arch/x86/kernel/e820.c

   1 /*
   2  * Handle the memory map.
   3  * The functions here do the job until bootmem takes over.
   4  *
   5  *  Getting sanitize_e820_map() in sync with i386 version by applying change:
   6  *  -  Provisions for empty E820 memory regions (reported by certain BIOSes).
   7  *     Alex Achenbach <xela@slit.de>, December 2002.
   8  *  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
   9  *
  10  */
  11 #include <linux/kernel.h>
  12 #include <linux/types.h>
  13 #include <linux/init.h>
  14 #include <linux/bootmem.h>
  15 #include <linux/ioport.h>
  16 #include <linux/string.h>
  17 #include <linux/kexec.h>
  18 #include <linux/module.h>
  19 #include <linux/mm.h>
  20 #include <linux/pfn.h>
  21
  22 #include <asm/pgtable.h>
  23 #include <asm/page.h>
  24 #include <asm/e820.h>
  25 #include <asm/setup.h>
  26
  27 struct e820map e820;
  28
  29 /* For PCI or other memory-mapped resources */
  30 unsigned long pci_mem_start = 0xaeedbabe;
  31 #ifdef CONFIG_PCI
  32 EXPORT_SYMBOL(pci_mem_start);
  33 #endif
  34
  35 /*
  36  * This function checks if any part of the range <start,end> is mapped
  37  * with type.
  38  */
  39 int
  40 e820_any_mapped(u64 start, u64 end, unsigned type)
  41 {
  42         int i;
  43
  44         for (i = 0; i < e820.nr_map; i++) {
  45                 struct e820entry *ei = &e820.map[i];
  46
  47                 if (type && ei->type != type)
  48                         continue;
  49                 if (ei->addr >= end || ei->addr + ei->size <= start)
  50                         continue;
  51                 return 1;
  52         }
  53         return 0;
  54 }
  55 EXPORT_SYMBOL_GPL(e820_any_mapped);
  56
  57 /*
  58  * This function checks if the entire range <start,end> is mapped with type.
  59  *
  60  * Note: this function only works correct if the e820 table is sorted and
  61  * not-overlapping, which is the case
  62  */
  63 int __init e820_all_mapped(u64 start, u64 end, unsigned type)
  64 {
  65         int i;
  66
  67         for (i = 0; i < e820.nr_map; i++) {
  68                 struct e820entry *ei = &e820.map[i];
  69
  70                 if (type && ei->type != type)
  71                         continue;
  72                 /* is the region (part) in overlap with the current region ?*/
  73                 if (ei->addr >= end || ei->addr + ei->size <= start)
  74                         continue;
  75
  76                 /* if the region is at the beginning of <start,end> we move
  77                  * start to the end of the region since it's ok until there
  78                  */
  79                 if (ei->addr <= start)
  80                         start = ei->addr + ei->size;
  81                 /*
  82                  * if start is now at or beyond end, we're done, full
  83                  * coverage
  84                  */
  85                 if (start >= end)
  86                         return 1;
  87         }
  88         return 0;
  89 }
  90
  91 /*
  92  * Add a memory region to the kernel e820 map.
  93  */
  94 void __init add_memory_region(u64 start, u64 size, int type)
  95 {
  96         int x = e820.nr_map;
  97
  98         if (x == ARRAY_SIZE(e820.map)) {
  99                 printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 100                 return;
 101         }
 102
 103         e820.map[x].addr = start;
 104         e820.map[x].size = size;
 105         e820.map[x].type = type;
 106         e820.nr_map++;
 107 }
 108
 109 void __init e820_print_map(char *who)
 110 {
 111         int i;
 112
 113         for (i = 0; i < e820.nr_map; i++) {
 114                 printk(KERN_INFO " %s: %016Lx - %016Lx ", who,
 115                        (unsigned long long) e820.map[i].addr,
 116                        (unsigned long long)
 117                        (e820.map[i].addr + e820.map[i].size));
 118                 switch (e820.map[i].type) {
 119                 case E820_RAM:
 120                         printk(KERN_CONT "(usable)\n");
 121                         break;
 122                 case E820_RESERVED:
 123                         printk(KERN_CONT "(reserved)\n");
 124                         break;
 125                 case E820_ACPI:
 126                         printk(KERN_CONT "(ACPI data)\n");
 127                         break;
 128                 case E820_NVS:
 129                         printk(KERN_CONT "(ACPI NVS)\n");
 130                         break;
 131                 default:
 132                         printk(KERN_CONT "type %u\n", e820.map[i].type);
 133                         break;
 134                 }
 135         }
 136 }
 137
 138 /*
 139  * Sanitize the BIOS e820 map.
 140  *
 141  * Some e820 responses include overlapping entries. The following
 142  * replaces the original e820 map with a new one, removing overlaps.
 143  *
 144  */
 145 int __init sanitize_e820_map(struct e820entry *biosmap, int max_nr_map,
 146                                 int *pnr_map)
 147 {
 148         struct change_member {
 149                 struct e820entry *pbios; /* pointer to original bios entry */
 150                 unsigned long long addr; /* address for this change point */
 151         };
 152 static struct change_member change_point_list[2*E820_X_MAX] __initdata;
 153 static struct change_member *change_point[2*E820_X_MAX] __initdata;
 154 static struct e820entry *overlap_list[E820_X_MAX] __initdata;
 155 static struct e820entry new_bios[E820_X_MAX] __initdata;
 156         struct change_member *change_tmp;
 157         unsigned long current_type, last_type;
 158         unsigned long long last_addr;
 159         int chgidx, still_changing;
 160         int overlap_entries;
 161         int new_bios_entry;
 162         int old_nr, new_nr, chg_nr;
 163         int i;
 164
 165         /*
 166                 Visually we're performing the following
 167                 (1,2,3,4 = memory types)...
 168
 169                 Sample memory map (w/overlaps):
 170                    ____22__________________
 171                    ______________________4_
 172                    ____1111________________
 173                    _44_____________________
 174                    11111111________________
 175                    ____________________33__
 176                    ___________44___________
 177                    __________33333_________
 178                    ______________22________
 179                    ___________________2222_
 180                    _________111111111______
 181                    _____________________11_
 182                    _________________4______
 183
 184                 Sanitized equivalent (no overlap):
 185                    1_______________________
 186                    _44_____________________
 187                    ___1____________________
 188                    ____22__________________
 189                    ______11________________
 190                    _________1______________
 191                    __________3_____________
 192                    ___________44___________
 193                    _____________33_________
 194                    _______________2________
 195                    ________________1_______
 196                    _________________4______
 197                    ___________________2____
 198                    ____________________33__
 199                    ______________________4_
 200         */
 201
 202         /* if there's only one memory region, don't bother */
 203         if (*pnr_map < 2)
 204                 return -1;
 205
 206         old_nr = *pnr_map;
 207         BUG_ON(old_nr > max_nr_map);
 208
 209         /* bail out if we find any unreasonable addresses in bios map */
 210         for (i = 0; i < old_nr; i++)
 211                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr)
 212                         return -1;
 213
 214         /* create pointers for initial change-point information (for sorting) */
 215         for (i = 0; i < 2 * old_nr; i++)
 216                 change_point[i] = &change_point_list[i];
 217
 218         /* record all known change-points (starting and ending addresses),
 219            omitting those that are for empty memory regions */
 220         chgidx = 0;
 221         for (i = 0; i < old_nr; i++)    {
 222                 if (biosmap[i].size != 0) {
 223                         change_point[chgidx]->addr = biosmap[i].addr;
 224                         change_point[chgidx++]->pbios = &biosmap[i];
 225                         change_point[chgidx]->addr = biosmap[i].addr +
 226                                 biosmap[i].size;
 227                         change_point[chgidx++]->pbios = &biosmap[i];
 228                 }
 229         }
 230         chg_nr = chgidx;
 231
 232         /* sort change-point list by memory addresses (low -> high) */
 233         still_changing = 1;
 234         while (still_changing)  {
 235                 still_changing = 0;
 236                 for (i = 1; i < chg_nr; i++)  {
 237                         unsigned long long curaddr, lastaddr;
 238                         unsigned long long curpbaddr, lastpbaddr;
 239
 240                         curaddr = change_point[i]->addr;
 241                         lastaddr = change_point[i - 1]->addr;
 242                         curpbaddr = change_point[i]->pbios->addr;
 243                         lastpbaddr = change_point[i - 1]->pbios->addr;
 244
 245                         /*
 246                          * swap entries, when:
 247                          *
 248                          * curaddr > lastaddr or
 249                          * curaddr == lastaddr and curaddr == curpbaddr and
 250                          * lastaddr != lastpbaddr
 251                          */
 252                         if (curaddr < lastaddr ||
 253                             (curaddr == lastaddr && curaddr == curpbaddr &&
 254                              lastaddr != lastpbaddr)) {
 255                                 change_tmp = change_point[i];
 256                                 change_point[i] = change_point[i-1];
 257                                 change_point[i-1] = change_tmp;
 258                                 still_changing = 1;
 259                         }
 260                 }
 261         }
 262
 263         /* create a new bios memory map, removing overlaps */
 264         overlap_entries = 0;     /* number of entries in the overlap table */
 265         new_bios_entry = 0;      /* index for creating new bios map entries */
 266         last_type = 0;           /* start with undefined memory type */
 267         last_addr = 0;           /* start with 0 as last starting address */
 268
 269         /* loop through change-points, determining affect on the new bios map */
 270         for (chgidx = 0; chgidx < chg_nr; chgidx++) {
 271                 /* keep track of all overlapping bios entries */
 272                 if (change_point[chgidx]->addr ==
 273                     change_point[chgidx]->pbios->addr) {
 274                         /*
 275                          * add map entry to overlap list (> 1 entry
 276                          * implies an overlap)
 277                          */
 278                         overlap_list[overlap_entries++] =
 279                                 change_point[chgidx]->pbios;
 280                 } else {
 281                         /*
 282                          * remove entry from list (order independent,
 283                          * so swap with last)
 284                          */
 285                         for (i = 0; i < overlap_entries; i++) {
 286                                 if (overlap_list[i] ==
 287                                     change_point[chgidx]->pbios)
 288                                         overlap_list[i] =
 289                                                 overlap_list[overlap_entries-1];
 290                         }
 291                         overlap_entries--;
 292                 }
 293                 /*
 294                  * if there are overlapping entries, decide which
 295                  * "type" to use (larger value takes precedence --
 296                  * 1=usable, 2,3,4,4+=unusable)
 297                  */
 298                 current_type = 0;
 299                 for (i = 0; i < overlap_entries; i++)
 300                         if (overlap_list[i]->type > current_type)
 301                                 current_type = overlap_list[i]->type;
 302                 /*
 303                  * continue building up new bios map based on this
 304                  * information
 305                  */
 306                 if (current_type != last_type)  {
 307                         if (last_type != 0)      {
 308                                 new_bios[new_bios_entry].size =
 309                                         change_point[chgidx]->addr - last_addr;
 310                                 /*
 311                                  * move forward only if the new size
 312                                  * was non-zero
 313                                  */
 314                                 if (new_bios[new_bios_entry].size != 0)
 315                                         /*
 316                                          * no more space left for new
 317                                          * bios entries ?
 318                                          */
 319                                         if (++new_bios_entry >= max_nr_map)
 320                                                 break;
 321                         }
 322                         if (current_type != 0)  {
 323                                 new_bios[new_bios_entry].addr =
 324                                         change_point[chgidx]->addr;
 325                                 new_bios[new_bios_entry].type = current_type;
 326                                 last_addr = change_point[chgidx]->addr;
 327                         }
 328                         last_type = current_type;
 329                 }
 330         }
 331         /* retain count for new bios entries */
 332         new_nr = new_bios_entry;
 333
 334         /* copy new bios mapping into original location */
 335         memcpy(biosmap, new_bios, new_nr * sizeof(struct e820entry));
 336         *pnr_map = new_nr;
 337
 338         return 0;
 339 }
 340
 341 /*
 342  * Copy the BIOS e820 map into a safe place.
 343  *
 344  * Sanity-check it while we're at it..
 345  *
 346  * If we're lucky and live on a modern system, the setup code
 347  * will have given us a memory map that we can use to properly
 348  * set up memory.  If we aren't, we'll fake a memory map.
 349  */
 350 int __init copy_e820_map(struct e820entry *biosmap, int nr_map)
 351 {
 352         /* Only one memory region (or negative)? Ignore it */
 353         if (nr_map < 2)
 354                 return -1;
 355
 356         do {
 357                 u64 start = biosmap->addr;
 358                 u64 size = biosmap->size;
 359                 u64 end = start + size;
 360                 u32 type = biosmap->type;
 361
 362                 /* Overflow in 64 bits? Ignore the memory map. */
 363                 if (start > end)
 364                         return -1;
 365
 366                 add_memory_region(start, size, type);
 367         } while (biosmap++, --nr_map);
 368         return 0;
 369 }
 370
 371 u64 __init update_memory_range(u64 start, u64 size, unsigned old_type,
 372                                 unsigned new_type)
 373 {
 374         int i;
 375         u64 real_updated_size = 0;
 376
 377         BUG_ON(old_type == new_type);
 378
 379         for (i = 0; i < e820.nr_map; i++) {
 380                 struct e820entry *ei = &e820.map[i];
 381                 u64 final_start, final_end;
 382                 if (ei->type != old_type)
 383                         continue;
 384                 /* totally covered? */
 385                 if (ei->addr >= start &&
 386                     (ei->addr + ei->size) <= (start + size)) {
 387                         ei->type = new_type;
 388                         real_updated_size += ei->size;
 389                         continue;
 390                 }
 391                 /* partially covered */
 392                 final_start = max(start, ei->addr);
 393                 final_end = min(start + size, ei->addr + ei->size);
 394                 if (final_start >= final_end)
 395                         continue;
 396                 add_memory_region(final_start, final_end - final_start,
 397                                          new_type);
 398                 real_updated_size += final_end - final_start;
 399         }
 400         return real_updated_size;
 401 }
 402
 403 void __init update_e820(void)
 404 {
 405         int nr_map;
 406
 407         nr_map = e820.nr_map;
 408         if (sanitize_e820_map(e820.map, ARRAY_SIZE(e820.map), &nr_map))
 409                 return;
 410         e820.nr_map = nr_map;
 411         printk(KERN_INFO "modified physical RAM map:\n");
 412         e820_print_map("modified");
 413 }
 414
 415 /*
 416  * Search for the biggest gap in the low 32 bits of the e820
 417  * memory space.  We pass this space to PCI to assign MMIO resources
 418  * for hotplug or unconfigured devices in.
 419  * Hopefully the BIOS let enough space left.
 420  */
 421 __init void e820_setup_gap(void)
 422 {
 423         unsigned long gapstart, gapsize, round;
 424         unsigned long long last;
 425         int i;
 426         int found = 0;
 427
 428         last = 0x100000000ull;
 429         gapstart = 0x10000000;
 430         gapsize = 0x400000;
 431         i = e820.nr_map;
 432         while (--i >= 0) {
 433                 unsigned long long start = e820.map[i].addr;
 434                 unsigned long long end = start + e820.map[i].size;
 435
 436                 /*
 437                  * Since "last" is at most 4GB, we know we'll
 438                  * fit in 32 bits if this condition is true
 439                  */
 440                 if (last > end) {
 441                         unsigned long gap = last - end;
 442
 443                         if (gap > gapsize) {
 444                                 gapsize = gap;
 445                                 gapstart = end;
 446                                 found = 1;
 447                         }
 448                 }
 449                 if (start < last)
 450                         last = start;
 451         }
 452
 453 #ifdef CONFIG_X86_64
 454         if (!found) {
 455                 gapstart = (end_pfn << PAGE_SHIFT) + 1024*1024;
 456                 printk(KERN_ERR "PCI: Warning: Cannot find a gap in the 32bit "
 457                        "address range\n"
 458                        KERN_ERR "PCI: Unassigned devices with 32bit resource "
 459                        "registers may break!\n");
 460         }
 461 #endif
 462
 463         /*
 464          * See how much we want to round up: start off with
 465          * rounding to the next 1MB area.
 466          */
 467         round = 0x100000;
 468         while ((gapsize >> 4) > round)
 469                 round += round;
 470         /* Fun with two's complement */
 471         pci_mem_start = (gapstart + round) & -round;
 472
 473         printk(KERN_INFO
 474                "Allocating PCI resources starting at %lx (gap: %lx:%lx)\n",
 475                pci_mem_start, gapstart, gapsize);
 476 }
 477