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/pci.h>
20 #include <linux/file.h>
21 #include <linux/poll.h>
22 #include <linux/vfio.h>
23 #include <linux/wait.h>
24 #include <linux/workqueue.h>
25 #include <linux/slab.h>
27 #include "vfio_pci_private.h"
33 struct vfio_pci_device *vdev;
34 struct eventfd_ctx *eventfd;
35 int (*handler)(struct vfio_pci_device *, void *);
36 void (*thread)(struct vfio_pci_device *, void *);
38 struct work_struct inject;
41 struct work_struct shutdown;
42 struct virqfd **pvirqfd;
45 static struct workqueue_struct *vfio_irqfd_cleanup_wq;
47 int __init vfio_pci_virqfd_init(void)
49 vfio_irqfd_cleanup_wq =
50 create_singlethread_workqueue("vfio-irqfd-cleanup");
51 if (!vfio_irqfd_cleanup_wq)
57 void vfio_pci_virqfd_exit(void)
59 destroy_workqueue(vfio_irqfd_cleanup_wq);
62 static void virqfd_deactivate(struct virqfd *virqfd)
64 queue_work(vfio_irqfd_cleanup_wq, &virqfd->shutdown);
67 static int virqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
69 struct virqfd *virqfd = container_of(wait, struct virqfd, wait);
70 unsigned long flags = (unsigned long)key;
73 /* An event has been signaled, call function */
74 if ((!virqfd->handler ||
75 virqfd->handler(virqfd->vdev, virqfd->data)) &&
77 schedule_work(&virqfd->inject);
80 if (flags & POLLHUP) {
82 spin_lock_irqsave(&virqfd->vdev->irqlock, flags);
85 * The eventfd is closing, if the virqfd has not yet been
86 * queued for release, as determined by testing whether the
87 * vdev pointer to it is still valid, queue it now. As
88 * with kvm irqfds, we know we won't race against the virqfd
89 * going away because we hold wqh->lock to get here.
91 if (*(virqfd->pvirqfd) == virqfd) {
92 *(virqfd->pvirqfd) = NULL;
93 virqfd_deactivate(virqfd);
96 spin_unlock_irqrestore(&virqfd->vdev->irqlock, flags);
102 static void virqfd_ptable_queue_proc(struct file *file,
103 wait_queue_head_t *wqh, poll_table *pt)
105 struct virqfd *virqfd = container_of(pt, struct virqfd, pt);
106 add_wait_queue(wqh, &virqfd->wait);
109 static void virqfd_shutdown(struct work_struct *work)
111 struct virqfd *virqfd = container_of(work, struct virqfd, shutdown);
114 eventfd_ctx_remove_wait_queue(virqfd->eventfd, &virqfd->wait, &cnt);
115 flush_work(&virqfd->inject);
116 eventfd_ctx_put(virqfd->eventfd);
121 static void virqfd_inject(struct work_struct *work)
123 struct virqfd *virqfd = container_of(work, struct virqfd, inject);
125 virqfd->thread(virqfd->vdev, virqfd->data);
128 static int virqfd_enable(struct vfio_pci_device *vdev,
129 int (*handler)(struct vfio_pci_device *, void *),
130 void (*thread)(struct vfio_pci_device *, void *),
131 void *data, struct virqfd **pvirqfd, int fd)
134 struct eventfd_ctx *ctx;
135 struct virqfd *virqfd;
139 virqfd = kzalloc(sizeof(*virqfd), GFP_KERNEL);
143 virqfd->pvirqfd = pvirqfd;
145 virqfd->handler = handler;
146 virqfd->thread = thread;
149 INIT_WORK(&virqfd->shutdown, virqfd_shutdown);
150 INIT_WORK(&virqfd->inject, virqfd_inject);
158 ctx = eventfd_ctx_fileget(irqfd.file);
164 virqfd->eventfd = ctx;
167 * virqfds can be released by closing the eventfd or directly
168 * through ioctl. These are both done through a workqueue, so
169 * we update the pointer to the virqfd under lock to avoid
170 * pushing multiple jobs to release the same virqfd.
172 spin_lock_irq(&vdev->irqlock);
175 spin_unlock_irq(&vdev->irqlock);
181 spin_unlock_irq(&vdev->irqlock);
184 * Install our own custom wake-up handling so we are notified via
185 * a callback whenever someone signals the underlying eventfd.
187 init_waitqueue_func_entry(&virqfd->wait, virqfd_wakeup);
188 init_poll_funcptr(&virqfd->pt, virqfd_ptable_queue_proc);
190 events = irqfd.file->f_op->poll(irqfd.file, &virqfd->pt);
193 * Check if there was an event already pending on the eventfd
194 * before we registered and trigger it as if we didn't miss it.
196 if (events & POLLIN) {
197 if ((!handler || handler(vdev, data)) && thread)
198 schedule_work(&virqfd->inject);
202 * Do not drop the file until the irqfd is fully initialized,
203 * otherwise we might race against the POLLHUP.
209 eventfd_ctx_put(ctx);
218 static void virqfd_disable(struct vfio_pci_device *vdev,
219 struct virqfd **pvirqfd)
223 spin_lock_irqsave(&vdev->irqlock, flags);
226 virqfd_deactivate(*pvirqfd);
230 spin_unlock_irqrestore(&vdev->irqlock, flags);
233 * Block until we know all outstanding shutdown jobs have completed.
234 * Even if we don't queue the job, flush the wq to be sure it's
237 flush_workqueue(vfio_irqfd_cleanup_wq);
243 static void vfio_send_intx_eventfd(struct vfio_pci_device *vdev, void *unused)
245 if (likely(is_intx(vdev) && !vdev->virq_disabled))
246 eventfd_signal(vdev->ctx[0].trigger, 1);
249 void vfio_pci_intx_mask(struct vfio_pci_device *vdev)
251 struct pci_dev *pdev = vdev->pdev;
254 spin_lock_irqsave(&vdev->irqlock, flags);
257 * Masking can come from interrupt, ioctl, or config space
258 * via INTx disable. The latter means this can get called
259 * even when not using intx delivery. In this case, just
260 * try to have the physical bit follow the virtual bit.
262 if (unlikely(!is_intx(vdev))) {
265 } else if (!vdev->ctx[0].masked) {
267 * Can't use check_and_mask here because we always want to
268 * mask, not just when something is pending.
273 disable_irq_nosync(pdev->irq);
275 vdev->ctx[0].masked = true;
278 spin_unlock_irqrestore(&vdev->irqlock, flags);
282 * If this is triggered by an eventfd, we can't call eventfd_signal
283 * or else we'll deadlock on the eventfd wait queue. Return >0 when
284 * a signal is necessary, which can then be handled via a work queue
285 * or directly depending on the caller.
287 static int vfio_pci_intx_unmask_handler(struct vfio_pci_device *vdev,
290 struct pci_dev *pdev = vdev->pdev;
294 spin_lock_irqsave(&vdev->irqlock, flags);
297 * Unmasking comes from ioctl or config, so again, have the
298 * physical bit follow the virtual even when not using INTx.
300 if (unlikely(!is_intx(vdev))) {
303 } else if (vdev->ctx[0].masked && !vdev->virq_disabled) {
305 * A pending interrupt here would immediately trigger,
306 * but we can avoid that overhead by just re-sending
307 * the interrupt to the user.
310 if (!pci_check_and_unmask_intx(pdev))
313 enable_irq(pdev->irq);
315 vdev->ctx[0].masked = (ret > 0);
318 spin_unlock_irqrestore(&vdev->irqlock, flags);
323 void vfio_pci_intx_unmask(struct vfio_pci_device *vdev)
325 if (vfio_pci_intx_unmask_handler(vdev, NULL) > 0)
326 vfio_send_intx_eventfd(vdev, NULL);
329 static irqreturn_t vfio_intx_handler(int irq, void *dev_id)
331 struct vfio_pci_device *vdev = dev_id;
335 spin_lock_irqsave(&vdev->irqlock, flags);
337 if (!vdev->pci_2_3) {
338 disable_irq_nosync(vdev->pdev->irq);
339 vdev->ctx[0].masked = true;
341 } else if (!vdev->ctx[0].masked && /* may be shared */
342 pci_check_and_mask_intx(vdev->pdev)) {
343 vdev->ctx[0].masked = true;
347 spin_unlock_irqrestore(&vdev->irqlock, flags);
349 if (ret == IRQ_HANDLED)
350 vfio_send_intx_eventfd(vdev, NULL);
355 static int vfio_intx_enable(struct vfio_pci_device *vdev)
357 if (!is_irq_none(vdev))
360 if (!vdev->pdev->irq)
363 vdev->ctx = kzalloc(sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
370 * If the virtual interrupt is masked, restore it. Devices
371 * supporting DisINTx can be masked at the hardware level
372 * here, non-PCI-2.3 devices will have to wait until the
373 * interrupt is enabled.
375 vdev->ctx[0].masked = vdev->virq_disabled;
377 pci_intx(vdev->pdev, !vdev->ctx[0].masked);
379 vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX;
384 static int vfio_intx_set_signal(struct vfio_pci_device *vdev, int fd)
386 struct pci_dev *pdev = vdev->pdev;
387 unsigned long irqflags = IRQF_SHARED;
388 struct eventfd_ctx *trigger;
392 if (vdev->ctx[0].trigger) {
393 free_irq(pdev->irq, vdev);
394 kfree(vdev->ctx[0].name);
395 eventfd_ctx_put(vdev->ctx[0].trigger);
396 vdev->ctx[0].trigger = NULL;
399 if (fd < 0) /* Disable only */
402 vdev->ctx[0].name = kasprintf(GFP_KERNEL, "vfio-intx(%s)",
404 if (!vdev->ctx[0].name)
407 trigger = eventfd_ctx_fdget(fd);
408 if (IS_ERR(trigger)) {
409 kfree(vdev->ctx[0].name);
410 return PTR_ERR(trigger);
413 vdev->ctx[0].trigger = trigger;
418 ret = request_irq(pdev->irq, vfio_intx_handler,
419 irqflags, vdev->ctx[0].name, vdev);
421 vdev->ctx[0].trigger = NULL;
422 kfree(vdev->ctx[0].name);
423 eventfd_ctx_put(trigger);
428 * INTx disable will stick across the new irq setup,
431 spin_lock_irqsave(&vdev->irqlock, flags);
432 if (!vdev->pci_2_3 && vdev->ctx[0].masked)
433 disable_irq_nosync(pdev->irq);
434 spin_unlock_irqrestore(&vdev->irqlock, flags);
439 static void vfio_intx_disable(struct vfio_pci_device *vdev)
441 vfio_intx_set_signal(vdev, -1);
442 virqfd_disable(vdev, &vdev->ctx[0].unmask);
443 virqfd_disable(vdev, &vdev->ctx[0].mask);
444 vdev->irq_type = VFIO_PCI_NUM_IRQS;
452 static irqreturn_t vfio_msihandler(int irq, void *arg)
454 struct eventfd_ctx *trigger = arg;
456 eventfd_signal(trigger, 1);
460 static int vfio_msi_enable(struct vfio_pci_device *vdev, int nvec, bool msix)
462 struct pci_dev *pdev = vdev->pdev;
465 if (!is_irq_none(vdev))
468 vdev->ctx = kzalloc(nvec * sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
475 vdev->msix = kzalloc(nvec * sizeof(struct msix_entry),
482 for (i = 0; i < nvec; i++)
483 vdev->msix[i].entry = i;
485 ret = pci_enable_msix_range(pdev, vdev->msix, 1, nvec);
488 pci_disable_msix(pdev);
494 ret = pci_enable_msi_range(pdev, 1, nvec);
497 pci_disable_msi(pdev);
503 vdev->num_ctx = nvec;
504 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX :
505 VFIO_PCI_MSI_IRQ_INDEX;
509 * Compute the virtual hardware field for max msi vectors -
510 * it is the log base 2 of the number of vectors.
512 vdev->msi_qmax = fls(nvec * 2 - 1) - 1;
518 static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
519 int vector, int fd, bool msix)
521 struct pci_dev *pdev = vdev->pdev;
522 int irq = msix ? vdev->msix[vector].vector : pdev->irq + vector;
523 char *name = msix ? "vfio-msix" : "vfio-msi";
524 struct eventfd_ctx *trigger;
527 if (vector >= vdev->num_ctx)
530 if (vdev->ctx[vector].trigger) {
531 free_irq(irq, vdev->ctx[vector].trigger);
532 kfree(vdev->ctx[vector].name);
533 eventfd_ctx_put(vdev->ctx[vector].trigger);
534 vdev->ctx[vector].trigger = NULL;
540 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "%s[%d](%s)",
541 name, vector, pci_name(pdev));
542 if (!vdev->ctx[vector].name)
545 trigger = eventfd_ctx_fdget(fd);
546 if (IS_ERR(trigger)) {
547 kfree(vdev->ctx[vector].name);
548 return PTR_ERR(trigger);
551 ret = request_irq(irq, vfio_msihandler, 0,
552 vdev->ctx[vector].name, trigger);
554 kfree(vdev->ctx[vector].name);
555 eventfd_ctx_put(trigger);
559 vdev->ctx[vector].trigger = trigger;
564 static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
565 unsigned count, int32_t *fds, bool msix)
569 if (start + count > vdev->num_ctx)
572 for (i = 0, j = start; i < count && !ret; i++, j++) {
573 int fd = fds ? fds[i] : -1;
574 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix);
578 for (--j; j >= start; j--)
579 vfio_msi_set_vector_signal(vdev, j, -1, msix);
585 static void vfio_msi_disable(struct vfio_pci_device *vdev, bool msix)
587 struct pci_dev *pdev = vdev->pdev;
590 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix);
592 for (i = 0; i < vdev->num_ctx; i++) {
593 virqfd_disable(vdev, &vdev->ctx[i].unmask);
594 virqfd_disable(vdev, &vdev->ctx[i].mask);
598 pci_disable_msix(vdev->pdev);
601 pci_disable_msi(pdev);
603 vdev->irq_type = VFIO_PCI_NUM_IRQS;
611 static int vfio_pci_set_intx_unmask(struct vfio_pci_device *vdev,
612 unsigned index, unsigned start,
613 unsigned count, uint32_t flags, void *data)
615 if (!is_intx(vdev) || start != 0 || count != 1)
618 if (flags & VFIO_IRQ_SET_DATA_NONE) {
619 vfio_pci_intx_unmask(vdev);
620 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
621 uint8_t unmask = *(uint8_t *)data;
623 vfio_pci_intx_unmask(vdev);
624 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
625 int32_t fd = *(int32_t *)data;
627 return virqfd_enable(vdev, vfio_pci_intx_unmask_handler,
628 vfio_send_intx_eventfd, NULL,
629 &vdev->ctx[0].unmask, fd);
631 virqfd_disable(vdev, &vdev->ctx[0].unmask);
637 static int vfio_pci_set_intx_mask(struct vfio_pci_device *vdev,
638 unsigned index, unsigned start,
639 unsigned count, uint32_t flags, void *data)
641 if (!is_intx(vdev) || start != 0 || count != 1)
644 if (flags & VFIO_IRQ_SET_DATA_NONE) {
645 vfio_pci_intx_mask(vdev);
646 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
647 uint8_t mask = *(uint8_t *)data;
649 vfio_pci_intx_mask(vdev);
650 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
651 return -ENOTTY; /* XXX implement me */
657 static int vfio_pci_set_intx_trigger(struct vfio_pci_device *vdev,
658 unsigned index, unsigned start,
659 unsigned count, uint32_t flags, void *data)
661 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
662 vfio_intx_disable(vdev);
666 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1)
669 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
670 int32_t fd = *(int32_t *)data;
674 return vfio_intx_set_signal(vdev, fd);
676 ret = vfio_intx_enable(vdev);
680 ret = vfio_intx_set_signal(vdev, fd);
682 vfio_intx_disable(vdev);
690 if (flags & VFIO_IRQ_SET_DATA_NONE) {
691 vfio_send_intx_eventfd(vdev, NULL);
692 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
693 uint8_t trigger = *(uint8_t *)data;
695 vfio_send_intx_eventfd(vdev, NULL);
700 static int vfio_pci_set_msi_trigger(struct vfio_pci_device *vdev,
701 unsigned index, unsigned start,
702 unsigned count, uint32_t flags, void *data)
705 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false;
707 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
708 vfio_msi_disable(vdev, msix);
712 if (!(irq_is(vdev, index) || is_irq_none(vdev)))
715 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
719 if (vdev->irq_type == index)
720 return vfio_msi_set_block(vdev, start, count,
723 ret = vfio_msi_enable(vdev, start + count, msix);
727 ret = vfio_msi_set_block(vdev, start, count, fds, msix);
729 vfio_msi_disable(vdev, msix);
734 if (!irq_is(vdev, index) || start + count > vdev->num_ctx)
737 for (i = start; i < start + count; i++) {
738 if (!vdev->ctx[i].trigger)
740 if (flags & VFIO_IRQ_SET_DATA_NONE) {
741 eventfd_signal(vdev->ctx[i].trigger, 1);
742 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
743 uint8_t *bools = data;
744 if (bools[i - start])
745 eventfd_signal(vdev->ctx[i].trigger, 1);
751 static int vfio_pci_set_err_trigger(struct vfio_pci_device *vdev,
752 unsigned index, unsigned start,
753 unsigned count, uint32_t flags, void *data)
755 int32_t fd = *(int32_t *)data;
757 if ((index != VFIO_PCI_ERR_IRQ_INDEX) ||
758 !(flags & VFIO_IRQ_SET_DATA_TYPE_MASK))
761 /* DATA_NONE/DATA_BOOL enables loopback testing */
762 if (flags & VFIO_IRQ_SET_DATA_NONE) {
763 if (vdev->err_trigger)
764 eventfd_signal(vdev->err_trigger, 1);
766 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
767 uint8_t trigger = *(uint8_t *)data;
768 if (trigger && vdev->err_trigger)
769 eventfd_signal(vdev->err_trigger, 1);
773 /* Handle SET_DATA_EVENTFD */
775 if (vdev->err_trigger)
776 eventfd_ctx_put(vdev->err_trigger);
777 vdev->err_trigger = NULL;
779 } else if (fd >= 0) {
780 struct eventfd_ctx *efdctx;
781 efdctx = eventfd_ctx_fdget(fd);
783 return PTR_ERR(efdctx);
784 if (vdev->err_trigger)
785 eventfd_ctx_put(vdev->err_trigger);
786 vdev->err_trigger = efdctx;
791 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags,
792 unsigned index, unsigned start, unsigned count,
795 int (*func)(struct vfio_pci_device *vdev, unsigned index,
796 unsigned start, unsigned count, uint32_t flags,
800 case VFIO_PCI_INTX_IRQ_INDEX:
801 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
802 case VFIO_IRQ_SET_ACTION_MASK:
803 func = vfio_pci_set_intx_mask;
805 case VFIO_IRQ_SET_ACTION_UNMASK:
806 func = vfio_pci_set_intx_unmask;
808 case VFIO_IRQ_SET_ACTION_TRIGGER:
809 func = vfio_pci_set_intx_trigger;
813 case VFIO_PCI_MSI_IRQ_INDEX:
814 case VFIO_PCI_MSIX_IRQ_INDEX:
815 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
816 case VFIO_IRQ_SET_ACTION_MASK:
817 case VFIO_IRQ_SET_ACTION_UNMASK:
818 /* XXX Need masking support exported */
820 case VFIO_IRQ_SET_ACTION_TRIGGER:
821 func = vfio_pci_set_msi_trigger;
825 case VFIO_PCI_ERR_IRQ_INDEX:
826 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
827 case VFIO_IRQ_SET_ACTION_TRIGGER:
828 if (pci_is_pcie(vdev->pdev))
829 func = vfio_pci_set_err_trigger;
837 return func(vdev, index, start, count, flags, data);