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 bootmem_data_t bootmem_node_data[MAX_NUMNODES] __initdata;
  40
  41 /* return the number of _pages_ that will be allocated for the boot bitmap */
  42 unsigned long __init bootmem_bootmap_pages(unsigned long pages)
  43 {
  44         unsigned long mapsize;
  45
  46         mapsize = (pages+7)/8;
  47         mapsize = (mapsize + ~PAGE_MASK) & PAGE_MASK;
  48         mapsize >>= PAGE_SHIFT;
  49
  50         return mapsize;
  51 }
  52
  53 /*
  54  * link bdata in order
  55  */
  56 static void __init link_bootmem(bootmem_data_t *bdata)
  57 {
  58         bootmem_data_t *ent;
  59
  60         if (list_empty(&bdata_list)) {
  61                 list_add(&bdata->list, &bdata_list);
  62                 return;
  63         }
  64         /* insert in order */
  65         list_for_each_entry(ent, &bdata_list, list) {
  66                 if (bdata->node_boot_start < ent->node_boot_start) {
  67                         list_add_tail(&bdata->list, &ent->list);
  68                         return;
  69                 }
  70         }
  71         list_add_tail(&bdata->list, &bdata_list);
  72 }
  73
  74 /*
  75  * Given an initialised bdata, it returns the size of the boot bitmap
  76  */
  77 static unsigned long __init get_mapsize(bootmem_data_t *bdata)
  78 {
  79         unsigned long mapsize;
  80         unsigned long start = PFN_DOWN(bdata->node_boot_start);
  81         unsigned long end = bdata->node_low_pfn;
  82
  83         mapsize = ((end - start) + 7) / 8;
  84         return ALIGN(mapsize, sizeof(long));
  85 }
  86
  87 /*
  88  * Called once to set up the allocator itself.
  89  */
  90 static unsigned long __init init_bootmem_core(pg_data_t *pgdat,
  91         unsigned long mapstart, unsigned long start, unsigned long end)
  92 {
  93         bootmem_data_t *bdata = pgdat->bdata;
  94         unsigned long mapsize;
  95
  96         mminit_validate_memmodel_limits(&start, &end);
  97         bdata->node_bootmem_map = phys_to_virt(PFN_PHYS(mapstart));
  98         bdata->node_boot_start = PFN_PHYS(start);
  99         bdata->node_low_pfn = end;
 100         link_bootmem(bdata);
 101
 102         /*
 103          * Initially all pages are reserved - setup_arch() has to
 104          * register free RAM areas explicitly.
 105          */
 106         mapsize = get_mapsize(bdata);
 107         memset(bdata->node_bootmem_map, 0xff, mapsize);
 108
 109         return mapsize;
 110 }
 111
 112 /*
 113  * Marks a particular physical memory range as unallocatable. Usable RAM
 114  * might be used for boot-time allocations - or it might get added
 115  * to the free page pool later on.
 116  */
 117 static int __init can_reserve_bootmem_core(bootmem_data_t *bdata,
 118                         unsigned long addr, unsigned long size, int flags)
 119 {
 120         unsigned long sidx, eidx;
 121         unsigned long i;
 122
 123         BUG_ON(!size);
 124
 125         /* out of range, don't hold other */
 126         if (addr + size < bdata->node_boot_start ||
 127                 PFN_DOWN(addr) > bdata->node_low_pfn)
 128                 return 0;
 129
 130         /*
 131          * Round up to index to the range.
 132          */
 133         if (addr > bdata->node_boot_start)
 134                 sidx= PFN_DOWN(addr - bdata->node_boot_start);
 135         else
 136                 sidx = 0;
 137
 138         eidx = PFN_UP(addr + size - bdata->node_boot_start);
 139         if (eidx > bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start))
 140                 eidx = bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start);
 141
 142         for (i = sidx; i < eidx; i++) {
 143                 if (test_bit(i, bdata->node_bootmem_map)) {
 144                         if (flags & BOOTMEM_EXCLUSIVE)
 145                                 return -EBUSY;
 146                 }
 147         }
 148
 149         return 0;
 150
 151 }
 152
 153 static void __init reserve_bootmem_core(bootmem_data_t *bdata,
 154                         unsigned long addr, unsigned long size, int flags)
 155 {
 156         unsigned long sidx, eidx;
 157         unsigned long i;
 158
 159         BUG_ON(!size);
 160
 161         /* out of range */
 162         if (addr + size < bdata->node_boot_start ||
 163                 PFN_DOWN(addr) > bdata->node_low_pfn)
 164                 return;
 165
 166         /*
 167          * Round up to index to the range.
 168          */
 169         if (addr > bdata->node_boot_start)
 170                 sidx= PFN_DOWN(addr - bdata->node_boot_start);
 171         else
 172                 sidx = 0;
 173
 174         eidx = PFN_UP(addr + size - bdata->node_boot_start);
 175         if (eidx > bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start))
 176                 eidx = bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start);
 177
 178         for (i = sidx; i < eidx; i++) {
 179                 if (test_and_set_bit(i, bdata->node_bootmem_map)) {
 180 #ifdef CONFIG_DEBUG_BOOTMEM
 181                         printk("hm, page %08lx reserved twice.\n", i*PAGE_SIZE);
 182 #endif
 183                 }
 184         }
 185 }
 186
 187 static void __init free_bootmem_core(bootmem_data_t *bdata, unsigned long addr,
 188                                      unsigned long size)
 189 {
 190         unsigned long sidx, eidx;
 191         unsigned long i;
 192
 193         BUG_ON(!size);
 194
 195         /* out range */
 196         if (addr + size < bdata->node_boot_start ||
 197                 PFN_DOWN(addr) > bdata->node_low_pfn)
 198                 return;
 199         /*
 200          * round down end of usable mem, partially free pages are
 201          * considered reserved.
 202          */
 203
 204         if (addr >= bdata->node_boot_start && addr < bdata->last_success)
 205                 bdata->last_success = addr;
 206
 207         /*
 208          * Round up to index to the range.
 209          */
 210         if (PFN_UP(addr) > PFN_DOWN(bdata->node_boot_start))
 211                 sidx = PFN_UP(addr) - PFN_DOWN(bdata->node_boot_start);
 212         else
 213                 sidx = 0;
 214
 215         eidx = PFN_DOWN(addr + size - bdata->node_boot_start);
 216         if (eidx > bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start))
 217                 eidx = bdata->node_low_pfn - PFN_DOWN(bdata->node_boot_start);
 218
 219         for (i = sidx; i < eidx; i++) {
 220                 if (unlikely(!test_and_clear_bit(i, bdata->node_bootmem_map)))
 221                         BUG();
 222         }
 223 }
 224
 225 /*
 226  * We 'merge' subsequent allocations to save space. We might 'lose'
 227  * some fraction of a page if allocations cannot be satisfied due to
 228  * size constraints on boxes where there is physical RAM space
 229  * fragmentation - in these cases (mostly large memory boxes) this
 230  * is not a problem.
 231  *
 232  * On low memory boxes we get it right in 100% of the cases.
 233  *
 234  * alignment has to be a power of 2 value.
 235  *
 236  * NOTE:  This function is _not_ reentrant.
 237  */
 238 void * __init
 239 __alloc_bootmem_core(struct bootmem_data *bdata, unsigned long size,
 240               unsigned long align, unsigned long goal, unsigned long limit)
 241 {
 242         unsigned long areasize, preferred;
 243         unsigned long i, start = 0, incr, eidx, end_pfn;
 244         void *ret;
 245         unsigned long node_boot_start;
 246         void *node_bootmem_map;
 247
 248         if (!size) {
 249                 printk("__alloc_bootmem_core(): zero-sized request\n");
 250                 BUG();
 251         }
 252         BUG_ON(align & (align-1));
 253
 254         /* on nodes without memory - bootmem_map is NULL */
 255         if (!bdata->node_bootmem_map)
 256                 return NULL;
 257
 258         /* bdata->node_boot_start is supposed to be (12+6)bits alignment on x86_64 ? */
 259         node_boot_start = bdata->node_boot_start;
 260         node_bootmem_map = bdata->node_bootmem_map;
 261         if (align) {
 262                 node_boot_start = ALIGN(bdata->node_boot_start, align);
 263                 if (node_boot_start > bdata->node_boot_start)
 264                         node_bootmem_map = (unsigned long *)bdata->node_bootmem_map +
 265                             PFN_DOWN(node_boot_start - bdata->node_boot_start)/BITS_PER_LONG;
 266         }
 267
 268         if (limit && node_boot_start >= limit)
 269                 return NULL;
 270
 271         end_pfn = bdata->node_low_pfn;
 272         limit = PFN_DOWN(limit);
 273         if (limit && end_pfn > limit)
 274                 end_pfn = limit;
 275
 276         eidx = end_pfn - PFN_DOWN(node_boot_start);
 277
 278         /*
 279          * We try to allocate bootmem pages above 'goal'
 280          * first, then we try to allocate lower pages.
 281          */
 282         preferred = 0;
 283         if (goal && PFN_DOWN(goal) < end_pfn) {
 284                 if (goal > node_boot_start)
 285                         preferred = goal - node_boot_start;
 286
 287                 if (bdata->last_success > node_boot_start &&
 288                         bdata->last_success - node_boot_start >= preferred)
 289                         if (!limit || (limit && limit > bdata->last_success))
 290                                 preferred = bdata->last_success - node_boot_start;
 291         }
 292
 293         preferred = PFN_DOWN(ALIGN(preferred, align));
 294         areasize = (size + PAGE_SIZE-1) / PAGE_SIZE;
 295         incr = align >> PAGE_SHIFT ? : 1;
 296
 297 restart_scan:
 298         for (i = preferred; i < eidx;) {
 299                 unsigned long j;
 300
 301                 i = find_next_zero_bit(node_bootmem_map, eidx, i);
 302                 i = ALIGN(i, incr);
 303                 if (i >= eidx)
 304                         break;
 305                 if (test_bit(i, node_bootmem_map)) {
 306                         i += incr;
 307                         continue;
 308                 }
 309                 for (j = i + 1; j < i + areasize; ++j) {
 310                         if (j >= eidx)
 311                                 goto fail_block;
 312                         if (test_bit(j, node_bootmem_map))
 313                                 goto fail_block;
 314                 }
 315                 start = i;
 316                 goto found;
 317         fail_block:
 318                 i = ALIGN(j, incr);
 319                 if (i == j)
 320                         i += incr;
 321         }
 322
 323         if (preferred > 0) {
 324                 preferred = 0;
 325                 goto restart_scan;
 326         }
 327         return NULL;
 328
 329 found:
 330         bdata->last_success = PFN_PHYS(start) + node_boot_start;
 331         BUG_ON(start >= eidx);
 332
 333         /*
 334          * Is the next page of the previous allocation-end the start
 335          * of this allocation's buffer? If yes then we can 'merge'
 336          * the previous partial page with this allocation.
 337          */
 338         if (align < PAGE_SIZE &&
 339             bdata->last_offset && bdata->last_pos+1 == start) {
 340                 unsigned long offset, remaining_size;
 341                 offset = ALIGN(bdata->last_offset, align);
 342                 BUG_ON(offset > PAGE_SIZE);
 343                 remaining_size = PAGE_SIZE - offset;
 344                 if (size < remaining_size) {
 345                         areasize = 0;
 346                         /* last_pos unchanged */
 347                         bdata->last_offset = offset + size;
 348                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 349                                            offset + node_boot_start);
 350                 } else {
 351                         remaining_size = size - remaining_size;
 352                         areasize = (remaining_size + PAGE_SIZE-1) / PAGE_SIZE;
 353                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 354                                            offset + node_boot_start);
 355                         bdata->last_pos = start + areasize - 1;
 356                         bdata->last_offset = remaining_size;
 357                 }
 358                 bdata->last_offset &= ~PAGE_MASK;
 359         } else {
 360                 bdata->last_pos = start + areasize - 1;
 361                 bdata->last_offset = size & ~PAGE_MASK;
 362                 ret = phys_to_virt(start * PAGE_SIZE + node_boot_start);
 363         }
 364
 365         /*
 366          * Reserve the area now:
 367          */
 368         for (i = start; i < start + areasize; i++)
 369                 if (unlikely(test_and_set_bit(i, node_bootmem_map)))
 370                         BUG();
 371         memset(ret, 0, size);
 372         return ret;
 373 }
 374
 375 static unsigned long __init free_all_bootmem_core(pg_data_t *pgdat)
 376 {
 377         struct page *page;
 378         unsigned long pfn;
 379         bootmem_data_t *bdata = pgdat->bdata;
 380         unsigned long i, count, total = 0;
 381         unsigned long idx;
 382         unsigned long *map;
 383         int gofast = 0;
 384
 385         BUG_ON(!bdata->node_bootmem_map);
 386
 387         count = 0;
 388         /* first extant page of the node */
 389         pfn = PFN_DOWN(bdata->node_boot_start);
 390         idx = bdata->node_low_pfn - pfn;
 391         map = bdata->node_bootmem_map;
 392         /* Check physaddr is O(LOG2(BITS_PER_LONG)) page aligned */
 393         if (bdata->node_boot_start == 0 ||
 394             ffs(bdata->node_boot_start) - PAGE_SHIFT > ffs(BITS_PER_LONG))
 395                 gofast = 1;
 396         for (i = 0; i < idx; ) {
 397                 unsigned long v = ~map[i / BITS_PER_LONG];
 398
 399                 if (gofast && v == ~0UL) {
 400                         int order;
 401
 402                         page = pfn_to_page(pfn);
 403                         count += BITS_PER_LONG;
 404                         order = ffs(BITS_PER_LONG) - 1;
 405                         __free_pages_bootmem(page, order);
 406                         i += BITS_PER_LONG;
 407                         page += BITS_PER_LONG;
 408                 } else if (v) {
 409                         unsigned long m;
 410
 411                         page = pfn_to_page(pfn);
 412                         for (m = 1; m && i < idx; m<<=1, page++, i++) {
 413                                 if (v & m) {
 414                                         count++;
 415                                         __free_pages_bootmem(page, 0);
 416                                 }
 417                         }
 418                 } else {
 419                         i += BITS_PER_LONG;
 420                 }
 421                 pfn += BITS_PER_LONG;
 422         }
 423         total += count;
 424
 425         /*
 426          * Now free the allocator bitmap itself, it's not
 427          * needed anymore:
 428          */
 429         page = virt_to_page(bdata->node_bootmem_map);
 430         count = 0;
 431         idx = (get_mapsize(bdata) + PAGE_SIZE-1) >> PAGE_SHIFT;
 432         for (i = 0; i < idx; i++, page++) {
 433                 __free_pages_bootmem(page, 0);
 434                 count++;
 435         }
 436         total += count;
 437         bdata->node_bootmem_map = NULL;
 438
 439         return total;
 440 }
 441
 442 unsigned long __init init_bootmem_node(pg_data_t *pgdat, unsigned long freepfn,
 443                                 unsigned long startpfn, unsigned long endpfn)
 444 {
 445         return init_bootmem_core(pgdat, freepfn, startpfn, endpfn);
 446 }
 447
 448 int __init reserve_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 449                                  unsigned long size, int flags)
 450 {
 451         int ret;
 452
 453         ret = can_reserve_bootmem_core(pgdat->bdata, physaddr, size, flags);
 454         if (ret < 0)
 455                 return -ENOMEM;
 456         reserve_bootmem_core(pgdat->bdata, physaddr, size, flags);
 457
 458         return 0;
 459 }
 460
 461 void __init free_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 462                               unsigned long size)
 463 {
 464         free_bootmem_core(pgdat->bdata, physaddr, size);
 465 }
 466
 467 unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
 468 {
 469         register_page_bootmem_info_node(pgdat);
 470         return free_all_bootmem_core(pgdat);
 471 }
 472
 473 unsigned long __init init_bootmem(unsigned long start, unsigned long pages)
 474 {
 475         max_low_pfn = pages;
 476         min_low_pfn = start;
 477         return init_bootmem_core(NODE_DATA(0), start, 0, pages);
 478 }
 479
 480 #ifndef CONFIG_HAVE_ARCH_BOOTMEM_NODE
 481 int __init reserve_bootmem(unsigned long addr, unsigned long size,
 482                             int flags)
 483 {
 484         bootmem_data_t *bdata;
 485         int ret;
 486
 487         list_for_each_entry(bdata, &bdata_list, list) {
 488                 ret = can_reserve_bootmem_core(bdata, addr, size, flags);
 489                 if (ret < 0)
 490                         return ret;
 491         }
 492         list_for_each_entry(bdata, &bdata_list, list)
 493                 reserve_bootmem_core(bdata, addr, size, flags);
 494
 495         return 0;
 496 }
 497 #endif /* !CONFIG_HAVE_ARCH_BOOTMEM_NODE */
 498
 499 void __init free_bootmem(unsigned long addr, unsigned long size)
 500 {
 501         bootmem_data_t *bdata;
 502         list_for_each_entry(bdata, &bdata_list, list)
 503                 free_bootmem_core(bdata, addr, size);
 504 }
 505
 506 unsigned long __init free_all_bootmem(void)
 507 {
 508         return free_all_bootmem_core(NODE_DATA(0));
 509 }
 510
 511 void * __init __alloc_bootmem_nopanic(unsigned long size, unsigned long align,
 512                                       unsigned long goal)
 513 {
 514         bootmem_data_t *bdata;
 515         void *ptr;
 516
 517         list_for_each_entry(bdata, &bdata_list, list) {
 518                 ptr = __alloc_bootmem_core(bdata, size, align, goal, 0);
 519                 if (ptr)
 520                         return ptr;
 521         }
 522         return NULL;
 523 }
 524
 525 void * __init __alloc_bootmem(unsigned long size, unsigned long align,
 526                               unsigned long goal)
 527 {
 528         void *mem = __alloc_bootmem_nopanic(size,align,goal);
 529
 530         if (mem)
 531                 return mem;
 532         /*
 533          * Whoops, we cannot satisfy the allocation request.
 534          */
 535         printk(KERN_ALERT "bootmem alloc of %lu bytes failed!\n", size);
 536         panic("Out of memory");
 537         return NULL;
 538 }
 539
 540
 541 void * __init __alloc_bootmem_node(pg_data_t *pgdat, unsigned long size,
 542                                    unsigned long align, unsigned long goal)
 543 {
 544         void *ptr;
 545
 546         ptr = __alloc_bootmem_core(pgdat->bdata, size, align, goal, 0);
 547         if (ptr)
 548                 return ptr;
 549
 550         return __alloc_bootmem(size, align, goal);
 551 }
 552
 553 #ifdef CONFIG_SPARSEMEM
 554 void * __init alloc_bootmem_section(unsigned long size,
 555                                     unsigned long section_nr)
 556 {
 557         void *ptr;
 558         unsigned long limit, goal, start_nr, end_nr, pfn;
 559         struct pglist_data *pgdat;
 560
 561         pfn = section_nr_to_pfn(section_nr);
 562         goal = PFN_PHYS(pfn);
 563         limit = PFN_PHYS(section_nr_to_pfn(section_nr + 1)) - 1;
 564         pgdat = NODE_DATA(early_pfn_to_nid(pfn));
 565         ptr = __alloc_bootmem_core(pgdat->bdata, size, SMP_CACHE_BYTES, goal,
 566                                    limit);
 567
 568         if (!ptr)
 569                 return NULL;
 570
 571         start_nr = pfn_to_section_nr(PFN_DOWN(__pa(ptr)));
 572         end_nr = pfn_to_section_nr(PFN_DOWN(__pa(ptr) + size));
 573         if (start_nr != section_nr || end_nr != section_nr) {
 574                 printk(KERN_WARNING "alloc_bootmem failed on section %ld.\n",
 575                        section_nr);
 576                 free_bootmem_core(pgdat->bdata, __pa(ptr), size);
 577                 ptr = NULL;
 578         }
 579
 580         return ptr;
 581 }
 582 #endif
 583
 584 #ifndef ARCH_LOW_ADDRESS_LIMIT
 585 #define ARCH_LOW_ADDRESS_LIMIT  0xffffffffUL
 586 #endif
 587
 588 void * __init __alloc_bootmem_low(unsigned long size, unsigned long align,
 589                                   unsigned long goal)
 590 {
 591         bootmem_data_t *bdata;
 592         void *ptr;
 593
 594         list_for_each_entry(bdata, &bdata_list, list) {
 595                 ptr = __alloc_bootmem_core(bdata, size, align, goal,
 596                                                 ARCH_LOW_ADDRESS_LIMIT);
 597                 if (ptr)
 598                         return ptr;
 599         }
 600
 601         /*
 602          * Whoops, we cannot satisfy the allocation request.
 603          */
 604         printk(KERN_ALERT "low bootmem alloc of %lu bytes failed!\n", size);
 605         panic("Out of low memory");
 606         return NULL;
 607 }
 608
 609 void * __init __alloc_bootmem_low_node(pg_data_t *pgdat, unsigned long size,
 610                                        unsigned long align, unsigned long goal)
 611 {
 612         return __alloc_bootmem_core(pgdat->bdata, size, align, goal,
 613                                     ARCH_LOW_ADDRESS_LIMIT);
 614 }