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 "qemu/osdep.h" 22 #include <linux/vfio.h> 23 #include <sys/ioctl.h> 24 25 #include "hw/pci/msi.h" 26 #include "hw/pci/msix.h" 27 #include "hw/pci/pci_bridge.h" 28 #include "qemu/error-report.h" 29 #include "qemu/range.h" 30 #include "sysemu/kvm.h" 31 #include "sysemu/sysemu.h" 32 #include "pci.h" 33 #include "trace.h" 34 #include "qapi/error.h" 35 36 #define MSIX_CAP_LENGTH 12 37 38 static void vfio_disable_interrupts(VFIOPCIDevice *vdev); 39 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled); 40 41 /* 42 * Disabling BAR mmaping can be slow, but toggling it around INTx can 43 * also be a huge overhead. We try to get the best of both worlds by 44 * waiting until an interrupt to disable mmaps (subsequent transitions 45 * to the same state are effectively no overhead). If the interrupt has 46 * been serviced and the time gap is long enough, we re-enable mmaps for 47 * performance. This works well for things like graphics cards, which 48 * may not use their interrupt at all and are penalized to an unusable 49 * level by read/write BAR traps. Other devices, like NICs, have more 50 * regular interrupts and see much better latency by staying in non-mmap 51 * mode. We therefore set the default mmap_timeout such that a ping 52 * is just enough to keep the mmap disabled. Users can experiment with 53 * other options with the x-intx-mmap-timeout-ms parameter (a value of 54 * zero disables the timer). 55 */ 56 static void vfio_intx_mmap_enable(void *opaque) 57 { 58 VFIOPCIDevice *vdev = opaque; 59 60 if (vdev->intx.pending) { 61 timer_mod(vdev->intx.mmap_timer, 62 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout); 63 return; 64 } 65 66 vfio_mmap_set_enabled(vdev, true); 67 } 68 69 static void vfio_intx_interrupt(void *opaque) 70 { 71 VFIOPCIDevice *vdev = opaque; 72 73 if (!event_notifier_test_and_clear(&vdev->intx.interrupt)) { 74 return; 75 } 76 77 trace_vfio_intx_interrupt(vdev->vbasedev.name, 'A' + vdev->intx.pin); 78 79 vdev->intx.pending = true; 80 pci_irq_assert(&vdev->pdev); 81 vfio_mmap_set_enabled(vdev, false); 82 if (vdev->intx.mmap_timeout) { 83 timer_mod(vdev->intx.mmap_timer, 84 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL) + vdev->intx.mmap_timeout); 85 } 86 } 87 88 static void vfio_intx_eoi(VFIODevice *vbasedev) 89 { 90 VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev); 91 92 if (!vdev->intx.pending) { 93 return; 94 } 95 96 trace_vfio_intx_eoi(vbasedev->name); 97 98 vdev->intx.pending = false; 99 pci_irq_deassert(&vdev->pdev); 100 vfio_unmask_single_irqindex(vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 101 } 102 103 static void vfio_intx_enable_kvm(VFIOPCIDevice *vdev, Error **errp) 104 { 105 #ifdef CONFIG_KVM 106 struct kvm_irqfd irqfd = { 107 .fd = event_notifier_get_fd(&vdev->intx.interrupt), 108 .gsi = vdev->intx.route.irq, 109 .flags = KVM_IRQFD_FLAG_RESAMPLE, 110 }; 111 struct vfio_irq_set *irq_set; 112 int ret, argsz; 113 int32_t *pfd; 114 115 if (vdev->no_kvm_intx || !kvm_irqfds_enabled() || 116 vdev->intx.route.mode != PCI_INTX_ENABLED || 117 !kvm_resamplefds_enabled()) { 118 return; 119 } 120 121 /* Get to a known interrupt state */ 122 qemu_set_fd_handler(irqfd.fd, NULL, NULL, vdev); 123 vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 124 vdev->intx.pending = false; 125 pci_irq_deassert(&vdev->pdev); 126 127 /* Get an eventfd for resample/unmask */ 128 if (event_notifier_init(&vdev->intx.unmask, 0)) { 129 error_setg(errp, "event_notifier_init failed eoi"); 130 goto fail; 131 } 132 133 /* KVM triggers it, VFIO listens for it */ 134 irqfd.resamplefd = event_notifier_get_fd(&vdev->intx.unmask); 135 136 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) { 137 error_setg_errno(errp, errno, "failed to setup resample irqfd"); 138 goto fail_irqfd; 139 } 140 141 argsz = sizeof(*irq_set) + sizeof(*pfd); 142 143 irq_set = g_malloc0(argsz); 144 irq_set->argsz = argsz; 145 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_UNMASK; 146 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX; 147 irq_set->start = 0; 148 irq_set->count = 1; 149 pfd = (int32_t *)&irq_set->data; 150 151 *pfd = irqfd.resamplefd; 152 153 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 154 g_free(irq_set); 155 if (ret) { 156 error_setg_errno(errp, -ret, "failed to setup INTx unmask fd"); 157 goto fail_vfio; 158 } 159 160 /* Let'em rip */ 161 vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 162 163 vdev->intx.kvm_accel = true; 164 165 trace_vfio_intx_enable_kvm(vdev->vbasedev.name); 166 167 return; 168 169 fail_vfio: 170 irqfd.flags = KVM_IRQFD_FLAG_DEASSIGN; 171 kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd); 172 fail_irqfd: 173 event_notifier_cleanup(&vdev->intx.unmask); 174 fail: 175 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev); 176 vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 177 #endif 178 } 179 180 static void vfio_intx_disable_kvm(VFIOPCIDevice *vdev) 181 { 182 #ifdef CONFIG_KVM 183 struct kvm_irqfd irqfd = { 184 .fd = event_notifier_get_fd(&vdev->intx.interrupt), 185 .gsi = vdev->intx.route.irq, 186 .flags = KVM_IRQFD_FLAG_DEASSIGN, 187 }; 188 189 if (!vdev->intx.kvm_accel) { 190 return; 191 } 192 193 /* 194 * Get to a known state, hardware masked, QEMU ready to accept new 195 * interrupts, QEMU IRQ de-asserted. 196 */ 197 vfio_mask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 198 vdev->intx.pending = false; 199 pci_irq_deassert(&vdev->pdev); 200 201 /* Tell KVM to stop listening for an INTx irqfd */ 202 if (kvm_vm_ioctl(kvm_state, KVM_IRQFD, &irqfd)) { 203 error_report("vfio: Error: Failed to disable INTx irqfd: %m"); 204 } 205 206 /* We only need to close the eventfd for VFIO to cleanup the kernel side */ 207 event_notifier_cleanup(&vdev->intx.unmask); 208 209 /* QEMU starts listening for interrupt events. */ 210 qemu_set_fd_handler(irqfd.fd, vfio_intx_interrupt, NULL, vdev); 211 212 vdev->intx.kvm_accel = false; 213 214 /* If we've missed an event, let it re-fire through QEMU */ 215 vfio_unmask_single_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 216 217 trace_vfio_intx_disable_kvm(vdev->vbasedev.name); 218 #endif 219 } 220 221 static void vfio_intx_update(PCIDevice *pdev) 222 { 223 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 224 PCIINTxRoute route; 225 Error *err = NULL; 226 227 if (vdev->interrupt != VFIO_INT_INTx) { 228 return; 229 } 230 231 route = pci_device_route_intx_to_irq(&vdev->pdev, vdev->intx.pin); 232 233 if (!pci_intx_route_changed(&vdev->intx.route, &route)) { 234 return; /* Nothing changed */ 235 } 236 237 trace_vfio_intx_update(vdev->vbasedev.name, 238 vdev->intx.route.irq, route.irq); 239 240 vfio_intx_disable_kvm(vdev); 241 242 vdev->intx.route = route; 243 244 if (route.mode != PCI_INTX_ENABLED) { 245 return; 246 } 247 248 vfio_intx_enable_kvm(vdev, &err); 249 if (err) { 250 error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name); 251 } 252 253 /* Re-enable the interrupt in cased we missed an EOI */ 254 vfio_intx_eoi(&vdev->vbasedev); 255 } 256 257 static int vfio_intx_enable(VFIOPCIDevice *vdev, Error **errp) 258 { 259 uint8_t pin = vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1); 260 int ret, argsz, retval = 0; 261 struct vfio_irq_set *irq_set; 262 int32_t *pfd; 263 Error *err = NULL; 264 265 if (!pin) { 266 return 0; 267 } 268 269 vfio_disable_interrupts(vdev); 270 271 vdev->intx.pin = pin - 1; /* Pin A (1) -> irq[0] */ 272 pci_config_set_interrupt_pin(vdev->pdev.config, pin); 273 274 #ifdef CONFIG_KVM 275 /* 276 * Only conditional to avoid generating error messages on platforms 277 * where we won't actually use the result anyway. 278 */ 279 if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) { 280 vdev->intx.route = pci_device_route_intx_to_irq(&vdev->pdev, 281 vdev->intx.pin); 282 } 283 #endif 284 285 ret = event_notifier_init(&vdev->intx.interrupt, 0); 286 if (ret) { 287 error_setg_errno(errp, -ret, "event_notifier_init failed"); 288 return ret; 289 } 290 291 argsz = sizeof(*irq_set) + sizeof(*pfd); 292 293 irq_set = g_malloc0(argsz); 294 irq_set->argsz = argsz; 295 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER; 296 irq_set->index = VFIO_PCI_INTX_IRQ_INDEX; 297 irq_set->start = 0; 298 irq_set->count = 1; 299 pfd = (int32_t *)&irq_set->data; 300 301 *pfd = event_notifier_get_fd(&vdev->intx.interrupt); 302 qemu_set_fd_handler(*pfd, vfio_intx_interrupt, NULL, vdev); 303 304 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 305 if (ret) { 306 error_setg_errno(errp, -ret, "failed to setup INTx fd"); 307 qemu_set_fd_handler(*pfd, NULL, NULL, vdev); 308 event_notifier_cleanup(&vdev->intx.interrupt); 309 retval = -errno; 310 goto cleanup; 311 } 312 313 vfio_intx_enable_kvm(vdev, &err); 314 if (err) { 315 error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name); 316 } 317 318 vdev->interrupt = VFIO_INT_INTx; 319 320 trace_vfio_intx_enable(vdev->vbasedev.name); 321 322 cleanup: 323 g_free(irq_set); 324 325 return retval; 326 } 327 328 static void vfio_intx_disable(VFIOPCIDevice *vdev) 329 { 330 int fd; 331 332 timer_del(vdev->intx.mmap_timer); 333 vfio_intx_disable_kvm(vdev); 334 vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_INTX_IRQ_INDEX); 335 vdev->intx.pending = false; 336 pci_irq_deassert(&vdev->pdev); 337 vfio_mmap_set_enabled(vdev, true); 338 339 fd = event_notifier_get_fd(&vdev->intx.interrupt); 340 qemu_set_fd_handler(fd, NULL, NULL, vdev); 341 event_notifier_cleanup(&vdev->intx.interrupt); 342 343 vdev->interrupt = VFIO_INT_NONE; 344 345 trace_vfio_intx_disable(vdev->vbasedev.name); 346 } 347 348 /* 349 * MSI/X 350 */ 351 static void vfio_msi_interrupt(void *opaque) 352 { 353 VFIOMSIVector *vector = opaque; 354 VFIOPCIDevice *vdev = vector->vdev; 355 MSIMessage (*get_msg)(PCIDevice *dev, unsigned vector); 356 void (*notify)(PCIDevice *dev, unsigned vector); 357 MSIMessage msg; 358 int nr = vector - vdev->msi_vectors; 359 360 if (!event_notifier_test_and_clear(&vector->interrupt)) { 361 return; 362 } 363 364 if (vdev->interrupt == VFIO_INT_MSIX) { 365 get_msg = msix_get_message; 366 notify = msix_notify; 367 368 /* A masked vector firing needs to use the PBA, enable it */ 369 if (msix_is_masked(&vdev->pdev, nr)) { 370 set_bit(nr, vdev->msix->pending); 371 memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, true); 372 trace_vfio_msix_pba_enable(vdev->vbasedev.name); 373 } 374 } else if (vdev->interrupt == VFIO_INT_MSI) { 375 get_msg = msi_get_message; 376 notify = msi_notify; 377 } else { 378 abort(); 379 } 380 381 msg = get_msg(&vdev->pdev, nr); 382 trace_vfio_msi_interrupt(vdev->vbasedev.name, nr, msg.address, msg.data); 383 notify(&vdev->pdev, nr); 384 } 385 386 static int vfio_enable_vectors(VFIOPCIDevice *vdev, bool msix) 387 { 388 struct vfio_irq_set *irq_set; 389 int ret = 0, i, argsz; 390 int32_t *fds; 391 392 argsz = sizeof(*irq_set) + (vdev->nr_vectors * sizeof(*fds)); 393 394 irq_set = g_malloc0(argsz); 395 irq_set->argsz = argsz; 396 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER; 397 irq_set->index = msix ? VFIO_PCI_MSIX_IRQ_INDEX : VFIO_PCI_MSI_IRQ_INDEX; 398 irq_set->start = 0; 399 irq_set->count = vdev->nr_vectors; 400 fds = (int32_t *)&irq_set->data; 401 402 for (i = 0; i < vdev->nr_vectors; i++) { 403 int fd = -1; 404 405 /* 406 * MSI vs MSI-X - The guest has direct access to MSI mask and pending 407 * bits, therefore we always use the KVM signaling path when setup. 408 * MSI-X mask and pending bits are emulated, so we want to use the 409 * KVM signaling path only when configured and unmasked. 410 */ 411 if (vdev->msi_vectors[i].use) { 412 if (vdev->msi_vectors[i].virq < 0 || 413 (msix && msix_is_masked(&vdev->pdev, i))) { 414 fd = event_notifier_get_fd(&vdev->msi_vectors[i].interrupt); 415 } else { 416 fd = event_notifier_get_fd(&vdev->msi_vectors[i].kvm_interrupt); 417 } 418 } 419 420 fds[i] = fd; 421 } 422 423 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 424 425 g_free(irq_set); 426 427 return ret; 428 } 429 430 static void vfio_add_kvm_msi_virq(VFIOPCIDevice *vdev, VFIOMSIVector *vector, 431 int vector_n, bool msix) 432 { 433 int virq; 434 435 if ((msix && vdev->no_kvm_msix) || (!msix && vdev->no_kvm_msi)) { 436 return; 437 } 438 439 if (event_notifier_init(&vector->kvm_interrupt, 0)) { 440 return; 441 } 442 443 virq = kvm_irqchip_add_msi_route(kvm_state, vector_n, &vdev->pdev); 444 if (virq < 0) { 445 event_notifier_cleanup(&vector->kvm_interrupt); 446 return; 447 } 448 449 if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt, 450 NULL, virq) < 0) { 451 kvm_irqchip_release_virq(kvm_state, virq); 452 event_notifier_cleanup(&vector->kvm_interrupt); 453 return; 454 } 455 456 vector->virq = virq; 457 } 458 459 static void vfio_remove_kvm_msi_virq(VFIOMSIVector *vector) 460 { 461 kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state, &vector->kvm_interrupt, 462 vector->virq); 463 kvm_irqchip_release_virq(kvm_state, vector->virq); 464 vector->virq = -1; 465 event_notifier_cleanup(&vector->kvm_interrupt); 466 } 467 468 static void vfio_update_kvm_msi_virq(VFIOMSIVector *vector, MSIMessage msg, 469 PCIDevice *pdev) 470 { 471 kvm_irqchip_update_msi_route(kvm_state, vector->virq, msg, pdev); 472 kvm_irqchip_commit_routes(kvm_state); 473 } 474 475 static int vfio_msix_vector_do_use(PCIDevice *pdev, unsigned int nr, 476 MSIMessage *msg, IOHandler *handler) 477 { 478 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 479 VFIOMSIVector *vector; 480 int ret; 481 482 trace_vfio_msix_vector_do_use(vdev->vbasedev.name, nr); 483 484 vector = &vdev->msi_vectors[nr]; 485 486 if (!vector->use) { 487 vector->vdev = vdev; 488 vector->virq = -1; 489 if (event_notifier_init(&vector->interrupt, 0)) { 490 error_report("vfio: Error: event_notifier_init failed"); 491 } 492 vector->use = true; 493 msix_vector_use(pdev, nr); 494 } 495 496 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), 497 handler, NULL, vector); 498 499 /* 500 * Attempt to enable route through KVM irqchip, 501 * default to userspace handling if unavailable. 502 */ 503 if (vector->virq >= 0) { 504 if (!msg) { 505 vfio_remove_kvm_msi_virq(vector); 506 } else { 507 vfio_update_kvm_msi_virq(vector, *msg, pdev); 508 } 509 } else { 510 if (msg) { 511 vfio_add_kvm_msi_virq(vdev, vector, nr, true); 512 } 513 } 514 515 /* 516 * We don't want to have the host allocate all possible MSI vectors 517 * for a device if they're not in use, so we shutdown and incrementally 518 * increase them as needed. 519 */ 520 if (vdev->nr_vectors < nr + 1) { 521 vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX); 522 vdev->nr_vectors = nr + 1; 523 ret = vfio_enable_vectors(vdev, true); 524 if (ret) { 525 error_report("vfio: failed to enable vectors, %d", ret); 526 } 527 } else { 528 int argsz; 529 struct vfio_irq_set *irq_set; 530 int32_t *pfd; 531 532 argsz = sizeof(*irq_set) + sizeof(*pfd); 533 534 irq_set = g_malloc0(argsz); 535 irq_set->argsz = argsz; 536 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 537 VFIO_IRQ_SET_ACTION_TRIGGER; 538 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX; 539 irq_set->start = nr; 540 irq_set->count = 1; 541 pfd = (int32_t *)&irq_set->data; 542 543 if (vector->virq >= 0) { 544 *pfd = event_notifier_get_fd(&vector->kvm_interrupt); 545 } else { 546 *pfd = event_notifier_get_fd(&vector->interrupt); 547 } 548 549 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 550 g_free(irq_set); 551 if (ret) { 552 error_report("vfio: failed to modify vector, %d", ret); 553 } 554 } 555 556 /* Disable PBA emulation when nothing more is pending. */ 557 clear_bit(nr, vdev->msix->pending); 558 if (find_first_bit(vdev->msix->pending, 559 vdev->nr_vectors) == vdev->nr_vectors) { 560 memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false); 561 trace_vfio_msix_pba_disable(vdev->vbasedev.name); 562 } 563 564 return 0; 565 } 566 567 static int vfio_msix_vector_use(PCIDevice *pdev, 568 unsigned int nr, MSIMessage msg) 569 { 570 return vfio_msix_vector_do_use(pdev, nr, &msg, vfio_msi_interrupt); 571 } 572 573 static void vfio_msix_vector_release(PCIDevice *pdev, unsigned int nr) 574 { 575 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 576 VFIOMSIVector *vector = &vdev->msi_vectors[nr]; 577 578 trace_vfio_msix_vector_release(vdev->vbasedev.name, nr); 579 580 /* 581 * There are still old guests that mask and unmask vectors on every 582 * interrupt. If we're using QEMU bypass with a KVM irqfd, leave all of 583 * the KVM setup in place, simply switch VFIO to use the non-bypass 584 * eventfd. We'll then fire the interrupt through QEMU and the MSI-X 585 * core will mask the interrupt and set pending bits, allowing it to 586 * be re-asserted on unmask. Nothing to do if already using QEMU mode. 587 */ 588 if (vector->virq >= 0) { 589 int argsz; 590 struct vfio_irq_set *irq_set; 591 int32_t *pfd; 592 593 argsz = sizeof(*irq_set) + sizeof(*pfd); 594 595 irq_set = g_malloc0(argsz); 596 irq_set->argsz = argsz; 597 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 598 VFIO_IRQ_SET_ACTION_TRIGGER; 599 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX; 600 irq_set->start = nr; 601 irq_set->count = 1; 602 pfd = (int32_t *)&irq_set->data; 603 604 *pfd = event_notifier_get_fd(&vector->interrupt); 605 606 ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 607 608 g_free(irq_set); 609 } 610 } 611 612 static void vfio_msix_enable(VFIOPCIDevice *vdev) 613 { 614 vfio_disable_interrupts(vdev); 615 616 vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->msix->entries); 617 618 vdev->interrupt = VFIO_INT_MSIX; 619 620 /* 621 * Some communication channels between VF & PF or PF & fw rely on the 622 * physical state of the device and expect that enabling MSI-X from the 623 * guest enables the same on the host. When our guest is Linux, the 624 * guest driver call to pci_enable_msix() sets the enabling bit in the 625 * MSI-X capability, but leaves the vector table masked. We therefore 626 * can't rely on a vector_use callback (from request_irq() in the guest) 627 * to switch the physical device into MSI-X mode because that may come a 628 * long time after pci_enable_msix(). This code enables vector 0 with 629 * triggering to userspace, then immediately release the vector, leaving 630 * the physical device with no vectors enabled, but MSI-X enabled, just 631 * like the guest view. 632 */ 633 vfio_msix_vector_do_use(&vdev->pdev, 0, NULL, NULL); 634 vfio_msix_vector_release(&vdev->pdev, 0); 635 636 if (msix_set_vector_notifiers(&vdev->pdev, vfio_msix_vector_use, 637 vfio_msix_vector_release, NULL)) { 638 error_report("vfio: msix_set_vector_notifiers failed"); 639 } 640 641 trace_vfio_msix_enable(vdev->vbasedev.name); 642 } 643 644 static void vfio_msi_enable(VFIOPCIDevice *vdev) 645 { 646 int ret, i; 647 648 vfio_disable_interrupts(vdev); 649 650 vdev->nr_vectors = msi_nr_vectors_allocated(&vdev->pdev); 651 retry: 652 vdev->msi_vectors = g_new0(VFIOMSIVector, vdev->nr_vectors); 653 654 for (i = 0; i < vdev->nr_vectors; i++) { 655 VFIOMSIVector *vector = &vdev->msi_vectors[i]; 656 657 vector->vdev = vdev; 658 vector->virq = -1; 659 vector->use = true; 660 661 if (event_notifier_init(&vector->interrupt, 0)) { 662 error_report("vfio: Error: event_notifier_init failed"); 663 } 664 665 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), 666 vfio_msi_interrupt, NULL, vector); 667 668 /* 669 * Attempt to enable route through KVM irqchip, 670 * default to userspace handling if unavailable. 671 */ 672 vfio_add_kvm_msi_virq(vdev, vector, i, false); 673 } 674 675 /* Set interrupt type prior to possible interrupts */ 676 vdev->interrupt = VFIO_INT_MSI; 677 678 ret = vfio_enable_vectors(vdev, false); 679 if (ret) { 680 if (ret < 0) { 681 error_report("vfio: Error: Failed to setup MSI fds: %m"); 682 } else if (ret != vdev->nr_vectors) { 683 error_report("vfio: Error: Failed to enable %d " 684 "MSI vectors, retry with %d", vdev->nr_vectors, ret); 685 } 686 687 for (i = 0; i < vdev->nr_vectors; i++) { 688 VFIOMSIVector *vector = &vdev->msi_vectors[i]; 689 if (vector->virq >= 0) { 690 vfio_remove_kvm_msi_virq(vector); 691 } 692 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), 693 NULL, NULL, NULL); 694 event_notifier_cleanup(&vector->interrupt); 695 } 696 697 g_free(vdev->msi_vectors); 698 699 if (ret > 0 && ret != vdev->nr_vectors) { 700 vdev->nr_vectors = ret; 701 goto retry; 702 } 703 vdev->nr_vectors = 0; 704 705 /* 706 * Failing to setup MSI doesn't really fall within any specification. 707 * Let's try leaving interrupts disabled and hope the guest figures 708 * out to fall back to INTx for this device. 709 */ 710 error_report("vfio: Error: Failed to enable MSI"); 711 vdev->interrupt = VFIO_INT_NONE; 712 713 return; 714 } 715 716 trace_vfio_msi_enable(vdev->vbasedev.name, vdev->nr_vectors); 717 } 718 719 static void vfio_msi_disable_common(VFIOPCIDevice *vdev) 720 { 721 Error *err = NULL; 722 int i; 723 724 for (i = 0; i < vdev->nr_vectors; i++) { 725 VFIOMSIVector *vector = &vdev->msi_vectors[i]; 726 if (vdev->msi_vectors[i].use) { 727 if (vector->virq >= 0) { 728 vfio_remove_kvm_msi_virq(vector); 729 } 730 qemu_set_fd_handler(event_notifier_get_fd(&vector->interrupt), 731 NULL, NULL, NULL); 732 event_notifier_cleanup(&vector->interrupt); 733 } 734 } 735 736 g_free(vdev->msi_vectors); 737 vdev->msi_vectors = NULL; 738 vdev->nr_vectors = 0; 739 vdev->interrupt = VFIO_INT_NONE; 740 741 vfio_intx_enable(vdev, &err); 742 if (err) { 743 error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name); 744 } 745 } 746 747 static void vfio_msix_disable(VFIOPCIDevice *vdev) 748 { 749 int i; 750 751 msix_unset_vector_notifiers(&vdev->pdev); 752 753 /* 754 * MSI-X will only release vectors if MSI-X is still enabled on the 755 * device, check through the rest and release it ourselves if necessary. 756 */ 757 for (i = 0; i < vdev->nr_vectors; i++) { 758 if (vdev->msi_vectors[i].use) { 759 vfio_msix_vector_release(&vdev->pdev, i); 760 msix_vector_unuse(&vdev->pdev, i); 761 } 762 } 763 764 if (vdev->nr_vectors) { 765 vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSIX_IRQ_INDEX); 766 } 767 768 vfio_msi_disable_common(vdev); 769 770 memset(vdev->msix->pending, 0, 771 BITS_TO_LONGS(vdev->msix->entries) * sizeof(unsigned long)); 772 773 trace_vfio_msix_disable(vdev->vbasedev.name); 774 } 775 776 static void vfio_msi_disable(VFIOPCIDevice *vdev) 777 { 778 vfio_disable_irqindex(&vdev->vbasedev, VFIO_PCI_MSI_IRQ_INDEX); 779 vfio_msi_disable_common(vdev); 780 781 trace_vfio_msi_disable(vdev->vbasedev.name); 782 } 783 784 static void vfio_update_msi(VFIOPCIDevice *vdev) 785 { 786 int i; 787 788 for (i = 0; i < vdev->nr_vectors; i++) { 789 VFIOMSIVector *vector = &vdev->msi_vectors[i]; 790 MSIMessage msg; 791 792 if (!vector->use || vector->virq < 0) { 793 continue; 794 } 795 796 msg = msi_get_message(&vdev->pdev, i); 797 vfio_update_kvm_msi_virq(vector, msg, &vdev->pdev); 798 } 799 } 800 801 static void vfio_pci_load_rom(VFIOPCIDevice *vdev) 802 { 803 struct vfio_region_info *reg_info; 804 uint64_t size; 805 off_t off = 0; 806 ssize_t bytes; 807 808 if (vfio_get_region_info(&vdev->vbasedev, 809 VFIO_PCI_ROM_REGION_INDEX, ®_info)) { 810 error_report("vfio: Error getting ROM info: %m"); 811 return; 812 } 813 814 trace_vfio_pci_load_rom(vdev->vbasedev.name, (unsigned long)reg_info->size, 815 (unsigned long)reg_info->offset, 816 (unsigned long)reg_info->flags); 817 818 vdev->rom_size = size = reg_info->size; 819 vdev->rom_offset = reg_info->offset; 820 821 g_free(reg_info); 822 823 if (!vdev->rom_size) { 824 vdev->rom_read_failed = true; 825 error_report("vfio-pci: Cannot read device rom at " 826 "%s", vdev->vbasedev.name); 827 error_printf("Device option ROM contents are probably invalid " 828 "(check dmesg).\nSkip option ROM probe with rombar=0, " 829 "or load from file with romfile=\n"); 830 return; 831 } 832 833 vdev->rom = g_malloc(size); 834 memset(vdev->rom, 0xff, size); 835 836 while (size) { 837 bytes = pread(vdev->vbasedev.fd, vdev->rom + off, 838 size, vdev->rom_offset + off); 839 if (bytes == 0) { 840 break; 841 } else if (bytes > 0) { 842 off += bytes; 843 size -= bytes; 844 } else { 845 if (errno == EINTR || errno == EAGAIN) { 846 continue; 847 } 848 error_report("vfio: Error reading device ROM: %m"); 849 break; 850 } 851 } 852 853 /* 854 * Test the ROM signature against our device, if the vendor is correct 855 * but the device ID doesn't match, store the correct device ID and 856 * recompute the checksum. Intel IGD devices need this and are known 857 * to have bogus checksums so we can't simply adjust the checksum. 858 */ 859 if (pci_get_word(vdev->rom) == 0xaa55 && 860 pci_get_word(vdev->rom + 0x18) + 8 < vdev->rom_size && 861 !memcmp(vdev->rom + pci_get_word(vdev->rom + 0x18), "PCIR", 4)) { 862 uint16_t vid, did; 863 864 vid = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 4); 865 did = pci_get_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6); 866 867 if (vid == vdev->vendor_id && did != vdev->device_id) { 868 int i; 869 uint8_t csum, *data = vdev->rom; 870 871 pci_set_word(vdev->rom + pci_get_word(vdev->rom + 0x18) + 6, 872 vdev->device_id); 873 data[6] = 0; 874 875 for (csum = 0, i = 0; i < vdev->rom_size; i++) { 876 csum += data[i]; 877 } 878 879 data[6] = -csum; 880 } 881 } 882 } 883 884 static uint64_t vfio_rom_read(void *opaque, hwaddr addr, unsigned size) 885 { 886 VFIOPCIDevice *vdev = opaque; 887 union { 888 uint8_t byte; 889 uint16_t word; 890 uint32_t dword; 891 uint64_t qword; 892 } val; 893 uint64_t data = 0; 894 895 /* Load the ROM lazily when the guest tries to read it */ 896 if (unlikely(!vdev->rom && !vdev->rom_read_failed)) { 897 vfio_pci_load_rom(vdev); 898 } 899 900 memcpy(&val, vdev->rom + addr, 901 (addr < vdev->rom_size) ? MIN(size, vdev->rom_size - addr) : 0); 902 903 switch (size) { 904 case 1: 905 data = val.byte; 906 break; 907 case 2: 908 data = le16_to_cpu(val.word); 909 break; 910 case 4: 911 data = le32_to_cpu(val.dword); 912 break; 913 default: 914 hw_error("vfio: unsupported read size, %d bytes\n", size); 915 break; 916 } 917 918 trace_vfio_rom_read(vdev->vbasedev.name, addr, size, data); 919 920 return data; 921 } 922 923 static void vfio_rom_write(void *opaque, hwaddr addr, 924 uint64_t data, unsigned size) 925 { 926 } 927 928 static const MemoryRegionOps vfio_rom_ops = { 929 .read = vfio_rom_read, 930 .write = vfio_rom_write, 931 .endianness = DEVICE_LITTLE_ENDIAN, 932 }; 933 934 static void vfio_pci_size_rom(VFIOPCIDevice *vdev) 935 { 936 uint32_t orig, size = cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK); 937 off_t offset = vdev->config_offset + PCI_ROM_ADDRESS; 938 DeviceState *dev = DEVICE(vdev); 939 char *name; 940 int fd = vdev->vbasedev.fd; 941 942 if (vdev->pdev.romfile || !vdev->pdev.rom_bar) { 943 /* Since pci handles romfile, just print a message and return */ 944 if (vfio_blacklist_opt_rom(vdev) && vdev->pdev.romfile) { 945 error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified romfile\n", 946 vdev->vbasedev.name); 947 } 948 return; 949 } 950 951 /* 952 * Use the same size ROM BAR as the physical device. The contents 953 * will get filled in later when the guest tries to read it. 954 */ 955 if (pread(fd, &orig, 4, offset) != 4 || 956 pwrite(fd, &size, 4, offset) != 4 || 957 pread(fd, &size, 4, offset) != 4 || 958 pwrite(fd, &orig, 4, offset) != 4) { 959 error_report("%s(%s) failed: %m", __func__, vdev->vbasedev.name); 960 return; 961 } 962 963 size = ~(le32_to_cpu(size) & PCI_ROM_ADDRESS_MASK) + 1; 964 965 if (!size) { 966 return; 967 } 968 969 if (vfio_blacklist_opt_rom(vdev)) { 970 if (dev->opts && qemu_opt_get(dev->opts, "rombar")) { 971 error_printf("Warning : Device at %s is known to cause system instability issues during option rom execution. Proceeding anyway since user specified non zero value for rombar\n", 972 vdev->vbasedev.name); 973 } else { 974 error_printf("Warning : Rom loading for device at %s has been disabled due to system instability issues. Specify rombar=1 or romfile to force\n", 975 vdev->vbasedev.name); 976 return; 977 } 978 } 979 980 trace_vfio_pci_size_rom(vdev->vbasedev.name, size); 981 982 name = g_strdup_printf("vfio[%s].rom", vdev->vbasedev.name); 983 984 memory_region_init_io(&vdev->pdev.rom, OBJECT(vdev), 985 &vfio_rom_ops, vdev, name, size); 986 g_free(name); 987 988 pci_register_bar(&vdev->pdev, PCI_ROM_SLOT, 989 PCI_BASE_ADDRESS_SPACE_MEMORY, &vdev->pdev.rom); 990 991 vdev->pdev.has_rom = true; 992 vdev->rom_read_failed = false; 993 } 994 995 void vfio_vga_write(void *opaque, hwaddr addr, 996 uint64_t data, unsigned size) 997 { 998 VFIOVGARegion *region = opaque; 999 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]); 1000 union { 1001 uint8_t byte; 1002 uint16_t word; 1003 uint32_t dword; 1004 uint64_t qword; 1005 } buf; 1006 off_t offset = vga->fd_offset + region->offset + addr; 1007 1008 switch (size) { 1009 case 1: 1010 buf.byte = data; 1011 break; 1012 case 2: 1013 buf.word = cpu_to_le16(data); 1014 break; 1015 case 4: 1016 buf.dword = cpu_to_le32(data); 1017 break; 1018 default: 1019 hw_error("vfio: unsupported write size, %d bytes", size); 1020 break; 1021 } 1022 1023 if (pwrite(vga->fd, &buf, size, offset) != size) { 1024 error_report("%s(,0x%"HWADDR_PRIx", 0x%"PRIx64", %d) failed: %m", 1025 __func__, region->offset + addr, data, size); 1026 } 1027 1028 trace_vfio_vga_write(region->offset + addr, data, size); 1029 } 1030 1031 uint64_t vfio_vga_read(void *opaque, hwaddr addr, unsigned size) 1032 { 1033 VFIOVGARegion *region = opaque; 1034 VFIOVGA *vga = container_of(region, VFIOVGA, region[region->nr]); 1035 union { 1036 uint8_t byte; 1037 uint16_t word; 1038 uint32_t dword; 1039 uint64_t qword; 1040 } buf; 1041 uint64_t data = 0; 1042 off_t offset = vga->fd_offset + region->offset + addr; 1043 1044 if (pread(vga->fd, &buf, size, offset) != size) { 1045 error_report("%s(,0x%"HWADDR_PRIx", %d) failed: %m", 1046 __func__, region->offset + addr, size); 1047 return (uint64_t)-1; 1048 } 1049 1050 switch (size) { 1051 case 1: 1052 data = buf.byte; 1053 break; 1054 case 2: 1055 data = le16_to_cpu(buf.word); 1056 break; 1057 case 4: 1058 data = le32_to_cpu(buf.dword); 1059 break; 1060 default: 1061 hw_error("vfio: unsupported read size, %d bytes", size); 1062 break; 1063 } 1064 1065 trace_vfio_vga_read(region->offset + addr, size, data); 1066 1067 return data; 1068 } 1069 1070 static const MemoryRegionOps vfio_vga_ops = { 1071 .read = vfio_vga_read, 1072 .write = vfio_vga_write, 1073 .endianness = DEVICE_LITTLE_ENDIAN, 1074 }; 1075 1076 /* 1077 * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page 1078 * size if the BAR is in an exclusive page in host so that we could map 1079 * this BAR to guest. But this sub-page BAR may not occupy an exclusive 1080 * page in guest. So we should set the priority of the expanded memory 1081 * region to zero in case of overlap with BARs which share the same page 1082 * with the sub-page BAR in guest. Besides, we should also recover the 1083 * size of this sub-page BAR when its base address is changed in guest 1084 * and not page aligned any more. 1085 */ 1086 static void vfio_sub_page_bar_update_mapping(PCIDevice *pdev, int bar) 1087 { 1088 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 1089 VFIORegion *region = &vdev->bars[bar].region; 1090 MemoryRegion *mmap_mr, *mr; 1091 PCIIORegion *r; 1092 pcibus_t bar_addr; 1093 uint64_t size = region->size; 1094 1095 /* Make sure that the whole region is allowed to be mmapped */ 1096 if (region->nr_mmaps != 1 || !region->mmaps[0].mmap || 1097 region->mmaps[0].size != region->size) { 1098 return; 1099 } 1100 1101 r = &pdev->io_regions[bar]; 1102 bar_addr = r->addr; 1103 mr = region->mem; 1104 mmap_mr = ®ion->mmaps[0].mem; 1105 1106 /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */ 1107 if (bar_addr != PCI_BAR_UNMAPPED && 1108 !(bar_addr & ~qemu_real_host_page_mask)) { 1109 size = qemu_real_host_page_size; 1110 } 1111 1112 memory_region_transaction_begin(); 1113 1114 memory_region_set_size(mr, size); 1115 memory_region_set_size(mmap_mr, size); 1116 if (size != region->size && memory_region_is_mapped(mr)) { 1117 memory_region_del_subregion(r->address_space, mr); 1118 memory_region_add_subregion_overlap(r->address_space, 1119 bar_addr, mr, 0); 1120 } 1121 1122 memory_region_transaction_commit(); 1123 } 1124 1125 /* 1126 * PCI config space 1127 */ 1128 uint32_t vfio_pci_read_config(PCIDevice *pdev, uint32_t addr, int len) 1129 { 1130 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 1131 uint32_t emu_bits = 0, emu_val = 0, phys_val = 0, val; 1132 1133 memcpy(&emu_bits, vdev->emulated_config_bits + addr, len); 1134 emu_bits = le32_to_cpu(emu_bits); 1135 1136 if (emu_bits) { 1137 emu_val = pci_default_read_config(pdev, addr, len); 1138 } 1139 1140 if (~emu_bits & (0xffffffffU >> (32 - len * 8))) { 1141 ssize_t ret; 1142 1143 ret = pread(vdev->vbasedev.fd, &phys_val, len, 1144 vdev->config_offset + addr); 1145 if (ret != len) { 1146 error_report("%s(%s, 0x%x, 0x%x) failed: %m", 1147 __func__, vdev->vbasedev.name, addr, len); 1148 return -errno; 1149 } 1150 phys_val = le32_to_cpu(phys_val); 1151 } 1152 1153 val = (emu_val & emu_bits) | (phys_val & ~emu_bits); 1154 1155 trace_vfio_pci_read_config(vdev->vbasedev.name, addr, len, val); 1156 1157 return val; 1158 } 1159 1160 void vfio_pci_write_config(PCIDevice *pdev, 1161 uint32_t addr, uint32_t val, int len) 1162 { 1163 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 1164 uint32_t val_le = cpu_to_le32(val); 1165 1166 trace_vfio_pci_write_config(vdev->vbasedev.name, addr, val, len); 1167 1168 /* Write everything to VFIO, let it filter out what we can't write */ 1169 if (pwrite(vdev->vbasedev.fd, &val_le, len, vdev->config_offset + addr) 1170 != len) { 1171 error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m", 1172 __func__, vdev->vbasedev.name, addr, val, len); 1173 } 1174 1175 /* MSI/MSI-X Enabling/Disabling */ 1176 if (pdev->cap_present & QEMU_PCI_CAP_MSI && 1177 ranges_overlap(addr, len, pdev->msi_cap, vdev->msi_cap_size)) { 1178 int is_enabled, was_enabled = msi_enabled(pdev); 1179 1180 pci_default_write_config(pdev, addr, val, len); 1181 1182 is_enabled = msi_enabled(pdev); 1183 1184 if (!was_enabled) { 1185 if (is_enabled) { 1186 vfio_msi_enable(vdev); 1187 } 1188 } else { 1189 if (!is_enabled) { 1190 vfio_msi_disable(vdev); 1191 } else { 1192 vfio_update_msi(vdev); 1193 } 1194 } 1195 } else if (pdev->cap_present & QEMU_PCI_CAP_MSIX && 1196 ranges_overlap(addr, len, pdev->msix_cap, MSIX_CAP_LENGTH)) { 1197 int is_enabled, was_enabled = msix_enabled(pdev); 1198 1199 pci_default_write_config(pdev, addr, val, len); 1200 1201 is_enabled = msix_enabled(pdev); 1202 1203 if (!was_enabled && is_enabled) { 1204 vfio_msix_enable(vdev); 1205 } else if (was_enabled && !is_enabled) { 1206 vfio_msix_disable(vdev); 1207 } 1208 } else if (ranges_overlap(addr, len, PCI_BASE_ADDRESS_0, 24) || 1209 range_covers_byte(addr, len, PCI_COMMAND)) { 1210 pcibus_t old_addr[PCI_NUM_REGIONS - 1]; 1211 int bar; 1212 1213 for (bar = 0; bar < PCI_ROM_SLOT; bar++) { 1214 old_addr[bar] = pdev->io_regions[bar].addr; 1215 } 1216 1217 pci_default_write_config(pdev, addr, val, len); 1218 1219 for (bar = 0; bar < PCI_ROM_SLOT; bar++) { 1220 if (old_addr[bar] != pdev->io_regions[bar].addr && 1221 pdev->io_regions[bar].size > 0 && 1222 pdev->io_regions[bar].size < qemu_real_host_page_size) { 1223 vfio_sub_page_bar_update_mapping(pdev, bar); 1224 } 1225 } 1226 } else { 1227 /* Write everything to QEMU to keep emulated bits correct */ 1228 pci_default_write_config(pdev, addr, val, len); 1229 } 1230 } 1231 1232 /* 1233 * Interrupt setup 1234 */ 1235 static void vfio_disable_interrupts(VFIOPCIDevice *vdev) 1236 { 1237 /* 1238 * More complicated than it looks. Disabling MSI/X transitions the 1239 * device to INTx mode (if supported). Therefore we need to first 1240 * disable MSI/X and then cleanup by disabling INTx. 1241 */ 1242 if (vdev->interrupt == VFIO_INT_MSIX) { 1243 vfio_msix_disable(vdev); 1244 } else if (vdev->interrupt == VFIO_INT_MSI) { 1245 vfio_msi_disable(vdev); 1246 } 1247 1248 if (vdev->interrupt == VFIO_INT_INTx) { 1249 vfio_intx_disable(vdev); 1250 } 1251 } 1252 1253 static int vfio_msi_setup(VFIOPCIDevice *vdev, int pos, Error **errp) 1254 { 1255 uint16_t ctrl; 1256 bool msi_64bit, msi_maskbit; 1257 int ret, entries; 1258 Error *err = NULL; 1259 1260 if (pread(vdev->vbasedev.fd, &ctrl, sizeof(ctrl), 1261 vdev->config_offset + pos + PCI_CAP_FLAGS) != sizeof(ctrl)) { 1262 error_setg_errno(errp, errno, "failed reading MSI PCI_CAP_FLAGS"); 1263 return -errno; 1264 } 1265 ctrl = le16_to_cpu(ctrl); 1266 1267 msi_64bit = !!(ctrl & PCI_MSI_FLAGS_64BIT); 1268 msi_maskbit = !!(ctrl & PCI_MSI_FLAGS_MASKBIT); 1269 entries = 1 << ((ctrl & PCI_MSI_FLAGS_QMASK) >> 1); 1270 1271 trace_vfio_msi_setup(vdev->vbasedev.name, pos); 1272 1273 ret = msi_init(&vdev->pdev, pos, entries, msi_64bit, msi_maskbit, &err); 1274 if (ret < 0) { 1275 if (ret == -ENOTSUP) { 1276 return 0; 1277 } 1278 error_prepend(&err, "msi_init failed: "); 1279 error_propagate(errp, err); 1280 return ret; 1281 } 1282 vdev->msi_cap_size = 0xa + (msi_maskbit ? 0xa : 0) + (msi_64bit ? 0x4 : 0); 1283 1284 return 0; 1285 } 1286 1287 static void vfio_pci_fixup_msix_region(VFIOPCIDevice *vdev) 1288 { 1289 off_t start, end; 1290 VFIORegion *region = &vdev->bars[vdev->msix->table_bar].region; 1291 1292 /* 1293 * We expect to find a single mmap covering the whole BAR, anything else 1294 * means it's either unsupported or already setup. 1295 */ 1296 if (region->nr_mmaps != 1 || region->mmaps[0].offset || 1297 region->size != region->mmaps[0].size) { 1298 return; 1299 } 1300 1301 /* MSI-X table start and end aligned to host page size */ 1302 start = vdev->msix->table_offset & qemu_real_host_page_mask; 1303 end = REAL_HOST_PAGE_ALIGN((uint64_t)vdev->msix->table_offset + 1304 (vdev->msix->entries * PCI_MSIX_ENTRY_SIZE)); 1305 1306 /* 1307 * Does the MSI-X table cover the beginning of the BAR? The whole BAR? 1308 * NB - Host page size is necessarily a power of two and so is the PCI 1309 * BAR (not counting EA yet), therefore if we have host page aligned 1310 * @start and @end, then any remainder of the BAR before or after those 1311 * must be at least host page sized and therefore mmap'able. 1312 */ 1313 if (!start) { 1314 if (end >= region->size) { 1315 region->nr_mmaps = 0; 1316 g_free(region->mmaps); 1317 region->mmaps = NULL; 1318 trace_vfio_msix_fixup(vdev->vbasedev.name, 1319 vdev->msix->table_bar, 0, 0); 1320 } else { 1321 region->mmaps[0].offset = end; 1322 region->mmaps[0].size = region->size - end; 1323 trace_vfio_msix_fixup(vdev->vbasedev.name, 1324 vdev->msix->table_bar, region->mmaps[0].offset, 1325 region->mmaps[0].offset + region->mmaps[0].size); 1326 } 1327 1328 /* Maybe it's aligned at the end of the BAR */ 1329 } else if (end >= region->size) { 1330 region->mmaps[0].size = start; 1331 trace_vfio_msix_fixup(vdev->vbasedev.name, 1332 vdev->msix->table_bar, region->mmaps[0].offset, 1333 region->mmaps[0].offset + region->mmaps[0].size); 1334 1335 /* Otherwise it must split the BAR */ 1336 } else { 1337 region->nr_mmaps = 2; 1338 region->mmaps = g_renew(VFIOMmap, region->mmaps, 2); 1339 1340 memcpy(®ion->mmaps[1], ®ion->mmaps[0], sizeof(VFIOMmap)); 1341 1342 region->mmaps[0].size = start; 1343 trace_vfio_msix_fixup(vdev->vbasedev.name, 1344 vdev->msix->table_bar, region->mmaps[0].offset, 1345 region->mmaps[0].offset + region->mmaps[0].size); 1346 1347 region->mmaps[1].offset = end; 1348 region->mmaps[1].size = region->size - end; 1349 trace_vfio_msix_fixup(vdev->vbasedev.name, 1350 vdev->msix->table_bar, region->mmaps[1].offset, 1351 region->mmaps[1].offset + region->mmaps[1].size); 1352 } 1353 } 1354 1355 /* 1356 * We don't have any control over how pci_add_capability() inserts 1357 * capabilities into the chain. In order to setup MSI-X we need a 1358 * MemoryRegion for the BAR. In order to setup the BAR and not 1359 * attempt to mmap the MSI-X table area, which VFIO won't allow, we 1360 * need to first look for where the MSI-X table lives. So we 1361 * unfortunately split MSI-X setup across two functions. 1362 */ 1363 static void vfio_msix_early_setup(VFIOPCIDevice *vdev, Error **errp) 1364 { 1365 uint8_t pos; 1366 uint16_t ctrl; 1367 uint32_t table, pba; 1368 int fd = vdev->vbasedev.fd; 1369 VFIOMSIXInfo *msix; 1370 1371 pos = pci_find_capability(&vdev->pdev, PCI_CAP_ID_MSIX); 1372 if (!pos) { 1373 return; 1374 } 1375 1376 if (pread(fd, &ctrl, sizeof(ctrl), 1377 vdev->config_offset + pos + PCI_MSIX_FLAGS) != sizeof(ctrl)) { 1378 error_setg_errno(errp, errno, "failed to read PCI MSIX FLAGS"); 1379 return; 1380 } 1381 1382 if (pread(fd, &table, sizeof(table), 1383 vdev->config_offset + pos + PCI_MSIX_TABLE) != sizeof(table)) { 1384 error_setg_errno(errp, errno, "failed to read PCI MSIX TABLE"); 1385 return; 1386 } 1387 1388 if (pread(fd, &pba, sizeof(pba), 1389 vdev->config_offset + pos + PCI_MSIX_PBA) != sizeof(pba)) { 1390 error_setg_errno(errp, errno, "failed to read PCI MSIX PBA"); 1391 return; 1392 } 1393 1394 ctrl = le16_to_cpu(ctrl); 1395 table = le32_to_cpu(table); 1396 pba = le32_to_cpu(pba); 1397 1398 msix = g_malloc0(sizeof(*msix)); 1399 msix->table_bar = table & PCI_MSIX_FLAGS_BIRMASK; 1400 msix->table_offset = table & ~PCI_MSIX_FLAGS_BIRMASK; 1401 msix->pba_bar = pba & PCI_MSIX_FLAGS_BIRMASK; 1402 msix->pba_offset = pba & ~PCI_MSIX_FLAGS_BIRMASK; 1403 msix->entries = (ctrl & PCI_MSIX_FLAGS_QSIZE) + 1; 1404 1405 /* 1406 * Test the size of the pba_offset variable and catch if it extends outside 1407 * of the specified BAR. If it is the case, we need to apply a hardware 1408 * specific quirk if the device is known or we have a broken configuration. 1409 */ 1410 if (msix->pba_offset >= vdev->bars[msix->pba_bar].region.size) { 1411 /* 1412 * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5 1413 * adapters. The T5 hardware returns an incorrect value of 0x8000 for 1414 * the VF PBA offset while the BAR itself is only 8k. The correct value 1415 * is 0x1000, so we hard code that here. 1416 */ 1417 if (vdev->vendor_id == PCI_VENDOR_ID_CHELSIO && 1418 (vdev->device_id & 0xff00) == 0x5800) { 1419 msix->pba_offset = 0x1000; 1420 } else { 1421 error_setg(errp, "hardware reports invalid configuration, " 1422 "MSIX PBA outside of specified BAR"); 1423 g_free(msix); 1424 return; 1425 } 1426 } 1427 1428 trace_vfio_msix_early_setup(vdev->vbasedev.name, pos, msix->table_bar, 1429 msix->table_offset, msix->entries); 1430 vdev->msix = msix; 1431 1432 vfio_pci_fixup_msix_region(vdev); 1433 } 1434 1435 static int vfio_msix_setup(VFIOPCIDevice *vdev, int pos, Error **errp) 1436 { 1437 int ret; 1438 Error *err = NULL; 1439 1440 vdev->msix->pending = g_malloc0(BITS_TO_LONGS(vdev->msix->entries) * 1441 sizeof(unsigned long)); 1442 ret = msix_init(&vdev->pdev, vdev->msix->entries, 1443 vdev->bars[vdev->msix->table_bar].region.mem, 1444 vdev->msix->table_bar, vdev->msix->table_offset, 1445 vdev->bars[vdev->msix->pba_bar].region.mem, 1446 vdev->msix->pba_bar, vdev->msix->pba_offset, pos, 1447 &err); 1448 if (ret < 0) { 1449 if (ret == -ENOTSUP) { 1450 error_report_err(err); 1451 return 0; 1452 } 1453 1454 error_propagate(errp, err); 1455 return ret; 1456 } 1457 1458 /* 1459 * The PCI spec suggests that devices provide additional alignment for 1460 * MSI-X structures and avoid overlapping non-MSI-X related registers. 1461 * For an assigned device, this hopefully means that emulation of MSI-X 1462 * structures does not affect the performance of the device. If devices 1463 * fail to provide that alignment, a significant performance penalty may 1464 * result, for instance Mellanox MT27500 VFs: 1465 * http://www.spinics.net/lists/kvm/msg125881.html 1466 * 1467 * The PBA is simply not that important for such a serious regression and 1468 * most drivers do not appear to look at it. The solution for this is to 1469 * disable the PBA MemoryRegion unless it's being used. We disable it 1470 * here and only enable it if a masked vector fires through QEMU. As the 1471 * vector-use notifier is called, which occurs on unmask, we test whether 1472 * PBA emulation is needed and again disable if not. 1473 */ 1474 memory_region_set_enabled(&vdev->pdev.msix_pba_mmio, false); 1475 1476 return 0; 1477 } 1478 1479 static void vfio_teardown_msi(VFIOPCIDevice *vdev) 1480 { 1481 msi_uninit(&vdev->pdev); 1482 1483 if (vdev->msix) { 1484 msix_uninit(&vdev->pdev, 1485 vdev->bars[vdev->msix->table_bar].region.mem, 1486 vdev->bars[vdev->msix->pba_bar].region.mem); 1487 g_free(vdev->msix->pending); 1488 } 1489 } 1490 1491 /* 1492 * Resource setup 1493 */ 1494 static void vfio_mmap_set_enabled(VFIOPCIDevice *vdev, bool enabled) 1495 { 1496 int i; 1497 1498 for (i = 0; i < PCI_ROM_SLOT; i++) { 1499 vfio_region_mmaps_set_enabled(&vdev->bars[i].region, enabled); 1500 } 1501 } 1502 1503 static void vfio_bar_setup(VFIOPCIDevice *vdev, int nr) 1504 { 1505 VFIOBAR *bar = &vdev->bars[nr]; 1506 1507 uint32_t pci_bar; 1508 uint8_t type; 1509 int ret; 1510 1511 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */ 1512 if (!bar->region.size) { 1513 return; 1514 } 1515 1516 /* Determine what type of BAR this is for registration */ 1517 ret = pread(vdev->vbasedev.fd, &pci_bar, sizeof(pci_bar), 1518 vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr)); 1519 if (ret != sizeof(pci_bar)) { 1520 error_report("vfio: Failed to read BAR %d (%m)", nr); 1521 return; 1522 } 1523 1524 pci_bar = le32_to_cpu(pci_bar); 1525 bar->ioport = (pci_bar & PCI_BASE_ADDRESS_SPACE_IO); 1526 bar->mem64 = bar->ioport ? 0 : (pci_bar & PCI_BASE_ADDRESS_MEM_TYPE_64); 1527 type = pci_bar & (bar->ioport ? ~PCI_BASE_ADDRESS_IO_MASK : 1528 ~PCI_BASE_ADDRESS_MEM_MASK); 1529 1530 if (vfio_region_mmap(&bar->region)) { 1531 error_report("Failed to mmap %s BAR %d. Performance may be slow", 1532 vdev->vbasedev.name, nr); 1533 } 1534 1535 pci_register_bar(&vdev->pdev, nr, type, bar->region.mem); 1536 } 1537 1538 static void vfio_bars_setup(VFIOPCIDevice *vdev) 1539 { 1540 int i; 1541 1542 for (i = 0; i < PCI_ROM_SLOT; i++) { 1543 vfio_bar_setup(vdev, i); 1544 } 1545 } 1546 1547 static void vfio_bars_exit(VFIOPCIDevice *vdev) 1548 { 1549 int i; 1550 1551 for (i = 0; i < PCI_ROM_SLOT; i++) { 1552 vfio_bar_quirk_exit(vdev, i); 1553 vfio_region_exit(&vdev->bars[i].region); 1554 } 1555 1556 if (vdev->vga) { 1557 pci_unregister_vga(&vdev->pdev); 1558 vfio_vga_quirk_exit(vdev); 1559 } 1560 } 1561 1562 static void vfio_bars_finalize(VFIOPCIDevice *vdev) 1563 { 1564 int i; 1565 1566 for (i = 0; i < PCI_ROM_SLOT; i++) { 1567 vfio_bar_quirk_finalize(vdev, i); 1568 vfio_region_finalize(&vdev->bars[i].region); 1569 } 1570 1571 if (vdev->vga) { 1572 vfio_vga_quirk_finalize(vdev); 1573 for (i = 0; i < ARRAY_SIZE(vdev->vga->region); i++) { 1574 object_unparent(OBJECT(&vdev->vga->region[i].mem)); 1575 } 1576 g_free(vdev->vga); 1577 } 1578 } 1579 1580 /* 1581 * General setup 1582 */ 1583 static uint8_t vfio_std_cap_max_size(PCIDevice *pdev, uint8_t pos) 1584 { 1585 uint8_t tmp; 1586 uint16_t next = PCI_CONFIG_SPACE_SIZE; 1587 1588 for (tmp = pdev->config[PCI_CAPABILITY_LIST]; tmp; 1589 tmp = pdev->config[tmp + PCI_CAP_LIST_NEXT]) { 1590 if (tmp > pos && tmp < next) { 1591 next = tmp; 1592 } 1593 } 1594 1595 return next - pos; 1596 } 1597 1598 1599 static uint16_t vfio_ext_cap_max_size(const uint8_t *config, uint16_t pos) 1600 { 1601 uint16_t tmp, next = PCIE_CONFIG_SPACE_SIZE; 1602 1603 for (tmp = PCI_CONFIG_SPACE_SIZE; tmp; 1604 tmp = PCI_EXT_CAP_NEXT(pci_get_long(config + tmp))) { 1605 if (tmp > pos && tmp < next) { 1606 next = tmp; 1607 } 1608 } 1609 1610 return next - pos; 1611 } 1612 1613 static void vfio_set_word_bits(uint8_t *buf, uint16_t val, uint16_t mask) 1614 { 1615 pci_set_word(buf, (pci_get_word(buf) & ~mask) | val); 1616 } 1617 1618 static void vfio_add_emulated_word(VFIOPCIDevice *vdev, int pos, 1619 uint16_t val, uint16_t mask) 1620 { 1621 vfio_set_word_bits(vdev->pdev.config + pos, val, mask); 1622 vfio_set_word_bits(vdev->pdev.wmask + pos, ~mask, mask); 1623 vfio_set_word_bits(vdev->emulated_config_bits + pos, mask, mask); 1624 } 1625 1626 static void vfio_set_long_bits(uint8_t *buf, uint32_t val, uint32_t mask) 1627 { 1628 pci_set_long(buf, (pci_get_long(buf) & ~mask) | val); 1629 } 1630 1631 static void vfio_add_emulated_long(VFIOPCIDevice *vdev, int pos, 1632 uint32_t val, uint32_t mask) 1633 { 1634 vfio_set_long_bits(vdev->pdev.config + pos, val, mask); 1635 vfio_set_long_bits(vdev->pdev.wmask + pos, ~mask, mask); 1636 vfio_set_long_bits(vdev->emulated_config_bits + pos, mask, mask); 1637 } 1638 1639 static int vfio_setup_pcie_cap(VFIOPCIDevice *vdev, int pos, uint8_t size, 1640 Error **errp) 1641 { 1642 uint16_t flags; 1643 uint8_t type; 1644 1645 flags = pci_get_word(vdev->pdev.config + pos + PCI_CAP_FLAGS); 1646 type = (flags & PCI_EXP_FLAGS_TYPE) >> 4; 1647 1648 if (type != PCI_EXP_TYPE_ENDPOINT && 1649 type != PCI_EXP_TYPE_LEG_END && 1650 type != PCI_EXP_TYPE_RC_END) { 1651 1652 error_setg(errp, "assignment of PCIe type 0x%x " 1653 "devices is not currently supported", type); 1654 return -EINVAL; 1655 } 1656 1657 if (!pci_bus_is_express(vdev->pdev.bus)) { 1658 PCIBus *bus = vdev->pdev.bus; 1659 PCIDevice *bridge; 1660 1661 /* 1662 * Traditionally PCI device assignment exposes the PCIe capability 1663 * as-is on non-express buses. The reason being that some drivers 1664 * simply assume that it's there, for example tg3. However when 1665 * we're running on a native PCIe machine type, like Q35, we need 1666 * to hide the PCIe capability. The reason for this is twofold; 1667 * first Windows guests get a Code 10 error when the PCIe capability 1668 * is exposed in this configuration. Therefore express devices won't 1669 * work at all unless they're attached to express buses in the VM. 1670 * Second, a native PCIe machine introduces the possibility of fine 1671 * granularity IOMMUs supporting both translation and isolation. 1672 * Guest code to discover the IOMMU visibility of a device, such as 1673 * IOMMU grouping code on Linux, is very aware of device types and 1674 * valid transitions between bus types. An express device on a non- 1675 * express bus is not a valid combination on bare metal systems. 1676 * 1677 * Drivers that require a PCIe capability to make the device 1678 * functional are simply going to need to have their devices placed 1679 * on a PCIe bus in the VM. 1680 */ 1681 while (!pci_bus_is_root(bus)) { 1682 bridge = pci_bridge_get_device(bus); 1683 bus = bridge->bus; 1684 } 1685 1686 if (pci_bus_is_express(bus)) { 1687 return 0; 1688 } 1689 1690 } else if (pci_bus_is_root(vdev->pdev.bus)) { 1691 /* 1692 * On a Root Complex bus Endpoints become Root Complex Integrated 1693 * Endpoints, which changes the type and clears the LNK & LNK2 fields. 1694 */ 1695 if (type == PCI_EXP_TYPE_ENDPOINT) { 1696 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS, 1697 PCI_EXP_TYPE_RC_END << 4, 1698 PCI_EXP_FLAGS_TYPE); 1699 1700 /* Link Capabilities, Status, and Control goes away */ 1701 if (size > PCI_EXP_LNKCTL) { 1702 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 0, ~0); 1703 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0); 1704 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 0, ~0); 1705 1706 #ifndef PCI_EXP_LNKCAP2 1707 #define PCI_EXP_LNKCAP2 44 1708 #endif 1709 #ifndef PCI_EXP_LNKSTA2 1710 #define PCI_EXP_LNKSTA2 50 1711 #endif 1712 /* Link 2 Capabilities, Status, and Control goes away */ 1713 if (size > PCI_EXP_LNKCAP2) { 1714 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP2, 0, ~0); 1715 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL2, 0, ~0); 1716 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA2, 0, ~0); 1717 } 1718 } 1719 1720 } else if (type == PCI_EXP_TYPE_LEG_END) { 1721 /* 1722 * Legacy endpoints don't belong on the root complex. Windows 1723 * seems to be happier with devices if we skip the capability. 1724 */ 1725 return 0; 1726 } 1727 1728 } else { 1729 /* 1730 * Convert Root Complex Integrated Endpoints to regular endpoints. 1731 * These devices don't support LNK/LNK2 capabilities, so make them up. 1732 */ 1733 if (type == PCI_EXP_TYPE_RC_END) { 1734 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS, 1735 PCI_EXP_TYPE_ENDPOINT << 4, 1736 PCI_EXP_FLAGS_TYPE); 1737 vfio_add_emulated_long(vdev, pos + PCI_EXP_LNKCAP, 1738 PCI_EXP_LNK_MLW_1 | PCI_EXP_LNK_LS_25, ~0); 1739 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKCTL, 0, ~0); 1740 } 1741 1742 /* Mark the Link Status bits as emulated to allow virtual negotiation */ 1743 vfio_add_emulated_word(vdev, pos + PCI_EXP_LNKSTA, 1744 pci_get_word(vdev->pdev.config + pos + 1745 PCI_EXP_LNKSTA), 1746 PCI_EXP_LNKCAP_MLW | PCI_EXP_LNKCAP_SLS); 1747 } 1748 1749 /* 1750 * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0 1751 * (Niantic errate #35) causing Windows to error with a Code 10 for the 1752 * device on Q35. Fixup any such devices to report version 1. If we 1753 * were to remove the capability entirely the guest would lose extended 1754 * config space. 1755 */ 1756 if ((flags & PCI_EXP_FLAGS_VERS) == 0) { 1757 vfio_add_emulated_word(vdev, pos + PCI_CAP_FLAGS, 1758 1, PCI_EXP_FLAGS_VERS); 1759 } 1760 1761 pos = pci_add_capability(&vdev->pdev, PCI_CAP_ID_EXP, pos, size, 1762 errp); 1763 if (pos < 0) { 1764 return pos; 1765 } 1766 1767 vdev->pdev.exp.exp_cap = pos; 1768 1769 return pos; 1770 } 1771 1772 static void vfio_check_pcie_flr(VFIOPCIDevice *vdev, uint8_t pos) 1773 { 1774 uint32_t cap = pci_get_long(vdev->pdev.config + pos + PCI_EXP_DEVCAP); 1775 1776 if (cap & PCI_EXP_DEVCAP_FLR) { 1777 trace_vfio_check_pcie_flr(vdev->vbasedev.name); 1778 vdev->has_flr = true; 1779 } 1780 } 1781 1782 static void vfio_check_pm_reset(VFIOPCIDevice *vdev, uint8_t pos) 1783 { 1784 uint16_t csr = pci_get_word(vdev->pdev.config + pos + PCI_PM_CTRL); 1785 1786 if (!(csr & PCI_PM_CTRL_NO_SOFT_RESET)) { 1787 trace_vfio_check_pm_reset(vdev->vbasedev.name); 1788 vdev->has_pm_reset = true; 1789 } 1790 } 1791 1792 static void vfio_check_af_flr(VFIOPCIDevice *vdev, uint8_t pos) 1793 { 1794 uint8_t cap = pci_get_byte(vdev->pdev.config + pos + PCI_AF_CAP); 1795 1796 if ((cap & PCI_AF_CAP_TP) && (cap & PCI_AF_CAP_FLR)) { 1797 trace_vfio_check_af_flr(vdev->vbasedev.name); 1798 vdev->has_flr = true; 1799 } 1800 } 1801 1802 static int vfio_add_std_cap(VFIOPCIDevice *vdev, uint8_t pos, Error **errp) 1803 { 1804 PCIDevice *pdev = &vdev->pdev; 1805 uint8_t cap_id, next, size; 1806 int ret; 1807 1808 cap_id = pdev->config[pos]; 1809 next = pdev->config[pos + PCI_CAP_LIST_NEXT]; 1810 1811 /* 1812 * If it becomes important to configure capabilities to their actual 1813 * size, use this as the default when it's something we don't recognize. 1814 * Since QEMU doesn't actually handle many of the config accesses, 1815 * exact size doesn't seem worthwhile. 1816 */ 1817 size = vfio_std_cap_max_size(pdev, pos); 1818 1819 /* 1820 * pci_add_capability always inserts the new capability at the head 1821 * of the chain. Therefore to end up with a chain that matches the 1822 * physical device, we insert from the end by making this recursive. 1823 * This is also why we pre-calculate size above as cached config space 1824 * will be changed as we unwind the stack. 1825 */ 1826 if (next) { 1827 ret = vfio_add_std_cap(vdev, next, errp); 1828 if (ret) { 1829 goto out; 1830 } 1831 } else { 1832 /* Begin the rebuild, use QEMU emulated list bits */ 1833 pdev->config[PCI_CAPABILITY_LIST] = 0; 1834 vdev->emulated_config_bits[PCI_CAPABILITY_LIST] = 0xff; 1835 vdev->emulated_config_bits[PCI_STATUS] |= PCI_STATUS_CAP_LIST; 1836 } 1837 1838 /* Use emulated next pointer to allow dropping caps */ 1839 pci_set_byte(vdev->emulated_config_bits + pos + PCI_CAP_LIST_NEXT, 0xff); 1840 1841 switch (cap_id) { 1842 case PCI_CAP_ID_MSI: 1843 ret = vfio_msi_setup(vdev, pos, errp); 1844 break; 1845 case PCI_CAP_ID_EXP: 1846 vfio_check_pcie_flr(vdev, pos); 1847 ret = vfio_setup_pcie_cap(vdev, pos, size, errp); 1848 break; 1849 case PCI_CAP_ID_MSIX: 1850 ret = vfio_msix_setup(vdev, pos, errp); 1851 break; 1852 case PCI_CAP_ID_PM: 1853 vfio_check_pm_reset(vdev, pos); 1854 vdev->pm_cap = pos; 1855 ret = pci_add_capability(pdev, cap_id, pos, size, errp); 1856 break; 1857 case PCI_CAP_ID_AF: 1858 vfio_check_af_flr(vdev, pos); 1859 ret = pci_add_capability(pdev, cap_id, pos, size, errp); 1860 break; 1861 default: 1862 ret = pci_add_capability(pdev, cap_id, pos, size, errp); 1863 break; 1864 } 1865 out: 1866 if (ret < 0) { 1867 error_prepend(errp, 1868 "failed to add PCI capability 0x%x[0x%x]@0x%x: ", 1869 cap_id, size, pos); 1870 return ret; 1871 } 1872 1873 return 0; 1874 } 1875 1876 static void vfio_add_ext_cap(VFIOPCIDevice *vdev) 1877 { 1878 PCIDevice *pdev = &vdev->pdev; 1879 uint32_t header; 1880 uint16_t cap_id, next, size; 1881 uint8_t cap_ver; 1882 uint8_t *config; 1883 1884 /* Only add extended caps if we have them and the guest can see them */ 1885 if (!pci_is_express(pdev) || !pci_bus_is_express(pdev->bus) || 1886 !pci_get_long(pdev->config + PCI_CONFIG_SPACE_SIZE)) { 1887 return; 1888 } 1889 1890 /* 1891 * pcie_add_capability always inserts the new capability at the tail 1892 * of the chain. Therefore to end up with a chain that matches the 1893 * physical device, we cache the config space to avoid overwriting 1894 * the original config space when we parse the extended capabilities. 1895 */ 1896 config = g_memdup(pdev->config, vdev->config_size); 1897 1898 /* 1899 * Extended capabilities are chained with each pointing to the next, so we 1900 * can drop anything other than the head of the chain simply by modifying 1901 * the previous next pointer. Seed the head of the chain here such that 1902 * we can simply skip any capabilities we want to drop below, regardless 1903 * of their position in the chain. If this stub capability still exists 1904 * after we add the capabilities we want to expose, update the capability 1905 * ID to zero. Note that we cannot seed with the capability header being 1906 * zero as this conflicts with definition of an absent capability chain 1907 * and prevents capabilities beyond the head of the list from being added. 1908 * By replacing the dummy capability ID with zero after walking the device 1909 * chain, we also transparently mark extended capabilities as absent if 1910 * no capabilities were added. Note that the PCIe spec defines an absence 1911 * of extended capabilities to be determined by a value of zero for the 1912 * capability ID, version, AND next pointer. A non-zero next pointer 1913 * should be sufficient to indicate additional capabilities are present, 1914 * which will occur if we call pcie_add_capability() below. The entire 1915 * first dword is emulated to support this. 1916 * 1917 * NB. The kernel side does similar masking, so be prepared that our 1918 * view of the device may also contain a capability ID zero in the head 1919 * of the chain. Skip it for the same reason that we cannot seed the 1920 * chain with a zero capability. 1921 */ 1922 pci_set_long(pdev->config + PCI_CONFIG_SPACE_SIZE, 1923 PCI_EXT_CAP(0xFFFF, 0, 0)); 1924 pci_set_long(pdev->wmask + PCI_CONFIG_SPACE_SIZE, 0); 1925 pci_set_long(vdev->emulated_config_bits + PCI_CONFIG_SPACE_SIZE, ~0); 1926 1927 for (next = PCI_CONFIG_SPACE_SIZE; next; 1928 next = PCI_EXT_CAP_NEXT(pci_get_long(config + next))) { 1929 header = pci_get_long(config + next); 1930 cap_id = PCI_EXT_CAP_ID(header); 1931 cap_ver = PCI_EXT_CAP_VER(header); 1932 1933 /* 1934 * If it becomes important to configure extended capabilities to their 1935 * actual size, use this as the default when it's something we don't 1936 * recognize. Since QEMU doesn't actually handle many of the config 1937 * accesses, exact size doesn't seem worthwhile. 1938 */ 1939 size = vfio_ext_cap_max_size(config, next); 1940 1941 /* Use emulated next pointer to allow dropping extended caps */ 1942 pci_long_test_and_set_mask(vdev->emulated_config_bits + next, 1943 PCI_EXT_CAP_NEXT_MASK); 1944 1945 switch (cap_id) { 1946 case 0: /* kernel masked capability */ 1947 case PCI_EXT_CAP_ID_SRIOV: /* Read-only VF BARs confuse OVMF */ 1948 case PCI_EXT_CAP_ID_ARI: /* XXX Needs next function virtualization */ 1949 trace_vfio_add_ext_cap_dropped(vdev->vbasedev.name, cap_id, next); 1950 break; 1951 default: 1952 pcie_add_capability(pdev, cap_id, cap_ver, next, size); 1953 } 1954 1955 } 1956 1957 /* Cleanup chain head ID if necessary */ 1958 if (pci_get_word(pdev->config + PCI_CONFIG_SPACE_SIZE) == 0xFFFF) { 1959 pci_set_word(pdev->config + PCI_CONFIG_SPACE_SIZE, 0); 1960 } 1961 1962 g_free(config); 1963 return; 1964 } 1965 1966 static int vfio_add_capabilities(VFIOPCIDevice *vdev, Error **errp) 1967 { 1968 PCIDevice *pdev = &vdev->pdev; 1969 int ret; 1970 1971 if (!(pdev->config[PCI_STATUS] & PCI_STATUS_CAP_LIST) || 1972 !pdev->config[PCI_CAPABILITY_LIST]) { 1973 return 0; /* Nothing to add */ 1974 } 1975 1976 ret = vfio_add_std_cap(vdev, pdev->config[PCI_CAPABILITY_LIST], errp); 1977 if (ret) { 1978 return ret; 1979 } 1980 1981 vfio_add_ext_cap(vdev); 1982 return 0; 1983 } 1984 1985 static void vfio_pci_pre_reset(VFIOPCIDevice *vdev) 1986 { 1987 PCIDevice *pdev = &vdev->pdev; 1988 uint16_t cmd; 1989 1990 vfio_disable_interrupts(vdev); 1991 1992 /* Make sure the device is in D0 */ 1993 if (vdev->pm_cap) { 1994 uint16_t pmcsr; 1995 uint8_t state; 1996 1997 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2); 1998 state = pmcsr & PCI_PM_CTRL_STATE_MASK; 1999 if (state) { 2000 pmcsr &= ~PCI_PM_CTRL_STATE_MASK; 2001 vfio_pci_write_config(pdev, vdev->pm_cap + PCI_PM_CTRL, pmcsr, 2); 2002 /* vfio handles the necessary delay here */ 2003 pmcsr = vfio_pci_read_config(pdev, vdev->pm_cap + PCI_PM_CTRL, 2); 2004 state = pmcsr & PCI_PM_CTRL_STATE_MASK; 2005 if (state) { 2006 error_report("vfio: Unable to power on device, stuck in D%d", 2007 state); 2008 } 2009 } 2010 } 2011 2012 /* 2013 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master. 2014 * Also put INTx Disable in known state. 2015 */ 2016 cmd = vfio_pci_read_config(pdev, PCI_COMMAND, 2); 2017 cmd &= ~(PCI_COMMAND_IO | PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER | 2018 PCI_COMMAND_INTX_DISABLE); 2019 vfio_pci_write_config(pdev, PCI_COMMAND, cmd, 2); 2020 } 2021 2022 static void vfio_pci_post_reset(VFIOPCIDevice *vdev) 2023 { 2024 Error *err = NULL; 2025 int nr; 2026 2027 vfio_intx_enable(vdev, &err); 2028 if (err) { 2029 error_reportf_err(err, ERR_PREFIX, vdev->vbasedev.name); 2030 } 2031 2032 for (nr = 0; nr < PCI_NUM_REGIONS - 1; ++nr) { 2033 off_t addr = vdev->config_offset + PCI_BASE_ADDRESS_0 + (4 * nr); 2034 uint32_t val = 0; 2035 uint32_t len = sizeof(val); 2036 2037 if (pwrite(vdev->vbasedev.fd, &val, len, addr) != len) { 2038 error_report("%s(%s) reset bar %d failed: %m", __func__, 2039 vdev->vbasedev.name, nr); 2040 } 2041 } 2042 } 2043 2044 static bool vfio_pci_host_match(PCIHostDeviceAddress *addr, const char *name) 2045 { 2046 char tmp[13]; 2047 2048 sprintf(tmp, "%04x:%02x:%02x.%1x", addr->domain, 2049 addr->bus, addr->slot, addr->function); 2050 2051 return (strcmp(tmp, name) == 0); 2052 } 2053 2054 static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool single) 2055 { 2056 VFIOGroup *group; 2057 struct vfio_pci_hot_reset_info *info; 2058 struct vfio_pci_dependent_device *devices; 2059 struct vfio_pci_hot_reset *reset; 2060 int32_t *fds; 2061 int ret, i, count; 2062 bool multi = false; 2063 2064 trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi"); 2065 2066 if (!single) { 2067 vfio_pci_pre_reset(vdev); 2068 } 2069 vdev->vbasedev.needs_reset = false; 2070 2071 info = g_malloc0(sizeof(*info)); 2072 info->argsz = sizeof(*info); 2073 2074 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info); 2075 if (ret && errno != ENOSPC) { 2076 ret = -errno; 2077 if (!vdev->has_pm_reset) { 2078 error_report("vfio: Cannot reset device %s, " 2079 "no available reset mechanism.", vdev->vbasedev.name); 2080 } 2081 goto out_single; 2082 } 2083 2084 count = info->count; 2085 info = g_realloc(info, sizeof(*info) + (count * sizeof(*devices))); 2086 info->argsz = sizeof(*info) + (count * sizeof(*devices)); 2087 devices = &info->devices[0]; 2088 2089 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO, info); 2090 if (ret) { 2091 ret = -errno; 2092 error_report("vfio: hot reset info failed: %m"); 2093 goto out_single; 2094 } 2095 2096 trace_vfio_pci_hot_reset_has_dep_devices(vdev->vbasedev.name); 2097 2098 /* Verify that we have all the groups required */ 2099 for (i = 0; i < info->count; i++) { 2100 PCIHostDeviceAddress host; 2101 VFIOPCIDevice *tmp; 2102 VFIODevice *vbasedev_iter; 2103 2104 host.domain = devices[i].segment; 2105 host.bus = devices[i].bus; 2106 host.slot = PCI_SLOT(devices[i].devfn); 2107 host.function = PCI_FUNC(devices[i].devfn); 2108 2109 trace_vfio_pci_hot_reset_dep_devices(host.domain, 2110 host.bus, host.slot, host.function, devices[i].group_id); 2111 2112 if (vfio_pci_host_match(&host, vdev->vbasedev.name)) { 2113 continue; 2114 } 2115 2116 QLIST_FOREACH(group, &vfio_group_list, next) { 2117 if (group->groupid == devices[i].group_id) { 2118 break; 2119 } 2120 } 2121 2122 if (!group) { 2123 if (!vdev->has_pm_reset) { 2124 error_report("vfio: Cannot reset device %s, " 2125 "depends on group %d which is not owned.", 2126 vdev->vbasedev.name, devices[i].group_id); 2127 } 2128 ret = -EPERM; 2129 goto out; 2130 } 2131 2132 /* Prep dependent devices for reset and clear our marker. */ 2133 QLIST_FOREACH(vbasedev_iter, &group->device_list, next) { 2134 if (!vbasedev_iter->dev->realized || 2135 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) { 2136 continue; 2137 } 2138 tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev); 2139 if (vfio_pci_host_match(&host, tmp->vbasedev.name)) { 2140 if (single) { 2141 ret = -EINVAL; 2142 goto out_single; 2143 } 2144 vfio_pci_pre_reset(tmp); 2145 tmp->vbasedev.needs_reset = false; 2146 multi = true; 2147 break; 2148 } 2149 } 2150 } 2151 2152 if (!single && !multi) { 2153 ret = -EINVAL; 2154 goto out_single; 2155 } 2156 2157 /* Determine how many group fds need to be passed */ 2158 count = 0; 2159 QLIST_FOREACH(group, &vfio_group_list, next) { 2160 for (i = 0; i < info->count; i++) { 2161 if (group->groupid == devices[i].group_id) { 2162 count++; 2163 break; 2164 } 2165 } 2166 } 2167 2168 reset = g_malloc0(sizeof(*reset) + (count * sizeof(*fds))); 2169 reset->argsz = sizeof(*reset) + (count * sizeof(*fds)); 2170 fds = &reset->group_fds[0]; 2171 2172 /* Fill in group fds */ 2173 QLIST_FOREACH(group, &vfio_group_list, next) { 2174 for (i = 0; i < info->count; i++) { 2175 if (group->groupid == devices[i].group_id) { 2176 fds[reset->count++] = group->fd; 2177 break; 2178 } 2179 } 2180 } 2181 2182 /* Bus reset! */ 2183 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_PCI_HOT_RESET, reset); 2184 g_free(reset); 2185 2186 trace_vfio_pci_hot_reset_result(vdev->vbasedev.name, 2187 ret ? "%m" : "Success"); 2188 2189 out: 2190 /* Re-enable INTx on affected devices */ 2191 for (i = 0; i < info->count; i++) { 2192 PCIHostDeviceAddress host; 2193 VFIOPCIDevice *tmp; 2194 VFIODevice *vbasedev_iter; 2195 2196 host.domain = devices[i].segment; 2197 host.bus = devices[i].bus; 2198 host.slot = PCI_SLOT(devices[i].devfn); 2199 host.function = PCI_FUNC(devices[i].devfn); 2200 2201 if (vfio_pci_host_match(&host, vdev->vbasedev.name)) { 2202 continue; 2203 } 2204 2205 QLIST_FOREACH(group, &vfio_group_list, next) { 2206 if (group->groupid == devices[i].group_id) { 2207 break; 2208 } 2209 } 2210 2211 if (!group) { 2212 break; 2213 } 2214 2215 QLIST_FOREACH(vbasedev_iter, &group->device_list, next) { 2216 if (!vbasedev_iter->dev->realized || 2217 vbasedev_iter->type != VFIO_DEVICE_TYPE_PCI) { 2218 continue; 2219 } 2220 tmp = container_of(vbasedev_iter, VFIOPCIDevice, vbasedev); 2221 if (vfio_pci_host_match(&host, tmp->vbasedev.name)) { 2222 vfio_pci_post_reset(tmp); 2223 break; 2224 } 2225 } 2226 } 2227 out_single: 2228 if (!single) { 2229 vfio_pci_post_reset(vdev); 2230 } 2231 g_free(info); 2232 2233 return ret; 2234 } 2235 2236 /* 2237 * We want to differentiate hot reset of mulitple in-use devices vs hot reset 2238 * of a single in-use device. VFIO_DEVICE_RESET will already handle the case 2239 * of doing hot resets when there is only a single device per bus. The in-use 2240 * here refers to how many VFIODevices are affected. A hot reset that affects 2241 * multiple devices, but only a single in-use device, means that we can call 2242 * it from our bus ->reset() callback since the extent is effectively a single 2243 * device. This allows us to make use of it in the hotplug path. When there 2244 * are multiple in-use devices, we can only trigger the hot reset during a 2245 * system reset and thus from our reset handler. We separate _one vs _multi 2246 * here so that we don't overlap and do a double reset on the system reset 2247 * path where both our reset handler and ->reset() callback are used. Calling 2248 * _one() will only do a hot reset for the one in-use devices case, calling 2249 * _multi() will do nothing if a _one() would have been sufficient. 2250 */ 2251 static int vfio_pci_hot_reset_one(VFIOPCIDevice *vdev) 2252 { 2253 return vfio_pci_hot_reset(vdev, true); 2254 } 2255 2256 static int vfio_pci_hot_reset_multi(VFIODevice *vbasedev) 2257 { 2258 VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev); 2259 return vfio_pci_hot_reset(vdev, false); 2260 } 2261 2262 static void vfio_pci_compute_needs_reset(VFIODevice *vbasedev) 2263 { 2264 VFIOPCIDevice *vdev = container_of(vbasedev, VFIOPCIDevice, vbasedev); 2265 if (!vbasedev->reset_works || (!vdev->has_flr && vdev->has_pm_reset)) { 2266 vbasedev->needs_reset = true; 2267 } 2268 } 2269 2270 static VFIODeviceOps vfio_pci_ops = { 2271 .vfio_compute_needs_reset = vfio_pci_compute_needs_reset, 2272 .vfio_hot_reset_multi = vfio_pci_hot_reset_multi, 2273 .vfio_eoi = vfio_intx_eoi, 2274 }; 2275 2276 int vfio_populate_vga(VFIOPCIDevice *vdev, Error **errp) 2277 { 2278 VFIODevice *vbasedev = &vdev->vbasedev; 2279 struct vfio_region_info *reg_info; 2280 int ret; 2281 2282 ret = vfio_get_region_info(vbasedev, VFIO_PCI_VGA_REGION_INDEX, ®_info); 2283 if (ret) { 2284 error_setg_errno(errp, -ret, 2285 "failed getting region info for VGA region index %d", 2286 VFIO_PCI_VGA_REGION_INDEX); 2287 return ret; 2288 } 2289 2290 if (!(reg_info->flags & VFIO_REGION_INFO_FLAG_READ) || 2291 !(reg_info->flags & VFIO_REGION_INFO_FLAG_WRITE) || 2292 reg_info->size < 0xbffff + 1) { 2293 error_setg(errp, "unexpected VGA info, flags 0x%lx, size 0x%lx", 2294 (unsigned long)reg_info->flags, 2295 (unsigned long)reg_info->size); 2296 g_free(reg_info); 2297 return -EINVAL; 2298 } 2299 2300 vdev->vga = g_new0(VFIOVGA, 1); 2301 2302 vdev->vga->fd_offset = reg_info->offset; 2303 vdev->vga->fd = vdev->vbasedev.fd; 2304 2305 g_free(reg_info); 2306 2307 vdev->vga->region[QEMU_PCI_VGA_MEM].offset = QEMU_PCI_VGA_MEM_BASE; 2308 vdev->vga->region[QEMU_PCI_VGA_MEM].nr = QEMU_PCI_VGA_MEM; 2309 QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_MEM].quirks); 2310 2311 memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_MEM].mem, 2312 OBJECT(vdev), &vfio_vga_ops, 2313 &vdev->vga->region[QEMU_PCI_VGA_MEM], 2314 "vfio-vga-mmio@0xa0000", 2315 QEMU_PCI_VGA_MEM_SIZE); 2316 2317 vdev->vga->region[QEMU_PCI_VGA_IO_LO].offset = QEMU_PCI_VGA_IO_LO_BASE; 2318 vdev->vga->region[QEMU_PCI_VGA_IO_LO].nr = QEMU_PCI_VGA_IO_LO; 2319 QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].quirks); 2320 2321 memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem, 2322 OBJECT(vdev), &vfio_vga_ops, 2323 &vdev->vga->region[QEMU_PCI_VGA_IO_LO], 2324 "vfio-vga-io@0x3b0", 2325 QEMU_PCI_VGA_IO_LO_SIZE); 2326 2327 vdev->vga->region[QEMU_PCI_VGA_IO_HI].offset = QEMU_PCI_VGA_IO_HI_BASE; 2328 vdev->vga->region[QEMU_PCI_VGA_IO_HI].nr = QEMU_PCI_VGA_IO_HI; 2329 QLIST_INIT(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].quirks); 2330 2331 memory_region_init_io(&vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem, 2332 OBJECT(vdev), &vfio_vga_ops, 2333 &vdev->vga->region[QEMU_PCI_VGA_IO_HI], 2334 "vfio-vga-io@0x3c0", 2335 QEMU_PCI_VGA_IO_HI_SIZE); 2336 2337 pci_register_vga(&vdev->pdev, &vdev->vga->region[QEMU_PCI_VGA_MEM].mem, 2338 &vdev->vga->region[QEMU_PCI_VGA_IO_LO].mem, 2339 &vdev->vga->region[QEMU_PCI_VGA_IO_HI].mem); 2340 2341 return 0; 2342 } 2343 2344 static void vfio_populate_device(VFIOPCIDevice *vdev, Error **errp) 2345 { 2346 VFIODevice *vbasedev = &vdev->vbasedev; 2347 struct vfio_region_info *reg_info; 2348 struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info) }; 2349 int i, ret = -1; 2350 2351 /* Sanity check device */ 2352 if (!(vbasedev->flags & VFIO_DEVICE_FLAGS_PCI)) { 2353 error_setg(errp, "this isn't a PCI device"); 2354 return; 2355 } 2356 2357 if (vbasedev->num_regions < VFIO_PCI_CONFIG_REGION_INDEX + 1) { 2358 error_setg(errp, "unexpected number of io regions %u", 2359 vbasedev->num_regions); 2360 return; 2361 } 2362 2363 if (vbasedev->num_irqs < VFIO_PCI_MSIX_IRQ_INDEX + 1) { 2364 error_setg(errp, "unexpected number of irqs %u", vbasedev->num_irqs); 2365 return; 2366 } 2367 2368 for (i = VFIO_PCI_BAR0_REGION_INDEX; i < VFIO_PCI_ROM_REGION_INDEX; i++) { 2369 char *name = g_strdup_printf("%s BAR %d", vbasedev->name, i); 2370 2371 ret = vfio_region_setup(OBJECT(vdev), vbasedev, 2372 &vdev->bars[i].region, i, name); 2373 g_free(name); 2374 2375 if (ret) { 2376 error_setg_errno(errp, -ret, "failed to get region %d info", i); 2377 return; 2378 } 2379 2380 QLIST_INIT(&vdev->bars[i].quirks); 2381 } 2382 2383 ret = vfio_get_region_info(vbasedev, 2384 VFIO_PCI_CONFIG_REGION_INDEX, ®_info); 2385 if (ret) { 2386 error_setg_errno(errp, -ret, "failed to get config info"); 2387 return; 2388 } 2389 2390 trace_vfio_populate_device_config(vdev->vbasedev.name, 2391 (unsigned long)reg_info->size, 2392 (unsigned long)reg_info->offset, 2393 (unsigned long)reg_info->flags); 2394 2395 vdev->config_size = reg_info->size; 2396 if (vdev->config_size == PCI_CONFIG_SPACE_SIZE) { 2397 vdev->pdev.cap_present &= ~QEMU_PCI_CAP_EXPRESS; 2398 } 2399 vdev->config_offset = reg_info->offset; 2400 2401 g_free(reg_info); 2402 2403 if (vdev->features & VFIO_FEATURE_ENABLE_VGA) { 2404 ret = vfio_populate_vga(vdev, errp); 2405 if (ret) { 2406 error_append_hint(errp, "device does not support " 2407 "requested feature x-vga\n"); 2408 return; 2409 } 2410 } 2411 2412 irq_info.index = VFIO_PCI_ERR_IRQ_INDEX; 2413 2414 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_GET_IRQ_INFO, &irq_info); 2415 if (ret) { 2416 /* This can fail for an old kernel or legacy PCI dev */ 2417 trace_vfio_populate_device_get_irq_info_failure(); 2418 } else if (irq_info.count == 1) { 2419 vdev->pci_aer = true; 2420 } else { 2421 error_report(WARN_PREFIX 2422 "Could not enable error recovery for the device", 2423 vbasedev->name); 2424 } 2425 } 2426 2427 static void vfio_put_device(VFIOPCIDevice *vdev) 2428 { 2429 g_free(vdev->vbasedev.name); 2430 g_free(vdev->msix); 2431 2432 vfio_put_base_device(&vdev->vbasedev); 2433 } 2434 2435 static void vfio_err_notifier_handler(void *opaque) 2436 { 2437 VFIOPCIDevice *vdev = opaque; 2438 2439 if (!event_notifier_test_and_clear(&vdev->err_notifier)) { 2440 return; 2441 } 2442 2443 /* 2444 * TBD. Retrieve the error details and decide what action 2445 * needs to be taken. One of the actions could be to pass 2446 * the error to the guest and have the guest driver recover 2447 * from the error. This requires that PCIe capabilities be 2448 * exposed to the guest. For now, we just terminate the 2449 * guest to contain the error. 2450 */ 2451 2452 error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__, vdev->vbasedev.name); 2453 2454 vm_stop(RUN_STATE_INTERNAL_ERROR); 2455 } 2456 2457 /* 2458 * Registers error notifier for devices supporting error recovery. 2459 * If we encounter a failure in this function, we report an error 2460 * and continue after disabling error recovery support for the 2461 * device. 2462 */ 2463 static void vfio_register_err_notifier(VFIOPCIDevice *vdev) 2464 { 2465 int ret; 2466 int argsz; 2467 struct vfio_irq_set *irq_set; 2468 int32_t *pfd; 2469 2470 if (!vdev->pci_aer) { 2471 return; 2472 } 2473 2474 if (event_notifier_init(&vdev->err_notifier, 0)) { 2475 error_report("vfio: Unable to init event notifier for error detection"); 2476 vdev->pci_aer = false; 2477 return; 2478 } 2479 2480 argsz = sizeof(*irq_set) + sizeof(*pfd); 2481 2482 irq_set = g_malloc0(argsz); 2483 irq_set->argsz = argsz; 2484 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 2485 VFIO_IRQ_SET_ACTION_TRIGGER; 2486 irq_set->index = VFIO_PCI_ERR_IRQ_INDEX; 2487 irq_set->start = 0; 2488 irq_set->count = 1; 2489 pfd = (int32_t *)&irq_set->data; 2490 2491 *pfd = event_notifier_get_fd(&vdev->err_notifier); 2492 qemu_set_fd_handler(*pfd, vfio_err_notifier_handler, NULL, vdev); 2493 2494 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 2495 if (ret) { 2496 error_report("vfio: Failed to set up error notification"); 2497 qemu_set_fd_handler(*pfd, NULL, NULL, vdev); 2498 event_notifier_cleanup(&vdev->err_notifier); 2499 vdev->pci_aer = false; 2500 } 2501 g_free(irq_set); 2502 } 2503 2504 static void vfio_unregister_err_notifier(VFIOPCIDevice *vdev) 2505 { 2506 int argsz; 2507 struct vfio_irq_set *irq_set; 2508 int32_t *pfd; 2509 int ret; 2510 2511 if (!vdev->pci_aer) { 2512 return; 2513 } 2514 2515 argsz = sizeof(*irq_set) + sizeof(*pfd); 2516 2517 irq_set = g_malloc0(argsz); 2518 irq_set->argsz = argsz; 2519 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 2520 VFIO_IRQ_SET_ACTION_TRIGGER; 2521 irq_set->index = VFIO_PCI_ERR_IRQ_INDEX; 2522 irq_set->start = 0; 2523 irq_set->count = 1; 2524 pfd = (int32_t *)&irq_set->data; 2525 *pfd = -1; 2526 2527 ret = ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set); 2528 if (ret) { 2529 error_report("vfio: Failed to de-assign error fd: %m"); 2530 } 2531 g_free(irq_set); 2532 qemu_set_fd_handler(event_notifier_get_fd(&vdev->err_notifier), 2533 NULL, NULL, vdev); 2534 event_notifier_cleanup(&vdev->err_notifier); 2535 } 2536 2537 static void vfio_req_notifier_handler(void *opaque) 2538 { 2539 VFIOPCIDevice *vdev = opaque; 2540 Error *err = NULL; 2541 2542 if (!event_notifier_test_and_clear(&vdev->req_notifier)) { 2543 return; 2544 } 2545 2546 qdev_unplug(&vdev->pdev.qdev, &err); 2547 if (err) { 2548 error_reportf_err(err, WARN_PREFIX, vdev->vbasedev.name); 2549 } 2550 } 2551 2552 static void vfio_register_req_notifier(VFIOPCIDevice *vdev) 2553 { 2554 struct vfio_irq_info irq_info = { .argsz = sizeof(irq_info), 2555 .index = VFIO_PCI_REQ_IRQ_INDEX }; 2556 int argsz; 2557 struct vfio_irq_set *irq_set; 2558 int32_t *pfd; 2559 2560 if (!(vdev->features & VFIO_FEATURE_ENABLE_REQ)) { 2561 return; 2562 } 2563 2564 if (ioctl(vdev->vbasedev.fd, 2565 VFIO_DEVICE_GET_IRQ_INFO, &irq_info) < 0 || irq_info.count < 1) { 2566 return; 2567 } 2568 2569 if (event_notifier_init(&vdev->req_notifier, 0)) { 2570 error_report("vfio: Unable to init event notifier for device request"); 2571 return; 2572 } 2573 2574 argsz = sizeof(*irq_set) + sizeof(*pfd); 2575 2576 irq_set = g_malloc0(argsz); 2577 irq_set->argsz = argsz; 2578 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 2579 VFIO_IRQ_SET_ACTION_TRIGGER; 2580 irq_set->index = VFIO_PCI_REQ_IRQ_INDEX; 2581 irq_set->start = 0; 2582 irq_set->count = 1; 2583 pfd = (int32_t *)&irq_set->data; 2584 2585 *pfd = event_notifier_get_fd(&vdev->req_notifier); 2586 qemu_set_fd_handler(*pfd, vfio_req_notifier_handler, NULL, vdev); 2587 2588 if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) { 2589 error_report("vfio: Failed to set up device request notification"); 2590 qemu_set_fd_handler(*pfd, NULL, NULL, vdev); 2591 event_notifier_cleanup(&vdev->req_notifier); 2592 } else { 2593 vdev->req_enabled = true; 2594 } 2595 2596 g_free(irq_set); 2597 } 2598 2599 static void vfio_unregister_req_notifier(VFIOPCIDevice *vdev) 2600 { 2601 int argsz; 2602 struct vfio_irq_set *irq_set; 2603 int32_t *pfd; 2604 2605 if (!vdev->req_enabled) { 2606 return; 2607 } 2608 2609 argsz = sizeof(*irq_set) + sizeof(*pfd); 2610 2611 irq_set = g_malloc0(argsz); 2612 irq_set->argsz = argsz; 2613 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | 2614 VFIO_IRQ_SET_ACTION_TRIGGER; 2615 irq_set->index = VFIO_PCI_REQ_IRQ_INDEX; 2616 irq_set->start = 0; 2617 irq_set->count = 1; 2618 pfd = (int32_t *)&irq_set->data; 2619 *pfd = -1; 2620 2621 if (ioctl(vdev->vbasedev.fd, VFIO_DEVICE_SET_IRQS, irq_set)) { 2622 error_report("vfio: Failed to de-assign device request fd: %m"); 2623 } 2624 g_free(irq_set); 2625 qemu_set_fd_handler(event_notifier_get_fd(&vdev->req_notifier), 2626 NULL, NULL, vdev); 2627 event_notifier_cleanup(&vdev->req_notifier); 2628 2629 vdev->req_enabled = false; 2630 } 2631 2632 static void vfio_realize(PCIDevice *pdev, Error **errp) 2633 { 2634 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 2635 VFIODevice *vbasedev_iter; 2636 VFIOGroup *group; 2637 char *tmp, group_path[PATH_MAX], *group_name; 2638 Error *err = NULL; 2639 ssize_t len; 2640 struct stat st; 2641 int groupid; 2642 int i, ret; 2643 2644 if (!vdev->vbasedev.sysfsdev) { 2645 if (!(~vdev->host.domain || ~vdev->host.bus || 2646 ~vdev->host.slot || ~vdev->host.function)) { 2647 error_setg(errp, "No provided host device"); 2648 error_append_hint(errp, "Use -device vfio-pci,host=DDDD:BB:DD.F " 2649 "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n"); 2650 return; 2651 } 2652 vdev->vbasedev.sysfsdev = 2653 g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x", 2654 vdev->host.domain, vdev->host.bus, 2655 vdev->host.slot, vdev->host.function); 2656 } 2657 2658 if (stat(vdev->vbasedev.sysfsdev, &st) < 0) { 2659 error_setg_errno(errp, errno, "no such host device"); 2660 error_prepend(errp, ERR_PREFIX, vdev->vbasedev.sysfsdev); 2661 return; 2662 } 2663 2664 vdev->vbasedev.name = g_strdup(basename(vdev->vbasedev.sysfsdev)); 2665 vdev->vbasedev.ops = &vfio_pci_ops; 2666 vdev->vbasedev.type = VFIO_DEVICE_TYPE_PCI; 2667 vdev->vbasedev.dev = &vdev->pdev.qdev; 2668 2669 tmp = g_strdup_printf("%s/iommu_group", vdev->vbasedev.sysfsdev); 2670 len = readlink(tmp, group_path, sizeof(group_path)); 2671 g_free(tmp); 2672 2673 if (len <= 0 || len >= sizeof(group_path)) { 2674 error_setg_errno(errp, len < 0 ? errno : ENAMETOOLONG, 2675 "no iommu_group found"); 2676 goto error; 2677 } 2678 2679 group_path[len] = 0; 2680 2681 group_name = basename(group_path); 2682 if (sscanf(group_name, "%d", &groupid) != 1) { 2683 error_setg_errno(errp, errno, "failed to read %s", group_path); 2684 goto error; 2685 } 2686 2687 trace_vfio_realize(vdev->vbasedev.name, groupid); 2688 2689 group = vfio_get_group(groupid, pci_device_iommu_address_space(pdev), errp); 2690 if (!group) { 2691 goto error; 2692 } 2693 2694 QLIST_FOREACH(vbasedev_iter, &group->device_list, next) { 2695 if (strcmp(vbasedev_iter->name, vdev->vbasedev.name) == 0) { 2696 error_setg(errp, "device is already attached"); 2697 vfio_put_group(group); 2698 goto error; 2699 } 2700 } 2701 2702 ret = vfio_get_device(group, vdev->vbasedev.name, &vdev->vbasedev, errp); 2703 if (ret) { 2704 vfio_put_group(group); 2705 goto error; 2706 } 2707 2708 vfio_populate_device(vdev, &err); 2709 if (err) { 2710 error_propagate(errp, err); 2711 goto error; 2712 } 2713 2714 /* Get a copy of config space */ 2715 ret = pread(vdev->vbasedev.fd, vdev->pdev.config, 2716 MIN(pci_config_size(&vdev->pdev), vdev->config_size), 2717 vdev->config_offset); 2718 if (ret < (int)MIN(pci_config_size(&vdev->pdev), vdev->config_size)) { 2719 ret = ret < 0 ? -errno : -EFAULT; 2720 error_setg_errno(errp, -ret, "failed to read device config space"); 2721 goto error; 2722 } 2723 2724 /* vfio emulates a lot for us, but some bits need extra love */ 2725 vdev->emulated_config_bits = g_malloc0(vdev->config_size); 2726 2727 /* QEMU can choose to expose the ROM or not */ 2728 memset(vdev->emulated_config_bits + PCI_ROM_ADDRESS, 0xff, 4); 2729 2730 /* 2731 * The PCI spec reserves vendor ID 0xffff as an invalid value. The 2732 * device ID is managed by the vendor and need only be a 16-bit value. 2733 * Allow any 16-bit value for subsystem so they can be hidden or changed. 2734 */ 2735 if (vdev->vendor_id != PCI_ANY_ID) { 2736 if (vdev->vendor_id >= 0xffff) { 2737 error_setg(errp, "invalid PCI vendor ID provided"); 2738 goto error; 2739 } 2740 vfio_add_emulated_word(vdev, PCI_VENDOR_ID, vdev->vendor_id, ~0); 2741 trace_vfio_pci_emulated_vendor_id(vdev->vbasedev.name, vdev->vendor_id); 2742 } else { 2743 vdev->vendor_id = pci_get_word(pdev->config + PCI_VENDOR_ID); 2744 } 2745 2746 if (vdev->device_id != PCI_ANY_ID) { 2747 if (vdev->device_id > 0xffff) { 2748 error_setg(errp, "invalid PCI device ID provided"); 2749 goto error; 2750 } 2751 vfio_add_emulated_word(vdev, PCI_DEVICE_ID, vdev->device_id, ~0); 2752 trace_vfio_pci_emulated_device_id(vdev->vbasedev.name, vdev->device_id); 2753 } else { 2754 vdev->device_id = pci_get_word(pdev->config + PCI_DEVICE_ID); 2755 } 2756 2757 if (vdev->sub_vendor_id != PCI_ANY_ID) { 2758 if (vdev->sub_vendor_id > 0xffff) { 2759 error_setg(errp, "invalid PCI subsystem vendor ID provided"); 2760 goto error; 2761 } 2762 vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_VENDOR_ID, 2763 vdev->sub_vendor_id, ~0); 2764 trace_vfio_pci_emulated_sub_vendor_id(vdev->vbasedev.name, 2765 vdev->sub_vendor_id); 2766 } 2767 2768 if (vdev->sub_device_id != PCI_ANY_ID) { 2769 if (vdev->sub_device_id > 0xffff) { 2770 error_setg(errp, "invalid PCI subsystem device ID provided"); 2771 goto error; 2772 } 2773 vfio_add_emulated_word(vdev, PCI_SUBSYSTEM_ID, vdev->sub_device_id, ~0); 2774 trace_vfio_pci_emulated_sub_device_id(vdev->vbasedev.name, 2775 vdev->sub_device_id); 2776 } 2777 2778 /* QEMU can change multi-function devices to single function, or reverse */ 2779 vdev->emulated_config_bits[PCI_HEADER_TYPE] = 2780 PCI_HEADER_TYPE_MULTI_FUNCTION; 2781 2782 /* Restore or clear multifunction, this is always controlled by QEMU */ 2783 if (vdev->pdev.cap_present & QEMU_PCI_CAP_MULTIFUNCTION) { 2784 vdev->pdev.config[PCI_HEADER_TYPE] |= PCI_HEADER_TYPE_MULTI_FUNCTION; 2785 } else { 2786 vdev->pdev.config[PCI_HEADER_TYPE] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION; 2787 } 2788 2789 /* 2790 * Clear host resource mapping info. If we choose not to register a 2791 * BAR, such as might be the case with the option ROM, we can get 2792 * confusing, unwritable, residual addresses from the host here. 2793 */ 2794 memset(&vdev->pdev.config[PCI_BASE_ADDRESS_0], 0, 24); 2795 memset(&vdev->pdev.config[PCI_ROM_ADDRESS], 0, 4); 2796 2797 vfio_pci_size_rom(vdev); 2798 2799 vfio_msix_early_setup(vdev, &err); 2800 if (err) { 2801 error_propagate(errp, err); 2802 goto error; 2803 } 2804 2805 vfio_bars_setup(vdev); 2806 2807 ret = vfio_add_capabilities(vdev, errp); 2808 if (ret) { 2809 goto out_teardown; 2810 } 2811 2812 if (vdev->vga) { 2813 vfio_vga_quirk_setup(vdev); 2814 } 2815 2816 for (i = 0; i < PCI_ROM_SLOT; i++) { 2817 vfio_bar_quirk_setup(vdev, i); 2818 } 2819 2820 if (!vdev->igd_opregion && 2821 vdev->features & VFIO_FEATURE_ENABLE_IGD_OPREGION) { 2822 struct vfio_region_info *opregion; 2823 2824 if (vdev->pdev.qdev.hotplugged) { 2825 error_setg(errp, 2826 "cannot support IGD OpRegion feature on hotplugged " 2827 "device"); 2828 goto out_teardown; 2829 } 2830 2831 ret = vfio_get_dev_region_info(&vdev->vbasedev, 2832 VFIO_REGION_TYPE_PCI_VENDOR_TYPE | PCI_VENDOR_ID_INTEL, 2833 VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION, &opregion); 2834 if (ret) { 2835 error_setg_errno(errp, -ret, 2836 "does not support requested IGD OpRegion feature"); 2837 goto out_teardown; 2838 } 2839 2840 ret = vfio_pci_igd_opregion_init(vdev, opregion, errp); 2841 g_free(opregion); 2842 if (ret) { 2843 goto out_teardown; 2844 } 2845 } 2846 2847 /* QEMU emulates all of MSI & MSIX */ 2848 if (pdev->cap_present & QEMU_PCI_CAP_MSIX) { 2849 memset(vdev->emulated_config_bits + pdev->msix_cap, 0xff, 2850 MSIX_CAP_LENGTH); 2851 } 2852 2853 if (pdev->cap_present & QEMU_PCI_CAP_MSI) { 2854 memset(vdev->emulated_config_bits + pdev->msi_cap, 0xff, 2855 vdev->msi_cap_size); 2856 } 2857 2858 if (vfio_pci_read_config(&vdev->pdev, PCI_INTERRUPT_PIN, 1)) { 2859 vdev->intx.mmap_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL, 2860 vfio_intx_mmap_enable, vdev); 2861 pci_device_set_intx_routing_notifier(&vdev->pdev, vfio_intx_update); 2862 ret = vfio_intx_enable(vdev, errp); 2863 if (ret) { 2864 goto out_teardown; 2865 } 2866 } 2867 2868 vfio_register_err_notifier(vdev); 2869 vfio_register_req_notifier(vdev); 2870 vfio_setup_resetfn_quirk(vdev); 2871 2872 return; 2873 2874 out_teardown: 2875 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL); 2876 vfio_teardown_msi(vdev); 2877 vfio_bars_exit(vdev); 2878 error: 2879 error_prepend(errp, ERR_PREFIX, vdev->vbasedev.name); 2880 } 2881 2882 static void vfio_instance_finalize(Object *obj) 2883 { 2884 PCIDevice *pci_dev = PCI_DEVICE(obj); 2885 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pci_dev); 2886 VFIOGroup *group = vdev->vbasedev.group; 2887 2888 vfio_bars_finalize(vdev); 2889 g_free(vdev->emulated_config_bits); 2890 g_free(vdev->rom); 2891 /* 2892 * XXX Leaking igd_opregion is not an oversight, we can't remove the 2893 * fw_cfg entry therefore leaking this allocation seems like the safest 2894 * option. 2895 * 2896 * g_free(vdev->igd_opregion); 2897 */ 2898 vfio_put_device(vdev); 2899 vfio_put_group(group); 2900 } 2901 2902 static void vfio_exitfn(PCIDevice *pdev) 2903 { 2904 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 2905 2906 vfio_unregister_req_notifier(vdev); 2907 vfio_unregister_err_notifier(vdev); 2908 pci_device_set_intx_routing_notifier(&vdev->pdev, NULL); 2909 vfio_disable_interrupts(vdev); 2910 if (vdev->intx.mmap_timer) { 2911 timer_free(vdev->intx.mmap_timer); 2912 } 2913 vfio_teardown_msi(vdev); 2914 vfio_bars_exit(vdev); 2915 } 2916 2917 static void vfio_pci_reset(DeviceState *dev) 2918 { 2919 PCIDevice *pdev = DO_UPCAST(PCIDevice, qdev, dev); 2920 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, pdev); 2921 2922 trace_vfio_pci_reset(vdev->vbasedev.name); 2923 2924 vfio_pci_pre_reset(vdev); 2925 2926 if (vdev->resetfn && !vdev->resetfn(vdev)) { 2927 goto post_reset; 2928 } 2929 2930 if (vdev->vbasedev.reset_works && 2931 (vdev->has_flr || !vdev->has_pm_reset) && 2932 !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) { 2933 trace_vfio_pci_reset_flr(vdev->vbasedev.name); 2934 goto post_reset; 2935 } 2936 2937 /* See if we can do our own bus reset */ 2938 if (!vfio_pci_hot_reset_one(vdev)) { 2939 goto post_reset; 2940 } 2941 2942 /* If nothing else works and the device supports PM reset, use it */ 2943 if (vdev->vbasedev.reset_works && vdev->has_pm_reset && 2944 !ioctl(vdev->vbasedev.fd, VFIO_DEVICE_RESET)) { 2945 trace_vfio_pci_reset_pm(vdev->vbasedev.name); 2946 goto post_reset; 2947 } 2948 2949 post_reset: 2950 vfio_pci_post_reset(vdev); 2951 } 2952 2953 static void vfio_instance_init(Object *obj) 2954 { 2955 PCIDevice *pci_dev = PCI_DEVICE(obj); 2956 VFIOPCIDevice *vdev = DO_UPCAST(VFIOPCIDevice, pdev, PCI_DEVICE(obj)); 2957 2958 device_add_bootindex_property(obj, &vdev->bootindex, 2959 "bootindex", NULL, 2960 &pci_dev->qdev, NULL); 2961 vdev->host.domain = ~0U; 2962 vdev->host.bus = ~0U; 2963 vdev->host.slot = ~0U; 2964 vdev->host.function = ~0U; 2965 } 2966 2967 static Property vfio_pci_dev_properties[] = { 2968 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice, host), 2969 DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice, vbasedev.sysfsdev), 2970 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice, 2971 intx.mmap_timeout, 1100), 2972 DEFINE_PROP_BIT("x-vga", VFIOPCIDevice, features, 2973 VFIO_FEATURE_ENABLE_VGA_BIT, false), 2974 DEFINE_PROP_BIT("x-req", VFIOPCIDevice, features, 2975 VFIO_FEATURE_ENABLE_REQ_BIT, true), 2976 DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice, features, 2977 VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT, false), 2978 DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice, vbasedev.no_mmap, false), 2979 DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice, no_kvm_intx, false), 2980 DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice, no_kvm_msi, false), 2981 DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice, no_kvm_msix, false), 2982 DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice, vendor_id, PCI_ANY_ID), 2983 DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice, device_id, PCI_ANY_ID), 2984 DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice, 2985 sub_vendor_id, PCI_ANY_ID), 2986 DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice, 2987 sub_device_id, PCI_ANY_ID), 2988 DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice, igd_gms, 0), 2989 /* 2990 * TODO - support passed fds... is this necessary? 2991 * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name), 2992 * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name), 2993 */ 2994 DEFINE_PROP_END_OF_LIST(), 2995 }; 2996 2997 static const VMStateDescription vfio_pci_vmstate = { 2998 .name = "vfio-pci", 2999 .unmigratable = 1, 3000 }; 3001 3002 static void vfio_pci_dev_class_init(ObjectClass *klass, void *data) 3003 { 3004 DeviceClass *dc = DEVICE_CLASS(klass); 3005 PCIDeviceClass *pdc = PCI_DEVICE_CLASS(klass); 3006 3007 dc->reset = vfio_pci_reset; 3008 dc->props = vfio_pci_dev_properties; 3009 dc->vmsd = &vfio_pci_vmstate; 3010 dc->desc = "VFIO-based PCI device assignment"; 3011 set_bit(DEVICE_CATEGORY_MISC, dc->categories); 3012 pdc->realize = vfio_realize; 3013 pdc->exit = vfio_exitfn; 3014 pdc->config_read = vfio_pci_read_config; 3015 pdc->config_write = vfio_pci_write_config; 3016 pdc->is_express = 1; /* We might be */ 3017 } 3018 3019 static const TypeInfo vfio_pci_dev_info = { 3020 .name = "vfio-pci", 3021 .parent = TYPE_PCI_DEVICE, 3022 .instance_size = sizeof(VFIOPCIDevice), 3023 .class_init = vfio_pci_dev_class_init, 3024 .instance_init = vfio_instance_init, 3025 .instance_finalize = vfio_instance_finalize, 3026 }; 3027 3028 static void register_vfio_pci_dev_type(void) 3029 { 3030 type_register_static(&vfio_pci_dev_info); 3031 } 3032 3033 type_init(register_vfio_pci_dev_type) 3034