1/* 2 * mmconfig-shared.c - Low-level direct PCI config space access via 3 * MMCONFIG - common code between i386 and x86-64. 4 * 5 * This code does: 6 * - known chipset handling 7 * - ACPI decoding and validation 8 * 9 * Per-architecture code takes care of the mappings and accesses 10 * themselves. 11 */ 12 13#include <linux/pci.h> 14#include <linux/init.h> 15#include <linux/acpi.h> 16#include <linux/sfi_acpi.h> 17#include <linux/bitmap.h> 18#include <linux/dmi.h> 19#include <linux/slab.h> 20#include <asm/e820.h> 21#include <asm/pci_x86.h> 22#include <asm/acpi.h> 23 24#define PREFIX "PCI: " 25 26/* Indicate if the mmcfg resources have been placed into the resource table. */ 27static int __initdata pci_mmcfg_resources_inserted; 28 29LIST_HEAD(pci_mmcfg_list); 30 31static __init void pci_mmconfig_remove(struct pci_mmcfg_region *cfg) 32{ 33 if (cfg->res.parent) 34 release_resource(&cfg->res); 35 list_del(&cfg->list); 36 kfree(cfg); 37} 38 39static __init void free_all_mmcfg(void) 40{ 41 struct pci_mmcfg_region *cfg, *tmp; 42 43 pci_mmcfg_arch_free(); 44 list_for_each_entry_safe(cfg, tmp, &pci_mmcfg_list, list) 45 pci_mmconfig_remove(cfg); 46} 47 48static __init void list_add_sorted(struct pci_mmcfg_region *new) 49{ 50 struct pci_mmcfg_region *cfg; 51 52 /* keep list sorted by segment and starting bus number */ 53 list_for_each_entry(cfg, &pci_mmcfg_list, list) { 54 if (cfg->segment > new->segment || 55 (cfg->segment == new->segment && 56 cfg->start_bus >= new->start_bus)) { 57 list_add_tail(&new->list, &cfg->list); 58 return; 59 } 60 } 61 list_add_tail(&new->list, &pci_mmcfg_list); 62} 63 64static __init struct pci_mmcfg_region *pci_mmconfig_add(int segment, int start, 65 int end, u64 addr) 66{ 67 struct pci_mmcfg_region *new; 68 int num_buses; 69 struct resource *res; 70 71 if (addr == 0) 72 return NULL; 73 74 new = kzalloc(sizeof(*new), GFP_KERNEL); 75 if (!new) 76 return NULL; 77 78 new->address = addr; 79 new->segment = segment; 80 new->start_bus = start; 81 new->end_bus = end; 82 83 list_add_sorted(new); 84 85 num_buses = end - start + 1; 86 res = &new->res; 87 res->start = addr + PCI_MMCFG_BUS_OFFSET(start); 88 res->end = addr + PCI_MMCFG_BUS_OFFSET(num_buses) - 1; 89 res->flags = IORESOURCE_MEM | IORESOURCE_BUSY; 90 snprintf(new->name, PCI_MMCFG_RESOURCE_NAME_LEN, 91 "PCI MMCONFIG %04x [bus %02x-%02x]", segment, start, end); 92 res->name = new->name; 93 94 printk(KERN_INFO PREFIX "MMCONFIG for domain %04x [bus %02x-%02x] at " 95 "%pR (base %#lx)\n", segment, start, end, &new->res, 96 (unsigned long) addr); 97 98 return new; 99} 100 101struct pci_mmcfg_region *pci_mmconfig_lookup(int segment, int bus) 102{ 103 struct pci_mmcfg_region *cfg; 104 105 list_for_each_entry(cfg, &pci_mmcfg_list, list) 106 if (cfg->segment == segment && 107 cfg->start_bus <= bus && bus <= cfg->end_bus) 108 return cfg; 109 110 return NULL; 111} 112 113static const char __init *pci_mmcfg_e7520(void) 114{ 115 u32 win; 116 raw_pci_ops->read(0, 0, PCI_DEVFN(0, 0), 0xce, 2, &win); 117 118 win = win & 0xf000; 119 if (win == 0x0000 || win == 0xf000) 120 return NULL; 121 122 if (pci_mmconfig_add(0, 0, 255, win << 16) == NULL) 123 return NULL; 124 125 return "Intel Corporation E7520 Memory Controller Hub"; 126} 127 128static const char __init *pci_mmcfg_intel_945(void) 129{ 130 u32 pciexbar, mask = 0, len = 0; 131 132 raw_pci_ops->read(0, 0, PCI_DEVFN(0, 0), 0x48, 4, &pciexbar); 133 134 /* Enable bit */ 135 if (!(pciexbar & 1)) 136 return NULL; 137 138 /* Size bits */ 139 switch ((pciexbar >> 1) & 3) { 140 case 0: 141 mask = 0xf0000000U; 142 len = 0x10000000U; 143 break; 144 case 1: 145 mask = 0xf8000000U; 146 len = 0x08000000U; 147 break; 148 case 2: 149 mask = 0xfc000000U; 150 len = 0x04000000U; 151 break; 152 default: 153 return NULL; 154 } 155 156 /* Errata #2, things break when not aligned on a 256Mb boundary */ 157 /* Can only happen in 64M/128M mode */ 158 159 if ((pciexbar & mask) & 0x0fffffffU) 160 return NULL; 161 162 /* Don't hit the APIC registers and their friends */ 163 if ((pciexbar & mask) >= 0xf0000000U) 164 return NULL; 165 166 if (pci_mmconfig_add(0, 0, (len >> 20) - 1, pciexbar & mask) == NULL) 167 return NULL; 168 169 return "Intel Corporation 945G/GZ/P/PL Express Memory Controller Hub"; 170} 171 172static const char __init *pci_mmcfg_amd_fam10h(void) 173{ 174 u32 low, high, address; 175 u64 base, msr; 176 int i; 177 unsigned segnbits = 0, busnbits, end_bus; 178 179 if (!(pci_probe & PCI_CHECK_ENABLE_AMD_MMCONF)) 180 return NULL; 181 182 address = MSR_FAM10H_MMIO_CONF_BASE; 183 if (rdmsr_safe(address, &low, &high)) 184 return NULL; 185 186 msr = high; 187 msr <<= 32; 188 msr |= low; 189 190 /* mmconfig is not enable */ 191 if (!(msr & FAM10H_MMIO_CONF_ENABLE)) 192 return NULL; 193 194 base = msr & (FAM10H_MMIO_CONF_BASE_MASK<<FAM10H_MMIO_CONF_BASE_SHIFT); 195 196 busnbits = (msr >> FAM10H_MMIO_CONF_BUSRANGE_SHIFT) & 197 FAM10H_MMIO_CONF_BUSRANGE_MASK; 198 199 /* 200 * only handle bus 0 ? 201 * need to skip it 202 */ 203 if (!busnbits) 204 return NULL; 205 206 if (busnbits > 8) { 207 segnbits = busnbits - 8; 208 busnbits = 8; 209 } 210 211 end_bus = (1 << busnbits) - 1; 212 for (i = 0; i < (1 << segnbits); i++) 213 if (pci_mmconfig_add(i, 0, end_bus, 214 base + (1<<28) * i) == NULL) { 215 free_all_mmcfg(); 216 return NULL; 217 } 218 219 return "AMD Family 10h NB"; 220} 221 222static bool __initdata mcp55_checked; 223static const char __init *pci_mmcfg_nvidia_mcp55(void) 224{ 225 int bus; 226 int mcp55_mmconf_found = 0; 227 228 static const u32 extcfg_regnum = 0x90; 229 static const u32 extcfg_regsize = 4; 230 static const u32 extcfg_enable_mask = 1<<31; 231 static const u32 extcfg_start_mask = 0xff<<16; 232 static const int extcfg_start_shift = 16; 233 static const u32 extcfg_size_mask = 0x3<<28; 234 static const int extcfg_size_shift = 28; 235 static const int extcfg_sizebus[] = {0x100, 0x80, 0x40, 0x20}; 236 static const u32 extcfg_base_mask[] = {0x7ff8, 0x7ffc, 0x7ffe, 0x7fff}; 237 static const int extcfg_base_lshift = 25; 238 239 /* 240 * do check if amd fam10h already took over 241 */ 242 if (!acpi_disabled || !list_empty(&pci_mmcfg_list) || mcp55_checked) 243 return NULL; 244 245 mcp55_checked = true; 246 for (bus = 0; bus < 256; bus++) { 247 u64 base; 248 u32 l, extcfg; 249 u16 vendor, device; 250 int start, size_index, end; 251 252 raw_pci_ops->read(0, bus, PCI_DEVFN(0, 0), 0, 4, &l); 253 vendor = l & 0xffff; 254 device = (l >> 16) & 0xffff; 255 256 if (PCI_VENDOR_ID_NVIDIA != vendor || 0x0369 != device) 257 continue; 258 259 raw_pci_ops->read(0, bus, PCI_DEVFN(0, 0), extcfg_regnum, 260 extcfg_regsize, &extcfg); 261 262 if (!(extcfg & extcfg_enable_mask)) 263 continue; 264 265 size_index = (extcfg & extcfg_size_mask) >> extcfg_size_shift; 266 base = extcfg & extcfg_base_mask[size_index]; 267 /* base could > 4G */ 268 base <<= extcfg_base_lshift; 269 start = (extcfg & extcfg_start_mask) >> extcfg_start_shift; 270 end = start + extcfg_sizebus[size_index] - 1; 271 if (pci_mmconfig_add(0, start, end, base) == NULL) 272 continue; 273 mcp55_mmconf_found++; 274 } 275 276 if (!mcp55_mmconf_found) 277 return NULL; 278 279 return "nVidia MCP55"; 280} 281 282struct pci_mmcfg_hostbridge_probe { 283 u32 bus; 284 u32 devfn; 285 u32 vendor; 286 u32 device; 287 const char *(*probe)(void); 288}; 289 290static struct pci_mmcfg_hostbridge_probe pci_mmcfg_probes[] __initdata = { 291 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_INTEL, 292 PCI_DEVICE_ID_INTEL_E7520_MCH, pci_mmcfg_e7520 }, 293 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_INTEL, 294 PCI_DEVICE_ID_INTEL_82945G_HB, pci_mmcfg_intel_945 }, 295 { 0, PCI_DEVFN(0x18, 0), PCI_VENDOR_ID_AMD, 296 0x1200, pci_mmcfg_amd_fam10h }, 297 { 0xff, PCI_DEVFN(0, 0), PCI_VENDOR_ID_AMD, 298 0x1200, pci_mmcfg_amd_fam10h }, 299 { 0, PCI_DEVFN(0, 0), PCI_VENDOR_ID_NVIDIA, 300 0x0369, pci_mmcfg_nvidia_mcp55 }, 301}; 302 303static void __init pci_mmcfg_check_end_bus_number(void) 304{ 305 struct pci_mmcfg_region *cfg, *cfgx; 306 307 /* Fixup overlaps */ 308 list_for_each_entry(cfg, &pci_mmcfg_list, list) { 309 if (cfg->end_bus < cfg->start_bus) 310 cfg->end_bus = 255; 311 312 /* Don't access the list head ! */ 313 if (cfg->list.next == &pci_mmcfg_list) 314 break; 315 316 cfgx = list_entry(cfg->list.next, typeof(*cfg), list); 317 if (cfg->end_bus >= cfgx->start_bus) 318 cfg->end_bus = cfgx->start_bus - 1; 319 } 320} 321 322static int __init pci_mmcfg_check_hostbridge(void) 323{ 324 u32 l; 325 u32 bus, devfn; 326 u16 vendor, device; 327 int i; 328 const char *name; 329 330 if (!raw_pci_ops) 331 return 0; 332 333 free_all_mmcfg(); 334 335 for (i = 0; i < ARRAY_SIZE(pci_mmcfg_probes); i++) { 336 bus = pci_mmcfg_probes[i].bus; 337 devfn = pci_mmcfg_probes[i].devfn; 338 raw_pci_ops->read(0, bus, devfn, 0, 4, &l); 339 vendor = l & 0xffff; 340 device = (l >> 16) & 0xffff; 341 342 name = NULL; 343 if (pci_mmcfg_probes[i].vendor == vendor && 344 pci_mmcfg_probes[i].device == device) 345 name = pci_mmcfg_probes[i].probe(); 346 347 if (name) 348 printk(KERN_INFO PREFIX "%s with MMCONFIG support\n", 349 name); 350 } 351 352 /* some end_bus_number is crazy, fix it */ 353 pci_mmcfg_check_end_bus_number(); 354 355 return !list_empty(&pci_mmcfg_list); 356} 357 358static void __init pci_mmcfg_insert_resources(void) 359{ 360 struct pci_mmcfg_region *cfg; 361 362 list_for_each_entry(cfg, &pci_mmcfg_list, list) 363 insert_resource(&iomem_resource, &cfg->res); 364 365 /* Mark that the resources have been inserted. */ 366 pci_mmcfg_resources_inserted = 1; 367} 368 369static acpi_status __init check_mcfg_resource(struct acpi_resource *res, 370 void *data) 371{ 372 struct resource *mcfg_res = data; 373 struct acpi_resource_address64 address; 374 acpi_status status; 375 376 if (res->type == ACPI_RESOURCE_TYPE_FIXED_MEMORY32) { 377 struct acpi_resource_fixed_memory32 *fixmem32 = 378 &res->data.fixed_memory32; 379 if (!fixmem32) 380 return AE_OK; 381 if ((mcfg_res->start >= fixmem32->address) && 382 (mcfg_res->end < (fixmem32->address + 383 fixmem32->address_length))) { 384 mcfg_res->flags = 1; 385 return AE_CTRL_TERMINATE; 386 } 387 } 388 if ((res->type != ACPI_RESOURCE_TYPE_ADDRESS32) && 389 (res->type != ACPI_RESOURCE_TYPE_ADDRESS64)) 390 return AE_OK; 391 392 status = acpi_resource_to_address64(res, &address); 393 if (ACPI_FAILURE(status) || 394 (address.address_length <= 0) || 395 (address.resource_type != ACPI_MEMORY_RANGE)) 396 return AE_OK; 397 398 if ((mcfg_res->start >= address.minimum) && 399 (mcfg_res->end < (address.minimum + address.address_length))) { 400 mcfg_res->flags = 1; 401 return AE_CTRL_TERMINATE; 402 } 403 return AE_OK; 404} 405 406static acpi_status __init find_mboard_resource(acpi_handle handle, u32 lvl, 407 void *context, void **rv) 408{ 409 struct resource *mcfg_res = context; 410 411 acpi_walk_resources(handle, METHOD_NAME__CRS, 412 check_mcfg_resource, context); 413 414 if (mcfg_res->flags) 415 return AE_CTRL_TERMINATE; 416 417 return AE_OK; 418} 419 420static int __init is_acpi_reserved(u64 start, u64 end, unsigned not_used) 421{ 422 struct resource mcfg_res; 423 424 mcfg_res.start = start; 425 mcfg_res.end = end - 1; 426 mcfg_res.flags = 0; 427 428 acpi_get_devices("PNP0C01", find_mboard_resource, &mcfg_res, NULL); 429 430 if (!mcfg_res.flags) 431 acpi_get_devices("PNP0C02", find_mboard_resource, &mcfg_res, 432 NULL); 433 434 return mcfg_res.flags; 435} 436 437typedef int (*check_reserved_t)(u64 start, u64 end, unsigned type); 438 439static int __init is_mmconf_reserved(check_reserved_t is_reserved, 440 struct pci_mmcfg_region *cfg, int with_e820) 441{ 442 u64 addr = cfg->res.start; 443 u64 size = resource_size(&cfg->res); 444 u64 old_size = size; 445 int valid = 0, num_buses; 446 447 while (!is_reserved(addr, addr + size, E820_RESERVED)) { 448 size >>= 1; 449 if (size < (16UL<<20)) 450 break; 451 } 452 453 if (size >= (16UL<<20) || size == old_size) { 454 printk(KERN_INFO PREFIX "MMCONFIG at %pR reserved in %s\n", 455 &cfg->res, 456 with_e820 ? "E820" : "ACPI motherboard resources"); 457 valid = 1; 458 459 if (old_size != size) { 460 /* update end_bus */ 461 cfg->end_bus = cfg->start_bus + ((size>>20) - 1); 462 num_buses = cfg->end_bus - cfg->start_bus + 1; 463 cfg->res.end = cfg->res.start + 464 PCI_MMCFG_BUS_OFFSET(num_buses) - 1; 465 snprintf(cfg->name, PCI_MMCFG_RESOURCE_NAME_LEN, 466 "PCI MMCONFIG %04x [bus %02x-%02x]", 467 cfg->segment, cfg->start_bus, cfg->end_bus); 468 printk(KERN_INFO PREFIX 469 "MMCONFIG for %04x [bus%02x-%02x] " 470 "at %pR (base %#lx) (size reduced!)\n", 471 cfg->segment, cfg->start_bus, cfg->end_bus, 472 &cfg->res, (unsigned long) cfg->address); 473 } 474 } 475 476 return valid; 477} 478 479static void __init pci_mmcfg_reject_broken(int early) 480{ 481 struct pci_mmcfg_region *cfg; 482 483 list_for_each_entry(cfg, &pci_mmcfg_list, list) { 484 int valid = 0; 485 486 if (!early && !acpi_disabled) { 487 valid = is_mmconf_reserved(is_acpi_reserved, cfg, 0); 488 489 if (valid) 490 continue; 491 else 492 printk(KERN_ERR FW_BUG PREFIX 493 "MMCONFIG at %pR not reserved in " 494 "ACPI motherboard resources\n", 495 &cfg->res); 496 } 497 498 /* Don't try to do this check unless configuration 499 type 1 is available. how about type 2 ?*/ 500 if (raw_pci_ops) 501 valid = is_mmconf_reserved(e820_all_mapped, cfg, 1); 502 503 if (!valid) 504 goto reject; 505 } 506 507 return; 508 509reject: 510 printk(KERN_INFO PREFIX "not using MMCONFIG\n"); 511 free_all_mmcfg(); 512} 513 514static int __initdata known_bridge; 515 516static int __init acpi_mcfg_check_entry(struct acpi_table_mcfg *mcfg, 517 struct acpi_mcfg_allocation *cfg) 518{ 519 int year; 520 521 if (cfg->address < 0xFFFFFFFF) 522 return 0; 523 524 if (!strcmp(mcfg->header.oem_id, "SGI")) 525 return 0; 526 527 if (mcfg->header.revision >= 1) { 528 if (dmi_get_date(DMI_BIOS_DATE, &year, NULL, NULL) && 529 year >= 2010) 530 return 0; 531 } 532 533 printk(KERN_ERR PREFIX "MCFG region for %04x [bus %02x-%02x] at %#llx " 534 "is above 4GB, ignored\n", cfg->pci_segment, 535 cfg->start_bus_number, cfg->end_bus_number, cfg->address); 536 return -EINVAL; 537} 538 539static int __init pci_parse_mcfg(struct acpi_table_header *header) 540{ 541 struct acpi_table_mcfg *mcfg; 542 struct acpi_mcfg_allocation *cfg_table, *cfg; 543 unsigned long i; 544 int entries; 545 546 if (!header) 547 return -EINVAL; 548 549 mcfg = (struct acpi_table_mcfg *)header; 550 551 /* how many config structures do we have */ 552 free_all_mmcfg(); 553 entries = 0; 554 i = header->length - sizeof(struct acpi_table_mcfg); 555 while (i >= sizeof(struct acpi_mcfg_allocation)) { 556 entries++; 557 i -= sizeof(struct acpi_mcfg_allocation); 558 }; 559 if (entries == 0) { 560 printk(KERN_ERR PREFIX "MMCONFIG has no entries\n"); 561 return -ENODEV; 562 } 563 564 cfg_table = (struct acpi_mcfg_allocation *) &mcfg[1]; 565 for (i = 0; i < entries; i++) { 566 cfg = &cfg_table[i]; 567 if (acpi_mcfg_check_entry(mcfg, cfg)) { 568 free_all_mmcfg(); 569 return -ENODEV; 570 } 571 572 if (pci_mmconfig_add(cfg->pci_segment, cfg->start_bus_number, 573 cfg->end_bus_number, cfg->address) == NULL) { 574 printk(KERN_WARNING PREFIX 575 "no memory for MCFG entries\n"); 576 free_all_mmcfg(); 577 return -ENOMEM; 578 } 579 } 580 581 return 0; 582} 583 584static void __init __pci_mmcfg_init(int early) 585{ 586 /* MMCONFIG disabled */ 587 if ((pci_probe & PCI_PROBE_MMCONF) == 0) 588 return; 589 590 /* MMCONFIG already enabled */ 591 if (!early && !(pci_probe & PCI_PROBE_MASK & ~PCI_PROBE_MMCONF)) 592 return; 593 594 /* for late to exit */ 595 if (known_bridge) 596 return; 597 598 if (early) { 599 if (pci_mmcfg_check_hostbridge()) 600 known_bridge = 1; 601 } 602 603 if (!known_bridge) 604 acpi_sfi_table_parse(ACPI_SIG_MCFG, pci_parse_mcfg); 605 606 pci_mmcfg_reject_broken(early); 607 608 if (list_empty(&pci_mmcfg_list)) 609 return; 610 611 if (pci_mmcfg_arch_init()) 612 pci_probe = (pci_probe & ~PCI_PROBE_MASK) | PCI_PROBE_MMCONF; 613 else { 614 /* 615 * Signal not to attempt to insert mmcfg resources because 616 * the architecture mmcfg setup could not initialize. 617 */ 618 pci_mmcfg_resources_inserted = 1; 619 } 620} 621 622void __init pci_mmcfg_early_init(void) 623{ 624 __pci_mmcfg_init(1); 625} 626 627void __init pci_mmcfg_late_init(void) 628{ 629 __pci_mmcfg_init(0); 630} 631 632static int __init pci_mmcfg_late_insert_resources(void) 633{ 634 /* 635 * If resources are already inserted or we are not using MMCONFIG, 636 * don't insert the resources. 637 */ 638 if ((pci_mmcfg_resources_inserted == 1) || 639 (pci_probe & PCI_PROBE_MMCONF) == 0 || 640 list_empty(&pci_mmcfg_list)) 641 return 1; 642 643 /* 644 * Attempt to insert the mmcfg resources but not with the busy flag 645 * marked so it won't cause request errors when __request_region is 646 * called. 647 */ 648 pci_mmcfg_insert_resources(); 649 650 return 0; 651} 652 653/* 654 * Perform MMCONFIG resource insertion after PCI initialization to allow for 655 * misprogrammed MCFG tables that state larger sizes but actually conflict 656 * with other system resources. 657 */ 658late_initcall(pci_mmcfg_late_insert_resources); 659