1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright (C) 2007-2010 Advanced Micro Devices, Inc. 4 * Author: Joerg Roedel <jroedel@suse.de> 5 * Leo Duran <leo.duran@amd.com> 6 */ 7 8 #define pr_fmt(fmt) "AMD-Vi: " fmt 9 #define dev_fmt(fmt) pr_fmt(fmt) 10 11 #include <linux/pci.h> 12 #include <linux/acpi.h> 13 #include <linux/list.h> 14 #include <linux/bitmap.h> 15 #include <linux/slab.h> 16 #include <linux/syscore_ops.h> 17 #include <linux/interrupt.h> 18 #include <linux/msi.h> 19 #include <linux/irq.h> 20 #include <linux/amd-iommu.h> 21 #include <linux/export.h> 22 #include <linux/kmemleak.h> 23 #include <linux/cc_platform.h> 24 #include <linux/iopoll.h> 25 #include <asm/pci-direct.h> 26 #include <asm/iommu.h> 27 #include <asm/apic.h> 28 #include <asm/gart.h> 29 #include <asm/x86_init.h> 30 #include <asm/io_apic.h> 31 #include <asm/irq_remapping.h> 32 #include <asm/set_memory.h> 33 34 #include <linux/crash_dump.h> 35 36 #include "amd_iommu.h" 37 #include "../irq_remapping.h" 38 39 /* 40 * definitions for the ACPI scanning code 41 */ 42 #define IVRS_HEADER_LENGTH 48 43 44 #define ACPI_IVHD_TYPE_MAX_SUPPORTED 0x40 45 #define ACPI_IVMD_TYPE_ALL 0x20 46 #define ACPI_IVMD_TYPE 0x21 47 #define ACPI_IVMD_TYPE_RANGE 0x22 48 49 #define IVHD_DEV_ALL 0x01 50 #define IVHD_DEV_SELECT 0x02 51 #define IVHD_DEV_SELECT_RANGE_START 0x03 52 #define IVHD_DEV_RANGE_END 0x04 53 #define IVHD_DEV_ALIAS 0x42 54 #define IVHD_DEV_ALIAS_RANGE 0x43 55 #define IVHD_DEV_EXT_SELECT 0x46 56 #define IVHD_DEV_EXT_SELECT_RANGE 0x47 57 #define IVHD_DEV_SPECIAL 0x48 58 #define IVHD_DEV_ACPI_HID 0xf0 59 60 #define UID_NOT_PRESENT 0 61 #define UID_IS_INTEGER 1 62 #define UID_IS_CHARACTER 2 63 64 #define IVHD_SPECIAL_IOAPIC 1 65 #define IVHD_SPECIAL_HPET 2 66 67 #define IVHD_FLAG_HT_TUN_EN_MASK 0x01 68 #define IVHD_FLAG_PASSPW_EN_MASK 0x02 69 #define IVHD_FLAG_RESPASSPW_EN_MASK 0x04 70 #define IVHD_FLAG_ISOC_EN_MASK 0x08 71 72 #define IVMD_FLAG_EXCL_RANGE 0x08 73 #define IVMD_FLAG_IW 0x04 74 #define IVMD_FLAG_IR 0x02 75 #define IVMD_FLAG_UNITY_MAP 0x01 76 77 #define ACPI_DEVFLAG_INITPASS 0x01 78 #define ACPI_DEVFLAG_EXTINT 0x02 79 #define ACPI_DEVFLAG_NMI 0x04 80 #define ACPI_DEVFLAG_SYSMGT1 0x10 81 #define ACPI_DEVFLAG_SYSMGT2 0x20 82 #define ACPI_DEVFLAG_LINT0 0x40 83 #define ACPI_DEVFLAG_LINT1 0x80 84 #define ACPI_DEVFLAG_ATSDIS 0x10000000 85 86 #define LOOP_TIMEOUT 2000000 87 88 #define IVRS_GET_SBDF_ID(seg, bus, dev, fn) (((seg & 0xffff) << 16) | ((bus & 0xff) << 8) \ 89 | ((dev & 0x1f) << 3) | (fn & 0x7)) 90 91 /* 92 * ACPI table definitions 93 * 94 * These data structures are laid over the table to parse the important values 95 * out of it. 96 */ 97 98 /* 99 * structure describing one IOMMU in the ACPI table. Typically followed by one 100 * or more ivhd_entrys. 101 */ 102 struct ivhd_header { 103 u8 type; 104 u8 flags; 105 u16 length; 106 u16 devid; 107 u16 cap_ptr; 108 u64 mmio_phys; 109 u16 pci_seg; 110 u16 info; 111 u32 efr_attr; 112 113 /* Following only valid on IVHD type 11h and 40h */ 114 u64 efr_reg; /* Exact copy of MMIO_EXT_FEATURES */ 115 u64 efr_reg2; 116 } __attribute__((packed)); 117 118 /* 119 * A device entry describing which devices a specific IOMMU translates and 120 * which requestor ids they use. 121 */ 122 struct ivhd_entry { 123 u8 type; 124 u16 devid; 125 u8 flags; 126 struct_group(ext_hid, 127 u32 ext; 128 u32 hidh; 129 ); 130 u64 cid; 131 u8 uidf; 132 u8 uidl; 133 u8 uid; 134 } __attribute__((packed)); 135 136 /* 137 * An AMD IOMMU memory definition structure. It defines things like exclusion 138 * ranges for devices and regions that should be unity mapped. 139 */ 140 struct ivmd_header { 141 u8 type; 142 u8 flags; 143 u16 length; 144 u16 devid; 145 u16 aux; 146 u16 pci_seg; 147 u8 resv[6]; 148 u64 range_start; 149 u64 range_length; 150 } __attribute__((packed)); 151 152 bool amd_iommu_dump; 153 bool amd_iommu_irq_remap __read_mostly; 154 155 enum io_pgtable_fmt amd_iommu_pgtable = AMD_IOMMU_V1; 156 /* Guest page table level */ 157 int amd_iommu_gpt_level = PAGE_MODE_4_LEVEL; 158 159 int amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_VAPIC; 160 static int amd_iommu_xt_mode = IRQ_REMAP_XAPIC_MODE; 161 162 static bool amd_iommu_detected; 163 static bool amd_iommu_disabled __initdata; 164 static bool amd_iommu_force_enable __initdata; 165 static int amd_iommu_target_ivhd_type; 166 167 /* Global EFR and EFR2 registers */ 168 u64 amd_iommu_efr; 169 u64 amd_iommu_efr2; 170 171 /* SNP is enabled on the system? */ 172 bool amd_iommu_snp_en; 173 EXPORT_SYMBOL(amd_iommu_snp_en); 174 175 LIST_HEAD(amd_iommu_pci_seg_list); /* list of all PCI segments */ 176 LIST_HEAD(amd_iommu_list); /* list of all AMD IOMMUs in the 177 system */ 178 179 /* Array to assign indices to IOMMUs*/ 180 struct amd_iommu *amd_iommus[MAX_IOMMUS]; 181 182 /* Number of IOMMUs present in the system */ 183 static int amd_iommus_present; 184 185 /* IOMMUs have a non-present cache? */ 186 bool amd_iommu_np_cache __read_mostly; 187 bool amd_iommu_iotlb_sup __read_mostly = true; 188 189 u32 amd_iommu_max_pasid __read_mostly = ~0; 190 191 bool amd_iommu_v2_present __read_mostly; 192 static bool amd_iommu_pc_present __read_mostly; 193 bool amdr_ivrs_remap_support __read_mostly; 194 195 bool amd_iommu_force_isolation __read_mostly; 196 197 /* 198 * AMD IOMMU allows up to 2^16 different protection domains. This is a bitmap 199 * to know which ones are already in use. 200 */ 201 unsigned long *amd_iommu_pd_alloc_bitmap; 202 203 enum iommu_init_state { 204 IOMMU_START_STATE, 205 IOMMU_IVRS_DETECTED, 206 IOMMU_ACPI_FINISHED, 207 IOMMU_ENABLED, 208 IOMMU_PCI_INIT, 209 IOMMU_INTERRUPTS_EN, 210 IOMMU_INITIALIZED, 211 IOMMU_NOT_FOUND, 212 IOMMU_INIT_ERROR, 213 IOMMU_CMDLINE_DISABLED, 214 }; 215 216 /* Early ioapic and hpet maps from kernel command line */ 217 #define EARLY_MAP_SIZE 4 218 static struct devid_map __initdata early_ioapic_map[EARLY_MAP_SIZE]; 219 static struct devid_map __initdata early_hpet_map[EARLY_MAP_SIZE]; 220 static struct acpihid_map_entry __initdata early_acpihid_map[EARLY_MAP_SIZE]; 221 222 static int __initdata early_ioapic_map_size; 223 static int __initdata early_hpet_map_size; 224 static int __initdata early_acpihid_map_size; 225 226 static bool __initdata cmdline_maps; 227 228 static enum iommu_init_state init_state = IOMMU_START_STATE; 229 230 static int amd_iommu_enable_interrupts(void); 231 static int __init iommu_go_to_state(enum iommu_init_state state); 232 static void init_device_table_dma(struct amd_iommu_pci_seg *pci_seg); 233 234 static bool amd_iommu_pre_enabled = true; 235 236 static u32 amd_iommu_ivinfo __initdata; 237 238 bool translation_pre_enabled(struct amd_iommu *iommu) 239 { 240 return (iommu->flags & AMD_IOMMU_FLAG_TRANS_PRE_ENABLED); 241 } 242 243 static void clear_translation_pre_enabled(struct amd_iommu *iommu) 244 { 245 iommu->flags &= ~AMD_IOMMU_FLAG_TRANS_PRE_ENABLED; 246 } 247 248 static void init_translation_status(struct amd_iommu *iommu) 249 { 250 u64 ctrl; 251 252 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET); 253 if (ctrl & (1<<CONTROL_IOMMU_EN)) 254 iommu->flags |= AMD_IOMMU_FLAG_TRANS_PRE_ENABLED; 255 } 256 257 static inline unsigned long tbl_size(int entry_size, int last_bdf) 258 { 259 unsigned shift = PAGE_SHIFT + 260 get_order((last_bdf + 1) * entry_size); 261 262 return 1UL << shift; 263 } 264 265 int amd_iommu_get_num_iommus(void) 266 { 267 return amd_iommus_present; 268 } 269 270 /* 271 * Iterate through all the IOMMUs to get common EFR 272 * masks among all IOMMUs and warn if found inconsistency. 273 */ 274 static void get_global_efr(void) 275 { 276 struct amd_iommu *iommu; 277 278 for_each_iommu(iommu) { 279 u64 tmp = iommu->features; 280 u64 tmp2 = iommu->features2; 281 282 if (list_is_first(&iommu->list, &amd_iommu_list)) { 283 amd_iommu_efr = tmp; 284 amd_iommu_efr2 = tmp2; 285 continue; 286 } 287 288 if (amd_iommu_efr == tmp && 289 amd_iommu_efr2 == tmp2) 290 continue; 291 292 pr_err(FW_BUG 293 "Found inconsistent EFR/EFR2 %#llx,%#llx (global %#llx,%#llx) on iommu%d (%04x:%02x:%02x.%01x).\n", 294 tmp, tmp2, amd_iommu_efr, amd_iommu_efr2, 295 iommu->index, iommu->pci_seg->id, 296 PCI_BUS_NUM(iommu->devid), PCI_SLOT(iommu->devid), 297 PCI_FUNC(iommu->devid)); 298 299 amd_iommu_efr &= tmp; 300 amd_iommu_efr2 &= tmp2; 301 } 302 303 pr_info("Using global IVHD EFR:%#llx, EFR2:%#llx\n", amd_iommu_efr, amd_iommu_efr2); 304 } 305 306 static bool check_feature_on_all_iommus(u64 mask) 307 { 308 return !!(amd_iommu_efr & mask); 309 } 310 311 static inline int check_feature_gpt_level(void) 312 { 313 return ((amd_iommu_efr >> FEATURE_GATS_SHIFT) & FEATURE_GATS_MASK); 314 } 315 316 /* 317 * For IVHD type 0x11/0x40, EFR is also available via IVHD. 318 * Default to IVHD EFR since it is available sooner 319 * (i.e. before PCI init). 320 */ 321 static void __init early_iommu_features_init(struct amd_iommu *iommu, 322 struct ivhd_header *h) 323 { 324 if (amd_iommu_ivinfo & IOMMU_IVINFO_EFRSUP) { 325 iommu->features = h->efr_reg; 326 iommu->features2 = h->efr_reg2; 327 } 328 if (amd_iommu_ivinfo & IOMMU_IVINFO_DMA_REMAP) 329 amdr_ivrs_remap_support = true; 330 } 331 332 /* Access to l1 and l2 indexed register spaces */ 333 334 static u32 iommu_read_l1(struct amd_iommu *iommu, u16 l1, u8 address) 335 { 336 u32 val; 337 338 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16)); 339 pci_read_config_dword(iommu->dev, 0xfc, &val); 340 return val; 341 } 342 343 static void iommu_write_l1(struct amd_iommu *iommu, u16 l1, u8 address, u32 val) 344 { 345 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16 | 1 << 31)); 346 pci_write_config_dword(iommu->dev, 0xfc, val); 347 pci_write_config_dword(iommu->dev, 0xf8, (address | l1 << 16)); 348 } 349 350 static u32 iommu_read_l2(struct amd_iommu *iommu, u8 address) 351 { 352 u32 val; 353 354 pci_write_config_dword(iommu->dev, 0xf0, address); 355 pci_read_config_dword(iommu->dev, 0xf4, &val); 356 return val; 357 } 358 359 static void iommu_write_l2(struct amd_iommu *iommu, u8 address, u32 val) 360 { 361 pci_write_config_dword(iommu->dev, 0xf0, (address | 1 << 8)); 362 pci_write_config_dword(iommu->dev, 0xf4, val); 363 } 364 365 /**************************************************************************** 366 * 367 * AMD IOMMU MMIO register space handling functions 368 * 369 * These functions are used to program the IOMMU device registers in 370 * MMIO space required for that driver. 371 * 372 ****************************************************************************/ 373 374 /* 375 * This function set the exclusion range in the IOMMU. DMA accesses to the 376 * exclusion range are passed through untranslated 377 */ 378 static void iommu_set_exclusion_range(struct amd_iommu *iommu) 379 { 380 u64 start = iommu->exclusion_start & PAGE_MASK; 381 u64 limit = (start + iommu->exclusion_length - 1) & PAGE_MASK; 382 u64 entry; 383 384 if (!iommu->exclusion_start) 385 return; 386 387 entry = start | MMIO_EXCL_ENABLE_MASK; 388 memcpy_toio(iommu->mmio_base + MMIO_EXCL_BASE_OFFSET, 389 &entry, sizeof(entry)); 390 391 entry = limit; 392 memcpy_toio(iommu->mmio_base + MMIO_EXCL_LIMIT_OFFSET, 393 &entry, sizeof(entry)); 394 } 395 396 static void iommu_set_cwwb_range(struct amd_iommu *iommu) 397 { 398 u64 start = iommu_virt_to_phys((void *)iommu->cmd_sem); 399 u64 entry = start & PM_ADDR_MASK; 400 401 if (!check_feature_on_all_iommus(FEATURE_SNP)) 402 return; 403 404 /* Note: 405 * Re-purpose Exclusion base/limit registers for Completion wait 406 * write-back base/limit. 407 */ 408 memcpy_toio(iommu->mmio_base + MMIO_EXCL_BASE_OFFSET, 409 &entry, sizeof(entry)); 410 411 /* Note: 412 * Default to 4 Kbytes, which can be specified by setting base 413 * address equal to the limit address. 414 */ 415 memcpy_toio(iommu->mmio_base + MMIO_EXCL_LIMIT_OFFSET, 416 &entry, sizeof(entry)); 417 } 418 419 /* Programs the physical address of the device table into the IOMMU hardware */ 420 static void iommu_set_device_table(struct amd_iommu *iommu) 421 { 422 u64 entry; 423 u32 dev_table_size = iommu->pci_seg->dev_table_size; 424 void *dev_table = (void *)get_dev_table(iommu); 425 426 BUG_ON(iommu->mmio_base == NULL); 427 428 entry = iommu_virt_to_phys(dev_table); 429 entry |= (dev_table_size >> 12) - 1; 430 memcpy_toio(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET, 431 &entry, sizeof(entry)); 432 } 433 434 /* Generic functions to enable/disable certain features of the IOMMU. */ 435 static void iommu_feature_enable(struct amd_iommu *iommu, u8 bit) 436 { 437 u64 ctrl; 438 439 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET); 440 ctrl |= (1ULL << bit); 441 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET); 442 } 443 444 static void iommu_feature_disable(struct amd_iommu *iommu, u8 bit) 445 { 446 u64 ctrl; 447 448 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET); 449 ctrl &= ~(1ULL << bit); 450 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET); 451 } 452 453 static void iommu_set_inv_tlb_timeout(struct amd_iommu *iommu, int timeout) 454 { 455 u64 ctrl; 456 457 ctrl = readq(iommu->mmio_base + MMIO_CONTROL_OFFSET); 458 ctrl &= ~CTRL_INV_TO_MASK; 459 ctrl |= (timeout << CONTROL_INV_TIMEOUT) & CTRL_INV_TO_MASK; 460 writeq(ctrl, iommu->mmio_base + MMIO_CONTROL_OFFSET); 461 } 462 463 /* Function to enable the hardware */ 464 static void iommu_enable(struct amd_iommu *iommu) 465 { 466 iommu_feature_enable(iommu, CONTROL_IOMMU_EN); 467 } 468 469 static void iommu_disable(struct amd_iommu *iommu) 470 { 471 if (!iommu->mmio_base) 472 return; 473 474 /* Disable command buffer */ 475 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN); 476 477 /* Disable event logging and event interrupts */ 478 iommu_feature_disable(iommu, CONTROL_EVT_INT_EN); 479 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN); 480 481 /* Disable IOMMU GA_LOG */ 482 iommu_feature_disable(iommu, CONTROL_GALOG_EN); 483 iommu_feature_disable(iommu, CONTROL_GAINT_EN); 484 485 /* Disable IOMMU hardware itself */ 486 iommu_feature_disable(iommu, CONTROL_IOMMU_EN); 487 } 488 489 /* 490 * mapping and unmapping functions for the IOMMU MMIO space. Each AMD IOMMU in 491 * the system has one. 492 */ 493 static u8 __iomem * __init iommu_map_mmio_space(u64 address, u64 end) 494 { 495 if (!request_mem_region(address, end, "amd_iommu")) { 496 pr_err("Can not reserve memory region %llx-%llx for mmio\n", 497 address, end); 498 pr_err("This is a BIOS bug. Please contact your hardware vendor\n"); 499 return NULL; 500 } 501 502 return (u8 __iomem *)ioremap(address, end); 503 } 504 505 static void __init iommu_unmap_mmio_space(struct amd_iommu *iommu) 506 { 507 if (iommu->mmio_base) 508 iounmap(iommu->mmio_base); 509 release_mem_region(iommu->mmio_phys, iommu->mmio_phys_end); 510 } 511 512 static inline u32 get_ivhd_header_size(struct ivhd_header *h) 513 { 514 u32 size = 0; 515 516 switch (h->type) { 517 case 0x10: 518 size = 24; 519 break; 520 case 0x11: 521 case 0x40: 522 size = 40; 523 break; 524 } 525 return size; 526 } 527 528 /**************************************************************************** 529 * 530 * The functions below belong to the first pass of AMD IOMMU ACPI table 531 * parsing. In this pass we try to find out the highest device id this 532 * code has to handle. Upon this information the size of the shared data 533 * structures is determined later. 534 * 535 ****************************************************************************/ 536 537 /* 538 * This function calculates the length of a given IVHD entry 539 */ 540 static inline int ivhd_entry_length(u8 *ivhd) 541 { 542 u32 type = ((struct ivhd_entry *)ivhd)->type; 543 544 if (type < 0x80) { 545 return 0x04 << (*ivhd >> 6); 546 } else if (type == IVHD_DEV_ACPI_HID) { 547 /* For ACPI_HID, offset 21 is uid len */ 548 return *((u8 *)ivhd + 21) + 22; 549 } 550 return 0; 551 } 552 553 /* 554 * After reading the highest device id from the IOMMU PCI capability header 555 * this function looks if there is a higher device id defined in the ACPI table 556 */ 557 static int __init find_last_devid_from_ivhd(struct ivhd_header *h) 558 { 559 u8 *p = (void *)h, *end = (void *)h; 560 struct ivhd_entry *dev; 561 int last_devid = -EINVAL; 562 563 u32 ivhd_size = get_ivhd_header_size(h); 564 565 if (!ivhd_size) { 566 pr_err("Unsupported IVHD type %#x\n", h->type); 567 return -EINVAL; 568 } 569 570 p += ivhd_size; 571 end += h->length; 572 573 while (p < end) { 574 dev = (struct ivhd_entry *)p; 575 switch (dev->type) { 576 case IVHD_DEV_ALL: 577 /* Use maximum BDF value for DEV_ALL */ 578 return 0xffff; 579 case IVHD_DEV_SELECT: 580 case IVHD_DEV_RANGE_END: 581 case IVHD_DEV_ALIAS: 582 case IVHD_DEV_EXT_SELECT: 583 /* all the above subfield types refer to device ids */ 584 if (dev->devid > last_devid) 585 last_devid = dev->devid; 586 break; 587 default: 588 break; 589 } 590 p += ivhd_entry_length(p); 591 } 592 593 WARN_ON(p != end); 594 595 return last_devid; 596 } 597 598 static int __init check_ivrs_checksum(struct acpi_table_header *table) 599 { 600 int i; 601 u8 checksum = 0, *p = (u8 *)table; 602 603 for (i = 0; i < table->length; ++i) 604 checksum += p[i]; 605 if (checksum != 0) { 606 /* ACPI table corrupt */ 607 pr_err(FW_BUG "IVRS invalid checksum\n"); 608 return -ENODEV; 609 } 610 611 return 0; 612 } 613 614 /* 615 * Iterate over all IVHD entries in the ACPI table and find the highest device 616 * id which we need to handle. This is the first of three functions which parse 617 * the ACPI table. So we check the checksum here. 618 */ 619 static int __init find_last_devid_acpi(struct acpi_table_header *table, u16 pci_seg) 620 { 621 u8 *p = (u8 *)table, *end = (u8 *)table; 622 struct ivhd_header *h; 623 int last_devid, last_bdf = 0; 624 625 p += IVRS_HEADER_LENGTH; 626 627 end += table->length; 628 while (p < end) { 629 h = (struct ivhd_header *)p; 630 if (h->pci_seg == pci_seg && 631 h->type == amd_iommu_target_ivhd_type) { 632 last_devid = find_last_devid_from_ivhd(h); 633 634 if (last_devid < 0) 635 return -EINVAL; 636 if (last_devid > last_bdf) 637 last_bdf = last_devid; 638 } 639 p += h->length; 640 } 641 WARN_ON(p != end); 642 643 return last_bdf; 644 } 645 646 /**************************************************************************** 647 * 648 * The following functions belong to the code path which parses the ACPI table 649 * the second time. In this ACPI parsing iteration we allocate IOMMU specific 650 * data structures, initialize the per PCI segment device/alias/rlookup table 651 * and also basically initialize the hardware. 652 * 653 ****************************************************************************/ 654 655 /* Allocate per PCI segment device table */ 656 static inline int __init alloc_dev_table(struct amd_iommu_pci_seg *pci_seg) 657 { 658 pci_seg->dev_table = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO | GFP_DMA32, 659 get_order(pci_seg->dev_table_size)); 660 if (!pci_seg->dev_table) 661 return -ENOMEM; 662 663 return 0; 664 } 665 666 static inline void free_dev_table(struct amd_iommu_pci_seg *pci_seg) 667 { 668 free_pages((unsigned long)pci_seg->dev_table, 669 get_order(pci_seg->dev_table_size)); 670 pci_seg->dev_table = NULL; 671 } 672 673 /* Allocate per PCI segment IOMMU rlookup table. */ 674 static inline int __init alloc_rlookup_table(struct amd_iommu_pci_seg *pci_seg) 675 { 676 pci_seg->rlookup_table = (void *)__get_free_pages( 677 GFP_KERNEL | __GFP_ZERO, 678 get_order(pci_seg->rlookup_table_size)); 679 if (pci_seg->rlookup_table == NULL) 680 return -ENOMEM; 681 682 return 0; 683 } 684 685 static inline void free_rlookup_table(struct amd_iommu_pci_seg *pci_seg) 686 { 687 free_pages((unsigned long)pci_seg->rlookup_table, 688 get_order(pci_seg->rlookup_table_size)); 689 pci_seg->rlookup_table = NULL; 690 } 691 692 static inline int __init alloc_irq_lookup_table(struct amd_iommu_pci_seg *pci_seg) 693 { 694 pci_seg->irq_lookup_table = (void *)__get_free_pages( 695 GFP_KERNEL | __GFP_ZERO, 696 get_order(pci_seg->rlookup_table_size)); 697 kmemleak_alloc(pci_seg->irq_lookup_table, 698 pci_seg->rlookup_table_size, 1, GFP_KERNEL); 699 if (pci_seg->irq_lookup_table == NULL) 700 return -ENOMEM; 701 702 return 0; 703 } 704 705 static inline void free_irq_lookup_table(struct amd_iommu_pci_seg *pci_seg) 706 { 707 kmemleak_free(pci_seg->irq_lookup_table); 708 free_pages((unsigned long)pci_seg->irq_lookup_table, 709 get_order(pci_seg->rlookup_table_size)); 710 pci_seg->irq_lookup_table = NULL; 711 } 712 713 static int __init alloc_alias_table(struct amd_iommu_pci_seg *pci_seg) 714 { 715 int i; 716 717 pci_seg->alias_table = (void *)__get_free_pages(GFP_KERNEL, 718 get_order(pci_seg->alias_table_size)); 719 if (!pci_seg->alias_table) 720 return -ENOMEM; 721 722 /* 723 * let all alias entries point to itself 724 */ 725 for (i = 0; i <= pci_seg->last_bdf; ++i) 726 pci_seg->alias_table[i] = i; 727 728 return 0; 729 } 730 731 static void __init free_alias_table(struct amd_iommu_pci_seg *pci_seg) 732 { 733 free_pages((unsigned long)pci_seg->alias_table, 734 get_order(pci_seg->alias_table_size)); 735 pci_seg->alias_table = NULL; 736 } 737 738 /* 739 * Allocates the command buffer. This buffer is per AMD IOMMU. We can 740 * write commands to that buffer later and the IOMMU will execute them 741 * asynchronously 742 */ 743 static int __init alloc_command_buffer(struct amd_iommu *iommu) 744 { 745 iommu->cmd_buf = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 746 get_order(CMD_BUFFER_SIZE)); 747 748 return iommu->cmd_buf ? 0 : -ENOMEM; 749 } 750 751 /* 752 * This function restarts event logging in case the IOMMU experienced 753 * an event log buffer overflow. 754 */ 755 void amd_iommu_restart_event_logging(struct amd_iommu *iommu) 756 { 757 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN); 758 iommu_feature_enable(iommu, CONTROL_EVT_LOG_EN); 759 } 760 761 /* 762 * This function resets the command buffer if the IOMMU stopped fetching 763 * commands from it. 764 */ 765 static void amd_iommu_reset_cmd_buffer(struct amd_iommu *iommu) 766 { 767 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN); 768 769 writel(0x00, iommu->mmio_base + MMIO_CMD_HEAD_OFFSET); 770 writel(0x00, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET); 771 iommu->cmd_buf_head = 0; 772 iommu->cmd_buf_tail = 0; 773 774 iommu_feature_enable(iommu, CONTROL_CMDBUF_EN); 775 } 776 777 /* 778 * This function writes the command buffer address to the hardware and 779 * enables it. 780 */ 781 static void iommu_enable_command_buffer(struct amd_iommu *iommu) 782 { 783 u64 entry; 784 785 BUG_ON(iommu->cmd_buf == NULL); 786 787 entry = iommu_virt_to_phys(iommu->cmd_buf); 788 entry |= MMIO_CMD_SIZE_512; 789 790 memcpy_toio(iommu->mmio_base + MMIO_CMD_BUF_OFFSET, 791 &entry, sizeof(entry)); 792 793 amd_iommu_reset_cmd_buffer(iommu); 794 } 795 796 /* 797 * This function disables the command buffer 798 */ 799 static void iommu_disable_command_buffer(struct amd_iommu *iommu) 800 { 801 iommu_feature_disable(iommu, CONTROL_CMDBUF_EN); 802 } 803 804 static void __init free_command_buffer(struct amd_iommu *iommu) 805 { 806 free_pages((unsigned long)iommu->cmd_buf, get_order(CMD_BUFFER_SIZE)); 807 } 808 809 static void *__init iommu_alloc_4k_pages(struct amd_iommu *iommu, 810 gfp_t gfp, size_t size) 811 { 812 int order = get_order(size); 813 void *buf = (void *)__get_free_pages(gfp, order); 814 815 if (buf && 816 check_feature_on_all_iommus(FEATURE_SNP) && 817 set_memory_4k((unsigned long)buf, (1 << order))) { 818 free_pages((unsigned long)buf, order); 819 buf = NULL; 820 } 821 822 return buf; 823 } 824 825 /* allocates the memory where the IOMMU will log its events to */ 826 static int __init alloc_event_buffer(struct amd_iommu *iommu) 827 { 828 iommu->evt_buf = iommu_alloc_4k_pages(iommu, GFP_KERNEL | __GFP_ZERO, 829 EVT_BUFFER_SIZE); 830 831 return iommu->evt_buf ? 0 : -ENOMEM; 832 } 833 834 static void iommu_enable_event_buffer(struct amd_iommu *iommu) 835 { 836 u64 entry; 837 838 BUG_ON(iommu->evt_buf == NULL); 839 840 entry = iommu_virt_to_phys(iommu->evt_buf) | EVT_LEN_MASK; 841 842 memcpy_toio(iommu->mmio_base + MMIO_EVT_BUF_OFFSET, 843 &entry, sizeof(entry)); 844 845 /* set head and tail to zero manually */ 846 writel(0x00, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET); 847 writel(0x00, iommu->mmio_base + MMIO_EVT_TAIL_OFFSET); 848 849 iommu_feature_enable(iommu, CONTROL_EVT_LOG_EN); 850 } 851 852 /* 853 * This function disables the event log buffer 854 */ 855 static void iommu_disable_event_buffer(struct amd_iommu *iommu) 856 { 857 iommu_feature_disable(iommu, CONTROL_EVT_LOG_EN); 858 } 859 860 static void __init free_event_buffer(struct amd_iommu *iommu) 861 { 862 free_pages((unsigned long)iommu->evt_buf, get_order(EVT_BUFFER_SIZE)); 863 } 864 865 /* allocates the memory where the IOMMU will log its events to */ 866 static int __init alloc_ppr_log(struct amd_iommu *iommu) 867 { 868 iommu->ppr_log = iommu_alloc_4k_pages(iommu, GFP_KERNEL | __GFP_ZERO, 869 PPR_LOG_SIZE); 870 871 return iommu->ppr_log ? 0 : -ENOMEM; 872 } 873 874 static void iommu_enable_ppr_log(struct amd_iommu *iommu) 875 { 876 u64 entry; 877 878 if (iommu->ppr_log == NULL) 879 return; 880 881 entry = iommu_virt_to_phys(iommu->ppr_log) | PPR_LOG_SIZE_512; 882 883 memcpy_toio(iommu->mmio_base + MMIO_PPR_LOG_OFFSET, 884 &entry, sizeof(entry)); 885 886 /* set head and tail to zero manually */ 887 writel(0x00, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET); 888 writel(0x00, iommu->mmio_base + MMIO_PPR_TAIL_OFFSET); 889 890 iommu_feature_enable(iommu, CONTROL_PPRLOG_EN); 891 iommu_feature_enable(iommu, CONTROL_PPR_EN); 892 } 893 894 static void __init free_ppr_log(struct amd_iommu *iommu) 895 { 896 free_pages((unsigned long)iommu->ppr_log, get_order(PPR_LOG_SIZE)); 897 } 898 899 static void free_ga_log(struct amd_iommu *iommu) 900 { 901 #ifdef CONFIG_IRQ_REMAP 902 free_pages((unsigned long)iommu->ga_log, get_order(GA_LOG_SIZE)); 903 free_pages((unsigned long)iommu->ga_log_tail, get_order(8)); 904 #endif 905 } 906 907 #ifdef CONFIG_IRQ_REMAP 908 static int iommu_ga_log_enable(struct amd_iommu *iommu) 909 { 910 u32 status, i; 911 u64 entry; 912 913 if (!iommu->ga_log) 914 return -EINVAL; 915 916 entry = iommu_virt_to_phys(iommu->ga_log) | GA_LOG_SIZE_512; 917 memcpy_toio(iommu->mmio_base + MMIO_GA_LOG_BASE_OFFSET, 918 &entry, sizeof(entry)); 919 entry = (iommu_virt_to_phys(iommu->ga_log_tail) & 920 (BIT_ULL(52)-1)) & ~7ULL; 921 memcpy_toio(iommu->mmio_base + MMIO_GA_LOG_TAIL_OFFSET, 922 &entry, sizeof(entry)); 923 writel(0x00, iommu->mmio_base + MMIO_GA_HEAD_OFFSET); 924 writel(0x00, iommu->mmio_base + MMIO_GA_TAIL_OFFSET); 925 926 927 iommu_feature_enable(iommu, CONTROL_GAINT_EN); 928 iommu_feature_enable(iommu, CONTROL_GALOG_EN); 929 930 for (i = 0; i < LOOP_TIMEOUT; ++i) { 931 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); 932 if (status & (MMIO_STATUS_GALOG_RUN_MASK)) 933 break; 934 udelay(10); 935 } 936 937 if (WARN_ON(i >= LOOP_TIMEOUT)) 938 return -EINVAL; 939 940 return 0; 941 } 942 943 static int iommu_init_ga_log(struct amd_iommu *iommu) 944 { 945 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) 946 return 0; 947 948 iommu->ga_log = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 949 get_order(GA_LOG_SIZE)); 950 if (!iommu->ga_log) 951 goto err_out; 952 953 iommu->ga_log_tail = (u8 *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 954 get_order(8)); 955 if (!iommu->ga_log_tail) 956 goto err_out; 957 958 return 0; 959 err_out: 960 free_ga_log(iommu); 961 return -EINVAL; 962 } 963 #endif /* CONFIG_IRQ_REMAP */ 964 965 static int __init alloc_cwwb_sem(struct amd_iommu *iommu) 966 { 967 iommu->cmd_sem = iommu_alloc_4k_pages(iommu, GFP_KERNEL | __GFP_ZERO, 1); 968 969 return iommu->cmd_sem ? 0 : -ENOMEM; 970 } 971 972 static void __init free_cwwb_sem(struct amd_iommu *iommu) 973 { 974 if (iommu->cmd_sem) 975 free_page((unsigned long)iommu->cmd_sem); 976 } 977 978 static void iommu_enable_xt(struct amd_iommu *iommu) 979 { 980 #ifdef CONFIG_IRQ_REMAP 981 /* 982 * XT mode (32-bit APIC destination ID) requires 983 * GA mode (128-bit IRTE support) as a prerequisite. 984 */ 985 if (AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir) && 986 amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE) 987 iommu_feature_enable(iommu, CONTROL_XT_EN); 988 #endif /* CONFIG_IRQ_REMAP */ 989 } 990 991 static void iommu_enable_gt(struct amd_iommu *iommu) 992 { 993 if (!iommu_feature(iommu, FEATURE_GT)) 994 return; 995 996 iommu_feature_enable(iommu, CONTROL_GT_EN); 997 } 998 999 /* sets a specific bit in the device table entry. */ 1000 static void __set_dev_entry_bit(struct dev_table_entry *dev_table, 1001 u16 devid, u8 bit) 1002 { 1003 int i = (bit >> 6) & 0x03; 1004 int _bit = bit & 0x3f; 1005 1006 dev_table[devid].data[i] |= (1UL << _bit); 1007 } 1008 1009 static void set_dev_entry_bit(struct amd_iommu *iommu, u16 devid, u8 bit) 1010 { 1011 struct dev_table_entry *dev_table = get_dev_table(iommu); 1012 1013 return __set_dev_entry_bit(dev_table, devid, bit); 1014 } 1015 1016 static int __get_dev_entry_bit(struct dev_table_entry *dev_table, 1017 u16 devid, u8 bit) 1018 { 1019 int i = (bit >> 6) & 0x03; 1020 int _bit = bit & 0x3f; 1021 1022 return (dev_table[devid].data[i] & (1UL << _bit)) >> _bit; 1023 } 1024 1025 static int get_dev_entry_bit(struct amd_iommu *iommu, u16 devid, u8 bit) 1026 { 1027 struct dev_table_entry *dev_table = get_dev_table(iommu); 1028 1029 return __get_dev_entry_bit(dev_table, devid, bit); 1030 } 1031 1032 static bool __copy_device_table(struct amd_iommu *iommu) 1033 { 1034 u64 int_ctl, int_tab_len, entry = 0; 1035 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 1036 struct dev_table_entry *old_devtb = NULL; 1037 u32 lo, hi, devid, old_devtb_size; 1038 phys_addr_t old_devtb_phys; 1039 u16 dom_id, dte_v, irq_v; 1040 gfp_t gfp_flag; 1041 u64 tmp; 1042 1043 /* Each IOMMU use separate device table with the same size */ 1044 lo = readl(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET); 1045 hi = readl(iommu->mmio_base + MMIO_DEV_TABLE_OFFSET + 4); 1046 entry = (((u64) hi) << 32) + lo; 1047 1048 old_devtb_size = ((entry & ~PAGE_MASK) + 1) << 12; 1049 if (old_devtb_size != pci_seg->dev_table_size) { 1050 pr_err("The device table size of IOMMU:%d is not expected!\n", 1051 iommu->index); 1052 return false; 1053 } 1054 1055 /* 1056 * When SME is enabled in the first kernel, the entry includes the 1057 * memory encryption mask(sme_me_mask), we must remove the memory 1058 * encryption mask to obtain the true physical address in kdump kernel. 1059 */ 1060 old_devtb_phys = __sme_clr(entry) & PAGE_MASK; 1061 1062 if (old_devtb_phys >= 0x100000000ULL) { 1063 pr_err("The address of old device table is above 4G, not trustworthy!\n"); 1064 return false; 1065 } 1066 old_devtb = (cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT) && is_kdump_kernel()) 1067 ? (__force void *)ioremap_encrypted(old_devtb_phys, 1068 pci_seg->dev_table_size) 1069 : memremap(old_devtb_phys, pci_seg->dev_table_size, MEMREMAP_WB); 1070 1071 if (!old_devtb) 1072 return false; 1073 1074 gfp_flag = GFP_KERNEL | __GFP_ZERO | GFP_DMA32; 1075 pci_seg->old_dev_tbl_cpy = (void *)__get_free_pages(gfp_flag, 1076 get_order(pci_seg->dev_table_size)); 1077 if (pci_seg->old_dev_tbl_cpy == NULL) { 1078 pr_err("Failed to allocate memory for copying old device table!\n"); 1079 memunmap(old_devtb); 1080 return false; 1081 } 1082 1083 for (devid = 0; devid <= pci_seg->last_bdf; ++devid) { 1084 pci_seg->old_dev_tbl_cpy[devid] = old_devtb[devid]; 1085 dom_id = old_devtb[devid].data[1] & DEV_DOMID_MASK; 1086 dte_v = old_devtb[devid].data[0] & DTE_FLAG_V; 1087 1088 if (dte_v && dom_id) { 1089 pci_seg->old_dev_tbl_cpy[devid].data[0] = old_devtb[devid].data[0]; 1090 pci_seg->old_dev_tbl_cpy[devid].data[1] = old_devtb[devid].data[1]; 1091 __set_bit(dom_id, amd_iommu_pd_alloc_bitmap); 1092 /* If gcr3 table existed, mask it out */ 1093 if (old_devtb[devid].data[0] & DTE_FLAG_GV) { 1094 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B; 1095 tmp |= DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C; 1096 pci_seg->old_dev_tbl_cpy[devid].data[1] &= ~tmp; 1097 tmp = DTE_GCR3_VAL_A(~0ULL) << DTE_GCR3_SHIFT_A; 1098 tmp |= DTE_FLAG_GV; 1099 pci_seg->old_dev_tbl_cpy[devid].data[0] &= ~tmp; 1100 } 1101 } 1102 1103 irq_v = old_devtb[devid].data[2] & DTE_IRQ_REMAP_ENABLE; 1104 int_ctl = old_devtb[devid].data[2] & DTE_IRQ_REMAP_INTCTL_MASK; 1105 int_tab_len = old_devtb[devid].data[2] & DTE_INTTABLEN_MASK; 1106 if (irq_v && (int_ctl || int_tab_len)) { 1107 if ((int_ctl != DTE_IRQ_REMAP_INTCTL) || 1108 (int_tab_len != DTE_INTTABLEN)) { 1109 pr_err("Wrong old irq remapping flag: %#x\n", devid); 1110 memunmap(old_devtb); 1111 return false; 1112 } 1113 1114 pci_seg->old_dev_tbl_cpy[devid].data[2] = old_devtb[devid].data[2]; 1115 } 1116 } 1117 memunmap(old_devtb); 1118 1119 return true; 1120 } 1121 1122 static bool copy_device_table(void) 1123 { 1124 struct amd_iommu *iommu; 1125 struct amd_iommu_pci_seg *pci_seg; 1126 1127 if (!amd_iommu_pre_enabled) 1128 return false; 1129 1130 pr_warn("Translation is already enabled - trying to copy translation structures\n"); 1131 1132 /* 1133 * All IOMMUs within PCI segment shares common device table. 1134 * Hence copy device table only once per PCI segment. 1135 */ 1136 for_each_pci_segment(pci_seg) { 1137 for_each_iommu(iommu) { 1138 if (pci_seg->id != iommu->pci_seg->id) 1139 continue; 1140 if (!__copy_device_table(iommu)) 1141 return false; 1142 break; 1143 } 1144 } 1145 1146 return true; 1147 } 1148 1149 void amd_iommu_apply_erratum_63(struct amd_iommu *iommu, u16 devid) 1150 { 1151 int sysmgt; 1152 1153 sysmgt = get_dev_entry_bit(iommu, devid, DEV_ENTRY_SYSMGT1) | 1154 (get_dev_entry_bit(iommu, devid, DEV_ENTRY_SYSMGT2) << 1); 1155 1156 if (sysmgt == 0x01) 1157 set_dev_entry_bit(iommu, devid, DEV_ENTRY_IW); 1158 } 1159 1160 /* 1161 * This function takes the device specific flags read from the ACPI 1162 * table and sets up the device table entry with that information 1163 */ 1164 static void __init set_dev_entry_from_acpi(struct amd_iommu *iommu, 1165 u16 devid, u32 flags, u32 ext_flags) 1166 { 1167 if (flags & ACPI_DEVFLAG_INITPASS) 1168 set_dev_entry_bit(iommu, devid, DEV_ENTRY_INIT_PASS); 1169 if (flags & ACPI_DEVFLAG_EXTINT) 1170 set_dev_entry_bit(iommu, devid, DEV_ENTRY_EINT_PASS); 1171 if (flags & ACPI_DEVFLAG_NMI) 1172 set_dev_entry_bit(iommu, devid, DEV_ENTRY_NMI_PASS); 1173 if (flags & ACPI_DEVFLAG_SYSMGT1) 1174 set_dev_entry_bit(iommu, devid, DEV_ENTRY_SYSMGT1); 1175 if (flags & ACPI_DEVFLAG_SYSMGT2) 1176 set_dev_entry_bit(iommu, devid, DEV_ENTRY_SYSMGT2); 1177 if (flags & ACPI_DEVFLAG_LINT0) 1178 set_dev_entry_bit(iommu, devid, DEV_ENTRY_LINT0_PASS); 1179 if (flags & ACPI_DEVFLAG_LINT1) 1180 set_dev_entry_bit(iommu, devid, DEV_ENTRY_LINT1_PASS); 1181 1182 amd_iommu_apply_erratum_63(iommu, devid); 1183 1184 amd_iommu_set_rlookup_table(iommu, devid); 1185 } 1186 1187 int __init add_special_device(u8 type, u8 id, u32 *devid, bool cmd_line) 1188 { 1189 struct devid_map *entry; 1190 struct list_head *list; 1191 1192 if (type == IVHD_SPECIAL_IOAPIC) 1193 list = &ioapic_map; 1194 else if (type == IVHD_SPECIAL_HPET) 1195 list = &hpet_map; 1196 else 1197 return -EINVAL; 1198 1199 list_for_each_entry(entry, list, list) { 1200 if (!(entry->id == id && entry->cmd_line)) 1201 continue; 1202 1203 pr_info("Command-line override present for %s id %d - ignoring\n", 1204 type == IVHD_SPECIAL_IOAPIC ? "IOAPIC" : "HPET", id); 1205 1206 *devid = entry->devid; 1207 1208 return 0; 1209 } 1210 1211 entry = kzalloc(sizeof(*entry), GFP_KERNEL); 1212 if (!entry) 1213 return -ENOMEM; 1214 1215 entry->id = id; 1216 entry->devid = *devid; 1217 entry->cmd_line = cmd_line; 1218 1219 list_add_tail(&entry->list, list); 1220 1221 return 0; 1222 } 1223 1224 static int __init add_acpi_hid_device(u8 *hid, u8 *uid, u32 *devid, 1225 bool cmd_line) 1226 { 1227 struct acpihid_map_entry *entry; 1228 struct list_head *list = &acpihid_map; 1229 1230 list_for_each_entry(entry, list, list) { 1231 if (strcmp(entry->hid, hid) || 1232 (*uid && *entry->uid && strcmp(entry->uid, uid)) || 1233 !entry->cmd_line) 1234 continue; 1235 1236 pr_info("Command-line override for hid:%s uid:%s\n", 1237 hid, uid); 1238 *devid = entry->devid; 1239 return 0; 1240 } 1241 1242 entry = kzalloc(sizeof(*entry), GFP_KERNEL); 1243 if (!entry) 1244 return -ENOMEM; 1245 1246 memcpy(entry->uid, uid, strlen(uid)); 1247 memcpy(entry->hid, hid, strlen(hid)); 1248 entry->devid = *devid; 1249 entry->cmd_line = cmd_line; 1250 entry->root_devid = (entry->devid & (~0x7)); 1251 1252 pr_info("%s, add hid:%s, uid:%s, rdevid:%d\n", 1253 entry->cmd_line ? "cmd" : "ivrs", 1254 entry->hid, entry->uid, entry->root_devid); 1255 1256 list_add_tail(&entry->list, list); 1257 return 0; 1258 } 1259 1260 static int __init add_early_maps(void) 1261 { 1262 int i, ret; 1263 1264 for (i = 0; i < early_ioapic_map_size; ++i) { 1265 ret = add_special_device(IVHD_SPECIAL_IOAPIC, 1266 early_ioapic_map[i].id, 1267 &early_ioapic_map[i].devid, 1268 early_ioapic_map[i].cmd_line); 1269 if (ret) 1270 return ret; 1271 } 1272 1273 for (i = 0; i < early_hpet_map_size; ++i) { 1274 ret = add_special_device(IVHD_SPECIAL_HPET, 1275 early_hpet_map[i].id, 1276 &early_hpet_map[i].devid, 1277 early_hpet_map[i].cmd_line); 1278 if (ret) 1279 return ret; 1280 } 1281 1282 for (i = 0; i < early_acpihid_map_size; ++i) { 1283 ret = add_acpi_hid_device(early_acpihid_map[i].hid, 1284 early_acpihid_map[i].uid, 1285 &early_acpihid_map[i].devid, 1286 early_acpihid_map[i].cmd_line); 1287 if (ret) 1288 return ret; 1289 } 1290 1291 return 0; 1292 } 1293 1294 /* 1295 * Takes a pointer to an AMD IOMMU entry in the ACPI table and 1296 * initializes the hardware and our data structures with it. 1297 */ 1298 static int __init init_iommu_from_acpi(struct amd_iommu *iommu, 1299 struct ivhd_header *h) 1300 { 1301 u8 *p = (u8 *)h; 1302 u8 *end = p, flags = 0; 1303 u16 devid = 0, devid_start = 0, devid_to = 0, seg_id; 1304 u32 dev_i, ext_flags = 0; 1305 bool alias = false; 1306 struct ivhd_entry *e; 1307 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 1308 u32 ivhd_size; 1309 int ret; 1310 1311 1312 ret = add_early_maps(); 1313 if (ret) 1314 return ret; 1315 1316 amd_iommu_apply_ivrs_quirks(); 1317 1318 /* 1319 * First save the recommended feature enable bits from ACPI 1320 */ 1321 iommu->acpi_flags = h->flags; 1322 1323 /* 1324 * Done. Now parse the device entries 1325 */ 1326 ivhd_size = get_ivhd_header_size(h); 1327 if (!ivhd_size) { 1328 pr_err("Unsupported IVHD type %#x\n", h->type); 1329 return -EINVAL; 1330 } 1331 1332 p += ivhd_size; 1333 1334 end += h->length; 1335 1336 1337 while (p < end) { 1338 e = (struct ivhd_entry *)p; 1339 seg_id = pci_seg->id; 1340 1341 switch (e->type) { 1342 case IVHD_DEV_ALL: 1343 1344 DUMP_printk(" DEV_ALL\t\t\tflags: %02x\n", e->flags); 1345 1346 for (dev_i = 0; dev_i <= pci_seg->last_bdf; ++dev_i) 1347 set_dev_entry_from_acpi(iommu, dev_i, e->flags, 0); 1348 break; 1349 case IVHD_DEV_SELECT: 1350 1351 DUMP_printk(" DEV_SELECT\t\t\t devid: %04x:%02x:%02x.%x " 1352 "flags: %02x\n", 1353 seg_id, PCI_BUS_NUM(e->devid), 1354 PCI_SLOT(e->devid), 1355 PCI_FUNC(e->devid), 1356 e->flags); 1357 1358 devid = e->devid; 1359 set_dev_entry_from_acpi(iommu, devid, e->flags, 0); 1360 break; 1361 case IVHD_DEV_SELECT_RANGE_START: 1362 1363 DUMP_printk(" DEV_SELECT_RANGE_START\t " 1364 "devid: %04x:%02x:%02x.%x flags: %02x\n", 1365 seg_id, PCI_BUS_NUM(e->devid), 1366 PCI_SLOT(e->devid), 1367 PCI_FUNC(e->devid), 1368 e->flags); 1369 1370 devid_start = e->devid; 1371 flags = e->flags; 1372 ext_flags = 0; 1373 alias = false; 1374 break; 1375 case IVHD_DEV_ALIAS: 1376 1377 DUMP_printk(" DEV_ALIAS\t\t\t devid: %04x:%02x:%02x.%x " 1378 "flags: %02x devid_to: %02x:%02x.%x\n", 1379 seg_id, PCI_BUS_NUM(e->devid), 1380 PCI_SLOT(e->devid), 1381 PCI_FUNC(e->devid), 1382 e->flags, 1383 PCI_BUS_NUM(e->ext >> 8), 1384 PCI_SLOT(e->ext >> 8), 1385 PCI_FUNC(e->ext >> 8)); 1386 1387 devid = e->devid; 1388 devid_to = e->ext >> 8; 1389 set_dev_entry_from_acpi(iommu, devid , e->flags, 0); 1390 set_dev_entry_from_acpi(iommu, devid_to, e->flags, 0); 1391 pci_seg->alias_table[devid] = devid_to; 1392 break; 1393 case IVHD_DEV_ALIAS_RANGE: 1394 1395 DUMP_printk(" DEV_ALIAS_RANGE\t\t " 1396 "devid: %04x:%02x:%02x.%x flags: %02x " 1397 "devid_to: %04x:%02x:%02x.%x\n", 1398 seg_id, PCI_BUS_NUM(e->devid), 1399 PCI_SLOT(e->devid), 1400 PCI_FUNC(e->devid), 1401 e->flags, 1402 seg_id, PCI_BUS_NUM(e->ext >> 8), 1403 PCI_SLOT(e->ext >> 8), 1404 PCI_FUNC(e->ext >> 8)); 1405 1406 devid_start = e->devid; 1407 flags = e->flags; 1408 devid_to = e->ext >> 8; 1409 ext_flags = 0; 1410 alias = true; 1411 break; 1412 case IVHD_DEV_EXT_SELECT: 1413 1414 DUMP_printk(" DEV_EXT_SELECT\t\t devid: %04x:%02x:%02x.%x " 1415 "flags: %02x ext: %08x\n", 1416 seg_id, PCI_BUS_NUM(e->devid), 1417 PCI_SLOT(e->devid), 1418 PCI_FUNC(e->devid), 1419 e->flags, e->ext); 1420 1421 devid = e->devid; 1422 set_dev_entry_from_acpi(iommu, devid, e->flags, 1423 e->ext); 1424 break; 1425 case IVHD_DEV_EXT_SELECT_RANGE: 1426 1427 DUMP_printk(" DEV_EXT_SELECT_RANGE\t devid: " 1428 "%04x:%02x:%02x.%x flags: %02x ext: %08x\n", 1429 seg_id, PCI_BUS_NUM(e->devid), 1430 PCI_SLOT(e->devid), 1431 PCI_FUNC(e->devid), 1432 e->flags, e->ext); 1433 1434 devid_start = e->devid; 1435 flags = e->flags; 1436 ext_flags = e->ext; 1437 alias = false; 1438 break; 1439 case IVHD_DEV_RANGE_END: 1440 1441 DUMP_printk(" DEV_RANGE_END\t\t devid: %04x:%02x:%02x.%x\n", 1442 seg_id, PCI_BUS_NUM(e->devid), 1443 PCI_SLOT(e->devid), 1444 PCI_FUNC(e->devid)); 1445 1446 devid = e->devid; 1447 for (dev_i = devid_start; dev_i <= devid; ++dev_i) { 1448 if (alias) { 1449 pci_seg->alias_table[dev_i] = devid_to; 1450 set_dev_entry_from_acpi(iommu, 1451 devid_to, flags, ext_flags); 1452 } 1453 set_dev_entry_from_acpi(iommu, dev_i, 1454 flags, ext_flags); 1455 } 1456 break; 1457 case IVHD_DEV_SPECIAL: { 1458 u8 handle, type; 1459 const char *var; 1460 u32 devid; 1461 int ret; 1462 1463 handle = e->ext & 0xff; 1464 devid = PCI_SEG_DEVID_TO_SBDF(seg_id, (e->ext >> 8)); 1465 type = (e->ext >> 24) & 0xff; 1466 1467 if (type == IVHD_SPECIAL_IOAPIC) 1468 var = "IOAPIC"; 1469 else if (type == IVHD_SPECIAL_HPET) 1470 var = "HPET"; 1471 else 1472 var = "UNKNOWN"; 1473 1474 DUMP_printk(" DEV_SPECIAL(%s[%d])\t\tdevid: %04x:%02x:%02x.%x\n", 1475 var, (int)handle, 1476 seg_id, PCI_BUS_NUM(devid), 1477 PCI_SLOT(devid), 1478 PCI_FUNC(devid)); 1479 1480 ret = add_special_device(type, handle, &devid, false); 1481 if (ret) 1482 return ret; 1483 1484 /* 1485 * add_special_device might update the devid in case a 1486 * command-line override is present. So call 1487 * set_dev_entry_from_acpi after add_special_device. 1488 */ 1489 set_dev_entry_from_acpi(iommu, devid, e->flags, 0); 1490 1491 break; 1492 } 1493 case IVHD_DEV_ACPI_HID: { 1494 u32 devid; 1495 u8 hid[ACPIHID_HID_LEN]; 1496 u8 uid[ACPIHID_UID_LEN]; 1497 int ret; 1498 1499 if (h->type != 0x40) { 1500 pr_err(FW_BUG "Invalid IVHD device type %#x\n", 1501 e->type); 1502 break; 1503 } 1504 1505 BUILD_BUG_ON(sizeof(e->ext_hid) != ACPIHID_HID_LEN - 1); 1506 memcpy(hid, &e->ext_hid, ACPIHID_HID_LEN - 1); 1507 hid[ACPIHID_HID_LEN - 1] = '\0'; 1508 1509 if (!(*hid)) { 1510 pr_err(FW_BUG "Invalid HID.\n"); 1511 break; 1512 } 1513 1514 uid[0] = '\0'; 1515 switch (e->uidf) { 1516 case UID_NOT_PRESENT: 1517 1518 if (e->uidl != 0) 1519 pr_warn(FW_BUG "Invalid UID length.\n"); 1520 1521 break; 1522 case UID_IS_INTEGER: 1523 1524 sprintf(uid, "%d", e->uid); 1525 1526 break; 1527 case UID_IS_CHARACTER: 1528 1529 memcpy(uid, &e->uid, e->uidl); 1530 uid[e->uidl] = '\0'; 1531 1532 break; 1533 default: 1534 break; 1535 } 1536 1537 devid = PCI_SEG_DEVID_TO_SBDF(seg_id, e->devid); 1538 DUMP_printk(" DEV_ACPI_HID(%s[%s])\t\tdevid: %04x:%02x:%02x.%x\n", 1539 hid, uid, seg_id, 1540 PCI_BUS_NUM(devid), 1541 PCI_SLOT(devid), 1542 PCI_FUNC(devid)); 1543 1544 flags = e->flags; 1545 1546 ret = add_acpi_hid_device(hid, uid, &devid, false); 1547 if (ret) 1548 return ret; 1549 1550 /* 1551 * add_special_device might update the devid in case a 1552 * command-line override is present. So call 1553 * set_dev_entry_from_acpi after add_special_device. 1554 */ 1555 set_dev_entry_from_acpi(iommu, devid, e->flags, 0); 1556 1557 break; 1558 } 1559 default: 1560 break; 1561 } 1562 1563 p += ivhd_entry_length(p); 1564 } 1565 1566 return 0; 1567 } 1568 1569 /* Allocate PCI segment data structure */ 1570 static struct amd_iommu_pci_seg *__init alloc_pci_segment(u16 id, 1571 struct acpi_table_header *ivrs_base) 1572 { 1573 struct amd_iommu_pci_seg *pci_seg; 1574 int last_bdf; 1575 1576 /* 1577 * First parse ACPI tables to find the largest Bus/Dev/Func we need to 1578 * handle in this PCI segment. Upon this information the shared data 1579 * structures for the PCI segments in the system will be allocated. 1580 */ 1581 last_bdf = find_last_devid_acpi(ivrs_base, id); 1582 if (last_bdf < 0) 1583 return NULL; 1584 1585 pci_seg = kzalloc(sizeof(struct amd_iommu_pci_seg), GFP_KERNEL); 1586 if (pci_seg == NULL) 1587 return NULL; 1588 1589 pci_seg->last_bdf = last_bdf; 1590 DUMP_printk("PCI segment : 0x%0x, last bdf : 0x%04x\n", id, last_bdf); 1591 pci_seg->dev_table_size = tbl_size(DEV_TABLE_ENTRY_SIZE, last_bdf); 1592 pci_seg->alias_table_size = tbl_size(ALIAS_TABLE_ENTRY_SIZE, last_bdf); 1593 pci_seg->rlookup_table_size = tbl_size(RLOOKUP_TABLE_ENTRY_SIZE, last_bdf); 1594 1595 pci_seg->id = id; 1596 init_llist_head(&pci_seg->dev_data_list); 1597 INIT_LIST_HEAD(&pci_seg->unity_map); 1598 list_add_tail(&pci_seg->list, &amd_iommu_pci_seg_list); 1599 1600 if (alloc_dev_table(pci_seg)) 1601 return NULL; 1602 if (alloc_alias_table(pci_seg)) 1603 return NULL; 1604 if (alloc_rlookup_table(pci_seg)) 1605 return NULL; 1606 1607 return pci_seg; 1608 } 1609 1610 static struct amd_iommu_pci_seg *__init get_pci_segment(u16 id, 1611 struct acpi_table_header *ivrs_base) 1612 { 1613 struct amd_iommu_pci_seg *pci_seg; 1614 1615 for_each_pci_segment(pci_seg) { 1616 if (pci_seg->id == id) 1617 return pci_seg; 1618 } 1619 1620 return alloc_pci_segment(id, ivrs_base); 1621 } 1622 1623 static void __init free_pci_segments(void) 1624 { 1625 struct amd_iommu_pci_seg *pci_seg, *next; 1626 1627 for_each_pci_segment_safe(pci_seg, next) { 1628 list_del(&pci_seg->list); 1629 free_irq_lookup_table(pci_seg); 1630 free_rlookup_table(pci_seg); 1631 free_alias_table(pci_seg); 1632 free_dev_table(pci_seg); 1633 kfree(pci_seg); 1634 } 1635 } 1636 1637 static void __init free_iommu_one(struct amd_iommu *iommu) 1638 { 1639 free_cwwb_sem(iommu); 1640 free_command_buffer(iommu); 1641 free_event_buffer(iommu); 1642 free_ppr_log(iommu); 1643 free_ga_log(iommu); 1644 iommu_unmap_mmio_space(iommu); 1645 } 1646 1647 static void __init free_iommu_all(void) 1648 { 1649 struct amd_iommu *iommu, *next; 1650 1651 for_each_iommu_safe(iommu, next) { 1652 list_del(&iommu->list); 1653 free_iommu_one(iommu); 1654 kfree(iommu); 1655 } 1656 } 1657 1658 /* 1659 * Family15h Model 10h-1fh erratum 746 (IOMMU Logging May Stall Translations) 1660 * Workaround: 1661 * BIOS should disable L2B micellaneous clock gating by setting 1662 * L2_L2B_CK_GATE_CONTROL[CKGateL2BMiscDisable](D0F2xF4_x90[2]) = 1b 1663 */ 1664 static void amd_iommu_erratum_746_workaround(struct amd_iommu *iommu) 1665 { 1666 u32 value; 1667 1668 if ((boot_cpu_data.x86 != 0x15) || 1669 (boot_cpu_data.x86_model < 0x10) || 1670 (boot_cpu_data.x86_model > 0x1f)) 1671 return; 1672 1673 pci_write_config_dword(iommu->dev, 0xf0, 0x90); 1674 pci_read_config_dword(iommu->dev, 0xf4, &value); 1675 1676 if (value & BIT(2)) 1677 return; 1678 1679 /* Select NB indirect register 0x90 and enable writing */ 1680 pci_write_config_dword(iommu->dev, 0xf0, 0x90 | (1 << 8)); 1681 1682 pci_write_config_dword(iommu->dev, 0xf4, value | 0x4); 1683 pci_info(iommu->dev, "Applying erratum 746 workaround\n"); 1684 1685 /* Clear the enable writing bit */ 1686 pci_write_config_dword(iommu->dev, 0xf0, 0x90); 1687 } 1688 1689 /* 1690 * Family15h Model 30h-3fh (IOMMU Mishandles ATS Write Permission) 1691 * Workaround: 1692 * BIOS should enable ATS write permission check by setting 1693 * L2_DEBUG_3[AtsIgnoreIWDis](D0F2xF4_x47[0]) = 1b 1694 */ 1695 static void amd_iommu_ats_write_check_workaround(struct amd_iommu *iommu) 1696 { 1697 u32 value; 1698 1699 if ((boot_cpu_data.x86 != 0x15) || 1700 (boot_cpu_data.x86_model < 0x30) || 1701 (boot_cpu_data.x86_model > 0x3f)) 1702 return; 1703 1704 /* Test L2_DEBUG_3[AtsIgnoreIWDis] == 1 */ 1705 value = iommu_read_l2(iommu, 0x47); 1706 1707 if (value & BIT(0)) 1708 return; 1709 1710 /* Set L2_DEBUG_3[AtsIgnoreIWDis] = 1 */ 1711 iommu_write_l2(iommu, 0x47, value | BIT(0)); 1712 1713 pci_info(iommu->dev, "Applying ATS write check workaround\n"); 1714 } 1715 1716 /* 1717 * This function glues the initialization function for one IOMMU 1718 * together and also allocates the command buffer and programs the 1719 * hardware. It does NOT enable the IOMMU. This is done afterwards. 1720 */ 1721 static int __init init_iommu_one(struct amd_iommu *iommu, struct ivhd_header *h, 1722 struct acpi_table_header *ivrs_base) 1723 { 1724 struct amd_iommu_pci_seg *pci_seg; 1725 1726 pci_seg = get_pci_segment(h->pci_seg, ivrs_base); 1727 if (pci_seg == NULL) 1728 return -ENOMEM; 1729 iommu->pci_seg = pci_seg; 1730 1731 raw_spin_lock_init(&iommu->lock); 1732 iommu->cmd_sem_val = 0; 1733 1734 /* Add IOMMU to internal data structures */ 1735 list_add_tail(&iommu->list, &amd_iommu_list); 1736 iommu->index = amd_iommus_present++; 1737 1738 if (unlikely(iommu->index >= MAX_IOMMUS)) { 1739 WARN(1, "System has more IOMMUs than supported by this driver\n"); 1740 return -ENOSYS; 1741 } 1742 1743 /* Index is fine - add IOMMU to the array */ 1744 amd_iommus[iommu->index] = iommu; 1745 1746 /* 1747 * Copy data from ACPI table entry to the iommu struct 1748 */ 1749 iommu->devid = h->devid; 1750 iommu->cap_ptr = h->cap_ptr; 1751 iommu->mmio_phys = h->mmio_phys; 1752 1753 switch (h->type) { 1754 case 0x10: 1755 /* Check if IVHD EFR contains proper max banks/counters */ 1756 if ((h->efr_attr != 0) && 1757 ((h->efr_attr & (0xF << 13)) != 0) && 1758 ((h->efr_attr & (0x3F << 17)) != 0)) 1759 iommu->mmio_phys_end = MMIO_REG_END_OFFSET; 1760 else 1761 iommu->mmio_phys_end = MMIO_CNTR_CONF_OFFSET; 1762 1763 /* 1764 * Note: GA (128-bit IRTE) mode requires cmpxchg16b supports. 1765 * GAM also requires GA mode. Therefore, we need to 1766 * check cmpxchg16b support before enabling it. 1767 */ 1768 if (!boot_cpu_has(X86_FEATURE_CX16) || 1769 ((h->efr_attr & (0x1 << IOMMU_FEAT_GASUP_SHIFT)) == 0)) 1770 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY; 1771 break; 1772 case 0x11: 1773 case 0x40: 1774 if (h->efr_reg & (1 << 9)) 1775 iommu->mmio_phys_end = MMIO_REG_END_OFFSET; 1776 else 1777 iommu->mmio_phys_end = MMIO_CNTR_CONF_OFFSET; 1778 1779 /* 1780 * Note: GA (128-bit IRTE) mode requires cmpxchg16b supports. 1781 * XT, GAM also requires GA mode. Therefore, we need to 1782 * check cmpxchg16b support before enabling them. 1783 */ 1784 if (!boot_cpu_has(X86_FEATURE_CX16) || 1785 ((h->efr_reg & (0x1 << IOMMU_EFR_GASUP_SHIFT)) == 0)) { 1786 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY; 1787 break; 1788 } 1789 1790 if (h->efr_reg & BIT(IOMMU_EFR_XTSUP_SHIFT)) 1791 amd_iommu_xt_mode = IRQ_REMAP_X2APIC_MODE; 1792 1793 early_iommu_features_init(iommu, h); 1794 1795 break; 1796 default: 1797 return -EINVAL; 1798 } 1799 1800 iommu->mmio_base = iommu_map_mmio_space(iommu->mmio_phys, 1801 iommu->mmio_phys_end); 1802 if (!iommu->mmio_base) 1803 return -ENOMEM; 1804 1805 return init_iommu_from_acpi(iommu, h); 1806 } 1807 1808 static int __init init_iommu_one_late(struct amd_iommu *iommu) 1809 { 1810 int ret; 1811 1812 if (alloc_cwwb_sem(iommu)) 1813 return -ENOMEM; 1814 1815 if (alloc_command_buffer(iommu)) 1816 return -ENOMEM; 1817 1818 if (alloc_event_buffer(iommu)) 1819 return -ENOMEM; 1820 1821 iommu->int_enabled = false; 1822 1823 init_translation_status(iommu); 1824 if (translation_pre_enabled(iommu) && !is_kdump_kernel()) { 1825 iommu_disable(iommu); 1826 clear_translation_pre_enabled(iommu); 1827 pr_warn("Translation was enabled for IOMMU:%d but we are not in kdump mode\n", 1828 iommu->index); 1829 } 1830 if (amd_iommu_pre_enabled) 1831 amd_iommu_pre_enabled = translation_pre_enabled(iommu); 1832 1833 if (amd_iommu_irq_remap) { 1834 ret = amd_iommu_create_irq_domain(iommu); 1835 if (ret) 1836 return ret; 1837 } 1838 1839 /* 1840 * Make sure IOMMU is not considered to translate itself. The IVRS 1841 * table tells us so, but this is a lie! 1842 */ 1843 iommu->pci_seg->rlookup_table[iommu->devid] = NULL; 1844 1845 return 0; 1846 } 1847 1848 /** 1849 * get_highest_supported_ivhd_type - Look up the appropriate IVHD type 1850 * @ivrs: Pointer to the IVRS header 1851 * 1852 * This function search through all IVDB of the maximum supported IVHD 1853 */ 1854 static u8 get_highest_supported_ivhd_type(struct acpi_table_header *ivrs) 1855 { 1856 u8 *base = (u8 *)ivrs; 1857 struct ivhd_header *ivhd = (struct ivhd_header *) 1858 (base + IVRS_HEADER_LENGTH); 1859 u8 last_type = ivhd->type; 1860 u16 devid = ivhd->devid; 1861 1862 while (((u8 *)ivhd - base < ivrs->length) && 1863 (ivhd->type <= ACPI_IVHD_TYPE_MAX_SUPPORTED)) { 1864 u8 *p = (u8 *) ivhd; 1865 1866 if (ivhd->devid == devid) 1867 last_type = ivhd->type; 1868 ivhd = (struct ivhd_header *)(p + ivhd->length); 1869 } 1870 1871 return last_type; 1872 } 1873 1874 /* 1875 * Iterates over all IOMMU entries in the ACPI table, allocates the 1876 * IOMMU structure and initializes it with init_iommu_one() 1877 */ 1878 static int __init init_iommu_all(struct acpi_table_header *table) 1879 { 1880 u8 *p = (u8 *)table, *end = (u8 *)table; 1881 struct ivhd_header *h; 1882 struct amd_iommu *iommu; 1883 int ret; 1884 1885 end += table->length; 1886 p += IVRS_HEADER_LENGTH; 1887 1888 /* Phase 1: Process all IVHD blocks */ 1889 while (p < end) { 1890 h = (struct ivhd_header *)p; 1891 if (*p == amd_iommu_target_ivhd_type) { 1892 1893 DUMP_printk("device: %04x:%02x:%02x.%01x cap: %04x " 1894 "flags: %01x info %04x\n", 1895 h->pci_seg, PCI_BUS_NUM(h->devid), 1896 PCI_SLOT(h->devid), PCI_FUNC(h->devid), 1897 h->cap_ptr, h->flags, h->info); 1898 DUMP_printk(" mmio-addr: %016llx\n", 1899 h->mmio_phys); 1900 1901 iommu = kzalloc(sizeof(struct amd_iommu), GFP_KERNEL); 1902 if (iommu == NULL) 1903 return -ENOMEM; 1904 1905 ret = init_iommu_one(iommu, h, table); 1906 if (ret) 1907 return ret; 1908 } 1909 p += h->length; 1910 1911 } 1912 WARN_ON(p != end); 1913 1914 /* Phase 2 : Early feature support check */ 1915 get_global_efr(); 1916 1917 /* Phase 3 : Enabling IOMMU features */ 1918 for_each_iommu(iommu) { 1919 ret = init_iommu_one_late(iommu); 1920 if (ret) 1921 return ret; 1922 } 1923 1924 return 0; 1925 } 1926 1927 static void init_iommu_perf_ctr(struct amd_iommu *iommu) 1928 { 1929 u64 val; 1930 struct pci_dev *pdev = iommu->dev; 1931 1932 if (!iommu_feature(iommu, FEATURE_PC)) 1933 return; 1934 1935 amd_iommu_pc_present = true; 1936 1937 pci_info(pdev, "IOMMU performance counters supported\n"); 1938 1939 val = readl(iommu->mmio_base + MMIO_CNTR_CONF_OFFSET); 1940 iommu->max_banks = (u8) ((val >> 12) & 0x3f); 1941 iommu->max_counters = (u8) ((val >> 7) & 0xf); 1942 1943 return; 1944 } 1945 1946 static ssize_t amd_iommu_show_cap(struct device *dev, 1947 struct device_attribute *attr, 1948 char *buf) 1949 { 1950 struct amd_iommu *iommu = dev_to_amd_iommu(dev); 1951 return sysfs_emit(buf, "%x\n", iommu->cap); 1952 } 1953 static DEVICE_ATTR(cap, S_IRUGO, amd_iommu_show_cap, NULL); 1954 1955 static ssize_t amd_iommu_show_features(struct device *dev, 1956 struct device_attribute *attr, 1957 char *buf) 1958 { 1959 struct amd_iommu *iommu = dev_to_amd_iommu(dev); 1960 return sysfs_emit(buf, "%llx:%llx\n", iommu->features2, iommu->features); 1961 } 1962 static DEVICE_ATTR(features, S_IRUGO, amd_iommu_show_features, NULL); 1963 1964 static struct attribute *amd_iommu_attrs[] = { 1965 &dev_attr_cap.attr, 1966 &dev_attr_features.attr, 1967 NULL, 1968 }; 1969 1970 static struct attribute_group amd_iommu_group = { 1971 .name = "amd-iommu", 1972 .attrs = amd_iommu_attrs, 1973 }; 1974 1975 static const struct attribute_group *amd_iommu_groups[] = { 1976 &amd_iommu_group, 1977 NULL, 1978 }; 1979 1980 /* 1981 * Note: IVHD 0x11 and 0x40 also contains exact copy 1982 * of the IOMMU Extended Feature Register [MMIO Offset 0030h]. 1983 * Default to EFR in IVHD since it is available sooner (i.e. before PCI init). 1984 */ 1985 static void __init late_iommu_features_init(struct amd_iommu *iommu) 1986 { 1987 u64 features, features2; 1988 1989 if (!(iommu->cap & (1 << IOMMU_CAP_EFR))) 1990 return; 1991 1992 /* read extended feature bits */ 1993 features = readq(iommu->mmio_base + MMIO_EXT_FEATURES); 1994 features2 = readq(iommu->mmio_base + MMIO_EXT_FEATURES2); 1995 1996 if (!iommu->features) { 1997 iommu->features = features; 1998 iommu->features2 = features2; 1999 return; 2000 } 2001 2002 /* 2003 * Sanity check and warn if EFR values from 2004 * IVHD and MMIO conflict. 2005 */ 2006 if (features != iommu->features || 2007 features2 != iommu->features2) { 2008 pr_warn(FW_WARN 2009 "EFR mismatch. Use IVHD EFR (%#llx : %#llx), EFR2 (%#llx : %#llx).\n", 2010 features, iommu->features, 2011 features2, iommu->features2); 2012 } 2013 } 2014 2015 static int __init iommu_init_pci(struct amd_iommu *iommu) 2016 { 2017 int cap_ptr = iommu->cap_ptr; 2018 int ret; 2019 2020 iommu->dev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, 2021 PCI_BUS_NUM(iommu->devid), 2022 iommu->devid & 0xff); 2023 if (!iommu->dev) 2024 return -ENODEV; 2025 2026 /* Prevent binding other PCI device drivers to IOMMU devices */ 2027 iommu->dev->match_driver = false; 2028 2029 pci_read_config_dword(iommu->dev, cap_ptr + MMIO_CAP_HDR_OFFSET, 2030 &iommu->cap); 2031 2032 if (!(iommu->cap & (1 << IOMMU_CAP_IOTLB))) 2033 amd_iommu_iotlb_sup = false; 2034 2035 late_iommu_features_init(iommu); 2036 2037 if (iommu_feature(iommu, FEATURE_GT)) { 2038 int glxval; 2039 u32 max_pasid; 2040 u64 pasmax; 2041 2042 pasmax = iommu->features & FEATURE_PASID_MASK; 2043 pasmax >>= FEATURE_PASID_SHIFT; 2044 max_pasid = (1 << (pasmax + 1)) - 1; 2045 2046 amd_iommu_max_pasid = min(amd_iommu_max_pasid, max_pasid); 2047 2048 BUG_ON(amd_iommu_max_pasid & ~PASID_MASK); 2049 2050 glxval = iommu->features & FEATURE_GLXVAL_MASK; 2051 glxval >>= FEATURE_GLXVAL_SHIFT; 2052 2053 if (amd_iommu_max_glx_val == -1) 2054 amd_iommu_max_glx_val = glxval; 2055 else 2056 amd_iommu_max_glx_val = min(amd_iommu_max_glx_val, glxval); 2057 } 2058 2059 if (iommu_feature(iommu, FEATURE_GT) && 2060 iommu_feature(iommu, FEATURE_PPR)) { 2061 iommu->is_iommu_v2 = true; 2062 amd_iommu_v2_present = true; 2063 } 2064 2065 if (iommu_feature(iommu, FEATURE_PPR) && alloc_ppr_log(iommu)) 2066 return -ENOMEM; 2067 2068 if (iommu->cap & (1UL << IOMMU_CAP_NPCACHE)) { 2069 pr_info("Using strict mode due to virtualization\n"); 2070 iommu_set_dma_strict(); 2071 amd_iommu_np_cache = true; 2072 } 2073 2074 init_iommu_perf_ctr(iommu); 2075 2076 if (amd_iommu_pgtable == AMD_IOMMU_V2) { 2077 if (!iommu_feature(iommu, FEATURE_GIOSUP) || 2078 !iommu_feature(iommu, FEATURE_GT)) { 2079 pr_warn("Cannot enable v2 page table for DMA-API. Fallback to v1.\n"); 2080 amd_iommu_pgtable = AMD_IOMMU_V1; 2081 } else if (iommu_default_passthrough()) { 2082 pr_warn("V2 page table doesn't support passthrough mode. Fallback to v1.\n"); 2083 amd_iommu_pgtable = AMD_IOMMU_V1; 2084 } 2085 } 2086 2087 if (is_rd890_iommu(iommu->dev)) { 2088 int i, j; 2089 2090 iommu->root_pdev = 2091 pci_get_domain_bus_and_slot(iommu->pci_seg->id, 2092 iommu->dev->bus->number, 2093 PCI_DEVFN(0, 0)); 2094 2095 /* 2096 * Some rd890 systems may not be fully reconfigured by the 2097 * BIOS, so it's necessary for us to store this information so 2098 * it can be reprogrammed on resume 2099 */ 2100 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 4, 2101 &iommu->stored_addr_lo); 2102 pci_read_config_dword(iommu->dev, iommu->cap_ptr + 8, 2103 &iommu->stored_addr_hi); 2104 2105 /* Low bit locks writes to configuration space */ 2106 iommu->stored_addr_lo &= ~1; 2107 2108 for (i = 0; i < 6; i++) 2109 for (j = 0; j < 0x12; j++) 2110 iommu->stored_l1[i][j] = iommu_read_l1(iommu, i, j); 2111 2112 for (i = 0; i < 0x83; i++) 2113 iommu->stored_l2[i] = iommu_read_l2(iommu, i); 2114 } 2115 2116 amd_iommu_erratum_746_workaround(iommu); 2117 amd_iommu_ats_write_check_workaround(iommu); 2118 2119 ret = iommu_device_sysfs_add(&iommu->iommu, &iommu->dev->dev, 2120 amd_iommu_groups, "ivhd%d", iommu->index); 2121 if (ret) 2122 return ret; 2123 2124 iommu_device_register(&iommu->iommu, &amd_iommu_ops, NULL); 2125 2126 return pci_enable_device(iommu->dev); 2127 } 2128 2129 static void print_iommu_info(void) 2130 { 2131 static const char * const feat_str[] = { 2132 "PreF", "PPR", "X2APIC", "NX", "GT", "[5]", 2133 "IA", "GA", "HE", "PC" 2134 }; 2135 struct amd_iommu *iommu; 2136 2137 for_each_iommu(iommu) { 2138 struct pci_dev *pdev = iommu->dev; 2139 int i; 2140 2141 pci_info(pdev, "Found IOMMU cap 0x%x\n", iommu->cap_ptr); 2142 2143 if (iommu->cap & (1 << IOMMU_CAP_EFR)) { 2144 pr_info("Extended features (%#llx, %#llx):", iommu->features, iommu->features2); 2145 2146 for (i = 0; i < ARRAY_SIZE(feat_str); ++i) { 2147 if (iommu_feature(iommu, (1ULL << i))) 2148 pr_cont(" %s", feat_str[i]); 2149 } 2150 2151 if (iommu->features & FEATURE_GAM_VAPIC) 2152 pr_cont(" GA_vAPIC"); 2153 2154 if (iommu->features & FEATURE_SNP) 2155 pr_cont(" SNP"); 2156 2157 pr_cont("\n"); 2158 } 2159 } 2160 if (irq_remapping_enabled) { 2161 pr_info("Interrupt remapping enabled\n"); 2162 if (amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE) 2163 pr_info("X2APIC enabled\n"); 2164 } 2165 if (amd_iommu_pgtable == AMD_IOMMU_V2) { 2166 pr_info("V2 page table enabled (Paging mode : %d level)\n", 2167 amd_iommu_gpt_level); 2168 } 2169 } 2170 2171 static int __init amd_iommu_init_pci(void) 2172 { 2173 struct amd_iommu *iommu; 2174 struct amd_iommu_pci_seg *pci_seg; 2175 int ret; 2176 2177 for_each_iommu(iommu) { 2178 ret = iommu_init_pci(iommu); 2179 if (ret) { 2180 pr_err("IOMMU%d: Failed to initialize IOMMU Hardware (error=%d)!\n", 2181 iommu->index, ret); 2182 goto out; 2183 } 2184 /* Need to setup range after PCI init */ 2185 iommu_set_cwwb_range(iommu); 2186 } 2187 2188 /* 2189 * Order is important here to make sure any unity map requirements are 2190 * fulfilled. The unity mappings are created and written to the device 2191 * table during the iommu_init_pci() call. 2192 * 2193 * After that we call init_device_table_dma() to make sure any 2194 * uninitialized DTE will block DMA, and in the end we flush the caches 2195 * of all IOMMUs to make sure the changes to the device table are 2196 * active. 2197 */ 2198 for_each_pci_segment(pci_seg) 2199 init_device_table_dma(pci_seg); 2200 2201 for_each_iommu(iommu) 2202 iommu_flush_all_caches(iommu); 2203 2204 print_iommu_info(); 2205 2206 out: 2207 return ret; 2208 } 2209 2210 /**************************************************************************** 2211 * 2212 * The following functions initialize the MSI interrupts for all IOMMUs 2213 * in the system. It's a bit challenging because there could be multiple 2214 * IOMMUs per PCI BDF but we can call pci_enable_msi(x) only once per 2215 * pci_dev. 2216 * 2217 ****************************************************************************/ 2218 2219 static int iommu_setup_msi(struct amd_iommu *iommu) 2220 { 2221 int r; 2222 2223 r = pci_enable_msi(iommu->dev); 2224 if (r) 2225 return r; 2226 2227 r = request_threaded_irq(iommu->dev->irq, 2228 amd_iommu_int_handler, 2229 amd_iommu_int_thread, 2230 0, "AMD-Vi", 2231 iommu); 2232 2233 if (r) { 2234 pci_disable_msi(iommu->dev); 2235 return r; 2236 } 2237 2238 return 0; 2239 } 2240 2241 union intcapxt { 2242 u64 capxt; 2243 struct { 2244 u64 reserved_0 : 2, 2245 dest_mode_logical : 1, 2246 reserved_1 : 5, 2247 destid_0_23 : 24, 2248 vector : 8, 2249 reserved_2 : 16, 2250 destid_24_31 : 8; 2251 }; 2252 } __attribute__ ((packed)); 2253 2254 2255 static struct irq_chip intcapxt_controller; 2256 2257 static int intcapxt_irqdomain_activate(struct irq_domain *domain, 2258 struct irq_data *irqd, bool reserve) 2259 { 2260 return 0; 2261 } 2262 2263 static void intcapxt_irqdomain_deactivate(struct irq_domain *domain, 2264 struct irq_data *irqd) 2265 { 2266 } 2267 2268 2269 static int intcapxt_irqdomain_alloc(struct irq_domain *domain, unsigned int virq, 2270 unsigned int nr_irqs, void *arg) 2271 { 2272 struct irq_alloc_info *info = arg; 2273 int i, ret; 2274 2275 if (!info || info->type != X86_IRQ_ALLOC_TYPE_AMDVI) 2276 return -EINVAL; 2277 2278 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg); 2279 if (ret < 0) 2280 return ret; 2281 2282 for (i = virq; i < virq + nr_irqs; i++) { 2283 struct irq_data *irqd = irq_domain_get_irq_data(domain, i); 2284 2285 irqd->chip = &intcapxt_controller; 2286 irqd->chip_data = info->data; 2287 __irq_set_handler(i, handle_edge_irq, 0, "edge"); 2288 } 2289 2290 return ret; 2291 } 2292 2293 static void intcapxt_irqdomain_free(struct irq_domain *domain, unsigned int virq, 2294 unsigned int nr_irqs) 2295 { 2296 irq_domain_free_irqs_top(domain, virq, nr_irqs); 2297 } 2298 2299 2300 static void intcapxt_unmask_irq(struct irq_data *irqd) 2301 { 2302 struct amd_iommu *iommu = irqd->chip_data; 2303 struct irq_cfg *cfg = irqd_cfg(irqd); 2304 union intcapxt xt; 2305 2306 xt.capxt = 0ULL; 2307 xt.dest_mode_logical = apic->dest_mode_logical; 2308 xt.vector = cfg->vector; 2309 xt.destid_0_23 = cfg->dest_apicid & GENMASK(23, 0); 2310 xt.destid_24_31 = cfg->dest_apicid >> 24; 2311 2312 /** 2313 * Current IOMMU implementation uses the same IRQ for all 2314 * 3 IOMMU interrupts. 2315 */ 2316 writeq(xt.capxt, iommu->mmio_base + MMIO_INTCAPXT_EVT_OFFSET); 2317 writeq(xt.capxt, iommu->mmio_base + MMIO_INTCAPXT_PPR_OFFSET); 2318 writeq(xt.capxt, iommu->mmio_base + MMIO_INTCAPXT_GALOG_OFFSET); 2319 } 2320 2321 static void intcapxt_mask_irq(struct irq_data *irqd) 2322 { 2323 struct amd_iommu *iommu = irqd->chip_data; 2324 2325 writeq(0, iommu->mmio_base + MMIO_INTCAPXT_EVT_OFFSET); 2326 writeq(0, iommu->mmio_base + MMIO_INTCAPXT_PPR_OFFSET); 2327 writeq(0, iommu->mmio_base + MMIO_INTCAPXT_GALOG_OFFSET); 2328 } 2329 2330 2331 static int intcapxt_set_affinity(struct irq_data *irqd, 2332 const struct cpumask *mask, bool force) 2333 { 2334 struct irq_data *parent = irqd->parent_data; 2335 int ret; 2336 2337 ret = parent->chip->irq_set_affinity(parent, mask, force); 2338 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE) 2339 return ret; 2340 return 0; 2341 } 2342 2343 static int intcapxt_set_wake(struct irq_data *irqd, unsigned int on) 2344 { 2345 return on ? -EOPNOTSUPP : 0; 2346 } 2347 2348 static struct irq_chip intcapxt_controller = { 2349 .name = "IOMMU-MSI", 2350 .irq_unmask = intcapxt_unmask_irq, 2351 .irq_mask = intcapxt_mask_irq, 2352 .irq_ack = irq_chip_ack_parent, 2353 .irq_retrigger = irq_chip_retrigger_hierarchy, 2354 .irq_set_affinity = intcapxt_set_affinity, 2355 .irq_set_wake = intcapxt_set_wake, 2356 .flags = IRQCHIP_MASK_ON_SUSPEND, 2357 }; 2358 2359 static const struct irq_domain_ops intcapxt_domain_ops = { 2360 .alloc = intcapxt_irqdomain_alloc, 2361 .free = intcapxt_irqdomain_free, 2362 .activate = intcapxt_irqdomain_activate, 2363 .deactivate = intcapxt_irqdomain_deactivate, 2364 }; 2365 2366 2367 static struct irq_domain *iommu_irqdomain; 2368 2369 static struct irq_domain *iommu_get_irqdomain(void) 2370 { 2371 struct fwnode_handle *fn; 2372 2373 /* No need for locking here (yet) as the init is single-threaded */ 2374 if (iommu_irqdomain) 2375 return iommu_irqdomain; 2376 2377 fn = irq_domain_alloc_named_fwnode("AMD-Vi-MSI"); 2378 if (!fn) 2379 return NULL; 2380 2381 iommu_irqdomain = irq_domain_create_hierarchy(x86_vector_domain, 0, 0, 2382 fn, &intcapxt_domain_ops, 2383 NULL); 2384 if (!iommu_irqdomain) 2385 irq_domain_free_fwnode(fn); 2386 2387 return iommu_irqdomain; 2388 } 2389 2390 static int iommu_setup_intcapxt(struct amd_iommu *iommu) 2391 { 2392 struct irq_domain *domain; 2393 struct irq_alloc_info info; 2394 int irq, ret; 2395 int node = dev_to_node(&iommu->dev->dev); 2396 2397 domain = iommu_get_irqdomain(); 2398 if (!domain) 2399 return -ENXIO; 2400 2401 init_irq_alloc_info(&info, NULL); 2402 info.type = X86_IRQ_ALLOC_TYPE_AMDVI; 2403 info.data = iommu; 2404 2405 irq = irq_domain_alloc_irqs(domain, 1, node, &info); 2406 if (irq < 0) { 2407 irq_domain_remove(domain); 2408 return irq; 2409 } 2410 2411 ret = request_threaded_irq(irq, amd_iommu_int_handler, 2412 amd_iommu_int_thread, 0, "AMD-Vi", iommu); 2413 if (ret) { 2414 irq_domain_free_irqs(irq, 1); 2415 irq_domain_remove(domain); 2416 return ret; 2417 } 2418 2419 return 0; 2420 } 2421 2422 static int iommu_init_irq(struct amd_iommu *iommu) 2423 { 2424 int ret; 2425 2426 if (iommu->int_enabled) 2427 goto enable_faults; 2428 2429 if (amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE) 2430 ret = iommu_setup_intcapxt(iommu); 2431 else if (iommu->dev->msi_cap) 2432 ret = iommu_setup_msi(iommu); 2433 else 2434 ret = -ENODEV; 2435 2436 if (ret) 2437 return ret; 2438 2439 iommu->int_enabled = true; 2440 enable_faults: 2441 2442 if (amd_iommu_xt_mode == IRQ_REMAP_X2APIC_MODE) 2443 iommu_feature_enable(iommu, CONTROL_INTCAPXT_EN); 2444 2445 iommu_feature_enable(iommu, CONTROL_EVT_INT_EN); 2446 2447 if (iommu->ppr_log != NULL) 2448 iommu_feature_enable(iommu, CONTROL_PPRINT_EN); 2449 return 0; 2450 } 2451 2452 /**************************************************************************** 2453 * 2454 * The next functions belong to the third pass of parsing the ACPI 2455 * table. In this last pass the memory mapping requirements are 2456 * gathered (like exclusion and unity mapping ranges). 2457 * 2458 ****************************************************************************/ 2459 2460 static void __init free_unity_maps(void) 2461 { 2462 struct unity_map_entry *entry, *next; 2463 struct amd_iommu_pci_seg *p, *pci_seg; 2464 2465 for_each_pci_segment_safe(pci_seg, p) { 2466 list_for_each_entry_safe(entry, next, &pci_seg->unity_map, list) { 2467 list_del(&entry->list); 2468 kfree(entry); 2469 } 2470 } 2471 } 2472 2473 /* called for unity map ACPI definition */ 2474 static int __init init_unity_map_range(struct ivmd_header *m, 2475 struct acpi_table_header *ivrs_base) 2476 { 2477 struct unity_map_entry *e = NULL; 2478 struct amd_iommu_pci_seg *pci_seg; 2479 char *s; 2480 2481 pci_seg = get_pci_segment(m->pci_seg, ivrs_base); 2482 if (pci_seg == NULL) 2483 return -ENOMEM; 2484 2485 e = kzalloc(sizeof(*e), GFP_KERNEL); 2486 if (e == NULL) 2487 return -ENOMEM; 2488 2489 switch (m->type) { 2490 default: 2491 kfree(e); 2492 return 0; 2493 case ACPI_IVMD_TYPE: 2494 s = "IVMD_TYPEi\t\t\t"; 2495 e->devid_start = e->devid_end = m->devid; 2496 break; 2497 case ACPI_IVMD_TYPE_ALL: 2498 s = "IVMD_TYPE_ALL\t\t"; 2499 e->devid_start = 0; 2500 e->devid_end = pci_seg->last_bdf; 2501 break; 2502 case ACPI_IVMD_TYPE_RANGE: 2503 s = "IVMD_TYPE_RANGE\t\t"; 2504 e->devid_start = m->devid; 2505 e->devid_end = m->aux; 2506 break; 2507 } 2508 e->address_start = PAGE_ALIGN(m->range_start); 2509 e->address_end = e->address_start + PAGE_ALIGN(m->range_length); 2510 e->prot = m->flags >> 1; 2511 2512 /* 2513 * Treat per-device exclusion ranges as r/w unity-mapped regions 2514 * since some buggy BIOSes might lead to the overwritten exclusion 2515 * range (exclusion_start and exclusion_length members). This 2516 * happens when there are multiple exclusion ranges (IVMD entries) 2517 * defined in ACPI table. 2518 */ 2519 if (m->flags & IVMD_FLAG_EXCL_RANGE) 2520 e->prot = (IVMD_FLAG_IW | IVMD_FLAG_IR) >> 1; 2521 2522 DUMP_printk("%s devid_start: %04x:%02x:%02x.%x devid_end: " 2523 "%04x:%02x:%02x.%x range_start: %016llx range_end: %016llx" 2524 " flags: %x\n", s, m->pci_seg, 2525 PCI_BUS_NUM(e->devid_start), PCI_SLOT(e->devid_start), 2526 PCI_FUNC(e->devid_start), m->pci_seg, 2527 PCI_BUS_NUM(e->devid_end), 2528 PCI_SLOT(e->devid_end), PCI_FUNC(e->devid_end), 2529 e->address_start, e->address_end, m->flags); 2530 2531 list_add_tail(&e->list, &pci_seg->unity_map); 2532 2533 return 0; 2534 } 2535 2536 /* iterates over all memory definitions we find in the ACPI table */ 2537 static int __init init_memory_definitions(struct acpi_table_header *table) 2538 { 2539 u8 *p = (u8 *)table, *end = (u8 *)table; 2540 struct ivmd_header *m; 2541 2542 end += table->length; 2543 p += IVRS_HEADER_LENGTH; 2544 2545 while (p < end) { 2546 m = (struct ivmd_header *)p; 2547 if (m->flags & (IVMD_FLAG_UNITY_MAP | IVMD_FLAG_EXCL_RANGE)) 2548 init_unity_map_range(m, table); 2549 2550 p += m->length; 2551 } 2552 2553 return 0; 2554 } 2555 2556 /* 2557 * Init the device table to not allow DMA access for devices 2558 */ 2559 static void init_device_table_dma(struct amd_iommu_pci_seg *pci_seg) 2560 { 2561 u32 devid; 2562 struct dev_table_entry *dev_table = pci_seg->dev_table; 2563 2564 if (dev_table == NULL) 2565 return; 2566 2567 for (devid = 0; devid <= pci_seg->last_bdf; ++devid) { 2568 __set_dev_entry_bit(dev_table, devid, DEV_ENTRY_VALID); 2569 if (!amd_iommu_snp_en) 2570 __set_dev_entry_bit(dev_table, devid, DEV_ENTRY_TRANSLATION); 2571 } 2572 } 2573 2574 static void __init uninit_device_table_dma(struct amd_iommu_pci_seg *pci_seg) 2575 { 2576 u32 devid; 2577 struct dev_table_entry *dev_table = pci_seg->dev_table; 2578 2579 if (dev_table == NULL) 2580 return; 2581 2582 for (devid = 0; devid <= pci_seg->last_bdf; ++devid) { 2583 dev_table[devid].data[0] = 0ULL; 2584 dev_table[devid].data[1] = 0ULL; 2585 } 2586 } 2587 2588 static void init_device_table(void) 2589 { 2590 struct amd_iommu_pci_seg *pci_seg; 2591 u32 devid; 2592 2593 if (!amd_iommu_irq_remap) 2594 return; 2595 2596 for_each_pci_segment(pci_seg) { 2597 for (devid = 0; devid <= pci_seg->last_bdf; ++devid) 2598 __set_dev_entry_bit(pci_seg->dev_table, 2599 devid, DEV_ENTRY_IRQ_TBL_EN); 2600 } 2601 } 2602 2603 static void iommu_init_flags(struct amd_iommu *iommu) 2604 { 2605 iommu->acpi_flags & IVHD_FLAG_HT_TUN_EN_MASK ? 2606 iommu_feature_enable(iommu, CONTROL_HT_TUN_EN) : 2607 iommu_feature_disable(iommu, CONTROL_HT_TUN_EN); 2608 2609 iommu->acpi_flags & IVHD_FLAG_PASSPW_EN_MASK ? 2610 iommu_feature_enable(iommu, CONTROL_PASSPW_EN) : 2611 iommu_feature_disable(iommu, CONTROL_PASSPW_EN); 2612 2613 iommu->acpi_flags & IVHD_FLAG_RESPASSPW_EN_MASK ? 2614 iommu_feature_enable(iommu, CONTROL_RESPASSPW_EN) : 2615 iommu_feature_disable(iommu, CONTROL_RESPASSPW_EN); 2616 2617 iommu->acpi_flags & IVHD_FLAG_ISOC_EN_MASK ? 2618 iommu_feature_enable(iommu, CONTROL_ISOC_EN) : 2619 iommu_feature_disable(iommu, CONTROL_ISOC_EN); 2620 2621 /* 2622 * make IOMMU memory accesses cache coherent 2623 */ 2624 iommu_feature_enable(iommu, CONTROL_COHERENT_EN); 2625 2626 /* Set IOTLB invalidation timeout to 1s */ 2627 iommu_set_inv_tlb_timeout(iommu, CTRL_INV_TO_1S); 2628 } 2629 2630 static void iommu_apply_resume_quirks(struct amd_iommu *iommu) 2631 { 2632 int i, j; 2633 u32 ioc_feature_control; 2634 struct pci_dev *pdev = iommu->root_pdev; 2635 2636 /* RD890 BIOSes may not have completely reconfigured the iommu */ 2637 if (!is_rd890_iommu(iommu->dev) || !pdev) 2638 return; 2639 2640 /* 2641 * First, we need to ensure that the iommu is enabled. This is 2642 * controlled by a register in the northbridge 2643 */ 2644 2645 /* Select Northbridge indirect register 0x75 and enable writing */ 2646 pci_write_config_dword(pdev, 0x60, 0x75 | (1 << 7)); 2647 pci_read_config_dword(pdev, 0x64, &ioc_feature_control); 2648 2649 /* Enable the iommu */ 2650 if (!(ioc_feature_control & 0x1)) 2651 pci_write_config_dword(pdev, 0x64, ioc_feature_control | 1); 2652 2653 /* Restore the iommu BAR */ 2654 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4, 2655 iommu->stored_addr_lo); 2656 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 8, 2657 iommu->stored_addr_hi); 2658 2659 /* Restore the l1 indirect regs for each of the 6 l1s */ 2660 for (i = 0; i < 6; i++) 2661 for (j = 0; j < 0x12; j++) 2662 iommu_write_l1(iommu, i, j, iommu->stored_l1[i][j]); 2663 2664 /* Restore the l2 indirect regs */ 2665 for (i = 0; i < 0x83; i++) 2666 iommu_write_l2(iommu, i, iommu->stored_l2[i]); 2667 2668 /* Lock PCI setup registers */ 2669 pci_write_config_dword(iommu->dev, iommu->cap_ptr + 4, 2670 iommu->stored_addr_lo | 1); 2671 } 2672 2673 static void iommu_enable_ga(struct amd_iommu *iommu) 2674 { 2675 #ifdef CONFIG_IRQ_REMAP 2676 switch (amd_iommu_guest_ir) { 2677 case AMD_IOMMU_GUEST_IR_VAPIC: 2678 case AMD_IOMMU_GUEST_IR_LEGACY_GA: 2679 iommu_feature_enable(iommu, CONTROL_GA_EN); 2680 iommu->irte_ops = &irte_128_ops; 2681 break; 2682 default: 2683 iommu->irte_ops = &irte_32_ops; 2684 break; 2685 } 2686 #endif 2687 } 2688 2689 static void early_enable_iommu(struct amd_iommu *iommu) 2690 { 2691 iommu_disable(iommu); 2692 iommu_init_flags(iommu); 2693 iommu_set_device_table(iommu); 2694 iommu_enable_command_buffer(iommu); 2695 iommu_enable_event_buffer(iommu); 2696 iommu_set_exclusion_range(iommu); 2697 iommu_enable_ga(iommu); 2698 iommu_enable_xt(iommu); 2699 iommu_enable(iommu); 2700 iommu_flush_all_caches(iommu); 2701 } 2702 2703 /* 2704 * This function finally enables all IOMMUs found in the system after 2705 * they have been initialized. 2706 * 2707 * Or if in kdump kernel and IOMMUs are all pre-enabled, try to copy 2708 * the old content of device table entries. Not this case or copy failed, 2709 * just continue as normal kernel does. 2710 */ 2711 static void early_enable_iommus(void) 2712 { 2713 struct amd_iommu *iommu; 2714 struct amd_iommu_pci_seg *pci_seg; 2715 2716 if (!copy_device_table()) { 2717 /* 2718 * If come here because of failure in copying device table from old 2719 * kernel with all IOMMUs enabled, print error message and try to 2720 * free allocated old_dev_tbl_cpy. 2721 */ 2722 if (amd_iommu_pre_enabled) 2723 pr_err("Failed to copy DEV table from previous kernel.\n"); 2724 2725 for_each_pci_segment(pci_seg) { 2726 if (pci_seg->old_dev_tbl_cpy != NULL) { 2727 free_pages((unsigned long)pci_seg->old_dev_tbl_cpy, 2728 get_order(pci_seg->dev_table_size)); 2729 pci_seg->old_dev_tbl_cpy = NULL; 2730 } 2731 } 2732 2733 for_each_iommu(iommu) { 2734 clear_translation_pre_enabled(iommu); 2735 early_enable_iommu(iommu); 2736 } 2737 } else { 2738 pr_info("Copied DEV table from previous kernel.\n"); 2739 2740 for_each_pci_segment(pci_seg) { 2741 free_pages((unsigned long)pci_seg->dev_table, 2742 get_order(pci_seg->dev_table_size)); 2743 pci_seg->dev_table = pci_seg->old_dev_tbl_cpy; 2744 } 2745 2746 for_each_iommu(iommu) { 2747 iommu_disable_command_buffer(iommu); 2748 iommu_disable_event_buffer(iommu); 2749 iommu_enable_command_buffer(iommu); 2750 iommu_enable_event_buffer(iommu); 2751 iommu_enable_ga(iommu); 2752 iommu_enable_xt(iommu); 2753 iommu_set_device_table(iommu); 2754 iommu_flush_all_caches(iommu); 2755 } 2756 } 2757 } 2758 2759 static void enable_iommus_v2(void) 2760 { 2761 struct amd_iommu *iommu; 2762 2763 for_each_iommu(iommu) { 2764 iommu_enable_ppr_log(iommu); 2765 iommu_enable_gt(iommu); 2766 } 2767 } 2768 2769 static void enable_iommus_vapic(void) 2770 { 2771 #ifdef CONFIG_IRQ_REMAP 2772 u32 status, i; 2773 struct amd_iommu *iommu; 2774 2775 for_each_iommu(iommu) { 2776 /* 2777 * Disable GALog if already running. It could have been enabled 2778 * in the previous boot before kdump. 2779 */ 2780 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); 2781 if (!(status & MMIO_STATUS_GALOG_RUN_MASK)) 2782 continue; 2783 2784 iommu_feature_disable(iommu, CONTROL_GALOG_EN); 2785 iommu_feature_disable(iommu, CONTROL_GAINT_EN); 2786 2787 /* 2788 * Need to set and poll check the GALOGRun bit to zero before 2789 * we can set/ modify GA Log registers safely. 2790 */ 2791 for (i = 0; i < LOOP_TIMEOUT; ++i) { 2792 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); 2793 if (!(status & MMIO_STATUS_GALOG_RUN_MASK)) 2794 break; 2795 udelay(10); 2796 } 2797 2798 if (WARN_ON(i >= LOOP_TIMEOUT)) 2799 return; 2800 } 2801 2802 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) && 2803 !check_feature_on_all_iommus(FEATURE_GAM_VAPIC)) { 2804 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY_GA; 2805 return; 2806 } 2807 2808 if (amd_iommu_snp_en && 2809 !FEATURE_SNPAVICSUP_GAM(amd_iommu_efr2)) { 2810 pr_warn("Force to disable Virtual APIC due to SNP\n"); 2811 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY_GA; 2812 return; 2813 } 2814 2815 /* Enabling GAM and SNPAVIC support */ 2816 for_each_iommu(iommu) { 2817 if (iommu_init_ga_log(iommu) || 2818 iommu_ga_log_enable(iommu)) 2819 return; 2820 2821 iommu_feature_enable(iommu, CONTROL_GAM_EN); 2822 if (amd_iommu_snp_en) 2823 iommu_feature_enable(iommu, CONTROL_SNPAVIC_EN); 2824 } 2825 2826 amd_iommu_irq_ops.capability |= (1 << IRQ_POSTING_CAP); 2827 pr_info("Virtual APIC enabled\n"); 2828 #endif 2829 } 2830 2831 static void enable_iommus(void) 2832 { 2833 early_enable_iommus(); 2834 enable_iommus_vapic(); 2835 enable_iommus_v2(); 2836 } 2837 2838 static void disable_iommus(void) 2839 { 2840 struct amd_iommu *iommu; 2841 2842 for_each_iommu(iommu) 2843 iommu_disable(iommu); 2844 2845 #ifdef CONFIG_IRQ_REMAP 2846 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) 2847 amd_iommu_irq_ops.capability &= ~(1 << IRQ_POSTING_CAP); 2848 #endif 2849 } 2850 2851 /* 2852 * Suspend/Resume support 2853 * disable suspend until real resume implemented 2854 */ 2855 2856 static void amd_iommu_resume(void) 2857 { 2858 struct amd_iommu *iommu; 2859 2860 for_each_iommu(iommu) 2861 iommu_apply_resume_quirks(iommu); 2862 2863 /* re-load the hardware */ 2864 enable_iommus(); 2865 2866 amd_iommu_enable_interrupts(); 2867 } 2868 2869 static int amd_iommu_suspend(void) 2870 { 2871 /* disable IOMMUs to go out of the way for BIOS */ 2872 disable_iommus(); 2873 2874 return 0; 2875 } 2876 2877 static struct syscore_ops amd_iommu_syscore_ops = { 2878 .suspend = amd_iommu_suspend, 2879 .resume = amd_iommu_resume, 2880 }; 2881 2882 static void __init free_iommu_resources(void) 2883 { 2884 kmem_cache_destroy(amd_iommu_irq_cache); 2885 amd_iommu_irq_cache = NULL; 2886 2887 free_iommu_all(); 2888 free_pci_segments(); 2889 } 2890 2891 /* SB IOAPIC is always on this device in AMD systems */ 2892 #define IOAPIC_SB_DEVID ((0x00 << 8) | PCI_DEVFN(0x14, 0)) 2893 2894 static bool __init check_ioapic_information(void) 2895 { 2896 const char *fw_bug = FW_BUG; 2897 bool ret, has_sb_ioapic; 2898 int idx; 2899 2900 has_sb_ioapic = false; 2901 ret = false; 2902 2903 /* 2904 * If we have map overrides on the kernel command line the 2905 * messages in this function might not describe firmware bugs 2906 * anymore - so be careful 2907 */ 2908 if (cmdline_maps) 2909 fw_bug = ""; 2910 2911 for (idx = 0; idx < nr_ioapics; idx++) { 2912 int devid, id = mpc_ioapic_id(idx); 2913 2914 devid = get_ioapic_devid(id); 2915 if (devid < 0) { 2916 pr_err("%s: IOAPIC[%d] not in IVRS table\n", 2917 fw_bug, id); 2918 ret = false; 2919 } else if (devid == IOAPIC_SB_DEVID) { 2920 has_sb_ioapic = true; 2921 ret = true; 2922 } 2923 } 2924 2925 if (!has_sb_ioapic) { 2926 /* 2927 * We expect the SB IOAPIC to be listed in the IVRS 2928 * table. The system timer is connected to the SB IOAPIC 2929 * and if we don't have it in the list the system will 2930 * panic at boot time. This situation usually happens 2931 * when the BIOS is buggy and provides us the wrong 2932 * device id for the IOAPIC in the system. 2933 */ 2934 pr_err("%s: No southbridge IOAPIC found\n", fw_bug); 2935 } 2936 2937 if (!ret) 2938 pr_err("Disabling interrupt remapping\n"); 2939 2940 return ret; 2941 } 2942 2943 static void __init free_dma_resources(void) 2944 { 2945 free_pages((unsigned long)amd_iommu_pd_alloc_bitmap, 2946 get_order(MAX_DOMAIN_ID/8)); 2947 amd_iommu_pd_alloc_bitmap = NULL; 2948 2949 free_unity_maps(); 2950 } 2951 2952 static void __init ivinfo_init(void *ivrs) 2953 { 2954 amd_iommu_ivinfo = *((u32 *)(ivrs + IOMMU_IVINFO_OFFSET)); 2955 } 2956 2957 /* 2958 * This is the hardware init function for AMD IOMMU in the system. 2959 * This function is called either from amd_iommu_init or from the interrupt 2960 * remapping setup code. 2961 * 2962 * This function basically parses the ACPI table for AMD IOMMU (IVRS) 2963 * four times: 2964 * 2965 * 1 pass) Discover the most comprehensive IVHD type to use. 2966 * 2967 * 2 pass) Find the highest PCI device id the driver has to handle. 2968 * Upon this information the size of the data structures is 2969 * determined that needs to be allocated. 2970 * 2971 * 3 pass) Initialize the data structures just allocated with the 2972 * information in the ACPI table about available AMD IOMMUs 2973 * in the system. It also maps the PCI devices in the 2974 * system to specific IOMMUs 2975 * 2976 * 4 pass) After the basic data structures are allocated and 2977 * initialized we update them with information about memory 2978 * remapping requirements parsed out of the ACPI table in 2979 * this last pass. 2980 * 2981 * After everything is set up the IOMMUs are enabled and the necessary 2982 * hotplug and suspend notifiers are registered. 2983 */ 2984 static int __init early_amd_iommu_init(void) 2985 { 2986 struct acpi_table_header *ivrs_base; 2987 int remap_cache_sz, ret; 2988 acpi_status status; 2989 2990 if (!amd_iommu_detected) 2991 return -ENODEV; 2992 2993 status = acpi_get_table("IVRS", 0, &ivrs_base); 2994 if (status == AE_NOT_FOUND) 2995 return -ENODEV; 2996 else if (ACPI_FAILURE(status)) { 2997 const char *err = acpi_format_exception(status); 2998 pr_err("IVRS table error: %s\n", err); 2999 return -EINVAL; 3000 } 3001 3002 /* 3003 * Validate checksum here so we don't need to do it when 3004 * we actually parse the table 3005 */ 3006 ret = check_ivrs_checksum(ivrs_base); 3007 if (ret) 3008 goto out; 3009 3010 ivinfo_init(ivrs_base); 3011 3012 amd_iommu_target_ivhd_type = get_highest_supported_ivhd_type(ivrs_base); 3013 DUMP_printk("Using IVHD type %#x\n", amd_iommu_target_ivhd_type); 3014 3015 /* Device table - directly used by all IOMMUs */ 3016 ret = -ENOMEM; 3017 3018 amd_iommu_pd_alloc_bitmap = (void *)__get_free_pages( 3019 GFP_KERNEL | __GFP_ZERO, 3020 get_order(MAX_DOMAIN_ID/8)); 3021 if (amd_iommu_pd_alloc_bitmap == NULL) 3022 goto out; 3023 3024 /* 3025 * never allocate domain 0 because its used as the non-allocated and 3026 * error value placeholder 3027 */ 3028 __set_bit(0, amd_iommu_pd_alloc_bitmap); 3029 3030 /* 3031 * now the data structures are allocated and basically initialized 3032 * start the real acpi table scan 3033 */ 3034 ret = init_iommu_all(ivrs_base); 3035 if (ret) 3036 goto out; 3037 3038 /* 5 level guest page table */ 3039 if (cpu_feature_enabled(X86_FEATURE_LA57) && 3040 check_feature_gpt_level() == GUEST_PGTABLE_5_LEVEL) 3041 amd_iommu_gpt_level = PAGE_MODE_5_LEVEL; 3042 3043 /* Disable any previously enabled IOMMUs */ 3044 if (!is_kdump_kernel() || amd_iommu_disabled) 3045 disable_iommus(); 3046 3047 if (amd_iommu_irq_remap) 3048 amd_iommu_irq_remap = check_ioapic_information(); 3049 3050 if (amd_iommu_irq_remap) { 3051 struct amd_iommu_pci_seg *pci_seg; 3052 /* 3053 * Interrupt remapping enabled, create kmem_cache for the 3054 * remapping tables. 3055 */ 3056 ret = -ENOMEM; 3057 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir)) 3058 remap_cache_sz = MAX_IRQS_PER_TABLE * sizeof(u32); 3059 else 3060 remap_cache_sz = MAX_IRQS_PER_TABLE * (sizeof(u64) * 2); 3061 amd_iommu_irq_cache = kmem_cache_create("irq_remap_cache", 3062 remap_cache_sz, 3063 DTE_INTTAB_ALIGNMENT, 3064 0, NULL); 3065 if (!amd_iommu_irq_cache) 3066 goto out; 3067 3068 for_each_pci_segment(pci_seg) { 3069 if (alloc_irq_lookup_table(pci_seg)) 3070 goto out; 3071 } 3072 } 3073 3074 ret = init_memory_definitions(ivrs_base); 3075 if (ret) 3076 goto out; 3077 3078 /* init the device table */ 3079 init_device_table(); 3080 3081 out: 3082 /* Don't leak any ACPI memory */ 3083 acpi_put_table(ivrs_base); 3084 3085 return ret; 3086 } 3087 3088 static int amd_iommu_enable_interrupts(void) 3089 { 3090 struct amd_iommu *iommu; 3091 int ret = 0; 3092 3093 for_each_iommu(iommu) { 3094 ret = iommu_init_irq(iommu); 3095 if (ret) 3096 goto out; 3097 } 3098 3099 out: 3100 return ret; 3101 } 3102 3103 static bool __init detect_ivrs(void) 3104 { 3105 struct acpi_table_header *ivrs_base; 3106 acpi_status status; 3107 int i; 3108 3109 status = acpi_get_table("IVRS", 0, &ivrs_base); 3110 if (status == AE_NOT_FOUND) 3111 return false; 3112 else if (ACPI_FAILURE(status)) { 3113 const char *err = acpi_format_exception(status); 3114 pr_err("IVRS table error: %s\n", err); 3115 return false; 3116 } 3117 3118 acpi_put_table(ivrs_base); 3119 3120 if (amd_iommu_force_enable) 3121 goto out; 3122 3123 /* Don't use IOMMU if there is Stoney Ridge graphics */ 3124 for (i = 0; i < 32; i++) { 3125 u32 pci_id; 3126 3127 pci_id = read_pci_config(0, i, 0, 0); 3128 if ((pci_id & 0xffff) == 0x1002 && (pci_id >> 16) == 0x98e4) { 3129 pr_info("Disable IOMMU on Stoney Ridge\n"); 3130 return false; 3131 } 3132 } 3133 3134 out: 3135 /* Make sure ACS will be enabled during PCI probe */ 3136 pci_request_acs(); 3137 3138 return true; 3139 } 3140 3141 /**************************************************************************** 3142 * 3143 * AMD IOMMU Initialization State Machine 3144 * 3145 ****************************************************************************/ 3146 3147 static int __init state_next(void) 3148 { 3149 int ret = 0; 3150 3151 switch (init_state) { 3152 case IOMMU_START_STATE: 3153 if (!detect_ivrs()) { 3154 init_state = IOMMU_NOT_FOUND; 3155 ret = -ENODEV; 3156 } else { 3157 init_state = IOMMU_IVRS_DETECTED; 3158 } 3159 break; 3160 case IOMMU_IVRS_DETECTED: 3161 if (amd_iommu_disabled) { 3162 init_state = IOMMU_CMDLINE_DISABLED; 3163 ret = -EINVAL; 3164 } else { 3165 ret = early_amd_iommu_init(); 3166 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_ACPI_FINISHED; 3167 } 3168 break; 3169 case IOMMU_ACPI_FINISHED: 3170 early_enable_iommus(); 3171 x86_platform.iommu_shutdown = disable_iommus; 3172 init_state = IOMMU_ENABLED; 3173 break; 3174 case IOMMU_ENABLED: 3175 register_syscore_ops(&amd_iommu_syscore_ops); 3176 ret = amd_iommu_init_pci(); 3177 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_PCI_INIT; 3178 enable_iommus_vapic(); 3179 enable_iommus_v2(); 3180 break; 3181 case IOMMU_PCI_INIT: 3182 ret = amd_iommu_enable_interrupts(); 3183 init_state = ret ? IOMMU_INIT_ERROR : IOMMU_INTERRUPTS_EN; 3184 break; 3185 case IOMMU_INTERRUPTS_EN: 3186 init_state = IOMMU_INITIALIZED; 3187 break; 3188 case IOMMU_INITIALIZED: 3189 /* Nothing to do */ 3190 break; 3191 case IOMMU_NOT_FOUND: 3192 case IOMMU_INIT_ERROR: 3193 case IOMMU_CMDLINE_DISABLED: 3194 /* Error states => do nothing */ 3195 ret = -EINVAL; 3196 break; 3197 default: 3198 /* Unknown state */ 3199 BUG(); 3200 } 3201 3202 if (ret) { 3203 free_dma_resources(); 3204 if (!irq_remapping_enabled) { 3205 disable_iommus(); 3206 free_iommu_resources(); 3207 } else { 3208 struct amd_iommu *iommu; 3209 struct amd_iommu_pci_seg *pci_seg; 3210 3211 for_each_pci_segment(pci_seg) 3212 uninit_device_table_dma(pci_seg); 3213 3214 for_each_iommu(iommu) 3215 iommu_flush_all_caches(iommu); 3216 } 3217 } 3218 return ret; 3219 } 3220 3221 static int __init iommu_go_to_state(enum iommu_init_state state) 3222 { 3223 int ret = -EINVAL; 3224 3225 while (init_state != state) { 3226 if (init_state == IOMMU_NOT_FOUND || 3227 init_state == IOMMU_INIT_ERROR || 3228 init_state == IOMMU_CMDLINE_DISABLED) 3229 break; 3230 ret = state_next(); 3231 } 3232 3233 return ret; 3234 } 3235 3236 #ifdef CONFIG_IRQ_REMAP 3237 int __init amd_iommu_prepare(void) 3238 { 3239 int ret; 3240 3241 amd_iommu_irq_remap = true; 3242 3243 ret = iommu_go_to_state(IOMMU_ACPI_FINISHED); 3244 if (ret) { 3245 amd_iommu_irq_remap = false; 3246 return ret; 3247 } 3248 3249 return amd_iommu_irq_remap ? 0 : -ENODEV; 3250 } 3251 3252 int __init amd_iommu_enable(void) 3253 { 3254 int ret; 3255 3256 ret = iommu_go_to_state(IOMMU_ENABLED); 3257 if (ret) 3258 return ret; 3259 3260 irq_remapping_enabled = 1; 3261 return amd_iommu_xt_mode; 3262 } 3263 3264 void amd_iommu_disable(void) 3265 { 3266 amd_iommu_suspend(); 3267 } 3268 3269 int amd_iommu_reenable(int mode) 3270 { 3271 amd_iommu_resume(); 3272 3273 return 0; 3274 } 3275 3276 int __init amd_iommu_enable_faulting(void) 3277 { 3278 /* We enable MSI later when PCI is initialized */ 3279 return 0; 3280 } 3281 #endif 3282 3283 /* 3284 * This is the core init function for AMD IOMMU hardware in the system. 3285 * This function is called from the generic x86 DMA layer initialization 3286 * code. 3287 */ 3288 static int __init amd_iommu_init(void) 3289 { 3290 struct amd_iommu *iommu; 3291 int ret; 3292 3293 ret = iommu_go_to_state(IOMMU_INITIALIZED); 3294 #ifdef CONFIG_GART_IOMMU 3295 if (ret && list_empty(&amd_iommu_list)) { 3296 /* 3297 * We failed to initialize the AMD IOMMU - try fallback 3298 * to GART if possible. 3299 */ 3300 gart_iommu_init(); 3301 } 3302 #endif 3303 3304 for_each_iommu(iommu) 3305 amd_iommu_debugfs_setup(iommu); 3306 3307 return ret; 3308 } 3309 3310 static bool amd_iommu_sme_check(void) 3311 { 3312 if (!cc_platform_has(CC_ATTR_HOST_MEM_ENCRYPT) || 3313 (boot_cpu_data.x86 != 0x17)) 3314 return true; 3315 3316 /* For Fam17h, a specific level of support is required */ 3317 if (boot_cpu_data.microcode >= 0x08001205) 3318 return true; 3319 3320 if ((boot_cpu_data.microcode >= 0x08001126) && 3321 (boot_cpu_data.microcode <= 0x080011ff)) 3322 return true; 3323 3324 pr_notice("IOMMU not currently supported when SME is active\n"); 3325 3326 return false; 3327 } 3328 3329 /**************************************************************************** 3330 * 3331 * Early detect code. This code runs at IOMMU detection time in the DMA 3332 * layer. It just looks if there is an IVRS ACPI table to detect AMD 3333 * IOMMUs 3334 * 3335 ****************************************************************************/ 3336 int __init amd_iommu_detect(void) 3337 { 3338 int ret; 3339 3340 if (no_iommu || (iommu_detected && !gart_iommu_aperture)) 3341 return -ENODEV; 3342 3343 if (!amd_iommu_sme_check()) 3344 return -ENODEV; 3345 3346 ret = iommu_go_to_state(IOMMU_IVRS_DETECTED); 3347 if (ret) 3348 return ret; 3349 3350 amd_iommu_detected = true; 3351 iommu_detected = 1; 3352 x86_init.iommu.iommu_init = amd_iommu_init; 3353 3354 return 1; 3355 } 3356 3357 /**************************************************************************** 3358 * 3359 * Parsing functions for the AMD IOMMU specific kernel command line 3360 * options. 3361 * 3362 ****************************************************************************/ 3363 3364 static int __init parse_amd_iommu_dump(char *str) 3365 { 3366 amd_iommu_dump = true; 3367 3368 return 1; 3369 } 3370 3371 static int __init parse_amd_iommu_intr(char *str) 3372 { 3373 for (; *str; ++str) { 3374 if (strncmp(str, "legacy", 6) == 0) { 3375 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_LEGACY_GA; 3376 break; 3377 } 3378 if (strncmp(str, "vapic", 5) == 0) { 3379 amd_iommu_guest_ir = AMD_IOMMU_GUEST_IR_VAPIC; 3380 break; 3381 } 3382 } 3383 return 1; 3384 } 3385 3386 static int __init parse_amd_iommu_options(char *str) 3387 { 3388 if (!str) 3389 return -EINVAL; 3390 3391 while (*str) { 3392 if (strncmp(str, "fullflush", 9) == 0) { 3393 pr_warn("amd_iommu=fullflush deprecated; use iommu.strict=1 instead\n"); 3394 iommu_set_dma_strict(); 3395 } else if (strncmp(str, "force_enable", 12) == 0) { 3396 amd_iommu_force_enable = true; 3397 } else if (strncmp(str, "off", 3) == 0) { 3398 amd_iommu_disabled = true; 3399 } else if (strncmp(str, "force_isolation", 15) == 0) { 3400 amd_iommu_force_isolation = true; 3401 } else if (strncmp(str, "pgtbl_v1", 8) == 0) { 3402 amd_iommu_pgtable = AMD_IOMMU_V1; 3403 } else if (strncmp(str, "pgtbl_v2", 8) == 0) { 3404 amd_iommu_pgtable = AMD_IOMMU_V2; 3405 } else { 3406 pr_notice("Unknown option - '%s'\n", str); 3407 } 3408 3409 str += strcspn(str, ","); 3410 while (*str == ',') 3411 str++; 3412 } 3413 3414 return 1; 3415 } 3416 3417 static int __init parse_ivrs_ioapic(char *str) 3418 { 3419 u32 seg = 0, bus, dev, fn; 3420 int id, i; 3421 u32 devid; 3422 3423 if (sscanf(str, "=%d@%x:%x.%x", &id, &bus, &dev, &fn) == 4 || 3424 sscanf(str, "=%d@%x:%x:%x.%x", &id, &seg, &bus, &dev, &fn) == 5) 3425 goto found; 3426 3427 if (sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn) == 4 || 3428 sscanf(str, "[%d]=%x:%x:%x.%x", &id, &seg, &bus, &dev, &fn) == 5) { 3429 pr_warn("ivrs_ioapic%s option format deprecated; use ivrs_ioapic=%d@%04x:%02x:%02x.%d instead\n", 3430 str, id, seg, bus, dev, fn); 3431 goto found; 3432 } 3433 3434 pr_err("Invalid command line: ivrs_ioapic%s\n", str); 3435 return 1; 3436 3437 found: 3438 if (early_ioapic_map_size == EARLY_MAP_SIZE) { 3439 pr_err("Early IOAPIC map overflow - ignoring ivrs_ioapic%s\n", 3440 str); 3441 return 1; 3442 } 3443 3444 devid = IVRS_GET_SBDF_ID(seg, bus, dev, fn); 3445 3446 cmdline_maps = true; 3447 i = early_ioapic_map_size++; 3448 early_ioapic_map[i].id = id; 3449 early_ioapic_map[i].devid = devid; 3450 early_ioapic_map[i].cmd_line = true; 3451 3452 return 1; 3453 } 3454 3455 static int __init parse_ivrs_hpet(char *str) 3456 { 3457 u32 seg = 0, bus, dev, fn; 3458 int id, i; 3459 u32 devid; 3460 3461 if (sscanf(str, "=%d@%x:%x.%x", &id, &bus, &dev, &fn) == 4 || 3462 sscanf(str, "=%d@%x:%x:%x.%x", &id, &seg, &bus, &dev, &fn) == 5) 3463 goto found; 3464 3465 if (sscanf(str, "[%d]=%x:%x.%x", &id, &bus, &dev, &fn) == 4 || 3466 sscanf(str, "[%d]=%x:%x:%x.%x", &id, &seg, &bus, &dev, &fn) == 5) { 3467 pr_warn("ivrs_hpet%s option format deprecated; use ivrs_hpet=%d@%04x:%02x:%02x.%d instead\n", 3468 str, id, seg, bus, dev, fn); 3469 goto found; 3470 } 3471 3472 pr_err("Invalid command line: ivrs_hpet%s\n", str); 3473 return 1; 3474 3475 found: 3476 if (early_hpet_map_size == EARLY_MAP_SIZE) { 3477 pr_err("Early HPET map overflow - ignoring ivrs_hpet%s\n", 3478 str); 3479 return 1; 3480 } 3481 3482 devid = IVRS_GET_SBDF_ID(seg, bus, dev, fn); 3483 3484 cmdline_maps = true; 3485 i = early_hpet_map_size++; 3486 early_hpet_map[i].id = id; 3487 early_hpet_map[i].devid = devid; 3488 early_hpet_map[i].cmd_line = true; 3489 3490 return 1; 3491 } 3492 3493 #define ACPIID_LEN (ACPIHID_UID_LEN + ACPIHID_HID_LEN) 3494 3495 static int __init parse_ivrs_acpihid(char *str) 3496 { 3497 u32 seg = 0, bus, dev, fn; 3498 char *hid, *uid, *p, *addr; 3499 char acpiid[ACPIID_LEN] = {0}; 3500 int i; 3501 3502 addr = strchr(str, '@'); 3503 if (!addr) { 3504 addr = strchr(str, '='); 3505 if (!addr) 3506 goto not_found; 3507 3508 ++addr; 3509 3510 if (strlen(addr) > ACPIID_LEN) 3511 goto not_found; 3512 3513 if (sscanf(str, "[%x:%x.%x]=%s", &bus, &dev, &fn, acpiid) == 4 || 3514 sscanf(str, "[%x:%x:%x.%x]=%s", &seg, &bus, &dev, &fn, acpiid) == 5) { 3515 pr_warn("ivrs_acpihid%s option format deprecated; use ivrs_acpihid=%s@%04x:%02x:%02x.%d instead\n", 3516 str, acpiid, seg, bus, dev, fn); 3517 goto found; 3518 } 3519 goto not_found; 3520 } 3521 3522 /* We have the '@', make it the terminator to get just the acpiid */ 3523 *addr++ = 0; 3524 3525 if (strlen(str) > ACPIID_LEN + 1) 3526 goto not_found; 3527 3528 if (sscanf(str, "=%s", acpiid) != 1) 3529 goto not_found; 3530 3531 if (sscanf(addr, "%x:%x.%x", &bus, &dev, &fn) == 3 || 3532 sscanf(addr, "%x:%x:%x.%x", &seg, &bus, &dev, &fn) == 4) 3533 goto found; 3534 3535 not_found: 3536 pr_err("Invalid command line: ivrs_acpihid%s\n", str); 3537 return 1; 3538 3539 found: 3540 p = acpiid; 3541 hid = strsep(&p, ":"); 3542 uid = p; 3543 3544 if (!hid || !(*hid) || !uid) { 3545 pr_err("Invalid command line: hid or uid\n"); 3546 return 1; 3547 } 3548 3549 /* 3550 * Ignore leading zeroes after ':', so e.g., AMDI0095:00 3551 * will match AMDI0095:0 in the second strcmp in acpi_dev_hid_uid_match 3552 */ 3553 while (*uid == '0' && *(uid + 1)) 3554 uid++; 3555 3556 i = early_acpihid_map_size++; 3557 memcpy(early_acpihid_map[i].hid, hid, strlen(hid)); 3558 memcpy(early_acpihid_map[i].uid, uid, strlen(uid)); 3559 early_acpihid_map[i].devid = IVRS_GET_SBDF_ID(seg, bus, dev, fn); 3560 early_acpihid_map[i].cmd_line = true; 3561 3562 return 1; 3563 } 3564 3565 __setup("amd_iommu_dump", parse_amd_iommu_dump); 3566 __setup("amd_iommu=", parse_amd_iommu_options); 3567 __setup("amd_iommu_intr=", parse_amd_iommu_intr); 3568 __setup("ivrs_ioapic", parse_ivrs_ioapic); 3569 __setup("ivrs_hpet", parse_ivrs_hpet); 3570 __setup("ivrs_acpihid", parse_ivrs_acpihid); 3571 3572 bool amd_iommu_v2_supported(void) 3573 { 3574 /* CPU page table size should match IOMMU guest page table size */ 3575 if (cpu_feature_enabled(X86_FEATURE_LA57) && 3576 amd_iommu_gpt_level != PAGE_MODE_5_LEVEL) 3577 return false; 3578 3579 /* 3580 * Since DTE[Mode]=0 is prohibited on SNP-enabled system 3581 * (i.e. EFR[SNPSup]=1), IOMMUv2 page table cannot be used without 3582 * setting up IOMMUv1 page table. 3583 */ 3584 return amd_iommu_v2_present && !amd_iommu_snp_en; 3585 } 3586 EXPORT_SYMBOL(amd_iommu_v2_supported); 3587 3588 struct amd_iommu *get_amd_iommu(unsigned int idx) 3589 { 3590 unsigned int i = 0; 3591 struct amd_iommu *iommu; 3592 3593 for_each_iommu(iommu) 3594 if (i++ == idx) 3595 return iommu; 3596 return NULL; 3597 } 3598 3599 /**************************************************************************** 3600 * 3601 * IOMMU EFR Performance Counter support functionality. This code allows 3602 * access to the IOMMU PC functionality. 3603 * 3604 ****************************************************************************/ 3605 3606 u8 amd_iommu_pc_get_max_banks(unsigned int idx) 3607 { 3608 struct amd_iommu *iommu = get_amd_iommu(idx); 3609 3610 if (iommu) 3611 return iommu->max_banks; 3612 3613 return 0; 3614 } 3615 EXPORT_SYMBOL(amd_iommu_pc_get_max_banks); 3616 3617 bool amd_iommu_pc_supported(void) 3618 { 3619 return amd_iommu_pc_present; 3620 } 3621 EXPORT_SYMBOL(amd_iommu_pc_supported); 3622 3623 u8 amd_iommu_pc_get_max_counters(unsigned int idx) 3624 { 3625 struct amd_iommu *iommu = get_amd_iommu(idx); 3626 3627 if (iommu) 3628 return iommu->max_counters; 3629 3630 return 0; 3631 } 3632 EXPORT_SYMBOL(amd_iommu_pc_get_max_counters); 3633 3634 static int iommu_pc_get_set_reg(struct amd_iommu *iommu, u8 bank, u8 cntr, 3635 u8 fxn, u64 *value, bool is_write) 3636 { 3637 u32 offset; 3638 u32 max_offset_lim; 3639 3640 /* Make sure the IOMMU PC resource is available */ 3641 if (!amd_iommu_pc_present) 3642 return -ENODEV; 3643 3644 /* Check for valid iommu and pc register indexing */ 3645 if (WARN_ON(!iommu || (fxn > 0x28) || (fxn & 7))) 3646 return -ENODEV; 3647 3648 offset = (u32)(((0x40 | bank) << 12) | (cntr << 8) | fxn); 3649 3650 /* Limit the offset to the hw defined mmio region aperture */ 3651 max_offset_lim = (u32)(((0x40 | iommu->max_banks) << 12) | 3652 (iommu->max_counters << 8) | 0x28); 3653 if ((offset < MMIO_CNTR_REG_OFFSET) || 3654 (offset > max_offset_lim)) 3655 return -EINVAL; 3656 3657 if (is_write) { 3658 u64 val = *value & GENMASK_ULL(47, 0); 3659 3660 writel((u32)val, iommu->mmio_base + offset); 3661 writel((val >> 32), iommu->mmio_base + offset + 4); 3662 } else { 3663 *value = readl(iommu->mmio_base + offset + 4); 3664 *value <<= 32; 3665 *value |= readl(iommu->mmio_base + offset); 3666 *value &= GENMASK_ULL(47, 0); 3667 } 3668 3669 return 0; 3670 } 3671 3672 int amd_iommu_pc_get_reg(struct amd_iommu *iommu, u8 bank, u8 cntr, u8 fxn, u64 *value) 3673 { 3674 if (!iommu) 3675 return -EINVAL; 3676 3677 return iommu_pc_get_set_reg(iommu, bank, cntr, fxn, value, false); 3678 } 3679 3680 int amd_iommu_pc_set_reg(struct amd_iommu *iommu, u8 bank, u8 cntr, u8 fxn, u64 *value) 3681 { 3682 if (!iommu) 3683 return -EINVAL; 3684 3685 return iommu_pc_get_set_reg(iommu, bank, cntr, fxn, value, true); 3686 } 3687 3688 #ifdef CONFIG_AMD_MEM_ENCRYPT 3689 int amd_iommu_snp_enable(void) 3690 { 3691 /* 3692 * The SNP support requires that IOMMU must be enabled, and is 3693 * not configured in the passthrough mode. 3694 */ 3695 if (no_iommu || iommu_default_passthrough()) { 3696 pr_err("SNP: IOMMU is disabled or configured in passthrough mode, SNP cannot be supported"); 3697 return -EINVAL; 3698 } 3699 3700 /* 3701 * Prevent enabling SNP after IOMMU_ENABLED state because this process 3702 * affect how IOMMU driver sets up data structures and configures 3703 * IOMMU hardware. 3704 */ 3705 if (init_state > IOMMU_ENABLED) { 3706 pr_err("SNP: Too late to enable SNP for IOMMU.\n"); 3707 return -EINVAL; 3708 } 3709 3710 amd_iommu_snp_en = check_feature_on_all_iommus(FEATURE_SNP); 3711 if (!amd_iommu_snp_en) 3712 return -EINVAL; 3713 3714 pr_info("SNP enabled\n"); 3715 3716 /* Enforce IOMMU v1 pagetable when SNP is enabled. */ 3717 if (amd_iommu_pgtable != AMD_IOMMU_V1) { 3718 pr_warn("Force to using AMD IOMMU v1 page table due to SNP\n"); 3719 amd_iommu_pgtable = AMD_IOMMU_V1; 3720 } 3721 3722 return 0; 3723 } 3724 #endif 3725