arch/x86_64/kernel/e820.c

   1 /*
   2  * Handle the memory map.
   3  * The functions here do the job until bootmem takes over.
   4  * $Id: e820.c,v 1.4 2002/09/19 19:25:32 ak Exp $
   5  *
   6  *  Getting sanitize_e820_map() in sync with i386 version by applying change:
   7  *  -  Provisions for empty E820 memory regions (reported by certain BIOSes).
   8  *     Alex Achenbach <xela@slit.de>, December 2002.
   9  *  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *
  11  */
  12 #include <linux/config.h>
  13 #include <linux/kernel.h>
  14 #include <linux/types.h>
  15 #include <linux/init.h>
  16 #include <linux/bootmem.h>
  17 #include <linux/ioport.h>
  18 #include <linux/string.h>
  19 #include <linux/kexec.h>
  20 #include <asm/page.h>
  21 #include <asm/e820.h>
  22 #include <asm/proto.h>
  23 #include <asm/bootsetup.h>
  24
  25 extern char _end[];
  26
  27 /*
  28  * PFN of last memory page.
  29  */
  30 unsigned long end_pfn;
  31
  32 /*
  33  * end_pfn only includes RAM, while end_pfn_map includes all e820 entries.
  34  * The direct mapping extends to end_pfn_map, so that we can directly access
  35  * apertures, ACPI and other tables without having to play with fixmaps.
  36  */
  37 unsigned long end_pfn_map;
  38
  39 /*
  40  * Last pfn which the user wants to use.
  41  */
  42 unsigned long end_user_pfn = MAXMEM>>PAGE_SHIFT;
  43
  44 extern struct resource code_resource, data_resource;
  45
  46 /* Check for some hardcoded bad areas that early boot is not allowed to touch */
  47 static inline int bad_addr(unsigned long *addrp, unsigned long size)
  48 {
  49         unsigned long addr = *addrp, last = addr + size;
  50
  51         /* various gunk below that needed for SMP startup */
  52         if (addr < 0x8000) {
  53                 *addrp = 0x8000;
  54                 return 1;
  55         }
  56
  57         /* direct mapping tables of the kernel */
  58         if (last >= table_start<<PAGE_SHIFT && addr < table_end<<PAGE_SHIFT) {
  59                 *addrp = table_end << PAGE_SHIFT;
  60                 return 1;
  61         }
  62
  63         /* initrd */
  64 #ifdef CONFIG_BLK_DEV_INITRD
  65         if (LOADER_TYPE && INITRD_START && last >= INITRD_START &&
  66             addr < INITRD_START+INITRD_SIZE) {
  67                 *addrp = INITRD_START + INITRD_SIZE;
  68                 return 1;
  69         }
  70 #endif
  71         /* kernel code + 640k memory hole (later should not be needed, but
  72            be paranoid for now) */
  73         if (last >= 640*1024 && addr < __pa_symbol(&_end)) {
  74                 *addrp = __pa_symbol(&_end);
  75                 return 1;
  76         }
  77         /* XXX ramdisk image here? */
  78         return 0;
  79 }
  80
  81 int __init e820_mapped(unsigned long start, unsigned long end, unsigned type)
  82 {
  83         int i;
  84         for (i = 0; i < e820.nr_map; i++) {
  85                 struct e820entry *ei = &e820.map[i];
  86                 if (type && ei->type != type)
  87                         continue;
  88                 if (ei->addr >= end || ei->addr + ei->size < start)
  89                         continue;
  90                 return 1;
  91         }
  92         return 0;
  93 }
  94
  95 /*
  96  * Find a free area in a specific range.
  97  */
  98 unsigned long __init find_e820_area(unsigned long start, unsigned long end, unsigned size)
  99 {
 100         int i;
 101         for (i = 0; i < e820.nr_map; i++) {
 102                 struct e820entry *ei = &e820.map[i];
 103                 unsigned long addr = ei->addr, last;
 104                 if (ei->type != E820_RAM)
 105                         continue;
 106                 if (addr < start)
 107                         addr = start;
 108                 if (addr > ei->addr + ei->size)
 109                         continue;
 110                 while (bad_addr(&addr, size) && addr+size < ei->addr + ei->size)
 111                         ;
 112                 last = addr + size;
 113                 if (last > ei->addr + ei->size)
 114                         continue;
 115                 if (last > end)
 116                         continue;
 117                 return addr;
 118         }
 119         return -1UL;
 120 }
 121
 122 /*
 123  * Free bootmem based on the e820 table for a node.
 124  */
 125 void __init e820_bootmem_free(pg_data_t *pgdat, unsigned long start,unsigned long end)
 126 {
 127         int i;
 128         for (i = 0; i < e820.nr_map; i++) {
 129                 struct e820entry *ei = &e820.map[i];
 130                 unsigned long last, addr;
 131
 132                 if (ei->type != E820_RAM ||
 133                     ei->addr+ei->size <= start ||
 134                     ei->addr > end)
 135                         continue;
 136
 137                 addr = round_up(ei->addr, PAGE_SIZE);
 138                 if (addr < start)
 139                         addr = start;
 140
 141                 last = round_down(ei->addr + ei->size, PAGE_SIZE);
 142                 if (last >= end)
 143                         last = end;
 144
 145                 if (last > addr && last-addr >= PAGE_SIZE)
 146                         free_bootmem_node(pgdat, addr, last-addr);
 147         }
 148 }
 149
 150 /*
 151  * Find the highest page frame number we have available
 152  */
 153 unsigned long __init e820_end_of_ram(void)
 154 {
 155         int i;
 156         unsigned long end_pfn = 0;
 157
 158         for (i = 0; i < e820.nr_map; i++) {
 159                 struct e820entry *ei = &e820.map[i];
 160                 unsigned long start, end;
 161
 162                 start = round_up(ei->addr, PAGE_SIZE);
 163                 end = round_down(ei->addr + ei->size, PAGE_SIZE);
 164                 if (start >= end)
 165                         continue;
 166                 if (ei->type == E820_RAM) {
 167                 if (end > end_pfn<<PAGE_SHIFT)
 168                         end_pfn = end>>PAGE_SHIFT;
 169                 } else {
 170                         if (end > end_pfn_map<<PAGE_SHIFT)
 171                                 end_pfn_map = end>>PAGE_SHIFT;
 172                 }
 173         }
 174
 175         if (end_pfn > end_pfn_map)
 176                 end_pfn_map = end_pfn;
 177         if (end_pfn_map > MAXMEM>>PAGE_SHIFT)
 178                 end_pfn_map = MAXMEM>>PAGE_SHIFT;
 179         if (end_pfn > end_user_pfn)
 180                 end_pfn = end_user_pfn;
 181         if (end_pfn > end_pfn_map)
 182                 end_pfn = end_pfn_map;
 183
 184         return end_pfn;
 185 }
 186
 187 /*
 188  * Mark e820 reserved areas as busy for the resource manager.
 189  */
 190 void __init e820_reserve_resources(void)
 191 {
 192         int i;
 193         for (i = 0; i < e820.nr_map; i++) {
 194                 struct resource *res;
 195                 res = alloc_bootmem_low(sizeof(struct resource));
 196                 switch (e820.map[i].type) {
 197                 case E820_RAM:  res->name = "System RAM"; break;
 198                 case E820_ACPI: res->name = "ACPI Tables"; break;
 199                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 200                 default:        res->name = "reserved";
 201                 }
 202                 res->start = e820.map[i].addr;
 203                 res->end = res->start + e820.map[i].size - 1;
 204                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 205                 request_resource(&iomem_resource, res);
 206                 if (e820.map[i].type == E820_RAM) {
 207                         /*
 208                          *  We don't know which RAM region contains kernel data,
 209                          *  so we try it repeatedly and let the resource manager
 210                          *  test it.
 211                          */
 212                         request_resource(res, &code_resource);
 213                         request_resource(res, &data_resource);
 214 #ifdef CONFIG_KEXEC
 215                         request_resource(res, &crashk_res);
 216 #endif
 217                 }
 218         }
 219 }
 220
 221 /*
 222  * Add a memory region to the kernel e820 map.
 223  */
 224 void __init add_memory_region(unsigned long start, unsigned long size, int type)
 225 {
 226         int x = e820.nr_map;
 227
 228         if (x == E820MAX) {
 229                 printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 230                 return;
 231         }
 232
 233         e820.map[x].addr = start;
 234         e820.map[x].size = size;
 235         e820.map[x].type = type;
 236         e820.nr_map++;
 237 }
 238
 239 void __init e820_print_map(char *who)
 240 {
 241         int i;
 242
 243         for (i = 0; i < e820.nr_map; i++) {
 244                 printk(" %s: %016Lx - %016Lx ", who,
 245                         (unsigned long long) e820.map[i].addr,
 246                         (unsigned long long) (e820.map[i].addr + e820.map[i].size));
 247                 switch (e820.map[i].type) {
 248                 case E820_RAM:  printk("(usable)\n");
 249                                 break;
 250                 case E820_RESERVED:
 251                                 printk("(reserved)\n");
 252                                 break;
 253                 case E820_ACPI:
 254                                 printk("(ACPI data)\n");
 255                                 break;
 256                 case E820_NVS:
 257                                 printk("(ACPI NVS)\n");
 258                                 break;
 259                 default:        printk("type %u\n", e820.map[i].type);
 260                                 break;
 261                 }
 262         }
 263 }
 264
 265 /*
 266  * Sanitize the BIOS e820 map.
 267  *
 268  * Some e820 responses include overlapping entries.  The following
 269  * replaces the original e820 map with a new one, removing overlaps.
 270  *
 271  */
 272 static int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 273 {
 274         struct change_member {
 275                 struct e820entry *pbios; /* pointer to original bios entry */
 276                 unsigned long long addr; /* address for this change point */
 277         };
 278         static struct change_member change_point_list[2*E820MAX] __initdata;
 279         static struct change_member *change_point[2*E820MAX] __initdata;
 280         static struct e820entry *overlap_list[E820MAX] __initdata;
 281         static struct e820entry new_bios[E820MAX] __initdata;
 282         struct change_member *change_tmp;
 283         unsigned long current_type, last_type;
 284         unsigned long long last_addr;
 285         int chgidx, still_changing;
 286         int overlap_entries;
 287         int new_bios_entry;
 288         int old_nr, new_nr, chg_nr;
 289         int i;
 290
 291         /*
 292                 Visually we're performing the following (1,2,3,4 = memory types)...
 293
 294                 Sample memory map (w/overlaps):
 295                    ____22__________________
 296                    ______________________4_
 297                    ____1111________________
 298                    _44_____________________
 299                    11111111________________
 300                    ____________________33__
 301                    ___________44___________
 302                    __________33333_________
 303                    ______________22________
 304                    ___________________2222_
 305                    _________111111111______
 306                    _____________________11_
 307                    _________________4______
 308
 309                 Sanitized equivalent (no overlap):
 310                    1_______________________
 311                    _44_____________________
 312                    ___1____________________
 313                    ____22__________________
 314                    ______11________________
 315                    _________1______________
 316                    __________3_____________
 317                    ___________44___________
 318                    _____________33_________
 319                    _______________2________
 320                    ________________1_______
 321                    _________________4______
 322                    ___________________2____
 323                    ____________________33__
 324                    ______________________4_
 325         */
 326
 327         /* if there's only one memory region, don't bother */
 328         if (*pnr_map < 2)
 329                 return -1;
 330
 331         old_nr = *pnr_map;
 332
 333         /* bail out if we find any unreasonable addresses in bios map */
 334         for (i=0; i<old_nr; i++)
 335                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr)
 336                         return -1;
 337
 338         /* create pointers for initial change-point information (for sorting) */
 339         for (i=0; i < 2*old_nr; i++)
 340                 change_point[i] = &change_point_list[i];
 341
 342         /* record all known change-points (starting and ending addresses),
 343            omitting those that are for empty memory regions */
 344         chgidx = 0;
 345         for (i=0; i < old_nr; i++)      {
 346                 if (biosmap[i].size != 0) {
 347                         change_point[chgidx]->addr = biosmap[i].addr;
 348                         change_point[chgidx++]->pbios = &biosmap[i];
 349                         change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 350                         change_point[chgidx++]->pbios = &biosmap[i];
 351                 }
 352         }
 353         chg_nr = chgidx;
 354
 355         /* sort change-point list by memory addresses (low -> high) */
 356         still_changing = 1;
 357         while (still_changing)  {
 358                 still_changing = 0;
 359                 for (i=1; i < chg_nr; i++)  {
 360                         /* if <current_addr> > <last_addr>, swap */
 361                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 362                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 363                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 364                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 365                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 366                            )
 367                         {
 368                                 change_tmp = change_point[i];
 369                                 change_point[i] = change_point[i-1];
 370                                 change_point[i-1] = change_tmp;
 371                                 still_changing=1;
 372                         }
 373                 }
 374         }
 375
 376         /* create a new bios memory map, removing overlaps */
 377         overlap_entries=0;       /* number of entries in the overlap table */
 378         new_bios_entry=0;        /* index for creating new bios map entries */
 379         last_type = 0;           /* start with undefined memory type */
 380         last_addr = 0;           /* start with 0 as last starting address */
 381         /* loop through change-points, determining affect on the new bios map */
 382         for (chgidx=0; chgidx < chg_nr; chgidx++)
 383         {
 384                 /* keep track of all overlapping bios entries */
 385                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 386                 {
 387                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 388                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 389                 }
 390                 else
 391                 {
 392                         /* remove entry from list (order independent, so swap with last) */
 393                         for (i=0; i<overlap_entries; i++)
 394                         {
 395                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 396                                         overlap_list[i] = overlap_list[overlap_entries-1];
 397                         }
 398                         overlap_entries--;
 399                 }
 400                 /* if there are overlapping entries, decide which "type" to use */
 401                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 402                 current_type = 0;
 403                 for (i=0; i<overlap_entries; i++)
 404                         if (overlap_list[i]->type > current_type)
 405                                 current_type = overlap_list[i]->type;
 406                 /* continue building up new bios map based on this information */
 407                 if (current_type != last_type)  {
 408                         if (last_type != 0)      {
 409                                 new_bios[new_bios_entry].size =
 410                                         change_point[chgidx]->addr - last_addr;
 411                                 /* move forward only if the new size was non-zero */
 412                                 if (new_bios[new_bios_entry].size != 0)
 413                                         if (++new_bios_entry >= E820MAX)
 414                                                 break;  /* no more space left for new bios entries */
 415                         }
 416                         if (current_type != 0)  {
 417                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 418                                 new_bios[new_bios_entry].type = current_type;
 419                                 last_addr=change_point[chgidx]->addr;
 420                         }
 421                         last_type = current_type;
 422                 }
 423         }
 424         new_nr = new_bios_entry;   /* retain count for new bios entries */
 425
 426         /* copy new bios mapping into original location */
 427         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 428         *pnr_map = new_nr;
 429
 430         return 0;
 431 }
 432
 433 /*
 434  * Copy the BIOS e820 map into a safe place.
 435  *
 436  * Sanity-check it while we're at it..
 437  *
 438  * If we're lucky and live on a modern system, the setup code
 439  * will have given us a memory map that we can use to properly
 440  * set up memory.  If we aren't, we'll fake a memory map.
 441  *
 442  * We check to see that the memory map contains at least 2 elements
 443  * before we'll use it, because the detection code in setup.S may
 444  * not be perfect and most every PC known to man has two memory
 445  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 446  * thinkpad 560x, for example, does not cooperate with the memory
 447  * detection code.)
 448  */
 449 static int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 450 {
 451         /* Only one memory region (or negative)? Ignore it */
 452         if (nr_map < 2)
 453                 return -1;
 454
 455         do {
 456                 unsigned long start = biosmap->addr;
 457                 unsigned long size = biosmap->size;
 458                 unsigned long end = start + size;
 459                 unsigned long type = biosmap->type;
 460
 461                 /* Overflow in 64 bits? Ignore the memory map. */
 462                 if (start > end)
 463                         return -1;
 464
 465                 /*
 466                  * Some BIOSes claim RAM in the 640k - 1M region.
 467                  * Not right. Fix it up.
 468                  *
 469                  * This should be removed on Hammer which is supposed to not
 470                  * have non e820 covered ISA mappings there, but I had some strange
 471                  * problems so it stays for now.  -AK
 472                  */
 473                 if (type == E820_RAM) {
 474                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 475                                 if (start < 0xA0000ULL)
 476                                         add_memory_region(start, 0xA0000ULL-start, type);
 477                                 if (end <= 0x100000ULL)
 478                                         continue;
 479                                 start = 0x100000ULL;
 480                                 size = end - start;
 481                         }
 482                 }
 483
 484                 add_memory_region(start, size, type);
 485         } while (biosmap++,--nr_map);
 486         return 0;
 487 }
 488
 489 void __init setup_memory_region(void)
 490 {
 491         char *who = "BIOS-e820";
 492
 493         /*
 494          * Try to copy the BIOS-supplied E820-map.
 495          *
 496          * Otherwise fake a memory map; one section from 0k->640k,
 497          * the next section from 1mb->appropriate_mem_k
 498          */
 499         sanitize_e820_map(E820_MAP, &E820_MAP_NR);
 500         if (copy_e820_map(E820_MAP, E820_MAP_NR) < 0) {
 501                 unsigned long mem_size;
 502
 503                 /* compare results from other methods and take the greater */
 504                 if (ALT_MEM_K < EXT_MEM_K) {
 505                         mem_size = EXT_MEM_K;
 506                         who = "BIOS-88";
 507                 } else {
 508                         mem_size = ALT_MEM_K;
 509                         who = "BIOS-e801";
 510                 }
 511
 512                 e820.nr_map = 0;
 513                 add_memory_region(0, LOWMEMSIZE(), E820_RAM);
 514                 add_memory_region(HIGH_MEMORY, mem_size << 10, E820_RAM);
 515         }
 516         printk(KERN_INFO "BIOS-provided physical RAM map:\n");
 517         e820_print_map(who);
 518 }
 519
 520 void __init parse_memopt(char *p, char **from)
 521 {
 522         end_user_pfn = memparse(p, from);
 523         end_user_pfn >>= PAGE_SHIFT;
 524 }
 525
 526 unsigned long pci_mem_start = 0xaeedbabe;
 527
 528 /*
 529  * Search for the biggest gap in the low 32 bits of the e820
 530  * memory space.  We pass this space to PCI to assign MMIO resources
 531  * for hotplug or unconfigured devices in.
 532  * Hopefully the BIOS let enough space left.
 533  */
 534 __init void e820_setup_gap(void)
 535 {
 536         unsigned long gapstart, gapsize;
 537         unsigned long last;
 538         int i;
 539         int found = 0;
 540
 541         last = 0x100000000ull;
 542         gapstart = 0x10000000;
 543         gapsize = 0x400000;
 544         i = e820.nr_map;
 545         while (--i >= 0) {
 546                 unsigned long long start = e820.map[i].addr;
 547                 unsigned long long end = start + e820.map[i].size;
 548
 549                 /*
 550                  * Since "last" is at most 4GB, we know we'll
 551                  * fit in 32 bits if this condition is true
 552                  */
 553                 if (last > end) {
 554                         unsigned long gap = last - end;
 555
 556                         if (gap > gapsize) {
 557                                 gapsize = gap;
 558                                 gapstart = end;
 559                                 found = 1;
 560                         }
 561                 }
 562                 if (start < last)
 563                         last = start;
 564         }
 565
 566         if (!found) {
 567                 gapstart = (end_pfn << PAGE_SHIFT) + 1024*1024;
 568                 printk(KERN_ERR "PCI: Warning: Cannot find a gap in the 32bit address range\n"
 569                        KERN_ERR "PCI: Unassigned devices with 32bit resource registers may break!\n");
 570         }
 571
 572         /*
 573          * Start allocating dynamic PCI memory a bit into the gap,
 574          * aligned up to the nearest megabyte.
 575          *
 576          * Question: should we try to pad it up a bit (do something
 577          * like " + (gapsize >> 3)" in there too?). We now have the
 578          * technology.
 579          */
 580         pci_mem_start = (gapstart + 0xfffff) & ~0xfffff;
 581
 582         printk(KERN_INFO "Allocating PCI resources starting at %lx (gap: %lx:%lx)\n",
 583                 pci_mem_start, gapstart, gapsize);
 584 }