arch/x86/kernel/e820_32.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 #include <linux/uaccess.h>
  13 #include <linux/suspend.h>
  14
  15 #include <asm/pgtable.h>
  16 #include <asm/page.h>
  17 #include <asm/e820.h>
  18 #include <asm/setup.h>
  19
  20 #ifdef CONFIG_EFI
  21 int efi_enabled = 0;
  22 EXPORT_SYMBOL(efi_enabled);
  23 #endif
  24
  25 struct e820map e820;
  26 struct change_member {
  27         struct e820entry *pbios; /* pointer to original bios entry */
  28         unsigned long long addr; /* address for this change point */
  29 };
  30 static struct change_member change_point_list[2*E820MAX] __initdata;
  31 static struct change_member *change_point[2*E820MAX] __initdata;
  32 static struct e820entry *overlap_list[E820MAX] __initdata;
  33 static struct e820entry new_bios[E820MAX] __initdata;
  34 /* For PCI or other memory-mapped resources */
  35 unsigned long pci_mem_start = 0x10000000;
  36 #ifdef CONFIG_PCI
  37 EXPORT_SYMBOL(pci_mem_start);
  38 #endif
  39 extern int user_defined_memmap;
  40 struct resource data_resource = {
  41         .name   = "Kernel data",
  42         .start  = 0,
  43         .end    = 0,
  44         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  45 };
  46
  47 struct resource code_resource = {
  48         .name   = "Kernel code",
  49         .start  = 0,
  50         .end    = 0,
  51         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  52 };
  53
  54 static struct resource system_rom_resource = {
  55         .name   = "System ROM",
  56         .start  = 0xf0000,
  57         .end    = 0xfffff,
  58         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  59 };
  60
  61 static struct resource extension_rom_resource = {
  62         .name   = "Extension ROM",
  63         .start  = 0xe0000,
  64         .end    = 0xeffff,
  65         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  66 };
  67
  68 static struct resource adapter_rom_resources[] = { {
  69         .name   = "Adapter ROM",
  70         .start  = 0xc8000,
  71         .end    = 0,
  72         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  73 }, {
  74         .name   = "Adapter ROM",
  75         .start  = 0,
  76         .end    = 0,
  77         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  78 }, {
  79         .name   = "Adapter ROM",
  80         .start  = 0,
  81         .end    = 0,
  82         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  83 }, {
  84         .name   = "Adapter ROM",
  85         .start  = 0,
  86         .end    = 0,
  87         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  88 }, {
  89         .name   = "Adapter ROM",
  90         .start  = 0,
  91         .end    = 0,
  92         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  93 }, {
  94         .name   = "Adapter ROM",
  95         .start  = 0,
  96         .end    = 0,
  97         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  98 } };
  99
 100 static struct resource video_rom_resource = {
 101         .name   = "Video ROM",
 102         .start  = 0xc0000,
 103         .end    = 0xc7fff,
 104         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
 105 };
 106
 107 static struct resource video_ram_resource = {
 108         .name   = "Video RAM area",
 109         .start  = 0xa0000,
 110         .end    = 0xbffff,
 111         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
 112 };
 113
 114 static struct resource standard_io_resources[] = { {
 115         .name   = "dma1",
 116         .start  = 0x0000,
 117         .end    = 0x001f,
 118         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 119 }, {
 120         .name   = "pic1",
 121         .start  = 0x0020,
 122         .end    = 0x0021,
 123         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 124 }, {
 125         .name   = "timer0",
 126         .start  = 0x0040,
 127         .end    = 0x0043,
 128         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 129 }, {
 130         .name   = "timer1",
 131         .start  = 0x0050,
 132         .end    = 0x0053,
 133         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 134 }, {
 135         .name   = "keyboard",
 136         .start  = 0x0060,
 137         .end    = 0x006f,
 138         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 139 }, {
 140         .name   = "dma page reg",
 141         .start  = 0x0080,
 142         .end    = 0x008f,
 143         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 144 }, {
 145         .name   = "pic2",
 146         .start  = 0x00a0,
 147         .end    = 0x00a1,
 148         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 149 }, {
 150         .name   = "dma2",
 151         .start  = 0x00c0,
 152         .end    = 0x00df,
 153         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 154 }, {
 155         .name   = "fpu",
 156         .start  = 0x00f0,
 157         .end    = 0x00ff,
 158         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 159 } };
 160
 161 #define ROMSIGNATURE 0xaa55
 162
 163 static int __init romsignature(const unsigned char *rom)
 164 {
 165         const unsigned short * const ptr = (const unsigned short *)rom;
 166         unsigned short sig;
 167
 168         return probe_kernel_address(ptr, sig) == 0 && sig == ROMSIGNATURE;
 169 }
 170
 171 static int __init romchecksum(const unsigned char *rom, unsigned long length)
 172 {
 173         unsigned char sum, c;
 174
 175         for (sum = 0; length && probe_kernel_address(rom++, c) == 0; length--)
 176                 sum += c;
 177         return !length && !sum;
 178 }
 179
 180 static void __init probe_roms(void)
 181 {
 182         const unsigned char *rom;
 183         unsigned long start, length, upper;
 184         unsigned char c;
 185         int i;
 186
 187         /* video rom */
 188         upper = adapter_rom_resources[0].start;
 189         for (start = video_rom_resource.start; start < upper; start += 2048) {
 190                 rom = isa_bus_to_virt(start);
 191                 if (!romsignature(rom))
 192                         continue;
 193
 194                 video_rom_resource.start = start;
 195
 196                 if (probe_kernel_address(rom + 2, c) != 0)
 197                         continue;
 198
 199                 /* 0 < length <= 0x7f * 512, historically */
 200                 length = c * 512;
 201
 202                 /* if checksum okay, trust length byte */
 203                 if (length && romchecksum(rom, length))
 204                         video_rom_resource.end = start + length - 1;
 205
 206                 request_resource(&iomem_resource, &video_rom_resource);
 207                 break;
 208         }
 209
 210         start = (video_rom_resource.end + 1 + 2047) & ~2047UL;
 211         if (start < upper)
 212                 start = upper;
 213
 214         /* system rom */
 215         request_resource(&iomem_resource, &system_rom_resource);
 216         upper = system_rom_resource.start;
 217
 218         /* check for extension rom (ignore length byte!) */
 219         rom = isa_bus_to_virt(extension_rom_resource.start);
 220         if (romsignature(rom)) {
 221                 length = extension_rom_resource.end - extension_rom_resource.start + 1;
 222                 if (romchecksum(rom, length)) {
 223                         request_resource(&iomem_resource, &extension_rom_resource);
 224                         upper = extension_rom_resource.start;
 225                 }
 226         }
 227
 228         /* check for adapter roms on 2k boundaries */
 229         for (i = 0; i < ARRAY_SIZE(adapter_rom_resources) && start < upper; start += 2048) {
 230                 rom = isa_bus_to_virt(start);
 231                 if (!romsignature(rom))
 232                         continue;
 233
 234                 if (probe_kernel_address(rom + 2, c) != 0)
 235                         continue;
 236
 237                 /* 0 < length <= 0x7f * 512, historically */
 238                 length = c * 512;
 239
 240                 /* but accept any length that fits if checksum okay */
 241                 if (!length || start + length > upper || !romchecksum(rom, length))
 242                         continue;
 243
 244                 adapter_rom_resources[i].start = start;
 245                 adapter_rom_resources[i].end = start + length - 1;
 246                 request_resource(&iomem_resource, &adapter_rom_resources[i]);
 247
 248                 start = adapter_rom_resources[i++].end & ~2047UL;
 249         }
 250 }
 251
 252 /*
 253  * Request address space for all standard RAM and ROM resources
 254  * and also for regions reported as reserved by the e820.
 255  */
 256 static void __init
 257 legacy_init_iomem_resources(struct resource *code_resource, struct resource *data_resource)
 258 {
 259         int i;
 260
 261         probe_roms();
 262         for (i = 0; i < e820.nr_map; i++) {
 263                 struct resource *res;
 264 #ifndef CONFIG_RESOURCES_64BIT
 265                 if (e820.map[i].addr + e820.map[i].size > 0x100000000ULL)
 266                         continue;
 267 #endif
 268                 res = kzalloc(sizeof(struct resource), GFP_ATOMIC);
 269                 switch (e820.map[i].type) {
 270                 case E820_RAM:  res->name = "System RAM"; break;
 271                 case E820_ACPI: res->name = "ACPI Tables"; break;
 272                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 273                 default:        res->name = "reserved";
 274                 }
 275                 res->start = e820.map[i].addr;
 276                 res->end = res->start + e820.map[i].size - 1;
 277                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 278                 if (request_resource(&iomem_resource, res)) {
 279                         kfree(res);
 280                         continue;
 281                 }
 282                 if (e820.map[i].type == E820_RAM) {
 283                         /*
 284                          *  We don't know which RAM region contains kernel data,
 285                          *  so we try it repeatedly and let the resource manager
 286                          *  test it.
 287                          */
 288                         request_resource(res, code_resource);
 289                         request_resource(res, data_resource);
 290 #ifdef CONFIG_KEXEC
 291                         if (crashk_res.start != crashk_res.end)
 292                                 request_resource(res, &crashk_res);
 293 #endif
 294                 }
 295         }
 296 }
 297
 298 /*
 299  * Request address space for all standard resources
 300  *
 301  * This is called just before pcibios_init(), which is also a
 302  * subsys_initcall, but is linked in later (in arch/i386/pci/common.c).
 303  */
 304 static int __init request_standard_resources(void)
 305 {
 306         int i;
 307
 308         printk("Setting up standard PCI resources\n");
 309         if (efi_enabled)
 310                 efi_initialize_iomem_resources(&code_resource, &data_resource);
 311         else
 312                 legacy_init_iomem_resources(&code_resource, &data_resource);
 313
 314         /* EFI systems may still have VGA */
 315         request_resource(&iomem_resource, &video_ram_resource);
 316
 317         /* request I/O space for devices used on all i[345]86 PCs */
 318         for (i = 0; i < ARRAY_SIZE(standard_io_resources); i++)
 319                 request_resource(&ioport_resource, &standard_io_resources[i]);
 320         return 0;
 321 }
 322
 323 subsys_initcall(request_standard_resources);
 324
 325 #if defined(CONFIG_PM) && defined(CONFIG_HIBERNATION)
 326 /**
 327  * e820_mark_nosave_regions - Find the ranges of physical addresses that do not
 328  * correspond to e820 RAM areas and mark the corresponding pages as nosave for
 329  * hibernation.
 330  *
 331  * This function requires the e820 map to be sorted and without any
 332  * overlapping entries and assumes the first e820 area to be RAM.
 333  */
 334 void __init e820_mark_nosave_regions(void)
 335 {
 336         int i;
 337         unsigned long pfn;
 338
 339         pfn = PFN_DOWN(e820.map[0].addr + e820.map[0].size);
 340         for (i = 1; i < e820.nr_map; i++) {
 341                 struct e820entry *ei = &e820.map[i];
 342
 343                 if (pfn < PFN_UP(ei->addr))
 344                         register_nosave_region(pfn, PFN_UP(ei->addr));
 345
 346                 pfn = PFN_DOWN(ei->addr + ei->size);
 347                 if (ei->type != E820_RAM)
 348                         register_nosave_region(PFN_UP(ei->addr), pfn);
 349
 350                 if (pfn >= max_low_pfn)
 351                         break;
 352         }
 353 }
 354 #endif
 355
 356 void __init add_memory_region(unsigned long long start,
 357                               unsigned long long size, int type)
 358 {
 359         int x;
 360
 361         if (!efi_enabled) {
 362                 x = e820.nr_map;
 363
 364                 if (x == E820MAX) {
 365                     printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 366                     return;
 367                 }
 368
 369                 e820.map[x].addr = start;
 370                 e820.map[x].size = size;
 371                 e820.map[x].type = type;
 372                 e820.nr_map++;
 373         }
 374 } /* add_memory_region */
 375
 376 /*
 377  * Sanitize the BIOS e820 map.
 378  *
 379  * Some e820 responses include overlapping entries.  The following
 380  * replaces the original e820 map with a new one, removing overlaps.
 381  *
 382  */
 383 int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 384 {
 385         struct change_member *change_tmp;
 386         unsigned long current_type, last_type;
 387         unsigned long long last_addr;
 388         int chgidx, still_changing;
 389         int overlap_entries;
 390         int new_bios_entry;
 391         int old_nr, new_nr, chg_nr;
 392         int i;
 393
 394         /*
 395                 Visually we're performing the following (1,2,3,4 = memory types)...
 396
 397                 Sample memory map (w/overlaps):
 398                    ____22__________________
 399                    ______________________4_
 400                    ____1111________________
 401                    _44_____________________
 402                    11111111________________
 403                    ____________________33__
 404                    ___________44___________
 405                    __________33333_________
 406                    ______________22________
 407                    ___________________2222_
 408                    _________111111111______
 409                    _____________________11_
 410                    _________________4______
 411
 412                 Sanitized equivalent (no overlap):
 413                    1_______________________
 414                    _44_____________________
 415                    ___1____________________
 416                    ____22__________________
 417                    ______11________________
 418                    _________1______________
 419                    __________3_____________
 420                    ___________44___________
 421                    _____________33_________
 422                    _______________2________
 423                    ________________1_______
 424                    _________________4______
 425                    ___________________2____
 426                    ____________________33__
 427                    ______________________4_
 428         */
 429         /* if there's only one memory region, don't bother */
 430         if (*pnr_map < 2) {
 431                 return -1;
 432         }
 433
 434         old_nr = *pnr_map;
 435
 436         /* bail out if we find any unreasonable addresses in bios map */
 437         for (i=0; i<old_nr; i++)
 438                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr) {
 439                         return -1;
 440                 }
 441
 442         /* create pointers for initial change-point information (for sorting) */
 443         for (i=0; i < 2*old_nr; i++)
 444                 change_point[i] = &change_point_list[i];
 445
 446         /* record all known change-points (starting and ending addresses),
 447            omitting those that are for empty memory regions */
 448         chgidx = 0;
 449         for (i=0; i < old_nr; i++)      {
 450                 if (biosmap[i].size != 0) {
 451                         change_point[chgidx]->addr = biosmap[i].addr;
 452                         change_point[chgidx++]->pbios = &biosmap[i];
 453                         change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 454                         change_point[chgidx++]->pbios = &biosmap[i];
 455                 }
 456         }
 457         chg_nr = chgidx;        /* true number of change-points */
 458
 459         /* sort change-point list by memory addresses (low -> high) */
 460         still_changing = 1;
 461         while (still_changing)  {
 462                 still_changing = 0;
 463                 for (i=1; i < chg_nr; i++)  {
 464                         /* if <current_addr> > <last_addr>, swap */
 465                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 466                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 467                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 468                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 469                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 470                            )
 471                         {
 472                                 change_tmp = change_point[i];
 473                                 change_point[i] = change_point[i-1];
 474                                 change_point[i-1] = change_tmp;
 475                                 still_changing=1;
 476                         }
 477                 }
 478         }
 479
 480         /* create a new bios memory map, removing overlaps */
 481         overlap_entries=0;       /* number of entries in the overlap table */
 482         new_bios_entry=0;        /* index for creating new bios map entries */
 483         last_type = 0;           /* start with undefined memory type */
 484         last_addr = 0;           /* start with 0 as last starting address */
 485         /* loop through change-points, determining affect on the new bios map */
 486         for (chgidx=0; chgidx < chg_nr; chgidx++)
 487         {
 488                 /* keep track of all overlapping bios entries */
 489                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 490                 {
 491                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 492                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 493                 }
 494                 else
 495                 {
 496                         /* remove entry from list (order independent, so swap with last) */
 497                         for (i=0; i<overlap_entries; i++)
 498                         {
 499                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 500                                         overlap_list[i] = overlap_list[overlap_entries-1];
 501                         }
 502                         overlap_entries--;
 503                 }
 504                 /* if there are overlapping entries, decide which "type" to use */
 505                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 506                 current_type = 0;
 507                 for (i=0; i<overlap_entries; i++)
 508                         if (overlap_list[i]->type > current_type)
 509                                 current_type = overlap_list[i]->type;
 510                 /* continue building up new bios map based on this information */
 511                 if (current_type != last_type)  {
 512                         if (last_type != 0)      {
 513                                 new_bios[new_bios_entry].size =
 514                                         change_point[chgidx]->addr - last_addr;
 515                                 /* move forward only if the new size was non-zero */
 516                                 if (new_bios[new_bios_entry].size != 0)
 517                                         if (++new_bios_entry >= E820MAX)
 518                                                 break;  /* no more space left for new bios entries */
 519                         }
 520                         if (current_type != 0)  {
 521                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 522                                 new_bios[new_bios_entry].type = current_type;
 523                                 last_addr=change_point[chgidx]->addr;
 524                         }
 525                         last_type = current_type;
 526                 }
 527         }
 528         new_nr = new_bios_entry;   /* retain count for new bios entries */
 529
 530         /* copy new bios mapping into original location */
 531         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 532         *pnr_map = new_nr;
 533
 534         return 0;
 535 }
 536
 537 /*
 538  * Copy the BIOS e820 map into a safe place.
 539  *
 540  * Sanity-check it while we're at it..
 541  *
 542  * If we're lucky and live on a modern system, the setup code
 543  * will have given us a memory map that we can use to properly
 544  * set up memory.  If we aren't, we'll fake a memory map.
 545  *
 546  * We check to see that the memory map contains at least 2 elements
 547  * before we'll use it, because the detection code in setup.S may
 548  * not be perfect and most every PC known to man has two memory
 549  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 550  * thinkpad 560x, for example, does not cooperate with the memory
 551  * detection code.)
 552  */
 553 int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 554 {
 555         /* Only one memory region (or negative)? Ignore it */
 556         if (nr_map < 2)
 557                 return -1;
 558
 559         do {
 560                 unsigned long long start = biosmap->addr;
 561                 unsigned long long size = biosmap->size;
 562                 unsigned long long end = start + size;
 563                 unsigned long type = biosmap->type;
 564
 565                 /* Overflow in 64 bits? Ignore the memory map. */
 566                 if (start > end)
 567                         return -1;
 568
 569                 /*
 570                  * Some BIOSes claim RAM in the 640k - 1M region.
 571                  * Not right. Fix it up.
 572                  */
 573                 if (type == E820_RAM) {
 574                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 575                                 if (start < 0xA0000ULL)
 576                                         add_memory_region(start, 0xA0000ULL-start, type);
 577                                 if (end <= 0x100000ULL)
 578                                         continue;
 579                                 start = 0x100000ULL;
 580                                 size = end - start;
 581                         }
 582                 }
 583                 add_memory_region(start, size, type);
 584         } while (biosmap++,--nr_map);
 585         return 0;
 586 }
 587
 588 /*
 589  * Callback for efi_memory_walk.
 590  */
 591 static int __init
 592 efi_find_max_pfn(unsigned long start, unsigned long end, void *arg)
 593 {
 594         unsigned long *max_pfn = arg, pfn;
 595
 596         if (start < end) {
 597                 pfn = PFN_UP(end -1);
 598                 if (pfn > *max_pfn)
 599                         *max_pfn = pfn;
 600         }
 601         return 0;
 602 }
 603
 604 static int __init
 605 efi_memory_present_wrapper(unsigned long start, unsigned long end, void *arg)
 606 {
 607         memory_present(0, PFN_UP(start), PFN_DOWN(end));
 608         return 0;
 609 }
 610
 611 /*
 612  * Find the highest page frame number we have available
 613  */
 614 void __init find_max_pfn(void)
 615 {
 616         int i;
 617
 618         max_pfn = 0;
 619         if (efi_enabled) {
 620                 efi_memmap_walk(efi_find_max_pfn, &max_pfn);
 621                 efi_memmap_walk(efi_memory_present_wrapper, NULL);
 622                 return;
 623         }
 624
 625         for (i = 0; i < e820.nr_map; i++) {
 626                 unsigned long start, end;
 627                 /* RAM? */
 628                 if (e820.map[i].type != E820_RAM)
 629                         continue;
 630                 start = PFN_UP(e820.map[i].addr);
 631                 end = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
 632                 if (start >= end)
 633                         continue;
 634                 if (end > max_pfn)
 635                         max_pfn = end;
 636                 memory_present(0, start, end);
 637         }
 638 }
 639
 640 /*
 641  * Free all available memory for boot time allocation.  Used
 642  * as a callback function by efi_memory_walk()
 643  */
 644
 645 static int __init
 646 free_available_memory(unsigned long start, unsigned long end, void *arg)
 647 {
 648         /* check max_low_pfn */
 649         if (start >= (max_low_pfn << PAGE_SHIFT))
 650                 return 0;
 651         if (end >= (max_low_pfn << PAGE_SHIFT))
 652                 end = max_low_pfn << PAGE_SHIFT;
 653         if (start < end)
 654                 free_bootmem(start, end - start);
 655
 656         return 0;
 657 }
 658 /*
 659  * Register fully available low RAM pages with the bootmem allocator.
 660  */
 661 void __init register_bootmem_low_pages(unsigned long max_low_pfn)
 662 {
 663         int i;
 664
 665         if (efi_enabled) {
 666                 efi_memmap_walk(free_available_memory, NULL);
 667                 return;
 668         }
 669         for (i = 0; i < e820.nr_map; i++) {
 670                 unsigned long curr_pfn, last_pfn, size;
 671                 /*
 672                  * Reserve usable low memory
 673                  */
 674                 if (e820.map[i].type != E820_RAM)
 675                         continue;
 676                 /*
 677                  * We are rounding up the start address of usable memory:
 678                  */
 679                 curr_pfn = PFN_UP(e820.map[i].addr);
 680                 if (curr_pfn >= max_low_pfn)
 681                         continue;
 682                 /*
 683                  * ... and at the end of the usable range downwards:
 684                  */
 685                 last_pfn = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
 686
 687                 if (last_pfn > max_low_pfn)
 688                         last_pfn = max_low_pfn;
 689
 690                 /*
 691                  * .. finally, did all the rounding and playing
 692                  * around just make the area go away?
 693                  */
 694                 if (last_pfn <= curr_pfn)
 695                         continue;
 696
 697                 size = last_pfn - curr_pfn;
 698                 free_bootmem(PFN_PHYS(curr_pfn), PFN_PHYS(size));
 699         }
 700 }
 701
 702 void __init e820_register_memory(void)
 703 {
 704         unsigned long gapstart, gapsize, round;
 705         unsigned long long last;
 706         int i;
 707
 708         /*
 709          * Search for the bigest gap in the low 32 bits of the e820
 710          * memory space.
 711          */
 712         last = 0x100000000ull;
 713         gapstart = 0x10000000;
 714         gapsize = 0x400000;
 715         i = e820.nr_map;
 716         while (--i >= 0) {
 717                 unsigned long long start = e820.map[i].addr;
 718                 unsigned long long end = start + e820.map[i].size;
 719
 720                 /*
 721                  * Since "last" is at most 4GB, we know we'll
 722                  * fit in 32 bits if this condition is true
 723                  */
 724                 if (last > end) {
 725                         unsigned long gap = last - end;
 726
 727                         if (gap > gapsize) {
 728                                 gapsize = gap;
 729                                 gapstart = end;
 730                         }
 731                 }
 732                 if (start < last)
 733                         last = start;
 734         }
 735
 736         /*
 737          * See how much we want to round up: start off with
 738          * rounding to the next 1MB area.
 739          */
 740         round = 0x100000;
 741         while ((gapsize >> 4) > round)
 742                 round += round;
 743         /* Fun with two's complement */
 744         pci_mem_start = (gapstart + round) & -round;
 745
 746         printk("Allocating PCI resources starting at %08lx (gap: %08lx:%08lx)\n",
 747                 pci_mem_start, gapstart, gapsize);
 748 }
 749
 750 void __init print_memory_map(char *who)
 751 {
 752         int i;
 753
 754         for (i = 0; i < e820.nr_map; i++) {
 755                 printk(" %s: %016Lx - %016Lx ", who,
 756                         e820.map[i].addr,
 757                         e820.map[i].addr + e820.map[i].size);
 758                 switch (e820.map[i].type) {
 759                 case E820_RAM:  printk("(usable)\n");
 760                                 break;
 761                 case E820_RESERVED:
 762                                 printk("(reserved)\n");
 763                                 break;
 764                 case E820_ACPI:
 765                                 printk("(ACPI data)\n");
 766                                 break;
 767                 case E820_NVS:
 768                                 printk("(ACPI NVS)\n");
 769                                 break;
 770                 default:        printk("type %u\n", e820.map[i].type);
 771                                 break;
 772                 }
 773         }
 774 }
 775
 776 static __init __always_inline void efi_limit_regions(unsigned long long size)
 777 {
 778         unsigned long long current_addr = 0;
 779         efi_memory_desc_t *md, *next_md;
 780         void *p, *p1;
 781         int i, j;
 782
 783         j = 0;
 784         p1 = memmap.map;
 785         for (p = p1, i = 0; p < memmap.map_end; p += memmap.desc_size, i++) {
 786                 md = p;
 787                 next_md = p1;
 788                 current_addr = md->phys_addr +
 789                         PFN_PHYS(md->num_pages);
 790                 if (is_available_memory(md)) {
 791                         if (md->phys_addr >= size) continue;
 792                         memcpy(next_md, md, memmap.desc_size);
 793                         if (current_addr >= size) {
 794                                 next_md->num_pages -=
 795                                         PFN_UP(current_addr-size);
 796                         }
 797                         p1 += memmap.desc_size;
 798                         next_md = p1;
 799                         j++;
 800                 } else if ((md->attribute & EFI_MEMORY_RUNTIME) ==
 801                            EFI_MEMORY_RUNTIME) {
 802                         /* In order to make runtime services
 803                          * available we have to include runtime
 804                          * memory regions in memory map */
 805                         memcpy(next_md, md, memmap.desc_size);
 806                         p1 += memmap.desc_size;
 807                         next_md = p1;
 808                         j++;
 809                 }
 810         }
 811         memmap.nr_map = j;
 812         memmap.map_end = memmap.map +
 813                 (memmap.nr_map * memmap.desc_size);
 814 }
 815
 816 void __init limit_regions(unsigned long long size)
 817 {
 818         unsigned long long current_addr;
 819         int i;
 820
 821         print_memory_map("limit_regions start");
 822         if (efi_enabled) {
 823                 efi_limit_regions(size);
 824                 return;
 825         }
 826         for (i = 0; i < e820.nr_map; i++) {
 827                 current_addr = e820.map[i].addr + e820.map[i].size;
 828                 if (current_addr < size)
 829                         continue;
 830
 831                 if (e820.map[i].type != E820_RAM)
 832                         continue;
 833
 834                 if (e820.map[i].addr >= size) {
 835                         /*
 836                          * This region starts past the end of the
 837                          * requested size, skip it completely.
 838                          */
 839                         e820.nr_map = i;
 840                 } else {
 841                         e820.nr_map = i + 1;
 842                         e820.map[i].size -= current_addr - size;
 843                 }
 844                 print_memory_map("limit_regions endfor");
 845                 return;
 846         }
 847         print_memory_map("limit_regions endfunc");
 848 }
 849
 850 /*
 851  * This function checks if any part of the range <start,end> is mapped
 852  * with type.
 853  */
 854 int
 855 e820_any_mapped(u64 start, u64 end, unsigned type)
 856 {
 857         int i;
 858         for (i = 0; i < e820.nr_map; i++) {
 859                 const struct e820entry *ei = &e820.map[i];
 860                 if (type && ei->type != type)
 861                         continue;
 862                 if (ei->addr >= end || ei->addr + ei->size <= start)
 863                         continue;
 864                 return 1;
 865         }
 866         return 0;
 867 }
 868 EXPORT_SYMBOL_GPL(e820_any_mapped);
 869
 870  /*
 871   * This function checks if the entire range <start,end> is mapped with type.
 872   *
 873   * Note: this function only works correct if the e820 table is sorted and
 874   * not-overlapping, which is the case
 875   */
 876 int __init
 877 e820_all_mapped(unsigned long s, unsigned long e, unsigned type)
 878 {
 879         u64 start = s;
 880         u64 end = e;
 881         int i;
 882         for (i = 0; i < e820.nr_map; i++) {
 883                 struct e820entry *ei = &e820.map[i];
 884                 if (type && ei->type != type)
 885                         continue;
 886                 /* is the region (part) in overlap with the current region ?*/
 887                 if (ei->addr >= end || ei->addr + ei->size <= start)
 888                         continue;
 889                 /* if the region is at the beginning of <start,end> we move
 890                  * start to the end of the region since it's ok until there
 891                  */
 892                 if (ei->addr <= start)
 893                         start = ei->addr + ei->size;
 894                 /* if start is now at or beyond end, we're done, full
 895                  * coverage */
 896                 if (start >= end)
 897                         return 1; /* we're done */
 898         }
 899         return 0;
 900 }
 901
 902 static int __init parse_memmap(char *arg)
 903 {
 904         if (!arg)
 905                 return -EINVAL;
 906
 907         if (strcmp(arg, "exactmap") == 0) {
 908 #ifdef CONFIG_CRASH_DUMP
 909                 /* If we are doing a crash dump, we
 910                  * still need to know the real mem
 911                  * size before original memory map is
 912                  * reset.
 913                  */
 914                 find_max_pfn();
 915                 saved_max_pfn = max_pfn;
 916 #endif
 917                 e820.nr_map = 0;
 918                 user_defined_memmap = 1;
 919         } else {
 920                 /* If the user specifies memory size, we
 921                  * limit the BIOS-provided memory map to
 922                  * that size. exactmap can be used to specify
 923                  * the exact map. mem=number can be used to
 924                  * trim the existing memory map.
 925                  */
 926                 unsigned long long start_at, mem_size;
 927
 928                 mem_size = memparse(arg, &arg);
 929                 if (*arg == '@') {
 930                         start_at = memparse(arg+1, &arg);
 931                         add_memory_region(start_at, mem_size, E820_RAM);
 932                 } else if (*arg == '#') {
 933                         start_at = memparse(arg+1, &arg);
 934                         add_memory_region(start_at, mem_size, E820_ACPI);
 935                 } else if (*arg == '$') {
 936                         start_at = memparse(arg+1, &arg);
 937                         add_memory_region(start_at, mem_size, E820_RESERVED);
 938                 } else {
 939                         limit_regions(mem_size);
 940                         user_defined_memmap = 1;
 941                 }
 942         }
 943         return 0;
 944 }
 945 early_param("memmap", parse_memmap);