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  * Compute how much memory is missing in a range.
 189  * Unlike the other functions in this file the arguments are in page numbers.
 190  */
 191 unsigned long __init
 192 e820_hole_size(unsigned long start_pfn, unsigned long end_pfn)
 193 {
 194         unsigned long ram = 0;
 195         unsigned long start = start_pfn << PAGE_SHIFT;
 196         unsigned long end = end_pfn << PAGE_SHIFT;
 197         int i;
 198         for (i = 0; i < e820.nr_map; i++) {
 199                 struct e820entry *ei = &e820.map[i];
 200                 unsigned long last, addr;
 201
 202                 if (ei->type != E820_RAM ||
 203                     ei->addr+ei->size <= start ||
 204                     ei->addr >= end)
 205                         continue;
 206
 207                 addr = round_up(ei->addr, PAGE_SIZE);
 208                 if (addr < start)
 209                         addr = start;
 210
 211                 last = round_down(ei->addr + ei->size, PAGE_SIZE);
 212                 if (last >= end)
 213                         last = end;
 214
 215                 if (last > addr)
 216                         ram += last - addr;
 217         }
 218         return ((end - start) - ram) >> PAGE_SHIFT;
 219 }
 220
 221 /*
 222  * Mark e820 reserved areas as busy for the resource manager.
 223  */
 224 void __init e820_reserve_resources(void)
 225 {
 226         int i;
 227         for (i = 0; i < e820.nr_map; i++) {
 228                 struct resource *res;
 229                 res = alloc_bootmem_low(sizeof(struct resource));
 230                 switch (e820.map[i].type) {
 231                 case E820_RAM:  res->name = "System RAM"; break;
 232                 case E820_ACPI: res->name = "ACPI Tables"; break;
 233                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 234                 default:        res->name = "reserved";
 235                 }
 236                 res->start = e820.map[i].addr;
 237                 res->end = res->start + e820.map[i].size - 1;
 238                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 239                 request_resource(&iomem_resource, res);
 240                 if (e820.map[i].type == E820_RAM) {
 241                         /*
 242                          *  We don't know which RAM region contains kernel data,
 243                          *  so we try it repeatedly and let the resource manager
 244                          *  test it.
 245                          */
 246                         request_resource(res, &code_resource);
 247                         request_resource(res, &data_resource);
 248 #ifdef CONFIG_KEXEC
 249                         request_resource(res, &crashk_res);
 250 #endif
 251                 }
 252         }
 253 }
 254
 255 /*
 256  * Add a memory region to the kernel e820 map.
 257  */
 258 void __init add_memory_region(unsigned long start, unsigned long size, int type)
 259 {
 260         int x = e820.nr_map;
 261
 262         if (x == E820MAX) {
 263                 printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 264                 return;
 265         }
 266
 267         e820.map[x].addr = start;
 268         e820.map[x].size = size;
 269         e820.map[x].type = type;
 270         e820.nr_map++;
 271 }
 272
 273 void __init e820_print_map(char *who)
 274 {
 275         int i;
 276
 277         for (i = 0; i < e820.nr_map; i++) {
 278                 printk(" %s: %016Lx - %016Lx ", who,
 279                         (unsigned long long) e820.map[i].addr,
 280                         (unsigned long long) (e820.map[i].addr + e820.map[i].size));
 281                 switch (e820.map[i].type) {
 282                 case E820_RAM:  printk("(usable)\n");
 283                                 break;
 284                 case E820_RESERVED:
 285                                 printk("(reserved)\n");
 286                                 break;
 287                 case E820_ACPI:
 288                                 printk("(ACPI data)\n");
 289                                 break;
 290                 case E820_NVS:
 291                                 printk("(ACPI NVS)\n");
 292                                 break;
 293                 default:        printk("type %u\n", e820.map[i].type);
 294                                 break;
 295                 }
 296         }
 297 }
 298
 299 /*
 300  * Sanitize the BIOS e820 map.
 301  *
 302  * Some e820 responses include overlapping entries.  The following
 303  * replaces the original e820 map with a new one, removing overlaps.
 304  *
 305  */
 306 static int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 307 {
 308         struct change_member {
 309                 struct e820entry *pbios; /* pointer to original bios entry */
 310                 unsigned long long addr; /* address for this change point */
 311         };
 312         static struct change_member change_point_list[2*E820MAX] __initdata;
 313         static struct change_member *change_point[2*E820MAX] __initdata;
 314         static struct e820entry *overlap_list[E820MAX] __initdata;
 315         static struct e820entry new_bios[E820MAX] __initdata;
 316         struct change_member *change_tmp;
 317         unsigned long current_type, last_type;
 318         unsigned long long last_addr;
 319         int chgidx, still_changing;
 320         int overlap_entries;
 321         int new_bios_entry;
 322         int old_nr, new_nr, chg_nr;
 323         int i;
 324
 325         /*
 326                 Visually we're performing the following (1,2,3,4 = memory types)...
 327
 328                 Sample memory map (w/overlaps):
 329                    ____22__________________
 330                    ______________________4_
 331                    ____1111________________
 332                    _44_____________________
 333                    11111111________________
 334                    ____________________33__
 335                    ___________44___________
 336                    __________33333_________
 337                    ______________22________
 338                    ___________________2222_
 339                    _________111111111______
 340                    _____________________11_
 341                    _________________4______
 342
 343                 Sanitized equivalent (no overlap):
 344                    1_______________________
 345                    _44_____________________
 346                    ___1____________________
 347                    ____22__________________
 348                    ______11________________
 349                    _________1______________
 350                    __________3_____________
 351                    ___________44___________
 352                    _____________33_________
 353                    _______________2________
 354                    ________________1_______
 355                    _________________4______
 356                    ___________________2____
 357                    ____________________33__
 358                    ______________________4_
 359         */
 360
 361         /* if there's only one memory region, don't bother */
 362         if (*pnr_map < 2)
 363                 return -1;
 364
 365         old_nr = *pnr_map;
 366
 367         /* bail out if we find any unreasonable addresses in bios map */
 368         for (i=0; i<old_nr; i++)
 369                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr)
 370                         return -1;
 371
 372         /* create pointers for initial change-point information (for sorting) */
 373         for (i=0; i < 2*old_nr; i++)
 374                 change_point[i] = &change_point_list[i];
 375
 376         /* record all known change-points (starting and ending addresses),
 377            omitting those that are for empty memory regions */
 378         chgidx = 0;
 379         for (i=0; i < old_nr; i++)      {
 380                 if (biosmap[i].size != 0) {
 381                         change_point[chgidx]->addr = biosmap[i].addr;
 382                         change_point[chgidx++]->pbios = &biosmap[i];
 383                         change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 384                         change_point[chgidx++]->pbios = &biosmap[i];
 385                 }
 386         }
 387         chg_nr = chgidx;
 388
 389         /* sort change-point list by memory addresses (low -> high) */
 390         still_changing = 1;
 391         while (still_changing)  {
 392                 still_changing = 0;
 393                 for (i=1; i < chg_nr; i++)  {
 394                         /* if <current_addr> > <last_addr>, swap */
 395                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 396                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 397                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 398                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 399                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 400                            )
 401                         {
 402                                 change_tmp = change_point[i];
 403                                 change_point[i] = change_point[i-1];
 404                                 change_point[i-1] = change_tmp;
 405                                 still_changing=1;
 406                         }
 407                 }
 408         }
 409
 410         /* create a new bios memory map, removing overlaps */
 411         overlap_entries=0;       /* number of entries in the overlap table */
 412         new_bios_entry=0;        /* index for creating new bios map entries */
 413         last_type = 0;           /* start with undefined memory type */
 414         last_addr = 0;           /* start with 0 as last starting address */
 415         /* loop through change-points, determining affect on the new bios map */
 416         for (chgidx=0; chgidx < chg_nr; chgidx++)
 417         {
 418                 /* keep track of all overlapping bios entries */
 419                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 420                 {
 421                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 422                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 423                 }
 424                 else
 425                 {
 426                         /* remove entry from list (order independent, so swap with last) */
 427                         for (i=0; i<overlap_entries; i++)
 428                         {
 429                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 430                                         overlap_list[i] = overlap_list[overlap_entries-1];
 431                         }
 432                         overlap_entries--;
 433                 }
 434                 /* if there are overlapping entries, decide which "type" to use */
 435                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 436                 current_type = 0;
 437                 for (i=0; i<overlap_entries; i++)
 438                         if (overlap_list[i]->type > current_type)
 439                                 current_type = overlap_list[i]->type;
 440                 /* continue building up new bios map based on this information */
 441                 if (current_type != last_type)  {
 442                         if (last_type != 0)      {
 443                                 new_bios[new_bios_entry].size =
 444                                         change_point[chgidx]->addr - last_addr;
 445                                 /* move forward only if the new size was non-zero */
 446                                 if (new_bios[new_bios_entry].size != 0)
 447                                         if (++new_bios_entry >= E820MAX)
 448                                                 break;  /* no more space left for new bios entries */
 449                         }
 450                         if (current_type != 0)  {
 451                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 452                                 new_bios[new_bios_entry].type = current_type;
 453                                 last_addr=change_point[chgidx]->addr;
 454                         }
 455                         last_type = current_type;
 456                 }
 457         }
 458         new_nr = new_bios_entry;   /* retain count for new bios entries */
 459
 460         /* copy new bios mapping into original location */
 461         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 462         *pnr_map = new_nr;
 463
 464         return 0;
 465 }
 466
 467 /*
 468  * Copy the BIOS e820 map into a safe place.
 469  *
 470  * Sanity-check it while we're at it..
 471  *
 472  * If we're lucky and live on a modern system, the setup code
 473  * will have given us a memory map that we can use to properly
 474  * set up memory.  If we aren't, we'll fake a memory map.
 475  *
 476  * We check to see that the memory map contains at least 2 elements
 477  * before we'll use it, because the detection code in setup.S may
 478  * not be perfect and most every PC known to man has two memory
 479  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 480  * thinkpad 560x, for example, does not cooperate with the memory
 481  * detection code.)
 482  */
 483 static int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 484 {
 485         /* Only one memory region (or negative)? Ignore it */
 486         if (nr_map < 2)
 487                 return -1;
 488
 489         do {
 490                 unsigned long start = biosmap->addr;
 491                 unsigned long size = biosmap->size;
 492                 unsigned long end = start + size;
 493                 unsigned long type = biosmap->type;
 494
 495                 /* Overflow in 64 bits? Ignore the memory map. */
 496                 if (start > end)
 497                         return -1;
 498
 499                 /*
 500                  * Some BIOSes claim RAM in the 640k - 1M region.
 501                  * Not right. Fix it up.
 502                  *
 503                  * This should be removed on Hammer which is supposed to not
 504                  * have non e820 covered ISA mappings there, but I had some strange
 505                  * problems so it stays for now.  -AK
 506                  */
 507                 if (type == E820_RAM) {
 508                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 509                                 if (start < 0xA0000ULL)
 510                                         add_memory_region(start, 0xA0000ULL-start, type);
 511                                 if (end <= 0x100000ULL)
 512                                         continue;
 513                                 start = 0x100000ULL;
 514                                 size = end - start;
 515                         }
 516                 }
 517
 518                 add_memory_region(start, size, type);
 519         } while (biosmap++,--nr_map);
 520         return 0;
 521 }
 522
 523 void __init setup_memory_region(void)
 524 {
 525         char *who = "BIOS-e820";
 526
 527         /*
 528          * Try to copy the BIOS-supplied E820-map.
 529          *
 530          * Otherwise fake a memory map; one section from 0k->640k,
 531          * the next section from 1mb->appropriate_mem_k
 532          */
 533         sanitize_e820_map(E820_MAP, &E820_MAP_NR);
 534         if (copy_e820_map(E820_MAP, E820_MAP_NR) < 0) {
 535                 unsigned long mem_size;
 536
 537                 /* compare results from other methods and take the greater */
 538                 if (ALT_MEM_K < EXT_MEM_K) {
 539                         mem_size = EXT_MEM_K;
 540                         who = "BIOS-88";
 541                 } else {
 542                         mem_size = ALT_MEM_K;
 543                         who = "BIOS-e801";
 544                 }
 545
 546                 e820.nr_map = 0;
 547                 add_memory_region(0, LOWMEMSIZE(), E820_RAM);
 548                 add_memory_region(HIGH_MEMORY, mem_size << 10, E820_RAM);
 549         }
 550         printk(KERN_INFO "BIOS-provided physical RAM map:\n");
 551         e820_print_map(who);
 552 }
 553
 554 void __init parse_memopt(char *p, char **from)
 555 {
 556         end_user_pfn = memparse(p, from);
 557         end_user_pfn >>= PAGE_SHIFT;
 558 }
 559
 560 unsigned long pci_mem_start = 0xaeedbabe;
 561
 562 /*
 563  * Search for the biggest gap in the low 32 bits of the e820
 564  * memory space.  We pass this space to PCI to assign MMIO resources
 565  * for hotplug or unconfigured devices in.
 566  * Hopefully the BIOS let enough space left.
 567  */
 568 __init void e820_setup_gap(void)
 569 {
 570         unsigned long gapstart, gapsize;
 571         unsigned long last;
 572         int i;
 573         int found = 0;
 574
 575         last = 0x100000000ull;
 576         gapstart = 0x10000000;
 577         gapsize = 0x400000;
 578         i = e820.nr_map;
 579         while (--i >= 0) {
 580                 unsigned long long start = e820.map[i].addr;
 581                 unsigned long long end = start + e820.map[i].size;
 582
 583                 /*
 584                  * Since "last" is at most 4GB, we know we'll
 585                  * fit in 32 bits if this condition is true
 586                  */
 587                 if (last > end) {
 588                         unsigned long gap = last - end;
 589
 590                         if (gap > gapsize) {
 591                                 gapsize = gap;
 592                                 gapstart = end;
 593                                 found = 1;
 594                         }
 595                 }
 596                 if (start < last)
 597                         last = start;
 598         }
 599
 600         if (!found) {
 601                 gapstart = (end_pfn << PAGE_SHIFT) + 1024*1024;
 602                 printk(KERN_ERR "PCI: Warning: Cannot find a gap in the 32bit address range\n"
 603                        KERN_ERR "PCI: Unassigned devices with 32bit resource registers may break!\n");
 604         }
 605
 606         /*
 607          * Start allocating dynamic PCI memory a bit into the gap,
 608          * aligned up to the nearest megabyte.
 609          *
 610          * Question: should we try to pad it up a bit (do something
 611          * like " + (gapsize >> 3)" in there too?). We now have the
 612          * technology.
 613          */
 614         pci_mem_start = (gapstart + 0xfffff) & ~0xfffff;
 615
 616         printk(KERN_INFO "Allocating PCI resources starting at %lx (gap: %lx:%lx)\n",
 617                 pci_mem_start, gapstart, gapsize);
 618 }