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, listener);
 319     hwaddr iova, end;
 320     Int128 llend;
 321     void *vaddr;
 322     int ret;
 323
 324     if (vfio_listener_skipped_section(section)) {
 325         trace_vfio_listener_region_add_skip(
 326                 section->offset_within_address_space,
 327                 section->offset_within_address_space +
 328                 int128_get64(int128_sub(section->size, int128_one())));
 329         return;
 330     }
 331
 332     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 333                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 334         error_report("%s received unaligned region", __func__);
 335         return;
 336     }
 337
 338     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 339     llend = int128_make64(section->offset_within_address_space);
 340     llend = int128_add(llend, section->size);
 341     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 342
 343     if (int128_ge(int128_make64(iova), llend)) {
 344         return;
 345     }
 346
 347     memory_region_ref(section->mr);
 348
 349     if (memory_region_is_iommu(section->mr)) {
 350         VFIOGuestIOMMU *giommu;
 351
 352         trace_vfio_listener_region_add_iommu(iova,
 353                     int128_get64(int128_sub(llend, int128_one())));
 354         /*
 355          * FIXME: We should do some checking to see if the
 356          * capabilities of the host VFIO IOMMU are adequate to model
 357          * the guest IOMMU
 358          *
 359          * FIXME: For VFIO iommu types which have KVM acceleration to
 360          * avoid bouncing all map/unmaps through qemu this way, this
 361          * would be the right place to wire that up (tell the KVM
 362          * device emulation the VFIO iommu handles to use).
 363          */
 364         /*
 365          * This assumes that the guest IOMMU is empty of
 366          * mappings at this point.
 367          *
 368          * One way of doing this is:
 369          * 1. Avoid sharing IOMMUs between emulated devices or different
 370          * IOMMU groups.
 371          * 2. Implement VFIO_IOMMU_ENABLE in the host kernel to fail if
 372          * there are some mappings in IOMMU.
 373          *
 374          * VFIO on SPAPR does that. Other IOMMU models may do that different,
 375          * they must make sure there are no existing mappings or
 376          * loop through existing mappings to map them into VFIO.
 377          */
 378         giommu = g_malloc0(sizeof(*giommu));
 379         giommu->iommu = section->mr;
 380         giommu->container = container;
 381         giommu->n.notify = vfio_iommu_map_notify;
 382         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
 383         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
 384
 385         return;
 386     }
 387
 388     /* Here we assume that memory_region_is_ram(section->mr)==true */
 389
 390     end = int128_get64(llend);
 391     vaddr = memory_region_get_ram_ptr(section->mr) +
 392             section->offset_within_region +
 393             (iova - section->offset_within_address_space);
 394
 395     trace_vfio_listener_region_add_ram(iova, end - 1, vaddr);
 396
 397     ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
 398     if (ret) {
 399         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 400                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
 401                      container, iova, end - iova, vaddr, ret);
 402         goto fail;
 403     }
 404
 405     return;
 406
 407 fail:
 408     /*
 409      * On the initfn path, store the first error in the container so we
 410      * can gracefully fail.  Runtime, there's not much we can do other
 411      * than throw a hardware error.
 412      */
 413     if (!container->initialized) {
 414         if (!container->error) {
 415             container->error = ret;
 416         }
 417     } else {
 418         hw_error("vfio: DMA mapping failed, unable to continue");
 419     }
 420 }
 421
 422 static void vfio_listener_region_del(MemoryListener *listener,
 423                                      MemoryRegionSection *section)
 424 {
 425     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
 426     hwaddr iova, end;
 427     int ret;
 428
 429     if (vfio_listener_skipped_section(section)) {
 430         trace_vfio_listener_region_del_skip(
 431                 section->offset_within_address_space,
 432                 section->offset_within_address_space +
 433                 int128_get64(int128_sub(section->size, int128_one())));
 434         return;
 435     }
 436
 437     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 438                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 439         error_report("%s received unaligned region", __func__);
 440         return;
 441     }
 442
 443     if (memory_region_is_iommu(section->mr)) {
 444         VFIOGuestIOMMU *giommu;
 445
 446         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
 447             if (giommu->iommu == section->mr) {
 448                 memory_region_unregister_iommu_notifier(&giommu->n);
 449                 QLIST_REMOVE(giommu, giommu_next);
 450                 g_free(giommu);
 451                 break;
 452             }
 453         }
 454
 455         /*
 456          * FIXME: We assume the one big unmap below is adequate to
 457          * remove any individual page mappings in the IOMMU which
 458          * might have been copied into VFIO. This works for a page table
 459          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
 460          * That may not be true for all IOMMU types.
 461          */
 462     }
 463
 464     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 465     end = (section->offset_within_address_space + int128_get64(section->size)) &
 466           TARGET_PAGE_MASK;
 467
 468     if (iova >= end) {
 469         return;
 470     }
 471
 472     trace_vfio_listener_region_del(iova, end - 1);
 473
 474     ret = vfio_dma_unmap(container, iova, end - iova);
 475     memory_region_unref(section->mr);
 476     if (ret) {
 477         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 478                      "0x%"HWADDR_PRIx") = %d (%m)",
 479                      container, iova, end - iova, ret);
 480     }
 481 }
 482
 483 static const MemoryListener vfio_memory_listener = {
 484     .region_add = vfio_listener_region_add,
 485     .region_del = vfio_listener_region_del,
 486 };
 487
 488 static void vfio_listener_release(VFIOContainer *container)
 489 {
 490     memory_listener_unregister(&container->listener);
 491 }
 492
 493 int vfio_mmap_region(Object *obj, VFIORegion *region,
 494                      MemoryRegion *mem, MemoryRegion *submem,
 495                      void **map, size_t size, off_t offset,
 496                      const char *name)
 497 {
 498     int ret = 0;
 499     VFIODevice *vbasedev = region->vbasedev;
 500
 501     if (!vbasedev->no_mmap && size && region->flags &
 502         VFIO_REGION_INFO_FLAG_MMAP) {
 503         int prot = 0;
 504
 505         if (region->flags & VFIO_REGION_INFO_FLAG_READ) {
 506             prot |= PROT_READ;
 507         }
 508
 509         if (region->flags & VFIO_REGION_INFO_FLAG_WRITE) {
 510             prot |= PROT_WRITE;
 511         }
 512
 513         *map = mmap(NULL, size, prot, MAP_SHARED,
 514                     vbasedev->fd,
 515                     region->fd_offset + offset);
 516         if (*map == MAP_FAILED) {
 517             *map = NULL;
 518             ret = -errno;
 519             goto empty_region;
 520         }
 521
 522         memory_region_init_ram_ptr(submem, obj, name, size, *map);
 523         memory_region_set_skip_dump(submem);
 524     } else {
 525 empty_region:
 526         /* Create a zero sized sub-region to make cleanup easy. */
 527         memory_region_init(submem, obj, name, 0);
 528     }
 529
 530     memory_region_add_subregion(mem, offset, submem);
 531
 532     return ret;
 533 }
 534
 535 void vfio_reset_handler(void *opaque)
 536 {
 537     VFIOGroup *group;
 538     VFIODevice *vbasedev;
 539
 540     QLIST_FOREACH(group, &vfio_group_list, next) {
 541         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 542             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
 543         }
 544     }
 545
 546     QLIST_FOREACH(group, &vfio_group_list, next) {
 547         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 548             if (vbasedev->needs_reset) {
 549                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
 550             }
 551         }
 552     }
 553 }
 554
 555 static void vfio_kvm_device_add_group(VFIOGroup *group)
 556 {
 557 #ifdef CONFIG_KVM
 558     struct kvm_device_attr attr = {
 559         .group = KVM_DEV_VFIO_GROUP,
 560         .attr = KVM_DEV_VFIO_GROUP_ADD,
 561         .addr = (uint64_t)(unsigned long)&group->fd,
 562     };
 563
 564     if (!kvm_enabled()) {
 565         return;
 566     }
 567
 568     if (vfio_kvm_device_fd < 0) {
 569         struct kvm_create_device cd = {
 570             .type = KVM_DEV_TYPE_VFIO,
 571         };
 572
 573         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
 574             error_report("Failed to create KVM VFIO device: %m");
 575             return;
 576         }
 577
 578         vfio_kvm_device_fd = cd.fd;
 579     }
 580
 581     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 582         error_report("Failed to add group %d to KVM VFIO device: %m",
 583                      group->groupid);
 584     }
 585 #endif
 586 }
 587
 588 static void vfio_kvm_device_del_group(VFIOGroup *group)
 589 {
 590 #ifdef CONFIG_KVM
 591     struct kvm_device_attr attr = {
 592         .group = KVM_DEV_VFIO_GROUP,
 593         .attr = KVM_DEV_VFIO_GROUP_DEL,
 594         .addr = (uint64_t)(unsigned long)&group->fd,
 595     };
 596
 597     if (vfio_kvm_device_fd < 0) {
 598         return;
 599     }
 600
 601     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 602         error_report("Failed to remove group %d from KVM VFIO device: %m",
 603                      group->groupid);
 604     }
 605 #endif
 606 }
 607
 608 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
 609 {
 610     VFIOAddressSpace *space;
 611
 612     QLIST_FOREACH(space, &vfio_address_spaces, list) {
 613         if (space->as == as) {
 614             return space;
 615         }
 616     }
 617
 618     /* No suitable VFIOAddressSpace, create a new one */
 619     space = g_malloc0(sizeof(*space));
 620     space->as = as;
 621     QLIST_INIT(&space->containers);
 622
 623     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
 624
 625     return space;
 626 }
 627
 628 static void vfio_put_address_space(VFIOAddressSpace *space)
 629 {
 630     if (QLIST_EMPTY(&space->containers)) {
 631         QLIST_REMOVE(space, list);
 632         g_free(space);
 633     }
 634 }
 635
 636 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
 637 {
 638     VFIOContainer *container;
 639     int ret, fd;
 640     VFIOAddressSpace *space;
 641
 642     space = vfio_get_address_space(as);
 643
 644     QLIST_FOREACH(container, &space->containers, next) {
 645         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
 646             group->container = container;
 647             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 648             return 0;
 649         }
 650     }
 651
 652     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
 653     if (fd < 0) {
 654         error_report("vfio: failed to open /dev/vfio/vfio: %m");
 655         ret = -errno;
 656         goto put_space_exit;
 657     }
 658
 659     ret = ioctl(fd, VFIO_GET_API_VERSION);
 660     if (ret != VFIO_API_VERSION) {
 661         error_report("vfio: supported vfio version: %d, "
 662                      "reported version: %d", VFIO_API_VERSION, ret);
 663         ret = -EINVAL;
 664         goto close_fd_exit;
 665     }
 666
 667     container = g_malloc0(sizeof(*container));
 668     container->space = space;
 669     container->fd = fd;
 670     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
 671         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
 672         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
 673
 674         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 675         if (ret) {
 676             error_report("vfio: failed to set group container: %m");
 677             ret = -errno;
 678             goto free_container_exit;
 679         }
 680
 681         ret = ioctl(fd, VFIO_SET_IOMMU,
 682                     v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU);
 683         if (ret) {
 684             error_report("vfio: failed to set iommu for container: %m");
 685             ret = -errno;
 686             goto free_container_exit;
 687         }
 688     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) {
 689         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 690         if (ret) {
 691             error_report("vfio: failed to set group container: %m");
 692             ret = -errno;
 693             goto free_container_exit;
 694         }
 695         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU);
 696         if (ret) {
 697             error_report("vfio: failed to set iommu for container: %m");
 698             ret = -errno;
 699             goto free_container_exit;
 700         }
 701
 702         /*
 703          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
 704          * when container fd is closed so we do not call it explicitly
 705          * in this file.
 706          */
 707         ret = ioctl(fd, VFIO_IOMMU_ENABLE);
 708         if (ret) {
 709             error_report("vfio: failed to enable container: %m");
 710             ret = -errno;
 711             goto free_container_exit;
 712         }
 713     } else {
 714         error_report("vfio: No available IOMMU models");
 715         ret = -EINVAL;
 716         goto free_container_exit;
 717     }
 718
 719     container->listener = vfio_memory_listener;
 720
 721     memory_listener_register(&container->listener, container->space->as);
 722
 723     if (container->error) {
 724         ret = container->error;
 725         error_report("vfio: memory listener initialization failed for container");
 726         goto listener_release_exit;
 727     }
 728
 729     container->initialized = true;
 730
 731     QLIST_INIT(&container->group_list);
 732     QLIST_INSERT_HEAD(&space->containers, container, next);
 733
 734     group->container = container;
 735     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 736
 737     return 0;
 738 listener_release_exit:
 739     vfio_listener_release(container);
 740
 741 free_container_exit:
 742     g_free(container);
 743
 744 close_fd_exit:
 745     close(fd);
 746
 747 put_space_exit:
 748     vfio_put_address_space(space);
 749
 750     return ret;
 751 }
 752
 753 static void vfio_disconnect_container(VFIOGroup *group)
 754 {
 755     VFIOContainer *container = group->container;
 756
 757     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
 758         error_report("vfio: error disconnecting group %d from container",
 759                      group->groupid);
 760     }
 761
 762     QLIST_REMOVE(group, container_next);
 763     group->container = NULL;
 764
 765     if (QLIST_EMPTY(&container->group_list)) {
 766         VFIOAddressSpace *space = container->space;
 767         VFIOGuestIOMMU *giommu, *tmp;
 768
 769         vfio_listener_release(container);
 770         QLIST_REMOVE(container, next);
 771
 772         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
 773             memory_region_unregister_iommu_notifier(&giommu->n);
 774             QLIST_REMOVE(giommu, giommu_next);
 775             g_free(giommu);
 776         }
 777
 778         trace_vfio_disconnect_container(container->fd);
 779         close(container->fd);
 780         g_free(container);
 781
 782         vfio_put_address_space(space);
 783     }
 784 }
 785
 786 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
 787 {
 788     VFIOGroup *group;
 789     char path[32];
 790     struct vfio_group_status status = { .argsz = sizeof(status) };
 791
 792     QLIST_FOREACH(group, &vfio_group_list, next) {
 793         if (group->groupid == groupid) {
 794             /* Found it.  Now is it already in the right context? */
 795             if (group->container->space->as == as) {
 796                 return group;
 797             } else {
 798                 error_report("vfio: group %d used in multiple address spaces",
 799                              group->groupid);
 800                 return NULL;
 801             }
 802         }
 803     }
 804
 805     group = g_malloc0(sizeof(*group));
 806
 807     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
 808     group->fd = qemu_open(path, O_RDWR);
 809     if (group->fd < 0) {
 810         error_report("vfio: error opening %s: %m", path);
 811         goto free_group_exit;
 812     }
 813
 814     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
 815         error_report("vfio: error getting group status: %m");
 816         goto close_fd_exit;
 817     }
 818
 819     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
 820         error_report("vfio: error, group %d is not viable, please ensure "
 821                      "all devices within the iommu_group are bound to their "
 822                      "vfio bus driver.", groupid);
 823         goto close_fd_exit;
 824     }
 825
 826     group->groupid = groupid;
 827     QLIST_INIT(&group->device_list);
 828
 829     if (vfio_connect_container(group, as)) {
 830         error_report("vfio: failed to setup container for group %d", groupid);
 831         goto close_fd_exit;
 832     }
 833
 834     if (QLIST_EMPTY(&vfio_group_list)) {
 835         qemu_register_reset(vfio_reset_handler, NULL);
 836     }
 837
 838     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
 839
 840     vfio_kvm_device_add_group(group);
 841
 842     return group;
 843
 844 close_fd_exit:
 845     close(group->fd);
 846
 847 free_group_exit:
 848     g_free(group);
 849
 850     return NULL;
 851 }
 852
 853 void vfio_put_group(VFIOGroup *group)
 854 {
 855     if (!group || !QLIST_EMPTY(&group->device_list)) {
 856         return;
 857     }
 858
 859     vfio_kvm_device_del_group(group);
 860     vfio_disconnect_container(group);
 861     QLIST_REMOVE(group, next);
 862     trace_vfio_put_group(group->fd);
 863     close(group->fd);
 864     g_free(group);
 865
 866     if (QLIST_EMPTY(&vfio_group_list)) {
 867         qemu_unregister_reset(vfio_reset_handler, NULL);
 868     }
 869 }
 870
 871 int vfio_get_device(VFIOGroup *group, const char *name,
 872                        VFIODevice *vbasedev)
 873 {
 874     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
 875     int ret, fd;
 876
 877     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
 878     if (fd < 0) {
 879         error_report("vfio: error getting device %s from group %d: %m",
 880                      name, group->groupid);
 881         error_printf("Verify all devices in group %d are bound to vfio-<bus> "
 882                      "or pci-stub and not already in use\n", group->groupid);
 883         return fd;
 884     }
 885
 886     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
 887     if (ret) {
 888         error_report("vfio: error getting device info: %m");
 889         close(fd);
 890         return ret;
 891     }
 892
 893     vbasedev->fd = fd;
 894     vbasedev->group = group;
 895     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
 896
 897     vbasedev->num_irqs = dev_info.num_irqs;
 898     vbasedev->num_regions = dev_info.num_regions;
 899     vbasedev->flags = dev_info.flags;
 900
 901     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
 902                           dev_info.num_irqs);
 903
 904     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
 905     return 0;
 906 }
 907
 908 void vfio_put_base_device(VFIODevice *vbasedev)
 909 {
 910     if (!vbasedev->group) {
 911         return;
 912     }
 913     QLIST_REMOVE(vbasedev, next);
 914     vbasedev->group = NULL;
 915     trace_vfio_put_base_device(vbasedev->fd);
 916     close(vbasedev->fd);
 917 }
 918
 919 static int vfio_container_do_ioctl(AddressSpace *as, int32_t groupid,
 920                                    int req, void *param)
 921 {
 922     VFIOGroup *group;
 923     VFIOContainer *container;
 924     int ret = -1;
 925
 926     group = vfio_get_group(groupid, as);
 927     if (!group) {
 928         error_report("vfio: group %d not registered", groupid);
 929         return ret;
 930     }
 931
 932     container = group->container;
 933     if (group->container) {
 934         ret = ioctl(container->fd, req, param);
 935         if (ret < 0) {
 936             error_report("vfio: failed to ioctl %d to container: ret=%d, %s",
 937                          _IOC_NR(req) - VFIO_BASE, ret, strerror(errno));
 938         }
 939     }
 940
 941     vfio_put_group(group);
 942
 943     return ret;
 944 }
 945
 946 int vfio_container_ioctl(AddressSpace *as, int32_t groupid,
 947                          int req, void *param)
 948 {
 949     /* We allow only certain ioctls to the container */
 950     switch (req) {
 951     case VFIO_CHECK_EXTENSION:
 952     case VFIO_IOMMU_SPAPR_TCE_GET_INFO:
 953     case VFIO_EEH_PE_OP:
 954         break;
 955     default:
 956         /* Return an error on unknown requests */
 957         error_report("vfio: unsupported ioctl %X", req);
 958         return -1;
 959     }
 960
 961     return vfio_container_do_ioctl(as, groupid, req, param);
 962 }