pte = pte_alloc_kernel(pmd, addr);
  93         if (!pte)
  94                 return -ENOMEM;
  95         do {
  96                 struct page *page = **pages;
  97                 WARN_ON(!pte_none(*pte));
  98                 if (!page)
  99                         return -ENOMEM;
 100                 set_pte_at(&init_mm, addr, pte, mk_pte(page, prot));
 101                 (*pages)++;
 102         } while (pte++, addr += PAGE_SIZE, addr != end);
 103         return 0;
 104 }
 105
 106 static inline int vmap_pmd_range(pud_t *pud, unsigned long addr,
 107                         unsigned long end, pgprot_t prot, struct page ***pages)
 108 {
 109         pmd_t *pmd;
 110         unsigned long next;
 111
 112         pmd = pmd_alloc(&init_mm, pud, addr);
 113         if (!pmd)
 114                 return -ENOMEM;
 115         do {
 116                 next = pmd_addr_end(addr, end);
 117                 if (vmap_pte_range(pmd, addr, next, prot, pages))
 118                         return -ENOMEM;
 119         } while (pmd++, addr = next, addr != end);
 120         return 0;
 121 }
 122
 123 static inline int vmap_pud_range(pgd_t *pgd, unsigned long addr,
 124                         unsigned long end, pgprot_t prot, struct page ***pages)
 125 {
 126         pud_t *pud;
 127         unsigned long next;
 128
 129         pud = pud_alloc(&init_mm, pgd, addr);
 130         if (!pud)
 131                 return -ENOMEM;
 132         do {
 133                 next = pud_addr_end(addr, end);
 134                 if (vmap_pmd_range(pud, addr, next, prot, pages))
 135                         return -ENOMEM;
 136         } while (pud++, addr = next, addr != end);
 137         return 0;
 138 }
 139
 140 int map_vm_area(struct vm_struct *area, pgprot_t prot, struct page ***pages)
 141 {
 142         pgd_t *pgd;
 143         unsigned long next;
 144         unsigned long addr = (unsigned long) area->addr;
 145         unsigned long end = addr + area->size - PAGE_SIZE;
 146         int err;
 147
 148         BUG_ON(addr >= end);
 149         pgd = pgd_offset_k(addr);
 150         do {
 151                 next = pgd_addr_end(addr, end);
 152                 err = vmap_pud_range(pgd, addr, next, prot, pages);
 153                 if (err)
 154                         break;
 155         } while (pgd++, addr = next, addr != end);
 156         flush_cache_vmap((unsigned long) area->addr, end);
 157         return err;
 158 }
 159
 160 struct vm_struct *__get_vm_area_node(unsigned long size, unsigned long flags,
 161                                 unsigned long start, unsigned long end, int node)
 162 {
 163         struct vm_struct **p, *tmp, *area;
 164         unsigned long align = 1;
 165         unsigned long addr;
 166
 167         if (flags & VM_IOREMAP) {
 168                 int bit = fls(size);
 169
 170                 if (bit > IOREMAP_MAX_ORDER)
 171                         bit = IOREMAP_MAX_ORDER;
 172                 else if (bit < PAGE_SHIFT)
 173                         bit = PAGE_SHIFT;
 174
 175                 align = 1ul << bit;
 176         }
 177         addr = ALIGN(start, align);
 178         size = PAGE_ALIGN(size);
 179
 180         area = kmalloc_node(sizeof(*area), GFP_KERNEL, node);
 181         if (unlikely(!area))
 182                 return NULL;
 183
 184         if (unlikely(!size)) {
 185                 kfree (area);
 186                 return NULL;
 187         }
 188
 189         /*
 190          * We always allocate a guard page.
 191          */
 192         size += PAGE_SIZE;
 193
 194         write_lock(&vmlist_lock);
 195         for (p = &vmlist; (tmp = *p) != NULL ;p = &tmp->next) {
 196                 if ((unsigned long)tmp->addr < addr) {
 197                         if((unsigned long)tmp->addr + tmp->size >= addr)
 198                                 addr = ALIGN(tmp->size +
 199                                              (unsigned long)tmp->addr, align);
 200                         continue;
 201                 }
 202                 if ((size + addr) < addr)
 203                         goto out;
 204                 if (size + addr <= (unsigned long)tmp->addr)
 205                         goto found;
 206                 addr = ALIGN(tmp->size + (unsigned long)tmp->addr, align);
 207                 if (addr > end - size)
 208                         goto out;
 209         }
 210
 211 found:
 212         area->next = *p;
 213         *p = area;
 214
 215         area->flags = flags;
 216         area->addr = (void *)addr;
 217         area->size = size;
 218         area->pages = NULL;
 219         area->nr_pages = 0;
 220         area->phys_addr = 0;
 221         write_unlock(&vmlist_lock);
 222
 223         return area;
 224
 225 out:
 226         write_unlock(&vmlist_lock);
 227         kfree(area);
 228         if (printk_ratelimit())
 229                 printk(KERN_WARNING "allocation failed: out of vmalloc space - use vmalloc=<size> to increase size.\n");
 230         return NULL;
 231 }
 232
 233 struct vm_struct *__get_vm_area(unsigned long size, unsigned long flags,
 234                                 unsigned long start, unsigned long end)
 235 {
 236         return __get_vm_area_node(size, flags, start, end, -1);
 237 }
 238
 239 /**
 240  *      get_vm_area  -  reserve a contingous kernel virtual area
 241  *
 242  *      @size:          size of the area
 243  *      @flags:         %VM_IOREMAP for I/O mappings or VM_ALLOC
 244  *
 245  *      Search an area of @size in the kernel virtual mapping area,
 246  *      and reserved it for out purposes.  Returns the area descriptor
 247  *      on success or %NULL on failure.
 248  */
 249 struct vm_struct *get_vm_area(unsigned long size, unsigned long flags)
 250 {
 251         return __get_vm_area(size, flags, VMALLOC_START, VMALLOC_END);
 252 }
 253
 254 struct vm_struct *get_vm_area_node(unsigned long size, unsigned long flags, int node)
 255 {
 256         return __get_vm_area_node(size, flags, VMALLOC_START, VMALLOC_END, node);
 257 }
 258
 259 /* Caller must hold vmlist_lock */
 260 struct vm_struct *__remove_vm_area(void *addr)
 261 {
 262         struct vm_struct **p, *tmp;
 263
 264         for (p = &vmlist ; (tmp = *p) != NULL ;p = &tmp->next) {
 265                  if (tmp->addr == addr)
 266                          goto found;
 267         }
 268         return NULL;
 269
 270 found:
 271         unmap_vm_area(tmp);
 272         *p = tmp->next;
 273
 274         /*
 275          * Remove the guard page.
 276          */
 277         tmp->size -= PAGE_SIZE;
 278         return tmp;
 279 }
 280
 281 /**
 282  *      remove_vm_area  -  find and remove a contingous kernel virtual area
 283  *
 284  *      @addr:          base address
 285  *
 286  *      Search for the kernel VM area starting at @addr, and remove it.
 287  *      This function returns the found VM area, but using it is NOT safe
 288  *      on SMP machines, except for its size or flags.
 289  */
 290 struct vm_struct *remove_vm_area(void *addr)
 291 {
 292         struct vm_struct *v;
 293         write_lock(&vmlist_lock);
 294         v = __remove_vm_area(addr);
 295         write_unlock(&vmlist_lock);
 296         return v;
 297 }
 298
 299 void __vunmap(void *addr, int deallocate_pages)
 300 {
 301         struct vm_struct *area;
 302
 303         if (!addr)
 304                 return;
 305
 306         if ((PAGE_SIZE-1) & (unsigned long)addr) {
 307                 printk(KERN_ERR "Trying to vfree() bad address (%p)\n", addr);
 308                 WARN_ON(1);
 309                 return;
 310         }
 311
 312         area = remove_vm_area(addr);
 313         if (unlikely(!area)) {
 314                 printk(KERN_ERR "Trying to vfree() nonexistent vm area (%p)\n",
 315                                 addr);
 316                 WARN_ON(1);
 317                 return;
 318         }
 319
 320         if (deallocate_pages) {
 321                 int i;
 322
 323                 for (i = 0; i < area->nr_pages; i++) {
 324                         if (unlikely(!area->pages[i]))
 325                                 BUG();
 326                         __free_page(area->pages[i]);
 327                 }
 328
 329                 if (area->nr_pages > PAGE_SIZE/sizeof(struct page *))
 330                         vfree(area->pages);
 331                 else
 332                         kfree(area->pages);
 333         }
 334
 335         kfree(area);
 336         return;
 337 }
 338
 339 /**
 340  *      vfree  -  release memory allocated by vmalloc()
 341  *
 342  *      @addr:          memory base address
 343  *
 344  *      Free the virtually contiguous memory area starting at @addr, as
 345  *      obtained from vmalloc(), vmalloc_32() or __vmalloc(). If @addr is
 346  *      NULL, no operation is performed.
 347  *
 348  *      Must not be called in interrupt context.
 349  */
 350 void vfree(void *addr)
 351 {
 352         BUG_ON(in_interrupt());
 353         __vunmap(addr, 1);
 354 }
 355 EXPORT_SYMBOL(vfree);
 356
 357 /**
 358  *      vunmap  -  release virtual mapping obtained by vmap()
 359  *
 360  *      @addr:          memory base address
 361  *
 362  *      Free the virtually contiguous memory area starting at @addr,
 363  *      which was created from the page array passed to vmap().
 364  *
 365  *      Must not be called in interrupt context.
 366  */
 367 void vunmap(void *addr)
 368 {
 369         BUG_ON(in_interrupt());
 370         __vunmap(addr, 0);
 371 }
 372 EXPORT_SYMBOL(vunmap);
 373
 374 /**
 375  *      vmap  -  map an array of pages into virtually contiguous space
 376  *
 377  *      @pages:         array of page pointers
 378  *      @count:         number of pages to map
 379  *      @flags:         vm_area->flags
 380  *      @prot:          page protection for the mapping
 381  *
 382  *      Maps @count pages from @pages into contiguous kernel virtual
 383  *      space.
 384  */
 385 void *vmap(struct page **pages, unsigned int count,
 386                 unsigned long flags, pgprot_t prot)
 387 {
 388         struct vm_struct *area;
 389
 390         if (count > num_physpages)
 391                 return NULL;
 392
 393         area = get_vm_area((count << PAGE_SHIFT), flags);
 394         if (!area)
 395                 return NULL;
 396         if (map_vm_area(area, prot, &pages)) {
 397                 vunmap(area->addr);
 398                 return NULL;
 399         }
 400
 401         return area->addr;
 402 }
 403 EXPORT_SYMBOL(vmap);
 404
 405 void *__vmalloc_area_node(struct vm_struct *area, gfp_t gfp_mask,
 406                                 pgprot_t prot, int node)
 407 {
 408         struct page **pages;
 409         unsigned int nr_pages, array_size, i;
 410
 411         nr_pages = (area->size - PAGE_SIZE) >> PAGE_SHIFT;
 412         array_size = (nr_pages * sizeof(struct page *));
 413
 414         area->nr_pages = nr_pages;
 415         /* Please note that the recursion is strictly bounded. */
 416         if (array_size > PAGE_SIZE)
 417                 pages = __vmalloc_node(array_size, gfp_mask, PAGE_KERNEL, node);
 418         else
 419                 pages = kmalloc_node(array_size, (gfp_mask & ~__GFP_HIGHMEM), node);
 420         area->pages = pages;
 421         if (!area->pages) {
 422                 remove_vm_area(area->addr);
 423                 kfree(area);
 424                 return NULL;
 425         }
 426         memset(area->pages, 0, array_size);
 427
 428         for (i = 0; i < area->nr_pages; i++) {
 429                 if (node < 0)
 430                         area->pages[i] = alloc_page(gfp_mask);
 431                 else
 432                         area->pages[i] = alloc_pages_node(node, gfp_mask, 0);
 433                 if (unlikely(!area->pages[i])) {
 434                         /* Successfully allocated i pages, free them in __vunmap() */
 435                         area->nr_pages = i;
 436                         goto fail;
 437                 }
 438         }
 439
 440         if (map_vm_area(area, prot, &pages))
 441                 goto fail;
 442         return area->addr;
 443
 444 fail:
 445         vfree(area->addr);
 446         return NULL;
 447 }
 448
 449 void *__vmalloc_area(struct vm_struct *area, gfp_t gfp_mask, pgprot_t prot)
 450 {
 451         return __vmalloc_area_node(area, gfp_mask, prot, -1);
 452 }
 453
 454 /**
 455  *      __vmalloc_node  -  allocate virtually contiguous memory
 456  *
 457  *      @size:          allocation size
 458  *      @gfp_mask:      flags for the page level allocator
 459  *      @prot:          protection mask for the allocated pages
 460  *      @node:          node to use for allocation or -1
 461  *
 462  *      Allocate enough pages to cover @size from the page level
 463  *      allocator with @gfp_mask flags.  Map them into contiguous
 464  *      kernel virtual space, using a pagetable protection of @prot.
 465  */
 466 void *__vmalloc_node(unsigned long size, gfp_t gfp_mask, pgprot_t prot,
 467                         int node)
 468 {
 469         struct vm_struct *area;
 470
 471         size = PAGE_ALIGN(size);
 472         if (!size || (size >> PAGE_SHIFT) > num_physpages)
 473                 return NULL;
 474
 475         area = get_vm_area_node(size, VM_ALLOC, node);
 476         if (!area)
 477                 return NULL;
 478
 479         return __vmalloc_area_node(area, gfp_mask, prot, node);
 480 }
 481 EXPORT_SYMBOL(__vmalloc_node);
 482
 483 void *__vmalloc(unsigned long size, gfp_t gfp_mask, pgprot_t prot)
 484 {
 485         return __vmalloc_node(size, gfp_mask, prot, -1);
 486 }
 487 EXPORT_SYMBOL(__vmalloc);
 488
 489 /**
 490  *      vmalloc  -  allocate virtually contiguous memory
 491  *
 492  *      @size:          allocation size
 493  *
 494  *      Allocate enough pages to cover @size from the page level
 495  *      allocator and map them into contiguous kernel virtual space.
 496  *
 497  *      For tight cotrol over page level allocator and protection flags
 498  *      use __vmalloc() instead.
 499  */
 500 void *vmalloc(unsigned long size)
 501 {
 502        return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL);
 503 }
 504 EXPORT_SYMBOL(vmalloc);
 505
 506 /**
 507  *      vmalloc_node  -  allocate memory on a specific node
 508  *
 509  *      @size:          allocation size
 510  *      @node:          numa node
 511  *
 512  *      Allocate enough pages to cover @size from the page level
 513  *      allocator and map them into contiguous kernel virtual space.
 514  *
 515  *      For tight cotrol over page level allocator and protection flags
 516  *      use __vmalloc() instead.
 517  */
 518 void *vmalloc_node(unsigned long size, int node)
 519 {
 520        return __vmalloc_node(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL, node);
 521 }
 522 EXPORT_SYMBOL(vmalloc_node);
 523
 524 #ifndef PAGE_KERNEL_EXEC
 525 # define PAGE_KERNEL_EXEC PAGE_KERNEL
 526 #endif
 527
 528 /**
 529  *      vmalloc_exec  -  allocate virtually contiguous, executable memory
 530  *
 531  *      @size:          allocation size
 532  *
 533  *      Kernel-internal function to allocate enough pages to cover @size
 534  *      the page level allocator and map them into contiguous and
 535  *      executable kernel virtual space.
 536  *
 537  *      For tight cotrol over page level allocator and protection flags
 538  *      use __vmalloc() instead.
 539  */
 540
 541 void *vmalloc_exec(unsigned long size)
 542 {
 543         return __vmalloc(size, GFP_KERNEL | __GFP_HIGHMEM, PAGE_KERNEL_EXEC);
 544 }
 545
 546 /**
 547  *      vmalloc_32  -  allocate virtually contiguous memory (32bit addressable)
 548  *
 549  *      @size:          allocation size
 550  *
 551  *      Allocate enough 32bit PA addressable pages to cover @size from the
 552  *      page level allocator and map them into contiguous kernel virtual space.
 553  */
 554 void *vmalloc_32(unsigned long size)
 555 {
 556         return __vmalloc(size, GFP_KERNEL, PAGE_KERNEL);
 557 }
 558 EXPORT_SYMBOL(vmalloc_32);
 559
 560 long vread(char *buf, char *addr, unsigned long count)
 561 {
 562         struct vm_struct *tmp;
 563         char *vaddr, *buf_start = buf;
 564         unsigned long n;
 565
 566         /* Don't allow overflow */
 567         if ((unsigned long) addr + count < count)
 568                 count = -(unsigned long) addr;
 569
 570         read_lock(&vmlist_lock);
 571         for (tmp = vmlist; tmp; tmp = tmp->next) {
 572                 vaddr = (char *) tmp->addr;
 573                 if (addr >= vaddr + tmp->size - PAGE_SIZE)
 574                         continue;
 575                 while (addr < vaddr) {
 576                         if (count == 0)
 577                                 goto finished;
 578                         *buf = '\0';
 579                         buf++;
 580                         addr++;
 581                         count--;
 582                 }
 583                 n = vaddr + tmp->size - PAGE_SIZE - addr;
 584                 do {
 585                         if (count == 0)
 586                                 goto finished;
 587                         *buf = *addr;
 588                         buf++;
 589                         addr++;
 590                         count--;
 591                 } while (--n > 0);
 592         }
 593 finished:
 594         read_unlock(&vmlist_lock);
 595         return buf - buf_start;
 596 }
 597
 598 long vwrite(char *buf, char *addr, unsigned long count)
 599 {
 600         struct vm_struct *tmp;
 601         char *vaddr, *buf_start = buf;
 602         unsigned long n;
 603
 604         /* Don't allow overflow */
 605         if ((unsigned long) addr + count < count)
 606                 count = -(unsigned long) addr;
 607
 608         read_lock(&vmlist_lock);
 609         for (tmp = vmlist; tmp; tmp = tmp->next) {
 610                 vaddr = (char *) tmp->addr;
 611                 if (addr >= vaddr + tmp->size - PAGE_SIZE)
 612                         continue;
 613                 while (addr < vaddr) {
 614                         if (count == 0)
 615                                 goto finished;
 616                         buf++;
 617                         addr++;
 618                         count--;
 619                 }
 620                 n = vaddr + tmp->size - PAGE_SIZE - addr;
 621                 do {
 622                         if (count == 0)
 623                                 goto finished;
 624                         *addr = *buf;
 625                         buf++;
 626                         addr++;
 627                         count--;
 628                 } while (--n > 0);
 629         }
 630 finished:
 631         read_unlock(&vmlist_lock);
 632         return buf - buf_start;
 633 }