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)
22 #include <linux/vfio.h>
23 #include <sys/ioctl.h>
26 #include <sys/types.h>
30 #include "exec/address-spaces.h"
31 #include "exec/memory.h"
32 #include "hw/pci/msi.h"
33 #include "hw/pci/msix.h"
34 #include "hw/pci/pci.h"
35 #include "qemu-common.h"
36 #include "qemu/error-report.h"
37 #include "qemu/event_notifier.h"
38 #include "qemu/queue.h"
39 #include "qemu/range.h"
40 #include "sysemu/kvm.h"
41 #include "sysemu/sysemu.h"
43 /* #define DEBUG_VFIO */
45 #define DPRINTF(fmt, ...) \
46 do { fprintf(stderr, "vfio: " fmt, ## __VA_ARGS__); } while (0)
48 #define DPRINTF(fmt, ...) \
52 /* Extra debugging, trap acceleration paths for more logging */
53 #define VFIO_ALLOW_MMAP 1
54 #define VFIO_ALLOW_KVM_INTX 1
58 typedef struct VFIOQuirk
{
60 struct VFIODevice
*vdev
;
61 QLIST_ENTRY(VFIOQuirk
) next
;
63 uint32_t base_offset
:TARGET_PAGE_BITS
;
64 uint32_t address_offset
:TARGET_PAGE_BITS
;
65 uint32_t address_size
:3;
68 uint32_t address_match
;
69 uint32_t address_mask
;
71 uint32_t address_val
:TARGET_PAGE_BITS
;
72 uint32_t data_offset
:TARGET_PAGE_BITS
;
81 typedef struct VFIOBAR
{
82 off_t fd_offset
; /* offset of BAR within device fd */
83 int fd
; /* device fd, allows us to pass VFIOBAR as opaque data */
84 MemoryRegion mem
; /* slow, read/write access */
85 MemoryRegion mmap_mem
; /* direct mapped access */
88 uint32_t flags
; /* VFIO region flags (rd/wr/mmap) */
89 uint8_t nr
; /* cache the BAR number for debug */
92 QLIST_HEAD(, VFIOQuirk
) quirks
;
95 typedef struct VFIOVGARegion
{
99 QLIST_HEAD(, VFIOQuirk
) quirks
;
102 typedef struct VFIOVGA
{
105 VFIOVGARegion region
[QEMU_PCI_VGA_NUM_REGIONS
];
108 typedef struct VFIOINTx
{
109 bool pending
; /* interrupt pending */
110 bool kvm_accel
; /* set when QEMU bypass through KVM enabled */
111 uint8_t pin
; /* which pin to pull for qemu_set_irq */
112 EventNotifier interrupt
; /* eventfd triggered on interrupt */
113 EventNotifier unmask
; /* eventfd for unmask on QEMU bypass */
114 PCIINTxRoute route
; /* routing info for QEMU bypass */
115 uint32_t mmap_timeout
; /* delay to re-enable mmaps after interrupt */
116 QEMUTimer
*mmap_timer
; /* enable mmaps after periods w/o interrupts */
119 typedef struct VFIOMSIVector
{
120 EventNotifier interrupt
; /* eventfd triggered on interrupt */
121 struct VFIODevice
*vdev
; /* back pointer to device */
122 MSIMessage msg
; /* cache the MSI message so we know when it changes */
123 int virq
; /* KVM irqchip route for QEMU bypass */
136 typedef struct VFIOContainer
{
137 int fd
; /* /dev/vfio/vfio, empowered by the attached groups */
139 /* enable abstraction to support various iommu backends */
141 MemoryListener listener
; /* Used by type1 iommu */
143 void (*release
)(struct VFIOContainer
*);
145 QLIST_HEAD(, VFIOGroup
) group_list
;
146 QLIST_ENTRY(VFIOContainer
) next
;
149 /* Cache of MSI-X setup plus extra mmap and memory region for split BAR map */
150 typedef struct VFIOMSIXInfo
{
154 uint32_t table_offset
;
156 MemoryRegion mmap_mem
;
160 typedef struct VFIODevice
{
164 unsigned int config_size
;
165 uint8_t *emulated_config_bits
; /* QEMU emulated bits, little-endian */
166 off_t config_offset
; /* Offset of config space region within device fd */
167 unsigned int rom_size
;
168 off_t rom_offset
; /* Offset of ROM region within device fd */
171 VFIOMSIVector
*msi_vectors
;
173 int nr_vectors
; /* Number of MSI/MSIX vectors currently in use */
174 int interrupt
; /* Current interrupt type */
175 VFIOBAR bars
[PCI_NUM_REGIONS
- 1]; /* No ROM */
176 VFIOVGA vga
; /* 0xa0000, 0x3b0, 0x3c0 */
177 PCIHostDeviceAddress host
;
178 QLIST_ENTRY(VFIODevice
) next
;
179 struct VFIOGroup
*group
;
180 EventNotifier err_notifier
;
182 #define VFIO_FEATURE_ENABLE_VGA_BIT 0
183 #define VFIO_FEATURE_ENABLE_VGA (1 << VFIO_FEATURE_ENABLE_VGA_BIT)
194 typedef struct VFIOGroup
{
197 VFIOContainer
*container
;
198 QLIST_HEAD(, VFIODevice
) device_list
;
199 QLIST_ENTRY(VFIOGroup
) next
;
200 QLIST_ENTRY(VFIOGroup
) container_next
;
203 #define MSIX_CAP_LENGTH 12
205 static QLIST_HEAD(, VFIOContainer
)
206 container_list
= QLIST_HEAD_INITIALIZER(container_list
);
208 static QLIST_HEAD(, VFIOGroup
)
209 group_list
= QLIST_HEAD_INITIALIZER(group_list
);
211 static void vfio_disable_interrupts(VFIODevice
*vdev
);
212 static uint32_t vfio_pci_read_config(PCIDevice
*pdev
, uint32_t addr
, int len
);
213 static void vfio_pci_write_config(PCIDevice
*pdev
, uint32_t addr
,
214 uint32_t val
, int len
);
215 static void vfio_mmap_set_enabled(VFIODevice
*vdev
, bool enabled
);
218 * Common VFIO interrupt disable
220 static void vfio_disable_irqindex(VFIODevice
*vdev
, int index
)
222 struct vfio_irq_set irq_set
= {
223 .argsz
= sizeof(irq_set
),
224 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_TRIGGER
,
230 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
236 static void vfio_unmask_intx(VFIODevice
*vdev
)
238 struct vfio_irq_set irq_set
= {
239 .argsz
= sizeof(irq_set
),
240 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_UNMASK
,
241 .index
= VFIO_PCI_INTX_IRQ_INDEX
,
246 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
249 #ifdef CONFIG_KVM /* Unused outside of CONFIG_KVM code */
250 static void vfio_mask_intx(VFIODevice
*vdev
)
252 struct vfio_irq_set irq_set
= {
253 .argsz
= sizeof(irq_set
),
254 .flags
= VFIO_IRQ_SET_DATA_NONE
| VFIO_IRQ_SET_ACTION_MASK
,
255 .index
= VFIO_PCI_INTX_IRQ_INDEX
,
260 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, &irq_set
);
265 * Disabling BAR mmaping can be slow, but toggling it around INTx can
266 * also be a huge overhead. We try to get the best of both worlds by
267 * waiting until an interrupt to disable mmaps (subsequent transitions
268 * to the same state are effectively no overhead). If the interrupt has
269 * been serviced and the time gap is long enough, we re-enable mmaps for
270 * performance. This works well for things like graphics cards, which
271 * may not use their interrupt at all and are penalized to an unusable
272 * level by read/write BAR traps. Other devices, like NICs, have more
273 * regular interrupts and see much better latency by staying in non-mmap
274 * mode. We therefore set the default mmap_timeout such that a ping
275 * is just enough to keep the mmap disabled. Users can experiment with
276 * other options with the x-intx-mmap-timeout-ms parameter (a value of
277 * zero disables the timer).
279 static void vfio_intx_mmap_enable(void *opaque
)
281 VFIODevice
*vdev
= opaque
;
283 if (vdev
->intx
.pending
) {
284 timer_mod(vdev
->intx
.mmap_timer
,
285 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) + vdev
->intx
.mmap_timeout
);
289 vfio_mmap_set_enabled(vdev
, true);
292 static void vfio_intx_interrupt(void *opaque
)
294 VFIODevice
*vdev
= opaque
;
296 if (!event_notifier_test_and_clear(&vdev
->intx
.interrupt
)) {
300 DPRINTF("%s(%04x:%02x:%02x.%x) Pin %c\n", __func__
, vdev
->host
.domain
,
301 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
302 'A' + vdev
->intx
.pin
);
304 vdev
->intx
.pending
= true;
305 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 1);
306 vfio_mmap_set_enabled(vdev
, false);
307 if (vdev
->intx
.mmap_timeout
) {
308 timer_mod(vdev
->intx
.mmap_timer
,
309 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) + vdev
->intx
.mmap_timeout
);
313 static void vfio_eoi(VFIODevice
*vdev
)
315 if (!vdev
->intx
.pending
) {
319 DPRINTF("%s(%04x:%02x:%02x.%x) EOI\n", __func__
, vdev
->host
.domain
,
320 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
322 vdev
->intx
.pending
= false;
323 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
324 vfio_unmask_intx(vdev
);
327 static void vfio_enable_intx_kvm(VFIODevice
*vdev
)
330 struct kvm_irqfd irqfd
= {
331 .fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
),
332 .gsi
= vdev
->intx
.route
.irq
,
333 .flags
= KVM_IRQFD_FLAG_RESAMPLE
,
335 struct vfio_irq_set
*irq_set
;
339 if (!VFIO_ALLOW_KVM_INTX
|| !kvm_irqfds_enabled() ||
340 vdev
->intx
.route
.mode
!= PCI_INTX_ENABLED
||
341 !kvm_check_extension(kvm_state
, KVM_CAP_IRQFD_RESAMPLE
)) {
345 /* Get to a known interrupt state */
346 qemu_set_fd_handler(irqfd
.fd
, NULL
, NULL
, vdev
);
347 vfio_mask_intx(vdev
);
348 vdev
->intx
.pending
= false;
349 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
351 /* Get an eventfd for resample/unmask */
352 if (event_notifier_init(&vdev
->intx
.unmask
, 0)) {
353 error_report("vfio: Error: event_notifier_init failed eoi");
357 /* KVM triggers it, VFIO listens for it */
358 irqfd
.resamplefd
= event_notifier_get_fd(&vdev
->intx
.unmask
);
360 if (kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
)) {
361 error_report("vfio: Error: Failed to setup resample irqfd: %m");
365 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
367 irq_set
= g_malloc0(argsz
);
368 irq_set
->argsz
= argsz
;
369 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_UNMASK
;
370 irq_set
->index
= VFIO_PCI_INTX_IRQ_INDEX
;
373 pfd
= (int32_t *)&irq_set
->data
;
375 *pfd
= irqfd
.resamplefd
;
377 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
380 error_report("vfio: Error: Failed to setup INTx unmask fd: %m");
385 vfio_unmask_intx(vdev
);
387 vdev
->intx
.kvm_accel
= true;
389 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel enabled\n",
390 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
391 vdev
->host
.slot
, vdev
->host
.function
);
396 irqfd
.flags
= KVM_IRQFD_FLAG_DEASSIGN
;
397 kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
);
399 event_notifier_cleanup(&vdev
->intx
.unmask
);
401 qemu_set_fd_handler(irqfd
.fd
, vfio_intx_interrupt
, NULL
, vdev
);
402 vfio_unmask_intx(vdev
);
406 static void vfio_disable_intx_kvm(VFIODevice
*vdev
)
409 struct kvm_irqfd irqfd
= {
410 .fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
),
411 .gsi
= vdev
->intx
.route
.irq
,
412 .flags
= KVM_IRQFD_FLAG_DEASSIGN
,
415 if (!vdev
->intx
.kvm_accel
) {
420 * Get to a known state, hardware masked, QEMU ready to accept new
421 * interrupts, QEMU IRQ de-asserted.
423 vfio_mask_intx(vdev
);
424 vdev
->intx
.pending
= false;
425 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
427 /* Tell KVM to stop listening for an INTx irqfd */
428 if (kvm_vm_ioctl(kvm_state
, KVM_IRQFD
, &irqfd
)) {
429 error_report("vfio: Error: Failed to disable INTx irqfd: %m");
432 /* We only need to close the eventfd for VFIO to cleanup the kernel side */
433 event_notifier_cleanup(&vdev
->intx
.unmask
);
435 /* QEMU starts listening for interrupt events. */
436 qemu_set_fd_handler(irqfd
.fd
, vfio_intx_interrupt
, NULL
, vdev
);
438 vdev
->intx
.kvm_accel
= false;
440 /* If we've missed an event, let it re-fire through QEMU */
441 vfio_unmask_intx(vdev
);
443 DPRINTF("%s(%04x:%02x:%02x.%x) KVM INTx accel disabled\n",
444 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
445 vdev
->host
.slot
, vdev
->host
.function
);
449 static void vfio_update_irq(PCIDevice
*pdev
)
451 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
454 if (vdev
->interrupt
!= VFIO_INT_INTx
) {
458 route
= pci_device_route_intx_to_irq(&vdev
->pdev
, vdev
->intx
.pin
);
460 if (!pci_intx_route_changed(&vdev
->intx
.route
, &route
)) {
461 return; /* Nothing changed */
464 DPRINTF("%s(%04x:%02x:%02x.%x) IRQ moved %d -> %d\n", __func__
,
465 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
466 vdev
->host
.function
, vdev
->intx
.route
.irq
, route
.irq
);
468 vfio_disable_intx_kvm(vdev
);
470 vdev
->intx
.route
= route
;
472 if (route
.mode
!= PCI_INTX_ENABLED
) {
476 vfio_enable_intx_kvm(vdev
);
478 /* Re-enable the interrupt in cased we missed an EOI */
482 static int vfio_enable_intx(VFIODevice
*vdev
)
484 uint8_t pin
= vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1);
486 struct vfio_irq_set
*irq_set
;
493 vfio_disable_interrupts(vdev
);
495 vdev
->intx
.pin
= pin
- 1; /* Pin A (1) -> irq[0] */
499 * Only conditional to avoid generating error messages on platforms
500 * where we won't actually use the result anyway.
502 if (kvm_irqfds_enabled() &&
503 kvm_check_extension(kvm_state
, KVM_CAP_IRQFD_RESAMPLE
)) {
504 vdev
->intx
.route
= pci_device_route_intx_to_irq(&vdev
->pdev
,
509 ret
= event_notifier_init(&vdev
->intx
.interrupt
, 0);
511 error_report("vfio: Error: event_notifier_init failed");
515 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
517 irq_set
= g_malloc0(argsz
);
518 irq_set
->argsz
= argsz
;
519 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_TRIGGER
;
520 irq_set
->index
= VFIO_PCI_INTX_IRQ_INDEX
;
523 pfd
= (int32_t *)&irq_set
->data
;
525 *pfd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
526 qemu_set_fd_handler(*pfd
, vfio_intx_interrupt
, NULL
, vdev
);
528 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
531 error_report("vfio: Error: Failed to setup INTx fd: %m");
532 qemu_set_fd_handler(*pfd
, NULL
, NULL
, vdev
);
533 event_notifier_cleanup(&vdev
->intx
.interrupt
);
537 vfio_enable_intx_kvm(vdev
);
539 vdev
->interrupt
= VFIO_INT_INTx
;
541 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
542 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
547 static void vfio_disable_intx(VFIODevice
*vdev
)
551 timer_del(vdev
->intx
.mmap_timer
);
552 vfio_disable_intx_kvm(vdev
);
553 vfio_disable_irqindex(vdev
, VFIO_PCI_INTX_IRQ_INDEX
);
554 vdev
->intx
.pending
= false;
555 qemu_set_irq(vdev
->pdev
.irq
[vdev
->intx
.pin
], 0);
556 vfio_mmap_set_enabled(vdev
, true);
558 fd
= event_notifier_get_fd(&vdev
->intx
.interrupt
);
559 qemu_set_fd_handler(fd
, NULL
, NULL
, vdev
);
560 event_notifier_cleanup(&vdev
->intx
.interrupt
);
562 vdev
->interrupt
= VFIO_INT_NONE
;
564 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
565 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
571 static void vfio_msi_interrupt(void *opaque
)
573 VFIOMSIVector
*vector
= opaque
;
574 VFIODevice
*vdev
= vector
->vdev
;
575 int nr
= vector
- vdev
->msi_vectors
;
577 if (!event_notifier_test_and_clear(&vector
->interrupt
)) {
581 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d\n", __func__
,
582 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
583 vdev
->host
.function
, nr
);
585 if (vdev
->interrupt
== VFIO_INT_MSIX
) {
586 msix_notify(&vdev
->pdev
, nr
);
587 } else if (vdev
->interrupt
== VFIO_INT_MSI
) {
588 msi_notify(&vdev
->pdev
, nr
);
590 error_report("vfio: MSI interrupt receieved, but not enabled?");
594 static int vfio_enable_vectors(VFIODevice
*vdev
, bool msix
)
596 struct vfio_irq_set
*irq_set
;
597 int ret
= 0, i
, argsz
;
600 argsz
= sizeof(*irq_set
) + (vdev
->nr_vectors
* sizeof(*fds
));
602 irq_set
= g_malloc0(argsz
);
603 irq_set
->argsz
= argsz
;
604 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
| VFIO_IRQ_SET_ACTION_TRIGGER
;
605 irq_set
->index
= msix
? VFIO_PCI_MSIX_IRQ_INDEX
: VFIO_PCI_MSI_IRQ_INDEX
;
607 irq_set
->count
= vdev
->nr_vectors
;
608 fds
= (int32_t *)&irq_set
->data
;
610 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
611 if (!vdev
->msi_vectors
[i
].use
) {
616 fds
[i
] = event_notifier_get_fd(&vdev
->msi_vectors
[i
].interrupt
);
619 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
626 static int vfio_msix_vector_do_use(PCIDevice
*pdev
, unsigned int nr
,
627 MSIMessage
*msg
, IOHandler
*handler
)
629 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
630 VFIOMSIVector
*vector
;
633 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d used\n", __func__
,
634 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
635 vdev
->host
.function
, nr
);
637 vector
= &vdev
->msi_vectors
[nr
];
641 msix_vector_use(pdev
, nr
);
643 if (event_notifier_init(&vector
->interrupt
, 0)) {
644 error_report("vfio: Error: event_notifier_init failed");
648 * Attempt to enable route through KVM irqchip,
649 * default to userspace handling if unavailable.
651 vector
->virq
= msg
? kvm_irqchip_add_msi_route(kvm_state
, *msg
) : -1;
652 if (vector
->virq
< 0 ||
653 kvm_irqchip_add_irqfd_notifier(kvm_state
, &vector
->interrupt
,
654 NULL
, vector
->virq
) < 0) {
655 if (vector
->virq
>= 0) {
656 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
659 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
660 handler
, NULL
, vector
);
664 * We don't want to have the host allocate all possible MSI vectors
665 * for a device if they're not in use, so we shutdown and incrementally
666 * increase them as needed.
668 if (vdev
->nr_vectors
< nr
+ 1) {
669 vfio_disable_irqindex(vdev
, VFIO_PCI_MSIX_IRQ_INDEX
);
670 vdev
->nr_vectors
= nr
+ 1;
671 ret
= vfio_enable_vectors(vdev
, true);
673 error_report("vfio: failed to enable vectors, %d", ret
);
677 struct vfio_irq_set
*irq_set
;
680 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
682 irq_set
= g_malloc0(argsz
);
683 irq_set
->argsz
= argsz
;
684 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
685 VFIO_IRQ_SET_ACTION_TRIGGER
;
686 irq_set
->index
= VFIO_PCI_MSIX_IRQ_INDEX
;
689 pfd
= (int32_t *)&irq_set
->data
;
691 *pfd
= event_notifier_get_fd(&vector
->interrupt
);
693 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
696 error_report("vfio: failed to modify vector, %d", ret
);
703 static int vfio_msix_vector_use(PCIDevice
*pdev
,
704 unsigned int nr
, MSIMessage msg
)
706 return vfio_msix_vector_do_use(pdev
, nr
, &msg
, vfio_msi_interrupt
);
709 static void vfio_msix_vector_release(PCIDevice
*pdev
, unsigned int nr
)
711 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
712 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[nr
];
714 struct vfio_irq_set
*irq_set
;
717 DPRINTF("%s(%04x:%02x:%02x.%x) vector %d released\n", __func__
,
718 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
719 vdev
->host
.function
, nr
);
722 * XXX What's the right thing to do here? This turns off the interrupt
723 * completely, but do we really just want to switch the interrupt to
724 * bouncing through userspace and let msix.c drop it? Not sure.
726 msix_vector_unuse(pdev
, nr
);
728 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
730 irq_set
= g_malloc0(argsz
);
731 irq_set
->argsz
= argsz
;
732 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
733 VFIO_IRQ_SET_ACTION_TRIGGER
;
734 irq_set
->index
= VFIO_PCI_MSIX_IRQ_INDEX
;
737 pfd
= (int32_t *)&irq_set
->data
;
741 ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
745 if (vector
->virq
< 0) {
746 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
749 kvm_irqchip_remove_irqfd_notifier(kvm_state
, &vector
->interrupt
,
751 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
755 event_notifier_cleanup(&vector
->interrupt
);
759 static void vfio_enable_msix(VFIODevice
*vdev
)
761 vfio_disable_interrupts(vdev
);
763 vdev
->msi_vectors
= g_malloc0(vdev
->msix
->entries
* sizeof(VFIOMSIVector
));
765 vdev
->interrupt
= VFIO_INT_MSIX
;
768 * Some communication channels between VF & PF or PF & fw rely on the
769 * physical state of the device and expect that enabling MSI-X from the
770 * guest enables the same on the host. When our guest is Linux, the
771 * guest driver call to pci_enable_msix() sets the enabling bit in the
772 * MSI-X capability, but leaves the vector table masked. We therefore
773 * can't rely on a vector_use callback (from request_irq() in the guest)
774 * to switch the physical device into MSI-X mode because that may come a
775 * long time after pci_enable_msix(). This code enables vector 0 with
776 * triggering to userspace, then immediately release the vector, leaving
777 * the physical device with no vectors enabled, but MSI-X enabled, just
778 * like the guest view.
780 vfio_msix_vector_do_use(&vdev
->pdev
, 0, NULL
, NULL
);
781 vfio_msix_vector_release(&vdev
->pdev
, 0);
783 if (msix_set_vector_notifiers(&vdev
->pdev
, vfio_msix_vector_use
,
784 vfio_msix_vector_release
, NULL
)) {
785 error_report("vfio: msix_set_vector_notifiers failed");
788 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
789 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
792 static void vfio_enable_msi(VFIODevice
*vdev
)
796 vfio_disable_interrupts(vdev
);
798 vdev
->nr_vectors
= msi_nr_vectors_allocated(&vdev
->pdev
);
800 vdev
->msi_vectors
= g_malloc0(vdev
->nr_vectors
* sizeof(VFIOMSIVector
));
802 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
803 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
808 if (event_notifier_init(&vector
->interrupt
, 0)) {
809 error_report("vfio: Error: event_notifier_init failed");
812 vector
->msg
= msi_get_message(&vdev
->pdev
, i
);
815 * Attempt to enable route through KVM irqchip,
816 * default to userspace handling if unavailable.
818 vector
->virq
= kvm_irqchip_add_msi_route(kvm_state
, vector
->msg
);
819 if (vector
->virq
< 0 ||
820 kvm_irqchip_add_irqfd_notifier(kvm_state
, &vector
->interrupt
,
821 NULL
, vector
->virq
) < 0) {
822 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
823 vfio_msi_interrupt
, NULL
, vector
);
827 ret
= vfio_enable_vectors(vdev
, false);
830 error_report("vfio: Error: Failed to setup MSI fds: %m");
831 } else if (ret
!= vdev
->nr_vectors
) {
832 error_report("vfio: Error: Failed to enable %d "
833 "MSI vectors, retry with %d", vdev
->nr_vectors
, ret
);
836 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
837 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
838 if (vector
->virq
>= 0) {
839 kvm_irqchip_remove_irqfd_notifier(kvm_state
, &vector
->interrupt
,
841 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
844 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
847 event_notifier_cleanup(&vector
->interrupt
);
850 g_free(vdev
->msi_vectors
);
852 if (ret
> 0 && ret
!= vdev
->nr_vectors
) {
853 vdev
->nr_vectors
= ret
;
856 vdev
->nr_vectors
= 0;
861 vdev
->interrupt
= VFIO_INT_MSI
;
863 DPRINTF("%s(%04x:%02x:%02x.%x) Enabled %d MSI vectors\n", __func__
,
864 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
865 vdev
->host
.function
, vdev
->nr_vectors
);
868 static void vfio_disable_msi_common(VFIODevice
*vdev
)
870 g_free(vdev
->msi_vectors
);
871 vdev
->msi_vectors
= NULL
;
872 vdev
->nr_vectors
= 0;
873 vdev
->interrupt
= VFIO_INT_NONE
;
875 vfio_enable_intx(vdev
);
878 static void vfio_disable_msix(VFIODevice
*vdev
)
880 msix_unset_vector_notifiers(&vdev
->pdev
);
882 if (vdev
->nr_vectors
) {
883 vfio_disable_irqindex(vdev
, VFIO_PCI_MSIX_IRQ_INDEX
);
886 vfio_disable_msi_common(vdev
);
888 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
889 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
892 static void vfio_disable_msi(VFIODevice
*vdev
)
896 vfio_disable_irqindex(vdev
, VFIO_PCI_MSI_IRQ_INDEX
);
898 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
899 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
905 if (vector
->virq
>= 0) {
906 kvm_irqchip_remove_irqfd_notifier(kvm_state
,
907 &vector
->interrupt
, vector
->virq
);
908 kvm_irqchip_release_virq(kvm_state
, vector
->virq
);
911 qemu_set_fd_handler(event_notifier_get_fd(&vector
->interrupt
),
915 event_notifier_cleanup(&vector
->interrupt
);
918 vfio_disable_msi_common(vdev
);
920 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
921 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
924 static void vfio_update_msi(VFIODevice
*vdev
)
928 for (i
= 0; i
< vdev
->nr_vectors
; i
++) {
929 VFIOMSIVector
*vector
= &vdev
->msi_vectors
[i
];
932 if (!vector
->use
|| vector
->virq
< 0) {
936 msg
= msi_get_message(&vdev
->pdev
, i
);
938 if (msg
.address
!= vector
->msg
.address
||
939 msg
.data
!= vector
->msg
.data
) {
941 DPRINTF("%s(%04x:%02x:%02x.%x) MSI vector %d changed\n",
942 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
943 vdev
->host
.slot
, vdev
->host
.function
, i
);
945 kvm_irqchip_update_msi_route(kvm_state
, vector
->virq
, msg
);
952 * IO Port/MMIO - Beware of the endians, VFIO is always little endian
954 static void vfio_bar_write(void *opaque
, hwaddr addr
,
955 uint64_t data
, unsigned size
)
957 VFIOBAR
*bar
= opaque
;
970 buf
.word
= cpu_to_le16(data
);
973 buf
.dword
= cpu_to_le32(data
);
976 hw_error("vfio: unsupported write size, %d bytes\n", size
);
980 if (pwrite(bar
->fd
, &buf
, size
, bar
->fd_offset
+ addr
) != size
) {
981 error_report("%s(,0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d) failed: %m",
982 __func__
, addr
, data
, size
);
987 VFIODevice
*vdev
= container_of(bar
, VFIODevice
, bars
[bar
->nr
]);
989 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
", 0x%"PRIx64
990 ", %d)\n", __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
991 vdev
->host
.slot
, vdev
->host
.function
, bar
->nr
, addr
,
997 * A read or write to a BAR always signals an INTx EOI. This will
998 * do nothing if not pending (including not in INTx mode). We assume
999 * that a BAR access is in response to an interrupt and that BAR
1000 * accesses will service the interrupt. Unfortunately, we don't know
1001 * which access will service the interrupt, so we're potentially
1002 * getting quite a few host interrupts per guest interrupt.
1004 vfio_eoi(container_of(bar
, VFIODevice
, bars
[bar
->nr
]));
1007 static uint64_t vfio_bar_read(void *opaque
,
1008 hwaddr addr
, unsigned size
)
1010 VFIOBAR
*bar
= opaque
;
1019 if (pread(bar
->fd
, &buf
, size
, bar
->fd_offset
+ addr
) != size
) {
1020 error_report("%s(,0x%"HWADDR_PRIx
", %d) failed: %m",
1021 __func__
, addr
, size
);
1022 return (uint64_t)-1;
1030 data
= le16_to_cpu(buf
.word
);
1033 data
= le32_to_cpu(buf
.dword
);
1036 hw_error("vfio: unsupported read size, %d bytes\n", size
);
1042 VFIODevice
*vdev
= container_of(bar
, VFIODevice
, bars
[bar
->nr
]);
1044 DPRINTF("%s(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
1045 ", %d) = 0x%"PRIx64
"\n", __func__
, vdev
->host
.domain
,
1046 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
1047 bar
->nr
, addr
, size
, data
);
1051 /* Same as write above */
1052 vfio_eoi(container_of(bar
, VFIODevice
, bars
[bar
->nr
]));
1057 static const MemoryRegionOps vfio_bar_ops
= {
1058 .read
= vfio_bar_read
,
1059 .write
= vfio_bar_write
,
1060 .endianness
= DEVICE_LITTLE_ENDIAN
,
1063 static void vfio_pci_load_rom(VFIODevice
*vdev
)
1065 struct vfio_region_info reg_info
= {
1066 .argsz
= sizeof(reg_info
),
1067 .index
= VFIO_PCI_ROM_REGION_INDEX
1073 if (ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
)) {
1074 error_report("vfio: Error getting ROM info: %m");
1078 DPRINTF("Device %04x:%02x:%02x.%x ROM:\n", vdev
->host
.domain
,
1079 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
1080 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
1081 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
1082 (unsigned long)reg_info
.flags
);
1084 vdev
->rom_size
= size
= reg_info
.size
;
1085 vdev
->rom_offset
= reg_info
.offset
;
1087 if (!vdev
->rom_size
) {
1091 vdev
->rom
= g_malloc(size
);
1092 memset(vdev
->rom
, 0xff, size
);
1095 bytes
= pread(vdev
->fd
, vdev
->rom
+ off
, size
, vdev
->rom_offset
+ off
);
1098 } else if (bytes
> 0) {
1102 if (errno
== EINTR
|| errno
== EAGAIN
) {
1105 error_report("vfio: Error reading device ROM: %m");
1111 static uint64_t vfio_rom_read(void *opaque
, hwaddr addr
, unsigned size
)
1113 VFIODevice
*vdev
= opaque
;
1114 uint64_t val
= ((uint64_t)1 << (size
* 8)) - 1;
1116 /* Load the ROM lazily when the guest tries to read it */
1117 if (unlikely(!vdev
->rom
)) {
1118 vfio_pci_load_rom(vdev
);
1121 memcpy(&val
, vdev
->rom
+ addr
,
1122 (addr
< vdev
->rom_size
) ? MIN(size
, vdev
->rom_size
- addr
) : 0);
1124 DPRINTF("%s(%04x:%02x:%02x.%x, 0x%"HWADDR_PRIx
", 0x%x) = 0x%"PRIx64
"\n",
1125 __func__
, vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1126 vdev
->host
.function
, addr
, size
, val
);
1131 static void vfio_rom_write(void *opaque
, hwaddr addr
,
1132 uint64_t data
, unsigned size
)
1136 static const MemoryRegionOps vfio_rom_ops
= {
1137 .read
= vfio_rom_read
,
1138 .write
= vfio_rom_write
,
1139 .endianness
= DEVICE_LITTLE_ENDIAN
,
1142 static void vfio_pci_size_rom(VFIODevice
*vdev
)
1144 uint32_t orig
, size
= (uint32_t)PCI_ROM_ADDRESS_MASK
;
1145 off_t offset
= vdev
->config_offset
+ PCI_ROM_ADDRESS
;
1148 if (vdev
->pdev
.romfile
|| !vdev
->pdev
.rom_bar
) {
1153 * Use the same size ROM BAR as the physical device. The contents
1154 * will get filled in later when the guest tries to read it.
1156 if (pread(vdev
->fd
, &orig
, 4, offset
) != 4 ||
1157 pwrite(vdev
->fd
, &size
, 4, offset
) != 4 ||
1158 pread(vdev
->fd
, &size
, 4, offset
) != 4 ||
1159 pwrite(vdev
->fd
, &orig
, 4, offset
) != 4) {
1160 error_report("%s(%04x:%02x:%02x.%x) failed: %m",
1161 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
1162 vdev
->host
.slot
, vdev
->host
.function
);
1166 size
= ~(size
& PCI_ROM_ADDRESS_MASK
) + 1;
1172 DPRINTF("%04x:%02x:%02x.%x ROM size 0x%x\n", vdev
->host
.domain
,
1173 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
, size
);
1175 snprintf(name
, sizeof(name
), "vfio[%04x:%02x:%02x.%x].rom",
1176 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1177 vdev
->host
.function
);
1179 memory_region_init_io(&vdev
->pdev
.rom
, OBJECT(vdev
),
1180 &vfio_rom_ops
, vdev
, name
, size
);
1182 pci_register_bar(&vdev
->pdev
, PCI_ROM_SLOT
,
1183 PCI_BASE_ADDRESS_SPACE_MEMORY
, &vdev
->pdev
.rom
);
1185 vdev
->pdev
.has_rom
= true;
1188 static void vfio_vga_write(void *opaque
, hwaddr addr
,
1189 uint64_t data
, unsigned size
)
1191 VFIOVGARegion
*region
= opaque
;
1192 VFIOVGA
*vga
= container_of(region
, VFIOVGA
, region
[region
->nr
]);
1199 off_t offset
= vga
->fd_offset
+ region
->offset
+ addr
;
1206 buf
.word
= cpu_to_le16(data
);
1209 buf
.dword
= cpu_to_le32(data
);
1212 hw_error("vfio: unsupported write size, %d bytes\n", size
);
1216 if (pwrite(vga
->fd
, &buf
, size
, offset
) != size
) {
1217 error_report("%s(,0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d) failed: %m",
1218 __func__
, region
->offset
+ addr
, data
, size
);
1221 DPRINTF("%s(0x%"HWADDR_PRIx
", 0x%"PRIx64
", %d)\n",
1222 __func__
, region
->offset
+ addr
, data
, size
);
1225 static uint64_t vfio_vga_read(void *opaque
, hwaddr addr
, unsigned size
)
1227 VFIOVGARegion
*region
= opaque
;
1228 VFIOVGA
*vga
= container_of(region
, VFIOVGA
, region
[region
->nr
]);
1236 off_t offset
= vga
->fd_offset
+ region
->offset
+ addr
;
1238 if (pread(vga
->fd
, &buf
, size
, offset
) != size
) {
1239 error_report("%s(,0x%"HWADDR_PRIx
", %d) failed: %m",
1240 __func__
, region
->offset
+ addr
, size
);
1241 return (uint64_t)-1;
1249 data
= le16_to_cpu(buf
.word
);
1252 data
= le32_to_cpu(buf
.dword
);
1255 hw_error("vfio: unsupported read size, %d bytes\n", size
);
1259 DPRINTF("%s(0x%"HWADDR_PRIx
", %d) = 0x%"PRIx64
"\n",
1260 __func__
, region
->offset
+ addr
, size
, data
);
1265 static const MemoryRegionOps vfio_vga_ops
= {
1266 .read
= vfio_vga_read
,
1267 .write
= vfio_vga_write
,
1268 .endianness
= DEVICE_LITTLE_ENDIAN
,
1272 * Device specific quirks
1275 /* Is range1 fully contained within range2? */
1276 static bool vfio_range_contained(uint64_t first1
, uint64_t len1
,
1277 uint64_t first2
, uint64_t len2
) {
1278 return (first1
>= first2
&& first1
+ len1
<= first2
+ len2
);
1281 static bool vfio_flags_enabled(uint8_t flags
, uint8_t mask
)
1283 return (mask
&& (flags
& mask
) == mask
);
1286 static uint64_t vfio_generic_window_quirk_read(void *opaque
,
1287 hwaddr addr
, unsigned size
)
1289 VFIOQuirk
*quirk
= opaque
;
1290 VFIODevice
*vdev
= quirk
->vdev
;
1293 if (vfio_flags_enabled(quirk
->data
.flags
, quirk
->data
.read_flags
) &&
1294 ranges_overlap(addr
, size
,
1295 quirk
->data
.data_offset
, quirk
->data
.data_size
)) {
1296 hwaddr offset
= addr
- quirk
->data
.data_offset
;
1298 if (!vfio_range_contained(addr
, size
, quirk
->data
.data_offset
,
1299 quirk
->data
.data_size
)) {
1300 hw_error("%s: window data read not fully contained: %s\n",
1301 __func__
, memory_region_name(&quirk
->mem
));
1304 data
= vfio_pci_read_config(&vdev
->pdev
,
1305 quirk
->data
.address_val
+ offset
, size
);
1307 DPRINTF("%s read(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
", %d) = 0x%"
1308 PRIx64
"\n", memory_region_name(&quirk
->mem
), vdev
->host
.domain
,
1309 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
1310 quirk
->data
.bar
, addr
, size
, data
);
1312 data
= vfio_bar_read(&vdev
->bars
[quirk
->data
.bar
],
1313 addr
+ quirk
->data
.base_offset
, size
);
1319 static void vfio_generic_window_quirk_write(void *opaque
, hwaddr addr
,
1320 uint64_t data
, unsigned size
)
1322 VFIOQuirk
*quirk
= opaque
;
1323 VFIODevice
*vdev
= quirk
->vdev
;
1325 if (ranges_overlap(addr
, size
,
1326 quirk
->data
.address_offset
, quirk
->data
.address_size
)) {
1328 if (addr
!= quirk
->data
.address_offset
) {
1329 hw_error("%s: offset write into address window: %s\n",
1330 __func__
, memory_region_name(&quirk
->mem
));
1333 if ((data
& ~quirk
->data
.address_mask
) == quirk
->data
.address_match
) {
1334 quirk
->data
.flags
|= quirk
->data
.write_flags
|
1335 quirk
->data
.read_flags
;
1336 quirk
->data
.address_val
= data
& quirk
->data
.address_mask
;
1338 quirk
->data
.flags
&= ~(quirk
->data
.write_flags
|
1339 quirk
->data
.read_flags
);
1343 if (vfio_flags_enabled(quirk
->data
.flags
, quirk
->data
.write_flags
) &&
1344 ranges_overlap(addr
, size
,
1345 quirk
->data
.data_offset
, quirk
->data
.data_size
)) {
1346 hwaddr offset
= addr
- quirk
->data
.data_offset
;
1348 if (!vfio_range_contained(addr
, size
, quirk
->data
.data_offset
,
1349 quirk
->data
.data_size
)) {
1350 hw_error("%s: window data write not fully contained: %s\n",
1351 __func__
, memory_region_name(&quirk
->mem
));
1354 vfio_pci_write_config(&vdev
->pdev
,
1355 quirk
->data
.address_val
+ offset
, data
, size
);
1356 DPRINTF("%s write(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
", 0x%"
1357 PRIx64
", %d)\n", memory_region_name(&quirk
->mem
),
1358 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1359 vdev
->host
.function
, quirk
->data
.bar
, addr
, data
, size
);
1363 vfio_bar_write(&vdev
->bars
[quirk
->data
.bar
],
1364 addr
+ quirk
->data
.base_offset
, data
, size
);
1367 static const MemoryRegionOps vfio_generic_window_quirk
= {
1368 .read
= vfio_generic_window_quirk_read
,
1369 .write
= vfio_generic_window_quirk_write
,
1370 .endianness
= DEVICE_LITTLE_ENDIAN
,
1373 static uint64_t vfio_generic_quirk_read(void *opaque
,
1374 hwaddr addr
, unsigned size
)
1376 VFIOQuirk
*quirk
= opaque
;
1377 VFIODevice
*vdev
= quirk
->vdev
;
1378 hwaddr base
= quirk
->data
.address_match
& TARGET_PAGE_MASK
;
1379 hwaddr offset
= quirk
->data
.address_match
& ~TARGET_PAGE_MASK
;
1382 if (vfio_flags_enabled(quirk
->data
.flags
, quirk
->data
.read_flags
) &&
1383 ranges_overlap(addr
, size
, offset
, quirk
->data
.address_mask
+ 1)) {
1384 if (!vfio_range_contained(addr
, size
, offset
,
1385 quirk
->data
.address_mask
+ 1)) {
1386 hw_error("%s: read not fully contained: %s\n",
1387 __func__
, memory_region_name(&quirk
->mem
));
1390 data
= vfio_pci_read_config(&vdev
->pdev
, addr
- offset
, size
);
1392 DPRINTF("%s read(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
", %d) = 0x%"
1393 PRIx64
"\n", memory_region_name(&quirk
->mem
), vdev
->host
.domain
,
1394 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
1395 quirk
->data
.bar
, addr
+ base
, size
, data
);
1397 data
= vfio_bar_read(&vdev
->bars
[quirk
->data
.bar
], addr
+ base
, size
);
1403 static void vfio_generic_quirk_write(void *opaque
, hwaddr addr
,
1404 uint64_t data
, unsigned size
)
1406 VFIOQuirk
*quirk
= opaque
;
1407 VFIODevice
*vdev
= quirk
->vdev
;
1408 hwaddr base
= quirk
->data
.address_match
& TARGET_PAGE_MASK
;
1409 hwaddr offset
= quirk
->data
.address_match
& ~TARGET_PAGE_MASK
;
1411 if (vfio_flags_enabled(quirk
->data
.flags
, quirk
->data
.write_flags
) &&
1412 ranges_overlap(addr
, size
, offset
, quirk
->data
.address_mask
+ 1)) {
1413 if (!vfio_range_contained(addr
, size
, offset
,
1414 quirk
->data
.address_mask
+ 1)) {
1415 hw_error("%s: write not fully contained: %s\n",
1416 __func__
, memory_region_name(&quirk
->mem
));
1419 vfio_pci_write_config(&vdev
->pdev
, addr
- offset
, data
, size
);
1421 DPRINTF("%s write(%04x:%02x:%02x.%x:BAR%d+0x%"HWADDR_PRIx
", 0x%"
1422 PRIx64
", %d)\n", memory_region_name(&quirk
->mem
),
1423 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1424 vdev
->host
.function
, quirk
->data
.bar
, addr
+ base
, data
, size
);
1426 vfio_bar_write(&vdev
->bars
[quirk
->data
.bar
], addr
+ base
, data
, size
);
1430 static const MemoryRegionOps vfio_generic_quirk
= {
1431 .read
= vfio_generic_quirk_read
,
1432 .write
= vfio_generic_quirk_write
,
1433 .endianness
= DEVICE_LITTLE_ENDIAN
,
1436 #define PCI_VENDOR_ID_ATI 0x1002
1439 * Radeon HD cards (HD5450 & HD7850) report the upper byte of the I/O port BAR
1440 * through VGA register 0x3c3. On newer cards, the I/O port BAR is always
1441 * BAR4 (older cards like the X550 used BAR1, but we don't care to support
1442 * those). Note that on bare metal, a read of 0x3c3 doesn't always return the
1443 * I/O port BAR address. Originally this was coded to return the virtual BAR
1444 * address only if the physical register read returns the actual BAR address,
1445 * but users have reported greater success if we return the virtual address
1448 static uint64_t vfio_ati_3c3_quirk_read(void *opaque
,
1449 hwaddr addr
, unsigned size
)
1451 VFIOQuirk
*quirk
= opaque
;
1452 VFIODevice
*vdev
= quirk
->vdev
;
1453 uint64_t data
= vfio_pci_read_config(&vdev
->pdev
,
1454 PCI_BASE_ADDRESS_0
+ (4 * 4) + 1,
1456 DPRINTF("%s(0x3c3, 1) = 0x%"PRIx64
"\n", __func__
, data
);
1461 static const MemoryRegionOps vfio_ati_3c3_quirk
= {
1462 .read
= vfio_ati_3c3_quirk_read
,
1463 .endianness
= DEVICE_LITTLE_ENDIAN
,
1466 static void vfio_vga_probe_ati_3c3_quirk(VFIODevice
*vdev
)
1468 PCIDevice
*pdev
= &vdev
->pdev
;
1471 if (pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_ATI
) {
1476 * As long as the BAR is >= 256 bytes it will be aligned such that the
1477 * lower byte is always zero. Filter out anything else, if it exists.
1479 if (!vdev
->bars
[4].ioport
|| vdev
->bars
[4].size
< 256) {
1483 quirk
= g_malloc0(sizeof(*quirk
));
1486 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
), &vfio_ati_3c3_quirk
, quirk
,
1487 "vfio-ati-3c3-quirk", 1);
1488 memory_region_add_subregion(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].mem
,
1489 3 /* offset 3 bytes from 0x3c0 */, &quirk
->mem
);
1491 QLIST_INSERT_HEAD(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].quirks
,
1494 DPRINTF("Enabled ATI/AMD quirk 0x3c3 BAR4for device %04x:%02x:%02x.%x\n",
1495 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1496 vdev
->host
.function
);
1500 * Newer ATI/AMD devices, including HD5450 and HD7850, have a window to PCI
1501 * config space through MMIO BAR2 at offset 0x4000. Nothing seems to access
1502 * the MMIO space directly, but a window to this space is provided through
1503 * I/O port BAR4. Offset 0x0 is the address register and offset 0x4 is the
1504 * data register. When the address is programmed to a range of 0x4000-0x4fff
1505 * PCI configuration space is available. Experimentation seems to indicate
1506 * that only read-only access is provided, but we drop writes when the window
1507 * is enabled to config space nonetheless.
1509 static void vfio_probe_ati_bar4_window_quirk(VFIODevice
*vdev
, int nr
)
1511 PCIDevice
*pdev
= &vdev
->pdev
;
1514 if (!vdev
->has_vga
|| nr
!= 4 ||
1515 pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_ATI
) {
1519 quirk
= g_malloc0(sizeof(*quirk
));
1521 quirk
->data
.address_size
= 4;
1522 quirk
->data
.data_offset
= 4;
1523 quirk
->data
.data_size
= 4;
1524 quirk
->data
.address_match
= 0x4000;
1525 quirk
->data
.address_mask
= PCIE_CONFIG_SPACE_SIZE
- 1;
1526 quirk
->data
.bar
= nr
;
1527 quirk
->data
.read_flags
= quirk
->data
.write_flags
= 1;
1529 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
),
1530 &vfio_generic_window_quirk
, quirk
,
1531 "vfio-ati-bar4-window-quirk", 8);
1532 memory_region_add_subregion_overlap(&vdev
->bars
[nr
].mem
,
1533 quirk
->data
.base_offset
, &quirk
->mem
, 1);
1535 QLIST_INSERT_HEAD(&vdev
->bars
[nr
].quirks
, quirk
, next
);
1537 DPRINTF("Enabled ATI/AMD BAR4 window quirk for device %04x:%02x:%02x.%x\n",
1538 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1539 vdev
->host
.function
);
1543 * Trap the BAR2 MMIO window to config space as well.
1545 static void vfio_probe_ati_bar2_4000_quirk(VFIODevice
*vdev
, int nr
)
1547 PCIDevice
*pdev
= &vdev
->pdev
;
1550 /* Only enable on newer devices where BAR2 is 64bit */
1551 if (!vdev
->has_vga
|| nr
!= 2 || !vdev
->bars
[2].mem64
||
1552 pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_ATI
) {
1556 quirk
= g_malloc0(sizeof(*quirk
));
1558 quirk
->data
.flags
= quirk
->data
.read_flags
= quirk
->data
.write_flags
= 1;
1559 quirk
->data
.address_match
= 0x4000;
1560 quirk
->data
.address_mask
= PCIE_CONFIG_SPACE_SIZE
- 1;
1561 quirk
->data
.bar
= nr
;
1563 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
), &vfio_generic_quirk
, quirk
,
1564 "vfio-ati-bar2-4000-quirk",
1565 TARGET_PAGE_ALIGN(quirk
->data
.address_mask
+ 1));
1566 memory_region_add_subregion_overlap(&vdev
->bars
[nr
].mem
,
1567 quirk
->data
.address_match
& TARGET_PAGE_MASK
,
1570 QLIST_INSERT_HEAD(&vdev
->bars
[nr
].quirks
, quirk
, next
);
1572 DPRINTF("Enabled ATI/AMD BAR2 0x4000 quirk for device %04x:%02x:%02x.%x\n",
1573 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1574 vdev
->host
.function
);
1578 * Older ATI/AMD cards like the X550 have a similar window to that above.
1579 * I/O port BAR1 provides a window to a mirror of PCI config space located
1580 * in BAR2 at offset 0xf00. We don't care to support such older cards, but
1581 * note it for future reference.
1584 #define PCI_VENDOR_ID_NVIDIA 0x10de
1587 * Nvidia has several different methods to get to config space, the
1588 * nouveu project has several of these documented here:
1589 * https://github.com/pathscale/envytools/tree/master/hwdocs
1591 * The first quirk is actually not documented in envytools and is found
1592 * on 10de:01d1 (NVIDIA Corporation G72 [GeForce 7300 LE]). This is an
1593 * NV46 chipset. The backdoor uses the legacy VGA I/O ports to access
1594 * the mirror of PCI config space found at BAR0 offset 0x1800. The access
1595 * sequence first writes 0x338 to I/O port 0x3d4. The target offset is
1596 * then written to 0x3d0. Finally 0x538 is written for a read and 0x738
1597 * is written for a write to 0x3d4. The BAR0 offset is then accessible
1598 * through 0x3d0. This quirk doesn't seem to be necessary on newer cards
1599 * that use the I/O port BAR5 window but it doesn't hurt to leave it.
1609 static uint64_t vfio_nvidia_3d0_quirk_read(void *opaque
,
1610 hwaddr addr
, unsigned size
)
1612 VFIOQuirk
*quirk
= opaque
;
1613 VFIODevice
*vdev
= quirk
->vdev
;
1614 PCIDevice
*pdev
= &vdev
->pdev
;
1615 uint64_t data
= vfio_vga_read(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
],
1616 addr
+ quirk
->data
.base_offset
, size
);
1618 if (quirk
->data
.flags
== NV_3D0_READ
&& addr
== quirk
->data
.data_offset
) {
1619 data
= vfio_pci_read_config(pdev
, quirk
->data
.address_val
, size
);
1620 DPRINTF("%s(0x3d0, %d) = 0x%"PRIx64
"\n", __func__
, size
, data
);
1623 quirk
->data
.flags
= NV_3D0_NONE
;
1628 static void vfio_nvidia_3d0_quirk_write(void *opaque
, hwaddr addr
,
1629 uint64_t data
, unsigned size
)
1631 VFIOQuirk
*quirk
= opaque
;
1632 VFIODevice
*vdev
= quirk
->vdev
;
1633 PCIDevice
*pdev
= &vdev
->pdev
;
1635 switch (quirk
->data
.flags
) {
1637 if (addr
== quirk
->data
.address_offset
&& data
== 0x338) {
1638 quirk
->data
.flags
= NV_3D0_SELECT
;
1642 quirk
->data
.flags
= NV_3D0_NONE
;
1643 if (addr
== quirk
->data
.data_offset
&&
1644 (data
& ~quirk
->data
.address_mask
) == quirk
->data
.address_match
) {
1645 quirk
->data
.flags
= NV_3D0_WINDOW
;
1646 quirk
->data
.address_val
= data
& quirk
->data
.address_mask
;
1650 quirk
->data
.flags
= NV_3D0_NONE
;
1651 if (addr
== quirk
->data
.address_offset
) {
1652 if (data
== 0x538) {
1653 quirk
->data
.flags
= NV_3D0_READ
;
1654 } else if (data
== 0x738) {
1655 quirk
->data
.flags
= NV_3D0_WRITE
;
1660 quirk
->data
.flags
= NV_3D0_NONE
;
1661 if (addr
== quirk
->data
.data_offset
) {
1662 vfio_pci_write_config(pdev
, quirk
->data
.address_val
, data
, size
);
1663 DPRINTF("%s(0x3d0, 0x%"PRIx64
", %d)\n", __func__
, data
, size
);
1669 vfio_vga_write(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
],
1670 addr
+ quirk
->data
.base_offset
, data
, size
);
1673 static const MemoryRegionOps vfio_nvidia_3d0_quirk
= {
1674 .read
= vfio_nvidia_3d0_quirk_read
,
1675 .write
= vfio_nvidia_3d0_quirk_write
,
1676 .endianness
= DEVICE_LITTLE_ENDIAN
,
1679 static void vfio_vga_probe_nvidia_3d0_quirk(VFIODevice
*vdev
)
1681 PCIDevice
*pdev
= &vdev
->pdev
;
1684 if (pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_NVIDIA
||
1685 !vdev
->bars
[1].size
) {
1689 quirk
= g_malloc0(sizeof(*quirk
));
1691 quirk
->data
.base_offset
= 0x10;
1692 quirk
->data
.address_offset
= 4;
1693 quirk
->data
.address_size
= 2;
1694 quirk
->data
.address_match
= 0x1800;
1695 quirk
->data
.address_mask
= PCI_CONFIG_SPACE_SIZE
- 1;
1696 quirk
->data
.data_offset
= 0;
1697 quirk
->data
.data_size
= 4;
1699 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
), &vfio_nvidia_3d0_quirk
,
1700 quirk
, "vfio-nvidia-3d0-quirk", 6);
1701 memory_region_add_subregion(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].mem
,
1702 quirk
->data
.base_offset
, &quirk
->mem
);
1704 QLIST_INSERT_HEAD(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].quirks
,
1707 DPRINTF("Enabled NVIDIA VGA 0x3d0 quirk for device %04x:%02x:%02x.%x\n",
1708 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1709 vdev
->host
.function
);
1713 * The second quirk is documented in envytools. The I/O port BAR5 is just
1714 * a set of address/data ports to the MMIO BARs. The BAR we care about is
1715 * again BAR0. This backdoor is apparently a bit newer than the one above
1716 * so we need to not only trap 256 bytes @0x1800, but all of PCI config
1717 * space, including extended space is available at the 4k @0x88000.
1720 NV_BAR5_ADDRESS
= 0x1,
1721 NV_BAR5_ENABLE
= 0x2,
1722 NV_BAR5_MASTER
= 0x4,
1723 NV_BAR5_VALID
= 0x7,
1726 static void vfio_nvidia_bar5_window_quirk_write(void *opaque
, hwaddr addr
,
1727 uint64_t data
, unsigned size
)
1729 VFIOQuirk
*quirk
= opaque
;
1734 quirk
->data
.flags
|= NV_BAR5_MASTER
;
1736 quirk
->data
.flags
&= ~NV_BAR5_MASTER
;
1741 quirk
->data
.flags
|= NV_BAR5_ENABLE
;
1743 quirk
->data
.flags
&= ~NV_BAR5_ENABLE
;
1747 if (quirk
->data
.flags
& NV_BAR5_MASTER
) {
1748 if ((data
& ~0xfff) == 0x88000) {
1749 quirk
->data
.flags
|= NV_BAR5_ADDRESS
;
1750 quirk
->data
.address_val
= data
& 0xfff;
1751 } else if ((data
& ~0xff) == 0x1800) {
1752 quirk
->data
.flags
|= NV_BAR5_ADDRESS
;
1753 quirk
->data
.address_val
= data
& 0xff;
1755 quirk
->data
.flags
&= ~NV_BAR5_ADDRESS
;
1761 vfio_generic_window_quirk_write(opaque
, addr
, data
, size
);
1764 static const MemoryRegionOps vfio_nvidia_bar5_window_quirk
= {
1765 .read
= vfio_generic_window_quirk_read
,
1766 .write
= vfio_nvidia_bar5_window_quirk_write
,
1767 .valid
.min_access_size
= 4,
1768 .endianness
= DEVICE_LITTLE_ENDIAN
,
1771 static void vfio_probe_nvidia_bar5_window_quirk(VFIODevice
*vdev
, int nr
)
1773 PCIDevice
*pdev
= &vdev
->pdev
;
1776 if (!vdev
->has_vga
|| nr
!= 5 ||
1777 pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_NVIDIA
) {
1781 quirk
= g_malloc0(sizeof(*quirk
));
1783 quirk
->data
.read_flags
= quirk
->data
.write_flags
= NV_BAR5_VALID
;
1784 quirk
->data
.address_offset
= 0x8;
1785 quirk
->data
.address_size
= 0; /* actually 4, but avoids generic code */
1786 quirk
->data
.data_offset
= 0xc;
1787 quirk
->data
.data_size
= 4;
1788 quirk
->data
.bar
= nr
;
1790 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
),
1791 &vfio_nvidia_bar5_window_quirk
, quirk
,
1792 "vfio-nvidia-bar5-window-quirk", 16);
1793 memory_region_add_subregion_overlap(&vdev
->bars
[nr
].mem
, 0, &quirk
->mem
, 1);
1795 QLIST_INSERT_HEAD(&vdev
->bars
[nr
].quirks
, quirk
, next
);
1797 DPRINTF("Enabled NVIDIA BAR5 window quirk for device %04x:%02x:%02x.%x\n",
1798 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1799 vdev
->host
.function
);
1803 * Finally, BAR0 itself. We want to redirect any accesses to either
1804 * 0x1800 or 0x88000 through the PCI config space access functions.
1806 * NB - quirk at a page granularity or else they don't seem to work when
1809 * Here's offset 0x88000...
1811 static void vfio_probe_nvidia_bar0_88000_quirk(VFIODevice
*vdev
, int nr
)
1813 PCIDevice
*pdev
= &vdev
->pdev
;
1816 if (!vdev
->has_vga
|| nr
!= 0 ||
1817 pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_NVIDIA
) {
1821 quirk
= g_malloc0(sizeof(*quirk
));
1823 quirk
->data
.flags
= quirk
->data
.read_flags
= quirk
->data
.write_flags
= 1;
1824 quirk
->data
.address_match
= 0x88000;
1825 quirk
->data
.address_mask
= PCIE_CONFIG_SPACE_SIZE
- 1;
1826 quirk
->data
.bar
= nr
;
1828 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
), &vfio_generic_quirk
,
1829 quirk
, "vfio-nvidia-bar0-88000-quirk",
1830 TARGET_PAGE_ALIGN(quirk
->data
.address_mask
+ 1));
1831 memory_region_add_subregion_overlap(&vdev
->bars
[nr
].mem
,
1832 quirk
->data
.address_match
& TARGET_PAGE_MASK
,
1835 QLIST_INSERT_HEAD(&vdev
->bars
[nr
].quirks
, quirk
, next
);
1837 DPRINTF("Enabled NVIDIA BAR0 0x88000 quirk for device %04x:%02x:%02x.%x\n",
1838 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1839 vdev
->host
.function
);
1843 * And here's the same for BAR0 offset 0x1800...
1845 static void vfio_probe_nvidia_bar0_1800_quirk(VFIODevice
*vdev
, int nr
)
1847 PCIDevice
*pdev
= &vdev
->pdev
;
1850 if (!vdev
->has_vga
|| nr
!= 0 ||
1851 pci_get_word(pdev
->config
+ PCI_VENDOR_ID
) != PCI_VENDOR_ID_NVIDIA
) {
1855 /* Log the chipset ID */
1856 DPRINTF("Nvidia NV%02x\n",
1857 (unsigned int)(vfio_bar_read(&vdev
->bars
[0], 0, 4) >> 20) & 0xff);
1859 quirk
= g_malloc0(sizeof(*quirk
));
1861 quirk
->data
.flags
= quirk
->data
.read_flags
= quirk
->data
.write_flags
= 1;
1862 quirk
->data
.address_match
= 0x1800;
1863 quirk
->data
.address_mask
= PCI_CONFIG_SPACE_SIZE
- 1;
1864 quirk
->data
.bar
= nr
;
1866 memory_region_init_io(&quirk
->mem
, OBJECT(vdev
), &vfio_generic_quirk
, quirk
,
1867 "vfio-nvidia-bar0-1800-quirk",
1868 TARGET_PAGE_ALIGN(quirk
->data
.address_mask
+ 1));
1869 memory_region_add_subregion_overlap(&vdev
->bars
[nr
].mem
,
1870 quirk
->data
.address_match
& TARGET_PAGE_MASK
,
1873 QLIST_INSERT_HEAD(&vdev
->bars
[nr
].quirks
, quirk
, next
);
1875 DPRINTF("Enabled NVIDIA BAR0 0x1800 quirk for device %04x:%02x:%02x.%x\n",
1876 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1877 vdev
->host
.function
);
1881 * TODO - Some Nvidia devices provide config access to their companion HDA
1882 * device and even to their parent bridge via these config space mirrors.
1883 * Add quirks for those regions.
1887 * Common quirk probe entry points.
1889 static void vfio_vga_quirk_setup(VFIODevice
*vdev
)
1891 vfio_vga_probe_ati_3c3_quirk(vdev
);
1892 vfio_vga_probe_nvidia_3d0_quirk(vdev
);
1895 static void vfio_vga_quirk_teardown(VFIODevice
*vdev
)
1899 for (i
= 0; i
< ARRAY_SIZE(vdev
->vga
.region
); i
++) {
1900 while (!QLIST_EMPTY(&vdev
->vga
.region
[i
].quirks
)) {
1901 VFIOQuirk
*quirk
= QLIST_FIRST(&vdev
->vga
.region
[i
].quirks
);
1902 memory_region_del_subregion(&vdev
->vga
.region
[i
].mem
, &quirk
->mem
);
1903 QLIST_REMOVE(quirk
, next
);
1909 static void vfio_bar_quirk_setup(VFIODevice
*vdev
, int nr
)
1911 vfio_probe_ati_bar4_window_quirk(vdev
, nr
);
1912 vfio_probe_ati_bar2_4000_quirk(vdev
, nr
);
1913 vfio_probe_nvidia_bar5_window_quirk(vdev
, nr
);
1914 vfio_probe_nvidia_bar0_88000_quirk(vdev
, nr
);
1915 vfio_probe_nvidia_bar0_1800_quirk(vdev
, nr
);
1918 static void vfio_bar_quirk_teardown(VFIODevice
*vdev
, int nr
)
1920 VFIOBAR
*bar
= &vdev
->bars
[nr
];
1922 while (!QLIST_EMPTY(&bar
->quirks
)) {
1923 VFIOQuirk
*quirk
= QLIST_FIRST(&bar
->quirks
);
1924 memory_region_del_subregion(&bar
->mem
, &quirk
->mem
);
1925 QLIST_REMOVE(quirk
, next
);
1933 static uint32_t vfio_pci_read_config(PCIDevice
*pdev
, uint32_t addr
, int len
)
1935 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
1936 uint32_t emu_bits
= 0, emu_val
= 0, phys_val
= 0, val
;
1938 memcpy(&emu_bits
, vdev
->emulated_config_bits
+ addr
, len
);
1939 emu_bits
= le32_to_cpu(emu_bits
);
1942 emu_val
= pci_default_read_config(pdev
, addr
, len
);
1945 if (~emu_bits
& (0xffffffffU
>> (32 - len
* 8))) {
1948 ret
= pread(vdev
->fd
, &phys_val
, len
, vdev
->config_offset
+ addr
);
1950 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x) failed: %m",
1951 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
1952 vdev
->host
.slot
, vdev
->host
.function
, addr
, len
);
1955 phys_val
= le32_to_cpu(phys_val
);
1958 val
= (emu_val
& emu_bits
) | (phys_val
& ~emu_bits
);
1960 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, len=0x%x) %x\n", __func__
,
1961 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1962 vdev
->host
.function
, addr
, len
, val
);
1967 static void vfio_pci_write_config(PCIDevice
*pdev
, uint32_t addr
,
1968 uint32_t val
, int len
)
1970 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
1971 uint32_t val_le
= cpu_to_le32(val
);
1973 DPRINTF("%s(%04x:%02x:%02x.%x, @0x%x, 0x%x, len=0x%x)\n", __func__
,
1974 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
1975 vdev
->host
.function
, addr
, val
, len
);
1977 /* Write everything to VFIO, let it filter out what we can't write */
1978 if (pwrite(vdev
->fd
, &val_le
, len
, vdev
->config_offset
+ addr
) != len
) {
1979 error_report("%s(%04x:%02x:%02x.%x, 0x%x, 0x%x, 0x%x) failed: %m",
1980 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
1981 vdev
->host
.slot
, vdev
->host
.function
, addr
, val
, len
);
1984 /* MSI/MSI-X Enabling/Disabling */
1985 if (pdev
->cap_present
& QEMU_PCI_CAP_MSI
&&
1986 ranges_overlap(addr
, len
, pdev
->msi_cap
, vdev
->msi_cap_size
)) {
1987 int is_enabled
, was_enabled
= msi_enabled(pdev
);
1989 pci_default_write_config(pdev
, addr
, val
, len
);
1991 is_enabled
= msi_enabled(pdev
);
1995 vfio_enable_msi(vdev
);
1999 vfio_disable_msi(vdev
);
2001 vfio_update_msi(vdev
);
2004 } else if (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
&&
2005 ranges_overlap(addr
, len
, pdev
->msix_cap
, MSIX_CAP_LENGTH
)) {
2006 int is_enabled
, was_enabled
= msix_enabled(pdev
);
2008 pci_default_write_config(pdev
, addr
, val
, len
);
2010 is_enabled
= msix_enabled(pdev
);
2012 if (!was_enabled
&& is_enabled
) {
2013 vfio_enable_msix(vdev
);
2014 } else if (was_enabled
&& !is_enabled
) {
2015 vfio_disable_msix(vdev
);
2018 /* Write everything to QEMU to keep emulated bits correct */
2019 pci_default_write_config(pdev
, addr
, val
, len
);
2024 * DMA - Mapping and unmapping for the "type1" IOMMU interface used on x86
2026 static int vfio_dma_unmap(VFIOContainer
*container
,
2027 hwaddr iova
, ram_addr_t size
)
2029 struct vfio_iommu_type1_dma_unmap unmap
= {
2030 .argsz
= sizeof(unmap
),
2036 if (ioctl(container
->fd
, VFIO_IOMMU_UNMAP_DMA
, &unmap
)) {
2037 DPRINTF("VFIO_UNMAP_DMA: %d\n", -errno
);
2044 static int vfio_dma_map(VFIOContainer
*container
, hwaddr iova
,
2045 ram_addr_t size
, void *vaddr
, bool readonly
)
2047 struct vfio_iommu_type1_dma_map map
= {
2048 .argsz
= sizeof(map
),
2049 .flags
= VFIO_DMA_MAP_FLAG_READ
,
2050 .vaddr
= (__u64
)(uintptr_t)vaddr
,
2056 map
.flags
|= VFIO_DMA_MAP_FLAG_WRITE
;
2060 * Try the mapping, if it fails with EBUSY, unmap the region and try
2061 * again. This shouldn't be necessary, but we sometimes see it in
2062 * the the VGA ROM space.
2064 if (ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0 ||
2065 (errno
== EBUSY
&& vfio_dma_unmap(container
, iova
, size
) == 0 &&
2066 ioctl(container
->fd
, VFIO_IOMMU_MAP_DMA
, &map
) == 0)) {
2070 DPRINTF("VFIO_MAP_DMA: %d\n", -errno
);
2074 static bool vfio_listener_skipped_section(MemoryRegionSection
*section
)
2076 return !memory_region_is_ram(section
->mr
);
2079 static void vfio_listener_region_add(MemoryListener
*listener
,
2080 MemoryRegionSection
*section
)
2082 VFIOContainer
*container
= container_of(listener
, VFIOContainer
,
2083 iommu_data
.listener
);
2088 assert(!memory_region_is_iommu(section
->mr
));
2090 if (vfio_listener_skipped_section(section
)) {
2091 DPRINTF("SKIPPING region_add %"HWADDR_PRIx
" - %"PRIx64
"\n",
2092 section
->offset_within_address_space
,
2093 section
->offset_within_address_space
+
2094 int128_get64(int128_sub(section
->size
, int128_one())));
2098 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
2099 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
2100 error_report("%s received unaligned region", __func__
);
2104 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
2105 end
= (section
->offset_within_address_space
+ int128_get64(section
->size
)) &
2112 vaddr
= memory_region_get_ram_ptr(section
->mr
) +
2113 section
->offset_within_region
+
2114 (iova
- section
->offset_within_address_space
);
2116 DPRINTF("region_add %"HWADDR_PRIx
" - %"HWADDR_PRIx
" [%p]\n",
2117 iova
, end
- 1, vaddr
);
2119 memory_region_ref(section
->mr
);
2120 ret
= vfio_dma_map(container
, iova
, end
- iova
, vaddr
, section
->readonly
);
2122 error_report("vfio_dma_map(%p, 0x%"HWADDR_PRIx
", "
2123 "0x%"HWADDR_PRIx
", %p) = %d (%m)",
2124 container
, iova
, end
- iova
, vaddr
, ret
);
2128 static void vfio_listener_region_del(MemoryListener
*listener
,
2129 MemoryRegionSection
*section
)
2131 VFIOContainer
*container
= container_of(listener
, VFIOContainer
,
2132 iommu_data
.listener
);
2136 if (vfio_listener_skipped_section(section
)) {
2137 DPRINTF("SKIPPING region_del %"HWADDR_PRIx
" - %"PRIx64
"\n",
2138 section
->offset_within_address_space
,
2139 section
->offset_within_address_space
+
2140 int128_get64(int128_sub(section
->size
, int128_one())));
2144 if (unlikely((section
->offset_within_address_space
& ~TARGET_PAGE_MASK
) !=
2145 (section
->offset_within_region
& ~TARGET_PAGE_MASK
))) {
2146 error_report("%s received unaligned region", __func__
);
2150 iova
= TARGET_PAGE_ALIGN(section
->offset_within_address_space
);
2151 end
= (section
->offset_within_address_space
+ int128_get64(section
->size
)) &
2158 DPRINTF("region_del %"HWADDR_PRIx
" - %"HWADDR_PRIx
"\n",
2161 ret
= vfio_dma_unmap(container
, iova
, end
- iova
);
2162 memory_region_unref(section
->mr
);
2164 error_report("vfio_dma_unmap(%p, 0x%"HWADDR_PRIx
", "
2165 "0x%"HWADDR_PRIx
") = %d (%m)",
2166 container
, iova
, end
- iova
, ret
);
2170 static MemoryListener vfio_memory_listener
= {
2171 .region_add
= vfio_listener_region_add
,
2172 .region_del
= vfio_listener_region_del
,
2175 static void vfio_listener_release(VFIOContainer
*container
)
2177 memory_listener_unregister(&container
->iommu_data
.listener
);
2183 static void vfio_disable_interrupts(VFIODevice
*vdev
)
2185 switch (vdev
->interrupt
) {
2187 vfio_disable_intx(vdev
);
2190 vfio_disable_msi(vdev
);
2193 vfio_disable_msix(vdev
);
2198 static int vfio_setup_msi(VFIODevice
*vdev
, int pos
)
2201 bool msi_64bit
, msi_maskbit
;
2204 if (pread(vdev
->fd
, &ctrl
, sizeof(ctrl
),
2205 vdev
->config_offset
+ pos
+ PCI_CAP_FLAGS
) != sizeof(ctrl
)) {
2208 ctrl
= le16_to_cpu(ctrl
);
2210 msi_64bit
= !!(ctrl
& PCI_MSI_FLAGS_64BIT
);
2211 msi_maskbit
= !!(ctrl
& PCI_MSI_FLAGS_MASKBIT
);
2212 entries
= 1 << ((ctrl
& PCI_MSI_FLAGS_QMASK
) >> 1);
2214 DPRINTF("%04x:%02x:%02x.%x PCI MSI CAP @0x%x\n", vdev
->host
.domain
,
2215 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
, pos
);
2217 ret
= msi_init(&vdev
->pdev
, pos
, entries
, msi_64bit
, msi_maskbit
);
2219 if (ret
== -ENOTSUP
) {
2222 error_report("vfio: msi_init failed");
2225 vdev
->msi_cap_size
= 0xa + (msi_maskbit
? 0xa : 0) + (msi_64bit
? 0x4 : 0);
2231 * We don't have any control over how pci_add_capability() inserts
2232 * capabilities into the chain. In order to setup MSI-X we need a
2233 * MemoryRegion for the BAR. In order to setup the BAR and not
2234 * attempt to mmap the MSI-X table area, which VFIO won't allow, we
2235 * need to first look for where the MSI-X table lives. So we
2236 * unfortunately split MSI-X setup across two functions.
2238 static int vfio_early_setup_msix(VFIODevice
*vdev
)
2242 uint32_t table
, pba
;
2244 pos
= pci_find_capability(&vdev
->pdev
, PCI_CAP_ID_MSIX
);
2249 if (pread(vdev
->fd
, &ctrl
, sizeof(ctrl
),
2250 vdev
->config_offset
+ pos
+ PCI_CAP_FLAGS
) != sizeof(ctrl
)) {
2254 if (pread(vdev
->fd
, &table
, sizeof(table
),
2255 vdev
->config_offset
+ pos
+ PCI_MSIX_TABLE
) != sizeof(table
)) {
2259 if (pread(vdev
->fd
, &pba
, sizeof(pba
),
2260 vdev
->config_offset
+ pos
+ PCI_MSIX_PBA
) != sizeof(pba
)) {
2264 ctrl
= le16_to_cpu(ctrl
);
2265 table
= le32_to_cpu(table
);
2266 pba
= le32_to_cpu(pba
);
2268 vdev
->msix
= g_malloc0(sizeof(*(vdev
->msix
)));
2269 vdev
->msix
->table_bar
= table
& PCI_MSIX_FLAGS_BIRMASK
;
2270 vdev
->msix
->table_offset
= table
& ~PCI_MSIX_FLAGS_BIRMASK
;
2271 vdev
->msix
->pba_bar
= pba
& PCI_MSIX_FLAGS_BIRMASK
;
2272 vdev
->msix
->pba_offset
= pba
& ~PCI_MSIX_FLAGS_BIRMASK
;
2273 vdev
->msix
->entries
= (ctrl
& PCI_MSIX_FLAGS_QSIZE
) + 1;
2275 DPRINTF("%04x:%02x:%02x.%x "
2276 "PCI MSI-X CAP @0x%x, BAR %d, offset 0x%x, entries %d\n",
2277 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2278 vdev
->host
.function
, pos
, vdev
->msix
->table_bar
,
2279 vdev
->msix
->table_offset
, vdev
->msix
->entries
);
2284 static int vfio_setup_msix(VFIODevice
*vdev
, int pos
)
2288 ret
= msix_init(&vdev
->pdev
, vdev
->msix
->entries
,
2289 &vdev
->bars
[vdev
->msix
->table_bar
].mem
,
2290 vdev
->msix
->table_bar
, vdev
->msix
->table_offset
,
2291 &vdev
->bars
[vdev
->msix
->pba_bar
].mem
,
2292 vdev
->msix
->pba_bar
, vdev
->msix
->pba_offset
, pos
);
2294 if (ret
== -ENOTSUP
) {
2297 error_report("vfio: msix_init failed");
2304 static void vfio_teardown_msi(VFIODevice
*vdev
)
2306 msi_uninit(&vdev
->pdev
);
2309 msix_uninit(&vdev
->pdev
, &vdev
->bars
[vdev
->msix
->table_bar
].mem
,
2310 &vdev
->bars
[vdev
->msix
->pba_bar
].mem
);
2317 static void vfio_mmap_set_enabled(VFIODevice
*vdev
, bool enabled
)
2321 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
2322 VFIOBAR
*bar
= &vdev
->bars
[i
];
2328 memory_region_set_enabled(&bar
->mmap_mem
, enabled
);
2329 if (vdev
->msix
&& vdev
->msix
->table_bar
== i
) {
2330 memory_region_set_enabled(&vdev
->msix
->mmap_mem
, enabled
);
2335 static void vfio_unmap_bar(VFIODevice
*vdev
, int nr
)
2337 VFIOBAR
*bar
= &vdev
->bars
[nr
];
2343 vfio_bar_quirk_teardown(vdev
, nr
);
2345 memory_region_del_subregion(&bar
->mem
, &bar
->mmap_mem
);
2346 munmap(bar
->mmap
, memory_region_size(&bar
->mmap_mem
));
2348 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
2349 memory_region_del_subregion(&bar
->mem
, &vdev
->msix
->mmap_mem
);
2350 munmap(vdev
->msix
->mmap
, memory_region_size(&vdev
->msix
->mmap_mem
));
2353 memory_region_destroy(&bar
->mem
);
2356 static int vfio_mmap_bar(VFIODevice
*vdev
, VFIOBAR
*bar
,
2357 MemoryRegion
*mem
, MemoryRegion
*submem
,
2358 void **map
, size_t size
, off_t offset
,
2363 if (VFIO_ALLOW_MMAP
&& size
&& bar
->flags
& VFIO_REGION_INFO_FLAG_MMAP
) {
2366 if (bar
->flags
& VFIO_REGION_INFO_FLAG_READ
) {
2370 if (bar
->flags
& VFIO_REGION_INFO_FLAG_WRITE
) {
2374 *map
= mmap(NULL
, size
, prot
, MAP_SHARED
,
2375 bar
->fd
, bar
->fd_offset
+ offset
);
2376 if (*map
== MAP_FAILED
) {
2382 memory_region_init_ram_ptr(submem
, OBJECT(vdev
), name
, size
, *map
);
2385 /* Create a zero sized sub-region to make cleanup easy. */
2386 memory_region_init(submem
, OBJECT(vdev
), name
, 0);
2389 memory_region_add_subregion(mem
, offset
, submem
);
2394 static void vfio_map_bar(VFIODevice
*vdev
, int nr
)
2396 VFIOBAR
*bar
= &vdev
->bars
[nr
];
2397 unsigned size
= bar
->size
;
2403 /* Skip both unimplemented BARs and the upper half of 64bit BARS. */
2408 snprintf(name
, sizeof(name
), "VFIO %04x:%02x:%02x.%x BAR %d",
2409 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2410 vdev
->host
.function
, nr
);
2412 /* Determine what type of BAR this is for registration */
2413 ret
= pread(vdev
->fd
, &pci_bar
, sizeof(pci_bar
),
2414 vdev
->config_offset
+ PCI_BASE_ADDRESS_0
+ (4 * nr
));
2415 if (ret
!= sizeof(pci_bar
)) {
2416 error_report("vfio: Failed to read BAR %d (%m)", nr
);
2420 pci_bar
= le32_to_cpu(pci_bar
);
2421 bar
->ioport
= (pci_bar
& PCI_BASE_ADDRESS_SPACE_IO
);
2422 bar
->mem64
= bar
->ioport
? 0 : (pci_bar
& PCI_BASE_ADDRESS_MEM_TYPE_64
);
2423 type
= pci_bar
& (bar
->ioport
? ~PCI_BASE_ADDRESS_IO_MASK
:
2424 ~PCI_BASE_ADDRESS_MEM_MASK
);
2426 /* A "slow" read/write mapping underlies all BARs */
2427 memory_region_init_io(&bar
->mem
, OBJECT(vdev
), &vfio_bar_ops
,
2429 pci_register_bar(&vdev
->pdev
, nr
, type
, &bar
->mem
);
2432 * We can't mmap areas overlapping the MSIX vector table, so we
2433 * potentially insert a direct-mapped subregion before and after it.
2435 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
2436 size
= vdev
->msix
->table_offset
& TARGET_PAGE_MASK
;
2439 strncat(name
, " mmap", sizeof(name
) - strlen(name
) - 1);
2440 if (vfio_mmap_bar(vdev
, bar
, &bar
->mem
,
2441 &bar
->mmap_mem
, &bar
->mmap
, size
, 0, name
)) {
2442 error_report("%s unsupported. Performance may be slow", name
);
2445 if (vdev
->msix
&& vdev
->msix
->table_bar
== nr
) {
2448 start
= TARGET_PAGE_ALIGN(vdev
->msix
->table_offset
+
2449 (vdev
->msix
->entries
* PCI_MSIX_ENTRY_SIZE
));
2451 size
= start
< bar
->size
? bar
->size
- start
: 0;
2452 strncat(name
, " msix-hi", sizeof(name
) - strlen(name
) - 1);
2453 /* VFIOMSIXInfo contains another MemoryRegion for this mapping */
2454 if (vfio_mmap_bar(vdev
, bar
, &bar
->mem
, &vdev
->msix
->mmap_mem
,
2455 &vdev
->msix
->mmap
, size
, start
, name
)) {
2456 error_report("%s unsupported. Performance may be slow", name
);
2460 vfio_bar_quirk_setup(vdev
, nr
);
2463 static void vfio_map_bars(VFIODevice
*vdev
)
2467 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
2468 vfio_map_bar(vdev
, i
);
2471 if (vdev
->has_vga
) {
2472 memory_region_init_io(&vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].mem
,
2473 OBJECT(vdev
), &vfio_vga_ops
,
2474 &vdev
->vga
.region
[QEMU_PCI_VGA_MEM
],
2475 "vfio-vga-mmio@0xa0000",
2476 QEMU_PCI_VGA_MEM_SIZE
);
2477 memory_region_init_io(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].mem
,
2478 OBJECT(vdev
), &vfio_vga_ops
,
2479 &vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
],
2480 "vfio-vga-io@0x3b0",
2481 QEMU_PCI_VGA_IO_LO_SIZE
);
2482 memory_region_init_io(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].mem
,
2483 OBJECT(vdev
), &vfio_vga_ops
,
2484 &vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
],
2485 "vfio-vga-io@0x3c0",
2486 QEMU_PCI_VGA_IO_HI_SIZE
);
2488 pci_register_vga(&vdev
->pdev
, &vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].mem
,
2489 &vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].mem
,
2490 &vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].mem
);
2491 vfio_vga_quirk_setup(vdev
);
2495 static void vfio_unmap_bars(VFIODevice
*vdev
)
2499 for (i
= 0; i
< PCI_ROM_SLOT
; i
++) {
2500 vfio_unmap_bar(vdev
, i
);
2503 if (vdev
->has_vga
) {
2504 vfio_vga_quirk_teardown(vdev
);
2505 pci_unregister_vga(&vdev
->pdev
);
2506 memory_region_destroy(&vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].mem
);
2507 memory_region_destroy(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].mem
);
2508 memory_region_destroy(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].mem
);
2515 static uint8_t vfio_std_cap_max_size(PCIDevice
*pdev
, uint8_t pos
)
2517 uint8_t tmp
, next
= 0xff;
2519 for (tmp
= pdev
->config
[PCI_CAPABILITY_LIST
]; tmp
;
2520 tmp
= pdev
->config
[tmp
+ 1]) {
2521 if (tmp
> pos
&& tmp
< next
) {
2529 static void vfio_set_word_bits(uint8_t *buf
, uint16_t val
, uint16_t mask
)
2531 pci_set_word(buf
, (pci_get_word(buf
) & ~mask
) | val
);
2534 static void vfio_add_emulated_word(VFIODevice
*vdev
, int pos
,
2535 uint16_t val
, uint16_t mask
)
2537 vfio_set_word_bits(vdev
->pdev
.config
+ pos
, val
, mask
);
2538 vfio_set_word_bits(vdev
->pdev
.wmask
+ pos
, ~mask
, mask
);
2539 vfio_set_word_bits(vdev
->emulated_config_bits
+ pos
, mask
, mask
);
2542 static void vfio_set_long_bits(uint8_t *buf
, uint32_t val
, uint32_t mask
)
2544 pci_set_long(buf
, (pci_get_long(buf
) & ~mask
) | val
);
2547 static void vfio_add_emulated_long(VFIODevice
*vdev
, int pos
,
2548 uint32_t val
, uint32_t mask
)
2550 vfio_set_long_bits(vdev
->pdev
.config
+ pos
, val
, mask
);
2551 vfio_set_long_bits(vdev
->pdev
.wmask
+ pos
, ~mask
, mask
);
2552 vfio_set_long_bits(vdev
->emulated_config_bits
+ pos
, mask
, mask
);
2555 static int vfio_setup_pcie_cap(VFIODevice
*vdev
, int pos
, uint8_t size
)
2560 flags
= pci_get_word(vdev
->pdev
.config
+ pos
+ PCI_CAP_FLAGS
);
2561 type
= (flags
& PCI_EXP_FLAGS_TYPE
) >> 4;
2563 if (type
!= PCI_EXP_TYPE_ENDPOINT
&&
2564 type
!= PCI_EXP_TYPE_LEG_END
&&
2565 type
!= PCI_EXP_TYPE_RC_END
) {
2567 error_report("vfio: Assignment of PCIe type 0x%x "
2568 "devices is not currently supported", type
);
2572 if (!pci_bus_is_express(vdev
->pdev
.bus
)) {
2574 * Use express capability as-is on PCI bus. It doesn't make much
2575 * sense to even expose, but some drivers (ex. tg3) depend on it
2576 * and guests don't seem to be particular about it. We'll need
2577 * to revist this or force express devices to express buses if we
2578 * ever expose an IOMMU to the guest.
2580 } else if (pci_bus_is_root(vdev
->pdev
.bus
)) {
2582 * On a Root Complex bus Endpoints become Root Complex Integrated
2583 * Endpoints, which changes the type and clears the LNK & LNK2 fields.
2585 if (type
== PCI_EXP_TYPE_ENDPOINT
) {
2586 vfio_add_emulated_word(vdev
, pos
+ PCI_CAP_FLAGS
,
2587 PCI_EXP_TYPE_RC_END
<< 4,
2588 PCI_EXP_FLAGS_TYPE
);
2590 /* Link Capabilities, Status, and Control goes away */
2591 if (size
> PCI_EXP_LNKCTL
) {
2592 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP
, 0, ~0);
2593 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL
, 0, ~0);
2594 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKSTA
, 0, ~0);
2596 #ifndef PCI_EXP_LNKCAP2
2597 #define PCI_EXP_LNKCAP2 44
2599 #ifndef PCI_EXP_LNKSTA2
2600 #define PCI_EXP_LNKSTA2 50
2602 /* Link 2 Capabilities, Status, and Control goes away */
2603 if (size
> PCI_EXP_LNKCAP2
) {
2604 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP2
, 0, ~0);
2605 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL2
, 0, ~0);
2606 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKSTA2
, 0, ~0);
2610 } else if (type
== PCI_EXP_TYPE_LEG_END
) {
2612 * Legacy endpoints don't belong on the root complex. Windows
2613 * seems to be happier with devices if we skip the capability.
2620 * Convert Root Complex Integrated Endpoints to regular endpoints.
2621 * These devices don't support LNK/LNK2 capabilities, so make them up.
2623 if (type
== PCI_EXP_TYPE_RC_END
) {
2624 vfio_add_emulated_word(vdev
, pos
+ PCI_CAP_FLAGS
,
2625 PCI_EXP_TYPE_ENDPOINT
<< 4,
2626 PCI_EXP_FLAGS_TYPE
);
2627 vfio_add_emulated_long(vdev
, pos
+ PCI_EXP_LNKCAP
,
2628 PCI_EXP_LNK_MLW_1
| PCI_EXP_LNK_LS_25
, ~0);
2629 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKCTL
, 0, ~0);
2632 /* Mark the Link Status bits as emulated to allow virtual negotiation */
2633 vfio_add_emulated_word(vdev
, pos
+ PCI_EXP_LNKSTA
,
2634 pci_get_word(vdev
->pdev
.config
+ pos
+
2636 PCI_EXP_LNKCAP_MLW
| PCI_EXP_LNKCAP_SLS
);
2639 pos
= pci_add_capability(&vdev
->pdev
, PCI_CAP_ID_EXP
, pos
, size
);
2641 vdev
->pdev
.exp
.exp_cap
= pos
;
2647 static void vfio_check_pcie_flr(VFIODevice
*vdev
, uint8_t pos
)
2649 uint32_t cap
= pci_get_long(vdev
->pdev
.config
+ pos
+ PCI_EXP_DEVCAP
);
2651 if (cap
& PCI_EXP_DEVCAP_FLR
) {
2652 DPRINTF("%04x:%02x:%02x.%x Supports FLR via PCIe cap\n",
2653 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2654 vdev
->host
.function
);
2655 vdev
->has_flr
= true;
2659 static void vfio_check_pm_reset(VFIODevice
*vdev
, uint8_t pos
)
2661 uint16_t csr
= pci_get_word(vdev
->pdev
.config
+ pos
+ PCI_PM_CTRL
);
2663 if (!(csr
& PCI_PM_CTRL_NO_SOFT_RESET
)) {
2664 DPRINTF("%04x:%02x:%02x.%x Supports PM reset\n",
2665 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2666 vdev
->host
.function
);
2667 vdev
->has_pm_reset
= true;
2671 static void vfio_check_af_flr(VFIODevice
*vdev
, uint8_t pos
)
2673 uint8_t cap
= pci_get_byte(vdev
->pdev
.config
+ pos
+ PCI_AF_CAP
);
2675 if ((cap
& PCI_AF_CAP_TP
) && (cap
& PCI_AF_CAP_FLR
)) {
2676 DPRINTF("%04x:%02x:%02x.%x Supports FLR via AF cap\n",
2677 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2678 vdev
->host
.function
);
2679 vdev
->has_flr
= true;
2683 static int vfio_add_std_cap(VFIODevice
*vdev
, uint8_t pos
)
2685 PCIDevice
*pdev
= &vdev
->pdev
;
2686 uint8_t cap_id
, next
, size
;
2689 cap_id
= pdev
->config
[pos
];
2690 next
= pdev
->config
[pos
+ 1];
2693 * If it becomes important to configure capabilities to their actual
2694 * size, use this as the default when it's something we don't recognize.
2695 * Since QEMU doesn't actually handle many of the config accesses,
2696 * exact size doesn't seem worthwhile.
2698 size
= vfio_std_cap_max_size(pdev
, pos
);
2701 * pci_add_capability always inserts the new capability at the head
2702 * of the chain. Therefore to end up with a chain that matches the
2703 * physical device, we insert from the end by making this recursive.
2704 * This is also why we pre-caclulate size above as cached config space
2705 * will be changed as we unwind the stack.
2708 ret
= vfio_add_std_cap(vdev
, next
);
2713 /* Begin the rebuild, use QEMU emulated list bits */
2714 pdev
->config
[PCI_CAPABILITY_LIST
] = 0;
2715 vdev
->emulated_config_bits
[PCI_CAPABILITY_LIST
] = 0xff;
2716 vdev
->emulated_config_bits
[PCI_STATUS
] |= PCI_STATUS_CAP_LIST
;
2719 /* Use emulated next pointer to allow dropping caps */
2720 pci_set_byte(vdev
->emulated_config_bits
+ pos
+ 1, 0xff);
2723 case PCI_CAP_ID_MSI
:
2724 ret
= vfio_setup_msi(vdev
, pos
);
2726 case PCI_CAP_ID_EXP
:
2727 vfio_check_pcie_flr(vdev
, pos
);
2728 ret
= vfio_setup_pcie_cap(vdev
, pos
, size
);
2730 case PCI_CAP_ID_MSIX
:
2731 ret
= vfio_setup_msix(vdev
, pos
);
2734 vfio_check_pm_reset(vdev
, pos
);
2736 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
);
2739 vfio_check_af_flr(vdev
, pos
);
2740 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
);
2743 ret
= pci_add_capability(pdev
, cap_id
, pos
, size
);
2748 error_report("vfio: %04x:%02x:%02x.%x Error adding PCI capability "
2749 "0x%x[0x%x]@0x%x: %d", vdev
->host
.domain
,
2750 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
2751 cap_id
, size
, pos
, ret
);
2758 static int vfio_add_capabilities(VFIODevice
*vdev
)
2760 PCIDevice
*pdev
= &vdev
->pdev
;
2762 if (!(pdev
->config
[PCI_STATUS
] & PCI_STATUS_CAP_LIST
) ||
2763 !pdev
->config
[PCI_CAPABILITY_LIST
]) {
2764 return 0; /* Nothing to add */
2767 return vfio_add_std_cap(vdev
, pdev
->config
[PCI_CAPABILITY_LIST
]);
2770 static void vfio_pci_pre_reset(VFIODevice
*vdev
)
2772 PCIDevice
*pdev
= &vdev
->pdev
;
2775 vfio_disable_interrupts(vdev
);
2777 /* Make sure the device is in D0 */
2782 pmcsr
= vfio_pci_read_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, 2);
2783 state
= pmcsr
& PCI_PM_CTRL_STATE_MASK
;
2785 pmcsr
&= ~PCI_PM_CTRL_STATE_MASK
;
2786 vfio_pci_write_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, pmcsr
, 2);
2787 /* vfio handles the necessary delay here */
2788 pmcsr
= vfio_pci_read_config(pdev
, vdev
->pm_cap
+ PCI_PM_CTRL
, 2);
2789 state
= pmcsr
& PCI_PM_CTRL_STATE_MASK
;
2791 error_report("vfio: Unable to power on device, stuck in D%d\n",
2798 * Stop any ongoing DMA by disconecting I/O, MMIO, and bus master.
2799 * Also put INTx Disable in known state.
2801 cmd
= vfio_pci_read_config(pdev
, PCI_COMMAND
, 2);
2802 cmd
&= ~(PCI_COMMAND_IO
| PCI_COMMAND_MEMORY
| PCI_COMMAND_MASTER
|
2803 PCI_COMMAND_INTX_DISABLE
);
2804 vfio_pci_write_config(pdev
, PCI_COMMAND
, cmd
, 2);
2807 static void vfio_pci_post_reset(VFIODevice
*vdev
)
2809 vfio_enable_intx(vdev
);
2812 static bool vfio_pci_host_match(PCIHostDeviceAddress
*host1
,
2813 PCIHostDeviceAddress
*host2
)
2815 return (host1
->domain
== host2
->domain
&& host1
->bus
== host2
->bus
&&
2816 host1
->slot
== host2
->slot
&& host1
->function
== host2
->function
);
2819 static int vfio_pci_hot_reset(VFIODevice
*vdev
, bool single
)
2822 struct vfio_pci_hot_reset_info
*info
;
2823 struct vfio_pci_dependent_device
*devices
;
2824 struct vfio_pci_hot_reset
*reset
;
2829 DPRINTF("%s(%04x:%02x:%02x.%x) %s\n", __func__
, vdev
->host
.domain
,
2830 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
2831 single
? "one" : "multi");
2833 vfio_pci_pre_reset(vdev
);
2834 vdev
->needs_reset
= false;
2836 info
= g_malloc0(sizeof(*info
));
2837 info
->argsz
= sizeof(*info
);
2839 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO
, info
);
2840 if (ret
&& errno
!= ENOSPC
) {
2842 if (!vdev
->has_pm_reset
) {
2843 error_report("vfio: Cannot reset device %04x:%02x:%02x.%x, "
2844 "no available reset mechanism.", vdev
->host
.domain
,
2845 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
2850 count
= info
->count
;
2851 info
= g_realloc(info
, sizeof(*info
) + (count
* sizeof(*devices
)));
2852 info
->argsz
= sizeof(*info
) + (count
* sizeof(*devices
));
2853 devices
= &info
->devices
[0];
2855 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_PCI_HOT_RESET_INFO
, info
);
2858 error_report("vfio: hot reset info failed: %m");
2862 DPRINTF("%04x:%02x:%02x.%x: hot reset dependent devices:\n",
2863 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2864 vdev
->host
.function
);
2866 /* Verify that we have all the groups required */
2867 for (i
= 0; i
< info
->count
; i
++) {
2868 PCIHostDeviceAddress host
;
2871 host
.domain
= devices
[i
].segment
;
2872 host
.bus
= devices
[i
].bus
;
2873 host
.slot
= PCI_SLOT(devices
[i
].devfn
);
2874 host
.function
= PCI_FUNC(devices
[i
].devfn
);
2876 DPRINTF("\t%04x:%02x:%02x.%x group %d\n", host
.domain
,
2877 host
.bus
, host
.slot
, host
.function
, devices
[i
].group_id
);
2879 if (vfio_pci_host_match(&host
, &vdev
->host
)) {
2883 QLIST_FOREACH(group
, &group_list
, next
) {
2884 if (group
->groupid
== devices
[i
].group_id
) {
2890 if (!vdev
->has_pm_reset
) {
2891 error_report("vfio: Cannot reset device %04x:%02x:%02x.%x, "
2892 "depends on group %d which is not owned.",
2893 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
2894 vdev
->host
.function
, devices
[i
].group_id
);
2900 /* Prep dependent devices for reset and clear our marker. */
2901 QLIST_FOREACH(tmp
, &group
->device_list
, next
) {
2902 if (vfio_pci_host_match(&host
, &tmp
->host
)) {
2904 DPRINTF("vfio: found another in-use device "
2905 "%04x:%02x:%02x.%x\n", host
.domain
, host
.bus
,
2906 host
.slot
, host
.function
);
2910 vfio_pci_pre_reset(tmp
);
2911 tmp
->needs_reset
= false;
2918 if (!single
&& !multi
) {
2919 DPRINTF("vfio: No other in-use devices for multi hot reset\n");
2924 /* Determine how many group fds need to be passed */
2926 QLIST_FOREACH(group
, &group_list
, next
) {
2927 for (i
= 0; i
< info
->count
; i
++) {
2928 if (group
->groupid
== devices
[i
].group_id
) {
2935 reset
= g_malloc0(sizeof(*reset
) + (count
* sizeof(*fds
)));
2936 reset
->argsz
= sizeof(*reset
) + (count
* sizeof(*fds
));
2937 fds
= &reset
->group_fds
[0];
2939 /* Fill in group fds */
2940 QLIST_FOREACH(group
, &group_list
, next
) {
2941 for (i
= 0; i
< info
->count
; i
++) {
2942 if (group
->groupid
== devices
[i
].group_id
) {
2943 fds
[reset
->count
++] = group
->fd
;
2950 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_PCI_HOT_RESET
, reset
);
2953 DPRINTF("%04x:%02x:%02x.%x hot reset: %s\n", vdev
->host
.domain
,
2954 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
,
2955 ret
? "%m" : "Success");
2958 /* Re-enable INTx on affected devices */
2959 for (i
= 0; i
< info
->count
; i
++) {
2960 PCIHostDeviceAddress host
;
2963 host
.domain
= devices
[i
].segment
;
2964 host
.bus
= devices
[i
].bus
;
2965 host
.slot
= PCI_SLOT(devices
[i
].devfn
);
2966 host
.function
= PCI_FUNC(devices
[i
].devfn
);
2968 if (vfio_pci_host_match(&host
, &vdev
->host
)) {
2972 QLIST_FOREACH(group
, &group_list
, next
) {
2973 if (group
->groupid
== devices
[i
].group_id
) {
2982 QLIST_FOREACH(tmp
, &group
->device_list
, next
) {
2983 if (vfio_pci_host_match(&host
, &tmp
->host
)) {
2984 vfio_pci_post_reset(tmp
);
2990 vfio_pci_post_reset(vdev
);
2997 * We want to differentiate hot reset of mulitple in-use devices vs hot reset
2998 * of a single in-use device. VFIO_DEVICE_RESET will already handle the case
2999 * of doing hot resets when there is only a single device per bus. The in-use
3000 * here refers to how many VFIODevices are affected. A hot reset that affects
3001 * multiple devices, but only a single in-use device, means that we can call
3002 * it from our bus ->reset() callback since the extent is effectively a single
3003 * device. This allows us to make use of it in the hotplug path. When there
3004 * are multiple in-use devices, we can only trigger the hot reset during a
3005 * system reset and thus from our reset handler. We separate _one vs _multi
3006 * here so that we don't overlap and do a double reset on the system reset
3007 * path where both our reset handler and ->reset() callback are used. Calling
3008 * _one() will only do a hot reset for the one in-use devices case, calling
3009 * _multi() will do nothing if a _one() would have been sufficient.
3011 static int vfio_pci_hot_reset_one(VFIODevice
*vdev
)
3013 return vfio_pci_hot_reset(vdev
, true);
3016 static int vfio_pci_hot_reset_multi(VFIODevice
*vdev
)
3018 return vfio_pci_hot_reset(vdev
, false);
3021 static void vfio_pci_reset_handler(void *opaque
)
3026 QLIST_FOREACH(group
, &group_list
, next
) {
3027 QLIST_FOREACH(vdev
, &group
->device_list
, next
) {
3028 if (!vdev
->reset_works
|| (!vdev
->has_flr
&& vdev
->has_pm_reset
)) {
3029 vdev
->needs_reset
= true;
3034 QLIST_FOREACH(group
, &group_list
, next
) {
3035 QLIST_FOREACH(vdev
, &group
->device_list
, next
) {
3036 if (vdev
->needs_reset
) {
3037 vfio_pci_hot_reset_multi(vdev
);
3043 static int vfio_connect_container(VFIOGroup
*group
)
3045 VFIOContainer
*container
;
3048 if (group
->container
) {
3052 QLIST_FOREACH(container
, &container_list
, next
) {
3053 if (!ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &container
->fd
)) {
3054 group
->container
= container
;
3055 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
3060 fd
= qemu_open("/dev/vfio/vfio", O_RDWR
);
3062 error_report("vfio: failed to open /dev/vfio/vfio: %m");
3066 ret
= ioctl(fd
, VFIO_GET_API_VERSION
);
3067 if (ret
!= VFIO_API_VERSION
) {
3068 error_report("vfio: supported vfio version: %d, "
3069 "reported version: %d", VFIO_API_VERSION
, ret
);
3074 container
= g_malloc0(sizeof(*container
));
3077 if (ioctl(fd
, VFIO_CHECK_EXTENSION
, VFIO_TYPE1_IOMMU
)) {
3078 ret
= ioctl(group
->fd
, VFIO_GROUP_SET_CONTAINER
, &fd
);
3080 error_report("vfio: failed to set group container: %m");
3086 ret
= ioctl(fd
, VFIO_SET_IOMMU
, VFIO_TYPE1_IOMMU
);
3088 error_report("vfio: failed to set iommu for container: %m");
3094 container
->iommu_data
.listener
= vfio_memory_listener
;
3095 container
->iommu_data
.release
= vfio_listener_release
;
3097 memory_listener_register(&container
->iommu_data
.listener
, &address_space_memory
);
3099 error_report("vfio: No available IOMMU models");
3105 QLIST_INIT(&container
->group_list
);
3106 QLIST_INSERT_HEAD(&container_list
, container
, next
);
3108 group
->container
= container
;
3109 QLIST_INSERT_HEAD(&container
->group_list
, group
, container_next
);
3114 static void vfio_disconnect_container(VFIOGroup
*group
)
3116 VFIOContainer
*container
= group
->container
;
3118 if (ioctl(group
->fd
, VFIO_GROUP_UNSET_CONTAINER
, &container
->fd
)) {
3119 error_report("vfio: error disconnecting group %d from container",
3123 QLIST_REMOVE(group
, container_next
);
3124 group
->container
= NULL
;
3126 if (QLIST_EMPTY(&container
->group_list
)) {
3127 if (container
->iommu_data
.release
) {
3128 container
->iommu_data
.release(container
);
3130 QLIST_REMOVE(container
, next
);
3131 DPRINTF("vfio_disconnect_container: close container->fd\n");
3132 close(container
->fd
);
3137 static VFIOGroup
*vfio_get_group(int groupid
)
3141 struct vfio_group_status status
= { .argsz
= sizeof(status
) };
3143 QLIST_FOREACH(group
, &group_list
, next
) {
3144 if (group
->groupid
== groupid
) {
3149 group
= g_malloc0(sizeof(*group
));
3151 snprintf(path
, sizeof(path
), "/dev/vfio/%d", groupid
);
3152 group
->fd
= qemu_open(path
, O_RDWR
);
3153 if (group
->fd
< 0) {
3154 error_report("vfio: error opening %s: %m", path
);
3159 if (ioctl(group
->fd
, VFIO_GROUP_GET_STATUS
, &status
)) {
3160 error_report("vfio: error getting group status: %m");
3166 if (!(status
.flags
& VFIO_GROUP_FLAGS_VIABLE
)) {
3167 error_report("vfio: error, group %d is not viable, please ensure "
3168 "all devices within the iommu_group are bound to their "
3169 "vfio bus driver.", groupid
);
3175 group
->groupid
= groupid
;
3176 QLIST_INIT(&group
->device_list
);
3178 if (vfio_connect_container(group
)) {
3179 error_report("vfio: failed to setup container for group %d", groupid
);
3185 if (QLIST_EMPTY(&group_list
)) {
3186 qemu_register_reset(vfio_pci_reset_handler
, NULL
);
3189 QLIST_INSERT_HEAD(&group_list
, group
, next
);
3194 static void vfio_put_group(VFIOGroup
*group
)
3196 if (!QLIST_EMPTY(&group
->device_list
)) {
3200 vfio_disconnect_container(group
);
3201 QLIST_REMOVE(group
, next
);
3202 DPRINTF("vfio_put_group: close group->fd\n");
3206 if (QLIST_EMPTY(&group_list
)) {
3207 qemu_unregister_reset(vfio_pci_reset_handler
, NULL
);
3211 static int vfio_get_device(VFIOGroup
*group
, const char *name
, VFIODevice
*vdev
)
3213 struct vfio_device_info dev_info
= { .argsz
= sizeof(dev_info
) };
3214 struct vfio_region_info reg_info
= { .argsz
= sizeof(reg_info
) };
3215 struct vfio_irq_info irq_info
= { .argsz
= sizeof(irq_info
) };
3218 ret
= ioctl(group
->fd
, VFIO_GROUP_GET_DEVICE_FD
, name
);
3220 error_report("vfio: error getting device %s from group %d: %m",
3221 name
, group
->groupid
);
3222 error_printf("Verify all devices in group %d are bound to vfio-pci "
3223 "or pci-stub and not already in use\n", group
->groupid
);
3228 vdev
->group
= group
;
3229 QLIST_INSERT_HEAD(&group
->device_list
, vdev
, next
);
3231 /* Sanity check device */
3232 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_INFO
, &dev_info
);
3234 error_report("vfio: error getting device info: %m");
3238 DPRINTF("Device %s flags: %u, regions: %u, irgs: %u\n", name
,
3239 dev_info
.flags
, dev_info
.num_regions
, dev_info
.num_irqs
);
3241 if (!(dev_info
.flags
& VFIO_DEVICE_FLAGS_PCI
)) {
3242 error_report("vfio: Um, this isn't a PCI device");
3246 vdev
->reset_works
= !!(dev_info
.flags
& VFIO_DEVICE_FLAGS_RESET
);
3248 if (dev_info
.num_regions
< VFIO_PCI_CONFIG_REGION_INDEX
+ 1) {
3249 error_report("vfio: unexpected number of io regions %u",
3250 dev_info
.num_regions
);
3254 if (dev_info
.num_irqs
< VFIO_PCI_MSIX_IRQ_INDEX
+ 1) {
3255 error_report("vfio: unexpected number of irqs %u", dev_info
.num_irqs
);
3259 for (i
= VFIO_PCI_BAR0_REGION_INDEX
; i
< VFIO_PCI_ROM_REGION_INDEX
; i
++) {
3262 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
);
3264 error_report("vfio: Error getting region %d info: %m", i
);
3268 DPRINTF("Device %s region %d:\n", name
, i
);
3269 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
3270 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
3271 (unsigned long)reg_info
.flags
);
3273 vdev
->bars
[i
].flags
= reg_info
.flags
;
3274 vdev
->bars
[i
].size
= reg_info
.size
;
3275 vdev
->bars
[i
].fd_offset
= reg_info
.offset
;
3276 vdev
->bars
[i
].fd
= vdev
->fd
;
3277 vdev
->bars
[i
].nr
= i
;
3278 QLIST_INIT(&vdev
->bars
[i
].quirks
);
3281 reg_info
.index
= VFIO_PCI_CONFIG_REGION_INDEX
;
3283 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, ®_info
);
3285 error_report("vfio: Error getting config info: %m");
3289 DPRINTF("Device %s config:\n", name
);
3290 DPRINTF(" size: 0x%lx, offset: 0x%lx, flags: 0x%lx\n",
3291 (unsigned long)reg_info
.size
, (unsigned long)reg_info
.offset
,
3292 (unsigned long)reg_info
.flags
);
3294 vdev
->config_size
= reg_info
.size
;
3295 if (vdev
->config_size
== PCI_CONFIG_SPACE_SIZE
) {
3296 vdev
->pdev
.cap_present
&= ~QEMU_PCI_CAP_EXPRESS
;
3298 vdev
->config_offset
= reg_info
.offset
;
3300 if ((vdev
->features
& VFIO_FEATURE_ENABLE_VGA
) &&
3301 dev_info
.num_regions
> VFIO_PCI_VGA_REGION_INDEX
) {
3302 struct vfio_region_info vga_info
= {
3303 .argsz
= sizeof(vga_info
),
3304 .index
= VFIO_PCI_VGA_REGION_INDEX
,
3307 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_REGION_INFO
, &vga_info
);
3310 "vfio: Device does not support requested feature x-vga");
3314 if (!(vga_info
.flags
& VFIO_REGION_INFO_FLAG_READ
) ||
3315 !(vga_info
.flags
& VFIO_REGION_INFO_FLAG_WRITE
) ||
3316 vga_info
.size
< 0xbffff + 1) {
3317 error_report("vfio: Unexpected VGA info, flags 0x%lx, size 0x%lx",
3318 (unsigned long)vga_info
.flags
,
3319 (unsigned long)vga_info
.size
);
3323 vdev
->vga
.fd_offset
= vga_info
.offset
;
3324 vdev
->vga
.fd
= vdev
->fd
;
3326 vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].offset
= QEMU_PCI_VGA_MEM_BASE
;
3327 vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].nr
= QEMU_PCI_VGA_MEM
;
3328 QLIST_INIT(&vdev
->vga
.region
[QEMU_PCI_VGA_MEM
].quirks
);
3330 vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].offset
= QEMU_PCI_VGA_IO_LO_BASE
;
3331 vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].nr
= QEMU_PCI_VGA_IO_LO
;
3332 QLIST_INIT(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_LO
].quirks
);
3334 vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].offset
= QEMU_PCI_VGA_IO_HI_BASE
;
3335 vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].nr
= QEMU_PCI_VGA_IO_HI
;
3336 QLIST_INIT(&vdev
->vga
.region
[QEMU_PCI_VGA_IO_HI
].quirks
);
3338 vdev
->has_vga
= true;
3340 irq_info
.index
= VFIO_PCI_ERR_IRQ_INDEX
;
3342 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_GET_IRQ_INFO
, &irq_info
);
3344 /* This can fail for an old kernel or legacy PCI dev */
3345 DPRINTF("VFIO_DEVICE_GET_IRQ_INFO failure: %m\n");
3347 } else if (irq_info
.count
== 1) {
3348 vdev
->pci_aer
= true;
3350 error_report("vfio: %04x:%02x:%02x.%x "
3351 "Could not enable error recovery for the device",
3352 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
3353 vdev
->host
.function
);
3358 QLIST_REMOVE(vdev
, next
);
3365 static void vfio_put_device(VFIODevice
*vdev
)
3367 QLIST_REMOVE(vdev
, next
);
3369 DPRINTF("vfio_put_device: close vdev->fd\n");
3377 static void vfio_err_notifier_handler(void *opaque
)
3379 VFIODevice
*vdev
= opaque
;
3381 if (!event_notifier_test_and_clear(&vdev
->err_notifier
)) {
3386 * TBD. Retrieve the error details and decide what action
3387 * needs to be taken. One of the actions could be to pass
3388 * the error to the guest and have the guest driver recover
3389 * from the error. This requires that PCIe capabilities be
3390 * exposed to the guest. For now, we just terminate the
3391 * guest to contain the error.
3394 error_report("%s(%04x:%02x:%02x.%x) Unrecoverable error detected. "
3395 "Please collect any data possible and then kill the guest",
3396 __func__
, vdev
->host
.domain
, vdev
->host
.bus
,
3397 vdev
->host
.slot
, vdev
->host
.function
);
3399 vm_stop(RUN_STATE_IO_ERROR
);
3403 * Registers error notifier for devices supporting error recovery.
3404 * If we encounter a failure in this function, we report an error
3405 * and continue after disabling error recovery support for the
3408 static void vfio_register_err_notifier(VFIODevice
*vdev
)
3412 struct vfio_irq_set
*irq_set
;
3415 if (!vdev
->pci_aer
) {
3419 if (event_notifier_init(&vdev
->err_notifier
, 0)) {
3420 error_report("vfio: Unable to init event notifier for error detection");
3421 vdev
->pci_aer
= false;
3425 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
3427 irq_set
= g_malloc0(argsz
);
3428 irq_set
->argsz
= argsz
;
3429 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
3430 VFIO_IRQ_SET_ACTION_TRIGGER
;
3431 irq_set
->index
= VFIO_PCI_ERR_IRQ_INDEX
;
3434 pfd
= (int32_t *)&irq_set
->data
;
3436 *pfd
= event_notifier_get_fd(&vdev
->err_notifier
);
3437 qemu_set_fd_handler(*pfd
, vfio_err_notifier_handler
, NULL
, vdev
);
3439 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
3441 error_report("vfio: Failed to set up error notification");
3442 qemu_set_fd_handler(*pfd
, NULL
, NULL
, vdev
);
3443 event_notifier_cleanup(&vdev
->err_notifier
);
3444 vdev
->pci_aer
= false;
3449 static void vfio_unregister_err_notifier(VFIODevice
*vdev
)
3452 struct vfio_irq_set
*irq_set
;
3456 if (!vdev
->pci_aer
) {
3460 argsz
= sizeof(*irq_set
) + sizeof(*pfd
);
3462 irq_set
= g_malloc0(argsz
);
3463 irq_set
->argsz
= argsz
;
3464 irq_set
->flags
= VFIO_IRQ_SET_DATA_EVENTFD
|
3465 VFIO_IRQ_SET_ACTION_TRIGGER
;
3466 irq_set
->index
= VFIO_PCI_ERR_IRQ_INDEX
;
3469 pfd
= (int32_t *)&irq_set
->data
;
3472 ret
= ioctl(vdev
->fd
, VFIO_DEVICE_SET_IRQS
, irq_set
);
3474 error_report("vfio: Failed to de-assign error fd: %m");
3477 qemu_set_fd_handler(event_notifier_get_fd(&vdev
->err_notifier
),
3479 event_notifier_cleanup(&vdev
->err_notifier
);
3482 static int vfio_initfn(PCIDevice
*pdev
)
3484 VFIODevice
*pvdev
, *vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
3486 char path
[PATH_MAX
], iommu_group_path
[PATH_MAX
], *group_name
;
3492 /* Check that the host device exists */
3493 snprintf(path
, sizeof(path
),
3494 "/sys/bus/pci/devices/%04x:%02x:%02x.%01x/",
3495 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
3496 vdev
->host
.function
);
3497 if (stat(path
, &st
) < 0) {
3498 error_report("vfio: error: no such host device: %s", path
);
3502 strncat(path
, "iommu_group", sizeof(path
) - strlen(path
) - 1);
3504 len
= readlink(path
, iommu_group_path
, PATH_MAX
);
3506 error_report("vfio: error no iommu_group for device");
3510 iommu_group_path
[len
] = 0;
3511 group_name
= basename(iommu_group_path
);
3513 if (sscanf(group_name
, "%d", &groupid
) != 1) {
3514 error_report("vfio: error reading %s: %m", path
);
3518 DPRINTF("%s(%04x:%02x:%02x.%x) group %d\n", __func__
, vdev
->host
.domain
,
3519 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
, groupid
);
3521 group
= vfio_get_group(groupid
);
3523 error_report("vfio: failed to get group %d", groupid
);
3527 snprintf(path
, sizeof(path
), "%04x:%02x:%02x.%01x",
3528 vdev
->host
.domain
, vdev
->host
.bus
, vdev
->host
.slot
,
3529 vdev
->host
.function
);
3531 QLIST_FOREACH(pvdev
, &group
->device_list
, next
) {
3532 if (pvdev
->host
.domain
== vdev
->host
.domain
&&
3533 pvdev
->host
.bus
== vdev
->host
.bus
&&
3534 pvdev
->host
.slot
== vdev
->host
.slot
&&
3535 pvdev
->host
.function
== vdev
->host
.function
) {
3537 error_report("vfio: error: device %s is already attached", path
);
3538 vfio_put_group(group
);
3543 ret
= vfio_get_device(group
, path
, vdev
);
3545 error_report("vfio: failed to get device %s", path
);
3546 vfio_put_group(group
);
3550 /* Get a copy of config space */
3551 ret
= pread(vdev
->fd
, vdev
->pdev
.config
,
3552 MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
),
3553 vdev
->config_offset
);
3554 if (ret
< (int)MIN(pci_config_size(&vdev
->pdev
), vdev
->config_size
)) {
3555 ret
= ret
< 0 ? -errno
: -EFAULT
;
3556 error_report("vfio: Failed to read device config space");
3560 /* vfio emulates a lot for us, but some bits need extra love */
3561 vdev
->emulated_config_bits
= g_malloc0(vdev
->config_size
);
3563 /* QEMU can choose to expose the ROM or not */
3564 memset(vdev
->emulated_config_bits
+ PCI_ROM_ADDRESS
, 0xff, 4);
3566 /* QEMU can change multi-function devices to single function, or reverse */
3567 vdev
->emulated_config_bits
[PCI_HEADER_TYPE
] =
3568 PCI_HEADER_TYPE_MULTI_FUNCTION
;
3571 * Clear host resource mapping info. If we choose not to register a
3572 * BAR, such as might be the case with the option ROM, we can get
3573 * confusing, unwritable, residual addresses from the host here.
3575 memset(&vdev
->pdev
.config
[PCI_BASE_ADDRESS_0
], 0, 24);
3576 memset(&vdev
->pdev
.config
[PCI_ROM_ADDRESS
], 0, 4);
3578 vfio_pci_size_rom(vdev
);
3580 ret
= vfio_early_setup_msix(vdev
);
3585 vfio_map_bars(vdev
);
3587 ret
= vfio_add_capabilities(vdev
);
3592 /* QEMU emulates all of MSI & MSIX */
3593 if (pdev
->cap_present
& QEMU_PCI_CAP_MSIX
) {
3594 memset(vdev
->emulated_config_bits
+ pdev
->msix_cap
, 0xff,
3598 if (pdev
->cap_present
& QEMU_PCI_CAP_MSI
) {
3599 memset(vdev
->emulated_config_bits
+ pdev
->msi_cap
, 0xff,
3600 vdev
->msi_cap_size
);
3603 if (vfio_pci_read_config(&vdev
->pdev
, PCI_INTERRUPT_PIN
, 1)) {
3604 vdev
->intx
.mmap_timer
= timer_new_ms(QEMU_CLOCK_VIRTUAL
,
3605 vfio_intx_mmap_enable
, vdev
);
3606 pci_device_set_intx_routing_notifier(&vdev
->pdev
, vfio_update_irq
);
3607 ret
= vfio_enable_intx(vdev
);
3613 add_boot_device_path(vdev
->bootindex
, &pdev
->qdev
, NULL
);
3614 vfio_register_err_notifier(vdev
);
3619 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
3620 vfio_teardown_msi(vdev
);
3621 vfio_unmap_bars(vdev
);
3623 g_free(vdev
->emulated_config_bits
);
3624 vfio_put_device(vdev
);
3625 vfio_put_group(group
);
3629 static void vfio_exitfn(PCIDevice
*pdev
)
3631 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
3632 VFIOGroup
*group
= vdev
->group
;
3634 vfio_unregister_err_notifier(vdev
);
3635 pci_device_set_intx_routing_notifier(&vdev
->pdev
, NULL
);
3636 vfio_disable_interrupts(vdev
);
3637 if (vdev
->intx
.mmap_timer
) {
3638 timer_free(vdev
->intx
.mmap_timer
);
3640 vfio_teardown_msi(vdev
);
3641 vfio_unmap_bars(vdev
);
3642 g_free(vdev
->emulated_config_bits
);
3644 vfio_put_device(vdev
);
3645 vfio_put_group(group
);
3648 static void vfio_pci_reset(DeviceState
*dev
)
3650 PCIDevice
*pdev
= DO_UPCAST(PCIDevice
, qdev
, dev
);
3651 VFIODevice
*vdev
= DO_UPCAST(VFIODevice
, pdev
, pdev
);
3653 DPRINTF("%s(%04x:%02x:%02x.%x)\n", __func__
, vdev
->host
.domain
,
3654 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
3656 vfio_pci_pre_reset(vdev
);
3658 if (vdev
->reset_works
&& (vdev
->has_flr
|| !vdev
->has_pm_reset
) &&
3659 !ioctl(vdev
->fd
, VFIO_DEVICE_RESET
)) {
3660 DPRINTF("%04x:%02x:%02x.%x FLR/VFIO_DEVICE_RESET\n", vdev
->host
.domain
,
3661 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
3665 /* See if we can do our own bus reset */
3666 if (!vfio_pci_hot_reset_one(vdev
)) {
3670 /* If nothing else works and the device supports PM reset, use it */
3671 if (vdev
->reset_works
&& vdev
->has_pm_reset
&&
3672 !ioctl(vdev
->fd
, VFIO_DEVICE_RESET
)) {
3673 DPRINTF("%04x:%02x:%02x.%x PCI PM Reset\n", vdev
->host
.domain
,
3674 vdev
->host
.bus
, vdev
->host
.slot
, vdev
->host
.function
);
3679 vfio_pci_post_reset(vdev
);
3682 static Property vfio_pci_dev_properties
[] = {
3683 DEFINE_PROP_PCI_HOST_DEVADDR("host", VFIODevice
, host
),
3684 DEFINE_PROP_UINT32("x-intx-mmap-timeout-ms", VFIODevice
,
3685 intx
.mmap_timeout
, 1100),
3686 DEFINE_PROP_BIT("x-vga", VFIODevice
, features
,
3687 VFIO_FEATURE_ENABLE_VGA_BIT
, false),
3688 DEFINE_PROP_INT32("bootindex", VFIODevice
, bootindex
, -1),
3690 * TODO - support passed fds... is this necessary?
3691 * DEFINE_PROP_STRING("vfiofd", VFIODevice, vfiofd_name),
3692 * DEFINE_PROP_STRING("vfiogroupfd, VFIODevice, vfiogroupfd_name),
3694 DEFINE_PROP_END_OF_LIST(),
3697 static const VMStateDescription vfio_pci_vmstate
= {
3702 static void vfio_pci_dev_class_init(ObjectClass
*klass
, void *data
)
3704 DeviceClass
*dc
= DEVICE_CLASS(klass
);
3705 PCIDeviceClass
*pdc
= PCI_DEVICE_CLASS(klass
);
3707 dc
->reset
= vfio_pci_reset
;
3708 dc
->props
= vfio_pci_dev_properties
;
3709 dc
->vmsd
= &vfio_pci_vmstate
;
3710 dc
->desc
= "VFIO-based PCI device assignment";
3711 set_bit(DEVICE_CATEGORY_MISC
, dc
->categories
);
3712 pdc
->init
= vfio_initfn
;
3713 pdc
->exit
= vfio_exitfn
;
3714 pdc
->config_read
= vfio_pci_read_config
;
3715 pdc
->config_write
= vfio_pci_write_config
;
3716 pdc
->is_express
= 1; /* We might be */
3719 static const TypeInfo vfio_pci_dev_info
= {
3721 .parent
= TYPE_PCI_DEVICE
,
3722 .instance_size
= sizeof(VFIODevice
),
3723 .class_init
= vfio_pci_dev_class_init
,
3726 static void register_vfio_pci_dev_type(void)
3728 type_register_static(&vfio_pci_dev_info
);
3731 type_init(register_vfio_pci_dev_type
)