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)
23 #include <sys/ioctl.h>
26 #include <sys/types.h>
27 #include <linux/vfio.h>
30 #include "event_notifier.h"
31 #include "exec-memory.h"
37 #include "qemu-common.h"
38 #include "qemu-error.h"
39 #include "qemu-queue.h"
42 /* #define DEBUG_VFIO */
44 #define DPRINTF(fmt, ...) \
45 do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
47 #define DPRINTF(fmt, ...) \
51 typedef struct VFIOBAR
{
52 off_t fd_offset
; /* offset of BAR within device fd */
53 int fd
; /* device fd, allows us to pass VFIOBAR as opaque data */
54 MemoryRegion mem
; /* slow, read/write access */
55 MemoryRegion mmap_mem
; /* direct mapped access */
58 uint32_t flags
; /* VFIO region flags (rd/wr/mmap) */
59 uint8_t nr
; /* cache the BAR number for debug */
62 typedef struct VFIOINTx
{
63 bool pending
; /* interrupt pending */
64 bool kvm_accel
; /* set when QEMU bypass through KVM enabled */
65 uint8_t pin
; /* which pin to pull for qemu_set_irq */
66 EventNotifier interrupt
; /* eventfd triggered on interrupt */
67 EventNotifier unmask
; /* eventfd for unmask on QEMU bypass */
68 PCIINTxRoute route
; /* routing info for QEMU bypass */
69 uint32_t mmap_timeout
; /* delay to re-enable mmaps after interrupt */
70 QEMUTimer
*mmap_timer
; /* enable mmaps after periods w/o interrupts */
75 typedef struct VFIOMSIVector
{
76 EventNotifier interrupt
; /* eventfd triggered on interrupt */
77 struct VFIODevice
*vdev
; /* back pointer to device */
78 int virq
; /* KVM irqchip route for QEMU bypass */
91 typedef struct VFIOContainer
{
92 int fd
; /* /dev/vfio/vfio, empowered by the attached groups */
94 /* enable abstraction to support various iommu backends */
96 MemoryListener listener
; /* Used by type1 iommu */
98 void (*release
)(struct VFIOContainer
*);
100 QLIST_HEAD(, VFIOGroup
) group_list
;
101 QLIST_ENTRY(VFIOContainer
) next
;
104 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
105 typedef struct VFIOMSIXInfo
{
109 uint32_t table_offset
;
111 MemoryRegion mmap_mem
;
115 typedef struct VFIODevice
{
119 unsigned int config_size
;
120 off_t config_offset
; /* Offset of config space region within device fd */
121 unsigned int rom_size
;
122 off_t rom_offset
; /* Offset of ROM region within device fd */
124 VFIOMSIVector
*msi_vectors
;
126 int nr_vectors
; /* Number of MSI/MSIX vectors currently in use */
127 int interrupt
; /* Current interrupt type */
128 VFIOBAR bars
[PCI_NUM_REGIONS
- 1]; /* No ROM */
129 PCIHostDeviceAddress host
;
130 QLIST_ENTRY(VFIODevice
) next
;
131 struct VFIOGroup
*group
;
135 typedef struct VFIOGroup
{
138 VFIOContainer
*container
;
139 QLIST_HEAD(, VFIODevice
) device_list
;
140 QLIST_ENTRY(VFIOGroup
) next
;
141 QLIST_ENTRY(VFIOGroup
) container_next
;
144 #define MSIX_CAP_LENGTH 12
146 static QLIST_HEAD(, VFIOContainer
)
147 container_list
= QLIST_HEAD_INITIALIZER(container_list
);
149 static QLIST_HEAD(, VFIOGroup
)
150 group_list
= QLIST_HEAD_INITIALIZER(group_list
);
152 static void vfio_disable_interrupts(VFIODevice
*vdev
);
153 static uint32_t vfio_pci_read_config(PCIDevice
*pdev
, uint32_t addr
, int len
);
154 static void vfio_mmap_set_enabled(VFIODevice
*vdev
, bool enabled
);
157 * Common VFIO interrupt disable
159 static void vfio_disable_irqindex(VFIODevice
*vdev
, int index
)
161 struct vfio_irq_set irq_set
= {
162 .argsz
= sizeof(irq_set
),
163 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_TRIGGER
,
169 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
175 static void vfio_unmask_intx(VFIODevice
*vdev
)
177 struct vfio_irq_set irq_set
= {
178 .argsz
= sizeof(irq_set
),
179 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_UNMASK
,
180 .index
= VFIO_PCI_INTX_IRQ_INDEX
,
185 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
188 #ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
189 static void vfio_mask_intx(VFIODevice
*vdev
)
191 struct vfio_irq_set irq_set
= {
192 .argsz
= sizeof(irq_set
),
193 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_MASK
,
194 .index
= VFIO_PCI_INTX_IRQ_INDEX
,
199 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
204 * Disabling BAR mmaping can be slow, but toggling it around INTx can
205 * also be a huge overhead. We try to get the best of both worlds by
206 * waiting until an interrupt to disable mmaps (subsequent transitions
207 * to the same state are effectively no overhead). If the interrupt has
208 * been serviced and the time gap is long enough, we re-enable mmaps for
209 * performance. This works well for things like graphics cards, which
210 * may not use their interrupt at all and are penalized to an unusable
211 * level by read/write BAR traps. Other devices, like NICs, have more
212 * regular interrupts and see much better latency by staying in non-mmap
213 * mode. We therefore set the default mmap_timeout such that a ping
214 * is just enough to keep the mmap disabled. Users can experiment with
215 * other options with the x-intx-mmap-timeout-ms parameter (a value of
216 * zero disables the timer).
218 static void vfio_intx_mmap_enable(void *opaque
)
220 VFIODevice
*vdev
= opaque
;
222 if (vdev
->intx
.pending
) {
223 qemu_mod_timer(vdev
->intx
.mmap_timer
,
224 qemu_get_clock_ms(vm_clock
) + vdev
->intx
.mmap_timeout
);
228 vfio_mmap_set_enabled(vdev
, true);
231 static void vfio_intx_interrupt(void *opaque
)
233 VFIODevice
*vdev
= opaque
;
235 if (!event_notifier_test_and_clear(&vdev
->intx
.interrupt
)) {
239 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__
, vdev
->host
.domain
,
240 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
241 'A' + vdev
->intx
.pin
);
243 vdev
->intx
.pending
= true;
244 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 1);
245 vfio_mmap_set_enabled(vdev
, false);
246 if (vdev
->intx
.mmap_timeout
) {
247 qemu_mod_timer(vdev
->intx
.mmap_timer
,
248 qemu_get_clock_ms(vm_clock
) + vdev
->intx
.mmap_timeout
);
252 static void vfio_eoi(VFIODevice
*vdev
)
254 if (!vdev
->intx
.pending
) {
258 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__
, vdev
->host
.domain
,
259 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
261 vdev
->intx
.pending
= false;
262 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
263 vfio_unmask_intx(vdev
);
266 static void vfio_enable_intx_kvm(VFIODevice
*vdev
)
269 struct kvm_irqfd irqfd
= {
270 .fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
),
271 .gsi
= vdev
->intx
.route
.irq
,
272 .flags
= KVM_IRQFD_FLAG_RESAMPLE
,
274 struct vfio_irq_set
*irq_set
;
278 if (!kvm_irqfds_enabled() ||
279 vdev
->intx
.route
.mode
!= PCI_INTX_ENABLED
||
280 !kvm_check_extension(kvm_state
, KVM_CAP_IRQFD_RESAMPLE
)) {
284 /* Get to a known interrupt state */
285 qemu_set_fd_handler(irqfd
.fd
, NULL
, NULL
, vdev
);
286 vfio_mask_intx(vdev
);
287 vdev
->intx
.pending
= false;
288 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
290 /* Get an eventfd for resample/unmask */
291 if (event_notifier_init(&vdev
->intx
.unmask
, 0)) {
292 error_report("vfio: Error: event_notifier_init failed eoi\n");
296 /* KVM triggers it, VFIO listens for it */
297 irqfd
.resamplefd
= event_notifier_get_fd(&vdev
->intx
.unmask
);
299 if (kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
)) {
300 error_report("vfio: Error: Failed to setup resample irqfd: %m\n");
304 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
306 irq_set
= g_malloc0(argsz
);
307 irq_set
->argsz
= argsz
;
308 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_UNMASK
;
309 irq_set
->index
= VFIO_PCI_INTX_IRQ_INDEX
;
312 pfd
= (int32_t *)&irq_set
->data
;
314 *pfd
= irqfd
.resamplefd
;
316 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
319 error_report("vfio: Error: Failed to setup INTx unmask fd: %m\n");
324 vfio_unmask_intx(vdev
);
326 vdev
->intx
.kvm_accel
= true;
328 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
329 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
330 vdev
->host
.slot
, vdev
->host
.function
);
335 irqfd
.flags
= KVM_IRQFD_FLAG_DEASSIGN
;
336 kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
);
338 event_notifier_cleanup(&vdev
->intx
.unmask
);
340 qemu_set_fd_handler(irqfd
.fd
, vfio_intx_interrupt
, NULL
, vdev
);
341 vfio_unmask_intx(vdev
);
345 static void vfio_disable_intx_kvm(VFIODevice
*vdev
)
348 struct kvm_irqfd irqfd
= {
349 .fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
),
350 .gsi
= vdev
->intx
.route
.irq
,
351 .flags
= KVM_IRQFD_FLAG_DEASSIGN
,
354 if (!vdev
->intx
.kvm_accel
) {
359 * Get to a known state, hardware masked, QEMU ready to accept new
360 * interrupts, QEMU IRQ de-asserted.
362 vfio_mask_intx(vdev
);
363 vdev
->intx
.pending
= false;
364 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
366 /* Tell KVM to stop listening for an INTx irqfd */
367 if (kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
)) {
368 error_report("vfio: Error: Failed to disable INTx irqfd: %m\n");
371 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
372 event_notifier_cleanup(&vdev
->intx
.unmask
);
374 /* QEMU starts listening for interrupt events. */
375 qemu_set_fd_handler(irqfd
.fd
, vfio_intx_interrupt
, NULL
, vdev
);
377 vdev
->intx
.kvm_accel
= false;
379 /* If we've missed an event, let it re-fire through QEMU */
380 vfio_unmask_intx(vdev
);
382 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
383 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
384 vdev
->host
.slot
, vdev
->host
.function
);
388 static void vfio_update_irq(PCIDevice
*pdev
)
390 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
393 if (vdev
->interrupt
!= VFIO_INT_INTx
) {
397 route
= pci_device_route_intx_to_irq(&vdev
->pdev
, vdev
->intx
.pin
);
399 if (!pci_intx_route_changed(&vdev
->intx
.route
, &route
)) {
400 return; /* Nothing changed */
403 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__
,
404 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
405 vdev
->host
.function
, vdev
->intx
.route
.irq
, route
.irq
);
407 vfio_disable_intx_kvm(vdev
);
409 vdev
->intx
.route
= route
;
411 if (route
.mode
!= PCI_INTX_ENABLED
) {
415 vfio_enable_intx_kvm(vdev
);
417 /* Re-enable the interrupt in cased we missed an EOI */
421 static int vfio_enable_intx(VFIODevice
*vdev
)
423 uint8_t pin
= vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1);
425 struct vfio_irq_set
*irq_set
;
432 vfio_disable_interrupts(vdev
);
434 vdev
->intx
.pin
= pin
- 1; /* Pin A (1) -> irq[0] */
438 * Only conditional to avoid generating error messages on platforms
439 * where we won't actually use the result anyway.
441 if (kvm_irqfds_enabled() &&
442 kvm_check_extension(kvm_state
, KVM_CAP_IRQFD_RESAMPLE
)) {
443 vdev
->intx
.route
= pci_device_route_intx_to_irq(&vdev
->pdev
,
448 ret
= event_notifier_init(&vdev
->intx
.interrupt
, 0);
450 error_report("vfio: Error: event_notifier_init failed\n");
454 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
456 irq_set
= g_malloc0(argsz
);
457 irq_set
->argsz
= argsz
;
458 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_TRIGGER
;
459 irq_set
->index
= VFIO_PCI_INTX_IRQ_INDEX
;
462 pfd
= (int32_t *)&irq_set
->data
;
464 *pfd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
465 qemu_set_fd_handler(*pfd
, vfio_intx_interrupt
, NULL
, vdev
);
467 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
470 error_report("vfio: Error: Failed to setup INTx fd: %m\n");
471 qemu_set_fd_handler(*pfd
, NULL
, NULL
, vdev
);
472 event_notifier_cleanup(&vdev
->intx
.interrupt
);
476 vfio_enable_intx_kvm(vdev
);
478 vdev
->interrupt
= VFIO_INT_INTx
;
480 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
481 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
486 static void vfio_disable_intx(VFIODevice
*vdev
)
490 qemu_del_timer(vdev
->intx
.mmap_timer
);
491 vfio_disable_intx_kvm(vdev
);
492 vfio_disable_irqindex(vdev
, VFIO_PCI_INTX_IRQ_INDEX
);
493 vdev
->intx
.pending
= false;
494 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
495 vfio_mmap_set_enabled(vdev
, true);
497 fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
498 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
499 event_notifier_cleanup(&vdev
->intx
.interrupt
);
501 vdev
->interrupt
= VFIO_INT_NONE
;
503 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
504 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
510 static void vfio_msi_interrupt(void *opaque
)
512 VFIOMSIVector
*vector
= opaque
;
513 VFIODevice
*vdev
= vector
->vdev
;
514 int nr
= vector
- vdev
->msi_vectors
;
516 if (!event_notifier_test_and_clear(&vector
->interrupt
)) {
520 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__
,
521 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
522 vdev
->host
.function
, nr
);
524 if (vdev
->interrupt
== VFIO_INT_MSIX
) {
525 msix_notify(&vdev
->pdev
, nr
);
526 } else if (vdev
->interrupt
== VFIO_INT_MSI
) {
527 msi_notify(&vdev
->pdev
, nr
);
529 error_report("vfio: MSI interrupt receieved, but not enabled?\n");
533 static int vfio_enable_vectors(VFIODevice
*vdev
, bool msix
)
535 struct vfio_irq_set
*irq_set
;
536 int ret
= 0, i
, argsz
;
539 argsz
= sizeof(*irq_set
) + (vdev
->nr_vectors
* sizeof(*fds
));
541 irq_set
= g_malloc0(argsz
);
542 irq_set
->argsz
= argsz
;
543 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_TRIGGER
;
544 irq_set
->index
= msix
? VFIO_PCI_MSIX_IRQ_INDEX
: VFIO_PCI_MSI_IRQ_INDEX
;
546 irq_set
->count
= vdev
->nr_vectors
;
547 fds
= (int32_t *)&irq_set
->data
;
549 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
550 if (!vdev
->msi_vectors
[i
].use
) {
555 fds
[i
] = event_notifier_get_fd(&vdev
->msi_vectors
[i
].interrupt
);
558 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
565 static int vfio_msix_vector_use(PCIDevice
*pdev
,
566 unsigned int nr
, MSIMessage msg
)
568 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
569 VFIOMSIVector
*vector
;
572 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__
,
573 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
574 vdev
->host
.function
, nr
);
576 vector
= &vdev
->msi_vectors
[nr
];
580 msix_vector_use(pdev
, nr
);
582 if (event_notifier_init(&vector
->interrupt
, 0)) {
583 error_report("vfio: Error: event_notifier_init failed\n");
587 * Attempt to enable route through KVM irqchip,
588 * default to userspace handling if unavailable.
590 vector
->virq
= kvm_irqchip_add_msi_route(kvm_state
, msg
);
591 if (vector
->virq
< 0 ||
592 kvm_irqchip_add_irqfd_notifier(kvm_state
, &vector
->interrupt
,
594 if (vector
->virq
>= 0) {
595 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
598 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
599 vfio_msi_interrupt
, NULL
, vector
);
603 * We don't want to have the host allocate all possible MSI vectors
604 * for a device if they're not in use, so we shutdown and incrementally
605 * increase them as needed.
607 if (vdev
->nr_vectors
< nr
+ 1) {
608 vfio_disable_irqindex(vdev
, VFIO_PCI_MSIX_IRQ_INDEX
);
609 vdev
->nr_vectors
= nr
+ 1;
610 ret
= vfio_enable_vectors(vdev
, true);
612 error_report("vfio: failed to enable vectors, %d\n", ret
);
616 struct vfio_irq_set
*irq_set
;
619 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
621 irq_set
= g_malloc0(argsz
);
622 irq_set
->argsz
= argsz
;
623 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
624 VFIO_IRQ_SET_ACTION_TRIGGER
;
625 irq_set
->index
= VFIO_PCI_MSIX_IRQ_INDEX
;
628 pfd
= (int32_t *)&irq_set
->data
;
630 *pfd
= event_notifier_get_fd(&vector
->interrupt
);
632 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
635 error_report("vfio: failed to modify vector, %d\n", ret
);
642 static void vfio_msix_vector_release(PCIDevice
*pdev
, unsigned int nr
)
644 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
645 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[nr
];
647 struct vfio_irq_set
*irq_set
;
650 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__
,
651 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
652 vdev
->host
.function
, nr
);
655 * XXX What's the right thing to do here? This turns off the interrupt
656 * completely, but do we really just want to switch the interrupt to
657 * bouncing through userspace and let msix.c drop it? Not sure.
659 msix_vector_unuse(pdev
, nr
);
661 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
663 irq_set
= g_malloc0(argsz
);
664 irq_set
->argsz
= argsz
;
665 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
666 VFIO_IRQ_SET_ACTION_TRIGGER
;
667 irq_set
->index
= VFIO_PCI_MSIX_IRQ_INDEX
;
670 pfd
= (int32_t *)&irq_set
->data
;
674 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
678 if (vector
->virq
< 0) {
679 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
682 kvm_irqchip_remove_irqfd_notifier(kvm_state
, &vector
->interrupt
,
684 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
688 event_notifier_cleanup(&vector
->interrupt
);
692 static void vfio_enable_msix(VFIODevice
*vdev
)
694 vfio_disable_interrupts(vdev
);
696 vdev
->msi_vectors
= g_malloc0(vdev
->msix
->entries
* sizeof(VFIOMSIVector
));
698 vdev
->interrupt
= VFIO_INT_MSIX
;
700 if (msix_set_vector_notifiers(&vdev
->pdev
, vfio_msix_vector_use
,
701 vfio_msix_vector_release
)) {
702 error_report("vfio: msix_set_vector_notifiers failed\n");
705 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
706 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
709 static void vfio_enable_msi(VFIODevice
*vdev
)
713 vfio_disable_interrupts(vdev
);
715 vdev
->nr_vectors
= msi_nr_vectors_allocated(&vdev
->pdev
);
717 vdev
->msi_vectors
= g_malloc0(vdev
->nr_vectors
* sizeof(VFIOMSIVector
));
719 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
721 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
726 if (event_notifier_init(&vector
->interrupt
, 0)) {
727 error_report("vfio: Error: event_notifier_init failed\n");
730 msg
= msi_get_message(&vdev
->pdev
, i
);
733 * Attempt to enable route through KVM irqchip,
734 * default to userspace handling if unavailable.
736 vector
->virq
= kvm_irqchip_add_msi_route(kvm_state
, msg
);
737 if (vector
->virq
< 0 ||
738 kvm_irqchip_add_irqfd_notifier(kvm_state
, &vector
->interrupt
,
740 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
741 vfio_msi_interrupt
, NULL
, vector
);
745 ret
= vfio_enable_vectors(vdev
, false);
748 error_report("vfio: Error: Failed to setup MSI fds: %m\n");
749 } else if (ret
!= vdev
->nr_vectors
) {
750 error_report("vfio: Error: Failed to enable %d "
751 "MSI vectors, retry with %d\n", vdev
->nr_vectors
, ret
);
754 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
755 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
756 if (vector
->virq
>= 0) {
757 kvm_irqchip_remove_irqfd_notifier(kvm_state
, &vector
->interrupt
,
759 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
762 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
765 event_notifier_cleanup(&vector
->interrupt
);
768 g_free(vdev
->msi_vectors
);
770 if (ret
> 0 && ret
!= vdev
->nr_vectors
) {
771 vdev
->nr_vectors
= ret
;
774 vdev
->nr_vectors
= 0;
779 vdev
->interrupt
= VFIO_INT_MSI
;
781 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__
,
782 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
783 vdev
->host
.function
, vdev
->nr_vectors
);
786 static void vfio_disable_msi_common(VFIODevice
*vdev
)
788 g_free(vdev
->msi_vectors
);
789 vdev
->msi_vectors
= NULL
;
790 vdev
->nr_vectors
= 0;
791 vdev
->interrupt
= VFIO_INT_NONE
;
793 vfio_enable_intx(vdev
);
796 static void vfio_disable_msix(VFIODevice
*vdev
)
798 msix_unset_vector_notifiers(&vdev
->pdev
);
800 if (vdev
->nr_vectors
) {
801 vfio_disable_irqindex(vdev
, VFIO_PCI_MSIX_IRQ_INDEX
);
804 vfio_disable_msi_common(vdev
);
806 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
807 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
810 static void vfio_disable_msi(VFIODevice
*vdev
)
814 vfio_disable_irqindex(vdev
, VFIO_PCI_MSI_IRQ_INDEX
);
816 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
817 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
823 if (vector
->virq
>= 0) {
824 kvm_irqchip_remove_irqfd_notifier(kvm_state
,
825 &vector
->interrupt
, vector
->virq
);
826 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
829 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
833 event_notifier_cleanup(&vector
->interrupt
);
836 vfio_disable_msi_common(vdev
);
838 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
839 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
843 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
845 static void vfio_bar_write(void *opaque
, hwaddr addr
,
846 uint64_t data
, unsigned size
)
848 VFIOBAR
*bar
= opaque
;
861 buf
.word
= cpu_to_le16(data
);
864 buf
.dword
= cpu_to_le32(data
);
867 hw_error("vfio: unsupported write size, %d bytes\n", size
);
871 if (pwrite(bar
->fd
, &buf
, size
, bar
->fd_offset
+ addr
) != size
) {
872 error_report("%s(,0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d) failed: %m\n",
873 __func__
, addr
, data
, size
);
876 DPRINTF("%s(BAR%d+0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d)\n",
877 __func__
, bar
->nr
, addr
, data
, size
);
880 * A read or write to a BAR always signals an INTx EOI. This will
881 * do nothing if not pending (including not in INTx mode). We assume
882 * that a BAR access is in response to an interrupt and that BAR
883 * accesses will service the interrupt. Unfortunately, we don't know
884 * which access will service the interrupt, so we're potentially
885 * getting quite a few host interrupts per guest interrupt.
887 vfio_eoi(container_of(bar
, VFIODevice
, bars
[bar
->nr
]));
890 static uint64_t vfio_bar_read(void *opaque
,
891 hwaddr addr
, unsigned size
)
893 VFIOBAR
*bar
= opaque
;
902 if (pread(bar
->fd
, &buf
, size
, bar
->fd_offset
+ addr
) != size
) {
903 error_report("%s(,0x%"HWADDR_PRIx
", %d) failed: %m\n",
904 __func__
, addr
, size
);
913 data
= le16_to_cpu(buf
.word
);
916 data
= le32_to_cpu(buf
.dword
);
919 hw_error("vfio: unsupported read size, %d bytes\n", size
);
923 DPRINTF("%s(BAR%d+0x%"HWADDR_PRIx
", %d) = 0x%"PRIx64
"\n",
924 __func__
, bar
->nr
, addr
, size
, data
);
926 /* Same as write above */
927 vfio_eoi(container_of(bar
, VFIODevice
, bars
[bar
->nr
]));
932 static const MemoryRegionOps vfio_bar_ops
= {
933 .read
= vfio_bar_read
,
934 .write
= vfio_bar_write
,
935 .endianness
= DEVICE_LITTLE_ENDIAN
,
941 static uint32_t vfio_pci_read_config(PCIDevice
*pdev
, uint32_t addr
, int len
)
943 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
947 * We only need QEMU PCI config support for the ROM BAR, the MSI and MSIX
948 * capabilities, and the multifunction bit below. We let VFIO handle
949 * virtualizing everything else. Performance is not a concern here.
951 if (ranges_overlap(addr
, len
, PCI_ROM_ADDRESS
, 4) ||
952 (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
&&
953 ranges_overlap(addr
, len
, pdev
->msix_cap
, MSIX_CAP_LENGTH
)) ||
954 (pdev
->cap_present
& QEMU_PCI_CAP_MSI
&&
955 ranges_overlap(addr
, len
, pdev
->msi_cap
, vdev
->msi_cap_size
))) {
957 val
= pci_default_read_config(pdev
, addr
, len
);
959 if (pread(vdev
->fd
, &val
, len
, vdev
->config_offset
+ addr
) != len
) {
960 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m\n",
961 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
962 vdev
->host
.slot
, vdev
->host
.function
, addr
, len
);
965 val
= le32_to_cpu(val
);
968 /* Multifunction bit is virualized in QEMU */
969 if (unlikely(ranges_overlap(addr
, len
, PCI_HEADER_TYPE
, 1))) {
970 uint32_t mask
= PCI_HEADER_TYPE_MULTI_FUNCTION
;
976 if (pdev
->cap_present
& QEMU_PCI_CAP_MULTIFUNCTION
) {
983 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__
,
984 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
985 vdev
->host
.function
, addr
, len
, val
);
990 static void vfio_pci_write_config(PCIDevice
*pdev
, uint32_t addr
,
991 uint32_t val
, int len
)
993 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
994 uint32_t val_le
= cpu_to_le32(val
);
996 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__
,
997 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
998 vdev
->host
.function
, addr
, val
, len
);
1000 /* Write everything to VFIO, let it filter out what we can't write */
1001 if (pwrite(vdev
->fd
, &val_le
, len
, vdev
->config_offset
+ addr
) != len
) {
1002 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m\n",
1003 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
1004 vdev
->host
.slot
, vdev
->host
.function
, addr
, val
, len
);
1007 /* Write standard header bits to emulation */
1008 if (addr
< PCI_CONFIG_HEADER_SIZE
) {
1009 pci_default_write_config(pdev
, addr
, val
, len
);
1013 /* MSI/MSI-X Enabling/Disabling */
1014 if (pdev
->cap_present
& QEMU_PCI_CAP_MSI
&&
1015 ranges_overlap(addr
, len
, pdev
->msi_cap
, vdev
->msi_cap_size
)) {
1016 int is_enabled
, was_enabled
= msi_enabled(pdev
);
1018 pci_default_write_config(pdev
, addr
, val
, len
);
1020 is_enabled
= msi_enabled(pdev
);
1022 if (!was_enabled
&& is_enabled
) {
1023 vfio_enable_msi(vdev
);
1024 } else if (was_enabled
&& !is_enabled
) {
1025 vfio_disable_msi(vdev
);
1029 if (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
&&
1030 ranges_overlap(addr
, len
, pdev
->msix_cap
, MSIX_CAP_LENGTH
)) {
1031 int is_enabled
, was_enabled
= msix_enabled(pdev
);
1033 pci_default_write_config(pdev
, addr
, val
, len
);
1035 is_enabled
= msix_enabled(pdev
);
1037 if (!was_enabled
&& is_enabled
) {
1038 vfio_enable_msix(vdev
);
1039 } else if (was_enabled
&& !is_enabled
) {
1040 vfio_disable_msix(vdev
);
1046 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
1048 static int vfio_dma_unmap(VFIOContainer
*container
,
1049 hwaddr iova
, ram_addr_t size
)
1051 struct vfio_iommu_type1_dma_unmap unmap
= {
1052 .argsz
= sizeof(unmap
),
1058 if (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
1059 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno
);
1066 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
1067 ram_addr_t size
, void *vaddr
, bool readonly
)
1069 struct vfio_iommu_type1_dma_map map
= {
1070 .argsz
= sizeof(map
),
1071 .flags
= VFIO_DMA_MAP_FLAG_READ
,
1072 .vaddr
= (__u64
)(uintptr_t)vaddr
,
1078 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
1082 * Try the mapping, if it fails with EBUSY, unmap the region and try
1083 * again. This shouldn't be necessary, but we sometimes see it in
1084 * the the VGA ROM space.
1086 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
1087 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
) == 0 &&
1088 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
1092 DPRINTF("VFIO_MAP_DMA: %d\n", -errno
);
1096 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
1098 return !memory_region_is_ram(section
->mr
);
1101 static void vfio_listener_region_add(MemoryListener
*listener
,
1102 MemoryRegionSection
*section
)
1104 VFIOContainer
*container
= container_of(listener
, VFIOContainer
,
1105 iommu_data
.listener
);
1110 if (vfio_listener_skipped_section(section
)) {
1111 DPRINTF("vfio: SKIPPING region_add %"HWADDR_PRIx
" - %"PRIx64
"\n",
1112 section
->offset_within_address_space
,
1113 section
->offset_within_address_space
+ section
->size
- 1);
1117 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
1118 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
1119 error_report("%s received unaligned region\n", __func__
);
1123 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
1124 end
= (section
->offset_within_address_space
+ section
->size
) &
1131 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
1132 section
->offset_within_region
+
1133 (iova
- section
->offset_within_address_space
);
1135 DPRINTF("vfio: region_add %"HWADDR_PRIx
" - %"HWADDR_PRIx
" [%p]\n",
1136 iova
, end
- 1, vaddr
);
1138 ret
= vfio_dma_map(container
, iova
, end
- iova
, vaddr
, section
->readonly
);
1140 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
1141 "0x%"HWADDR_PRIx
", %p) = %d (%m)\n",
1142 container
, iova
, end
- iova
, vaddr
, ret
);
1146 static void vfio_listener_region_del(MemoryListener
*listener
,
1147 MemoryRegionSection
*section
)
1149 VFIOContainer
*container
= container_of(listener
, VFIOContainer
,
1150 iommu_data
.listener
);
1154 if (vfio_listener_skipped_section(section
)) {
1155 DPRINTF("vfio: SKIPPING region_del %"HWADDR_PRIx
" - %"PRIx64
"\n",
1156 section
->offset_within_address_space
,
1157 section
->offset_within_address_space
+ section
->size
- 1);
1161 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
1162 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
1163 error_report("%s received unaligned region\n", __func__
);
1167 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
1168 end
= (section
->offset_within_address_space
+ section
->size
) &
1175 DPRINTF("vfio: region_del %"HWADDR_PRIx
" - %"HWADDR_PRIx
"\n",
1178 ret
= vfio_dma_unmap(container
, iova
, end
- iova
);
1180 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
1181 "0x%"HWADDR_PRIx
") = %d (%m)\n",
1182 container
, iova
, end
- iova
, ret
);
1186 static MemoryListener vfio_memory_listener
= {
1187 .region_add
= vfio_listener_region_add
,
1188 .region_del
= vfio_listener_region_del
,
1191 static void vfio_listener_release(VFIOContainer
*container
)
1193 memory_listener_unregister(&container
->iommu_data
.listener
);
1199 static void vfio_disable_interrupts(VFIODevice
*vdev
)
1201 switch (vdev
->interrupt
) {
1203 vfio_disable_intx(vdev
);
1206 vfio_disable_msi(vdev
);
1209 vfio_disable_msix(vdev
);
1214 static int vfio_setup_msi(VFIODevice
*vdev
, int pos
)
1217 bool msi_64bit
, msi_maskbit
;
1220 if (pread(vdev
->fd
, &ctrl
, sizeof(ctrl
),
1221 vdev
->config_offset
+ pos
+ PCI_CAP_FLAGS
) != sizeof(ctrl
)) {
1224 ctrl
= le16_to_cpu(ctrl
);
1226 msi_64bit
= !!(ctrl
& PCI_MSI_FLAGS_64BIT
);
1227 msi_maskbit
= !!(ctrl
& PCI_MSI_FLAGS_MASKBIT
);
1228 entries
= 1 << ((ctrl
& PCI_MSI_FLAGS_QMASK
) >> 1);
1230 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev
->host
.domain
,
1231 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
, pos
);
1233 ret
= msi_init(&vdev
->pdev
, pos
, entries
, msi_64bit
, msi_maskbit
);
1235 if (ret
== -ENOTSUP
) {
1238 error_report("vfio: msi_init failed\n");
1241 vdev
->msi_cap_size
= 0xa + (msi_maskbit
? 0xa : 0) + (msi_64bit
? 0x4 : 0);
1247 * We don't have any control over how pci_add_capability() inserts
1248 * capabilities into the chain. In order to setup MSI-X we need a
1249 * MemoryRegion for the BAR. In order to setup the BAR and not
1250 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
1251 * need to first look for where the MSI-X table lives. So we
1252 * unfortunately split MSI-X setup across two functions.
1254 static int vfio_early_setup_msix(VFIODevice
*vdev
)
1258 uint32_t table
, pba
;
1260 pos
= pci_find_capability(&vdev
->pdev
, PCI_CAP_ID_MSIX
);
1265 if (pread(vdev
->fd
, &ctrl
, sizeof(ctrl
),
1266 vdev
->config_offset
+ pos
+ PCI_CAP_FLAGS
) != sizeof(ctrl
)) {
1270 if (pread(vdev
->fd
, &table
, sizeof(table
),
1271 vdev
->config_offset
+ pos
+ PCI_MSIX_TABLE
) != sizeof(table
)) {
1275 if (pread(vdev
->fd
, &pba
, sizeof(pba
),
1276 vdev
->config_offset
+ pos
+ PCI_MSIX_PBA
) != sizeof(pba
)) {
1280 ctrl
= le16_to_cpu(ctrl
);
1281 table
= le32_to_cpu(table
);
1282 pba
= le32_to_cpu(pba
);
1284 vdev
->msix
= g_malloc0(sizeof(*(vdev
->msix
)));
1285 vdev
->msix
->table_bar
= table
& PCI_MSIX_FLAGS_BIRMASK
;
1286 vdev
->msix
->table_offset
= table
& ~PCI_MSIX_FLAGS_BIRMASK
;
1287 vdev
->msix
->pba_bar
= pba
& PCI_MSIX_FLAGS_BIRMASK
;
1288 vdev
->msix
->pba_offset
= pba
& ~PCI_MSIX_FLAGS_BIRMASK
;
1289 vdev
->msix
->entries
= (ctrl
& PCI_MSIX_FLAGS_QSIZE
) + 1;
1291 DPRINTF("%04x:%02x:%02x.%x "
1292 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
1293 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1294 vdev
->host
.function
, pos
, vdev
->msix
->table_bar
,
1295 vdev
->msix
->table_offset
, vdev
->msix
->entries
);
1300 static int vfio_setup_msix(VFIODevice
*vdev
, int pos
)
1304 ret
= msix_init(&vdev
->pdev
, vdev
->msix
->entries
,
1305 &vdev
->bars
[vdev
->msix
->table_bar
].mem
,
1306 vdev
->msix
->table_bar
, vdev
->msix
->table_offset
,
1307 &vdev
->bars
[vdev
->msix
->pba_bar
].mem
,
1308 vdev
->msix
->pba_bar
, vdev
->msix
->pba_offset
, pos
);
1310 if (ret
== -ENOTSUP
) {
1313 error_report("vfio: msix_init failed\n");
1320 static void vfio_teardown_msi(VFIODevice
*vdev
)
1322 msi_uninit(&vdev
->pdev
);
1325 msix_uninit(&vdev
->pdev
, &vdev
->bars
[vdev
->msix
->table_bar
].mem
,
1326 &vdev
->bars
[vdev
->msix
->pba_bar
].mem
);
1333 static void vfio_mmap_set_enabled(VFIODevice
*vdev
, bool enabled
)
1337 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1338 VFIOBAR
*bar
= &vdev
->bars
[i
];
1344 memory_region_set_enabled(&bar
->mmap_mem
, enabled
);
1345 if (vdev
->msix
&& vdev
->msix
->table_bar
== i
) {
1346 memory_region_set_enabled(&vdev
->msix
->mmap_mem
, enabled
);
1351 static void vfio_unmap_bar(VFIODevice
*vdev
, int nr
)
1353 VFIOBAR
*bar
= &vdev
->bars
[nr
];
1359 memory_region_del_subregion(&bar
->mem
, &bar
->mmap_mem
);
1360 munmap(bar
->mmap
, memory_region_size(&bar
->mmap_mem
));
1362 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
1363 memory_region_del_subregion(&bar
->mem
, &vdev
->msix
->mmap_mem
);
1364 munmap(vdev
->msix
->mmap
, memory_region_size(&vdev
->msix
->mmap_mem
));
1367 memory_region_destroy(&bar
->mem
);
1370 static int vfio_mmap_bar(VFIOBAR
*bar
, MemoryRegion
*mem
, MemoryRegion
*submem
,
1371 void **map
, size_t size
, off_t offset
,
1376 if (size
&& bar
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
1379 if (bar
->flags
& VFIO_REGION_INFO_FLAG_READ
) {
1383 if (bar
->flags
& VFIO_REGION_INFO_FLAG_WRITE
) {
1387 *map
= mmap(NULL
, size
, prot
, MAP_SHARED
,
1388 bar
->fd
, bar
->fd_offset
+ offset
);
1389 if (*map
== MAP_FAILED
) {
1395 memory_region_init_ram_ptr(submem
, name
, size
, *map
);
1398 /* Create a zero sized sub-region to make cleanup easy. */
1399 memory_region_init(submem
, name
, 0);
1402 memory_region_add_subregion(mem
, offset
, submem
);
1407 static void vfio_map_bar(VFIODevice
*vdev
, int nr
)
1409 VFIOBAR
*bar
= &vdev
->bars
[nr
];
1410 unsigned size
= bar
->size
;
1416 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
1421 snprintf(name
, sizeof(name
), "VFIO %04x:%02x:%02x.%x BAR %d",
1422 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1423 vdev
->host
.function
, nr
);
1425 /* Determine what type of BAR this is for registration */
1426 ret
= pread(vdev
->fd
, &pci_bar
, sizeof(pci_bar
),
1427 vdev
->config_offset
+ PCI_BASE_ADDRESS_0
+ (4 * nr
));
1428 if (ret
!= sizeof(pci_bar
)) {
1429 error_report("vfio: Failed to read BAR %d (%m)\n", nr
);
1433 pci_bar
= le32_to_cpu(pci_bar
);
1434 type
= pci_bar
& (pci_bar
& PCI_BASE_ADDRESS_SPACE_IO
?
1435 ~PCI_BASE_ADDRESS_IO_MASK
: ~PCI_BASE_ADDRESS_MEM_MASK
);
1437 /* A "slow" read/write mapping underlies all BARs */
1438 memory_region_init_io(&bar
->mem
, &vfio_bar_ops
, bar
, name
, size
);
1439 pci_register_bar(&vdev
->pdev
, nr
, type
, &bar
->mem
);
1442 * We can't mmap areas overlapping the MSIX vector table, so we
1443 * potentially insert a direct-mapped subregion before and after it.
1445 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
1446 size
= vdev
->msix
->table_offset
& TARGET_PAGE_MASK
;
1449 strncat(name
, " mmap", sizeof(name
) - strlen(name
) - 1);
1450 if (vfio_mmap_bar(bar
, &bar
->mem
,
1451 &bar
->mmap_mem
, &bar
->mmap
, size
, 0, name
)) {
1452 error_report("%s unsupported. Performance may be slow\n", name
);
1455 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
1458 start
= TARGET_PAGE_ALIGN(vdev
->msix
->table_offset
+
1459 (vdev
->msix
->entries
* PCI_MSIX_ENTRY_SIZE
));
1461 size
= start
< bar
->size
? bar
->size
- start
: 0;
1462 strncat(name
, " msix-hi", sizeof(name
) - strlen(name
) - 1);
1463 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
1464 if (vfio_mmap_bar(bar
, &bar
->mem
, &vdev
->msix
->mmap_mem
,
1465 &vdev
->msix
->mmap
, size
, start
, name
)) {
1466 error_report("%s unsupported. Performance may be slow\n", name
);
1471 static void vfio_map_bars(VFIODevice
*vdev
)
1475 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1476 vfio_map_bar(vdev
, i
);
1480 static void vfio_unmap_bars(VFIODevice
*vdev
)
1484 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
1485 vfio_unmap_bar(vdev
, i
);
1492 static uint8_t vfio_std_cap_max_size(PCIDevice
*pdev
, uint8_t pos
)
1494 uint8_t tmp
, next
= 0xff;
1496 for (tmp
= pdev
->config
[PCI_CAPABILITY_LIST
]; tmp
;
1497 tmp
= pdev
->config
[tmp
+ 1]) {
1498 if (tmp
> pos
&& tmp
< next
) {
1506 static int vfio_add_std_cap(VFIODevice
*vdev
, uint8_t pos
)
1508 PCIDevice
*pdev
= &vdev
->pdev
;
1509 uint8_t cap_id
, next
, size
;
1512 cap_id
= pdev
->config
[pos
];
1513 next
= pdev
->config
[pos
+ 1];
1516 * If it becomes important to configure capabilities to their actual
1517 * size, use this as the default when it's something we don't recognize.
1518 * Since QEMU doesn't actually handle many of the config accesses,
1519 * exact size doesn't seem worthwhile.
1521 size
= vfio_std_cap_max_size(pdev
, pos
);
1524 * pci_add_capability always inserts the new capability at the head
1525 * of the chain. Therefore to end up with a chain that matches the
1526 * physical device, we insert from the end by making this recursive.
1527 * This is also why we pre-caclulate size above as cached config space
1528 * will be changed as we unwind the stack.
1531 ret
= vfio_add_std_cap(vdev
, next
);
1536 pdev
->config
[PCI_CAPABILITY_LIST
] = 0; /* Begin the rebuild */
1540 case PCI_CAP_ID_MSI
:
1541 ret
= vfio_setup_msi(vdev
, pos
);
1543 case PCI_CAP_ID_MSIX
:
1544 ret
= vfio_setup_msix(vdev
, pos
);
1547 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
);
1552 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
1553 "0x%x[0x%x]@0x%x: %d\n", vdev
->host
.domain
,
1554 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
1555 cap_id
, size
, pos
, ret
);
1562 static int vfio_add_capabilities(VFIODevice
*vdev
)
1564 PCIDevice
*pdev
= &vdev
->pdev
;
1566 if (!(pdev
->config
[PCI_STATUS
] & PCI_STATUS_CAP_LIST
) ||
1567 !pdev
->config
[PCI_CAPABILITY_LIST
]) {
1568 return 0; /* Nothing to add */
1571 return vfio_add_std_cap(vdev
, pdev
->config
[PCI_CAPABILITY_LIST
]);
1574 static int vfio_load_rom(VFIODevice
*vdev
)
1576 uint64_t size
= vdev
->rom_size
;
1578 off_t off
= 0, voff
= vdev
->rom_offset
;
1582 /* If loading ROM from file, pci handles it */
1583 if (vdev
->pdev
.romfile
|| !vdev
->pdev
.rom_bar
|| !size
) {
1587 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
1588 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
1590 snprintf(name
, sizeof(name
), "vfio[%04x:%02x:%02x.%x].rom",
1591 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1592 vdev
->host
.function
);
1593 memory_region_init_ram(&vdev
->pdev
.rom
, name
, size
);
1594 ptr
= memory_region_get_ram_ptr(&vdev
->pdev
.rom
);
1595 memset(ptr
, 0xff, size
);
1598 bytes
= pread(vdev
->fd
, ptr
+ off
, size
, voff
+ off
);
1600 break; /* expect that we could get back less than the ROM BAR */
1601 } else if (bytes
> 0) {
1605 if (errno
== EINTR
|| errno
== EAGAIN
) {
1608 error_report("vfio: Error reading device ROM: %m\n");
1609 memory_region_destroy(&vdev
->pdev
.rom
);
1614 pci_register_bar(&vdev
->pdev
, PCI_ROM_SLOT
, 0, &vdev
->pdev
.rom
);
1615 vdev
->pdev
.has_rom
= true;
1619 static int vfio_connect_container(VFIOGroup
*group
)
1621 VFIOContainer
*container
;
1624 if (group
->container
) {
1628 QLIST_FOREACH(container
, &container_list
, next
) {
1629 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
1630 group
->container
= container
;
1631 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1636 fd
= qemu_open("/dev/vfio/vfio", O_RDWR
);
1638 error_report("vfio: failed to open /dev/vfio/vfio: %m\n");
1642 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
1643 if (ret
!= VFIO_API_VERSION
) {
1644 error_report("vfio: supported vfio version: %d, "
1645 "reported version: %d\n", VFIO_API_VERSION
, ret
);
1650 container
= g_malloc0(sizeof(*container
));
1653 if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1_IOMMU
)) {
1654 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
1656 error_report("vfio: failed to set group container: %m\n");
1662 ret
= ioctl(fd
, VFIO_SET_IOMMU
, VFIO_TYPE1_IOMMU
);
1664 error_report("vfio: failed to set iommu for container: %m\n");
1670 container
->iommu_data
.listener
= vfio_memory_listener
;
1671 container
->iommu_data
.release
= vfio_listener_release
;
1673 memory_listener_register(&container
->iommu_data
.listener
, &address_space_memory
);
1675 error_report("vfio: No available IOMMU models\n");
1681 QLIST_INIT(&container
->group_list
);
1682 QLIST_INSERT_HEAD(&container_list
, container
, next
);
1684 group
->container
= container
;
1685 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
1690 static void vfio_disconnect_container(VFIOGroup
*group
)
1692 VFIOContainer
*container
= group
->container
;
1694 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
1695 error_report("vfio: error disconnecting group %d from container\n",
1699 QLIST_REMOVE(group
, container_next
);
1700 group
->container
= NULL
;
1702 if (QLIST_EMPTY(&container
->group_list
)) {
1703 if (container
->iommu_data
.release
) {
1704 container
->iommu_data
.release(container
);
1706 QLIST_REMOVE(container
, next
);
1707 DPRINTF("vfio_disconnect_container: close container->fd\n");
1708 close(container
->fd
);
1713 static VFIOGroup
*vfio_get_group(int groupid
)
1717 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
1719 QLIST_FOREACH(group
, &group_list
, next
) {
1720 if (group
->groupid
== groupid
) {
1725 group
= g_malloc0(sizeof(*group
));
1727 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
1728 group
->fd
= qemu_open(path
, O_RDWR
);
1729 if (group
->fd
< 0) {
1730 error_report("vfio: error opening %s: %m\n", path
);
1735 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
1736 error_report("vfio: error getting group status: %m\n");
1742 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
1743 error_report("vfio: error, group %d is not viable, please ensure "
1744 "all devices within the iommu_group are bound to their "
1745 "vfio bus driver.\n", groupid
);
1751 group
->groupid
= groupid
;
1752 QLIST_INIT(&group
->device_list
);
1754 if (vfio_connect_container(group
)) {
1755 error_report("vfio: failed to setup container for group %d\n", groupid
);
1761 QLIST_INSERT_HEAD(&group_list
, group
, next
);
1766 static void vfio_put_group(VFIOGroup
*group
)
1768 if (!QLIST_EMPTY(&group
->device_list
)) {
1772 vfio_disconnect_container(group
);
1773 QLIST_REMOVE(group
, next
);
1774 DPRINTF("vfio_put_group: close group->fd\n");
1779 static int vfio_get_device(VFIOGroup
*group
, const char *name
, VFIODevice
*vdev
)
1781 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
1782 struct vfio_region_info reg_info
= { .argsz
= sizeof(reg_info
) };
1785 ret
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
1787 error_report("vfio: error getting device %s from group %d: %m\n",
1788 name
, group
->groupid
);
1789 error_report("Verify all devices in group %d are bound to vfio-pci "
1790 "or pci-stub and not already in use\n", group
->groupid
);
1795 vdev
->group
= group
;
1796 QLIST_INSERT_HEAD(&group
->device_list
, vdev
, next
);
1798 /* Sanity check device */
1799 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
1801 error_report("vfio: error getting device info: %m\n");
1805 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name
,
1806 dev_info
.flags
, dev_info
.num_regions
, dev_info
.num_irqs
);
1808 if (!(dev_info
.flags
& VFIO_DEVICE_FLAGS_PCI
)) {
1809 error_report("vfio: Um, this isn't a PCI device\n");
1813 vdev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
1814 if (!vdev
->reset_works
) {
1815 error_report("Warning, device %s does not support reset\n", name
);
1818 if (dev_info
.num_regions
!= VFIO_PCI_NUM_REGIONS
) {
1819 error_report("vfio: unexpected number of io regions %u\n",
1820 dev_info
.num_regions
);
1824 if (dev_info
.num_irqs
!= VFIO_PCI_NUM_IRQS
) {
1825 error_report("vfio: unexpected number of irqs %u\n", dev_info
.num_irqs
);
1829 for (i
= VFIO_PCI_BAR0_REGION_INDEX
; i
< VFIO_PCI_ROM_REGION_INDEX
; i
++) {
1832 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
);
1834 error_report("vfio: Error getting region %d info: %m\n", i
);
1838 DPRINTF("Device %s region %d:\n", name
, i
);
1839 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1840 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
1841 (unsigned long)reg_info
.flags
);
1843 vdev
->bars
[i
].flags
= reg_info
.flags
;
1844 vdev
->bars
[i
].size
= reg_info
.size
;
1845 vdev
->bars
[i
].fd_offset
= reg_info
.offset
;
1846 vdev
->bars
[i
].fd
= vdev
->fd
;
1847 vdev
->bars
[i
].nr
= i
;
1850 reg_info
.index
= VFIO_PCI_ROM_REGION_INDEX
;
1852 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
);
1854 error_report("vfio: Error getting ROM info: %m\n");
1858 DPRINTF("Device %s ROM:\n", name
);
1859 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1860 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
1861 (unsigned long)reg_info
.flags
);
1863 vdev
->rom_size
= reg_info
.size
;
1864 vdev
->rom_offset
= reg_info
.offset
;
1866 reg_info
.index
= VFIO_PCI_CONFIG_REGION_INDEX
;
1868 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
);
1870 error_report("vfio: Error getting config info: %m\n");
1874 DPRINTF("Device %s config:\n", name
);
1875 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1876 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
1877 (unsigned long)reg_info
.flags
);
1879 vdev
->config_size
= reg_info
.size
;
1880 vdev
->config_offset
= reg_info
.offset
;
1884 QLIST_REMOVE(vdev
, next
);
1891 static void vfio_put_device(VFIODevice
*vdev
)
1893 QLIST_REMOVE(vdev
, next
);
1895 DPRINTF("vfio_put_device: close vdev->fd\n");
1903 static int vfio_initfn(PCIDevice
*pdev
)
1905 VFIODevice
*pvdev
, *vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
1907 char path
[PATH_MAX
], iommu_group_path
[PATH_MAX
], *group_name
;
1913 /* Check that the host device exists */
1914 snprintf(path
, sizeof(path
),
1915 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
1916 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1917 vdev
->host
.function
);
1918 if (stat(path
, &st
) < 0) {
1919 error_report("vfio: error: no such host device: %s\n", path
);
1923 strncat(path
, "iommu_group", sizeof(path
) - strlen(path
) - 1);
1925 len
= readlink(path
, iommu_group_path
, PATH_MAX
);
1927 error_report("vfio: error no iommu_group for device\n");
1931 iommu_group_path
[len
] = 0;
1932 group_name
= basename(iommu_group_path
);
1934 if (sscanf(group_name
, "%d", &groupid
) != 1) {
1935 error_report("vfio: error reading %s: %m\n", path
);
1939 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__
, vdev
->host
.domain
,
1940 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
, groupid
);
1942 group
= vfio_get_group(groupid
);
1944 error_report("vfio: failed to get group %d\n", groupid
);
1948 snprintf(path
, sizeof(path
), "%04x:%02x:%02x.%01x",
1949 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1950 vdev
->host
.function
);
1952 QLIST_FOREACH(pvdev
, &group
->device_list
, next
) {
1953 if (pvdev
->host
.domain
== vdev
->host
.domain
&&
1954 pvdev
->host
.bus
== vdev
->host
.bus
&&
1955 pvdev
->host
.slot
== vdev
->host
.slot
&&
1956 pvdev
->host
.function
== vdev
->host
.function
) {
1958 error_report("vfio: error: device %s is already attached\n", path
);
1959 vfio_put_group(group
);
1964 ret
= vfio_get_device(group
, path
, vdev
);
1966 error_report("vfio: failed to get device %s\n", path
);
1967 vfio_put_group(group
);
1971 /* Get a copy of config space */
1972 ret
= pread(vdev
->fd
, vdev
->pdev
.config
,
1973 MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
),
1974 vdev
->config_offset
);
1975 if (ret
< (int)MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
)) {
1976 ret
= ret
< 0 ? -errno
: -EFAULT
;
1977 error_report("vfio: Failed to read device config space\n");
1982 * Clear host resource mapping info. If we choose not to register a
1983 * BAR, such as might be the case with the option ROM, we can get
1984 * confusing, unwritable, residual addresses from the host here.
1986 memset(&vdev
->pdev
.config
[PCI_BASE_ADDRESS_0
], 0, 24);
1987 memset(&vdev
->pdev
.config
[PCI_ROM_ADDRESS
], 0, 4);
1989 vfio_load_rom(vdev
);
1991 ret
= vfio_early_setup_msix(vdev
);
1996 vfio_map_bars(vdev
);
1998 ret
= vfio_add_capabilities(vdev
);
2003 if (vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1)) {
2004 vdev
->intx
.mmap_timer
= qemu_new_timer_ms(vm_clock
,
2005 vfio_intx_mmap_enable
, vdev
);
2006 pci_device_set_intx_routing_notifier(&vdev
->pdev
, vfio_update_irq
);
2007 ret
= vfio_enable_intx(vdev
);
2016 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
2017 vfio_teardown_msi(vdev
);
2018 vfio_unmap_bars(vdev
);
2020 vfio_put_device(vdev
);
2021 vfio_put_group(group
);
2025 static void vfio_exitfn(PCIDevice
*pdev
)
2027 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
2028 VFIOGroup
*group
= vdev
->group
;
2030 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
2031 vfio_disable_interrupts(vdev
);
2032 if (vdev
->intx
.mmap_timer
) {
2033 qemu_free_timer(vdev
->intx
.mmap_timer
);
2035 vfio_teardown_msi(vdev
);
2036 vfio_unmap_bars(vdev
);
2037 vfio_put_device(vdev
);
2038 vfio_put_group(group
);
2041 static void vfio_pci_reset(DeviceState
*dev
)
2043 PCIDevice
*pdev
= DO_UPCAST(PCIDevice
, qdev
, dev
);
2044 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
2047 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
2048 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
2050 vfio_disable_interrupts(vdev
);
2053 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2054 * Also put INTx Disable in known state.
2056 cmd
= vfio_pci_read_config(pdev
, PCI_COMMAND
, 2);
2057 cmd
&= ~(PCI_COMMAND_IO
| PCI_COMMAND_MEMORY
| PCI_COMMAND_MASTER
|
2058 PCI_COMMAND_INTX_DISABLE
);
2059 vfio_pci_write_config(pdev
, PCI_COMMAND
, cmd
, 2);
2061 if (vdev
->reset_works
) {
2062 if (ioctl(vdev
->fd
, VFIO_DEVICE_RESET
)) {
2063 error_report("vfio: Error unable to reset physical device "
2064 "(%04x:%02x:%02x.%x): %m\n", vdev
->host
.domain
,
2065 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
2069 vfio_enable_intx(vdev
);
2072 static Property vfio_pci_dev_properties
[] = {
2073 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice
, host
),
2074 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice
,
2075 intx
.mmap_timeout
, 1100),
2077 * TODO - support passed fds... is this necessary?
2078 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
2079 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
2081 DEFINE_PROP_END_OF_LIST(),
2084 static const VMStateDescription vfio_pci_vmstate
= {
2089 static void vfio_pci_dev_class_init(ObjectClass
*klass
, void *data
)
2091 DeviceClass
*dc
= DEVICE_CLASS(klass
);
2092 PCIDeviceClass
*pdc
= PCI_DEVICE_CLASS(klass
);
2094 dc
->reset
= vfio_pci_reset
;
2095 dc
->props
= vfio_pci_dev_properties
;
2096 dc
->vmsd
= &vfio_pci_vmstate
;
2097 dc
->desc
= "VFIO-based PCI device assignment";
2098 pdc
->init
= vfio_initfn
;
2099 pdc
->exit
= vfio_exitfn
;
2100 pdc
->config_read
= vfio_pci_read_config
;
2101 pdc
->config_write
= vfio_pci_write_config
;
2104 static const TypeInfo vfio_pci_dev_info
= {
2106 .parent
= TYPE_PCI_DEVICE
,
2107 .instance_size
= sizeof(VFIODevice
),
2108 .class_init
= vfio_pci_dev_class_init
,
2111 static void register_vfio_pci_dev_type(void)
2113 type_register_static(&vfio_pci_dev_info
);
2116 type_init(register_vfio_pci_dev_type
)