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/ratelimit.h> 12 #include <linux/pci.h> 13 #include <linux/acpi.h> 14 #include <linux/pci-ats.h> 15 #include <linux/bitmap.h> 16 #include <linux/slab.h> 17 #include <linux/debugfs.h> 18 #include <linux/scatterlist.h> 19 #include <linux/dma-map-ops.h> 20 #include <linux/dma-direct.h> 21 #include <linux/iommu-helper.h> 22 #include <linux/delay.h> 23 #include <linux/amd-iommu.h> 24 #include <linux/notifier.h> 25 #include <linux/export.h> 26 #include <linux/irq.h> 27 #include <linux/msi.h> 28 #include <linux/irqdomain.h> 29 #include <linux/percpu.h> 30 #include <linux/io-pgtable.h> 31 #include <linux/cc_platform.h> 32 #include <asm/irq_remapping.h> 33 #include <asm/io_apic.h> 34 #include <asm/apic.h> 35 #include <asm/hw_irq.h> 36 #include <asm/proto.h> 37 #include <asm/iommu.h> 38 #include <asm/gart.h> 39 #include <asm/dma.h> 40 41 #include "amd_iommu.h" 42 #include "../dma-iommu.h" 43 #include "../irq_remapping.h" 44 45 #define CMD_SET_TYPE(cmd, t) ((cmd)->data[1] |= ((t) << 28)) 46 47 #define LOOP_TIMEOUT 100000 48 49 /* IO virtual address start page frame number */ 50 #define IOVA_START_PFN (1) 51 #define IOVA_PFN(addr) ((addr) >> PAGE_SHIFT) 52 53 /* Reserved IOVA ranges */ 54 #define MSI_RANGE_START (0xfee00000) 55 #define MSI_RANGE_END (0xfeefffff) 56 #define HT_RANGE_START (0xfd00000000ULL) 57 #define HT_RANGE_END (0xffffffffffULL) 58 59 #define DEFAULT_PGTABLE_LEVEL PAGE_MODE_3_LEVEL 60 61 static DEFINE_SPINLOCK(pd_bitmap_lock); 62 63 LIST_HEAD(ioapic_map); 64 LIST_HEAD(hpet_map); 65 LIST_HEAD(acpihid_map); 66 67 const struct iommu_ops amd_iommu_ops; 68 69 static ATOMIC_NOTIFIER_HEAD(ppr_notifier); 70 int amd_iommu_max_glx_val = -1; 71 72 /* 73 * general struct to manage commands send to an IOMMU 74 */ 75 struct iommu_cmd { 76 u32 data[4]; 77 }; 78 79 struct kmem_cache *amd_iommu_irq_cache; 80 81 static void detach_device(struct device *dev); 82 static int domain_enable_v2(struct protection_domain *domain, int pasids); 83 84 /**************************************************************************** 85 * 86 * Helper functions 87 * 88 ****************************************************************************/ 89 90 static inline int get_acpihid_device_id(struct device *dev, 91 struct acpihid_map_entry **entry) 92 { 93 struct acpi_device *adev = ACPI_COMPANION(dev); 94 struct acpihid_map_entry *p; 95 96 if (!adev) 97 return -ENODEV; 98 99 list_for_each_entry(p, &acpihid_map, list) { 100 if (acpi_dev_hid_uid_match(adev, p->hid, 101 p->uid[0] ? p->uid : NULL)) { 102 if (entry) 103 *entry = p; 104 return p->devid; 105 } 106 } 107 return -EINVAL; 108 } 109 110 static inline int get_device_sbdf_id(struct device *dev) 111 { 112 int sbdf; 113 114 if (dev_is_pci(dev)) 115 sbdf = get_pci_sbdf_id(to_pci_dev(dev)); 116 else 117 sbdf = get_acpihid_device_id(dev, NULL); 118 119 return sbdf; 120 } 121 122 struct dev_table_entry *get_dev_table(struct amd_iommu *iommu) 123 { 124 struct dev_table_entry *dev_table; 125 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 126 127 BUG_ON(pci_seg == NULL); 128 dev_table = pci_seg->dev_table; 129 BUG_ON(dev_table == NULL); 130 131 return dev_table; 132 } 133 134 static inline u16 get_device_segment(struct device *dev) 135 { 136 u16 seg; 137 138 if (dev_is_pci(dev)) { 139 struct pci_dev *pdev = to_pci_dev(dev); 140 141 seg = pci_domain_nr(pdev->bus); 142 } else { 143 u32 devid = get_acpihid_device_id(dev, NULL); 144 145 seg = PCI_SBDF_TO_SEGID(devid); 146 } 147 148 return seg; 149 } 150 151 /* Writes the specific IOMMU for a device into the PCI segment rlookup table */ 152 void amd_iommu_set_rlookup_table(struct amd_iommu *iommu, u16 devid) 153 { 154 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 155 156 pci_seg->rlookup_table[devid] = iommu; 157 } 158 159 static struct amd_iommu *__rlookup_amd_iommu(u16 seg, u16 devid) 160 { 161 struct amd_iommu_pci_seg *pci_seg; 162 163 for_each_pci_segment(pci_seg) { 164 if (pci_seg->id == seg) 165 return pci_seg->rlookup_table[devid]; 166 } 167 return NULL; 168 } 169 170 static struct amd_iommu *rlookup_amd_iommu(struct device *dev) 171 { 172 u16 seg = get_device_segment(dev); 173 int devid = get_device_sbdf_id(dev); 174 175 if (devid < 0) 176 return NULL; 177 return __rlookup_amd_iommu(seg, PCI_SBDF_TO_DEVID(devid)); 178 } 179 180 static struct protection_domain *to_pdomain(struct iommu_domain *dom) 181 { 182 return container_of(dom, struct protection_domain, domain); 183 } 184 185 static struct iommu_dev_data *alloc_dev_data(struct amd_iommu *iommu, u16 devid) 186 { 187 struct iommu_dev_data *dev_data; 188 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 189 190 dev_data = kzalloc(sizeof(*dev_data), GFP_KERNEL); 191 if (!dev_data) 192 return NULL; 193 194 spin_lock_init(&dev_data->lock); 195 dev_data->devid = devid; 196 ratelimit_default_init(&dev_data->rs); 197 198 llist_add(&dev_data->dev_data_list, &pci_seg->dev_data_list); 199 return dev_data; 200 } 201 202 static struct iommu_dev_data *search_dev_data(struct amd_iommu *iommu, u16 devid) 203 { 204 struct iommu_dev_data *dev_data; 205 struct llist_node *node; 206 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 207 208 if (llist_empty(&pci_seg->dev_data_list)) 209 return NULL; 210 211 node = pci_seg->dev_data_list.first; 212 llist_for_each_entry(dev_data, node, dev_data_list) { 213 if (dev_data->devid == devid) 214 return dev_data; 215 } 216 217 return NULL; 218 } 219 220 static int clone_alias(struct pci_dev *pdev, u16 alias, void *data) 221 { 222 struct amd_iommu *iommu; 223 struct dev_table_entry *dev_table; 224 u16 devid = pci_dev_id(pdev); 225 226 if (devid == alias) 227 return 0; 228 229 iommu = rlookup_amd_iommu(&pdev->dev); 230 if (!iommu) 231 return 0; 232 233 amd_iommu_set_rlookup_table(iommu, alias); 234 dev_table = get_dev_table(iommu); 235 memcpy(dev_table[alias].data, 236 dev_table[devid].data, 237 sizeof(dev_table[alias].data)); 238 239 return 0; 240 } 241 242 static void clone_aliases(struct amd_iommu *iommu, struct device *dev) 243 { 244 struct pci_dev *pdev; 245 246 if (!dev_is_pci(dev)) 247 return; 248 pdev = to_pci_dev(dev); 249 250 /* 251 * The IVRS alias stored in the alias table may not be 252 * part of the PCI DMA aliases if it's bus differs 253 * from the original device. 254 */ 255 clone_alias(pdev, iommu->pci_seg->alias_table[pci_dev_id(pdev)], NULL); 256 257 pci_for_each_dma_alias(pdev, clone_alias, NULL); 258 } 259 260 static void setup_aliases(struct amd_iommu *iommu, struct device *dev) 261 { 262 struct pci_dev *pdev = to_pci_dev(dev); 263 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 264 u16 ivrs_alias; 265 266 /* For ACPI HID devices, there are no aliases */ 267 if (!dev_is_pci(dev)) 268 return; 269 270 /* 271 * Add the IVRS alias to the pci aliases if it is on the same 272 * bus. The IVRS table may know about a quirk that we don't. 273 */ 274 ivrs_alias = pci_seg->alias_table[pci_dev_id(pdev)]; 275 if (ivrs_alias != pci_dev_id(pdev) && 276 PCI_BUS_NUM(ivrs_alias) == pdev->bus->number) 277 pci_add_dma_alias(pdev, ivrs_alias & 0xff, 1); 278 279 clone_aliases(iommu, dev); 280 } 281 282 static struct iommu_dev_data *find_dev_data(struct amd_iommu *iommu, u16 devid) 283 { 284 struct iommu_dev_data *dev_data; 285 286 dev_data = search_dev_data(iommu, devid); 287 288 if (dev_data == NULL) { 289 dev_data = alloc_dev_data(iommu, devid); 290 if (!dev_data) 291 return NULL; 292 293 if (translation_pre_enabled(iommu)) 294 dev_data->defer_attach = true; 295 } 296 297 return dev_data; 298 } 299 300 /* 301 * Find or create an IOMMU group for a acpihid device. 302 */ 303 static struct iommu_group *acpihid_device_group(struct device *dev) 304 { 305 struct acpihid_map_entry *p, *entry = NULL; 306 int devid; 307 308 devid = get_acpihid_device_id(dev, &entry); 309 if (devid < 0) 310 return ERR_PTR(devid); 311 312 list_for_each_entry(p, &acpihid_map, list) { 313 if ((devid == p->devid) && p->group) 314 entry->group = p->group; 315 } 316 317 if (!entry->group) 318 entry->group = generic_device_group(dev); 319 else 320 iommu_group_ref_get(entry->group); 321 322 return entry->group; 323 } 324 325 static bool pci_iommuv2_capable(struct pci_dev *pdev) 326 { 327 static const int caps[] = { 328 PCI_EXT_CAP_ID_PRI, 329 PCI_EXT_CAP_ID_PASID, 330 }; 331 int i, pos; 332 333 if (!pci_ats_supported(pdev)) 334 return false; 335 336 for (i = 0; i < 2; ++i) { 337 pos = pci_find_ext_capability(pdev, caps[i]); 338 if (pos == 0) 339 return false; 340 } 341 342 return true; 343 } 344 345 /* 346 * This function checks if the driver got a valid device from the caller to 347 * avoid dereferencing invalid pointers. 348 */ 349 static bool check_device(struct device *dev) 350 { 351 struct amd_iommu_pci_seg *pci_seg; 352 struct amd_iommu *iommu; 353 int devid, sbdf; 354 355 if (!dev) 356 return false; 357 358 sbdf = get_device_sbdf_id(dev); 359 if (sbdf < 0) 360 return false; 361 devid = PCI_SBDF_TO_DEVID(sbdf); 362 363 iommu = rlookup_amd_iommu(dev); 364 if (!iommu) 365 return false; 366 367 /* Out of our scope? */ 368 pci_seg = iommu->pci_seg; 369 if (devid > pci_seg->last_bdf) 370 return false; 371 372 return true; 373 } 374 375 static int iommu_init_device(struct amd_iommu *iommu, struct device *dev) 376 { 377 struct iommu_dev_data *dev_data; 378 int devid, sbdf; 379 380 if (dev_iommu_priv_get(dev)) 381 return 0; 382 383 sbdf = get_device_sbdf_id(dev); 384 if (sbdf < 0) 385 return sbdf; 386 387 devid = PCI_SBDF_TO_DEVID(sbdf); 388 dev_data = find_dev_data(iommu, devid); 389 if (!dev_data) 390 return -ENOMEM; 391 392 dev_data->dev = dev; 393 setup_aliases(iommu, dev); 394 395 /* 396 * By default we use passthrough mode for IOMMUv2 capable device. 397 * But if amd_iommu=force_isolation is set (e.g. to debug DMA to 398 * invalid address), we ignore the capability for the device so 399 * it'll be forced to go into translation mode. 400 */ 401 if ((iommu_default_passthrough() || !amd_iommu_force_isolation) && 402 dev_is_pci(dev) && pci_iommuv2_capable(to_pci_dev(dev))) { 403 dev_data->iommu_v2 = iommu->is_iommu_v2; 404 } 405 406 dev_iommu_priv_set(dev, dev_data); 407 408 return 0; 409 } 410 411 static void iommu_ignore_device(struct amd_iommu *iommu, struct device *dev) 412 { 413 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 414 struct dev_table_entry *dev_table = get_dev_table(iommu); 415 int devid, sbdf; 416 417 sbdf = get_device_sbdf_id(dev); 418 if (sbdf < 0) 419 return; 420 421 devid = PCI_SBDF_TO_DEVID(sbdf); 422 pci_seg->rlookup_table[devid] = NULL; 423 memset(&dev_table[devid], 0, sizeof(struct dev_table_entry)); 424 425 setup_aliases(iommu, dev); 426 } 427 428 static void amd_iommu_uninit_device(struct device *dev) 429 { 430 struct iommu_dev_data *dev_data; 431 432 dev_data = dev_iommu_priv_get(dev); 433 if (!dev_data) 434 return; 435 436 if (dev_data->domain) 437 detach_device(dev); 438 439 dev_iommu_priv_set(dev, NULL); 440 441 /* 442 * We keep dev_data around for unplugged devices and reuse it when the 443 * device is re-plugged - not doing so would introduce a ton of races. 444 */ 445 } 446 447 /**************************************************************************** 448 * 449 * Interrupt handling functions 450 * 451 ****************************************************************************/ 452 453 static void dump_dte_entry(struct amd_iommu *iommu, u16 devid) 454 { 455 int i; 456 struct dev_table_entry *dev_table = get_dev_table(iommu); 457 458 for (i = 0; i < 4; ++i) 459 pr_err("DTE[%d]: %016llx\n", i, dev_table[devid].data[i]); 460 } 461 462 static void dump_command(unsigned long phys_addr) 463 { 464 struct iommu_cmd *cmd = iommu_phys_to_virt(phys_addr); 465 int i; 466 467 for (i = 0; i < 4; ++i) 468 pr_err("CMD[%d]: %08x\n", i, cmd->data[i]); 469 } 470 471 static void amd_iommu_report_rmp_hw_error(struct amd_iommu *iommu, volatile u32 *event) 472 { 473 struct iommu_dev_data *dev_data = NULL; 474 int devid, vmg_tag, flags; 475 struct pci_dev *pdev; 476 u64 spa; 477 478 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK; 479 vmg_tag = (event[1]) & 0xFFFF; 480 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK; 481 spa = ((u64)event[3] << 32) | (event[2] & 0xFFFFFFF8); 482 483 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid), 484 devid & 0xff); 485 if (pdev) 486 dev_data = dev_iommu_priv_get(&pdev->dev); 487 488 if (dev_data) { 489 if (__ratelimit(&dev_data->rs)) { 490 pci_err(pdev, "Event logged [RMP_HW_ERROR vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n", 491 vmg_tag, spa, flags); 492 } 493 } else { 494 pr_err_ratelimited("Event logged [RMP_HW_ERROR device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, spa=0x%llx, flags=0x%04x]\n", 495 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 496 vmg_tag, spa, flags); 497 } 498 499 if (pdev) 500 pci_dev_put(pdev); 501 } 502 503 static void amd_iommu_report_rmp_fault(struct amd_iommu *iommu, volatile u32 *event) 504 { 505 struct iommu_dev_data *dev_data = NULL; 506 int devid, flags_rmp, vmg_tag, flags; 507 struct pci_dev *pdev; 508 u64 gpa; 509 510 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK; 511 flags_rmp = (event[0] >> EVENT_FLAGS_SHIFT) & 0xFF; 512 vmg_tag = (event[1]) & 0xFFFF; 513 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK; 514 gpa = ((u64)event[3] << 32) | event[2]; 515 516 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid), 517 devid & 0xff); 518 if (pdev) 519 dev_data = dev_iommu_priv_get(&pdev->dev); 520 521 if (dev_data) { 522 if (__ratelimit(&dev_data->rs)) { 523 pci_err(pdev, "Event logged [RMP_PAGE_FAULT vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n", 524 vmg_tag, gpa, flags_rmp, flags); 525 } 526 } else { 527 pr_err_ratelimited("Event logged [RMP_PAGE_FAULT device=%04x:%02x:%02x.%x, vmg_tag=0x%04x, gpa=0x%llx, flags_rmp=0x%04x, flags=0x%04x]\n", 528 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 529 vmg_tag, gpa, flags_rmp, flags); 530 } 531 532 if (pdev) 533 pci_dev_put(pdev); 534 } 535 536 #define IS_IOMMU_MEM_TRANSACTION(flags) \ 537 (((flags) & EVENT_FLAG_I) == 0) 538 539 #define IS_WRITE_REQUEST(flags) \ 540 ((flags) & EVENT_FLAG_RW) 541 542 static void amd_iommu_report_page_fault(struct amd_iommu *iommu, 543 u16 devid, u16 domain_id, 544 u64 address, int flags) 545 { 546 struct iommu_dev_data *dev_data = NULL; 547 struct pci_dev *pdev; 548 549 pdev = pci_get_domain_bus_and_slot(iommu->pci_seg->id, PCI_BUS_NUM(devid), 550 devid & 0xff); 551 if (pdev) 552 dev_data = dev_iommu_priv_get(&pdev->dev); 553 554 if (dev_data) { 555 /* 556 * If this is a DMA fault (for which the I(nterrupt) 557 * bit will be unset), allow report_iommu_fault() to 558 * prevent logging it. 559 */ 560 if (IS_IOMMU_MEM_TRANSACTION(flags)) { 561 /* Device not attached to domain properly */ 562 if (dev_data->domain == NULL) { 563 pr_err_ratelimited("Event logged [Device not attached to domain properly]\n"); 564 pr_err_ratelimited(" device=%04x:%02x:%02x.%x domain=0x%04x\n", 565 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), 566 PCI_FUNC(devid), domain_id); 567 goto out; 568 } 569 570 if (!report_iommu_fault(&dev_data->domain->domain, 571 &pdev->dev, address, 572 IS_WRITE_REQUEST(flags) ? 573 IOMMU_FAULT_WRITE : 574 IOMMU_FAULT_READ)) 575 goto out; 576 } 577 578 if (__ratelimit(&dev_data->rs)) { 579 pci_err(pdev, "Event logged [IO_PAGE_FAULT domain=0x%04x address=0x%llx flags=0x%04x]\n", 580 domain_id, address, flags); 581 } 582 } else { 583 pr_err_ratelimited("Event logged [IO_PAGE_FAULT device=%04x:%02x:%02x.%x domain=0x%04x address=0x%llx flags=0x%04x]\n", 584 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 585 domain_id, address, flags); 586 } 587 588 out: 589 if (pdev) 590 pci_dev_put(pdev); 591 } 592 593 static void iommu_print_event(struct amd_iommu *iommu, void *__evt) 594 { 595 struct device *dev = iommu->iommu.dev; 596 int type, devid, flags, tag; 597 volatile u32 *event = __evt; 598 int count = 0; 599 u64 address; 600 u32 pasid; 601 602 retry: 603 type = (event[1] >> EVENT_TYPE_SHIFT) & EVENT_TYPE_MASK; 604 devid = (event[0] >> EVENT_DEVID_SHIFT) & EVENT_DEVID_MASK; 605 pasid = (event[0] & EVENT_DOMID_MASK_HI) | 606 (event[1] & EVENT_DOMID_MASK_LO); 607 flags = (event[1] >> EVENT_FLAGS_SHIFT) & EVENT_FLAGS_MASK; 608 address = (u64)(((u64)event[3]) << 32) | event[2]; 609 610 if (type == 0) { 611 /* Did we hit the erratum? */ 612 if (++count == LOOP_TIMEOUT) { 613 pr_err("No event written to event log\n"); 614 return; 615 } 616 udelay(1); 617 goto retry; 618 } 619 620 if (type == EVENT_TYPE_IO_FAULT) { 621 amd_iommu_report_page_fault(iommu, devid, pasid, address, flags); 622 return; 623 } 624 625 switch (type) { 626 case EVENT_TYPE_ILL_DEV: 627 dev_err(dev, "Event logged [ILLEGAL_DEV_TABLE_ENTRY device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n", 628 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 629 pasid, address, flags); 630 dump_dte_entry(iommu, devid); 631 break; 632 case EVENT_TYPE_DEV_TAB_ERR: 633 dev_err(dev, "Event logged [DEV_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x " 634 "address=0x%llx flags=0x%04x]\n", 635 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 636 address, flags); 637 break; 638 case EVENT_TYPE_PAGE_TAB_ERR: 639 dev_err(dev, "Event logged [PAGE_TAB_HARDWARE_ERROR device=%04x:%02x:%02x.%x pasid=0x%04x address=0x%llx flags=0x%04x]\n", 640 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 641 pasid, address, flags); 642 break; 643 case EVENT_TYPE_ILL_CMD: 644 dev_err(dev, "Event logged [ILLEGAL_COMMAND_ERROR address=0x%llx]\n", address); 645 dump_command(address); 646 break; 647 case EVENT_TYPE_CMD_HARD_ERR: 648 dev_err(dev, "Event logged [COMMAND_HARDWARE_ERROR address=0x%llx flags=0x%04x]\n", 649 address, flags); 650 break; 651 case EVENT_TYPE_IOTLB_INV_TO: 652 dev_err(dev, "Event logged [IOTLB_INV_TIMEOUT device=%04x:%02x:%02x.%x address=0x%llx]\n", 653 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 654 address); 655 break; 656 case EVENT_TYPE_INV_DEV_REQ: 657 dev_err(dev, "Event logged [INVALID_DEVICE_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x]\n", 658 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 659 pasid, address, flags); 660 break; 661 case EVENT_TYPE_RMP_FAULT: 662 amd_iommu_report_rmp_fault(iommu, event); 663 break; 664 case EVENT_TYPE_RMP_HW_ERR: 665 amd_iommu_report_rmp_hw_error(iommu, event); 666 break; 667 case EVENT_TYPE_INV_PPR_REQ: 668 pasid = PPR_PASID(*((u64 *)__evt)); 669 tag = event[1] & 0x03FF; 670 dev_err(dev, "Event logged [INVALID_PPR_REQUEST device=%04x:%02x:%02x.%x pasid=0x%05x address=0x%llx flags=0x%04x tag=0x%03x]\n", 671 iommu->pci_seg->id, PCI_BUS_NUM(devid), PCI_SLOT(devid), PCI_FUNC(devid), 672 pasid, address, flags, tag); 673 break; 674 default: 675 dev_err(dev, "Event logged [UNKNOWN event[0]=0x%08x event[1]=0x%08x event[2]=0x%08x event[3]=0x%08x\n", 676 event[0], event[1], event[2], event[3]); 677 } 678 679 /* 680 * To detect the hardware errata 732 we need to clear the 681 * entry back to zero. This issue does not exist on SNP 682 * enabled system. Also this buffer is not writeable on 683 * SNP enabled system. 684 */ 685 if (!amd_iommu_snp_en) 686 memset(__evt, 0, 4 * sizeof(u32)); 687 } 688 689 static void iommu_poll_events(struct amd_iommu *iommu) 690 { 691 u32 head, tail; 692 693 head = readl(iommu->mmio_base + MMIO_EVT_HEAD_OFFSET); 694 tail = readl(iommu->mmio_base + MMIO_EVT_TAIL_OFFSET); 695 696 while (head != tail) { 697 iommu_print_event(iommu, iommu->evt_buf + head); 698 head = (head + EVENT_ENTRY_SIZE) % EVT_BUFFER_SIZE; 699 } 700 701 writel(head, iommu->mmio_base + MMIO_EVT_HEAD_OFFSET); 702 } 703 704 static void iommu_handle_ppr_entry(struct amd_iommu *iommu, u64 *raw) 705 { 706 struct amd_iommu_fault fault; 707 708 if (PPR_REQ_TYPE(raw[0]) != PPR_REQ_FAULT) { 709 pr_err_ratelimited("Unknown PPR request received\n"); 710 return; 711 } 712 713 fault.address = raw[1]; 714 fault.pasid = PPR_PASID(raw[0]); 715 fault.sbdf = PCI_SEG_DEVID_TO_SBDF(iommu->pci_seg->id, PPR_DEVID(raw[0])); 716 fault.tag = PPR_TAG(raw[0]); 717 fault.flags = PPR_FLAGS(raw[0]); 718 719 atomic_notifier_call_chain(&ppr_notifier, 0, &fault); 720 } 721 722 static void iommu_poll_ppr_log(struct amd_iommu *iommu) 723 { 724 u32 head, tail; 725 726 if (iommu->ppr_log == NULL) 727 return; 728 729 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET); 730 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET); 731 732 while (head != tail) { 733 volatile u64 *raw; 734 u64 entry[2]; 735 int i; 736 737 raw = (u64 *)(iommu->ppr_log + head); 738 739 /* 740 * Hardware bug: Interrupt may arrive before the entry is 741 * written to memory. If this happens we need to wait for the 742 * entry to arrive. 743 */ 744 for (i = 0; i < LOOP_TIMEOUT; ++i) { 745 if (PPR_REQ_TYPE(raw[0]) != 0) 746 break; 747 udelay(1); 748 } 749 750 /* Avoid memcpy function-call overhead */ 751 entry[0] = raw[0]; 752 entry[1] = raw[1]; 753 754 /* 755 * To detect the hardware errata 733 we need to clear the 756 * entry back to zero. This issue does not exist on SNP 757 * enabled system. Also this buffer is not writeable on 758 * SNP enabled system. 759 */ 760 if (!amd_iommu_snp_en) 761 raw[0] = raw[1] = 0UL; 762 763 /* Update head pointer of hardware ring-buffer */ 764 head = (head + PPR_ENTRY_SIZE) % PPR_LOG_SIZE; 765 writel(head, iommu->mmio_base + MMIO_PPR_HEAD_OFFSET); 766 767 /* Handle PPR entry */ 768 iommu_handle_ppr_entry(iommu, entry); 769 770 /* Refresh ring-buffer information */ 771 head = readl(iommu->mmio_base + MMIO_PPR_HEAD_OFFSET); 772 tail = readl(iommu->mmio_base + MMIO_PPR_TAIL_OFFSET); 773 } 774 } 775 776 #ifdef CONFIG_IRQ_REMAP 777 static int (*iommu_ga_log_notifier)(u32); 778 779 int amd_iommu_register_ga_log_notifier(int (*notifier)(u32)) 780 { 781 iommu_ga_log_notifier = notifier; 782 783 return 0; 784 } 785 EXPORT_SYMBOL(amd_iommu_register_ga_log_notifier); 786 787 static void iommu_poll_ga_log(struct amd_iommu *iommu) 788 { 789 u32 head, tail; 790 791 if (iommu->ga_log == NULL) 792 return; 793 794 head = readl(iommu->mmio_base + MMIO_GA_HEAD_OFFSET); 795 tail = readl(iommu->mmio_base + MMIO_GA_TAIL_OFFSET); 796 797 while (head != tail) { 798 volatile u64 *raw; 799 u64 log_entry; 800 801 raw = (u64 *)(iommu->ga_log + head); 802 803 /* Avoid memcpy function-call overhead */ 804 log_entry = *raw; 805 806 /* Update head pointer of hardware ring-buffer */ 807 head = (head + GA_ENTRY_SIZE) % GA_LOG_SIZE; 808 writel(head, iommu->mmio_base + MMIO_GA_HEAD_OFFSET); 809 810 /* Handle GA entry */ 811 switch (GA_REQ_TYPE(log_entry)) { 812 case GA_GUEST_NR: 813 if (!iommu_ga_log_notifier) 814 break; 815 816 pr_debug("%s: devid=%#x, ga_tag=%#x\n", 817 __func__, GA_DEVID(log_entry), 818 GA_TAG(log_entry)); 819 820 if (iommu_ga_log_notifier(GA_TAG(log_entry)) != 0) 821 pr_err("GA log notifier failed.\n"); 822 break; 823 default: 824 break; 825 } 826 } 827 } 828 829 static void 830 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) 831 { 832 if (!irq_remapping_enabled || !dev_is_pci(dev) || 833 !pci_dev_has_default_msi_parent_domain(to_pci_dev(dev))) 834 return; 835 836 dev_set_msi_domain(dev, iommu->ir_domain); 837 } 838 839 #else /* CONFIG_IRQ_REMAP */ 840 static inline void 841 amd_iommu_set_pci_msi_domain(struct device *dev, struct amd_iommu *iommu) { } 842 #endif /* !CONFIG_IRQ_REMAP */ 843 844 #define AMD_IOMMU_INT_MASK \ 845 (MMIO_STATUS_EVT_OVERFLOW_INT_MASK | \ 846 MMIO_STATUS_EVT_INT_MASK | \ 847 MMIO_STATUS_PPR_INT_MASK | \ 848 MMIO_STATUS_GALOG_OVERFLOW_MASK | \ 849 MMIO_STATUS_GALOG_INT_MASK) 850 851 irqreturn_t amd_iommu_int_thread(int irq, void *data) 852 { 853 struct amd_iommu *iommu = (struct amd_iommu *) data; 854 u32 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); 855 856 while (status & AMD_IOMMU_INT_MASK) { 857 /* Enable interrupt sources again */ 858 writel(AMD_IOMMU_INT_MASK, 859 iommu->mmio_base + MMIO_STATUS_OFFSET); 860 861 if (status & MMIO_STATUS_EVT_INT_MASK) { 862 pr_devel("Processing IOMMU Event Log\n"); 863 iommu_poll_events(iommu); 864 } 865 866 if (status & MMIO_STATUS_PPR_INT_MASK) { 867 pr_devel("Processing IOMMU PPR Log\n"); 868 iommu_poll_ppr_log(iommu); 869 } 870 871 #ifdef CONFIG_IRQ_REMAP 872 if (status & (MMIO_STATUS_GALOG_INT_MASK | 873 MMIO_STATUS_GALOG_OVERFLOW_MASK)) { 874 pr_devel("Processing IOMMU GA Log\n"); 875 iommu_poll_ga_log(iommu); 876 } 877 878 if (status & MMIO_STATUS_GALOG_OVERFLOW_MASK) { 879 pr_info_ratelimited("IOMMU GA Log overflow\n"); 880 amd_iommu_restart_ga_log(iommu); 881 } 882 #endif 883 884 if (status & MMIO_STATUS_EVT_OVERFLOW_INT_MASK) { 885 pr_info_ratelimited("IOMMU event log overflow\n"); 886 amd_iommu_restart_event_logging(iommu); 887 } 888 889 /* 890 * Hardware bug: ERBT1312 891 * When re-enabling interrupt (by writing 1 892 * to clear the bit), the hardware might also try to set 893 * the interrupt bit in the event status register. 894 * In this scenario, the bit will be set, and disable 895 * subsequent interrupts. 896 * 897 * Workaround: The IOMMU driver should read back the 898 * status register and check if the interrupt bits are cleared. 899 * If not, driver will need to go through the interrupt handler 900 * again and re-clear the bits 901 */ 902 status = readl(iommu->mmio_base + MMIO_STATUS_OFFSET); 903 } 904 return IRQ_HANDLED; 905 } 906 907 irqreturn_t amd_iommu_int_handler(int irq, void *data) 908 { 909 return IRQ_WAKE_THREAD; 910 } 911 912 /**************************************************************************** 913 * 914 * IOMMU command queuing functions 915 * 916 ****************************************************************************/ 917 918 static int wait_on_sem(struct amd_iommu *iommu, u64 data) 919 { 920 int i = 0; 921 922 while (*iommu->cmd_sem != data && i < LOOP_TIMEOUT) { 923 udelay(1); 924 i += 1; 925 } 926 927 if (i == LOOP_TIMEOUT) { 928 pr_alert("Completion-Wait loop timed out\n"); 929 return -EIO; 930 } 931 932 return 0; 933 } 934 935 static void copy_cmd_to_buffer(struct amd_iommu *iommu, 936 struct iommu_cmd *cmd) 937 { 938 u8 *target; 939 u32 tail; 940 941 /* Copy command to buffer */ 942 tail = iommu->cmd_buf_tail; 943 target = iommu->cmd_buf + tail; 944 memcpy(target, cmd, sizeof(*cmd)); 945 946 tail = (tail + sizeof(*cmd)) % CMD_BUFFER_SIZE; 947 iommu->cmd_buf_tail = tail; 948 949 /* Tell the IOMMU about it */ 950 writel(tail, iommu->mmio_base + MMIO_CMD_TAIL_OFFSET); 951 } 952 953 static void build_completion_wait(struct iommu_cmd *cmd, 954 struct amd_iommu *iommu, 955 u64 data) 956 { 957 u64 paddr = iommu_virt_to_phys((void *)iommu->cmd_sem); 958 959 memset(cmd, 0, sizeof(*cmd)); 960 cmd->data[0] = lower_32_bits(paddr) | CMD_COMPL_WAIT_STORE_MASK; 961 cmd->data[1] = upper_32_bits(paddr); 962 cmd->data[2] = lower_32_bits(data); 963 cmd->data[3] = upper_32_bits(data); 964 CMD_SET_TYPE(cmd, CMD_COMPL_WAIT); 965 } 966 967 static void build_inv_dte(struct iommu_cmd *cmd, u16 devid) 968 { 969 memset(cmd, 0, sizeof(*cmd)); 970 cmd->data[0] = devid; 971 CMD_SET_TYPE(cmd, CMD_INV_DEV_ENTRY); 972 } 973 974 /* 975 * Builds an invalidation address which is suitable for one page or multiple 976 * pages. Sets the size bit (S) as needed is more than one page is flushed. 977 */ 978 static inline u64 build_inv_address(u64 address, size_t size) 979 { 980 u64 pages, end, msb_diff; 981 982 pages = iommu_num_pages(address, size, PAGE_SIZE); 983 984 if (pages == 1) 985 return address & PAGE_MASK; 986 987 end = address + size - 1; 988 989 /* 990 * msb_diff would hold the index of the most significant bit that 991 * flipped between the start and end. 992 */ 993 msb_diff = fls64(end ^ address) - 1; 994 995 /* 996 * Bits 63:52 are sign extended. If for some reason bit 51 is different 997 * between the start and the end, invalidate everything. 998 */ 999 if (unlikely(msb_diff > 51)) { 1000 address = CMD_INV_IOMMU_ALL_PAGES_ADDRESS; 1001 } else { 1002 /* 1003 * The msb-bit must be clear on the address. Just set all the 1004 * lower bits. 1005 */ 1006 address |= (1ull << msb_diff) - 1; 1007 } 1008 1009 /* Clear bits 11:0 */ 1010 address &= PAGE_MASK; 1011 1012 /* Set the size bit - we flush more than one 4kb page */ 1013 return address | CMD_INV_IOMMU_PAGES_SIZE_MASK; 1014 } 1015 1016 static void build_inv_iommu_pages(struct iommu_cmd *cmd, u64 address, 1017 size_t size, u16 domid, int pde) 1018 { 1019 u64 inv_address = build_inv_address(address, size); 1020 1021 memset(cmd, 0, sizeof(*cmd)); 1022 cmd->data[1] |= domid; 1023 cmd->data[2] = lower_32_bits(inv_address); 1024 cmd->data[3] = upper_32_bits(inv_address); 1025 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES); 1026 if (pde) /* PDE bit - we want to flush everything, not only the PTEs */ 1027 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK; 1028 } 1029 1030 static void build_inv_iotlb_pages(struct iommu_cmd *cmd, u16 devid, int qdep, 1031 u64 address, size_t size) 1032 { 1033 u64 inv_address = build_inv_address(address, size); 1034 1035 memset(cmd, 0, sizeof(*cmd)); 1036 cmd->data[0] = devid; 1037 cmd->data[0] |= (qdep & 0xff) << 24; 1038 cmd->data[1] = devid; 1039 cmd->data[2] = lower_32_bits(inv_address); 1040 cmd->data[3] = upper_32_bits(inv_address); 1041 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES); 1042 } 1043 1044 static void build_inv_iommu_pasid(struct iommu_cmd *cmd, u16 domid, u32 pasid, 1045 u64 address, bool size) 1046 { 1047 memset(cmd, 0, sizeof(*cmd)); 1048 1049 address &= ~(0xfffULL); 1050 1051 cmd->data[0] = pasid; 1052 cmd->data[1] = domid; 1053 cmd->data[2] = lower_32_bits(address); 1054 cmd->data[3] = upper_32_bits(address); 1055 cmd->data[2] |= CMD_INV_IOMMU_PAGES_PDE_MASK; 1056 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK; 1057 if (size) 1058 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK; 1059 CMD_SET_TYPE(cmd, CMD_INV_IOMMU_PAGES); 1060 } 1061 1062 static void build_inv_iotlb_pasid(struct iommu_cmd *cmd, u16 devid, u32 pasid, 1063 int qdep, u64 address, bool size) 1064 { 1065 memset(cmd, 0, sizeof(*cmd)); 1066 1067 address &= ~(0xfffULL); 1068 1069 cmd->data[0] = devid; 1070 cmd->data[0] |= ((pasid >> 8) & 0xff) << 16; 1071 cmd->data[0] |= (qdep & 0xff) << 24; 1072 cmd->data[1] = devid; 1073 cmd->data[1] |= (pasid & 0xff) << 16; 1074 cmd->data[2] = lower_32_bits(address); 1075 cmd->data[2] |= CMD_INV_IOMMU_PAGES_GN_MASK; 1076 cmd->data[3] = upper_32_bits(address); 1077 if (size) 1078 cmd->data[2] |= CMD_INV_IOMMU_PAGES_SIZE_MASK; 1079 CMD_SET_TYPE(cmd, CMD_INV_IOTLB_PAGES); 1080 } 1081 1082 static void build_complete_ppr(struct iommu_cmd *cmd, u16 devid, u32 pasid, 1083 int status, int tag, bool gn) 1084 { 1085 memset(cmd, 0, sizeof(*cmd)); 1086 1087 cmd->data[0] = devid; 1088 if (gn) { 1089 cmd->data[1] = pasid; 1090 cmd->data[2] = CMD_INV_IOMMU_PAGES_GN_MASK; 1091 } 1092 cmd->data[3] = tag & 0x1ff; 1093 cmd->data[3] |= (status & PPR_STATUS_MASK) << PPR_STATUS_SHIFT; 1094 1095 CMD_SET_TYPE(cmd, CMD_COMPLETE_PPR); 1096 } 1097 1098 static void build_inv_all(struct iommu_cmd *cmd) 1099 { 1100 memset(cmd, 0, sizeof(*cmd)); 1101 CMD_SET_TYPE(cmd, CMD_INV_ALL); 1102 } 1103 1104 static void build_inv_irt(struct iommu_cmd *cmd, u16 devid) 1105 { 1106 memset(cmd, 0, sizeof(*cmd)); 1107 cmd->data[0] = devid; 1108 CMD_SET_TYPE(cmd, CMD_INV_IRT); 1109 } 1110 1111 /* 1112 * Writes the command to the IOMMUs command buffer and informs the 1113 * hardware about the new command. 1114 */ 1115 static int __iommu_queue_command_sync(struct amd_iommu *iommu, 1116 struct iommu_cmd *cmd, 1117 bool sync) 1118 { 1119 unsigned int count = 0; 1120 u32 left, next_tail; 1121 1122 next_tail = (iommu->cmd_buf_tail + sizeof(*cmd)) % CMD_BUFFER_SIZE; 1123 again: 1124 left = (iommu->cmd_buf_head - next_tail) % CMD_BUFFER_SIZE; 1125 1126 if (left <= 0x20) { 1127 /* Skip udelay() the first time around */ 1128 if (count++) { 1129 if (count == LOOP_TIMEOUT) { 1130 pr_err("Command buffer timeout\n"); 1131 return -EIO; 1132 } 1133 1134 udelay(1); 1135 } 1136 1137 /* Update head and recheck remaining space */ 1138 iommu->cmd_buf_head = readl(iommu->mmio_base + 1139 MMIO_CMD_HEAD_OFFSET); 1140 1141 goto again; 1142 } 1143 1144 copy_cmd_to_buffer(iommu, cmd); 1145 1146 /* Do we need to make sure all commands are processed? */ 1147 iommu->need_sync = sync; 1148 1149 return 0; 1150 } 1151 1152 static int iommu_queue_command_sync(struct amd_iommu *iommu, 1153 struct iommu_cmd *cmd, 1154 bool sync) 1155 { 1156 unsigned long flags; 1157 int ret; 1158 1159 raw_spin_lock_irqsave(&iommu->lock, flags); 1160 ret = __iommu_queue_command_sync(iommu, cmd, sync); 1161 raw_spin_unlock_irqrestore(&iommu->lock, flags); 1162 1163 return ret; 1164 } 1165 1166 static int iommu_queue_command(struct amd_iommu *iommu, struct iommu_cmd *cmd) 1167 { 1168 return iommu_queue_command_sync(iommu, cmd, true); 1169 } 1170 1171 /* 1172 * This function queues a completion wait command into the command 1173 * buffer of an IOMMU 1174 */ 1175 static int iommu_completion_wait(struct amd_iommu *iommu) 1176 { 1177 struct iommu_cmd cmd; 1178 unsigned long flags; 1179 int ret; 1180 u64 data; 1181 1182 if (!iommu->need_sync) 1183 return 0; 1184 1185 data = atomic64_add_return(1, &iommu->cmd_sem_val); 1186 build_completion_wait(&cmd, iommu, data); 1187 1188 raw_spin_lock_irqsave(&iommu->lock, flags); 1189 1190 ret = __iommu_queue_command_sync(iommu, &cmd, false); 1191 if (ret) 1192 goto out_unlock; 1193 1194 ret = wait_on_sem(iommu, data); 1195 1196 out_unlock: 1197 raw_spin_unlock_irqrestore(&iommu->lock, flags); 1198 1199 return ret; 1200 } 1201 1202 static int iommu_flush_dte(struct amd_iommu *iommu, u16 devid) 1203 { 1204 struct iommu_cmd cmd; 1205 1206 build_inv_dte(&cmd, devid); 1207 1208 return iommu_queue_command(iommu, &cmd); 1209 } 1210 1211 static void amd_iommu_flush_dte_all(struct amd_iommu *iommu) 1212 { 1213 u32 devid; 1214 u16 last_bdf = iommu->pci_seg->last_bdf; 1215 1216 for (devid = 0; devid <= last_bdf; ++devid) 1217 iommu_flush_dte(iommu, devid); 1218 1219 iommu_completion_wait(iommu); 1220 } 1221 1222 /* 1223 * This function uses heavy locking and may disable irqs for some time. But 1224 * this is no issue because it is only called during resume. 1225 */ 1226 static void amd_iommu_flush_tlb_all(struct amd_iommu *iommu) 1227 { 1228 u32 dom_id; 1229 u16 last_bdf = iommu->pci_seg->last_bdf; 1230 1231 for (dom_id = 0; dom_id <= last_bdf; ++dom_id) { 1232 struct iommu_cmd cmd; 1233 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1234 dom_id, 1); 1235 iommu_queue_command(iommu, &cmd); 1236 } 1237 1238 iommu_completion_wait(iommu); 1239 } 1240 1241 static void amd_iommu_flush_tlb_domid(struct amd_iommu *iommu, u32 dom_id) 1242 { 1243 struct iommu_cmd cmd; 1244 1245 build_inv_iommu_pages(&cmd, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1246 dom_id, 1); 1247 iommu_queue_command(iommu, &cmd); 1248 1249 iommu_completion_wait(iommu); 1250 } 1251 1252 static void amd_iommu_flush_all(struct amd_iommu *iommu) 1253 { 1254 struct iommu_cmd cmd; 1255 1256 build_inv_all(&cmd); 1257 1258 iommu_queue_command(iommu, &cmd); 1259 iommu_completion_wait(iommu); 1260 } 1261 1262 static void iommu_flush_irt(struct amd_iommu *iommu, u16 devid) 1263 { 1264 struct iommu_cmd cmd; 1265 1266 build_inv_irt(&cmd, devid); 1267 1268 iommu_queue_command(iommu, &cmd); 1269 } 1270 1271 static void amd_iommu_flush_irt_all(struct amd_iommu *iommu) 1272 { 1273 u32 devid; 1274 u16 last_bdf = iommu->pci_seg->last_bdf; 1275 1276 if (iommu->irtcachedis_enabled) 1277 return; 1278 1279 for (devid = 0; devid <= last_bdf; devid++) 1280 iommu_flush_irt(iommu, devid); 1281 1282 iommu_completion_wait(iommu); 1283 } 1284 1285 void iommu_flush_all_caches(struct amd_iommu *iommu) 1286 { 1287 if (iommu_feature(iommu, FEATURE_IA)) { 1288 amd_iommu_flush_all(iommu); 1289 } else { 1290 amd_iommu_flush_dte_all(iommu); 1291 amd_iommu_flush_irt_all(iommu); 1292 amd_iommu_flush_tlb_all(iommu); 1293 } 1294 } 1295 1296 /* 1297 * Command send function for flushing on-device TLB 1298 */ 1299 static int device_flush_iotlb(struct iommu_dev_data *dev_data, 1300 u64 address, size_t size) 1301 { 1302 struct amd_iommu *iommu; 1303 struct iommu_cmd cmd; 1304 int qdep; 1305 1306 qdep = dev_data->ats.qdep; 1307 iommu = rlookup_amd_iommu(dev_data->dev); 1308 if (!iommu) 1309 return -EINVAL; 1310 1311 build_inv_iotlb_pages(&cmd, dev_data->devid, qdep, address, size); 1312 1313 return iommu_queue_command(iommu, &cmd); 1314 } 1315 1316 static int device_flush_dte_alias(struct pci_dev *pdev, u16 alias, void *data) 1317 { 1318 struct amd_iommu *iommu = data; 1319 1320 return iommu_flush_dte(iommu, alias); 1321 } 1322 1323 /* 1324 * Command send function for invalidating a device table entry 1325 */ 1326 static int device_flush_dte(struct iommu_dev_data *dev_data) 1327 { 1328 struct amd_iommu *iommu; 1329 struct pci_dev *pdev = NULL; 1330 struct amd_iommu_pci_seg *pci_seg; 1331 u16 alias; 1332 int ret; 1333 1334 iommu = rlookup_amd_iommu(dev_data->dev); 1335 if (!iommu) 1336 return -EINVAL; 1337 1338 if (dev_is_pci(dev_data->dev)) 1339 pdev = to_pci_dev(dev_data->dev); 1340 1341 if (pdev) 1342 ret = pci_for_each_dma_alias(pdev, 1343 device_flush_dte_alias, iommu); 1344 else 1345 ret = iommu_flush_dte(iommu, dev_data->devid); 1346 if (ret) 1347 return ret; 1348 1349 pci_seg = iommu->pci_seg; 1350 alias = pci_seg->alias_table[dev_data->devid]; 1351 if (alias != dev_data->devid) { 1352 ret = iommu_flush_dte(iommu, alias); 1353 if (ret) 1354 return ret; 1355 } 1356 1357 if (dev_data->ats.enabled) 1358 ret = device_flush_iotlb(dev_data, 0, ~0UL); 1359 1360 return ret; 1361 } 1362 1363 /* 1364 * TLB invalidation function which is called from the mapping functions. 1365 * It invalidates a single PTE if the range to flush is within a single 1366 * page. Otherwise it flushes the whole TLB of the IOMMU. 1367 */ 1368 static void __domain_flush_pages(struct protection_domain *domain, 1369 u64 address, size_t size, int pde) 1370 { 1371 struct iommu_dev_data *dev_data; 1372 struct iommu_cmd cmd; 1373 int ret = 0, i; 1374 1375 build_inv_iommu_pages(&cmd, address, size, domain->id, pde); 1376 1377 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) { 1378 if (!domain->dev_iommu[i]) 1379 continue; 1380 1381 /* 1382 * Devices of this domain are behind this IOMMU 1383 * We need a TLB flush 1384 */ 1385 ret |= iommu_queue_command(amd_iommus[i], &cmd); 1386 } 1387 1388 list_for_each_entry(dev_data, &domain->dev_list, list) { 1389 1390 if (!dev_data->ats.enabled) 1391 continue; 1392 1393 ret |= device_flush_iotlb(dev_data, address, size); 1394 } 1395 1396 WARN_ON(ret); 1397 } 1398 1399 static void domain_flush_pages(struct protection_domain *domain, 1400 u64 address, size_t size, int pde) 1401 { 1402 if (likely(!amd_iommu_np_cache)) { 1403 __domain_flush_pages(domain, address, size, pde); 1404 return; 1405 } 1406 1407 /* 1408 * When NpCache is on, we infer that we run in a VM and use a vIOMMU. 1409 * In such setups it is best to avoid flushes of ranges which are not 1410 * naturally aligned, since it would lead to flushes of unmodified 1411 * PTEs. Such flushes would require the hypervisor to do more work than 1412 * necessary. Therefore, perform repeated flushes of aligned ranges 1413 * until you cover the range. Each iteration flushes the smaller 1414 * between the natural alignment of the address that we flush and the 1415 * greatest naturally aligned region that fits in the range. 1416 */ 1417 while (size != 0) { 1418 int addr_alignment = __ffs(address); 1419 int size_alignment = __fls(size); 1420 int min_alignment; 1421 size_t flush_size; 1422 1423 /* 1424 * size is always non-zero, but address might be zero, causing 1425 * addr_alignment to be negative. As the casting of the 1426 * argument in __ffs(address) to long might trim the high bits 1427 * of the address on x86-32, cast to long when doing the check. 1428 */ 1429 if (likely((unsigned long)address != 0)) 1430 min_alignment = min(addr_alignment, size_alignment); 1431 else 1432 min_alignment = size_alignment; 1433 1434 flush_size = 1ul << min_alignment; 1435 1436 __domain_flush_pages(domain, address, flush_size, pde); 1437 address += flush_size; 1438 size -= flush_size; 1439 } 1440 } 1441 1442 /* Flush the whole IO/TLB for a given protection domain - including PDE */ 1443 void amd_iommu_domain_flush_tlb_pde(struct protection_domain *domain) 1444 { 1445 domain_flush_pages(domain, 0, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 1); 1446 } 1447 1448 void amd_iommu_domain_flush_complete(struct protection_domain *domain) 1449 { 1450 int i; 1451 1452 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) { 1453 if (domain && !domain->dev_iommu[i]) 1454 continue; 1455 1456 /* 1457 * Devices of this domain are behind this IOMMU 1458 * We need to wait for completion of all commands. 1459 */ 1460 iommu_completion_wait(amd_iommus[i]); 1461 } 1462 } 1463 1464 /* Flush the not present cache if it exists */ 1465 static void domain_flush_np_cache(struct protection_domain *domain, 1466 dma_addr_t iova, size_t size) 1467 { 1468 if (unlikely(amd_iommu_np_cache)) { 1469 unsigned long flags; 1470 1471 spin_lock_irqsave(&domain->lock, flags); 1472 domain_flush_pages(domain, iova, size, 1); 1473 amd_iommu_domain_flush_complete(domain); 1474 spin_unlock_irqrestore(&domain->lock, flags); 1475 } 1476 } 1477 1478 1479 /* 1480 * This function flushes the DTEs for all devices in domain 1481 */ 1482 static void domain_flush_devices(struct protection_domain *domain) 1483 { 1484 struct iommu_dev_data *dev_data; 1485 1486 list_for_each_entry(dev_data, &domain->dev_list, list) 1487 device_flush_dte(dev_data); 1488 } 1489 1490 /**************************************************************************** 1491 * 1492 * The next functions belong to the domain allocation. A domain is 1493 * allocated for every IOMMU as the default domain. If device isolation 1494 * is enabled, every device get its own domain. The most important thing 1495 * about domains is the page table mapping the DMA address space they 1496 * contain. 1497 * 1498 ****************************************************************************/ 1499 1500 static u16 domain_id_alloc(void) 1501 { 1502 int id; 1503 1504 spin_lock(&pd_bitmap_lock); 1505 id = find_first_zero_bit(amd_iommu_pd_alloc_bitmap, MAX_DOMAIN_ID); 1506 BUG_ON(id == 0); 1507 if (id > 0 && id < MAX_DOMAIN_ID) 1508 __set_bit(id, amd_iommu_pd_alloc_bitmap); 1509 else 1510 id = 0; 1511 spin_unlock(&pd_bitmap_lock); 1512 1513 return id; 1514 } 1515 1516 static void domain_id_free(int id) 1517 { 1518 spin_lock(&pd_bitmap_lock); 1519 if (id > 0 && id < MAX_DOMAIN_ID) 1520 __clear_bit(id, amd_iommu_pd_alloc_bitmap); 1521 spin_unlock(&pd_bitmap_lock); 1522 } 1523 1524 static void free_gcr3_tbl_level1(u64 *tbl) 1525 { 1526 u64 *ptr; 1527 int i; 1528 1529 for (i = 0; i < 512; ++i) { 1530 if (!(tbl[i] & GCR3_VALID)) 1531 continue; 1532 1533 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK); 1534 1535 free_page((unsigned long)ptr); 1536 } 1537 } 1538 1539 static void free_gcr3_tbl_level2(u64 *tbl) 1540 { 1541 u64 *ptr; 1542 int i; 1543 1544 for (i = 0; i < 512; ++i) { 1545 if (!(tbl[i] & GCR3_VALID)) 1546 continue; 1547 1548 ptr = iommu_phys_to_virt(tbl[i] & PAGE_MASK); 1549 1550 free_gcr3_tbl_level1(ptr); 1551 } 1552 } 1553 1554 static void free_gcr3_table(struct protection_domain *domain) 1555 { 1556 if (domain->glx == 2) 1557 free_gcr3_tbl_level2(domain->gcr3_tbl); 1558 else if (domain->glx == 1) 1559 free_gcr3_tbl_level1(domain->gcr3_tbl); 1560 else 1561 BUG_ON(domain->glx != 0); 1562 1563 free_page((unsigned long)domain->gcr3_tbl); 1564 } 1565 1566 static void set_dte_entry(struct amd_iommu *iommu, u16 devid, 1567 struct protection_domain *domain, bool ats, bool ppr) 1568 { 1569 u64 pte_root = 0; 1570 u64 flags = 0; 1571 u32 old_domid; 1572 struct dev_table_entry *dev_table = get_dev_table(iommu); 1573 1574 if (domain->iop.mode != PAGE_MODE_NONE) 1575 pte_root = iommu_virt_to_phys(domain->iop.root); 1576 1577 pte_root |= (domain->iop.mode & DEV_ENTRY_MODE_MASK) 1578 << DEV_ENTRY_MODE_SHIFT; 1579 1580 pte_root |= DTE_FLAG_IR | DTE_FLAG_IW | DTE_FLAG_V; 1581 1582 /* 1583 * When SNP is enabled, Only set TV bit when IOMMU 1584 * page translation is in use. 1585 */ 1586 if (!amd_iommu_snp_en || (domain->id != 0)) 1587 pte_root |= DTE_FLAG_TV; 1588 1589 flags = dev_table[devid].data[1]; 1590 1591 if (ats) 1592 flags |= DTE_FLAG_IOTLB; 1593 1594 if (ppr) { 1595 if (iommu_feature(iommu, FEATURE_EPHSUP)) 1596 pte_root |= 1ULL << DEV_ENTRY_PPR; 1597 } 1598 1599 if (domain->flags & PD_IOMMUV2_MASK) { 1600 u64 gcr3 = iommu_virt_to_phys(domain->gcr3_tbl); 1601 u64 glx = domain->glx; 1602 u64 tmp; 1603 1604 pte_root |= DTE_FLAG_GV; 1605 pte_root |= (glx & DTE_GLX_MASK) << DTE_GLX_SHIFT; 1606 1607 /* First mask out possible old values for GCR3 table */ 1608 tmp = DTE_GCR3_VAL_B(~0ULL) << DTE_GCR3_SHIFT_B; 1609 flags &= ~tmp; 1610 1611 tmp = DTE_GCR3_VAL_C(~0ULL) << DTE_GCR3_SHIFT_C; 1612 flags &= ~tmp; 1613 1614 /* Encode GCR3 table into DTE */ 1615 tmp = DTE_GCR3_VAL_A(gcr3) << DTE_GCR3_SHIFT_A; 1616 pte_root |= tmp; 1617 1618 tmp = DTE_GCR3_VAL_B(gcr3) << DTE_GCR3_SHIFT_B; 1619 flags |= tmp; 1620 1621 tmp = DTE_GCR3_VAL_C(gcr3) << DTE_GCR3_SHIFT_C; 1622 flags |= tmp; 1623 1624 if (amd_iommu_gpt_level == PAGE_MODE_5_LEVEL) { 1625 dev_table[devid].data[2] |= 1626 ((u64)GUEST_PGTABLE_5_LEVEL << DTE_GPT_LEVEL_SHIFT); 1627 } 1628 1629 if (domain->flags & PD_GIOV_MASK) 1630 pte_root |= DTE_FLAG_GIOV; 1631 } 1632 1633 flags &= ~DEV_DOMID_MASK; 1634 flags |= domain->id; 1635 1636 old_domid = dev_table[devid].data[1] & DEV_DOMID_MASK; 1637 dev_table[devid].data[1] = flags; 1638 dev_table[devid].data[0] = pte_root; 1639 1640 /* 1641 * A kdump kernel might be replacing a domain ID that was copied from 1642 * the previous kernel--if so, it needs to flush the translation cache 1643 * entries for the old domain ID that is being overwritten 1644 */ 1645 if (old_domid) { 1646 amd_iommu_flush_tlb_domid(iommu, old_domid); 1647 } 1648 } 1649 1650 static void clear_dte_entry(struct amd_iommu *iommu, u16 devid) 1651 { 1652 struct dev_table_entry *dev_table = get_dev_table(iommu); 1653 1654 /* remove entry from the device table seen by the hardware */ 1655 dev_table[devid].data[0] = DTE_FLAG_V; 1656 1657 if (!amd_iommu_snp_en) 1658 dev_table[devid].data[0] |= DTE_FLAG_TV; 1659 1660 dev_table[devid].data[1] &= DTE_FLAG_MASK; 1661 1662 amd_iommu_apply_erratum_63(iommu, devid); 1663 } 1664 1665 static void do_attach(struct iommu_dev_data *dev_data, 1666 struct protection_domain *domain) 1667 { 1668 struct amd_iommu *iommu; 1669 bool ats; 1670 1671 iommu = rlookup_amd_iommu(dev_data->dev); 1672 if (!iommu) 1673 return; 1674 ats = dev_data->ats.enabled; 1675 1676 /* Update data structures */ 1677 dev_data->domain = domain; 1678 list_add(&dev_data->list, &domain->dev_list); 1679 1680 /* Update NUMA Node ID */ 1681 if (domain->nid == NUMA_NO_NODE) 1682 domain->nid = dev_to_node(dev_data->dev); 1683 1684 /* Do reference counting */ 1685 domain->dev_iommu[iommu->index] += 1; 1686 domain->dev_cnt += 1; 1687 1688 /* Update device table */ 1689 set_dte_entry(iommu, dev_data->devid, domain, 1690 ats, dev_data->iommu_v2); 1691 clone_aliases(iommu, dev_data->dev); 1692 1693 device_flush_dte(dev_data); 1694 } 1695 1696 static void do_detach(struct iommu_dev_data *dev_data) 1697 { 1698 struct protection_domain *domain = dev_data->domain; 1699 struct amd_iommu *iommu; 1700 1701 iommu = rlookup_amd_iommu(dev_data->dev); 1702 if (!iommu) 1703 return; 1704 1705 /* Update data structures */ 1706 dev_data->domain = NULL; 1707 list_del(&dev_data->list); 1708 clear_dte_entry(iommu, dev_data->devid); 1709 clone_aliases(iommu, dev_data->dev); 1710 1711 /* Flush the DTE entry */ 1712 device_flush_dte(dev_data); 1713 1714 /* Flush IOTLB */ 1715 amd_iommu_domain_flush_tlb_pde(domain); 1716 1717 /* Wait for the flushes to finish */ 1718 amd_iommu_domain_flush_complete(domain); 1719 1720 /* decrease reference counters - needs to happen after the flushes */ 1721 domain->dev_iommu[iommu->index] -= 1; 1722 domain->dev_cnt -= 1; 1723 } 1724 1725 static void pdev_iommuv2_disable(struct pci_dev *pdev) 1726 { 1727 pci_disable_ats(pdev); 1728 pci_disable_pri(pdev); 1729 pci_disable_pasid(pdev); 1730 } 1731 1732 static int pdev_pri_ats_enable(struct pci_dev *pdev) 1733 { 1734 int ret; 1735 1736 /* Only allow access to user-accessible pages */ 1737 ret = pci_enable_pasid(pdev, 0); 1738 if (ret) 1739 return ret; 1740 1741 /* First reset the PRI state of the device */ 1742 ret = pci_reset_pri(pdev); 1743 if (ret) 1744 goto out_err_pasid; 1745 1746 /* Enable PRI */ 1747 /* FIXME: Hardcode number of outstanding requests for now */ 1748 ret = pci_enable_pri(pdev, 32); 1749 if (ret) 1750 goto out_err_pasid; 1751 1752 ret = pci_enable_ats(pdev, PAGE_SHIFT); 1753 if (ret) 1754 goto out_err_pri; 1755 1756 return 0; 1757 1758 out_err_pri: 1759 pci_disable_pri(pdev); 1760 1761 out_err_pasid: 1762 pci_disable_pasid(pdev); 1763 1764 return ret; 1765 } 1766 1767 /* 1768 * If a device is not yet associated with a domain, this function makes the 1769 * device visible in the domain 1770 */ 1771 static int attach_device(struct device *dev, 1772 struct protection_domain *domain) 1773 { 1774 struct iommu_dev_data *dev_data; 1775 struct pci_dev *pdev; 1776 unsigned long flags; 1777 int ret; 1778 1779 spin_lock_irqsave(&domain->lock, flags); 1780 1781 dev_data = dev_iommu_priv_get(dev); 1782 1783 spin_lock(&dev_data->lock); 1784 1785 ret = -EBUSY; 1786 if (dev_data->domain != NULL) 1787 goto out; 1788 1789 if (!dev_is_pci(dev)) 1790 goto skip_ats_check; 1791 1792 pdev = to_pci_dev(dev); 1793 if (domain->flags & PD_IOMMUV2_MASK) { 1794 struct iommu_domain *def_domain = iommu_get_dma_domain(dev); 1795 1796 ret = -EINVAL; 1797 1798 /* 1799 * In case of using AMD_IOMMU_V1 page table mode and the device 1800 * is enabling for PPR/ATS support (using v2 table), 1801 * we need to make sure that the domain type is identity map. 1802 */ 1803 if ((amd_iommu_pgtable == AMD_IOMMU_V1) && 1804 def_domain->type != IOMMU_DOMAIN_IDENTITY) { 1805 goto out; 1806 } 1807 1808 if (dev_data->iommu_v2) { 1809 if (pdev_pri_ats_enable(pdev) != 0) 1810 goto out; 1811 1812 dev_data->ats.enabled = true; 1813 dev_data->ats.qdep = pci_ats_queue_depth(pdev); 1814 dev_data->pri_tlp = pci_prg_resp_pasid_required(pdev); 1815 } 1816 } else if (amd_iommu_iotlb_sup && 1817 pci_enable_ats(pdev, PAGE_SHIFT) == 0) { 1818 dev_data->ats.enabled = true; 1819 dev_data->ats.qdep = pci_ats_queue_depth(pdev); 1820 } 1821 1822 skip_ats_check: 1823 ret = 0; 1824 1825 do_attach(dev_data, domain); 1826 1827 /* 1828 * We might boot into a crash-kernel here. The crashed kernel 1829 * left the caches in the IOMMU dirty. So we have to flush 1830 * here to evict all dirty stuff. 1831 */ 1832 amd_iommu_domain_flush_tlb_pde(domain); 1833 1834 amd_iommu_domain_flush_complete(domain); 1835 1836 out: 1837 spin_unlock(&dev_data->lock); 1838 1839 spin_unlock_irqrestore(&domain->lock, flags); 1840 1841 return ret; 1842 } 1843 1844 /* 1845 * Removes a device from a protection domain (with devtable_lock held) 1846 */ 1847 static void detach_device(struct device *dev) 1848 { 1849 struct protection_domain *domain; 1850 struct iommu_dev_data *dev_data; 1851 unsigned long flags; 1852 1853 dev_data = dev_iommu_priv_get(dev); 1854 domain = dev_data->domain; 1855 1856 spin_lock_irqsave(&domain->lock, flags); 1857 1858 spin_lock(&dev_data->lock); 1859 1860 /* 1861 * First check if the device is still attached. It might already 1862 * be detached from its domain because the generic 1863 * iommu_detach_group code detached it and we try again here in 1864 * our alias handling. 1865 */ 1866 if (WARN_ON(!dev_data->domain)) 1867 goto out; 1868 1869 do_detach(dev_data); 1870 1871 if (!dev_is_pci(dev)) 1872 goto out; 1873 1874 if (domain->flags & PD_IOMMUV2_MASK && dev_data->iommu_v2) 1875 pdev_iommuv2_disable(to_pci_dev(dev)); 1876 else if (dev_data->ats.enabled) 1877 pci_disable_ats(to_pci_dev(dev)); 1878 1879 dev_data->ats.enabled = false; 1880 1881 out: 1882 spin_unlock(&dev_data->lock); 1883 1884 spin_unlock_irqrestore(&domain->lock, flags); 1885 } 1886 1887 static struct iommu_device *amd_iommu_probe_device(struct device *dev) 1888 { 1889 struct iommu_device *iommu_dev; 1890 struct amd_iommu *iommu; 1891 int ret; 1892 1893 if (!check_device(dev)) 1894 return ERR_PTR(-ENODEV); 1895 1896 iommu = rlookup_amd_iommu(dev); 1897 if (!iommu) 1898 return ERR_PTR(-ENODEV); 1899 1900 /* Not registered yet? */ 1901 if (!iommu->iommu.ops) 1902 return ERR_PTR(-ENODEV); 1903 1904 if (dev_iommu_priv_get(dev)) 1905 return &iommu->iommu; 1906 1907 ret = iommu_init_device(iommu, dev); 1908 if (ret) { 1909 if (ret != -ENOTSUPP) 1910 dev_err(dev, "Failed to initialize - trying to proceed anyway\n"); 1911 iommu_dev = ERR_PTR(ret); 1912 iommu_ignore_device(iommu, dev); 1913 } else { 1914 amd_iommu_set_pci_msi_domain(dev, iommu); 1915 iommu_dev = &iommu->iommu; 1916 } 1917 1918 iommu_completion_wait(iommu); 1919 1920 return iommu_dev; 1921 } 1922 1923 static void amd_iommu_probe_finalize(struct device *dev) 1924 { 1925 /* Domains are initialized for this device - have a look what we ended up with */ 1926 set_dma_ops(dev, NULL); 1927 iommu_setup_dma_ops(dev, 0, U64_MAX); 1928 } 1929 1930 static void amd_iommu_release_device(struct device *dev) 1931 { 1932 struct amd_iommu *iommu; 1933 1934 if (!check_device(dev)) 1935 return; 1936 1937 iommu = rlookup_amd_iommu(dev); 1938 if (!iommu) 1939 return; 1940 1941 amd_iommu_uninit_device(dev); 1942 iommu_completion_wait(iommu); 1943 } 1944 1945 static struct iommu_group *amd_iommu_device_group(struct device *dev) 1946 { 1947 if (dev_is_pci(dev)) 1948 return pci_device_group(dev); 1949 1950 return acpihid_device_group(dev); 1951 } 1952 1953 /***************************************************************************** 1954 * 1955 * The next functions belong to the dma_ops mapping/unmapping code. 1956 * 1957 *****************************************************************************/ 1958 1959 static void update_device_table(struct protection_domain *domain) 1960 { 1961 struct iommu_dev_data *dev_data; 1962 1963 list_for_each_entry(dev_data, &domain->dev_list, list) { 1964 struct amd_iommu *iommu = rlookup_amd_iommu(dev_data->dev); 1965 1966 if (!iommu) 1967 continue; 1968 set_dte_entry(iommu, dev_data->devid, domain, 1969 dev_data->ats.enabled, dev_data->iommu_v2); 1970 clone_aliases(iommu, dev_data->dev); 1971 } 1972 } 1973 1974 void amd_iommu_update_and_flush_device_table(struct protection_domain *domain) 1975 { 1976 update_device_table(domain); 1977 domain_flush_devices(domain); 1978 } 1979 1980 void amd_iommu_domain_update(struct protection_domain *domain) 1981 { 1982 /* Update device table */ 1983 amd_iommu_update_and_flush_device_table(domain); 1984 1985 /* Flush domain TLB(s) and wait for completion */ 1986 amd_iommu_domain_flush_tlb_pde(domain); 1987 amd_iommu_domain_flush_complete(domain); 1988 } 1989 1990 /***************************************************************************** 1991 * 1992 * The following functions belong to the exported interface of AMD IOMMU 1993 * 1994 * This interface allows access to lower level functions of the IOMMU 1995 * like protection domain handling and assignement of devices to domains 1996 * which is not possible with the dma_ops interface. 1997 * 1998 *****************************************************************************/ 1999 2000 static void cleanup_domain(struct protection_domain *domain) 2001 { 2002 struct iommu_dev_data *entry; 2003 unsigned long flags; 2004 2005 spin_lock_irqsave(&domain->lock, flags); 2006 2007 while (!list_empty(&domain->dev_list)) { 2008 entry = list_first_entry(&domain->dev_list, 2009 struct iommu_dev_data, list); 2010 BUG_ON(!entry->domain); 2011 do_detach(entry); 2012 } 2013 2014 spin_unlock_irqrestore(&domain->lock, flags); 2015 } 2016 2017 static void protection_domain_free(struct protection_domain *domain) 2018 { 2019 if (!domain) 2020 return; 2021 2022 if (domain->iop.pgtbl_cfg.tlb) 2023 free_io_pgtable_ops(&domain->iop.iop.ops); 2024 2025 if (domain->id) 2026 domain_id_free(domain->id); 2027 2028 kfree(domain); 2029 } 2030 2031 static int protection_domain_init_v1(struct protection_domain *domain, int mode) 2032 { 2033 u64 *pt_root = NULL; 2034 2035 BUG_ON(mode < PAGE_MODE_NONE || mode > PAGE_MODE_6_LEVEL); 2036 2037 spin_lock_init(&domain->lock); 2038 domain->id = domain_id_alloc(); 2039 if (!domain->id) 2040 return -ENOMEM; 2041 INIT_LIST_HEAD(&domain->dev_list); 2042 2043 if (mode != PAGE_MODE_NONE) { 2044 pt_root = (void *)get_zeroed_page(GFP_KERNEL); 2045 if (!pt_root) { 2046 domain_id_free(domain->id); 2047 return -ENOMEM; 2048 } 2049 } 2050 2051 amd_iommu_domain_set_pgtable(domain, pt_root, mode); 2052 2053 return 0; 2054 } 2055 2056 static int protection_domain_init_v2(struct protection_domain *domain) 2057 { 2058 spin_lock_init(&domain->lock); 2059 domain->id = domain_id_alloc(); 2060 if (!domain->id) 2061 return -ENOMEM; 2062 INIT_LIST_HEAD(&domain->dev_list); 2063 2064 domain->flags |= PD_GIOV_MASK; 2065 2066 domain->domain.pgsize_bitmap = AMD_IOMMU_PGSIZES_V2; 2067 2068 if (domain_enable_v2(domain, 1)) { 2069 domain_id_free(domain->id); 2070 return -ENOMEM; 2071 } 2072 2073 return 0; 2074 } 2075 2076 static struct protection_domain *protection_domain_alloc(unsigned int type) 2077 { 2078 struct io_pgtable_ops *pgtbl_ops; 2079 struct protection_domain *domain; 2080 int pgtable; 2081 int mode = DEFAULT_PGTABLE_LEVEL; 2082 int ret; 2083 2084 /* 2085 * Force IOMMU v1 page table when iommu=pt and 2086 * when allocating domain for pass-through devices. 2087 */ 2088 if (type == IOMMU_DOMAIN_IDENTITY) { 2089 pgtable = AMD_IOMMU_V1; 2090 mode = PAGE_MODE_NONE; 2091 } else if (type == IOMMU_DOMAIN_UNMANAGED) { 2092 pgtable = AMD_IOMMU_V1; 2093 } else if (type == IOMMU_DOMAIN_DMA || type == IOMMU_DOMAIN_DMA_FQ) { 2094 pgtable = amd_iommu_pgtable; 2095 } else { 2096 return NULL; 2097 } 2098 2099 domain = kzalloc(sizeof(*domain), GFP_KERNEL); 2100 if (!domain) 2101 return NULL; 2102 2103 switch (pgtable) { 2104 case AMD_IOMMU_V1: 2105 ret = protection_domain_init_v1(domain, mode); 2106 break; 2107 case AMD_IOMMU_V2: 2108 ret = protection_domain_init_v2(domain); 2109 break; 2110 default: 2111 ret = -EINVAL; 2112 } 2113 2114 if (ret) 2115 goto out_err; 2116 2117 /* No need to allocate io pgtable ops in passthrough mode */ 2118 if (type == IOMMU_DOMAIN_IDENTITY) 2119 return domain; 2120 2121 domain->nid = NUMA_NO_NODE; 2122 2123 pgtbl_ops = alloc_io_pgtable_ops(pgtable, &domain->iop.pgtbl_cfg, domain); 2124 if (!pgtbl_ops) { 2125 domain_id_free(domain->id); 2126 goto out_err; 2127 } 2128 2129 return domain; 2130 out_err: 2131 kfree(domain); 2132 return NULL; 2133 } 2134 2135 static inline u64 dma_max_address(void) 2136 { 2137 if (amd_iommu_pgtable == AMD_IOMMU_V1) 2138 return ~0ULL; 2139 2140 /* V2 with 4/5 level page table */ 2141 return ((1ULL << PM_LEVEL_SHIFT(amd_iommu_gpt_level)) - 1); 2142 } 2143 2144 static struct iommu_domain *amd_iommu_domain_alloc(unsigned type) 2145 { 2146 struct protection_domain *domain; 2147 2148 /* 2149 * Since DTE[Mode]=0 is prohibited on SNP-enabled system, 2150 * default to use IOMMU_DOMAIN_DMA[_FQ]. 2151 */ 2152 if (amd_iommu_snp_en && (type == IOMMU_DOMAIN_IDENTITY)) 2153 return NULL; 2154 2155 domain = protection_domain_alloc(type); 2156 if (!domain) 2157 return NULL; 2158 2159 domain->domain.geometry.aperture_start = 0; 2160 domain->domain.geometry.aperture_end = dma_max_address(); 2161 domain->domain.geometry.force_aperture = true; 2162 2163 return &domain->domain; 2164 } 2165 2166 static void amd_iommu_domain_free(struct iommu_domain *dom) 2167 { 2168 struct protection_domain *domain; 2169 2170 domain = to_pdomain(dom); 2171 2172 if (domain->dev_cnt > 0) 2173 cleanup_domain(domain); 2174 2175 BUG_ON(domain->dev_cnt != 0); 2176 2177 if (!dom) 2178 return; 2179 2180 if (domain->flags & PD_IOMMUV2_MASK) 2181 free_gcr3_table(domain); 2182 2183 protection_domain_free(domain); 2184 } 2185 2186 static int amd_iommu_attach_device(struct iommu_domain *dom, 2187 struct device *dev) 2188 { 2189 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev); 2190 struct protection_domain *domain = to_pdomain(dom); 2191 struct amd_iommu *iommu = rlookup_amd_iommu(dev); 2192 int ret; 2193 2194 /* 2195 * Skip attach device to domain if new domain is same as 2196 * devices current domain 2197 */ 2198 if (dev_data->domain == domain) 2199 return 0; 2200 2201 dev_data->defer_attach = false; 2202 2203 if (dev_data->domain) 2204 detach_device(dev); 2205 2206 ret = attach_device(dev, domain); 2207 2208 #ifdef CONFIG_IRQ_REMAP 2209 if (AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) { 2210 if (dom->type == IOMMU_DOMAIN_UNMANAGED) 2211 dev_data->use_vapic = 1; 2212 else 2213 dev_data->use_vapic = 0; 2214 } 2215 #endif 2216 2217 iommu_completion_wait(iommu); 2218 2219 return ret; 2220 } 2221 2222 static void amd_iommu_iotlb_sync_map(struct iommu_domain *dom, 2223 unsigned long iova, size_t size) 2224 { 2225 struct protection_domain *domain = to_pdomain(dom); 2226 struct io_pgtable_ops *ops = &domain->iop.iop.ops; 2227 2228 if (ops->map_pages) 2229 domain_flush_np_cache(domain, iova, size); 2230 } 2231 2232 static int amd_iommu_map_pages(struct iommu_domain *dom, unsigned long iova, 2233 phys_addr_t paddr, size_t pgsize, size_t pgcount, 2234 int iommu_prot, gfp_t gfp, size_t *mapped) 2235 { 2236 struct protection_domain *domain = to_pdomain(dom); 2237 struct io_pgtable_ops *ops = &domain->iop.iop.ops; 2238 int prot = 0; 2239 int ret = -EINVAL; 2240 2241 if ((amd_iommu_pgtable == AMD_IOMMU_V1) && 2242 (domain->iop.mode == PAGE_MODE_NONE)) 2243 return -EINVAL; 2244 2245 if (iommu_prot & IOMMU_READ) 2246 prot |= IOMMU_PROT_IR; 2247 if (iommu_prot & IOMMU_WRITE) 2248 prot |= IOMMU_PROT_IW; 2249 2250 if (ops->map_pages) { 2251 ret = ops->map_pages(ops, iova, paddr, pgsize, 2252 pgcount, prot, gfp, mapped); 2253 } 2254 2255 return ret; 2256 } 2257 2258 static void amd_iommu_iotlb_gather_add_page(struct iommu_domain *domain, 2259 struct iommu_iotlb_gather *gather, 2260 unsigned long iova, size_t size) 2261 { 2262 /* 2263 * AMD's IOMMU can flush as many pages as necessary in a single flush. 2264 * Unless we run in a virtual machine, which can be inferred according 2265 * to whether "non-present cache" is on, it is probably best to prefer 2266 * (potentially) too extensive TLB flushing (i.e., more misses) over 2267 * mutliple TLB flushes (i.e., more flushes). For virtual machines the 2268 * hypervisor needs to synchronize the host IOMMU PTEs with those of 2269 * the guest, and the trade-off is different: unnecessary TLB flushes 2270 * should be avoided. 2271 */ 2272 if (amd_iommu_np_cache && 2273 iommu_iotlb_gather_is_disjoint(gather, iova, size)) 2274 iommu_iotlb_sync(domain, gather); 2275 2276 iommu_iotlb_gather_add_range(gather, iova, size); 2277 } 2278 2279 static size_t amd_iommu_unmap_pages(struct iommu_domain *dom, unsigned long iova, 2280 size_t pgsize, size_t pgcount, 2281 struct iommu_iotlb_gather *gather) 2282 { 2283 struct protection_domain *domain = to_pdomain(dom); 2284 struct io_pgtable_ops *ops = &domain->iop.iop.ops; 2285 size_t r; 2286 2287 if ((amd_iommu_pgtable == AMD_IOMMU_V1) && 2288 (domain->iop.mode == PAGE_MODE_NONE)) 2289 return 0; 2290 2291 r = (ops->unmap_pages) ? ops->unmap_pages(ops, iova, pgsize, pgcount, NULL) : 0; 2292 2293 if (r) 2294 amd_iommu_iotlb_gather_add_page(dom, gather, iova, r); 2295 2296 return r; 2297 } 2298 2299 static phys_addr_t amd_iommu_iova_to_phys(struct iommu_domain *dom, 2300 dma_addr_t iova) 2301 { 2302 struct protection_domain *domain = to_pdomain(dom); 2303 struct io_pgtable_ops *ops = &domain->iop.iop.ops; 2304 2305 return ops->iova_to_phys(ops, iova); 2306 } 2307 2308 static bool amd_iommu_capable(struct device *dev, enum iommu_cap cap) 2309 { 2310 switch (cap) { 2311 case IOMMU_CAP_CACHE_COHERENCY: 2312 return true; 2313 case IOMMU_CAP_NOEXEC: 2314 return false; 2315 case IOMMU_CAP_PRE_BOOT_PROTECTION: 2316 return amdr_ivrs_remap_support; 2317 case IOMMU_CAP_ENFORCE_CACHE_COHERENCY: 2318 return true; 2319 case IOMMU_CAP_DEFERRED_FLUSH: 2320 return true; 2321 default: 2322 break; 2323 } 2324 2325 return false; 2326 } 2327 2328 static void amd_iommu_get_resv_regions(struct device *dev, 2329 struct list_head *head) 2330 { 2331 struct iommu_resv_region *region; 2332 struct unity_map_entry *entry; 2333 struct amd_iommu *iommu; 2334 struct amd_iommu_pci_seg *pci_seg; 2335 int devid, sbdf; 2336 2337 sbdf = get_device_sbdf_id(dev); 2338 if (sbdf < 0) 2339 return; 2340 2341 devid = PCI_SBDF_TO_DEVID(sbdf); 2342 iommu = rlookup_amd_iommu(dev); 2343 if (!iommu) 2344 return; 2345 pci_seg = iommu->pci_seg; 2346 2347 list_for_each_entry(entry, &pci_seg->unity_map, list) { 2348 int type, prot = 0; 2349 size_t length; 2350 2351 if (devid < entry->devid_start || devid > entry->devid_end) 2352 continue; 2353 2354 type = IOMMU_RESV_DIRECT; 2355 length = entry->address_end - entry->address_start; 2356 if (entry->prot & IOMMU_PROT_IR) 2357 prot |= IOMMU_READ; 2358 if (entry->prot & IOMMU_PROT_IW) 2359 prot |= IOMMU_WRITE; 2360 if (entry->prot & IOMMU_UNITY_MAP_FLAG_EXCL_RANGE) 2361 /* Exclusion range */ 2362 type = IOMMU_RESV_RESERVED; 2363 2364 region = iommu_alloc_resv_region(entry->address_start, 2365 length, prot, type, 2366 GFP_KERNEL); 2367 if (!region) { 2368 dev_err(dev, "Out of memory allocating dm-regions\n"); 2369 return; 2370 } 2371 list_add_tail(®ion->list, head); 2372 } 2373 2374 region = iommu_alloc_resv_region(MSI_RANGE_START, 2375 MSI_RANGE_END - MSI_RANGE_START + 1, 2376 0, IOMMU_RESV_MSI, GFP_KERNEL); 2377 if (!region) 2378 return; 2379 list_add_tail(®ion->list, head); 2380 2381 region = iommu_alloc_resv_region(HT_RANGE_START, 2382 HT_RANGE_END - HT_RANGE_START + 1, 2383 0, IOMMU_RESV_RESERVED, GFP_KERNEL); 2384 if (!region) 2385 return; 2386 list_add_tail(®ion->list, head); 2387 } 2388 2389 bool amd_iommu_is_attach_deferred(struct device *dev) 2390 { 2391 struct iommu_dev_data *dev_data = dev_iommu_priv_get(dev); 2392 2393 return dev_data->defer_attach; 2394 } 2395 EXPORT_SYMBOL_GPL(amd_iommu_is_attach_deferred); 2396 2397 static void amd_iommu_flush_iotlb_all(struct iommu_domain *domain) 2398 { 2399 struct protection_domain *dom = to_pdomain(domain); 2400 unsigned long flags; 2401 2402 spin_lock_irqsave(&dom->lock, flags); 2403 amd_iommu_domain_flush_tlb_pde(dom); 2404 amd_iommu_domain_flush_complete(dom); 2405 spin_unlock_irqrestore(&dom->lock, flags); 2406 } 2407 2408 static void amd_iommu_iotlb_sync(struct iommu_domain *domain, 2409 struct iommu_iotlb_gather *gather) 2410 { 2411 struct protection_domain *dom = to_pdomain(domain); 2412 unsigned long flags; 2413 2414 spin_lock_irqsave(&dom->lock, flags); 2415 domain_flush_pages(dom, gather->start, gather->end - gather->start + 1, 1); 2416 amd_iommu_domain_flush_complete(dom); 2417 spin_unlock_irqrestore(&dom->lock, flags); 2418 } 2419 2420 static int amd_iommu_def_domain_type(struct device *dev) 2421 { 2422 struct iommu_dev_data *dev_data; 2423 2424 dev_data = dev_iommu_priv_get(dev); 2425 if (!dev_data) 2426 return 0; 2427 2428 /* 2429 * Do not identity map IOMMUv2 capable devices when: 2430 * - memory encryption is active, because some of those devices 2431 * (AMD GPUs) don't have the encryption bit in their DMA-mask 2432 * and require remapping. 2433 * - SNP is enabled, because it prohibits DTE[Mode]=0. 2434 */ 2435 if (dev_data->iommu_v2 && 2436 !cc_platform_has(CC_ATTR_MEM_ENCRYPT) && 2437 !amd_iommu_snp_en) { 2438 return IOMMU_DOMAIN_IDENTITY; 2439 } 2440 2441 return 0; 2442 } 2443 2444 static bool amd_iommu_enforce_cache_coherency(struct iommu_domain *domain) 2445 { 2446 /* IOMMU_PTE_FC is always set */ 2447 return true; 2448 } 2449 2450 const struct iommu_ops amd_iommu_ops = { 2451 .capable = amd_iommu_capable, 2452 .domain_alloc = amd_iommu_domain_alloc, 2453 .probe_device = amd_iommu_probe_device, 2454 .release_device = amd_iommu_release_device, 2455 .probe_finalize = amd_iommu_probe_finalize, 2456 .device_group = amd_iommu_device_group, 2457 .get_resv_regions = amd_iommu_get_resv_regions, 2458 .is_attach_deferred = amd_iommu_is_attach_deferred, 2459 .pgsize_bitmap = AMD_IOMMU_PGSIZES, 2460 .def_domain_type = amd_iommu_def_domain_type, 2461 .default_domain_ops = &(const struct iommu_domain_ops) { 2462 .attach_dev = amd_iommu_attach_device, 2463 .map_pages = amd_iommu_map_pages, 2464 .unmap_pages = amd_iommu_unmap_pages, 2465 .iotlb_sync_map = amd_iommu_iotlb_sync_map, 2466 .iova_to_phys = amd_iommu_iova_to_phys, 2467 .flush_iotlb_all = amd_iommu_flush_iotlb_all, 2468 .iotlb_sync = amd_iommu_iotlb_sync, 2469 .free = amd_iommu_domain_free, 2470 .enforce_cache_coherency = amd_iommu_enforce_cache_coherency, 2471 } 2472 }; 2473 2474 /***************************************************************************** 2475 * 2476 * The next functions do a basic initialization of IOMMU for pass through 2477 * mode 2478 * 2479 * In passthrough mode the IOMMU is initialized and enabled but not used for 2480 * DMA-API translation. 2481 * 2482 *****************************************************************************/ 2483 2484 /* IOMMUv2 specific functions */ 2485 int amd_iommu_register_ppr_notifier(struct notifier_block *nb) 2486 { 2487 return atomic_notifier_chain_register(&ppr_notifier, nb); 2488 } 2489 EXPORT_SYMBOL(amd_iommu_register_ppr_notifier); 2490 2491 int amd_iommu_unregister_ppr_notifier(struct notifier_block *nb) 2492 { 2493 return atomic_notifier_chain_unregister(&ppr_notifier, nb); 2494 } 2495 EXPORT_SYMBOL(amd_iommu_unregister_ppr_notifier); 2496 2497 void amd_iommu_domain_direct_map(struct iommu_domain *dom) 2498 { 2499 struct protection_domain *domain = to_pdomain(dom); 2500 unsigned long flags; 2501 2502 spin_lock_irqsave(&domain->lock, flags); 2503 2504 if (domain->iop.pgtbl_cfg.tlb) 2505 free_io_pgtable_ops(&domain->iop.iop.ops); 2506 2507 spin_unlock_irqrestore(&domain->lock, flags); 2508 } 2509 EXPORT_SYMBOL(amd_iommu_domain_direct_map); 2510 2511 /* Note: This function expects iommu_domain->lock to be held prior calling the function. */ 2512 static int domain_enable_v2(struct protection_domain *domain, int pasids) 2513 { 2514 int levels; 2515 2516 /* Number of GCR3 table levels required */ 2517 for (levels = 0; (pasids - 1) & ~0x1ff; pasids >>= 9) 2518 levels += 1; 2519 2520 if (levels > amd_iommu_max_glx_val) 2521 return -EINVAL; 2522 2523 domain->gcr3_tbl = (void *)get_zeroed_page(GFP_ATOMIC); 2524 if (domain->gcr3_tbl == NULL) 2525 return -ENOMEM; 2526 2527 domain->glx = levels; 2528 domain->flags |= PD_IOMMUV2_MASK; 2529 2530 amd_iommu_domain_update(domain); 2531 2532 return 0; 2533 } 2534 2535 int amd_iommu_domain_enable_v2(struct iommu_domain *dom, int pasids) 2536 { 2537 struct protection_domain *pdom = to_pdomain(dom); 2538 unsigned long flags; 2539 int ret; 2540 2541 spin_lock_irqsave(&pdom->lock, flags); 2542 2543 /* 2544 * Save us all sanity checks whether devices already in the 2545 * domain support IOMMUv2. Just force that the domain has no 2546 * devices attached when it is switched into IOMMUv2 mode. 2547 */ 2548 ret = -EBUSY; 2549 if (pdom->dev_cnt > 0 || pdom->flags & PD_IOMMUV2_MASK) 2550 goto out; 2551 2552 if (!pdom->gcr3_tbl) 2553 ret = domain_enable_v2(pdom, pasids); 2554 2555 out: 2556 spin_unlock_irqrestore(&pdom->lock, flags); 2557 return ret; 2558 } 2559 EXPORT_SYMBOL(amd_iommu_domain_enable_v2); 2560 2561 static int __flush_pasid(struct protection_domain *domain, u32 pasid, 2562 u64 address, bool size) 2563 { 2564 struct iommu_dev_data *dev_data; 2565 struct iommu_cmd cmd; 2566 int i, ret; 2567 2568 if (!(domain->flags & PD_IOMMUV2_MASK)) 2569 return -EINVAL; 2570 2571 build_inv_iommu_pasid(&cmd, domain->id, pasid, address, size); 2572 2573 /* 2574 * IOMMU TLB needs to be flushed before Device TLB to 2575 * prevent device TLB refill from IOMMU TLB 2576 */ 2577 for (i = 0; i < amd_iommu_get_num_iommus(); ++i) { 2578 if (domain->dev_iommu[i] == 0) 2579 continue; 2580 2581 ret = iommu_queue_command(amd_iommus[i], &cmd); 2582 if (ret != 0) 2583 goto out; 2584 } 2585 2586 /* Wait until IOMMU TLB flushes are complete */ 2587 amd_iommu_domain_flush_complete(domain); 2588 2589 /* Now flush device TLBs */ 2590 list_for_each_entry(dev_data, &domain->dev_list, list) { 2591 struct amd_iommu *iommu; 2592 int qdep; 2593 2594 /* 2595 There might be non-IOMMUv2 capable devices in an IOMMUv2 2596 * domain. 2597 */ 2598 if (!dev_data->ats.enabled) 2599 continue; 2600 2601 qdep = dev_data->ats.qdep; 2602 iommu = rlookup_amd_iommu(dev_data->dev); 2603 if (!iommu) 2604 continue; 2605 build_inv_iotlb_pasid(&cmd, dev_data->devid, pasid, 2606 qdep, address, size); 2607 2608 ret = iommu_queue_command(iommu, &cmd); 2609 if (ret != 0) 2610 goto out; 2611 } 2612 2613 /* Wait until all device TLBs are flushed */ 2614 amd_iommu_domain_flush_complete(domain); 2615 2616 ret = 0; 2617 2618 out: 2619 2620 return ret; 2621 } 2622 2623 static int __amd_iommu_flush_page(struct protection_domain *domain, u32 pasid, 2624 u64 address) 2625 { 2626 return __flush_pasid(domain, pasid, address, false); 2627 } 2628 2629 int amd_iommu_flush_page(struct iommu_domain *dom, u32 pasid, 2630 u64 address) 2631 { 2632 struct protection_domain *domain = to_pdomain(dom); 2633 unsigned long flags; 2634 int ret; 2635 2636 spin_lock_irqsave(&domain->lock, flags); 2637 ret = __amd_iommu_flush_page(domain, pasid, address); 2638 spin_unlock_irqrestore(&domain->lock, flags); 2639 2640 return ret; 2641 } 2642 EXPORT_SYMBOL(amd_iommu_flush_page); 2643 2644 static int __amd_iommu_flush_tlb(struct protection_domain *domain, u32 pasid) 2645 { 2646 return __flush_pasid(domain, pasid, CMD_INV_IOMMU_ALL_PAGES_ADDRESS, 2647 true); 2648 } 2649 2650 int amd_iommu_flush_tlb(struct iommu_domain *dom, u32 pasid) 2651 { 2652 struct protection_domain *domain = to_pdomain(dom); 2653 unsigned long flags; 2654 int ret; 2655 2656 spin_lock_irqsave(&domain->lock, flags); 2657 ret = __amd_iommu_flush_tlb(domain, pasid); 2658 spin_unlock_irqrestore(&domain->lock, flags); 2659 2660 return ret; 2661 } 2662 EXPORT_SYMBOL(amd_iommu_flush_tlb); 2663 2664 static u64 *__get_gcr3_pte(u64 *root, int level, u32 pasid, bool alloc) 2665 { 2666 int index; 2667 u64 *pte; 2668 2669 while (true) { 2670 2671 index = (pasid >> (9 * level)) & 0x1ff; 2672 pte = &root[index]; 2673 2674 if (level == 0) 2675 break; 2676 2677 if (!(*pte & GCR3_VALID)) { 2678 if (!alloc) 2679 return NULL; 2680 2681 root = (void *)get_zeroed_page(GFP_ATOMIC); 2682 if (root == NULL) 2683 return NULL; 2684 2685 *pte = iommu_virt_to_phys(root) | GCR3_VALID; 2686 } 2687 2688 root = iommu_phys_to_virt(*pte & PAGE_MASK); 2689 2690 level -= 1; 2691 } 2692 2693 return pte; 2694 } 2695 2696 static int __set_gcr3(struct protection_domain *domain, u32 pasid, 2697 unsigned long cr3) 2698 { 2699 u64 *pte; 2700 2701 if (domain->iop.mode != PAGE_MODE_NONE) 2702 return -EINVAL; 2703 2704 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, true); 2705 if (pte == NULL) 2706 return -ENOMEM; 2707 2708 *pte = (cr3 & PAGE_MASK) | GCR3_VALID; 2709 2710 return __amd_iommu_flush_tlb(domain, pasid); 2711 } 2712 2713 static int __clear_gcr3(struct protection_domain *domain, u32 pasid) 2714 { 2715 u64 *pte; 2716 2717 if (domain->iop.mode != PAGE_MODE_NONE) 2718 return -EINVAL; 2719 2720 pte = __get_gcr3_pte(domain->gcr3_tbl, domain->glx, pasid, false); 2721 if (pte == NULL) 2722 return 0; 2723 2724 *pte = 0; 2725 2726 return __amd_iommu_flush_tlb(domain, pasid); 2727 } 2728 2729 int amd_iommu_domain_set_gcr3(struct iommu_domain *dom, u32 pasid, 2730 unsigned long cr3) 2731 { 2732 struct protection_domain *domain = to_pdomain(dom); 2733 unsigned long flags; 2734 int ret; 2735 2736 spin_lock_irqsave(&domain->lock, flags); 2737 ret = __set_gcr3(domain, pasid, cr3); 2738 spin_unlock_irqrestore(&domain->lock, flags); 2739 2740 return ret; 2741 } 2742 EXPORT_SYMBOL(amd_iommu_domain_set_gcr3); 2743 2744 int amd_iommu_domain_clear_gcr3(struct iommu_domain *dom, u32 pasid) 2745 { 2746 struct protection_domain *domain = to_pdomain(dom); 2747 unsigned long flags; 2748 int ret; 2749 2750 spin_lock_irqsave(&domain->lock, flags); 2751 ret = __clear_gcr3(domain, pasid); 2752 spin_unlock_irqrestore(&domain->lock, flags); 2753 2754 return ret; 2755 } 2756 EXPORT_SYMBOL(amd_iommu_domain_clear_gcr3); 2757 2758 int amd_iommu_complete_ppr(struct pci_dev *pdev, u32 pasid, 2759 int status, int tag) 2760 { 2761 struct iommu_dev_data *dev_data; 2762 struct amd_iommu *iommu; 2763 struct iommu_cmd cmd; 2764 2765 dev_data = dev_iommu_priv_get(&pdev->dev); 2766 iommu = rlookup_amd_iommu(&pdev->dev); 2767 if (!iommu) 2768 return -ENODEV; 2769 2770 build_complete_ppr(&cmd, dev_data->devid, pasid, status, 2771 tag, dev_data->pri_tlp); 2772 2773 return iommu_queue_command(iommu, &cmd); 2774 } 2775 EXPORT_SYMBOL(amd_iommu_complete_ppr); 2776 2777 int amd_iommu_device_info(struct pci_dev *pdev, 2778 struct amd_iommu_device_info *info) 2779 { 2780 int max_pasids; 2781 int pos; 2782 2783 if (pdev == NULL || info == NULL) 2784 return -EINVAL; 2785 2786 if (!amd_iommu_v2_supported()) 2787 return -EINVAL; 2788 2789 memset(info, 0, sizeof(*info)); 2790 2791 if (pci_ats_supported(pdev)) 2792 info->flags |= AMD_IOMMU_DEVICE_FLAG_ATS_SUP; 2793 2794 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PRI); 2795 if (pos) 2796 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRI_SUP; 2797 2798 pos = pci_find_ext_capability(pdev, PCI_EXT_CAP_ID_PASID); 2799 if (pos) { 2800 int features; 2801 2802 max_pasids = 1 << (9 * (amd_iommu_max_glx_val + 1)); 2803 max_pasids = min(max_pasids, (1 << 20)); 2804 2805 info->flags |= AMD_IOMMU_DEVICE_FLAG_PASID_SUP; 2806 info->max_pasids = min(pci_max_pasids(pdev), max_pasids); 2807 2808 features = pci_pasid_features(pdev); 2809 if (features & PCI_PASID_CAP_EXEC) 2810 info->flags |= AMD_IOMMU_DEVICE_FLAG_EXEC_SUP; 2811 if (features & PCI_PASID_CAP_PRIV) 2812 info->flags |= AMD_IOMMU_DEVICE_FLAG_PRIV_SUP; 2813 } 2814 2815 return 0; 2816 } 2817 EXPORT_SYMBOL(amd_iommu_device_info); 2818 2819 #ifdef CONFIG_IRQ_REMAP 2820 2821 /***************************************************************************** 2822 * 2823 * Interrupt Remapping Implementation 2824 * 2825 *****************************************************************************/ 2826 2827 static struct irq_chip amd_ir_chip; 2828 static DEFINE_SPINLOCK(iommu_table_lock); 2829 2830 static void iommu_flush_irt_and_complete(struct amd_iommu *iommu, u16 devid) 2831 { 2832 int ret; 2833 u64 data; 2834 unsigned long flags; 2835 struct iommu_cmd cmd, cmd2; 2836 2837 if (iommu->irtcachedis_enabled) 2838 return; 2839 2840 build_inv_irt(&cmd, devid); 2841 data = atomic64_add_return(1, &iommu->cmd_sem_val); 2842 build_completion_wait(&cmd2, iommu, data); 2843 2844 raw_spin_lock_irqsave(&iommu->lock, flags); 2845 ret = __iommu_queue_command_sync(iommu, &cmd, true); 2846 if (ret) 2847 goto out; 2848 ret = __iommu_queue_command_sync(iommu, &cmd2, false); 2849 if (ret) 2850 goto out; 2851 wait_on_sem(iommu, data); 2852 out: 2853 raw_spin_unlock_irqrestore(&iommu->lock, flags); 2854 } 2855 2856 static void set_dte_irq_entry(struct amd_iommu *iommu, u16 devid, 2857 struct irq_remap_table *table) 2858 { 2859 u64 dte; 2860 struct dev_table_entry *dev_table = get_dev_table(iommu); 2861 2862 dte = dev_table[devid].data[2]; 2863 dte &= ~DTE_IRQ_PHYS_ADDR_MASK; 2864 dte |= iommu_virt_to_phys(table->table); 2865 dte |= DTE_IRQ_REMAP_INTCTL; 2866 dte |= DTE_INTTABLEN; 2867 dte |= DTE_IRQ_REMAP_ENABLE; 2868 2869 dev_table[devid].data[2] = dte; 2870 } 2871 2872 static struct irq_remap_table *get_irq_table(struct amd_iommu *iommu, u16 devid) 2873 { 2874 struct irq_remap_table *table; 2875 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 2876 2877 if (WARN_ONCE(!pci_seg->rlookup_table[devid], 2878 "%s: no iommu for devid %x:%x\n", 2879 __func__, pci_seg->id, devid)) 2880 return NULL; 2881 2882 table = pci_seg->irq_lookup_table[devid]; 2883 if (WARN_ONCE(!table, "%s: no table for devid %x:%x\n", 2884 __func__, pci_seg->id, devid)) 2885 return NULL; 2886 2887 return table; 2888 } 2889 2890 static struct irq_remap_table *__alloc_irq_table(void) 2891 { 2892 struct irq_remap_table *table; 2893 2894 table = kzalloc(sizeof(*table), GFP_KERNEL); 2895 if (!table) 2896 return NULL; 2897 2898 table->table = kmem_cache_alloc(amd_iommu_irq_cache, GFP_KERNEL); 2899 if (!table->table) { 2900 kfree(table); 2901 return NULL; 2902 } 2903 raw_spin_lock_init(&table->lock); 2904 2905 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir)) 2906 memset(table->table, 0, 2907 MAX_IRQS_PER_TABLE * sizeof(u32)); 2908 else 2909 memset(table->table, 0, 2910 (MAX_IRQS_PER_TABLE * (sizeof(u64) * 2))); 2911 return table; 2912 } 2913 2914 static void set_remap_table_entry(struct amd_iommu *iommu, u16 devid, 2915 struct irq_remap_table *table) 2916 { 2917 struct amd_iommu_pci_seg *pci_seg = iommu->pci_seg; 2918 2919 pci_seg->irq_lookup_table[devid] = table; 2920 set_dte_irq_entry(iommu, devid, table); 2921 iommu_flush_dte(iommu, devid); 2922 } 2923 2924 static int set_remap_table_entry_alias(struct pci_dev *pdev, u16 alias, 2925 void *data) 2926 { 2927 struct irq_remap_table *table = data; 2928 struct amd_iommu_pci_seg *pci_seg; 2929 struct amd_iommu *iommu = rlookup_amd_iommu(&pdev->dev); 2930 2931 if (!iommu) 2932 return -EINVAL; 2933 2934 pci_seg = iommu->pci_seg; 2935 pci_seg->irq_lookup_table[alias] = table; 2936 set_dte_irq_entry(iommu, alias, table); 2937 iommu_flush_dte(pci_seg->rlookup_table[alias], alias); 2938 2939 return 0; 2940 } 2941 2942 static struct irq_remap_table *alloc_irq_table(struct amd_iommu *iommu, 2943 u16 devid, struct pci_dev *pdev) 2944 { 2945 struct irq_remap_table *table = NULL; 2946 struct irq_remap_table *new_table = NULL; 2947 struct amd_iommu_pci_seg *pci_seg; 2948 unsigned long flags; 2949 u16 alias; 2950 2951 spin_lock_irqsave(&iommu_table_lock, flags); 2952 2953 pci_seg = iommu->pci_seg; 2954 table = pci_seg->irq_lookup_table[devid]; 2955 if (table) 2956 goto out_unlock; 2957 2958 alias = pci_seg->alias_table[devid]; 2959 table = pci_seg->irq_lookup_table[alias]; 2960 if (table) { 2961 set_remap_table_entry(iommu, devid, table); 2962 goto out_wait; 2963 } 2964 spin_unlock_irqrestore(&iommu_table_lock, flags); 2965 2966 /* Nothing there yet, allocate new irq remapping table */ 2967 new_table = __alloc_irq_table(); 2968 if (!new_table) 2969 return NULL; 2970 2971 spin_lock_irqsave(&iommu_table_lock, flags); 2972 2973 table = pci_seg->irq_lookup_table[devid]; 2974 if (table) 2975 goto out_unlock; 2976 2977 table = pci_seg->irq_lookup_table[alias]; 2978 if (table) { 2979 set_remap_table_entry(iommu, devid, table); 2980 goto out_wait; 2981 } 2982 2983 table = new_table; 2984 new_table = NULL; 2985 2986 if (pdev) 2987 pci_for_each_dma_alias(pdev, set_remap_table_entry_alias, 2988 table); 2989 else 2990 set_remap_table_entry(iommu, devid, table); 2991 2992 if (devid != alias) 2993 set_remap_table_entry(iommu, alias, table); 2994 2995 out_wait: 2996 iommu_completion_wait(iommu); 2997 2998 out_unlock: 2999 spin_unlock_irqrestore(&iommu_table_lock, flags); 3000 3001 if (new_table) { 3002 kmem_cache_free(amd_iommu_irq_cache, new_table->table); 3003 kfree(new_table); 3004 } 3005 return table; 3006 } 3007 3008 static int alloc_irq_index(struct amd_iommu *iommu, u16 devid, int count, 3009 bool align, struct pci_dev *pdev) 3010 { 3011 struct irq_remap_table *table; 3012 int index, c, alignment = 1; 3013 unsigned long flags; 3014 3015 table = alloc_irq_table(iommu, devid, pdev); 3016 if (!table) 3017 return -ENODEV; 3018 3019 if (align) 3020 alignment = roundup_pow_of_two(count); 3021 3022 raw_spin_lock_irqsave(&table->lock, flags); 3023 3024 /* Scan table for free entries */ 3025 for (index = ALIGN(table->min_index, alignment), c = 0; 3026 index < MAX_IRQS_PER_TABLE;) { 3027 if (!iommu->irte_ops->is_allocated(table, index)) { 3028 c += 1; 3029 } else { 3030 c = 0; 3031 index = ALIGN(index + 1, alignment); 3032 continue; 3033 } 3034 3035 if (c == count) { 3036 for (; c != 0; --c) 3037 iommu->irte_ops->set_allocated(table, index - c + 1); 3038 3039 index -= count - 1; 3040 goto out; 3041 } 3042 3043 index++; 3044 } 3045 3046 index = -ENOSPC; 3047 3048 out: 3049 raw_spin_unlock_irqrestore(&table->lock, flags); 3050 3051 return index; 3052 } 3053 3054 static int modify_irte_ga(struct amd_iommu *iommu, u16 devid, int index, 3055 struct irte_ga *irte) 3056 { 3057 struct irq_remap_table *table; 3058 struct irte_ga *entry; 3059 unsigned long flags; 3060 u128 old; 3061 3062 table = get_irq_table(iommu, devid); 3063 if (!table) 3064 return -ENOMEM; 3065 3066 raw_spin_lock_irqsave(&table->lock, flags); 3067 3068 entry = (struct irte_ga *)table->table; 3069 entry = &entry[index]; 3070 3071 /* 3072 * We use cmpxchg16 to atomically update the 128-bit IRTE, 3073 * and it cannot be updated by the hardware or other processors 3074 * behind us, so the return value of cmpxchg16 should be the 3075 * same as the old value. 3076 */ 3077 old = entry->irte; 3078 WARN_ON(!try_cmpxchg128(&entry->irte, &old, irte->irte)); 3079 3080 raw_spin_unlock_irqrestore(&table->lock, flags); 3081 3082 iommu_flush_irt_and_complete(iommu, devid); 3083 3084 return 0; 3085 } 3086 3087 static int modify_irte(struct amd_iommu *iommu, 3088 u16 devid, int index, union irte *irte) 3089 { 3090 struct irq_remap_table *table; 3091 unsigned long flags; 3092 3093 table = get_irq_table(iommu, devid); 3094 if (!table) 3095 return -ENOMEM; 3096 3097 raw_spin_lock_irqsave(&table->lock, flags); 3098 table->table[index] = irte->val; 3099 raw_spin_unlock_irqrestore(&table->lock, flags); 3100 3101 iommu_flush_irt_and_complete(iommu, devid); 3102 3103 return 0; 3104 } 3105 3106 static void free_irte(struct amd_iommu *iommu, u16 devid, int index) 3107 { 3108 struct irq_remap_table *table; 3109 unsigned long flags; 3110 3111 table = get_irq_table(iommu, devid); 3112 if (!table) 3113 return; 3114 3115 raw_spin_lock_irqsave(&table->lock, flags); 3116 iommu->irte_ops->clear_allocated(table, index); 3117 raw_spin_unlock_irqrestore(&table->lock, flags); 3118 3119 iommu_flush_irt_and_complete(iommu, devid); 3120 } 3121 3122 static void irte_prepare(void *entry, 3123 u32 delivery_mode, bool dest_mode, 3124 u8 vector, u32 dest_apicid, int devid) 3125 { 3126 union irte *irte = (union irte *) entry; 3127 3128 irte->val = 0; 3129 irte->fields.vector = vector; 3130 irte->fields.int_type = delivery_mode; 3131 irte->fields.destination = dest_apicid; 3132 irte->fields.dm = dest_mode; 3133 irte->fields.valid = 1; 3134 } 3135 3136 static void irte_ga_prepare(void *entry, 3137 u32 delivery_mode, bool dest_mode, 3138 u8 vector, u32 dest_apicid, int devid) 3139 { 3140 struct irte_ga *irte = (struct irte_ga *) entry; 3141 3142 irte->lo.val = 0; 3143 irte->hi.val = 0; 3144 irte->lo.fields_remap.int_type = delivery_mode; 3145 irte->lo.fields_remap.dm = dest_mode; 3146 irte->hi.fields.vector = vector; 3147 irte->lo.fields_remap.destination = APICID_TO_IRTE_DEST_LO(dest_apicid); 3148 irte->hi.fields.destination = APICID_TO_IRTE_DEST_HI(dest_apicid); 3149 irte->lo.fields_remap.valid = 1; 3150 } 3151 3152 static void irte_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index) 3153 { 3154 union irte *irte = (union irte *) entry; 3155 3156 irte->fields.valid = 1; 3157 modify_irte(iommu, devid, index, irte); 3158 } 3159 3160 static void irte_ga_activate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index) 3161 { 3162 struct irte_ga *irte = (struct irte_ga *) entry; 3163 3164 irte->lo.fields_remap.valid = 1; 3165 modify_irte_ga(iommu, devid, index, irte); 3166 } 3167 3168 static void irte_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index) 3169 { 3170 union irte *irte = (union irte *) entry; 3171 3172 irte->fields.valid = 0; 3173 modify_irte(iommu, devid, index, irte); 3174 } 3175 3176 static void irte_ga_deactivate(struct amd_iommu *iommu, void *entry, u16 devid, u16 index) 3177 { 3178 struct irte_ga *irte = (struct irte_ga *) entry; 3179 3180 irte->lo.fields_remap.valid = 0; 3181 modify_irte_ga(iommu, devid, index, irte); 3182 } 3183 3184 static void irte_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index, 3185 u8 vector, u32 dest_apicid) 3186 { 3187 union irte *irte = (union irte *) entry; 3188 3189 irte->fields.vector = vector; 3190 irte->fields.destination = dest_apicid; 3191 modify_irte(iommu, devid, index, irte); 3192 } 3193 3194 static void irte_ga_set_affinity(struct amd_iommu *iommu, void *entry, u16 devid, u16 index, 3195 u8 vector, u32 dest_apicid) 3196 { 3197 struct irte_ga *irte = (struct irte_ga *) entry; 3198 3199 if (!irte->lo.fields_remap.guest_mode) { 3200 irte->hi.fields.vector = vector; 3201 irte->lo.fields_remap.destination = 3202 APICID_TO_IRTE_DEST_LO(dest_apicid); 3203 irte->hi.fields.destination = 3204 APICID_TO_IRTE_DEST_HI(dest_apicid); 3205 modify_irte_ga(iommu, devid, index, irte); 3206 } 3207 } 3208 3209 #define IRTE_ALLOCATED (~1U) 3210 static void irte_set_allocated(struct irq_remap_table *table, int index) 3211 { 3212 table->table[index] = IRTE_ALLOCATED; 3213 } 3214 3215 static void irte_ga_set_allocated(struct irq_remap_table *table, int index) 3216 { 3217 struct irte_ga *ptr = (struct irte_ga *)table->table; 3218 struct irte_ga *irte = &ptr[index]; 3219 3220 memset(&irte->lo.val, 0, sizeof(u64)); 3221 memset(&irte->hi.val, 0, sizeof(u64)); 3222 irte->hi.fields.vector = 0xff; 3223 } 3224 3225 static bool irte_is_allocated(struct irq_remap_table *table, int index) 3226 { 3227 union irte *ptr = (union irte *)table->table; 3228 union irte *irte = &ptr[index]; 3229 3230 return irte->val != 0; 3231 } 3232 3233 static bool irte_ga_is_allocated(struct irq_remap_table *table, int index) 3234 { 3235 struct irte_ga *ptr = (struct irte_ga *)table->table; 3236 struct irte_ga *irte = &ptr[index]; 3237 3238 return irte->hi.fields.vector != 0; 3239 } 3240 3241 static void irte_clear_allocated(struct irq_remap_table *table, int index) 3242 { 3243 table->table[index] = 0; 3244 } 3245 3246 static void irte_ga_clear_allocated(struct irq_remap_table *table, int index) 3247 { 3248 struct irte_ga *ptr = (struct irte_ga *)table->table; 3249 struct irte_ga *irte = &ptr[index]; 3250 3251 memset(&irte->lo.val, 0, sizeof(u64)); 3252 memset(&irte->hi.val, 0, sizeof(u64)); 3253 } 3254 3255 static int get_devid(struct irq_alloc_info *info) 3256 { 3257 switch (info->type) { 3258 case X86_IRQ_ALLOC_TYPE_IOAPIC: 3259 return get_ioapic_devid(info->devid); 3260 case X86_IRQ_ALLOC_TYPE_HPET: 3261 return get_hpet_devid(info->devid); 3262 case X86_IRQ_ALLOC_TYPE_PCI_MSI: 3263 case X86_IRQ_ALLOC_TYPE_PCI_MSIX: 3264 return get_device_sbdf_id(msi_desc_to_dev(info->desc)); 3265 default: 3266 WARN_ON_ONCE(1); 3267 return -1; 3268 } 3269 } 3270 3271 struct irq_remap_ops amd_iommu_irq_ops = { 3272 .prepare = amd_iommu_prepare, 3273 .enable = amd_iommu_enable, 3274 .disable = amd_iommu_disable, 3275 .reenable = amd_iommu_reenable, 3276 .enable_faulting = amd_iommu_enable_faulting, 3277 }; 3278 3279 static void fill_msi_msg(struct msi_msg *msg, u32 index) 3280 { 3281 msg->data = index; 3282 msg->address_lo = 0; 3283 msg->arch_addr_lo.base_address = X86_MSI_BASE_ADDRESS_LOW; 3284 msg->address_hi = X86_MSI_BASE_ADDRESS_HIGH; 3285 } 3286 3287 static void irq_remapping_prepare_irte(struct amd_ir_data *data, 3288 struct irq_cfg *irq_cfg, 3289 struct irq_alloc_info *info, 3290 int devid, int index, int sub_handle) 3291 { 3292 struct irq_2_irte *irte_info = &data->irq_2_irte; 3293 struct amd_iommu *iommu = data->iommu; 3294 3295 if (!iommu) 3296 return; 3297 3298 data->irq_2_irte.devid = devid; 3299 data->irq_2_irte.index = index + sub_handle; 3300 iommu->irte_ops->prepare(data->entry, apic->delivery_mode, 3301 apic->dest_mode_logical, irq_cfg->vector, 3302 irq_cfg->dest_apicid, devid); 3303 3304 switch (info->type) { 3305 case X86_IRQ_ALLOC_TYPE_IOAPIC: 3306 case X86_IRQ_ALLOC_TYPE_HPET: 3307 case X86_IRQ_ALLOC_TYPE_PCI_MSI: 3308 case X86_IRQ_ALLOC_TYPE_PCI_MSIX: 3309 fill_msi_msg(&data->msi_entry, irte_info->index); 3310 break; 3311 3312 default: 3313 BUG_ON(1); 3314 break; 3315 } 3316 } 3317 3318 struct amd_irte_ops irte_32_ops = { 3319 .prepare = irte_prepare, 3320 .activate = irte_activate, 3321 .deactivate = irte_deactivate, 3322 .set_affinity = irte_set_affinity, 3323 .set_allocated = irte_set_allocated, 3324 .is_allocated = irte_is_allocated, 3325 .clear_allocated = irte_clear_allocated, 3326 }; 3327 3328 struct amd_irte_ops irte_128_ops = { 3329 .prepare = irte_ga_prepare, 3330 .activate = irte_ga_activate, 3331 .deactivate = irte_ga_deactivate, 3332 .set_affinity = irte_ga_set_affinity, 3333 .set_allocated = irte_ga_set_allocated, 3334 .is_allocated = irte_ga_is_allocated, 3335 .clear_allocated = irte_ga_clear_allocated, 3336 }; 3337 3338 static int irq_remapping_alloc(struct irq_domain *domain, unsigned int virq, 3339 unsigned int nr_irqs, void *arg) 3340 { 3341 struct irq_alloc_info *info = arg; 3342 struct irq_data *irq_data; 3343 struct amd_ir_data *data = NULL; 3344 struct amd_iommu *iommu; 3345 struct irq_cfg *cfg; 3346 int i, ret, devid, seg, sbdf; 3347 int index; 3348 3349 if (!info) 3350 return -EINVAL; 3351 if (nr_irqs > 1 && info->type != X86_IRQ_ALLOC_TYPE_PCI_MSI) 3352 return -EINVAL; 3353 3354 sbdf = get_devid(info); 3355 if (sbdf < 0) 3356 return -EINVAL; 3357 3358 seg = PCI_SBDF_TO_SEGID(sbdf); 3359 devid = PCI_SBDF_TO_DEVID(sbdf); 3360 iommu = __rlookup_amd_iommu(seg, devid); 3361 if (!iommu) 3362 return -EINVAL; 3363 3364 ret = irq_domain_alloc_irqs_parent(domain, virq, nr_irqs, arg); 3365 if (ret < 0) 3366 return ret; 3367 3368 if (info->type == X86_IRQ_ALLOC_TYPE_IOAPIC) { 3369 struct irq_remap_table *table; 3370 3371 table = alloc_irq_table(iommu, devid, NULL); 3372 if (table) { 3373 if (!table->min_index) { 3374 /* 3375 * Keep the first 32 indexes free for IOAPIC 3376 * interrupts. 3377 */ 3378 table->min_index = 32; 3379 for (i = 0; i < 32; ++i) 3380 iommu->irte_ops->set_allocated(table, i); 3381 } 3382 WARN_ON(table->min_index != 32); 3383 index = info->ioapic.pin; 3384 } else { 3385 index = -ENOMEM; 3386 } 3387 } else if (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI || 3388 info->type == X86_IRQ_ALLOC_TYPE_PCI_MSIX) { 3389 bool align = (info->type == X86_IRQ_ALLOC_TYPE_PCI_MSI); 3390 3391 index = alloc_irq_index(iommu, devid, nr_irqs, align, 3392 msi_desc_to_pci_dev(info->desc)); 3393 } else { 3394 index = alloc_irq_index(iommu, devid, nr_irqs, false, NULL); 3395 } 3396 3397 if (index < 0) { 3398 pr_warn("Failed to allocate IRTE\n"); 3399 ret = index; 3400 goto out_free_parent; 3401 } 3402 3403 for (i = 0; i < nr_irqs; i++) { 3404 irq_data = irq_domain_get_irq_data(domain, virq + i); 3405 cfg = irq_data ? irqd_cfg(irq_data) : NULL; 3406 if (!cfg) { 3407 ret = -EINVAL; 3408 goto out_free_data; 3409 } 3410 3411 ret = -ENOMEM; 3412 data = kzalloc(sizeof(*data), GFP_KERNEL); 3413 if (!data) 3414 goto out_free_data; 3415 3416 if (!AMD_IOMMU_GUEST_IR_GA(amd_iommu_guest_ir)) 3417 data->entry = kzalloc(sizeof(union irte), GFP_KERNEL); 3418 else 3419 data->entry = kzalloc(sizeof(struct irte_ga), 3420 GFP_KERNEL); 3421 if (!data->entry) { 3422 kfree(data); 3423 goto out_free_data; 3424 } 3425 3426 data->iommu = iommu; 3427 irq_data->hwirq = (devid << 16) + i; 3428 irq_data->chip_data = data; 3429 irq_data->chip = &amd_ir_chip; 3430 irq_remapping_prepare_irte(data, cfg, info, devid, index, i); 3431 irq_set_status_flags(virq + i, IRQ_MOVE_PCNTXT); 3432 } 3433 3434 return 0; 3435 3436 out_free_data: 3437 for (i--; i >= 0; i--) { 3438 irq_data = irq_domain_get_irq_data(domain, virq + i); 3439 if (irq_data) 3440 kfree(irq_data->chip_data); 3441 } 3442 for (i = 0; i < nr_irqs; i++) 3443 free_irte(iommu, devid, index + i); 3444 out_free_parent: 3445 irq_domain_free_irqs_common(domain, virq, nr_irqs); 3446 return ret; 3447 } 3448 3449 static void irq_remapping_free(struct irq_domain *domain, unsigned int virq, 3450 unsigned int nr_irqs) 3451 { 3452 struct irq_2_irte *irte_info; 3453 struct irq_data *irq_data; 3454 struct amd_ir_data *data; 3455 int i; 3456 3457 for (i = 0; i < nr_irqs; i++) { 3458 irq_data = irq_domain_get_irq_data(domain, virq + i); 3459 if (irq_data && irq_data->chip_data) { 3460 data = irq_data->chip_data; 3461 irte_info = &data->irq_2_irte; 3462 free_irte(data->iommu, irte_info->devid, irte_info->index); 3463 kfree(data->entry); 3464 kfree(data); 3465 } 3466 } 3467 irq_domain_free_irqs_common(domain, virq, nr_irqs); 3468 } 3469 3470 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu, 3471 struct amd_ir_data *ir_data, 3472 struct irq_2_irte *irte_info, 3473 struct irq_cfg *cfg); 3474 3475 static int irq_remapping_activate(struct irq_domain *domain, 3476 struct irq_data *irq_data, bool reserve) 3477 { 3478 struct amd_ir_data *data = irq_data->chip_data; 3479 struct irq_2_irte *irte_info = &data->irq_2_irte; 3480 struct amd_iommu *iommu = data->iommu; 3481 struct irq_cfg *cfg = irqd_cfg(irq_data); 3482 3483 if (!iommu) 3484 return 0; 3485 3486 iommu->irte_ops->activate(iommu, data->entry, irte_info->devid, 3487 irte_info->index); 3488 amd_ir_update_irte(irq_data, iommu, data, irte_info, cfg); 3489 return 0; 3490 } 3491 3492 static void irq_remapping_deactivate(struct irq_domain *domain, 3493 struct irq_data *irq_data) 3494 { 3495 struct amd_ir_data *data = irq_data->chip_data; 3496 struct irq_2_irte *irte_info = &data->irq_2_irte; 3497 struct amd_iommu *iommu = data->iommu; 3498 3499 if (iommu) 3500 iommu->irte_ops->deactivate(iommu, data->entry, irte_info->devid, 3501 irte_info->index); 3502 } 3503 3504 static int irq_remapping_select(struct irq_domain *d, struct irq_fwspec *fwspec, 3505 enum irq_domain_bus_token bus_token) 3506 { 3507 struct amd_iommu *iommu; 3508 int devid = -1; 3509 3510 if (!amd_iommu_irq_remap) 3511 return 0; 3512 3513 if (x86_fwspec_is_ioapic(fwspec)) 3514 devid = get_ioapic_devid(fwspec->param[0]); 3515 else if (x86_fwspec_is_hpet(fwspec)) 3516 devid = get_hpet_devid(fwspec->param[0]); 3517 3518 if (devid < 0) 3519 return 0; 3520 iommu = __rlookup_amd_iommu((devid >> 16), (devid & 0xffff)); 3521 3522 return iommu && iommu->ir_domain == d; 3523 } 3524 3525 static const struct irq_domain_ops amd_ir_domain_ops = { 3526 .select = irq_remapping_select, 3527 .alloc = irq_remapping_alloc, 3528 .free = irq_remapping_free, 3529 .activate = irq_remapping_activate, 3530 .deactivate = irq_remapping_deactivate, 3531 }; 3532 3533 int amd_iommu_activate_guest_mode(void *data) 3534 { 3535 struct amd_ir_data *ir_data = (struct amd_ir_data *)data; 3536 struct irte_ga *entry = (struct irte_ga *) ir_data->entry; 3537 u64 valid; 3538 3539 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || !entry) 3540 return 0; 3541 3542 valid = entry->lo.fields_vapic.valid; 3543 3544 entry->lo.val = 0; 3545 entry->hi.val = 0; 3546 3547 entry->lo.fields_vapic.valid = valid; 3548 entry->lo.fields_vapic.guest_mode = 1; 3549 entry->lo.fields_vapic.ga_log_intr = 1; 3550 entry->hi.fields.ga_root_ptr = ir_data->ga_root_ptr; 3551 entry->hi.fields.vector = ir_data->ga_vector; 3552 entry->lo.fields_vapic.ga_tag = ir_data->ga_tag; 3553 3554 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid, 3555 ir_data->irq_2_irte.index, entry); 3556 } 3557 EXPORT_SYMBOL(amd_iommu_activate_guest_mode); 3558 3559 int amd_iommu_deactivate_guest_mode(void *data) 3560 { 3561 struct amd_ir_data *ir_data = (struct amd_ir_data *)data; 3562 struct irte_ga *entry = (struct irte_ga *) ir_data->entry; 3563 struct irq_cfg *cfg = ir_data->cfg; 3564 u64 valid; 3565 3566 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || 3567 !entry || !entry->lo.fields_vapic.guest_mode) 3568 return 0; 3569 3570 valid = entry->lo.fields_remap.valid; 3571 3572 entry->lo.val = 0; 3573 entry->hi.val = 0; 3574 3575 entry->lo.fields_remap.valid = valid; 3576 entry->lo.fields_remap.dm = apic->dest_mode_logical; 3577 entry->lo.fields_remap.int_type = apic->delivery_mode; 3578 entry->hi.fields.vector = cfg->vector; 3579 entry->lo.fields_remap.destination = 3580 APICID_TO_IRTE_DEST_LO(cfg->dest_apicid); 3581 entry->hi.fields.destination = 3582 APICID_TO_IRTE_DEST_HI(cfg->dest_apicid); 3583 3584 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid, 3585 ir_data->irq_2_irte.index, entry); 3586 } 3587 EXPORT_SYMBOL(amd_iommu_deactivate_guest_mode); 3588 3589 static int amd_ir_set_vcpu_affinity(struct irq_data *data, void *vcpu_info) 3590 { 3591 int ret; 3592 struct amd_iommu_pi_data *pi_data = vcpu_info; 3593 struct vcpu_data *vcpu_pi_info = pi_data->vcpu_data; 3594 struct amd_ir_data *ir_data = data->chip_data; 3595 struct irq_2_irte *irte_info = &ir_data->irq_2_irte; 3596 struct iommu_dev_data *dev_data; 3597 3598 if (ir_data->iommu == NULL) 3599 return -EINVAL; 3600 3601 dev_data = search_dev_data(ir_data->iommu, irte_info->devid); 3602 3603 /* Note: 3604 * This device has never been set up for guest mode. 3605 * we should not modify the IRTE 3606 */ 3607 if (!dev_data || !dev_data->use_vapic) 3608 return 0; 3609 3610 ir_data->cfg = irqd_cfg(data); 3611 pi_data->ir_data = ir_data; 3612 3613 /* Note: 3614 * SVM tries to set up for VAPIC mode, but we are in 3615 * legacy mode. So, we force legacy mode instead. 3616 */ 3617 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir)) { 3618 pr_debug("%s: Fall back to using intr legacy remap\n", 3619 __func__); 3620 pi_data->is_guest_mode = false; 3621 } 3622 3623 pi_data->prev_ga_tag = ir_data->cached_ga_tag; 3624 if (pi_data->is_guest_mode) { 3625 ir_data->ga_root_ptr = (pi_data->base >> 12); 3626 ir_data->ga_vector = vcpu_pi_info->vector; 3627 ir_data->ga_tag = pi_data->ga_tag; 3628 ret = amd_iommu_activate_guest_mode(ir_data); 3629 if (!ret) 3630 ir_data->cached_ga_tag = pi_data->ga_tag; 3631 } else { 3632 ret = amd_iommu_deactivate_guest_mode(ir_data); 3633 3634 /* 3635 * This communicates the ga_tag back to the caller 3636 * so that it can do all the necessary clean up. 3637 */ 3638 if (!ret) 3639 ir_data->cached_ga_tag = 0; 3640 } 3641 3642 return ret; 3643 } 3644 3645 3646 static void amd_ir_update_irte(struct irq_data *irqd, struct amd_iommu *iommu, 3647 struct amd_ir_data *ir_data, 3648 struct irq_2_irte *irte_info, 3649 struct irq_cfg *cfg) 3650 { 3651 3652 /* 3653 * Atomically updates the IRTE with the new destination, vector 3654 * and flushes the interrupt entry cache. 3655 */ 3656 iommu->irte_ops->set_affinity(iommu, ir_data->entry, irte_info->devid, 3657 irte_info->index, cfg->vector, 3658 cfg->dest_apicid); 3659 } 3660 3661 static int amd_ir_set_affinity(struct irq_data *data, 3662 const struct cpumask *mask, bool force) 3663 { 3664 struct amd_ir_data *ir_data = data->chip_data; 3665 struct irq_2_irte *irte_info = &ir_data->irq_2_irte; 3666 struct irq_cfg *cfg = irqd_cfg(data); 3667 struct irq_data *parent = data->parent_data; 3668 struct amd_iommu *iommu = ir_data->iommu; 3669 int ret; 3670 3671 if (!iommu) 3672 return -ENODEV; 3673 3674 ret = parent->chip->irq_set_affinity(parent, mask, force); 3675 if (ret < 0 || ret == IRQ_SET_MASK_OK_DONE) 3676 return ret; 3677 3678 amd_ir_update_irte(data, iommu, ir_data, irte_info, cfg); 3679 /* 3680 * After this point, all the interrupts will start arriving 3681 * at the new destination. So, time to cleanup the previous 3682 * vector allocation. 3683 */ 3684 vector_schedule_cleanup(cfg); 3685 3686 return IRQ_SET_MASK_OK_DONE; 3687 } 3688 3689 static void ir_compose_msi_msg(struct irq_data *irq_data, struct msi_msg *msg) 3690 { 3691 struct amd_ir_data *ir_data = irq_data->chip_data; 3692 3693 *msg = ir_data->msi_entry; 3694 } 3695 3696 static struct irq_chip amd_ir_chip = { 3697 .name = "AMD-IR", 3698 .irq_ack = apic_ack_irq, 3699 .irq_set_affinity = amd_ir_set_affinity, 3700 .irq_set_vcpu_affinity = amd_ir_set_vcpu_affinity, 3701 .irq_compose_msi_msg = ir_compose_msi_msg, 3702 }; 3703 3704 static const struct msi_parent_ops amdvi_msi_parent_ops = { 3705 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED | 3706 MSI_FLAG_MULTI_PCI_MSI | 3707 MSI_FLAG_PCI_IMS, 3708 .prefix = "IR-", 3709 .init_dev_msi_info = msi_parent_init_dev_msi_info, 3710 }; 3711 3712 static const struct msi_parent_ops virt_amdvi_msi_parent_ops = { 3713 .supported_flags = X86_VECTOR_MSI_FLAGS_SUPPORTED | 3714 MSI_FLAG_MULTI_PCI_MSI, 3715 .prefix = "vIR-", 3716 .init_dev_msi_info = msi_parent_init_dev_msi_info, 3717 }; 3718 3719 int amd_iommu_create_irq_domain(struct amd_iommu *iommu) 3720 { 3721 struct fwnode_handle *fn; 3722 3723 fn = irq_domain_alloc_named_id_fwnode("AMD-IR", iommu->index); 3724 if (!fn) 3725 return -ENOMEM; 3726 iommu->ir_domain = irq_domain_create_hierarchy(arch_get_ir_parent_domain(), 0, 0, 3727 fn, &amd_ir_domain_ops, iommu); 3728 if (!iommu->ir_domain) { 3729 irq_domain_free_fwnode(fn); 3730 return -ENOMEM; 3731 } 3732 3733 irq_domain_update_bus_token(iommu->ir_domain, DOMAIN_BUS_AMDVI); 3734 iommu->ir_domain->flags |= IRQ_DOMAIN_FLAG_MSI_PARENT | 3735 IRQ_DOMAIN_FLAG_ISOLATED_MSI; 3736 3737 if (amd_iommu_np_cache) 3738 iommu->ir_domain->msi_parent_ops = &virt_amdvi_msi_parent_ops; 3739 else 3740 iommu->ir_domain->msi_parent_ops = &amdvi_msi_parent_ops; 3741 3742 return 0; 3743 } 3744 3745 int amd_iommu_update_ga(int cpu, bool is_run, void *data) 3746 { 3747 struct amd_ir_data *ir_data = (struct amd_ir_data *)data; 3748 struct irte_ga *entry = (struct irte_ga *) ir_data->entry; 3749 3750 if (!AMD_IOMMU_GUEST_IR_VAPIC(amd_iommu_guest_ir) || 3751 !entry || !entry->lo.fields_vapic.guest_mode) 3752 return 0; 3753 3754 if (!ir_data->iommu) 3755 return -ENODEV; 3756 3757 if (cpu >= 0) { 3758 entry->lo.fields_vapic.destination = 3759 APICID_TO_IRTE_DEST_LO(cpu); 3760 entry->hi.fields.destination = 3761 APICID_TO_IRTE_DEST_HI(cpu); 3762 } 3763 entry->lo.fields_vapic.is_run = is_run; 3764 3765 return modify_irte_ga(ir_data->iommu, ir_data->irq_2_irte.devid, 3766 ir_data->irq_2_irte.index, entry); 3767 } 3768 EXPORT_SYMBOL(amd_iommu_update_ga); 3769 #endif 3770