2 * Kernel-based Virtual Machine - device assignment support
4 * Copyright (C) 2006-9 Red Hat, Inc
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;
64 assigned_dev = container_of(work, struct kvm_assigned_dev_kernel,
66 kvm = assigned_dev->kvm;
68 spin_lock_irq(&assigned_dev->assigned_dev_lock);
69 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
70 struct kvm_guest_msix_entry *guest_entries =
71 assigned_dev->guest_msix_entries;
72 for (i = 0; i < assigned_dev->entries_nr; i++) {
73 if (!(guest_entries[i].flags &
74 KVM_ASSIGNED_MSIX_PENDING))
76 guest_entries[i].flags &= ~KVM_ASSIGNED_MSIX_PENDING;
77 kvm_set_irq(assigned_dev->kvm,
78 assigned_dev->irq_source_id,
79 guest_entries[i].vector, 1);
82 kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
83 assigned_dev->guest_irq, 1);
85 spin_unlock_irq(&assigned_dev->assigned_dev_lock);
88 static irqreturn_t kvm_assigned_dev_intr(int irq, void *dev_id)
91 struct kvm_assigned_dev_kernel *assigned_dev =
92 (struct kvm_assigned_dev_kernel *) dev_id;
94 spin_lock_irqsave(&assigned_dev->assigned_dev_lock, flags);
95 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
96 int index = find_index_from_host_irq(assigned_dev, irq);
99 assigned_dev->guest_msix_entries[index].flags |=
100 KVM_ASSIGNED_MSIX_PENDING;
103 schedule_work(&assigned_dev->interrupt_work);
105 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_GUEST_INTX) {
106 disable_irq_nosync(irq);
107 assigned_dev->host_irq_disabled = true;
111 spin_unlock_irqrestore(&assigned_dev->assigned_dev_lock, flags);
115 /* Ack the irq line for an assigned device */
116 static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
118 struct kvm_assigned_dev_kernel *dev;
124 dev = container_of(kian, struct kvm_assigned_dev_kernel,
127 kvm_set_irq(dev->kvm, dev->irq_source_id, dev->guest_irq, 0);
129 /* The guest irq may be shared so this ack may be
130 * from another device.
132 spin_lock_irqsave(&dev->assigned_dev_lock, flags);
133 if (dev->host_irq_disabled) {
134 enable_irq(dev->host_irq);
135 dev->host_irq_disabled = false;
137 spin_unlock_irqrestore(&dev->assigned_dev_lock, flags);
140 static void deassign_guest_irq(struct kvm *kvm,
141 struct kvm_assigned_dev_kernel *assigned_dev)
143 kvm_unregister_irq_ack_notifier(kvm, &assigned_dev->ack_notifier);
144 assigned_dev->ack_notifier.gsi = -1;
146 if (assigned_dev->irq_source_id != -1)
147 kvm_free_irq_source_id(kvm, assigned_dev->irq_source_id);
148 assigned_dev->irq_source_id = -1;
149 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_GUEST_MASK);
152 /* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
153 static void deassign_host_irq(struct kvm *kvm,
154 struct kvm_assigned_dev_kernel *assigned_dev)
157 * In kvm_free_device_irq, cancel_work_sync return true if:
158 * 1. work is scheduled, and then cancelled.
159 * 2. work callback is executed.
161 * The first one ensured that the irq is disabled and no more events
162 * would happen. But for the second one, the irq may be enabled (e.g.
163 * for MSI). So we disable irq here to prevent further events.
165 * Notice this maybe result in nested disable if the interrupt type is
166 * INTx, but it's OK for we are going to free it.
168 * If this function is a part of VM destroy, please ensure that till
169 * now, the kvm state is still legal for probably we also have to wait
170 * interrupt_work done.
172 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
174 for (i = 0; i < assigned_dev->entries_nr; i++)
175 disable_irq_nosync(assigned_dev->
176 host_msix_entries[i].vector);
178 cancel_work_sync(&assigned_dev->interrupt_work);
180 for (i = 0; i < assigned_dev->entries_nr; i++)
181 free_irq(assigned_dev->host_msix_entries[i].vector,
182 (void *)assigned_dev);
184 assigned_dev->entries_nr = 0;
185 kfree(assigned_dev->host_msix_entries);
186 kfree(assigned_dev->guest_msix_entries);
187 pci_disable_msix(assigned_dev->dev);
189 /* Deal with MSI and INTx */
190 disable_irq_nosync(assigned_dev->host_irq);
191 cancel_work_sync(&assigned_dev->interrupt_work);
193 free_irq(assigned_dev->host_irq, (void *)assigned_dev);
195 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSI)
196 pci_disable_msi(assigned_dev->dev);
199 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_HOST_MASK);
202 static int kvm_deassign_irq(struct kvm *kvm,
203 struct kvm_assigned_dev_kernel *assigned_dev,
204 unsigned long irq_requested_type)
206 unsigned long guest_irq_type, host_irq_type;
208 if (!irqchip_in_kernel(kvm))
210 /* no irq assignment to deassign */
211 if (!assigned_dev->irq_requested_type)
214 host_irq_type = irq_requested_type & KVM_DEV_IRQ_HOST_MASK;
215 guest_irq_type = irq_requested_type & KVM_DEV_IRQ_GUEST_MASK;
218 deassign_host_irq(kvm, assigned_dev);
220 deassign_guest_irq(kvm, assigned_dev);
225 static void kvm_free_assigned_irq(struct kvm *kvm,
226 struct kvm_assigned_dev_kernel *assigned_dev)
228 kvm_deassign_irq(kvm, assigned_dev, assigned_dev->irq_requested_type);
231 static void kvm_free_assigned_device(struct kvm *kvm,
232 struct kvm_assigned_dev_kernel
235 kvm_free_assigned_irq(kvm, assigned_dev);
237 pci_reset_function(assigned_dev->dev);
239 pci_release_regions(assigned_dev->dev);
240 pci_disable_device(assigned_dev->dev);
241 pci_dev_put(assigned_dev->dev);
243 list_del(&assigned_dev->list);
247 void kvm_free_all_assigned_devices(struct kvm *kvm)
249 struct list_head *ptr, *ptr2;
250 struct kvm_assigned_dev_kernel *assigned_dev;
252 list_for_each_safe(ptr, ptr2, &kvm->arch.assigned_dev_head) {
253 assigned_dev = list_entry(ptr,
254 struct kvm_assigned_dev_kernel,
257 kvm_free_assigned_device(kvm, assigned_dev);
261 static int assigned_device_enable_host_intx(struct kvm *kvm,
262 struct kvm_assigned_dev_kernel *dev)
264 dev->host_irq = dev->dev->irq;
265 /* Even though this is PCI, we don't want to use shared
266 * interrupts. Sharing host devices with guest-assigned devices
267 * on the same interrupt line is not a happy situation: there
268 * are going to be long delays in accepting, acking, etc.
270 if (request_irq(dev->host_irq, kvm_assigned_dev_intr,
271 0, "kvm_assigned_intx_device", (void *)dev))
276 #ifdef __KVM_HAVE_MSI
277 static int assigned_device_enable_host_msi(struct kvm *kvm,
278 struct kvm_assigned_dev_kernel *dev)
282 if (!dev->dev->msi_enabled) {
283 r = pci_enable_msi(dev->dev);
288 dev->host_irq = dev->dev->irq;
289 if (request_irq(dev->host_irq, kvm_assigned_dev_intr, 0,
290 "kvm_assigned_msi_device", (void *)dev)) {
291 pci_disable_msi(dev->dev);
299 #ifdef __KVM_HAVE_MSIX
300 static int assigned_device_enable_host_msix(struct kvm *kvm,
301 struct kvm_assigned_dev_kernel *dev)
305 /* host_msix_entries and guest_msix_entries should have been
307 if (dev->entries_nr == 0)
310 r = pci_enable_msix(dev->dev, dev->host_msix_entries, dev->entries_nr);
314 for (i = 0; i < dev->entries_nr; i++) {
315 r = request_irq(dev->host_msix_entries[i].vector,
316 kvm_assigned_dev_intr, 0,
317 "kvm_assigned_msix_device",
325 for (i -= 1; i >= 0; i--)
326 free_irq(dev->host_msix_entries[i].vector, (void *)dev);
327 pci_disable_msix(dev->dev);
333 static int assigned_device_enable_guest_intx(struct kvm *kvm,
334 struct kvm_assigned_dev_kernel *dev,
335 struct kvm_assigned_irq *irq)
337 dev->guest_irq = irq->guest_irq;
338 dev->ack_notifier.gsi = irq->guest_irq;
342 #ifdef __KVM_HAVE_MSI
343 static int assigned_device_enable_guest_msi(struct kvm *kvm,
344 struct kvm_assigned_dev_kernel *dev,
345 struct kvm_assigned_irq *irq)
347 dev->guest_irq = irq->guest_irq;
348 dev->ack_notifier.gsi = -1;
349 dev->host_irq_disabled = false;
354 #ifdef __KVM_HAVE_MSIX
355 static int assigned_device_enable_guest_msix(struct kvm *kvm,
356 struct kvm_assigned_dev_kernel *dev,
357 struct kvm_assigned_irq *irq)
359 dev->guest_irq = irq->guest_irq;
360 dev->ack_notifier.gsi = -1;
361 dev->host_irq_disabled = false;
366 static int assign_host_irq(struct kvm *kvm,
367 struct kvm_assigned_dev_kernel *dev,
372 if (dev->irq_requested_type & KVM_DEV_IRQ_HOST_MASK)
375 switch (host_irq_type) {
376 case KVM_DEV_IRQ_HOST_INTX:
377 r = assigned_device_enable_host_intx(kvm, dev);
379 #ifdef __KVM_HAVE_MSI
380 case KVM_DEV_IRQ_HOST_MSI:
381 r = assigned_device_enable_host_msi(kvm, dev);
384 #ifdef __KVM_HAVE_MSIX
385 case KVM_DEV_IRQ_HOST_MSIX:
386 r = assigned_device_enable_host_msix(kvm, dev);
394 dev->irq_requested_type |= host_irq_type;
399 static int assign_guest_irq(struct kvm *kvm,
400 struct kvm_assigned_dev_kernel *dev,
401 struct kvm_assigned_irq *irq,
402 unsigned long guest_irq_type)
407 if (dev->irq_requested_type & KVM_DEV_IRQ_GUEST_MASK)
410 id = kvm_request_irq_source_id(kvm);
414 dev->irq_source_id = id;
416 switch (guest_irq_type) {
417 case KVM_DEV_IRQ_GUEST_INTX:
418 r = assigned_device_enable_guest_intx(kvm, dev, irq);
420 #ifdef __KVM_HAVE_MSI
421 case KVM_DEV_IRQ_GUEST_MSI:
422 r = assigned_device_enable_guest_msi(kvm, dev, irq);
425 #ifdef __KVM_HAVE_MSIX
426 case KVM_DEV_IRQ_GUEST_MSIX:
427 r = assigned_device_enable_guest_msix(kvm, dev, irq);
435 dev->irq_requested_type |= guest_irq_type;
436 kvm_register_irq_ack_notifier(kvm, &dev->ack_notifier);
438 kvm_free_irq_source_id(kvm, dev->irq_source_id);
443 /* TODO Deal with KVM_DEV_IRQ_ASSIGNED_MASK_MSIX */
444 static int kvm_vm_ioctl_assign_irq(struct kvm *kvm,
445 struct kvm_assigned_irq *assigned_irq)
448 struct kvm_assigned_dev_kernel *match;
449 unsigned long host_irq_type, guest_irq_type;
451 if (!capable(CAP_SYS_RAWIO))
454 if (!irqchip_in_kernel(kvm))
457 mutex_lock(&kvm->lock);
459 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
460 assigned_irq->assigned_dev_id);
464 host_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_HOST_MASK);
465 guest_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_GUEST_MASK);
468 /* can only assign one type at a time */
469 if (hweight_long(host_irq_type) > 1)
471 if (hweight_long(guest_irq_type) > 1)
473 if (host_irq_type == 0 && guest_irq_type == 0)
478 r = assign_host_irq(kvm, match, host_irq_type);
483 r = assign_guest_irq(kvm, match, assigned_irq, guest_irq_type);
485 mutex_unlock(&kvm->lock);
489 static int kvm_vm_ioctl_deassign_dev_irq(struct kvm *kvm,
490 struct kvm_assigned_irq
494 struct kvm_assigned_dev_kernel *match;
496 mutex_lock(&kvm->lock);
498 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
499 assigned_irq->assigned_dev_id);
503 r = kvm_deassign_irq(kvm, match, assigned_irq->flags);
505 mutex_unlock(&kvm->lock);
509 static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
510 struct kvm_assigned_pci_dev *assigned_dev)
513 struct kvm_assigned_dev_kernel *match;
516 mutex_lock(&kvm->lock);
517 idx = srcu_read_lock(&kvm->srcu);
519 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
520 assigned_dev->assigned_dev_id);
522 /* device already assigned */
527 match = kzalloc(sizeof(struct kvm_assigned_dev_kernel), GFP_KERNEL);
529 printk(KERN_INFO "%s: Couldn't allocate memory\n",
534 dev = pci_get_domain_bus_and_slot(assigned_dev->segnr,
536 assigned_dev->devfn);
538 printk(KERN_INFO "%s: host device not found\n", __func__);
542 if (pci_enable_device(dev)) {
543 printk(KERN_INFO "%s: Could not enable PCI device\n", __func__);
547 r = pci_request_regions(dev, "kvm_assigned_device");
549 printk(KERN_INFO "%s: Could not get access to device regions\n",
554 pci_reset_function(dev);
556 match->assigned_dev_id = assigned_dev->assigned_dev_id;
557 match->host_segnr = assigned_dev->segnr;
558 match->host_busnr = assigned_dev->busnr;
559 match->host_devfn = assigned_dev->devfn;
560 match->flags = assigned_dev->flags;
562 spin_lock_init(&match->assigned_dev_lock);
563 match->irq_source_id = -1;
565 match->ack_notifier.irq_acked = kvm_assigned_dev_ack_irq;
566 INIT_WORK(&match->interrupt_work,
567 kvm_assigned_dev_interrupt_work_handler);
569 list_add(&match->list, &kvm->arch.assigned_dev_head);
571 if (assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU) {
572 if (!kvm->arch.iommu_domain) {
573 r = kvm_iommu_map_guest(kvm);
577 r = kvm_assign_device(kvm, match);
583 srcu_read_unlock(&kvm->srcu, idx);
584 mutex_unlock(&kvm->lock);
587 list_del(&match->list);
588 pci_release_regions(dev);
590 pci_disable_device(dev);
595 srcu_read_unlock(&kvm->srcu, idx);
596 mutex_unlock(&kvm->lock);
600 static int kvm_vm_ioctl_deassign_device(struct kvm *kvm,
601 struct kvm_assigned_pci_dev *assigned_dev)
604 struct kvm_assigned_dev_kernel *match;
606 mutex_lock(&kvm->lock);
608 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
609 assigned_dev->assigned_dev_id);
611 printk(KERN_INFO "%s: device hasn't been assigned before, "
612 "so cannot be deassigned\n", __func__);
617 if (match->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU)
618 kvm_deassign_device(kvm, match);
620 kvm_free_assigned_device(kvm, match);
623 mutex_unlock(&kvm->lock);
628 #ifdef __KVM_HAVE_MSIX
629 static int kvm_vm_ioctl_set_msix_nr(struct kvm *kvm,
630 struct kvm_assigned_msix_nr *entry_nr)
633 struct kvm_assigned_dev_kernel *adev;
635 mutex_lock(&kvm->lock);
637 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
638 entry_nr->assigned_dev_id);
644 if (adev->entries_nr == 0) {
645 adev->entries_nr = entry_nr->entry_nr;
646 if (adev->entries_nr == 0 ||
647 adev->entries_nr >= KVM_MAX_MSIX_PER_DEV) {
652 adev->host_msix_entries = kzalloc(sizeof(struct msix_entry) *
655 if (!adev->host_msix_entries) {
659 adev->guest_msix_entries = kzalloc(
660 sizeof(struct kvm_guest_msix_entry) *
661 entry_nr->entry_nr, GFP_KERNEL);
662 if (!adev->guest_msix_entries) {
663 kfree(adev->host_msix_entries);
667 } else /* Not allowed set MSI-X number twice */
670 mutex_unlock(&kvm->lock);
674 static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
675 struct kvm_assigned_msix_entry *entry)
678 struct kvm_assigned_dev_kernel *adev;
680 mutex_lock(&kvm->lock);
682 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
683 entry->assigned_dev_id);
690 for (i = 0; i < adev->entries_nr; i++)
691 if (adev->guest_msix_entries[i].vector == 0 ||
692 adev->guest_msix_entries[i].entry == entry->entry) {
693 adev->guest_msix_entries[i].entry = entry->entry;
694 adev->guest_msix_entries[i].vector = entry->gsi;
695 adev->host_msix_entries[i].entry = entry->entry;
698 if (i == adev->entries_nr) {
704 mutex_unlock(&kvm->lock);
710 long kvm_vm_ioctl_assigned_device(struct kvm *kvm, unsigned ioctl,
713 void __user *argp = (void __user *)arg;
717 case KVM_ASSIGN_PCI_DEVICE: {
718 struct kvm_assigned_pci_dev assigned_dev;
721 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
723 r = kvm_vm_ioctl_assign_device(kvm, &assigned_dev);
728 case KVM_ASSIGN_IRQ: {
732 #ifdef KVM_CAP_ASSIGN_DEV_IRQ
733 case KVM_ASSIGN_DEV_IRQ: {
734 struct kvm_assigned_irq assigned_irq;
737 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
739 r = kvm_vm_ioctl_assign_irq(kvm, &assigned_irq);
744 case KVM_DEASSIGN_DEV_IRQ: {
745 struct kvm_assigned_irq assigned_irq;
748 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
750 r = kvm_vm_ioctl_deassign_dev_irq(kvm, &assigned_irq);
756 #ifdef KVM_CAP_DEVICE_DEASSIGNMENT
757 case KVM_DEASSIGN_PCI_DEVICE: {
758 struct kvm_assigned_pci_dev assigned_dev;
761 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
763 r = kvm_vm_ioctl_deassign_device(kvm, &assigned_dev);
769 #ifdef KVM_CAP_IRQ_ROUTING
770 case KVM_SET_GSI_ROUTING: {
771 struct kvm_irq_routing routing;
772 struct kvm_irq_routing __user *urouting;
773 struct kvm_irq_routing_entry *entries;
776 if (copy_from_user(&routing, argp, sizeof(routing)))
779 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
784 entries = vmalloc(routing.nr * sizeof(*entries));
789 if (copy_from_user(entries, urouting->entries,
790 routing.nr * sizeof(*entries)))
791 goto out_free_irq_routing;
792 r = kvm_set_irq_routing(kvm, entries, routing.nr,
794 out_free_irq_routing:
798 #endif /* KVM_CAP_IRQ_ROUTING */
799 #ifdef __KVM_HAVE_MSIX
800 case KVM_ASSIGN_SET_MSIX_NR: {
801 struct kvm_assigned_msix_nr entry_nr;
803 if (copy_from_user(&entry_nr, argp, sizeof entry_nr))
805 r = kvm_vm_ioctl_set_msix_nr(kvm, &entry_nr);
810 case KVM_ASSIGN_SET_MSIX_ENTRY: {
811 struct kvm_assigned_msix_entry entry;
813 if (copy_from_user(&entry, argp, sizeof entry))
815 r = kvm_vm_ioctl_set_msix_entry(kvm, &entry);