hw/vfio_pci.c

   1 /*
   2  * vfio based device assignment support
   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 <dirent.h>
  22 #include <unistd.h>
  23 #include <sys/ioctl.h>
  24 #include <sys/mman.h>
  25 #include <sys/stat.h>
  26 #include <sys/types.h>
  27 #include <linux/vfio.h>
  28
  29 #include "config.h"
  30 #include "event_notifier.h"
  31 #include "exec-memory.h"
  32 #include "kvm.h"
  33 #include "memory.h"
  34 #include "msi.h"
  35 #include "msix.h"
  36 #include "pci.h"
  37 #include "qemu-common.h"
  38 #include "qemu-error.h"
  39 #include "qemu-queue.h"
  40 #include "range.h"
  41
  42 /* #define DEBUG_VFIO */
  43 #ifdef DEBUG_VFIO
  44 #define DPRINTF(fmt, ...) \
  45     do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
  46 #else
  47 #define DPRINTF(fmt, ...) \
  48     do { } while (0)
  49 #endif
  50
  51 typedef struct VFIOBAR {
  52     off_t fd_offset; /* offset of BAR within device fd */
  53     int fd; /* device fd, allows us to pass VFIOBAR as opaque data */
  54     MemoryRegion mem; /* slow, read/write access */
  55     MemoryRegion mmap_mem; /* direct mapped access */
  56     void *mmap;
  57     size_t size;
  58     uint32_t flags; /* VFIO region flags (rd/wr/mmap) */
  59     uint8_t nr; /* cache the BAR number for debug */
  60 } VFIOBAR;
  61
  62 typedef struct VFIOINTx {
  63     bool pending; /* interrupt pending */
  64     bool kvm_accel; /* set when QEMU bypass through KVM enabled */
  65     uint8_t pin; /* which pin to pull for qemu_set_irq */
  66     EventNotifier interrupt; /* eventfd triggered on interrupt */
  67     EventNotifier unmask; /* eventfd for unmask on QEMU bypass */
  68     PCIINTxRoute route; /* routing info for QEMU bypass */
  69     uint32_t mmap_timeout; /* delay to re-enable mmaps after interrupt */
  70     QEMUTimer *mmap_timer; /* enable mmaps after periods w/o interrupts */
  71 } VFIOINTx;
  72
  73 struct VFIODevice;
  74
  75 typedef struct VFIOMSIVector {
  76     EventNotifier interrupt; /* eventfd triggered on interrupt */
  77     struct VFIODevice *vdev; /* back pointer to device */
  78     int virq; /* KVM irqchip route for QEMU bypass */
  79     bool use;
  80 } VFIOMSIVector;
  81
  82 enum {
  83     VFIO_INT_NONE = 0,
  84     VFIO_INT_INTx = 1,
  85     VFIO_INT_MSI  = 2,
  86     VFIO_INT_MSIX = 3,
  87 };
  88
  89 struct VFIOGroup;
  90
  91 typedef struct VFIOContainer {
  92     int fd; /* /dev/vfio/vfio, empowered by the attached groups */
  93     struct {
  94         /* enable abstraction to support various iommu backends */
  95         union {
  96             MemoryListener listener; /* Used by type1 iommu */
  97         };
  98         void (*release)(struct VFIOContainer *);
  99     } iommu_data;
 100     QLIST_HEAD(, VFIOGroup) group_list;
 101     QLIST_ENTRY(VFIOContainer) next;
 102 } VFIOContainer;
 103
 104 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
 105 typedef struct VFIOMSIXInfo {
 106     uint8_t table_bar;
 107     uint8_t pba_bar;
 108     uint16_t entries;
 109     uint32_t table_offset;
 110     uint32_t pba_offset;
 111     MemoryRegion mmap_mem;
 112     void *mmap;
 113 } VFIOMSIXInfo;
 114
 115 typedef struct VFIODevice {
 116     PCIDevice pdev;
 117     int fd;
 118     VFIOINTx intx;
 119     unsigned int config_size;
 120     off_t config_offset; /* Offset of config space region within device fd */
 121     unsigned int rom_size;
 122     off_t rom_offset; /* Offset of ROM region within device fd */
 123     int msi_cap_size;
 124     VFIOMSIVector *msi_vectors;
 125     VFIOMSIXInfo *msix;
 126     int nr_vectors; /* Number of MSI/MSIX vectors currently in use */
 127     int interrupt; /* Current interrupt type */
 128     VFIOBAR bars[PCI_NUM_REGIONS - 1]; /* No ROM */
 129     PCIHostDeviceAddress host;
 130     QLIST_ENTRY(VFIODevice) next;
 131     struct VFIOGroup *group;
 132     bool reset_works;
 133 } VFIODevice;
 134
 135 typedef struct VFIOGroup {
 136     int fd;
 137     int groupid;
 138     VFIOContainer *container;
 139     QLIST_HEAD(, VFIODevice) device_list;
 140     QLIST_ENTRY(VFIOGroup) next;
 141     QLIST_ENTRY(VFIOGroup) container_next;
 142 } VFIOGroup;
 143
 144 #define MSIX_CAP_LENGTH 12
 145
 146 static QLIST_HEAD(, VFIOContainer)
 147     container_list = QLIST_HEAD_INITIALIZER(container_list);
 148
 149 static QLIST_HEAD(, VFIOGroup)
 150     group_list = QLIST_HEAD_INITIALIZER(group_list);
 151
 152 static void vfio_disable_interrupts(VFIODevice *vdev);
 153 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len);
 154 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled);
 155
 156 /*
 157  * Common VFIO interrupt disable
 158  */
 159 static void vfio_disable_irqindex(VFIODevice *vdev, int index)
 160 {
 161     struct vfio_irq_set irq_set = {
 162         .argsz = sizeof(irq_set),
 163         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER,
 164         .index = index,
 165         .start = 0,
 166         .count = 0,
 167     };
 168
 169     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
 170 }
 171
 172 /*
 173  * INTx
 174  */
 175 static void vfio_unmask_intx(VFIODevice *vdev)
 176 {
 177     struct vfio_irq_set irq_set = {
 178         .argsz = sizeof(irq_set),
 179         .flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK,
 180         .index = VFIO_PCI_INTX_IRQ_INDEX,
 181         .start = 0,
 182         .count = 1,
 183     };
 184
 185     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, &irq_set);
 186 }
 187
 188 /*
 189  * Disabling BAR mmaping can be slow, but toggling it around INTx can
 190  * also be a huge overhead.  We try to get the best of both worlds by
 191  * waiting until an interrupt to disable mmaps (subsequent transitions
 192  * to the same state are effectively no overhead).  If the interrupt has
 193  * been serviced and the time gap is long enough, we re-enable mmaps for
 194  * performance.  This works well for things like graphics cards, which
 195  * may not use their interrupt at all and are penalized to an unusable
 196  * level by read/write BAR traps.  Other devices, like NICs, have more
 197  * regular interrupts and see much better latency by staying in non-mmap
 198  * mode.  We therefore set the default mmap_timeout such that a ping
 199  * is just enough to keep the mmap disabled.  Users can experiment with
 200  * other options with the x-intx-mmap-timeout-ms parameter (a value of
 201  * zero disables the timer).
 202  */
 203 static void vfio_intx_mmap_enable(void *opaque)
 204 {
 205     VFIODevice *vdev = opaque;
 206
 207     if (vdev->intx.pending) {
 208         qemu_mod_timer(vdev->intx.mmap_timer,
 209                        qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
 210         return;
 211     }
 212
 213     vfio_mmap_set_enabled(vdev, true);
 214 }
 215
 216 static void vfio_intx_interrupt(void *opaque)
 217 {
 218     VFIODevice *vdev = opaque;
 219
 220     if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) {
 221         return;
 222     }
 223
 224     DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__, vdev->host.domain,
 225             vdev->host.bus, vdev->host.slot, vdev->host.function,
 226             'A' + vdev->intx.pin);
 227
 228     vdev->intx.pending = true;
 229     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 1);
 230     vfio_mmap_set_enabled(vdev, false);
 231     if (vdev->intx.mmap_timeout) {
 232         qemu_mod_timer(vdev->intx.mmap_timer,
 233                        qemu_get_clock_ms(vm_clock) + vdev->intx.mmap_timeout);
 234     }
 235 }
 236
 237 static void vfio_eoi(VFIODevice *vdev)
 238 {
 239     if (!vdev->intx.pending) {
 240         return;
 241     }
 242
 243     DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__, vdev->host.domain,
 244             vdev->host.bus, vdev->host.slot, vdev->host.function);
 245
 246     vdev->intx.pending = false;
 247     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
 248     vfio_unmask_intx(vdev);
 249 }
 250
 251 static int vfio_enable_intx(VFIODevice *vdev)
 252 {
 253     uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1);
 254     int ret, argsz;
 255     struct vfio_irq_set *irq_set;
 256     int32_t *pfd;
 257
 258     if (!pin) {
 259         return 0;
 260     }
 261
 262     vfio_disable_interrupts(vdev);
 263
 264     vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */
 265     ret = event_notifier_init(&vdev->intx.interrupt, 0);
 266     if (ret) {
 267         error_report("vfio: Error: event_notifier_init failed\n");
 268         return ret;
 269     }
 270
 271     argsz = sizeof(*irq_set) + sizeof(*pfd);
 272
 273     irq_set = g_malloc0(argsz);
 274     irq_set->argsz = argsz;
 275     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 276     irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 277     irq_set->start = 0;
 278     irq_set->count = 1;
 279     pfd = (int32_t *)&irq_set->data;
 280
 281     *pfd = event_notifier_get_fd(&vdev->intx.interrupt);
 282     qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev);
 283
 284     ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
 285     g_free(irq_set);
 286     if (ret) {
 287         error_report("vfio: Error: Failed to setup INTx fd: %m\n");
 288         qemu_set_fd_handler(*pfd, NULL, NULL, vdev);
 289         event_notifier_cleanup(&vdev->intx.interrupt);
 290         return -errno;
 291     }
 292
 293     vdev->interrupt = VFIO_INT_INTx;
 294
 295     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 296             vdev->host.bus, vdev->host.slot, vdev->host.function);
 297
 298     return 0;
 299 }
 300
 301 static void vfio_disable_intx(VFIODevice *vdev)
 302 {
 303     int fd;
 304
 305     qemu_del_timer(vdev->intx.mmap_timer);
 306     vfio_disable_irqindex(vdev, VFIO_PCI_INTX_IRQ_INDEX);
 307     vdev->intx.pending = false;
 308     qemu_set_irq(vdev->pdev.irq[vdev->intx.pin], 0);
 309     vfio_mmap_set_enabled(vdev, true);
 310
 311     fd = event_notifier_get_fd(&vdev->intx.interrupt);
 312     qemu_set_fd_handler(fd, NULL, NULL, vdev);
 313     event_notifier_cleanup(&vdev->intx.interrupt);
 314
 315     vdev->interrupt = VFIO_INT_NONE;
 316
 317     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 318             vdev->host.bus, vdev->host.slot, vdev->host.function);
 319 }
 320
 321 /*
 322  * MSI/X
 323  */
 324 static void vfio_msi_interrupt(void *opaque)
 325 {
 326     VFIOMSIVector *vector = opaque;
 327     VFIODevice *vdev = vector->vdev;
 328     int nr = vector - vdev->msi_vectors;
 329
 330     if (!event_notifier_test_and_clear(&vector->interrupt)) {
 331         return;
 332     }
 333
 334     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__,
 335             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 336             vdev->host.function, nr);
 337
 338     if (vdev->interrupt == VFIO_INT_MSIX) {
 339         msix_notify(&vdev->pdev, nr);
 340     } else if (vdev->interrupt == VFIO_INT_MSI) {
 341         msi_notify(&vdev->pdev, nr);
 342     } else {
 343         error_report("vfio: MSI interrupt receieved, but not enabled?\n");
 344     }
 345 }
 346
 347 static int vfio_enable_vectors(VFIODevice *vdev, bool msix)
 348 {
 349     struct vfio_irq_set *irq_set;
 350     int ret = 0, i, argsz;
 351     int32_t *fds;
 352
 353     argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds));
 354
 355     irq_set = g_malloc0(argsz);
 356     irq_set->argsz = argsz;
 357     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 358     irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX;
 359     irq_set->start = 0;
 360     irq_set->count = vdev->nr_vectors;
 361     fds = (int32_t *)&irq_set->data;
 362
 363     for (i = 0; i < vdev->nr_vectors; i++) {
 364         if (!vdev->msi_vectors[i].use) {
 365             fds[i] = -1;
 366             continue;
 367         }
 368
 369         fds[i] = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt);
 370     }
 371
 372     ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
 373
 374     g_free(irq_set);
 375
 376     return ret;
 377 }
 378
 379 static int vfio_msix_vector_use(PCIDevice *pdev,
 380                                 unsigned int nr, MSIMessage msg)
 381 {
 382     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 383     VFIOMSIVector *vector;
 384     int ret;
 385
 386     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__,
 387             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 388             vdev->host.function, nr);
 389
 390     vector = &vdev->msi_vectors[nr];
 391     vector->vdev = vdev;
 392     vector->use = true;
 393
 394     msix_vector_use(pdev, nr);
 395
 396     if (event_notifier_init(&vector->interrupt, 0)) {
 397         error_report("vfio: Error: event_notifier_init failed\n");
 398     }
 399
 400     /*
 401      * Attempt to enable route through KVM irqchip,
 402      * default to userspace handling if unavailable.
 403      */
 404     vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
 405     if (vector->virq < 0 ||
 406         kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
 407                                        vector->virq) < 0) {
 408         if (vector->virq >= 0) {
 409             kvm_irqchip_release_virq(kvm_state, vector->virq);
 410             vector->virq = -1;
 411         }
 412         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 413                             vfio_msi_interrupt, NULL, vector);
 414     }
 415
 416     /*
 417      * We don't want to have the host allocate all possible MSI vectors
 418      * for a device if they're not in use, so we shutdown and incrementally
 419      * increase them as needed.
 420      */
 421     if (vdev->nr_vectors < nr + 1) {
 422         vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
 423         vdev->nr_vectors = nr + 1;
 424         ret = vfio_enable_vectors(vdev, true);
 425         if (ret) {
 426             error_report("vfio: failed to enable vectors, %d\n", ret);
 427         }
 428     } else {
 429         int argsz;
 430         struct vfio_irq_set *irq_set;
 431         int32_t *pfd;
 432
 433         argsz = sizeof(*irq_set) + sizeof(*pfd);
 434
 435         irq_set = g_malloc0(argsz);
 436         irq_set->argsz = argsz;
 437         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 438                          VFIO_IRQ_SET_ACTION_TRIGGER;
 439         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 440         irq_set->start = nr;
 441         irq_set->count = 1;
 442         pfd = (int32_t *)&irq_set->data;
 443
 444         *pfd = event_notifier_get_fd(&vector->interrupt);
 445
 446         ret = ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
 447         g_free(irq_set);
 448         if (ret) {
 449             error_report("vfio: failed to modify vector, %d\n", ret);
 450         }
 451     }
 452
 453     return 0;
 454 }
 455
 456 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr)
 457 {
 458     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 459     VFIOMSIVector *vector = &vdev->msi_vectors[nr];
 460     int argsz;
 461     struct vfio_irq_set *irq_set;
 462     int32_t *pfd;
 463
 464     DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__,
 465             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 466             vdev->host.function, nr);
 467
 468     /*
 469      * XXX What's the right thing to do here?  This turns off the interrupt
 470      * completely, but do we really just want to switch the interrupt to
 471      * bouncing through userspace and let msix.c drop it?  Not sure.
 472      */
 473     msix_vector_unuse(pdev, nr);
 474
 475     argsz = sizeof(*irq_set) + sizeof(*pfd);
 476
 477     irq_set = g_malloc0(argsz);
 478     irq_set->argsz = argsz;
 479     irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 480                      VFIO_IRQ_SET_ACTION_TRIGGER;
 481     irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 482     irq_set->start = nr;
 483     irq_set->count = 1;
 484     pfd = (int32_t *)&irq_set->data;
 485
 486     *pfd = -1;
 487
 488     ioctl(vdev->fd, VFIO_DEVICE_SET_IRQS, irq_set);
 489
 490     g_free(irq_set);
 491
 492     if (vector->virq < 0) {
 493         qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 494                             NULL, NULL, NULL);
 495     } else {
 496         kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
 497                                           vector->virq);
 498         kvm_irqchip_release_virq(kvm_state, vector->virq);
 499         vector->virq = -1;
 500     }
 501
 502     event_notifier_cleanup(&vector->interrupt);
 503     vector->use = false;
 504 }
 505
 506 /* TODO This should move to msi.c */
 507 static MSIMessage msi_get_msg(PCIDevice *pdev, unsigned int vector)
 508 {
 509     uint16_t flags = pci_get_word(pdev->config + pdev->msi_cap + PCI_MSI_FLAGS);
 510     bool msi64bit = flags & PCI_MSI_FLAGS_64BIT;
 511     MSIMessage msg;
 512
 513     if (msi64bit) {
 514         msg.address = pci_get_quad(pdev->config +
 515                                    pdev->msi_cap + PCI_MSI_ADDRESS_LO);
 516     } else {
 517         msg.address = pci_get_long(pdev->config +
 518                                    pdev->msi_cap + PCI_MSI_ADDRESS_LO);
 519     }
 520
 521     msg.data = pci_get_word(pdev->config + pdev->msi_cap +
 522                             (msi64bit ? PCI_MSI_DATA_64 : PCI_MSI_DATA_32));
 523     msg.data += vector;
 524
 525     return msg;
 526 }
 527
 528 static void vfio_enable_msix(VFIODevice *vdev)
 529 {
 530     vfio_disable_interrupts(vdev);
 531
 532     vdev->msi_vectors = g_malloc0(vdev->msix->entries * sizeof(VFIOMSIVector));
 533
 534     vdev->interrupt = VFIO_INT_MSIX;
 535
 536     if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use,
 537                                   vfio_msix_vector_release)) {
 538         error_report("vfio: msix_set_vector_notifiers failed\n");
 539     }
 540
 541     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 542             vdev->host.bus, vdev->host.slot, vdev->host.function);
 543 }
 544
 545 static void vfio_enable_msi(VFIODevice *vdev)
 546 {
 547     int ret, i;
 548
 549     vfio_disable_interrupts(vdev);
 550
 551     vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev);
 552 retry:
 553     vdev->msi_vectors = g_malloc0(vdev->nr_vectors * sizeof(VFIOMSIVector));
 554
 555     for (i = 0; i < vdev->nr_vectors; i++) {
 556         MSIMessage msg;
 557         VFIOMSIVector *vector = &vdev->msi_vectors[i];
 558
 559         vector->vdev = vdev;
 560         vector->use = true;
 561
 562         if (event_notifier_init(&vector->interrupt, 0)) {
 563             error_report("vfio: Error: event_notifier_init failed\n");
 564         }
 565
 566         msg = msi_get_msg(&vdev->pdev, i);
 567
 568         /*
 569          * Attempt to enable route through KVM irqchip,
 570          * default to userspace handling if unavailable.
 571          */
 572         vector->virq = kvm_irqchip_add_msi_route(kvm_state, msg);
 573         if (vector->virq < 0 ||
 574             kvm_irqchip_add_irqfd_notifier(kvm_state, &vector->interrupt,
 575                                            vector->virq) < 0) {
 576             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 577                                 vfio_msi_interrupt, NULL, vector);
 578         }
 579     }
 580
 581     ret = vfio_enable_vectors(vdev, false);
 582     if (ret) {
 583         if (ret < 0) {
 584             error_report("vfio: Error: Failed to setup MSI fds: %m\n");
 585         } else if (ret != vdev->nr_vectors) {
 586             error_report("vfio: Error: Failed to enable %d "
 587                          "MSI vectors, retry with %d\n", vdev->nr_vectors, ret);
 588         }
 589
 590         for (i = 0; i < vdev->nr_vectors; i++) {
 591             VFIOMSIVector *vector = &vdev->msi_vectors[i];
 592             if (vector->virq >= 0) {
 593                 kvm_irqchip_remove_irqfd_notifier(kvm_state, &vector->interrupt,
 594                                                   vector->virq);
 595                 kvm_irqchip_release_virq(kvm_state, vector->virq);
 596                 vector->virq = -1;
 597             } else {
 598                 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 599                                     NULL, NULL, NULL);
 600             }
 601             event_notifier_cleanup(&vector->interrupt);
 602         }
 603
 604         g_free(vdev->msi_vectors);
 605
 606         if (ret > 0 && ret != vdev->nr_vectors) {
 607             vdev->nr_vectors = ret;
 608             goto retry;
 609         }
 610         vdev->nr_vectors = 0;
 611
 612         return;
 613     }
 614
 615     vdev->interrupt = VFIO_INT_MSI;
 616
 617     DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__,
 618             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 619             vdev->host.function, vdev->nr_vectors);
 620 }
 621
 622 static void vfio_disable_msi_common(VFIODevice *vdev)
 623 {
 624     g_free(vdev->msi_vectors);
 625     vdev->msi_vectors = NULL;
 626     vdev->nr_vectors = 0;
 627     vdev->interrupt = VFIO_INT_NONE;
 628
 629     vfio_enable_intx(vdev);
 630 }
 631
 632 static void vfio_disable_msix(VFIODevice *vdev)
 633 {
 634     msix_unset_vector_notifiers(&vdev->pdev);
 635
 636     if (vdev->nr_vectors) {
 637         vfio_disable_irqindex(vdev, VFIO_PCI_MSIX_IRQ_INDEX);
 638     }
 639
 640     vfio_disable_msi_common(vdev);
 641
 642     DPRINTF("%s(%04x:%02x:%02x.%x, msi%s)\n", __func__,
 643             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 644             vdev->host.function, msix ? "x" : "");
 645 }
 646
 647 static void vfio_disable_msi(VFIODevice *vdev)
 648 {
 649     int i;
 650
 651     vfio_disable_irqindex(vdev, VFIO_PCI_MSI_IRQ_INDEX);
 652
 653     for (i = 0; i < vdev->nr_vectors; i++) {
 654         VFIOMSIVector *vector = &vdev->msi_vectors[i];
 655
 656         if (!vector->use) {
 657             continue;
 658         }
 659
 660         if (vector->virq >= 0) {
 661             kvm_irqchip_remove_irqfd_notifier(kvm_state,
 662                                               &vector->interrupt, vector->virq);
 663             kvm_irqchip_release_virq(kvm_state, vector->virq);
 664             vector->virq = -1;
 665         } else {
 666             qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt),
 667                                 NULL, NULL, NULL);
 668         }
 669
 670         event_notifier_cleanup(&vector->interrupt);
 671     }
 672
 673     vfio_disable_msi_common(vdev);
 674
 675     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
 676             vdev->host.bus, vdev->host.slot, vdev->host.function);
 677 }
 678
 679 /*
 680  * IO Port/MMIO - Beware of the endians, VFIO is always little endian
 681  */
 682 static void vfio_bar_write(void *opaque, target_phys_addr_t addr,
 683                            uint64_t data, unsigned size)
 684 {
 685     VFIOBAR *bar = opaque;
 686     union {
 687         uint8_t byte;
 688         uint16_t word;
 689         uint32_t dword;
 690         uint64_t qword;
 691     } buf;
 692
 693     switch (size) {
 694     case 1:
 695         buf.byte = data;
 696         break;
 697     case 2:
 698         buf.word = cpu_to_le16(data);
 699         break;
 700     case 4:
 701         buf.dword = cpu_to_le32(data);
 702         break;
 703     default:
 704         hw_error("vfio: unsupported write size, %d bytes\n", size);
 705         break;
 706     }
 707
 708     if (pwrite(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
 709         error_report("%s(,0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d) failed: %m\n",
 710                      __func__, addr, data, size);
 711     }
 712
 713     DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", 0x%"PRIx64", %d)\n",
 714             __func__, bar->nr, addr, data, size);
 715
 716     /*
 717      * A read or write to a BAR always signals an INTx EOI.  This will
 718      * do nothing if not pending (including not in INTx mode).  We assume
 719      * that a BAR access is in response to an interrupt and that BAR
 720      * accesses will service the interrupt.  Unfortunately, we don't know
 721      * which access will service the interrupt, so we're potentially
 722      * getting quite a few host interrupts per guest interrupt.
 723      */
 724     vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
 725 }
 726
 727 static uint64_t vfio_bar_read(void *opaque,
 728                               target_phys_addr_t addr, unsigned size)
 729 {
 730     VFIOBAR *bar = opaque;
 731     union {
 732         uint8_t byte;
 733         uint16_t word;
 734         uint32_t dword;
 735         uint64_t qword;
 736     } buf;
 737     uint64_t data = 0;
 738
 739     if (pread(bar->fd, &buf, size, bar->fd_offset + addr) != size) {
 740         error_report("%s(,0x%"TARGET_PRIxPHYS", %d) failed: %m\n",
 741                      __func__, addr, size);
 742         return (uint64_t)-1;
 743     }
 744
 745     switch (size) {
 746     case 1:
 747         data = buf.byte;
 748         break;
 749     case 2:
 750         data = le16_to_cpu(buf.word);
 751         break;
 752     case 4:
 753         data = le32_to_cpu(buf.dword);
 754         break;
 755     default:
 756         hw_error("vfio: unsupported read size, %d bytes\n", size);
 757         break;
 758     }
 759
 760     DPRINTF("%s(BAR%d+0x%"TARGET_PRIxPHYS", %d) = 0x%"PRIx64"\n",
 761             __func__, bar->nr, addr, size, data);
 762
 763     /* Same as write above */
 764     vfio_eoi(container_of(bar, VFIODevice, bars[bar->nr]));
 765
 766     return data;
 767 }
 768
 769 static const MemoryRegionOps vfio_bar_ops = {
 770     .read = vfio_bar_read,
 771     .write = vfio_bar_write,
 772     .endianness = DEVICE_LITTLE_ENDIAN,
 773 };
 774
 775 /*
 776  * PCI config space
 777  */
 778 static uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len)
 779 {
 780     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 781     uint32_t val = 0;
 782
 783     /*
 784      * We only need QEMU PCI config support for the ROM BAR, the MSI and MSIX
 785      * capabilities, and the multifunction bit below.  We let VFIO handle
 786      * virtualizing everything else.  Performance is not a concern here.
 787      */
 788     if (ranges_overlap(addr, len, PCI_ROM_ADDRESS, 4) ||
 789         (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
 790          ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) ||
 791         (pdev->cap_present & QEMU_PCI_CAP_MSI &&
 792          ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size))) {
 793
 794         val = pci_default_read_config(pdev, addr, len);
 795     } else {
 796         if (pread(vdev->fd, &val, len, vdev->config_offset + addr) != len) {
 797             error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m\n",
 798                          __func__, vdev->host.domain, vdev->host.bus,
 799                          vdev->host.slot, vdev->host.function, addr, len);
 800             return -errno;
 801         }
 802         val = le32_to_cpu(val);
 803     }
 804
 805     /* Multifunction bit is virualized in QEMU */
 806     if (unlikely(ranges_overlap(addr, len, PCI_HEADER_TYPE, 1))) {
 807         uint32_t mask = PCI_HEADER_TYPE_MULTI_FUNCTION;
 808
 809         if (len == 4) {
 810             mask <<= 16;
 811         }
 812
 813         if (pdev->cap_present & QEMU_PCI_CAP_MULTIFUNCTION) {
 814             val |= mask;
 815         } else {
 816             val &= ~mask;
 817         }
 818     }
 819
 820     DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__,
 821             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 822             vdev->host.function, addr, len, val);
 823
 824     return val;
 825 }
 826
 827 static void vfio_pci_write_config(PCIDevice *pdev, uint32_t addr,
 828                                   uint32_t val, int len)
 829 {
 830     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
 831     uint32_t val_le = cpu_to_le32(val);
 832
 833     DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__,
 834             vdev->host.domain, vdev->host.bus, vdev->host.slot,
 835             vdev->host.function, addr, val, len);
 836
 837     /* Write everything to VFIO, let it filter out what we can't write */
 838     if (pwrite(vdev->fd, &val_le, len, vdev->config_offset + addr) != len) {
 839         error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m\n",
 840                      __func__, vdev->host.domain, vdev->host.bus,
 841                      vdev->host.slot, vdev->host.function, addr, val, len);
 842     }
 843
 844     /* Write standard header bits to emulation */
 845     if (addr < PCI_CONFIG_HEADER_SIZE) {
 846         pci_default_write_config(pdev, addr, val, len);
 847         return;
 848     }
 849
 850     /* MSI/MSI-X Enabling/Disabling */
 851     if (pdev->cap_present & QEMU_PCI_CAP_MSI &&
 852         ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) {
 853         int is_enabled, was_enabled = msi_enabled(pdev);
 854
 855         pci_default_write_config(pdev, addr, val, len);
 856
 857         is_enabled = msi_enabled(pdev);
 858
 859         if (!was_enabled && is_enabled) {
 860             vfio_enable_msi(vdev);
 861         } else if (was_enabled && !is_enabled) {
 862             vfio_disable_msi(vdev);
 863         }
 864     }
 865
 866     if (pdev->cap_present & QEMU_PCI_CAP_MSIX &&
 867         ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) {
 868         int is_enabled, was_enabled = msix_enabled(pdev);
 869
 870         pci_default_write_config(pdev, addr, val, len);
 871
 872         is_enabled = msix_enabled(pdev);
 873
 874         if (!was_enabled && is_enabled) {
 875             vfio_enable_msix(vdev);
 876         } else if (was_enabled && !is_enabled) {
 877             vfio_disable_msix(vdev);
 878         }
 879     }
 880 }
 881
 882 /*
 883  * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
 884  */
 885 static int vfio_dma_unmap(VFIOContainer *container,
 886                           target_phys_addr_t iova, ram_addr_t size)
 887 {
 888     struct vfio_iommu_type1_dma_unmap unmap = {
 889         .argsz = sizeof(unmap),
 890         .flags = 0,
 891         .iova = iova,
 892         .size = size,
 893     };
 894
 895     if (ioctl(container->fd, VFIO_IOMMU_UNMAP_DMA, &unmap)) {
 896         DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno);
 897         return -errno;
 898     }
 899
 900     return 0;
 901 }
 902
 903 static int vfio_dma_map(VFIOContainer *container, target_phys_addr_t iova,
 904                         ram_addr_t size, void *vaddr, bool readonly)
 905 {
 906     struct vfio_iommu_type1_dma_map map = {
 907         .argsz = sizeof(map),
 908         .flags = VFIO_DMA_MAP_FLAG_READ,
 909         .vaddr = (__u64)(uintptr_t)vaddr,
 910         .iova = iova,
 911         .size = size,
 912     };
 913
 914     if (!readonly) {
 915         map.flags |= VFIO_DMA_MAP_FLAG_WRITE;
 916     }
 917
 918     /*
 919      * Try the mapping, if it fails with EBUSY, unmap the region and try
 920      * again.  This shouldn't be necessary, but we sometimes see it in
 921      * the the VGA ROM space.
 922      */
 923     if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 ||
 924         (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 &&
 925          ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) {
 926         return 0;
 927     }
 928
 929     DPRINTF("VFIO_MAP_DMA: %d\n", -errno);
 930     return -errno;
 931 }
 932
 933 static void vfio_listener_dummy1(MemoryListener *listener)
 934 {
 935     /* We don't do batching (begin/commit) or care about logging */
 936 }
 937
 938 static void vfio_listener_dummy2(MemoryListener *listener,
 939                                  MemoryRegionSection *section)
 940 {
 941     /* We don't do logging or care about nops */
 942 }
 943
 944 static void vfio_listener_dummy3(MemoryListener *listener,
 945                                  MemoryRegionSection *section,
 946                                  bool match_data, uint64_t data,
 947                                  EventNotifier *e)
 948 {
 949     /* We don't care about eventfds */
 950 }
 951
 952 static bool vfio_listener_skipped_section(MemoryRegionSection *section)
 953 {
 954     return !memory_region_is_ram(section->mr);
 955 }
 956
 957 static void vfio_listener_region_add(MemoryListener *listener,
 958                                      MemoryRegionSection *section)
 959 {
 960     VFIOContainer *container = container_of(listener, VFIOContainer,
 961                                             iommu_data.listener);
 962     target_phys_addr_t iova, end;
 963     void *vaddr;
 964     int ret;
 965
 966     if (vfio_listener_skipped_section(section)) {
 967         DPRINTF("vfio: SKIPPING region_add %"TARGET_PRIxPHYS" - %"PRIx64"\n",
 968                 section->offset_within_address_space,
 969                 section->offset_within_address_space + section->size - 1);
 970         return;
 971     }
 972
 973     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
 974                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
 975         error_report("%s received unaligned region\n", __func__);
 976         return;
 977     }
 978
 979     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
 980     end = (section->offset_within_address_space + section->size) &
 981           TARGET_PAGE_MASK;
 982
 983     if (iova >= end) {
 984         return;
 985     }
 986
 987     vaddr = memory_region_get_ram_ptr(section->mr) +
 988             section->offset_within_region +
 989             (iova - section->offset_within_address_space);
 990
 991     DPRINTF("vfio: region_add %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS" [%p]\n",
 992             iova, end - 1, vaddr);
 993
 994     ret = vfio_dma_map(container, iova, end - iova, vaddr, section->readonly);
 995     if (ret) {
 996         error_report("vfio_dma_map(%p, 0x%"TARGET_PRIxPHYS", "
 997                      "0x%"TARGET_PRIxPHYS", %p) = %d (%m)\n",
 998                      container, iova, end - iova, vaddr, ret);
 999     }
1000 }
1001
1002 static void vfio_listener_region_del(MemoryListener *listener,
1003                                      MemoryRegionSection *section)
1004 {
1005     VFIOContainer *container = container_of(listener, VFIOContainer,
1006                                             iommu_data.listener);
1007     target_phys_addr_t iova, end;
1008     int ret;
1009
1010     if (vfio_listener_skipped_section(section)) {
1011         DPRINTF("vfio: SKIPPING region_del %"TARGET_PRIxPHYS" - %"PRIx64"\n",
1012                 section->offset_within_address_space,
1013                 section->offset_within_address_space + section->size - 1);
1014         return;
1015     }
1016
1017     if (unlikely((section->offset_within_address_space & ~TARGET_PAGE_MASK) !=
1018                  (section->offset_within_region & ~TARGET_PAGE_MASK))) {
1019         error_report("%s received unaligned region\n", __func__);
1020         return;
1021     }
1022
1023     iova = TARGET_PAGE_ALIGN(section->offset_within_address_space);
1024     end = (section->offset_within_address_space + section->size) &
1025           TARGET_PAGE_MASK;
1026
1027     if (iova >= end) {
1028         return;
1029     }
1030
1031     DPRINTF("vfio: region_del %"TARGET_PRIxPHYS" - %"TARGET_PRIxPHYS"\n",
1032             iova, end - 1);
1033
1034     ret = vfio_dma_unmap(container, iova, end - iova);
1035     if (ret) {
1036         error_report("vfio_dma_unmap(%p, 0x%"TARGET_PRIxPHYS", "
1037                      "0x%"TARGET_PRIxPHYS") = %d (%m)\n",
1038                      container, iova, end - iova, ret);
1039     }
1040 }
1041
1042 static MemoryListener vfio_memory_listener = {
1043     .begin = vfio_listener_dummy1,
1044     .commit = vfio_listener_dummy1,
1045     .region_add = vfio_listener_region_add,
1046     .region_del = vfio_listener_region_del,
1047     .region_nop = vfio_listener_dummy2,
1048     .log_start = vfio_listener_dummy2,
1049     .log_stop = vfio_listener_dummy2,
1050     .log_sync = vfio_listener_dummy2,
1051     .log_global_start = vfio_listener_dummy1,
1052     .log_global_stop = vfio_listener_dummy1,
1053     .eventfd_add = vfio_listener_dummy3,
1054     .eventfd_del = vfio_listener_dummy3,
1055 };
1056
1057 static void vfio_listener_release(VFIOContainer *container)
1058 {
1059     memory_listener_unregister(&container->iommu_data.listener);
1060 }
1061
1062 /*
1063  * Interrupt setup
1064  */
1065 static void vfio_disable_interrupts(VFIODevice *vdev)
1066 {
1067     switch (vdev->interrupt) {
1068     case VFIO_INT_INTx:
1069         vfio_disable_intx(vdev);
1070         break;
1071     case VFIO_INT_MSI:
1072         vfio_disable_msi(vdev);
1073         break;
1074     case VFIO_INT_MSIX:
1075         vfio_disable_msix(vdev);
1076         break;
1077     }
1078 }
1079
1080 static int vfio_setup_msi(VFIODevice *vdev, int pos)
1081 {
1082     uint16_t ctrl;
1083     bool msi_64bit, msi_maskbit;
1084     int ret, entries;
1085
1086     if (pread(vdev->fd, &ctrl, sizeof(ctrl),
1087               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1088         return -errno;
1089     }
1090     ctrl = le16_to_cpu(ctrl);
1091
1092     msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT);
1093     msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT);
1094     entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1);
1095
1096     DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev->host.domain,
1097             vdev->host.bus, vdev->host.slot, vdev->host.function, pos);
1098
1099     ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit);
1100     if (ret < 0) {
1101         if (ret == -ENOTSUP) {
1102             return 0;
1103         }
1104         error_report("vfio: msi_init failed\n");
1105         return ret;
1106     }
1107     vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0);
1108
1109     return 0;
1110 }
1111
1112 /*
1113  * We don't have any control over how pci_add_capability() inserts
1114  * capabilities into the chain.  In order to setup MSI-X we need a
1115  * MemoryRegion for the BAR.  In order to setup the BAR and not
1116  * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1117  * need to first look for where the MSI-X table lives.  So we
1118  * unfortunately split MSI-X setup across two functions.
1119  */
1120 static int vfio_early_setup_msix(VFIODevice *vdev)
1121 {
1122     uint8_t pos;
1123     uint16_t ctrl;
1124     uint32_t table, pba;
1125
1126     pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX);
1127     if (!pos) {
1128         return 0;
1129     }
1130
1131     if (pread(vdev->fd, &ctrl, sizeof(ctrl),
1132               vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) {
1133         return -errno;
1134     }
1135
1136     if (pread(vdev->fd, &table, sizeof(table),
1137               vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) {
1138         return -errno;
1139     }
1140
1141     if (pread(vdev->fd, &pba, sizeof(pba),
1142               vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) {
1143         return -errno;
1144     }
1145
1146     ctrl = le16_to_cpu(ctrl);
1147     table = le32_to_cpu(table);
1148     pba = le32_to_cpu(pba);
1149
1150     vdev->msix = g_malloc0(sizeof(*(vdev->msix)));
1151     vdev->msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK;
1152     vdev->msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK;
1153     vdev->msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK;
1154     vdev->msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK;
1155     vdev->msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1;
1156
1157     DPRINTF("%04x:%02x:%02x.%x "
1158             "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
1159             vdev->host.domain, vdev->host.bus, vdev->host.slot,
1160             vdev->host.function, pos, vdev->msix->table_bar,
1161             vdev->msix->table_offset, vdev->msix->entries);
1162
1163     return 0;
1164 }
1165
1166 static int vfio_setup_msix(VFIODevice *vdev, int pos)
1167 {
1168     int ret;
1169
1170     ret = msix_init(&vdev->pdev, vdev->msix->entries,
1171                     &vdev->bars[vdev->msix->table_bar].mem,
1172                     vdev->msix->table_bar, vdev->msix->table_offset,
1173                     &vdev->bars[vdev->msix->pba_bar].mem,
1174                     vdev->msix->pba_bar, vdev->msix->pba_offset, pos);
1175     if (ret < 0) {
1176         if (ret == -ENOTSUP) {
1177             return 0;
1178         }
1179         error_report("vfio: msix_init failed\n");
1180         return ret;
1181     }
1182
1183     return 0;
1184 }
1185
1186 static void vfio_teardown_msi(VFIODevice *vdev)
1187 {
1188     msi_uninit(&vdev->pdev);
1189
1190     if (vdev->msix) {
1191         msix_uninit(&vdev->pdev, &vdev->bars[vdev->msix->table_bar].mem,
1192                     &vdev->bars[vdev->msix->pba_bar].mem);
1193     }
1194 }
1195
1196 /*
1197  * Resource setup
1198  */
1199 static void vfio_mmap_set_enabled(VFIODevice *vdev, bool enabled)
1200 {
1201     int i;
1202
1203     for (i = 0; i < PCI_ROM_SLOT; i++) {
1204         VFIOBAR *bar = &vdev->bars[i];
1205
1206         if (!bar->size) {
1207             continue;
1208         }
1209
1210         memory_region_set_enabled(&bar->mmap_mem, enabled);
1211         if (vdev->msix && vdev->msix->table_bar == i) {
1212             memory_region_set_enabled(&vdev->msix->mmap_mem, enabled);
1213         }
1214     }
1215 }
1216
1217 static void vfio_unmap_bar(VFIODevice *vdev, int nr)
1218 {
1219     VFIOBAR *bar = &vdev->bars[nr];
1220
1221     if (!bar->size) {
1222         return;
1223     }
1224
1225     memory_region_del_subregion(&bar->mem, &bar->mmap_mem);
1226     munmap(bar->mmap, memory_region_size(&bar->mmap_mem));
1227
1228     if (vdev->msix && vdev->msix->table_bar == nr) {
1229         memory_region_del_subregion(&bar->mem, &vdev->msix->mmap_mem);
1230         munmap(vdev->msix->mmap, memory_region_size(&vdev->msix->mmap_mem));
1231     }
1232
1233     memory_region_destroy(&bar->mem);
1234 }
1235
1236 static int vfio_mmap_bar(VFIOBAR *bar, MemoryRegion *mem, MemoryRegion *submem,
1237                          void **map, size_t size, off_t offset,
1238                          const char *name)
1239 {
1240     int ret = 0;
1241
1242     if (size && bar->flags & VFIO_REGION_INFO_FLAG_MMAP) {
1243         int prot = 0;
1244
1245         if (bar->flags & VFIO_REGION_INFO_FLAG_READ) {
1246             prot |= PROT_READ;
1247         }
1248
1249         if (bar->flags & VFIO_REGION_INFO_FLAG_WRITE) {
1250             prot |= PROT_WRITE;
1251         }
1252
1253         *map = mmap(NULL, size, prot, MAP_SHARED,
1254                     bar->fd, bar->fd_offset + offset);
1255         if (*map == MAP_FAILED) {
1256             *map = NULL;
1257             ret = -errno;
1258             goto empty_region;
1259         }
1260
1261         memory_region_init_ram_ptr(submem, name, size, *map);
1262     } else {
1263 empty_region:
1264         /* Create a zero sized sub-region to make cleanup easy. */
1265         memory_region_init(submem, name, 0);
1266     }
1267
1268     memory_region_add_subregion(mem, offset, submem);
1269
1270     return ret;
1271 }
1272
1273 static void vfio_map_bar(VFIODevice *vdev, int nr)
1274 {
1275     VFIOBAR *bar = &vdev->bars[nr];
1276     unsigned size = bar->size;
1277     char name[64];
1278     uint32_t pci_bar;
1279     uint8_t type;
1280     int ret;
1281
1282     /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1283     if (!size) {
1284         return;
1285     }
1286
1287     snprintf(name, sizeof(name), "VFIO %04x:%02x:%02x.%x BAR %d",
1288              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1289              vdev->host.function, nr);
1290
1291     /* Determine what type of BAR this is for registration */
1292     ret = pread(vdev->fd, &pci_bar, sizeof(pci_bar),
1293                 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr));
1294     if (ret != sizeof(pci_bar)) {
1295         error_report("vfio: Failed to read BAR %d (%m)\n", nr);
1296         return;
1297     }
1298
1299     pci_bar = le32_to_cpu(pci_bar);
1300     type = pci_bar & (pci_bar & PCI_BASE_ADDRESS_SPACE_IO ?
1301            ~PCI_BASE_ADDRESS_IO_MASK : ~PCI_BASE_ADDRESS_MEM_MASK);
1302
1303     /* A "slow" read/write mapping underlies all BARs */
1304     memory_region_init_io(&bar->mem, &vfio_bar_ops, bar, name, size);
1305     pci_register_bar(&vdev->pdev, nr, type, &bar->mem);
1306
1307     /*
1308      * We can't mmap areas overlapping the MSIX vector table, so we
1309      * potentially insert a direct-mapped subregion before and after it.
1310      */
1311     if (vdev->msix && vdev->msix->table_bar == nr) {
1312         size = vdev->msix->table_offset & TARGET_PAGE_MASK;
1313     }
1314
1315     strncat(name, " mmap", sizeof(name) - strlen(name) - 1);
1316     if (vfio_mmap_bar(bar, &bar->mem,
1317                       &bar->mmap_mem, &bar->mmap, size, 0, name)) {
1318         error_report("%s unsupported. Performance may be slow\n", name);
1319     }
1320
1321     if (vdev->msix && vdev->msix->table_bar == nr) {
1322         unsigned start;
1323
1324         start = TARGET_PAGE_ALIGN(vdev->msix->table_offset +
1325                                   (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE));
1326
1327         size = start < bar->size ? bar->size - start : 0;
1328         strncat(name, " msix-hi", sizeof(name) - strlen(name) - 1);
1329         /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
1330         if (vfio_mmap_bar(bar, &bar->mem, &vdev->msix->mmap_mem,
1331                           &vdev->msix->mmap, size, start, name)) {
1332             error_report("%s unsupported. Performance may be slow\n", name);
1333         }
1334     }
1335 }
1336
1337 static void vfio_map_bars(VFIODevice *vdev)
1338 {
1339     int i;
1340
1341     for (i = 0; i < PCI_ROM_SLOT; i++) {
1342         vfio_map_bar(vdev, i);
1343     }
1344 }
1345
1346 static void vfio_unmap_bars(VFIODevice *vdev)
1347 {
1348     int i;
1349
1350     for (i = 0; i < PCI_ROM_SLOT; i++) {
1351         vfio_unmap_bar(vdev, i);
1352     }
1353 }
1354
1355 /*
1356  * General setup
1357  */
1358 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos)
1359 {
1360     uint8_t tmp, next = 0xff;
1361
1362     for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp;
1363          tmp = pdev->config[tmp + 1]) {
1364         if (tmp > pos && tmp < next) {
1365             next = tmp;
1366         }
1367     }
1368
1369     return next - pos;
1370 }
1371
1372 static int vfio_add_std_cap(VFIODevice *vdev, uint8_t pos)
1373 {
1374     PCIDevice *pdev = &vdev->pdev;
1375     uint8_t cap_id, next, size;
1376     int ret;
1377
1378     cap_id = pdev->config[pos];
1379     next = pdev->config[pos + 1];
1380
1381     /*
1382      * If it becomes important to configure capabilities to their actual
1383      * size, use this as the default when it's something we don't recognize.
1384      * Since QEMU doesn't actually handle many of the config accesses,
1385      * exact size doesn't seem worthwhile.
1386      */
1387     size = vfio_std_cap_max_size(pdev, pos);
1388
1389     /*
1390      * pci_add_capability always inserts the new capability at the head
1391      * of the chain.  Therefore to end up with a chain that matches the
1392      * physical device, we insert from the end by making this recursive.
1393      * This is also why we pre-caclulate size above as cached config space
1394      * will be changed as we unwind the stack.
1395      */
1396     if (next) {
1397         ret = vfio_add_std_cap(vdev, next);
1398         if (ret) {
1399             return ret;
1400         }
1401     } else {
1402         pdev->config[PCI_CAPABILITY_LIST] = 0; /* Begin the rebuild */
1403     }
1404
1405     switch (cap_id) {
1406     case PCI_CAP_ID_MSI:
1407         ret = vfio_setup_msi(vdev, pos);
1408         break;
1409     case PCI_CAP_ID_MSIX:
1410         ret = vfio_setup_msix(vdev, pos);
1411         break;
1412     default:
1413         ret = pci_add_capability(pdev, cap_id, pos, size);
1414         break;
1415     }
1416
1417     if (ret < 0) {
1418         error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
1419                      "0x%x[0x%x]@0x%x: %d\n", vdev->host.domain,
1420                      vdev->host.bus, vdev->host.slot, vdev->host.function,
1421                      cap_id, size, pos, ret);
1422         return ret;
1423     }
1424
1425     return 0;
1426 }
1427
1428 static int vfio_add_capabilities(VFIODevice *vdev)
1429 {
1430     PCIDevice *pdev = &vdev->pdev;
1431
1432     if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) ||
1433         !pdev->config[PCI_CAPABILITY_LIST]) {
1434         return 0; /* Nothing to add */
1435     }
1436
1437     return vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST]);
1438 }
1439
1440 static int vfio_load_rom(VFIODevice *vdev)
1441 {
1442     uint64_t size = vdev->rom_size;
1443     char name[32];
1444     off_t off = 0, voff = vdev->rom_offset;
1445     ssize_t bytes;
1446     void *ptr;
1447
1448     /* If loading ROM from file, pci handles it */
1449     if (vdev->pdev.romfile || !vdev->pdev.rom_bar || !size) {
1450         return 0;
1451     }
1452
1453     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
1454             vdev->host.bus, vdev->host.slot, vdev->host.function);
1455
1456     snprintf(name, sizeof(name), "vfio[%04x:%02x:%02x.%x].rom",
1457              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1458              vdev->host.function);
1459     memory_region_init_ram(&vdev->pdev.rom, name, size);
1460     ptr = memory_region_get_ram_ptr(&vdev->pdev.rom);
1461     memset(ptr, 0xff, size);
1462
1463     while (size) {
1464         bytes = pread(vdev->fd, ptr + off, size, voff + off);
1465         if (bytes == 0) {
1466             break; /* expect that we could get back less than the ROM BAR */
1467         } else if (bytes > 0) {
1468             off += bytes;
1469             size -= bytes;
1470         } else {
1471             if (errno == EINTR || errno == EAGAIN) {
1472                 continue;
1473             }
1474             error_report("vfio: Error reading device ROM: %m\n");
1475             memory_region_destroy(&vdev->pdev.rom);
1476             return -errno;
1477         }
1478     }
1479
1480     pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 0, &vdev->pdev.rom);
1481     vdev->pdev.has_rom = true;
1482     return 0;
1483 }
1484
1485 static int vfio_connect_container(VFIOGroup *group)
1486 {
1487     VFIOContainer *container;
1488     int ret, fd;
1489
1490     if (group->container) {
1491         return 0;
1492     }
1493
1494     QLIST_FOREACH(container, &container_list, next) {
1495         if (!ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &container->fd)) {
1496             group->container = container;
1497             QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1498             return 0;
1499         }
1500     }
1501
1502     fd = qemu_open("/dev/vfio/vfio", O_RDWR);
1503     if (fd < 0) {
1504         error_report("vfio: failed to open /dev/vfio/vfio: %m\n");
1505         return -errno;
1506     }
1507
1508     ret = ioctl(fd, VFIO_GET_API_VERSION);
1509     if (ret != VFIO_API_VERSION) {
1510         error_report("vfio: supported vfio version: %d, "
1511                      "reported version: %d\n", VFIO_API_VERSION, ret);
1512         close(fd);
1513         return -EINVAL;
1514     }
1515
1516     container = g_malloc0(sizeof(*container));
1517     container->fd = fd;
1518
1519     if (ioctl(fd, VFIO_CHECK_EXTENSION, VFIO_TYPE1_IOMMU)) {
1520         ret = ioctl(group->fd, VFIO_GROUP_SET_CONTAINER, &fd);
1521         if (ret) {
1522             error_report("vfio: failed to set group container: %m\n");
1523             g_free(container);
1524             close(fd);
1525             return -errno;
1526         }
1527
1528         ret = ioctl(fd, VFIO_SET_IOMMU, VFIO_TYPE1_IOMMU);
1529         if (ret) {
1530             error_report("vfio: failed to set iommu for container: %m\n");
1531             g_free(container);
1532             close(fd);
1533             return -errno;
1534         }
1535
1536         container->iommu_data.listener = vfio_memory_listener;
1537         container->iommu_data.release = vfio_listener_release;
1538
1539         memory_listener_register(&container->iommu_data.listener,
1540                                  get_system_memory());
1541     } else {
1542         error_report("vfio: No available IOMMU models\n");
1543         g_free(container);
1544         close(fd);
1545         return -EINVAL;
1546     }
1547
1548     QLIST_INIT(&container->group_list);
1549     QLIST_INSERT_HEAD(&container_list, container, next);
1550
1551     group->container = container;
1552     QLIST_INSERT_HEAD(&container->group_list, group, container_next);
1553
1554     return 0;
1555 }
1556
1557 static void vfio_disconnect_container(VFIOGroup *group)
1558 {
1559     VFIOContainer *container = group->container;
1560
1561     if (ioctl(group->fd, VFIO_GROUP_UNSET_CONTAINER, &container->fd)) {
1562         error_report("vfio: error disconnecting group %d from container\n",
1563                      group->groupid);
1564     }
1565
1566     QLIST_REMOVE(group, container_next);
1567     group->container = NULL;
1568
1569     if (QLIST_EMPTY(&container->group_list)) {
1570         if (container->iommu_data.release) {
1571             container->iommu_data.release(container);
1572         }
1573         QLIST_REMOVE(container, next);
1574         DPRINTF("vfio_disconnect_container: close container->fd\n");
1575         close(container->fd);
1576         g_free(container);
1577     }
1578 }
1579
1580 static VFIOGroup *vfio_get_group(int groupid)
1581 {
1582     VFIOGroup *group;
1583     char path[32];
1584     struct vfio_group_status status = { .argsz = sizeof(status) };
1585
1586     QLIST_FOREACH(group, &group_list, next) {
1587         if (group->groupid == groupid) {
1588             return group;
1589         }
1590     }
1591
1592     group = g_malloc0(sizeof(*group));
1593
1594     snprintf(path, sizeof(path), "/dev/vfio/%d", groupid);
1595     group->fd = qemu_open(path, O_RDWR);
1596     if (group->fd < 0) {
1597         error_report("vfio: error opening %s: %m\n", path);
1598         g_free(group);
1599         return NULL;
1600     }
1601
1602     if (ioctl(group->fd, VFIO_GROUP_GET_STATUS, &status)) {
1603         error_report("vfio: error getting group status: %m\n");
1604         close(group->fd);
1605         g_free(group);
1606         return NULL;
1607     }
1608
1609     if (!(status.flags & VFIO_GROUP_FLAGS_VIABLE)) {
1610         error_report("vfio: error, group %d is not viable, please ensure "
1611                      "all devices within the iommu_group are bound to their "
1612                      "vfio bus driver.\n", groupid);
1613         close(group->fd);
1614         g_free(group);
1615         return NULL;
1616     }
1617
1618     group->groupid = groupid;
1619     QLIST_INIT(&group->device_list);
1620
1621     if (vfio_connect_container(group)) {
1622         error_report("vfio: failed to setup container for group %d\n", groupid);
1623         close(group->fd);
1624         g_free(group);
1625         return NULL;
1626     }
1627
1628     QLIST_INSERT_HEAD(&group_list, group, next);
1629
1630     return group;
1631 }
1632
1633 static void vfio_put_group(VFIOGroup *group)
1634 {
1635     if (!QLIST_EMPTY(&group->device_list)) {
1636         return;
1637     }
1638
1639     vfio_disconnect_container(group);
1640     QLIST_REMOVE(group, next);
1641     DPRINTF("vfio_put_group: close group->fd\n");
1642     close(group->fd);
1643     g_free(group);
1644 }
1645
1646 static int vfio_get_device(VFIOGroup *group, const char *name, VFIODevice *vdev)
1647 {
1648     struct vfio_device_info dev_info = { .argsz = sizeof(dev_info) };
1649     struct vfio_region_info reg_info = { .argsz = sizeof(reg_info) };
1650     int ret, i;
1651
1652     ret = ioctl(group->fd, VFIO_GROUP_GET_DEVICE_FD, name);
1653     if (ret < 0) {
1654         error_report("vfio: error getting device %s from group %d: %m\n",
1655                      name, group->groupid);
1656         error_report("Verify all devices in group %d are bound to vfio-pci "
1657                      "or pci-stub and not already in use\n", group->groupid);
1658         return ret;
1659     }
1660
1661     vdev->fd = ret;
1662     vdev->group = group;
1663     QLIST_INSERT_HEAD(&group->device_list, vdev, next);
1664
1665     /* Sanity check device */
1666     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_INFO, &dev_info);
1667     if (ret) {
1668         error_report("vfio: error getting device info: %m\n");
1669         goto error;
1670     }
1671
1672     DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name,
1673             dev_info.flags, dev_info.num_regions, dev_info.num_irqs);
1674
1675     if (!(dev_info.flags & VFIO_DEVICE_FLAGS_PCI)) {
1676         error_report("vfio: Um, this isn't a PCI device\n");
1677         goto error;
1678     }
1679
1680     vdev->reset_works = !!(dev_info.flags & VFIO_DEVICE_FLAGS_RESET);
1681     if (!vdev->reset_works) {
1682         error_report("Warning, device %s does not support reset\n", name);
1683     }
1684
1685     if (dev_info.num_regions != VFIO_PCI_NUM_REGIONS) {
1686         error_report("vfio: unexpected number of io regions %u\n",
1687                      dev_info.num_regions);
1688         goto error;
1689     }
1690
1691     if (dev_info.num_irqs != VFIO_PCI_NUM_IRQS) {
1692         error_report("vfio: unexpected number of irqs %u\n", dev_info.num_irqs);
1693         goto error;
1694     }
1695
1696     for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) {
1697         reg_info.index = i;
1698
1699         ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1700         if (ret) {
1701             error_report("vfio: Error getting region %d info: %m\n", i);
1702             goto error;
1703         }
1704
1705         DPRINTF("Device %s region %d:\n", name, i);
1706         DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1707                 (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1708                 (unsigned long)reg_info.flags);
1709
1710         vdev->bars[i].flags = reg_info.flags;
1711         vdev->bars[i].size = reg_info.size;
1712         vdev->bars[i].fd_offset = reg_info.offset;
1713         vdev->bars[i].fd = vdev->fd;
1714         vdev->bars[i].nr = i;
1715     }
1716
1717     reg_info.index = VFIO_PCI_ROM_REGION_INDEX;
1718
1719     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1720     if (ret) {
1721         error_report("vfio: Error getting ROM info: %m\n");
1722         goto error;
1723     }
1724
1725     DPRINTF("Device %s ROM:\n", name);
1726     DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1727             (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1728             (unsigned long)reg_info.flags);
1729
1730     vdev->rom_size = reg_info.size;
1731     vdev->rom_offset = reg_info.offset;
1732
1733     reg_info.index = VFIO_PCI_CONFIG_REGION_INDEX;
1734
1735     ret = ioctl(vdev->fd, VFIO_DEVICE_GET_REGION_INFO, &reg_info);
1736     if (ret) {
1737         error_report("vfio: Error getting config info: %m\n");
1738         goto error;
1739     }
1740
1741     DPRINTF("Device %s config:\n", name);
1742     DPRINTF("  size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1743             (unsigned long)reg_info.size, (unsigned long)reg_info.offset,
1744             (unsigned long)reg_info.flags);
1745
1746     vdev->config_size = reg_info.size;
1747     vdev->config_offset = reg_info.offset;
1748
1749 error:
1750     if (ret) {
1751         QLIST_REMOVE(vdev, next);
1752         vdev->group = NULL;
1753         close(vdev->fd);
1754     }
1755     return ret;
1756 }
1757
1758 static void vfio_put_device(VFIODevice *vdev)
1759 {
1760     QLIST_REMOVE(vdev, next);
1761     vdev->group = NULL;
1762     DPRINTF("vfio_put_device: close vdev->fd\n");
1763     close(vdev->fd);
1764     if (vdev->msix) {
1765         g_free(vdev->msix);
1766         vdev->msix = NULL;
1767     }
1768 }
1769
1770 static int vfio_initfn(PCIDevice *pdev)
1771 {
1772     VFIODevice *pvdev, *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1773     VFIOGroup *group;
1774     char path[PATH_MAX], iommu_group_path[PATH_MAX], *group_name;
1775     ssize_t len;
1776     struct stat st;
1777     int groupid;
1778     int ret;
1779
1780     /* Check that the host device exists */
1781     snprintf(path, sizeof(path),
1782              "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
1783              vdev->host.domain, vdev->host.bus, vdev->host.slot,
1784              vdev->host.function);
1785     if (stat(path, &st) < 0) {
1786         error_report("vfio: error: no such host device: %s\n", path);
1787         return -errno;
1788     }
1789
1790     strncat(path, "iommu_group", sizeof(path) - strlen(path) - 1);
1791
1792     len = readlink(path, iommu_group_path, PATH_MAX);
1793     if (len <= 0) {
1794         error_report("vfio: error no iommu_group for device\n");
1795         return -errno;
1796     }
1797
1798     iommu_group_path[len] = 0;
1799     group_name = basename(iommu_group_path);
1800
1801     if (sscanf(group_name, "%d", &groupid) != 1) {
1802         error_report("vfio: error reading %s: %m\n", path);
1803         return -errno;
1804     }
1805
1806     DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__, vdev->host.domain,
1807             vdev->host.bus, vdev->host.slot, vdev->host.function, groupid);
1808
1809     group = vfio_get_group(groupid);
1810     if (!group) {
1811         error_report("vfio: failed to get group %d\n", groupid);
1812         return -ENOENT;
1813     }
1814
1815     snprintf(path, sizeof(path), "%04x:%02x:%02x.%01x",
1816             vdev->host.domain, vdev->host.bus, vdev->host.slot,
1817             vdev->host.function);
1818
1819     QLIST_FOREACH(pvdev, &group->device_list, next) {
1820         if (pvdev->host.domain == vdev->host.domain &&
1821             pvdev->host.bus == vdev->host.bus &&
1822             pvdev->host.slot == vdev->host.slot &&
1823             pvdev->host.function == vdev->host.function) {
1824
1825             error_report("vfio: error: device %s is already attached\n", path);
1826             vfio_put_group(group);
1827             return -EBUSY;
1828         }
1829     }
1830
1831     ret = vfio_get_device(group, path, vdev);
1832     if (ret) {
1833         error_report("vfio: failed to get device %s\n", path);
1834         vfio_put_group(group);
1835         return ret;
1836     }
1837
1838     /* Get a copy of config space */
1839     ret = pread(vdev->fd, vdev->pdev.config,
1840                 MIN(pci_config_size(&vdev->pdev), vdev->config_size),
1841                 vdev->config_offset);
1842     if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) {
1843         ret = ret < 0 ? -errno : -EFAULT;
1844         error_report("vfio: Failed to read device config space\n");
1845         goto out_put;
1846     }
1847
1848     /*
1849      * Clear host resource mapping info.  If we choose not to register a
1850      * BAR, such as might be the case with the option ROM, we can get
1851      * confusing, unwritable, residual addresses from the host here.
1852      */
1853     memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24);
1854     memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4);
1855
1856     vfio_load_rom(vdev);
1857
1858     ret = vfio_early_setup_msix(vdev);
1859     if (ret) {
1860         goto out_put;
1861     }
1862
1863     vfio_map_bars(vdev);
1864
1865     ret = vfio_add_capabilities(vdev);
1866     if (ret) {
1867         goto out_teardown;
1868     }
1869
1870     if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) {
1871         vdev->intx.mmap_timer = qemu_new_timer_ms(vm_clock,
1872                                                   vfio_intx_mmap_enable, vdev);
1873         ret = vfio_enable_intx(vdev);
1874         if (ret) {
1875             goto out_teardown;
1876         }
1877     }
1878
1879     return 0;
1880
1881 out_teardown:
1882     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
1883     vfio_teardown_msi(vdev);
1884     vfio_unmap_bars(vdev);
1885 out_put:
1886     vfio_put_device(vdev);
1887     vfio_put_group(group);
1888     return ret;
1889 }
1890
1891 static void vfio_exitfn(PCIDevice *pdev)
1892 {
1893     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1894     VFIOGroup *group = vdev->group;
1895
1896     pci_device_set_intx_routing_notifier(&vdev->pdev, NULL);
1897     vfio_disable_interrupts(vdev);
1898     if (vdev->intx.mmap_timer) {
1899         qemu_free_timer(vdev->intx.mmap_timer);
1900     }
1901     vfio_teardown_msi(vdev);
1902     vfio_unmap_bars(vdev);
1903     vfio_put_device(vdev);
1904     vfio_put_group(group);
1905 }
1906
1907 static void vfio_pci_reset(DeviceState *dev)
1908 {
1909     PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev);
1910     VFIODevice *vdev = DO_UPCAST(VFIODevice, pdev, pdev);
1911     uint16_t cmd;
1912
1913     DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__, vdev->host.domain,
1914             vdev->host.bus, vdev->host.slot, vdev->host.function);
1915
1916     vfio_disable_interrupts(vdev);
1917
1918     /*
1919      * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
1920      * Also put INTx Disable in known state.
1921      */
1922     cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2);
1923     cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER |
1924              PCI_COMMAND_INTX_DISABLE);
1925     vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2);
1926
1927     if (vdev->reset_works) {
1928         if (ioctl(vdev->fd, VFIO_DEVICE_RESET)) {
1929             error_report("vfio: Error unable to reset physical device "
1930                          "(%04x:%02x:%02x.%x): %m\n", vdev->host.domain,
1931                          vdev->host.bus, vdev->host.slot, vdev->host.function);
1932         }
1933     }
1934
1935     vfio_enable_intx(vdev);
1936 }
1937
1938 static Property vfio_pci_dev_properties[] = {
1939     DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice, host),
1940     DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice,
1941                        intx.mmap_timeout, 1100),
1942     /*
1943      * TODO - support passed fds... is this necessary?
1944      * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
1945      * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
1946      */
1947     DEFINE_PROP_END_OF_LIST(),
1948 };
1949
1950
1951 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data)
1952 {
1953     DeviceClass *dc = DEVICE_CLASS(klass);
1954     PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass);
1955
1956     dc->reset = vfio_pci_reset;
1957     dc->props = vfio_pci_dev_properties;
1958     pdc->init = vfio_initfn;
1959     pdc->exit = vfio_exitfn;
1960     pdc->config_read = vfio_pci_read_config;
1961     pdc->config_write = vfio_pci_write_config;
1962 }
1963
1964 static const TypeInfo vfio_pci_dev_info = {
1965     .name = "vfio-pci",
1966     .parent = TYPE_PCI_DEVICE,
1967     .instance_size = sizeof(VFIODevice),
1968     .class_init = vfio_pci_dev_class_init,
1969 };
1970
1971 static void register_vfio_pci_dev_type(void)
1972 {
1973     type_register_static(&vfio_pci_dev_info);
1974 }
1975
1976 type_init(register_vfio_pci_dev_type)