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 "qemu/osdep.h"
  22 #include <sys/ioctl.h>
  23 #include <sys/mman.h>
  24 #include <linux/vfio.h>
  25
  26 #include "hw/vfio/vfio-common.h"
  27 #include "hw/vfio/vfio.h"
  28 #include "exec/address-spaces.h"
  29 #include "exec/memory.h"
  30 #include "hw/hw.h"
  31 #include "qemu/error-report.h"
  32 #include "sysemu/kvm.h"
  33 #ifdef CONFIG_KVM
  34 #include "linux/kvm.h"
  35 #endif
  36 #include "trace.h"
  37
  38 struct vfio_group_head vfio_group_list =
  39     QLIST_HEAD_INITIALIZER(vfio_group_list);
  40 struct vfio_as_head vfio_address_spaces =
  41     QLIST_HEAD_INITIALIZER(vfio_address_spaces);
  42
  43 #ifdef CONFIG_KVM
  44 /*
  45  * We have a single VFIO pseudo device per KVM VM.  Once created it lives
  46  * for the life of the VM.  Closing the file descriptor only drops our
  47  * reference to it and the device's reference to kvm.  Therefore once
  48  * initialized, this file descriptor is only released on QEMU exit and
  49  * we'll re-use it should another vfio device be attached before then.
  50  */
  51 static int vfio_kvm_device_fd = -1;
  52 #endif
  53
  54 /*
  55  * Common VFIO interrupt disable
  56  */
  57 void vfio_disable_irqindex(VFIODevice *vbasedev, int index)
  58 {
  59     struct vfio_irq_set irq_set = {
  60         .argsz = sizeof(irq_set),
  61         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
  62         .index = index,
  63         .start = 0,
  64         .count = 0,
  65     };
  66
  67     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  68 }
  69
  70 void vfio_unmask_single_irqindex(VFIODevice *vbasedev, int index)
  71 {
  72     struct vfio_irq_set irq_set = {
  73         .argsz = sizeof(irq_set),
  74         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
  75         .index = index,
  76         .start = 0,
  77         .count = 1,
  78     };
  79
  80     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  81 }
  82
  83 void vfio_mask_single_irqindex(VFIODevice *vbasedev, int index)
  84 {
  85     struct vfio_irq_set irq_set = {
  86         .argsz = sizeof(irq_set),
  87         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK,
  88         .index = index,
  89         .start = 0,
  90         .count = 1,
  91     };
  92
  93     ioctl(vbasedev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
  94 }
  95
  96 /*
  97  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
  98  */
  99 void vfio_region_write(void *opaque, hwaddr addr,
 100                        uint64_t data, unsigned size)
 101 {
 102     VFIORegion *region = opaque;
 103     VFIODevice *vbasedev = region->vbasedev;
 104     union {
 105         uint8_t byte;
 106         uint16_t word;
 107         uint32_t dword;
 108         uint64_t qword;
 109     } buf;
 110
 111     switch (size) {
 112     case 1:
 113         buf.byte = data;
 114         break;
 115     case 2:
 116         buf.word = cpu_to_le16(data);
 117         break;
 118     case 4:
 119         buf.dword = cpu_to_le32(data);
 120         break;
 121     default:
 122         hw_error("vfio: unsupported write size, %d bytes", size);
 123         break;
 124     }
 125
 126     if (pwrite(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 127         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", 0x%"PRIx64
 128                      ",%d) failed: %m",
 129                      __func__, vbasedev->name, region->nr,
 130                      addr, data, size);
 131     }
 132
 133     trace_vfio_region_write(vbasedev->name, region->nr, addr, data, size);
 134
 135     /*
 136      * A read or write to a BAR always signals an INTx EOI.  This will
 137      * do nothing if not pending (including not in INTx mode).  We assume
 138      * that a BAR access is in response to an interrupt and that BAR
 139      * accesses will service the interrupt.  Unfortunately, we don't know
 140      * which access will service the interrupt, so we're potentially
 141      * getting quite a few host interrupts per guest interrupt.
 142      */
 143     vbasedev->ops->vfio_eoi(vbasedev);
 144 }
 145
 146 uint64_t vfio_region_read(void *opaque,
 147                           hwaddr addr, unsigned size)
 148 {
 149     VFIORegion *region = opaque;
 150     VFIODevice *vbasedev = region->vbasedev;
 151     union {
 152         uint8_t byte;
 153         uint16_t word;
 154         uint32_t dword;
 155         uint64_t qword;
 156     } buf;
 157     uint64_t data = 0;
 158
 159     if (pread(vbasedev->fd, &buf, size, region->fd_offset + addr) != size) {
 160         error_report("%s(%s:region%d+0x%"HWADDR_PRIx", %d) failed: %m",
 161                      __func__, vbasedev->name, region->nr,
 162                      addr, size);
 163         return (uint64_t)-1;
 164     }
 165     switch (size) {
 166     case 1:
 167         data = buf.byte;
 168         break;
 169     case 2:
 170         data = le16_to_cpu(buf.word);
 171         break;
 172     case 4:
 173         data = le32_to_cpu(buf.dword);
 174         break;
 175     default:
 176         hw_error("vfio: unsupported read size, %d bytes", size);
 177         break;
 178     }
 179
 180     trace_vfio_region_read(vbasedev->name, region->nr, addr, size, data);
 181
 182     /* Same as write above */
 183     vbasedev->ops->vfio_eoi(vbasedev);
 184
 185     return data;
 186 }
 187
 188 const MemoryRegionOps vfio_region_ops = {
 189     .read = vfio_region_read,
 190     .write = vfio_region_write,
 191     .endianness = DEVICE_LITTLE_ENDIAN,
 192 };
 193
 194 /*
 195  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
 196  */
 197 static int vfio_dma_unmap(VFIOContainer *container,
 198                           hwaddr iova, ram_addr_t size)
 199 {
 200     struct vfio_iommu_type1_dma_unmap unmap = {
 201         .argsz = sizeof(unmap),
 202         .flags = 0,
 203         .iova = iova,
 204         .size = size,
 205     };
 206
 207     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
 208         error_report("VFIO_UNMAP_DMA: %d", -errno);
 209         return -errno;
 210     }
 211
 212     return 0;
 213 }
 214
 215 static int vfio_dma_map(VFIOContainer *container, hwaddr iova,
 216                         ram_addr_t size, void *vaddr, bool readonly)
 217 {
 218     struct vfio_iommu_type1_dma_map map = {
 219         .argsz = sizeof(map),
 220         .flags = VFIO_DMA_MAP_FLAG_READ,
 221         .vaddr = (__u64)(uintptr_t)vaddr,
 222         .iova = iova,
 223         .size = size,
 224     };
 225
 226     if (!readonly) {
 227         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
 228     }
 229
 230     /*
 231      * Try the mapping, if it fails with EBUSY, unmap the region and try
 232      * again.  This shouldn't be necessary, but we sometimes see it in
 233      * the VGA ROM space.
 234      */
 235     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
 236         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
 237          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
 238         return 0;
 239     }
 240
 241     error_report("VFIO_MAP_DMA: %d", -errno);
 242     return -errno;
 243 }
 244
 245 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 246 {
 247     return (!memory_region_is_ram(section->mr) &&
 248             !memory_region_is_iommu(section->mr)) ||
 249            /*
 250             * Sizing an enabled 64-bit BAR can cause spurious mappings to
 251             * addresses in the upper part of the 64-bit address space.  These
 252             * are never accessed by the CPU and beyond the address width of
 253             * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
 254             */
 255            section->offset_within_address_space & (1ULL << 63);
 256 }
 257
 258 static void vfio_iommu_map_notify(Notifier *n, void *data)
 259 {
 260     VFIOGuestIOMMU *giommu = container_of(n, VFIOGuestIOMMU, n);
 261     VFIOContainer *container = giommu->container;
 262     IOMMUTLBEntry *iotlb = data;
 263     MemoryRegion *mr;
 264     hwaddr xlat;
 265     hwaddr len = iotlb->addr_mask + 1;
 266     void *vaddr;
 267     int ret;
 268
 269     trace_vfio_iommu_map_notify(iotlb->iova,
 270                                 iotlb->iova + iotlb->addr_mask);
 271
 272     /*
 273      * The IOMMU TLB entry we have just covers translation through
 274      * this IOMMU to its immediate target.  We need to translate
 275      * it the rest of the way through to memory.
 276      */
 277     rcu_read_lock();
 278     mr = address_space_translate(&address_space_memory,
 279                                  iotlb->translated_addr,
 280                                  &xlat, &len, iotlb->perm & IOMMU_WO);
 281     if (!memory_region_is_ram(mr)) {
 282         error_report("iommu map to non memory area %"HWADDR_PRIx"",
 283                      xlat);
 284         goto out;
 285     }
 286     /*
 287      * Translation truncates length to the IOMMU page size,
 288      * check that it did not truncate too much.
 289      */
 290     if (len & iotlb->addr_mask) {
 291         error_report("iommu has granularity incompatible with target AS");
 292         goto out;
 293     }
 294
 295     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
 296         vaddr = memory_region_get_ram_ptr(mr) + xlat;
 297         ret = vfio_dma_map(container, iotlb->iova,
 298                            iotlb->addr_mask + 1, vaddr,
 299                            !(iotlb->perm & IOMMU_WO) || mr->readonly);
 300         if (ret) {
 301             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 302                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
 303                          container, iotlb->iova,
 304                          iotlb->addr_mask + 1, vaddr, ret);
 305         }
 306     } else {
 307         ret = vfio_dma_unmap(container, iotlb->iova, iotlb->addr_mask + 1);
 308         if (ret) {
 309             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 310                          "0x%"HWADDR_PRIx") = %d (%m)",
 311                          container, iotlb->iova,
 312                          iotlb->addr_mask + 1, ret);
 313         }
 314     }
 315 out:
 316     rcu_read_unlock();
 317 }
 318
 319 static hwaddr vfio_container_granularity(VFIOContainer *container)
 320 {
 321     return (hwaddr)1 << ctz64(container->iova_pgsizes);
 322 }
 323
 324 static void vfio_listener_region_add(MemoryListener *listener,
 325                                      MemoryRegionSection *section)
 326 {
 327     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
 328     hwaddr iova, end;
 329     Int128 llend, llsize;
 330     void *vaddr;
 331     int ret;
 332
 333     if (vfio_listener_skipped_section(section)) {
 334         trace_vfio_listener_region_add_skip(
 335                 section->offset_within_address_space,
 336                 section->offset_within_address_space +
 337                 int128_get64(int128_sub(section->size, int128_one())));
 338         return;
 339     }
 340
 341     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 342                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 343         error_report("%s received unaligned region", __func__);
 344         return;
 345     }
 346
 347     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 348     llend = int128_make64(section->offset_within_address_space);
 349     llend = int128_add(llend, section->size);
 350     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 351
 352     if (int128_ge(int128_make64(iova), llend)) {
 353         return;
 354     }
 355     end = int128_get64(int128_sub(llend, int128_one()));
 356
 357     if ((iova < container->min_iova) || (end > container->max_iova)) {
 358         error_report("vfio: IOMMU container %p can't map guest IOVA region"
 359                      " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
 360                      container, iova, end);
 361         ret = -EFAULT;
 362         goto fail;
 363     }
 364
 365     memory_region_ref(section->mr);
 366
 367     if (memory_region_is_iommu(section->mr)) {
 368         VFIOGuestIOMMU *giommu;
 369
 370         trace_vfio_listener_region_add_iommu(iova, end);
 371         /*
 372          * FIXME: We should do some checking to see if the
 373          * capabilities of the host VFIO IOMMU are adequate to model
 374          * the guest IOMMU
 375          *
 376          * FIXME: For VFIO iommu types which have KVM acceleration to
 377          * avoid bouncing all map/unmaps through qemu this way, this
 378          * would be the right place to wire that up (tell the KVM
 379          * device emulation the VFIO iommu handles to use).
 380          */
 381         giommu = g_malloc0(sizeof(*giommu));
 382         giommu->iommu = section->mr;
 383         giommu->container = container;
 384         giommu->n.notify = vfio_iommu_map_notify;
 385         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
 386
 387         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
 388         memory_region_iommu_replay(giommu->iommu, &giommu->n,
 389                                    vfio_container_granularity(container),
 390                                    false);
 391
 392         return;
 393     }
 394
 395     /* Here we assume that memory_region_is_ram(section->mr)==true */
 396
 397     vaddr = memory_region_get_ram_ptr(section->mr) +
 398             section->offset_within_region +
 399             (iova - section->offset_within_address_space);
 400
 401     trace_vfio_listener_region_add_ram(iova, end, vaddr);
 402
 403     llsize = int128_sub(llend, int128_make64(iova));
 404
 405     ret = vfio_dma_map(container, iova, int128_get64(llsize),
 406                        vaddr, section->readonly);
 407     if (ret) {
 408         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 409                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
 410                      container, iova, int128_get64(llsize), vaddr, ret);
 411         goto fail;
 412     }
 413
 414     return;
 415
 416 fail:
 417     /*
 418      * On the initfn path, store the first error in the container so we
 419      * can gracefully fail.  Runtime, there's not much we can do other
 420      * than throw a hardware error.
 421      */
 422     if (!container->initialized) {
 423         if (!container->error) {
 424             container->error = ret;
 425         }
 426     } else {
 427         hw_error("vfio: DMA mapping failed, unable to continue");
 428     }
 429 }
 430
 431 static void vfio_listener_region_del(MemoryListener *listener,
 432                                      MemoryRegionSection *section)
 433 {
 434     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
 435     hwaddr iova, end;
 436     int ret;
 437
 438     if (vfio_listener_skipped_section(section)) {
 439         trace_vfio_listener_region_del_skip(
 440                 section->offset_within_address_space,
 441                 section->offset_within_address_space +
 442                 int128_get64(int128_sub(section->size, int128_one())));
 443         return;
 444     }
 445
 446     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 447                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 448         error_report("%s received unaligned region", __func__);
 449         return;
 450     }
 451
 452     if (memory_region_is_iommu(section->mr)) {
 453         VFIOGuestIOMMU *giommu;
 454
 455         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
 456             if (giommu->iommu == section->mr) {
 457                 memory_region_unregister_iommu_notifier(&giommu->n);
 458                 QLIST_REMOVE(giommu, giommu_next);
 459                 g_free(giommu);
 460                 break;
 461             }
 462         }
 463
 464         /*
 465          * FIXME: We assume the one big unmap below is adequate to
 466          * remove any individual page mappings in the IOMMU which
 467          * might have been copied into VFIO. This works for a page table
 468          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
 469          * That may not be true for all IOMMU types.
 470          */
 471     }
 472
 473     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 474     end = (section->offset_within_address_space + int128_get64(section->size)) &
 475           TARGET_PAGE_MASK;
 476
 477     if (iova >= end) {
 478         return;
 479     }
 480
 481     trace_vfio_listener_region_del(iova, end - 1);
 482
 483     ret = vfio_dma_unmap(container, iova, end - iova);
 484     memory_region_unref(section->mr);
 485     if (ret) {
 486         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 487                      "0x%"HWADDR_PRIx") = %d (%m)",
 488                      container, iova, end - iova, ret);
 489     }
 490 }
 491
 492 static const MemoryListener vfio_memory_listener = {
 493     .region_add = vfio_listener_region_add,
 494     .region_del = vfio_listener_region_del,
 495 };
 496
 497 static void vfio_listener_release(VFIOContainer *container)
 498 {
 499     memory_listener_unregister(&container->listener);
 500 }
 501
 502 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
 503                       int index, const char *name)
 504 {
 505     struct vfio_region_info *info;
 506     int ret;
 507
 508     ret = vfio_get_region_info(vbasedev, index, &info);
 509     if (ret) {
 510         return ret;
 511     }
 512
 513     region->vbasedev = vbasedev;
 514     region->flags = info->flags;
 515     region->size = info->size;
 516     region->fd_offset = info->offset;
 517     region->nr = index;
 518
 519     if (region->size) {
 520         region->mem = g_new0(MemoryRegion, 1);
 521         memory_region_init_io(region->mem, obj, &vfio_region_ops,
 522                               region, name, region->size);
 523
 524         if (!vbasedev->no_mmap &&
 525             region->flags & VFIO_REGION_INFO_FLAG_MMAP &&
 526             !(region->size & ~qemu_real_host_page_mask)) {
 527
 528             region->nr_mmaps = 1;
 529             region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
 530
 531             region->mmaps[0].offset = 0;
 532             region->mmaps[0].size = region->size;
 533         }
 534     }
 535
 536     g_free(info);
 537
 538     trace_vfio_region_setup(vbasedev->name, index, name,
 539                             region->flags, region->fd_offset, region->size);
 540     return 0;
 541 }
 542
 543 int vfio_region_mmap(VFIORegion *region)
 544 {
 545     int i, prot = 0;
 546     char *name;
 547
 548     if (!region->mem) {
 549         return 0;
 550     }
 551
 552     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
 553     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
 554
 555     for (i = 0; i < region->nr_mmaps; i++) {
 556         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
 557                                      MAP_SHARED, region->vbasedev->fd,
 558                                      region->fd_offset +
 559                                      region->mmaps[i].offset);
 560         if (region->mmaps[i].mmap == MAP_FAILED) {
 561             int ret = -errno;
 562
 563             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
 564                                          region->fd_offset +
 565                                          region->mmaps[i].offset,
 566                                          region->fd_offset +
 567                                          region->mmaps[i].offset +
 568                                          region->mmaps[i].size - 1, ret);
 569
 570             region->mmaps[i].mmap = NULL;
 571
 572             for (i--; i >= 0; i--) {
 573                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
 574                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
 575                 object_unparent(OBJECT(&region->mmaps[i].mem));
 576                 region->mmaps[i].mmap = NULL;
 577             }
 578
 579             return ret;
 580         }
 581
 582         name = g_strdup_printf("%s mmaps[%d]",
 583                                memory_region_name(region->mem), i);
 584         memory_region_init_ram_ptr(&region->mmaps[i].mem,
 585                                    memory_region_owner(region->mem),
 586                                    name, region->mmaps[i].size,
 587                                    region->mmaps[i].mmap);
 588         g_free(name);
 589         memory_region_set_skip_dump(&region->mmaps[i].mem);
 590         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
 591                                     &region->mmaps[i].mem);
 592
 593         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
 594                                region->mmaps[i].offset,
 595                                region->mmaps[i].offset +
 596                                region->mmaps[i].size - 1);
 597     }
 598
 599     return 0;
 600 }
 601
 602 void vfio_region_exit(VFIORegion *region)
 603 {
 604     int i;
 605
 606     if (!region->mem) {
 607         return;
 608     }
 609
 610     for (i = 0; i < region->nr_mmaps; i++) {
 611         if (region->mmaps[i].mmap) {
 612             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
 613         }
 614     }
 615
 616     trace_vfio_region_exit(region->vbasedev->name, region->nr);
 617 }
 618
 619 void vfio_region_finalize(VFIORegion *region)
 620 {
 621     int i;
 622
 623     if (!region->mem) {
 624         return;
 625     }
 626
 627     for (i = 0; i < region->nr_mmaps; i++) {
 628         if (region->mmaps[i].mmap) {
 629             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
 630             object_unparent(OBJECT(&region->mmaps[i].mem));
 631         }
 632     }
 633
 634     object_unparent(OBJECT(region->mem));
 635
 636     g_free(region->mem);
 637     g_free(region->mmaps);
 638
 639     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
 640 }
 641
 642 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
 643 {
 644     int i;
 645
 646     if (!region->mem) {
 647         return;
 648     }
 649
 650     for (i = 0; i < region->nr_mmaps; i++) {
 651         if (region->mmaps[i].mmap) {
 652             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
 653         }
 654     }
 655
 656     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
 657                                         enabled);
 658 }
 659
 660 void vfio_reset_handler(void *opaque)
 661 {
 662     VFIOGroup *group;
 663     VFIODevice *vbasedev;
 664
 665     QLIST_FOREACH(group, &vfio_group_list, next) {
 666         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 667             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
 668         }
 669     }
 670
 671     QLIST_FOREACH(group, &vfio_group_list, next) {
 672         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 673             if (vbasedev->needs_reset) {
 674                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
 675             }
 676         }
 677     }
 678 }
 679
 680 static void vfio_kvm_device_add_group(VFIOGroup *group)
 681 {
 682 #ifdef CONFIG_KVM
 683     struct kvm_device_attr attr = {
 684         .group = KVM_DEV_VFIO_GROUP,
 685         .attr = KVM_DEV_VFIO_GROUP_ADD,
 686         .addr = (uint64_t)(unsigned long)&group->fd,
 687     };
 688
 689     if (!kvm_enabled()) {
 690         return;
 691     }
 692
 693     if (vfio_kvm_device_fd < 0) {
 694         struct kvm_create_device cd = {
 695             .type = KVM_DEV_TYPE_VFIO,
 696         };
 697
 698         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
 699             error_report("Failed to create KVM VFIO device: %m");
 700             return;
 701         }
 702
 703         vfio_kvm_device_fd = cd.fd;
 704     }
 705
 706     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 707         error_report("Failed to add group %d to KVM VFIO device: %m",
 708                      group->groupid);
 709     }
 710 #endif
 711 }
 712
 713 static void vfio_kvm_device_del_group(VFIOGroup *group)
 714 {
 715 #ifdef CONFIG_KVM
 716     struct kvm_device_attr attr = {
 717         .group = KVM_DEV_VFIO_GROUP,
 718         .attr = KVM_DEV_VFIO_GROUP_DEL,
 719         .addr = (uint64_t)(unsigned long)&group->fd,
 720     };
 721
 722     if (vfio_kvm_device_fd < 0) {
 723         return;
 724     }
 725
 726     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 727         error_report("Failed to remove group %d from KVM VFIO device: %m",
 728                      group->groupid);
 729     }
 730 #endif
 731 }
 732
 733 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
 734 {
 735     VFIOAddressSpace *space;
 736
 737     QLIST_FOREACH(space, &vfio_address_spaces, list) {
 738         if (space->as == as) {
 739             return space;
 740         }
 741     }
 742
 743     /* No suitable VFIOAddressSpace, create a new one */
 744     space = g_malloc0(sizeof(*space));
 745     space->as = as;
 746     QLIST_INIT(&space->containers);
 747
 748     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
 749
 750     return space;
 751 }
 752
 753 static void vfio_put_address_space(VFIOAddressSpace *space)
 754 {
 755     if (QLIST_EMPTY(&space->containers)) {
 756         QLIST_REMOVE(space, list);
 757         g_free(space);
 758     }
 759 }
 760
 761 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
 762 {
 763     VFIOContainer *container;
 764     int ret, fd;
 765     VFIOAddressSpace *space;
 766
 767     space = vfio_get_address_space(as);
 768
 769     QLIST_FOREACH(container, &space->containers, next) {
 770         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
 771             group->container = container;
 772             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 773             return 0;
 774         }
 775     }
 776
 777     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
 778     if (fd < 0) {
 779         error_report("vfio: failed to open /dev/vfio/vfio: %m");
 780         ret = -errno;
 781         goto put_space_exit;
 782     }
 783
 784     ret = ioctl(fd, VFIO_GET_API_VERSION);
 785     if (ret != VFIO_API_VERSION) {
 786         error_report("vfio: supported vfio version: %d, "
 787                      "reported version: %d", VFIO_API_VERSION, ret);
 788         ret = -EINVAL;
 789         goto close_fd_exit;
 790     }
 791
 792     container = g_malloc0(sizeof(*container));
 793     container->space = space;
 794     container->fd = fd;
 795     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
 796         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
 797         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
 798         struct vfio_iommu_type1_info info;
 799
 800         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 801         if (ret) {
 802             error_report("vfio: failed to set group container: %m");
 803             ret = -errno;
 804             goto free_container_exit;
 805         }
 806
 807         ret = ioctl(fd, VFIO_SET_IOMMU,
 808                     v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU);
 809         if (ret) {
 810             error_report("vfio: failed to set iommu for container: %m");
 811             ret = -errno;
 812             goto free_container_exit;
 813         }
 814
 815         /*
 816          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
 817          * IOVA whatsoever.  That's not actually true, but the current
 818          * kernel interface doesn't tell us what it can map, and the
 819          * existing Type1 IOMMUs generally support any IOVA we're
 820          * going to actually try in practice.
 821          */
 822         container->min_iova = 0;
 823         container->max_iova = (hwaddr)-1;
 824
 825         /* Assume just 4K IOVA page size */
 826         container->iova_pgsizes = 0x1000;
 827         info.argsz = sizeof(info);
 828         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
 829         /* Ignore errors */
 830         if ((ret == 0) && (info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
 831             container->iova_pgsizes = info.iova_pgsizes;
 832         }
 833     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) {
 834         struct vfio_iommu_spapr_tce_info info;
 835
 836         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 837         if (ret) {
 838             error_report("vfio: failed to set group container: %m");
 839             ret = -errno;
 840             goto free_container_exit;
 841         }
 842         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU);
 843         if (ret) {
 844             error_report("vfio: failed to set iommu for container: %m");
 845             ret = -errno;
 846             goto free_container_exit;
 847         }
 848
 849         /*
 850          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
 851          * when container fd is closed so we do not call it explicitly
 852          * in this file.
 853          */
 854         ret = ioctl(fd, VFIO_IOMMU_ENABLE);
 855         if (ret) {
 856             error_report("vfio: failed to enable container: %m");
 857             ret = -errno;
 858             goto free_container_exit;
 859         }
 860
 861         /*
 862          * This only considers the host IOMMU's 32-bit window.  At
 863          * some point we need to add support for the optional 64-bit
 864          * window and dynamic windows
 865          */
 866         info.argsz = sizeof(info);
 867         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
 868         if (ret) {
 869             error_report("vfio: VFIO_IOMMU_SPAPR_TCE_GET_INFO failed: %m");
 870             ret = -errno;
 871             goto free_container_exit;
 872         }
 873         container->min_iova = info.dma32_window_start;
 874         container->max_iova = container->min_iova + info.dma32_window_size - 1;
 875
 876         /* Assume just 4K IOVA pages for now */
 877         container->iova_pgsizes = 0x1000;
 878     } else {
 879         error_report("vfio: No available IOMMU models");
 880         ret = -EINVAL;
 881         goto free_container_exit;
 882     }
 883
 884     container->listener = vfio_memory_listener;
 885
 886     memory_listener_register(&container->listener, container->space->as);
 887
 888     if (container->error) {
 889         ret = container->error;
 890         error_report("vfio: memory listener initialization failed for container");
 891         goto listener_release_exit;
 892     }
 893
 894     container->initialized = true;
 895
 896     QLIST_INIT(&container->group_list);
 897     QLIST_INSERT_HEAD(&space->containers, container, next);
 898
 899     group->container = container;
 900     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 901
 902     return 0;
 903 listener_release_exit:
 904     vfio_listener_release(container);
 905
 906 free_container_exit:
 907     g_free(container);
 908
 909 close_fd_exit:
 910     close(fd);
 911
 912 put_space_exit:
 913     vfio_put_address_space(space);
 914
 915     return ret;
 916 }
 917
 918 static void vfio_disconnect_container(VFIOGroup *group)
 919 {
 920     VFIOContainer *container = group->container;
 921
 922     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
 923         error_report("vfio: error disconnecting group %d from container",
 924                      group->groupid);
 925     }
 926
 927     QLIST_REMOVE(group, container_next);
 928     group->container = NULL;
 929
 930     if (QLIST_EMPTY(&container->group_list)) {
 931         VFIOAddressSpace *space = container->space;
 932         VFIOGuestIOMMU *giommu, *tmp;
 933
 934         vfio_listener_release(container);
 935         QLIST_REMOVE(container, next);
 936
 937         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
 938             memory_region_unregister_iommu_notifier(&giommu->n);
 939             QLIST_REMOVE(giommu, giommu_next);
 940             g_free(giommu);
 941         }
 942
 943         trace_vfio_disconnect_container(container->fd);
 944         close(container->fd);
 945         g_free(container);
 946
 947         vfio_put_address_space(space);
 948     }
 949 }
 950
 951 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
 952 {
 953     VFIOGroup *group;
 954     char path[32];
 955     struct vfio_group_status status = { .argsz = sizeof(status) };
 956
 957     QLIST_FOREACH(group, &vfio_group_list, next) {
 958         if (group->groupid == groupid) {
 959             /* Found it.  Now is it already in the right context? */
 960             if (group->container->space->as == as) {
 961                 return group;
 962             } else {
 963                 error_report("vfio: group %d used in multiple address spaces",
 964                              group->groupid);
 965                 return NULL;
 966             }
 967         }
 968     }
 969
 970     group = g_malloc0(sizeof(*group));
 971
 972     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
 973     group->fd = qemu_open(path, O_RDWR);
 974     if (group->fd < 0) {
 975         error_report("vfio: error opening %s: %m", path);
 976         goto free_group_exit;
 977     }
 978
 979     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
 980         error_report("vfio: error getting group status: %m");
 981         goto close_fd_exit;
 982     }
 983
 984     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
 985         error_report("vfio: error, group %d is not viable, please ensure "
 986                      "all devices within the iommu_group are bound to their "
 987                      "vfio bus driver.", groupid);
 988         goto close_fd_exit;
 989     }
 990
 991     group->groupid = groupid;
 992     QLIST_INIT(&group->device_list);
 993
 994     if (vfio_connect_container(group, as)) {
 995         error_report("vfio: failed to setup container for group %d", groupid);
 996         goto close_fd_exit;
 997     }
 998
 999     if (QLIST_EMPTY(&vfio_group_list)) {
1000         qemu_register_reset(vfio_reset_handler, NULL);
1001     }
1002
1003     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1004
1005     vfio_kvm_device_add_group(group);
1006
1007     return group;
1008
1009 close_fd_exit:
1010     close(group->fd);
1011
1012 free_group_exit:
1013     g_free(group);
1014
1015     return NULL;
1016 }
1017
1018 void vfio_put_group(VFIOGroup *group)
1019 {
1020     if (!group || !QLIST_EMPTY(&group->device_list)) {
1021         return;
1022     }
1023
1024     vfio_kvm_device_del_group(group);
1025     vfio_disconnect_container(group);
1026     QLIST_REMOVE(group, next);
1027     trace_vfio_put_group(group->fd);
1028     close(group->fd);
1029     g_free(group);
1030
1031     if (QLIST_EMPTY(&vfio_group_list)) {
1032         qemu_unregister_reset(vfio_reset_handler, NULL);
1033     }
1034 }
1035
1036 int vfio_get_device(VFIOGroup *group, const char *name,
1037                        VFIODevice *vbasedev)
1038 {
1039     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1040     int ret, fd;
1041
1042     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1043     if (fd < 0) {
1044         error_report("vfio: error getting device %s from group %d: %m",
1045                      name, group->groupid);
1046         error_printf("Verify all devices in group %d are bound to vfio-<bus> "
1047                      "or pci-stub and not already in use\n", group->groupid);
1048         return fd;
1049     }
1050
1051     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1052     if (ret) {
1053         error_report("vfio: error getting device info: %m");
1054         close(fd);
1055         return ret;
1056     }
1057
1058     vbasedev->fd = fd;
1059     vbasedev->group = group;
1060     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1061
1062     vbasedev->num_irqs = dev_info.num_irqs;
1063     vbasedev->num_regions = dev_info.num_regions;
1064     vbasedev->flags = dev_info.flags;
1065
1066     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1067                           dev_info.num_irqs);
1068
1069     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1070     return 0;
1071 }
1072
1073 void vfio_put_base_device(VFIODevice *vbasedev)
1074 {
1075     if (!vbasedev->group) {
1076         return;
1077     }
1078     QLIST_REMOVE(vbasedev, next);
1079     vbasedev->group = NULL;
1080     trace_vfio_put_base_device(vbasedev->fd);
1081     close(vbasedev->fd);
1082 }
1083
1084 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1085                          struct vfio_region_info **info)
1086 {
1087     size_t argsz = sizeof(struct vfio_region_info);
1088
1089     *info = g_malloc0(argsz);
1090
1091     (*info)->index = index;
1092     (*info)->argsz = argsz;
1093
1094     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1095         g_free(*info);
1096         return -errno;
1097     }
1098
1099     return 0;
1100 }
1101
1102 /*
1103  * Interfaces for IBM EEH (Enhanced Error Handling)
1104  */
1105 static bool vfio_eeh_container_ok(VFIOContainer *container)
1106 {
1107     /*
1108      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1109      * implementation is broken if there are multiple groups in a
1110      * container.  The hardware works in units of Partitionable
1111      * Endpoints (== IOMMU groups) and the EEH operations naively
1112      * iterate across all groups in the container, without any logic
1113      * to make sure the groups have their state synchronized.  For
1114      * certain operations (ENABLE) that might be ok, until an error
1115      * occurs, but for others (GET_STATE) it's clearly broken.
1116      */
1117
1118     /*
1119      * XXX Once fixed kernels exist, test for them here
1120      */
1121
1122     if (QLIST_EMPTY(&container->group_list)) {
1123         return false;
1124     }
1125
1126     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1127         return false;
1128     }
1129
1130     return true;
1131 }
1132
1133 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1134 {
1135     struct vfio_eeh_pe_op pe_op = {
1136         .argsz = sizeof(pe_op),
1137         .op = op,
1138     };
1139     int ret;
1140
1141     if (!vfio_eeh_container_ok(container)) {
1142         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1143                      "kernel requires a container with exactly one group", op);
1144         return -EPERM;
1145     }
1146
1147     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1148     if (ret < 0) {
1149         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1150         return -errno;
1151     }
1152
1153     return 0;
1154 }
1155
1156 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1157 {
1158     VFIOAddressSpace *space = vfio_get_address_space(as);
1159     VFIOContainer *container = NULL;
1160
1161     if (QLIST_EMPTY(&space->containers)) {
1162         /* No containers to act on */
1163         goto out;
1164     }
1165
1166     container = QLIST_FIRST(&space->containers);
1167
1168     if (QLIST_NEXT(container, next)) {
1169         /* We don't yet have logic to synchronize EEH state across
1170          * multiple containers */
1171         container = NULL;
1172         goto out;
1173     }
1174
1175 out:
1176     vfio_put_address_space(space);
1177     return container;
1178 }
1179
1180 bool vfio_eeh_as_ok(AddressSpace *as)
1181 {
1182     VFIOContainer *container = vfio_eeh_as_container(as);
1183
1184     return (container != NULL) && vfio_eeh_container_ok(container);
1185 }
1186
1187 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1188 {
1189     VFIOContainer *container = vfio_eeh_as_container(as);
1190
1191     if (!container) {
1192         return -ENODEV;
1193     }
1194     return vfio_eeh_container_op(container, op);
1195 }