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 #include <linux/config.h>
   7 #include <linux/kernel.h>
   8 #include <linux/types.h>
   9 #include <linux/init.h>
  10 #include <linux/bootmem.h>
  11 #include <linux/ioport.h>
  12 #include <linux/string.h>
  13 #include <asm/page.h>
  14 #include <asm/e820.h>
  15 #include <asm/proto.h>
  16 #include <asm/bootsetup.h>
  17
  18 extern char _end[];
  19
  20 /*
  21  * PFN of last memory page.
  22  */
  23 unsigned long end_pfn;
  24
  25 /*
  26  * end_pfn only includes RAM, while end_pfn_map includes all e820 entries.
  27  * The direct mapping extends to end_pfn_map, so that we can directly access
  28  * apertures, ACPI and other tables without having to play with fixmaps.
  29  */
  30 unsigned long end_pfn_map;
  31
  32 /*
  33  * Last pfn which the user wants to use.
  34  */
  35 unsigned long end_user_pfn = MAXMEM>>PAGE_SHIFT;
  36
  37 extern struct resource code_resource, data_resource;
  38
  39 /* Check for some hardcoded bad areas that early boot is not allowed to touch */
  40 static inline int bad_addr(unsigned long *addrp, unsigned long size)
  41 {
  42         unsigned long addr = *addrp, last = addr + size;
  43
  44         /* various gunk below that needed for SMP startup */
  45         if (addr < 0x8000) {
  46                 *addrp = 0x8000;
  47                 return 1;
  48         }
  49
  50         /* direct mapping tables of the kernel */
  51         if (last >= table_start<<PAGE_SHIFT && addr < table_end<<PAGE_SHIFT) {
  52                 *addrp = table_end << PAGE_SHIFT;
  53                 return 1;
  54         }
  55
  56         /* initrd */
  57 #ifdef CONFIG_BLK_DEV_INITRD
  58         if (LOADER_TYPE && INITRD_START && last >= INITRD_START &&
  59             addr < INITRD_START+INITRD_SIZE) {
  60                 *addrp = INITRD_START + INITRD_SIZE;
  61                 return 1;
  62         }
  63 #endif
  64         /* kernel code + 640k memory hole (later should not be needed, but
  65            be paranoid for now) */
  66         if (last >= 640*1024 && addr < __pa_symbol(&_end)) {
  67                 *addrp = __pa_symbol(&_end);
  68                 return 1;
  69         }
  70         /* XXX ramdisk image here? */
  71         return 0;
  72 }
  73
  74 int __init e820_mapped(unsigned long start, unsigned long end, unsigned type)
  75 {
  76         int i;
  77         for (i = 0; i < e820.nr_map; i++) {
  78                 struct e820entry *ei = &e820.map[i];
  79                 if (type && ei->type != type)
  80                         continue;
  81                 if (ei->addr >= end || ei->addr + ei->size < start)
  82                         continue;
  83                 return 1;
  84         }
  85         return 0;
  86 }
  87
  88 /*
  89  * Find a free area in a specific range.
  90  */
  91 unsigned long __init find_e820_area(unsigned long start, unsigned long end, unsigned size)
  92 {
  93         int i;
  94         for (i = 0; i < e820.nr_map; i++) {
  95                 struct e820entry *ei = &e820.map[i];
  96                 unsigned long addr = ei->addr, last;
  97                 if (ei->type != E820_RAM)
  98                         continue;
  99                 if (addr < start)
 100                         addr = start;
 101                 if (addr > ei->addr + ei->size)
 102                         continue;
 103                 while (bad_addr(&addr, size) && addr+size < ei->addr + ei->size)
 104                         ;
 105                 last = addr + size;
 106                 if (last > ei->addr + ei->size)
 107                         continue;
 108                 if (last > end)
 109                         continue;
 110                 return addr;
 111         }
 112         return -1UL;
 113 }
 114
 115 /*
 116  * Free bootmem based on the e820 table for a node.
 117  */
 118 void __init e820_bootmem_free(pg_data_t *pgdat, unsigned long start,unsigned long end)
 119 {
 120         int i;
 121         for (i = 0; i < e820.nr_map; i++) {
 122                 struct e820entry *ei = &e820.map[i];
 123                 unsigned long last, addr;
 124
 125                 if (ei->type != E820_RAM ||
 126                     ei->addr+ei->size <= start ||
 127                     ei->addr > end)
 128                         continue;
 129
 130                 addr = round_up(ei->addr, PAGE_SIZE);
 131                 if (addr < start)
 132                         addr = start;
 133
 134                 last = round_down(ei->addr + ei->size, PAGE_SIZE);
 135                 if (last >= end)
 136                         last = end;
 137
 138                 if (last > addr && last-addr >= PAGE_SIZE)
 139                         free_bootmem_node(pgdat, addr, last-addr);
 140         }
 141 }
 142
 143 int __get_cpu_vendor(void)
 144 {
 145         char v[16];
 146         int cpuid_level;
 147         memset(v, 0, 16);
 148         cpuid(0x00000000, &cpuid_level,
 149                 (int *)&v[0],
 150                 (int *)&v[8],
 151                 (int *)&v[4]);
 152         if (!strcmp(v, "GenuineIntel"))
 153                 return X86_VENDOR_INTEL;
 154         else if (!strcmp(v, "AuthenticAMD"))
 155                 return X86_VENDOR_AMD;
 156         else if (!strcmp(v, "CyrixInstead"))
 157                 return X86_VENDOR_CYRIX;
 158 //      else if (!strcmp(v, "Geode by NSC"))
 159 //              return X86_VENDOR_NSC;
 160         else if (!strcmp(v, "UMC UMC UMC "))
 161                 return X86_VENDOR_UMC;
 162         else if (!strcmp(v, "CentaurHauls"))
 163                 return X86_VENDOR_CENTAUR;
 164         else if (!strcmp(v, "NexGenDriven"))
 165                 return X86_VENDOR_NEXGEN;
 166         else if (!strcmp(v, "RiseRiseRise"))
 167                 return X86_VENDOR_RISE;
 168         else if (!strcmp(v, "GenuineTMx86") ||
 169                  !strcmp(v, "TransmetaCPU"))
 170                 return X86_VENDOR_TRANSMETA;
 171 //      else if (!strcmp(v, "SiS SiS SiS "))
 172 //              return X86_VENDOR_SIS;
 173         else
 174                 return X86_VENDOR_UNKNOWN;
 175 }
 176
 177
 178 /*
 179  * Find the highest page frame number we have available
 180  */
 181 unsigned long __init e820_end_of_ram(void)
 182 {
 183         int i;
 184         unsigned long end_pfn = 0;
 185
 186         for (i = 0; i < e820.nr_map; i++) {
 187                 struct e820entry *ei = &e820.map[i];
 188                 unsigned long start, end;
 189
 190                 start = round_up(ei->addr, PAGE_SIZE);
 191                 end = round_down(ei->addr + ei->size, PAGE_SIZE);
 192                 if (__get_cpu_vendor() != X86_VENDOR_AMD)
 193                         end = round_down(min(end, 0xffffffff),PAGE_SIZE);
 194
 195                 if (start >= end)
 196                         continue;
 197                 if (ei->type == E820_RAM) {
 198                 if (end > end_pfn<<PAGE_SHIFT)
 199                         end_pfn = end>>PAGE_SHIFT;
 200                 } else {
 201                         if (end > end_pfn_map<<PAGE_SHIFT)
 202                                 end_pfn_map = end>>PAGE_SHIFT;
 203                 }
 204         }
 205
 206         if (end_pfn > end_pfn_map)
 207                 end_pfn_map = end_pfn;
 208         if (end_pfn_map > MAXMEM>>PAGE_SHIFT)
 209                 end_pfn_map = MAXMEM>>PAGE_SHIFT;
 210         if (end_pfn > end_user_pfn)
 211                 end_pfn = end_user_pfn;
 212         if (end_pfn > end_pfn_map)
 213                 end_pfn = end_pfn_map;
 214
 215         return end_pfn;
 216 }
 217
 218 /*
 219  * Mark e820 reserved areas as busy for the resource manager.
 220  */
 221 void __init e820_reserve_resources(void)
 222 {
 223         int i;
 224         for (i = 0; i < e820.nr_map; i++) {
 225                 struct resource *res;
 226                 if (e820.map[i].addr + e820.map[i].size > 0x100000000ULL)
 227                         continue;
 228                 res = alloc_bootmem_low(sizeof(struct resource));
 229                 switch (e820.map[i].type) {
 230                 case E820_RAM:  res->name = "System RAM"; break;
 231                 case E820_ACPI: res->name = "ACPI Tables"; break;
 232                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 233                 default:        res->name = "reserved";
 234                 }
 235                 res->start = e820.map[i].addr;
 236                 res->end = res->start + e820.map[i].size - 1;
 237                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 238                 request_resource(&iomem_resource, res);
 239                 if (e820.map[i].type == E820_RAM) {
 240                         /*
 241                          *  We don't know which RAM region contains kernel data,
 242                          *  so we try it repeatedly and let the resource manager
 243                          *  test it.
 244                          */
 245                         request_resource(res, &code_resource);
 246                         request_resource(res, &data_resource);
 247                 }
 248         }
 249 }
 250
 251 /*
 252  * Add a memory region to the kernel e820 map.
 253  */
 254 void __init add_memory_region(unsigned long start, unsigned long size, int type)
 255 {
 256         int x = e820.nr_map;
 257
 258         if (x == E820MAX) {
 259                 printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 260                 return;
 261         }
 262
 263         e820.map[x].addr = start;
 264         e820.map[x].size = size;
 265         e820.map[x].type = type;
 266         e820.nr_map++;
 267 }
 268
 269 void __init e820_print_map(char *who)
 270 {
 271         int i;
 272
 273         for (i = 0; i < e820.nr_map; i++) {
 274                 printk(" %s: %016Lx - %016Lx ", who,
 275                         (unsigned long long) e820.map[i].addr,
 276                         (unsigned long long) (e820.map[i].addr + e820.map[i].size));
 277                 switch (e820.map[i].type) {
 278                 case E820_RAM:  printk("(usable)\n");
 279                                 break;
 280                 case E820_RESERVED:
 281                                 printk("(reserved)\n");
 282                                 break;
 283                 case E820_ACPI:
 284                                 printk("(ACPI data)\n");
 285                                 break;
 286                 case E820_NVS:
 287                                 printk("(ACPI NVS)\n");
 288                                 break;
 289                 default:        printk("type %u\n", e820.map[i].type);
 290                                 break;
 291                 }
 292         }
 293 }
 294
 295 /*
 296  * Sanitize the BIOS e820 map.
 297  *
 298  * Some e820 responses include overlapping entries.  The following
 299  * replaces the original e820 map with a new one, removing overlaps.
 300  *
 301  */
 302 static int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 303 {
 304         struct change_member {
 305                 struct e820entry *pbios; /* pointer to original bios entry */
 306                 unsigned long long addr; /* address for this change point */
 307         };
 308         static struct change_member change_point_list[2*E820MAX] __initdata;
 309         static struct change_member *change_point[2*E820MAX] __initdata;
 310         static struct e820entry *overlap_list[E820MAX] __initdata;
 311         static struct e820entry new_bios[E820MAX] __initdata;
 312         struct change_member *change_tmp;
 313         unsigned long current_type, last_type;
 314         unsigned long long last_addr;
 315         int chgidx, still_changing;
 316         int overlap_entries;
 317         int new_bios_entry;
 318         int old_nr, new_nr;
 319         int i;
 320
 321         /*
 322                 Visually we're performing the following (1,2,3,4 = memory types)...
 323
 324                 Sample memory map (w/overlaps):
 325                    ____22__________________
 326                    ______________________4_
 327                    ____1111________________
 328                    _44_____________________
 329                    11111111________________
 330                    ____________________33__
 331                    ___________44___________
 332                    __________33333_________
 333                    ______________22________
 334                    ___________________2222_
 335                    _________111111111______
 336                    _____________________11_
 337                    _________________4______
 338
 339                 Sanitized equivalent (no overlap):
 340                    1_______________________
 341                    _44_____________________
 342                    ___1____________________
 343                    ____22__________________
 344                    ______11________________
 345                    _________1______________
 346                    __________3_____________
 347                    ___________44___________
 348                    _____________33_________
 349                    _______________2________
 350                    ________________1_______
 351                    _________________4______
 352                    ___________________2____
 353                    ____________________33__
 354                    ______________________4_
 355         */
 356
 357         /* if there's only one memory region, don't bother */
 358         if (*pnr_map < 2)
 359                 return -1;
 360
 361         old_nr = *pnr_map;
 362
 363         /* bail out if we find any unreasonable addresses in bios map */
 364         for (i=0; i<old_nr; i++)
 365                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr)
 366                         return -1;
 367
 368         /* create pointers for initial change-point information (for sorting) */
 369         for (i=0; i < 2*old_nr; i++)
 370                 change_point[i] = &change_point_list[i];
 371
 372         /* record all known change-points (starting and ending addresses) */
 373         chgidx = 0;
 374         for (i=0; i < old_nr; i++)      {
 375                 change_point[chgidx]->addr = biosmap[i].addr;
 376                 change_point[chgidx++]->pbios = &biosmap[i];
 377                 change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 378                 change_point[chgidx++]->pbios = &biosmap[i];
 379         }
 380
 381         /* sort change-point list by memory addresses (low -> high) */
 382         still_changing = 1;
 383         while (still_changing)  {
 384                 still_changing = 0;
 385                 for (i=1; i < 2*old_nr; i++)  {
 386                         /* if <current_addr> > <last_addr>, swap */
 387                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 388                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 389                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 390                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 391                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 392                            )
 393                         {
 394                                 change_tmp = change_point[i];
 395                                 change_point[i] = change_point[i-1];
 396                                 change_point[i-1] = change_tmp;
 397                                 still_changing=1;
 398                         }
 399                 }
 400         }
 401
 402         /* create a new bios memory map, removing overlaps */
 403         overlap_entries=0;       /* number of entries in the overlap table */
 404         new_bios_entry=0;        /* index for creating new bios map entries */
 405         last_type = 0;           /* start with undefined memory type */
 406         last_addr = 0;           /* start with 0 as last starting address */
 407         /* loop through change-points, determining affect on the new bios map */
 408         for (chgidx=0; chgidx < 2*old_nr; chgidx++)
 409         {
 410                 /* keep track of all overlapping bios entries */
 411                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 412                 {
 413                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 414                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 415                 }
 416                 else
 417                 {
 418                         /* remove entry from list (order independent, so swap with last) */
 419                         for (i=0; i<overlap_entries; i++)
 420                         {
 421                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 422                                         overlap_list[i] = overlap_list[overlap_entries-1];
 423                         }
 424                         overlap_entries--;
 425                 }
 426                 /* if there are overlapping entries, decide which "type" to use */
 427                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 428                 current_type = 0;
 429                 for (i=0; i<overlap_entries; i++)
 430                         if (overlap_list[i]->type > current_type)
 431                                 current_type = overlap_list[i]->type;
 432                 /* continue building up new bios map based on this information */
 433                 if (current_type != last_type)  {
 434                         if (last_type != 0)      {
 435                                 new_bios[new_bios_entry].size =
 436                                         change_point[chgidx]->addr - last_addr;
 437                                 /* move forward only if the new size was non-zero */
 438                                 if (new_bios[new_bios_entry].size != 0)
 439                                         if (++new_bios_entry >= E820MAX)
 440                                                 break;  /* no more space left for new bios entries */
 441                         }
 442                         if (current_type != 0)  {
 443                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 444                                 new_bios[new_bios_entry].type = current_type;
 445                                 last_addr=change_point[chgidx]->addr;
 446                         }
 447                         last_type = current_type;
 448                 }
 449         }
 450         new_nr = new_bios_entry;   /* retain count for new bios entries */
 451
 452         /* copy new bios mapping into original location */
 453         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 454         *pnr_map = new_nr;
 455
 456         return 0;
 457 }
 458
 459 /*
 460  * Copy the BIOS e820 map into a safe place.
 461  *
 462  * Sanity-check it while we're at it..
 463  *
 464  * If we're lucky and live on a modern system, the setup code
 465  * will have given us a memory map that we can use to properly
 466  * set up memory.  If we aren't, we'll fake a memory map.
 467  *
 468  * We check to see that the memory map contains at least 2 elements
 469  * before we'll use it, because the detection code in setup.S may
 470  * not be perfect and most every PC known to man has two memory
 471  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 472  * thinkpad 560x, for example, does not cooperate with the memory
 473  * detection code.)
 474  */
 475 static int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 476 {
 477         /* Only one memory region (or negative)? Ignore it */
 478         if (nr_map < 2)
 479                 return -1;
 480
 481         do {
 482                 unsigned long start = biosmap->addr;
 483                 unsigned long size = biosmap->size;
 484                 unsigned long end = start + size;
 485                 unsigned long type = biosmap->type;
 486
 487                 /* Overflow in 64 bits? Ignore the memory map. */
 488                 if (start > end)
 489                         return -1;
 490
 491                 /*
 492                  * Some BIOSes claim RAM in the 640k - 1M region.
 493                  * Not right. Fix it up.
 494                  *
 495                  * This should be removed on Hammer which is supposed to not
 496                  * have non e820 covered ISA mappings there, but I had some strange
 497                  * problems so it stays for now.  -AK
 498                  */
 499                 if (type == E820_RAM) {
 500                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 501                                 if (start < 0xA0000ULL)
 502                                         add_memory_region(start, 0xA0000ULL-start, type);
 503                                 if (end <= 0x100000ULL)
 504                                         continue;
 505                                 start = 0x100000ULL;
 506                                 size = end - start;
 507                         }
 508                 }
 509
 510                 add_memory_region(start, size, type);
 511         } while (biosmap++,--nr_map);
 512         return 0;
 513 }
 514
 515 void __init setup_memory_region(void)
 516 {
 517         char *who = "BIOS-e820";
 518
 519         /*
 520          * Try to copy the BIOS-supplied E820-map.
 521          *
 522          * Otherwise fake a memory map; one section from 0k->640k,
 523          * the next section from 1mb->appropriate_mem_k
 524          */
 525         sanitize_e820_map(E820_MAP, &E820_MAP_NR);
 526         if (copy_e820_map(E820_MAP, E820_MAP_NR) < 0) {
 527                 unsigned long mem_size;
 528
 529                 /* compare results from other methods and take the greater */
 530                 if (ALT_MEM_K < EXT_MEM_K) {
 531                         mem_size = EXT_MEM_K;
 532                         who = "BIOS-88";
 533                 } else {
 534                         mem_size = ALT_MEM_K;
 535                         who = "BIOS-e801";
 536                 }
 537
 538                 e820.nr_map = 0;
 539                 add_memory_region(0, LOWMEMSIZE(), E820_RAM);
 540                 add_memory_region(HIGH_MEMORY, mem_size << 10, E820_RAM);
 541         }
 542         printk(KERN_INFO "BIOS-provided physical RAM map:\n");
 543         e820_print_map(who);
 544 }
 545
 546 void __init parse_memopt(char *p, char **from)
 547 {
 548         /*
 549          * mem=XXX[kKmM] limits kernel memory to XXX+1MB
 550          *
 551          * It would be more logical to count from 0 instead of from
 552          * HIGH_MEMORY, but we keep that for now for i386 compatibility.
 553          *
 554          * No support for custom mapping like i386.  The reason is
 555          * that we need to read the e820 map anyways to handle the
 556          * ACPI mappings in the direct map.  Also on x86-64 there
 557          * should be always a good e820 map. This is only an upper
 558          * limit, you cannot force usage of memory not in e820.
 559          *
 560          * -AK
 561                          */
 562         end_user_pfn = memparse(p, from) + HIGH_MEMORY;
 563         end_user_pfn >>= PAGE_SHIFT;
 564 }
 565