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