hw/virtio/virtio-iommu.c

   1 /*
   2  * virtio-iommu device
   3  *
   4  * Copyright (c) 2020 Red Hat, Inc.
   5  *
   6  * This program is free software; you can redistribute it and/or modify it
   7  * under the terms and conditions of the GNU General Public License,
   8  * version 2 or later, as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope it will be useful, but WITHOUT
  11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  13  * more details.
  14  *
  15  * You should have received a copy of the GNU General Public License along with
  16  * this program.  If not, see <http://www.gnu.org/licenses/>.
  17  *
  18  */
  19
  20 #include "qemu/osdep.h"
  21 #include "qemu/log.h"
  22 #include "qemu/iov.h"
  23 #include "hw/qdev-properties.h"
  24 #include "hw/virtio/virtio.h"
  25 #include "sysemu/kvm.h"
  26 #include "sysemu/reset.h"
  27 #include "qapi/error.h"
  28 #include "qemu/error-report.h"
  29 #include "trace.h"
  30
  31 #include "standard-headers/linux/virtio_ids.h"
  32
  33 #include "hw/virtio/virtio-bus.h"
  34 #include "hw/virtio/virtio-access.h"
  35 #include "hw/virtio/virtio-iommu.h"
  36 #include "hw/pci/pci_bus.h"
  37 #include "hw/pci/pci.h"
  38
  39 /* Max size */
  40 #define VIOMMU_DEFAULT_QUEUE_SIZE 256
  41 #define VIOMMU_PROBE_SIZE 512
  42
  43 typedef struct VirtIOIOMMUDomain {
  44     uint32_t id;
  45     bool bypass;
  46     GTree *mappings;
  47     QLIST_HEAD(, VirtIOIOMMUEndpoint) endpoint_list;
  48 } VirtIOIOMMUDomain;
  49
  50 typedef struct VirtIOIOMMUEndpoint {
  51     uint32_t id;
  52     VirtIOIOMMUDomain *domain;
  53     IOMMUMemoryRegion *iommu_mr;
  54     QLIST_ENTRY(VirtIOIOMMUEndpoint) next;
  55 } VirtIOIOMMUEndpoint;
  56
  57 typedef struct VirtIOIOMMUInterval {
  58     uint64_t low;
  59     uint64_t high;
  60 } VirtIOIOMMUInterval;
  61
  62 typedef struct VirtIOIOMMUMapping {
  63     uint64_t phys_addr;
  64     uint32_t flags;
  65 } VirtIOIOMMUMapping;
  66
  67 static inline uint16_t virtio_iommu_get_bdf(IOMMUDevice *dev)
  68 {
  69     return PCI_BUILD_BDF(pci_bus_num(dev->bus), dev->devfn);
  70 }
  71
  72 static bool virtio_iommu_device_bypassed(IOMMUDevice *sdev)
  73 {
  74     uint32_t sid;
  75     bool bypassed;
  76     VirtIOIOMMU *s = sdev->viommu;
  77     VirtIOIOMMUEndpoint *ep;
  78
  79     sid = virtio_iommu_get_bdf(sdev);
  80
  81     qemu_rec_mutex_lock(&s->mutex);
  82     /* need to check bypass before system reset */
  83     if (!s->endpoints) {
  84         bypassed = s->config.bypass;
  85         goto unlock;
  86     }
  87
  88     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
  89     if (!ep || !ep->domain) {
  90         bypassed = s->config.bypass;
  91     } else {
  92         bypassed = ep->domain->bypass;
  93     }
  94
  95 unlock:
  96     qemu_rec_mutex_unlock(&s->mutex);
  97     return bypassed;
  98 }
  99
 100 /* Return whether the device is using IOMMU translation. */
 101 static bool virtio_iommu_switch_address_space(IOMMUDevice *sdev)
 102 {
 103     bool use_remapping;
 104
 105     assert(sdev);
 106
 107     use_remapping = !virtio_iommu_device_bypassed(sdev);
 108
 109     trace_virtio_iommu_switch_address_space(pci_bus_num(sdev->bus),
 110                                             PCI_SLOT(sdev->devfn),
 111                                             PCI_FUNC(sdev->devfn),
 112                                             use_remapping);
 113
 114     /* Turn off first then on the other */
 115     if (use_remapping) {
 116         memory_region_set_enabled(&sdev->bypass_mr, false);
 117         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), true);
 118     } else {
 119         memory_region_set_enabled(MEMORY_REGION(&sdev->iommu_mr), false);
 120         memory_region_set_enabled(&sdev->bypass_mr, true);
 121     }
 122
 123     return use_remapping;
 124 }
 125
 126 static void virtio_iommu_switch_address_space_all(VirtIOIOMMU *s)
 127 {
 128     GHashTableIter iter;
 129     IOMMUPciBus *iommu_pci_bus;
 130     int i;
 131
 132     g_hash_table_iter_init(&iter, s->as_by_busptr);
 133     while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
 134         for (i = 0; i < PCI_DEVFN_MAX; i++) {
 135             if (!iommu_pci_bus->pbdev[i]) {
 136                 continue;
 137             }
 138             virtio_iommu_switch_address_space(iommu_pci_bus->pbdev[i]);
 139         }
 140     }
 141 }
 142
 143 /**
 144  * The bus number is used for lookup when SID based operations occur.
 145  * In that case we lazily populate the IOMMUPciBus array from the bus hash
 146  * table. At the time the IOMMUPciBus is created (iommu_find_add_as), the bus
 147  * numbers may not be always initialized yet.
 148  */
 149 static IOMMUPciBus *iommu_find_iommu_pcibus(VirtIOIOMMU *s, uint8_t bus_num)
 150 {
 151     IOMMUPciBus *iommu_pci_bus = s->iommu_pcibus_by_bus_num[bus_num];
 152
 153     if (!iommu_pci_bus) {
 154         GHashTableIter iter;
 155
 156         g_hash_table_iter_init(&iter, s->as_by_busptr);
 157         while (g_hash_table_iter_next(&iter, NULL, (void **)&iommu_pci_bus)) {
 158             if (pci_bus_num(iommu_pci_bus->bus) == bus_num) {
 159                 s->iommu_pcibus_by_bus_num[bus_num] = iommu_pci_bus;
 160                 return iommu_pci_bus;
 161             }
 162         }
 163         return NULL;
 164     }
 165     return iommu_pci_bus;
 166 }
 167
 168 static IOMMUMemoryRegion *virtio_iommu_mr(VirtIOIOMMU *s, uint32_t sid)
 169 {
 170     uint8_t bus_n, devfn;
 171     IOMMUPciBus *iommu_pci_bus;
 172     IOMMUDevice *dev;
 173
 174     bus_n = PCI_BUS_NUM(sid);
 175     iommu_pci_bus = iommu_find_iommu_pcibus(s, bus_n);
 176     if (iommu_pci_bus) {
 177         devfn = sid & (PCI_DEVFN_MAX - 1);
 178         dev = iommu_pci_bus->pbdev[devfn];
 179         if (dev) {
 180             return &dev->iommu_mr;
 181         }
 182     }
 183     return NULL;
 184 }
 185
 186 static gint interval_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
 187 {
 188     VirtIOIOMMUInterval *inta = (VirtIOIOMMUInterval *)a;
 189     VirtIOIOMMUInterval *intb = (VirtIOIOMMUInterval *)b;
 190
 191     if (inta->high < intb->low) {
 192         return -1;
 193     } else if (intb->high < inta->low) {
 194         return 1;
 195     } else {
 196         return 0;
 197     }
 198 }
 199
 200 static void virtio_iommu_notify_map_unmap(IOMMUMemoryRegion *mr,
 201                                           IOMMUTLBEvent *event,
 202                                           hwaddr virt_start, hwaddr virt_end)
 203 {
 204     uint64_t delta = virt_end - virt_start;
 205
 206     event->entry.iova = virt_start;
 207     event->entry.addr_mask = delta;
 208
 209     if (delta == UINT64_MAX) {
 210         memory_region_notify_iommu(mr, 0, *event);
 211     }
 212
 213     while (virt_start != virt_end + 1) {
 214         uint64_t mask = dma_aligned_pow2_mask(virt_start, virt_end, 64);
 215
 216         event->entry.addr_mask = mask;
 217         event->entry.iova = virt_start;
 218         memory_region_notify_iommu(mr, 0, *event);
 219         virt_start += mask + 1;
 220         if (event->entry.perm != IOMMU_NONE) {
 221             event->entry.translated_addr += mask + 1;
 222         }
 223     }
 224 }
 225
 226 static void virtio_iommu_notify_map(IOMMUMemoryRegion *mr, hwaddr virt_start,
 227                                     hwaddr virt_end, hwaddr paddr,
 228                                     uint32_t flags)
 229 {
 230     IOMMUTLBEvent event;
 231     IOMMUAccessFlags perm = IOMMU_ACCESS_FLAG(flags & VIRTIO_IOMMU_MAP_F_READ,
 232                                               flags & VIRTIO_IOMMU_MAP_F_WRITE);
 233
 234     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_MAP) ||
 235         (flags & VIRTIO_IOMMU_MAP_F_MMIO) || !perm) {
 236         return;
 237     }
 238
 239     trace_virtio_iommu_notify_map(mr->parent_obj.name, virt_start, virt_end,
 240                                   paddr, perm);
 241
 242     event.type = IOMMU_NOTIFIER_MAP;
 243     event.entry.target_as = &address_space_memory;
 244     event.entry.perm = perm;
 245     event.entry.translated_addr = paddr;
 246
 247     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
 248 }
 249
 250 static void virtio_iommu_notify_unmap(IOMMUMemoryRegion *mr, hwaddr virt_start,
 251                                       hwaddr virt_end)
 252 {
 253     IOMMUTLBEvent event;
 254
 255     if (!(mr->iommu_notify_flags & IOMMU_NOTIFIER_UNMAP)) {
 256         return;
 257     }
 258
 259     trace_virtio_iommu_notify_unmap(mr->parent_obj.name, virt_start, virt_end);
 260
 261     event.type = IOMMU_NOTIFIER_UNMAP;
 262     event.entry.target_as = &address_space_memory;
 263     event.entry.perm = IOMMU_NONE;
 264     event.entry.translated_addr = 0;
 265
 266     virtio_iommu_notify_map_unmap(mr, &event, virt_start, virt_end);
 267 }
 268
 269 static gboolean virtio_iommu_notify_unmap_cb(gpointer key, gpointer value,
 270                                              gpointer data)
 271 {
 272     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
 273     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
 274
 275     virtio_iommu_notify_unmap(mr, interval->low, interval->high);
 276
 277     return false;
 278 }
 279
 280 static gboolean virtio_iommu_notify_map_cb(gpointer key, gpointer value,
 281                                            gpointer data)
 282 {
 283     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
 284     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
 285     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
 286
 287     virtio_iommu_notify_map(mr, interval->low, interval->high,
 288                             mapping->phys_addr, mapping->flags);
 289
 290     return false;
 291 }
 292
 293 static void virtio_iommu_detach_endpoint_from_domain(VirtIOIOMMUEndpoint *ep)
 294 {
 295     VirtIOIOMMUDomain *domain = ep->domain;
 296     IOMMUDevice *sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
 297
 298     if (!ep->domain) {
 299         return;
 300     }
 301     g_tree_foreach(domain->mappings, virtio_iommu_notify_unmap_cb,
 302                    ep->iommu_mr);
 303     QLIST_REMOVE(ep, next);
 304     ep->domain = NULL;
 305     virtio_iommu_switch_address_space(sdev);
 306 }
 307
 308 static VirtIOIOMMUEndpoint *virtio_iommu_get_endpoint(VirtIOIOMMU *s,
 309                                                       uint32_t ep_id)
 310 {
 311     VirtIOIOMMUEndpoint *ep;
 312     IOMMUMemoryRegion *mr;
 313
 314     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
 315     if (ep) {
 316         return ep;
 317     }
 318     mr = virtio_iommu_mr(s, ep_id);
 319     if (!mr) {
 320         return NULL;
 321     }
 322     ep = g_malloc0(sizeof(*ep));
 323     ep->id = ep_id;
 324     ep->iommu_mr = mr;
 325     trace_virtio_iommu_get_endpoint(ep_id);
 326     g_tree_insert(s->endpoints, GUINT_TO_POINTER(ep_id), ep);
 327     return ep;
 328 }
 329
 330 static void virtio_iommu_put_endpoint(gpointer data)
 331 {
 332     VirtIOIOMMUEndpoint *ep = (VirtIOIOMMUEndpoint *)data;
 333
 334     if (ep->domain) {
 335         virtio_iommu_detach_endpoint_from_domain(ep);
 336     }
 337
 338     trace_virtio_iommu_put_endpoint(ep->id);
 339     g_free(ep);
 340 }
 341
 342 static VirtIOIOMMUDomain *virtio_iommu_get_domain(VirtIOIOMMU *s,
 343                                                   uint32_t domain_id,
 344                                                   bool bypass)
 345 {
 346     VirtIOIOMMUDomain *domain;
 347
 348     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
 349     if (domain) {
 350         if (domain->bypass != bypass) {
 351             return NULL;
 352         }
 353         return domain;
 354     }
 355     domain = g_malloc0(sizeof(*domain));
 356     domain->id = domain_id;
 357     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
 358                                    NULL, (GDestroyNotify)g_free,
 359                                    (GDestroyNotify)g_free);
 360     domain->bypass = bypass;
 361     g_tree_insert(s->domains, GUINT_TO_POINTER(domain_id), domain);
 362     QLIST_INIT(&domain->endpoint_list);
 363     trace_virtio_iommu_get_domain(domain_id);
 364     return domain;
 365 }
 366
 367 static void virtio_iommu_put_domain(gpointer data)
 368 {
 369     VirtIOIOMMUDomain *domain = (VirtIOIOMMUDomain *)data;
 370     VirtIOIOMMUEndpoint *iter, *tmp;
 371
 372     QLIST_FOREACH_SAFE(iter, &domain->endpoint_list, next, tmp) {
 373         virtio_iommu_detach_endpoint_from_domain(iter);
 374     }
 375     g_tree_destroy(domain->mappings);
 376     trace_virtio_iommu_put_domain(domain->id);
 377     g_free(domain);
 378 }
 379
 380 static AddressSpace *virtio_iommu_find_add_as(PCIBus *bus, void *opaque,
 381                                               int devfn)
 382 {
 383     VirtIOIOMMU *s = opaque;
 384     IOMMUPciBus *sbus = g_hash_table_lookup(s->as_by_busptr, bus);
 385     static uint32_t mr_index;
 386     IOMMUDevice *sdev;
 387
 388     if (!sbus) {
 389         sbus = g_malloc0(sizeof(IOMMUPciBus) +
 390                          sizeof(IOMMUDevice *) * PCI_DEVFN_MAX);
 391         sbus->bus = bus;
 392         g_hash_table_insert(s->as_by_busptr, bus, sbus);
 393     }
 394
 395     sdev = sbus->pbdev[devfn];
 396     if (!sdev) {
 397         char *name = g_strdup_printf("%s-%d-%d",
 398                                      TYPE_VIRTIO_IOMMU_MEMORY_REGION,
 399                                      mr_index++, devfn);
 400         sdev = sbus->pbdev[devfn] = g_new0(IOMMUDevice, 1);
 401
 402         sdev->viommu = s;
 403         sdev->bus = bus;
 404         sdev->devfn = devfn;
 405
 406         trace_virtio_iommu_init_iommu_mr(name);
 407
 408         memory_region_init(&sdev->root, OBJECT(s), name, UINT64_MAX);
 409         address_space_init(&sdev->as, &sdev->root, TYPE_VIRTIO_IOMMU);
 410
 411         /*
 412          * Build the IOMMU disabled container with aliases to the
 413          * shared MRs.  Note that aliasing to a shared memory region
 414          * could help the memory API to detect same FlatViews so we
 415          * can have devices to share the same FlatView when in bypass
 416          * mode. (either by not configuring virtio-iommu driver or with
 417          * "iommu=pt").  It will greatly reduce the total number of
 418          * FlatViews of the system hence VM runs faster.
 419          */
 420         memory_region_init_alias(&sdev->bypass_mr, OBJECT(s),
 421                                  "system", get_system_memory(), 0,
 422                                  memory_region_size(get_system_memory()));
 423
 424         memory_region_init_iommu(&sdev->iommu_mr, sizeof(sdev->iommu_mr),
 425                                  TYPE_VIRTIO_IOMMU_MEMORY_REGION,
 426                                  OBJECT(s), name,
 427                                  UINT64_MAX);
 428
 429         /*
 430          * Hook both the containers under the root container, we
 431          * switch between iommu & bypass MRs by enable/disable
 432          * corresponding sub-containers
 433          */
 434         memory_region_add_subregion_overlap(&sdev->root, 0,
 435                                             MEMORY_REGION(&sdev->iommu_mr),
 436                                             0);
 437         memory_region_add_subregion_overlap(&sdev->root, 0,
 438                                             &sdev->bypass_mr, 0);
 439
 440         virtio_iommu_switch_address_space(sdev);
 441         g_free(name);
 442     }
 443     return &sdev->as;
 444 }
 445
 446 static int virtio_iommu_attach(VirtIOIOMMU *s,
 447                                struct virtio_iommu_req_attach *req)
 448 {
 449     uint32_t domain_id = le32_to_cpu(req->domain);
 450     uint32_t ep_id = le32_to_cpu(req->endpoint);
 451     uint32_t flags = le32_to_cpu(req->flags);
 452     VirtIOIOMMUDomain *domain;
 453     VirtIOIOMMUEndpoint *ep;
 454     IOMMUDevice *sdev;
 455
 456     trace_virtio_iommu_attach(domain_id, ep_id);
 457
 458     if (flags & ~VIRTIO_IOMMU_ATTACH_F_BYPASS) {
 459         return VIRTIO_IOMMU_S_INVAL;
 460     }
 461
 462     ep = virtio_iommu_get_endpoint(s, ep_id);
 463     if (!ep) {
 464         return VIRTIO_IOMMU_S_NOENT;
 465     }
 466
 467     if (ep->domain) {
 468         VirtIOIOMMUDomain *previous_domain = ep->domain;
 469         /*
 470          * the device is already attached to a domain,
 471          * detach it first
 472          */
 473         virtio_iommu_detach_endpoint_from_domain(ep);
 474         if (QLIST_EMPTY(&previous_domain->endpoint_list)) {
 475             g_tree_remove(s->domains, GUINT_TO_POINTER(previous_domain->id));
 476         }
 477     }
 478
 479     domain = virtio_iommu_get_domain(s, domain_id,
 480                                      flags & VIRTIO_IOMMU_ATTACH_F_BYPASS);
 481     if (!domain) {
 482         /* Incompatible bypass flag */
 483         return VIRTIO_IOMMU_S_INVAL;
 484     }
 485     QLIST_INSERT_HEAD(&domain->endpoint_list, ep, next);
 486
 487     ep->domain = domain;
 488     sdev = container_of(ep->iommu_mr, IOMMUDevice, iommu_mr);
 489     virtio_iommu_switch_address_space(sdev);
 490
 491     /* Replay domain mappings on the associated memory region */
 492     g_tree_foreach(domain->mappings, virtio_iommu_notify_map_cb,
 493                    ep->iommu_mr);
 494
 495     return VIRTIO_IOMMU_S_OK;
 496 }
 497
 498 static int virtio_iommu_detach(VirtIOIOMMU *s,
 499                                struct virtio_iommu_req_detach *req)
 500 {
 501     uint32_t domain_id = le32_to_cpu(req->domain);
 502     uint32_t ep_id = le32_to_cpu(req->endpoint);
 503     VirtIOIOMMUDomain *domain;
 504     VirtIOIOMMUEndpoint *ep;
 505
 506     trace_virtio_iommu_detach(domain_id, ep_id);
 507
 508     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(ep_id));
 509     if (!ep) {
 510         return VIRTIO_IOMMU_S_NOENT;
 511     }
 512
 513     domain = ep->domain;
 514
 515     if (!domain || domain->id != domain_id) {
 516         return VIRTIO_IOMMU_S_INVAL;
 517     }
 518
 519     virtio_iommu_detach_endpoint_from_domain(ep);
 520
 521     if (QLIST_EMPTY(&domain->endpoint_list)) {
 522         g_tree_remove(s->domains, GUINT_TO_POINTER(domain->id));
 523     }
 524     return VIRTIO_IOMMU_S_OK;
 525 }
 526
 527 static int virtio_iommu_map(VirtIOIOMMU *s,
 528                             struct virtio_iommu_req_map *req)
 529 {
 530     uint32_t domain_id = le32_to_cpu(req->domain);
 531     uint64_t phys_start = le64_to_cpu(req->phys_start);
 532     uint64_t virt_start = le64_to_cpu(req->virt_start);
 533     uint64_t virt_end = le64_to_cpu(req->virt_end);
 534     uint32_t flags = le32_to_cpu(req->flags);
 535     VirtIOIOMMUDomain *domain;
 536     VirtIOIOMMUInterval *interval;
 537     VirtIOIOMMUMapping *mapping;
 538     VirtIOIOMMUEndpoint *ep;
 539
 540     if (flags & ~VIRTIO_IOMMU_MAP_F_MASK) {
 541         return VIRTIO_IOMMU_S_INVAL;
 542     }
 543
 544     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
 545     if (!domain) {
 546         return VIRTIO_IOMMU_S_NOENT;
 547     }
 548
 549     if (domain->bypass) {
 550         return VIRTIO_IOMMU_S_INVAL;
 551     }
 552
 553     interval = g_malloc0(sizeof(*interval));
 554
 555     interval->low = virt_start;
 556     interval->high = virt_end;
 557
 558     mapping = g_tree_lookup(domain->mappings, (gpointer)interval);
 559     if (mapping) {
 560         g_free(interval);
 561         return VIRTIO_IOMMU_S_INVAL;
 562     }
 563
 564     trace_virtio_iommu_map(domain_id, virt_start, virt_end, phys_start, flags);
 565
 566     mapping = g_malloc0(sizeof(*mapping));
 567     mapping->phys_addr = phys_start;
 568     mapping->flags = flags;
 569
 570     g_tree_insert(domain->mappings, interval, mapping);
 571
 572     QLIST_FOREACH(ep, &domain->endpoint_list, next) {
 573         virtio_iommu_notify_map(ep->iommu_mr, virt_start, virt_end, phys_start,
 574                                 flags);
 575     }
 576
 577     return VIRTIO_IOMMU_S_OK;
 578 }
 579
 580 static int virtio_iommu_unmap(VirtIOIOMMU *s,
 581                               struct virtio_iommu_req_unmap *req)
 582 {
 583     uint32_t domain_id = le32_to_cpu(req->domain);
 584     uint64_t virt_start = le64_to_cpu(req->virt_start);
 585     uint64_t virt_end = le64_to_cpu(req->virt_end);
 586     VirtIOIOMMUMapping *iter_val;
 587     VirtIOIOMMUInterval interval, *iter_key;
 588     VirtIOIOMMUDomain *domain;
 589     VirtIOIOMMUEndpoint *ep;
 590     int ret = VIRTIO_IOMMU_S_OK;
 591
 592     trace_virtio_iommu_unmap(domain_id, virt_start, virt_end);
 593
 594     domain = g_tree_lookup(s->domains, GUINT_TO_POINTER(domain_id));
 595     if (!domain) {
 596         return VIRTIO_IOMMU_S_NOENT;
 597     }
 598
 599     if (domain->bypass) {
 600         return VIRTIO_IOMMU_S_INVAL;
 601     }
 602
 603     interval.low = virt_start;
 604     interval.high = virt_end;
 605
 606     while (g_tree_lookup_extended(domain->mappings, &interval,
 607                                   (void **)&iter_key, (void**)&iter_val)) {
 608         uint64_t current_low = iter_key->low;
 609         uint64_t current_high = iter_key->high;
 610
 611         if (interval.low <= current_low && interval.high >= current_high) {
 612             QLIST_FOREACH(ep, &domain->endpoint_list, next) {
 613                 virtio_iommu_notify_unmap(ep->iommu_mr, current_low,
 614                                           current_high);
 615             }
 616             g_tree_remove(domain->mappings, iter_key);
 617             trace_virtio_iommu_unmap_done(domain_id, current_low, current_high);
 618         } else {
 619             ret = VIRTIO_IOMMU_S_RANGE;
 620             break;
 621         }
 622     }
 623     return ret;
 624 }
 625
 626 static ssize_t virtio_iommu_fill_resv_mem_prop(VirtIOIOMMU *s, uint32_t ep,
 627                                                uint8_t *buf, size_t free)
 628 {
 629     struct virtio_iommu_probe_resv_mem prop = {};
 630     size_t size = sizeof(prop), length = size - sizeof(prop.head), total;
 631     int i;
 632
 633     total = size * s->nb_reserved_regions;
 634
 635     if (total > free) {
 636         return -ENOSPC;
 637     }
 638
 639     for (i = 0; i < s->nb_reserved_regions; i++) {
 640         unsigned subtype = s->reserved_regions[i].type;
 641
 642         assert(subtype == VIRTIO_IOMMU_RESV_MEM_T_RESERVED ||
 643                subtype == VIRTIO_IOMMU_RESV_MEM_T_MSI);
 644         prop.head.type = cpu_to_le16(VIRTIO_IOMMU_PROBE_T_RESV_MEM);
 645         prop.head.length = cpu_to_le16(length);
 646         prop.subtype = subtype;
 647         prop.start = cpu_to_le64(s->reserved_regions[i].low);
 648         prop.end = cpu_to_le64(s->reserved_regions[i].high);
 649
 650         memcpy(buf, &prop, size);
 651
 652         trace_virtio_iommu_fill_resv_property(ep, prop.subtype,
 653                                               prop.start, prop.end);
 654         buf += size;
 655     }
 656     return total;
 657 }
 658
 659 /**
 660  * virtio_iommu_probe - Fill the probe request buffer with
 661  * the properties the device is able to return
 662  */
 663 static int virtio_iommu_probe(VirtIOIOMMU *s,
 664                               struct virtio_iommu_req_probe *req,
 665                               uint8_t *buf)
 666 {
 667     uint32_t ep_id = le32_to_cpu(req->endpoint);
 668     size_t free = VIOMMU_PROBE_SIZE;
 669     ssize_t count;
 670
 671     if (!virtio_iommu_mr(s, ep_id)) {
 672         return VIRTIO_IOMMU_S_NOENT;
 673     }
 674
 675     count = virtio_iommu_fill_resv_mem_prop(s, ep_id, buf, free);
 676     if (count < 0) {
 677         return VIRTIO_IOMMU_S_INVAL;
 678     }
 679     buf += count;
 680     free -= count;
 681
 682     return VIRTIO_IOMMU_S_OK;
 683 }
 684
 685 static int virtio_iommu_iov_to_req(struct iovec *iov,
 686                                    unsigned int iov_cnt,
 687                                    void *req, size_t payload_sz)
 688 {
 689     size_t sz = iov_to_buf(iov, iov_cnt, 0, req, payload_sz);
 690
 691     if (unlikely(sz != payload_sz)) {
 692         return VIRTIO_IOMMU_S_INVAL;
 693     }
 694     return 0;
 695 }
 696
 697 #define virtio_iommu_handle_req(__req)                                  \
 698 static int virtio_iommu_handle_ ## __req(VirtIOIOMMU *s,                \
 699                                          struct iovec *iov,             \
 700                                          unsigned int iov_cnt)          \
 701 {                                                                       \
 702     struct virtio_iommu_req_ ## __req req;                              \
 703     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req,               \
 704                     sizeof(req) - sizeof(struct virtio_iommu_req_tail));\
 705                                                                         \
 706     return ret ? ret : virtio_iommu_ ## __req(s, &req);                 \
 707 }
 708
 709 virtio_iommu_handle_req(attach)
 710 virtio_iommu_handle_req(detach)
 711 virtio_iommu_handle_req(map)
 712 virtio_iommu_handle_req(unmap)
 713
 714 static int virtio_iommu_handle_probe(VirtIOIOMMU *s,
 715                                      struct iovec *iov,
 716                                      unsigned int iov_cnt,
 717                                      uint8_t *buf)
 718 {
 719     struct virtio_iommu_req_probe req;
 720     int ret = virtio_iommu_iov_to_req(iov, iov_cnt, &req, sizeof(req));
 721
 722     return ret ? ret : virtio_iommu_probe(s, &req, buf);
 723 }
 724
 725 static void virtio_iommu_handle_command(VirtIODevice *vdev, VirtQueue *vq)
 726 {
 727     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
 728     struct virtio_iommu_req_head head;
 729     struct virtio_iommu_req_tail tail = {};
 730     size_t output_size = sizeof(tail), sz;
 731     VirtQueueElement *elem;
 732     unsigned int iov_cnt;
 733     struct iovec *iov;
 734     void *buf = NULL;
 735
 736     for (;;) {
 737         elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
 738         if (!elem) {
 739             return;
 740         }
 741
 742         if (iov_size(elem->in_sg, elem->in_num) < sizeof(tail) ||
 743             iov_size(elem->out_sg, elem->out_num) < sizeof(head)) {
 744             virtio_error(vdev, "virtio-iommu bad head/tail size");
 745             virtqueue_detach_element(vq, elem, 0);
 746             g_free(elem);
 747             break;
 748         }
 749
 750         iov_cnt = elem->out_num;
 751         iov = elem->out_sg;
 752         sz = iov_to_buf(iov, iov_cnt, 0, &head, sizeof(head));
 753         if (unlikely(sz != sizeof(head))) {
 754             tail.status = VIRTIO_IOMMU_S_DEVERR;
 755             goto out;
 756         }
 757         qemu_rec_mutex_lock(&s->mutex);
 758         switch (head.type) {
 759         case VIRTIO_IOMMU_T_ATTACH:
 760             tail.status = virtio_iommu_handle_attach(s, iov, iov_cnt);
 761             break;
 762         case VIRTIO_IOMMU_T_DETACH:
 763             tail.status = virtio_iommu_handle_detach(s, iov, iov_cnt);
 764             break;
 765         case VIRTIO_IOMMU_T_MAP:
 766             tail.status = virtio_iommu_handle_map(s, iov, iov_cnt);
 767             break;
 768         case VIRTIO_IOMMU_T_UNMAP:
 769             tail.status = virtio_iommu_handle_unmap(s, iov, iov_cnt);
 770             break;
 771         case VIRTIO_IOMMU_T_PROBE:
 772         {
 773             struct virtio_iommu_req_tail *ptail;
 774
 775             output_size = s->config.probe_size + sizeof(tail);
 776             buf = g_malloc0(output_size);
 777
 778             ptail = (struct virtio_iommu_req_tail *)
 779                         (buf + s->config.probe_size);
 780             ptail->status = virtio_iommu_handle_probe(s, iov, iov_cnt, buf);
 781             break;
 782         }
 783         default:
 784             tail.status = VIRTIO_IOMMU_S_UNSUPP;
 785         }
 786         qemu_rec_mutex_unlock(&s->mutex);
 787
 788 out:
 789         sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
 790                           buf ? buf : &tail, output_size);
 791         assert(sz == output_size);
 792
 793         virtqueue_push(vq, elem, sz);
 794         virtio_notify(vdev, vq);
 795         g_free(elem);
 796         g_free(buf);
 797         buf = NULL;
 798     }
 799 }
 800
 801 static void virtio_iommu_report_fault(VirtIOIOMMU *viommu, uint8_t reason,
 802                                       int flags, uint32_t endpoint,
 803                                       uint64_t address)
 804 {
 805     VirtIODevice *vdev = &viommu->parent_obj;
 806     VirtQueue *vq = viommu->event_vq;
 807     struct virtio_iommu_fault fault;
 808     VirtQueueElement *elem;
 809     size_t sz;
 810
 811     memset(&fault, 0, sizeof(fault));
 812     fault.reason = reason;
 813     fault.flags = cpu_to_le32(flags);
 814     fault.endpoint = cpu_to_le32(endpoint);
 815     fault.address = cpu_to_le64(address);
 816
 817     elem = virtqueue_pop(vq, sizeof(VirtQueueElement));
 818
 819     if (!elem) {
 820         error_report_once(
 821             "no buffer available in event queue to report event");
 822         return;
 823     }
 824
 825     if (iov_size(elem->in_sg, elem->in_num) < sizeof(fault)) {
 826         virtio_error(vdev, "error buffer of wrong size");
 827         virtqueue_detach_element(vq, elem, 0);
 828         g_free(elem);
 829         return;
 830     }
 831
 832     sz = iov_from_buf(elem->in_sg, elem->in_num, 0,
 833                       &fault, sizeof(fault));
 834     assert(sz == sizeof(fault));
 835
 836     trace_virtio_iommu_report_fault(reason, flags, endpoint, address);
 837     virtqueue_push(vq, elem, sz);
 838     virtio_notify(vdev, vq);
 839     g_free(elem);
 840
 841 }
 842
 843 static IOMMUTLBEntry virtio_iommu_translate(IOMMUMemoryRegion *mr, hwaddr addr,
 844                                             IOMMUAccessFlags flag,
 845                                             int iommu_idx)
 846 {
 847     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
 848     VirtIOIOMMUInterval interval, *mapping_key;
 849     VirtIOIOMMUMapping *mapping_value;
 850     VirtIOIOMMU *s = sdev->viommu;
 851     bool read_fault, write_fault;
 852     VirtIOIOMMUEndpoint *ep;
 853     uint32_t sid, flags;
 854     bool bypass_allowed;
 855     bool found;
 856     int i;
 857
 858     interval.low = addr;
 859     interval.high = addr + 1;
 860
 861     IOMMUTLBEntry entry = {
 862         .target_as = &address_space_memory,
 863         .iova = addr,
 864         .translated_addr = addr,
 865         .addr_mask = (1 << ctz32(s->config.page_size_mask)) - 1,
 866         .perm = IOMMU_NONE,
 867     };
 868
 869     bypass_allowed = s->config.bypass;
 870
 871     sid = virtio_iommu_get_bdf(sdev);
 872
 873     trace_virtio_iommu_translate(mr->parent_obj.name, sid, addr, flag);
 874     qemu_rec_mutex_lock(&s->mutex);
 875
 876     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
 877
 878     if (bypass_allowed)
 879         assert(ep && ep->domain && !ep->domain->bypass);
 880
 881     if (!ep) {
 882         if (!bypass_allowed) {
 883             error_report_once("%s sid=%d is not known!!", __func__, sid);
 884             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_UNKNOWN,
 885                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
 886                                       sid, addr);
 887         } else {
 888             entry.perm = flag;
 889         }
 890         goto unlock;
 891     }
 892
 893     for (i = 0; i < s->nb_reserved_regions; i++) {
 894         ReservedRegion *reg = &s->reserved_regions[i];
 895
 896         if (addr >= reg->low && addr <= reg->high) {
 897             switch (reg->type) {
 898             case VIRTIO_IOMMU_RESV_MEM_T_MSI:
 899                 entry.perm = flag;
 900                 break;
 901             case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
 902             default:
 903                 virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
 904                                           VIRTIO_IOMMU_FAULT_F_ADDRESS,
 905                                           sid, addr);
 906                 break;
 907             }
 908             goto unlock;
 909         }
 910     }
 911
 912     if (!ep->domain) {
 913         if (!bypass_allowed) {
 914             error_report_once("%s %02x:%02x.%01x not attached to any domain",
 915                               __func__, PCI_BUS_NUM(sid),
 916                               PCI_SLOT(sid), PCI_FUNC(sid));
 917             virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_DOMAIN,
 918                                       VIRTIO_IOMMU_FAULT_F_ADDRESS,
 919                                       sid, addr);
 920         } else {
 921             entry.perm = flag;
 922         }
 923         goto unlock;
 924     } else if (ep->domain->bypass) {
 925         entry.perm = flag;
 926         goto unlock;
 927     }
 928
 929     found = g_tree_lookup_extended(ep->domain->mappings, (gpointer)(&interval),
 930                                    (void **)&mapping_key,
 931                                    (void **)&mapping_value);
 932     if (!found) {
 933         error_report_once("%s no mapping for 0x%"PRIx64" for sid=%d",
 934                           __func__, addr, sid);
 935         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
 936                                   VIRTIO_IOMMU_FAULT_F_ADDRESS,
 937                                   sid, addr);
 938         goto unlock;
 939     }
 940
 941     read_fault = (flag & IOMMU_RO) &&
 942                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_READ);
 943     write_fault = (flag & IOMMU_WO) &&
 944                     !(mapping_value->flags & VIRTIO_IOMMU_MAP_F_WRITE);
 945
 946     flags = read_fault ? VIRTIO_IOMMU_FAULT_F_READ : 0;
 947     flags |= write_fault ? VIRTIO_IOMMU_FAULT_F_WRITE : 0;
 948     if (flags) {
 949         error_report_once("%s permission error on 0x%"PRIx64"(%d): allowed=%d",
 950                           __func__, addr, flag, mapping_value->flags);
 951         flags |= VIRTIO_IOMMU_FAULT_F_ADDRESS;
 952         virtio_iommu_report_fault(s, VIRTIO_IOMMU_FAULT_R_MAPPING,
 953                                   flags | VIRTIO_IOMMU_FAULT_F_ADDRESS,
 954                                   sid, addr);
 955         goto unlock;
 956     }
 957     entry.translated_addr = addr - mapping_key->low + mapping_value->phys_addr;
 958     entry.perm = flag;
 959     trace_virtio_iommu_translate_out(addr, entry.translated_addr, sid);
 960
 961 unlock:
 962     qemu_rec_mutex_unlock(&s->mutex);
 963     return entry;
 964 }
 965
 966 static void virtio_iommu_get_config(VirtIODevice *vdev, uint8_t *config_data)
 967 {
 968     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
 969     struct virtio_iommu_config *dev_config = &dev->config;
 970     struct virtio_iommu_config *out_config = (void *)config_data;
 971
 972     out_config->page_size_mask = cpu_to_le64(dev_config->page_size_mask);
 973     out_config->input_range.start = cpu_to_le64(dev_config->input_range.start);
 974     out_config->input_range.end = cpu_to_le64(dev_config->input_range.end);
 975     out_config->domain_range.start = cpu_to_le32(dev_config->domain_range.start);
 976     out_config->domain_range.end = cpu_to_le32(dev_config->domain_range.end);
 977     out_config->probe_size = cpu_to_le32(dev_config->probe_size);
 978     out_config->bypass = dev_config->bypass;
 979
 980     trace_virtio_iommu_get_config(dev_config->page_size_mask,
 981                                   dev_config->input_range.start,
 982                                   dev_config->input_range.end,
 983                                   dev_config->domain_range.start,
 984                                   dev_config->domain_range.end,
 985                                   dev_config->probe_size,
 986                                   dev_config->bypass);
 987 }
 988
 989 static void virtio_iommu_set_config(VirtIODevice *vdev,
 990                                     const uint8_t *config_data)
 991 {
 992     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
 993     struct virtio_iommu_config *dev_config = &dev->config;
 994     const struct virtio_iommu_config *in_config = (void *)config_data;
 995
 996     if (in_config->bypass != dev_config->bypass) {
 997         if (!virtio_vdev_has_feature(vdev, VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
 998             virtio_error(vdev, "cannot set config.bypass");
 999             return;
1000         } else if (in_config->bypass != 0 && in_config->bypass != 1) {
1001             virtio_error(vdev, "invalid config.bypass value '%u'",
1002                          in_config->bypass);
1003             return;
1004         }
1005         dev_config->bypass = in_config->bypass;
1006         virtio_iommu_switch_address_space_all(dev);
1007     }
1008
1009     trace_virtio_iommu_set_config(in_config->bypass);
1010 }
1011
1012 static uint64_t virtio_iommu_get_features(VirtIODevice *vdev, uint64_t f,
1013                                           Error **errp)
1014 {
1015     VirtIOIOMMU *dev = VIRTIO_IOMMU(vdev);
1016
1017     f |= dev->features;
1018     trace_virtio_iommu_get_features(f);
1019     return f;
1020 }
1021
1022 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
1023 {
1024     guint ua = GPOINTER_TO_UINT(a);
1025     guint ub = GPOINTER_TO_UINT(b);
1026     return (ua > ub) - (ua < ub);
1027 }
1028
1029 static gboolean virtio_iommu_remap(gpointer key, gpointer value, gpointer data)
1030 {
1031     VirtIOIOMMUMapping *mapping = (VirtIOIOMMUMapping *) value;
1032     VirtIOIOMMUInterval *interval = (VirtIOIOMMUInterval *) key;
1033     IOMMUMemoryRegion *mr = (IOMMUMemoryRegion *) data;
1034
1035     trace_virtio_iommu_remap(mr->parent_obj.name, interval->low, interval->high,
1036                              mapping->phys_addr);
1037     virtio_iommu_notify_map(mr, interval->low, interval->high,
1038                             mapping->phys_addr, mapping->flags);
1039     return false;
1040 }
1041
1042 static void virtio_iommu_replay(IOMMUMemoryRegion *mr, IOMMUNotifier *n)
1043 {
1044     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1045     VirtIOIOMMU *s = sdev->viommu;
1046     uint32_t sid;
1047     VirtIOIOMMUEndpoint *ep;
1048
1049     sid = virtio_iommu_get_bdf(sdev);
1050
1051     qemu_rec_mutex_lock(&s->mutex);
1052
1053     if (!s->endpoints) {
1054         goto unlock;
1055     }
1056
1057     ep = g_tree_lookup(s->endpoints, GUINT_TO_POINTER(sid));
1058     if (!ep || !ep->domain) {
1059         goto unlock;
1060     }
1061
1062     g_tree_foreach(ep->domain->mappings, virtio_iommu_remap, mr);
1063
1064 unlock:
1065     qemu_rec_mutex_unlock(&s->mutex);
1066 }
1067
1068 static int virtio_iommu_notify_flag_changed(IOMMUMemoryRegion *iommu_mr,
1069                                             IOMMUNotifierFlag old,
1070                                             IOMMUNotifierFlag new,
1071                                             Error **errp)
1072 {
1073     if (new & IOMMU_NOTIFIER_DEVIOTLB_UNMAP) {
1074         error_setg(errp, "Virtio-iommu does not support dev-iotlb yet");
1075         return -EINVAL;
1076     }
1077
1078     if (old == IOMMU_NOTIFIER_NONE) {
1079         trace_virtio_iommu_notify_flag_add(iommu_mr->parent_obj.name);
1080     } else if (new == IOMMU_NOTIFIER_NONE) {
1081         trace_virtio_iommu_notify_flag_del(iommu_mr->parent_obj.name);
1082     }
1083     return 0;
1084 }
1085
1086 /*
1087  * The default mask (TARGET_PAGE_MASK) is the smallest supported guest granule,
1088  * for example 0xfffffffffffff000. When an assigned device has page size
1089  * restrictions due to the hardware IOMMU configuration, apply this restriction
1090  * to the mask.
1091  */
1092 static int virtio_iommu_set_page_size_mask(IOMMUMemoryRegion *mr,
1093                                            uint64_t new_mask,
1094                                            Error **errp)
1095 {
1096     IOMMUDevice *sdev = container_of(mr, IOMMUDevice, iommu_mr);
1097     VirtIOIOMMU *s = sdev->viommu;
1098     uint64_t cur_mask = s->config.page_size_mask;
1099
1100     trace_virtio_iommu_set_page_size_mask(mr->parent_obj.name, cur_mask,
1101                                           new_mask);
1102
1103     if ((cur_mask & new_mask) == 0) {
1104         error_setg(errp, "virtio-iommu page mask 0x%"PRIx64
1105                    " is incompatible with mask 0x%"PRIx64, cur_mask, new_mask);
1106         return -1;
1107     }
1108
1109     /*
1110      * After the machine is finalized, we can't change the mask anymore. If by
1111      * chance the hotplugged device supports the same granule, we can still
1112      * accept it. Having a different masks is possible but the guest will use
1113      * sub-optimal block sizes, so warn about it.
1114      */
1115     if (phase_check(PHASE_MACHINE_READY)) {
1116         int new_granule = ctz64(new_mask);
1117         int cur_granule = ctz64(cur_mask);
1118
1119         if (new_granule != cur_granule) {
1120             error_setg(errp, "virtio-iommu page mask 0x%"PRIx64
1121                        " is incompatible with mask 0x%"PRIx64, cur_mask,
1122                        new_mask);
1123             return -1;
1124         } else if (new_mask != cur_mask) {
1125             warn_report("virtio-iommu page mask 0x%"PRIx64
1126                         " does not match 0x%"PRIx64, cur_mask, new_mask);
1127         }
1128         return 0;
1129     }
1130
1131     s->config.page_size_mask &= new_mask;
1132     return 0;
1133 }
1134
1135 static void virtio_iommu_system_reset(void *opaque)
1136 {
1137     VirtIOIOMMU *s = opaque;
1138
1139     trace_virtio_iommu_system_reset();
1140
1141     /*
1142      * config.bypass is sticky across device reset, but should be restored on
1143      * system reset
1144      */
1145     s->config.bypass = s->boot_bypass;
1146     virtio_iommu_switch_address_space_all(s);
1147
1148 }
1149
1150 static void virtio_iommu_device_realize(DeviceState *dev, Error **errp)
1151 {
1152     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1153     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1154
1155     virtio_init(vdev, VIRTIO_ID_IOMMU, sizeof(struct virtio_iommu_config));
1156
1157     memset(s->iommu_pcibus_by_bus_num, 0, sizeof(s->iommu_pcibus_by_bus_num));
1158
1159     s->req_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE,
1160                              virtio_iommu_handle_command);
1161     s->event_vq = virtio_add_queue(vdev, VIOMMU_DEFAULT_QUEUE_SIZE, NULL);
1162
1163     /*
1164      * config.bypass is needed to get initial address space early, such as
1165      * in vfio realize
1166      */
1167     s->config.bypass = s->boot_bypass;
1168     s->config.page_size_mask = TARGET_PAGE_MASK;
1169     s->config.input_range.end = UINT64_MAX;
1170     s->config.domain_range.end = UINT32_MAX;
1171     s->config.probe_size = VIOMMU_PROBE_SIZE;
1172
1173     virtio_add_feature(&s->features, VIRTIO_RING_F_EVENT_IDX);
1174     virtio_add_feature(&s->features, VIRTIO_RING_F_INDIRECT_DESC);
1175     virtio_add_feature(&s->features, VIRTIO_F_VERSION_1);
1176     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_INPUT_RANGE);
1177     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_DOMAIN_RANGE);
1178     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MAP_UNMAP);
1179     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_MMIO);
1180     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_PROBE);
1181     virtio_add_feature(&s->features, VIRTIO_IOMMU_F_BYPASS_CONFIG);
1182
1183     qemu_rec_mutex_init(&s->mutex);
1184
1185     s->as_by_busptr = g_hash_table_new_full(NULL, NULL, NULL, g_free);
1186
1187     if (s->primary_bus) {
1188         pci_setup_iommu(s->primary_bus, virtio_iommu_find_add_as, s);
1189     } else {
1190         error_setg(errp, "VIRTIO-IOMMU is not attached to any PCI bus!");
1191     }
1192
1193     qemu_register_reset(virtio_iommu_system_reset, s);
1194 }
1195
1196 static void virtio_iommu_device_unrealize(DeviceState *dev)
1197 {
1198     VirtIODevice *vdev = VIRTIO_DEVICE(dev);
1199     VirtIOIOMMU *s = VIRTIO_IOMMU(dev);
1200
1201     qemu_unregister_reset(virtio_iommu_system_reset, s);
1202
1203     g_hash_table_destroy(s->as_by_busptr);
1204     if (s->domains) {
1205         g_tree_destroy(s->domains);
1206     }
1207     if (s->endpoints) {
1208         g_tree_destroy(s->endpoints);
1209     }
1210
1211     qemu_rec_mutex_destroy(&s->mutex);
1212
1213     virtio_delete_queue(s->req_vq);
1214     virtio_delete_queue(s->event_vq);
1215     virtio_cleanup(vdev);
1216 }
1217
1218 static void virtio_iommu_device_reset(VirtIODevice *vdev)
1219 {
1220     VirtIOIOMMU *s = VIRTIO_IOMMU(vdev);
1221
1222     trace_virtio_iommu_device_reset();
1223
1224     if (s->domains) {
1225         g_tree_destroy(s->domains);
1226     }
1227     if (s->endpoints) {
1228         g_tree_destroy(s->endpoints);
1229     }
1230     s->domains = g_tree_new_full((GCompareDataFunc)int_cmp,
1231                                  NULL, NULL, virtio_iommu_put_domain);
1232     s->endpoints = g_tree_new_full((GCompareDataFunc)int_cmp,
1233                                    NULL, NULL, virtio_iommu_put_endpoint);
1234 }
1235
1236 static void virtio_iommu_set_status(VirtIODevice *vdev, uint8_t status)
1237 {
1238     trace_virtio_iommu_device_status(status);
1239 }
1240
1241 static void virtio_iommu_instance_init(Object *obj)
1242 {
1243 }
1244
1245 #define VMSTATE_INTERVAL                               \
1246 {                                                      \
1247     .name = "interval",                                \
1248     .version_id = 1,                                   \
1249     .minimum_version_id = 1,                           \
1250     .fields = (VMStateField[]) {                       \
1251         VMSTATE_UINT64(low, VirtIOIOMMUInterval),      \
1252         VMSTATE_UINT64(high, VirtIOIOMMUInterval),     \
1253         VMSTATE_END_OF_LIST()                          \
1254     }                                                  \
1255 }
1256
1257 #define VMSTATE_MAPPING                               \
1258 {                                                     \
1259     .name = "mapping",                                \
1260     .version_id = 1,                                  \
1261     .minimum_version_id = 1,                          \
1262     .fields = (VMStateField[]) {                      \
1263         VMSTATE_UINT64(phys_addr, VirtIOIOMMUMapping),\
1264         VMSTATE_UINT32(flags, VirtIOIOMMUMapping),    \
1265         VMSTATE_END_OF_LIST()                         \
1266     },                                                \
1267 }
1268
1269 static const VMStateDescription vmstate_interval_mapping[2] = {
1270     VMSTATE_MAPPING,   /* value */
1271     VMSTATE_INTERVAL   /* key   */
1272 };
1273
1274 static int domain_preload(void *opaque)
1275 {
1276     VirtIOIOMMUDomain *domain = opaque;
1277
1278     domain->mappings = g_tree_new_full((GCompareDataFunc)interval_cmp,
1279                                        NULL, g_free, g_free);
1280     return 0;
1281 }
1282
1283 static const VMStateDescription vmstate_endpoint = {
1284     .name = "endpoint",
1285     .version_id = 1,
1286     .minimum_version_id = 1,
1287     .fields = (VMStateField[]) {
1288         VMSTATE_UINT32(id, VirtIOIOMMUEndpoint),
1289         VMSTATE_END_OF_LIST()
1290     }
1291 };
1292
1293 static const VMStateDescription vmstate_domain = {
1294     .name = "domain",
1295     .version_id = 2,
1296     .minimum_version_id = 2,
1297     .pre_load = domain_preload,
1298     .fields = (VMStateField[]) {
1299         VMSTATE_UINT32(id, VirtIOIOMMUDomain),
1300         VMSTATE_GTREE_V(mappings, VirtIOIOMMUDomain, 1,
1301                         vmstate_interval_mapping,
1302                         VirtIOIOMMUInterval, VirtIOIOMMUMapping),
1303         VMSTATE_QLIST_V(endpoint_list, VirtIOIOMMUDomain, 1,
1304                         vmstate_endpoint, VirtIOIOMMUEndpoint, next),
1305         VMSTATE_BOOL_V(bypass, VirtIOIOMMUDomain, 2),
1306         VMSTATE_END_OF_LIST()
1307     }
1308 };
1309
1310 static gboolean reconstruct_endpoints(gpointer key, gpointer value,
1311                                       gpointer data)
1312 {
1313     VirtIOIOMMU *s = (VirtIOIOMMU *)data;
1314     VirtIOIOMMUDomain *d = (VirtIOIOMMUDomain *)value;
1315     VirtIOIOMMUEndpoint *iter;
1316     IOMMUMemoryRegion *mr;
1317
1318     QLIST_FOREACH(iter, &d->endpoint_list, next) {
1319         mr = virtio_iommu_mr(s, iter->id);
1320         assert(mr);
1321
1322         iter->domain = d;
1323         iter->iommu_mr = mr;
1324         g_tree_insert(s->endpoints, GUINT_TO_POINTER(iter->id), iter);
1325     }
1326     return false; /* continue the domain traversal */
1327 }
1328
1329 static int iommu_post_load(void *opaque, int version_id)
1330 {
1331     VirtIOIOMMU *s = opaque;
1332
1333     g_tree_foreach(s->domains, reconstruct_endpoints, s);
1334
1335     /*
1336      * Memory regions are dynamically turned on/off depending on
1337      * 'config.bypass' and attached domain type if there is. After
1338      * migration, we need to make sure the memory regions are
1339      * still correct.
1340      */
1341     virtio_iommu_switch_address_space_all(s);
1342     return 0;
1343 }
1344
1345 static const VMStateDescription vmstate_virtio_iommu_device = {
1346     .name = "virtio-iommu-device",
1347     .minimum_version_id = 2,
1348     .version_id = 2,
1349     .post_load = iommu_post_load,
1350     .fields = (VMStateField[]) {
1351         VMSTATE_GTREE_DIRECT_KEY_V(domains, VirtIOIOMMU, 2,
1352                                    &vmstate_domain, VirtIOIOMMUDomain),
1353         VMSTATE_UINT8_V(config.bypass, VirtIOIOMMU, 2),
1354         VMSTATE_END_OF_LIST()
1355     },
1356 };
1357
1358 static const VMStateDescription vmstate_virtio_iommu = {
1359     .name = "virtio-iommu",
1360     .minimum_version_id = 2,
1361     .priority = MIG_PRI_IOMMU,
1362     .version_id = 2,
1363     .fields = (VMStateField[]) {
1364         VMSTATE_VIRTIO_DEVICE,
1365         VMSTATE_END_OF_LIST()
1366     },
1367 };
1368
1369 static Property virtio_iommu_properties[] = {
1370     DEFINE_PROP_LINK("primary-bus", VirtIOIOMMU, primary_bus, "PCI", PCIBus *),
1371     DEFINE_PROP_BOOL("boot-bypass", VirtIOIOMMU, boot_bypass, true),
1372     DEFINE_PROP_END_OF_LIST(),
1373 };
1374
1375 static void virtio_iommu_class_init(ObjectClass *klass, void *data)
1376 {
1377     DeviceClass *dc = DEVICE_CLASS(klass);
1378     VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass);
1379
1380     device_class_set_props(dc, virtio_iommu_properties);
1381     dc->vmsd = &vmstate_virtio_iommu;
1382
1383     set_bit(DEVICE_CATEGORY_MISC, dc->categories);
1384     vdc->realize = virtio_iommu_device_realize;
1385     vdc->unrealize = virtio_iommu_device_unrealize;
1386     vdc->reset = virtio_iommu_device_reset;
1387     vdc->get_config = virtio_iommu_get_config;
1388     vdc->set_config = virtio_iommu_set_config;
1389     vdc->get_features = virtio_iommu_get_features;
1390     vdc->set_status = virtio_iommu_set_status;
1391     vdc->vmsd = &vmstate_virtio_iommu_device;
1392 }
1393
1394 static void virtio_iommu_memory_region_class_init(ObjectClass *klass,
1395                                                   void *data)
1396 {
1397     IOMMUMemoryRegionClass *imrc = IOMMU_MEMORY_REGION_CLASS(klass);
1398
1399     imrc->translate = virtio_iommu_translate;
1400     imrc->replay = virtio_iommu_replay;
1401     imrc->notify_flag_changed = virtio_iommu_notify_flag_changed;
1402     imrc->iommu_set_page_size_mask = virtio_iommu_set_page_size_mask;
1403 }
1404
1405 static const TypeInfo virtio_iommu_info = {
1406     .name = TYPE_VIRTIO_IOMMU,
1407     .parent = TYPE_VIRTIO_DEVICE,
1408     .instance_size = sizeof(VirtIOIOMMU),
1409     .instance_init = virtio_iommu_instance_init,
1410     .class_init = virtio_iommu_class_init,
1411 };
1412
1413 static const TypeInfo virtio_iommu_memory_region_info = {
1414     .parent = TYPE_IOMMU_MEMORY_REGION,
1415     .name = TYPE_VIRTIO_IOMMU_MEMORY_REGION,
1416     .class_init = virtio_iommu_memory_region_class_init,
1417 };
1418
1419 static void virtio_register_types(void)
1420 {
1421     type_register_static(&virtio_iommu_info);
1422     type_register_static(&virtio_iommu_memory_region_info);
1423 }
1424
1425 type_init(virtio_register_types)