2 * vfio based device assignment support
4 * Copyright Red Hat, Inc. 2012
7 * Alex Williamson <alex.williamson@redhat.com>
9 * This work is licensed under the terms of the GNU GPL, version 2. See
10 * the COPYING file in the top-level directory.
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)
21 #include "qemu/osdep.h"
22 #include <linux/vfio.h>
23 #include <sys/ioctl.h>
26 #include "hw/pci/msi.h"
27 #include "hw/pci/msix.h"
28 #include "hw/pci/pci_bridge.h"
29 #include "hw/qdev-properties.h"
30 #include "hw/qdev-properties-system.h"
31 #include "migration/vmstate.h"
32 #include "qapi/qmp/qdict.h"
33 #include "qemu/error-report.h"
34 #include "qemu/main-loop.h"
35 #include "qemu/module.h"
36 #include "qemu/range.h"
37 #include "qemu/units.h"
38 #include "sysemu/kvm.h"
39 #include "sysemu/runstate.h"
42 #include "qapi/error.h"
43 #include "migration/blocker.h"
44 #include "migration/qemu-file.h"
46 #define TYPE_VFIO_PCI_NOHOTPLUG "vfio-pci-nohotplug"
48 static void vfio_disable_interrupts(VFIOPCIDevice
*vdev
);
49 static void vfio_mmap_set_enabled(VFIOPCIDevice
*vdev
, bool enabled
);
52 * Disabling BAR mmaping can be slow, but toggling it around INTx can
53 * also be a huge overhead. We try to get the best of both worlds by
54 * waiting until an interrupt to disable mmaps (subsequent transitions
55 * to the same state are effectively no overhead). If the interrupt has
56 * been serviced and the time gap is long enough, we re-enable mmaps for
57 * performance. This works well for things like graphics cards, which
58 * may not use their interrupt at all and are penalized to an unusable
59 * level by read/write BAR traps. Other devices, like NICs, have more
60 * regular interrupts and see much better latency by staying in non-mmap
61 * mode. We therefore set the default mmap_timeout such that a ping
62 * is just enough to keep the mmap disabled. Users can experiment with
63 * other options with the x-intx-mmap-timeout-ms parameter (a value of
64 * zero disables the timer).
66 static void vfio_intx_mmap_enable(void *opaque
)
68 VFIOPCIDevice
*vdev
= opaque
;
70 if (vdev
->intx
.pending
) {
71 timer_mod(vdev
->intx
.mmap_timer
,
72 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) + vdev
->intx
.mmap_timeout
);
76 vfio_mmap_set_enabled(vdev
, true);
79 static void vfio_intx_interrupt(void *opaque
)
81 VFIOPCIDevice
*vdev
= opaque
;
83 if (!event_notifier_test_and_clear(&vdev
->intx
.interrupt
)) {
87 trace_vfio_intx_interrupt(vdev
->vbasedev
.name
, 'A' + vdev
->intx
.pin
);
89 vdev
->intx
.pending
= true;
90 pci_irq_assert(&vdev
->pdev
);
91 vfio_mmap_set_enabled(vdev
, false);
92 if (vdev
->intx
.mmap_timeout
) {
93 timer_mod(vdev
->intx
.mmap_timer
,
94 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) + vdev
->intx
.mmap_timeout
);
98 static void vfio_intx_eoi(VFIODevice
*vbasedev
)
100 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
102 if (!vdev
->intx
.pending
) {
106 trace_vfio_intx_eoi(vbasedev
->name
);
108 vdev
->intx
.pending
= false;
109 pci_irq_deassert(&vdev
->pdev
);
110 vfio_unmask_single_irqindex(vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
113 static void vfio_intx_enable_kvm(VFIOPCIDevice
*vdev
, Error
**errp
)
116 int irq_fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
118 if (vdev
->no_kvm_intx
|| !kvm_irqfds_enabled() ||
119 vdev
->intx
.route
.mode
!= PCI_INTX_ENABLED
||
120 !kvm_resamplefds_enabled()) {
124 /* Get to a known interrupt state */
125 qemu_set_fd_handler(irq_fd
, NULL
, NULL
, vdev
);
126 vfio_mask_single_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
127 vdev
->intx
.pending
= false;
128 pci_irq_deassert(&vdev
->pdev
);
130 /* Get an eventfd for resample/unmask */
131 if (event_notifier_init(&vdev
->intx
.unmask
, 0)) {
132 error_setg(errp
, "event_notifier_init failed eoi");
136 if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state
,
137 &vdev
->intx
.interrupt
,
139 vdev
->intx
.route
.irq
)) {
140 error_setg_errno(errp
, errno
, "failed to setup resample irqfd");
144 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
, 0,
145 VFIO_IRQ_SET_ACTION_UNMASK
,
146 event_notifier_get_fd(&vdev
->intx
.unmask
),
152 vfio_unmask_single_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
154 vdev
->intx
.kvm_accel
= true;
156 trace_vfio_intx_enable_kvm(vdev
->vbasedev
.name
);
161 kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state
, &vdev
->intx
.interrupt
,
162 vdev
->intx
.route
.irq
);
164 event_notifier_cleanup(&vdev
->intx
.unmask
);
166 qemu_set_fd_handler(irq_fd
, vfio_intx_interrupt
, NULL
, vdev
);
167 vfio_unmask_single_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
171 static void vfio_intx_disable_kvm(VFIOPCIDevice
*vdev
)
174 if (!vdev
->intx
.kvm_accel
) {
179 * Get to a known state, hardware masked, QEMU ready to accept new
180 * interrupts, QEMU IRQ de-asserted.
182 vfio_mask_single_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
183 vdev
->intx
.pending
= false;
184 pci_irq_deassert(&vdev
->pdev
);
186 /* Tell KVM to stop listening for an INTx irqfd */
187 if (kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state
, &vdev
->intx
.interrupt
,
188 vdev
->intx
.route
.irq
)) {
189 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
192 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
193 event_notifier_cleanup(&vdev
->intx
.unmask
);
195 /* QEMU starts listening for interrupt events. */
196 qemu_set_fd_handler(event_notifier_get_fd(&vdev
->intx
.interrupt
),
197 vfio_intx_interrupt
, NULL
, vdev
);
199 vdev
->intx
.kvm_accel
= false;
201 /* If we've missed an event, let it re-fire through QEMU */
202 vfio_unmask_single_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
204 trace_vfio_intx_disable_kvm(vdev
->vbasedev
.name
);
208 static void vfio_intx_update(VFIOPCIDevice
*vdev
, PCIINTxRoute
*route
)
212 trace_vfio_intx_update(vdev
->vbasedev
.name
,
213 vdev
->intx
.route
.irq
, route
->irq
);
215 vfio_intx_disable_kvm(vdev
);
217 vdev
->intx
.route
= *route
;
219 if (route
->mode
!= PCI_INTX_ENABLED
) {
223 vfio_intx_enable_kvm(vdev
, &err
);
225 warn_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
228 /* Re-enable the interrupt in cased we missed an EOI */
229 vfio_intx_eoi(&vdev
->vbasedev
);
232 static void vfio_intx_routing_notifier(PCIDevice
*pdev
)
234 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
237 if (vdev
->interrupt
!= VFIO_INT_INTx
) {
241 route
= pci_device_route_intx_to_irq(&vdev
->pdev
, vdev
->intx
.pin
);
243 if (pci_intx_route_changed(&vdev
->intx
.route
, &route
)) {
244 vfio_intx_update(vdev
, &route
);
248 static void vfio_irqchip_change(Notifier
*notify
, void *data
)
250 VFIOPCIDevice
*vdev
= container_of(notify
, VFIOPCIDevice
,
251 irqchip_change_notifier
);
253 vfio_intx_update(vdev
, &vdev
->intx
.route
);
256 static int vfio_intx_enable(VFIOPCIDevice
*vdev
, Error
**errp
)
258 uint8_t pin
= vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1);
268 vfio_disable_interrupts(vdev
);
270 vdev
->intx
.pin
= pin
- 1; /* Pin A (1) -> irq[0] */
271 pci_config_set_interrupt_pin(vdev
->pdev
.config
, pin
);
275 * Only conditional to avoid generating error messages on platforms
276 * where we won't actually use the result anyway.
278 if (kvm_irqfds_enabled() && kvm_resamplefds_enabled()) {
279 vdev
->intx
.route
= pci_device_route_intx_to_irq(&vdev
->pdev
,
284 ret
= event_notifier_init(&vdev
->intx
.interrupt
, 0);
286 error_setg_errno(errp
, -ret
, "event_notifier_init failed");
289 fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
290 qemu_set_fd_handler(fd
, vfio_intx_interrupt
, NULL
, vdev
);
292 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
, 0,
293 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, errp
)) {
294 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
295 event_notifier_cleanup(&vdev
->intx
.interrupt
);
299 vfio_intx_enable_kvm(vdev
, &err
);
301 warn_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
304 vdev
->interrupt
= VFIO_INT_INTx
;
306 trace_vfio_intx_enable(vdev
->vbasedev
.name
);
310 static void vfio_intx_disable(VFIOPCIDevice
*vdev
)
314 timer_del(vdev
->intx
.mmap_timer
);
315 vfio_intx_disable_kvm(vdev
);
316 vfio_disable_irqindex(&vdev
->vbasedev
, VFIO_PCI_INTX_IRQ_INDEX
);
317 vdev
->intx
.pending
= false;
318 pci_irq_deassert(&vdev
->pdev
);
319 vfio_mmap_set_enabled(vdev
, true);
321 fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
322 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
323 event_notifier_cleanup(&vdev
->intx
.interrupt
);
325 vdev
->interrupt
= VFIO_INT_NONE
;
327 trace_vfio_intx_disable(vdev
->vbasedev
.name
);
333 static void vfio_msi_interrupt(void *opaque
)
335 VFIOMSIVector
*vector
= opaque
;
336 VFIOPCIDevice
*vdev
= vector
->vdev
;
337 MSIMessage (*get_msg
)(PCIDevice
*dev
, unsigned vector
);
338 void (*notify
)(PCIDevice
*dev
, unsigned vector
);
340 int nr
= vector
- vdev
->msi_vectors
;
342 if (!event_notifier_test_and_clear(&vector
->interrupt
)) {
346 if (vdev
->interrupt
== VFIO_INT_MSIX
) {
347 get_msg
= msix_get_message
;
348 notify
= msix_notify
;
350 /* A masked vector firing needs to use the PBA, enable it */
351 if (msix_is_masked(&vdev
->pdev
, nr
)) {
352 set_bit(nr
, vdev
->msix
->pending
);
353 memory_region_set_enabled(&vdev
->pdev
.msix_pba_mmio
, true);
354 trace_vfio_msix_pba_enable(vdev
->vbasedev
.name
);
356 } else if (vdev
->interrupt
== VFIO_INT_MSI
) {
357 get_msg
= msi_get_message
;
363 msg
= get_msg(&vdev
->pdev
, nr
);
364 trace_vfio_msi_interrupt(vdev
->vbasedev
.name
, nr
, msg
.address
, msg
.data
);
365 notify(&vdev
->pdev
, nr
);
368 static int vfio_enable_vectors(VFIOPCIDevice
*vdev
, bool msix
)
370 struct vfio_irq_set
*irq_set
;
371 int ret
= 0, i
, argsz
;
374 argsz
= sizeof(*irq_set
) + (vdev
->nr_vectors
* sizeof(*fds
));
376 irq_set
= g_malloc0(argsz
);
377 irq_set
->argsz
= argsz
;
378 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_TRIGGER
;
379 irq_set
->index
= msix
? VFIO_PCI_MSIX_IRQ_INDEX
: VFIO_PCI_MSI_IRQ_INDEX
;
381 irq_set
->count
= vdev
->nr_vectors
;
382 fds
= (int32_t *)&irq_set
->data
;
384 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
388 * MSI vs MSI-X - The guest has direct access to MSI mask and pending
389 * bits, therefore we always use the KVM signaling path when setup.
390 * MSI-X mask and pending bits are emulated, so we want to use the
391 * KVM signaling path only when configured and unmasked.
393 if (vdev
->msi_vectors
[i
].use
) {
394 if (vdev
->msi_vectors
[i
].virq
< 0 ||
395 (msix
&& msix_is_masked(&vdev
->pdev
, i
))) {
396 fd
= event_notifier_get_fd(&vdev
->msi_vectors
[i
].interrupt
);
398 fd
= event_notifier_get_fd(&vdev
->msi_vectors
[i
].kvm_interrupt
);
405 ret
= ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
412 static void vfio_add_kvm_msi_virq(VFIOPCIDevice
*vdev
, VFIOMSIVector
*vector
,
413 int vector_n
, bool msix
)
418 if ((msix
&& vdev
->no_kvm_msix
) || (!msix
&& vdev
->no_kvm_msi
)) {
422 if (event_notifier_init(&vector
->kvm_interrupt
, 0)) {
426 c
= kvm_irqchip_begin_route_changes(kvm_state
);
427 virq
= kvm_irqchip_add_msi_route(&c
, vector_n
, &vdev
->pdev
);
429 event_notifier_cleanup(&vector
->kvm_interrupt
);
432 kvm_irqchip_commit_route_changes(&c
);
434 if (kvm_irqchip_add_irqfd_notifier_gsi(kvm_state
, &vector
->kvm_interrupt
,
436 kvm_irqchip_release_virq(kvm_state
, virq
);
437 event_notifier_cleanup(&vector
->kvm_interrupt
);
444 static void vfio_remove_kvm_msi_virq(VFIOMSIVector
*vector
)
446 kvm_irqchip_remove_irqfd_notifier_gsi(kvm_state
, &vector
->kvm_interrupt
,
448 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
450 event_notifier_cleanup(&vector
->kvm_interrupt
);
453 static void vfio_update_kvm_msi_virq(VFIOMSIVector
*vector
, MSIMessage msg
,
456 kvm_irqchip_update_msi_route(kvm_state
, vector
->virq
, msg
, pdev
);
457 kvm_irqchip_commit_routes(kvm_state
);
460 static int vfio_msix_vector_do_use(PCIDevice
*pdev
, unsigned int nr
,
461 MSIMessage
*msg
, IOHandler
*handler
)
463 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
464 VFIOMSIVector
*vector
;
467 trace_vfio_msix_vector_do_use(vdev
->vbasedev
.name
, nr
);
469 vector
= &vdev
->msi_vectors
[nr
];
474 if (event_notifier_init(&vector
->interrupt
, 0)) {
475 error_report("vfio: Error: event_notifier_init failed");
478 msix_vector_use(pdev
, nr
);
481 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
482 handler
, NULL
, vector
);
485 * Attempt to enable route through KVM irqchip,
486 * default to userspace handling if unavailable.
488 if (vector
->virq
>= 0) {
490 vfio_remove_kvm_msi_virq(vector
);
492 vfio_update_kvm_msi_virq(vector
, *msg
, pdev
);
496 vfio_add_kvm_msi_virq(vdev
, vector
, nr
, true);
501 * We don't want to have the host allocate all possible MSI vectors
502 * for a device if they're not in use, so we shutdown and incrementally
503 * increase them as needed.
505 if (vdev
->nr_vectors
< nr
+ 1) {
506 vfio_disable_irqindex(&vdev
->vbasedev
, VFIO_PCI_MSIX_IRQ_INDEX
);
507 vdev
->nr_vectors
= nr
+ 1;
508 ret
= vfio_enable_vectors(vdev
, true);
510 error_report("vfio: failed to enable vectors, %d", ret
);
516 if (vector
->virq
>= 0) {
517 fd
= event_notifier_get_fd(&vector
->kvm_interrupt
);
519 fd
= event_notifier_get_fd(&vector
->interrupt
);
522 if (vfio_set_irq_signaling(&vdev
->vbasedev
,
523 VFIO_PCI_MSIX_IRQ_INDEX
, nr
,
524 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, &err
)) {
525 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
529 /* Disable PBA emulation when nothing more is pending. */
530 clear_bit(nr
, vdev
->msix
->pending
);
531 if (find_first_bit(vdev
->msix
->pending
,
532 vdev
->nr_vectors
) == vdev
->nr_vectors
) {
533 memory_region_set_enabled(&vdev
->pdev
.msix_pba_mmio
, false);
534 trace_vfio_msix_pba_disable(vdev
->vbasedev
.name
);
540 static int vfio_msix_vector_use(PCIDevice
*pdev
,
541 unsigned int nr
, MSIMessage msg
)
543 return vfio_msix_vector_do_use(pdev
, nr
, &msg
, vfio_msi_interrupt
);
546 static void vfio_msix_vector_release(PCIDevice
*pdev
, unsigned int nr
)
548 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
549 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[nr
];
551 trace_vfio_msix_vector_release(vdev
->vbasedev
.name
, nr
);
554 * There are still old guests that mask and unmask vectors on every
555 * interrupt. If we're using QEMU bypass with a KVM irqfd, leave all of
556 * the KVM setup in place, simply switch VFIO to use the non-bypass
557 * eventfd. We'll then fire the interrupt through QEMU and the MSI-X
558 * core will mask the interrupt and set pending bits, allowing it to
559 * be re-asserted on unmask. Nothing to do if already using QEMU mode.
561 if (vector
->virq
>= 0) {
562 int32_t fd
= event_notifier_get_fd(&vector
->interrupt
);
565 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_MSIX_IRQ_INDEX
, nr
,
566 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, &err
)) {
567 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
572 static void vfio_msix_enable(VFIOPCIDevice
*vdev
)
574 PCIDevice
*pdev
= &vdev
->pdev
;
575 unsigned int nr
, max_vec
= 0;
577 vfio_disable_interrupts(vdev
);
579 vdev
->msi_vectors
= g_new0(VFIOMSIVector
, vdev
->msix
->entries
);
581 vdev
->interrupt
= VFIO_INT_MSIX
;
584 * Some communication channels between VF & PF or PF & fw rely on the
585 * physical state of the device and expect that enabling MSI-X from the
586 * guest enables the same on the host. When our guest is Linux, the
587 * guest driver call to pci_enable_msix() sets the enabling bit in the
588 * MSI-X capability, but leaves the vector table masked. We therefore
589 * can't rely on a vector_use callback (from request_irq() in the guest)
590 * to switch the physical device into MSI-X mode because that may come a
591 * long time after pci_enable_msix(). This code enables vector 0 with
592 * triggering to userspace, then immediately release the vector, leaving
593 * the physical device with no vectors enabled, but MSI-X enabled, just
594 * like the guest view.
595 * If there are already unmasked vectors (in migration resume phase and
596 * some guest startups) which will be enabled soon, we can allocate all
597 * of them here to avoid inefficiently disabling and enabling vectors
600 if (!pdev
->msix_function_masked
) {
601 for (nr
= 0; nr
< msix_nr_vectors_allocated(pdev
); nr
++) {
602 if (!msix_is_masked(pdev
, nr
)) {
607 vfio_msix_vector_do_use(pdev
, max_vec
, NULL
, NULL
);
608 vfio_msix_vector_release(pdev
, max_vec
);
610 if (msix_set_vector_notifiers(pdev
, vfio_msix_vector_use
,
611 vfio_msix_vector_release
, NULL
)) {
612 error_report("vfio: msix_set_vector_notifiers failed");
615 trace_vfio_msix_enable(vdev
->vbasedev
.name
);
618 static void vfio_msi_enable(VFIOPCIDevice
*vdev
)
622 vfio_disable_interrupts(vdev
);
624 vdev
->nr_vectors
= msi_nr_vectors_allocated(&vdev
->pdev
);
626 vdev
->msi_vectors
= g_new0(VFIOMSIVector
, vdev
->nr_vectors
);
628 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
629 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
635 if (event_notifier_init(&vector
->interrupt
, 0)) {
636 error_report("vfio: Error: event_notifier_init failed");
639 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
640 vfio_msi_interrupt
, NULL
, vector
);
643 * Attempt to enable route through KVM irqchip,
644 * default to userspace handling if unavailable.
646 vfio_add_kvm_msi_virq(vdev
, vector
, i
, false);
649 /* Set interrupt type prior to possible interrupts */
650 vdev
->interrupt
= VFIO_INT_MSI
;
652 ret
= vfio_enable_vectors(vdev
, false);
655 error_report("vfio: Error: Failed to setup MSI fds: %m");
656 } else if (ret
!= vdev
->nr_vectors
) {
657 error_report("vfio: Error: Failed to enable %d "
658 "MSI vectors, retry with %d", vdev
->nr_vectors
, ret
);
661 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
662 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
663 if (vector
->virq
>= 0) {
664 vfio_remove_kvm_msi_virq(vector
);
666 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
668 event_notifier_cleanup(&vector
->interrupt
);
671 g_free(vdev
->msi_vectors
);
672 vdev
->msi_vectors
= NULL
;
674 if (ret
> 0 && ret
!= vdev
->nr_vectors
) {
675 vdev
->nr_vectors
= ret
;
678 vdev
->nr_vectors
= 0;
681 * Failing to setup MSI doesn't really fall within any specification.
682 * Let's try leaving interrupts disabled and hope the guest figures
683 * out to fall back to INTx for this device.
685 error_report("vfio: Error: Failed to enable MSI");
686 vdev
->interrupt
= VFIO_INT_NONE
;
691 trace_vfio_msi_enable(vdev
->vbasedev
.name
, vdev
->nr_vectors
);
694 static void vfio_msi_disable_common(VFIOPCIDevice
*vdev
)
699 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
700 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
701 if (vdev
->msi_vectors
[i
].use
) {
702 if (vector
->virq
>= 0) {
703 vfio_remove_kvm_msi_virq(vector
);
705 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
707 event_notifier_cleanup(&vector
->interrupt
);
711 g_free(vdev
->msi_vectors
);
712 vdev
->msi_vectors
= NULL
;
713 vdev
->nr_vectors
= 0;
714 vdev
->interrupt
= VFIO_INT_NONE
;
716 vfio_intx_enable(vdev
, &err
);
718 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
722 static void vfio_msix_disable(VFIOPCIDevice
*vdev
)
726 msix_unset_vector_notifiers(&vdev
->pdev
);
729 * MSI-X will only release vectors if MSI-X is still enabled on the
730 * device, check through the rest and release it ourselves if necessary.
732 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
733 if (vdev
->msi_vectors
[i
].use
) {
734 vfio_msix_vector_release(&vdev
->pdev
, i
);
735 msix_vector_unuse(&vdev
->pdev
, i
);
739 if (vdev
->nr_vectors
) {
740 vfio_disable_irqindex(&vdev
->vbasedev
, VFIO_PCI_MSIX_IRQ_INDEX
);
743 vfio_msi_disable_common(vdev
);
745 memset(vdev
->msix
->pending
, 0,
746 BITS_TO_LONGS(vdev
->msix
->entries
) * sizeof(unsigned long));
748 trace_vfio_msix_disable(vdev
->vbasedev
.name
);
751 static void vfio_msi_disable(VFIOPCIDevice
*vdev
)
753 vfio_disable_irqindex(&vdev
->vbasedev
, VFIO_PCI_MSI_IRQ_INDEX
);
754 vfio_msi_disable_common(vdev
);
756 trace_vfio_msi_disable(vdev
->vbasedev
.name
);
759 static void vfio_update_msi(VFIOPCIDevice
*vdev
)
763 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
764 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
767 if (!vector
->use
|| vector
->virq
< 0) {
771 msg
= msi_get_message(&vdev
->pdev
, i
);
772 vfio_update_kvm_msi_virq(vector
, msg
, &vdev
->pdev
);
776 static void vfio_pci_load_rom(VFIOPCIDevice
*vdev
)
778 struct vfio_region_info
*reg_info
;
783 if (vfio_get_region_info(&vdev
->vbasedev
,
784 VFIO_PCI_ROM_REGION_INDEX
, ®_info
)) {
785 error_report("vfio: Error getting ROM info: %m");
789 trace_vfio_pci_load_rom(vdev
->vbasedev
.name
, (unsigned long)reg_info
->size
,
790 (unsigned long)reg_info
->offset
,
791 (unsigned long)reg_info
->flags
);
793 vdev
->rom_size
= size
= reg_info
->size
;
794 vdev
->rom_offset
= reg_info
->offset
;
798 if (!vdev
->rom_size
) {
799 vdev
->rom_read_failed
= true;
800 error_report("vfio-pci: Cannot read device rom at "
801 "%s", vdev
->vbasedev
.name
);
802 error_printf("Device option ROM contents are probably invalid "
803 "(check dmesg).\nSkip option ROM probe with rombar=0, "
804 "or load from file with romfile=\n");
808 vdev
->rom
= g_malloc(size
);
809 memset(vdev
->rom
, 0xff, size
);
812 bytes
= pread(vdev
->vbasedev
.fd
, vdev
->rom
+ off
,
813 size
, vdev
->rom_offset
+ off
);
816 } else if (bytes
> 0) {
820 if (errno
== EINTR
|| errno
== EAGAIN
) {
823 error_report("vfio: Error reading device ROM: %m");
829 * Test the ROM signature against our device, if the vendor is correct
830 * but the device ID doesn't match, store the correct device ID and
831 * recompute the checksum. Intel IGD devices need this and are known
832 * to have bogus checksums so we can't simply adjust the checksum.
834 if (pci_get_word(vdev
->rom
) == 0xaa55 &&
835 pci_get_word(vdev
->rom
+ 0x18) + 8 < vdev
->rom_size
&&
836 !memcmp(vdev
->rom
+ pci_get_word(vdev
->rom
+ 0x18), "PCIR", 4)) {
839 vid
= pci_get_word(vdev
->rom
+ pci_get_word(vdev
->rom
+ 0x18) + 4);
840 did
= pci_get_word(vdev
->rom
+ pci_get_word(vdev
->rom
+ 0x18) + 6);
842 if (vid
== vdev
->vendor_id
&& did
!= vdev
->device_id
) {
844 uint8_t csum
, *data
= vdev
->rom
;
846 pci_set_word(vdev
->rom
+ pci_get_word(vdev
->rom
+ 0x18) + 6,
850 for (csum
= 0, i
= 0; i
< vdev
->rom_size
; i
++) {
859 static uint64_t vfio_rom_read(void *opaque
, hwaddr addr
, unsigned size
)
861 VFIOPCIDevice
*vdev
= opaque
;
870 /* Load the ROM lazily when the guest tries to read it */
871 if (unlikely(!vdev
->rom
&& !vdev
->rom_read_failed
)) {
872 vfio_pci_load_rom(vdev
);
875 memcpy(&val
, vdev
->rom
+ addr
,
876 (addr
< vdev
->rom_size
) ? MIN(size
, vdev
->rom_size
- addr
) : 0);
883 data
= le16_to_cpu(val
.word
);
886 data
= le32_to_cpu(val
.dword
);
889 hw_error("vfio: unsupported read size, %d bytes\n", size
);
893 trace_vfio_rom_read(vdev
->vbasedev
.name
, addr
, size
, data
);
898 static void vfio_rom_write(void *opaque
, hwaddr addr
,
899 uint64_t data
, unsigned size
)
903 static const MemoryRegionOps vfio_rom_ops
= {
904 .read
= vfio_rom_read
,
905 .write
= vfio_rom_write
,
906 .endianness
= DEVICE_LITTLE_ENDIAN
,
909 static void vfio_pci_size_rom(VFIOPCIDevice
*vdev
)
911 uint32_t orig
, size
= cpu_to_le32((uint32_t)PCI_ROM_ADDRESS_MASK
);
912 off_t offset
= vdev
->config_offset
+ PCI_ROM_ADDRESS
;
913 DeviceState
*dev
= DEVICE(vdev
);
915 int fd
= vdev
->vbasedev
.fd
;
917 if (vdev
->pdev
.romfile
|| !vdev
->pdev
.rom_bar
) {
918 /* Since pci handles romfile, just print a message and return */
919 if (vfio_opt_rom_in_denylist(vdev
) && vdev
->pdev
.romfile
) {
920 warn_report("Device at %s is known to cause system instability"
921 " issues during option rom execution",
922 vdev
->vbasedev
.name
);
923 error_printf("Proceeding anyway since user specified romfile\n");
929 * Use the same size ROM BAR as the physical device. The contents
930 * will get filled in later when the guest tries to read it.
932 if (pread(fd
, &orig
, 4, offset
) != 4 ||
933 pwrite(fd
, &size
, 4, offset
) != 4 ||
934 pread(fd
, &size
, 4, offset
) != 4 ||
935 pwrite(fd
, &orig
, 4, offset
) != 4) {
936 error_report("%s(%s) failed: %m", __func__
, vdev
->vbasedev
.name
);
940 size
= ~(le32_to_cpu(size
) & PCI_ROM_ADDRESS_MASK
) + 1;
946 if (vfio_opt_rom_in_denylist(vdev
)) {
947 if (dev
->opts
&& qdict_haskey(dev
->opts
, "rombar")) {
948 warn_report("Device at %s is known to cause system instability"
949 " issues during option rom execution",
950 vdev
->vbasedev
.name
);
951 error_printf("Proceeding anyway since user specified"
952 " non zero value for rombar\n");
954 warn_report("Rom loading for device at %s has been disabled"
955 " due to system instability issues",
956 vdev
->vbasedev
.name
);
957 error_printf("Specify rombar=1 or romfile to force\n");
962 trace_vfio_pci_size_rom(vdev
->vbasedev
.name
, size
);
964 name
= g_strdup_printf("vfio[%s].rom", vdev
->vbasedev
.name
);
966 memory_region_init_io(&vdev
->pdev
.rom
, OBJECT(vdev
),
967 &vfio_rom_ops
, vdev
, name
, size
);
970 pci_register_bar(&vdev
->pdev
, PCI_ROM_SLOT
,
971 PCI_BASE_ADDRESS_SPACE_MEMORY
, &vdev
->pdev
.rom
);
973 vdev
->rom_read_failed
= false;
976 void vfio_vga_write(void *opaque
, hwaddr addr
,
977 uint64_t data
, unsigned size
)
979 VFIOVGARegion
*region
= opaque
;
980 VFIOVGA
*vga
= container_of(region
, VFIOVGA
, region
[region
->nr
]);
987 off_t offset
= vga
->fd_offset
+ region
->offset
+ addr
;
994 buf
.word
= cpu_to_le16(data
);
997 buf
.dword
= cpu_to_le32(data
);
1000 hw_error("vfio: unsupported write size, %d bytes", size
);
1004 if (pwrite(vga
->fd
, &buf
, size
, offset
) != size
) {
1005 error_report("%s(,0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d) failed: %m",
1006 __func__
, region
->offset
+ addr
, data
, size
);
1009 trace_vfio_vga_write(region
->offset
+ addr
, data
, size
);
1012 uint64_t vfio_vga_read(void *opaque
, hwaddr addr
, unsigned size
)
1014 VFIOVGARegion
*region
= opaque
;
1015 VFIOVGA
*vga
= container_of(region
, VFIOVGA
, region
[region
->nr
]);
1023 off_t offset
= vga
->fd_offset
+ region
->offset
+ addr
;
1025 if (pread(vga
->fd
, &buf
, size
, offset
) != size
) {
1026 error_report("%s(,0x%"HWADDR_PRIx
", %d) failed: %m",
1027 __func__
, region
->offset
+ addr
, size
);
1028 return (uint64_t)-1;
1036 data
= le16_to_cpu(buf
.word
);
1039 data
= le32_to_cpu(buf
.dword
);
1042 hw_error("vfio: unsupported read size, %d bytes", size
);
1046 trace_vfio_vga_read(region
->offset
+ addr
, size
, data
);
1051 static const MemoryRegionOps vfio_vga_ops
= {
1052 .read
= vfio_vga_read
,
1053 .write
= vfio_vga_write
,
1054 .endianness
= DEVICE_LITTLE_ENDIAN
,
1058 * Expand memory region of sub-page(size < PAGE_SIZE) MMIO BAR to page
1059 * size if the BAR is in an exclusive page in host so that we could map
1060 * this BAR to guest. But this sub-page BAR may not occupy an exclusive
1061 * page in guest. So we should set the priority of the expanded memory
1062 * region to zero in case of overlap with BARs which share the same page
1063 * with the sub-page BAR in guest. Besides, we should also recover the
1064 * size of this sub-page BAR when its base address is changed in guest
1065 * and not page aligned any more.
1067 static void vfio_sub_page_bar_update_mapping(PCIDevice
*pdev
, int bar
)
1069 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
1070 VFIORegion
*region
= &vdev
->bars
[bar
].region
;
1071 MemoryRegion
*mmap_mr
, *region_mr
, *base_mr
;
1074 uint64_t size
= region
->size
;
1076 /* Make sure that the whole region is allowed to be mmapped */
1077 if (region
->nr_mmaps
!= 1 || !region
->mmaps
[0].mmap
||
1078 region
->mmaps
[0].size
!= region
->size
) {
1082 r
= &pdev
->io_regions
[bar
];
1084 base_mr
= vdev
->bars
[bar
].mr
;
1085 region_mr
= region
->mem
;
1086 mmap_mr
= ®ion
->mmaps
[0].mem
;
1088 /* If BAR is mapped and page aligned, update to fill PAGE_SIZE */
1089 if (bar_addr
!= PCI_BAR_UNMAPPED
&&
1090 !(bar_addr
& ~qemu_real_host_page_mask())) {
1091 size
= qemu_real_host_page_size();
1094 memory_region_transaction_begin();
1096 if (vdev
->bars
[bar
].size
< size
) {
1097 memory_region_set_size(base_mr
, size
);
1099 memory_region_set_size(region_mr
, size
);
1100 memory_region_set_size(mmap_mr
, size
);
1101 if (size
!= vdev
->bars
[bar
].size
&& memory_region_is_mapped(base_mr
)) {
1102 memory_region_del_subregion(r
->address_space
, base_mr
);
1103 memory_region_add_subregion_overlap(r
->address_space
,
1104 bar_addr
, base_mr
, 0);
1107 memory_region_transaction_commit();
1113 uint32_t vfio_pci_read_config(PCIDevice
*pdev
, uint32_t addr
, int len
)
1115 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
1116 uint32_t emu_bits
= 0, emu_val
= 0, phys_val
= 0, val
;
1118 memcpy(&emu_bits
, vdev
->emulated_config_bits
+ addr
, len
);
1119 emu_bits
= le32_to_cpu(emu_bits
);
1122 emu_val
= pci_default_read_config(pdev
, addr
, len
);
1125 if (~emu_bits
& (0xffffffffU
>> (32 - len
* 8))) {
1128 ret
= pread(vdev
->vbasedev
.fd
, &phys_val
, len
,
1129 vdev
->config_offset
+ addr
);
1131 error_report("%s(%s, 0x%x, 0x%x) failed: %m",
1132 __func__
, vdev
->vbasedev
.name
, addr
, len
);
1135 phys_val
= le32_to_cpu(phys_val
);
1138 val
= (emu_val
& emu_bits
) | (phys_val
& ~emu_bits
);
1140 trace_vfio_pci_read_config(vdev
->vbasedev
.name
, addr
, len
, val
);
1145 void vfio_pci_write_config(PCIDevice
*pdev
,
1146 uint32_t addr
, uint32_t val
, int len
)
1148 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
1149 uint32_t val_le
= cpu_to_le32(val
);
1151 trace_vfio_pci_write_config(vdev
->vbasedev
.name
, addr
, val
, len
);
1153 /* Write everything to VFIO, let it filter out what we can't write */
1154 if (pwrite(vdev
->vbasedev
.fd
, &val_le
, len
, vdev
->config_offset
+ addr
)
1156 error_report("%s(%s, 0x%x, 0x%x, 0x%x) failed: %m",
1157 __func__
, vdev
->vbasedev
.name
, addr
, val
, len
);
1160 /* MSI/MSI-X Enabling/Disabling */
1161 if (pdev
->cap_present
& QEMU_PCI_CAP_MSI
&&
1162 ranges_overlap(addr
, len
, pdev
->msi_cap
, vdev
->msi_cap_size
)) {
1163 int is_enabled
, was_enabled
= msi_enabled(pdev
);
1165 pci_default_write_config(pdev
, addr
, val
, len
);
1167 is_enabled
= msi_enabled(pdev
);
1171 vfio_msi_enable(vdev
);
1175 vfio_msi_disable(vdev
);
1177 vfio_update_msi(vdev
);
1180 } else if (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
&&
1181 ranges_overlap(addr
, len
, pdev
->msix_cap
, MSIX_CAP_LENGTH
)) {
1182 int is_enabled
, was_enabled
= msix_enabled(pdev
);
1184 pci_default_write_config(pdev
, addr
, val
, len
);
1186 is_enabled
= msix_enabled(pdev
);
1188 if (!was_enabled
&& is_enabled
) {
1189 vfio_msix_enable(vdev
);
1190 } else if (was_enabled
&& !is_enabled
) {
1191 vfio_msix_disable(vdev
);
1193 } else if (ranges_overlap(addr
, len
, PCI_BASE_ADDRESS_0
, 24) ||
1194 range_covers_byte(addr
, len
, PCI_COMMAND
)) {
1195 pcibus_t old_addr
[PCI_NUM_REGIONS
- 1];
1198 for (bar
= 0; bar
< PCI_ROM_SLOT
; bar
++) {
1199 old_addr
[bar
] = pdev
->io_regions
[bar
].addr
;
1202 pci_default_write_config(pdev
, addr
, val
, len
);
1204 for (bar
= 0; bar
< PCI_ROM_SLOT
; bar
++) {
1205 if (old_addr
[bar
] != pdev
->io_regions
[bar
].addr
&&
1206 vdev
->bars
[bar
].region
.size
> 0 &&
1207 vdev
->bars
[bar
].region
.size
< qemu_real_host_page_size()) {
1208 vfio_sub_page_bar_update_mapping(pdev
, bar
);
1212 /* Write everything to QEMU to keep emulated bits correct */
1213 pci_default_write_config(pdev
, addr
, val
, len
);
1220 static void vfio_disable_interrupts(VFIOPCIDevice
*vdev
)
1223 * More complicated than it looks. Disabling MSI/X transitions the
1224 * device to INTx mode (if supported). Therefore we need to first
1225 * disable MSI/X and then cleanup by disabling INTx.
1227 if (vdev
->interrupt
== VFIO_INT_MSIX
) {
1228 vfio_msix_disable(vdev
);
1229 } else if (vdev
->interrupt
== VFIO_INT_MSI
) {
1230 vfio_msi_disable(vdev
);
1233 if (vdev
->interrupt
== VFIO_INT_INTx
) {
1234 vfio_intx_disable(vdev
);
1238 static int vfio_msi_setup(VFIOPCIDevice
*vdev
, int pos
, Error
**errp
)
1241 bool msi_64bit
, msi_maskbit
;
1245 if (pread(vdev
->vbasedev
.fd
, &ctrl
, sizeof(ctrl
),
1246 vdev
->config_offset
+ pos
+ PCI_CAP_FLAGS
) != sizeof(ctrl
)) {
1247 error_setg_errno(errp
, errno
, "failed reading MSI PCI_CAP_FLAGS");
1250 ctrl
= le16_to_cpu(ctrl
);
1252 msi_64bit
= !!(ctrl
& PCI_MSI_FLAGS_64BIT
);
1253 msi_maskbit
= !!(ctrl
& PCI_MSI_FLAGS_MASKBIT
);
1254 entries
= 1 << ((ctrl
& PCI_MSI_FLAGS_QMASK
) >> 1);
1256 trace_vfio_msi_setup(vdev
->vbasedev
.name
, pos
);
1258 ret
= msi_init(&vdev
->pdev
, pos
, entries
, msi_64bit
, msi_maskbit
, &err
);
1260 if (ret
== -ENOTSUP
) {
1263 error_propagate_prepend(errp
, err
, "msi_init failed: ");
1266 vdev
->msi_cap_size
= 0xa + (msi_maskbit
? 0xa : 0) + (msi_64bit
? 0x4 : 0);
1271 static void vfio_pci_fixup_msix_region(VFIOPCIDevice
*vdev
)
1274 VFIORegion
*region
= &vdev
->bars
[vdev
->msix
->table_bar
].region
;
1277 * If the host driver allows mapping of a MSIX data, we are going to
1278 * do map the entire BAR and emulate MSIX table on top of that.
1280 if (vfio_has_region_cap(&vdev
->vbasedev
, region
->nr
,
1281 VFIO_REGION_INFO_CAP_MSIX_MAPPABLE
)) {
1286 * We expect to find a single mmap covering the whole BAR, anything else
1287 * means it's either unsupported or already setup.
1289 if (region
->nr_mmaps
!= 1 || region
->mmaps
[0].offset
||
1290 region
->size
!= region
->mmaps
[0].size
) {
1294 /* MSI-X table start and end aligned to host page size */
1295 start
= vdev
->msix
->table_offset
& qemu_real_host_page_mask();
1296 end
= REAL_HOST_PAGE_ALIGN((uint64_t)vdev
->msix
->table_offset
+
1297 (vdev
->msix
->entries
* PCI_MSIX_ENTRY_SIZE
));
1300 * Does the MSI-X table cover the beginning of the BAR? The whole BAR?
1301 * NB - Host page size is necessarily a power of two and so is the PCI
1302 * BAR (not counting EA yet), therefore if we have host page aligned
1303 * @start and @end, then any remainder of the BAR before or after those
1304 * must be at least host page sized and therefore mmap'able.
1307 if (end
>= region
->size
) {
1308 region
->nr_mmaps
= 0;
1309 g_free(region
->mmaps
);
1310 region
->mmaps
= NULL
;
1311 trace_vfio_msix_fixup(vdev
->vbasedev
.name
,
1312 vdev
->msix
->table_bar
, 0, 0);
1314 region
->mmaps
[0].offset
= end
;
1315 region
->mmaps
[0].size
= region
->size
- end
;
1316 trace_vfio_msix_fixup(vdev
->vbasedev
.name
,
1317 vdev
->msix
->table_bar
, region
->mmaps
[0].offset
,
1318 region
->mmaps
[0].offset
+ region
->mmaps
[0].size
);
1321 /* Maybe it's aligned at the end of the BAR */
1322 } else if (end
>= region
->size
) {
1323 region
->mmaps
[0].size
= start
;
1324 trace_vfio_msix_fixup(vdev
->vbasedev
.name
,
1325 vdev
->msix
->table_bar
, region
->mmaps
[0].offset
,
1326 region
->mmaps
[0].offset
+ region
->mmaps
[0].size
);
1328 /* Otherwise it must split the BAR */
1330 region
->nr_mmaps
= 2;
1331 region
->mmaps
= g_renew(VFIOMmap
, region
->mmaps
, 2);
1333 memcpy(®ion
->mmaps
[1], ®ion
->mmaps
[0], sizeof(VFIOMmap
));
1335 region
->mmaps
[0].size
= start
;
1336 trace_vfio_msix_fixup(vdev
->vbasedev
.name
,
1337 vdev
->msix
->table_bar
, region
->mmaps
[0].offset
,
1338 region
->mmaps
[0].offset
+ region
->mmaps
[0].size
);
1340 region
->mmaps
[1].offset
= end
;
1341 region
->mmaps
[1].size
= region
->size
- end
;
1342 trace_vfio_msix_fixup(vdev
->vbasedev
.name
,
1343 vdev
->msix
->table_bar
, region
->mmaps
[1].offset
,
1344 region
->mmaps
[1].offset
+ region
->mmaps
[1].size
);
1348 static void vfio_pci_relocate_msix(VFIOPCIDevice
*vdev
, Error
**errp
)
1350 int target_bar
= -1;
1353 if (!vdev
->msix
|| vdev
->msix_relo
== OFF_AUTOPCIBAR_OFF
) {
1357 /* The actual minimum size of MSI-X structures */
1358 msix_sz
= (vdev
->msix
->entries
* PCI_MSIX_ENTRY_SIZE
) +
1359 (QEMU_ALIGN_UP(vdev
->msix
->entries
, 64) / 8);
1360 /* Round up to host pages, we don't want to share a page */
1361 msix_sz
= REAL_HOST_PAGE_ALIGN(msix_sz
);
1362 /* PCI BARs must be a power of 2 */
1363 msix_sz
= pow2ceil(msix_sz
);
1365 if (vdev
->msix_relo
== OFF_AUTOPCIBAR_AUTO
) {
1367 * TODO: Lookup table for known devices.
1369 * Logically we might use an algorithm here to select the BAR adding
1370 * the least additional MMIO space, but we cannot programmatically
1371 * predict the driver dependency on BAR ordering or sizing, therefore
1372 * 'auto' becomes a lookup for combinations reported to work.
1374 if (target_bar
< 0) {
1375 error_setg(errp
, "No automatic MSI-X relocation available for "
1376 "device %04x:%04x", vdev
->vendor_id
, vdev
->device_id
);
1380 target_bar
= (int)(vdev
->msix_relo
- OFF_AUTOPCIBAR_BAR0
);
1383 /* I/O port BARs cannot host MSI-X structures */
1384 if (vdev
->bars
[target_bar
].ioport
) {
1385 error_setg(errp
, "Invalid MSI-X relocation BAR %d, "
1386 "I/O port BAR", target_bar
);
1390 /* Cannot use a BAR in the "shadow" of a 64-bit BAR */
1391 if (!vdev
->bars
[target_bar
].size
&&
1392 target_bar
> 0 && vdev
->bars
[target_bar
- 1].mem64
) {
1393 error_setg(errp
, "Invalid MSI-X relocation BAR %d, "
1394 "consumed by 64-bit BAR %d", target_bar
, target_bar
- 1);
1398 /* 2GB max size for 32-bit BARs, cannot double if already > 1G */
1399 if (vdev
->bars
[target_bar
].size
> 1 * GiB
&&
1400 !vdev
->bars
[target_bar
].mem64
) {
1401 error_setg(errp
, "Invalid MSI-X relocation BAR %d, "
1402 "no space to extend 32-bit BAR", target_bar
);
1407 * If adding a new BAR, test if we can make it 64bit. We make it
1408 * prefetchable since QEMU MSI-X emulation has no read side effects
1409 * and doing so makes mapping more flexible.
1411 if (!vdev
->bars
[target_bar
].size
) {
1412 if (target_bar
< (PCI_ROM_SLOT
- 1) &&
1413 !vdev
->bars
[target_bar
+ 1].size
) {
1414 vdev
->bars
[target_bar
].mem64
= true;
1415 vdev
->bars
[target_bar
].type
= PCI_BASE_ADDRESS_MEM_TYPE_64
;
1417 vdev
->bars
[target_bar
].type
|= PCI_BASE_ADDRESS_MEM_PREFETCH
;
1418 vdev
->bars
[target_bar
].size
= msix_sz
;
1419 vdev
->msix
->table_offset
= 0;
1421 vdev
->bars
[target_bar
].size
= MAX(vdev
->bars
[target_bar
].size
* 2,
1424 * Due to above size calc, MSI-X always starts halfway into the BAR,
1425 * which will always be a separate host page.
1427 vdev
->msix
->table_offset
= vdev
->bars
[target_bar
].size
/ 2;
1430 vdev
->msix
->table_bar
= target_bar
;
1431 vdev
->msix
->pba_bar
= target_bar
;
1432 /* Requires 8-byte alignment, but PCI_MSIX_ENTRY_SIZE guarantees that */
1433 vdev
->msix
->pba_offset
= vdev
->msix
->table_offset
+
1434 (vdev
->msix
->entries
* PCI_MSIX_ENTRY_SIZE
);
1436 trace_vfio_msix_relo(vdev
->vbasedev
.name
,
1437 vdev
->msix
->table_bar
, vdev
->msix
->table_offset
);
1441 * We don't have any control over how pci_add_capability() inserts
1442 * capabilities into the chain. In order to setup MSI-X we need a
1443 * MemoryRegion for the BAR. In order to setup the BAR and not
1444 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1445 * need to first look for where the MSI-X table lives. So we
1446 * unfortunately split MSI-X setup across two functions.
1448 static void vfio_msix_early_setup(VFIOPCIDevice
*vdev
, Error
**errp
)
1452 uint32_t table
, pba
;
1453 int fd
= vdev
->vbasedev
.fd
;
1456 pos
= pci_find_capability(&vdev
->pdev
, PCI_CAP_ID_MSIX
);
1461 if (pread(fd
, &ctrl
, sizeof(ctrl
),
1462 vdev
->config_offset
+ pos
+ PCI_MSIX_FLAGS
) != sizeof(ctrl
)) {
1463 error_setg_errno(errp
, errno
, "failed to read PCI MSIX FLAGS");
1467 if (pread(fd
, &table
, sizeof(table
),
1468 vdev
->config_offset
+ pos
+ PCI_MSIX_TABLE
) != sizeof(table
)) {
1469 error_setg_errno(errp
, errno
, "failed to read PCI MSIX TABLE");
1473 if (pread(fd
, &pba
, sizeof(pba
),
1474 vdev
->config_offset
+ pos
+ PCI_MSIX_PBA
) != sizeof(pba
)) {
1475 error_setg_errno(errp
, errno
, "failed to read PCI MSIX PBA");
1479 ctrl
= le16_to_cpu(ctrl
);
1480 table
= le32_to_cpu(table
);
1481 pba
= le32_to_cpu(pba
);
1483 msix
= g_malloc0(sizeof(*msix
));
1484 msix
->table_bar
= table
& PCI_MSIX_FLAGS_BIRMASK
;
1485 msix
->table_offset
= table
& ~PCI_MSIX_FLAGS_BIRMASK
;
1486 msix
->pba_bar
= pba
& PCI_MSIX_FLAGS_BIRMASK
;
1487 msix
->pba_offset
= pba
& ~PCI_MSIX_FLAGS_BIRMASK
;
1488 msix
->entries
= (ctrl
& PCI_MSIX_FLAGS_QSIZE
) + 1;
1491 * Test the size of the pba_offset variable and catch if it extends outside
1492 * of the specified BAR. If it is the case, we need to apply a hardware
1493 * specific quirk if the device is known or we have a broken configuration.
1495 if (msix
->pba_offset
>= vdev
->bars
[msix
->pba_bar
].region
.size
) {
1497 * Chelsio T5 Virtual Function devices are encoded as 0x58xx for T5
1498 * adapters. The T5 hardware returns an incorrect value of 0x8000 for
1499 * the VF PBA offset while the BAR itself is only 8k. The correct value
1500 * is 0x1000, so we hard code that here.
1502 if (vdev
->vendor_id
== PCI_VENDOR_ID_CHELSIO
&&
1503 (vdev
->device_id
& 0xff00) == 0x5800) {
1504 msix
->pba_offset
= 0x1000;
1506 * BAIDU KUNLUN Virtual Function devices for KUNLUN AI processor
1507 * return an incorrect value of 0x460000 for the VF PBA offset while
1508 * the BAR itself is only 0x10000. The correct value is 0xb400.
1510 } else if (vfio_pci_is(vdev
, PCI_VENDOR_ID_BAIDU
,
1511 PCI_DEVICE_ID_KUNLUN_VF
)) {
1512 msix
->pba_offset
= 0xb400;
1513 } else if (vdev
->msix_relo
== OFF_AUTOPCIBAR_OFF
) {
1514 error_setg(errp
, "hardware reports invalid configuration, "
1515 "MSIX PBA outside of specified BAR");
1521 trace_vfio_msix_early_setup(vdev
->vbasedev
.name
, pos
, msix
->table_bar
,
1522 msix
->table_offset
, msix
->entries
);
1525 vfio_pci_fixup_msix_region(vdev
);
1527 vfio_pci_relocate_msix(vdev
, errp
);
1530 static int vfio_msix_setup(VFIOPCIDevice
*vdev
, int pos
, Error
**errp
)
1535 vdev
->msix
->pending
= g_new0(unsigned long,
1536 BITS_TO_LONGS(vdev
->msix
->entries
));
1537 ret
= msix_init(&vdev
->pdev
, vdev
->msix
->entries
,
1538 vdev
->bars
[vdev
->msix
->table_bar
].mr
,
1539 vdev
->msix
->table_bar
, vdev
->msix
->table_offset
,
1540 vdev
->bars
[vdev
->msix
->pba_bar
].mr
,
1541 vdev
->msix
->pba_bar
, vdev
->msix
->pba_offset
, pos
,
1544 if (ret
== -ENOTSUP
) {
1545 warn_report_err(err
);
1549 error_propagate(errp
, err
);
1554 * The PCI spec suggests that devices provide additional alignment for
1555 * MSI-X structures and avoid overlapping non-MSI-X related registers.
1556 * For an assigned device, this hopefully means that emulation of MSI-X
1557 * structures does not affect the performance of the device. If devices
1558 * fail to provide that alignment, a significant performance penalty may
1559 * result, for instance Mellanox MT27500 VFs:
1560 * http://www.spinics.net/lists/kvm/msg125881.html
1562 * The PBA is simply not that important for such a serious regression and
1563 * most drivers do not appear to look at it. The solution for this is to
1564 * disable the PBA MemoryRegion unless it's being used. We disable it
1565 * here and only enable it if a masked vector fires through QEMU. As the
1566 * vector-use notifier is called, which occurs on unmask, we test whether
1567 * PBA emulation is needed and again disable if not.
1569 memory_region_set_enabled(&vdev
->pdev
.msix_pba_mmio
, false);
1572 * The emulated machine may provide a paravirt interface for MSIX setup
1573 * so it is not strictly necessary to emulate MSIX here. This becomes
1574 * helpful when frequently accessed MMIO registers are located in
1575 * subpages adjacent to the MSIX table but the MSIX data containing page
1576 * cannot be mapped because of a host page size bigger than the MSIX table
1579 if (object_property_get_bool(OBJECT(qdev_get_machine()),
1580 "vfio-no-msix-emulation", NULL
)) {
1581 memory_region_set_enabled(&vdev
->pdev
.msix_table_mmio
, false);
1587 static void vfio_teardown_msi(VFIOPCIDevice
*vdev
)
1589 msi_uninit(&vdev
->pdev
);
1592 msix_uninit(&vdev
->pdev
,
1593 vdev
->bars
[vdev
->msix
->table_bar
].mr
,
1594 vdev
->bars
[vdev
->msix
->pba_bar
].mr
);
1595 g_free(vdev
->msix
->pending
);
1602 static void vfio_mmap_set_enabled(VFIOPCIDevice
*vdev
, bool enabled
)
1606 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1607 vfio_region_mmaps_set_enabled(&vdev
->bars
[i
].region
, enabled
);
1611 static void vfio_bar_prepare(VFIOPCIDevice
*vdev
, int nr
)
1613 VFIOBAR
*bar
= &vdev
->bars
[nr
];
1618 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1619 if (!bar
->region
.size
) {
1623 /* Determine what type of BAR this is for registration */
1624 ret
= pread(vdev
->vbasedev
.fd
, &pci_bar
, sizeof(pci_bar
),
1625 vdev
->config_offset
+ PCI_BASE_ADDRESS_0
+ (4 * nr
));
1626 if (ret
!= sizeof(pci_bar
)) {
1627 error_report("vfio: Failed to read BAR %d (%m)", nr
);
1631 pci_bar
= le32_to_cpu(pci_bar
);
1632 bar
->ioport
= (pci_bar
& PCI_BASE_ADDRESS_SPACE_IO
);
1633 bar
->mem64
= bar
->ioport
? 0 : (pci_bar
& PCI_BASE_ADDRESS_MEM_TYPE_64
);
1634 bar
->type
= pci_bar
& (bar
->ioport
? ~PCI_BASE_ADDRESS_IO_MASK
:
1635 ~PCI_BASE_ADDRESS_MEM_MASK
);
1636 bar
->size
= bar
->region
.size
;
1639 static void vfio_bars_prepare(VFIOPCIDevice
*vdev
)
1643 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1644 vfio_bar_prepare(vdev
, i
);
1648 static void vfio_bar_register(VFIOPCIDevice
*vdev
, int nr
)
1650 VFIOBAR
*bar
= &vdev
->bars
[nr
];
1657 bar
->mr
= g_new0(MemoryRegion
, 1);
1658 name
= g_strdup_printf("%s base BAR %d", vdev
->vbasedev
.name
, nr
);
1659 memory_region_init_io(bar
->mr
, OBJECT(vdev
), NULL
, NULL
, name
, bar
->size
);
1662 if (bar
->region
.size
) {
1663 memory_region_add_subregion(bar
->mr
, 0, bar
->region
.mem
);
1665 if (vfio_region_mmap(&bar
->region
)) {
1666 error_report("Failed to mmap %s BAR %d. Performance may be slow",
1667 vdev
->vbasedev
.name
, nr
);
1671 pci_register_bar(&vdev
->pdev
, nr
, bar
->type
, bar
->mr
);
1674 static void vfio_bars_register(VFIOPCIDevice
*vdev
)
1678 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1679 vfio_bar_register(vdev
, i
);
1683 static void vfio_bars_exit(VFIOPCIDevice
*vdev
)
1687 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1688 VFIOBAR
*bar
= &vdev
->bars
[i
];
1690 vfio_bar_quirk_exit(vdev
, i
);
1691 vfio_region_exit(&bar
->region
);
1692 if (bar
->region
.size
) {
1693 memory_region_del_subregion(bar
->mr
, bar
->region
.mem
);
1698 pci_unregister_vga(&vdev
->pdev
);
1699 vfio_vga_quirk_exit(vdev
);
1703 static void vfio_bars_finalize(VFIOPCIDevice
*vdev
)
1707 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1708 VFIOBAR
*bar
= &vdev
->bars
[i
];
1710 vfio_bar_quirk_finalize(vdev
, i
);
1711 vfio_region_finalize(&bar
->region
);
1713 object_unparent(OBJECT(bar
->mr
));
1719 vfio_vga_quirk_finalize(vdev
);
1720 for (i
= 0; i
< ARRAY_SIZE(vdev
->vga
->region
); i
++) {
1721 object_unparent(OBJECT(&vdev
->vga
->region
[i
].mem
));
1730 static uint8_t vfio_std_cap_max_size(PCIDevice
*pdev
, uint8_t pos
)
1733 uint16_t next
= PCI_CONFIG_SPACE_SIZE
;
1735 for (tmp
= pdev
->config
[PCI_CAPABILITY_LIST
]; tmp
;
1736 tmp
= pdev
->config
[tmp
+ PCI_CAP_LIST_NEXT
]) {
1737 if (tmp
> pos
&& tmp
< next
) {
1746 static uint16_t vfio_ext_cap_max_size(const uint8_t *config
, uint16_t pos
)
1748 uint16_t tmp
, next
= PCIE_CONFIG_SPACE_SIZE
;
1750 for (tmp
= PCI_CONFIG_SPACE_SIZE
; tmp
;
1751 tmp
= PCI_EXT_CAP_NEXT(pci_get_long(config
+ tmp
))) {
1752 if (tmp
> pos
&& tmp
< next
) {
1760 static void vfio_set_word_bits(uint8_t *buf
, uint16_t val
, uint16_t mask
)
1762 pci_set_word(buf
, (pci_get_word(buf
) & ~mask
) | val
);
1765 static void vfio_add_emulated_word(VFIOPCIDevice
*vdev
, int pos
,
1766 uint16_t val
, uint16_t mask
)
1768 vfio_set_word_bits(vdev
->pdev
.config
+ pos
, val
, mask
);
1769 vfio_set_word_bits(vdev
->pdev
.wmask
+ pos
, ~mask
, mask
);
1770 vfio_set_word_bits(vdev
->emulated_config_bits
+ pos
, mask
, mask
);
1773 static void vfio_set_long_bits(uint8_t *buf
, uint32_t val
, uint32_t mask
)
1775 pci_set_long(buf
, (pci_get_long(buf
) & ~mask
) | val
);
1778 static void vfio_add_emulated_long(VFIOPCIDevice
*vdev
, int pos
,
1779 uint32_t val
, uint32_t mask
)
1781 vfio_set_long_bits(vdev
->pdev
.config
+ pos
, val
, mask
);
1782 vfio_set_long_bits(vdev
->pdev
.wmask
+ pos
, ~mask
, mask
);
1783 vfio_set_long_bits(vdev
->emulated_config_bits
+ pos
, mask
, mask
);
1786 static int vfio_setup_pcie_cap(VFIOPCIDevice
*vdev
, int pos
, uint8_t size
,
1792 flags
= pci_get_word(vdev
->pdev
.config
+ pos
+ PCI_CAP_FLAGS
);
1793 type
= (flags
& PCI_EXP_FLAGS_TYPE
) >> 4;
1795 if (type
!= PCI_EXP_TYPE_ENDPOINT
&&
1796 type
!= PCI_EXP_TYPE_LEG_END
&&
1797 type
!= PCI_EXP_TYPE_RC_END
) {
1799 error_setg(errp
, "assignment of PCIe type 0x%x "
1800 "devices is not currently supported", type
);
1804 if (!pci_bus_is_express(pci_get_bus(&vdev
->pdev
))) {
1805 PCIBus
*bus
= pci_get_bus(&vdev
->pdev
);
1809 * Traditionally PCI device assignment exposes the PCIe capability
1810 * as-is on non-express buses. The reason being that some drivers
1811 * simply assume that it's there, for example tg3. However when
1812 * we're running on a native PCIe machine type, like Q35, we need
1813 * to hide the PCIe capability. The reason for this is twofold;
1814 * first Windows guests get a Code 10 error when the PCIe capability
1815 * is exposed in this configuration. Therefore express devices won't
1816 * work at all unless they're attached to express buses in the VM.
1817 * Second, a native PCIe machine introduces the possibility of fine
1818 * granularity IOMMUs supporting both translation and isolation.
1819 * Guest code to discover the IOMMU visibility of a device, such as
1820 * IOMMU grouping code on Linux, is very aware of device types and
1821 * valid transitions between bus types. An express device on a non-
1822 * express bus is not a valid combination on bare metal systems.
1824 * Drivers that require a PCIe capability to make the device
1825 * functional are simply going to need to have their devices placed
1826 * on a PCIe bus in the VM.
1828 while (!pci_bus_is_root(bus
)) {
1829 bridge
= pci_bridge_get_device(bus
);
1830 bus
= pci_get_bus(bridge
);
1833 if (pci_bus_is_express(bus
)) {
1837 } else if (pci_bus_is_root(pci_get_bus(&vdev
->pdev
))) {
1839 * On a Root Complex bus Endpoints become Root Complex Integrated
1840 * Endpoints, which changes the type and clears the LNK & LNK2 fields.
1842 if (type
== PCI_EXP_TYPE_ENDPOINT
) {
1843 vfio_add_emulated_word(vdev
, pos
+ PCI_CAP_FLAGS
,
1844 PCI_EXP_TYPE_RC_END
<< 4,
1845 PCI_EXP_FLAGS_TYPE
);
1847 /* Link Capabilities, Status, and Control goes away */
1848 if (size
> PCI_EXP_LNKCTL
) {
1849 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP
, 0, ~0);
1850 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL
, 0, ~0);
1851 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKSTA
, 0, ~0);
1853 #ifndef PCI_EXP_LNKCAP2
1854 #define PCI_EXP_LNKCAP2 44
1856 #ifndef PCI_EXP_LNKSTA2
1857 #define PCI_EXP_LNKSTA2 50
1859 /* Link 2 Capabilities, Status, and Control goes away */
1860 if (size
> PCI_EXP_LNKCAP2
) {
1861 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP2
, 0, ~0);
1862 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL2
, 0, ~0);
1863 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKSTA2
, 0, ~0);
1867 } else if (type
== PCI_EXP_TYPE_LEG_END
) {
1869 * Legacy endpoints don't belong on the root complex. Windows
1870 * seems to be happier with devices if we skip the capability.
1877 * Convert Root Complex Integrated Endpoints to regular endpoints.
1878 * These devices don't support LNK/LNK2 capabilities, so make them up.
1880 if (type
== PCI_EXP_TYPE_RC_END
) {
1881 vfio_add_emulated_word(vdev
, pos
+ PCI_CAP_FLAGS
,
1882 PCI_EXP_TYPE_ENDPOINT
<< 4,
1883 PCI_EXP_FLAGS_TYPE
);
1884 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP
,
1885 QEMU_PCI_EXP_LNKCAP_MLW(QEMU_PCI_EXP_LNK_X1
) |
1886 QEMU_PCI_EXP_LNKCAP_MLS(QEMU_PCI_EXP_LNK_2_5GT
), ~0);
1887 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL
, 0, ~0);
1892 * Intel 82599 SR-IOV VFs report an invalid PCIe capability version 0
1893 * (Niantic errate #35) causing Windows to error with a Code 10 for the
1894 * device on Q35. Fixup any such devices to report version 1. If we
1895 * were to remove the capability entirely the guest would lose extended
1898 if ((flags
& PCI_EXP_FLAGS_VERS
) == 0) {
1899 vfio_add_emulated_word(vdev
, pos
+ PCI_CAP_FLAGS
,
1900 1, PCI_EXP_FLAGS_VERS
);
1903 pos
= pci_add_capability(&vdev
->pdev
, PCI_CAP_ID_EXP
, pos
, size
,
1909 vdev
->pdev
.exp
.exp_cap
= pos
;
1914 static void vfio_check_pcie_flr(VFIOPCIDevice
*vdev
, uint8_t pos
)
1916 uint32_t cap
= pci_get_long(vdev
->pdev
.config
+ pos
+ PCI_EXP_DEVCAP
);
1918 if (cap
& PCI_EXP_DEVCAP_FLR
) {
1919 trace_vfio_check_pcie_flr(vdev
->vbasedev
.name
);
1920 vdev
->has_flr
= true;
1924 static void vfio_check_pm_reset(VFIOPCIDevice
*vdev
, uint8_t pos
)
1926 uint16_t csr
= pci_get_word(vdev
->pdev
.config
+ pos
+ PCI_PM_CTRL
);
1928 if (!(csr
& PCI_PM_CTRL_NO_SOFT_RESET
)) {
1929 trace_vfio_check_pm_reset(vdev
->vbasedev
.name
);
1930 vdev
->has_pm_reset
= true;
1934 static void vfio_check_af_flr(VFIOPCIDevice
*vdev
, uint8_t pos
)
1936 uint8_t cap
= pci_get_byte(vdev
->pdev
.config
+ pos
+ PCI_AF_CAP
);
1938 if ((cap
& PCI_AF_CAP_TP
) && (cap
& PCI_AF_CAP_FLR
)) {
1939 trace_vfio_check_af_flr(vdev
->vbasedev
.name
);
1940 vdev
->has_flr
= true;
1944 static int vfio_add_std_cap(VFIOPCIDevice
*vdev
, uint8_t pos
, Error
**errp
)
1946 PCIDevice
*pdev
= &vdev
->pdev
;
1947 uint8_t cap_id
, next
, size
;
1950 cap_id
= pdev
->config
[pos
];
1951 next
= pdev
->config
[pos
+ PCI_CAP_LIST_NEXT
];
1954 * If it becomes important to configure capabilities to their actual
1955 * size, use this as the default when it's something we don't recognize.
1956 * Since QEMU doesn't actually handle many of the config accesses,
1957 * exact size doesn't seem worthwhile.
1959 size
= vfio_std_cap_max_size(pdev
, pos
);
1962 * pci_add_capability always inserts the new capability at the head
1963 * of the chain. Therefore to end up with a chain that matches the
1964 * physical device, we insert from the end by making this recursive.
1965 * This is also why we pre-calculate size above as cached config space
1966 * will be changed as we unwind the stack.
1969 ret
= vfio_add_std_cap(vdev
, next
, errp
);
1974 /* Begin the rebuild, use QEMU emulated list bits */
1975 pdev
->config
[PCI_CAPABILITY_LIST
] = 0;
1976 vdev
->emulated_config_bits
[PCI_CAPABILITY_LIST
] = 0xff;
1977 vdev
->emulated_config_bits
[PCI_STATUS
] |= PCI_STATUS_CAP_LIST
;
1979 ret
= vfio_add_virt_caps(vdev
, errp
);
1985 /* Scale down size, esp in case virt caps were added above */
1986 size
= MIN(size
, vfio_std_cap_max_size(pdev
, pos
));
1988 /* Use emulated next pointer to allow dropping caps */
1989 pci_set_byte(vdev
->emulated_config_bits
+ pos
+ PCI_CAP_LIST_NEXT
, 0xff);
1992 case PCI_CAP_ID_MSI
:
1993 ret
= vfio_msi_setup(vdev
, pos
, errp
);
1995 case PCI_CAP_ID_EXP
:
1996 vfio_check_pcie_flr(vdev
, pos
);
1997 ret
= vfio_setup_pcie_cap(vdev
, pos
, size
, errp
);
1999 case PCI_CAP_ID_MSIX
:
2000 ret
= vfio_msix_setup(vdev
, pos
, errp
);
2003 vfio_check_pm_reset(vdev
, pos
);
2005 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
, errp
);
2008 vfio_check_af_flr(vdev
, pos
);
2009 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
, errp
);
2012 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
, errp
);
2018 "failed to add PCI capability 0x%x[0x%x]@0x%x: ",
2026 static void vfio_add_ext_cap(VFIOPCIDevice
*vdev
)
2028 PCIDevice
*pdev
= &vdev
->pdev
;
2030 uint16_t cap_id
, next
, size
;
2034 /* Only add extended caps if we have them and the guest can see them */
2035 if (!pci_is_express(pdev
) || !pci_bus_is_express(pci_get_bus(pdev
)) ||
2036 !pci_get_long(pdev
->config
+ PCI_CONFIG_SPACE_SIZE
)) {
2041 * pcie_add_capability always inserts the new capability at the tail
2042 * of the chain. Therefore to end up with a chain that matches the
2043 * physical device, we cache the config space to avoid overwriting
2044 * the original config space when we parse the extended capabilities.
2046 config
= g_memdup(pdev
->config
, vdev
->config_size
);
2049 * Extended capabilities are chained with each pointing to the next, so we
2050 * can drop anything other than the head of the chain simply by modifying
2051 * the previous next pointer. Seed the head of the chain here such that
2052 * we can simply skip any capabilities we want to drop below, regardless
2053 * of their position in the chain. If this stub capability still exists
2054 * after we add the capabilities we want to expose, update the capability
2055 * ID to zero. Note that we cannot seed with the capability header being
2056 * zero as this conflicts with definition of an absent capability chain
2057 * and prevents capabilities beyond the head of the list from being added.
2058 * By replacing the dummy capability ID with zero after walking the device
2059 * chain, we also transparently mark extended capabilities as absent if
2060 * no capabilities were added. Note that the PCIe spec defines an absence
2061 * of extended capabilities to be determined by a value of zero for the
2062 * capability ID, version, AND next pointer. A non-zero next pointer
2063 * should be sufficient to indicate additional capabilities are present,
2064 * which will occur if we call pcie_add_capability() below. The entire
2065 * first dword is emulated to support this.
2067 * NB. The kernel side does similar masking, so be prepared that our
2068 * view of the device may also contain a capability ID zero in the head
2069 * of the chain. Skip it for the same reason that we cannot seed the
2070 * chain with a zero capability.
2072 pci_set_long(pdev
->config
+ PCI_CONFIG_SPACE_SIZE
,
2073 PCI_EXT_CAP(0xFFFF, 0, 0));
2074 pci_set_long(pdev
->wmask
+ PCI_CONFIG_SPACE_SIZE
, 0);
2075 pci_set_long(vdev
->emulated_config_bits
+ PCI_CONFIG_SPACE_SIZE
, ~0);
2077 for (next
= PCI_CONFIG_SPACE_SIZE
; next
;
2078 next
= PCI_EXT_CAP_NEXT(pci_get_long(config
+ next
))) {
2079 header
= pci_get_long(config
+ next
);
2080 cap_id
= PCI_EXT_CAP_ID(header
);
2081 cap_ver
= PCI_EXT_CAP_VER(header
);
2084 * If it becomes important to configure extended capabilities to their
2085 * actual size, use this as the default when it's something we don't
2086 * recognize. Since QEMU doesn't actually handle many of the config
2087 * accesses, exact size doesn't seem worthwhile.
2089 size
= vfio_ext_cap_max_size(config
, next
);
2091 /* Use emulated next pointer to allow dropping extended caps */
2092 pci_long_test_and_set_mask(vdev
->emulated_config_bits
+ next
,
2093 PCI_EXT_CAP_NEXT_MASK
);
2096 case 0: /* kernel masked capability */
2097 case PCI_EXT_CAP_ID_SRIOV
: /* Read-only VF BARs confuse OVMF */
2098 case PCI_EXT_CAP_ID_ARI
: /* XXX Needs next function virtualization */
2099 case PCI_EXT_CAP_ID_REBAR
: /* Can't expose read-only */
2100 trace_vfio_add_ext_cap_dropped(vdev
->vbasedev
.name
, cap_id
, next
);
2103 pcie_add_capability(pdev
, cap_id
, cap_ver
, next
, size
);
2108 /* Cleanup chain head ID if necessary */
2109 if (pci_get_word(pdev
->config
+ PCI_CONFIG_SPACE_SIZE
) == 0xFFFF) {
2110 pci_set_word(pdev
->config
+ PCI_CONFIG_SPACE_SIZE
, 0);
2117 static int vfio_add_capabilities(VFIOPCIDevice
*vdev
, Error
**errp
)
2119 PCIDevice
*pdev
= &vdev
->pdev
;
2122 if (!(pdev
->config
[PCI_STATUS
] & PCI_STATUS_CAP_LIST
) ||
2123 !pdev
->config
[PCI_CAPABILITY_LIST
]) {
2124 return 0; /* Nothing to add */
2127 ret
= vfio_add_std_cap(vdev
, pdev
->config
[PCI_CAPABILITY_LIST
], errp
);
2132 vfio_add_ext_cap(vdev
);
2136 static void vfio_pci_pre_reset(VFIOPCIDevice
*vdev
)
2138 PCIDevice
*pdev
= &vdev
->pdev
;
2141 vfio_disable_interrupts(vdev
);
2143 /* Make sure the device is in D0 */
2148 pmcsr
= vfio_pci_read_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, 2);
2149 state
= pmcsr
& PCI_PM_CTRL_STATE_MASK
;
2151 pmcsr
&= ~PCI_PM_CTRL_STATE_MASK
;
2152 vfio_pci_write_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, pmcsr
, 2);
2153 /* vfio handles the necessary delay here */
2154 pmcsr
= vfio_pci_read_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, 2);
2155 state
= pmcsr
& PCI_PM_CTRL_STATE_MASK
;
2157 error_report("vfio: Unable to power on device, stuck in D%d",
2164 * Stop any ongoing DMA by disconnecting I/O, MMIO, and bus master.
2165 * Also put INTx Disable in known state.
2167 cmd
= vfio_pci_read_config(pdev
, PCI_COMMAND
, 2);
2168 cmd
&= ~(PCI_COMMAND_IO
| PCI_COMMAND_MEMORY
| PCI_COMMAND_MASTER
|
2169 PCI_COMMAND_INTX_DISABLE
);
2170 vfio_pci_write_config(pdev
, PCI_COMMAND
, cmd
, 2);
2173 static void vfio_pci_post_reset(VFIOPCIDevice
*vdev
)
2178 vfio_intx_enable(vdev
, &err
);
2180 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2183 for (nr
= 0; nr
< PCI_NUM_REGIONS
- 1; ++nr
) {
2184 off_t addr
= vdev
->config_offset
+ PCI_BASE_ADDRESS_0
+ (4 * nr
);
2186 uint32_t len
= sizeof(val
);
2188 if (pwrite(vdev
->vbasedev
.fd
, &val
, len
, addr
) != len
) {
2189 error_report("%s(%s) reset bar %d failed: %m", __func__
,
2190 vdev
->vbasedev
.name
, nr
);
2194 vfio_quirk_reset(vdev
);
2197 static bool vfio_pci_host_match(PCIHostDeviceAddress
*addr
, const char *name
)
2201 sprintf(tmp
, "%04x:%02x:%02x.%1x", addr
->domain
,
2202 addr
->bus
, addr
->slot
, addr
->function
);
2204 return (strcmp(tmp
, name
) == 0);
2207 static int vfio_pci_hot_reset(VFIOPCIDevice
*vdev
, bool single
)
2210 struct vfio_pci_hot_reset_info
*info
;
2211 struct vfio_pci_dependent_device
*devices
;
2212 struct vfio_pci_hot_reset
*reset
;
2217 trace_vfio_pci_hot_reset(vdev
->vbasedev
.name
, single
? "one" : "multi");
2220 vfio_pci_pre_reset(vdev
);
2222 vdev
->vbasedev
.needs_reset
= false;
2224 info
= g_malloc0(sizeof(*info
));
2225 info
->argsz
= sizeof(*info
);
2227 ret
= ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO
, info
);
2228 if (ret
&& errno
!= ENOSPC
) {
2230 if (!vdev
->has_pm_reset
) {
2231 error_report("vfio: Cannot reset device %s, "
2232 "no available reset mechanism.", vdev
->vbasedev
.name
);
2237 count
= info
->count
;
2238 info
= g_realloc(info
, sizeof(*info
) + (count
* sizeof(*devices
)));
2239 info
->argsz
= sizeof(*info
) + (count
* sizeof(*devices
));
2240 devices
= &info
->devices
[0];
2242 ret
= ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO
, info
);
2245 error_report("vfio: hot reset info failed: %m");
2249 trace_vfio_pci_hot_reset_has_dep_devices(vdev
->vbasedev
.name
);
2251 /* Verify that we have all the groups required */
2252 for (i
= 0; i
< info
->count
; i
++) {
2253 PCIHostDeviceAddress host
;
2255 VFIODevice
*vbasedev_iter
;
2257 host
.domain
= devices
[i
].segment
;
2258 host
.bus
= devices
[i
].bus
;
2259 host
.slot
= PCI_SLOT(devices
[i
].devfn
);
2260 host
.function
= PCI_FUNC(devices
[i
].devfn
);
2262 trace_vfio_pci_hot_reset_dep_devices(host
.domain
,
2263 host
.bus
, host
.slot
, host
.function
, devices
[i
].group_id
);
2265 if (vfio_pci_host_match(&host
, vdev
->vbasedev
.name
)) {
2269 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
2270 if (group
->groupid
== devices
[i
].group_id
) {
2276 if (!vdev
->has_pm_reset
) {
2277 error_report("vfio: Cannot reset device %s, "
2278 "depends on group %d which is not owned.",
2279 vdev
->vbasedev
.name
, devices
[i
].group_id
);
2285 /* Prep dependent devices for reset and clear our marker. */
2286 QLIST_FOREACH(vbasedev_iter
, &group
->device_list
, next
) {
2287 if (!vbasedev_iter
->dev
->realized
||
2288 vbasedev_iter
->type
!= VFIO_DEVICE_TYPE_PCI
) {
2291 tmp
= container_of(vbasedev_iter
, VFIOPCIDevice
, vbasedev
);
2292 if (vfio_pci_host_match(&host
, tmp
->vbasedev
.name
)) {
2297 vfio_pci_pre_reset(tmp
);
2298 tmp
->vbasedev
.needs_reset
= false;
2305 if (!single
&& !multi
) {
2310 /* Determine how many group fds need to be passed */
2312 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
2313 for (i
= 0; i
< info
->count
; i
++) {
2314 if (group
->groupid
== devices
[i
].group_id
) {
2321 reset
= g_malloc0(sizeof(*reset
) + (count
* sizeof(*fds
)));
2322 reset
->argsz
= sizeof(*reset
) + (count
* sizeof(*fds
));
2323 fds
= &reset
->group_fds
[0];
2325 /* Fill in group fds */
2326 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
2327 for (i
= 0; i
< info
->count
; i
++) {
2328 if (group
->groupid
== devices
[i
].group_id
) {
2329 fds
[reset
->count
++] = group
->fd
;
2336 ret
= ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_PCI_HOT_RESET
, reset
);
2339 trace_vfio_pci_hot_reset_result(vdev
->vbasedev
.name
,
2340 ret
? "%m" : "Success");
2343 /* Re-enable INTx on affected devices */
2344 for (i
= 0; i
< info
->count
; i
++) {
2345 PCIHostDeviceAddress host
;
2347 VFIODevice
*vbasedev_iter
;
2349 host
.domain
= devices
[i
].segment
;
2350 host
.bus
= devices
[i
].bus
;
2351 host
.slot
= PCI_SLOT(devices
[i
].devfn
);
2352 host
.function
= PCI_FUNC(devices
[i
].devfn
);
2354 if (vfio_pci_host_match(&host
, vdev
->vbasedev
.name
)) {
2358 QLIST_FOREACH(group
, &vfio_group_list
, next
) {
2359 if (group
->groupid
== devices
[i
].group_id
) {
2368 QLIST_FOREACH(vbasedev_iter
, &group
->device_list
, next
) {
2369 if (!vbasedev_iter
->dev
->realized
||
2370 vbasedev_iter
->type
!= VFIO_DEVICE_TYPE_PCI
) {
2373 tmp
= container_of(vbasedev_iter
, VFIOPCIDevice
, vbasedev
);
2374 if (vfio_pci_host_match(&host
, tmp
->vbasedev
.name
)) {
2375 vfio_pci_post_reset(tmp
);
2382 vfio_pci_post_reset(vdev
);
2390 * We want to differentiate hot reset of multiple in-use devices vs hot reset
2391 * of a single in-use device. VFIO_DEVICE_RESET will already handle the case
2392 * of doing hot resets when there is only a single device per bus. The in-use
2393 * here refers to how many VFIODevices are affected. A hot reset that affects
2394 * multiple devices, but only a single in-use device, means that we can call
2395 * it from our bus ->reset() callback since the extent is effectively a single
2396 * device. This allows us to make use of it in the hotplug path. When there
2397 * are multiple in-use devices, we can only trigger the hot reset during a
2398 * system reset and thus from our reset handler. We separate _one vs _multi
2399 * here so that we don't overlap and do a double reset on the system reset
2400 * path where both our reset handler and ->reset() callback are used. Calling
2401 * _one() will only do a hot reset for the one in-use devices case, calling
2402 * _multi() will do nothing if a _one() would have been sufficient.
2404 static int vfio_pci_hot_reset_one(VFIOPCIDevice
*vdev
)
2406 return vfio_pci_hot_reset(vdev
, true);
2409 static int vfio_pci_hot_reset_multi(VFIODevice
*vbasedev
)
2411 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
2412 return vfio_pci_hot_reset(vdev
, false);
2415 static void vfio_pci_compute_needs_reset(VFIODevice
*vbasedev
)
2417 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
2418 if (!vbasedev
->reset_works
|| (!vdev
->has_flr
&& vdev
->has_pm_reset
)) {
2419 vbasedev
->needs_reset
= true;
2423 static Object
*vfio_pci_get_object(VFIODevice
*vbasedev
)
2425 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
2427 return OBJECT(vdev
);
2430 static bool vfio_msix_present(void *opaque
, int version_id
)
2432 PCIDevice
*pdev
= opaque
;
2434 return msix_present(pdev
);
2437 const VMStateDescription vmstate_vfio_pci_config
= {
2438 .name
= "VFIOPCIDevice",
2440 .minimum_version_id
= 1,
2441 .fields
= (VMStateField
[]) {
2442 VMSTATE_PCI_DEVICE(pdev
, VFIOPCIDevice
),
2443 VMSTATE_MSIX_TEST(pdev
, VFIOPCIDevice
, vfio_msix_present
),
2444 VMSTATE_END_OF_LIST()
2448 static void vfio_pci_save_config(VFIODevice
*vbasedev
, QEMUFile
*f
)
2450 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
2452 vmstate_save_state(f
, &vmstate_vfio_pci_config
, vdev
, NULL
);
2455 static int vfio_pci_load_config(VFIODevice
*vbasedev
, QEMUFile
*f
)
2457 VFIOPCIDevice
*vdev
= container_of(vbasedev
, VFIOPCIDevice
, vbasedev
);
2458 PCIDevice
*pdev
= &vdev
->pdev
;
2459 pcibus_t old_addr
[PCI_NUM_REGIONS
- 1];
2462 for (bar
= 0; bar
< PCI_ROM_SLOT
; bar
++) {
2463 old_addr
[bar
] = pdev
->io_regions
[bar
].addr
;
2466 ret
= vmstate_load_state(f
, &vmstate_vfio_pci_config
, vdev
, 1);
2471 vfio_pci_write_config(pdev
, PCI_COMMAND
,
2472 pci_get_word(pdev
->config
+ PCI_COMMAND
), 2);
2474 for (bar
= 0; bar
< PCI_ROM_SLOT
; bar
++) {
2476 * The address may not be changed in some scenarios
2477 * (e.g. the VF driver isn't loaded in VM).
2479 if (old_addr
[bar
] != pdev
->io_regions
[bar
].addr
&&
2480 vdev
->bars
[bar
].region
.size
> 0 &&
2481 vdev
->bars
[bar
].region
.size
< qemu_real_host_page_size()) {
2482 vfio_sub_page_bar_update_mapping(pdev
, bar
);
2486 if (msi_enabled(pdev
)) {
2487 vfio_msi_enable(vdev
);
2488 } else if (msix_enabled(pdev
)) {
2489 vfio_msix_enable(vdev
);
2495 static VFIODeviceOps vfio_pci_ops
= {
2496 .vfio_compute_needs_reset
= vfio_pci_compute_needs_reset
,
2497 .vfio_hot_reset_multi
= vfio_pci_hot_reset_multi
,
2498 .vfio_eoi
= vfio_intx_eoi
,
2499 .vfio_get_object
= vfio_pci_get_object
,
2500 .vfio_save_config
= vfio_pci_save_config
,
2501 .vfio_load_config
= vfio_pci_load_config
,
2504 int vfio_populate_vga(VFIOPCIDevice
*vdev
, Error
**errp
)
2506 VFIODevice
*vbasedev
= &vdev
->vbasedev
;
2507 struct vfio_region_info
*reg_info
;
2510 ret
= vfio_get_region_info(vbasedev
, VFIO_PCI_VGA_REGION_INDEX
, ®_info
);
2512 error_setg_errno(errp
, -ret
,
2513 "failed getting region info for VGA region index %d",
2514 VFIO_PCI_VGA_REGION_INDEX
);
2518 if (!(reg_info
->flags
& VFIO_REGION_INFO_FLAG_READ
) ||
2519 !(reg_info
->flags
& VFIO_REGION_INFO_FLAG_WRITE
) ||
2520 reg_info
->size
< 0xbffff + 1) {
2521 error_setg(errp
, "unexpected VGA info, flags 0x%lx, size 0x%lx",
2522 (unsigned long)reg_info
->flags
,
2523 (unsigned long)reg_info
->size
);
2528 vdev
->vga
= g_new0(VFIOVGA
, 1);
2530 vdev
->vga
->fd_offset
= reg_info
->offset
;
2531 vdev
->vga
->fd
= vdev
->vbasedev
.fd
;
2535 vdev
->vga
->region
[QEMU_PCI_VGA_MEM
].offset
= QEMU_PCI_VGA_MEM_BASE
;
2536 vdev
->vga
->region
[QEMU_PCI_VGA_MEM
].nr
= QEMU_PCI_VGA_MEM
;
2537 QLIST_INIT(&vdev
->vga
->region
[QEMU_PCI_VGA_MEM
].quirks
);
2539 memory_region_init_io(&vdev
->vga
->region
[QEMU_PCI_VGA_MEM
].mem
,
2540 OBJECT(vdev
), &vfio_vga_ops
,
2541 &vdev
->vga
->region
[QEMU_PCI_VGA_MEM
],
2542 "vfio-vga-mmio@0xa0000",
2543 QEMU_PCI_VGA_MEM_SIZE
);
2545 vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
].offset
= QEMU_PCI_VGA_IO_LO_BASE
;
2546 vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
].nr
= QEMU_PCI_VGA_IO_LO
;
2547 QLIST_INIT(&vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
].quirks
);
2549 memory_region_init_io(&vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
].mem
,
2550 OBJECT(vdev
), &vfio_vga_ops
,
2551 &vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
],
2552 "vfio-vga-io@0x3b0",
2553 QEMU_PCI_VGA_IO_LO_SIZE
);
2555 vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
].offset
= QEMU_PCI_VGA_IO_HI_BASE
;
2556 vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
].nr
= QEMU_PCI_VGA_IO_HI
;
2557 QLIST_INIT(&vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
].quirks
);
2559 memory_region_init_io(&vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
].mem
,
2560 OBJECT(vdev
), &vfio_vga_ops
,
2561 &vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
],
2562 "vfio-vga-io@0x3c0",
2563 QEMU_PCI_VGA_IO_HI_SIZE
);
2565 pci_register_vga(&vdev
->pdev
, &vdev
->vga
->region
[QEMU_PCI_VGA_MEM
].mem
,
2566 &vdev
->vga
->region
[QEMU_PCI_VGA_IO_LO
].mem
,
2567 &vdev
->vga
->region
[QEMU_PCI_VGA_IO_HI
].mem
);
2572 static void vfio_populate_device(VFIOPCIDevice
*vdev
, Error
**errp
)
2574 VFIODevice
*vbasedev
= &vdev
->vbasedev
;
2575 struct vfio_region_info
*reg_info
;
2576 struct vfio_irq_info irq_info
= { .argsz
= sizeof(irq_info
) };
2579 /* Sanity check device */
2580 if (!(vbasedev
->flags
& VFIO_DEVICE_FLAGS_PCI
)) {
2581 error_setg(errp
, "this isn't a PCI device");
2585 if (vbasedev
->num_regions
< VFIO_PCI_CONFIG_REGION_INDEX
+ 1) {
2586 error_setg(errp
, "unexpected number of io regions %u",
2587 vbasedev
->num_regions
);
2591 if (vbasedev
->num_irqs
< VFIO_PCI_MSIX_IRQ_INDEX
+ 1) {
2592 error_setg(errp
, "unexpected number of irqs %u", vbasedev
->num_irqs
);
2596 for (i
= VFIO_PCI_BAR0_REGION_INDEX
; i
< VFIO_PCI_ROM_REGION_INDEX
; i
++) {
2597 char *name
= g_strdup_printf("%s BAR %d", vbasedev
->name
, i
);
2599 ret
= vfio_region_setup(OBJECT(vdev
), vbasedev
,
2600 &vdev
->bars
[i
].region
, i
, name
);
2604 error_setg_errno(errp
, -ret
, "failed to get region %d info", i
);
2608 QLIST_INIT(&vdev
->bars
[i
].quirks
);
2611 ret
= vfio_get_region_info(vbasedev
,
2612 VFIO_PCI_CONFIG_REGION_INDEX
, ®_info
);
2614 error_setg_errno(errp
, -ret
, "failed to get config info");
2618 trace_vfio_populate_device_config(vdev
->vbasedev
.name
,
2619 (unsigned long)reg_info
->size
,
2620 (unsigned long)reg_info
->offset
,
2621 (unsigned long)reg_info
->flags
);
2623 vdev
->config_size
= reg_info
->size
;
2624 if (vdev
->config_size
== PCI_CONFIG_SPACE_SIZE
) {
2625 vdev
->pdev
.cap_present
&= ~QEMU_PCI_CAP_EXPRESS
;
2627 vdev
->config_offset
= reg_info
->offset
;
2631 if (vdev
->features
& VFIO_FEATURE_ENABLE_VGA
) {
2632 ret
= vfio_populate_vga(vdev
, errp
);
2634 error_append_hint(errp
, "device does not support "
2635 "requested feature x-vga\n");
2640 irq_info
.index
= VFIO_PCI_ERR_IRQ_INDEX
;
2642 ret
= ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_GET_IRQ_INFO
, &irq_info
);
2644 /* This can fail for an old kernel or legacy PCI dev */
2645 trace_vfio_populate_device_get_irq_info_failure(strerror(errno
));
2646 } else if (irq_info
.count
== 1) {
2647 vdev
->pci_aer
= true;
2649 warn_report(VFIO_MSG_PREFIX
2650 "Could not enable error recovery for the device",
2655 static void vfio_put_device(VFIOPCIDevice
*vdev
)
2657 g_free(vdev
->vbasedev
.name
);
2660 vfio_put_base_device(&vdev
->vbasedev
);
2663 static void vfio_err_notifier_handler(void *opaque
)
2665 VFIOPCIDevice
*vdev
= opaque
;
2667 if (!event_notifier_test_and_clear(&vdev
->err_notifier
)) {
2672 * TBD. Retrieve the error details and decide what action
2673 * needs to be taken. One of the actions could be to pass
2674 * the error to the guest and have the guest driver recover
2675 * from the error. This requires that PCIe capabilities be
2676 * exposed to the guest. For now, we just terminate the
2677 * guest to contain the error.
2680 error_report("%s(%s) Unrecoverable error detected. Please collect any data possible and then kill the guest", __func__
, vdev
->vbasedev
.name
);
2682 vm_stop(RUN_STATE_INTERNAL_ERROR
);
2686 * Registers error notifier for devices supporting error recovery.
2687 * If we encounter a failure in this function, we report an error
2688 * and continue after disabling error recovery support for the
2691 static void vfio_register_err_notifier(VFIOPCIDevice
*vdev
)
2696 if (!vdev
->pci_aer
) {
2700 if (event_notifier_init(&vdev
->err_notifier
, 0)) {
2701 error_report("vfio: Unable to init event notifier for error detection");
2702 vdev
->pci_aer
= false;
2706 fd
= event_notifier_get_fd(&vdev
->err_notifier
);
2707 qemu_set_fd_handler(fd
, vfio_err_notifier_handler
, NULL
, vdev
);
2709 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_ERR_IRQ_INDEX
, 0,
2710 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, &err
)) {
2711 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2712 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
2713 event_notifier_cleanup(&vdev
->err_notifier
);
2714 vdev
->pci_aer
= false;
2718 static void vfio_unregister_err_notifier(VFIOPCIDevice
*vdev
)
2722 if (!vdev
->pci_aer
) {
2726 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_ERR_IRQ_INDEX
, 0,
2727 VFIO_IRQ_SET_ACTION_TRIGGER
, -1, &err
)) {
2728 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2730 qemu_set_fd_handler(event_notifier_get_fd(&vdev
->err_notifier
),
2732 event_notifier_cleanup(&vdev
->err_notifier
);
2735 static void vfio_req_notifier_handler(void *opaque
)
2737 VFIOPCIDevice
*vdev
= opaque
;
2740 if (!event_notifier_test_and_clear(&vdev
->req_notifier
)) {
2744 qdev_unplug(DEVICE(vdev
), &err
);
2746 warn_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2750 static void vfio_register_req_notifier(VFIOPCIDevice
*vdev
)
2752 struct vfio_irq_info irq_info
= { .argsz
= sizeof(irq_info
),
2753 .index
= VFIO_PCI_REQ_IRQ_INDEX
};
2757 if (!(vdev
->features
& VFIO_FEATURE_ENABLE_REQ
)) {
2761 if (ioctl(vdev
->vbasedev
.fd
,
2762 VFIO_DEVICE_GET_IRQ_INFO
, &irq_info
) < 0 || irq_info
.count
< 1) {
2766 if (event_notifier_init(&vdev
->req_notifier
, 0)) {
2767 error_report("vfio: Unable to init event notifier for device request");
2771 fd
= event_notifier_get_fd(&vdev
->req_notifier
);
2772 qemu_set_fd_handler(fd
, vfio_req_notifier_handler
, NULL
, vdev
);
2774 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_REQ_IRQ_INDEX
, 0,
2775 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, &err
)) {
2776 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2777 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
2778 event_notifier_cleanup(&vdev
->req_notifier
);
2780 vdev
->req_enabled
= true;
2784 static void vfio_unregister_req_notifier(VFIOPCIDevice
*vdev
)
2788 if (!vdev
->req_enabled
) {
2792 if (vfio_set_irq_signaling(&vdev
->vbasedev
, VFIO_PCI_REQ_IRQ_INDEX
, 0,
2793 VFIO_IRQ_SET_ACTION_TRIGGER
, -1, &err
)) {
2794 error_reportf_err(err
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
2796 qemu_set_fd_handler(event_notifier_get_fd(&vdev
->req_notifier
),
2798 event_notifier_cleanup(&vdev
->req_notifier
);
2800 vdev
->req_enabled
= false;
2803 static void vfio_realize(PCIDevice
*pdev
, Error
**errp
)
2805 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
2806 VFIODevice
*vbasedev_iter
;
2808 char *tmp
, *subsys
, group_path
[PATH_MAX
], *group_name
;
2816 if (!vdev
->vbasedev
.sysfsdev
) {
2817 if (!(~vdev
->host
.domain
|| ~vdev
->host
.bus
||
2818 ~vdev
->host
.slot
|| ~vdev
->host
.function
)) {
2819 error_setg(errp
, "No provided host device");
2820 error_append_hint(errp
, "Use -device vfio-pci,host=DDDD:BB:DD.F "
2821 "or -device vfio-pci,sysfsdev=PATH_TO_DEVICE\n");
2824 vdev
->vbasedev
.sysfsdev
=
2825 g_strdup_printf("/sys/bus/pci/devices/%04x:%02x:%02x.%01x",
2826 vdev
->host
.domain
, vdev
->host
.bus
,
2827 vdev
->host
.slot
, vdev
->host
.function
);
2830 if (stat(vdev
->vbasedev
.sysfsdev
, &st
) < 0) {
2831 error_setg_errno(errp
, errno
, "no such host device");
2832 error_prepend(errp
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.sysfsdev
);
2836 vdev
->vbasedev
.name
= g_path_get_basename(vdev
->vbasedev
.sysfsdev
);
2837 vdev
->vbasedev
.ops
= &vfio_pci_ops
;
2838 vdev
->vbasedev
.type
= VFIO_DEVICE_TYPE_PCI
;
2839 vdev
->vbasedev
.dev
= DEVICE(vdev
);
2841 tmp
= g_strdup_printf("%s/iommu_group", vdev
->vbasedev
.sysfsdev
);
2842 len
= readlink(tmp
, group_path
, sizeof(group_path
));
2845 if (len
<= 0 || len
>= sizeof(group_path
)) {
2846 error_setg_errno(errp
, len
< 0 ? errno
: ENAMETOOLONG
,
2847 "no iommu_group found");
2851 group_path
[len
] = 0;
2853 group_name
= basename(group_path
);
2854 if (sscanf(group_name
, "%d", &groupid
) != 1) {
2855 error_setg_errno(errp
, errno
, "failed to read %s", group_path
);
2859 trace_vfio_realize(vdev
->vbasedev
.name
, groupid
);
2861 group
= vfio_get_group(groupid
, pci_device_iommu_address_space(pdev
), errp
);
2866 QLIST_FOREACH(vbasedev_iter
, &group
->device_list
, next
) {
2867 if (strcmp(vbasedev_iter
->name
, vdev
->vbasedev
.name
) == 0) {
2868 error_setg(errp
, "device is already attached");
2869 vfio_put_group(group
);
2875 * Mediated devices *might* operate compatibly with discarding of RAM, but
2876 * we cannot know for certain, it depends on whether the mdev vendor driver
2877 * stays in sync with the active working set of the guest driver. Prevent
2878 * the x-balloon-allowed option unless this is minimally an mdev device.
2880 tmp
= g_strdup_printf("%s/subsystem", vdev
->vbasedev
.sysfsdev
);
2881 subsys
= realpath(tmp
, NULL
);
2883 is_mdev
= subsys
&& (strcmp(subsys
, "/sys/bus/mdev") == 0);
2886 trace_vfio_mdev(vdev
->vbasedev
.name
, is_mdev
);
2888 if (vdev
->vbasedev
.ram_block_discard_allowed
&& !is_mdev
) {
2889 error_setg(errp
, "x-balloon-allowed only potentially compatible "
2890 "with mdev devices");
2891 vfio_put_group(group
);
2895 ret
= vfio_get_device(group
, vdev
->vbasedev
.name
, &vdev
->vbasedev
, errp
);
2897 vfio_put_group(group
);
2901 vfio_populate_device(vdev
, &err
);
2903 error_propagate(errp
, err
);
2907 /* Get a copy of config space */
2908 ret
= pread(vdev
->vbasedev
.fd
, vdev
->pdev
.config
,
2909 MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
),
2910 vdev
->config_offset
);
2911 if (ret
< (int)MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
)) {
2912 ret
= ret
< 0 ? -errno
: -EFAULT
;
2913 error_setg_errno(errp
, -ret
, "failed to read device config space");
2917 /* vfio emulates a lot for us, but some bits need extra love */
2918 vdev
->emulated_config_bits
= g_malloc0(vdev
->config_size
);
2920 /* QEMU can choose to expose the ROM or not */
2921 memset(vdev
->emulated_config_bits
+ PCI_ROM_ADDRESS
, 0xff, 4);
2922 /* QEMU can also add or extend BARs */
2923 memset(vdev
->emulated_config_bits
+ PCI_BASE_ADDRESS_0
, 0xff, 6 * 4);
2926 * The PCI spec reserves vendor ID 0xffff as an invalid value. The
2927 * device ID is managed by the vendor and need only be a 16-bit value.
2928 * Allow any 16-bit value for subsystem so they can be hidden or changed.
2930 if (vdev
->vendor_id
!= PCI_ANY_ID
) {
2931 if (vdev
->vendor_id
>= 0xffff) {
2932 error_setg(errp
, "invalid PCI vendor ID provided");
2935 vfio_add_emulated_word(vdev
, PCI_VENDOR_ID
, vdev
->vendor_id
, ~0);
2936 trace_vfio_pci_emulated_vendor_id(vdev
->vbasedev
.name
, vdev
->vendor_id
);
2938 vdev
->vendor_id
= pci_get_word(pdev
->config
+ PCI_VENDOR_ID
);
2941 if (vdev
->device_id
!= PCI_ANY_ID
) {
2942 if (vdev
->device_id
> 0xffff) {
2943 error_setg(errp
, "invalid PCI device ID provided");
2946 vfio_add_emulated_word(vdev
, PCI_DEVICE_ID
, vdev
->device_id
, ~0);
2947 trace_vfio_pci_emulated_device_id(vdev
->vbasedev
.name
, vdev
->device_id
);
2949 vdev
->device_id
= pci_get_word(pdev
->config
+ PCI_DEVICE_ID
);
2952 if (vdev
->sub_vendor_id
!= PCI_ANY_ID
) {
2953 if (vdev
->sub_vendor_id
> 0xffff) {
2954 error_setg(errp
, "invalid PCI subsystem vendor ID provided");
2957 vfio_add_emulated_word(vdev
, PCI_SUBSYSTEM_VENDOR_ID
,
2958 vdev
->sub_vendor_id
, ~0);
2959 trace_vfio_pci_emulated_sub_vendor_id(vdev
->vbasedev
.name
,
2960 vdev
->sub_vendor_id
);
2963 if (vdev
->sub_device_id
!= PCI_ANY_ID
) {
2964 if (vdev
->sub_device_id
> 0xffff) {
2965 error_setg(errp
, "invalid PCI subsystem device ID provided");
2968 vfio_add_emulated_word(vdev
, PCI_SUBSYSTEM_ID
, vdev
->sub_device_id
, ~0);
2969 trace_vfio_pci_emulated_sub_device_id(vdev
->vbasedev
.name
,
2970 vdev
->sub_device_id
);
2973 /* QEMU can change multi-function devices to single function, or reverse */
2974 vdev
->emulated_config_bits
[PCI_HEADER_TYPE
] =
2975 PCI_HEADER_TYPE_MULTI_FUNCTION
;
2977 /* Restore or clear multifunction, this is always controlled by QEMU */
2978 if (vdev
->pdev
.cap_present
& QEMU_PCI_CAP_MULTIFUNCTION
) {
2979 vdev
->pdev
.config
[PCI_HEADER_TYPE
] |= PCI_HEADER_TYPE_MULTI_FUNCTION
;
2981 vdev
->pdev
.config
[PCI_HEADER_TYPE
] &= ~PCI_HEADER_TYPE_MULTI_FUNCTION
;
2985 * Clear host resource mapping info. If we choose not to register a
2986 * BAR, such as might be the case with the option ROM, we can get
2987 * confusing, unwritable, residual addresses from the host here.
2989 memset(&vdev
->pdev
.config
[PCI_BASE_ADDRESS_0
], 0, 24);
2990 memset(&vdev
->pdev
.config
[PCI_ROM_ADDRESS
], 0, 4);
2992 vfio_pci_size_rom(vdev
);
2994 vfio_bars_prepare(vdev
);
2996 vfio_msix_early_setup(vdev
, &err
);
2998 error_propagate(errp
, err
);
3002 vfio_bars_register(vdev
);
3004 ret
= vfio_add_capabilities(vdev
, errp
);
3010 vfio_vga_quirk_setup(vdev
);
3013 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
3014 vfio_bar_quirk_setup(vdev
, i
);
3017 if (!vdev
->igd_opregion
&&
3018 vdev
->features
& VFIO_FEATURE_ENABLE_IGD_OPREGION
) {
3019 struct vfio_region_info
*opregion
;
3021 if (vdev
->pdev
.qdev
.hotplugged
) {
3023 "cannot support IGD OpRegion feature on hotplugged "
3028 ret
= vfio_get_dev_region_info(&vdev
->vbasedev
,
3029 VFIO_REGION_TYPE_PCI_VENDOR_TYPE
| PCI_VENDOR_ID_INTEL
,
3030 VFIO_REGION_SUBTYPE_INTEL_IGD_OPREGION
, &opregion
);
3032 error_setg_errno(errp
, -ret
,
3033 "does not support requested IGD OpRegion feature");
3037 ret
= vfio_pci_igd_opregion_init(vdev
, opregion
, errp
);
3044 /* QEMU emulates all of MSI & MSIX */
3045 if (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
) {
3046 memset(vdev
->emulated_config_bits
+ pdev
->msix_cap
, 0xff,
3050 if (pdev
->cap_present
& QEMU_PCI_CAP_MSI
) {
3051 memset(vdev
->emulated_config_bits
+ pdev
->msi_cap
, 0xff,
3052 vdev
->msi_cap_size
);
3055 if (vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1)) {
3056 vdev
->intx
.mmap_timer
= timer_new_ms(QEMU_CLOCK_VIRTUAL
,
3057 vfio_intx_mmap_enable
, vdev
);
3058 pci_device_set_intx_routing_notifier(&vdev
->pdev
,
3059 vfio_intx_routing_notifier
);
3060 vdev
->irqchip_change_notifier
.notify
= vfio_irqchip_change
;
3061 kvm_irqchip_add_change_notifier(&vdev
->irqchip_change_notifier
);
3062 ret
= vfio_intx_enable(vdev
, errp
);
3064 goto out_deregister
;
3068 if (vdev
->display
!= ON_OFF_AUTO_OFF
) {
3069 ret
= vfio_display_probe(vdev
, errp
);
3071 goto out_deregister
;
3074 if (vdev
->enable_ramfb
&& vdev
->dpy
== NULL
) {
3075 error_setg(errp
, "ramfb=on requires display=on");
3076 goto out_deregister
;
3078 if (vdev
->display_xres
|| vdev
->display_yres
) {
3079 if (vdev
->dpy
== NULL
) {
3080 error_setg(errp
, "xres and yres properties require display=on");
3081 goto out_deregister
;
3083 if (vdev
->dpy
->edid_regs
== NULL
) {
3084 error_setg(errp
, "xres and yres properties need edid support");
3085 goto out_deregister
;
3089 if (vfio_pci_is(vdev
, PCI_VENDOR_ID_NVIDIA
, PCI_ANY_ID
)) {
3090 ret
= vfio_pci_nvidia_v100_ram_init(vdev
, errp
);
3091 if (ret
&& ret
!= -ENODEV
) {
3092 error_report("Failed to setup NVIDIA V100 GPU RAM");
3096 if (vfio_pci_is(vdev
, PCI_VENDOR_ID_IBM
, PCI_ANY_ID
)) {
3097 ret
= vfio_pci_nvlink2_init(vdev
, errp
);
3098 if (ret
&& ret
!= -ENODEV
) {
3099 error_report("Failed to setup NVlink2 bridge");
3103 if (!pdev
->failover_pair_id
) {
3104 ret
= vfio_migration_probe(&vdev
->vbasedev
, errp
);
3106 error_report("%s: Migration disabled", vdev
->vbasedev
.name
);
3110 vfio_register_err_notifier(vdev
);
3111 vfio_register_req_notifier(vdev
);
3112 vfio_setup_resetfn_quirk(vdev
);
3117 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
3118 kvm_irqchip_remove_change_notifier(&vdev
->irqchip_change_notifier
);
3120 vfio_teardown_msi(vdev
);
3121 vfio_bars_exit(vdev
);
3123 error_prepend(errp
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
3126 static void vfio_instance_finalize(Object
*obj
)
3128 VFIOPCIDevice
*vdev
= VFIO_PCI(obj
);
3129 VFIOGroup
*group
= vdev
->vbasedev
.group
;
3131 vfio_display_finalize(vdev
);
3132 vfio_bars_finalize(vdev
);
3133 g_free(vdev
->emulated_config_bits
);
3136 * XXX Leaking igd_opregion is not an oversight, we can't remove the
3137 * fw_cfg entry therefore leaking this allocation seems like the safest
3140 * g_free(vdev->igd_opregion);
3142 vfio_put_device(vdev
);
3143 vfio_put_group(group
);
3146 static void vfio_exitfn(PCIDevice
*pdev
)
3148 VFIOPCIDevice
*vdev
= VFIO_PCI(pdev
);
3150 vfio_unregister_req_notifier(vdev
);
3151 vfio_unregister_err_notifier(vdev
);
3152 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
3153 if (vdev
->irqchip_change_notifier
.notify
) {
3154 kvm_irqchip_remove_change_notifier(&vdev
->irqchip_change_notifier
);
3156 vfio_disable_interrupts(vdev
);
3157 if (vdev
->intx
.mmap_timer
) {
3158 timer_free(vdev
->intx
.mmap_timer
);
3160 vfio_teardown_msi(vdev
);
3161 vfio_bars_exit(vdev
);
3162 vfio_migration_finalize(&vdev
->vbasedev
);
3165 static void vfio_pci_reset(DeviceState
*dev
)
3167 VFIOPCIDevice
*vdev
= VFIO_PCI(dev
);
3169 trace_vfio_pci_reset(vdev
->vbasedev
.name
);
3171 vfio_pci_pre_reset(vdev
);
3173 if (vdev
->display
!= ON_OFF_AUTO_OFF
) {
3174 vfio_display_reset(vdev
);
3177 if (vdev
->resetfn
&& !vdev
->resetfn(vdev
)) {
3181 if (vdev
->vbasedev
.reset_works
&&
3182 (vdev
->has_flr
|| !vdev
->has_pm_reset
) &&
3183 !ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_RESET
)) {
3184 trace_vfio_pci_reset_flr(vdev
->vbasedev
.name
);
3188 /* See if we can do our own bus reset */
3189 if (!vfio_pci_hot_reset_one(vdev
)) {
3193 /* If nothing else works and the device supports PM reset, use it */
3194 if (vdev
->vbasedev
.reset_works
&& vdev
->has_pm_reset
&&
3195 !ioctl(vdev
->vbasedev
.fd
, VFIO_DEVICE_RESET
)) {
3196 trace_vfio_pci_reset_pm(vdev
->vbasedev
.name
);
3201 vfio_pci_post_reset(vdev
);
3204 static void vfio_instance_init(Object
*obj
)
3206 PCIDevice
*pci_dev
= PCI_DEVICE(obj
);
3207 VFIOPCIDevice
*vdev
= VFIO_PCI(obj
);
3209 device_add_bootindex_property(obj
, &vdev
->bootindex
,
3212 vdev
->host
.domain
= ~0U;
3213 vdev
->host
.bus
= ~0U;
3214 vdev
->host
.slot
= ~0U;
3215 vdev
->host
.function
= ~0U;
3217 vdev
->nv_gpudirect_clique
= 0xFF;
3219 /* QEMU_PCI_CAP_EXPRESS initialization does not depend on QEMU command
3220 * line, therefore, no need to wait to realize like other devices */
3221 pci_dev
->cap_present
|= QEMU_PCI_CAP_EXPRESS
;
3224 static Property vfio_pci_dev_properties
[] = {
3225 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIOPCIDevice
, host
),
3226 DEFINE_PROP_STRING("sysfsdev", VFIOPCIDevice
, vbasedev
.sysfsdev
),
3227 DEFINE_PROP_ON_OFF_AUTO("x-pre-copy-dirty-page-tracking", VFIOPCIDevice
,
3228 vbasedev
.pre_copy_dirty_page_tracking
,
3230 DEFINE_PROP_ON_OFF_AUTO("display", VFIOPCIDevice
,
3231 display
, ON_OFF_AUTO_OFF
),
3232 DEFINE_PROP_UINT32("xres", VFIOPCIDevice
, display_xres
, 0),
3233 DEFINE_PROP_UINT32("yres", VFIOPCIDevice
, display_yres
, 0),
3234 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIOPCIDevice
,
3235 intx
.mmap_timeout
, 1100),
3236 DEFINE_PROP_BIT("x-vga", VFIOPCIDevice
, features
,
3237 VFIO_FEATURE_ENABLE_VGA_BIT
, false),
3238 DEFINE_PROP_BIT("x-req", VFIOPCIDevice
, features
,
3239 VFIO_FEATURE_ENABLE_REQ_BIT
, true),
3240 DEFINE_PROP_BIT("x-igd-opregion", VFIOPCIDevice
, features
,
3241 VFIO_FEATURE_ENABLE_IGD_OPREGION_BIT
, false),
3242 DEFINE_PROP_BOOL("x-enable-migration", VFIOPCIDevice
,
3243 vbasedev
.enable_migration
, false),
3244 DEFINE_PROP_BOOL("x-no-mmap", VFIOPCIDevice
, vbasedev
.no_mmap
, false),
3245 DEFINE_PROP_BOOL("x-balloon-allowed", VFIOPCIDevice
,
3246 vbasedev
.ram_block_discard_allowed
, false),
3247 DEFINE_PROP_BOOL("x-no-kvm-intx", VFIOPCIDevice
, no_kvm_intx
, false),
3248 DEFINE_PROP_BOOL("x-no-kvm-msi", VFIOPCIDevice
, no_kvm_msi
, false),
3249 DEFINE_PROP_BOOL("x-no-kvm-msix", VFIOPCIDevice
, no_kvm_msix
, false),
3250 DEFINE_PROP_BOOL("x-no-geforce-quirks", VFIOPCIDevice
,
3251 no_geforce_quirks
, false),
3252 DEFINE_PROP_BOOL("x-no-kvm-ioeventfd", VFIOPCIDevice
, no_kvm_ioeventfd
,
3254 DEFINE_PROP_BOOL("x-no-vfio-ioeventfd", VFIOPCIDevice
, no_vfio_ioeventfd
,
3256 DEFINE_PROP_UINT32("x-pci-vendor-id", VFIOPCIDevice
, vendor_id
, PCI_ANY_ID
),
3257 DEFINE_PROP_UINT32("x-pci-device-id", VFIOPCIDevice
, device_id
, PCI_ANY_ID
),
3258 DEFINE_PROP_UINT32("x-pci-sub-vendor-id", VFIOPCIDevice
,
3259 sub_vendor_id
, PCI_ANY_ID
),
3260 DEFINE_PROP_UINT32("x-pci-sub-device-id", VFIOPCIDevice
,
3261 sub_device_id
, PCI_ANY_ID
),
3262 DEFINE_PROP_UINT32("x-igd-gms", VFIOPCIDevice
, igd_gms
, 0),
3263 DEFINE_PROP_UNSIGNED_NODEFAULT("x-nv-gpudirect-clique", VFIOPCIDevice
,
3264 nv_gpudirect_clique
,
3265 qdev_prop_nv_gpudirect_clique
, uint8_t),
3266 DEFINE_PROP_OFF_AUTO_PCIBAR("x-msix-relocation", VFIOPCIDevice
, msix_relo
,
3267 OFF_AUTOPCIBAR_OFF
),
3269 * TODO - support passed fds... is this necessary?
3270 * DEFINE_PROP_STRING("vfiofd", VFIOPCIDevice, vfiofd_name),
3271 * DEFINE_PROP_STRING("vfiogroupfd, VFIOPCIDevice, vfiogroupfd_name),
3273 DEFINE_PROP_END_OF_LIST(),
3276 static void vfio_pci_dev_class_init(ObjectClass
*klass
, void *data
)
3278 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3279 PCIDeviceClass
*pdc
= PCI_DEVICE_CLASS(klass
);
3281 dc
->reset
= vfio_pci_reset
;
3282 device_class_set_props(dc
, vfio_pci_dev_properties
);
3283 dc
->desc
= "VFIO-based PCI device assignment";
3284 set_bit(DEVICE_CATEGORY_MISC
, dc
->categories
);
3285 pdc
->realize
= vfio_realize
;
3286 pdc
->exit
= vfio_exitfn
;
3287 pdc
->config_read
= vfio_pci_read_config
;
3288 pdc
->config_write
= vfio_pci_write_config
;
3291 static const TypeInfo vfio_pci_dev_info
= {
3292 .name
= TYPE_VFIO_PCI
,
3293 .parent
= TYPE_PCI_DEVICE
,
3294 .instance_size
= sizeof(VFIOPCIDevice
),
3295 .class_init
= vfio_pci_dev_class_init
,
3296 .instance_init
= vfio_instance_init
,
3297 .instance_finalize
= vfio_instance_finalize
,
3298 .interfaces
= (InterfaceInfo
[]) {
3299 { INTERFACE_PCIE_DEVICE
},
3300 { INTERFACE_CONVENTIONAL_PCI_DEVICE
},
3305 static Property vfio_pci_dev_nohotplug_properties
[] = {
3306 DEFINE_PROP_BOOL("ramfb", VFIOPCIDevice
, enable_ramfb
, false),
3307 DEFINE_PROP_END_OF_LIST(),
3310 static void vfio_pci_nohotplug_dev_class_init(ObjectClass
*klass
, void *data
)
3312 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3314 device_class_set_props(dc
, vfio_pci_dev_nohotplug_properties
);
3315 dc
->hotpluggable
= false;
3318 static const TypeInfo vfio_pci_nohotplug_dev_info
= {
3319 .name
= TYPE_VFIO_PCI_NOHOTPLUG
,
3320 .parent
= TYPE_VFIO_PCI
,
3321 .instance_size
= sizeof(VFIOPCIDevice
),
3322 .class_init
= vfio_pci_nohotplug_dev_class_init
,
3325 static void register_vfio_pci_dev_type(void)
3327 type_register_static(&vfio_pci_dev_info
);
3328 type_register_static(&vfio_pci_nohotplug_dev_info
);
3331 type_init(register_vfio_pci_dev_type
)