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>
26 #include "vfio_pci_private.h"
32 struct vfio_pci_device *vdev;
33 struct eventfd_ctx *eventfd;
34 int (*handler)(struct vfio_pci_device *, void *);
35 void (*thread)(struct vfio_pci_device *, void *);
37 struct work_struct inject;
40 struct work_struct shutdown;
41 struct virqfd **pvirqfd;
44 static struct workqueue_struct *vfio_irqfd_cleanup_wq;
46 int __init vfio_pci_virqfd_init(void)
48 vfio_irqfd_cleanup_wq =
49 create_singlethread_workqueue("vfio-irqfd-cleanup");
50 if (!vfio_irqfd_cleanup_wq)
56 void vfio_pci_virqfd_exit(void)
58 destroy_workqueue(vfio_irqfd_cleanup_wq);
61 static void virqfd_deactivate(struct virqfd *virqfd)
63 queue_work(vfio_irqfd_cleanup_wq, &virqfd->shutdown);
66 static int virqfd_wakeup(wait_queue_t *wait, unsigned mode, int sync, void *key)
68 struct virqfd *virqfd = container_of(wait, struct virqfd, wait);
69 unsigned long flags = (unsigned long)key;
72 /* An event has been signaled, call function */
73 if ((!virqfd->handler ||
74 virqfd->handler(virqfd->vdev, virqfd->data)) &&
76 schedule_work(&virqfd->inject);
79 if (flags & POLLHUP) {
81 spin_lock_irqsave(&virqfd->vdev->irqlock, flags);
84 * The eventfd is closing, if the virqfd has not yet been
85 * queued for release, as determined by testing whether the
86 * vdev pointer to it is still valid, queue it now. As
87 * with kvm irqfds, we know we won't race against the virqfd
88 * going away because we hold wqh->lock to get here.
90 if (*(virqfd->pvirqfd) == virqfd) {
91 *(virqfd->pvirqfd) = NULL;
92 virqfd_deactivate(virqfd);
95 spin_unlock_irqrestore(&virqfd->vdev->irqlock, flags);
101 static void virqfd_ptable_queue_proc(struct file *file,
102 wait_queue_head_t *wqh, poll_table *pt)
104 struct virqfd *virqfd = container_of(pt, struct virqfd, pt);
105 add_wait_queue(wqh, &virqfd->wait);
108 static void virqfd_shutdown(struct work_struct *work)
110 struct virqfd *virqfd = container_of(work, struct virqfd, shutdown);
113 eventfd_ctx_remove_wait_queue(virqfd->eventfd, &virqfd->wait, &cnt);
114 flush_work(&virqfd->inject);
115 eventfd_ctx_put(virqfd->eventfd);
120 static void virqfd_inject(struct work_struct *work)
122 struct virqfd *virqfd = container_of(work, struct virqfd, inject);
124 virqfd->thread(virqfd->vdev, virqfd->data);
127 static int virqfd_enable(struct vfio_pci_device *vdev,
128 int (*handler)(struct vfio_pci_device *, void *),
129 void (*thread)(struct vfio_pci_device *, void *),
130 void *data, struct virqfd **pvirqfd, int fd)
132 struct file *file = NULL;
133 struct eventfd_ctx *ctx = NULL;
134 struct virqfd *virqfd;
138 virqfd = kzalloc(sizeof(*virqfd), GFP_KERNEL);
142 virqfd->pvirqfd = pvirqfd;
144 virqfd->handler = handler;
145 virqfd->thread = thread;
148 INIT_WORK(&virqfd->shutdown, virqfd_shutdown);
149 INIT_WORK(&virqfd->inject, virqfd_inject);
151 file = eventfd_fget(fd);
157 ctx = eventfd_ctx_fileget(file);
163 virqfd->eventfd = ctx;
166 * virqfds can be released by closing the eventfd or directly
167 * through ioctl. These are both done through a workqueue, so
168 * we update the pointer to the virqfd under lock to avoid
169 * pushing multiple jobs to release the same virqfd.
171 spin_lock_irq(&vdev->irqlock);
174 spin_unlock_irq(&vdev->irqlock);
180 spin_unlock_irq(&vdev->irqlock);
183 * Install our own custom wake-up handling so we are notified via
184 * a callback whenever someone signals the underlying eventfd.
186 init_waitqueue_func_entry(&virqfd->wait, virqfd_wakeup);
187 init_poll_funcptr(&virqfd->pt, virqfd_ptable_queue_proc);
189 events = file->f_op->poll(file, &virqfd->pt);
192 * Check if there was an event already pending on the eventfd
193 * before we registered and trigger it as if we didn't miss it.
195 if (events & POLLIN) {
196 if ((!handler || handler(vdev, data)) && thread)
197 schedule_work(&virqfd->inject);
201 * Do not drop the file until the irqfd is fully initialized,
202 * otherwise we might race against the POLLHUP.
209 if (ctx && !IS_ERR(ctx))
210 eventfd_ctx_put(ctx);
212 if (file && !IS_ERR(file))
220 static void virqfd_disable(struct vfio_pci_device *vdev,
221 struct virqfd **pvirqfd)
225 spin_lock_irqsave(&vdev->irqlock, flags);
228 virqfd_deactivate(*pvirqfd);
232 spin_unlock_irqrestore(&vdev->irqlock, flags);
235 * Block until we know all outstanding shutdown jobs have completed.
236 * Even if we don't queue the job, flush the wq to be sure it's
239 flush_workqueue(vfio_irqfd_cleanup_wq);
245 static void vfio_send_intx_eventfd(struct vfio_pci_device *vdev, void *unused)
247 if (likely(is_intx(vdev) && !vdev->virq_disabled))
248 eventfd_signal(vdev->ctx[0].trigger, 1);
251 void vfio_pci_intx_mask(struct vfio_pci_device *vdev)
253 struct pci_dev *pdev = vdev->pdev;
256 spin_lock_irqsave(&vdev->irqlock, flags);
259 * Masking can come from interrupt, ioctl, or config space
260 * via INTx disable. The latter means this can get called
261 * even when not using intx delivery. In this case, just
262 * try to have the physical bit follow the virtual bit.
264 if (unlikely(!is_intx(vdev))) {
267 } else if (!vdev->ctx[0].masked) {
269 * Can't use check_and_mask here because we always want to
270 * mask, not just when something is pending.
275 disable_irq_nosync(pdev->irq);
277 vdev->ctx[0].masked = true;
280 spin_unlock_irqrestore(&vdev->irqlock, flags);
284 * If this is triggered by an eventfd, we can't call eventfd_signal
285 * or else we'll deadlock on the eventfd wait queue. Return >0 when
286 * a signal is necessary, which can then be handled via a work queue
287 * or directly depending on the caller.
289 int vfio_pci_intx_unmask_handler(struct vfio_pci_device *vdev, void *unused)
291 struct pci_dev *pdev = vdev->pdev;
295 spin_lock_irqsave(&vdev->irqlock, flags);
298 * Unmasking comes from ioctl or config, so again, have the
299 * physical bit follow the virtual even when not using INTx.
301 if (unlikely(!is_intx(vdev))) {
304 } else if (vdev->ctx[0].masked && !vdev->virq_disabled) {
306 * A pending interrupt here would immediately trigger,
307 * but we can avoid that overhead by just re-sending
308 * the interrupt to the user.
311 if (!pci_check_and_unmask_intx(pdev))
314 enable_irq(pdev->irq);
316 vdev->ctx[0].masked = (ret > 0);
319 spin_unlock_irqrestore(&vdev->irqlock, flags);
324 void vfio_pci_intx_unmask(struct vfio_pci_device *vdev)
326 if (vfio_pci_intx_unmask_handler(vdev, NULL) > 0)
327 vfio_send_intx_eventfd(vdev, NULL);
330 static irqreturn_t vfio_intx_handler(int irq, void *dev_id)
332 struct vfio_pci_device *vdev = dev_id;
336 spin_lock_irqsave(&vdev->irqlock, flags);
338 if (!vdev->pci_2_3) {
339 disable_irq_nosync(vdev->pdev->irq);
340 vdev->ctx[0].masked = true;
342 } else if (!vdev->ctx[0].masked && /* may be shared */
343 pci_check_and_mask_intx(vdev->pdev)) {
344 vdev->ctx[0].masked = true;
348 spin_unlock_irqrestore(&vdev->irqlock, flags);
350 if (ret == IRQ_HANDLED)
351 vfio_send_intx_eventfd(vdev, NULL);
356 static int vfio_intx_enable(struct vfio_pci_device *vdev)
358 if (!is_irq_none(vdev))
361 if (!vdev->pdev->irq)
364 vdev->ctx = kzalloc(sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
371 * If the virtual interrupt is masked, restore it. Devices
372 * supporting DisINTx can be masked at the hardware level
373 * here, non-PCI-2.3 devices will have to wait until the
374 * interrupt is enabled.
376 vdev->ctx[0].masked = vdev->virq_disabled;
378 pci_intx(vdev->pdev, !vdev->ctx[0].masked);
380 vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX;
385 static int vfio_intx_set_signal(struct vfio_pci_device *vdev, int fd)
387 struct pci_dev *pdev = vdev->pdev;
388 unsigned long irqflags = IRQF_SHARED;
389 struct eventfd_ctx *trigger;
393 if (vdev->ctx[0].trigger) {
394 free_irq(pdev->irq, vdev);
395 kfree(vdev->ctx[0].name);
396 eventfd_ctx_put(vdev->ctx[0].trigger);
397 vdev->ctx[0].trigger = NULL;
400 if (fd < 0) /* Disable only */
403 vdev->ctx[0].name = kasprintf(GFP_KERNEL, "vfio-intx(%s)",
405 if (!vdev->ctx[0].name)
408 trigger = eventfd_ctx_fdget(fd);
409 if (IS_ERR(trigger)) {
410 kfree(vdev->ctx[0].name);
411 return PTR_ERR(trigger);
414 vdev->ctx[0].trigger = trigger;
419 ret = request_irq(pdev->irq, vfio_intx_handler,
420 irqflags, vdev->ctx[0].name, vdev);
422 vdev->ctx[0].trigger = NULL;
423 kfree(vdev->ctx[0].name);
424 eventfd_ctx_put(trigger);
429 * INTx disable will stick across the new irq setup,
432 spin_lock_irqsave(&vdev->irqlock, flags);
433 if (!vdev->pci_2_3 && vdev->ctx[0].masked)
434 disable_irq_nosync(pdev->irq);
435 spin_unlock_irqrestore(&vdev->irqlock, flags);
440 static void vfio_intx_disable(struct vfio_pci_device *vdev)
442 vfio_intx_set_signal(vdev, -1);
443 virqfd_disable(vdev, &vdev->ctx[0].unmask);
444 virqfd_disable(vdev, &vdev->ctx[0].mask);
445 vdev->irq_type = VFIO_PCI_NUM_IRQS;
453 static irqreturn_t vfio_msihandler(int irq, void *arg)
455 struct eventfd_ctx *trigger = arg;
457 eventfd_signal(trigger, 1);
461 static int vfio_msi_enable(struct vfio_pci_device *vdev, int nvec, bool msix)
463 struct pci_dev *pdev = vdev->pdev;
466 if (!is_irq_none(vdev))
469 vdev->ctx = kzalloc(nvec * sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
476 vdev->msix = kzalloc(nvec * sizeof(struct msix_entry),
483 for (i = 0; i < nvec; i++)
484 vdev->msix[i].entry = i;
486 ret = pci_enable_msix(pdev, vdev->msix, nvec);
493 ret = pci_enable_msi_block(pdev, nvec);
500 vdev->num_ctx = nvec;
501 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX :
502 VFIO_PCI_MSI_IRQ_INDEX;
506 * Compute the virtual hardware field for max msi vectors -
507 * it is the log base 2 of the number of vectors.
509 vdev->msi_qmax = fls(nvec * 2 - 1) - 1;
515 static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
516 int vector, int fd, bool msix)
518 struct pci_dev *pdev = vdev->pdev;
519 int irq = msix ? vdev->msix[vector].vector : pdev->irq + vector;
520 char *name = msix ? "vfio-msix" : "vfio-msi";
521 struct eventfd_ctx *trigger;
524 if (vector >= vdev->num_ctx)
527 if (vdev->ctx[vector].trigger) {
528 free_irq(irq, vdev->ctx[vector].trigger);
529 kfree(vdev->ctx[vector].name);
530 eventfd_ctx_put(vdev->ctx[vector].trigger);
531 vdev->ctx[vector].trigger = NULL;
537 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "%s[%d](%s)",
538 name, vector, pci_name(pdev));
539 if (!vdev->ctx[vector].name)
542 trigger = eventfd_ctx_fdget(fd);
543 if (IS_ERR(trigger)) {
544 kfree(vdev->ctx[vector].name);
545 return PTR_ERR(trigger);
548 ret = request_irq(irq, vfio_msihandler, 0,
549 vdev->ctx[vector].name, trigger);
551 kfree(vdev->ctx[vector].name);
552 eventfd_ctx_put(trigger);
556 vdev->ctx[vector].trigger = trigger;
561 static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
562 unsigned count, int32_t *fds, bool msix)
566 if (start + count > vdev->num_ctx)
569 for (i = 0, j = start; i < count && !ret; i++, j++) {
570 int fd = fds ? fds[i] : -1;
571 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix);
575 for (--j; j >= start; j--)
576 vfio_msi_set_vector_signal(vdev, j, -1, msix);
582 static void vfio_msi_disable(struct vfio_pci_device *vdev, bool msix)
584 struct pci_dev *pdev = vdev->pdev;
587 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix);
589 for (i = 0; i < vdev->num_ctx; i++) {
590 virqfd_disable(vdev, &vdev->ctx[i].unmask);
591 virqfd_disable(vdev, &vdev->ctx[i].mask);
595 pci_disable_msix(vdev->pdev);
598 pci_disable_msi(pdev);
600 vdev->irq_type = VFIO_PCI_NUM_IRQS;
608 static int vfio_pci_set_intx_unmask(struct vfio_pci_device *vdev,
609 unsigned index, unsigned start,
610 unsigned count, uint32_t flags, void *data)
612 if (!is_intx(vdev) || start != 0 || count != 1)
615 if (flags & VFIO_IRQ_SET_DATA_NONE) {
616 vfio_pci_intx_unmask(vdev);
617 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
618 uint8_t unmask = *(uint8_t *)data;
620 vfio_pci_intx_unmask(vdev);
621 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
622 int32_t fd = *(int32_t *)data;
624 return virqfd_enable(vdev, vfio_pci_intx_unmask_handler,
625 vfio_send_intx_eventfd, NULL,
626 &vdev->ctx[0].unmask, fd);
628 virqfd_disable(vdev, &vdev->ctx[0].unmask);
634 static int vfio_pci_set_intx_mask(struct vfio_pci_device *vdev,
635 unsigned index, unsigned start,
636 unsigned count, uint32_t flags, void *data)
638 if (!is_intx(vdev) || start != 0 || count != 1)
641 if (flags & VFIO_IRQ_SET_DATA_NONE) {
642 vfio_pci_intx_mask(vdev);
643 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
644 uint8_t mask = *(uint8_t *)data;
646 vfio_pci_intx_mask(vdev);
647 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
648 return -ENOTTY; /* XXX implement me */
654 static int vfio_pci_set_intx_trigger(struct vfio_pci_device *vdev,
655 unsigned index, unsigned start,
656 unsigned count, uint32_t flags, void *data)
658 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
659 vfio_intx_disable(vdev);
663 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1)
666 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
667 int32_t fd = *(int32_t *)data;
671 return vfio_intx_set_signal(vdev, fd);
673 ret = vfio_intx_enable(vdev);
677 ret = vfio_intx_set_signal(vdev, fd);
679 vfio_intx_disable(vdev);
687 if (flags & VFIO_IRQ_SET_DATA_NONE) {
688 vfio_send_intx_eventfd(vdev, NULL);
689 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
690 uint8_t trigger = *(uint8_t *)data;
692 vfio_send_intx_eventfd(vdev, NULL);
697 static int vfio_pci_set_msi_trigger(struct vfio_pci_device *vdev,
698 unsigned index, unsigned start,
699 unsigned count, uint32_t flags, void *data)
702 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false;
704 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
705 vfio_msi_disable(vdev, msix);
709 if (!(irq_is(vdev, index) || is_irq_none(vdev)))
712 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
716 if (vdev->irq_type == index)
717 return vfio_msi_set_block(vdev, start, count,
720 ret = vfio_msi_enable(vdev, start + count, msix);
724 ret = vfio_msi_set_block(vdev, start, count, fds, msix);
726 vfio_msi_disable(vdev, msix);
731 if (!irq_is(vdev, index) || start + count > vdev->num_ctx)
734 for (i = start; i < start + count; i++) {
735 if (!vdev->ctx[i].trigger)
737 if (flags & VFIO_IRQ_SET_DATA_NONE) {
738 eventfd_signal(vdev->ctx[i].trigger, 1);
739 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
740 uint8_t *bools = data;
741 if (bools[i - start])
742 eventfd_signal(vdev->ctx[i].trigger, 1);
748 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags,
749 unsigned index, unsigned start, unsigned count,
752 int (*func)(struct vfio_pci_device *vdev, unsigned index,
753 unsigned start, unsigned count, uint32_t flags,
757 case VFIO_PCI_INTX_IRQ_INDEX:
758 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
759 case VFIO_IRQ_SET_ACTION_MASK:
760 func = vfio_pci_set_intx_mask;
762 case VFIO_IRQ_SET_ACTION_UNMASK:
763 func = vfio_pci_set_intx_unmask;
765 case VFIO_IRQ_SET_ACTION_TRIGGER:
766 func = vfio_pci_set_intx_trigger;
770 case VFIO_PCI_MSI_IRQ_INDEX:
771 case VFIO_PCI_MSIX_IRQ_INDEX:
772 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
773 case VFIO_IRQ_SET_ACTION_MASK:
774 case VFIO_IRQ_SET_ACTION_UNMASK:
775 /* XXX Need masking support exported */
777 case VFIO_IRQ_SET_ACTION_TRIGGER:
778 func = vfio_pci_set_msi_trigger;
787 return func(vdev, index, start, count, flags, data);