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     hwaddr iova = iotlb->iova + giommu->iommu_offset;
 264     MemoryRegion *mr;
 265     hwaddr xlat;
 266     hwaddr len = iotlb->addr_mask + 1;
 267     void *vaddr;
 268     int ret;
 269
 270     trace_vfio_iommu_map_notify(iova, iova + iotlb->addr_mask);
 271
 272     if (iotlb->target_as != &address_space_memory) {
 273         error_report("Wrong target AS \"%s\", only system memory is allowed",
 274                      iotlb->target_as->name ? iotlb->target_as->name : "none");
 275         return;
 276     }
 277
 278     /*
 279      * The IOMMU TLB entry we have just covers translation through
 280      * this IOMMU to its immediate target.  We need to translate
 281      * it the rest of the way through to memory.
 282      */
 283     rcu_read_lock();
 284     mr = address_space_translate(&address_space_memory,
 285                                  iotlb->translated_addr,
 286                                  &xlat, &len, iotlb->perm & IOMMU_WO);
 287     if (!memory_region_is_ram(mr)) {
 288         error_report("iommu map to non memory area %"HWADDR_PRIx"",
 289                      xlat);
 290         goto out;
 291     }
 292     /*
 293      * Translation truncates length to the IOMMU page size,
 294      * check that it did not truncate too much.
 295      */
 296     if (len & iotlb->addr_mask) {
 297         error_report("iommu has granularity incompatible with target AS");
 298         goto out;
 299     }
 300
 301     if ((iotlb->perm & IOMMU_RW) != IOMMU_NONE) {
 302         vaddr = memory_region_get_ram_ptr(mr) + xlat;
 303         ret = vfio_dma_map(container, iova,
 304                            iotlb->addr_mask + 1, vaddr,
 305                            !(iotlb->perm & IOMMU_WO) || mr->readonly);
 306         if (ret) {
 307             error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 308                          "0x%"HWADDR_PRIx", %p) = %d (%m)",
 309                          container, iova,
 310                          iotlb->addr_mask + 1, vaddr, ret);
 311         }
 312     } else {
 313         ret = vfio_dma_unmap(container, iova, iotlb->addr_mask + 1);
 314         if (ret) {
 315             error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 316                          "0x%"HWADDR_PRIx") = %d (%m)",
 317                          container, iova,
 318                          iotlb->addr_mask + 1, ret);
 319         }
 320     }
 321 out:
 322     rcu_read_unlock();
 323 }
 324
 325 static hwaddr vfio_container_granularity(VFIOContainer *container)
 326 {
 327     return (hwaddr)1 << ctz64(container->iova_pgsizes);
 328 }
 329
 330 static void vfio_listener_region_add(MemoryListener *listener,
 331                                      MemoryRegionSection *section)
 332 {
 333     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
 334     hwaddr iova, end;
 335     Int128 llend, llsize;
 336     void *vaddr;
 337     int ret;
 338
 339     if (vfio_listener_skipped_section(section)) {
 340         trace_vfio_listener_region_add_skip(
 341                 section->offset_within_address_space,
 342                 section->offset_within_address_space +
 343                 int128_get64(int128_sub(section->size, int128_one())));
 344         return;
 345     }
 346
 347     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 348                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 349         error_report("%s received unaligned region", __func__);
 350         return;
 351     }
 352
 353     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 354     llend = int128_make64(section->offset_within_address_space);
 355     llend = int128_add(llend, section->size);
 356     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 357
 358     if (int128_ge(int128_make64(iova), llend)) {
 359         return;
 360     }
 361     end = int128_get64(int128_sub(llend, int128_one()));
 362
 363     if ((iova < container->min_iova) || (end > container->max_iova)) {
 364         error_report("vfio: IOMMU container %p can't map guest IOVA region"
 365                      " 0x%"HWADDR_PRIx"..0x%"HWADDR_PRIx,
 366                      container, iova, end);
 367         ret = -EFAULT;
 368         goto fail;
 369     }
 370
 371     memory_region_ref(section->mr);
 372
 373     if (memory_region_is_iommu(section->mr)) {
 374         VFIOGuestIOMMU *giommu;
 375
 376         trace_vfio_listener_region_add_iommu(iova, end);
 377         /*
 378          * FIXME: We should do some checking to see if the
 379          * capabilities of the host VFIO IOMMU are adequate to model
 380          * the guest IOMMU
 381          *
 382          * FIXME: For VFIO iommu types which have KVM acceleration to
 383          * avoid bouncing all map/unmaps through qemu this way, this
 384          * would be the right place to wire that up (tell the KVM
 385          * device emulation the VFIO iommu handles to use).
 386          */
 387         giommu = g_malloc0(sizeof(*giommu));
 388         giommu->iommu = section->mr;
 389         giommu->iommu_offset = section->offset_within_address_space -
 390                                section->offset_within_region;
 391         giommu->container = container;
 392         giommu->n.notify = vfio_iommu_map_notify;
 393         QLIST_INSERT_HEAD(&container->giommu_list, giommu, giommu_next);
 394
 395         memory_region_register_iommu_notifier(giommu->iommu, &giommu->n);
 396         memory_region_iommu_replay(giommu->iommu, &giommu->n,
 397                                    vfio_container_granularity(container),
 398                                    false);
 399
 400         return;
 401     }
 402
 403     /* Here we assume that memory_region_is_ram(section->mr)==true */
 404
 405     vaddr = memory_region_get_ram_ptr(section->mr) +
 406             section->offset_within_region +
 407             (iova - section->offset_within_address_space);
 408
 409     trace_vfio_listener_region_add_ram(iova, end, vaddr);
 410
 411     llsize = int128_sub(llend, int128_make64(iova));
 412
 413     ret = vfio_dma_map(container, iova, int128_get64(llsize),
 414                        vaddr, section->readonly);
 415     if (ret) {
 416         error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx", "
 417                      "0x%"HWADDR_PRIx", %p) = %d (%m)",
 418                      container, iova, int128_get64(llsize), vaddr, ret);
 419         goto fail;
 420     }
 421
 422     return;
 423
 424 fail:
 425     /*
 426      * On the initfn path, store the first error in the container so we
 427      * can gracefully fail.  Runtime, there's not much we can do other
 428      * than throw a hardware error.
 429      */
 430     if (!container->initialized) {
 431         if (!container->error) {
 432             container->error = ret;
 433         }
 434     } else {
 435         hw_error("vfio: DMA mapping failed, unable to continue");
 436     }
 437 }
 438
 439 static void vfio_listener_region_del(MemoryListener *listener,
 440                                      MemoryRegionSection *section)
 441 {
 442     VFIOContainer *container = container_of(listener, VFIOContainer, listener);
 443     hwaddr iova, end;
 444     Int128 llend, llsize;
 445     int ret;
 446
 447     if (vfio_listener_skipped_section(section)) {
 448         trace_vfio_listener_region_del_skip(
 449                 section->offset_within_address_space,
 450                 section->offset_within_address_space +
 451                 int128_get64(int128_sub(section->size, int128_one())));
 452         return;
 453     }
 454
 455     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 456                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 457         error_report("%s received unaligned region", __func__);
 458         return;
 459     }
 460
 461     if (memory_region_is_iommu(section->mr)) {
 462         VFIOGuestIOMMU *giommu;
 463
 464         QLIST_FOREACH(giommu, &container->giommu_list, giommu_next) {
 465             if (giommu->iommu == section->mr) {
 466                 memory_region_unregister_iommu_notifier(&giommu->n);
 467                 QLIST_REMOVE(giommu, giommu_next);
 468                 g_free(giommu);
 469                 break;
 470             }
 471         }
 472
 473         /*
 474          * FIXME: We assume the one big unmap below is adequate to
 475          * remove any individual page mappings in the IOMMU which
 476          * might have been copied into VFIO. This works for a page table
 477          * based IOMMU where a big unmap flattens a large range of IO-PTEs.
 478          * That may not be true for all IOMMU types.
 479          */
 480     }
 481
 482     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 483     llend = int128_make64(section->offset_within_address_space);
 484     llend = int128_add(llend, section->size);
 485     llend = int128_and(llend, int128_exts64(TARGET_PAGE_MASK));
 486
 487     if (int128_ge(int128_make64(iova), llend)) {
 488         return;
 489     }
 490     end = int128_get64(int128_sub(llend, int128_one()));
 491
 492     llsize = int128_sub(llend, int128_make64(iova));
 493
 494     trace_vfio_listener_region_del(iova, end);
 495
 496     ret = vfio_dma_unmap(container, iova, int128_get64(llsize));
 497     memory_region_unref(section->mr);
 498     if (ret) {
 499         error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx", "
 500                      "0x%"HWADDR_PRIx") = %d (%m)",
 501                      container, iova, int128_get64(llsize), ret);
 502     }
 503 }
 504
 505 static const MemoryListener vfio_memory_listener = {
 506     .region_add = vfio_listener_region_add,
 507     .region_del = vfio_listener_region_del,
 508 };
 509
 510 static void vfio_listener_release(VFIOContainer *container)
 511 {
 512     memory_listener_unregister(&container->listener);
 513 }
 514
 515 static struct vfio_info_cap_header *
 516 vfio_get_region_info_cap(struct vfio_region_info *info, uint16_t id)
 517 {
 518     struct vfio_info_cap_header *hdr;
 519     void *ptr = info;
 520
 521     if (!(info->flags & VFIO_REGION_INFO_FLAG_CAPS)) {
 522         return NULL;
 523     }
 524
 525     for (hdr = ptr + info->cap_offset; hdr != ptr; hdr = ptr + hdr->next) {
 526         if (hdr->id == id) {
 527             return hdr;
 528         }
 529     }
 530
 531     return NULL;
 532 }
 533
 534 static void vfio_setup_region_sparse_mmaps(VFIORegion *region,
 535                                            struct vfio_region_info *info)
 536 {
 537     struct vfio_info_cap_header *hdr;
 538     struct vfio_region_info_cap_sparse_mmap *sparse;
 539     int i;
 540
 541     hdr = vfio_get_region_info_cap(info, VFIO_REGION_INFO_CAP_SPARSE_MMAP);
 542     if (!hdr) {
 543         return;
 544     }
 545
 546     sparse = container_of(hdr, struct vfio_region_info_cap_sparse_mmap, header);
 547
 548     trace_vfio_region_sparse_mmap_header(region->vbasedev->name,
 549                                          region->nr, sparse->nr_areas);
 550
 551     region->nr_mmaps = sparse->nr_areas;
 552     region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
 553
 554     for (i = 0; i < region->nr_mmaps; i++) {
 555         region->mmaps[i].offset = sparse->areas[i].offset;
 556         region->mmaps[i].size = sparse->areas[i].size;
 557         trace_vfio_region_sparse_mmap_entry(i, region->mmaps[i].offset,
 558                                             region->mmaps[i].offset +
 559                                             region->mmaps[i].size);
 560     }
 561 }
 562
 563 int vfio_region_setup(Object *obj, VFIODevice *vbasedev, VFIORegion *region,
 564                       int index, const char *name)
 565 {
 566     struct vfio_region_info *info;
 567     int ret;
 568
 569     ret = vfio_get_region_info(vbasedev, index, &info);
 570     if (ret) {
 571         return ret;
 572     }
 573
 574     region->vbasedev = vbasedev;
 575     region->flags = info->flags;
 576     region->size = info->size;
 577     region->fd_offset = info->offset;
 578     region->nr = index;
 579
 580     if (region->size) {
 581         region->mem = g_new0(MemoryRegion, 1);
 582         memory_region_init_io(region->mem, obj, &vfio_region_ops,
 583                               region, name, region->size);
 584
 585         if (!vbasedev->no_mmap &&
 586             region->flags & VFIO_REGION_INFO_FLAG_MMAP &&
 587             !(region->size & ~qemu_real_host_page_mask)) {
 588
 589             vfio_setup_region_sparse_mmaps(region, info);
 590
 591             if (!region->nr_mmaps) {
 592                 region->nr_mmaps = 1;
 593                 region->mmaps = g_new0(VFIOMmap, region->nr_mmaps);
 594                 region->mmaps[0].offset = 0;
 595                 region->mmaps[0].size = region->size;
 596             }
 597         }
 598     }
 599
 600     g_free(info);
 601
 602     trace_vfio_region_setup(vbasedev->name, index, name,
 603                             region->flags, region->fd_offset, region->size);
 604     return 0;
 605 }
 606
 607 int vfio_region_mmap(VFIORegion *region)
 608 {
 609     int i, prot = 0;
 610     char *name;
 611
 612     if (!region->mem) {
 613         return 0;
 614     }
 615
 616     prot |= region->flags & VFIO_REGION_INFO_FLAG_READ ? PROT_READ : 0;
 617     prot |= region->flags & VFIO_REGION_INFO_FLAG_WRITE ? PROT_WRITE : 0;
 618
 619     for (i = 0; i < region->nr_mmaps; i++) {
 620         region->mmaps[i].mmap = mmap(NULL, region->mmaps[i].size, prot,
 621                                      MAP_SHARED, region->vbasedev->fd,
 622                                      region->fd_offset +
 623                                      region->mmaps[i].offset);
 624         if (region->mmaps[i].mmap == MAP_FAILED) {
 625             int ret = -errno;
 626
 627             trace_vfio_region_mmap_fault(memory_region_name(region->mem), i,
 628                                          region->fd_offset +
 629                                          region->mmaps[i].offset,
 630                                          region->fd_offset +
 631                                          region->mmaps[i].offset +
 632                                          region->mmaps[i].size - 1, ret);
 633
 634             region->mmaps[i].mmap = NULL;
 635
 636             for (i--; i >= 0; i--) {
 637                 memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
 638                 munmap(region->mmaps[i].mmap, region->mmaps[i].size);
 639                 object_unparent(OBJECT(&region->mmaps[i].mem));
 640                 region->mmaps[i].mmap = NULL;
 641             }
 642
 643             return ret;
 644         }
 645
 646         name = g_strdup_printf("%s mmaps[%d]",
 647                                memory_region_name(region->mem), i);
 648         memory_region_init_ram_ptr(&region->mmaps[i].mem,
 649                                    memory_region_owner(region->mem),
 650                                    name, region->mmaps[i].size,
 651                                    region->mmaps[i].mmap);
 652         g_free(name);
 653         memory_region_set_skip_dump(&region->mmaps[i].mem);
 654         memory_region_add_subregion(region->mem, region->mmaps[i].offset,
 655                                     &region->mmaps[i].mem);
 656
 657         trace_vfio_region_mmap(memory_region_name(&region->mmaps[i].mem),
 658                                region->mmaps[i].offset,
 659                                region->mmaps[i].offset +
 660                                region->mmaps[i].size - 1);
 661     }
 662
 663     return 0;
 664 }
 665
 666 void vfio_region_exit(VFIORegion *region)
 667 {
 668     int i;
 669
 670     if (!region->mem) {
 671         return;
 672     }
 673
 674     for (i = 0; i < region->nr_mmaps; i++) {
 675         if (region->mmaps[i].mmap) {
 676             memory_region_del_subregion(region->mem, &region->mmaps[i].mem);
 677         }
 678     }
 679
 680     trace_vfio_region_exit(region->vbasedev->name, region->nr);
 681 }
 682
 683 void vfio_region_finalize(VFIORegion *region)
 684 {
 685     int i;
 686
 687     if (!region->mem) {
 688         return;
 689     }
 690
 691     for (i = 0; i < region->nr_mmaps; i++) {
 692         if (region->mmaps[i].mmap) {
 693             munmap(region->mmaps[i].mmap, region->mmaps[i].size);
 694             object_unparent(OBJECT(&region->mmaps[i].mem));
 695         }
 696     }
 697
 698     object_unparent(OBJECT(region->mem));
 699
 700     g_free(region->mem);
 701     g_free(region->mmaps);
 702
 703     trace_vfio_region_finalize(region->vbasedev->name, region->nr);
 704 }
 705
 706 void vfio_region_mmaps_set_enabled(VFIORegion *region, bool enabled)
 707 {
 708     int i;
 709
 710     if (!region->mem) {
 711         return;
 712     }
 713
 714     for (i = 0; i < region->nr_mmaps; i++) {
 715         if (region->mmaps[i].mmap) {
 716             memory_region_set_enabled(&region->mmaps[i].mem, enabled);
 717         }
 718     }
 719
 720     trace_vfio_region_mmaps_set_enabled(memory_region_name(region->mem),
 721                                         enabled);
 722 }
 723
 724 void vfio_reset_handler(void *opaque)
 725 {
 726     VFIOGroup *group;
 727     VFIODevice *vbasedev;
 728
 729     QLIST_FOREACH(group, &vfio_group_list, next) {
 730         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 731             vbasedev->ops->vfio_compute_needs_reset(vbasedev);
 732         }
 733     }
 734
 735     QLIST_FOREACH(group, &vfio_group_list, next) {
 736         QLIST_FOREACH(vbasedev, &group->device_list, next) {
 737             if (vbasedev->needs_reset) {
 738                 vbasedev->ops->vfio_hot_reset_multi(vbasedev);
 739             }
 740         }
 741     }
 742 }
 743
 744 static void vfio_kvm_device_add_group(VFIOGroup *group)
 745 {
 746 #ifdef CONFIG_KVM
 747     struct kvm_device_attr attr = {
 748         .group = KVM_DEV_VFIO_GROUP,
 749         .attr = KVM_DEV_VFIO_GROUP_ADD,
 750         .addr = (uint64_t)(unsigned long)&group->fd,
 751     };
 752
 753     if (!kvm_enabled()) {
 754         return;
 755     }
 756
 757     if (vfio_kvm_device_fd < 0) {
 758         struct kvm_create_device cd = {
 759             .type = KVM_DEV_TYPE_VFIO,
 760         };
 761
 762         if (kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &cd)) {
 763             error_report("Failed to create KVM VFIO device: %m");
 764             return;
 765         }
 766
 767         vfio_kvm_device_fd = cd.fd;
 768     }
 769
 770     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 771         error_report("Failed to add group %d to KVM VFIO device: %m",
 772                      group->groupid);
 773     }
 774 #endif
 775 }
 776
 777 static void vfio_kvm_device_del_group(VFIOGroup *group)
 778 {
 779 #ifdef CONFIG_KVM
 780     struct kvm_device_attr attr = {
 781         .group = KVM_DEV_VFIO_GROUP,
 782         .attr = KVM_DEV_VFIO_GROUP_DEL,
 783         .addr = (uint64_t)(unsigned long)&group->fd,
 784     };
 785
 786     if (vfio_kvm_device_fd < 0) {
 787         return;
 788     }
 789
 790     if (ioctl(vfio_kvm_device_fd, KVM_SET_DEVICE_ATTR, &attr)) {
 791         error_report("Failed to remove group %d from KVM VFIO device: %m",
 792                      group->groupid);
 793     }
 794 #endif
 795 }
 796
 797 static VFIOAddressSpace *vfio_get_address_space(AddressSpace *as)
 798 {
 799     VFIOAddressSpace *space;
 800
 801     QLIST_FOREACH(space, &vfio_address_spaces, list) {
 802         if (space->as == as) {
 803             return space;
 804         }
 805     }
 806
 807     /* No suitable VFIOAddressSpace, create a new one */
 808     space = g_malloc0(sizeof(*space));
 809     space->as = as;
 810     QLIST_INIT(&space->containers);
 811
 812     QLIST_INSERT_HEAD(&vfio_address_spaces, space, list);
 813
 814     return space;
 815 }
 816
 817 static void vfio_put_address_space(VFIOAddressSpace *space)
 818 {
 819     if (QLIST_EMPTY(&space->containers)) {
 820         QLIST_REMOVE(space, list);
 821         g_free(space);
 822     }
 823 }
 824
 825 static int vfio_connect_container(VFIOGroup *group, AddressSpace *as)
 826 {
 827     VFIOContainer *container;
 828     int ret, fd;
 829     VFIOAddressSpace *space;
 830
 831     space = vfio_get_address_space(as);
 832
 833     QLIST_FOREACH(container, &space->containers, next) {
 834         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
 835             group->container = container;
 836             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 837             return 0;
 838         }
 839     }
 840
 841     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
 842     if (fd < 0) {
 843         error_report("vfio: failed to open /dev/vfio/vfio: %m");
 844         ret = -errno;
 845         goto put_space_exit;
 846     }
 847
 848     ret = ioctl(fd, VFIO_GET_API_VERSION);
 849     if (ret != VFIO_API_VERSION) {
 850         error_report("vfio: supported vfio version: %d, "
 851                      "reported version: %d", VFIO_API_VERSION, ret);
 852         ret = -EINVAL;
 853         goto close_fd_exit;
 854     }
 855
 856     container = g_malloc0(sizeof(*container));
 857     container->space = space;
 858     container->fd = fd;
 859     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU) ||
 860         ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU)) {
 861         bool v2 = !!ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1v2_IOMMU);
 862         struct vfio_iommu_type1_info info;
 863
 864         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 865         if (ret) {
 866             error_report("vfio: failed to set group container: %m");
 867             ret = -errno;
 868             goto free_container_exit;
 869         }
 870
 871         ret = ioctl(fd, VFIO_SET_IOMMU,
 872                     v2 ? VFIO_TYPE1v2_IOMMU : VFIO_TYPE1_IOMMU);
 873         if (ret) {
 874             error_report("vfio: failed to set iommu for container: %m");
 875             ret = -errno;
 876             goto free_container_exit;
 877         }
 878
 879         /*
 880          * FIXME: This assumes that a Type1 IOMMU can map any 64-bit
 881          * IOVA whatsoever.  That's not actually true, but the current
 882          * kernel interface doesn't tell us what it can map, and the
 883          * existing Type1 IOMMUs generally support any IOVA we're
 884          * going to actually try in practice.
 885          */
 886         container->min_iova = 0;
 887         container->max_iova = (hwaddr)-1;
 888
 889         /* Assume just 4K IOVA page size */
 890         container->iova_pgsizes = 0x1000;
 891         info.argsz = sizeof(info);
 892         ret = ioctl(fd, VFIO_IOMMU_GET_INFO, &info);
 893         /* Ignore errors */
 894         if ((ret == 0) && (info.flags & VFIO_IOMMU_INFO_PGSIZES)) {
 895             container->iova_pgsizes = info.iova_pgsizes;
 896         }
 897     } else if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_SPAPR_TCE_IOMMU)) {
 898         struct vfio_iommu_spapr_tce_info info;
 899
 900         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
 901         if (ret) {
 902             error_report("vfio: failed to set group container: %m");
 903             ret = -errno;
 904             goto free_container_exit;
 905         }
 906         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_SPAPR_TCE_IOMMU);
 907         if (ret) {
 908             error_report("vfio: failed to set iommu for container: %m");
 909             ret = -errno;
 910             goto free_container_exit;
 911         }
 912
 913         /*
 914          * The host kernel code implementing VFIO_IOMMU_DISABLE is called
 915          * when container fd is closed so we do not call it explicitly
 916          * in this file.
 917          */
 918         ret = ioctl(fd, VFIO_IOMMU_ENABLE);
 919         if (ret) {
 920             error_report("vfio: failed to enable container: %m");
 921             ret = -errno;
 922             goto free_container_exit;
 923         }
 924
 925         /*
 926          * This only considers the host IOMMU's 32-bit window.  At
 927          * some point we need to add support for the optional 64-bit
 928          * window and dynamic windows
 929          */
 930         info.argsz = sizeof(info);
 931         ret = ioctl(fd, VFIO_IOMMU_SPAPR_TCE_GET_INFO, &info);
 932         if (ret) {
 933             error_report("vfio: VFIO_IOMMU_SPAPR_TCE_GET_INFO failed: %m");
 934             ret = -errno;
 935             goto free_container_exit;
 936         }
 937         container->min_iova = info.dma32_window_start;
 938         container->max_iova = container->min_iova + info.dma32_window_size - 1;
 939
 940         /* Assume just 4K IOVA pages for now */
 941         container->iova_pgsizes = 0x1000;
 942     } else {
 943         error_report("vfio: No available IOMMU models");
 944         ret = -EINVAL;
 945         goto free_container_exit;
 946     }
 947
 948     container->listener = vfio_memory_listener;
 949
 950     memory_listener_register(&container->listener, container->space->as);
 951
 952     if (container->error) {
 953         ret = container->error;
 954         error_report("vfio: memory listener initialization failed for container");
 955         goto listener_release_exit;
 956     }
 957
 958     container->initialized = true;
 959
 960     QLIST_INIT(&container->group_list);
 961     QLIST_INSERT_HEAD(&space->containers, container, next);
 962
 963     group->container = container;
 964     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
 965
 966     return 0;
 967 listener_release_exit:
 968     vfio_listener_release(container);
 969
 970 free_container_exit:
 971     g_free(container);
 972
 973 close_fd_exit:
 974     close(fd);
 975
 976 put_space_exit:
 977     vfio_put_address_space(space);
 978
 979     return ret;
 980 }
 981
 982 static void vfio_disconnect_container(VFIOGroup *group)
 983 {
 984     VFIOContainer *container = group->container;
 985
 986     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
 987         error_report("vfio: error disconnecting group %d from container",
 988                      group->groupid);
 989     }
 990
 991     QLIST_REMOVE(group, container_next);
 992     group->container = NULL;
 993
 994     if (QLIST_EMPTY(&container->group_list)) {
 995         VFIOAddressSpace *space = container->space;
 996         VFIOGuestIOMMU *giommu, *tmp;
 997
 998         vfio_listener_release(container);
 999         QLIST_REMOVE(container, next);
1000
1001         QLIST_FOREACH_SAFE(giommu, &container->giommu_list, giommu_next, tmp) {
1002             memory_region_unregister_iommu_notifier(&giommu->n);
1003             QLIST_REMOVE(giommu, giommu_next);
1004             g_free(giommu);
1005         }
1006
1007         trace_vfio_disconnect_container(container->fd);
1008         close(container->fd);
1009         g_free(container);
1010
1011         vfio_put_address_space(space);
1012     }
1013 }
1014
1015 VFIOGroup *vfio_get_group(int groupid, AddressSpace *as)
1016 {
1017     VFIOGroup *group;
1018     char path[32];
1019     struct vfio_group_status status = { .argsz = sizeof(status) };
1020
1021     QLIST_FOREACH(group, &vfio_group_list, next) {
1022         if (group->groupid == groupid) {
1023             /* Found it.  Now is it already in the right context? */
1024             if (group->container->space->as == as) {
1025                 return group;
1026             } else {
1027                 error_report("vfio: group %d used in multiple address spaces",
1028                              group->groupid);
1029                 return NULL;
1030             }
1031         }
1032     }
1033
1034     group = g_malloc0(sizeof(*group));
1035
1036     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1037     group->fd = qemu_open(path, O_RDWR);
1038     if (group->fd < 0) {
1039         error_report("vfio: error opening %s: %m", path);
1040         goto free_group_exit;
1041     }
1042
1043     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1044         error_report("vfio: error getting group status: %m");
1045         goto close_fd_exit;
1046     }
1047
1048     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1049         error_report("vfio: error, group %d is not viable, please ensure "
1050                      "all devices within the iommu_group are bound to their "
1051                      "vfio bus driver.", groupid);
1052         goto close_fd_exit;
1053     }
1054
1055     group->groupid = groupid;
1056     QLIST_INIT(&group->device_list);
1057
1058     if (vfio_connect_container(group, as)) {
1059         error_report("vfio: failed to setup container for group %d", groupid);
1060         goto close_fd_exit;
1061     }
1062
1063     if (QLIST_EMPTY(&vfio_group_list)) {
1064         qemu_register_reset(vfio_reset_handler, NULL);
1065     }
1066
1067     QLIST_INSERT_HEAD(&vfio_group_list, group, next);
1068
1069     vfio_kvm_device_add_group(group);
1070
1071     return group;
1072
1073 close_fd_exit:
1074     close(group->fd);
1075
1076 free_group_exit:
1077     g_free(group);
1078
1079     return NULL;
1080 }
1081
1082 void vfio_put_group(VFIOGroup *group)
1083 {
1084     if (!group || !QLIST_EMPTY(&group->device_list)) {
1085         return;
1086     }
1087
1088     vfio_kvm_device_del_group(group);
1089     vfio_disconnect_container(group);
1090     QLIST_REMOVE(group, next);
1091     trace_vfio_put_group(group->fd);
1092     close(group->fd);
1093     g_free(group);
1094
1095     if (QLIST_EMPTY(&vfio_group_list)) {
1096         qemu_unregister_reset(vfio_reset_handler, NULL);
1097     }
1098 }
1099
1100 int vfio_get_device(VFIOGroup *group, const char *name,
1101                        VFIODevice *vbasedev)
1102 {
1103     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1104     int ret, fd;
1105
1106     fd = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1107     if (fd < 0) {
1108         error_report("vfio: error getting device %s from group %d: %m",
1109                      name, group->groupid);
1110         error_printf("Verify all devices in group %d are bound to vfio-<bus> "
1111                      "or pci-stub and not already in use\n", group->groupid);
1112         return fd;
1113     }
1114
1115     ret = ioctl(fd, VFIO_DEVICE_GET_INFO, &dev_info);
1116     if (ret) {
1117         error_report("vfio: error getting device info: %m");
1118         close(fd);
1119         return ret;
1120     }
1121
1122     vbasedev->fd = fd;
1123     vbasedev->group = group;
1124     QLIST_INSERT_HEAD(&group->device_list, vbasedev, next);
1125
1126     vbasedev->num_irqs = dev_info.num_irqs;
1127     vbasedev->num_regions = dev_info.num_regions;
1128     vbasedev->flags = dev_info.flags;
1129
1130     trace_vfio_get_device(name, dev_info.flags, dev_info.num_regions,
1131                           dev_info.num_irqs);
1132
1133     vbasedev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1134     return 0;
1135 }
1136
1137 void vfio_put_base_device(VFIODevice *vbasedev)
1138 {
1139     if (!vbasedev->group) {
1140         return;
1141     }
1142     QLIST_REMOVE(vbasedev, next);
1143     vbasedev->group = NULL;
1144     trace_vfio_put_base_device(vbasedev->fd);
1145     close(vbasedev->fd);
1146 }
1147
1148 int vfio_get_region_info(VFIODevice *vbasedev, int index,
1149                          struct vfio_region_info **info)
1150 {
1151     size_t argsz = sizeof(struct vfio_region_info);
1152
1153     *info = g_malloc0(argsz);
1154
1155     (*info)->index = index;
1156 retry:
1157     (*info)->argsz = argsz;
1158
1159     if (ioctl(vbasedev->fd, VFIO_DEVICE_GET_REGION_INFO, *info)) {
1160         g_free(*info);
1161         *info = NULL;
1162         return -errno;
1163     }
1164
1165     if ((*info)->argsz > argsz) {
1166         argsz = (*info)->argsz;
1167         *info = g_realloc(*info, argsz);
1168
1169         goto retry;
1170     }
1171
1172     return 0;
1173 }
1174
1175 int vfio_get_dev_region_info(VFIODevice *vbasedev, uint32_t type,
1176                              uint32_t subtype, struct vfio_region_info **info)
1177 {
1178     int i;
1179
1180     for (i = 0; i < vbasedev->num_regions; i++) {
1181         struct vfio_info_cap_header *hdr;
1182         struct vfio_region_info_cap_type *cap_type;
1183
1184         if (vfio_get_region_info(vbasedev, i, info)) {
1185             continue;
1186         }
1187
1188         hdr = vfio_get_region_info_cap(*info, VFIO_REGION_INFO_CAP_TYPE);
1189         if (!hdr) {
1190             g_free(*info);
1191             continue;
1192         }
1193
1194         cap_type = container_of(hdr, struct vfio_region_info_cap_type, header);
1195
1196         trace_vfio_get_dev_region(vbasedev->name, i,
1197                                   cap_type->type, cap_type->subtype);
1198
1199         if (cap_type->type == type && cap_type->subtype == subtype) {
1200             return 0;
1201         }
1202
1203         g_free(*info);
1204     }
1205
1206     *info = NULL;
1207     return -ENODEV;
1208 }
1209
1210 /*
1211  * Interfaces for IBM EEH (Enhanced Error Handling)
1212  */
1213 static bool vfio_eeh_container_ok(VFIOContainer *container)
1214 {
1215     /*
1216      * As of 2016-03-04 (linux-4.5) the host kernel EEH/VFIO
1217      * implementation is broken if there are multiple groups in a
1218      * container.  The hardware works in units of Partitionable
1219      * Endpoints (== IOMMU groups) and the EEH operations naively
1220      * iterate across all groups in the container, without any logic
1221      * to make sure the groups have their state synchronized.  For
1222      * certain operations (ENABLE) that might be ok, until an error
1223      * occurs, but for others (GET_STATE) it's clearly broken.
1224      */
1225
1226     /*
1227      * XXX Once fixed kernels exist, test for them here
1228      */
1229
1230     if (QLIST_EMPTY(&container->group_list)) {
1231         return false;
1232     }
1233
1234     if (QLIST_NEXT(QLIST_FIRST(&container->group_list), container_next)) {
1235         return false;
1236     }
1237
1238     return true;
1239 }
1240
1241 static int vfio_eeh_container_op(VFIOContainer *container, uint32_t op)
1242 {
1243     struct vfio_eeh_pe_op pe_op = {
1244         .argsz = sizeof(pe_op),
1245         .op = op,
1246     };
1247     int ret;
1248
1249     if (!vfio_eeh_container_ok(container)) {
1250         error_report("vfio/eeh: EEH_PE_OP 0x%x: "
1251                      "kernel requires a container with exactly one group", op);
1252         return -EPERM;
1253     }
1254
1255     ret = ioctl(container->fd, VFIO_EEH_PE_OP, &pe_op);
1256     if (ret < 0) {
1257         error_report("vfio/eeh: EEH_PE_OP 0x%x failed: %m", op);
1258         return -errno;
1259     }
1260
1261     return 0;
1262 }
1263
1264 static VFIOContainer *vfio_eeh_as_container(AddressSpace *as)
1265 {
1266     VFIOAddressSpace *space = vfio_get_address_space(as);
1267     VFIOContainer *container = NULL;
1268
1269     if (QLIST_EMPTY(&space->containers)) {
1270         /* No containers to act on */
1271         goto out;
1272     }
1273
1274     container = QLIST_FIRST(&space->containers);
1275
1276     if (QLIST_NEXT(container, next)) {
1277         /* We don't yet have logic to synchronize EEH state across
1278          * multiple containers */
1279         container = NULL;
1280         goto out;
1281     }
1282
1283 out:
1284     vfio_put_address_space(space);
1285     return container;
1286 }
1287
1288 bool vfio_eeh_as_ok(AddressSpace *as)
1289 {
1290     VFIOContainer *container = vfio_eeh_as_container(as);
1291
1292     return (container != NULL) && vfio_eeh_container_ok(container);
1293 }
1294
1295 int vfio_eeh_as_op(AddressSpace *as, uint32_t op)
1296 {
1297     VFIOContainer *container = vfio_eeh_as_container(as);
1298
1299     if (!container) {
1300         return -ENODEV;
1301     }
1302     return vfio_eeh_container_op(container, op);
1303 }