KVM: Device Assignment with VT-d
authorBen-Ami Yassour <benami@il.ibm.com>
Sun, 14 Sep 2008 00:48:28 +0000 (03:48 +0300)
committerAvi Kivity <avi@redhat.com>
Wed, 15 Oct 2008 12:25:04 +0000 (14:25 +0200)
Based on a patch by: Kay, Allen M <allen.m.kay@intel.com>

This patch enables PCI device assignment based on VT-d support.
When a device is assigned to the guest, the guest memory is pinned and
the mapping is updated in the VT-d IOMMU.

[Amit: Expose KVM_CAP_IOMMU so we can check if an IOMMU is present
and also control enable/disable from userspace]

Signed-off-by: Kay, Allen M <allen.m.kay@intel.com>
Signed-off-by: Weidong Han <weidong.han@intel.com>
Signed-off-by: Ben-Ami Yassour <benami@il.ibm.com>
Signed-off-by: Amit Shah <amit.shah@qumranet.com>
Acked-by: Mark Gross <mgross@linux.intel.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
arch/x86/kvm/Makefile
arch/x86/kvm/vtd.c [new file with mode: 0644]
arch/x86/kvm/x86.c
include/asm-x86/kvm_host.h
include/linux/kvm.h
include/linux/kvm_host.h
virt/kvm/kvm_main.c

index d0e940b..3072b17 100644 (file)
@@ -12,6 +12,9 @@ EXTRA_CFLAGS += -Ivirt/kvm -Iarch/x86/kvm
 
 kvm-objs := $(common-objs) x86.o mmu.o x86_emulate.o i8259.o irq.o lapic.o \
        i8254.o
+ifeq ($(CONFIG_DMAR),y)
+kvm-objs += vtd.o
+endif
 obj-$(CONFIG_KVM) += kvm.o
 kvm-intel-objs = vmx.o
 obj-$(CONFIG_KVM_INTEL) += kvm-intel.o
diff --git a/arch/x86/kvm/vtd.c b/arch/x86/kvm/vtd.c
new file mode 100644 (file)
index 0000000..667bf3f
--- /dev/null
@@ -0,0 +1,198 @@
+/*
+ * Copyright (c) 2006, Intel Corporation.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ *
+ * You should have received a copy of the GNU General Public License along with
+ * this program; if not, write to the Free Software Foundation, Inc., 59 Temple
+ * Place - Suite 330, Boston, MA 02111-1307 USA.
+ *
+ * Copyright (C) 2006-2008 Intel Corporation
+ * Copyright IBM Corporation, 2008
+ * Author: Allen M. Kay <allen.m.kay@intel.com>
+ * Author: Weidong Han <weidong.han@intel.com>
+ * Author: Ben-Ami Yassour <benami@il.ibm.com>
+ */
+
+#include <linux/list.h>
+#include <linux/kvm_host.h>
+#include <linux/pci.h>
+#include <linux/dmar.h>
+#include <linux/intel-iommu.h>
+
+static int kvm_iommu_unmap_memslots(struct kvm *kvm);
+static void kvm_iommu_put_pages(struct kvm *kvm,
+                               gfn_t base_gfn, unsigned long npages);
+
+int kvm_iommu_map_pages(struct kvm *kvm,
+                       gfn_t base_gfn, unsigned long npages)
+{
+       gfn_t gfn = base_gfn;
+       pfn_t pfn;
+       int i, r;
+       struct dmar_domain *domain = kvm->arch.intel_iommu_domain;
+
+       /* check if iommu exists and in use */
+       if (!domain)
+               return 0;
+
+       r = -EINVAL;
+       for (i = 0; i < npages; i++) {
+               /* check if already mapped */
+               pfn = (pfn_t)intel_iommu_iova_to_pfn(domain,
+                                                    gfn_to_gpa(gfn));
+               if (pfn && !is_mmio_pfn(pfn))
+                       continue;
+
+               pfn = gfn_to_pfn(kvm, gfn);
+               if (!is_mmio_pfn(pfn)) {
+                       r = intel_iommu_page_mapping(domain,
+                                                    gfn_to_gpa(gfn),
+                                                    pfn_to_hpa(pfn),
+                                                    PAGE_SIZE,
+                                                    DMA_PTE_READ |
+                                                    DMA_PTE_WRITE);
+                       if (r) {
+                               printk(KERN_DEBUG "kvm_iommu_map_pages:"
+                                      "iommu failed to map pfn=%lx\n", pfn);
+                               goto unmap_pages;
+                       }
+               } else {
+                       printk(KERN_DEBUG "kvm_iommu_map_page:"
+                              "invalid pfn=%lx\n", pfn);
+                       goto unmap_pages;
+               }
+               gfn++;
+       }
+       return 0;
+
+unmap_pages:
+       kvm_iommu_put_pages(kvm, base_gfn, i);
+       return r;
+}
+
+static int kvm_iommu_map_memslots(struct kvm *kvm)
+{
+       int i, r;
+
+       down_read(&kvm->slots_lock);
+       for (i = 0; i < kvm->nmemslots; i++) {
+               r = kvm_iommu_map_pages(kvm, kvm->memslots[i].base_gfn,
+                                       kvm->memslots[i].npages);
+               if (r)
+                       break;
+       }
+       up_read(&kvm->slots_lock);
+       return r;
+}
+
+int kvm_iommu_map_guest(struct kvm *kvm,
+                       struct kvm_assigned_dev_kernel *assigned_dev)
+{
+       struct pci_dev *pdev = NULL;
+       int r;
+
+       if (!intel_iommu_found()) {
+               printk(KERN_ERR "%s: intel iommu not found\n", __func__);
+               return -ENODEV;
+       }
+
+       printk(KERN_DEBUG "VT-d direct map: host bdf = %x:%x:%x\n",
+              assigned_dev->host_busnr,
+              PCI_SLOT(assigned_dev->host_devfn),
+              PCI_FUNC(assigned_dev->host_devfn));
+
+       pdev = assigned_dev->dev;
+
+       if (pdev == NULL) {
+               if (kvm->arch.intel_iommu_domain) {
+                       intel_iommu_domain_exit(kvm->arch.intel_iommu_domain);
+                       kvm->arch.intel_iommu_domain = NULL;
+               }
+               return -ENODEV;
+       }
+
+       kvm->arch.intel_iommu_domain = intel_iommu_domain_alloc(pdev);
+       if (!kvm->arch.intel_iommu_domain)
+               return -ENODEV;
+
+       r = kvm_iommu_map_memslots(kvm);
+       if (r)
+               goto out_unmap;
+
+       intel_iommu_detach_dev(kvm->arch.intel_iommu_domain,
+                              pdev->bus->number, pdev->devfn);
+
+       r = intel_iommu_context_mapping(kvm->arch.intel_iommu_domain,
+                                       pdev);
+       if (r) {
+               printk(KERN_ERR "Domain context map for %s failed",
+                      pci_name(pdev));
+               goto out_unmap;
+       }
+       return 0;
+
+out_unmap:
+       kvm_iommu_unmap_memslots(kvm);
+       return r;
+}
+
+static void kvm_iommu_put_pages(struct kvm *kvm,
+                              gfn_t base_gfn, unsigned long npages)
+{
+       gfn_t gfn = base_gfn;
+       pfn_t pfn;
+       struct dmar_domain *domain = kvm->arch.intel_iommu_domain;
+       int i;
+
+       for (i = 0; i < npages; i++) {
+               pfn = (pfn_t)intel_iommu_iova_to_pfn(domain,
+                                                    gfn_to_gpa(gfn));
+               kvm_release_pfn_clean(pfn);
+               gfn++;
+       }
+}
+
+static int kvm_iommu_unmap_memslots(struct kvm *kvm)
+{
+       int i;
+       down_read(&kvm->slots_lock);
+       for (i = 0; i < kvm->nmemslots; i++) {
+               kvm_iommu_put_pages(kvm, kvm->memslots[i].base_gfn,
+                                   kvm->memslots[i].npages);
+       }
+       up_read(&kvm->slots_lock);
+
+       return 0;
+}
+
+int kvm_iommu_unmap_guest(struct kvm *kvm)
+{
+       struct kvm_assigned_dev_kernel *entry;
+       struct dmar_domain *domain = kvm->arch.intel_iommu_domain;
+
+       /* check if iommu exists and in use */
+       if (!domain)
+               return 0;
+
+       list_for_each_entry(entry, &kvm->arch.assigned_dev_head, list) {
+               printk(KERN_DEBUG "VT-d unmap: host bdf = %x:%x:%x\n",
+                      entry->host_busnr,
+                      PCI_SLOT(entry->host_devfn),
+                      PCI_FUNC(entry->host_devfn));
+
+               /* detach kvm dmar domain */
+               intel_iommu_detach_dev(domain, entry->host_busnr,
+                                      entry->host_devfn);
+       }
+       kvm_iommu_unmap_memslots(kvm);
+       intel_iommu_domain_exit(domain);
+       return 0;
+}
index 2134f3e..c8a2793 100644 (file)
@@ -35,6 +35,7 @@
 #include <linux/module.h>
 #include <linux/mman.h>
 #include <linux/highmem.h>
+#include <linux/intel-iommu.h>
 
 #include <asm/uaccess.h>
 #include <asm/msr.h>
@@ -277,9 +278,18 @@ static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
 
        list_add(&match->list, &kvm->arch.assigned_dev_head);
 
+       if (assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU) {
+               r = kvm_iommu_map_guest(kvm, match);
+               if (r)
+                       goto out_list_del;
+       }
+
 out:
        mutex_unlock(&kvm->lock);
        return r;
+out_list_del:
+       list_del(&match->list);
+       pci_release_regions(dev);
 out_disable:
        pci_disable_device(dev);
 out_put:
@@ -1147,6 +1157,9 @@ int kvm_dev_ioctl_check_extension(long ext)
        case KVM_CAP_PV_MMU:
                r = !tdp_enabled;
                break;
+       case KVM_CAP_IOMMU:
+               r = intel_iommu_found();
+               break;
        default:
                r = 0;
                break;
@@ -4282,6 +4295,7 @@ static void kvm_free_vcpus(struct kvm *kvm)
 
 void kvm_arch_destroy_vm(struct kvm *kvm)
 {
+       kvm_iommu_unmap_guest(kvm);
        kvm_free_assigned_devices(kvm);
        kvm_free_pit(kvm);
        kfree(kvm->arch.vpic);
index 68a3ac1..805629c 100644 (file)
@@ -331,26 +331,6 @@ struct kvm_mem_alias {
        gfn_t target_gfn;
 };
 
-struct kvm_irq_ack_notifier {
-       struct hlist_node link;
-       unsigned gsi;
-       void (*irq_acked)(struct kvm_irq_ack_notifier *kian);
-};
-
-struct kvm_assigned_dev_kernel {
-       struct kvm_irq_ack_notifier ack_notifier;
-       struct work_struct interrupt_work;
-       struct list_head list;
-       int assigned_dev_id;
-       int host_busnr;
-       int host_devfn;
-       int host_irq;
-       int guest_irq;
-       int irq_requested;
-       struct pci_dev *dev;
-       struct kvm *kvm;
-};
-
 struct kvm_arch{
        int naliases;
        struct kvm_mem_alias aliases[KVM_ALIAS_SLOTS];
@@ -364,6 +344,7 @@ struct kvm_arch{
         */
        struct list_head active_mmu_pages;
        struct list_head assigned_dev_head;
+       struct dmar_domain *intel_iommu_domain;
        struct kvm_pic *vpic;
        struct kvm_ioapic *vioapic;
        struct kvm_pit *vpit;
@@ -514,6 +495,8 @@ int emulator_write_phys(struct kvm_vcpu *vcpu, gpa_t gpa,
 int kvm_pv_mmu_op(struct kvm_vcpu *vcpu, unsigned long bytes,
                  gpa_t addr, unsigned long *ret);
 
+int is_mmio_pfn(pfn_t pfn);
+
 extern bool tdp_enabled;
 
 enum emulation_result {
index ef4bc6f..4269be1 100644 (file)
@@ -384,6 +384,7 @@ struct kvm_trace_rec {
 #define KVM_CAP_COALESCED_MMIO 15
 #define KVM_CAP_SYNC_MMU 16  /* Changes to host mmap are reflected in guest */
 #define KVM_CAP_DEVICE_ASSIGNMENT 17
+#define KVM_CAP_IOMMU 18
 
 /*
  * ioctls for VM fds
@@ -495,4 +496,6 @@ struct kvm_assigned_irq {
        __u32 flags;
 };
 
+#define KVM_DEV_ASSIGN_ENABLE_IOMMU    (1 << 0)
+
 #endif
index 4b03643..6252802 100644 (file)
@@ -286,6 +286,53 @@ int kvm_cpu_has_interrupt(struct kvm_vcpu *v);
 int kvm_cpu_has_pending_timer(struct kvm_vcpu *vcpu);
 void kvm_vcpu_kick(struct kvm_vcpu *vcpu);
 
+struct kvm_irq_ack_notifier {
+       struct hlist_node link;
+       unsigned gsi;
+       void (*irq_acked)(struct kvm_irq_ack_notifier *kian);
+};
+
+struct kvm_assigned_dev_kernel {
+       struct kvm_irq_ack_notifier ack_notifier;
+       struct work_struct interrupt_work;
+       struct list_head list;
+       int assigned_dev_id;
+       int host_busnr;
+       int host_devfn;
+       int host_irq;
+       int guest_irq;
+       int irq_requested;
+       struct pci_dev *dev;
+       struct kvm *kvm;
+};
+
+#ifdef CONFIG_DMAR
+int kvm_iommu_map_pages(struct kvm *kvm, gfn_t base_gfn,
+                       unsigned long npages);
+int kvm_iommu_map_guest(struct kvm *kvm,
+                       struct kvm_assigned_dev_kernel *assigned_dev);
+int kvm_iommu_unmap_guest(struct kvm *kvm);
+#else /* CONFIG_DMAR */
+static inline int kvm_iommu_map_pages(struct kvm *kvm,
+                                     gfn_t base_gfn,
+                                     unsigned long npages)
+{
+       return 0;
+}
+
+static inline int kvm_iommu_map_guest(struct kvm *kvm,
+                                     struct kvm_assigned_dev_kernel
+                                     *assigned_dev)
+{
+       return -ENODEV;
+}
+
+static inline int kvm_iommu_unmap_guest(struct kvm *kvm)
+{
+       return 0;
+}
+#endif /* CONFIG_DMAR */
+
 static inline void kvm_guest_enter(void)
 {
        account_system_vtime(current);
@@ -308,6 +355,11 @@ static inline gpa_t gfn_to_gpa(gfn_t gfn)
        return (gpa_t)gfn << PAGE_SHIFT;
 }
 
+static inline hpa_t pfn_to_hpa(pfn_t pfn)
+{
+       return (hpa_t)pfn << PAGE_SHIFT;
+}
+
 static inline void kvm_migrate_timers(struct kvm_vcpu *vcpu)
 {
        set_bit(KVM_REQ_MIGRATE_TIMER, &vcpu->requests);
index 63e661b..f42d5c2 100644 (file)
@@ -41,6 +41,7 @@
 #include <linux/pagemap.h>
 #include <linux/mman.h>
 #include <linux/swap.h>
+#include <linux/intel-iommu.h>
 
 #include <asm/processor.h>
 #include <asm/io.h>
@@ -76,7 +77,7 @@ static inline int valid_vcpu(int n)
        return likely(n >= 0 && n < KVM_MAX_VCPUS);
 }
 
-static inline int is_mmio_pfn(pfn_t pfn)
+inline int is_mmio_pfn(pfn_t pfn)
 {
        if (pfn_valid(pfn))
                return PageReserved(pfn_to_page(pfn));
@@ -578,6 +579,12 @@ int __kvm_set_memory_region(struct kvm *kvm,
        }
 
        kvm_free_physmem_slot(&old, &new);
+
+       /* map the pages in iommu page table */
+       r = kvm_iommu_map_pages(kvm, base_gfn, npages);
+       if (r)
+               goto out;
+
        return 0;
 
 out_free: