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 irqreturn_t kvm_assigned_dev_thread(int irq, void *dev_id)
60 struct kvm_assigned_dev_kernel *assigned_dev = dev_id;
64 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_INTX) {
65 spin_lock(&assigned_dev->intx_lock);
66 disable_irq_nosync(irq);
67 assigned_dev->host_irq_disabled = true;
68 spin_unlock(&assigned_dev->intx_lock);
71 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
72 index = find_index_from_host_irq(assigned_dev, irq);
74 vector = assigned_dev->
75 guest_msix_entries[index].vector;
76 kvm_set_irq(assigned_dev->kvm,
77 assigned_dev->irq_source_id, vector, 1);
80 kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
81 assigned_dev->guest_irq, 1);
86 /* Ack the irq line for an assigned device */
87 static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
89 struct kvm_assigned_dev_kernel *dev;
94 dev = container_of(kian, struct kvm_assigned_dev_kernel,
97 kvm_set_irq(dev->kvm, dev->irq_source_id, dev->guest_irq, 0);
99 /* The guest irq may be shared so this ack may be
100 * from another device.
102 spin_lock(&dev->intx_lock);
103 if (dev->host_irq_disabled) {
104 enable_irq(dev->host_irq);
105 dev->host_irq_disabled = false;
107 spin_unlock(&dev->intx_lock);
110 static void deassign_guest_irq(struct kvm *kvm,
111 struct kvm_assigned_dev_kernel *assigned_dev)
113 kvm_unregister_irq_ack_notifier(kvm, &assigned_dev->ack_notifier);
114 assigned_dev->ack_notifier.gsi = -1;
116 kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
117 assigned_dev->guest_irq, 0);
119 if (assigned_dev->irq_source_id != -1)
120 kvm_free_irq_source_id(kvm, assigned_dev->irq_source_id);
121 assigned_dev->irq_source_id = -1;
122 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_GUEST_MASK);
125 /* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
126 static void deassign_host_irq(struct kvm *kvm,
127 struct kvm_assigned_dev_kernel *assigned_dev)
130 * We disable irq here to prevent further events.
132 * Notice this maybe result in nested disable if the interrupt type is
133 * INTx, but it's OK for we are going to free it.
135 * If this function is a part of VM destroy, please ensure that till
136 * now, the kvm state is still legal for probably we also have to wait
137 * on a currently running IRQ handler.
139 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
141 for (i = 0; i < assigned_dev->entries_nr; i++)
142 disable_irq(assigned_dev->host_msix_entries[i].vector);
144 for (i = 0; i < assigned_dev->entries_nr; i++)
145 free_irq(assigned_dev->host_msix_entries[i].vector,
146 (void *)assigned_dev);
148 assigned_dev->entries_nr = 0;
149 kfree(assigned_dev->host_msix_entries);
150 kfree(assigned_dev->guest_msix_entries);
151 pci_disable_msix(assigned_dev->dev);
153 /* Deal with MSI and INTx */
154 disable_irq(assigned_dev->host_irq);
156 free_irq(assigned_dev->host_irq, (void *)assigned_dev);
158 if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSI)
159 pci_disable_msi(assigned_dev->dev);
162 assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_HOST_MASK);
165 static int kvm_deassign_irq(struct kvm *kvm,
166 struct kvm_assigned_dev_kernel *assigned_dev,
167 unsigned long irq_requested_type)
169 unsigned long guest_irq_type, host_irq_type;
171 if (!irqchip_in_kernel(kvm))
173 /* no irq assignment to deassign */
174 if (!assigned_dev->irq_requested_type)
177 host_irq_type = irq_requested_type & KVM_DEV_IRQ_HOST_MASK;
178 guest_irq_type = irq_requested_type & KVM_DEV_IRQ_GUEST_MASK;
181 deassign_host_irq(kvm, assigned_dev);
183 deassign_guest_irq(kvm, assigned_dev);
188 static void kvm_free_assigned_irq(struct kvm *kvm,
189 struct kvm_assigned_dev_kernel *assigned_dev)
191 kvm_deassign_irq(kvm, assigned_dev, assigned_dev->irq_requested_type);
194 static void kvm_free_assigned_device(struct kvm *kvm,
195 struct kvm_assigned_dev_kernel
198 kvm_free_assigned_irq(kvm, assigned_dev);
200 pci_reset_function(assigned_dev->dev);
201 if (pci_load_and_free_saved_state(assigned_dev->dev,
202 &assigned_dev->pci_saved_state))
203 printk(KERN_INFO "%s: Couldn't reload %s saved state\n",
204 __func__, dev_name(&assigned_dev->dev->dev));
206 pci_restore_state(assigned_dev->dev);
208 assigned_dev->dev->dev_flags &= ~PCI_DEV_FLAGS_ASSIGNED;
210 pci_release_regions(assigned_dev->dev);
211 pci_disable_device(assigned_dev->dev);
212 pci_dev_put(assigned_dev->dev);
214 list_del(&assigned_dev->list);
218 void kvm_free_all_assigned_devices(struct kvm *kvm)
220 struct list_head *ptr, *ptr2;
221 struct kvm_assigned_dev_kernel *assigned_dev;
223 list_for_each_safe(ptr, ptr2, &kvm->arch.assigned_dev_head) {
224 assigned_dev = list_entry(ptr,
225 struct kvm_assigned_dev_kernel,
228 kvm_free_assigned_device(kvm, assigned_dev);
232 static int assigned_device_enable_host_intx(struct kvm *kvm,
233 struct kvm_assigned_dev_kernel *dev)
235 dev->host_irq = dev->dev->irq;
236 /* Even though this is PCI, we don't want to use shared
237 * interrupts. Sharing host devices with guest-assigned devices
238 * on the same interrupt line is not a happy situation: there
239 * are going to be long delays in accepting, acking, etc.
241 if (request_threaded_irq(dev->host_irq, NULL, kvm_assigned_dev_thread,
242 IRQF_ONESHOT, dev->irq_name, (void *)dev))
247 #ifdef __KVM_HAVE_MSI
248 static int assigned_device_enable_host_msi(struct kvm *kvm,
249 struct kvm_assigned_dev_kernel *dev)
253 if (!dev->dev->msi_enabled) {
254 r = pci_enable_msi(dev->dev);
259 dev->host_irq = dev->dev->irq;
260 if (request_threaded_irq(dev->host_irq, NULL, kvm_assigned_dev_thread,
261 0, dev->irq_name, (void *)dev)) {
262 pci_disable_msi(dev->dev);
270 #ifdef __KVM_HAVE_MSIX
271 static int assigned_device_enable_host_msix(struct kvm *kvm,
272 struct kvm_assigned_dev_kernel *dev)
276 /* host_msix_entries and guest_msix_entries should have been
278 if (dev->entries_nr == 0)
281 r = pci_enable_msix(dev->dev, dev->host_msix_entries, dev->entries_nr);
285 for (i = 0; i < dev->entries_nr; i++) {
286 r = request_threaded_irq(dev->host_msix_entries[i].vector,
287 NULL, kvm_assigned_dev_thread,
288 0, dev->irq_name, (void *)dev);
295 for (i -= 1; i >= 0; i--)
296 free_irq(dev->host_msix_entries[i].vector, (void *)dev);
297 pci_disable_msix(dev->dev);
303 static int assigned_device_enable_guest_intx(struct kvm *kvm,
304 struct kvm_assigned_dev_kernel *dev,
305 struct kvm_assigned_irq *irq)
307 dev->guest_irq = irq->guest_irq;
308 dev->ack_notifier.gsi = irq->guest_irq;
312 #ifdef __KVM_HAVE_MSI
313 static int assigned_device_enable_guest_msi(struct kvm *kvm,
314 struct kvm_assigned_dev_kernel *dev,
315 struct kvm_assigned_irq *irq)
317 dev->guest_irq = irq->guest_irq;
318 dev->ack_notifier.gsi = -1;
319 dev->host_irq_disabled = false;
324 #ifdef __KVM_HAVE_MSIX
325 static int assigned_device_enable_guest_msix(struct kvm *kvm,
326 struct kvm_assigned_dev_kernel *dev,
327 struct kvm_assigned_irq *irq)
329 dev->guest_irq = irq->guest_irq;
330 dev->ack_notifier.gsi = -1;
331 dev->host_irq_disabled = false;
336 static int assign_host_irq(struct kvm *kvm,
337 struct kvm_assigned_dev_kernel *dev,
342 if (dev->irq_requested_type & KVM_DEV_IRQ_HOST_MASK)
345 snprintf(dev->irq_name, sizeof(dev->irq_name), "kvm:%s",
348 switch (host_irq_type) {
349 case KVM_DEV_IRQ_HOST_INTX:
350 r = assigned_device_enable_host_intx(kvm, dev);
352 #ifdef __KVM_HAVE_MSI
353 case KVM_DEV_IRQ_HOST_MSI:
354 r = assigned_device_enable_host_msi(kvm, dev);
357 #ifdef __KVM_HAVE_MSIX
358 case KVM_DEV_IRQ_HOST_MSIX:
359 r = assigned_device_enable_host_msix(kvm, dev);
367 dev->irq_requested_type |= host_irq_type;
372 static int assign_guest_irq(struct kvm *kvm,
373 struct kvm_assigned_dev_kernel *dev,
374 struct kvm_assigned_irq *irq,
375 unsigned long guest_irq_type)
380 if (dev->irq_requested_type & KVM_DEV_IRQ_GUEST_MASK)
383 id = kvm_request_irq_source_id(kvm);
387 dev->irq_source_id = id;
389 switch (guest_irq_type) {
390 case KVM_DEV_IRQ_GUEST_INTX:
391 r = assigned_device_enable_guest_intx(kvm, dev, irq);
393 #ifdef __KVM_HAVE_MSI
394 case KVM_DEV_IRQ_GUEST_MSI:
395 r = assigned_device_enable_guest_msi(kvm, dev, irq);
398 #ifdef __KVM_HAVE_MSIX
399 case KVM_DEV_IRQ_GUEST_MSIX:
400 r = assigned_device_enable_guest_msix(kvm, dev, irq);
408 dev->irq_requested_type |= guest_irq_type;
409 kvm_register_irq_ack_notifier(kvm, &dev->ack_notifier);
411 kvm_free_irq_source_id(kvm, dev->irq_source_id);
416 /* TODO Deal with KVM_DEV_IRQ_ASSIGNED_MASK_MSIX */
417 static int kvm_vm_ioctl_assign_irq(struct kvm *kvm,
418 struct kvm_assigned_irq *assigned_irq)
421 struct kvm_assigned_dev_kernel *match;
422 unsigned long host_irq_type, guest_irq_type;
424 if (!irqchip_in_kernel(kvm))
427 mutex_lock(&kvm->lock);
429 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
430 assigned_irq->assigned_dev_id);
434 host_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_HOST_MASK);
435 guest_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_GUEST_MASK);
438 /* can only assign one type at a time */
439 if (hweight_long(host_irq_type) > 1)
441 if (hweight_long(guest_irq_type) > 1)
443 if (host_irq_type == 0 && guest_irq_type == 0)
448 r = assign_host_irq(kvm, match, host_irq_type);
453 r = assign_guest_irq(kvm, match, assigned_irq, guest_irq_type);
455 mutex_unlock(&kvm->lock);
459 static int kvm_vm_ioctl_deassign_dev_irq(struct kvm *kvm,
460 struct kvm_assigned_irq
464 struct kvm_assigned_dev_kernel *match;
466 mutex_lock(&kvm->lock);
468 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
469 assigned_irq->assigned_dev_id);
473 r = kvm_deassign_irq(kvm, match, assigned_irq->flags);
475 mutex_unlock(&kvm->lock);
479 static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
480 struct kvm_assigned_pci_dev *assigned_dev)
483 struct kvm_assigned_dev_kernel *match;
486 mutex_lock(&kvm->lock);
487 idx = srcu_read_lock(&kvm->srcu);
489 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
490 assigned_dev->assigned_dev_id);
492 /* device already assigned */
497 match = kzalloc(sizeof(struct kvm_assigned_dev_kernel), GFP_KERNEL);
499 printk(KERN_INFO "%s: Couldn't allocate memory\n",
504 dev = pci_get_domain_bus_and_slot(assigned_dev->segnr,
506 assigned_dev->devfn);
508 printk(KERN_INFO "%s: host device not found\n", __func__);
512 if (pci_enable_device(dev)) {
513 printk(KERN_INFO "%s: Could not enable PCI device\n", __func__);
517 r = pci_request_regions(dev, "kvm_assigned_device");
519 printk(KERN_INFO "%s: Could not get access to device regions\n",
524 pci_reset_function(dev);
526 match->pci_saved_state = pci_store_saved_state(dev);
527 if (!match->pci_saved_state)
528 printk(KERN_DEBUG "%s: Couldn't store %s saved state\n",
529 __func__, dev_name(&dev->dev));
530 match->assigned_dev_id = assigned_dev->assigned_dev_id;
531 match->host_segnr = assigned_dev->segnr;
532 match->host_busnr = assigned_dev->busnr;
533 match->host_devfn = assigned_dev->devfn;
534 match->flags = assigned_dev->flags;
536 spin_lock_init(&match->intx_lock);
537 match->irq_source_id = -1;
539 match->ack_notifier.irq_acked = kvm_assigned_dev_ack_irq;
541 list_add(&match->list, &kvm->arch.assigned_dev_head);
543 if (assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU) {
544 if (!kvm->arch.iommu_domain) {
545 r = kvm_iommu_map_guest(kvm);
549 r = kvm_assign_device(kvm, match);
555 srcu_read_unlock(&kvm->srcu, idx);
556 mutex_unlock(&kvm->lock);
559 if (pci_load_and_free_saved_state(dev, &match->pci_saved_state))
560 printk(KERN_INFO "%s: Couldn't reload %s saved state\n",
561 __func__, dev_name(&dev->dev));
562 list_del(&match->list);
563 pci_release_regions(dev);
565 pci_disable_device(dev);
570 srcu_read_unlock(&kvm->srcu, idx);
571 mutex_unlock(&kvm->lock);
575 static int kvm_vm_ioctl_deassign_device(struct kvm *kvm,
576 struct kvm_assigned_pci_dev *assigned_dev)
579 struct kvm_assigned_dev_kernel *match;
581 mutex_lock(&kvm->lock);
583 match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
584 assigned_dev->assigned_dev_id);
586 printk(KERN_INFO "%s: device hasn't been assigned before, "
587 "so cannot be deassigned\n", __func__);
592 if (match->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU)
593 kvm_deassign_device(kvm, match);
595 kvm_free_assigned_device(kvm, match);
598 mutex_unlock(&kvm->lock);
603 #ifdef __KVM_HAVE_MSIX
604 static int kvm_vm_ioctl_set_msix_nr(struct kvm *kvm,
605 struct kvm_assigned_msix_nr *entry_nr)
608 struct kvm_assigned_dev_kernel *adev;
610 mutex_lock(&kvm->lock);
612 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
613 entry_nr->assigned_dev_id);
619 if (adev->entries_nr == 0) {
620 adev->entries_nr = entry_nr->entry_nr;
621 if (adev->entries_nr == 0 ||
622 adev->entries_nr > KVM_MAX_MSIX_PER_DEV) {
627 adev->host_msix_entries = kzalloc(sizeof(struct msix_entry) *
630 if (!adev->host_msix_entries) {
634 adev->guest_msix_entries =
635 kzalloc(sizeof(struct msix_entry) * entry_nr->entry_nr,
637 if (!adev->guest_msix_entries) {
638 kfree(adev->host_msix_entries);
642 } else /* Not allowed set MSI-X number twice */
645 mutex_unlock(&kvm->lock);
649 static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
650 struct kvm_assigned_msix_entry *entry)
653 struct kvm_assigned_dev_kernel *adev;
655 mutex_lock(&kvm->lock);
657 adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
658 entry->assigned_dev_id);
665 for (i = 0; i < adev->entries_nr; i++)
666 if (adev->guest_msix_entries[i].vector == 0 ||
667 adev->guest_msix_entries[i].entry == entry->entry) {
668 adev->guest_msix_entries[i].entry = entry->entry;
669 adev->guest_msix_entries[i].vector = entry->gsi;
670 adev->host_msix_entries[i].entry = entry->entry;
673 if (i == adev->entries_nr) {
679 mutex_unlock(&kvm->lock);
685 long kvm_vm_ioctl_assigned_device(struct kvm *kvm, unsigned ioctl,
688 void __user *argp = (void __user *)arg;
692 case KVM_ASSIGN_PCI_DEVICE: {
693 struct kvm_assigned_pci_dev assigned_dev;
696 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
698 r = kvm_vm_ioctl_assign_device(kvm, &assigned_dev);
703 case KVM_ASSIGN_IRQ: {
707 case KVM_ASSIGN_DEV_IRQ: {
708 struct kvm_assigned_irq assigned_irq;
711 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
713 r = kvm_vm_ioctl_assign_irq(kvm, &assigned_irq);
718 case KVM_DEASSIGN_DEV_IRQ: {
719 struct kvm_assigned_irq assigned_irq;
722 if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
724 r = kvm_vm_ioctl_deassign_dev_irq(kvm, &assigned_irq);
729 case KVM_DEASSIGN_PCI_DEVICE: {
730 struct kvm_assigned_pci_dev assigned_dev;
733 if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
735 r = kvm_vm_ioctl_deassign_device(kvm, &assigned_dev);
740 #ifdef KVM_CAP_IRQ_ROUTING
741 case KVM_SET_GSI_ROUTING: {
742 struct kvm_irq_routing routing;
743 struct kvm_irq_routing __user *urouting;
744 struct kvm_irq_routing_entry *entries;
747 if (copy_from_user(&routing, argp, sizeof(routing)))
750 if (routing.nr >= KVM_MAX_IRQ_ROUTES)
755 entries = vmalloc(routing.nr * sizeof(*entries));
760 if (copy_from_user(entries, urouting->entries,
761 routing.nr * sizeof(*entries)))
762 goto out_free_irq_routing;
763 r = kvm_set_irq_routing(kvm, entries, routing.nr,
765 out_free_irq_routing:
769 #endif /* KVM_CAP_IRQ_ROUTING */
770 #ifdef __KVM_HAVE_MSIX
771 case KVM_ASSIGN_SET_MSIX_NR: {
772 struct kvm_assigned_msix_nr entry_nr;
774 if (copy_from_user(&entry_nr, argp, sizeof entry_nr))
776 r = kvm_vm_ioctl_set_msix_nr(kvm, &entry_nr);
781 case KVM_ASSIGN_SET_MSIX_ENTRY: {
782 struct kvm_assigned_msix_entry entry;
784 if (copy_from_user(&entry, argp, sizeof entry))
786 r = kvm_vm_ioctl_set_msix_entry(kvm, &entry);