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