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 areasize, preferred;
 210         unsigned long i, start = 0, incr, eidx, end_pfn;
 211         void *ret;
 212         unsigned long node_boot_start;
 213         void *node_bootmem_map;
 214
 215         if (!size) {
 216                 printk("__alloc_bootmem_core(): zero-sized request\n");
 217                 BUG();
 218         }
 219         BUG_ON(align & (align-1));
 220
 221         /* on nodes without memory - bootmem_map is NULL */
 222         if (!bdata->node_bootmem_map)
 223                 return NULL;
 224
 225         /* bdata->node_boot_start is supposed to be (12+6)bits alignment on x86_64 ? */
 226         node_boot_start = bdata->node_boot_start;
 227         node_bootmem_map = bdata->node_bootmem_map;
 228         if (align) {
 229                 node_boot_start = ALIGN(bdata->node_boot_start, align);
 230                 if (node_boot_start > bdata->node_boot_start)
 231                         node_bootmem_map = (unsigned long *)bdata->node_bootmem_map +
 232                             PFN_DOWN(node_boot_start - bdata->node_boot_start)/BITS_PER_LONG;
 233         }
 234
 235         if (limit && node_boot_start >= limit)
 236                 return NULL;
 237
 238         end_pfn = bdata->node_low_pfn;
 239         limit = PFN_DOWN(limit);
 240         if (limit && end_pfn > limit)
 241                 end_pfn = limit;
 242
 243         eidx = end_pfn - PFN_DOWN(node_boot_start);
 244
 245         /*
 246          * We try to allocate bootmem pages above 'goal'
 247          * first, then we try to allocate lower pages.
 248          */
 249         preferred = 0;
 250         if (goal && PFN_DOWN(goal) < end_pfn) {
 251                 if (goal > node_boot_start)
 252                         preferred = goal - node_boot_start;
 253
 254                 if (bdata->last_success > node_boot_start &&
 255                         bdata->last_success - node_boot_start >= preferred)
 256                         if (!limit || (limit && limit > bdata->last_success))
 257                                 preferred = bdata->last_success - node_boot_start;
 258         }
 259
 260         preferred = PFN_DOWN(ALIGN(preferred, align));
 261         areasize = (size + PAGE_SIZE-1) / PAGE_SIZE;
 262         incr = align >> PAGE_SHIFT ? : 1;
 263
 264 restart_scan:
 265         for (i = preferred; i < eidx;) {
 266                 unsigned long j;
 267
 268                 i = find_next_zero_bit(node_bootmem_map, eidx, i);
 269                 i = ALIGN(i, incr);
 270                 if (i >= eidx)
 271                         break;
 272                 if (test_bit(i, node_bootmem_map)) {
 273                         i += incr;
 274                         continue;
 275                 }
 276                 for (j = i + 1; j < i + areasize; ++j) {
 277                         if (j >= eidx)
 278                                 goto fail_block;
 279                         if (test_bit(j, node_bootmem_map))
 280                                 goto fail_block;
 281                 }
 282                 start = i;
 283                 goto found;
 284         fail_block:
 285                 i = ALIGN(j, incr);
 286                 if (i == j)
 287                         i += incr;
 288         }
 289
 290         if (preferred > 0) {
 291                 preferred = 0;
 292                 goto restart_scan;
 293         }
 294         return NULL;
 295
 296 found:
 297         bdata->last_success = PFN_PHYS(start) + node_boot_start;
 298         BUG_ON(start >= eidx);
 299
 300         /*
 301          * Is the next page of the previous allocation-end the start
 302          * of this allocation's buffer? If yes then we can 'merge'
 303          * the previous partial page with this allocation.
 304          */
 305         if (align < PAGE_SIZE &&
 306             bdata->last_offset && bdata->last_pos+1 == start) {
 307                 unsigned long offset, remaining_size;
 308                 offset = ALIGN(bdata->last_offset, align);
 309                 BUG_ON(offset > PAGE_SIZE);
 310                 remaining_size = PAGE_SIZE - offset;
 311                 if (size < remaining_size) {
 312                         areasize = 0;
 313                         /* last_pos unchanged */
 314                         bdata->last_offset = offset + size;
 315                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 316                                            offset + node_boot_start);
 317                 } else {
 318                         remaining_size = size - remaining_size;
 319                         areasize = (remaining_size + PAGE_SIZE-1) / PAGE_SIZE;
 320                         ret = phys_to_virt(bdata->last_pos * PAGE_SIZE +
 321                                            offset + node_boot_start);
 322                         bdata->last_pos = start + areasize - 1;
 323                         bdata->last_offset = remaining_size;
 324                 }
 325                 bdata->last_offset &= ~PAGE_MASK;
 326         } else {
 327                 bdata->last_pos = start + areasize - 1;
 328                 bdata->last_offset = size & ~PAGE_MASK;
 329                 ret = phys_to_virt(start * PAGE_SIZE + node_boot_start);
 330         }
 331
 332         /*
 333          * Reserve the area now:
 334          */
 335         for (i = start; i < start + areasize; i++)
 336                 if (unlikely(test_and_set_bit(i, node_bootmem_map)))
 337                         BUG();
 338         memset(ret, 0, size);
 339         return ret;
 340 }
 341
 342 static unsigned long __init free_all_bootmem_core(pg_data_t *pgdat)
 343 {
 344         struct page *page;
 345         unsigned long pfn;
 346         bootmem_data_t *bdata = pgdat->bdata;
 347         unsigned long i, count, total = 0;
 348         unsigned long idx;
 349         unsigned long *map;
 350         int gofast = 0;
 351
 352         BUG_ON(!bdata->node_bootmem_map);
 353
 354         count = 0;
 355         /* first extant page of the node */
 356         pfn = PFN_DOWN(bdata->node_boot_start);
 357         idx = bdata->node_low_pfn - pfn;
 358         map = bdata->node_bootmem_map;
 359         /* Check physaddr is O(LOG2(BITS_PER_LONG)) page aligned */
 360         if (bdata->node_boot_start == 0 ||
 361             ffs(bdata->node_boot_start) - PAGE_SHIFT > ffs(BITS_PER_LONG))
 362                 gofast = 1;
 363         for (i = 0; i < idx; ) {
 364                 unsigned long v = ~map[i / BITS_PER_LONG];
 365
 366                 if (gofast && v == ~0UL) {
 367                         int order;
 368
 369                         page = pfn_to_page(pfn);
 370                         count += BITS_PER_LONG;
 371                         order = ffs(BITS_PER_LONG) - 1;
 372                         __free_pages_bootmem(page, order);
 373                         i += BITS_PER_LONG;
 374                         page += BITS_PER_LONG;
 375                 } else if (v) {
 376                         unsigned long m;
 377
 378                         page = pfn_to_page(pfn);
 379                         for (m = 1; m && i < idx; m<<=1, page++, i++) {
 380                                 if (v & m) {
 381                                         count++;
 382                                         __free_pages_bootmem(page, 0);
 383                                 }
 384                         }
 385                 } else {
 386                         i += BITS_PER_LONG;
 387                 }
 388                 pfn += BITS_PER_LONG;
 389         }
 390         total += count;
 391
 392         /*
 393          * Now free the allocator bitmap itself, it's not
 394          * needed anymore:
 395          */
 396         page = virt_to_page(bdata->node_bootmem_map);
 397         count = 0;
 398         idx = (get_mapsize(bdata) + PAGE_SIZE-1) >> PAGE_SHIFT;
 399         for (i = 0; i < idx; i++, page++) {
 400                 __free_pages_bootmem(page, 0);
 401                 count++;
 402         }
 403         total += count;
 404         bdata->node_bootmem_map = NULL;
 405
 406         return total;
 407 }
 408
 409 unsigned long __init init_bootmem_node(pg_data_t *pgdat, unsigned long freepfn,
 410                                 unsigned long startpfn, unsigned long endpfn)
 411 {
 412         return init_bootmem_core(pgdat, freepfn, startpfn, endpfn);
 413 }
 414
 415 void __init reserve_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 416                                  unsigned long size, int flags)
 417 {
 418         reserve_bootmem_core(pgdat->bdata, physaddr, size, flags);
 419 }
 420
 421 void __init free_bootmem_node(pg_data_t *pgdat, unsigned long physaddr,
 422                               unsigned long size)
 423 {
 424         free_bootmem_core(pgdat->bdata, physaddr, size);
 425 }
 426
 427 unsigned long __init free_all_bootmem_node(pg_data_t *pgdat)
 428 {
 429         return free_all_bootmem_core(pgdat);
 430 }
 431
 432 unsigned long __init init_bootmem(unsigned long start, unsigned long pages)
 433 {
 434         max_low_pfn = pages;
 435         min_low_pfn = start;
 436         return init_bootmem_core(NODE_DATA(0), start, 0, pages);
 437 }
 438
 439 #ifndef CONFIG_HAVE_ARCH_BOOTMEM_NODE
 440 int __init reserve_bootmem(unsigned long addr, unsigned long size,
 441                             int flags)
 442 {
 443         return reserve_bootmem_core(NODE_DATA(0)->bdata, addr, size, flags);
 444 }
 445 #endif /* !CONFIG_HAVE_ARCH_BOOTMEM_NODE */
 446
 447 void __init free_bootmem(unsigned long addr, unsigned long size)
 448 {
 449         bootmem_data_t *bdata;
 450         list_for_each_entry(bdata, &bdata_list, list)
 451                 free_bootmem_core(bdata, addr, size);
 452 }
 453
 454 unsigned long __init free_all_bootmem(void)
 455 {
 456         return free_all_bootmem_core(NODE_DATA(0));
 457 }
 458
 459 void * __init __alloc_bootmem_nopanic(unsigned long size, unsigned long align,
 460                                       unsigned long goal)
 461 {
 462         bootmem_data_t *bdata;
 463         void *ptr;
 464
 465         list_for_each_entry(bdata, &bdata_list, list) {
 466                 ptr = __alloc_bootmem_core(bdata, size, align, goal, 0);
 467                 if (ptr)
 468                         return ptr;
 469         }
 470         return NULL;
 471 }
 472
 473 void * __init __alloc_bootmem(unsigned long size, unsigned long align,
 474                               unsigned long goal)
 475 {
 476         void *mem = __alloc_bootmem_nopanic(size,align,goal);
 477
 478         if (mem)
 479                 return mem;
 480         /*
 481          * Whoops, we cannot satisfy the allocation request.
 482          */
 483         printk(KERN_ALERT "bootmem alloc of %lu bytes failed!\n", size);
 484         panic("Out of memory");
 485         return NULL;
 486 }
 487
 488
 489 void * __init __alloc_bootmem_node(pg_data_t *pgdat, unsigned long size,
 490                                    unsigned long align, unsigned long goal)
 491 {
 492         void *ptr;
 493
 494         ptr = __alloc_bootmem_core(pgdat->bdata, size, align, goal, 0);
 495         if (ptr)
 496                 return ptr;
 497
 498         return __alloc_bootmem(size, align, goal);
 499 }
 500
 501 #ifndef ARCH_LOW_ADDRESS_LIMIT
 502 #define ARCH_LOW_ADDRESS_LIMIT  0xffffffffUL
 503 #endif
 504
 505 void * __init __alloc_bootmem_low(unsigned long size, unsigned long align,
 506                                   unsigned long goal)
 507 {
 508         bootmem_data_t *bdata;
 509         void *ptr;
 510
 511         list_for_each_entry(bdata, &bdata_list, list) {
 512                 ptr = __alloc_bootmem_core(bdata, size, align, goal,
 513                                                 ARCH_LOW_ADDRESS_LIMIT);
 514                 if (ptr)
 515                         return ptr;
 516         }
 517
 518         /*
 519          * Whoops, we cannot satisfy the allocation request.
 520          */
 521         printk(KERN_ALERT "low bootmem alloc of %lu bytes failed!\n", size);
 522         panic("Out of low memory");
 523         return NULL;
 524 }
 525
 526 void * __init __alloc_bootmem_low_node(pg_data_t *pgdat, unsigned long size,
 527                                        unsigned long align, unsigned long goal)
 528 {
 529         return __alloc_bootmem_core(pgdat->bdata, size, align, goal,
 530                                     ARCH_LOW_ADDRESS_LIMIT);
 531 }