KVM: MMU: delay flush all tlbs on sync_page path
[linux-2.6.git] / virt / kvm / kvm_main.c
index b5e7e3f..ee99b77 100644 (file)
@@ -5,6 +5,7 @@
  * machines without emulation or binary translation.
  *
  * Copyright (C) 2006 Qumranet, Inc.
+ * Copyright 2010 Red Hat, Inc. and/or its affiliates.
  *
  * Authors:
  *   Avi Kivity   <avi@qumranet.com>
@@ -22,7 +23,6 @@
 #include <linux/module.h>
 #include <linux/errno.h>
 #include <linux/percpu.h>
-#include <linux/gfp.h>
 #include <linux/mm.h>
 #include <linux/miscdevice.h>
 #include <linux/vmalloc.h>
 #include <linux/swap.h>
 #include <linux/bitops.h>
 #include <linux/spinlock.h>
+#include <linux/compat.h>
+#include <linux/srcu.h>
+#include <linux/hugetlb.h>
+#include <linux/slab.h>
 
 #include <asm/processor.h>
 #include <asm/io.h>
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
+#include <asm-generic/bitops/le.h>
 
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
 #include "coalesced_mmio.h"
-#endif
-
-#ifdef KVM_CAP_DEVICE_ASSIGNMENT
-#include <linux/pci.h>
-#include <linux/interrupt.h>
-#include "irq.h"
-#endif
+#include "async_pf.h"
 
 #define CREATE_TRACE_POINTS
 #include <trace/events/kvm.h>
@@ -68,13 +66,15 @@ MODULE_LICENSE("GPL");
 /*
  * Ordering of locks:
  *
- *             kvm->slots_lock --> kvm->lock --> kvm->irq_lock
+ *             kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  */
 
 DEFINE_SPINLOCK(kvm_lock);
 LIST_HEAD(vm_list);
 
 static cpumask_var_t cpus_hardware_enabled;
+static int kvm_usage_count = 0;
+static atomic_t hardware_enable_failed;
 
 struct kmem_cache *kvm_vcpu_cache;
 EXPORT_SYMBOL_GPL(kvm_vcpu_cache);
@@ -85,614 +85,20 @@ struct dentry *kvm_debugfs_dir;
 
 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
                           unsigned long arg);
+static int hardware_enable_all(void);
+static void hardware_disable_all(void);
+
+static void kvm_io_bus_destroy(struct kvm_io_bus *bus);
 
 static bool kvm_rebooting;
 
 static bool largepages_enabled = true;
 
-#ifdef KVM_CAP_DEVICE_ASSIGNMENT
-static struct kvm_assigned_dev_kernel *kvm_find_assigned_dev(struct list_head *head,
-                                                     int assigned_dev_id)
-{
-       struct list_head *ptr;
-       struct kvm_assigned_dev_kernel *match;
-
-       list_for_each(ptr, head) {
-               match = list_entry(ptr, struct kvm_assigned_dev_kernel, list);
-               if (match->assigned_dev_id == assigned_dev_id)
-                       return match;
-       }
-       return NULL;
-}
-
-static int find_index_from_host_irq(struct kvm_assigned_dev_kernel
-                                   *assigned_dev, int irq)
-{
-       int i, index;
-       struct msix_entry *host_msix_entries;
-
-       host_msix_entries = assigned_dev->host_msix_entries;
-
-       index = -1;
-       for (i = 0; i < assigned_dev->entries_nr; i++)
-               if (irq == host_msix_entries[i].vector) {
-                       index = i;
-                       break;
-               }
-       if (index < 0) {
-               printk(KERN_WARNING "Fail to find correlated MSI-X entry!\n");
-               return 0;
-       }
-
-       return index;
-}
-
-static void kvm_assigned_dev_interrupt_work_handler(struct work_struct *work)
-{
-       struct kvm_assigned_dev_kernel *assigned_dev;
-       struct kvm *kvm;
-       int i;
-
-       assigned_dev = container_of(work, struct kvm_assigned_dev_kernel,
-                                   interrupt_work);
-       kvm = assigned_dev->kvm;
-
-       mutex_lock(&kvm->irq_lock);
-       spin_lock_irq(&assigned_dev->assigned_dev_lock);
-       if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
-               struct kvm_guest_msix_entry *guest_entries =
-                       assigned_dev->guest_msix_entries;
-               for (i = 0; i < assigned_dev->entries_nr; i++) {
-                       if (!(guest_entries[i].flags &
-                                       KVM_ASSIGNED_MSIX_PENDING))
-                               continue;
-                       guest_entries[i].flags &= ~KVM_ASSIGNED_MSIX_PENDING;
-                       kvm_set_irq(assigned_dev->kvm,
-                                   assigned_dev->irq_source_id,
-                                   guest_entries[i].vector, 1);
-               }
-       } else
-               kvm_set_irq(assigned_dev->kvm, assigned_dev->irq_source_id,
-                           assigned_dev->guest_irq, 1);
-
-       spin_unlock_irq(&assigned_dev->assigned_dev_lock);
-       mutex_unlock(&assigned_dev->kvm->irq_lock);
-}
-
-static irqreturn_t kvm_assigned_dev_intr(int irq, void *dev_id)
-{
-       unsigned long flags;
-       struct kvm_assigned_dev_kernel *assigned_dev =
-               (struct kvm_assigned_dev_kernel *) dev_id;
-
-       spin_lock_irqsave(&assigned_dev->assigned_dev_lock, flags);
-       if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
-               int index = find_index_from_host_irq(assigned_dev, irq);
-               if (index < 0)
-                       goto out;
-               assigned_dev->guest_msix_entries[index].flags |=
-                       KVM_ASSIGNED_MSIX_PENDING;
-       }
-
-       schedule_work(&assigned_dev->interrupt_work);
-
-       if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_GUEST_INTX) {
-               disable_irq_nosync(irq);
-               assigned_dev->host_irq_disabled = true;
-       }
-
-out:
-       spin_unlock_irqrestore(&assigned_dev->assigned_dev_lock, flags);
-       return IRQ_HANDLED;
-}
-
-/* Ack the irq line for an assigned device */
-static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
-{
-       struct kvm_assigned_dev_kernel *dev;
-       unsigned long flags;
-
-       if (kian->gsi == -1)
-               return;
-
-       dev = container_of(kian, struct kvm_assigned_dev_kernel,
-                          ack_notifier);
-
-       kvm_set_irq(dev->kvm, dev->irq_source_id, dev->guest_irq, 0);
-
-       /* The guest irq may be shared so this ack may be
-        * from another device.
-        */
-       spin_lock_irqsave(&dev->assigned_dev_lock, flags);
-       if (dev->host_irq_disabled) {
-               enable_irq(dev->host_irq);
-               dev->host_irq_disabled = false;
-       }
-       spin_unlock_irqrestore(&dev->assigned_dev_lock, flags);
-}
-
-static void deassign_guest_irq(struct kvm *kvm,
-                              struct kvm_assigned_dev_kernel *assigned_dev)
-{
-       kvm_unregister_irq_ack_notifier(kvm, &assigned_dev->ack_notifier);
-       assigned_dev->ack_notifier.gsi = -1;
-
-       if (assigned_dev->irq_source_id != -1)
-               kvm_free_irq_source_id(kvm, assigned_dev->irq_source_id);
-       assigned_dev->irq_source_id = -1;
-       assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_GUEST_MASK);
-}
-
-/* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
-static void deassign_host_irq(struct kvm *kvm,
-                             struct kvm_assigned_dev_kernel *assigned_dev)
-{
-       /*
-        * In kvm_free_device_irq, cancel_work_sync return true if:
-        * 1. work is scheduled, and then cancelled.
-        * 2. work callback is executed.
-        *
-        * The first one ensured that the irq is disabled and no more events
-        * would happen. But for the second one, the irq may be enabled (e.g.
-        * for MSI). So we disable irq here to prevent further events.
-        *
-        * Notice this maybe result in nested disable if the interrupt type is
-        * INTx, but it's OK for we are going to free it.
-        *
-        * If this function is a part of VM destroy, please ensure that till
-        * now, the kvm state is still legal for probably we also have to wait
-        * interrupt_work done.
-        */
-       if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSIX) {
-               int i;
-               for (i = 0; i < assigned_dev->entries_nr; i++)
-                       disable_irq_nosync(assigned_dev->
-                                          host_msix_entries[i].vector);
-
-               cancel_work_sync(&assigned_dev->interrupt_work);
-
-               for (i = 0; i < assigned_dev->entries_nr; i++)
-                       free_irq(assigned_dev->host_msix_entries[i].vector,
-                                (void *)assigned_dev);
-
-               assigned_dev->entries_nr = 0;
-               kfree(assigned_dev->host_msix_entries);
-               kfree(assigned_dev->guest_msix_entries);
-               pci_disable_msix(assigned_dev->dev);
-       } else {
-               /* Deal with MSI and INTx */
-               disable_irq_nosync(assigned_dev->host_irq);
-               cancel_work_sync(&assigned_dev->interrupt_work);
-
-               free_irq(assigned_dev->host_irq, (void *)assigned_dev);
-
-               if (assigned_dev->irq_requested_type & KVM_DEV_IRQ_HOST_MSI)
-                       pci_disable_msi(assigned_dev->dev);
-       }
-
-       assigned_dev->irq_requested_type &= ~(KVM_DEV_IRQ_HOST_MASK);
-}
-
-static int kvm_deassign_irq(struct kvm *kvm,
-                           struct kvm_assigned_dev_kernel *assigned_dev,
-                           unsigned long irq_requested_type)
-{
-       unsigned long guest_irq_type, host_irq_type;
-
-       if (!irqchip_in_kernel(kvm))
-               return -EINVAL;
-       /* no irq assignment to deassign */
-       if (!assigned_dev->irq_requested_type)
-               return -ENXIO;
-
-       host_irq_type = irq_requested_type & KVM_DEV_IRQ_HOST_MASK;
-       guest_irq_type = irq_requested_type & KVM_DEV_IRQ_GUEST_MASK;
-
-       if (host_irq_type)
-               deassign_host_irq(kvm, assigned_dev);
-       if (guest_irq_type)
-               deassign_guest_irq(kvm, assigned_dev);
-
-       return 0;
-}
-
-static void kvm_free_assigned_irq(struct kvm *kvm,
-                                 struct kvm_assigned_dev_kernel *assigned_dev)
-{
-       kvm_deassign_irq(kvm, assigned_dev, assigned_dev->irq_requested_type);
-}
-
-static void kvm_free_assigned_device(struct kvm *kvm,
-                                    struct kvm_assigned_dev_kernel
-                                    *assigned_dev)
-{
-       kvm_free_assigned_irq(kvm, assigned_dev);
-
-       pci_reset_function(assigned_dev->dev);
-
-       pci_release_regions(assigned_dev->dev);
-       pci_disable_device(assigned_dev->dev);
-       pci_dev_put(assigned_dev->dev);
-
-       list_del(&assigned_dev->list);
-       kfree(assigned_dev);
-}
-
-void kvm_free_all_assigned_devices(struct kvm *kvm)
-{
-       struct list_head *ptr, *ptr2;
-       struct kvm_assigned_dev_kernel *assigned_dev;
-
-       list_for_each_safe(ptr, ptr2, &kvm->arch.assigned_dev_head) {
-               assigned_dev = list_entry(ptr,
-                                         struct kvm_assigned_dev_kernel,
-                                         list);
-
-               kvm_free_assigned_device(kvm, assigned_dev);
-       }
-}
-
-static int assigned_device_enable_host_intx(struct kvm *kvm,
-                                           struct kvm_assigned_dev_kernel *dev)
-{
-       dev->host_irq = dev->dev->irq;
-       /* Even though this is PCI, we don't want to use shared
-        * interrupts. Sharing host devices with guest-assigned devices
-        * on the same interrupt line is not a happy situation: there
-        * are going to be long delays in accepting, acking, etc.
-        */
-       if (request_irq(dev->host_irq, kvm_assigned_dev_intr,
-                       0, "kvm_assigned_intx_device", (void *)dev))
-               return -EIO;
-       return 0;
-}
-
-#ifdef __KVM_HAVE_MSI
-static int assigned_device_enable_host_msi(struct kvm *kvm,
-                                          struct kvm_assigned_dev_kernel *dev)
-{
-       int r;
-
-       if (!dev->dev->msi_enabled) {
-               r = pci_enable_msi(dev->dev);
-               if (r)
-                       return r;
-       }
-
-       dev->host_irq = dev->dev->irq;
-       if (request_irq(dev->host_irq, kvm_assigned_dev_intr, 0,
-                       "kvm_assigned_msi_device", (void *)dev)) {
-               pci_disable_msi(dev->dev);
-               return -EIO;
-       }
-
-       return 0;
-}
-#endif
-
-#ifdef __KVM_HAVE_MSIX
-static int assigned_device_enable_host_msix(struct kvm *kvm,
-                                           struct kvm_assigned_dev_kernel *dev)
-{
-       int i, r = -EINVAL;
-
-       /* host_msix_entries and guest_msix_entries should have been
-        * initialized */
-       if (dev->entries_nr == 0)
-               return r;
-
-       r = pci_enable_msix(dev->dev, dev->host_msix_entries, dev->entries_nr);
-       if (r)
-               return r;
-
-       for (i = 0; i < dev->entries_nr; i++) {
-               r = request_irq(dev->host_msix_entries[i].vector,
-                               kvm_assigned_dev_intr, 0,
-                               "kvm_assigned_msix_device",
-                               (void *)dev);
-               /* FIXME: free requested_irq's on failure */
-               if (r)
-                       return r;
-       }
-
-       return 0;
-}
-
-#endif
-
-static int assigned_device_enable_guest_intx(struct kvm *kvm,
-                               struct kvm_assigned_dev_kernel *dev,
-                               struct kvm_assigned_irq *irq)
-{
-       dev->guest_irq = irq->guest_irq;
-       dev->ack_notifier.gsi = irq->guest_irq;
-       return 0;
-}
-
-#ifdef __KVM_HAVE_MSI
-static int assigned_device_enable_guest_msi(struct kvm *kvm,
-                       struct kvm_assigned_dev_kernel *dev,
-                       struct kvm_assigned_irq *irq)
-{
-       dev->guest_irq = irq->guest_irq;
-       dev->ack_notifier.gsi = -1;
-       dev->host_irq_disabled = false;
-       return 0;
-}
-#endif
-#ifdef __KVM_HAVE_MSIX
-static int assigned_device_enable_guest_msix(struct kvm *kvm,
-                       struct kvm_assigned_dev_kernel *dev,
-                       struct kvm_assigned_irq *irq)
-{
-       dev->guest_irq = irq->guest_irq;
-       dev->ack_notifier.gsi = -1;
-       dev->host_irq_disabled = false;
-       return 0;
-}
-#endif
-
-static int assign_host_irq(struct kvm *kvm,
-                          struct kvm_assigned_dev_kernel *dev,
-                          __u32 host_irq_type)
-{
-       int r = -EEXIST;
-
-       if (dev->irq_requested_type & KVM_DEV_IRQ_HOST_MASK)
-               return r;
-
-       switch (host_irq_type) {
-       case KVM_DEV_IRQ_HOST_INTX:
-               r = assigned_device_enable_host_intx(kvm, dev);
-               break;
-#ifdef __KVM_HAVE_MSI
-       case KVM_DEV_IRQ_HOST_MSI:
-               r = assigned_device_enable_host_msi(kvm, dev);
-               break;
-#endif
-#ifdef __KVM_HAVE_MSIX
-       case KVM_DEV_IRQ_HOST_MSIX:
-               r = assigned_device_enable_host_msix(kvm, dev);
-               break;
-#endif
-       default:
-               r = -EINVAL;
-       }
-
-       if (!r)
-               dev->irq_requested_type |= host_irq_type;
-
-       return r;
-}
-
-static int assign_guest_irq(struct kvm *kvm,
-                           struct kvm_assigned_dev_kernel *dev,
-                           struct kvm_assigned_irq *irq,
-                           unsigned long guest_irq_type)
-{
-       int id;
-       int r = -EEXIST;
-
-       if (dev->irq_requested_type & KVM_DEV_IRQ_GUEST_MASK)
-               return r;
-
-       id = kvm_request_irq_source_id(kvm);
-       if (id < 0)
-               return id;
-
-       dev->irq_source_id = id;
-
-       switch (guest_irq_type) {
-       case KVM_DEV_IRQ_GUEST_INTX:
-               r = assigned_device_enable_guest_intx(kvm, dev, irq);
-               break;
-#ifdef __KVM_HAVE_MSI
-       case KVM_DEV_IRQ_GUEST_MSI:
-               r = assigned_device_enable_guest_msi(kvm, dev, irq);
-               break;
-#endif
-#ifdef __KVM_HAVE_MSIX
-       case KVM_DEV_IRQ_GUEST_MSIX:
-               r = assigned_device_enable_guest_msix(kvm, dev, irq);
-               break;
-#endif
-       default:
-               r = -EINVAL;
-       }
-
-       if (!r) {
-               dev->irq_requested_type |= guest_irq_type;
-               kvm_register_irq_ack_notifier(kvm, &dev->ack_notifier);
-       } else
-               kvm_free_irq_source_id(kvm, dev->irq_source_id);
-
-       return r;
-}
-
-/* TODO Deal with KVM_DEV_IRQ_ASSIGNED_MASK_MSIX */
-static int kvm_vm_ioctl_assign_irq(struct kvm *kvm,
-                                  struct kvm_assigned_irq *assigned_irq)
-{
-       int r = -EINVAL;
-       struct kvm_assigned_dev_kernel *match;
-       unsigned long host_irq_type, guest_irq_type;
-
-       if (!capable(CAP_SYS_RAWIO))
-               return -EPERM;
-
-       if (!irqchip_in_kernel(kvm))
-               return r;
-
-       mutex_lock(&kvm->lock);
-       r = -ENODEV;
-       match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     assigned_irq->assigned_dev_id);
-       if (!match)
-               goto out;
-
-       host_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_HOST_MASK);
-       guest_irq_type = (assigned_irq->flags & KVM_DEV_IRQ_GUEST_MASK);
-
-       r = -EINVAL;
-       /* can only assign one type at a time */
-       if (hweight_long(host_irq_type) > 1)
-               goto out;
-       if (hweight_long(guest_irq_type) > 1)
-               goto out;
-       if (host_irq_type == 0 && guest_irq_type == 0)
-               goto out;
-
-       r = 0;
-       if (host_irq_type)
-               r = assign_host_irq(kvm, match, host_irq_type);
-       if (r)
-               goto out;
-
-       if (guest_irq_type)
-               r = assign_guest_irq(kvm, match, assigned_irq, guest_irq_type);
-out:
-       mutex_unlock(&kvm->lock);
-       return r;
-}
-
-static int kvm_vm_ioctl_deassign_dev_irq(struct kvm *kvm,
-                                        struct kvm_assigned_irq
-                                        *assigned_irq)
-{
-       int r = -ENODEV;
-       struct kvm_assigned_dev_kernel *match;
-
-       mutex_lock(&kvm->lock);
-
-       match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     assigned_irq->assigned_dev_id);
-       if (!match)
-               goto out;
-
-       r = kvm_deassign_irq(kvm, match, assigned_irq->flags);
-out:
-       mutex_unlock(&kvm->lock);
-       return r;
-}
-
-static int kvm_vm_ioctl_assign_device(struct kvm *kvm,
-                                     struct kvm_assigned_pci_dev *assigned_dev)
-{
-       int r = 0;
-       struct kvm_assigned_dev_kernel *match;
-       struct pci_dev *dev;
+static struct page *hwpoison_page;
+static pfn_t hwpoison_pfn;
 
-       down_read(&kvm->slots_lock);
-       mutex_lock(&kvm->lock);
-
-       match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     assigned_dev->assigned_dev_id);
-       if (match) {
-               /* device already assigned */
-               r = -EEXIST;
-               goto out;
-       }
-
-       match = kzalloc(sizeof(struct kvm_assigned_dev_kernel), GFP_KERNEL);
-       if (match == NULL) {
-               printk(KERN_INFO "%s: Couldn't allocate memory\n",
-                      __func__);
-               r = -ENOMEM;
-               goto out;
-       }
-       dev = pci_get_bus_and_slot(assigned_dev->busnr,
-                                  assigned_dev->devfn);
-       if (!dev) {
-               printk(KERN_INFO "%s: host device not found\n", __func__);
-               r = -EINVAL;
-               goto out_free;
-       }
-       if (pci_enable_device(dev)) {
-               printk(KERN_INFO "%s: Could not enable PCI device\n", __func__);
-               r = -EBUSY;
-               goto out_put;
-       }
-       r = pci_request_regions(dev, "kvm_assigned_device");
-       if (r) {
-               printk(KERN_INFO "%s: Could not get access to device regions\n",
-                      __func__);
-               goto out_disable;
-       }
-
-       pci_reset_function(dev);
-
-       match->assigned_dev_id = assigned_dev->assigned_dev_id;
-       match->host_busnr = assigned_dev->busnr;
-       match->host_devfn = assigned_dev->devfn;
-       match->flags = assigned_dev->flags;
-       match->dev = dev;
-       spin_lock_init(&match->assigned_dev_lock);
-       match->irq_source_id = -1;
-       match->kvm = kvm;
-       match->ack_notifier.irq_acked = kvm_assigned_dev_ack_irq;
-       INIT_WORK(&match->interrupt_work,
-                 kvm_assigned_dev_interrupt_work_handler);
-
-       list_add(&match->list, &kvm->arch.assigned_dev_head);
-
-       if (assigned_dev->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU) {
-               if (!kvm->arch.iommu_domain) {
-                       r = kvm_iommu_map_guest(kvm);
-                       if (r)
-                               goto out_list_del;
-               }
-               r = kvm_assign_device(kvm, match);
-               if (r)
-                       goto out_list_del;
-       }
-
-out:
-       mutex_unlock(&kvm->lock);
-       up_read(&kvm->slots_lock);
-       return r;
-out_list_del:
-       list_del(&match->list);
-       pci_release_regions(dev);
-out_disable:
-       pci_disable_device(dev);
-out_put:
-       pci_dev_put(dev);
-out_free:
-       kfree(match);
-       mutex_unlock(&kvm->lock);
-       up_read(&kvm->slots_lock);
-       return r;
-}
-#endif
-
-#ifdef KVM_CAP_DEVICE_DEASSIGNMENT
-static int kvm_vm_ioctl_deassign_device(struct kvm *kvm,
-               struct kvm_assigned_pci_dev *assigned_dev)
-{
-       int r = 0;
-       struct kvm_assigned_dev_kernel *match;
-
-       mutex_lock(&kvm->lock);
-
-       match = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     assigned_dev->assigned_dev_id);
-       if (!match) {
-               printk(KERN_INFO "%s: device hasn't been assigned before, "
-                 "so cannot be deassigned\n", __func__);
-               r = -EINVAL;
-               goto out;
-       }
-
-       if (match->flags & KVM_DEV_ASSIGN_ENABLE_IOMMU)
-               kvm_deassign_device(kvm, match);
-
-       kvm_free_assigned_device(kvm, match);
-
-out:
-       mutex_unlock(&kvm->lock);
-       return r;
-}
-#endif
+static struct page *fault_page;
+static pfn_t fault_pfn;
 
 inline int kvm_is_mmio_pfn(pfn_t pfn)
 {
@@ -740,10 +146,10 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 
        zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 
-       spin_lock(&kvm->requests_lock);
+       raw_spin_lock(&kvm->requests_lock);
        me = smp_processor_id();
        kvm_for_each_vcpu(i, vcpu, kvm) {
-               if (test_and_set_bit(req, &vcpu->requests))
+               if (kvm_make_check_request(req, vcpu))
                        continue;
                cpu = vcpu->cpu;
                if (cpus != NULL && cpu != -1 && cpu != me)
@@ -755,15 +161,19 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
                smp_call_function_many(cpus, ack_flush, NULL, 1);
        else
                called = false;
-       spin_unlock(&kvm->requests_lock);
+       raw_spin_unlock(&kvm->requests_lock);
        free_cpumask_var(cpus);
        return called;
 }
 
 void kvm_flush_remote_tlbs(struct kvm *kvm)
 {
+       int dirty_count = kvm->tlbs_dirty;
+
+       smp_mb();
        if (make_all_cpus_request(kvm, KVM_REQ_TLB_FLUSH))
                ++kvm->stat.remote_tlb_flush;
+       cmpxchg(&kvm->tlbs_dirty, dirty_count, 0);
 }
 
 void kvm_reload_remote_mmus(struct kvm *kvm)
@@ -781,6 +191,7 @@ int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
        vcpu->kvm = kvm;
        vcpu->vcpu_id = id;
        init_waitqueue_head(&vcpu->wq);
+       kvm_async_pf_vcpu_init(vcpu);
 
        page = alloc_page(GFP_KERNEL | __GFP_ZERO);
        if (!page) {
@@ -819,7 +230,7 @@ static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
                                             unsigned long address)
 {
        struct kvm *kvm = mmu_notifier_to_kvm(mn);
-       int need_tlb_flush;
+       int need_tlb_flush, idx;
 
        /*
         * When ->invalidate_page runs, the linux pte has been zapped
@@ -839,10 +250,12 @@ static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
         * pte after kvm_unmap_hva returned, without noticing the page
         * is going to be freed.
         */
+       idx = srcu_read_lock(&kvm->srcu);
        spin_lock(&kvm->mmu_lock);
        kvm->mmu_notifier_seq++;
-       need_tlb_flush = kvm_unmap_hva(kvm, address);
+       need_tlb_flush = kvm_unmap_hva(kvm, address) | kvm->tlbs_dirty;
        spin_unlock(&kvm->mmu_lock);
+       srcu_read_unlock(&kvm->srcu, idx);
 
        /* we've to flush the tlb before the pages can be freed */
        if (need_tlb_flush)
@@ -850,14 +263,31 @@ static void kvm_mmu_notifier_invalidate_page(struct mmu_notifier *mn,
 
 }
 
+static void kvm_mmu_notifier_change_pte(struct mmu_notifier *mn,
+                                       struct mm_struct *mm,
+                                       unsigned long address,
+                                       pte_t pte)
+{
+       struct kvm *kvm = mmu_notifier_to_kvm(mn);
+       int idx;
+
+       idx = srcu_read_lock(&kvm->srcu);
+       spin_lock(&kvm->mmu_lock);
+       kvm->mmu_notifier_seq++;
+       kvm_set_spte_hva(kvm, address, pte);
+       spin_unlock(&kvm->mmu_lock);
+       srcu_read_unlock(&kvm->srcu, idx);
+}
+
 static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
                                                    struct mm_struct *mm,
                                                    unsigned long start,
                                                    unsigned long end)
 {
        struct kvm *kvm = mmu_notifier_to_kvm(mn);
-       int need_tlb_flush = 0;
+       int need_tlb_flush = 0, idx;
 
+       idx = srcu_read_lock(&kvm->srcu);
        spin_lock(&kvm->mmu_lock);
        /*
         * The count increase must become visible at unlock time as no
@@ -867,7 +297,9 @@ static void kvm_mmu_notifier_invalidate_range_start(struct mmu_notifier *mn,
        kvm->mmu_notifier_count++;
        for (; start < end; start += PAGE_SIZE)
                need_tlb_flush |= kvm_unmap_hva(kvm, start);
+       need_tlb_flush |= kvm->tlbs_dirty;
        spin_unlock(&kvm->mmu_lock);
+       srcu_read_unlock(&kvm->srcu, idx);
 
        /* we've to flush the tlb before the pages can be freed */
        if (need_tlb_flush)
@@ -905,11 +337,13 @@ static int kvm_mmu_notifier_clear_flush_young(struct mmu_notifier *mn,
                                              unsigned long address)
 {
        struct kvm *kvm = mmu_notifier_to_kvm(mn);
-       int young;
+       int young, idx;
 
+       idx = srcu_read_lock(&kvm->srcu);
        spin_lock(&kvm->mmu_lock);
        young = kvm_age_hva(kvm, address);
        spin_unlock(&kvm->mmu_lock);
+       srcu_read_unlock(&kvm->srcu, idx);
 
        if (young)
                kvm_flush_remote_tlbs(kvm);
@@ -921,7 +355,11 @@ static void kvm_mmu_notifier_release(struct mmu_notifier *mn,
                                     struct mm_struct *mm)
 {
        struct kvm *kvm = mmu_notifier_to_kvm(mn);
+       int idx;
+
+       idx = srcu_read_lock(&kvm->srcu);
        kvm_arch_flush_shadow(kvm);
+       srcu_read_unlock(&kvm->srcu, idx);
 }
 
 static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
@@ -929,68 +367,102 @@ static const struct mmu_notifier_ops kvm_mmu_notifier_ops = {
        .invalidate_range_start = kvm_mmu_notifier_invalidate_range_start,
        .invalidate_range_end   = kvm_mmu_notifier_invalidate_range_end,
        .clear_flush_young      = kvm_mmu_notifier_clear_flush_young,
+       .change_pte             = kvm_mmu_notifier_change_pte,
        .release                = kvm_mmu_notifier_release,
 };
+
+static int kvm_init_mmu_notifier(struct kvm *kvm)
+{
+       kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
+       return mmu_notifier_register(&kvm->mmu_notifier, current->mm);
+}
+
+#else  /* !(CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER) */
+
+static int kvm_init_mmu_notifier(struct kvm *kvm)
+{
+       return 0;
+}
+
 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 
 static struct kvm *kvm_create_vm(void)
 {
-       struct kvm *kvm = kvm_arch_create_vm();
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
-       struct page *page;
-#endif
+       int r, i;
+       struct kvm *kvm = kvm_arch_alloc_vm();
+
+       if (!kvm)
+               return ERR_PTR(-ENOMEM);
+
+       r = kvm_arch_init_vm(kvm);
+       if (r)
+               goto out_err_nodisable;
+
+       r = hardware_enable_all();
+       if (r)
+               goto out_err_nodisable;
 
-       if (IS_ERR(kvm))
-               goto out;
 #ifdef CONFIG_HAVE_KVM_IRQCHIP
-       INIT_LIST_HEAD(&kvm->irq_routing);
        INIT_HLIST_HEAD(&kvm->mask_notifier_list);
+       INIT_HLIST_HEAD(&kvm->irq_ack_notifier_list);
 #endif
 
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
-       page = alloc_page(GFP_KERNEL | __GFP_ZERO);
-       if (!page) {
-               kfree(kvm);
-               return ERR_PTR(-ENOMEM);
-       }
-       kvm->coalesced_mmio_ring =
-                       (struct kvm_coalesced_mmio_ring *)page_address(page);
-#endif
-
-#if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
-       {
-               int err;
-               kvm->mmu_notifier.ops = &kvm_mmu_notifier_ops;
-               err = mmu_notifier_register(&kvm->mmu_notifier, current->mm);
-               if (err) {
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
-                       put_page(page);
-#endif
-                       kfree(kvm);
-                       return ERR_PTR(err);
-               }
-       }
-#endif
+       r = -ENOMEM;
+       kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
+       if (!kvm->memslots)
+               goto out_err_nosrcu;
+       if (init_srcu_struct(&kvm->srcu))
+               goto out_err_nosrcu;
+       for (i = 0; i < KVM_NR_BUSES; i++) {
+               kvm->buses[i] = kzalloc(sizeof(struct kvm_io_bus),
+                                       GFP_KERNEL);
+               if (!kvm->buses[i])
+                       goto out_err;
+       }
+
+       r = kvm_init_mmu_notifier(kvm);
+       if (r)
+               goto out_err;
 
        kvm->mm = current->mm;
        atomic_inc(&kvm->mm->mm_count);
        spin_lock_init(&kvm->mmu_lock);
-       spin_lock_init(&kvm->requests_lock);
-       kvm_io_bus_init(&kvm->pio_bus);
+       raw_spin_lock_init(&kvm->requests_lock);
        kvm_eventfd_init(kvm);
        mutex_init(&kvm->lock);
        mutex_init(&kvm->irq_lock);
-       kvm_io_bus_init(&kvm->mmio_bus);
-       init_rwsem(&kvm->slots_lock);
+       mutex_init(&kvm->slots_lock);
        atomic_set(&kvm->users_count, 1);
        spin_lock(&kvm_lock);
        list_add(&kvm->vm_list, &vm_list);
        spin_unlock(&kvm_lock);
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
-       kvm_coalesced_mmio_init(kvm);
-#endif
-out:
+
        return kvm;
+
+out_err:
+       cleanup_srcu_struct(&kvm->srcu);
+out_err_nosrcu:
+       hardware_disable_all();
+out_err_nodisable:
+       for (i = 0; i < KVM_NR_BUSES; i++)
+               kfree(kvm->buses[i]);
+       kfree(kvm->memslots);
+       kvm_arch_free_vm(kvm);
+       return ERR_PTR(r);
+}
+
+static void kvm_destroy_dirty_bitmap(struct kvm_memory_slot *memslot)
+{
+       if (!memslot->dirty_bitmap)
+               return;
+
+       if (2 * kvm_dirty_bitmap_bytes(memslot) > PAGE_SIZE)
+               vfree(memslot->dirty_bitmap_head);
+       else
+               kfree(memslot->dirty_bitmap_head);
+
+       memslot->dirty_bitmap = NULL;
+       memslot->dirty_bitmap_head = NULL;
 }
 
 /*
@@ -1005,7 +477,7 @@ static void kvm_free_physmem_slot(struct kvm_memory_slot *free,
                vfree(free->rmap);
 
        if (!dont || free->dirty_bitmap != dont->dirty_bitmap)
-               vfree(free->dirty_bitmap);
+               kvm_destroy_dirty_bitmap(free);
 
 
        for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i) {
@@ -1016,20 +488,23 @@ static void kvm_free_physmem_slot(struct kvm_memory_slot *free,
        }
 
        free->npages = 0;
-       free->dirty_bitmap = NULL;
        free->rmap = NULL;
 }
 
 void kvm_free_physmem(struct kvm *kvm)
 {
        int i;
+       struct kvm_memslots *slots = kvm->memslots;
+
+       for (i = 0; i < slots->nmemslots; ++i)
+               kvm_free_physmem_slot(&slots->memslots[i], NULL);
 
-       for (i = 0; i < kvm->nmemslots; ++i)
-               kvm_free_physmem_slot(&kvm->memslots[i], NULL);
+       kfree(kvm->memslots);
 }
 
 static void kvm_destroy_vm(struct kvm *kvm)
 {
+       int i;
        struct mm_struct *mm = kvm->mm;
 
        kvm_arch_sync_events(kvm);
@@ -1037,18 +512,19 @@ static void kvm_destroy_vm(struct kvm *kvm)
        list_del(&kvm->vm_list);
        spin_unlock(&kvm_lock);
        kvm_free_irq_routing(kvm);
-       kvm_io_bus_destroy(&kvm->pio_bus);
-       kvm_io_bus_destroy(&kvm->mmio_bus);
-#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
-       if (kvm->coalesced_mmio_ring != NULL)
-               free_page((unsigned long)kvm->coalesced_mmio_ring);
-#endif
+       for (i = 0; i < KVM_NR_BUSES; i++)
+               kvm_io_bus_destroy(kvm->buses[i]);
+       kvm_coalesced_mmio_free(kvm);
 #if defined(CONFIG_MMU_NOTIFIER) && defined(KVM_ARCH_WANT_MMU_NOTIFIER)
        mmu_notifier_unregister(&kvm->mmu_notifier, kvm->mm);
 #else
        kvm_arch_flush_shadow(kvm);
 #endif
        kvm_arch_destroy_vm(kvm);
+       kvm_free_physmem(kvm);
+       cleanup_srcu_struct(&kvm->srcu);
+       kvm_arch_free_vm(kvm);
+       hardware_disable_all();
        mmdrop(mm);
 }
 
@@ -1077,6 +553,27 @@ static int kvm_vm_release(struct inode *inode, struct file *filp)
 }
 
 /*
+ * Allocation size is twice as large as the actual dirty bitmap size.
+ * This makes it possible to do double buffering: see x86's
+ * kvm_vm_ioctl_get_dirty_log().
+ */
+static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
+{
+       unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
+
+       if (dirty_bytes > PAGE_SIZE)
+               memslot->dirty_bitmap = vzalloc(dirty_bytes);
+       else
+               memslot->dirty_bitmap = kzalloc(dirty_bytes, GFP_KERNEL);
+
+       if (!memslot->dirty_bitmap)
+               return -ENOMEM;
+
+       memslot->dirty_bitmap_head = memslot->dirty_bitmap;
+       return 0;
+}
+
+/*
  * Allocate some memory and give it an address in the guest physical address
  * space.
  *
@@ -1088,12 +585,13 @@ int __kvm_set_memory_region(struct kvm *kvm,
                            struct kvm_userspace_memory_region *mem,
                            int user_alloc)
 {
-       int r;
+       int r, flush_shadow = 0;
        gfn_t base_gfn;
        unsigned long npages;
        unsigned long i;
        struct kvm_memory_slot *memslot;
        struct kvm_memory_slot old, new;
+       struct kvm_memslots *slots, *old_memslots;
 
        r = -EINVAL;
        /* General sanity checks */
@@ -1108,15 +606,20 @@ int __kvm_set_memory_region(struct kvm *kvm,
        if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
                goto out;
 
-       memslot = &kvm->memslots[mem->slot];
+       memslot = &kvm->memslots->memslots[mem->slot];
        base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
        npages = mem->memory_size >> PAGE_SHIFT;
 
+       r = -EINVAL;
+       if (npages > KVM_MEM_MAX_NR_PAGES)
+               goto out;
+
        if (!npages)
                mem->flags &= ~KVM_MEM_LOG_DIRTY_PAGES;
 
        new = old = *memslot;
 
+       new.id = mem->slot;
        new.base_gfn = base_gfn;
        new.npages = npages;
        new.flags = mem->flags;
@@ -1129,7 +632,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
        /* Check for overlaps */
        r = -EEXIST;
        for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
-               struct kvm_memory_slot *s = &kvm->memslots[i];
+               struct kvm_memory_slot *s = &kvm->memslots->memslots[i];
 
                if (s == memslot || !s->npages)
                        continue;
@@ -1147,23 +650,13 @@ int __kvm_set_memory_region(struct kvm *kvm,
        /* Allocate if a slot is being created */
 #ifndef CONFIG_S390
        if (npages && !new.rmap) {
-               new.rmap = vmalloc(npages * sizeof(struct page *));
+               new.rmap = vzalloc(npages * sizeof(*new.rmap));
 
                if (!new.rmap)
                        goto out_free;
 
-               memset(new.rmap, 0, npages * sizeof(*new.rmap));
-
                new.user_alloc = user_alloc;
-               /*
-                * hva_to_rmmap() serialzies with the mmu_lock and to be
-                * safe it has to ignore memslots with !user_alloc &&
-                * !userspace_addr.
-                */
-               if (user_alloc)
-                       new.userspace_addr = mem->userspace_addr;
-               else
-                       new.userspace_addr = 0;
+               new.userspace_addr = mem->userspace_addr;
        }
        if (!npages)
                goto skip_lpage;
@@ -1180,21 +673,18 @@ int __kvm_set_memory_region(struct kvm *kvm,
                if (new.lpage_info[i])
                        continue;
 
-               lpages = 1 + (base_gfn + npages - 1) /
-                            KVM_PAGES_PER_HPAGE(level);
-               lpages -= base_gfn / KVM_PAGES_PER_HPAGE(level);
+               lpages = 1 + ((base_gfn + npages - 1)
+                            >> KVM_HPAGE_GFN_SHIFT(level));
+               lpages -= base_gfn >> KVM_HPAGE_GFN_SHIFT(level);
 
-               new.lpage_info[i] = vmalloc(lpages * sizeof(*new.lpage_info[i]));
+               new.lpage_info[i] = vzalloc(lpages * sizeof(*new.lpage_info[i]));
 
                if (!new.lpage_info[i])
                        goto out_free;
 
-               memset(new.lpage_info[i], 0,
-                      lpages * sizeof(*new.lpage_info[i]));
-
-               if (base_gfn % KVM_PAGES_PER_HPAGE(level))
+               if (base_gfn & (KVM_PAGES_PER_HPAGE(level) - 1))
                        new.lpage_info[i][0].write_count = 1;
-               if ((base_gfn+npages) % KVM_PAGES_PER_HPAGE(level))
+               if ((base_gfn+npages) & (KVM_PAGES_PER_HPAGE(level) - 1))
                        new.lpage_info[i][lpages - 1].write_count = 1;
                ugfn = new.userspace_addr >> PAGE_SHIFT;
                /*
@@ -1212,14 +702,11 @@ skip_lpage:
 
        /* Allocate page dirty bitmap if needed */
        if ((new.flags & KVM_MEM_LOG_DIRTY_PAGES) && !new.dirty_bitmap) {
-               unsigned dirty_bytes = ALIGN(npages, BITS_PER_LONG) / 8;
-
-               new.dirty_bitmap = vmalloc(dirty_bytes);
-               if (!new.dirty_bitmap)
+               if (kvm_create_dirty_bitmap(&new) < 0)
                        goto out_free;
-               memset(new.dirty_bitmap, 0, dirty_bytes);
+               /* destroy any largepage mappings for dirty tracking */
                if (old.npages)
-                       kvm_arch_flush_shadow(kvm);
+                       flush_shadow = 1;
        }
 #else  /* not defined CONFIG_S390 */
        new.user_alloc = user_alloc;
@@ -1227,36 +714,72 @@ skip_lpage:
                new.userspace_addr = mem->userspace_addr;
 #endif /* not defined CONFIG_S390 */
 
-       if (!npages)
+       if (!npages) {
+               r = -ENOMEM;
+               slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
+               if (!slots)
+                       goto out_free;
+               memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots));
+               if (mem->slot >= slots->nmemslots)
+                       slots->nmemslots = mem->slot + 1;
+               slots->generation++;
+               slots->memslots[mem->slot].flags |= KVM_MEMSLOT_INVALID;
+
+               old_memslots = kvm->memslots;
+               rcu_assign_pointer(kvm->memslots, slots);
+               synchronize_srcu_expedited(&kvm->srcu);
+               /* From this point no new shadow pages pointing to a deleted
+                * memslot will be created.
+                *
+                * validation of sp->gfn happens in:
+                *      - gfn_to_hva (kvm_read_guest, gfn_to_pfn)
+                *      - kvm_is_visible_gfn (mmu_check_roots)
+                */
                kvm_arch_flush_shadow(kvm);
+               kfree(old_memslots);
+       }
 
-       spin_lock(&kvm->mmu_lock);
-       if (mem->slot >= kvm->nmemslots)
-               kvm->nmemslots = mem->slot + 1;
+       r = kvm_arch_prepare_memory_region(kvm, &new, old, mem, user_alloc);
+       if (r)
+               goto out_free;
 
-       *memslot = new;
-       spin_unlock(&kvm->mmu_lock);
+       /* map the pages in iommu page table */
+       if (npages) {
+               r = kvm_iommu_map_pages(kvm, &new);
+               if (r)
+                       goto out_free;
+       }
 
-       r = kvm_arch_set_memory_region(kvm, mem, old, user_alloc);
-       if (r) {
-               spin_lock(&kvm->mmu_lock);
-               *memslot = old;
-               spin_unlock(&kvm->mmu_lock);
+       r = -ENOMEM;
+       slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
+       if (!slots)
                goto out_free;
+       memcpy(slots, kvm->memslots, sizeof(struct kvm_memslots));
+       if (mem->slot >= slots->nmemslots)
+               slots->nmemslots = mem->slot + 1;
+       slots->generation++;
+
+       /* actual memory is freed via old in kvm_free_physmem_slot below */
+       if (!npages) {
+               new.rmap = NULL;
+               new.dirty_bitmap = NULL;
+               for (i = 0; i < KVM_NR_PAGE_SIZES - 1; ++i)
+                       new.lpage_info[i] = NULL;
        }
 
-       kvm_free_physmem_slot(&old, npages ? &new : NULL);
-       /* Slot deletion case: we have to update the current slot */
-       spin_lock(&kvm->mmu_lock);
-       if (!npages)
-               *memslot = old;
-       spin_unlock(&kvm->mmu_lock);
-#ifdef CONFIG_DMAR
-       /* map the pages in iommu page table */
-       r = kvm_iommu_map_pages(kvm, base_gfn, npages);
-       if (r)
-               goto out;
-#endif
+       slots->memslots[mem->slot] = new;
+       old_memslots = kvm->memslots;
+       rcu_assign_pointer(kvm->memslots, slots);
+       synchronize_srcu_expedited(&kvm->srcu);
+
+       kvm_arch_commit_memory_region(kvm, mem, old, user_alloc);
+
+       kvm_free_physmem_slot(&old, &new);
+       kfree(old_memslots);
+
+       if (flush_shadow)
+               kvm_arch_flush_shadow(kvm);
+
        return 0;
 
 out_free:
@@ -1273,9 +796,9 @@ int kvm_set_memory_region(struct kvm *kvm,
 {
        int r;
 
-       down_write(&kvm->slots_lock);
+       mutex_lock(&kvm->slots_lock);
        r = __kvm_set_memory_region(kvm, mem, user_alloc);
-       up_write(&kvm->slots_lock);
+       mutex_unlock(&kvm->slots_lock);
        return r;
 }
 EXPORT_SYMBOL_GPL(kvm_set_memory_region);
@@ -1295,19 +818,19 @@ int kvm_get_dirty_log(struct kvm *kvm,
 {
        struct kvm_memory_slot *memslot;
        int r, i;
-       int n;
+       unsigned long n;
        unsigned long any = 0;
 
        r = -EINVAL;
        if (log->slot >= KVM_MEMORY_SLOTS)
                goto out;
 
-       memslot = &kvm->memslots[log->slot];
+       memslot = &kvm->memslots->memslots[log->slot];
        r = -ENOENT;
        if (!memslot->dirty_bitmap)
                goto out;
 
-       n = ALIGN(memslot->npages, BITS_PER_LONG) / 8;
+       n = kvm_dirty_bitmap_bytes(memslot);
 
        for (i = 0; !any && i < n/sizeof(long); ++i)
                any = memslot->dirty_bitmap[i];
@@ -1332,16 +855,28 @@ EXPORT_SYMBOL_GPL(kvm_disable_largepages);
 
 int is_error_page(struct page *page)
 {
-       return page == bad_page;
+       return page == bad_page || page == hwpoison_page || page == fault_page;
 }
 EXPORT_SYMBOL_GPL(is_error_page);
 
 int is_error_pfn(pfn_t pfn)
 {
-       return pfn == bad_pfn;
+       return pfn == bad_pfn || pfn == hwpoison_pfn || pfn == fault_pfn;
 }
 EXPORT_SYMBOL_GPL(is_error_pfn);
 
+int is_hwpoison_pfn(pfn_t pfn)
+{
+       return pfn == hwpoison_pfn;
+}
+EXPORT_SYMBOL_GPL(is_hwpoison_pfn);
+
+int is_fault_pfn(pfn_t pfn)
+{
+       return pfn == fault_pfn;
+}
+EXPORT_SYMBOL_GPL(is_fault_pfn);
+
 static inline unsigned long bad_hva(void)
 {
        return PAGE_OFFSET;
@@ -1353,12 +888,13 @@ int kvm_is_error_hva(unsigned long addr)
 }
 EXPORT_SYMBOL_GPL(kvm_is_error_hva);
 
-struct kvm_memory_slot *gfn_to_memslot_unaliased(struct kvm *kvm, gfn_t gfn)
+static struct kvm_memory_slot *__gfn_to_memslot(struct kvm_memslots *slots,
+                                               gfn_t gfn)
 {
        int i;
 
-       for (i = 0; i < kvm->nmemslots; ++i) {
-               struct kvm_memory_slot *memslot = &kvm->memslots[i];
+       for (i = 0; i < slots->nmemslots; ++i) {
+               struct kvm_memory_slot *memslot = &slots->memslots[i];
 
                if (gfn >= memslot->base_gfn
                    && gfn < memslot->base_gfn + memslot->npages)
@@ -1366,21 +902,23 @@ struct kvm_memory_slot *gfn_to_memslot_unaliased(struct kvm *kvm, gfn_t gfn)
        }
        return NULL;
 }
-EXPORT_SYMBOL_GPL(gfn_to_memslot_unaliased);
 
 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
 {
-       gfn = unalias_gfn(kvm, gfn);
-       return gfn_to_memslot_unaliased(kvm, gfn);
+       return __gfn_to_memslot(kvm_memslots(kvm), gfn);
 }
+EXPORT_SYMBOL_GPL(gfn_to_memslot);
 
 int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
 {
        int i;
+       struct kvm_memslots *slots = kvm_memslots(kvm);
 
-       gfn = unalias_gfn(kvm, gfn);
        for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
-               struct kvm_memory_slot *memslot = &kvm->memslots[i];
+               struct kvm_memory_slot *memslot = &slots->memslots[i];
+
+               if (memslot->flags & KVM_MEMSLOT_INVALID)
+                       continue;
 
                if (gfn >= memslot->base_gfn
                    && gfn < memslot->base_gfn + memslot->npages)
@@ -1390,59 +928,217 @@ int kvm_is_visible_gfn(struct kvm *kvm, gfn_t gfn)
 }
 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
 
-unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
+unsigned long kvm_host_page_size(struct kvm *kvm, gfn_t gfn)
+{
+       struct vm_area_struct *vma;
+       unsigned long addr, size;
+
+       size = PAGE_SIZE;
+
+       addr = gfn_to_hva(kvm, gfn);
+       if (kvm_is_error_hva(addr))
+               return PAGE_SIZE;
+
+       down_read(&current->mm->mmap_sem);
+       vma = find_vma(current->mm, addr);
+       if (!vma)
+               goto out;
+
+       size = vma_kernel_pagesize(vma);
+
+out:
+       up_read(&current->mm->mmap_sem);
+
+       return size;
+}
+
+int memslot_id(struct kvm *kvm, gfn_t gfn)
 {
-       struct kvm_memory_slot *slot;
+       int i;
+       struct kvm_memslots *slots = kvm_memslots(kvm);
+       struct kvm_memory_slot *memslot = NULL;
+
+       for (i = 0; i < slots->nmemslots; ++i) {
+               memslot = &slots->memslots[i];
+
+               if (gfn >= memslot->base_gfn
+                   && gfn < memslot->base_gfn + memslot->npages)
+                       break;
+       }
 
-       gfn = unalias_gfn(kvm, gfn);
-       slot = gfn_to_memslot_unaliased(kvm, gfn);
-       if (!slot)
+       return memslot - slots->memslots;
+}
+
+static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
+                                    gfn_t *nr_pages)
+{
+       if (!slot || slot->flags & KVM_MEMSLOT_INVALID)
                return bad_hva();
-       return (slot->userspace_addr + (gfn - slot->base_gfn) * PAGE_SIZE);
+
+       if (nr_pages)
+               *nr_pages = slot->npages - (gfn - slot->base_gfn);
+
+       return gfn_to_hva_memslot(slot, gfn);
+}
+
+unsigned long gfn_to_hva(struct kvm *kvm, gfn_t gfn)
+{
+       return gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, NULL);
 }
 EXPORT_SYMBOL_GPL(gfn_to_hva);
 
-pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
+static pfn_t get_fault_pfn(void)
+{
+       get_page(fault_page);
+       return fault_pfn;
+}
+
+static pfn_t hva_to_pfn(struct kvm *kvm, unsigned long addr, bool atomic,
+                       bool *async, bool write_fault, bool *writable)
 {
        struct page *page[1];
-       unsigned long addr;
-       int npages;
+       int npages = 0;
        pfn_t pfn;
 
-       might_sleep();
+       /* we can do it either atomically or asynchronously, not both */
+       BUG_ON(atomic && async);
 
-       addr = gfn_to_hva(kvm, gfn);
-       if (kvm_is_error_hva(addr)) {
-               get_page(bad_page);
-               return page_to_pfn(bad_page);
-       }
+       BUG_ON(!write_fault && !writable);
+
+       if (writable)
+               *writable = true;
+
+       if (atomic || async)
+               npages = __get_user_pages_fast(addr, 1, 1, page);
+
+       if (unlikely(npages != 1) && !atomic) {
+               might_sleep();
 
-       npages = get_user_pages_fast(addr, 1, 1, page);
+               if (writable)
+                       *writable = write_fault;
+
+               npages = get_user_pages_fast(addr, 1, write_fault, page);
+
+               /* map read fault as writable if possible */
+               if (unlikely(!write_fault) && npages == 1) {
+                       struct page *wpage[1];
+
+                       npages = __get_user_pages_fast(addr, 1, 1, wpage);
+                       if (npages == 1) {
+                               *writable = true;
+                               put_page(page[0]);
+                               page[0] = wpage[0];
+                       }
+                       npages = 1;
+               }
+       }
 
        if (unlikely(npages != 1)) {
                struct vm_area_struct *vma;
 
-               down_read(&current->mm->mmap_sem);
-               vma = find_vma(current->mm, addr);
+               if (atomic)
+                       return get_fault_pfn();
 
-               if (vma == NULL || addr < vma->vm_start ||
-                   !(vma->vm_flags & VM_PFNMAP)) {
+               down_read(&current->mm->mmap_sem);
+               if (is_hwpoison_address(addr)) {
                        up_read(&current->mm->mmap_sem);
-                       get_page(bad_page);
-                       return page_to_pfn(bad_page);
+                       get_page(hwpoison_page);
+                       return page_to_pfn(hwpoison_page);
                }
 
-               pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
+               vma = find_vma_intersection(current->mm, addr, addr+1);
+
+               if (vma == NULL)
+                       pfn = get_fault_pfn();
+               else if ((vma->vm_flags & VM_PFNMAP)) {
+                       pfn = ((addr - vma->vm_start) >> PAGE_SHIFT) +
+                               vma->vm_pgoff;
+                       BUG_ON(!kvm_is_mmio_pfn(pfn));
+               } else {
+                       if (async && (vma->vm_flags & VM_WRITE))
+                               *async = true;
+                       pfn = get_fault_pfn();
+               }
                up_read(&current->mm->mmap_sem);
-               BUG_ON(!kvm_is_mmio_pfn(pfn));
        } else
                pfn = page_to_pfn(page[0]);
 
        return pfn;
 }
 
+pfn_t hva_to_pfn_atomic(struct kvm *kvm, unsigned long addr)
+{
+       return hva_to_pfn(kvm, addr, true, NULL, true, NULL);
+}
+EXPORT_SYMBOL_GPL(hva_to_pfn_atomic);
+
+static pfn_t __gfn_to_pfn(struct kvm *kvm, gfn_t gfn, bool atomic, bool *async,
+                         bool write_fault, bool *writable)
+{
+       unsigned long addr;
+
+       if (async)
+               *async = false;
+
+       addr = gfn_to_hva(kvm, gfn);
+       if (kvm_is_error_hva(addr)) {
+               get_page(bad_page);
+               return page_to_pfn(bad_page);
+       }
+
+       return hva_to_pfn(kvm, addr, atomic, async, write_fault, writable);
+}
+
+pfn_t gfn_to_pfn_atomic(struct kvm *kvm, gfn_t gfn)
+{
+       return __gfn_to_pfn(kvm, gfn, true, NULL, true, NULL);
+}
+EXPORT_SYMBOL_GPL(gfn_to_pfn_atomic);
+
+pfn_t gfn_to_pfn_async(struct kvm *kvm, gfn_t gfn, bool *async,
+                      bool write_fault, bool *writable)
+{
+       return __gfn_to_pfn(kvm, gfn, false, async, write_fault, writable);
+}
+EXPORT_SYMBOL_GPL(gfn_to_pfn_async);
+
+pfn_t gfn_to_pfn(struct kvm *kvm, gfn_t gfn)
+{
+       return __gfn_to_pfn(kvm, gfn, false, NULL, true, NULL);
+}
 EXPORT_SYMBOL_GPL(gfn_to_pfn);
 
+pfn_t gfn_to_pfn_prot(struct kvm *kvm, gfn_t gfn, bool write_fault,
+                     bool *writable)
+{
+       return __gfn_to_pfn(kvm, gfn, false, NULL, write_fault, writable);
+}
+EXPORT_SYMBOL_GPL(gfn_to_pfn_prot);
+
+pfn_t gfn_to_pfn_memslot(struct kvm *kvm,
+                        struct kvm_memory_slot *slot, gfn_t gfn)
+{
+       unsigned long addr = gfn_to_hva_memslot(slot, gfn);
+       return hva_to_pfn(kvm, addr, false, NULL, true, NULL);
+}
+
+int gfn_to_page_many_atomic(struct kvm *kvm, gfn_t gfn, struct page **pages,
+                                                                 int nr_pages)
+{
+       unsigned long addr;
+       gfn_t entry;
+
+       addr = gfn_to_hva_many(gfn_to_memslot(kvm, gfn), gfn, &entry);
+       if (kvm_is_error_hva(addr))
+               return -1;
+
+       if (entry < nr_pages)
+               return 0;
+
+       return __get_user_pages_fast(addr, nr_pages, 1, pages);
+}
+EXPORT_SYMBOL_GPL(gfn_to_page_many_atomic);
+
 struct page *gfn_to_page(struct kvm *kvm, gfn_t gfn)
 {
        pfn_t pfn;
@@ -1616,9 +1312,51 @@ int kvm_write_guest(struct kvm *kvm, gpa_t gpa, const void *data,
        return 0;
 }
 
+int kvm_gfn_to_hva_cache_init(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
+                             gpa_t gpa)
+{
+       struct kvm_memslots *slots = kvm_memslots(kvm);
+       int offset = offset_in_page(gpa);
+       gfn_t gfn = gpa >> PAGE_SHIFT;
+
+       ghc->gpa = gpa;
+       ghc->generation = slots->generation;
+       ghc->memslot = __gfn_to_memslot(slots, gfn);
+       ghc->hva = gfn_to_hva_many(ghc->memslot, gfn, NULL);
+       if (!kvm_is_error_hva(ghc->hva))
+               ghc->hva += offset;
+       else
+               return -EFAULT;
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(kvm_gfn_to_hva_cache_init);
+
+int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
+                          void *data, unsigned long len)
+{
+       struct kvm_memslots *slots = kvm_memslots(kvm);
+       int r;
+
+       if (slots->generation != ghc->generation)
+               kvm_gfn_to_hva_cache_init(kvm, ghc, ghc->gpa);
+
+       if (kvm_is_error_hva(ghc->hva))
+               return -EFAULT;
+
+       r = copy_to_user((void __user *)ghc->hva, data, len);
+       if (r)
+               return -EFAULT;
+       mark_page_dirty_in_slot(kvm, ghc->memslot, ghc->gpa >> PAGE_SHIFT);
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
+
 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
 {
-       return kvm_write_guest_page(kvm, gfn, empty_zero_page, offset, len);
+       return kvm_write_guest_page(kvm, gfn, (const void *) empty_zero_page,
+                                   offset, len);
 }
 EXPORT_SYMBOL_GPL(kvm_clear_guest_page);
 
@@ -1641,21 +1379,24 @@ int kvm_clear_guest(struct kvm *kvm, gpa_t gpa, unsigned long len)
 }
 EXPORT_SYMBOL_GPL(kvm_clear_guest);
 
-void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
+void mark_page_dirty_in_slot(struct kvm *kvm, struct kvm_memory_slot *memslot,
+                            gfn_t gfn)
 {
-       struct kvm_memory_slot *memslot;
-
-       gfn = unalias_gfn(kvm, gfn);
-       memslot = gfn_to_memslot_unaliased(kvm, gfn);
        if (memslot && memslot->dirty_bitmap) {
                unsigned long rel_gfn = gfn - memslot->base_gfn;
 
-               /* avoid RMW */
-               if (!test_bit(rel_gfn, memslot->dirty_bitmap))
-                       set_bit(rel_gfn, memslot->dirty_bitmap);
+               generic___set_le_bit(rel_gfn, memslot->dirty_bitmap);
        }
 }
 
+void mark_page_dirty(struct kvm *kvm, gfn_t gfn)
+{
+       struct kvm_memory_slot *memslot;
+
+       memslot = gfn_to_memslot(kvm, gfn);
+       mark_page_dirty_in_slot(kvm, memslot, gfn);
+}
+
 /*
  * The vCPU has executed a HLT instruction with in-kernel mode enabled.
  */
@@ -1667,7 +1408,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
                prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
 
                if (kvm_arch_vcpu_runnable(vcpu)) {
-                       set_bit(KVM_REQ_UNHALT, &vcpu->requests);
+                       kvm_make_request(KVM_REQ_UNHALT, vcpu);
                        break;
                }
                if (kvm_cpu_has_pending_timer(vcpu))
@@ -1675,9 +1416,7 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
                if (signal_pending(current))
                        break;
 
-               vcpu_put(vcpu);
                schedule();
-               vcpu_load(vcpu);
        }
 
        finish_wait(&vcpu->wq, &wait);
@@ -1691,6 +1430,21 @@ void kvm_resched(struct kvm_vcpu *vcpu)
 }
 EXPORT_SYMBOL_GPL(kvm_resched);
 
+void kvm_vcpu_on_spin(struct kvm_vcpu *vcpu)
+{
+       ktime_t expires;
+       DEFINE_WAIT(wait);
+
+       prepare_to_wait(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
+
+       /* Sleep for 100 us, and hope lock-holder got scheduled */
+       expires = ktime_add_ns(ktime_get(), 100000UL);
+       schedule_hrtimeout(&expires, HRTIMER_MODE_ABS);
+
+       finish_wait(&vcpu->wq, &wait);
+}
+EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
+
 static int kvm_vcpu_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
        struct kvm_vcpu *vcpu = vma->vm_file->private_data;
@@ -1736,6 +1490,7 @@ static struct file_operations kvm_vcpu_fops = {
        .unlocked_ioctl = kvm_vcpu_ioctl,
        .compat_ioctl   = kvm_vcpu_ioctl,
        .mmap           = kvm_vcpu_mmap,
+       .llseek         = noop_llseek,
 };
 
 /*
@@ -1743,7 +1498,7 @@ static struct file_operations kvm_vcpu_fops = {
  */
 static int create_vcpu_fd(struct kvm_vcpu *vcpu)
 {
-       return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, 0);
+       return anon_inode_getfd("kvm-vcpu", &kvm_vcpu_fops, vcpu, O_RDWR);
 }
 
 /*
@@ -1814,88 +1569,6 @@ static int kvm_vcpu_ioctl_set_sigmask(struct kvm_vcpu *vcpu, sigset_t *sigset)
        return 0;
 }
 
-#ifdef __KVM_HAVE_MSIX
-static int kvm_vm_ioctl_set_msix_nr(struct kvm *kvm,
-                                   struct kvm_assigned_msix_nr *entry_nr)
-{
-       int r = 0;
-       struct kvm_assigned_dev_kernel *adev;
-
-       mutex_lock(&kvm->lock);
-
-       adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     entry_nr->assigned_dev_id);
-       if (!adev) {
-               r = -EINVAL;
-               goto msix_nr_out;
-       }
-
-       if (adev->entries_nr == 0) {
-               adev->entries_nr = entry_nr->entry_nr;
-               if (adev->entries_nr == 0 ||
-                   adev->entries_nr >= KVM_MAX_MSIX_PER_DEV) {
-                       r = -EINVAL;
-                       goto msix_nr_out;
-               }
-
-               adev->host_msix_entries = kzalloc(sizeof(struct msix_entry) *
-                                               entry_nr->entry_nr,
-                                               GFP_KERNEL);
-               if (!adev->host_msix_entries) {
-                       r = -ENOMEM;
-                       goto msix_nr_out;
-               }
-               adev->guest_msix_entries = kzalloc(
-                               sizeof(struct kvm_guest_msix_entry) *
-                               entry_nr->entry_nr, GFP_KERNEL);
-               if (!adev->guest_msix_entries) {
-                       kfree(adev->host_msix_entries);
-                       r = -ENOMEM;
-                       goto msix_nr_out;
-               }
-       } else /* Not allowed set MSI-X number twice */
-               r = -EINVAL;
-msix_nr_out:
-       mutex_unlock(&kvm->lock);
-       return r;
-}
-
-static int kvm_vm_ioctl_set_msix_entry(struct kvm *kvm,
-                                      struct kvm_assigned_msix_entry *entry)
-{
-       int r = 0, i;
-       struct kvm_assigned_dev_kernel *adev;
-
-       mutex_lock(&kvm->lock);
-
-       adev = kvm_find_assigned_dev(&kvm->arch.assigned_dev_head,
-                                     entry->assigned_dev_id);
-
-       if (!adev) {
-               r = -EINVAL;
-               goto msix_entry_out;
-       }
-
-       for (i = 0; i < adev->entries_nr; i++)
-               if (adev->guest_msix_entries[i].vector == 0 ||
-                   adev->guest_msix_entries[i].entry == entry->entry) {
-                       adev->guest_msix_entries[i].entry = entry->entry;
-                       adev->guest_msix_entries[i].vector = entry->gsi;
-                       adev->host_msix_entries[i].entry = entry->entry;
-                       break;
-               }
-       if (i == adev->entries_nr) {
-               r = -ENOSPC;
-               goto msix_entry_out;
-       }
-
-msix_entry_out:
-       mutex_unlock(&kvm->lock);
-
-       return r;
-}
-#endif
-
 static long kvm_vcpu_ioctl(struct file *filp,
                           unsigned int ioctl, unsigned long arg)
 {
@@ -1907,12 +1580,25 @@ static long kvm_vcpu_ioctl(struct file *filp,
 
        if (vcpu->kvm->mm != current->mm)
                return -EIO;
+
+#if defined(CONFIG_S390) || defined(CONFIG_PPC)
+       /*
+        * Special cases: vcpu ioctls that are asynchronous to vcpu execution,
+        * so vcpu_load() would break it.
+        */
+       if (ioctl == KVM_S390_INTERRUPT || ioctl == KVM_INTERRUPT)
+               return kvm_arch_vcpu_ioctl(filp, ioctl, arg);
+#endif
+
+
+       vcpu_load(vcpu);
        switch (ioctl) {
        case KVM_RUN:
                r = -EINVAL;
                if (arg)
                        goto out;
                r = kvm_arch_vcpu_ioctl_run(vcpu, vcpu->run);
+               trace_kvm_userspace_exit(vcpu->run->exit_reason, r);
                break;
        case KVM_GET_REGS: {
                struct kvm_regs *kvm_regs;
@@ -2049,7 +1735,7 @@ out_free2:
                                goto out;
                        p = &sigset;
                }
-               r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
+               r = kvm_vcpu_ioctl_set_sigmask(vcpu, p);
                break;
        }
        case KVM_GET_FPU: {
@@ -2084,6 +1770,7 @@ out_free2:
                r = kvm_arch_vcpu_ioctl(filp, ioctl, arg);
        }
 out:
+       vcpu_put(vcpu);
        kfree(fpu);
        kfree(kvm_sregs);
        return r;
@@ -2134,7 +1821,6 @@ static long kvm_vm_ioctl(struct file *filp,
                r = -EFAULT;
                if (copy_from_user(&zone, argp, sizeof zone))
                        goto out;
-               r = -ENXIO;
                r = kvm_vm_ioctl_register_coalesced_mmio(kvm, &zone);
                if (r)
                        goto out;
@@ -2146,7 +1832,6 @@ static long kvm_vm_ioctl(struct file *filp,
                r = -EFAULT;
                if (copy_from_user(&zone, argp, sizeof zone))
                        goto out;
-               r = -ENXIO;
                r = kvm_vm_ioctl_unregister_coalesced_mmio(kvm, &zone);
                if (r)
                        goto out;
@@ -2154,112 +1839,6 @@ static long kvm_vm_ioctl(struct file *filp,
                break;
        }
 #endif
-#ifdef KVM_CAP_DEVICE_ASSIGNMENT
-       case KVM_ASSIGN_PCI_DEVICE: {
-               struct kvm_assigned_pci_dev assigned_dev;
-
-               r = -EFAULT;
-               if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
-                       goto out;
-               r = kvm_vm_ioctl_assign_device(kvm, &assigned_dev);
-               if (r)
-                       goto out;
-               break;
-       }
-       case KVM_ASSIGN_IRQ: {
-               r = -EOPNOTSUPP;
-               break;
-       }
-#ifdef KVM_CAP_ASSIGN_DEV_IRQ
-       case KVM_ASSIGN_DEV_IRQ: {
-               struct kvm_assigned_irq assigned_irq;
-
-               r = -EFAULT;
-               if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
-                       goto out;
-               r = kvm_vm_ioctl_assign_irq(kvm, &assigned_irq);
-               if (r)
-                       goto out;
-               break;
-       }
-       case KVM_DEASSIGN_DEV_IRQ: {
-               struct kvm_assigned_irq assigned_irq;
-
-               r = -EFAULT;
-               if (copy_from_user(&assigned_irq, argp, sizeof assigned_irq))
-                       goto out;
-               r = kvm_vm_ioctl_deassign_dev_irq(kvm, &assigned_irq);
-               if (r)
-                       goto out;
-               break;
-       }
-#endif
-#endif
-#ifdef KVM_CAP_DEVICE_DEASSIGNMENT
-       case KVM_DEASSIGN_PCI_DEVICE: {
-               struct kvm_assigned_pci_dev assigned_dev;
-
-               r = -EFAULT;
-               if (copy_from_user(&assigned_dev, argp, sizeof assigned_dev))
-                       goto out;
-               r = kvm_vm_ioctl_deassign_device(kvm, &assigned_dev);
-               if (r)
-                       goto out;
-               break;
-       }
-#endif
-#ifdef KVM_CAP_IRQ_ROUTING
-       case KVM_SET_GSI_ROUTING: {
-               struct kvm_irq_routing routing;
-               struct kvm_irq_routing __user *urouting;
-               struct kvm_irq_routing_entry *entries;
-
-               r = -EFAULT;
-               if (copy_from_user(&routing, argp, sizeof(routing)))
-                       goto out;
-               r = -EINVAL;
-               if (routing.nr >= KVM_MAX_IRQ_ROUTES)
-                       goto out;
-               if (routing.flags)
-                       goto out;
-               r = -ENOMEM;
-               entries = vmalloc(routing.nr * sizeof(*entries));
-               if (!entries)
-                       goto out;
-               r = -EFAULT;
-               urouting = argp;
-               if (copy_from_user(entries, urouting->entries,
-                                  routing.nr * sizeof(*entries)))
-                       goto out_free_irq_routing;
-               r = kvm_set_irq_routing(kvm, entries, routing.nr,
-                                       routing.flags);
-       out_free_irq_routing:
-               vfree(entries);
-               break;
-       }
-#endif /* KVM_CAP_IRQ_ROUTING */
-#ifdef __KVM_HAVE_MSIX
-       case KVM_ASSIGN_SET_MSIX_NR: {
-               struct kvm_assigned_msix_nr entry_nr;
-               r = -EFAULT;
-               if (copy_from_user(&entry_nr, argp, sizeof entry_nr))
-                       goto out;
-               r = kvm_vm_ioctl_set_msix_nr(kvm, &entry_nr);
-               if (r)
-                       goto out;
-               break;
-       }
-       case KVM_ASSIGN_SET_MSIX_ENTRY: {
-               struct kvm_assigned_msix_entry entry;
-               r = -EFAULT;
-               if (copy_from_user(&entry, argp, sizeof entry))
-                       goto out;
-               r = kvm_vm_ioctl_set_msix_entry(kvm, &entry);
-               if (r)
-                       goto out;
-               break;
-       }
-#endif
        case KVM_IRQFD: {
                struct kvm_irqfd data;
 
@@ -2291,10 +1870,58 @@ static long kvm_vm_ioctl(struct file *filp,
 #endif
        default:
                r = kvm_arch_vm_ioctl(filp, ioctl, arg);
+               if (r == -ENOTTY)
+                       r = kvm_vm_ioctl_assigned_device(kvm, ioctl, arg);
+       }
+out:
+       return r;
+}
+
+#ifdef CONFIG_COMPAT
+struct compat_kvm_dirty_log {
+       __u32 slot;
+       __u32 padding1;
+       union {
+               compat_uptr_t dirty_bitmap; /* one bit per page */
+               __u64 padding2;
+       };
+};
+
+static long kvm_vm_compat_ioctl(struct file *filp,
+                          unsigned int ioctl, unsigned long arg)
+{
+       struct kvm *kvm = filp->private_data;
+       int r;
+
+       if (kvm->mm != current->mm)
+               return -EIO;
+       switch (ioctl) {
+       case KVM_GET_DIRTY_LOG: {
+               struct compat_kvm_dirty_log compat_log;
+               struct kvm_dirty_log log;
+
+               r = -EFAULT;
+               if (copy_from_user(&compat_log, (void __user *)arg,
+                                  sizeof(compat_log)))
+                       goto out;
+               log.slot         = compat_log.slot;
+               log.padding1     = compat_log.padding1;
+               log.padding2     = compat_log.padding2;
+               log.dirty_bitmap = compat_ptr(compat_log.dirty_bitmap);
+
+               r = kvm_vm_ioctl_get_dirty_log(kvm, &log);
+               if (r)
+                       goto out;
+               break;
+       }
+       default:
+               r = kvm_vm_ioctl(filp, ioctl, arg);
        }
+
 out:
        return r;
 }
+#endif
 
 static int kvm_vm_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 {
@@ -2330,23 +1957,33 @@ static int kvm_vm_mmap(struct file *file, struct vm_area_struct *vma)
 static struct file_operations kvm_vm_fops = {
        .release        = kvm_vm_release,
        .unlocked_ioctl = kvm_vm_ioctl,
-       .compat_ioctl   = kvm_vm_ioctl,
+#ifdef CONFIG_COMPAT
+       .compat_ioctl   = kvm_vm_compat_ioctl,
+#endif
        .mmap           = kvm_vm_mmap,
+       .llseek         = noop_llseek,
 };
 
 static int kvm_dev_ioctl_create_vm(void)
 {
-       int fd;
+       int r;
        struct kvm *kvm;
 
        kvm = kvm_create_vm();
        if (IS_ERR(kvm))
                return PTR_ERR(kvm);
-       fd = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, 0);
-       if (fd < 0)
+#ifdef KVM_COALESCED_MMIO_PAGE_OFFSET
+       r = kvm_coalesced_mmio_init(kvm);
+       if (r < 0) {
+               kvm_put_kvm(kvm);
+               return r;
+       }
+#endif
+       r = anon_inode_getfd("kvm-vm", &kvm_vm_fops, kvm, O_RDWR);
+       if (r < 0)
                kvm_put_kvm(kvm);
 
-       return fd;
+       return r;
 }
 
 static long kvm_dev_ioctl_check_extension_generic(long arg)
@@ -2358,6 +1995,7 @@ static long kvm_dev_ioctl_check_extension_generic(long arg)
 #ifdef CONFIG_KVM_APIC_ARCHITECTURE
        case KVM_CAP_SET_BOOT_CPU_ID:
 #endif
+       case KVM_CAP_INTERNAL_ERROR_DATA:
                return 1;
 #ifdef CONFIG_HAVE_KVM_IRQCHIP
        case KVM_CAP_IRQ_ROUTING:
@@ -2417,6 +2055,7 @@ out:
 static struct file_operations kvm_chardev_ops = {
        .unlocked_ioctl = kvm_dev_ioctl,
        .compat_ioctl   = kvm_dev_ioctl,
+       .llseek         = noop_llseek,
 };
 
 static struct miscdevice kvm_dev = {
@@ -2425,17 +2064,34 @@ static struct miscdevice kvm_dev = {
        &kvm_chardev_ops,
 };
 
-static void hardware_enable(void *junk)
+static void hardware_enable_nolock(void *junk)
 {
        int cpu = raw_smp_processor_id();
+       int r;
 
        if (cpumask_test_cpu(cpu, cpus_hardware_enabled))
                return;
+
        cpumask_set_cpu(cpu, cpus_hardware_enabled);
-       kvm_arch_hardware_enable(NULL);
+
+       r = kvm_arch_hardware_enable(NULL);
+
+       if (r) {
+               cpumask_clear_cpu(cpu, cpus_hardware_enabled);
+               atomic_inc(&hardware_enable_failed);
+               printk(KERN_INFO "kvm: enabling virtualization on "
+                                "CPU%d failed\n", cpu);
+       }
 }
 
-static void hardware_disable(void *junk)
+static void hardware_enable(void *junk)
+{
+       spin_lock(&kvm_lock);
+       hardware_enable_nolock(junk);
+       spin_unlock(&kvm_lock);
+}
+
+static void hardware_disable_nolock(void *junk)
 {
        int cpu = raw_smp_processor_id();
 
@@ -2445,11 +2101,59 @@ static void hardware_disable(void *junk)
        kvm_arch_hardware_disable(NULL);
 }
 
+static void hardware_disable(void *junk)
+{
+       spin_lock(&kvm_lock);
+       hardware_disable_nolock(junk);
+       spin_unlock(&kvm_lock);
+}
+
+static void hardware_disable_all_nolock(void)
+{
+       BUG_ON(!kvm_usage_count);
+
+       kvm_usage_count--;
+       if (!kvm_usage_count)
+               on_each_cpu(hardware_disable_nolock, NULL, 1);
+}
+
+static void hardware_disable_all(void)
+{
+       spin_lock(&kvm_lock);
+       hardware_disable_all_nolock();
+       spin_unlock(&kvm_lock);
+}
+
+static int hardware_enable_all(void)
+{
+       int r = 0;
+
+       spin_lock(&kvm_lock);
+
+       kvm_usage_count++;
+       if (kvm_usage_count == 1) {
+               atomic_set(&hardware_enable_failed, 0);
+               on_each_cpu(hardware_enable_nolock, NULL, 1);
+
+               if (atomic_read(&hardware_enable_failed)) {
+                       hardware_disable_all_nolock();
+                       r = -EBUSY;
+               }
+       }
+
+       spin_unlock(&kvm_lock);
+
+       return r;
+}
+
 static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
                           void *v)
 {
        int cpu = (long)v;
 
+       if (!kvm_usage_count)
+               return NOTIFY_OK;
+
        val &= ~CPU_TASKS_FROZEN;
        switch (val) {
        case CPU_DYING:
@@ -2457,15 +2161,10 @@ static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
                       cpu);
                hardware_disable(NULL);
                break;
-       case CPU_UP_CANCELED:
-               printk(KERN_INFO "kvm: disabling virtualization on CPU%d\n",
-                      cpu);
-               smp_call_function_single(cpu, hardware_disable, NULL, 1);
-               break;
-       case CPU_ONLINE:
+       case CPU_STARTING:
                printk(KERN_INFO "kvm: enabling virtualization on CPU%d\n",
                       cpu);
-               smp_call_function_single(cpu, hardware_enable, NULL, 1);
+               hardware_enable(NULL);
                break;
        }
        return NOTIFY_OK;
@@ -2474,10 +2173,12 @@ static int kvm_cpu_hotplug(struct notifier_block *notifier, unsigned long val,
 
 asmlinkage void kvm_handle_fault_on_reboot(void)
 {
-       if (kvm_rebooting)
+       if (kvm_rebooting) {
                /* spin while reset goes on */
+               local_irq_enable();
                while (true)
-                       ;
+                       cpu_relax();
+       }
        /* Fault while not rebooting.  We want the trace. */
        BUG();
 }
@@ -2494,7 +2195,7 @@ static int kvm_reboot(struct notifier_block *notifier, unsigned long val,
         */
        printk(KERN_INFO "kvm: exiting hardware virtualization\n");
        kvm_rebooting = true;
-       on_each_cpu(hardware_disable, NULL, 1);
+       on_each_cpu(hardware_disable_nolock, NULL, 1);
        return NOTIFY_OK;
 }
 
@@ -2503,12 +2204,7 @@ static struct notifier_block kvm_reboot_notifier = {
        .priority = 0,
 };
 
-void kvm_io_bus_init(struct kvm_io_bus *bus)
-{
-       memset(bus, 0, sizeof(*bus));
-}
-
-void kvm_io_bus_destroy(struct kvm_io_bus *bus)
+static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
 {
        int i;
 
@@ -2517,13 +2213,17 @@ void kvm_io_bus_destroy(struct kvm_io_bus *bus)
 
                kvm_iodevice_destructor(pos);
        }
+       kfree(bus);
 }
 
 /* kvm_io_bus_write - called under kvm->slots_lock */
-int kvm_io_bus_write(struct kvm_io_bus *bus, gpa_t addr,
+int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
                     int len, const void *val)
 {
        int i;
+       struct kvm_io_bus *bus;
+
+       bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
        for (i = 0; i < bus->dev_count; i++)
                if (!kvm_iodevice_write(bus->devs[i], addr, len, val))
                        return 0;
@@ -2531,64 +2231,76 @@ int kvm_io_bus_write(struct kvm_io_bus *bus, gpa_t addr,
 }
 
 /* kvm_io_bus_read - called under kvm->slots_lock */
-int kvm_io_bus_read(struct kvm_io_bus *bus, gpa_t addr, int len, void *val)
+int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
+                   int len, void *val)
 {
        int i;
+       struct kvm_io_bus *bus;
+
+       bus = srcu_dereference(kvm->buses[bus_idx], &kvm->srcu);
        for (i = 0; i < bus->dev_count; i++)
                if (!kvm_iodevice_read(bus->devs[i], addr, len, val))
                        return 0;
        return -EOPNOTSUPP;
 }
 
-int kvm_io_bus_register_dev(struct kvm *kvm, struct kvm_io_bus *bus,
-                            struct kvm_io_device *dev)
+/* Caller must hold slots_lock. */
+int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                           struct kvm_io_device *dev)
 {
-       int ret;
-
-       down_write(&kvm->slots_lock);
-       ret = __kvm_io_bus_register_dev(bus, dev);
-       up_write(&kvm->slots_lock);
-
-       return ret;
-}
+       struct kvm_io_bus *new_bus, *bus;
 
-/* An unlocked version. Caller must have write lock on slots_lock. */
-int __kvm_io_bus_register_dev(struct kvm_io_bus *bus,
-                             struct kvm_io_device *dev)
-{
+       bus = kvm->buses[bus_idx];
        if (bus->dev_count > NR_IOBUS_DEVS-1)
                return -ENOSPC;
 
-       bus->devs[bus->dev_count++] = dev;
+       new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL);
+       if (!new_bus)
+               return -ENOMEM;
+       memcpy(new_bus, bus, sizeof(struct kvm_io_bus));
+       new_bus->devs[new_bus->dev_count++] = dev;
+       rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
+       synchronize_srcu_expedited(&kvm->srcu);
+       kfree(bus);
 
        return 0;
 }
 
-void kvm_io_bus_unregister_dev(struct kvm *kvm,
-                              struct kvm_io_bus *bus,
-                              struct kvm_io_device *dev)
+/* Caller must hold slots_lock. */
+int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
+                             struct kvm_io_device *dev)
 {
-       down_write(&kvm->slots_lock);
-       __kvm_io_bus_unregister_dev(bus, dev);
-       up_write(&kvm->slots_lock);
-}
+       int i, r;
+       struct kvm_io_bus *new_bus, *bus;
 
-/* An unlocked version. Caller must have write lock on slots_lock. */
-void __kvm_io_bus_unregister_dev(struct kvm_io_bus *bus,
-                                struct kvm_io_device *dev)
-{
-       int i;
+       new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL);
+       if (!new_bus)
+               return -ENOMEM;
 
-       for (i = 0; i < bus->dev_count; i++)
-               if (bus->devs[i] == dev) {
-                       bus->devs[i] = bus->devs[--bus->dev_count];
+       bus = kvm->buses[bus_idx];
+       memcpy(new_bus, bus, sizeof(struct kvm_io_bus));
+
+       r = -ENOENT;
+       for (i = 0; i < new_bus->dev_count; i++)
+               if (new_bus->devs[i] == dev) {
+                       r = 0;
+                       new_bus->devs[i] = new_bus->devs[--new_bus->dev_count];
                        break;
                }
+
+       if (r) {
+               kfree(new_bus);
+               return r;
+       }
+
+       rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
+       synchronize_srcu_expedited(&kvm->srcu);
+       kfree(bus);
+       return r;
 }
 
 static struct notifier_block kvm_cpu_notifier = {
        .notifier_call = kvm_cpu_hotplug,
-       .priority = 20, /* must be > scheduler priority */
 };
 
 static int vm_stat_get(void *_offset, u64 *val)
@@ -2625,7 +2337,7 @@ static int vcpu_stat_get(void *_offset, u64 *val)
 
 DEFINE_SIMPLE_ATTRIBUTE(vcpu_stat_fops, vcpu_stat_get, NULL, "%llu\n");
 
-static struct file_operations *stat_fops[] = {
+static const struct file_operations *stat_fops[] = {
        [KVM_STAT_VCPU] = &vcpu_stat_fops,
        [KVM_STAT_VM]   = &vm_stat_fops,
 };
@@ -2652,13 +2364,17 @@ static void kvm_exit_debug(void)
 
 static int kvm_suspend(struct sys_device *dev, pm_message_t state)
 {
-       hardware_disable(NULL);
+       if (kvm_usage_count)
+               hardware_disable_nolock(NULL);
        return 0;
 }
 
 static int kvm_resume(struct sys_device *dev)
 {
-       hardware_enable(NULL);
+       if (kvm_usage_count) {
+               WARN_ON(spin_is_locked(&kvm_lock));
+               hardware_enable_nolock(NULL);
+       }
        return 0;
 }
 
@@ -2697,14 +2413,12 @@ static void kvm_sched_out(struct preempt_notifier *pn,
        kvm_arch_vcpu_put(vcpu);
 }
 
-int kvm_init(void *opaque, unsigned int vcpu_size,
+int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
                  struct module *module)
 {
        int r;
        int cpu;
 
-       kvm_init_debug();
-
        r = kvm_arch_init(opaque);
        if (r)
                goto out_fail;
@@ -2718,6 +2432,24 @@ int kvm_init(void *opaque, unsigned int vcpu_size,
 
        bad_pfn = page_to_pfn(bad_page);
 
+       hwpoison_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+
+       if (hwpoison_page == NULL) {
+               r = -ENOMEM;
+               goto out_free_0;
+       }
+
+       hwpoison_pfn = page_to_pfn(hwpoison_page);
+
+       fault_page = alloc_page(GFP_KERNEL | __GFP_ZERO);
+
+       if (fault_page == NULL) {
+               r = -ENOMEM;
+               goto out_free_0;
+       }
+
+       fault_pfn = page_to_pfn(fault_page);
+
        if (!zalloc_cpumask_var(&cpus_hardware_enabled, GFP_KERNEL)) {
                r = -ENOMEM;
                goto out_free_0;
@@ -2735,7 +2467,6 @@ int kvm_init(void *opaque, unsigned int vcpu_size,
                        goto out_free_1;
        }
 
-       on_each_cpu(hardware_enable, NULL, 1);
        r = register_cpu_notifier(&kvm_cpu_notifier);
        if (r)
                goto out_free_2;
@@ -2750,14 +2481,19 @@ int kvm_init(void *opaque, unsigned int vcpu_size,
                goto out_free_4;
 
        /* A kmem cache lets us meet the alignment requirements of fx_save. */
-       kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size,
-                                          __alignof__(struct kvm_vcpu),
+       if (!vcpu_align)
+               vcpu_align = __alignof__(struct kvm_vcpu);
+       kvm_vcpu_cache = kmem_cache_create("kvm_vcpu", vcpu_size, vcpu_align,
                                           0, NULL);
        if (!kvm_vcpu_cache) {
                r = -ENOMEM;
                goto out_free_5;
        }
 
+       r = kvm_async_pf_init();
+       if (r)
+               goto out_free;
+
        kvm_chardev_ops.owner = module;
        kvm_vm_fops.owner = module;
        kvm_vcpu_fops.owner = module;
@@ -2765,14 +2501,18 @@ int kvm_init(void *opaque, unsigned int vcpu_size,
        r = misc_register(&kvm_dev);
        if (r) {
                printk(KERN_ERR "kvm: misc device register failed\n");
-               goto out_free;
+               goto out_unreg;
        }
 
        kvm_preempt_ops.sched_in = kvm_sched_in;
        kvm_preempt_ops.sched_out = kvm_sched_out;
 
+       kvm_init_debug();
+
        return 0;
 
+out_unreg:
+       kvm_async_pf_deinit();
 out_free:
        kmem_cache_destroy(kvm_vcpu_cache);
 out_free_5:
@@ -2783,35 +2523,38 @@ out_free_3:
        unregister_reboot_notifier(&kvm_reboot_notifier);
        unregister_cpu_notifier(&kvm_cpu_notifier);
 out_free_2:
-       on_each_cpu(hardware_disable, NULL, 1);
 out_free_1:
        kvm_arch_hardware_unsetup();
 out_free_0a:
        free_cpumask_var(cpus_hardware_enabled);
 out_free_0:
+       if (fault_page)
+               __free_page(fault_page);
+       if (hwpoison_page)
+               __free_page(hwpoison_page);
        __free_page(bad_page);
 out:
        kvm_arch_exit();
 out_fail:
-       kvm_exit_debug();
        return r;
 }
 EXPORT_SYMBOL_GPL(kvm_init);
 
 void kvm_exit(void)
 {
-       tracepoint_synchronize_unregister();
+       kvm_exit_debug();
        misc_deregister(&kvm_dev);
        kmem_cache_destroy(kvm_vcpu_cache);
+       kvm_async_pf_deinit();
        sysdev_unregister(&kvm_sysdev);
        sysdev_class_unregister(&kvm_sysdev_class);
        unregister_reboot_notifier(&kvm_reboot_notifier);
        unregister_cpu_notifier(&kvm_cpu_notifier);
-       on_each_cpu(hardware_disable, NULL, 1);
+       on_each_cpu(hardware_disable_nolock, NULL, 1);
        kvm_arch_hardware_unsetup();
        kvm_arch_exit();
-       kvm_exit_debug();
        free_cpumask_var(cpus_hardware_enabled);
+       __free_page(hwpoison_page);
        __free_page(bad_page);
 }
 EXPORT_SYMBOL_GPL(kvm_exit);