hw/vfio/common.c

   1 /*
   2  * generic functions used by VFIO devices
   3  *
   4  * Copyright Red Hat, Inc. 2012
   5  *
   6  * Authors:
   7  *  Alex Williamson <alex.williamson@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Based on qemu-kvm device-assignment:
  13  *  Adapted for KVM by Qumranet.
  14  *  Copyright (c) 2007, Neocleus, Alex Novik (alex@neocleus.com)
  15  *  Copyright (c) 2007, Neocleus, Guy Zana (guy@neocleus.com)
  16  *  Copyright (C) 2008, Qumranet, Amit Shah (amit.shah@qumranet.com)
  17  *  Copyright (C) 2008, Red Hat, Amit Shah (amit.shah@redhat.com)
  18  *  Copyright (C) 2008, IBM, Muli Ben-Yehuda (muli@il.ibm.com)
  19  */
  20
  21 #include <sys/ioctl.h>
  22 #include <sys/mman.h>
  23 #include <linux/vfio.h>
  24
  25 #include "hw/vfio/vfio-common.h"
  26 #include "hw/vfio/vfio.h"
  27 #include "exec/address-spaces.h"
  28 #include "exec/memory.h"
  29 #include "hw/hw.h"
  30 #include "qemu/error-report.h"
  31 #include "sysemu/kvm.h"
  32 #include "trace.h"
  33
  34 struct vfio_group_head vfio_group_list =
  35     QLIST_HEAD_INITIALIZER(vfio_group_list);
  36 struct vfio_as_head vfio_address_spaces =
  37     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
  38
  39 #ifdef CONFIG_KVM
  40 /*
  41  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
  42  * for the life of the VM.  Closing the file descriptor only drops our
  43  * reference to it and the device's reference to kvm.  Therefore once
  44  * initialized, this file descriptor is only released on QEMU exit and
  45  * we'll re-use it should another vfio device be attached before then.
  46  */
  47 static int vfio_kvm_device_fd = -1;
  48 #endif
  49
  50 /*
  51  * Common VFIO interrupt disable
  52  */
  53 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
  54 {
  55     struct vfio_irq_set irq_set = {
  56         .argsz = sizeof(irq_set),
  57         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
  58         .index = index,
  59         .start = 0,
  60         .count = 0,
  61     };
  62
  63     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  64 }
  65
  66 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
  67 {
  68     struct vfio_irq_set irq_set = {
  69         .argsz = sizeof(irq_set),
  70         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
  71         .index = index,
  72         .start = 0,
  73         .count = 1,
  74     };
  75
  76     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  77 }
  78
  79 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
  80 {
  81     struct vfio_irq_set irq_set = {
  82         .argsz = sizeof(irq_set),
  83         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
  84         .index = index,
  85         .start = 0,
  86         .count = 1,
  87     };
  88
  89     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  90 }
  91
  92 /*
  93  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
  94  */
  95 void vfio_region_write(void *opaque, hwaddr addr,
  96                        uint64_t data, unsigned size)
  97 {
  98     VFIORegion *region = opaque;
  99     VFIODevice *vbasedev = region->vbasedev;
 100     union {
 101         uint8_t byte;
 102         uint16_t word;
 103         uint32_t dword;
 104         uint64_t qword;
 105     } buf;
 106
 107     switch (size) {
 108     case 1:
 109         buf.byte = data;
 110         break;
 111     case 2:
 112         buf.word = cpu_to_le16(data);
 113         break;
 114     case 4:
 115         buf.dword = cpu_to_le32(data);
 116         break;
 117     default:
 118         hw_error("vfio: unsupported write size, %d bytes", size);
 119         break;
 120     }
 121
 122     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 123         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
 124                      ",%d) failed: %m",
 125                      __func__, vbasedev->name, region->nr,
 126                      addr, data, size);
 127     }
 128
 129     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
 130
 131     /*
 132      * A read or write to a BAR always signals an INTx EOI.  This will
 133      * do nothing if not pending (including not in INTx mode).  We assume
 134      * that a BAR access is in response to an interrupt and that BAR
 135      * accesses will service the interrupt.  Unfortunately, we don't know
 136      * which access will service the interrupt, so we're potentially
 137      * getting quite a few host interrupts per guest interrupt.
 138      */
 139     vbasedev->ops->vfio_eoi(vbasedev);
 140 }
 141
 142 uint64_t vfio_region_read(void *opaque,
 143                           hwaddr addr, unsigned size)
 144 {
 145     VFIORegion *region = opaque;
 146     VFIODevice *vbasedev = region->vbasedev;
 147     union {
 148         uint8_t byte;
 149         uint16_t word;
 150         uint32_t dword;
 151         uint64_t qword;
 152     } buf;
 153     uint64_t data = 0;
 154
 155     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 156         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
 157                      __func__, vbasedev->name, region->nr,
 158                      addr, size);
 159         return (uint64_t)-1;
 160     }
 161     switch (size) {
 162     case 1:
 163         data = buf.byte;
 164         break;
 165     case 2:
 166         data = le16_to_cpu(buf.word);
 167         break;
 168     case 4:
 169         data = le32_to_cpu(buf.dword);
 170         break;
 171     default:
 172         hw_error("vfio: unsupported read size, %d bytes", size);
 173         break;
 174     }
 175
 176     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
 177
 178     /* Same as write above */
 179     vbasedev->ops->vfio_eoi(vbasedev);
 180
 181     return data;
 182 }
 183
 184 const MemoryRegionOps vfio_region_ops = {
 185     .read = vfio_region_read,
 186     .write = vfio_region_write,
 187     .endianness = DEVICE_LITTLE_ENDIAN,
 188 };
 189
 190 /*
 191  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
 192  */
 193 static int vfio_dma_unmap(VFIOContainer *container,
 194                           hwaddr iova, ram_addr_t size)
 195 {
 196     struct vfio_iommu_type1_dma_unmap unmap = {
 197         .argsz = sizeof(unmap),
 198         .flags = 0,
 199         .iova = iova,
 200         .size = size,
 201     };
 202
 203     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
 204         error_report("VFIO_UNMAP_DMA: %d", -errno);
 205         return -errno;
 206     }
 207
 208     return 0;
 209 }
 210
 211 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
 212                         ram_addr_t size, void *vaddr, bool readonly)
 213 {
 214     struct vfio_iommu_type1_dma_map map = {
 215         .argsz = sizeof(map),
 216         .flags = VFIO_DMA_MAP_FLAG_READ,
 217         .vaddr = (__u64)(uintptr_t)vaddr,
 218         .iova = iova,
 219         .size = size,
 220     };
 221
 222     if (!readonly) {
 223         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
 224     }
 225
 226     /*
 227      * Try the mapping, if it fails with EBUSY, unmap the region and try
 228      * again.  This shouldn't be necessary, but we sometimes see it in
 229      * the VGA ROM space.
 230      */
 231     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
 232         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
 233          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
 234         return 0;
 235     }
 236
 237     error_report("VFIO_MAP_DMA: %d", -errno);
 238     return -errno;
 239 }
 240
 241 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 242 {
 243     return (!memory_region_is_ram(section->mr) &&
 244             !memory_region_is_iommu(section->mr)) ||
 245            /*
 246             * Sizing an enabled 64-bit BAR can cause spurious mappings to
 247             * addresses in the upper part of the 64-bit address space.  These
 248             * are never accessed by the CPU and beyond the address width of
 249             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
 250             */
 251            section->offset_within_address_space & (1ULL << 63);
 252 }
 253
 254 static void vfio_iommu_map_notify(Notifier *n, void *data)
 255 {
 256     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
 257     VFIOContainer *container = giommu->container;
 258     IOMMUTLBEntry *iotlb = data;
 259     MemoryRegion *mr;
 260     hwaddr xlat;
 261     hwaddr len = iotlb->addr_mask + 1;
 262     void *vaddr;
 263     int ret;
 264
 265     trace_vfio_iommu_map_notify(iotlb->iova,
 266                                 iotlb->iova + iotlb->addr_mask);
 267
 268     /*
 269      * The IOMMU TLB entry we have just covers translation through
 270      * this IOMMU to its immediate target.  We need to translate
 271      * it the rest of the way through to memory.
 272      */
 273     rcu_read_lock();
 274     mr = address_space_translate(&address_space_memory,
 275                                  iotlb->translated_addr,
 276                                  &xlat, &len, iotlb->perm & IOMMU_WO);
 277     if (!memory_region_is_ram(mr)) {
 278         error_report("iommu map to non memory area %"HWADDR_PRIx"",
 279                      xlat);
 280         goto out;
 281     }
 282     /*
 283      * Translation truncates length to the IOMMU page size,
 284      * check that it did not truncate too much.
 285      */
 286     if (len & iotlb->addr_mask) {
 287         error_report("iommu has granularity incompatible with target AS");
 288         goto out;
 289     }
 290
 291     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
 292         vaddr = memory_region_get_ram_ptr(mr) + xlat;
 293         ret = vfio_dma_map(container, iotlb->iova,
 294                            iotlb->addr_mask + 1, vaddr,
 295                            !(iotlb->perm & IOMMU_WO) || mr->readonly);
 296         if (ret) {
 297             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 298                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
 299                          container, iotlb->iova,
 300                          iotlb->addr_mask + 1, vaddr, ret);
 301         }
 302     } else {
 303         ret = vfio_dma_unmap(container, iotlb->iova, iotlb->addr_mask + 1);
 304         if (ret) {
 305             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 306                          "0x%"HWADDR_PRIx") = %d (%m)",
 307                          container, iotlb->iova,
 308                          iotlb->addr_mask + 1, ret);
 309         }
 310     }
 311 out:
 312     rcu_read_unlock();
 313 }
 314
 315 static void vfio_listener_region_add(MemoryListener *listener,
 316                                      MemoryRegionSection *section)
 317 {
 318     VFIOContainer *container = container_of(listener, VFIOContainer,
 319                                             iommu_data.type1.listener);
 320     hwaddr iova, end;
 321     Int128 llend;
 322     void *vaddr;
 323     int ret;
 324
 325     if (vfio_listener_skipped_section(section)) {
 326         trace_vfio_listener_region_add_skip(
 327                 section->offset_within_address_space,
 328                 section->offset_within_address_space +
 329                 int128_get64(int128_sub(section->size, int128_one())));
 330         return;
 331     }
 332
 333     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 334                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 335         error_report("%s received unaligned region", __func__);
 336         return;
 337     }
 338
 339     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 340     llend = int128_make64(section->offset_within_address_space);
 341     llend = int128_add(llend, section->size);
 342     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 343
 344     if (int128_ge(int128_make64(iova), llend)) {
 345         return;
 346     }
 347
 348     memory_region_ref(section->mr);
 349
 350     if (memory_region_is_iommu(section->mr)) {
 351         VFIOGuestIOMMU *giommu;
 352
 353         trace_vfio_listener_region_add_iommu(iova,
 354                     int128_get64(int128_sub(llend, int128_one())));
 355         /*
 356          * FIXME: We should do some checking to see if the
 357          * capabilities of the host VFIO IOMMU are adequate to model
 358          * the guest IOMMU
 359          *
 360          * FIXME: For VFIO iommu types which have KVM acceleration to
 361          * avoid bouncing all map/unmaps through qemu this way, this
 362          * would be the right place to wire that up (tell the KVM
 363          * device emulation the VFIO iommu handles to use).
 364          */
 365         /*
 366          * This assumes that the guest IOMMU is empty of
 367          * mappings at this point.
 368          *
 369          * One way of doing this is:
 370          * 1. Avoid sharing IOMMUs between emulated devices or different
 371          * IOMMU groups.
 372          * 2. Implement VFIO_IOMMU_ENABLE in the host kernel to fail if
 373          * there are some mappings in IOMMU.
 374          *
 375          * VFIO on SPAPR does that. Other IOMMU models may do that different,
 376          * they must make sure there are no existing mappings or
 377          * loop through existing mappings to map them into VFIO.
 378          */
 379         giommu = g_malloc0(sizeof(*giommu));
 380         giommu->iommu = section->mr;
 381         giommu->container = container;
 382         giommu->n.notify = vfio_iommu_map_notify;
 383         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
 384         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
 385
 386         return;
 387     }
 388
 389     /* Here we assume that memory_region_is_ram(section->mr)==true */
 390
 391     end = int128_get64(llend);
 392     vaddr = memory_region_get_ram_ptr(section->mr) +
 393             section->offset_within_region +
 394             (iova - section->offset_within_address_space);
 395
 396     trace_vfio_listener_region_add_ram(iova, end - 1, vaddr);
 397
 398     ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
 399     if (ret) {
 400         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 401                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
 402                      container, iova, end - iova, vaddr, ret);
 403
 404         /*
 405          * On the initfn path, store the first error in the container so we
 406          * can gracefully fail.  Runtime, there's not much we can do other
 407          * than throw a hardware error.
 408          */
 409         if (!container->iommu_data.type1.initialized) {
 410             if (!container->iommu_data.type1.error) {
 411                 container->iommu_data.type1.error = ret;
 412             }
 413         } else {
 414             hw_error("vfio: DMA mapping failed, unable to continue");
 415         }
 416     }
 417 }
 418
 419 static void vfio_listener_region_del(MemoryListener *listener,
 420                                      MemoryRegionSection *section)
 421 {
 422     VFIOContainer *container = container_of(listener, VFIOContainer,
 423                                             iommu_data.type1.listener);
 424     hwaddr iova, end;
 425     int ret;
 426
 427     if (vfio_listener_skipped_section(section)) {
 428         trace_vfio_listener_region_del_skip(
 429                 section->offset_within_address_space,
 430                 section->offset_within_address_space +
 431                 int128_get64(int128_sub(section->size, int128_one())));
 432         return;
 433     }
 434
 435     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 436                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 437         error_report("%s received unaligned region", __func__);
 438         return;
 439     }
 440
 441     if (memory_region_is_iommu(section->mr)) {
 442         VFIOGuestIOMMU *giommu;
 443
 444         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
 445             if (giommu->iommu == section->mr) {
 446                 memory_region_unregister_iommu_notifier(&giommu->n);
 447                 QLIST_REMOVE(giommu, giommu_next);
 448                 g_free(giommu);
 449                 break;
 450             }
 451         }
 452
 453         /*
 454          * FIXME: We assume the one big unmap below is adequate to
 455          * remove any individual page mappings in the IOMMU which
 456          * might have been copied into VFIO. This works for a page table
 457          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
 458          * That may not be true for all IOMMU types.
 459          */
 460     }
 461
 462     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 463     end = (section->offset_within_address_space + int128_get64(section->size)) &
 464           TARGET_PAGE_MASK;
 465
 466     if (iova >= end) {
 467         return;
 468     }
 469
 470     trace_vfio_listener_region_del(iova, end - 1);
 471
 472     ret = vfio_dma_unmap(container, iova, end - iova);
 473     memory_region_unref(section->mr);
 474     if (ret) {
 475         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 476                      "0x%"HWADDR_PRIx") = %d (%m)",
 477                      container, iova, end - iova, ret);
 478     }
 479 }
 480
 481 static const MemoryListener vfio_memory_listener = {
 482     .region_add = vfio_listener_region_add,
 483     .region_del = vfio_listener_region_del,
 484 };
 485
 486 static void vfio_listener_release(VFIOContainer *container)
 487 {
 488     memory_listener_unregister(&container->iommu_data.type1.listener);
 489 }
 490
 491 int vfio_mmap_region(Object *obj, VFIORegion *region,
 492                      MemoryRegion *mem, MemoryRegion *submem,
 493                      void **map, size_t size, off_t offset,
 494                      const char *name)
 495 {
 496     int ret = 0;
 497     VFIODevice *vbasedev = region->vbasedev;
 498
 499     if (!vbasedev->no_mmap && size && region->flags &
 500         VFIO_REGION_INFO_FLAG_MMAP) {
 501         int prot = 0;
 502
 503         if (region->flags & VFIO_REGION_INFO_FLAG_READ) {
 504             prot |= PROT_READ;
 505         }
 506
 507         if (region->flags & VFIO_REGION_INFO_FLAG_WRITE) {
 508             prot |= PROT_WRITE;
 509         }
 510
 511         *map = mmap(NULL, size, prot, MAP_SHARED,
 512                     vbasedev->fd,
 513                     region->fd_offset + offset);
 514         if (*map == MAP_FAILED) {
 515             *map = NULL;
 516             ret = -errno;
 517             goto empty_region;
 518         }
 519
 520         memory_region_init_ram_ptr(submem, obj, name, size, *map);
 521         memory_region_set_skip_dump(submem);
 522     } else {
 523 empty_region:
 524         /* Create a zero sized sub-region to make cleanup easy. */
 525         memory_region_init(submem, obj, name, 0);
 526     }
 527
 528     memory_region_add_subregion(mem, offset, submem);
 529
 530     return ret;
 531 }
 532
 533 void vfio_reset_handler(void *opaque)
 534 {
 535     VFIOGroup *group;
 536     VFIODevice *vbasedev;
 537
 538     QLIST_FOREACH(group, &vfio_group_list, next) {
 539         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 540             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
 541         }
 542     }
 543
 544     QLIST_FOREACH(group, &vfio_group_list, next) {
 545         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 546             if (vbasedev->needs_reset) {
 547                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
 548             }
 549         }
 550     }
 551 }
 552
 553 static void vfio_kvm_device_add_group(VFIOGroup *group)
 554 {
 555 #ifdef CONFIG_KVM
 556     struct kvm_device_attr attr = {
 557         .group = KVM_DEV_VFIO_GROUP,
 558         .attr = KVM_DEV_VFIO_GROUP_ADD,
 559         .addr = (uint64_t)(unsigned long)&group->fd,
 560     };
 561
 562     if (!kvm_enabled()) {
 563         return;
 564     }
 565
 566     if (vfio_kvm_device_fd < 0) {
 567         struct kvm_create_device cd = {
 568             .type = KVM_DEV_TYPE_VFIO,
 569         };
 570
 571         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
 572             error_report("Failed to create KVM VFIO device: %m");
 573             return;
 574         }
 575
 576         vfio_kvm_device_fd = cd.fd;
 577     }
 578
 579     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 580         error_report("Failed to add group %d to KVM VFIO device: %m",
 581                      group->groupid);
 582     }
 583 #endif
 584 }
 585
 586 static void vfio_kvm_device_del_group(VFIOGroup *group)
 587 {
 588 #ifdef CONFIG_KVM
 589     struct kvm_device_attr attr = {
 590         .group = KVM_DEV_VFIO_GROUP,
 591         .attr = KVM_DEV_VFIO_GROUP_DEL,
 592         .addr = (uint64_t)(unsigned long)&group->fd,
 593     };
 594
 595     if (vfio_kvm_device_fd < 0) {
 596         return;
 597     }
 598
 599     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 600         error_report("Failed to remove group %d from KVM VFIO device: %m",
 601                      group->groupid);
 602     }
 603 #endif
 604 }
 605
 606 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
 607 {
 608     VFIOAddressSpace *space;
 609
 610     QLIST_FOREACH(space, &vfio_address_spaces, list) {
 611         if (space->as == as) {
 612             return space;
 613         }
 614     }
 615
 616     /* No suitable VFIOAddressSpace, create a new one */
 617     space = g_malloc0(sizeof(*space));
 618     space->as = as;
 619     QLIST_INIT(&space->containers);
 620
 621     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
 622
 623     return space;
 624 }
 625
 626 static void vfio_put_address_space(VFIOAddressSpace *space)
 627 {
 628     if (QLIST_EMPTY(&space->containers)) {
 629         QLIST_REMOVE(space, list);
 630         g_free(space);
 631     }
 632 }
 633
 634 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
 635 {
 636     VFIOContainer *container;
 637     int ret, fd;
 638     VFIOAddressSpace *space;
 639
 640     space = vfio_get_address_space(as);
 641
 642     QLIST_FOREACH(container, &space->containers, next) {
 643         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
 644             group->container = container;
 645             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 646             return 0;
 647         }
 648     }
 649
 650     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
 651     if (fd < 0) {
 652         error_report("vfio: failed to open /dev/vfio/vfio: %m");
 653         ret = -errno;
 654         goto put_space_exit;
 655     }
 656
 657     ret = ioctl(fd, VFIO_GET_API_VERSION);
 658     if (ret != VFIO_API_VERSION) {
 659         error_report("vfio: supported vfio version: %d, "
 660                      "reported version: %d", VFIO_API_VERSION, ret);
 661         ret = -EINVAL;
 662         goto close_fd_exit;
 663     }
 664
 665     container = g_malloc0(sizeof(*container));
 666     container->space = space;
 667     container->fd = fd;
 668     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
 669         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
 670         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
 671
 672         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 673         if (ret) {
 674             error_report("vfio: failed to set group container: %m");
 675             ret = -errno;
 676             goto free_container_exit;
 677         }
 678
 679         ret = ioctl(fd, VFIO_SET_IOMMU,
 680                     v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU);
 681         if (ret) {
 682             error_report("vfio: failed to set iommu for container: %m");
 683             ret = -errno;
 684             goto free_container_exit;
 685         }
 686
 687         container->iommu_data.type1.listener = vfio_memory_listener;
 688         container->iommu_data.release = vfio_listener_release;
 689
 690         memory_listener_register(&container->iommu_data.type1.listener,
 691                                  container->space->as);
 692
 693         if (container->iommu_data.type1.error) {
 694             ret = container->iommu_data.type1.error;
 695             error_report("vfio: memory listener initialization failed for container");
 696             goto listener_release_exit;
 697         }
 698
 699         container->iommu_data.type1.initialized = true;
 700
 701     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) {
 702         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 703         if (ret) {
 704             error_report("vfio: failed to set group container: %m");
 705             ret = -errno;
 706             goto free_container_exit;
 707         }
 708         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU);
 709         if (ret) {
 710             error_report("vfio: failed to set iommu for container: %m");
 711             ret = -errno;
 712             goto free_container_exit;
 713         }
 714
 715         /*
 716          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
 717          * when container fd is closed so we do not call it explicitly
 718          * in this file.
 719          */
 720         ret = ioctl(fd, VFIO_IOMMU_ENABLE);
 721         if (ret) {
 722             error_report("vfio: failed to enable container: %m");
 723             ret = -errno;
 724             goto free_container_exit;
 725         }
 726
 727         container->iommu_data.type1.listener = vfio_memory_listener;
 728         container->iommu_data.release = vfio_listener_release;
 729
 730         memory_listener_register(&container->iommu_data.type1.listener,
 731                                  container->space->as);
 732
 733     } else {
 734         error_report("vfio: No available IOMMU models");
 735         ret = -EINVAL;
 736         goto free_container_exit;
 737     }
 738
 739     QLIST_INIT(&container->group_list);
 740     QLIST_INSERT_HEAD(&space->containers, container, next);
 741
 742     group->container = container;
 743     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 744
 745     return 0;
 746 listener_release_exit:
 747     vfio_listener_release(container);
 748
 749 free_container_exit:
 750     g_free(container);
 751
 752 close_fd_exit:
 753     close(fd);
 754
 755 put_space_exit:
 756     vfio_put_address_space(space);
 757
 758     return ret;
 759 }
 760
 761 static void vfio_disconnect_container(VFIOGroup *group)
 762 {
 763     VFIOContainer *container = group->container;
 764
 765     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
 766         error_report("vfio: error disconnecting group %d from container",
 767                      group->groupid);
 768     }
 769
 770     QLIST_REMOVE(group, container_next);
 771     group->container = NULL;
 772
 773     if (QLIST_EMPTY(&container->group_list)) {
 774         VFIOAddressSpace *space = container->space;
 775         VFIOGuestIOMMU *giommu, *tmp;
 776
 777         if (container->iommu_data.release) {
 778             container->iommu_data.release(container);
 779         }
 780         QLIST_REMOVE(container, next);
 781
 782         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
 783             memory_region_unregister_iommu_notifier(&giommu->n);
 784             QLIST_REMOVE(giommu, giommu_next);
 785             g_free(giommu);
 786         }
 787
 788         trace_vfio_disconnect_container(container->fd);
 789         close(container->fd);
 790         g_free(container);
 791
 792         vfio_put_address_space(space);
 793     }
 794 }
 795
 796 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
 797 {
 798     VFIOGroup *group;
 799     char path[32];
 800     struct vfio_group_status status = { .argsz = sizeof(status) };
 801
 802     QLIST_FOREACH(group, &vfio_group_list, next) {
 803         if (group->groupid == groupid) {
 804             /* Found it.  Now is it already in the right context? */
 805             if (group->container->space->as == as) {
 806                 return group;
 807             } else {
 808                 error_report("vfio: group %d used in multiple address spaces",
 809                              group->groupid);
 810                 return NULL;
 811             }
 812         }
 813     }
 814
 815     group = g_malloc0(sizeof(*group));
 816
 817     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
 818     group->fd = qemu_open(path, O_RDWR);
 819     if (group->fd < 0) {
 820         error_report("vfio: error opening %s: %m", path);
 821         goto free_group_exit;
 822     }
 823
 824     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
 825         error_report("vfio: error getting group status: %m");
 826         goto close_fd_exit;
 827     }
 828
 829     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
 830         error_report("vfio: error, group %d is not viable, please ensure "
 831                      "all devices within the iommu_group are bound to their "
 832                      "vfio bus driver.", groupid);
 833         goto close_fd_exit;
 834     }
 835
 836     group->groupid = groupid;
 837     QLIST_INIT(&group->device_list);
 838
 839     if (vfio_connect_container(group, as)) {
 840         error_report("vfio: failed to setup container for group %d", groupid);
 841         goto close_fd_exit;
 842     }
 843
 844     if (QLIST_EMPTY(&vfio_group_list)) {
 845         qemu_register_reset(vfio_reset_handler, NULL);
 846     }
 847
 848     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
 849
 850     vfio_kvm_device_add_group(group);
 851
 852     return group;
 853
 854 close_fd_exit:
 855     close(group->fd);
 856
 857 free_group_exit:
 858     g_free(group);
 859
 860     return NULL;
 861 }
 862
 863 void vfio_put_group(VFIOGroup *group)
 864 {
 865     if (!group || !QLIST_EMPTY(&group->device_list)) {
 866         return;
 867     }
 868
 869     vfio_kvm_device_del_group(group);
 870     vfio_disconnect_container(group);
 871     QLIST_REMOVE(group, next);
 872     trace_vfio_put_group(group->fd);
 873     close(group->fd);
 874     g_free(group);
 875
 876     if (QLIST_EMPTY(&vfio_group_list)) {
 877         qemu_unregister_reset(vfio_reset_handler, NULL);
 878     }
 879 }
 880
 881 int vfio_get_device(VFIOGroup *group, const char *name,
 882                        VFIODevice *vbasedev)
 883 {
 884     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
 885     int ret, fd;
 886
 887     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
 888     if (fd < 0) {
 889         error_report("vfio: error getting device %s from group %d: %m",
 890                      name, group->groupid);
 891         error_printf("Verify all devices in group %d are bound to vfio-<bus> "
 892                      "or pci-stub and not already in use\n", group->groupid);
 893         return fd;
 894     }
 895
 896     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
 897     if (ret) {
 898         error_report("vfio: error getting device info: %m");
 899         close(fd);
 900         return ret;
 901     }
 902
 903     vbasedev->fd = fd;
 904     vbasedev->group = group;
 905     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
 906
 907     vbasedev->num_irqs = dev_info.num_irqs;
 908     vbasedev->num_regions = dev_info.num_regions;
 909     vbasedev->flags = dev_info.flags;
 910
 911     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
 912                           dev_info.num_irqs);
 913
 914     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
 915     return 0;
 916 }
 917
 918 void vfio_put_base_device(VFIODevice *vbasedev)
 919 {
 920     if (!vbasedev->group) {
 921         return;
 922     }
 923     QLIST_REMOVE(vbasedev, next);
 924     vbasedev->group = NULL;
 925     trace_vfio_put_base_device(vbasedev->fd);
 926     close(vbasedev->fd);
 927 }
 928
 929 static int vfio_container_do_ioctl(AddressSpace *as, int32_t groupid,
 930                                    int req, void *param)
 931 {
 932     VFIOGroup *group;
 933     VFIOContainer *container;
 934     int ret = -1;
 935
 936     group = vfio_get_group(groupid, as);
 937     if (!group) {
 938         error_report("vfio: group %d not registered", groupid);
 939         return ret;
 940     }
 941
 942     container = group->container;
 943     if (group->container) {
 944         ret = ioctl(container->fd, req, param);
 945         if (ret < 0) {
 946             error_report("vfio: failed to ioctl %d to container: ret=%d, %s",
 947                          _IOC_NR(req) - VFIO_BASE, ret, strerror(errno));
 948         }
 949     }
 950
 951     vfio_put_group(group);
 952
 953     return ret;
 954 }
 955
 956 int vfio_container_ioctl(AddressSpace *as, int32_t groupid,
 957                          int req, void *param)
 958 {
 959     /* We allow only certain ioctls to the container */
 960     switch (req) {
 961     case VFIO_CHECK_EXTENSION:
 962     case VFIO_IOMMU_SPAPR_TCE_GET_INFO:
 963     case VFIO_EEH_PE_OP:
 964         break;
 965     default:
 966         /* Return an error on unknown requests */
 967         error_report("vfio: unsupported ioctl %X", req);
 968         return -1;
 969     }
 970
 971     return vfio_container_do_ioctl(as, groupid, req, param);
 972 }