2 * vfio based device assignment support - platform devices
4 * Copyright Linaro Limited, 2014
7 * Kim Phillips <kim.phillips@linaro.org>
8 * Eric Auger <eric.auger@linaro.org>
10 * This work is licensed under the terms of the GNU GPL, version 2. See
11 * the COPYING file in the top-level directory.
13 * Based on vfio based PCI device assignment support:
14 * Copyright Red Hat, Inc. 2012
17 #include "qemu/osdep.h"
18 #include "qapi/error.h"
19 #include <sys/ioctl.h>
20 #include <linux/vfio.h>
22 #include "hw/vfio/vfio-platform.h"
23 #include "migration/vmstate.h"
24 #include "qemu/error-report.h"
25 #include "qemu/lockable.h"
26 #include "qemu/main-loop.h"
27 #include "qemu/module.h"
28 #include "qemu/range.h"
29 #include "exec/memory.h"
30 #include "exec/address-spaces.h"
31 #include "qemu/queue.h"
32 #include "hw/sysbus.h"
35 #include "hw/platform-bus.h"
36 #include "hw/qdev-properties.h"
37 #include "sysemu/kvm.h"
40 * Functions used whatever the injection method
43 static inline bool vfio_irq_is_automasked(VFIOINTp
*intp
)
45 return intp
->flags
& VFIO_IRQ_INFO_AUTOMASKED
;
49 * vfio_init_intp - allocate, initialize the IRQ struct pointer
50 * and add it into the list of IRQs
51 * @vbasedev: the VFIO device handle
52 * @info: irq info struct retrieved from VFIO driver
55 static VFIOINTp
*vfio_init_intp(VFIODevice
*vbasedev
,
56 struct vfio_irq_info info
, Error
**errp
)
59 VFIOPlatformDevice
*vdev
=
60 container_of(vbasedev
, VFIOPlatformDevice
, vbasedev
);
61 SysBusDevice
*sbdev
= SYS_BUS_DEVICE(vdev
);
64 intp
= g_malloc0(sizeof(*intp
));
66 intp
->pin
= info
.index
;
67 intp
->flags
= info
.flags
;
68 intp
->state
= VFIO_IRQ_INACTIVE
;
69 intp
->kvm_accel
= false;
71 sysbus_init_irq(sbdev
, &intp
->qemuirq
);
73 /* Get an eventfd for trigger */
74 intp
->interrupt
= g_new0(EventNotifier
, 1);
75 ret
= event_notifier_init(intp
->interrupt
, 0);
77 g_free(intp
->interrupt
);
79 error_setg_errno(errp
, -ret
,
80 "failed to initialize trigger eventfd notifier");
83 if (vfio_irq_is_automasked(intp
)) {
84 /* Get an eventfd for resample/unmask */
85 intp
->unmask
= g_new0(EventNotifier
, 1);
86 ret
= event_notifier_init(intp
->unmask
, 0);
88 g_free(intp
->interrupt
);
91 error_setg_errno(errp
, -ret
,
92 "failed to initialize resample eventfd notifier");
97 QLIST_INSERT_HEAD(&vdev
->intp_list
, intp
, next
);
102 * vfio_set_trigger_eventfd - set VFIO eventfd handling
104 * @intp: IRQ struct handle
105 * @handler: handler to be called on eventfd signaling
107 * Setup VFIO signaling and attach an optional user-side handler
110 static int vfio_set_trigger_eventfd(VFIOINTp
*intp
,
111 eventfd_user_side_handler_t handler
)
113 VFIODevice
*vbasedev
= &intp
->vdev
->vbasedev
;
114 int32_t fd
= event_notifier_get_fd(intp
->interrupt
);
118 qemu_set_fd_handler(fd
, (IOHandler
*)handler
, NULL
, intp
);
120 ret
= vfio_set_irq_signaling(vbasedev
, intp
->pin
, 0,
121 VFIO_IRQ_SET_ACTION_TRIGGER
, fd
, &err
);
123 error_reportf_err(err
, VFIO_MSG_PREFIX
, vbasedev
->name
);
124 qemu_set_fd_handler(fd
, NULL
, NULL
, NULL
);
131 * Functions only used when eventfds are handled on user-side
136 * vfio_mmap_set_enabled - enable/disable the fast path mode
137 * @vdev: the VFIO platform device
138 * @enabled: the target mmap state
140 * enabled = true ~ fast path = MMIO region is mmaped (no KVM TRAP);
141 * enabled = false ~ slow path = MMIO region is trapped and region callbacks
142 * are called; slow path enables to trap the device IRQ status register reset
145 static void vfio_mmap_set_enabled(VFIOPlatformDevice
*vdev
, bool enabled
)
149 for (i
= 0; i
< vdev
->vbasedev
.num_regions
; i
++) {
150 vfio_region_mmaps_set_enabled(vdev
->regions
[i
], enabled
);
155 * vfio_intp_mmap_enable - timer function, restores the fast path
156 * if there is no more active IRQ
157 * @opaque: actually points to the VFIO platform device
159 * Called on mmap timer timeout, this function checks whether the
160 * IRQ is still active and if not, restores the fast path.
161 * by construction a single eventfd is handled at a time.
162 * if the IRQ is still active, the timer is re-programmed.
164 static void vfio_intp_mmap_enable(void *opaque
)
167 VFIOPlatformDevice
*vdev
= (VFIOPlatformDevice
*)opaque
;
169 QEMU_LOCK_GUARD(&vdev
->intp_mutex
);
170 QLIST_FOREACH(tmp
, &vdev
->intp_list
, next
) {
171 if (tmp
->state
== VFIO_IRQ_ACTIVE
) {
172 trace_vfio_platform_intp_mmap_enable(tmp
->pin
);
173 /* re-program the timer to check active status later */
174 timer_mod(vdev
->mmap_timer
,
175 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) +
180 vfio_mmap_set_enabled(vdev
, true);
184 * vfio_intp_inject_pending_lockheld - Injects a pending IRQ
185 * @opaque: opaque pointer, in practice the VFIOINTp handle
187 * The function is called on a previous IRQ completion, from
188 * vfio_platform_eoi, while the intp_mutex is locked.
189 * Also in such situation, the slow path already is set and
190 * the mmap timer was already programmed.
192 static void vfio_intp_inject_pending_lockheld(VFIOINTp
*intp
)
194 trace_vfio_platform_intp_inject_pending_lockheld(intp
->pin
,
195 event_notifier_get_fd(intp
->interrupt
));
197 intp
->state
= VFIO_IRQ_ACTIVE
;
199 /* trigger the virtual IRQ */
200 qemu_set_irq(intp
->qemuirq
, 1);
204 * vfio_intp_interrupt - The user-side eventfd handler
205 * @opaque: opaque pointer which in practice is the VFIOINTp handle
207 * the function is entered in event handler context:
208 * the vIRQ is injected into the guest if there is no other active
211 static void vfio_intp_interrupt(VFIOINTp
*intp
)
215 VFIOPlatformDevice
*vdev
= intp
->vdev
;
216 bool delay_handling
= false;
218 QEMU_LOCK_GUARD(&vdev
->intp_mutex
);
219 if (intp
->state
== VFIO_IRQ_INACTIVE
) {
220 QLIST_FOREACH(tmp
, &vdev
->intp_list
, next
) {
221 if (tmp
->state
== VFIO_IRQ_ACTIVE
||
222 tmp
->state
== VFIO_IRQ_PENDING
) {
223 delay_handling
= true;
228 if (delay_handling
) {
230 * the new IRQ gets a pending status and is pushed in
233 intp
->state
= VFIO_IRQ_PENDING
;
234 trace_vfio_intp_interrupt_set_pending(intp
->pin
);
235 QSIMPLEQ_INSERT_TAIL(&vdev
->pending_intp_queue
,
237 event_notifier_test_and_clear(intp
->interrupt
);
241 trace_vfio_platform_intp_interrupt(intp
->pin
,
242 event_notifier_get_fd(intp
->interrupt
));
244 ret
= event_notifier_test_and_clear(intp
->interrupt
);
246 error_report("Error when clearing fd=%d (ret = %d)",
247 event_notifier_get_fd(intp
->interrupt
), ret
);
250 intp
->state
= VFIO_IRQ_ACTIVE
;
253 vfio_mmap_set_enabled(vdev
, false);
255 /* trigger the virtual IRQ */
256 qemu_set_irq(intp
->qemuirq
, 1);
259 * Schedule the mmap timer which will restore fastpath when no IRQ
262 if (vdev
->mmap_timeout
) {
263 timer_mod(vdev
->mmap_timer
,
264 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL
) +
270 * vfio_platform_eoi - IRQ completion routine
271 * @vbasedev: the VFIO device handle
273 * De-asserts the active virtual IRQ and unmasks the physical IRQ
274 * (effective for level sensitive IRQ auto-masked by the VFIO driver).
275 * Then it handles next pending IRQ if any.
276 * eoi function is called on the first access to any MMIO region
277 * after an IRQ was triggered, trapped since slow path was set.
278 * It is assumed this access corresponds to the IRQ status
279 * register reset. With such a mechanism, a single IRQ can be
280 * handled at a time since there is no way to know which IRQ
281 * was completed by the guest (we would need additional details
282 * about the IRQ status register mask).
284 static void vfio_platform_eoi(VFIODevice
*vbasedev
)
287 VFIOPlatformDevice
*vdev
=
288 container_of(vbasedev
, VFIOPlatformDevice
, vbasedev
);
290 QEMU_LOCK_GUARD(&vdev
->intp_mutex
);
291 QLIST_FOREACH(intp
, &vdev
->intp_list
, next
) {
292 if (intp
->state
== VFIO_IRQ_ACTIVE
) {
293 trace_vfio_platform_eoi(intp
->pin
,
294 event_notifier_get_fd(intp
->interrupt
));
295 intp
->state
= VFIO_IRQ_INACTIVE
;
297 /* deassert the virtual IRQ */
298 qemu_set_irq(intp
->qemuirq
, 0);
300 if (vfio_irq_is_automasked(intp
)) {
301 /* unmasks the physical level-sensitive IRQ */
302 vfio_unmask_single_irqindex(vbasedev
, intp
->pin
);
305 /* a single IRQ can be active at a time */
309 /* in case there are pending IRQs, handle the first one */
310 if (!QSIMPLEQ_EMPTY(&vdev
->pending_intp_queue
)) {
311 intp
= QSIMPLEQ_FIRST(&vdev
->pending_intp_queue
);
312 vfio_intp_inject_pending_lockheld(intp
);
313 QSIMPLEQ_REMOVE_HEAD(&vdev
->pending_intp_queue
, pqnext
);
318 * vfio_start_eventfd_injection - starts the virtual IRQ injection using
319 * user-side handled eventfds
320 * @sbdev: the sysbus device handle
321 * @irq: the qemu irq handle
324 static void vfio_start_eventfd_injection(SysBusDevice
*sbdev
, qemu_irq irq
)
326 VFIOPlatformDevice
*vdev
= VFIO_PLATFORM_DEVICE(sbdev
);
329 QLIST_FOREACH(intp
, &vdev
->intp_list
, next
) {
330 if (intp
->qemuirq
== irq
) {
336 if (vfio_set_trigger_eventfd(intp
, vfio_intp_interrupt
)) {
342 * Functions used for irqfd
346 * vfio_set_resample_eventfd - sets the resamplefd for an IRQ
347 * @intp: the IRQ struct handle
348 * programs the VFIO driver to unmask this IRQ when the
349 * intp->unmask eventfd is triggered
351 static int vfio_set_resample_eventfd(VFIOINTp
*intp
)
353 int32_t fd
= event_notifier_get_fd(intp
->unmask
);
354 VFIODevice
*vbasedev
= &intp
->vdev
->vbasedev
;
358 qemu_set_fd_handler(fd
, NULL
, NULL
, NULL
);
359 ret
= vfio_set_irq_signaling(vbasedev
, intp
->pin
, 0,
360 VFIO_IRQ_SET_ACTION_UNMASK
, fd
, &err
);
362 error_reportf_err(err
, VFIO_MSG_PREFIX
, vbasedev
->name
);
368 * vfio_start_irqfd_injection - starts the virtual IRQ injection using
371 * @sbdev: the sysbus device handle
372 * @irq: the qemu irq handle
374 * In case the irqfd setup fails, we fallback to userspace handled eventfd
376 static void vfio_start_irqfd_injection(SysBusDevice
*sbdev
, qemu_irq irq
)
378 VFIOPlatformDevice
*vdev
= VFIO_PLATFORM_DEVICE(sbdev
);
381 if (!kvm_irqfds_enabled() || !kvm_resamplefds_enabled() ||
382 !vdev
->irqfd_allowed
) {
386 QLIST_FOREACH(intp
, &vdev
->intp_list
, next
) {
387 if (intp
->qemuirq
== irq
) {
393 if (kvm_irqchip_add_irqfd_notifier(kvm_state
, intp
->interrupt
,
394 intp
->unmask
, irq
) < 0) {
398 if (vfio_set_trigger_eventfd(intp
, NULL
) < 0) {
401 if (vfio_irq_is_automasked(intp
)) {
402 if (vfio_set_resample_eventfd(intp
) < 0) {
405 trace_vfio_platform_start_level_irqfd_injection(intp
->pin
,
406 event_notifier_get_fd(intp
->interrupt
),
407 event_notifier_get_fd(intp
->unmask
));
409 trace_vfio_platform_start_edge_irqfd_injection(intp
->pin
,
410 event_notifier_get_fd(intp
->interrupt
));
413 intp
->kvm_accel
= true;
417 kvm_irqchip_remove_irqfd_notifier(kvm_state
, intp
->interrupt
, irq
);
420 vfio_start_eventfd_injection(sbdev
, irq
);
426 static void vfio_platform_compute_needs_reset(VFIODevice
*vbasedev
)
428 vbasedev
->needs_reset
= true;
431 /* not implemented yet */
432 static int vfio_platform_hot_reset_multi(VFIODevice
*vbasedev
)
438 * vfio_populate_device - Allocate and populate MMIO region
439 * and IRQ structs according to driver returned information
440 * @vbasedev: the VFIO device handle
441 * @errp: error object
444 static int vfio_populate_device(VFIODevice
*vbasedev
, Error
**errp
)
446 VFIOINTp
*intp
, *tmp
;
448 VFIOPlatformDevice
*vdev
=
449 container_of(vbasedev
, VFIOPlatformDevice
, vbasedev
);
451 if (!(vbasedev
->flags
& VFIO_DEVICE_FLAGS_PLATFORM
)) {
452 error_setg(errp
, "this isn't a platform device");
456 vdev
->regions
= g_new0(VFIORegion
*, vbasedev
->num_regions
);
458 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
459 char *name
= g_strdup_printf("VFIO %s region %d\n", vbasedev
->name
, i
);
461 vdev
->regions
[i
] = g_new0(VFIORegion
, 1);
462 ret
= vfio_region_setup(OBJECT(vdev
), vbasedev
,
463 vdev
->regions
[i
], i
, name
);
466 error_setg_errno(errp
, -ret
, "failed to get region %d info", i
);
471 vdev
->mmap_timer
= timer_new_ms(QEMU_CLOCK_VIRTUAL
,
472 vfio_intp_mmap_enable
, vdev
);
474 QSIMPLEQ_INIT(&vdev
->pending_intp_queue
);
476 for (i
= 0; i
< vbasedev
->num_irqs
; i
++) {
477 struct vfio_irq_info irq
= { .argsz
= sizeof(irq
) };
480 ret
= ioctl(vbasedev
->fd
, VFIO_DEVICE_GET_IRQ_INFO
, &irq
);
482 error_setg_errno(errp
, -ret
, "failed to get device irq info");
485 trace_vfio_platform_populate_interrupts(irq
.index
,
488 intp
= vfio_init_intp(vbasedev
, irq
, errp
);
497 timer_del(vdev
->mmap_timer
);
498 QLIST_FOREACH_SAFE(intp
, &vdev
->intp_list
, next
, tmp
) {
499 QLIST_REMOVE(intp
, next
);
503 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
504 if (vdev
->regions
[i
]) {
505 vfio_region_finalize(vdev
->regions
[i
]);
507 g_free(vdev
->regions
[i
]);
509 g_free(vdev
->regions
);
513 /* specialized functions for VFIO Platform devices */
514 static VFIODeviceOps vfio_platform_ops
= {
515 .vfio_compute_needs_reset
= vfio_platform_compute_needs_reset
,
516 .vfio_hot_reset_multi
= vfio_platform_hot_reset_multi
,
517 .vfio_eoi
= vfio_platform_eoi
,
521 * vfio_base_device_init - perform preliminary VFIO setup
522 * @vbasedev: the VFIO device handle
523 * @errp: error object
525 * Implement the VFIO command sequence that allows to discover
526 * assigned device resources: group extraction, device
527 * fd retrieval, resource query.
528 * Precondition: the device name must be initialized
530 static int vfio_base_device_init(VFIODevice
*vbasedev
, Error
**errp
)
533 VFIODevice
*vbasedev_iter
;
534 char *tmp
, group_path
[PATH_MAX
], *group_name
;
540 /* @sysfsdev takes precedence over @host */
541 if (vbasedev
->sysfsdev
) {
542 g_free(vbasedev
->name
);
543 vbasedev
->name
= g_path_get_basename(vbasedev
->sysfsdev
);
545 if (!vbasedev
->name
|| strchr(vbasedev
->name
, '/')) {
546 error_setg(errp
, "wrong host device name");
550 vbasedev
->sysfsdev
= g_strdup_printf("/sys/bus/platform/devices/%s",
554 if (stat(vbasedev
->sysfsdev
, &st
) < 0) {
555 error_setg_errno(errp
, errno
,
556 "failed to get the sysfs host device file status");
560 tmp
= g_strdup_printf("%s/iommu_group", vbasedev
->sysfsdev
);
561 len
= readlink(tmp
, group_path
, sizeof(group_path
));
564 if (len
< 0 || len
>= sizeof(group_path
)) {
565 ret
= len
< 0 ? -errno
: -ENAMETOOLONG
;
566 error_setg_errno(errp
, -ret
, "no iommu_group found");
572 group_name
= basename(group_path
);
573 if (sscanf(group_name
, "%d", &groupid
) != 1) {
574 error_setg_errno(errp
, errno
, "failed to read %s", group_path
);
578 trace_vfio_platform_base_device_init(vbasedev
->name
, groupid
);
580 group
= vfio_get_group(groupid
, &address_space_memory
, errp
);
585 QLIST_FOREACH(vbasedev_iter
, &group
->device_list
, next
) {
586 if (strcmp(vbasedev_iter
->name
, vbasedev
->name
) == 0) {
587 error_setg(errp
, "device is already attached");
588 vfio_put_group(group
);
592 ret
= vfio_get_device(group
, vbasedev
->name
, vbasedev
, errp
);
594 vfio_put_group(group
);
598 ret
= vfio_populate_device(vbasedev
, errp
);
600 vfio_put_group(group
);
607 * vfio_platform_realize - the device realize function
608 * @dev: device state pointer
611 * initialize the device, its memory regions and IRQ structures
612 * IRQ are started separately
614 static void vfio_platform_realize(DeviceState
*dev
, Error
**errp
)
616 VFIOPlatformDevice
*vdev
= VFIO_PLATFORM_DEVICE(dev
);
617 SysBusDevice
*sbdev
= SYS_BUS_DEVICE(dev
);
618 VFIODevice
*vbasedev
= &vdev
->vbasedev
;
621 vbasedev
->type
= VFIO_DEVICE_TYPE_PLATFORM
;
623 vbasedev
->ops
= &vfio_platform_ops
;
625 qemu_mutex_init(&vdev
->intp_mutex
);
627 trace_vfio_platform_realize(vbasedev
->sysfsdev
?
628 vbasedev
->sysfsdev
: vbasedev
->name
,
631 ret
= vfio_base_device_init(vbasedev
, errp
);
642 path
= g_strdup_printf("%s/of_node/compatible", vbasedev
->sysfsdev
);
643 if (!g_file_get_contents(path
, &contents
, &length
, &gerr
)) {
644 error_setg(errp
, "%s", gerr
->message
);
650 vdev
->compat
= contents
;
651 for (vdev
->num_compat
= 0; length
; vdev
->num_compat
++) {
652 size_t skip
= strlen(contents
) + 1;
658 for (i
= 0; i
< vbasedev
->num_regions
; i
++) {
659 if (vfio_region_mmap(vdev
->regions
[i
])) {
660 warn_report("%s mmap unsupported, performance may be slow",
661 memory_region_name(vdev
->regions
[i
]->mem
));
663 sysbus_init_mmio(sbdev
, vdev
->regions
[i
]->mem
);
670 if (vdev
->vbasedev
.name
) {
671 error_prepend(errp
, VFIO_MSG_PREFIX
, vdev
->vbasedev
.name
);
673 error_prepend(errp
, "vfio error: ");
677 static const VMStateDescription vfio_platform_vmstate
= {
678 .name
= "vfio-platform",
682 static Property vfio_platform_dev_properties
[] = {
683 DEFINE_PROP_STRING("host", VFIOPlatformDevice
, vbasedev
.name
),
684 DEFINE_PROP_STRING("sysfsdev", VFIOPlatformDevice
, vbasedev
.sysfsdev
),
685 DEFINE_PROP_BOOL("x-no-mmap", VFIOPlatformDevice
, vbasedev
.no_mmap
, false),
686 DEFINE_PROP_UINT32("mmap-timeout-ms", VFIOPlatformDevice
,
688 DEFINE_PROP_BOOL("x-irqfd", VFIOPlatformDevice
, irqfd_allowed
, true),
689 DEFINE_PROP_END_OF_LIST(),
692 static void vfio_platform_class_init(ObjectClass
*klass
, void *data
)
694 DeviceClass
*dc
= DEVICE_CLASS(klass
);
695 SysBusDeviceClass
*sbc
= SYS_BUS_DEVICE_CLASS(klass
);
697 dc
->realize
= vfio_platform_realize
;
698 device_class_set_props(dc
, vfio_platform_dev_properties
);
699 dc
->vmsd
= &vfio_platform_vmstate
;
700 dc
->desc
= "VFIO-based platform device assignment";
701 sbc
->connect_irq_notifier
= vfio_start_irqfd_injection
;
702 set_bit(DEVICE_CATEGORY_MISC
, dc
->categories
);
703 /* Supported by TYPE_VIRT_MACHINE */
704 dc
->user_creatable
= true;
707 static const TypeInfo vfio_platform_dev_info
= {
708 .name
= TYPE_VFIO_PLATFORM
,
709 .parent
= TYPE_SYS_BUS_DEVICE
,
710 .instance_size
= sizeof(VFIOPlatformDevice
),
711 .class_init
= vfio_platform_class_init
,
712 .class_size
= sizeof(VFIOPlatformDeviceClass
),
715 static void register_vfio_platform_dev_type(void)
717 type_register_static(&vfio_platform_dev_info
);
720 type_init(register_vfio_platform_dev_type
)