git.oblomov.eu Git - linux-2.6/blob - arch/i386/kernel/e820.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
  14 #include <asm/pgtable.h>
  15 #include <asm/page.h>
  16 #include <asm/e820.h>
  17
  18 #ifdef CONFIG_EFI
  19 int efi_enabled = 0;
  20 EXPORT_SYMBOL(efi_enabled);
  21 #endif
  22
  23 struct e820map e820;
  24 struct change_member {
  25         struct e820entry *pbios; /* pointer to original bios entry */
  26         unsigned long long addr; /* address for this change point */
  27 };
  28 static struct change_member change_point_list[2*E820MAX] __initdata;
  29 static struct change_member *change_point[2*E820MAX] __initdata;
  30 static struct e820entry *overlap_list[E820MAX] __initdata;
  31 static struct e820entry new_bios[E820MAX] __initdata;
  32 /* For PCI or other memory-mapped resources */
  33 unsigned long pci_mem_start = 0x10000000;
  34 #ifdef CONFIG_PCI
  35 EXPORT_SYMBOL(pci_mem_start);
  36 #endif
  37 extern int user_defined_memmap;
  38 struct resource data_resource = {
  39         .name   = "Kernel data",
  40         .start  = 0,
  41         .end    = 0,
  42         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  43 };
  44
  45 struct resource code_resource = {
  46         .name   = "Kernel code",
  47         .start  = 0,
  48         .end    = 0,
  49         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
  50 };
  51
  52 static struct resource system_rom_resource = {
  53         .name   = "System ROM",
  54         .start  = 0xf0000,
  55         .end    = 0xfffff,
  56         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  57 };
  58
  59 static struct resource extension_rom_resource = {
  60         .name   = "Extension ROM",
  61         .start  = 0xe0000,
  62         .end    = 0xeffff,
  63         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  64 };
  65
  66 static struct resource adapter_rom_resources[] = { {
  67         .name   = "Adapter ROM",
  68         .start  = 0xc8000,
  69         .end    = 0,
  70         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  71 }, {
  72         .name   = "Adapter ROM",
  73         .start  = 0,
  74         .end    = 0,
  75         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  76 }, {
  77         .name   = "Adapter ROM",
  78         .start  = 0,
  79         .end    = 0,
  80         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  81 }, {
  82         .name   = "Adapter ROM",
  83         .start  = 0,
  84         .end    = 0,
  85         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  86 }, {
  87         .name   = "Adapter ROM",
  88         .start  = 0,
  89         .end    = 0,
  90         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  91 }, {
  92         .name   = "Adapter ROM",
  93         .start  = 0,
  94         .end    = 0,
  95         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
  96 } };
  97
  98 static struct resource video_rom_resource = {
  99         .name   = "Video ROM",
 100         .start  = 0xc0000,
 101         .end    = 0xc7fff,
 102         .flags  = IORESOURCE_BUSY | IORESOURCE_READONLY | IORESOURCE_MEM
 103 };
 104
 105 static struct resource video_ram_resource = {
 106         .name   = "Video RAM area",
 107         .start  = 0xa0000,
 108         .end    = 0xbffff,
 109         .flags  = IORESOURCE_BUSY | IORESOURCE_MEM
 110 };
 111
 112 static struct resource standard_io_resources[] = { {
 113         .name   = "dma1",
 114         .start  = 0x0000,
 115         .end    = 0x001f,
 116         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 117 }, {
 118         .name   = "pic1",
 119         .start  = 0x0020,
 120         .end    = 0x0021,
 121         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 122 }, {
 123         .name   = "timer0",
 124         .start  = 0x0040,
 125         .end    = 0x0043,
 126         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 127 }, {
 128         .name   = "timer1",
 129         .start  = 0x0050,
 130         .end    = 0x0053,
 131         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 132 }, {
 133         .name   = "keyboard",
 134         .start  = 0x0060,
 135         .end    = 0x006f,
 136         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 137 }, {
 138         .name   = "dma page reg",
 139         .start  = 0x0080,
 140         .end    = 0x008f,
 141         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 142 }, {
 143         .name   = "pic2",
 144         .start  = 0x00a0,
 145         .end    = 0x00a1,
 146         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 147 }, {
 148         .name   = "dma2",
 149         .start  = 0x00c0,
 150         .end    = 0x00df,
 151         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 152 }, {
 153         .name   = "fpu",
 154         .start  = 0x00f0,
 155         .end    = 0x00ff,
 156         .flags  = IORESOURCE_BUSY | IORESOURCE_IO
 157 } };
 158
 159 static int romsignature(const unsigned char *x)
 160 {
 161         unsigned short sig;
 162         int ret = 0;
 163         if (probe_kernel_address((const unsigned short *)x, sig) == 0)
 164                 ret = (sig == 0xaa55);
 165         return ret;
 166 }
 167
 168 static int __init romchecksum(unsigned char *rom, unsigned long length)
 169 {
 170         unsigned char *p, sum = 0;
 171
 172         for (p = rom; p < rom + length; p++)
 173                 sum += *p;
 174         return sum == 0;
 175 }
 176
 177 static void __init probe_roms(void)
 178 {
 179         unsigned long start, length, upper;
 180         unsigned char *rom;
 181         int           i;
 182
 183         /* video rom */
 184         upper = adapter_rom_resources[0].start;
 185         for (start = video_rom_resource.start; start < upper; start += 2048) {
 186                 rom = isa_bus_to_virt(start);
 187                 if (!romsignature(rom))
 188                         continue;
 189
 190                 video_rom_resource.start = start;
 191
 192                 /* 0 < length <= 0x7f * 512, historically */
 193                 length = rom[2] * 512;
 194
 195                 /* if checksum okay, trust length byte */
 196                 if (length && romchecksum(rom, length))
 197                         video_rom_resource.end = start + length - 1;
 198
 199                 request_resource(&iomem_resource, &video_rom_resource);
 200                 break;
 201         }
 202
 203         start = (video_rom_resource.end + 1 + 2047) & ~2047UL;
 204         if (start < upper)
 205                 start = upper;
 206
 207         /* system rom */
 208         request_resource(&iomem_resource, &system_rom_resource);
 209         upper = system_rom_resource.start;
 210
 211         /* check for extension rom (ignore length byte!) */
 212         rom = isa_bus_to_virt(extension_rom_resource.start);
 213         if (romsignature(rom)) {
 214                 length = extension_rom_resource.end - extension_rom_resource.start + 1;
 215                 if (romchecksum(rom, length)) {
 216                         request_resource(&iomem_resource, &extension_rom_resource);
 217                         upper = extension_rom_resource.start;
 218                 }
 219         }
 220
 221         /* check for adapter roms on 2k boundaries */
 222         for (i = 0; i < ARRAY_SIZE(adapter_rom_resources) && start < upper; start += 2048) {
 223                 rom = isa_bus_to_virt(start);
 224                 if (!romsignature(rom))
 225                         continue;
 226
 227                 /* 0 < length <= 0x7f * 512, historically */
 228                 length = rom[2] * 512;
 229
 230                 /* but accept any length that fits if checksum okay */
 231                 if (!length || start + length > upper || !romchecksum(rom, length))
 232                         continue;
 233
 234                 adapter_rom_resources[i].start = start;
 235                 adapter_rom_resources[i].end = start + length - 1;
 236                 request_resource(&iomem_resource, &adapter_rom_resources[i]);
 237
 238                 start = adapter_rom_resources[i++].end & ~2047UL;
 239         }
 240 }
 241
 242 /*
 243  * Request address space for all standard RAM and ROM resources
 244  * and also for regions reported as reserved by the e820.
 245  */
 246 static void __init
 247 legacy_init_iomem_resources(struct resource *code_resource, struct resource *data_resource)
 248 {
 249         int i;
 250
 251         probe_roms();
 252         for (i = 0; i < e820.nr_map; i++) {
 253                 struct resource *res;
 254 #ifndef CONFIG_RESOURCES_64BIT
 255                 if (e820.map[i].addr + e820.map[i].size > 0x100000000ULL)
 256                         continue;
 257 #endif
 258                 res = kzalloc(sizeof(struct resource), GFP_ATOMIC);
 259                 switch (e820.map[i].type) {
 260                 case E820_RAM:  res->name = "System RAM"; break;
 261                 case E820_ACPI: res->name = "ACPI Tables"; break;
 262                 case E820_NVS:  res->name = "ACPI Non-volatile Storage"; break;
 263                 default:        res->name = "reserved";
 264                 }
 265                 res->start = e820.map[i].addr;
 266                 res->end = res->start + e820.map[i].size - 1;
 267                 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY;
 268                 if (request_resource(&iomem_resource, res)) {
 269                         kfree(res);
 270                         continue;
 271                 }
 272                 if (e820.map[i].type == E820_RAM) {
 273                         /*
 274                          *  We don't know which RAM region contains kernel data,
 275                          *  so we try it repeatedly and let the resource manager
 276                          *  test it.
 277                          */
 278                         request_resource(res, code_resource);
 279                         request_resource(res, data_resource);
 280 #ifdef CONFIG_KEXEC
 281                         request_resource(res, &crashk_res);
 282 #endif
 283                 }
 284         }
 285 }
 286
 287 /*
 288  * Request address space for all standard resources
 289  *
 290  * This is called just before pcibios_init(), which is also a
 291  * subsys_initcall, but is linked in later (in arch/i386/pci/common.c).
 292  */
 293 static int __init request_standard_resources(void)
 294 {
 295         int i;
 296
 297         printk("Setting up standard PCI resources\n");
 298         if (efi_enabled)
 299                 efi_initialize_iomem_resources(&code_resource, &data_resource);
 300         else
 301                 legacy_init_iomem_resources(&code_resource, &data_resource);
 302
 303         /* EFI systems may still have VGA */
 304         request_resource(&iomem_resource, &video_ram_resource);
 305
 306         /* request I/O space for devices used on all i[345]86 PCs */
 307         for (i = 0; i < ARRAY_SIZE(standard_io_resources); i++)
 308                 request_resource(&ioport_resource, &standard_io_resources[i]);
 309         return 0;
 310 }
 311
 312 subsys_initcall(request_standard_resources);
 313
 314 void __init add_memory_region(unsigned long long start,
 315                               unsigned long long size, int type)
 316 {
 317         int x;
 318
 319         if (!efi_enabled) {
 320                 x = e820.nr_map;
 321
 322                 if (x == E820MAX) {
 323                     printk(KERN_ERR "Ooops! Too many entries in the memory map!\n");
 324                     return;
 325                 }
 326
 327                 e820.map[x].addr = start;
 328                 e820.map[x].size = size;
 329                 e820.map[x].type = type;
 330                 e820.nr_map++;
 331         }
 332 } /* add_memory_region */
 333
 334 /*
 335  * Sanitize the BIOS e820 map.
 336  *
 337  * Some e820 responses include overlapping entries.  The following
 338  * replaces the original e820 map with a new one, removing overlaps.
 339  *
 340  */
 341 int __init sanitize_e820_map(struct e820entry * biosmap, char * pnr_map)
 342 {
 343         struct change_member *change_tmp;
 344         unsigned long current_type, last_type;
 345         unsigned long long last_addr;
 346         int chgidx, still_changing;
 347         int overlap_entries;
 348         int new_bios_entry;
 349         int old_nr, new_nr, chg_nr;
 350         int i;
 351
 352         /*
 353                 Visually we're performing the following (1,2,3,4 = memory types)...
 354
 355                 Sample memory map (w/overlaps):
 356                    ____22__________________
 357                    ______________________4_
 358                    ____1111________________
 359                    _44_____________________
 360                    11111111________________
 361                    ____________________33__
 362                    ___________44___________
 363                    __________33333_________
 364                    ______________22________
 365                    ___________________2222_
 366                    _________111111111______
 367                    _____________________11_
 368                    _________________4______
 369
 370                 Sanitized equivalent (no overlap):
 371                    1_______________________
 372                    _44_____________________
 373                    ___1____________________
 374                    ____22__________________
 375                    ______11________________
 376                    _________1______________
 377                    __________3_____________
 378                    ___________44___________
 379                    _____________33_________
 380                    _______________2________
 381                    ________________1_______
 382                    _________________4______
 383                    ___________________2____
 384                    ____________________33__
 385                    ______________________4_
 386         */
 387         printk("sanitize start\n");
 388         /* if there's only one memory region, don't bother */
 389         if (*pnr_map < 2) {
 390                 printk("sanitize bail 0\n");
 391                 return -1;
 392         }
 393
 394         old_nr = *pnr_map;
 395
 396         /* bail out if we find any unreasonable addresses in bios map */
 397         for (i=0; i<old_nr; i++)
 398                 if (biosmap[i].addr + biosmap[i].size < biosmap[i].addr) {
 399                         printk("sanitize bail 1\n");
 400                         return -1;
 401                 }
 402
 403         /* create pointers for initial change-point information (for sorting) */
 404         for (i=0; i < 2*old_nr; i++)
 405                 change_point[i] = &change_point_list[i];
 406
 407         /* record all known change-points (starting and ending addresses),
 408            omitting those that are for empty memory regions */
 409         chgidx = 0;
 410         for (i=0; i < old_nr; i++)      {
 411                 if (biosmap[i].size != 0) {
 412                         change_point[chgidx]->addr = biosmap[i].addr;
 413                         change_point[chgidx++]->pbios = &biosmap[i];
 414                         change_point[chgidx]->addr = biosmap[i].addr + biosmap[i].size;
 415                         change_point[chgidx++]->pbios = &biosmap[i];
 416                 }
 417         }
 418         chg_nr = chgidx;        /* true number of change-points */
 419
 420         /* sort change-point list by memory addresses (low -> high) */
 421         still_changing = 1;
 422         while (still_changing)  {
 423                 still_changing = 0;
 424                 for (i=1; i < chg_nr; i++)  {
 425                         /* if <current_addr> > <last_addr>, swap */
 426                         /* or, if current=<start_addr> & last=<end_addr>, swap */
 427                         if ((change_point[i]->addr < change_point[i-1]->addr) ||
 428                                 ((change_point[i]->addr == change_point[i-1]->addr) &&
 429                                  (change_point[i]->addr == change_point[i]->pbios->addr) &&
 430                                  (change_point[i-1]->addr != change_point[i-1]->pbios->addr))
 431                            )
 432                         {
 433                                 change_tmp = change_point[i];
 434                                 change_point[i] = change_point[i-1];
 435                                 change_point[i-1] = change_tmp;
 436                                 still_changing=1;
 437                         }
 438                 }
 439         }
 440
 441         /* create a new bios memory map, removing overlaps */
 442         overlap_entries=0;       /* number of entries in the overlap table */
 443         new_bios_entry=0;        /* index for creating new bios map entries */
 444         last_type = 0;           /* start with undefined memory type */
 445         last_addr = 0;           /* start with 0 as last starting address */
 446         /* loop through change-points, determining affect on the new bios map */
 447         for (chgidx=0; chgidx < chg_nr; chgidx++)
 448         {
 449                 /* keep track of all overlapping bios entries */
 450                 if (change_point[chgidx]->addr == change_point[chgidx]->pbios->addr)
 451                 {
 452                         /* add map entry to overlap list (> 1 entry implies an overlap) */
 453                         overlap_list[overlap_entries++]=change_point[chgidx]->pbios;
 454                 }
 455                 else
 456                 {
 457                         /* remove entry from list (order independent, so swap with last) */
 458                         for (i=0; i<overlap_entries; i++)
 459                         {
 460                                 if (overlap_list[i] == change_point[chgidx]->pbios)
 461                                         overlap_list[i] = overlap_list[overlap_entries-1];
 462                         }
 463                         overlap_entries--;
 464                 }
 465                 /* if there are overlapping entries, decide which "type" to use */
 466                 /* (larger value takes precedence -- 1=usable, 2,3,4,4+=unusable) */
 467                 current_type = 0;
 468                 for (i=0; i<overlap_entries; i++)
 469                         if (overlap_list[i]->type > current_type)
 470                                 current_type = overlap_list[i]->type;
 471                 /* continue building up new bios map based on this information */
 472                 if (current_type != last_type)  {
 473                         if (last_type != 0)      {
 474                                 new_bios[new_bios_entry].size =
 475                                         change_point[chgidx]->addr - last_addr;
 476                                 /* move forward only if the new size was non-zero */
 477                                 if (new_bios[new_bios_entry].size != 0)
 478                                         if (++new_bios_entry >= E820MAX)
 479                                                 break;  /* no more space left for new bios entries */
 480                         }
 481                         if (current_type != 0)  {
 482                                 new_bios[new_bios_entry].addr = change_point[chgidx]->addr;
 483                                 new_bios[new_bios_entry].type = current_type;
 484                                 last_addr=change_point[chgidx]->addr;
 485                         }
 486                         last_type = current_type;
 487                 }
 488         }
 489         new_nr = new_bios_entry;   /* retain count for new bios entries */
 490
 491         /* copy new bios mapping into original location */
 492         memcpy(biosmap, new_bios, new_nr*sizeof(struct e820entry));
 493         *pnr_map = new_nr;
 494
 495         printk("sanitize end\n");
 496         return 0;
 497 }
 498
 499 /*
 500  * Copy the BIOS e820 map into a safe place.
 501  *
 502  * Sanity-check it while we're at it..
 503  *
 504  * If we're lucky and live on a modern system, the setup code
 505  * will have given us a memory map that we can use to properly
 506  * set up memory.  If we aren't, we'll fake a memory map.
 507  *
 508  * We check to see that the memory map contains at least 2 elements
 509  * before we'll use it, because the detection code in setup.S may
 510  * not be perfect and most every PC known to man has two memory
 511  * regions: one from 0 to 640k, and one from 1mb up.  (The IBM
 512  * thinkpad 560x, for example, does not cooperate with the memory
 513  * detection code.)
 514  */
 515 int __init copy_e820_map(struct e820entry * biosmap, int nr_map)
 516 {
 517         /* Only one memory region (or negative)? Ignore it */
 518         if (nr_map < 2)
 519                 return -1;
 520
 521         do {
 522                 unsigned long long start = biosmap->addr;
 523                 unsigned long long size = biosmap->size;
 524                 unsigned long long end = start + size;
 525                 unsigned long type = biosmap->type;
 526                 printk("copy_e820_map() start: %016Lx size: %016Lx end: %016Lx type: %ld\n", start, size, end, type);
 527
 528                 /* Overflow in 64 bits? Ignore the memory map. */
 529                 if (start > end)
 530                         return -1;
 531
 532                 /*
 533                  * Some BIOSes claim RAM in the 640k - 1M region.
 534                  * Not right. Fix it up.
 535                  */
 536                 if (type == E820_RAM) {
 537                         printk("copy_e820_map() type is E820_RAM\n");
 538                         if (start < 0x100000ULL && end > 0xA0000ULL) {
 539                                 printk("copy_e820_map() lies in range...\n");
 540                                 if (start < 0xA0000ULL) {
 541                                         printk("copy_e820_map() start < 0xA0000ULL\n");
 542                                         add_memory_region(start, 0xA0000ULL-start, type);
 543                                 }
 544                                 if (end <= 0x100000ULL) {
 545                                         printk("copy_e820_map() end <= 0x100000ULL\n");
 546                                         continue;
 547                                 }
 548                                 start = 0x100000ULL;
 549                                 size = end - start;
 550                         }
 551                 }
 552                 add_memory_region(start, size, type);
 553         } while (biosmap++,--nr_map);
 554         return 0;
 555 }
 556
 557 /*
 558  * Callback for efi_memory_walk.
 559  */
 560 static int __init
 561 efi_find_max_pfn(unsigned long start, unsigned long end, void *arg)
 562 {
 563         unsigned long *max_pfn = arg, pfn;
 564
 565         if (start < end) {
 566                 pfn = PFN_UP(end -1);
 567                 if (pfn > *max_pfn)
 568                         *max_pfn = pfn;
 569         }
 570         return 0;
 571 }
 572
 573 static int __init
 574 efi_memory_present_wrapper(unsigned long start, unsigned long end, void *arg)
 575 {
 576         memory_present(0, PFN_UP(start), PFN_DOWN(end));
 577         return 0;
 578 }
 579
 580 /*
 581  * Find the highest page frame number we have available
 582  */
 583 void __init find_max_pfn(void)
 584 {
 585         int i;
 586
 587         max_pfn = 0;
 588         if (efi_enabled) {
 589                 efi_memmap_walk(efi_find_max_pfn, &max_pfn);
 590                 efi_memmap_walk(efi_memory_present_wrapper, NULL);
 591                 return;
 592         }
 593
 594         for (i = 0; i < e820.nr_map; i++) {
 595                 unsigned long start, end;
 596                 /* RAM? */
 597                 if (e820.map[i].type != E820_RAM)
 598                         continue;
 599                 start = PFN_UP(e820.map[i].addr);
 600                 end = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
 601                 if (start >= end)
 602                         continue;
 603                 if (end > max_pfn)
 604                         max_pfn = end;
 605                 memory_present(0, start, end);
 606         }
 607 }
 608
 609 /*
 610  * Free all available memory for boot time allocation.  Used
 611  * as a callback function by efi_memory_walk()
 612  */
 613
 614 static int __init
 615 free_available_memory(unsigned long start, unsigned long end, void *arg)
 616 {
 617         /* check max_low_pfn */
 618         if (start >= (max_low_pfn << PAGE_SHIFT))
 619                 return 0;
 620         if (end >= (max_low_pfn << PAGE_SHIFT))
 621                 end = max_low_pfn << PAGE_SHIFT;
 622         if (start < end)
 623                 free_bootmem(start, end - start);
 624
 625         return 0;
 626 }
 627 /*
 628  * Register fully available low RAM pages with the bootmem allocator.
 629  */
 630 void __init register_bootmem_low_pages(unsigned long max_low_pfn)
 631 {
 632         int i;
 633
 634         if (efi_enabled) {
 635                 efi_memmap_walk(free_available_memory, NULL);
 636                 return;
 637         }
 638         for (i = 0; i < e820.nr_map; i++) {
 639                 unsigned long curr_pfn, last_pfn, size;
 640                 /*
 641                  * Reserve usable low memory
 642                  */
 643                 if (e820.map[i].type != E820_RAM)
 644                         continue;
 645                 /*
 646                  * We are rounding up the start address of usable memory:
 647                  */
 648                 curr_pfn = PFN_UP(e820.map[i].addr);
 649                 if (curr_pfn >= max_low_pfn)
 650                         continue;
 651                 /*
 652                  * ... and at the end of the usable range downwards:
 653                  */
 654                 last_pfn = PFN_DOWN(e820.map[i].addr + e820.map[i].size);
 655
 656                 if (last_pfn > max_low_pfn)
 657                         last_pfn = max_low_pfn;
 658
 659                 /*
 660                  * .. finally, did all the rounding and playing
 661                  * around just make the area go away?
 662                  */
 663                 if (last_pfn <= curr_pfn)
 664                         continue;
 665
 666                 size = last_pfn - curr_pfn;
 667                 free_bootmem(PFN_PHYS(curr_pfn), PFN_PHYS(size));
 668         }
 669 }
 670
 671 void __init register_memory(void)
 672 {
 673         unsigned long gapstart, gapsize, round;
 674         unsigned long long last;
 675         int i;
 676
 677         /*
 678          * Search for the bigest gap in the low 32 bits of the e820
 679          * memory space.
 680          */
 681         last = 0x100000000ull;
 682         gapstart = 0x10000000;
 683         gapsize = 0x400000;
 684         i = e820.nr_map;
 685         while (--i >= 0) {
 686                 unsigned long long start = e820.map[i].addr;
 687                 unsigned long long end = start + e820.map[i].size;
 688
 689                 /*
 690                  * Since "last" is at most 4GB, we know we'll
 691                  * fit in 32 bits if this condition is true
 692                  */
 693                 if (last > end) {
 694                         unsigned long gap = last - end;
 695
 696                         if (gap > gapsize) {
 697                                 gapsize = gap;
 698                                 gapstart = end;
 699                         }
 700                 }
 701                 if (start < last)
 702                         last = start;
 703         }
 704
 705         /*
 706          * See how much we want to round up: start off with
 707          * rounding to the next 1MB area.
 708          */
 709         round = 0x100000;
 710         while ((gapsize >> 4) > round)
 711                 round += round;
 712         /* Fun with two's complement */
 713         pci_mem_start = (gapstart + round) & -round;
 714
 715         printk("Allocating PCI resources starting at %08lx (gap: %08lx:%08lx)\n",
 716                 pci_mem_start, gapstart, gapsize);
 717 }
 718
 719 void __init print_memory_map(char *who)
 720 {
 721         int i;
 722
 723         for (i = 0; i < e820.nr_map; i++) {
 724                 printk(" %s: %016Lx - %016Lx ", who,
 725                         e820.map[i].addr,
 726                         e820.map[i].addr + e820.map[i].size);
 727                 switch (e820.map[i].type) {
 728                 case E820_RAM:  printk("(usable)\n");
 729                                 break;
 730                 case E820_RESERVED:
 731                                 printk("(reserved)\n");
 732                                 break;
 733                 case E820_ACPI:
 734                                 printk("(ACPI data)\n");
 735                                 break;
 736                 case E820_NVS:
 737                                 printk("(ACPI NVS)\n");
 738                                 break;
 739                 default:        printk("type %lu\n", e820.map[i].type);
 740                                 break;
 741                 }
 742         }
 743 }
 744
 745 static __init __always_inline void efi_limit_regions(unsigned long long size)
 746 {
 747         unsigned long long current_addr = 0;
 748         efi_memory_desc_t *md, *next_md;
 749         void *p, *p1;
 750         int i, j;
 751
 752         j = 0;
 753         p1 = memmap.map;
 754         for (p = p1, i = 0; p < memmap.map_end; p += memmap.desc_size, i++) {
 755                 md = p;
 756                 next_md = p1;
 757                 current_addr = md->phys_addr +
 758                         PFN_PHYS(md->num_pages);
 759                 if (is_available_memory(md)) {
 760                         if (md->phys_addr >= size) continue;
 761                         memcpy(next_md, md, memmap.desc_size);
 762                         if (current_addr >= size) {
 763                                 next_md->num_pages -=
 764                                         PFN_UP(current_addr-size);
 765                         }
 766                         p1 += memmap.desc_size;
 767                         next_md = p1;
 768                         j++;
 769                 } else if ((md->attribute & EFI_MEMORY_RUNTIME) ==
 770                            EFI_MEMORY_RUNTIME) {
 771                         /* In order to make runtime services
 772                          * available we have to include runtime
 773                          * memory regions in memory map */
 774                         memcpy(next_md, md, memmap.desc_size);
 775                         p1 += memmap.desc_size;
 776                         next_md = p1;
 777                         j++;
 778                 }
 779         }
 780         memmap.nr_map = j;
 781         memmap.map_end = memmap.map +
 782                 (memmap.nr_map * memmap.desc_size);
 783 }
 784
 785 void __init limit_regions(unsigned long long size)
 786 {
 787         unsigned long long current_addr;
 788         int i;
 789
 790         print_memory_map("limit_regions start");
 791         if (efi_enabled) {
 792                 efi_limit_regions(size);
 793                 return;
 794         }
 795         for (i = 0; i < e820.nr_map; i++) {
 796                 current_addr = e820.map[i].addr + e820.map[i].size;
 797                 if (current_addr < size)
 798                         continue;
 799
 800                 if (e820.map[i].type != E820_RAM)
 801                         continue;
 802
 803                 if (e820.map[i].addr >= size) {
 804                         /*
 805                          * This region starts past the end of the
 806                          * requested size, skip it completely.
 807                          */
 808                         e820.nr_map = i;
 809                 } else {
 810                         e820.nr_map = i + 1;
 811                         e820.map[i].size -= current_addr - size;
 812                 }
 813                 print_memory_map("limit_regions endfor");
 814                 return;
 815         }
 816         print_memory_map("limit_regions endfunc");
 817 }
 818
 819  /*
 820   * This function checks if the entire range <start,end> is mapped with type.
 821   *
 822   * Note: this function only works correct if the e820 table is sorted and
 823   * not-overlapping, which is the case
 824   */
 825 int __init
 826 e820_all_mapped(unsigned long s, unsigned long e, unsigned type)
 827 {
 828         u64 start = s;
 829         u64 end = e;
 830         int i;
 831         for (i = 0; i < e820.nr_map; i++) {
 832                 struct e820entry *ei = &e820.map[i];
 833                 if (type && ei->type != type)
 834                         continue;
 835                 /* is the region (part) in overlap with the current region ?*/
 836                 if (ei->addr >= end || ei->addr + ei->size <= start)
 837                         continue;
 838                 /* if the region is at the beginning of <start,end> we move
 839                  * start to the end of the region since it's ok until there
 840                  */
 841                 if (ei->addr <= start)
 842                         start = ei->addr + ei->size;
 843                 /* if start is now at or beyond end, we're done, full
 844                  * coverage */
 845                 if (start >= end)
 846                         return 1; /* we're done */
 847         }
 848         return 0;
 849 }
 850
 851 static int __init parse_memmap(char *arg)
 852 {
 853         if (!arg)
 854                 return -EINVAL;
 855
 856         if (strcmp(arg, "exactmap") == 0) {
 857 #ifdef CONFIG_CRASH_DUMP
 858                 /* If we are doing a crash dump, we
 859                  * still need to know the real mem
 860                  * size before original memory map is
 861                  * reset.
 862                  */
 863                 find_max_pfn();
 864                 saved_max_pfn = max_pfn;
 865 #endif
 866                 e820.nr_map = 0;
 867                 user_defined_memmap = 1;
 868         } else {
 869                 /* If the user specifies memory size, we
 870                  * limit the BIOS-provided memory map to
 871                  * that size. exactmap can be used to specify
 872                  * the exact map. mem=number can be used to
 873                  * trim the existing memory map.
 874                  */
 875                 unsigned long long start_at, mem_size;
 876
 877                 mem_size = memparse(arg, &arg);
 878                 if (*arg == '@') {
 879                         start_at = memparse(arg+1, &arg);
 880                         add_memory_region(start_at, mem_size, E820_RAM);
 881                 } else if (*arg == '#') {
 882                         start_at = memparse(arg+1, &arg);
 883                         add_memory_region(start_at, mem_size, E820_ACPI);
 884                 } else if (*arg == '$') {
 885                         start_at = memparse(arg+1, &arg);
 886                         add_memory_region(start_at, mem_size, E820_RESERVED);
 887                 } else {
 888                         limit_regions(mem_size);
 889                         user_defined_memmap = 1;
 890                 }
 891         }
 892         return 0;
 893 }
 894 early_param("memmap", parse_memmap);