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