2 * Kernel-based Virtual Machine - device assignment support
4 * Copyright (C) 2010 Red Hat, Inc. and/or its affiliates.
6 * This work is licensed under the terms of the GNU GPL, version 2. See
7 * the COPYING file in the top-level directory.
11 #include <linux/kvm_host.h>
12 #include <linux/kvm.h>
13 #include <linux/uaccess.h>
14 #include <linux/vmalloc.h>
15 #include <linux/errno.h>
16 #include <linux/spinlock.h>
17 #include <linux/pci.h>
18 #include <linux/interrupt.h>
19 #include <linux/slab.h>
22 static struct kvm_assigned_dev_kernel *kvm_find_assigned_dev(struct list_head *head,
25 struct list_head *ptr;
26 struct kvm_assigned_dev_kernel *match;
28 list_for_each(ptr, head) {
29 match = list_entry(ptr, struct kvm_assigned_dev_kernel, list);
30 if (match->assigned_dev_id == assigned_dev_id)
36 static int find_index_from_host_irq(struct kvm_assigned_dev_kernel
37 *assigned_dev, int irq)
40 struct msix_entry *host_msix_entries;
42 host_msix_entries = assigned_dev->host_msix_entries;
45 for (i = 0; i < assigned_dev->entries_nr; i++)
46 if (irq == host_msix_entries[i].vector) {
51 printk(KERN_WARNING "Fail to find correlated MSI-X entry!\n");
58 static void kvm_assigned_dev_interrupt_work_handler(struct work_struct *work)
60 struct kvm_assigned_dev_kernel *assigned_dev;
63 assigned_dev = container_of(work, struct kvm_assigned_dev_kernel,
66 spin_lock_irq(&assigned_dev->assigned_dev_lock);
67 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
68 struct kvm_guest_msix_entry *guest_entries =
69 assigned_dev->guest_msix_entries;
70 for (i = 0; i < assigned_dev->entries_nr; i++) {
71 if (!(guest_entries[i].flags &
72 KVM_ASSIGNED_MSIX_PENDING))
74 guest_entries[i].flags &= ~KVM_ASSIGNED_MSIX_PENDING;
75 kvm_set_irq(assigned_dev->kvm,
76 assigned_dev->irq_source_id,
77 guest_entries[i].vector, 1);
80 kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
81 assigned_dev->guest_irq, 1);
83 spin_unlock_irq(&assigned_dev->assigned_dev_lock);
86 static irqreturn_t kvm_assigned_dev_intr(int irq, void *dev_id)
89 struct kvm_assigned_dev_kernel *assigned_dev =
90 (struct kvm_assigned_dev_kernel *) dev_id;
92 spin_lock_irqsave(&assigned_dev->assigned_dev_lock, flags);
93 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
94 int index = find_index_from_host_irq(assigned_dev, irq);
97 assigned_dev->guest_msix_entries[index].flags |=
98 KVM_ASSIGNED_MSIX_PENDING;
101 schedule_work(&assigned_dev->interrupt_work);
103 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_GUEST_INTX) {
104 disable_irq_nosync(irq);
105 assigned_dev->host_irq_disabled = true;
109 spin_unlock_irqrestore(&assigned_dev->assigned_dev_lock, flags);
113 /* Ack the irq line for an assigned device */
114 static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
116 struct kvm_assigned_dev_kernel *dev;
122 dev = container_of(kian, struct kvm_assigned_dev_kernel,
125 kvm_set_irq(dev->kvm, dev->irq_source_id, dev->guest_irq, 0);
127 /* The guest irq may be shared so this ack may be
128 * from another device.
130 spin_lock_irqsave(&dev->assigned_dev_lock, flags);
131 if (dev->host_irq_disabled) {
132 enable_irq(dev->host_irq);
133 dev->host_irq_disabled = false;
135 spin_unlock_irqrestore(&dev->assigned_dev_lock, flags);
138 static void deassign_guest_irq(struct kvm *kvm,
139 struct kvm_assigned_dev_kernel *assigned_dev)
141 kvm_unregister_irq_ack_notifier(kvm, &assigned_dev->ack_notifier);
142 assigned_dev->ack_notifier.gsi = -1;
144 if (assigned_dev->irq_source_id != -1)
145 kvm_free_irq_source_id(kvm, assigned_dev->irq_source_id);
146 assigned_dev->irq_source_id = -1;
147 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_GUEST_MASK);
150 /* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
151 static void deassign_host_irq(struct kvm *kvm,
152 struct kvm_assigned_dev_kernel *assigned_dev)
155 * In kvm_free_device_irq, cancel_work_sync return true if:
156 * 1. work is scheduled, and then cancelled.
157 * 2. work callback is executed.
159 * The first one ensured that the irq is disabled and no more events
160 * would happen. But for the second one, the irq may be enabled (e.g.
161 * for MSI). So we disable irq here to prevent further events.
163 * Notice this maybe result in nested disable if the interrupt type is
164 * INTx, but it's OK for we are going to free it.
166 * If this function is a part of VM destroy, please ensure that till
167 * now, the kvm state is still legal for probably we also have to wait
168 * interrupt_work done.
170 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
172 for (i = 0; i < assigned_dev->entries_nr; i++)
173 disable_irq_nosync(assigned_dev->
174 host_msix_entries[i].vector);
176 cancel_work_sync(&assigned_dev->interrupt_work);
178 for (i = 0; i < assigned_dev->entries_nr; i++)
179 free_irq(assigned_dev->host_msix_entries[i].vector,
180 (void *)assigned_dev);
182 assigned_dev->entries_nr = 0;
183 kfree(assigned_dev->host_msix_entries);
184 kfree(assigned_dev->guest_msix_entries);
185 pci_disable_msix(assigned_dev->dev);
187 /* Deal with MSI and INTx */
188 disable_irq_nosync(assigned_dev->host_irq);
189 cancel_work_sync(&assigned_dev->interrupt_work);
191 free_irq(assigned_dev->host_irq, (void *)assigned_dev);
193 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSI)
194 pci_disable_msi(assigned_dev->dev);
197 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_HOST_MASK);
200 static int kvm_deassign_irq(struct kvm *kvm,
201 struct kvm_assigned_dev_kernel *assigned_dev,
202 unsigned long irq_requested_type)
204 unsigned long guest_irq_type, host_irq_type;
206 if (!irqchip_in_kernel(kvm))
208 /* no irq assignment to deassign */
209 if (!assigned_dev->irq_requested_type)
212 host_irq_type = irq_requested_type & KVM_DEV_IRQ_HOST_MASK;
213 guest_irq_type = irq_requested_type & KVM_DEV_IRQ_GUEST_MASK;
216 deassign_host_irq(kvm, assigned_dev);
218 deassign_guest_irq(kvm, assigned_dev);
223 static void kvm_free_assigned_irq(struct kvm *kvm,
224 struct kvm_assigned_dev_kernel *assigned_dev)
226 kvm_deassign_irq(kvm, assigned_dev, assigned_dev->irq_requested_type);
229 static void kvm_free_assigned_device(struct kvm *kvm,
230 struct kvm_assigned_dev_kernel
233 kvm_free_assigned_irq(kvm, assigned_dev);
235 pci_reset_function(assigned_dev->dev);
237 pci_release_regions(assigned_dev->dev);
238 pci_disable_device(assigned_dev->dev);
239 pci_dev_put(assigned_dev->dev);
241 list_del(&assigned_dev->list);
245 void kvm_free_all_assigned_devices(struct kvm *kvm)
247 struct list_head *ptr, *ptr2;
248 struct kvm_assigned_dev_kernel *assigned_dev;
250 list_for_each_safe(ptr, ptr2, &kvm->arch.assigned_dev_head) {
251 assigned_dev = list_entry(ptr,
252 struct kvm_assigned_dev_kernel,
255 kvm_free_assigned_device(kvm, assigned_dev);
259 static int assigned_device_enable_host_intx(struct kvm *kvm,
260 struct kvm_assigned_dev_kernel *dev)
262 dev->host_irq = dev->dev->irq;
263 /* Even though this is PCI, we don't want to use shared
264 * interrupts. Sharing host devices with guest-assigned devices
265 * on the same interrupt line is not a happy situation: there
266 * are going to be long delays in accepting, acking, etc.
268 if (request_irq(dev->host_irq, kvm_assigned_dev_intr,
269 0, "kvm_assigned_intx_device", (void *)dev))
274 #ifdef __KVM_HAVE_MSI
275 static int assigned_device_enable_host_msi(struct kvm *kvm,
276 struct kvm_assigned_dev_kernel *dev)
280 if (!dev->dev->msi_enabled) {
281 r = pci_enable_msi(dev->dev);
286 dev->host_irq = dev->dev->irq;
287 if (request_irq(dev->host_irq, kvm_assigned_dev_intr, 0,
288 "kvm_assigned_msi_device", (void *)dev)) {
289 pci_disable_msi(dev->dev);
297 #ifdef __KVM_HAVE_MSIX
298 static int assigned_device_enable_host_msix(struct kvm *kvm,
299 struct kvm_assigned_dev_kernel *dev)
303 /* host_msix_entries and guest_msix_entries should have been
305 if (dev->entries_nr == 0)
308 r = pci_enable_msix(dev->dev, dev->host_msix_entries, dev->entries_nr);
312 for (i = 0; i < dev->entries_nr; i++) {
313 r = request_irq(dev->host_msix_entries[i].vector,
314 kvm_assigned_dev_intr, 0,
315 "kvm_assigned_msix_device",
323 for (i -= 1; i >= 0; i--)
324 free_irq(dev->host_msix_entries[i].vector, (void *)dev);
325 pci_disable_msix(dev->dev);
331 static int assigned_device_enable_guest_intx(struct kvm *kvm,
332 struct kvm_assigned_dev_kernel *dev,
333 struct kvm_assigned_irq *irq)
335 dev->guest_irq = irq->guest_irq;
336 dev->ack_notifier.gsi = irq->guest_irq;
340 #ifdef __KVM_HAVE_MSI
341 static int assigned_device_enable_guest_msi(struct kvm *kvm,
342 struct kvm_assigned_dev_kernel *dev,
343 struct kvm_assigned_irq *irq)
345 dev->guest_irq = irq->guest_irq;
346 dev->ack_notifier.gsi = -1;
347 dev->host_irq_disabled = false;
352 #ifdef __KVM_HAVE_MSIX
353 static int assigned_device_enable_guest_msix(struct kvm *kvm,
354 struct kvm_assigned_dev_kernel *dev,
355 struct kvm_assigned_irq *irq)
357 dev->guest_irq = irq->guest_irq;
358 dev->ack_notifier.gsi = -1;
359 dev->host_irq_disabled = false;
364 static int assign_host_irq(struct kvm *kvm,
365 struct kvm_assigned_dev_kernel *dev,
370 if (dev->irq_requested_type & KVM_DEV_IRQ_HOST_MASK)
373 switch (host_irq_type) {
374 case KVM_DEV_IRQ_HOST_INTX:
375 r = assigned_device_enable_host_intx(kvm, dev);
377 #ifdef __KVM_HAVE_MSI
378 case KVM_DEV_IRQ_HOST_MSI:
379 r = assigned_device_enable_host_msi(kvm, dev);
382 #ifdef __KVM_HAVE_MSIX
383 case KVM_DEV_IRQ_HOST_MSIX:
384 r = assigned_device_enable_host_msix(kvm, dev);
392 dev->irq_requested_type |= host_irq_type;
397 static int assign_guest_irq(struct kvm *kvm,
398 struct kvm_assigned_dev_kernel *dev,
399 struct kvm_assigned_irq *irq,
400 unsigned long guest_irq_type)
405 if (dev->irq_requested_type & KVM_DEV_IRQ_GUEST_MASK)
408 id = kvm_request_irq_source_id(kvm);
412 dev->irq_source_id = id;
414 switch (guest_irq_type) {
415 case KVM_DEV_IRQ_GUEST_INTX:
416 r = assigned_device_enable_guest_intx(kvm, dev, irq);
418 #ifdef __KVM_HAVE_MSI
419 case KVM_DEV_IRQ_GUEST_MSI:
420 r = assigned_device_enable_guest_msi(kvm, dev, irq);
423 #ifdef __KVM_HAVE_MSIX
424 case KVM_DEV_IRQ_GUEST_MSIX:
425 r = assigned_device_enable_guest_msix(kvm, dev, irq);
433 dev->irq_requested_type |= guest_irq_type;
434 kvm_register_irq_ack_notifier(kvm, &dev->ack_notifier);
436 kvm_free_irq_source_id(kvm, dev->irq_source_id);
441 /* TODO Deal with KVM_DEV_IRQ_ASSIGNED_MASK_MSIX */
442 static int kvm_vm_ioctl_assign_irq(struct kvm *kvm,
443 struct kvm_assigned_irq *assigned_irq)
446 struct kvm_assigned_dev_kernel *match;
447 unsigned long host_irq_type, guest_irq_type;
449 if (!irqchip_in_kernel(kvm))
452 mutex_lock(&kvm->lock);
454 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
455 assigned_irq->assigned_dev_id);
459 host_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_HOST_MASK);
460 guest_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_GUEST_MASK);
463 /* can only assign one type at a time */
464 if (hweight_long(host_irq_type) > 1)
466 if (hweight_long(guest_irq_type) > 1)
468 if (host_irq_type == 0 && guest_irq_type == 0)
473 r = assign_host_irq(kvm, match, host_irq_type);
478 r = assign_guest_irq(kvm, match, assigned_irq, guest_irq_type);
480 mutex_unlock(&kvm->lock);
484 static int kvm_vm_ioctl_deassign_dev_irq(struct kvm *kvm,
485 struct kvm_assigned_irq
489 struct kvm_assigned_dev_kernel *match;
491 mutex_lock(&kvm->lock);
493 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
494 assigned_irq->assigned_dev_id);
498 r = kvm_deassign_irq(kvm, match, assigned_irq->flags);
500 mutex_unlock(&kvm->lock);
504 static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
505 struct kvm_assigned_pci_dev *assigned_dev)
508 struct kvm_assigned_dev_kernel *match;
511 mutex_lock(&kvm->lock);
512 idx = srcu_read_lock(&kvm->srcu);
514 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
515 assigned_dev->assigned_dev_id);
517 /* device already assigned */
522 match = kzalloc(sizeof(struct kvm_assigned_dev_kernel), GFP_KERNEL);
524 printk(KERN_INFO "%s: Couldn't allocate memory\n",
529 dev = pci_get_domain_bus_and_slot(assigned_dev->segnr,
531 assigned_dev->devfn);
533 printk(KERN_INFO "%s: host device not found\n", __func__);
537 if (pci_enable_device(dev)) {
538 printk(KERN_INFO "%s: Could not enable PCI device\n", __func__);
542 r = pci_request_regions(dev, "kvm_assigned_device");
544 printk(KERN_INFO "%s: Could not get access to device regions\n",
549 pci_reset_function(dev);
551 match->assigned_dev_id = assigned_dev->assigned_dev_id;
552 match->host_segnr = assigned_dev->segnr;
553 match->host_busnr = assigned_dev->busnr;
554 match->host_devfn = assigned_dev->devfn;
555 match->flags = assigned_dev->flags;
557 spin_lock_init(&match->assigned_dev_lock);
558 match->irq_source_id = -1;
560 match->ack_notifier.irq_acked = kvm_assigned_dev_ack_irq;
561 INIT_WORK(&match->interrupt_work,
562 kvm_assigned_dev_interrupt_work_handler);
564 list_add(&match->list, &kvm->arch.assigned_dev_head);
566 if (assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU) {
567 if (!kvm->arch.iommu_domain) {
568 r = kvm_iommu_map_guest(kvm);
572 r = kvm_assign_device(kvm, match);
578 srcu_read_unlock(&kvm->srcu, idx);
579 mutex_unlock(&kvm->lock);
582 list_del(&match->list);
583 pci_release_regions(dev);
585 pci_disable_device(dev);
590 srcu_read_unlock(&kvm->srcu, idx);
591 mutex_unlock(&kvm->lock);
595 static int kvm_vm_ioctl_deassign_device(struct kvm *kvm,
596 struct kvm_assigned_pci_dev *assigned_dev)
599 struct kvm_assigned_dev_kernel *match;
601 mutex_lock(&kvm->lock);
603 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
604 assigned_dev->assigned_dev_id);
606 printk(KERN_INFO "%s: device hasn't been assigned before, "
607 "so cannot be deassigned\n", __func__);
612 if (match->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU)
613 kvm_deassign_device(kvm, match);
615 kvm_free_assigned_device(kvm, match);
618 mutex_unlock(&kvm->lock);
623 #ifdef __KVM_HAVE_MSIX
624 static int kvm_vm_ioctl_set_msix_nr(struct kvm *kvm,
625 struct kvm_assigned_msix_nr *entry_nr)
628 struct kvm_assigned_dev_kernel *adev;
630 mutex_lock(&kvm->lock);
632 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
633 entry_nr->assigned_dev_id);
639 if (adev->entries_nr == 0) {
640 adev->entries_nr = entry_nr->entry_nr;
641 if (adev->entries_nr == 0 ||
642 adev->entries_nr >= KVM_MAX_MSIX_PER_DEV) {
647 adev->host_msix_entries = kzalloc(sizeof(struct msix_entry) *
650 if (!adev->host_msix_entries) {
654 adev->guest_msix_entries = kzalloc(
655 sizeof(struct kvm_guest_msix_entry) *
656 entry_nr->entry_nr, GFP_KERNEL);
657 if (!adev->guest_msix_entries) {
658 kfree(adev->host_msix_entries);
662 } else /* Not allowed set MSI-X number twice */
665 mutex_unlock(&kvm->lock);
669 static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
670 struct kvm_assigned_msix_entry *entry)
673 struct kvm_assigned_dev_kernel *adev;
675 mutex_lock(&kvm->lock);
677 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
678 entry->assigned_dev_id);
685 for (i = 0; i < adev->entries_nr; i++)
686 if (adev->guest_msix_entries[i].vector == 0 ||
687 adev->guest_msix_entries[i].entry == entry->entry) {
688 adev->guest_msix_entries[i].entry = entry->entry;
689 adev->guest_msix_entries[i].vector = entry->gsi;
690 adev->host_msix_entries[i].entry = entry->entry;
693 if (i == adev->entries_nr) {
699 mutex_unlock(&kvm->lock);
705 long kvm_vm_ioctl_assigned_device(struct kvm *kvm, unsigned ioctl,
708 void __user *argp = (void __user *)arg;
712 case KVM_ASSIGN_PCI_DEVICE: {
713 struct kvm_assigned_pci_dev assigned_dev;
716 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
718 r = kvm_vm_ioctl_assign_device(kvm, &assigned_dev);
723 case KVM_ASSIGN_IRQ: {
727 #ifdef KVM_CAP_ASSIGN_DEV_IRQ
728 case KVM_ASSIGN_DEV_IRQ: {
729 struct kvm_assigned_irq assigned_irq;
732 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
734 r = kvm_vm_ioctl_assign_irq(kvm, &assigned_irq);
739 case KVM_DEASSIGN_DEV_IRQ: {
740 struct kvm_assigned_irq assigned_irq;
743 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
745 r = kvm_vm_ioctl_deassign_dev_irq(kvm, &assigned_irq);
751 #ifdef KVM_CAP_DEVICE_DEASSIGNMENT
752 case KVM_DEASSIGN_PCI_DEVICE: {
753 struct kvm_assigned_pci_dev assigned_dev;
756 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
758 r = kvm_vm_ioctl_deassign_device(kvm, &assigned_dev);
764 #ifdef KVM_CAP_IRQ_ROUTING
765 case KVM_SET_GSI_ROUTING: {
766 struct kvm_irq_routing routing;
767 struct kvm_irq_routing __user *urouting;
768 struct kvm_irq_routing_entry *entries;
771 if (copy_from_user(&routing, argp, sizeof(routing)))
774 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
779 entries = vmalloc(routing.nr * sizeof(*entries));
784 if (copy_from_user(entries, urouting->entries,
785 routing.nr * sizeof(*entries)))
786 goto out_free_irq_routing;
787 r = kvm_set_irq_routing(kvm, entries, routing.nr,
789 out_free_irq_routing:
793 #endif /* KVM_CAP_IRQ_ROUTING */
794 #ifdef __KVM_HAVE_MSIX
795 case KVM_ASSIGN_SET_MSIX_NR: {
796 struct kvm_assigned_msix_nr entry_nr;
798 if (copy_from_user(&entry_nr, argp, sizeof entry_nr))
800 r = kvm_vm_ioctl_set_msix_nr(kvm, &entry_nr);
805 case KVM_ASSIGN_SET_MSIX_ENTRY: {
806 struct kvm_assigned_msix_entry entry;
808 if (copy_from_user(&entry, argp, sizeof entry))
810 r = kvm_vm_ioctl_set_msix_entry(kvm, &entry);