1 // SPDX-License-Identifier: GPL-2.0-only
3 * VFIO PCI interrupt handling
5 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
8 * Derived from original vfio:
9 * Copyright 2010 Cisco Systems, Inc. All rights reserved.
13 #include <linux/device.h>
14 #include <linux/interrupt.h>
15 #include <linux/eventfd.h>
16 #include <linux/msi.h>
17 #include <linux/pci.h>
18 #include <linux/file.h>
19 #include <linux/vfio.h>
20 #include <linux/wait.h>
21 #include <linux/slab.h>
23 #include "vfio_pci_private.h"
28 static void vfio_send_intx_eventfd(void *opaque, void *unused)
30 struct vfio_pci_device *vdev = opaque;
32 if (likely(is_intx(vdev) && !vdev->virq_disabled))
33 eventfd_signal(vdev->ctx[0].trigger, 1);
36 void vfio_pci_intx_mask(struct vfio_pci_device *vdev)
38 struct pci_dev *pdev = vdev->pdev;
41 spin_lock_irqsave(&vdev->irqlock, flags);
44 * Masking can come from interrupt, ioctl, or config space
45 * via INTx disable. The latter means this can get called
46 * even when not using intx delivery. In this case, just
47 * try to have the physical bit follow the virtual bit.
49 if (unlikely(!is_intx(vdev))) {
52 } else if (!vdev->ctx[0].masked) {
54 * Can't use check_and_mask here because we always want to
55 * mask, not just when something is pending.
60 disable_irq_nosync(pdev->irq);
62 vdev->ctx[0].masked = true;
65 spin_unlock_irqrestore(&vdev->irqlock, flags);
69 * If this is triggered by an eventfd, we can't call eventfd_signal
70 * or else we'll deadlock on the eventfd wait queue. Return >0 when
71 * a signal is necessary, which can then be handled via a work queue
72 * or directly depending on the caller.
74 static int vfio_pci_intx_unmask_handler(void *opaque, void *unused)
76 struct vfio_pci_device *vdev = opaque;
77 struct pci_dev *pdev = vdev->pdev;
81 spin_lock_irqsave(&vdev->irqlock, flags);
84 * Unmasking comes from ioctl or config, so again, have the
85 * physical bit follow the virtual even when not using INTx.
87 if (unlikely(!is_intx(vdev))) {
90 } else if (vdev->ctx[0].masked && !vdev->virq_disabled) {
92 * A pending interrupt here would immediately trigger,
93 * but we can avoid that overhead by just re-sending
94 * the interrupt to the user.
97 if (!pci_check_and_unmask_intx(pdev))
100 enable_irq(pdev->irq);
102 vdev->ctx[0].masked = (ret > 0);
105 spin_unlock_irqrestore(&vdev->irqlock, flags);
110 void vfio_pci_intx_unmask(struct vfio_pci_device *vdev)
112 if (vfio_pci_intx_unmask_handler(vdev, NULL) > 0)
113 vfio_send_intx_eventfd(vdev, NULL);
116 static irqreturn_t vfio_intx_handler(int irq, void *dev_id)
118 struct vfio_pci_device *vdev = dev_id;
122 spin_lock_irqsave(&vdev->irqlock, flags);
124 if (!vdev->pci_2_3) {
125 disable_irq_nosync(vdev->pdev->irq);
126 vdev->ctx[0].masked = true;
128 } else if (!vdev->ctx[0].masked && /* may be shared */
129 pci_check_and_mask_intx(vdev->pdev)) {
130 vdev->ctx[0].masked = true;
134 spin_unlock_irqrestore(&vdev->irqlock, flags);
136 if (ret == IRQ_HANDLED)
137 vfio_send_intx_eventfd(vdev, NULL);
142 static int vfio_intx_enable(struct vfio_pci_device *vdev)
144 if (!is_irq_none(vdev))
147 if (!vdev->pdev->irq)
150 vdev->ctx = kzalloc(sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
157 * If the virtual interrupt is masked, restore it. Devices
158 * supporting DisINTx can be masked at the hardware level
159 * here, non-PCI-2.3 devices will have to wait until the
160 * interrupt is enabled.
162 vdev->ctx[0].masked = vdev->virq_disabled;
164 pci_intx(vdev->pdev, !vdev->ctx[0].masked);
166 vdev->irq_type = VFIO_PCI_INTX_IRQ_INDEX;
171 static int vfio_intx_set_signal(struct vfio_pci_device *vdev, int fd)
173 struct pci_dev *pdev = vdev->pdev;
174 unsigned long irqflags = IRQF_SHARED;
175 struct eventfd_ctx *trigger;
179 if (vdev->ctx[0].trigger) {
180 free_irq(pdev->irq, vdev);
181 kfree(vdev->ctx[0].name);
182 eventfd_ctx_put(vdev->ctx[0].trigger);
183 vdev->ctx[0].trigger = NULL;
186 if (fd < 0) /* Disable only */
189 vdev->ctx[0].name = kasprintf(GFP_KERNEL, "vfio-intx(%s)",
191 if (!vdev->ctx[0].name)
194 trigger = eventfd_ctx_fdget(fd);
195 if (IS_ERR(trigger)) {
196 kfree(vdev->ctx[0].name);
197 return PTR_ERR(trigger);
200 vdev->ctx[0].trigger = trigger;
205 ret = request_irq(pdev->irq, vfio_intx_handler,
206 irqflags, vdev->ctx[0].name, vdev);
208 vdev->ctx[0].trigger = NULL;
209 kfree(vdev->ctx[0].name);
210 eventfd_ctx_put(trigger);
215 * INTx disable will stick across the new irq setup,
218 spin_lock_irqsave(&vdev->irqlock, flags);
219 if (!vdev->pci_2_3 && vdev->ctx[0].masked)
220 disable_irq_nosync(pdev->irq);
221 spin_unlock_irqrestore(&vdev->irqlock, flags);
226 static void vfio_intx_disable(struct vfio_pci_device *vdev)
228 vfio_virqfd_disable(&vdev->ctx[0].unmask);
229 vfio_virqfd_disable(&vdev->ctx[0].mask);
230 vfio_intx_set_signal(vdev, -1);
231 vdev->irq_type = VFIO_PCI_NUM_IRQS;
239 static irqreturn_t vfio_msihandler(int irq, void *arg)
241 struct eventfd_ctx *trigger = arg;
243 eventfd_signal(trigger, 1);
247 static int vfio_msi_enable(struct vfio_pci_device *vdev, int nvec, bool msix)
249 struct pci_dev *pdev = vdev->pdev;
250 unsigned int flag = msix ? PCI_IRQ_MSIX : PCI_IRQ_MSI;
253 if (!is_irq_none(vdev))
256 vdev->ctx = kcalloc(nvec, sizeof(struct vfio_pci_irq_ctx), GFP_KERNEL);
260 /* return the number of supported vectors if we can't get all: */
261 ret = pci_alloc_irq_vectors(pdev, 1, nvec, flag);
264 pci_free_irq_vectors(pdev);
269 vdev->num_ctx = nvec;
270 vdev->irq_type = msix ? VFIO_PCI_MSIX_IRQ_INDEX :
271 VFIO_PCI_MSI_IRQ_INDEX;
275 * Compute the virtual hardware field for max msi vectors -
276 * it is the log base 2 of the number of vectors.
278 vdev->msi_qmax = fls(nvec * 2 - 1) - 1;
284 static int vfio_msi_set_vector_signal(struct vfio_pci_device *vdev,
285 int vector, int fd, bool msix)
287 struct pci_dev *pdev = vdev->pdev;
288 struct eventfd_ctx *trigger;
291 if (vector < 0 || vector >= vdev->num_ctx)
294 irq = pci_irq_vector(pdev, vector);
296 if (vdev->ctx[vector].trigger) {
297 irq_bypass_unregister_producer(&vdev->ctx[vector].producer);
298 free_irq(irq, vdev->ctx[vector].trigger);
299 kfree(vdev->ctx[vector].name);
300 eventfd_ctx_put(vdev->ctx[vector].trigger);
301 vdev->ctx[vector].trigger = NULL;
307 vdev->ctx[vector].name = kasprintf(GFP_KERNEL, "vfio-msi%s[%d](%s)",
308 msix ? "x" : "", vector,
310 if (!vdev->ctx[vector].name)
313 trigger = eventfd_ctx_fdget(fd);
314 if (IS_ERR(trigger)) {
315 kfree(vdev->ctx[vector].name);
316 return PTR_ERR(trigger);
320 * The MSIx vector table resides in device memory which may be cleared
321 * via backdoor resets. We don't allow direct access to the vector
322 * table so even if a userspace driver attempts to save/restore around
323 * such a reset it would be unsuccessful. To avoid this, restore the
324 * cached value of the message prior to enabling.
329 get_cached_msi_msg(irq, &msg);
330 pci_write_msi_msg(irq, &msg);
333 ret = request_irq(irq, vfio_msihandler, 0,
334 vdev->ctx[vector].name, trigger);
336 kfree(vdev->ctx[vector].name);
337 eventfd_ctx_put(trigger);
341 vdev->ctx[vector].producer.token = trigger;
342 vdev->ctx[vector].producer.irq = irq;
343 ret = irq_bypass_register_producer(&vdev->ctx[vector].producer);
346 "irq bypass producer (token %p) registration fails: %d\n",
347 vdev->ctx[vector].producer.token, ret);
349 vdev->ctx[vector].trigger = trigger;
354 static int vfio_msi_set_block(struct vfio_pci_device *vdev, unsigned start,
355 unsigned count, int32_t *fds, bool msix)
359 if (start >= vdev->num_ctx || start + count > vdev->num_ctx)
362 for (i = 0, j = start; i < count && !ret; i++, j++) {
363 int fd = fds ? fds[i] : -1;
364 ret = vfio_msi_set_vector_signal(vdev, j, fd, msix);
368 for (--j; j >= (int)start; j--)
369 vfio_msi_set_vector_signal(vdev, j, -1, msix);
375 static void vfio_msi_disable(struct vfio_pci_device *vdev, bool msix)
377 struct pci_dev *pdev = vdev->pdev;
380 for (i = 0; i < vdev->num_ctx; i++) {
381 vfio_virqfd_disable(&vdev->ctx[i].unmask);
382 vfio_virqfd_disable(&vdev->ctx[i].mask);
385 vfio_msi_set_block(vdev, 0, vdev->num_ctx, NULL, msix);
387 pci_free_irq_vectors(pdev);
390 * Both disable paths above use pci_intx_for_msi() to clear DisINTx
391 * via their shutdown paths. Restore for NoINTx devices.
396 vdev->irq_type = VFIO_PCI_NUM_IRQS;
404 static int vfio_pci_set_intx_unmask(struct vfio_pci_device *vdev,
405 unsigned index, unsigned start,
406 unsigned count, uint32_t flags, void *data)
408 if (!is_intx(vdev) || start != 0 || count != 1)
411 if (flags & VFIO_IRQ_SET_DATA_NONE) {
412 vfio_pci_intx_unmask(vdev);
413 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
414 uint8_t unmask = *(uint8_t *)data;
416 vfio_pci_intx_unmask(vdev);
417 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
418 int32_t fd = *(int32_t *)data;
420 return vfio_virqfd_enable((void *) vdev,
421 vfio_pci_intx_unmask_handler,
422 vfio_send_intx_eventfd, NULL,
423 &vdev->ctx[0].unmask, fd);
425 vfio_virqfd_disable(&vdev->ctx[0].unmask);
431 static int vfio_pci_set_intx_mask(struct vfio_pci_device *vdev,
432 unsigned index, unsigned start,
433 unsigned count, uint32_t flags, void *data)
435 if (!is_intx(vdev) || start != 0 || count != 1)
438 if (flags & VFIO_IRQ_SET_DATA_NONE) {
439 vfio_pci_intx_mask(vdev);
440 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
441 uint8_t mask = *(uint8_t *)data;
443 vfio_pci_intx_mask(vdev);
444 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
445 return -ENOTTY; /* XXX implement me */
451 static int vfio_pci_set_intx_trigger(struct vfio_pci_device *vdev,
452 unsigned index, unsigned start,
453 unsigned count, uint32_t flags, void *data)
455 if (is_intx(vdev) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
456 vfio_intx_disable(vdev);
460 if (!(is_intx(vdev) || is_irq_none(vdev)) || start != 0 || count != 1)
463 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
464 int32_t fd = *(int32_t *)data;
468 return vfio_intx_set_signal(vdev, fd);
470 ret = vfio_intx_enable(vdev);
474 ret = vfio_intx_set_signal(vdev, fd);
476 vfio_intx_disable(vdev);
484 if (flags & VFIO_IRQ_SET_DATA_NONE) {
485 vfio_send_intx_eventfd(vdev, NULL);
486 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
487 uint8_t trigger = *(uint8_t *)data;
489 vfio_send_intx_eventfd(vdev, NULL);
494 static int vfio_pci_set_msi_trigger(struct vfio_pci_device *vdev,
495 unsigned index, unsigned start,
496 unsigned count, uint32_t flags, void *data)
499 bool msix = (index == VFIO_PCI_MSIX_IRQ_INDEX) ? true : false;
501 if (irq_is(vdev, index) && !count && (flags & VFIO_IRQ_SET_DATA_NONE)) {
502 vfio_msi_disable(vdev, msix);
506 if (!(irq_is(vdev, index) || is_irq_none(vdev)))
509 if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
513 if (vdev->irq_type == index)
514 return vfio_msi_set_block(vdev, start, count,
517 ret = vfio_msi_enable(vdev, start + count, msix);
521 ret = vfio_msi_set_block(vdev, start, count, fds, msix);
523 vfio_msi_disable(vdev, msix);
528 if (!irq_is(vdev, index) || start + count > vdev->num_ctx)
531 for (i = start; i < start + count; i++) {
532 if (!vdev->ctx[i].trigger)
534 if (flags & VFIO_IRQ_SET_DATA_NONE) {
535 eventfd_signal(vdev->ctx[i].trigger, 1);
536 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
537 uint8_t *bools = data;
538 if (bools[i - start])
539 eventfd_signal(vdev->ctx[i].trigger, 1);
545 static int vfio_pci_set_ctx_trigger_single(struct eventfd_ctx **ctx,
546 unsigned int count, uint32_t flags,
549 /* DATA_NONE/DATA_BOOL enables loopback testing */
550 if (flags & VFIO_IRQ_SET_DATA_NONE) {
553 eventfd_signal(*ctx, 1);
555 eventfd_ctx_put(*ctx);
560 } else if (flags & VFIO_IRQ_SET_DATA_BOOL) {
566 trigger = *(uint8_t *)data;
568 eventfd_signal(*ctx, 1);
571 } else if (flags & VFIO_IRQ_SET_DATA_EVENTFD) {
577 fd = *(int32_t *)data;
580 eventfd_ctx_put(*ctx);
582 } else if (fd >= 0) {
583 struct eventfd_ctx *efdctx;
585 efdctx = eventfd_ctx_fdget(fd);
587 return PTR_ERR(efdctx);
590 eventfd_ctx_put(*ctx);
600 static int vfio_pci_set_err_trigger(struct vfio_pci_device *vdev,
601 unsigned index, unsigned start,
602 unsigned count, uint32_t flags, void *data)
604 if (index != VFIO_PCI_ERR_IRQ_INDEX || start != 0 || count > 1)
607 return vfio_pci_set_ctx_trigger_single(&vdev->err_trigger,
611 static int vfio_pci_set_req_trigger(struct vfio_pci_device *vdev,
612 unsigned index, unsigned start,
613 unsigned count, uint32_t flags, void *data)
615 if (index != VFIO_PCI_REQ_IRQ_INDEX || start != 0 || count > 1)
618 return vfio_pci_set_ctx_trigger_single(&vdev->req_trigger,
622 int vfio_pci_set_irqs_ioctl(struct vfio_pci_device *vdev, uint32_t flags,
623 unsigned index, unsigned start, unsigned count,
626 int (*func)(struct vfio_pci_device *vdev, unsigned index,
627 unsigned start, unsigned count, uint32_t flags,
631 case VFIO_PCI_INTX_IRQ_INDEX:
632 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
633 case VFIO_IRQ_SET_ACTION_MASK:
634 func = vfio_pci_set_intx_mask;
636 case VFIO_IRQ_SET_ACTION_UNMASK:
637 func = vfio_pci_set_intx_unmask;
639 case VFIO_IRQ_SET_ACTION_TRIGGER:
640 func = vfio_pci_set_intx_trigger;
644 case VFIO_PCI_MSI_IRQ_INDEX:
645 case VFIO_PCI_MSIX_IRQ_INDEX:
646 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
647 case VFIO_IRQ_SET_ACTION_MASK:
648 case VFIO_IRQ_SET_ACTION_UNMASK:
649 /* XXX Need masking support exported */
651 case VFIO_IRQ_SET_ACTION_TRIGGER:
652 func = vfio_pci_set_msi_trigger;
656 case VFIO_PCI_ERR_IRQ_INDEX:
657 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
658 case VFIO_IRQ_SET_ACTION_TRIGGER:
659 if (pci_is_pcie(vdev->pdev))
660 func = vfio_pci_set_err_trigger;
664 case VFIO_PCI_REQ_IRQ_INDEX:
665 switch (flags & VFIO_IRQ_SET_ACTION_TYPE_MASK) {
666 case VFIO_IRQ_SET_ACTION_TRIGGER:
667 func = vfio_pci_set_req_trigger;
676 return func(vdev, index, start, count, flags, data);