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_intx_set_signal(vdev, -1);
232 vfio_virqfd_disable(&vdev->ctx[0].unmask);
233 vfio_virqfd_disable(&vdev->ctx[0].mask);
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;
255 if (!is_irq_none(vdev))
258 vdev->ctx = kzalloc(nvec * sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
265 vdev->msix = kzalloc(nvec * sizeof(struct msix_entry),
272 for (i = 0; i < nvec; i++)
273 vdev->msix[i].entry = i;
275 ret = pci_enable_msix_range(pdev, vdev->msix, 1, nvec);
278 pci_disable_msix(pdev);
284 ret = pci_enable_msi_range(pdev, 1, nvec);
287 pci_disable_msi(pdev);
293 vdev->num_ctx = nvec;
294 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX :
295 VFIO_PCI_MSI_IRQ_INDEX;
299 * Compute the virtual hardware field for max msi vectors -
300 * it is the log base 2 of the number of vectors.
302 vdev->msi_qmax = fls(nvec * 2 - 1) - 1;
308 static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
309 int vector, int fd, bool msix)
311 struct pci_dev *pdev = vdev->pdev;
312 int irq = msix ? vdev->msix[vector].vector : pdev->irq + vector;
313 char *name = msix ? "vfio-msix" : "vfio-msi";
314 struct eventfd_ctx *trigger;
317 if (vector >= vdev->num_ctx)
320 if (vdev->ctx[vector].trigger) {
321 free_irq(irq, vdev->ctx[vector].trigger);
322 kfree(vdev->ctx[vector].name);
323 eventfd_ctx_put(vdev->ctx[vector].trigger);
324 vdev->ctx[vector].trigger = NULL;
330 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "%s[%d](%s)",
331 name, vector, pci_name(pdev));
332 if (!vdev->ctx[vector].name)
335 trigger = eventfd_ctx_fdget(fd);
336 if (IS_ERR(trigger)) {
337 kfree(vdev->ctx[vector].name);
338 return PTR_ERR(trigger);
342 * The MSIx vector table resides in device memory which may be cleared
343 * via backdoor resets. We don't allow direct access to the vector
344 * table so even if a userspace driver attempts to save/restore around
345 * such a reset it would be unsuccessful. To avoid this, restore the
346 * cached value of the message prior to enabling.
351 get_cached_msi_msg(irq, &msg);
352 pci_write_msi_msg(irq, &msg);
355 ret = request_irq(irq, vfio_msihandler, 0,
356 vdev->ctx[vector].name, trigger);
358 kfree(vdev->ctx[vector].name);
359 eventfd_ctx_put(trigger);
363 vdev->ctx[vector].trigger = trigger;
368 static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
369 unsigned count, int32_t *fds, bool msix)
373 if (start + count > vdev->num_ctx)
376 for (i = 0, j = start; i < count && !ret; i++, j++) {
377 int fd = fds ? fds[i] : -1;
378 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix);
382 for (--j; j >= start; j--)
383 vfio_msi_set_vector_signal(vdev, j, -1, msix);
389 static void vfio_msi_disable(struct vfio_pci_device *vdev, bool msix)
391 struct pci_dev *pdev = vdev->pdev;
394 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix);
396 for (i = 0; i < vdev->num_ctx; i++) {
397 vfio_virqfd_disable(&vdev->ctx[i].unmask);
398 vfio_virqfd_disable(&vdev->ctx[i].mask);
402 pci_disable_msix(vdev->pdev);
405 pci_disable_msi(pdev);
407 vdev->irq_type = VFIO_PCI_NUM_IRQS;
415 static int vfio_pci_set_intx_unmask(struct vfio_pci_device *vdev,
416 unsigned index, unsigned start,
417 unsigned count, uint32_t flags, void *data)
419 if (!is_intx(vdev) || start != 0 || count != 1)
422 if (flags & VFIO_IRQ_SET_DATA_NONE) {
423 vfio_pci_intx_unmask(vdev);
424 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
425 uint8_t unmask = *(uint8_t *)data;
427 vfio_pci_intx_unmask(vdev);
428 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
429 int32_t fd = *(int32_t *)data;
431 return vfio_virqfd_enable((void *) vdev,
432 vfio_pci_intx_unmask_handler,
433 vfio_send_intx_eventfd, NULL,
434 &vdev->ctx[0].unmask, fd);
436 vfio_virqfd_disable(&vdev->ctx[0].unmask);
442 static int vfio_pci_set_intx_mask(struct vfio_pci_device *vdev,
443 unsigned index, unsigned start,
444 unsigned count, uint32_t flags, void *data)
446 if (!is_intx(vdev) || start != 0 || count != 1)
449 if (flags & VFIO_IRQ_SET_DATA_NONE) {
450 vfio_pci_intx_mask(vdev);
451 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
452 uint8_t mask = *(uint8_t *)data;
454 vfio_pci_intx_mask(vdev);
455 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
456 return -ENOTTY; /* XXX implement me */
462 static int vfio_pci_set_intx_trigger(struct vfio_pci_device *vdev,
463 unsigned index, unsigned start,
464 unsigned count, uint32_t flags, void *data)
466 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
467 vfio_intx_disable(vdev);
471 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1)
474 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
475 int32_t fd = *(int32_t *)data;
479 return vfio_intx_set_signal(vdev, fd);
481 ret = vfio_intx_enable(vdev);
485 ret = vfio_intx_set_signal(vdev, fd);
487 vfio_intx_disable(vdev);
495 if (flags & VFIO_IRQ_SET_DATA_NONE) {
496 vfio_send_intx_eventfd(vdev, NULL);
497 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
498 uint8_t trigger = *(uint8_t *)data;
500 vfio_send_intx_eventfd(vdev, NULL);
505 static int vfio_pci_set_msi_trigger(struct vfio_pci_device *vdev,
506 unsigned index, unsigned start,
507 unsigned count, uint32_t flags, void *data)
510 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false;
512 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
513 vfio_msi_disable(vdev, msix);
517 if (!(irq_is(vdev, index) || is_irq_none(vdev)))
520 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
524 if (vdev->irq_type == index)
525 return vfio_msi_set_block(vdev, start, count,
528 ret = vfio_msi_enable(vdev, start + count, msix);
532 ret = vfio_msi_set_block(vdev, start, count, fds, msix);
534 vfio_msi_disable(vdev, msix);
539 if (!irq_is(vdev, index) || start + count > vdev->num_ctx)
542 for (i = start; i < start + count; i++) {
543 if (!vdev->ctx[i].trigger)
545 if (flags & VFIO_IRQ_SET_DATA_NONE) {
546 eventfd_signal(vdev->ctx[i].trigger, 1);
547 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
548 uint8_t *bools = data;
549 if (bools[i - start])
550 eventfd_signal(vdev->ctx[i].trigger, 1);
556 static int vfio_pci_set_ctx_trigger_single(struct eventfd_ctx **ctx,
557 uint32_t flags, void *data)
559 int32_t fd = *(int32_t *)data;
561 if (!(flags & VFIO_IRQ_SET_DATA_TYPE_MASK))
564 /* DATA_NONE/DATA_BOOL enables loopback testing */
565 if (flags & VFIO_IRQ_SET_DATA_NONE) {
567 eventfd_signal(*ctx, 1);
569 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
570 uint8_t trigger = *(uint8_t *)data;
572 eventfd_signal(*ctx, 1);
576 /* Handle SET_DATA_EVENTFD */
579 eventfd_ctx_put(*ctx);
582 } else if (fd >= 0) {
583 struct eventfd_ctx *efdctx;
584 efdctx = eventfd_ctx_fdget(fd);
586 return PTR_ERR(efdctx);
588 eventfd_ctx_put(*ctx);
595 static int vfio_pci_set_err_trigger(struct vfio_pci_device *vdev,
596 unsigned index, unsigned start,
597 unsigned count, uint32_t flags, void *data)
599 if (index != VFIO_PCI_ERR_IRQ_INDEX)
603 * We should sanitize start & count, but that wasn't caught
604 * originally, so this IRQ index must forever ignore them :-(
607 return vfio_pci_set_ctx_trigger_single(&vdev->err_trigger, flags, data);
610 static int vfio_pci_set_req_trigger(struct vfio_pci_device *vdev,
611 unsigned index, unsigned start,
612 unsigned count, uint32_t flags, void *data)
614 if (index != VFIO_PCI_REQ_IRQ_INDEX || start != 0 || count != 1)
617 return vfio_pci_set_ctx_trigger_single(&vdev->req_trigger, flags, data);
620 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags,
621 unsigned index, unsigned start, unsigned count,
624 int (*func)(struct vfio_pci_device *vdev, unsigned index,
625 unsigned start, unsigned count, uint32_t flags,
629 case VFIO_PCI_INTX_IRQ_INDEX:
630 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
631 case VFIO_IRQ_SET_ACTION_MASK:
632 func = vfio_pci_set_intx_mask;
634 case VFIO_IRQ_SET_ACTION_UNMASK:
635 func = vfio_pci_set_intx_unmask;
637 case VFIO_IRQ_SET_ACTION_TRIGGER:
638 func = vfio_pci_set_intx_trigger;
642 case VFIO_PCI_MSI_IRQ_INDEX:
643 case VFIO_PCI_MSIX_IRQ_INDEX:
644 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
645 case VFIO_IRQ_SET_ACTION_MASK:
646 case VFIO_IRQ_SET_ACTION_UNMASK:
647 /* XXX Need masking support exported */
649 case VFIO_IRQ_SET_ACTION_TRIGGER:
650 func = vfio_pci_set_msi_trigger;
654 case VFIO_PCI_ERR_IRQ_INDEX:
655 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
656 case VFIO_IRQ_SET_ACTION_TRIGGER:
657 if (pci_is_pcie(vdev->pdev))
658 func = vfio_pci_set_err_trigger;
662 case VFIO_PCI_REQ_IRQ_INDEX:
663 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
664 case VFIO_IRQ_SET_ACTION_TRIGGER:
665 func = vfio_pci_set_req_trigger;
674 return func(vdev, index, start, count, flags, data);