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 /*
 144  * Find the highest page frame number we have available
 145  */
 146 unsigned long __init e820_end_of_ram(void)
 147 {
 148         int i;
 149         unsigned long end_pfn = 0;
 150
 151         for (i = 0; i < e820.nr_map; i++) {
 152                 struct e820entry *ei = &e820.map[i];
 153                 unsigned long start, end;
 154
 155                 start = round_up(ei->addr, PAGE_SIZE);
 156                 end = round_down(ei->addr + ei->size, PAGE_SIZE);
 157                 if (start >= end)
 158                         continue;
 159                 if (ei->type == E820_RAM) {
 160                 if (end > end_pfn<<PAGE_SHIFT)
 161                         end_pfn = end>>PAGE_SHIFT;
 162                 } else {
 163                         if (end > end_pfn_map<<PAGE_SHIFT)
 164                                 end_pfn_map = end>>PAGE_SHIFT;
 165                 }
 166         }
 167
 168         if (end_pfn > end_pfn_map)
 169                 end_pfn_map = end_pfn;
 170         if (end_pfn_map > MAXMEM>>PAGE_SHIFT)
 171                 end_pfn_map = MAXMEM>>PAGE_SHIFT;
 172         if (end_pfn > end_user_pfn)
 173                 end_pfn = end_user_pfn;
 174         if (end_pfn > end_pfn_map)
 175                 end_pfn = end_pfn_map;
 176
 177         return end_pfn;
 178 }
 179
 180 /*
 181  * Mark e820 reserved areas as busy for the resource manager.
 182  */
 183 void __init e820_reserve_resources(void)
 184 {
 185         int i;
 186         for (i = 0; i < e820.nr_map; i++) {
 187                 struct resource *res;
 188                 res = alloc_bootmem_low(sizeof(struct resource));
 189                 switch (e820.map[i].type) {
 190                 case E820_RAM:  res->name = "System RAM"; break;
 191                 case E820_ACPI: res->name = "ACPI Tables"; break;
 192                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 193                 default:        res->name = "reserved";
 194                 }
 195                 res->start = e820.map[i].addr;
 196                 res->end = res->start + e820.map[i].size - 1;
 197                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 198                 request_resource(&iomem_resource, res);
 199                 if (e820.map[i].type == E820_RAM) {
 200                         /*
 201                          *  We don't know which RAM region contains kernel data,
 202                          *  so we try it repeatedly and let the resource manager
 203                          *  test it.
 204                          */
 205                         request_resource(res, &code_resource);
 206                         request_resource(res, &data_resource);
 207                 }
 208         }
 209 }
 210
 211 /*
 212  * Add a memory region to the kernel e820 map.
 213  */
 214 void __init add_memory_region(unsigned long start, unsigned long size, int type)
 215 {
 216         int x = e820.nr_map;
 217
 218         if (x == E820MAX) {
 219                 printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 220                 return;
 221         }
 222
 223         e820.map[x].addr = start;
 224         e820.map[x].size = size;
 225         e820.map[x].type = type;
 226         e820.nr_map++;
 227 }
 228
 229 void __init e820_print_map(char *who)
 230 {
 231         int i;
 232
 233         for (i = 0; i < e820.nr_map; i++) {
 234                 printk(" %s: %016Lx - %016Lx ", who,
 235                         (unsigned long long) e820.map[i].addr,
 236                         (unsigned long long) (e820.map[i].addr + e820.map[i].size));
 237                 switch (e820.map[i].type) {
 238                 case E820_RAM:  printk("(usable)\n");
 239                                 break;
 240                 case E820_RESERVED:
 241                                 printk("(reserved)\n");
 242                                 break;
 243                 case E820_ACPI:
 244                                 printk("(ACPI data)\n");
 245                                 break;
 246                 case E820_NVS:
 247                                 printk("(ACPI NVS)\n");
 248                                 break;
 249                 default:        printk("type %u\n", e820.map[i].type);
 250                                 break;
 251                 }
 252         }
 253 }
 254
 255 /*
 256  * Sanitize the BIOS e820 map.
 257  *
 258  * Some e820 responses include overlapping entries.  The following
 259  * replaces the original e820 map with a new one, removing overlaps.
 260  *
 261  */
 262 static int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 263 {
 264         struct change_member {
 265                 struct e820entry *pbios; /* pointer to original bios entry */
 266                 unsigned long long addr; /* address for this change point */
 267         };
 268         static struct change_member change_point_list[2*E820MAX] __initdata;
 269         static struct change_member *change_point[2*E820MAX] __initdata;
 270         static struct e820entry *overlap_list[E820MAX] __initdata;
 271         static struct e820entry new_bios[E820MAX] __initdata;
 272         struct change_member *change_tmp;
 273         unsigned long current_type, last_type;
 274         unsigned long long last_addr;
 275         int chgidx, still_changing;
 276         int overlap_entries;
 277         int new_bios_entry;
 278         int old_nr, new_nr;
 279         int i;
 280
 281         /*
 282                 Visually we're performing the following (1,2,3,4 = memory types)...
 283
 284                 Sample memory map (w/overlaps):
 285                    ____22__________________
 286                    ______________________4_
 287                    ____1111________________
 288                    _44_____________________
 289                    11111111________________
 290                    ____________________33__
 291                    ___________44___________
 292                    __________33333_________
 293                    ______________22________
 294                    ___________________2222_
 295                    _________111111111______
 296                    _____________________11_
 297                    _________________4______
 298
 299                 Sanitized equivalent (no overlap):
 300                    1_______________________
 301                    _44_____________________
 302                    ___1____________________
 303                    ____22__________________
 304                    ______11________________
 305                    _________1______________
 306                    __________3_____________
 307                    ___________44___________
 308                    _____________33_________
 309                    _______________2________
 310                    ________________1_______
 311                    _________________4______
 312                    ___________________2____
 313                    ____________________33__
 314                    ______________________4_
 315         */
 316
 317         /* if there's only one memory region, don't bother */
 318         if (*pnr_map < 2)
 319                 return -1;
 320
 321         old_nr = *pnr_map;
 322
 323         /* bail out if we find any unreasonable addresses in bios map */
 324         for (i=0; i<old_nr; i++)
 325                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr)
 326                         return -1;
 327
 328         /* create pointers for initial change-point information (for sorting) */
 329         for (i=0; i < 2*old_nr; i++)
 330                 change_point[i] = &change_point_list[i];
 331
 332         /* record all known change-points (starting and ending addresses) */
 333         chgidx = 0;
 334         for (i=0; i < old_nr; i++)      {
 335                 change_point[chgidx]->addr = biosmap[i].addr;
 336                 change_point[chgidx++]->pbios = &biosmap[i];
 337                 change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 338                 change_point[chgidx++]->pbios = &biosmap[i];
 339         }
 340
 341         /* sort change-point list by memory addresses (low -> high) */
 342         still_changing = 1;
 343         while (still_changing)  {
 344                 still_changing = 0;
 345                 for (i=1; i < 2*old_nr; i++)  {
 346                         /* if <current_addr> > <last_addr>, swap */
 347                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 348                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 349                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 350                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 351                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 352                            )
 353                         {
 354                                 change_tmp = change_point[i];
 355                                 change_point[i] = change_point[i-1];
 356                                 change_point[i-1] = change_tmp;
 357                                 still_changing=1;
 358                         }
 359                 }
 360         }
 361
 362         /* create a new bios memory map, removing overlaps */
 363         overlap_entries=0;       /* number of entries in the overlap table */
 364         new_bios_entry=0;        /* index for creating new bios map entries */
 365         last_type = 0;           /* start with undefined memory type */
 366         last_addr = 0;           /* start with 0 as last starting address */
 367         /* loop through change-points, determining affect on the new bios map */
 368         for (chgidx=0; chgidx < 2*old_nr; chgidx++)
 369         {
 370                 /* keep track of all overlapping bios entries */
 371                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 372                 {
 373                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 374                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 375                 }
 376                 else
 377                 {
 378                         /* remove entry from list (order independent, so swap with last) */
 379                         for (i=0; i<overlap_entries; i++)
 380                         {
 381                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 382                                         overlap_list[i] = overlap_list[overlap_entries-1];
 383                         }
 384                         overlap_entries--;
 385                 }
 386                 /* if there are overlapping entries, decide which "type" to use */
 387                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 388                 current_type = 0;
 389                 for (i=0; i<overlap_entries; i++)
 390                         if (overlap_list[i]->type > current_type)
 391                                 current_type = overlap_list[i]->type;
 392                 /* continue building up new bios map based on this information */
 393                 if (current_type != last_type)  {
 394                         if (last_type != 0)      {
 395                                 new_bios[new_bios_entry].size =
 396                                         change_point[chgidx]->addr - last_addr;
 397                                 /* move forward only if the new size was non-zero */
 398                                 if (new_bios[new_bios_entry].size != 0)
 399                                         if (++new_bios_entry >= E820MAX)
 400                                                 break;  /* no more space left for new bios entries */
 401                         }
 402                         if (current_type != 0)  {
 403                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 404                                 new_bios[new_bios_entry].type = current_type;
 405                                 last_addr=change_point[chgidx]->addr;
 406                         }
 407                         last_type = current_type;
 408                 }
 409         }
 410         new_nr = new_bios_entry;   /* retain count for new bios entries */
 411
 412         /* copy new bios mapping into original location */
 413         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 414         *pnr_map = new_nr;
 415
 416         return 0;
 417 }
 418
 419 /*
 420  * Copy the BIOS e820 map into a safe place.
 421  *
 422  * Sanity-check it while we're at it..
 423  *
 424  * If we're lucky and live on a modern system, the setup code
 425  * will have given us a memory map that we can use to properly
 426  * set up memory.  If we aren't, we'll fake a memory map.
 427  *
 428  * We check to see that the memory map contains at least 2 elements
 429  * before we'll use it, because the detection code in setup.S may
 430  * not be perfect and most every PC known to man has two memory
 431  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 432  * thinkpad 560x, for example, does not cooperate with the memory
 433  * detection code.)
 434  */
 435 static int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 436 {
 437         /* Only one memory region (or negative)? Ignore it */
 438         if (nr_map < 2)
 439                 return -1;
 440
 441         do {
 442                 unsigned long start = biosmap->addr;
 443                 unsigned long size = biosmap->size;
 444                 unsigned long end = start + size;
 445                 unsigned long type = biosmap->type;
 446
 447                 /* Overflow in 64 bits? Ignore the memory map. */
 448                 if (start > end)
 449                         return -1;
 450
 451                 /*
 452                  * Some BIOSes claim RAM in the 640k - 1M region.
 453                  * Not right. Fix it up.
 454                  *
 455                  * This should be removed on Hammer which is supposed to not
 456                  * have non e820 covered ISA mappings there, but I had some strange
 457                  * problems so it stays for now.  -AK
 458                  */
 459                 if (type == E820_RAM) {
 460                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 461                                 if (start < 0xA0000ULL)
 462                                         add_memory_region(start, 0xA0000ULL-start, type);
 463                                 if (end <= 0x100000ULL)
 464                                         continue;
 465                                 start = 0x100000ULL;
 466                                 size = end - start;
 467                         }
 468                 }
 469
 470                 add_memory_region(start, size, type);
 471         } while (biosmap++,--nr_map);
 472         return 0;
 473 }
 474
 475 void __init setup_memory_region(void)
 476 {
 477         char *who = "BIOS-e820";
 478
 479         /*
 480          * Try to copy the BIOS-supplied E820-map.
 481          *
 482          * Otherwise fake a memory map; one section from 0k->640k,
 483          * the next section from 1mb->appropriate_mem_k
 484          */
 485         sanitize_e820_map(E820_MAP, &E820_MAP_NR);
 486         if (copy_e820_map(E820_MAP, E820_MAP_NR) < 0) {
 487                 unsigned long mem_size;
 488
 489                 /* compare results from other methods and take the greater */
 490                 if (ALT_MEM_K < EXT_MEM_K) {
 491                         mem_size = EXT_MEM_K;
 492                         who = "BIOS-88";
 493                 } else {
 494                         mem_size = ALT_MEM_K;
 495                         who = "BIOS-e801";
 496                 }
 497
 498                 e820.nr_map = 0;
 499                 add_memory_region(0, LOWMEMSIZE(), E820_RAM);
 500                 add_memory_region(HIGH_MEMORY, mem_size << 10, E820_RAM);
 501         }
 502         printk(KERN_INFO "BIOS-provided physical RAM map:\n");
 503         e820_print_map(who);
 504 }
 505
 506 void __init parse_memopt(char *p, char **from)
 507 {
 508         /*
 509          * mem=XXX[kKmM] limits kernel memory to XXX+1MB
 510          *
 511          * It would be more logical to count from 0 instead of from
 512          * HIGH_MEMORY, but we keep that for now for i386 compatibility.
 513          *
 514          * No support for custom mapping like i386.  The reason is
 515          * that we need to read the e820 map anyways to handle the
 516          * ACPI mappings in the direct map.  Also on x86-64 there
 517          * should be always a good e820 map. This is only an upper
 518          * limit, you cannot force usage of memory not in e820.
 519          *
 520          * -AK
 521                          */
 522         end_user_pfn = memparse(p, from) + HIGH_MEMORY;
 523         end_user_pfn >>= PAGE_SHIFT;
 524 }
 525