drivers/iommu/dmar.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2006, Intel Corporation.
   4  *
   5  * Copyright (C) 2006-2008 Intel Corporation
   6  * Author: Ashok Raj <[email protected]>
   7  * Author: Shaohua Li <[email protected]>
   8  * Author: Anil S Keshavamurthy <[email protected]>
   9  *
  10  * This file implements early detection/parsing of Remapping Devices
  11  * reported to OS through BIOS via DMA remapping reporting (DMAR) ACPI
  12  * tables.
  13  *
  14  * These routines are used by both DMA-remapping and Interrupt-remapping
  15  */
  16
  17 #define pr_fmt(fmt)     "DMAR: " fmt
  18
  19 #include <linux/pci.h>
  20 #include <linux/dmar.h>
  21 #include <linux/iova.h>
  22 #include <linux/intel-iommu.h>
  23 #include <linux/timer.h>
  24 #include <linux/irq.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/tboot.h>
  27 #include <linux/dmi.h>
  28 #include <linux/slab.h>
  29 #include <linux/iommu.h>
  30 #include <linux/numa.h>
  31 #include <linux/limits.h>
  32 #include <asm/irq_remapping.h>
  33 #include <asm/iommu_table.h>
  34
  35 #include "irq_remapping.h"
  36
  37 typedef int (*dmar_res_handler_t)(struct acpi_dmar_header *, void *);
  38 struct dmar_res_callback {
  39         dmar_res_handler_t      cb[ACPI_DMAR_TYPE_RESERVED];
  40         void                    *arg[ACPI_DMAR_TYPE_RESERVED];
  41         bool                    ignore_unhandled;
  42         bool                    print_entry;
  43 };
  44
  45 /*
  46  * Assumptions:
  47  * 1) The hotplug framework guarentees that DMAR unit will be hot-added
  48  *    before IO devices managed by that unit.
  49  * 2) The hotplug framework guarantees that DMAR unit will be hot-removed
  50  *    after IO devices managed by that unit.
  51  * 3) Hotplug events are rare.
  52  *
  53  * Locking rules for DMA and interrupt remapping related global data structures:
  54  * 1) Use dmar_global_lock in process context
  55  * 2) Use RCU in interrupt context
  56  */
  57 DECLARE_RWSEM(dmar_global_lock);
  58 LIST_HEAD(dmar_drhd_units);
  59
  60 struct acpi_table_header * __initdata dmar_tbl;
  61 static int dmar_dev_scope_status = 1;
  62 static unsigned long dmar_seq_ids[BITS_TO_LONGS(DMAR_UNITS_SUPPORTED)];
  63
  64 static int alloc_iommu(struct dmar_drhd_unit *drhd);
  65 static void free_iommu(struct intel_iommu *iommu);
  66
  67 extern const struct iommu_ops intel_iommu_ops;
  68
  69 static void dmar_register_drhd_unit(struct dmar_drhd_unit *drhd)
  70 {
  71         /*
  72          * add INCLUDE_ALL at the tail, so scan the list will find it at
  73          * the very end.
  74          */
  75         if (drhd->include_all)
  76                 list_add_tail_rcu(&drhd->list, &dmar_drhd_units);
  77         else
  78                 list_add_rcu(&drhd->list, &dmar_drhd_units);
  79 }
  80
  81 void *dmar_alloc_dev_scope(void *start, void *end, int *cnt)
  82 {
  83         struct acpi_dmar_device_scope *scope;
  84
  85         *cnt = 0;
  86         while (start < end) {
  87                 scope = start;
  88                 if (scope->entry_type == ACPI_DMAR_SCOPE_TYPE_NAMESPACE ||
  89                     scope->entry_type == ACPI_DMAR_SCOPE_TYPE_ENDPOINT ||
  90                     scope->entry_type == ACPI_DMAR_SCOPE_TYPE_BRIDGE)
  91                         (*cnt)++;
  92                 else if (scope->entry_type != ACPI_DMAR_SCOPE_TYPE_IOAPIC &&
  93                         scope->entry_type != ACPI_DMAR_SCOPE_TYPE_HPET) {
  94                         pr_warn("Unsupported device scope\n");
  95                 }
  96                 start += scope->length;
  97         }
  98         if (*cnt == 0)
  99                 return NULL;
 100
 101         return kcalloc(*cnt, sizeof(struct dmar_dev_scope), GFP_KERNEL);
 102 }
 103
 104 void dmar_free_dev_scope(struct dmar_dev_scope **devices, int *cnt)
 105 {
 106         int i;
 107         struct device *tmp_dev;
 108
 109         if (*devices && *cnt) {
 110                 for_each_active_dev_scope(*devices, *cnt, i, tmp_dev)
 111                         put_device(tmp_dev);
 112                 kfree(*devices);
 113         }
 114
 115         *devices = NULL;
 116         *cnt = 0;
 117 }
 118
 119 /* Optimize out kzalloc()/kfree() for normal cases */
 120 static char dmar_pci_notify_info_buf[64];
 121
 122 static struct dmar_pci_notify_info *
 123 dmar_alloc_pci_notify_info(struct pci_dev *dev, unsigned long event)
 124 {
 125         int level = 0;
 126         size_t size;
 127         struct pci_dev *tmp;
 128         struct dmar_pci_notify_info *info;
 129
 130         BUG_ON(dev->is_virtfn);
 131
 132         /*
 133          * Ignore devices that have a domain number higher than what can
 134          * be looked up in DMAR, e.g. VMD subdevices with domain 0x10000
 135          */
 136         if (pci_domain_nr(dev->bus) > U16_MAX)
 137                 return NULL;
 138
 139         /* Only generate path[] for device addition event */
 140         if (event == BUS_NOTIFY_ADD_DEVICE)
 141                 for (tmp = dev; tmp; tmp = tmp->bus->self)
 142                         level++;
 143
 144         size = struct_size(info, path, level);
 145         if (size <= sizeof(dmar_pci_notify_info_buf)) {
 146                 info = (struct dmar_pci_notify_info *)dmar_pci_notify_info_buf;
 147         } else {
 148                 info = kzalloc(size, GFP_KERNEL);
 149                 if (!info) {
 150                         pr_warn("Out of memory when allocating notify_info "
 151                                 "for %s.\n", pci_name(dev));
 152                         if (dmar_dev_scope_status == 0)
 153                                 dmar_dev_scope_status = -ENOMEM;
 154                         return NULL;
 155                 }
 156         }
 157
 158         info->event = event;
 159         info->dev = dev;
 160         info->seg = pci_domain_nr(dev->bus);
 161         info->level = level;
 162         if (event == BUS_NOTIFY_ADD_DEVICE) {
 163                 for (tmp = dev; tmp; tmp = tmp->bus->self) {
 164                         level--;
 165                         info->path[level].bus = tmp->bus->number;
 166                         info->path[level].device = PCI_SLOT(tmp->devfn);
 167                         info->path[level].function = PCI_FUNC(tmp->devfn);
 168                         if (pci_is_root_bus(tmp->bus))
 169                                 info->bus = tmp->bus->number;
 170                 }
 171         }
 172
 173         return info;
 174 }
 175
 176 static inline void dmar_free_pci_notify_info(struct dmar_pci_notify_info *info)
 177 {
 178         if ((void *)info != dmar_pci_notify_info_buf)
 179                 kfree(info);
 180 }
 181
 182 static bool dmar_match_pci_path(struct dmar_pci_notify_info *info, int bus,
 183                                 struct acpi_dmar_pci_path *path, int count)
 184 {
 185         int i;
 186
 187         if (info->bus != bus)
 188                 goto fallback;
 189         if (info->level != count)
 190                 goto fallback;
 191
 192         for (i = 0; i < count; i++) {
 193                 if (path[i].device != info->path[i].device ||
 194                     path[i].function != info->path[i].function)
 195                         goto fallback;
 196         }
 197
 198         return true;
 199
 200 fallback:
 201
 202         if (count != 1)
 203                 return false;
 204
 205         i = info->level - 1;
 206         if (bus              == info->path[i].bus &&
 207             path[0].device   == info->path[i].device &&
 208             path[0].function == info->path[i].function) {
 209                 pr_info(FW_BUG "RMRR entry for device %02x:%02x.%x is broken - applying workaround\n",
 210                         bus, path[0].device, path[0].function);
 211                 return true;
 212         }
 213
 214         return false;
 215 }
 216
 217 /* Return: > 0 if match found, 0 if no match found, < 0 if error happens */
 218 int dmar_insert_dev_scope(struct dmar_pci_notify_info *info,
 219                           void *start, void*end, u16 segment,
 220                           struct dmar_dev_scope *devices,
 221                           int devices_cnt)
 222 {
 223         int i, level;
 224         struct device *tmp, *dev = &info->dev->dev;
 225         struct acpi_dmar_device_scope *scope;
 226         struct acpi_dmar_pci_path *path;
 227
 228         if (segment != info->seg)
 229                 return 0;
 230
 231         for (; start < end; start += scope->length) {
 232                 scope = start;
 233                 if (scope->entry_type != ACPI_DMAR_SCOPE_TYPE_ENDPOINT &&
 234                     scope->entry_type != ACPI_DMAR_SCOPE_TYPE_BRIDGE)
 235                         continue;
 236
 237                 path = (struct acpi_dmar_pci_path *)(scope + 1);
 238                 level = (scope->length - sizeof(*scope)) / sizeof(*path);
 239                 if (!dmar_match_pci_path(info, scope->bus, path, level))
 240                         continue;
 241
 242                 /*
 243                  * We expect devices with endpoint scope to have normal PCI
 244                  * headers, and devices with bridge scope to have bridge PCI
 245                  * headers.  However PCI NTB devices may be listed in the
 246                  * DMAR table with bridge scope, even though they have a
 247                  * normal PCI header.  NTB devices are identified by class
 248                  * "BRIDGE_OTHER" (0680h) - we don't declare a socpe mismatch
 249                  * for this special case.
 250                  */
 251                 if ((scope->entry_type == ACPI_DMAR_SCOPE_TYPE_ENDPOINT &&
 252                      info->dev->hdr_type != PCI_HEADER_TYPE_NORMAL) ||
 253                     (scope->entry_type == ACPI_DMAR_SCOPE_TYPE_BRIDGE &&
 254                      (info->dev->hdr_type == PCI_HEADER_TYPE_NORMAL &&
 255                       info->dev->class >> 16 != PCI_BASE_CLASS_BRIDGE))) {
 256                         pr_warn("Device scope type does not match for %s\n",
 257                                 pci_name(info->dev));
 258                         return -EINVAL;
 259                 }
 260
 261                 for_each_dev_scope(devices, devices_cnt, i, tmp)
 262                         if (tmp == NULL) {
 263                                 devices[i].bus = info->dev->bus->number;
 264                                 devices[i].devfn = info->dev->devfn;
 265                                 rcu_assign_pointer(devices[i].dev,
 266                                                    get_device(dev));
 267                                 return 1;
 268                         }
 269                 BUG_ON(i >= devices_cnt);
 270         }
 271
 272         return 0;
 273 }
 274
 275 int dmar_remove_dev_scope(struct dmar_pci_notify_info *info, u16 segment,
 276                           struct dmar_dev_scope *devices, int count)
 277 {
 278         int index;
 279         struct device *tmp;
 280
 281         if (info->seg != segment)
 282                 return 0;
 283
 284         for_each_active_dev_scope(devices, count, index, tmp)
 285                 if (tmp == &info->dev->dev) {
 286                         RCU_INIT_POINTER(devices[index].dev, NULL);
 287                         synchronize_rcu();
 288                         put_device(tmp);
 289                         return 1;
 290                 }
 291
 292         return 0;
 293 }
 294
 295 static int dmar_pci_bus_add_dev(struct dmar_pci_notify_info *info)
 296 {
 297         int ret = 0;
 298         struct dmar_drhd_unit *dmaru;
 299         struct acpi_dmar_hardware_unit *drhd;
 300
 301         for_each_drhd_unit(dmaru) {
 302                 if (dmaru->include_all)
 303                         continue;
 304
 305                 drhd = container_of(dmaru->hdr,
 306                                     struct acpi_dmar_hardware_unit, header);
 307                 ret = dmar_insert_dev_scope(info, (void *)(drhd + 1),
 308                                 ((void *)drhd) + drhd->header.length,
 309                                 dmaru->segment,
 310                                 dmaru->devices, dmaru->devices_cnt);
 311                 if (ret)
 312                         break;
 313         }
 314         if (ret >= 0)
 315                 ret = dmar_iommu_notify_scope_dev(info);
 316         if (ret < 0 && dmar_dev_scope_status == 0)
 317                 dmar_dev_scope_status = ret;
 318
 319         return ret;
 320 }
 321
 322 static void  dmar_pci_bus_del_dev(struct dmar_pci_notify_info *info)
 323 {
 324         struct dmar_drhd_unit *dmaru;
 325
 326         for_each_drhd_unit(dmaru)
 327                 if (dmar_remove_dev_scope(info, dmaru->segment,
 328                         dmaru->devices, dmaru->devices_cnt))
 329                         break;
 330         dmar_iommu_notify_scope_dev(info);
 331 }
 332
 333 static int dmar_pci_bus_notifier(struct notifier_block *nb,
 334                                  unsigned long action, void *data)
 335 {
 336         struct pci_dev *pdev = to_pci_dev(data);
 337         struct dmar_pci_notify_info *info;
 338
 339         /* Only care about add/remove events for physical functions.
 340          * For VFs we actually do the lookup based on the corresponding
 341          * PF in device_to_iommu() anyway. */
 342         if (pdev->is_virtfn)
 343                 return NOTIFY_DONE;
 344         if (action != BUS_NOTIFY_ADD_DEVICE &&
 345             action != BUS_NOTIFY_REMOVED_DEVICE)
 346                 return NOTIFY_DONE;
 347
 348         info = dmar_alloc_pci_notify_info(pdev, action);
 349         if (!info)
 350                 return NOTIFY_DONE;
 351
 352         down_write(&dmar_global_lock);
 353         if (action == BUS_NOTIFY_ADD_DEVICE)
 354                 dmar_pci_bus_add_dev(info);
 355         else if (action == BUS_NOTIFY_REMOVED_DEVICE)
 356                 dmar_pci_bus_del_dev(info);
 357         up_write(&dmar_global_lock);
 358
 359         dmar_free_pci_notify_info(info);
 360
 361         return NOTIFY_OK;
 362 }
 363
 364 static struct notifier_block dmar_pci_bus_nb = {
 365         .notifier_call = dmar_pci_bus_notifier,
 366         .priority = INT_MIN,
 367 };
 368
 369 static struct dmar_drhd_unit *
 370 dmar_find_dmaru(struct acpi_dmar_hardware_unit *drhd)
 371 {
 372         struct dmar_drhd_unit *dmaru;
 373
 374         list_for_each_entry_rcu(dmaru, &dmar_drhd_units, list,
 375                                 dmar_rcu_check())
 376                 if (dmaru->segment == drhd->segment &&
 377                     dmaru->reg_base_addr == drhd->address)
 378                         return dmaru;
 379
 380         return NULL;
 381 }
 382
 383 /**
 384  * dmar_parse_one_drhd - parses exactly one DMA remapping hardware definition
 385  * structure which uniquely represent one DMA remapping hardware unit
 386  * present in the platform
 387  */
 388 static int dmar_parse_one_drhd(struct acpi_dmar_header *header, void *arg)
 389 {
 390         struct acpi_dmar_hardware_unit *drhd;
 391         struct dmar_drhd_unit *dmaru;
 392         int ret;
 393
 394         drhd = (struct acpi_dmar_hardware_unit *)header;
 395         dmaru = dmar_find_dmaru(drhd);
 396         if (dmaru)
 397                 goto out;
 398
 399         dmaru = kzalloc(sizeof(*dmaru) + header->length, GFP_KERNEL);
 400         if (!dmaru)
 401                 return -ENOMEM;
 402
 403         /*
 404          * If header is allocated from slab by ACPI _DSM method, we need to
 405          * copy the content because the memory buffer will be freed on return.
 406          */
 407         dmaru->hdr = (void *)(dmaru + 1);
 408         memcpy(dmaru->hdr, header, header->length);
 409         dmaru->reg_base_addr = drhd->address;
 410         dmaru->segment = drhd->segment;
 411         dmaru->include_all = drhd->flags & 0x1; /* BIT0: INCLUDE_ALL */
 412         dmaru->devices = dmar_alloc_dev_scope((void *)(drhd + 1),
 413                                               ((void *)drhd) + drhd->header.length,
 414                                               &dmaru->devices_cnt);
 415         if (dmaru->devices_cnt && dmaru->devices == NULL) {
 416                 kfree(dmaru);
 417                 return -ENOMEM;
 418         }
 419
 420         ret = alloc_iommu(dmaru);
 421         if (ret) {
 422                 dmar_free_dev_scope(&dmaru->devices,
 423                                     &dmaru->devices_cnt);
 424                 kfree(dmaru);
 425                 return ret;
 426         }
 427         dmar_register_drhd_unit(dmaru);
 428
 429 out:
 430         if (arg)
 431                 (*(int *)arg)++;
 432
 433         return 0;
 434 }
 435
 436 static void dmar_free_drhd(struct dmar_drhd_unit *dmaru)
 437 {
 438         if (dmaru->devices && dmaru->devices_cnt)
 439                 dmar_free_dev_scope(&dmaru->devices, &dmaru->devices_cnt);
 440         if (dmaru->iommu)
 441                 free_iommu(dmaru->iommu);
 442         kfree(dmaru);
 443 }
 444
 445 static int __init dmar_parse_one_andd(struct acpi_dmar_header *header,
 446                                       void *arg)
 447 {
 448         struct acpi_dmar_andd *andd = (void *)header;
 449
 450         /* Check for NUL termination within the designated length */
 451         if (strnlen(andd->device_name, header->length - 8) == header->length - 8) {
 452                 pr_warn(FW_BUG
 453                            "Your BIOS is broken; ANDD object name is not NUL-terminated\n"
 454                            "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
 455                            dmi_get_system_info(DMI_BIOS_VENDOR),
 456                            dmi_get_system_info(DMI_BIOS_VERSION),
 457                            dmi_get_system_info(DMI_PRODUCT_VERSION));
 458                 add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
 459                 return -EINVAL;
 460         }
 461         pr_info("ANDD device: %x name: %s\n", andd->device_number,
 462                 andd->device_name);
 463
 464         return 0;
 465 }
 466
 467 #ifdef CONFIG_ACPI_NUMA
 468 static int dmar_parse_one_rhsa(struct acpi_dmar_header *header, void *arg)
 469 {
 470         struct acpi_dmar_rhsa *rhsa;
 471         struct dmar_drhd_unit *drhd;
 472
 473         rhsa = (struct acpi_dmar_rhsa *)header;
 474         for_each_drhd_unit(drhd) {
 475                 if (drhd->reg_base_addr == rhsa->base_address) {
 476                         int node = acpi_map_pxm_to_node(rhsa->proximity_domain);
 477
 478                         if (!node_online(node))
 479                                 node = NUMA_NO_NODE;
 480                         drhd->iommu->node = node;
 481                         return 0;
 482                 }
 483         }
 484         pr_warn(FW_BUG
 485                 "Your BIOS is broken; RHSA refers to non-existent DMAR unit at %llx\n"
 486                 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
 487                 rhsa->base_address,
 488                 dmi_get_system_info(DMI_BIOS_VENDOR),
 489                 dmi_get_system_info(DMI_BIOS_VERSION),
 490                 dmi_get_system_info(DMI_PRODUCT_VERSION));
 491         add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
 492
 493         return 0;
 494 }
 495 #else
 496 #define dmar_parse_one_rhsa             dmar_res_noop
 497 #endif
 498
 499 static void
 500 dmar_table_print_dmar_entry(struct acpi_dmar_header *header)
 501 {
 502         struct acpi_dmar_hardware_unit *drhd;
 503         struct acpi_dmar_reserved_memory *rmrr;
 504         struct acpi_dmar_atsr *atsr;
 505         struct acpi_dmar_rhsa *rhsa;
 506
 507         switch (header->type) {
 508         case ACPI_DMAR_TYPE_HARDWARE_UNIT:
 509                 drhd = container_of(header, struct acpi_dmar_hardware_unit,
 510                                     header);
 511                 pr_info("DRHD base: %#016Lx flags: %#x\n",
 512                         (unsigned long long)drhd->address, drhd->flags);
 513                 break;
 514         case ACPI_DMAR_TYPE_RESERVED_MEMORY:
 515                 rmrr = container_of(header, struct acpi_dmar_reserved_memory,
 516                                     header);
 517                 pr_info("RMRR base: %#016Lx end: %#016Lx\n",
 518                         (unsigned long long)rmrr->base_address,
 519                         (unsigned long long)rmrr->end_address);
 520                 break;
 521         case ACPI_DMAR_TYPE_ROOT_ATS:
 522                 atsr = container_of(header, struct acpi_dmar_atsr, header);
 523                 pr_info("ATSR flags: %#x\n", atsr->flags);
 524                 break;
 525         case ACPI_DMAR_TYPE_HARDWARE_AFFINITY:
 526                 rhsa = container_of(header, struct acpi_dmar_rhsa, header);
 527                 pr_info("RHSA base: %#016Lx proximity domain: %#x\n",
 528                        (unsigned long long)rhsa->base_address,
 529                        rhsa->proximity_domain);
 530                 break;
 531         case ACPI_DMAR_TYPE_NAMESPACE:
 532                 /* We don't print this here because we need to sanity-check
 533                    it first. So print it in dmar_parse_one_andd() instead. */
 534                 break;
 535         }
 536 }
 537
 538 /**
 539  * dmar_table_detect - checks to see if the platform supports DMAR devices
 540  */
 541 static int __init dmar_table_detect(void)
 542 {
 543         acpi_status status = AE_OK;
 544
 545         /* if we could find DMAR table, then there are DMAR devices */
 546         status = acpi_get_table(ACPI_SIG_DMAR, 0, &dmar_tbl);
 547
 548         if (ACPI_SUCCESS(status) && !dmar_tbl) {
 549                 pr_warn("Unable to map DMAR\n");
 550                 status = AE_NOT_FOUND;
 551         }
 552
 553         return ACPI_SUCCESS(status) ? 0 : -ENOENT;
 554 }
 555
 556 static int dmar_walk_remapping_entries(struct acpi_dmar_header *start,
 557                                        size_t len, struct dmar_res_callback *cb)
 558 {
 559         struct acpi_dmar_header *iter, *next;
 560         struct acpi_dmar_header *end = ((void *)start) + len;
 561
 562         for (iter = start; iter < end; iter = next) {
 563                 next = (void *)iter + iter->length;
 564                 if (iter->length == 0) {
 565                         /* Avoid looping forever on bad ACPI tables */
 566                         pr_debug(FW_BUG "Invalid 0-length structure\n");
 567                         break;
 568                 } else if (next > end) {
 569                         /* Avoid passing table end */
 570                         pr_warn(FW_BUG "Record passes table end\n");
 571                         return -EINVAL;
 572                 }
 573
 574                 if (cb->print_entry)
 575                         dmar_table_print_dmar_entry(iter);
 576
 577                 if (iter->type >= ACPI_DMAR_TYPE_RESERVED) {
 578                         /* continue for forward compatibility */
 579                         pr_debug("Unknown DMAR structure type %d\n",
 580                                  iter->type);
 581                 } else if (cb->cb[iter->type]) {
 582                         int ret;
 583
 584                         ret = cb->cb[iter->type](iter, cb->arg[iter->type]);
 585                         if (ret)
 586                                 return ret;
 587                 } else if (!cb->ignore_unhandled) {
 588                         pr_warn("No handler for DMAR structure type %d\n",
 589                                 iter->type);
 590                         return -EINVAL;
 591                 }
 592         }
 593
 594         return 0;
 595 }
 596
 597 static inline int dmar_walk_dmar_table(struct acpi_table_dmar *dmar,
 598                                        struct dmar_res_callback *cb)
 599 {
 600         return dmar_walk_remapping_entries((void *)(dmar + 1),
 601                         dmar->header.length - sizeof(*dmar), cb);
 602 }
 603
 604 /**
 605  * parse_dmar_table - parses the DMA reporting table
 606  */
 607 static int __init
 608 parse_dmar_table(void)
 609 {
 610         struct acpi_table_dmar *dmar;
 611         int drhd_count = 0;
 612         int ret;
 613         struct dmar_res_callback cb = {
 614                 .print_entry = true,
 615                 .ignore_unhandled = true,
 616                 .arg[ACPI_DMAR_TYPE_HARDWARE_UNIT] = &drhd_count,
 617                 .cb[ACPI_DMAR_TYPE_HARDWARE_UNIT] = &dmar_parse_one_drhd,
 618                 .cb[ACPI_DMAR_TYPE_RESERVED_MEMORY] = &dmar_parse_one_rmrr,
 619                 .cb[ACPI_DMAR_TYPE_ROOT_ATS] = &dmar_parse_one_atsr,
 620                 .cb[ACPI_DMAR_TYPE_HARDWARE_AFFINITY] = &dmar_parse_one_rhsa,
 621                 .cb[ACPI_DMAR_TYPE_NAMESPACE] = &dmar_parse_one_andd,
 622         };
 623
 624         /*
 625          * Do it again, earlier dmar_tbl mapping could be mapped with
 626          * fixed map.
 627          */
 628         dmar_table_detect();
 629
 630         /*
 631          * ACPI tables may not be DMA protected by tboot, so use DMAR copy
 632          * SINIT saved in SinitMleData in TXT heap (which is DMA protected)
 633          */
 634         dmar_tbl = tboot_get_dmar_table(dmar_tbl);
 635
 636         dmar = (struct acpi_table_dmar *)dmar_tbl;
 637         if (!dmar)
 638                 return -ENODEV;
 639
 640         if (dmar->width < PAGE_SHIFT - 1) {
 641                 pr_warn("Invalid DMAR haw\n");
 642                 return -EINVAL;
 643         }
 644
 645         pr_info("Host address width %d\n", dmar->width + 1);
 646         ret = dmar_walk_dmar_table(dmar, &cb);
 647         if (ret == 0 && drhd_count == 0)
 648                 pr_warn(FW_BUG "No DRHD structure found in DMAR table\n");
 649
 650         return ret;
 651 }
 652
 653 static int dmar_pci_device_match(struct dmar_dev_scope devices[],
 654                                  int cnt, struct pci_dev *dev)
 655 {
 656         int index;
 657         struct device *tmp;
 658
 659         while (dev) {
 660                 for_each_active_dev_scope(devices, cnt, index, tmp)
 661                         if (dev_is_pci(tmp) && dev == to_pci_dev(tmp))
 662                                 return 1;
 663
 664                 /* Check our parent */
 665                 dev = dev->bus->self;
 666         }
 667
 668         return 0;
 669 }
 670
 671 struct dmar_drhd_unit *
 672 dmar_find_matched_drhd_unit(struct pci_dev *dev)
 673 {
 674         struct dmar_drhd_unit *dmaru;
 675         struct acpi_dmar_hardware_unit *drhd;
 676
 677         dev = pci_physfn(dev);
 678
 679         rcu_read_lock();
 680         for_each_drhd_unit(dmaru) {
 681                 drhd = container_of(dmaru->hdr,
 682                                     struct acpi_dmar_hardware_unit,
 683                                     header);
 684
 685                 if (dmaru->include_all &&
 686                     drhd->segment == pci_domain_nr(dev->bus))
 687                         goto out;
 688
 689                 if (dmar_pci_device_match(dmaru->devices,
 690                                           dmaru->devices_cnt, dev))
 691                         goto out;
 692         }
 693         dmaru = NULL;
 694 out:
 695         rcu_read_unlock();
 696
 697         return dmaru;
 698 }
 699
 700 static void __init dmar_acpi_insert_dev_scope(u8 device_number,
 701                                               struct acpi_device *adev)
 702 {
 703         struct dmar_drhd_unit *dmaru;
 704         struct acpi_dmar_hardware_unit *drhd;
 705         struct acpi_dmar_device_scope *scope;
 706         struct device *tmp;
 707         int i;
 708         struct acpi_dmar_pci_path *path;
 709
 710         for_each_drhd_unit(dmaru) {
 711                 drhd = container_of(dmaru->hdr,
 712                                     struct acpi_dmar_hardware_unit,
 713                                     header);
 714
 715                 for (scope = (void *)(drhd + 1);
 716                      (unsigned long)scope < ((unsigned long)drhd) + drhd->header.length;
 717                      scope = ((void *)scope) + scope->length) {
 718                         if (scope->entry_type != ACPI_DMAR_SCOPE_TYPE_NAMESPACE)
 719                                 continue;
 720                         if (scope->enumeration_id != device_number)
 721                                 continue;
 722
 723                         path = (void *)(scope + 1);
 724                         pr_info("ACPI device \"%s\" under DMAR at %llx as %02x:%02x.%d\n",
 725                                 dev_name(&adev->dev), dmaru->reg_base_addr,
 726                                 scope->bus, path->device, path->function);
 727                         for_each_dev_scope(dmaru->devices, dmaru->devices_cnt, i, tmp)
 728                                 if (tmp == NULL) {
 729                                         dmaru->devices[i].bus = scope->bus;
 730                                         dmaru->devices[i].devfn = PCI_DEVFN(path->device,
 731                                                                             path->function);
 732                                         rcu_assign_pointer(dmaru->devices[i].dev,
 733                                                            get_device(&adev->dev));
 734                                         return;
 735                                 }
 736                         BUG_ON(i >= dmaru->devices_cnt);
 737                 }
 738         }
 739         pr_warn("No IOMMU scope found for ANDD enumeration ID %d (%s)\n",
 740                 device_number, dev_name(&adev->dev));
 741 }
 742
 743 static int __init dmar_acpi_dev_scope_init(void)
 744 {
 745         struct acpi_dmar_andd *andd;
 746
 747         if (dmar_tbl == NULL)
 748                 return -ENODEV;
 749
 750         for (andd = (void *)dmar_tbl + sizeof(struct acpi_table_dmar);
 751              ((unsigned long)andd) < ((unsigned long)dmar_tbl) + dmar_tbl->length;
 752              andd = ((void *)andd) + andd->header.length) {
 753                 if (andd->header.type == ACPI_DMAR_TYPE_NAMESPACE) {
 754                         acpi_handle h;
 755                         struct acpi_device *adev;
 756
 757                         if (!ACPI_SUCCESS(acpi_get_handle(ACPI_ROOT_OBJECT,
 758                                                           andd->device_name,
 759                                                           &h))) {
 760                                 pr_err("Failed to find handle for ACPI object %s\n",
 761                                        andd->device_name);
 762                                 continue;
 763                         }
 764                         if (acpi_bus_get_device(h, &adev)) {
 765                                 pr_err("Failed to get device for ACPI object %s\n",
 766                                        andd->device_name);
 767                                 continue;
 768                         }
 769                         dmar_acpi_insert_dev_scope(andd->device_number, adev);
 770                 }
 771         }
 772         return 0;
 773 }
 774
 775 int __init dmar_dev_scope_init(void)
 776 {
 777         struct pci_dev *dev = NULL;
 778         struct dmar_pci_notify_info *info;
 779
 780         if (dmar_dev_scope_status != 1)
 781                 return dmar_dev_scope_status;
 782
 783         if (list_empty(&dmar_drhd_units)) {
 784                 dmar_dev_scope_status = -ENODEV;
 785         } else {
 786                 dmar_dev_scope_status = 0;
 787
 788                 dmar_acpi_dev_scope_init();
 789
 790                 for_each_pci_dev(dev) {
 791                         if (dev->is_virtfn)
 792                                 continue;
 793
 794                         info = dmar_alloc_pci_notify_info(dev,
 795                                         BUS_NOTIFY_ADD_DEVICE);
 796                         if (!info) {
 797                                 return dmar_dev_scope_status;
 798                         } else {
 799                                 dmar_pci_bus_add_dev(info);
 800                                 dmar_free_pci_notify_info(info);
 801                         }
 802                 }
 803         }
 804
 805         return dmar_dev_scope_status;
 806 }
 807
 808 void __init dmar_register_bus_notifier(void)
 809 {
 810         bus_register_notifier(&pci_bus_type, &dmar_pci_bus_nb);
 811 }
 812
 813
 814 int __init dmar_table_init(void)
 815 {
 816         static int dmar_table_initialized;
 817         int ret;
 818
 819         if (dmar_table_initialized == 0) {
 820                 ret = parse_dmar_table();
 821                 if (ret < 0) {
 822                         if (ret != -ENODEV)
 823                                 pr_info("Parse DMAR table failure.\n");
 824                 } else  if (list_empty(&dmar_drhd_units)) {
 825                         pr_info("No DMAR devices found\n");
 826                         ret = -ENODEV;
 827                 }
 828
 829                 if (ret < 0)
 830                         dmar_table_initialized = ret;
 831                 else
 832                         dmar_table_initialized = 1;
 833         }
 834
 835         return dmar_table_initialized < 0 ? dmar_table_initialized : 0;
 836 }
 837
 838 static void warn_invalid_dmar(u64 addr, const char *message)
 839 {
 840         pr_warn_once(FW_BUG
 841                 "Your BIOS is broken; DMAR reported at address %llx%s!\n"
 842                 "BIOS vendor: %s; Ver: %s; Product Version: %s\n",
 843                 addr, message,
 844                 dmi_get_system_info(DMI_BIOS_VENDOR),
 845                 dmi_get_system_info(DMI_BIOS_VERSION),
 846                 dmi_get_system_info(DMI_PRODUCT_VERSION));
 847         add_taint(TAINT_FIRMWARE_WORKAROUND, LOCKDEP_STILL_OK);
 848 }
 849
 850 static int __ref
 851 dmar_validate_one_drhd(struct acpi_dmar_header *entry, void *arg)
 852 {
 853         struct acpi_dmar_hardware_unit *drhd;
 854         void __iomem *addr;
 855         u64 cap, ecap;
 856
 857         drhd = (void *)entry;
 858         if (!drhd->address) {
 859                 warn_invalid_dmar(0, "");
 860                 return -EINVAL;
 861         }
 862
 863         if (arg)
 864                 addr = ioremap(drhd->address, VTD_PAGE_SIZE);
 865         else
 866                 addr = early_ioremap(drhd->address, VTD_PAGE_SIZE);
 867         if (!addr) {
 868                 pr_warn("Can't validate DRHD address: %llx\n", drhd->address);
 869                 return -EINVAL;
 870         }
 871
 872         cap = dmar_readq(addr + DMAR_CAP_REG);
 873         ecap = dmar_readq(addr + DMAR_ECAP_REG);
 874
 875         if (arg)
 876                 iounmap(addr);
 877         else
 878                 early_iounmap(addr, VTD_PAGE_SIZE);
 879
 880         if (cap == (uint64_t)-1 && ecap == (uint64_t)-1) {
 881                 warn_invalid_dmar(drhd->address, " returns all ones");
 882                 return -EINVAL;
 883         }
 884
 885         return 0;
 886 }
 887
 888 int __init detect_intel_iommu(void)
 889 {
 890         int ret;
 891         struct dmar_res_callback validate_drhd_cb = {
 892                 .cb[ACPI_DMAR_TYPE_HARDWARE_UNIT] = &dmar_validate_one_drhd,
 893                 .ignore_unhandled = true,
 894         };
 895
 896         down_write(&dmar_global_lock);
 897         ret = dmar_table_detect();
 898         if (!ret)
 899                 ret = dmar_walk_dmar_table((struct acpi_table_dmar *)dmar_tbl,
 900                                            &validate_drhd_cb);
 901         if (!ret && !no_iommu && !iommu_detected && !dmar_disabled) {
 902                 iommu_detected = 1;
 903                 /* Make sure ACS will be enabled */
 904                 pci_request_acs();
 905         }
 906
 907 #ifdef CONFIG_X86
 908         if (!ret) {
 909                 x86_init.iommu.iommu_init = intel_iommu_init;
 910                 x86_platform.iommu_shutdown = intel_iommu_shutdown;
 911         }
 912
 913 #endif
 914
 915         if (dmar_tbl) {
 916                 acpi_put_table(dmar_tbl);
 917                 dmar_tbl = NULL;
 918         }
 919         up_write(&dmar_global_lock);
 920
 921         return ret ? ret : 1;
 922 }
 923
 924 static void unmap_iommu(struct intel_iommu *iommu)
 925 {
 926         iounmap(iommu->reg);
 927         release_mem_region(iommu->reg_phys, iommu->reg_size);
 928 }
 929
 930 /**
 931  * map_iommu: map the iommu's registers
 932  * @iommu: the iommu to map
 933  * @phys_addr: the physical address of the base resgister
 934  *
 935  * Memory map the iommu's registers.  Start w/ a single page, and
 936  * possibly expand if that turns out to be insufficent.
 937  */
 938 static int map_iommu(struct intel_iommu *iommu, u64 phys_addr)
 939 {
 940         int map_size, err=0;
 941
 942         iommu->reg_phys = phys_addr;
 943         iommu->reg_size = VTD_PAGE_SIZE;
 944
 945         if (!request_mem_region(iommu->reg_phys, iommu->reg_size, iommu->name)) {
 946                 pr_err("Can't reserve memory\n");
 947                 err = -EBUSY;
 948                 goto out;
 949         }
 950
 951         iommu->reg = ioremap(iommu->reg_phys, iommu->reg_size);
 952         if (!iommu->reg) {
 953                 pr_err("Can't map the region\n");
 954                 err = -ENOMEM;
 955                 goto release;
 956         }
 957
 958         iommu->cap = dmar_readq(iommu->reg + DMAR_CAP_REG);
 959         iommu->ecap = dmar_readq(iommu->reg + DMAR_ECAP_REG);
 960
 961         if (iommu->cap == (uint64_t)-1 && iommu->ecap == (uint64_t)-1) {
 962                 err = -EINVAL;
 963                 warn_invalid_dmar(phys_addr, " returns all ones");
 964                 goto unmap;
 965         }
 966         iommu->vccap = dmar_readq(iommu->reg + DMAR_VCCAP_REG);
 967
 968         /* the registers might be more than one page */
 969         map_size = max_t(int, ecap_max_iotlb_offset(iommu->ecap),
 970                          cap_max_fault_reg_offset(iommu->cap));
 971         map_size = VTD_PAGE_ALIGN(map_size);
 972         if (map_size > iommu->reg_size) {
 973                 iounmap(iommu->reg);
 974                 release_mem_region(iommu->reg_phys, iommu->reg_size);
 975                 iommu->reg_size = map_size;
 976                 if (!request_mem_region(iommu->reg_phys, iommu->reg_size,
 977                                         iommu->name)) {
 978                         pr_err("Can't reserve memory\n");
 979                         err = -EBUSY;
 980                         goto out;
 981                 }
 982                 iommu->reg = ioremap(iommu->reg_phys, iommu->reg_size);
 983                 if (!iommu->reg) {
 984                         pr_err("Can't map the region\n");
 985                         err = -ENOMEM;
 986                         goto release;
 987                 }
 988         }
 989         err = 0;
 990         goto out;
 991
 992 unmap:
 993         iounmap(iommu->reg);
 994 release:
 995         release_mem_region(iommu->reg_phys, iommu->reg_size);
 996 out:
 997         return err;
 998 }
 999
1000 static int dmar_alloc_seq_id(struct intel_iommu *iommu)
1001 {
1002         iommu->seq_id = find_first_zero_bit(dmar_seq_ids,
1003                                             DMAR_UNITS_SUPPORTED);
1004         if (iommu->seq_id >= DMAR_UNITS_SUPPORTED) {
1005                 iommu->seq_id = -1;
1006         } else {
1007                 set_bit(iommu->seq_id, dmar_seq_ids);
1008                 sprintf(iommu->name, "dmar%d", iommu->seq_id);
1009         }
1010
1011         return iommu->seq_id;
1012 }
1013
1014 static void dmar_free_seq_id(struct intel_iommu *iommu)
1015 {
1016         if (iommu->seq_id >= 0) {
1017                 clear_bit(iommu->seq_id, dmar_seq_ids);
1018                 iommu->seq_id = -1;
1019         }
1020 }
1021
1022 static int alloc_iommu(struct dmar_drhd_unit *drhd)
1023 {
1024         struct intel_iommu *iommu;
1025         u32 ver, sts;
1026         int agaw = 0;
1027         int msagaw = 0;
1028         int err;
1029
1030         if (!drhd->reg_base_addr) {
1031                 warn_invalid_dmar(0, "");
1032                 return -EINVAL;
1033         }
1034
1035         iommu = kzalloc(sizeof(*iommu), GFP_KERNEL);
1036         if (!iommu)
1037                 return -ENOMEM;
1038
1039         if (dmar_alloc_seq_id(iommu) < 0) {
1040                 pr_err("Failed to allocate seq_id\n");
1041                 err = -ENOSPC;
1042                 goto error;
1043         }
1044
1045         err = map_iommu(iommu, drhd->reg_base_addr);
1046         if (err) {
1047                 pr_err("Failed to map %s\n", iommu->name);
1048                 goto error_free_seq_id;
1049         }
1050
1051         err = -EINVAL;
1052         agaw = iommu_calculate_agaw(iommu);
1053         if (agaw < 0) {
1054                 pr_err("Cannot get a valid agaw for iommu (seq_id = %d)\n",
1055                         iommu->seq_id);
1056                 goto err_unmap;
1057         }
1058         msagaw = iommu_calculate_max_sagaw(iommu);
1059         if (msagaw < 0) {
1060                 pr_err("Cannot get a valid max agaw for iommu (seq_id = %d)\n",
1061                         iommu->seq_id);
1062                 goto err_unmap;
1063         }
1064         iommu->agaw = agaw;
1065         iommu->msagaw = msagaw;
1066         iommu->segment = drhd->segment;
1067
1068         iommu->node = NUMA_NO_NODE;
1069
1070         ver = readl(iommu->reg + DMAR_VER_REG);
1071         pr_info("%s: reg_base_addr %llx ver %d:%d cap %llx ecap %llx\n",
1072                 iommu->name,
1073                 (unsigned long long)drhd->reg_base_addr,
1074                 DMAR_VER_MAJOR(ver), DMAR_VER_MINOR(ver),
1075                 (unsigned long long)iommu->cap,
1076                 (unsigned long long)iommu->ecap);
1077
1078         /* Reflect status in gcmd */
1079         sts = readl(iommu->reg + DMAR_GSTS_REG);
1080         if (sts & DMA_GSTS_IRES)
1081                 iommu->gcmd |= DMA_GCMD_IRE;
1082         if (sts & DMA_GSTS_TES)
1083                 iommu->gcmd |= DMA_GCMD_TE;
1084         if (sts & DMA_GSTS_QIES)
1085                 iommu->gcmd |= DMA_GCMD_QIE;
1086
1087         raw_spin_lock_init(&iommu->register_lock);
1088
1089         if (intel_iommu_enabled) {
1090                 err = iommu_device_sysfs_add(&iommu->iommu, NULL,
1091                                              intel_iommu_groups,
1092                                              "%s", iommu->name);
1093                 if (err)
1094                         goto err_unmap;
1095
1096                 iommu_device_set_ops(&iommu->iommu, &intel_iommu_ops);
1097
1098                 err = iommu_device_register(&iommu->iommu);
1099                 if (err)
1100                         goto err_unmap;
1101         }
1102
1103         drhd->iommu = iommu;
1104
1105         return 0;
1106
1107 err_unmap:
1108         unmap_iommu(iommu);
1109 error_free_seq_id:
1110         dmar_free_seq_id(iommu);
1111 error:
1112         kfree(iommu);
1113         return err;
1114 }
1115
1116 static void free_iommu(struct intel_iommu *iommu)
1117 {
1118         if (intel_iommu_enabled) {
1119                 iommu_device_unregister(&iommu->iommu);
1120                 iommu_device_sysfs_remove(&iommu->iommu);
1121         }
1122
1123         if (iommu->irq) {
1124                 if (iommu->pr_irq) {
1125                         free_irq(iommu->pr_irq, iommu);
1126                         dmar_free_hwirq(iommu->pr_irq);
1127                         iommu->pr_irq = 0;
1128                 }
1129                 free_irq(iommu->irq, iommu);
1130                 dmar_free_hwirq(iommu->irq);
1131                 iommu->irq = 0;
1132         }
1133
1134         if (iommu->qi) {
1135                 free_page((unsigned long)iommu->qi->desc);
1136                 kfree(iommu->qi->desc_status);
1137                 kfree(iommu->qi);
1138         }
1139
1140         if (iommu->reg)
1141                 unmap_iommu(iommu);
1142
1143         dmar_free_seq_id(iommu);
1144         kfree(iommu);
1145 }
1146
1147 /*
1148  * Reclaim all the submitted descriptors which have completed its work.
1149  */
1150 static inline void reclaim_free_desc(struct q_inval *qi)
1151 {
1152         while (qi->desc_status[qi->free_tail] == QI_DONE ||
1153                qi->desc_status[qi->free_tail] == QI_ABORT) {
1154                 qi->desc_status[qi->free_tail] = QI_FREE;
1155                 qi->free_tail = (qi->free_tail + 1) % QI_LENGTH;
1156                 qi->free_cnt++;
1157         }
1158 }
1159
1160 static int qi_check_fault(struct intel_iommu *iommu, int index, int wait_index)
1161 {
1162         u32 fault;
1163         int head, tail;
1164         struct q_inval *qi = iommu->qi;
1165         int shift = qi_shift(iommu);
1166
1167         if (qi->desc_status[wait_index] == QI_ABORT)
1168                 return -EAGAIN;
1169
1170         fault = readl(iommu->reg + DMAR_FSTS_REG);
1171
1172         /*
1173          * If IQE happens, the head points to the descriptor associated
1174          * with the error. No new descriptors are fetched until the IQE
1175          * is cleared.
1176          */
1177         if (fault & DMA_FSTS_IQE) {
1178                 head = readl(iommu->reg + DMAR_IQH_REG);
1179                 if ((head >> shift) == index) {
1180                         struct qi_desc *desc = qi->desc + head;
1181
1182                         /*
1183                          * desc->qw2 and desc->qw3 are either reserved or
1184                          * used by software as private data. We won't print
1185                          * out these two qw's for security consideration.
1186                          */
1187                         pr_err("VT-d detected invalid descriptor: qw0 = %llx, qw1 = %llx\n",
1188                                (unsigned long long)desc->qw0,
1189                                (unsigned long long)desc->qw1);
1190                         memcpy(desc, qi->desc + (wait_index << shift),
1191                                1 << shift);
1192                         writel(DMA_FSTS_IQE, iommu->reg + DMAR_FSTS_REG);
1193                         return -EINVAL;
1194                 }
1195         }
1196
1197         /*
1198          * If ITE happens, all pending wait_desc commands are aborted.
1199          * No new descriptors are fetched until the ITE is cleared.
1200          */
1201         if (fault & DMA_FSTS_ITE) {
1202                 head = readl(iommu->reg + DMAR_IQH_REG);
1203                 head = ((head >> shift) - 1 + QI_LENGTH) % QI_LENGTH;
1204                 head |= 1;
1205                 tail = readl(iommu->reg + DMAR_IQT_REG);
1206                 tail = ((tail >> shift) - 1 + QI_LENGTH) % QI_LENGTH;
1207
1208                 writel(DMA_FSTS_ITE, iommu->reg + DMAR_FSTS_REG);
1209
1210                 do {
1211                         if (qi->desc_status[head] == QI_IN_USE)
1212                                 qi->desc_status[head] = QI_ABORT;
1213                         head = (head - 2 + QI_LENGTH) % QI_LENGTH;
1214                 } while (head != tail);
1215
1216                 if (qi->desc_status[wait_index] == QI_ABORT)
1217                         return -EAGAIN;
1218         }
1219
1220         if (fault & DMA_FSTS_ICE)
1221                 writel(DMA_FSTS_ICE, iommu->reg + DMAR_FSTS_REG);
1222
1223         return 0;
1224 }
1225
1226 /*
1227  * Function to submit invalidation descriptors of all types to the queued
1228  * invalidation interface(QI). Multiple descriptors can be submitted at a
1229  * time, a wait descriptor will be appended to each submission to ensure
1230  * hardware has completed the invalidation before return. Wait descriptors
1231  * can be part of the submission but it will not be polled for completion.
1232  */
1233 int qi_submit_sync(struct intel_iommu *iommu, struct qi_desc *desc,
1234                    unsigned int count, unsigned long options)
1235 {
1236         struct q_inval *qi = iommu->qi;
1237         struct qi_desc wait_desc;
1238         int wait_index, index;
1239         unsigned long flags;
1240         int offset, shift;
1241         int rc, i;
1242
1243         if (!qi)
1244                 return 0;
1245
1246 restart:
1247         rc = 0;
1248
1249         raw_spin_lock_irqsave(&qi->q_lock, flags);
1250         /*
1251          * Check if we have enough empty slots in the queue to submit,
1252          * the calculation is based on:
1253          * # of desc + 1 wait desc + 1 space between head and tail
1254          */
1255         while (qi->free_cnt < count + 2) {
1256                 raw_spin_unlock_irqrestore(&qi->q_lock, flags);
1257                 cpu_relax();
1258                 raw_spin_lock_irqsave(&qi->q_lock, flags);
1259         }
1260
1261         index = qi->free_head;
1262         wait_index = (index + count) % QI_LENGTH;
1263         shift = qi_shift(iommu);
1264
1265         for (i = 0; i < count; i++) {
1266                 offset = ((index + i) % QI_LENGTH) << shift;
1267                 memcpy(qi->desc + offset, &desc[i], 1 << shift);
1268                 qi->desc_status[(index + i) % QI_LENGTH] = QI_IN_USE;
1269         }
1270         qi->desc_status[wait_index] = QI_IN_USE;
1271
1272         wait_desc.qw0 = QI_IWD_STATUS_DATA(QI_DONE) |
1273                         QI_IWD_STATUS_WRITE | QI_IWD_TYPE;
1274         if (options & QI_OPT_WAIT_DRAIN)
1275                 wait_desc.qw0 |= QI_IWD_PRQ_DRAIN;
1276         wait_desc.qw1 = virt_to_phys(&qi->desc_status[wait_index]);
1277         wait_desc.qw2 = 0;
1278         wait_desc.qw3 = 0;
1279
1280         offset = wait_index << shift;
1281         memcpy(qi->desc + offset, &wait_desc, 1 << shift);
1282
1283         qi->free_head = (qi->free_head + count + 1) % QI_LENGTH;
1284         qi->free_cnt -= count + 1;
1285
1286         /*
1287          * update the HW tail register indicating the presence of
1288          * new descriptors.
1289          */
1290         writel(qi->free_head << shift, iommu->reg + DMAR_IQT_REG);
1291
1292         while (qi->desc_status[wait_index] != QI_DONE) {
1293                 /*
1294                  * We will leave the interrupts disabled, to prevent interrupt
1295                  * context to queue another cmd while a cmd is already submitted
1296                  * and waiting for completion on this cpu. This is to avoid
1297                  * a deadlock where the interrupt context can wait indefinitely
1298                  * for free slots in the queue.
1299                  */
1300                 rc = qi_check_fault(iommu, index, wait_index);
1301                 if (rc)
1302                         break;
1303
1304                 raw_spin_unlock(&qi->q_lock);
1305                 cpu_relax();
1306                 raw_spin_lock(&qi->q_lock);
1307         }
1308
1309         for (i = 0; i < count; i++)
1310                 qi->desc_status[(index + i) % QI_LENGTH] = QI_DONE;
1311
1312         reclaim_free_desc(qi);
1313         raw_spin_unlock_irqrestore(&qi->q_lock, flags);
1314
1315         if (rc == -EAGAIN)
1316                 goto restart;
1317
1318         return rc;
1319 }
1320
1321 /*
1322  * Flush the global interrupt entry cache.
1323  */
1324 void qi_global_iec(struct intel_iommu *iommu)
1325 {
1326         struct qi_desc desc;
1327
1328         desc.qw0 = QI_IEC_TYPE;
1329         desc.qw1 = 0;
1330         desc.qw2 = 0;
1331         desc.qw3 = 0;
1332
1333         /* should never fail */
1334         qi_submit_sync(iommu, &desc, 1, 0);
1335 }
1336
1337 void qi_flush_context(struct intel_iommu *iommu, u16 did, u16 sid, u8 fm,
1338                       u64 type)
1339 {
1340         struct qi_desc desc;
1341
1342         desc.qw0 = QI_CC_FM(fm) | QI_CC_SID(sid) | QI_CC_DID(did)
1343                         | QI_CC_GRAN(type) | QI_CC_TYPE;
1344         desc.qw1 = 0;
1345         desc.qw2 = 0;
1346         desc.qw3 = 0;
1347
1348         qi_submit_sync(iommu, &desc, 1, 0);
1349 }
1350
1351 void qi_flush_iotlb(struct intel_iommu *iommu, u16 did, u64 addr,
1352                     unsigned int size_order, u64 type)
1353 {
1354         u8 dw = 0, dr = 0;
1355
1356         struct qi_desc desc;
1357         int ih = 0;
1358
1359         if (cap_write_drain(iommu->cap))
1360                 dw = 1;
1361
1362         if (cap_read_drain(iommu->cap))
1363                 dr = 1;
1364
1365         desc.qw0 = QI_IOTLB_DID(did) | QI_IOTLB_DR(dr) | QI_IOTLB_DW(dw)
1366                 | QI_IOTLB_GRAN(type) | QI_IOTLB_TYPE;
1367         desc.qw1 = QI_IOTLB_ADDR(addr) | QI_IOTLB_IH(ih)
1368                 | QI_IOTLB_AM(size_order);
1369         desc.qw2 = 0;
1370         desc.qw3 = 0;
1371
1372         qi_submit_sync(iommu, &desc, 1, 0);
1373 }
1374
1375 void qi_flush_dev_iotlb(struct intel_iommu *iommu, u16 sid, u16 pfsid,
1376                         u16 qdep, u64 addr, unsigned mask)
1377 {
1378         struct qi_desc desc;
1379
1380         if (mask) {
1381                 addr |= (1ULL << (VTD_PAGE_SHIFT + mask - 1)) - 1;
1382                 desc.qw1 = QI_DEV_IOTLB_ADDR(addr) | QI_DEV_IOTLB_SIZE;
1383         } else
1384                 desc.qw1 = QI_DEV_IOTLB_ADDR(addr);
1385
1386         if (qdep >= QI_DEV_IOTLB_MAX_INVS)
1387                 qdep = 0;
1388
1389         desc.qw0 = QI_DEV_IOTLB_SID(sid) | QI_DEV_IOTLB_QDEP(qdep) |
1390                    QI_DIOTLB_TYPE | QI_DEV_IOTLB_PFSID(pfsid);
1391         desc.qw2 = 0;
1392         desc.qw3 = 0;
1393
1394         qi_submit_sync(iommu, &desc, 1, 0);
1395 }
1396
1397 /* PASID-based IOTLB invalidation */
1398 void qi_flush_piotlb(struct intel_iommu *iommu, u16 did, u32 pasid, u64 addr,
1399                      unsigned long npages, bool ih)
1400 {
1401         struct qi_desc desc = {.qw2 = 0, .qw3 = 0};
1402
1403         /*
1404          * npages == -1 means a PASID-selective invalidation, otherwise,
1405          * a positive value for Page-selective-within-PASID invalidation.
1406          * 0 is not a valid input.
1407          */
1408         if (WARN_ON(!npages)) {
1409                 pr_err("Invalid input npages = %ld\n", npages);
1410                 return;
1411         }
1412
1413         if (npages == -1) {
1414                 desc.qw0 = QI_EIOTLB_PASID(pasid) |
1415                                 QI_EIOTLB_DID(did) |
1416                                 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
1417                                 QI_EIOTLB_TYPE;
1418                 desc.qw1 = 0;
1419         } else {
1420                 int mask = ilog2(__roundup_pow_of_two(npages));
1421                 unsigned long align = (1ULL << (VTD_PAGE_SHIFT + mask));
1422
1423                 if (WARN_ON_ONCE(!ALIGN(addr, align)))
1424                         addr &= ~(align - 1);
1425
1426                 desc.qw0 = QI_EIOTLB_PASID(pasid) |
1427                                 QI_EIOTLB_DID(did) |
1428                                 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
1429                                 QI_EIOTLB_TYPE;
1430                 desc.qw1 = QI_EIOTLB_ADDR(addr) |
1431                                 QI_EIOTLB_IH(ih) |
1432                                 QI_EIOTLB_AM(mask);
1433         }
1434
1435         qi_submit_sync(iommu, &desc, 1, 0);
1436 }
1437
1438 /* PASID-based device IOTLB Invalidate */
1439 void qi_flush_dev_iotlb_pasid(struct intel_iommu *iommu, u16 sid, u16 pfsid,
1440                               u32 pasid,  u16 qdep, u64 addr,
1441                               unsigned int size_order, u64 granu)
1442 {
1443         unsigned long mask = 1UL << (VTD_PAGE_SHIFT + size_order - 1);
1444         struct qi_desc desc = {.qw1 = 0, .qw2 = 0, .qw3 = 0};
1445
1446         desc.qw0 = QI_DEV_EIOTLB_PASID(pasid) | QI_DEV_EIOTLB_SID(sid) |
1447                 QI_DEV_EIOTLB_QDEP(qdep) | QI_DEIOTLB_TYPE |
1448                 QI_DEV_IOTLB_PFSID(pfsid);
1449         desc.qw1 = QI_DEV_EIOTLB_GLOB(granu);
1450
1451         /*
1452          * If S bit is 0, we only flush a single page. If S bit is set,
1453          * The least significant zero bit indicates the invalidation address
1454          * range. VT-d spec 6.5.2.6.
1455          * e.g. address bit 12[0] indicates 8KB, 13[0] indicates 16KB.
1456          * size order = 0 is PAGE_SIZE 4KB
1457          * Max Invs Pending (MIP) is set to 0 for now until we have DIT in
1458          * ECAP.
1459          */
1460         desc.qw1 |= addr & ~mask;
1461         if (size_order)
1462                 desc.qw1 |= QI_DEV_EIOTLB_SIZE;
1463
1464         qi_submit_sync(iommu, &desc, 1, 0);
1465 }
1466
1467 void qi_flush_pasid_cache(struct intel_iommu *iommu, u16 did,
1468                           u64 granu, int pasid)
1469 {
1470         struct qi_desc desc = {.qw1 = 0, .qw2 = 0, .qw3 = 0};
1471
1472         desc.qw0 = QI_PC_PASID(pasid) | QI_PC_DID(did) |
1473                         QI_PC_GRAN(granu) | QI_PC_TYPE;
1474         qi_submit_sync(iommu, &desc, 1, 0);
1475 }
1476
1477 /*
1478  * Disable Queued Invalidation interface.
1479  */
1480 void dmar_disable_qi(struct intel_iommu *iommu)
1481 {
1482         unsigned long flags;
1483         u32 sts;
1484         cycles_t start_time = get_cycles();
1485
1486         if (!ecap_qis(iommu->ecap))
1487                 return;
1488
1489         raw_spin_lock_irqsave(&iommu->register_lock, flags);
1490
1491         sts =  readl(iommu->reg + DMAR_GSTS_REG);
1492         if (!(sts & DMA_GSTS_QIES))
1493                 goto end;
1494
1495         /*
1496          * Give a chance to HW to complete the pending invalidation requests.
1497          */
1498         while ((readl(iommu->reg + DMAR_IQT_REG) !=
1499                 readl(iommu->reg + DMAR_IQH_REG)) &&
1500                 (DMAR_OPERATION_TIMEOUT > (get_cycles() - start_time)))
1501                 cpu_relax();
1502
1503         iommu->gcmd &= ~DMA_GCMD_QIE;
1504         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1505
1506         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG, readl,
1507                       !(sts & DMA_GSTS_QIES), sts);
1508 end:
1509         raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1510 }
1511
1512 /*
1513  * Enable queued invalidation.
1514  */
1515 static void __dmar_enable_qi(struct intel_iommu *iommu)
1516 {
1517         u32 sts;
1518         unsigned long flags;
1519         struct q_inval *qi = iommu->qi;
1520         u64 val = virt_to_phys(qi->desc);
1521
1522         qi->free_head = qi->free_tail = 0;
1523         qi->free_cnt = QI_LENGTH;
1524
1525         /*
1526          * Set DW=1 and QS=1 in IQA_REG when Scalable Mode capability
1527          * is present.
1528          */
1529         if (ecap_smts(iommu->ecap))
1530                 val |= (1 << 11) | 1;
1531
1532         raw_spin_lock_irqsave(&iommu->register_lock, flags);
1533
1534         /* write zero to the tail reg */
1535         writel(0, iommu->reg + DMAR_IQT_REG);
1536
1537         dmar_writeq(iommu->reg + DMAR_IQA_REG, val);
1538
1539         iommu->gcmd |= DMA_GCMD_QIE;
1540         writel(iommu->gcmd, iommu->reg + DMAR_GCMD_REG);
1541
1542         /* Make sure hardware complete it */
1543         IOMMU_WAIT_OP(iommu, DMAR_GSTS_REG, readl, (sts & DMA_GSTS_QIES), sts);
1544
1545         raw_spin_unlock_irqrestore(&iommu->register_lock, flags);
1546 }
1547
1548 /*
1549  * Enable Queued Invalidation interface. This is a must to support
1550  * interrupt-remapping. Also used by DMA-remapping, which replaces
1551  * register based IOTLB invalidation.
1552  */
1553 int dmar_enable_qi(struct intel_iommu *iommu)
1554 {
1555         struct q_inval *qi;
1556         struct page *desc_page;
1557
1558         if (!ecap_qis(iommu->ecap))
1559                 return -ENOENT;
1560
1561         /*
1562          * queued invalidation is already setup and enabled.
1563          */
1564         if (iommu->qi)
1565                 return 0;
1566
1567         iommu->qi = kmalloc(sizeof(*qi), GFP_ATOMIC);
1568         if (!iommu->qi)
1569                 return -ENOMEM;
1570
1571         qi = iommu->qi;
1572
1573         /*
1574          * Need two pages to accommodate 256 descriptors of 256 bits each
1575          * if the remapping hardware supports scalable mode translation.
1576          */
1577         desc_page = alloc_pages_node(iommu->node, GFP_ATOMIC | __GFP_ZERO,
1578                                      !!ecap_smts(iommu->ecap));
1579         if (!desc_page) {
1580                 kfree(qi);
1581                 iommu->qi = NULL;
1582                 return -ENOMEM;
1583         }
1584
1585         qi->desc = page_address(desc_page);
1586
1587         qi->desc_status = kcalloc(QI_LENGTH, sizeof(int), GFP_ATOMIC);
1588         if (!qi->desc_status) {
1589                 free_page((unsigned long) qi->desc);
1590                 kfree(qi);
1591                 iommu->qi = NULL;
1592                 return -ENOMEM;
1593         }
1594
1595         raw_spin_lock_init(&qi->q_lock);
1596
1597         __dmar_enable_qi(iommu);
1598
1599         return 0;
1600 }
1601
1602 /* iommu interrupt handling. Most stuff are MSI-like. */
1603
1604 enum faulttype {
1605         DMA_REMAP,
1606         INTR_REMAP,
1607         UNKNOWN,
1608 };
1609
1610 static const char *dma_remap_fault_reasons[] =
1611 {
1612         "Software",
1613         "Present bit in root entry is clear",
1614         "Present bit in context entry is clear",
1615         "Invalid context entry",
1616         "Access beyond MGAW",
1617         "PTE Write access is not set",
1618         "PTE Read access is not set",
1619         "Next page table ptr is invalid",
1620         "Root table address invalid",
1621         "Context table ptr is invalid",
1622         "non-zero reserved fields in RTP",
1623         "non-zero reserved fields in CTP",
1624         "non-zero reserved fields in PTE",
1625         "PCE for translation request specifies blocking",
1626 };
1627
1628 static const char * const dma_remap_sm_fault_reasons[] = {
1629         "SM: Invalid Root Table Address",
1630         "SM: TTM 0 for request with PASID",
1631         "SM: TTM 0 for page group request",
1632         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x33-0x37 */
1633         "SM: Error attempting to access Root Entry",
1634         "SM: Present bit in Root Entry is clear",
1635         "SM: Non-zero reserved field set in Root Entry",
1636         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x3B-0x3F */
1637         "SM: Error attempting to access Context Entry",
1638         "SM: Present bit in Context Entry is clear",
1639         "SM: Non-zero reserved field set in the Context Entry",
1640         "SM: Invalid Context Entry",
1641         "SM: DTE field in Context Entry is clear",
1642         "SM: PASID Enable field in Context Entry is clear",
1643         "SM: PASID is larger than the max in Context Entry",
1644         "SM: PRE field in Context-Entry is clear",
1645         "SM: RID_PASID field error in Context-Entry",
1646         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x49-0x4F */
1647         "SM: Error attempting to access the PASID Directory Entry",
1648         "SM: Present bit in Directory Entry is clear",
1649         "SM: Non-zero reserved field set in PASID Directory Entry",
1650         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x53-0x57 */
1651         "SM: Error attempting to access PASID Table Entry",
1652         "SM: Present bit in PASID Table Entry is clear",
1653         "SM: Non-zero reserved field set in PASID Table Entry",
1654         "SM: Invalid Scalable-Mode PASID Table Entry",
1655         "SM: ERE field is clear in PASID Table Entry",
1656         "SM: SRE field is clear in PASID Table Entry",
1657         "Unknown", "Unknown",/* 0x5E-0x5F */
1658         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x60-0x67 */
1659         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x68-0x6F */
1660         "SM: Error attempting to access first-level paging entry",
1661         "SM: Present bit in first-level paging entry is clear",
1662         "SM: Non-zero reserved field set in first-level paging entry",
1663         "SM: Error attempting to access FL-PML4 entry",
1664         "SM: First-level entry address beyond MGAW in Nested translation",
1665         "SM: Read permission error in FL-PML4 entry in Nested translation",
1666         "SM: Read permission error in first-level paging entry in Nested translation",
1667         "SM: Write permission error in first-level paging entry in Nested translation",
1668         "SM: Error attempting to access second-level paging entry",
1669         "SM: Read/Write permission error in second-level paging entry",
1670         "SM: Non-zero reserved field set in second-level paging entry",
1671         "SM: Invalid second-level page table pointer",
1672         "SM: A/D bit update needed in second-level entry when set up in no snoop",
1673         "Unknown", "Unknown", "Unknown", /* 0x7D-0x7F */
1674         "SM: Address in first-level translation is not canonical",
1675         "SM: U/S set 0 for first-level translation with user privilege",
1676         "SM: No execute permission for request with PASID and ER=1",
1677         "SM: Address beyond the DMA hardware max",
1678         "SM: Second-level entry address beyond the max",
1679         "SM: No write permission for Write/AtomicOp request",
1680         "SM: No read permission for Read/AtomicOp request",
1681         "SM: Invalid address-interrupt address",
1682         "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", "Unknown", /* 0x88-0x8F */
1683         "SM: A/D bit update needed in first-level entry when set up in no snoop",
1684 };
1685
1686 static const char *irq_remap_fault_reasons[] =
1687 {
1688         "Detected reserved fields in the decoded interrupt-remapped request",
1689         "Interrupt index exceeded the interrupt-remapping table size",
1690         "Present field in the IRTE entry is clear",
1691         "Error accessing interrupt-remapping table pointed by IRTA_REG",
1692         "Detected reserved fields in the IRTE entry",
1693         "Blocked a compatibility format interrupt request",
1694         "Blocked an interrupt request due to source-id verification failure",
1695 };
1696
1697 static const char *dmar_get_fault_reason(u8 fault_reason, int *fault_type)
1698 {
1699         if (fault_reason >= 0x20 && (fault_reason - 0x20 <
1700                                         ARRAY_SIZE(irq_remap_fault_reasons))) {
1701                 *fault_type = INTR_REMAP;
1702                 return irq_remap_fault_reasons[fault_reason - 0x20];
1703         } else if (fault_reason >= 0x30 && (fault_reason - 0x30 <
1704                         ARRAY_SIZE(dma_remap_sm_fault_reasons))) {
1705                 *fault_type = DMA_REMAP;
1706                 return dma_remap_sm_fault_reasons[fault_reason - 0x30];
1707         } else if (fault_reason < ARRAY_SIZE(dma_remap_fault_reasons)) {
1708                 *fault_type = DMA_REMAP;
1709                 return dma_remap_fault_reasons[fault_reason];
1710         } else {
1711                 *fault_type = UNKNOWN;
1712                 return "Unknown";
1713         }
1714 }
1715
1716
1717 static inline int dmar_msi_reg(struct intel_iommu *iommu, int irq)
1718 {
1719         if (iommu->irq == irq)
1720                 return DMAR_FECTL_REG;
1721         else if (iommu->pr_irq == irq)
1722                 return DMAR_PECTL_REG;
1723         else
1724                 BUG();
1725 }
1726
1727 void dmar_msi_unmask(struct irq_data *data)
1728 {
1729         struct intel_iommu *iommu = irq_data_get_irq_handler_data(data);
1730         int reg = dmar_msi_reg(iommu, data->irq);
1731         unsigned long flag;
1732
1733         /* unmask it */
1734         raw_spin_lock_irqsave(&iommu->register_lock, flag);
1735         writel(0, iommu->reg + reg);
1736         /* Read a reg to force flush the post write */
1737         readl(iommu->reg + reg);
1738         raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1739 }
1740
1741 void dmar_msi_mask(struct irq_data *data)
1742 {
1743         struct intel_iommu *iommu = irq_data_get_irq_handler_data(data);
1744         int reg = dmar_msi_reg(iommu, data->irq);
1745         unsigned long flag;
1746
1747         /* mask it */
1748         raw_spin_lock_irqsave(&iommu->register_lock, flag);
1749         writel(DMA_FECTL_IM, iommu->reg + reg);
1750         /* Read a reg to force flush the post write */
1751         readl(iommu->reg + reg);
1752         raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1753 }
1754
1755 void dmar_msi_write(int irq, struct msi_msg *msg)
1756 {
1757         struct intel_iommu *iommu = irq_get_handler_data(irq);
1758         int reg = dmar_msi_reg(iommu, irq);
1759         unsigned long flag;
1760
1761         raw_spin_lock_irqsave(&iommu->register_lock, flag);
1762         writel(msg->data, iommu->reg + reg + 4);
1763         writel(msg->address_lo, iommu->reg + reg + 8);
1764         writel(msg->address_hi, iommu->reg + reg + 12);
1765         raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1766 }
1767
1768 void dmar_msi_read(int irq, struct msi_msg *msg)
1769 {
1770         struct intel_iommu *iommu = irq_get_handler_data(irq);
1771         int reg = dmar_msi_reg(iommu, irq);
1772         unsigned long flag;
1773
1774         raw_spin_lock_irqsave(&iommu->register_lock, flag);
1775         msg->data = readl(iommu->reg + reg + 4);
1776         msg->address_lo = readl(iommu->reg + reg + 8);
1777         msg->address_hi = readl(iommu->reg + reg + 12);
1778         raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1779 }
1780
1781 static int dmar_fault_do_one(struct intel_iommu *iommu, int type,
1782                 u8 fault_reason, int pasid, u16 source_id,
1783                 unsigned long long addr)
1784 {
1785         const char *reason;
1786         int fault_type;
1787
1788         reason = dmar_get_fault_reason(fault_reason, &fault_type);
1789
1790         if (fault_type == INTR_REMAP)
1791                 pr_err("[INTR-REMAP] Request device [%02x:%02x.%d] fault index %llx [fault reason %02d] %s\n",
1792                         source_id >> 8, PCI_SLOT(source_id & 0xFF),
1793                         PCI_FUNC(source_id & 0xFF), addr >> 48,
1794                         fault_reason, reason);
1795         else
1796                 pr_err("[%s] Request device [%02x:%02x.%d] PASID %x fault addr %llx [fault reason %02d] %s\n",
1797                        type ? "DMA Read" : "DMA Write",
1798                        source_id >> 8, PCI_SLOT(source_id & 0xFF),
1799                        PCI_FUNC(source_id & 0xFF), pasid, addr,
1800                        fault_reason, reason);
1801         return 0;
1802 }
1803
1804 #define PRIMARY_FAULT_REG_LEN (16)
1805 irqreturn_t dmar_fault(int irq, void *dev_id)
1806 {
1807         struct intel_iommu *iommu = dev_id;
1808         int reg, fault_index;
1809         u32 fault_status;
1810         unsigned long flag;
1811         static DEFINE_RATELIMIT_STATE(rs,
1812                                       DEFAULT_RATELIMIT_INTERVAL,
1813                                       DEFAULT_RATELIMIT_BURST);
1814
1815         raw_spin_lock_irqsave(&iommu->register_lock, flag);
1816         fault_status = readl(iommu->reg + DMAR_FSTS_REG);
1817         if (fault_status && __ratelimit(&rs))
1818                 pr_err("DRHD: handling fault status reg %x\n", fault_status);
1819
1820         /* TBD: ignore advanced fault log currently */
1821         if (!(fault_status & DMA_FSTS_PPF))
1822                 goto unlock_exit;
1823
1824         fault_index = dma_fsts_fault_record_index(fault_status);
1825         reg = cap_fault_reg_offset(iommu->cap);
1826         while (1) {
1827                 /* Disable printing, simply clear the fault when ratelimited */
1828                 bool ratelimited = !__ratelimit(&rs);
1829                 u8 fault_reason;
1830                 u16 source_id;
1831                 u64 guest_addr;
1832                 int type, pasid;
1833                 u32 data;
1834                 bool pasid_present;
1835
1836                 /* highest 32 bits */
1837                 data = readl(iommu->reg + reg +
1838                                 fault_index * PRIMARY_FAULT_REG_LEN + 12);
1839                 if (!(data & DMA_FRCD_F))
1840                         break;
1841
1842                 if (!ratelimited) {
1843                         fault_reason = dma_frcd_fault_reason(data);
1844                         type = dma_frcd_type(data);
1845
1846                         pasid = dma_frcd_pasid_value(data);
1847                         data = readl(iommu->reg + reg +
1848                                      fault_index * PRIMARY_FAULT_REG_LEN + 8);
1849                         source_id = dma_frcd_source_id(data);
1850
1851                         pasid_present = dma_frcd_pasid_present(data);
1852                         guest_addr = dmar_readq(iommu->reg + reg +
1853                                         fault_index * PRIMARY_FAULT_REG_LEN);
1854                         guest_addr = dma_frcd_page_addr(guest_addr);
1855                 }
1856
1857                 /* clear the fault */
1858                 writel(DMA_FRCD_F, iommu->reg + reg +
1859                         fault_index * PRIMARY_FAULT_REG_LEN + 12);
1860
1861                 raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1862
1863                 if (!ratelimited)
1864                         /* Using pasid -1 if pasid is not present */
1865                         dmar_fault_do_one(iommu, type, fault_reason,
1866                                           pasid_present ? pasid : -1,
1867                                           source_id, guest_addr);
1868
1869                 fault_index++;
1870                 if (fault_index >= cap_num_fault_regs(iommu->cap))
1871                         fault_index = 0;
1872                 raw_spin_lock_irqsave(&iommu->register_lock, flag);
1873         }
1874
1875         writel(DMA_FSTS_PFO | DMA_FSTS_PPF | DMA_FSTS_PRO,
1876                iommu->reg + DMAR_FSTS_REG);
1877
1878 unlock_exit:
1879         raw_spin_unlock_irqrestore(&iommu->register_lock, flag);
1880         return IRQ_HANDLED;
1881 }
1882
1883 int dmar_set_interrupt(struct intel_iommu *iommu)
1884 {
1885         int irq, ret;
1886
1887         /*
1888          * Check if the fault interrupt is already initialized.
1889          */
1890         if (iommu->irq)
1891                 return 0;
1892
1893         irq = dmar_alloc_hwirq(iommu->seq_id, iommu->node, iommu);
1894         if (irq > 0) {
1895                 iommu->irq = irq;
1896         } else {
1897                 pr_err("No free IRQ vectors\n");
1898                 return -EINVAL;
1899         }
1900
1901         ret = request_irq(irq, dmar_fault, IRQF_NO_THREAD, iommu->name, iommu);
1902         if (ret)
1903                 pr_err("Can't request irq\n");
1904         return ret;
1905 }
1906
1907 int __init enable_drhd_fault_handling(void)
1908 {
1909         struct dmar_drhd_unit *drhd;
1910         struct intel_iommu *iommu;
1911
1912         /*
1913          * Enable fault control interrupt.
1914          */
1915         for_each_iommu(iommu, drhd) {
1916                 u32 fault_status;
1917                 int ret = dmar_set_interrupt(iommu);
1918
1919                 if (ret) {
1920                         pr_err("DRHD %Lx: failed to enable fault, interrupt, ret %d\n",
1921                                (unsigned long long)drhd->reg_base_addr, ret);
1922                         return -1;
1923                 }
1924
1925                 /*
1926                  * Clear any previous faults.
1927                  */
1928                 dmar_fault(iommu->irq, iommu);
1929                 fault_status = readl(iommu->reg + DMAR_FSTS_REG);
1930                 writel(fault_status, iommu->reg + DMAR_FSTS_REG);
1931         }
1932
1933         return 0;
1934 }
1935
1936 /*
1937  * Re-enable Queued Invalidation interface.
1938  */
1939 int dmar_reenable_qi(struct intel_iommu *iommu)
1940 {
1941         if (!ecap_qis(iommu->ecap))
1942                 return -ENOENT;
1943
1944         if (!iommu->qi)
1945                 return -ENOENT;
1946
1947         /*
1948          * First disable queued invalidation.
1949          */
1950         dmar_disable_qi(iommu);
1951         /*
1952          * Then enable queued invalidation again. Since there is no pending
1953          * invalidation requests now, it's safe to re-enable queued
1954          * invalidation.
1955          */
1956         __dmar_enable_qi(iommu);
1957
1958         return 0;
1959 }
1960
1961 /*
1962  * Check interrupt remapping support in DMAR table description.
1963  */
1964 int __init dmar_ir_support(void)
1965 {
1966         struct acpi_table_dmar *dmar;
1967         dmar = (struct acpi_table_dmar *)dmar_tbl;
1968         if (!dmar)
1969                 return 0;
1970         return dmar->flags & 0x1;
1971 }
1972
1973 /* Check whether DMAR units are in use */
1974 static inline bool dmar_in_use(void)
1975 {
1976         return irq_remapping_enabled || intel_iommu_enabled;
1977 }
1978
1979 static int __init dmar_free_unused_resources(void)
1980 {
1981         struct dmar_drhd_unit *dmaru, *dmaru_n;
1982
1983         if (dmar_in_use())
1984                 return 0;
1985
1986         if (dmar_dev_scope_status != 1 && !list_empty(&dmar_drhd_units))
1987                 bus_unregister_notifier(&pci_bus_type, &dmar_pci_bus_nb);
1988
1989         down_write(&dmar_global_lock);
1990         list_for_each_entry_safe(dmaru, dmaru_n, &dmar_drhd_units, list) {
1991                 list_del(&dmaru->list);
1992                 dmar_free_drhd(dmaru);
1993         }
1994         up_write(&dmar_global_lock);
1995
1996         return 0;
1997 }
1998
1999 late_initcall(dmar_free_unused_resources);
2000 IOMMU_INIT_POST(detect_intel_iommu);
2001
2002 /*
2003  * DMAR Hotplug Support
2004  * For more details, please refer to Intel(R) Virtualization Technology
2005  * for Directed-IO Architecture Specifiction, Rev 2.2, Section 8.8
2006  * "Remapping Hardware Unit Hot Plug".
2007  */
2008 static guid_t dmar_hp_guid =
2009         GUID_INIT(0xD8C1A3A6, 0xBE9B, 0x4C9B,
2010                   0x91, 0xBF, 0xC3, 0xCB, 0x81, 0xFC, 0x5D, 0xAF);
2011
2012 /*
2013  * Currently there's only one revision and BIOS will not check the revision id,
2014  * so use 0 for safety.
2015  */
2016 #define DMAR_DSM_REV_ID                 0
2017 #define DMAR_DSM_FUNC_DRHD              1
2018 #define DMAR_DSM_FUNC_ATSR              2
2019 #define DMAR_DSM_FUNC_RHSA              3
2020
2021 static inline bool dmar_detect_dsm(acpi_handle handle, int func)
2022 {
2023         return acpi_check_dsm(handle, &dmar_hp_guid, DMAR_DSM_REV_ID, 1 << func);
2024 }
2025
2026 static int dmar_walk_dsm_resource(acpi_handle handle, int func,
2027                                   dmar_res_handler_t handler, void *arg)
2028 {
2029         int ret = -ENODEV;
2030         union acpi_object *obj;
2031         struct acpi_dmar_header *start;
2032         struct dmar_res_callback callback;
2033         static int res_type[] = {
2034                 [DMAR_DSM_FUNC_DRHD] = ACPI_DMAR_TYPE_HARDWARE_UNIT,
2035                 [DMAR_DSM_FUNC_ATSR] = ACPI_DMAR_TYPE_ROOT_ATS,
2036                 [DMAR_DSM_FUNC_RHSA] = ACPI_DMAR_TYPE_HARDWARE_AFFINITY,
2037         };
2038
2039         if (!dmar_detect_dsm(handle, func))
2040                 return 0;
2041
2042         obj = acpi_evaluate_dsm_typed(handle, &dmar_hp_guid, DMAR_DSM_REV_ID,
2043                                       func, NULL, ACPI_TYPE_BUFFER);
2044         if (!obj)
2045                 return -ENODEV;
2046
2047         memset(&callback, 0, sizeof(callback));
2048         callback.cb[res_type[func]] = handler;
2049         callback.arg[res_type[func]] = arg;
2050         start = (struct acpi_dmar_header *)obj->buffer.pointer;
2051         ret = dmar_walk_remapping_entries(start, obj->buffer.length, &callback);
2052
2053         ACPI_FREE(obj);
2054
2055         return ret;
2056 }
2057
2058 static int dmar_hp_add_drhd(struct acpi_dmar_header *header, void *arg)
2059 {
2060         int ret;
2061         struct dmar_drhd_unit *dmaru;
2062
2063         dmaru = dmar_find_dmaru((struct acpi_dmar_hardware_unit *)header);
2064         if (!dmaru)
2065                 return -ENODEV;
2066
2067         ret = dmar_ir_hotplug(dmaru, true);
2068         if (ret == 0)
2069                 ret = dmar_iommu_hotplug(dmaru, true);
2070
2071         return ret;
2072 }
2073
2074 static int dmar_hp_remove_drhd(struct acpi_dmar_header *header, void *arg)
2075 {
2076         int i, ret;
2077         struct device *dev;
2078         struct dmar_drhd_unit *dmaru;
2079
2080         dmaru = dmar_find_dmaru((struct acpi_dmar_hardware_unit *)header);
2081         if (!dmaru)
2082                 return 0;
2083
2084         /*
2085          * All PCI devices managed by this unit should have been destroyed.
2086          */
2087         if (!dmaru->include_all && dmaru->devices && dmaru->devices_cnt) {
2088                 for_each_active_dev_scope(dmaru->devices,
2089                                           dmaru->devices_cnt, i, dev)
2090                         return -EBUSY;
2091         }
2092
2093         ret = dmar_ir_hotplug(dmaru, false);
2094         if (ret == 0)
2095                 ret = dmar_iommu_hotplug(dmaru, false);
2096
2097         return ret;
2098 }
2099
2100 static int dmar_hp_release_drhd(struct acpi_dmar_header *header, void *arg)
2101 {
2102         struct dmar_drhd_unit *dmaru;
2103
2104         dmaru = dmar_find_dmaru((struct acpi_dmar_hardware_unit *)header);
2105         if (dmaru) {
2106                 list_del_rcu(&dmaru->list);
2107                 synchronize_rcu();
2108                 dmar_free_drhd(dmaru);
2109         }
2110
2111         return 0;
2112 }
2113
2114 static int dmar_hotplug_insert(acpi_handle handle)
2115 {
2116         int ret;
2117         int drhd_count = 0;
2118
2119         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2120                                      &dmar_validate_one_drhd, (void *)1);
2121         if (ret)
2122                 goto out;
2123
2124         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2125                                      &dmar_parse_one_drhd, (void *)&drhd_count);
2126         if (ret == 0 && drhd_count == 0) {
2127                 pr_warn(FW_BUG "No DRHD structures in buffer returned by _DSM method\n");
2128                 goto out;
2129         } else if (ret) {
2130                 goto release_drhd;
2131         }
2132
2133         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_RHSA,
2134                                      &dmar_parse_one_rhsa, NULL);
2135         if (ret)
2136                 goto release_drhd;
2137
2138         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_ATSR,
2139                                      &dmar_parse_one_atsr, NULL);
2140         if (ret)
2141                 goto release_atsr;
2142
2143         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2144                                      &dmar_hp_add_drhd, NULL);
2145         if (!ret)
2146                 return 0;
2147
2148         dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2149                                &dmar_hp_remove_drhd, NULL);
2150 release_atsr:
2151         dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_ATSR,
2152                                &dmar_release_one_atsr, NULL);
2153 release_drhd:
2154         dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2155                                &dmar_hp_release_drhd, NULL);
2156 out:
2157         return ret;
2158 }
2159
2160 static int dmar_hotplug_remove(acpi_handle handle)
2161 {
2162         int ret;
2163
2164         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_ATSR,
2165                                      &dmar_check_one_atsr, NULL);
2166         if (ret)
2167                 return ret;
2168
2169         ret = dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2170                                      &dmar_hp_remove_drhd, NULL);
2171         if (ret == 0) {
2172                 WARN_ON(dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_ATSR,
2173                                                &dmar_release_one_atsr, NULL));
2174                 WARN_ON(dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2175                                                &dmar_hp_release_drhd, NULL));
2176         } else {
2177                 dmar_walk_dsm_resource(handle, DMAR_DSM_FUNC_DRHD,
2178                                        &dmar_hp_add_drhd, NULL);
2179         }
2180
2181         return ret;
2182 }
2183
2184 static acpi_status dmar_get_dsm_handle(acpi_handle handle, u32 lvl,
2185                                        void *context, void **retval)
2186 {
2187         acpi_handle *phdl = retval;
2188
2189         if (dmar_detect_dsm(handle, DMAR_DSM_FUNC_DRHD)) {
2190                 *phdl = handle;
2191                 return AE_CTRL_TERMINATE;
2192         }
2193
2194         return AE_OK;
2195 }
2196
2197 static int dmar_device_hotplug(acpi_handle handle, bool insert)
2198 {
2199         int ret;
2200         acpi_handle tmp = NULL;
2201         acpi_status status;
2202
2203         if (!dmar_in_use())
2204                 return 0;
2205
2206         if (dmar_detect_dsm(handle, DMAR_DSM_FUNC_DRHD)) {
2207                 tmp = handle;
2208         } else {
2209                 status = acpi_walk_namespace(ACPI_TYPE_DEVICE, handle,
2210                                              ACPI_UINT32_MAX,
2211                                              dmar_get_dsm_handle,
2212                                              NULL, NULL, &tmp);
2213                 if (ACPI_FAILURE(status)) {
2214                         pr_warn("Failed to locate _DSM method.\n");
2215                         return -ENXIO;
2216                 }
2217         }
2218         if (tmp == NULL)
2219                 return 0;
2220
2221         down_write(&dmar_global_lock);
2222         if (insert)
2223                 ret = dmar_hotplug_insert(tmp);
2224         else
2225                 ret = dmar_hotplug_remove(tmp);
2226         up_write(&dmar_global_lock);
2227
2228         return ret;
2229 }
2230
2231 int dmar_device_add(acpi_handle handle)
2232 {
2233         return dmar_device_hotplug(handle, true);
2234 }
2235
2236 int dmar_device_remove(acpi_handle handle)
2237 {
2238         return dmar_device_hotplug(handle, false);
2239 }
2240
2241 /*
2242  * dmar_platform_optin - Is %DMA_CTRL_PLATFORM_OPT_IN_FLAG set in DMAR table
2243  *
2244  * Returns true if the platform has %DMA_CTRL_PLATFORM_OPT_IN_FLAG set in
2245  * the ACPI DMAR table. This means that the platform boot firmware has made
2246  * sure no device can issue DMA outside of RMRR regions.
2247  */
2248 bool dmar_platform_optin(void)
2249 {
2250         struct acpi_table_dmar *dmar;
2251         acpi_status status;
2252         bool ret;
2253
2254         status = acpi_get_table(ACPI_SIG_DMAR, 0,
2255                                 (struct acpi_table_header **)&dmar);
2256         if (ACPI_FAILURE(status))
2257                 return false;
2258
2259         ret = !!(dmar->flags & DMAR_PLATFORM_OPT_IN);
2260         acpi_put_table((struct acpi_table_header *)dmar);
2261
2262         return ret;
2263 }
2264 EXPORT_SYMBOL_GPL(dmar_platform_optin);