arch/i386/kernel/e820.c

   1 #include <linux/kernel.h>
   2 #include <linux/types.h>
   3 #include <linux/init.h>
   4 #include <linux/bootmem.h>
   5 #include <linux/ioport.h>
   6 #include <linux/string.h>
   7 #include <linux/kexec.h>
   8 #include <linux/module.h>
   9 #include <linux/mm.h>
  10 #include <linux/efi.h>
  11 #include <linux/pfn.h>
  12
  13 #include <asm/pgtable.h>
  14 #include <asm/page.h>
  15 #include <asm/e820.h>
  16
  17 #ifdef CONFIG_EFI
  18 int efi_enabled = 0;
  19 EXPORT_SYMBOL(efi_enabled);
  20 #endif
  21
  22 struct e820map e820;
  23 struct change_member {
  24         struct e820entry *pbios; /* pointer to original bios entry */
  25         unsigned long long addr; /* address for this change point */
  26 };
  27 static struct change_member change_point_list[2*E820MAX] __initdata;
  28 static struct change_member *change_point[2*E820MAX] __initdata;
  29 static struct e820entry *overlap_list[E820MAX] __initdata;
  30 static struct e820entry new_bios[E820MAX] __initdata;
  31 struct resource data_resource = {
  32         .name   = "Kernel data",
  33         .start  = 0,
  34         .end    = 0,
  35         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  36 };
  37
  38 struct resource code_resource = {
  39         .name   = "Kernel code",
  40         .start  = 0,
  41         .end    = 0,
  42         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  43 };
  44
  45 static struct resource system_rom_resource = {
  46         .name   = "System ROM",
  47         .start  = 0xf0000,
  48         .end    = 0xfffff,
  49         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  50 };
  51
  52 static struct resource extension_rom_resource = {
  53         .name   = "Extension ROM",
  54         .start  = 0xe0000,
  55         .end    = 0xeffff,
  56         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  57 };
  58
  59 static struct resource adapter_rom_resources[] = { {
  60         .name   = "Adapter ROM",
  61         .start  = 0xc8000,
  62         .end    = 0,
  63         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  64 }, {
  65         .name   = "Adapter ROM",
  66         .start  = 0,
  67         .end    = 0,
  68         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  69 }, {
  70         .name   = "Adapter ROM",
  71         .start  = 0,
  72         .end    = 0,
  73         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  74 }, {
  75         .name   = "Adapter ROM",
  76         .start  = 0,
  77         .end    = 0,
  78         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  79 }, {
  80         .name   = "Adapter ROM",
  81         .start  = 0,
  82         .end    = 0,
  83         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  84 }, {
  85         .name   = "Adapter ROM",
  86         .start  = 0,
  87         .end    = 0,
  88         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  89 } };
  90
  91 static struct resource video_rom_resource = {
  92         .name   = "Video ROM",
  93         .start  = 0xc0000,
  94         .end    = 0xc7fff,
  95         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  96 };
  97
  98 static struct resource video_ram_resource = {
  99         .name   = "Video RAM area",
 100         .start  = 0xa0000,
 101         .end    = 0xbffff,
 102         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
 103 };
 104
 105 static struct resource standard_io_resources[] = { {
 106         .name   = "dma1",
 107         .start  = 0x0000,
 108         .end    = 0x001f,
 109         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 110 }, {
 111         .name   = "pic1",
 112         .start  = 0x0020,
 113         .end    = 0x0021,
 114         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 115 }, {
 116         .name   = "timer0",
 117         .start  = 0x0040,
 118         .end    = 0x0043,
 119         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 120 }, {
 121         .name   = "timer1",
 122         .start  = 0x0050,
 123         .end    = 0x0053,
 124         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 125 }, {
 126         .name   = "keyboard",
 127         .start  = 0x0060,
 128         .end    = 0x006f,
 129         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 130 }, {
 131         .name   = "dma page reg",
 132         .start  = 0x0080,
 133         .end    = 0x008f,
 134         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 135 }, {
 136         .name   = "pic2",
 137         .start  = 0x00a0,
 138         .end    = 0x00a1,
 139         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 140 }, {
 141         .name   = "dma2",
 142         .start  = 0x00c0,
 143         .end    = 0x00df,
 144         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 145 }, {
 146         .name   = "fpu",
 147         .start  = 0x00f0,
 148         .end    = 0x00ff,
 149         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 150 } };
 151
 152 #define romsignature(x) (*(unsigned short *)(x) == 0xaa55)
 153
 154 static int __init romchecksum(unsigned char *rom, unsigned long length)
 155 {
 156         unsigned char *p, sum = 0;
 157
 158         for (p = rom; p < rom + length; p++)
 159                 sum += *p;
 160         return sum == 0;
 161 }
 162
 163 static void __init probe_roms(void)
 164 {
 165         unsigned long start, length, upper;
 166         unsigned char *rom;
 167         int           i;
 168
 169         /* video rom */
 170         upper = adapter_rom_resources[0].start;
 171         for (start = video_rom_resource.start; start < upper; start += 2048) {
 172                 rom = isa_bus_to_virt(start);
 173                 if (!romsignature(rom))
 174                         continue;
 175
 176                 video_rom_resource.start = start;
 177
 178                 /* 0 < length <= 0x7f * 512, historically */
 179                 length = rom[2] * 512;
 180
 181                 /* if checksum okay, trust length byte */
 182                 if (length && romchecksum(rom, length))
 183                         video_rom_resource.end = start + length - 1;
 184
 185                 request_resource(&iomem_resource, &video_rom_resource);
 186                 break;
 187         }
 188
 189         start = (video_rom_resource.end + 1 + 2047) & ~2047UL;
 190         if (start < upper)
 191                 start = upper;
 192
 193         /* system rom */
 194         request_resource(&iomem_resource, &system_rom_resource);
 195         upper = system_rom_resource.start;
 196
 197         /* check for extension rom (ignore length byte!) */
 198         rom = isa_bus_to_virt(extension_rom_resource.start);
 199         if (romsignature(rom)) {
 200                 length = extension_rom_resource.end - extension_rom_resource.start + 1;
 201                 if (romchecksum(rom, length)) {
 202                         request_resource(&iomem_resource, &extension_rom_resource);
 203                         upper = extension_rom_resource.start;
 204                 }
 205         }
 206
 207         /* check for adapter roms on 2k boundaries */
 208         for (i = 0; i < ARRAY_SIZE(adapter_rom_resources) && start < upper; start += 2048) {
 209                 rom = isa_bus_to_virt(start);
 210                 if (!romsignature(rom))
 211                         continue;
 212
 213                 /* 0 < length <= 0x7f * 512, historically */
 214                 length = rom[2] * 512;
 215
 216                 /* but accept any length that fits if checksum okay */
 217                 if (!length || start + length > upper || !romchecksum(rom, length))
 218                         continue;
 219
 220                 adapter_rom_resources[i].start = start;
 221                 adapter_rom_resources[i].end = start + length - 1;
 222                 request_resource(&iomem_resource, &adapter_rom_resources[i]);
 223
 224                 start = adapter_rom_resources[i++].end & ~2047UL;
 225         }
 226 }
 227
 228 /*
 229  * Request address space for all standard RAM and ROM resources
 230  * and also for regions reported as reserved by the e820.
 231  */
 232 static void __init
 233 legacy_init_iomem_resources(struct resource *code_resource, struct resource *data_resource)
 234 {
 235         int i;
 236
 237         probe_roms();
 238         for (i = 0; i < e820.nr_map; i++) {
 239                 struct resource *res;
 240 #ifndef CONFIG_RESOURCES_64BIT
 241                 if (e820.map[i].addr + e820.map[i].size > 0x100000000ULL)
 242                         continue;
 243 #endif
 244                 res = kzalloc(sizeof(struct resource), GFP_ATOMIC);
 245                 switch (e820.map[i].type) {
 246                 case E820_RAM:  res->name = "System RAM"; break;
 247                 case E820_ACPI: res->name = "ACPI Tables"; break;
 248                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 249                 default:        res->name = "reserved";
 250                 }
 251                 res->start = e820.map[i].addr;
 252                 res->end = res->start + e820.map[i].size - 1;
 253                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 254                 if (request_resource(&iomem_resource, res)) {
 255                         kfree(res);
 256                         continue;
 257                 }
 258                 if (e820.map[i].type == E820_RAM) {
 259                         /*
 260                          *  We don't know which RAM region contains kernel data,
 261                          *  so we try it repeatedly and let the resource manager
 262                          *  test it.
 263                          */
 264                         request_resource(res, code_resource);
 265                         request_resource(res, data_resource);
 266 #ifdef CONFIG_KEXEC
 267                         request_resource(res, &crashk_res);
 268 #endif
 269                 }
 270         }
 271 }
 272
 273 /*
 274  * Request address space for all standard resources
 275  *
 276  * This is called just before pcibios_init(), which is also a
 277  * subsys_initcall, but is linked in later (in arch/i386/pci/common.c).
 278  */
 279 static int __init request_standard_resources(void)
 280 {
 281         int i;
 282
 283         printk("Setting up standard PCI resources\n");
 284         if (efi_enabled)
 285                 efi_initialize_iomem_resources(&code_resource, &data_resource);
 286         else
 287                 legacy_init_iomem_resources(&code_resource, &data_resource);
 288
 289         /* EFI systems may still have VGA */
 290         request_resource(&iomem_resource, &video_ram_resource);
 291
 292         /* request I/O space for devices used on all i[345]86 PCs */
 293         for (i = 0; i < ARRAY_SIZE(standard_io_resources); i++)
 294                 request_resource(&ioport_resource, &standard_io_resources[i]);
 295         return 0;
 296 }
 297
 298 subsys_initcall(request_standard_resources);
 299
 300 void __init add_memory_region(unsigned long long start,
 301                               unsigned long long size, int type)
 302 {
 303         int x;
 304
 305         if (!efi_enabled) {
 306                 x = e820.nr_map;
 307
 308                 if (x == E820MAX) {
 309                     printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 310                     return;
 311                 }
 312
 313                 e820.map[x].addr = start;
 314                 e820.map[x].size = size;
 315                 e820.map[x].type = type;
 316                 e820.nr_map++;
 317         }
 318 } /* add_memory_region */
 319
 320 /*
 321  * Sanitize the BIOS e820 map.
 322  *
 323  * Some e820 responses include overlapping entries.  The following
 324  * replaces the original e820 map with a new one, removing overlaps.
 325  *
 326  */
 327 int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 328 {
 329         struct change_member *change_tmp;
 330         unsigned long current_type, last_type;
 331         unsigned long long last_addr;
 332         int chgidx, still_changing;
 333         int overlap_entries;
 334         int new_bios_entry;
 335         int old_nr, new_nr, chg_nr;
 336         int i;
 337
 338         /*
 339                 Visually we're performing the following (1,2,3,4 = memory types)...
 340
 341                 Sample memory map (w/overlaps):
 342                    ____22__________________
 343                    ______________________4_
 344                    ____1111________________
 345                    _44_____________________
 346                    11111111________________
 347                    ____________________33__
 348                    ___________44___________
 349                    __________33333_________
 350                    ______________22________
 351                    ___________________2222_
 352                    _________111111111______
 353                    _____________________11_
 354                    _________________4______
 355
 356                 Sanitized equivalent (no overlap):
 357                    1_______________________
 358                    _44_____________________
 359                    ___1____________________
 360                    ____22__________________
 361                    ______11________________
 362                    _________1______________
 363                    __________3_____________
 364                    ___________44___________
 365                    _____________33_________
 366                    _______________2________
 367                    ________________1_______
 368                    _________________4______
 369                    ___________________2____
 370                    ____________________33__
 371                    ______________________4_
 372         */
 373         printk("sanitize start\n");
 374         /* if there's only one memory region, don't bother */
 375         if (*pnr_map < 2) {
 376                 printk("sanitize bail 0\n");
 377                 return -1;
 378         }
 379
 380         old_nr = *pnr_map;
 381
 382         /* bail out if we find any unreasonable addresses in bios map */
 383         for (i=0; i<old_nr; i++)
 384                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr) {
 385                         printk("sanitize bail 1\n");
 386                         return -1;
 387                 }
 388
 389         /* create pointers for initial change-point information (for sorting) */
 390         for (i=0; i < 2*old_nr; i++)
 391                 change_point[i] = &change_point_list[i];
 392
 393         /* record all known change-points (starting and ending addresses),
 394            omitting those that are for empty memory regions */
 395         chgidx = 0;
 396         for (i=0; i < old_nr; i++)      {
 397                 if (biosmap[i].size != 0) {
 398                         change_point[chgidx]->addr = biosmap[i].addr;
 399                         change_point[chgidx++]->pbios = &biosmap[i];
 400                         change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 401                         change_point[chgidx++]->pbios = &biosmap[i];
 402                 }
 403         }
 404         chg_nr = chgidx;        /* true number of change-points */
 405
 406         /* sort change-point list by memory addresses (low -> high) */
 407         still_changing = 1;
 408         while (still_changing)  {
 409                 still_changing = 0;
 410                 for (i=1; i < chg_nr; i++)  {
 411                         /* if <current_addr> > <last_addr>, swap */
 412                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 413                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 414                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 415                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 416                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 417                            )
 418                         {
 419                                 change_tmp = change_point[i];
 420                                 change_point[i] = change_point[i-1];
 421                                 change_point[i-1] = change_tmp;
 422                                 still_changing=1;
 423                         }
 424                 }
 425         }
 426
 427         /* create a new bios memory map, removing overlaps */
 428         overlap_entries=0;       /* number of entries in the overlap table */
 429         new_bios_entry=0;        /* index for creating new bios map entries */
 430         last_type = 0;           /* start with undefined memory type */
 431         last_addr = 0;           /* start with 0 as last starting address */
 432         /* loop through change-points, determining affect on the new bios map */
 433         for (chgidx=0; chgidx < chg_nr; chgidx++)
 434         {
 435                 /* keep track of all overlapping bios entries */
 436                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 437                 {
 438                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 439                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 440                 }
 441                 else
 442                 {
 443                         /* remove entry from list (order independent, so swap with last) */
 444                         for (i=0; i<overlap_entries; i++)
 445                         {
 446                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 447                                         overlap_list[i] = overlap_list[overlap_entries-1];
 448                         }
 449                         overlap_entries--;
 450                 }
 451                 /* if there are overlapping entries, decide which "type" to use */
 452                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 453                 current_type = 0;
 454                 for (i=0; i<overlap_entries; i++)
 455                         if (overlap_list[i]->type > current_type)
 456                                 current_type = overlap_list[i]->type;
 457                 /* continue building up new bios map based on this information */
 458                 if (current_type != last_type)  {
 459                         if (last_type != 0)      {
 460                                 new_bios[new_bios_entry].size =
 461                                         change_point[chgidx]->addr - last_addr;
 462                                 /* move forward only if the new size was non-zero */
 463                                 if (new_bios[new_bios_entry].size != 0)
 464                                         if (++new_bios_entry >= E820MAX)
 465                                                 break;  /* no more space left for new bios entries */
 466                         }
 467                         if (current_type != 0)  {
 468                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 469                                 new_bios[new_bios_entry].type = current_type;
 470                                 last_addr=change_point[chgidx]->addr;
 471                         }
 472                         last_type = current_type;
 473                 }
 474         }
 475         new_nr = new_bios_entry;   /* retain count for new bios entries */
 476
 477         /* copy new bios mapping into original location */
 478         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 479         *pnr_map = new_nr;
 480
 481         printk("sanitize end\n");
 482         return 0;
 483 }
 484
 485 /*
 486  * Copy the BIOS e820 map into a safe place.
 487  *
 488  * Sanity-check it while we're at it..
 489  *
 490  * If we're lucky and live on a modern system, the setup code
 491  * will have given us a memory map that we can use to properly
 492  * set up memory.  If we aren't, we'll fake a memory map.
 493  *
 494  * We check to see that the memory map contains at least 2 elements
 495  * before we'll use it, because the detection code in setup.S may
 496  * not be perfect and most every PC known to man has two memory
 497  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 498  * thinkpad 560x, for example, does not cooperate with the memory
 499  * detection code.)
 500  */
 501 int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 502 {
 503         /* Only one memory region (or negative)? Ignore it */
 504         if (nr_map < 2)
 505                 return -1;
 506
 507         do {
 508                 unsigned long long start = biosmap->addr;
 509                 unsigned long long size = biosmap->size;
 510                 unsigned long long end = start + size;
 511                 unsigned long type = biosmap->type;
 512                 printk("copy_e820_map() start: %016Lx size: %016Lx end: %016Lx type: %ld\n", start, size, end, type);
 513
 514                 /* Overflow in 64 bits? Ignore the memory map. */
 515                 if (start > end)
 516                         return -1;
 517
 518                 /*
 519                  * Some BIOSes claim RAM in the 640k - 1M region.
 520                  * Not right. Fix it up.
 521                  */
 522                 if (type == E820_RAM) {
 523                         printk("copy_e820_map() type is E820_RAM\n");
 524                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 525                                 printk("copy_e820_map() lies in range...\n");
 526                                 if (start < 0xA0000ULL) {
 527                                         printk("copy_e820_map() start < 0xA0000ULL\n");
 528                                         add_memory_region(start, 0xA0000ULL-start, type);
 529                                 }
 530                                 if (end <= 0x100000ULL) {
 531                                         printk("copy_e820_map() end <= 0x100000ULL\n");
 532                                         continue;
 533                                 }
 534                                 start = 0x100000ULL;
 535                                 size = end - start;
 536                         }
 537                 }
 538                 add_memory_region(start, size, type);
 539         } while (biosmap++,--nr_map);
 540         return 0;
 541 }
 542
 543 /*
 544  * Callback for efi_memory_walk.
 545  */
 546 static int __init
 547 efi_find_max_pfn(unsigned long start, unsigned long end, void *arg)
 548 {
 549         unsigned long *max_pfn = arg, pfn;
 550
 551         if (start < end) {
 552                 pfn = PFN_UP(end -1);
 553                 if (pfn > *max_pfn)
 554                         *max_pfn = pfn;
 555         }
 556         return 0;
 557 }
 558
 559 static int __init
 560 efi_memory_present_wrapper(unsigned long start, unsigned long end, void *arg)
 561 {
 562         memory_present(0, PFN_UP(start), PFN_DOWN(end));
 563         return 0;
 564 }
 565
 566 /*
 567  * Find the highest page frame number we have available
 568  */
 569 void __init find_max_pfn(void)
 570 {
 571         int i;
 572
 573         max_pfn = 0;
 574         if (efi_enabled) {
 575                 efi_memmap_walk(efi_find_max_pfn, &max_pfn);
 576                 efi_memmap_walk(efi_memory_present_wrapper, NULL);
 577                 return;
 578         }
 579
 580         for (i = 0; i < e820.nr_map; i++) {
 581                 unsigned long start, end;
 582                 /* RAM? */
 583                 if (e820.map[i].type != E820_RAM)
 584                         continue;
 585                 start = PFN_UP(e820.map[i].addr);
 586                 end = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
 587                 if (start >= end)
 588                         continue;
 589                 if (end > max_pfn)
 590                         max_pfn = end;
 591                 memory_present(0, start, end);
 592         }
 593 }