2 * VFIO PCI interrupt handling
4 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
5 * Author: Alex Williamson <alex.williamson@redhat.com>
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
11 * Derived from original vfio:
12 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 * Author: Tom Lyon, pugs@cisco.com
16 #include <linux/device.h>
17 #include <linux/interrupt.h>
18 #include <linux/eventfd.h>
19 #include <linux/msi.h>
20 #include <linux/pci.h>
21 #include <linux/file.h>
22 #include <linux/vfio.h>
23 #include <linux/wait.h>
24 #include <linux/slab.h>
26 #include "vfio_pci_private.h"
31 static void vfio_send_intx_eventfd(void *opaque
, void *unused
)
33 struct vfio_pci_device
*vdev
= opaque
;
35 if (likely(is_intx(vdev
) && !vdev
->virq_disabled
))
36 eventfd_signal(vdev
->ctx
[0].trigger
, 1);
39 void vfio_pci_intx_mask(struct vfio_pci_device
*vdev
)
41 struct pci_dev
*pdev
= vdev
->pdev
;
44 spin_lock_irqsave(&vdev
->irqlock
, flags
);
47 * Masking can come from interrupt, ioctl, or config space
48 * via INTx disable. The latter means this can get called
49 * even when not using intx delivery. In this case, just
50 * try to have the physical bit follow the virtual bit.
52 if (unlikely(!is_intx(vdev
))) {
55 } else if (!vdev
->ctx
[0].masked
) {
57 * Can't use check_and_mask here because we always want to
58 * mask, not just when something is pending.
63 disable_irq_nosync(pdev
->irq
);
65 vdev
->ctx
[0].masked
= true;
68 spin_unlock_irqrestore(&vdev
->irqlock
, flags
);
72 * If this is triggered by an eventfd, we can't call eventfd_signal
73 * or else we'll deadlock on the eventfd wait queue. Return >0 when
74 * a signal is necessary, which can then be handled via a work queue
75 * or directly depending on the caller.
77 static int vfio_pci_intx_unmask_handler(void *opaque
, void *unused
)
79 struct vfio_pci_device
*vdev
= opaque
;
80 struct pci_dev
*pdev
= vdev
->pdev
;
84 spin_lock_irqsave(&vdev
->irqlock
, flags
);
87 * Unmasking comes from ioctl or config, so again, have the
88 * physical bit follow the virtual even when not using INTx.
90 if (unlikely(!is_intx(vdev
))) {
93 } else if (vdev
->ctx
[0].masked
&& !vdev
->virq_disabled
) {
95 * A pending interrupt here would immediately trigger,
96 * but we can avoid that overhead by just re-sending
97 * the interrupt to the user.
100 if (!pci_check_and_unmask_intx(pdev
))
103 enable_irq(pdev
->irq
);
105 vdev
->ctx
[0].masked
= (ret
> 0);
108 spin_unlock_irqrestore(&vdev
->irqlock
, flags
);
113 void vfio_pci_intx_unmask(struct vfio_pci_device
*vdev
)
115 if (vfio_pci_intx_unmask_handler(vdev
, NULL
) > 0)
116 vfio_send_intx_eventfd(vdev
, NULL
);
119 static irqreturn_t
vfio_intx_handler(int irq
, void *dev_id
)
121 struct vfio_pci_device
*vdev
= dev_id
;
125 spin_lock_irqsave(&vdev
->irqlock
, flags
);
127 if (!vdev
->pci_2_3
) {
128 disable_irq_nosync(vdev
->pdev
->irq
);
129 vdev
->ctx
[0].masked
= true;
131 } else if (!vdev
->ctx
[0].masked
&& /* may be shared */
132 pci_check_and_mask_intx(vdev
->pdev
)) {
133 vdev
->ctx
[0].masked
= true;
137 spin_unlock_irqrestore(&vdev
->irqlock
, flags
);
139 if (ret
== IRQ_HANDLED
)
140 vfio_send_intx_eventfd(vdev
, NULL
);
145 static int vfio_intx_enable(struct vfio_pci_device
*vdev
)
147 if (!is_irq_none(vdev
))
150 if (!vdev
->pdev
->irq
)
153 vdev
->ctx
= kzalloc(sizeof(struct vfio_pci_irq_ctx
), GFP_KERNEL
);
160 * If the virtual interrupt is masked, restore it. Devices
161 * supporting DisINTx can be masked at the hardware level
162 * here, non-PCI-2.3 devices will have to wait until the
163 * interrupt is enabled.
165 vdev
->ctx
[0].masked
= vdev
->virq_disabled
;
167 pci_intx(vdev
->pdev
, !vdev
->ctx
[0].masked
);
169 vdev
->irq_type
= VFIO_PCI_INTX_IRQ_INDEX
;
174 static int vfio_intx_set_signal(struct vfio_pci_device
*vdev
, int fd
)
176 struct pci_dev
*pdev
= vdev
->pdev
;
177 unsigned long irqflags
= IRQF_SHARED
;
178 struct eventfd_ctx
*trigger
;
182 if (vdev
->ctx
[0].trigger
) {
183 free_irq(pdev
->irq
, vdev
);
184 kfree(vdev
->ctx
[0].name
);
185 eventfd_ctx_put(vdev
->ctx
[0].trigger
);
186 vdev
->ctx
[0].trigger
= NULL
;
189 if (fd
< 0) /* Disable only */
192 vdev
->ctx
[0].name
= kasprintf(GFP_KERNEL
, "vfio-intx(%s)",
194 if (!vdev
->ctx
[0].name
)
197 trigger
= eventfd_ctx_fdget(fd
);
198 if (IS_ERR(trigger
)) {
199 kfree(vdev
->ctx
[0].name
);
200 return PTR_ERR(trigger
);
203 vdev
->ctx
[0].trigger
= trigger
;
208 ret
= request_irq(pdev
->irq
, vfio_intx_handler
,
209 irqflags
, vdev
->ctx
[0].name
, vdev
);
211 vdev
->ctx
[0].trigger
= NULL
;
212 kfree(vdev
->ctx
[0].name
);
213 eventfd_ctx_put(trigger
);
218 * INTx disable will stick across the new irq setup,
221 spin_lock_irqsave(&vdev
->irqlock
, flags
);
222 if (!vdev
->pci_2_3
&& vdev
->ctx
[0].masked
)
223 disable_irq_nosync(pdev
->irq
);
224 spin_unlock_irqrestore(&vdev
->irqlock
, flags
);
229 static void vfio_intx_disable(struct vfio_pci_device
*vdev
)
231 vfio_virqfd_disable(&vdev
->ctx
[0].unmask
);
232 vfio_virqfd_disable(&vdev
->ctx
[0].mask
);
233 vfio_intx_set_signal(vdev
, -1);
234 vdev
->irq_type
= VFIO_PCI_NUM_IRQS
;
242 static irqreturn_t
vfio_msihandler(int irq
, void *arg
)
244 struct eventfd_ctx
*trigger
= arg
;
246 eventfd_signal(trigger
, 1);
250 static int vfio_msi_enable(struct vfio_pci_device
*vdev
, int nvec
, bool msix
)
252 struct pci_dev
*pdev
= vdev
->pdev
;
253 unsigned int flag
= msix
? PCI_IRQ_MSIX
: PCI_IRQ_MSI
;
256 if (!is_irq_none(vdev
))
259 vdev
->ctx
= kcalloc(nvec
, sizeof(struct vfio_pci_irq_ctx
), GFP_KERNEL
);
263 /* return the number of supported vectors if we can't get all: */
264 ret
= pci_alloc_irq_vectors(pdev
, 1, nvec
, flag
);
267 pci_free_irq_vectors(pdev
);
272 vdev
->num_ctx
= nvec
;
273 vdev
->irq_type
= msix
? VFIO_PCI_MSIX_IRQ_INDEX
:
274 VFIO_PCI_MSI_IRQ_INDEX
;
278 * Compute the virtual hardware field for max msi vectors -
279 * it is the log base 2 of the number of vectors.
281 vdev
->msi_qmax
= fls(nvec
* 2 - 1) - 1;
287 static int vfio_msi_set_vector_signal(struct vfio_pci_device
*vdev
,
288 int vector
, int fd
, bool msix
)
290 struct pci_dev
*pdev
= vdev
->pdev
;
291 struct eventfd_ctx
*trigger
;
294 if (vector
< 0 || vector
>= vdev
->num_ctx
)
297 irq
= pci_irq_vector(pdev
, vector
);
299 if (vdev
->ctx
[vector
].trigger
) {
300 free_irq(irq
, vdev
->ctx
[vector
].trigger
);
301 irq_bypass_unregister_producer(&vdev
->ctx
[vector
].producer
);
302 kfree(vdev
->ctx
[vector
].name
);
303 eventfd_ctx_put(vdev
->ctx
[vector
].trigger
);
304 vdev
->ctx
[vector
].trigger
= NULL
;
310 vdev
->ctx
[vector
].name
= kasprintf(GFP_KERNEL
, "vfio-msi%s[%d](%s)",
311 msix
? "x" : "", vector
,
313 if (!vdev
->ctx
[vector
].name
)
316 trigger
= eventfd_ctx_fdget(fd
);
317 if (IS_ERR(trigger
)) {
318 kfree(vdev
->ctx
[vector
].name
);
319 return PTR_ERR(trigger
);
323 * The MSIx vector table resides in device memory which may be cleared
324 * via backdoor resets. We don't allow direct access to the vector
325 * table so even if a userspace driver attempts to save/restore around
326 * such a reset it would be unsuccessful. To avoid this, restore the
327 * cached value of the message prior to enabling.
332 get_cached_msi_msg(irq
, &msg
);
333 pci_write_msi_msg(irq
, &msg
);
336 ret
= request_irq(irq
, vfio_msihandler
, 0,
337 vdev
->ctx
[vector
].name
, trigger
);
339 kfree(vdev
->ctx
[vector
].name
);
340 eventfd_ctx_put(trigger
);
344 vdev
->ctx
[vector
].producer
.token
= trigger
;
345 vdev
->ctx
[vector
].producer
.irq
= irq
;
346 ret
= irq_bypass_register_producer(&vdev
->ctx
[vector
].producer
);
349 "irq bypass producer (token %p) registration fails: %d\n",
350 vdev
->ctx
[vector
].producer
.token
, ret
);
352 vdev
->ctx
[vector
].trigger
= trigger
;
357 static int vfio_msi_set_block(struct vfio_pci_device
*vdev
, unsigned start
,
358 unsigned count
, int32_t *fds
, bool msix
)
362 if (start
>= vdev
->num_ctx
|| start
+ count
> vdev
->num_ctx
)
365 for (i
= 0, j
= start
; i
< count
&& !ret
; i
++, j
++) {
366 int fd
= fds
? fds
[i
] : -1;
367 ret
= vfio_msi_set_vector_signal(vdev
, j
, fd
, msix
);
371 for (--j
; j
>= (int)start
; j
--)
372 vfio_msi_set_vector_signal(vdev
, j
, -1, msix
);
378 static void vfio_msi_disable(struct vfio_pci_device
*vdev
, bool msix
)
380 struct pci_dev
*pdev
= vdev
->pdev
;
383 for (i
= 0; i
< vdev
->num_ctx
; i
++) {
384 vfio_virqfd_disable(&vdev
->ctx
[i
].unmask
);
385 vfio_virqfd_disable(&vdev
->ctx
[i
].mask
);
388 vfio_msi_set_block(vdev
, 0, vdev
->num_ctx
, NULL
, msix
);
390 pci_free_irq_vectors(pdev
);
393 * Both disable paths above use pci_intx_for_msi() to clear DisINTx
394 * via their shutdown paths. Restore for NoINTx devices.
399 vdev
->irq_type
= VFIO_PCI_NUM_IRQS
;
407 static int vfio_pci_set_intx_unmask(struct vfio_pci_device
*vdev
,
408 unsigned index
, unsigned start
,
409 unsigned count
, uint32_t flags
, void *data
)
411 if (!is_intx(vdev
) || start
!= 0 || count
!= 1)
414 if (flags
& VFIO_IRQ_SET_DATA_NONE
) {
415 vfio_pci_intx_unmask(vdev
);
416 } else if (flags
& VFIO_IRQ_SET_DATA_BOOL
) {
417 uint8_t unmask
= *(uint8_t *)data
;
419 vfio_pci_intx_unmask(vdev
);
420 } else if (flags
& VFIO_IRQ_SET_DATA_EVENTFD
) {
421 int32_t fd
= *(int32_t *)data
;
423 return vfio_virqfd_enable((void *) vdev
,
424 vfio_pci_intx_unmask_handler
,
425 vfio_send_intx_eventfd
, NULL
,
426 &vdev
->ctx
[0].unmask
, fd
);
428 vfio_virqfd_disable(&vdev
->ctx
[0].unmask
);
434 static int vfio_pci_set_intx_mask(struct vfio_pci_device
*vdev
,
435 unsigned index
, unsigned start
,
436 unsigned count
, uint32_t flags
, void *data
)
438 if (!is_intx(vdev
) || start
!= 0 || count
!= 1)
441 if (flags
& VFIO_IRQ_SET_DATA_NONE
) {
442 vfio_pci_intx_mask(vdev
);
443 } else if (flags
& VFIO_IRQ_SET_DATA_BOOL
) {
444 uint8_t mask
= *(uint8_t *)data
;
446 vfio_pci_intx_mask(vdev
);
447 } else if (flags
& VFIO_IRQ_SET_DATA_EVENTFD
) {
448 return -ENOTTY
; /* XXX implement me */
454 static int vfio_pci_set_intx_trigger(struct vfio_pci_device
*vdev
,
455 unsigned index
, unsigned start
,
456 unsigned count
, uint32_t flags
, void *data
)
458 if (is_intx(vdev
) && !count
&& (flags
& VFIO_IRQ_SET_DATA_NONE
)) {
459 vfio_intx_disable(vdev
);
463 if (!(is_intx(vdev
) || is_irq_none(vdev
)) || start
!= 0 || count
!= 1)
466 if (flags
& VFIO_IRQ_SET_DATA_EVENTFD
) {
467 int32_t fd
= *(int32_t *)data
;
471 return vfio_intx_set_signal(vdev
, fd
);
473 ret
= vfio_intx_enable(vdev
);
477 ret
= vfio_intx_set_signal(vdev
, fd
);
479 vfio_intx_disable(vdev
);
487 if (flags
& VFIO_IRQ_SET_DATA_NONE
) {
488 vfio_send_intx_eventfd(vdev
, NULL
);
489 } else if (flags
& VFIO_IRQ_SET_DATA_BOOL
) {
490 uint8_t trigger
= *(uint8_t *)data
;
492 vfio_send_intx_eventfd(vdev
, NULL
);
497 static int vfio_pci_set_msi_trigger(struct vfio_pci_device
*vdev
,
498 unsigned index
, unsigned start
,
499 unsigned count
, uint32_t flags
, void *data
)
502 bool msix
= (index
== VFIO_PCI_MSIX_IRQ_INDEX
) ? true : false;
504 if (irq_is(vdev
, index
) && !count
&& (flags
& VFIO_IRQ_SET_DATA_NONE
)) {
505 vfio_msi_disable(vdev
, msix
);
509 if (!(irq_is(vdev
, index
) || is_irq_none(vdev
)))
512 if (flags
& VFIO_IRQ_SET_DATA_EVENTFD
) {
516 if (vdev
->irq_type
== index
)
517 return vfio_msi_set_block(vdev
, start
, count
,
520 ret
= vfio_msi_enable(vdev
, start
+ count
, msix
);
524 ret
= vfio_msi_set_block(vdev
, start
, count
, fds
, msix
);
526 vfio_msi_disable(vdev
, msix
);
531 if (!irq_is(vdev
, index
) || start
+ count
> vdev
->num_ctx
)
534 for (i
= start
; i
< start
+ count
; i
++) {
535 if (!vdev
->ctx
[i
].trigger
)
537 if (flags
& VFIO_IRQ_SET_DATA_NONE
) {
538 eventfd_signal(vdev
->ctx
[i
].trigger
, 1);
539 } else if (flags
& VFIO_IRQ_SET_DATA_BOOL
) {
540 uint8_t *bools
= data
;
541 if (bools
[i
- start
])
542 eventfd_signal(vdev
->ctx
[i
].trigger
, 1);
548 static int vfio_pci_set_ctx_trigger_single(struct eventfd_ctx
**ctx
,
549 unsigned int count
, uint32_t flags
,
552 /* DATA_NONE/DATA_BOOL enables loopback testing */
553 if (flags
& VFIO_IRQ_SET_DATA_NONE
) {
556 eventfd_signal(*ctx
, 1);
558 eventfd_ctx_put(*ctx
);
563 } else if (flags
& VFIO_IRQ_SET_DATA_BOOL
) {
569 trigger
= *(uint8_t *)data
;
571 eventfd_signal(*ctx
, 1);
574 } else if (flags
& VFIO_IRQ_SET_DATA_EVENTFD
) {
580 fd
= *(int32_t *)data
;
583 eventfd_ctx_put(*ctx
);
585 } else if (fd
>= 0) {
586 struct eventfd_ctx
*efdctx
;
588 efdctx
= eventfd_ctx_fdget(fd
);
590 return PTR_ERR(efdctx
);
593 eventfd_ctx_put(*ctx
);
603 static int vfio_pci_set_err_trigger(struct vfio_pci_device
*vdev
,
604 unsigned index
, unsigned start
,
605 unsigned count
, uint32_t flags
, void *data
)
607 if (index
!= VFIO_PCI_ERR_IRQ_INDEX
|| start
!= 0 || count
> 1)
610 return vfio_pci_set_ctx_trigger_single(&vdev
->err_trigger
,
614 static int vfio_pci_set_req_trigger(struct vfio_pci_device
*vdev
,
615 unsigned index
, unsigned start
,
616 unsigned count
, uint32_t flags
, void *data
)
618 if (index
!= VFIO_PCI_REQ_IRQ_INDEX
|| start
!= 0 || count
> 1)
621 return vfio_pci_set_ctx_trigger_single(&vdev
->req_trigger
,
625 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device
*vdev
, uint32_t flags
,
626 unsigned index
, unsigned start
, unsigned count
,
629 int (*func
)(struct vfio_pci_device
*vdev
, unsigned index
,
630 unsigned start
, unsigned count
, uint32_t flags
,
634 case VFIO_PCI_INTX_IRQ_INDEX
:
635 switch (flags
& VFIO_IRQ_SET_ACTION_TYPE_MASK
) {
636 case VFIO_IRQ_SET_ACTION_MASK
:
637 func
= vfio_pci_set_intx_mask
;
639 case VFIO_IRQ_SET_ACTION_UNMASK
:
640 func
= vfio_pci_set_intx_unmask
;
642 case VFIO_IRQ_SET_ACTION_TRIGGER
:
643 func
= vfio_pci_set_intx_trigger
;
647 case VFIO_PCI_MSI_IRQ_INDEX
:
648 case VFIO_PCI_MSIX_IRQ_INDEX
:
649 switch (flags
& VFIO_IRQ_SET_ACTION_TYPE_MASK
) {
650 case VFIO_IRQ_SET_ACTION_MASK
:
651 case VFIO_IRQ_SET_ACTION_UNMASK
:
652 /* XXX Need masking support exported */
654 case VFIO_IRQ_SET_ACTION_TRIGGER
:
655 func
= vfio_pci_set_msi_trigger
;
659 case VFIO_PCI_ERR_IRQ_INDEX
:
660 switch (flags
& VFIO_IRQ_SET_ACTION_TYPE_MASK
) {
661 case VFIO_IRQ_SET_ACTION_TRIGGER
:
662 if (pci_is_pcie(vdev
->pdev
))
663 func
= vfio_pci_set_err_trigger
;
667 case VFIO_PCI_REQ_IRQ_INDEX
:
668 switch (flags
& VFIO_IRQ_SET_ACTION_TYPE_MASK
) {
669 case VFIO_IRQ_SET_ACTION_TRIGGER
:
670 func
= vfio_pci_set_req_trigger
;
679 return func(vdev
, index
, start
, count
, flags
, data
);