mm/bootmem.c

   1 /*
   2  *  linux/mm/bootmem.c
   3  *
   4  *  Copyright (C) 1999 Ingo Molnar
   5  *  Discontiguous memory support, Kanoj Sarcar, SGI, Nov 1999
   6  *
   7  *  simple boot-time physical memory area allocator and
   8  *  free memory collector. It's used to deal with reserved
   9  *  system memory and memory holes as well.
  10  */
  11 #include <linux/init.h>
  12 #include <linux/pfn.h>
  13 #include <linux/bootmem.h>
  14 #include <linux/module.h>
  15
  16 #include <asm/bug.h>
  17 #include <asm/io.h>
  18 #include <asm/processor.h>
  19
  20 #include "internal.h"
  21
  22 /*
  23  * Access to this subsystem has to be serialized externally. (this is
  24  * true for the boot process anyway)
  25  */
  26 unsigned long max_low_pfn;
  27 unsigned long min_low_pfn;
  28 unsigned long max_pfn;
  29
  30 static LIST_HEAD(bdata_list);
  31 #ifdef CONFIG_CRASH_DUMP
  32 /*
  33  * If we have booted due to a crash, max_pfn will be a very low value. We need
  34  * to know the amount of memory that the previous kernel used.
  35  */
  36 unsigned long saved_max_pfn;
  37 #endif
  38
  39 /* return the number of _pages_ that will be allocated for the boot bitmap */
  40 unsigned long __init bootmem_bootmap_pages(unsigned long pages)
  41 {
  42         unsigned long mapsize;
  43
  44         mapsize = (pages+7)/8;
  45         mapsize = (mapsize + ~PAGE_MASK) & PAGE_MASK;
  46         mapsize >>= PAGE_SHIFT;
  47
  48         return mapsize;
  49 }
  50
  51 /*
  52  * link bdata in order
  53  */
  54 static void __init link_bootmem(bootmem_data_t *bdata)
  55 {
  56         bootmem_data_t *ent;
  57
  58         if (list_empty(&bdata_list)) {
  59                 list_add(&bdata->list, &bdata_list);
  60                 return;
  61         }
  62         /* insert in order */
  63         list_for_each_entry(ent, &bdata_list, list) {
  64                 if (bdata->node_boot_start < ent->node_boot_start) {
  65                         list_add_tail(&bdata->list, &ent->list);
  66                         return;
  67                 }
  68         }
  69         list_add_tail(&bdata->list, &bdata_list);
  70 }
  71
  72 /*
  73  * Given an initialised bdata, it returns the size of the boot bitmap
  74  */
  75 static unsigned long __init get_mapsize(bootmem_data_t *bdata)
  76 {
  77         unsigned long mapsize;
  78         unsigned long start = PFN_DOWN(bdata->node_boot_start);
  79         unsigned long end = bdata->node_low_pfn;
  80
  81         mapsize = ((end - start) + 7) / 8;
  82         return ALIGN(mapsize, sizeof(long));
  83 }
  84
  85 /*
  86  * Called once to set up the allocator itself.
  87  */
  88 static unsigned long __init init_bootmem_core(pg_data_t *pgdat,
  89         unsigned long mapstart, unsigned long start, unsigned long end)
  90 {
  91         bootmem_data_t *bdata = pgdat->bdata;
  92         unsigned long mapsize;
  93
  94         bdata->node_bootmem_map = phys_to_virt(PFN_PHYS(mapstart));
  95         bdata->node_boot_start = PFN_PHYS(start);
  96         bdata->node_low_pfn = end;
  97         link_bootmem(bdata);
  98
  99         /*
 100          * Initially all pages are reserved - setup_arch() has to
 101          * register free RAM areas explicitly.
 102          */
 103         mapsize = get_mapsize(bdata);
 104         memset(bdata->node_bootmem_map, 0xff, mapsize);
 105
 106         return mapsize;
 107 }
 108
 109 /*
 110  * Marks a particular physical memory range as unallocatable. Usable RAM
 111  * might be used for boot-time allocations - or it might get added
 112  * to the free page pool later on.
 113  */
 114 static int __init reserve_bootmem_core(bootmem_data_t *bdata,
 115                         unsigned long addr, unsigned long size, int flags)
 116 {
 117         unsigned long sidx, eidx;
 118         unsigned long i;
 119         int ret;
 120
 121         /*
 122          * round up, partially reserved pages are considered
 123          * fully reserved.
 124          */
 125         BUG_ON(!size);
 126         BUG_ON(PFN_DOWN(addr) >= bdata->node_low_pfn);
 127         BUG_ON(PFN_UP(addr + size) > bdata->node_low_pfn);
 128         BUG_ON(addr < bdata->node_boot_start);
 129
 130         sidx = PFN_DOWN(addr - bdata->node_boot_start);
 131         eidx = PFN_UP(addr + size - bdata->node_boot_start);
 132
 133         for (i = sidx; i < eidx; i++)
 134                 if (test_and_set_bit(i, bdata->node_bootmem_map)) {
 135 #ifdef CONFIG_DEBUG_BOOTMEM
 136                         printk("hm, page %08lx reserved twice.\n", i*PAGE_SIZE);
 137 #endif
 138                         if (flags & BOOTMEM_EXCLUSIVE) {
 139                                 ret = -EBUSY;
 140                                 goto err;
 141                         }
 142                 }
 143
 144         return 0;
 145
 146 err:
 147         /* unreserve memory we accidentally reserved */
 148         for (i--; i >= sidx; i--)
 149                 clear_bit(i, bdata->node_bootmem_map);
 150
 151         return ret;
 152 }
 153
 154 static void __init free_bootmem_core(bootmem_data_t *bdata, unsigned long addr,
 155                                      unsigned long size)
 156 {
 157         unsigned long sidx, eidx;
 158         unsigned long i;
 159
 160         BUG_ON(!size);
 161
 162         /* out range */
 163         if (addr + size < bdata->node_boot_start ||
 164                 PFN_DOWN(addr) > bdata->node_low_pfn)
 165                 return;
 166         /*
 167          * round down end of usable mem, partially free pages are
 168          * considered reserved.
 169          */
 170
 171         if (addr >= bdata->node_boot_start && addr < bdata->last_success)
 172                 bdata->last_success = addr;
 173
 174         /*
 175          * Round up to index to the range.
 176          */
 177         if (PFN_UP(addr) > PFN_DOWN(bdata->node_boot_start))
 178                 sidx = PFN_UP(addr) - PFN_DOWN(bdata->node_boot_start);
 179         else
 180                 sidx = 0;
 181
 182         eidx = PFN_DOWN(addr + size - bdata->node_boot_start);
 183         if (eidx > bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start))
 184                 eidx = bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start);
 185
 186         for (i = sidx; i < eidx; i++) {
 187                 if (unlikely(!test_and_clear_bit(i, bdata->node_bootmem_map)))
 188                         BUG();
 189         }
 190 }
 191
 192 /*
 193  * We 'merge' subsequent allocations to save space. We might 'lose'
 194  * some fraction of a page if allocations cannot be satisfied due to
 195  * size constraints on boxes where there is physical RAM space
 196  * fragmentation - in these cases (mostly large memory boxes) this
 197  * is not a problem.
 198  *
 199  * On low memory boxes we get it right in 100% of the cases.
 200  *
 201  * alignment has to be a power of 2 value.
 202  *
 203  * NOTE:  This function is _not_ reentrant.
 204  */
 205 void * __init
 206 __alloc_bootmem_core(struct bootmem_data *bdata, unsigned long size,
 207               unsigned long align, unsigned long goal, unsigned long limit)
 208 {
 209         unsigned long offset, remaining_size, areasize, preferred;
 210         unsigned long i, start = 0, incr, eidx, end_pfn;
 211         void *ret;
 212
 213         if (!size) {
 214                 printk("__alloc_bootmem_core(): zero-sized request\n");
 215                 BUG();
 216         }
 217         BUG_ON(align & (align-1));
 218
 219         if (limit && bdata->node_boot_start >= limit)
 220                 return NULL;
 221
 222         /* on nodes without memory - bootmem_map is NULL */
 223         if (!bdata->node_bootmem_map)
 224                 return NULL;
 225
 226         end_pfn = bdata->node_low_pfn;
 227         limit = PFN_DOWN(limit);
 228         if (limit && end_pfn > limit)
 229                 end_pfn = limit;
 230
 231         eidx = end_pfn - PFN_DOWN(bdata->node_boot_start);
 232         offset = 0;
 233         if (align && (bdata->node_boot_start & (align - 1UL)) != 0)
 234                 offset = align - (bdata->node_boot_start & (align - 1UL));
 235         offset = PFN_DOWN(offset);
 236
 237         /*
 238          * We try to allocate bootmem pages above 'goal'
 239          * first, then we try to allocate lower pages.
 240          */
 241         if (goal && goal >= bdata->node_boot_start && PFN_DOWN(goal) < end_pfn) {
 242                 preferred = goal - bdata->node_boot_start;
 243
 244                 if (bdata->last_success >= preferred)
 245                         if (!limit || (limit && limit > bdata->last_success))
 246                                 preferred = bdata->last_success;
 247         } else
 248                 preferred = 0;
 249
 250         preferred = PFN_DOWN(ALIGN(preferred, align)) + offset;
 251         areasize = (size + PAGE_SIZE-1) / PAGE_SIZE;
 252         incr = align >> PAGE_SHIFT ? : 1;
 253
 254 restart_scan:
 255         for (i = preferred; i < eidx; i += incr) {
 256                 unsigned long j;
 257                 i = find_next_zero_bit(bdata->node_bootmem_map, eidx, i);
 258                 i = ALIGN(i, incr);
 259                 if (i >= eidx)
 260                         break;
 261                 if (test_bit(i, bdata->node_bootmem_map))
 262                         continue;
 263                 for (j = i + 1; j < i + areasize; ++j) {
 264                         if (j >= eidx)
 265                                 goto fail_block;
 266                         if (test_bit(j, bdata->node_bootmem_map))
 267                                 goto fail_block;
 268                 }
 269                 start = i;
 270                 goto found;
 271         fail_block:
 272                 i = ALIGN(j, incr);
 273         }
 274
 275         if (preferred > offset) {
 276                 preferred = offset;
 277                 goto restart_scan;
 278         }
 279         return NULL;
 280
 281 found:
 282         bdata->last_success = PFN_PHYS(start);
 283         BUG_ON(start >= eidx);
 284
 285         /*
 286          * Is the next page of the previous allocation-end the start
 287          * of this allocation's buffer? If yes then we can 'merge'
 288          * the previous partial page with this allocation.
 289          */
 290         if (align < PAGE_SIZE &&
 291             bdata->last_offset && bdata->last_pos+1 == start) {
 292                 offset = ALIGN(bdata->last_offset, align);
 293                 BUG_ON(offset > PAGE_SIZE);
 294                 remaining_size = PAGE_SIZE - offset;
 295                 if (size < remaining_size) {
 296                         areasize = 0;
 297                         /* last_pos unchanged */
 298                         bdata->last_offset = offset + size;
 299                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 300                                            offset +
 301                                            bdata->node_boot_start);
 302                 } else {
 303                         remaining_size = size - remaining_size;
 304                         areasize = (remaining_size + PAGE_SIZE-1) / PAGE_SIZE;
 305                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 306                                            offset +
 307                                            bdata->node_boot_start);
 308                         bdata->last_pos = start + areasize - 1;
 309                         bdata->last_offset = remaining_size;
 310                 }
 311                 bdata->last_offset &= ~PAGE_MASK;
 312         } else {
 313                 bdata->last_pos = start + areasize - 1;
 314                 bdata->last_offset = size & ~PAGE_MASK;
 315                 ret = phys_to_virt(start * PAGE_SIZE + bdata->node_boot_start);
 316         }
 317
 318         /*
 319          * Reserve the area now:
 320          */
 321         for (i = start; i < start + areasize; i++)
 322                 if (unlikely(test_and_set_bit(i, bdata->node_bootmem_map)))
 323                         BUG();
 324         memset(ret, 0, size);
 325         return ret;
 326 }
 327
 328 static unsigned long __init free_all_bootmem_core(pg_data_t *pgdat)
 329 {
 330         struct page *page;
 331         unsigned long pfn;
 332         bootmem_data_t *bdata = pgdat->bdata;
 333         unsigned long i, count, total = 0;
 334         unsigned long idx;
 335         unsigned long *map;
 336         int gofast = 0;
 337
 338         BUG_ON(!bdata->node_bootmem_map);
 339
 340         count = 0;
 341         /* first extant page of the node */
 342         pfn = PFN_DOWN(bdata->node_boot_start);
 343         idx = bdata->node_low_pfn - pfn;
 344         map = bdata->node_bootmem_map;
 345         /* Check physaddr is O(LOG2(BITS_PER_LONG)) page aligned */
 346         if (bdata->node_boot_start == 0 ||
 347             ffs(bdata->node_boot_start) - PAGE_SHIFT > ffs(BITS_PER_LONG))
 348                 gofast = 1;
 349         for (i = 0; i < idx; ) {
 350                 unsigned long v = ~map[i / BITS_PER_LONG];
 351
 352                 if (gofast && v == ~0UL) {
 353                         int order;
 354
 355                         page = pfn_to_page(pfn);
 356                         count += BITS_PER_LONG;
 357                         order = ffs(BITS_PER_LONG) - 1;
 358                         __free_pages_bootmem(page, order);
 359                         i += BITS_PER_LONG;
 360                         page += BITS_PER_LONG;
 361                 } else if (v) {
 362                         unsigned long m;
 363
 364                         page = pfn_to_page(pfn);
 365                         for (m = 1; m && i < idx; m<<=1, page++, i++) {
 366                                 if (v & m) {
 367                                         count++;
 368                                         __free_pages_bootmem(page, 0);
 369                                 }
 370                         }
 371                 } else {
 372                         i += BITS_PER_LONG;
 373                 }
 374                 pfn += BITS_PER_LONG;
 375         }
 376         total += count;
 377
 378         /*
 379          * Now free the allocator bitmap itself, it's not
 380          * needed anymore:
 381          */
 382         page = virt_to_page(bdata->node_bootmem_map);
 383         count = 0;
 384         idx = (get_mapsize(bdata) + PAGE_SIZE-1) >> PAGE_SHIFT;
 385         for (i = 0; i < idx; i++, page++) {
 386                 __free_pages_bootmem(page, 0);
 387                 count++;
 388         }
 389         total += count;
 390         bdata->node_bootmem_map = NULL;
 391
 392         return total;
 393 }
 394
 395 unsigned long __init init_bootmem_node(pg_data_t *pgdat, unsigned long freepfn,
 396                                 unsigned long startpfn, unsigned long endpfn)
 397 {
 398         return init_bootmem_core(pgdat, freepfn, startpfn, endpfn);
 399 }
 400
 401 void __init reserve_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 402                                  unsigned long size, int flags)
 403 {
 404         reserve_bootmem_core(pgdat->bdata, physaddr, size, flags);
 405 }
 406
 407 void __init free_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 408                               unsigned long size)
 409 {
 410         free_bootmem_core(pgdat->bdata, physaddr, size);
 411 }
 412
 413 unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
 414 {
 415         return free_all_bootmem_core(pgdat);
 416 }
 417
 418 unsigned long __init init_bootmem(unsigned long start, unsigned long pages)
 419 {
 420         max_low_pfn = pages;
 421         min_low_pfn = start;
 422         return init_bootmem_core(NODE_DATA(0), start, 0, pages);
 423 }
 424
 425 #ifndef CONFIG_HAVE_ARCH_BOOTMEM_NODE
 426 int __init reserve_bootmem(unsigned long addr, unsigned long size,
 427                             int flags)
 428 {
 429         return reserve_bootmem_core(NODE_DATA(0)->bdata, addr, size, flags);
 430 }
 431 #endif /* !CONFIG_HAVE_ARCH_BOOTMEM_NODE */
 432
 433 void __init free_bootmem(unsigned long addr, unsigned long size)
 434 {
 435         bootmem_data_t *bdata;
 436         list_for_each_entry(bdata, &bdata_list, list)
 437                 free_bootmem_core(bdata, addr, size);
 438 }
 439
 440 unsigned long __init free_all_bootmem(void)
 441 {
 442         return free_all_bootmem_core(NODE_DATA(0));
 443 }
 444
 445 void * __init __alloc_bootmem_nopanic(unsigned long size, unsigned long align,
 446                                       unsigned long goal)
 447 {
 448         bootmem_data_t *bdata;
 449         void *ptr;
 450
 451         list_for_each_entry(bdata, &bdata_list, list) {
 452                 ptr = __alloc_bootmem_core(bdata, size, align, goal, 0);
 453                 if (ptr)
 454                         return ptr;
 455         }
 456         return NULL;
 457 }
 458
 459 void * __init __alloc_bootmem(unsigned long size, unsigned long align,
 460                               unsigned long goal)
 461 {
 462         void *mem = __alloc_bootmem_nopanic(size,align,goal);
 463
 464         if (mem)
 465                 return mem;
 466         /*
 467          * Whoops, we cannot satisfy the allocation request.
 468          */
 469         printk(KERN_ALERT "bootmem alloc of %lu bytes failed!\n", size);
 470         panic("Out of memory");
 471         return NULL;
 472 }
 473
 474
 475 void * __init __alloc_bootmem_node(pg_data_t *pgdat, unsigned long size,
 476                                    unsigned long align, unsigned long goal)
 477 {
 478         void *ptr;
 479
 480         ptr = __alloc_bootmem_core(pgdat->bdata, size, align, goal, 0);
 481         if (ptr)
 482                 return ptr;
 483
 484         return __alloc_bootmem(size, align, goal);
 485 }
 486
 487 #ifndef ARCH_LOW_ADDRESS_LIMIT
 488 #define ARCH_LOW_ADDRESS_LIMIT  0xffffffffUL
 489 #endif
 490
 491 void * __init __alloc_bootmem_low(unsigned long size, unsigned long align,
 492                                   unsigned long goal)
 493 {
 494         bootmem_data_t *bdata;
 495         void *ptr;
 496
 497         list_for_each_entry(bdata, &bdata_list, list) {
 498                 ptr = __alloc_bootmem_core(bdata, size, align, goal,
 499                                                 ARCH_LOW_ADDRESS_LIMIT);
 500                 if (ptr)
 501                         return ptr;
 502         }
 503
 504         /*
 505          * Whoops, we cannot satisfy the allocation request.
 506          */
 507         printk(KERN_ALERT "low bootmem alloc of %lu bytes failed!\n", size);
 508         panic("Out of low memory");
 509         return NULL;
 510 }
 511
 512 void * __init __alloc_bootmem_low_node(pg_data_t *pgdat, unsigned long size,
 513                                        unsigned long align, unsigned long goal)
 514 {
 515         return __alloc_bootmem_core(pgdat->bdata, size, align, goal,
 516                                     ARCH_LOW_ADDRESS_LIMIT);
 517 }