KVM: Use kmemdup() instead of kmalloc/memcpy
[linux-2.6.git] / virt / kvm / kvm_main.c
index f29abeb..a6e612f 100644 (file)
@@ -30,7 +30,7 @@
 #include <linux/debugfs.h>
 #include <linux/highmem.h>
 #include <linux/file.h>
-#include <linux/sysdev.h>
+#include <linux/syscore_ops.h>
 #include <linux/cpu.h>
 #include <linux/sched.h>
 #include <linux/cpumask.h>
 #include <linux/srcu.h>
 #include <linux/hugetlb.h>
 #include <linux/slab.h>
+#include <linux/sort.h>
+#include <linux/bsearch.h>
 
 #include <asm/processor.h>
 #include <asm/io.h>
 #include <asm/uaccess.h>
 #include <asm/pgtable.h>
-#include <asm-generic/bitops/le.h>
 
 #include "coalesced_mmio.h"
 #include "async_pf.h"
@@ -69,7 +70,7 @@ MODULE_LICENSE("GPL");
  *             kvm->lock --> kvm->slots_lock --> kvm->irq_lock
  */
 
-DEFINE_SPINLOCK(kvm_lock);
+DEFINE_RAW_SPINLOCK(kvm_lock);
 LIST_HEAD(vm_list);
 
 static cpumask_var_t cpus_hardware_enabled;
@@ -85,6 +86,10 @@ struct dentry *kvm_debugfs_dir;
 
 static long kvm_vcpu_ioctl(struct file *file, unsigned int ioctl,
                           unsigned long arg);
+#ifdef CONFIG_COMPAT
+static long kvm_vcpu_compat_ioctl(struct file *file, unsigned int ioctl,
+                                 unsigned long arg);
+#endif
 static int hardware_enable_all(void);
 static void hardware_disable_all(void);
 
@@ -98,8 +103,8 @@ static bool largepages_enabled = true;
 static struct page *hwpoison_page;
 static pfn_t hwpoison_pfn;
 
-static struct page *fault_page;
-static pfn_t fault_pfn;
+struct page *fault_page;
+pfn_t fault_pfn;
 
 inline int kvm_is_mmio_pfn(pfn_t pfn)
 {
@@ -137,6 +142,14 @@ void vcpu_load(struct kvm_vcpu *vcpu)
        int cpu;
 
        mutex_lock(&vcpu->mutex);
+       if (unlikely(vcpu->pid != current->pids[PIDTYPE_PID].pid)) {
+               /* The thread running this VCPU changed. */
+               struct pid *oldpid = vcpu->pid;
+               struct pid *newpid = get_task_pid(current, PIDTYPE_PID);
+               rcu_assign_pointer(vcpu->pid, newpid);
+               synchronize_rcu();
+               put_pid(oldpid);
+       }
        cpu = get_cpu();
        preempt_notifier_register(&vcpu->preempt_notifier);
        kvm_arch_vcpu_load(vcpu, cpu);
@@ -165,13 +178,16 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
 
        zalloc_cpumask_var(&cpus, GFP_ATOMIC);
 
-       raw_spin_lock(&kvm->requests_lock);
-       me = smp_processor_id();
+       me = get_cpu();
        kvm_for_each_vcpu(i, vcpu, kvm) {
-               if (kvm_make_check_request(req, vcpu))
-                       continue;
+               kvm_make_request(req, vcpu);
                cpu = vcpu->cpu;
-               if (cpus != NULL && cpu != -1 && cpu != me)
+
+               /* Set ->requests bit before we read ->mode */
+               smp_mb();
+
+               if (cpus != NULL && cpu != -1 && cpu != me &&
+                     kvm_vcpu_exiting_guest_mode(vcpu) != OUTSIDE_GUEST_MODE)
                        cpumask_set_cpu(cpu, cpus);
        }
        if (unlikely(cpus == NULL))
@@ -180,7 +196,7 @@ static bool make_all_cpus_request(struct kvm *kvm, unsigned int req)
                smp_call_function_many(cpus, ack_flush, NULL, 1);
        else
                called = false;
-       raw_spin_unlock(&kvm->requests_lock);
+       put_cpu();
        free_cpumask_var(cpus);
        return called;
 }
@@ -209,6 +225,7 @@ int kvm_vcpu_init(struct kvm_vcpu *vcpu, struct kvm *kvm, unsigned id)
        vcpu->cpu = -1;
        vcpu->kvm = kvm;
        vcpu->vcpu_id = id;
+       vcpu->pid = NULL;
        init_waitqueue_head(&vcpu->wq);
        kvm_async_pf_vcpu_init(vcpu);
 
@@ -233,6 +250,7 @@ EXPORT_SYMBOL_GPL(kvm_vcpu_init);
 
 void kvm_vcpu_uninit(struct kvm_vcpu *vcpu)
 {
+       put_pid(vcpu->pid);
        kvm_arch_vcpu_uninit(vcpu);
        free_page((unsigned long)vcpu->run);
 }
@@ -422,6 +440,15 @@ static int kvm_init_mmu_notifier(struct kvm *kvm)
 
 #endif /* CONFIG_MMU_NOTIFIER && KVM_ARCH_WANT_MMU_NOTIFIER */
 
+static void kvm_init_memslots_id(struct kvm *kvm)
+{
+       int i;
+       struct kvm_memslots *slots = kvm->memslots;
+
+       for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
+               slots->id_to_index[i] = slots->memslots[i].id = i;
+}
+
 static struct kvm *kvm_create_vm(void)
 {
        int r, i;
@@ -447,6 +474,7 @@ static struct kvm *kvm_create_vm(void)
        kvm->memslots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
        if (!kvm->memslots)
                goto out_err_nosrcu;
+       kvm_init_memslots_id(kvm);
        if (init_srcu_struct(&kvm->srcu))
                goto out_err_nosrcu;
        for (i = 0; i < KVM_NR_BUSES; i++) {
@@ -456,22 +484,22 @@ static struct kvm *kvm_create_vm(void)
                        goto out_err;
        }
 
-       r = kvm_init_mmu_notifier(kvm);
-       if (r)
-               goto out_err;
-
+       spin_lock_init(&kvm->mmu_lock);
        kvm->mm = current->mm;
        atomic_inc(&kvm->mm->mm_count);
-       spin_lock_init(&kvm->mmu_lock);
-       raw_spin_lock_init(&kvm->requests_lock);
        kvm_eventfd_init(kvm);
        mutex_init(&kvm->lock);
        mutex_init(&kvm->irq_lock);
        mutex_init(&kvm->slots_lock);
        atomic_set(&kvm->users_count, 1);
-       spin_lock(&kvm_lock);
+
+       r = kvm_init_mmu_notifier(kvm);
+       if (r)
+               goto out_err;
+
+       raw_spin_lock(&kvm_lock);
        list_add(&kvm->vm_list, &vm_list);
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
 
        return kvm;
 
@@ -529,11 +557,11 @@ static void kvm_free_physmem_slot(struct kvm_memory_slot *free,
 
 void kvm_free_physmem(struct kvm *kvm)
 {
-       int i;
        struct kvm_memslots *slots = kvm->memslots;
+       struct kvm_memory_slot *memslot;
 
-       for (i = 0; i < slots->nmemslots; ++i)
-               kvm_free_physmem_slot(&slots->memslots[i], NULL);
+       kvm_for_each_memslot(memslot, slots)
+               kvm_free_physmem_slot(memslot, NULL);
 
        kfree(kvm->memslots);
 }
@@ -544,9 +572,9 @@ static void kvm_destroy_vm(struct kvm *kvm)
        struct mm_struct *mm = kvm->mm;
 
        kvm_arch_sync_events(kvm);
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        list_del(&kvm->vm_list);
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
        kvm_free_irq_routing(kvm);
        for (i = 0; i < KVM_NR_BUSES; i++)
                kvm_io_bus_destroy(kvm->buses[i]);
@@ -588,6 +616,7 @@ static int kvm_vm_release(struct inode *inode, struct file *filp)
        return 0;
 }
 
+#ifndef CONFIG_S390
 /*
  * Allocation size is twice as large as the actual dirty bitmap size.
  * This makes it possible to do double buffering: see x86's
@@ -606,10 +635,70 @@ static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
                return -ENOMEM;
 
        memslot->dirty_bitmap_head = memslot->dirty_bitmap;
+       memslot->nr_dirty_pages = 0;
+       return 0;
+}
+#endif /* !CONFIG_S390 */
+
+static struct kvm_memory_slot *
+search_memslots(struct kvm_memslots *slots, gfn_t gfn)
+{
+       struct kvm_memory_slot *memslot;
+
+       kvm_for_each_memslot(memslot, slots)
+               if (gfn >= memslot->base_gfn &&
+                     gfn < memslot->base_gfn + memslot->npages)
+                       return memslot;
+
+       return NULL;
+}
+
+static int cmp_memslot(const void *slot1, const void *slot2)
+{
+       struct kvm_memory_slot *s1, *s2;
+
+       s1 = (struct kvm_memory_slot *)slot1;
+       s2 = (struct kvm_memory_slot *)slot2;
+
+       if (s1->npages < s2->npages)
+               return 1;
+       if (s1->npages > s2->npages)
+               return -1;
+
        return 0;
 }
 
 /*
+ * Sort the memslots base on its size, so the larger slots
+ * will get better fit.
+ */
+static void sort_memslots(struct kvm_memslots *slots)
+{
+       int i;
+
+       sort(slots->memslots, KVM_MEM_SLOTS_NUM,
+             sizeof(struct kvm_memory_slot), cmp_memslot, NULL);
+
+       for (i = 0; i < KVM_MEM_SLOTS_NUM; i++)
+               slots->id_to_index[slots->memslots[i].id] = i;
+}
+
+void update_memslots(struct kvm_memslots *slots, struct kvm_memory_slot *new)
+{
+       if (new) {
+               int id = new->id;
+               struct kvm_memory_slot *old = id_to_memslot(slots, id);
+               unsigned long npages = old->npages;
+
+               *old = *new;
+               if (new->npages != npages)
+                       sort_memslots(slots);
+       }
+
+       slots->generation++;
+}
+
+/*
  * Allocate some memory and give it an address in the guest physical address
  * space.
  *
@@ -621,7 +710,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
                            struct kvm_userspace_memory_region *mem,
                            int user_alloc)
 {
-       int r, flush_shadow = 0;
+       int r;
        gfn_t base_gfn;
        unsigned long npages;
        unsigned long i;
@@ -635,14 +724,19 @@ int __kvm_set_memory_region(struct kvm *kvm,
                goto out;
        if (mem->guest_phys_addr & (PAGE_SIZE - 1))
                goto out;
-       if (user_alloc && (mem->userspace_addr & (PAGE_SIZE - 1)))
+       /* We can read the guest memory with __xxx_user() later on. */
+       if (user_alloc &&
+           ((mem->userspace_addr & (PAGE_SIZE - 1)) ||
+            !access_ok(VERIFY_WRITE,
+                       (void __user *)(unsigned long)mem->userspace_addr,
+                       mem->memory_size)))
                goto out;
-       if (mem->slot >= KVM_MEMORY_SLOTS + KVM_PRIVATE_MEM_SLOTS)
+       if (mem->slot >= KVM_MEM_SLOTS_NUM)
                goto out;
        if (mem->guest_phys_addr + mem->memory_size < mem->guest_phys_addr)
                goto out;
 
-       memslot = &kvm->memslots->memslots[mem->slot];
+       memslot = id_to_memslot(kvm->memslots, mem->slot);
        base_gfn = mem->guest_phys_addr >> PAGE_SHIFT;
        npages = mem->memory_size >> PAGE_SHIFT;
 
@@ -741,8 +835,6 @@ skip_lpage:
                if (kvm_create_dirty_bitmap(&new) < 0)
                        goto out_free;
                /* destroy any largepage mappings for dirty tracking */
-               if (old.npages)
-                       flush_shadow = 1;
        }
 #else  /* not defined CONFIG_S390 */
        new.user_alloc = user_alloc;
@@ -751,15 +843,17 @@ skip_lpage:
 #endif /* not defined CONFIG_S390 */
 
        if (!npages) {
+               struct kvm_memory_slot *slot;
+
                r = -ENOMEM;
-               slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
+               slots = kmemdup(kvm->memslots, 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;
+               slot = id_to_memslot(slots, mem->slot);
+               slot->flags |= KVM_MEMSLOT_INVALID;
+
+               update_memslots(slots, NULL);
 
                old_memslots = kvm->memslots;
                rcu_assign_pointer(kvm->memslots, slots);
@@ -787,13 +881,10 @@ skip_lpage:
        }
 
        r = -ENOMEM;
-       slots = kzalloc(sizeof(struct kvm_memslots), GFP_KERNEL);
+       slots = kmemdup(kvm->memslots, 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) {
@@ -803,19 +894,23 @@ skip_lpage:
                        new.lpage_info[i] = NULL;
        }
 
-       slots->memslots[mem->slot] = new;
+       update_memslots(slots, &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);
 
+       /*
+        * If the new memory slot is created, we need to clear all
+        * mmio sptes.
+        */
+       if (npages && old.base_gfn != mem->guest_phys_addr >> PAGE_SHIFT)
+               kvm_arch_flush_shadow(kvm);
+
        kvm_free_physmem_slot(&old, &new);
        kfree(old_memslots);
 
-       if (flush_shadow)
-               kvm_arch_flush_shadow(kvm);
-
        return 0;
 
 out_free:
@@ -861,7 +956,7 @@ int kvm_get_dirty_log(struct kvm *kvm,
        if (log->slot >= KVM_MEMORY_SLOTS)
                goto out;
 
-       memslot = &kvm->memslots->memslots[log->slot];
+       memslot = id_to_memslot(kvm->memslots, log->slot);
        r = -ENOENT;
        if (!memslot->dirty_bitmap)
                goto out;
@@ -913,6 +1008,18 @@ int is_fault_pfn(pfn_t pfn)
 }
 EXPORT_SYMBOL_GPL(is_fault_pfn);
 
+int is_noslot_pfn(pfn_t pfn)
+{
+       return pfn == bad_pfn;
+}
+EXPORT_SYMBOL_GPL(is_noslot_pfn);
+
+int is_invalid_pfn(pfn_t pfn)
+{
+       return pfn == hwpoison_pfn || pfn == fault_pfn;
+}
+EXPORT_SYMBOL_GPL(is_invalid_pfn);
+
 static inline unsigned long bad_hva(void)
 {
        return PAGE_OFFSET;
@@ -927,16 +1034,7 @@ EXPORT_SYMBOL_GPL(kvm_is_error_hva);
 static struct kvm_memory_slot *__gfn_to_memslot(struct kvm_memslots *slots,
                                                gfn_t gfn)
 {
-       int 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)
-                       return memslot;
-       }
-       return NULL;
+       return search_memslots(slots, gfn);
 }
 
 struct kvm_memory_slot *gfn_to_memslot(struct kvm *kvm, gfn_t gfn)
@@ -947,20 +1045,13 @@ 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);
+       struct kvm_memory_slot *memslot = gfn_to_memslot(kvm, gfn);
 
-       for (i = 0; i < KVM_MEMORY_SLOTS; ++i) {
-               struct kvm_memory_slot *memslot = &slots->memslots[i];
-
-               if (memslot->flags & KVM_MEMSLOT_INVALID)
-                       continue;
+       if (!memslot || memslot->id >= KVM_MEMORY_SLOTS ||
+             memslot->flags & KVM_MEMSLOT_INVALID)
+               return 0;
 
-               if (gfn >= memslot->base_gfn
-                   && gfn < memslot->base_gfn + memslot->npages)
-                       return 1;
-       }
-       return 0;
+       return 1;
 }
 EXPORT_SYMBOL_GPL(kvm_is_visible_gfn);
 
@@ -988,23 +1079,6 @@ out:
        return size;
 }
 
-int memslot_id(struct kvm *kvm, gfn_t gfn)
-{
-       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;
-       }
-
-       return memslot - slots->memslots;
-}
-
 static unsigned long gfn_to_hva_many(struct kvm_memory_slot *slot, gfn_t gfn,
                                     gfn_t *nr_pages)
 {
@@ -1029,6 +1103,26 @@ static pfn_t get_fault_pfn(void)
        return fault_pfn;
 }
 
+int get_user_page_nowait(struct task_struct *tsk, struct mm_struct *mm,
+       unsigned long start, int write, struct page **page)
+{
+       int flags = FOLL_TOUCH | FOLL_NOWAIT | FOLL_HWPOISON | FOLL_GET;
+
+       if (write)
+               flags |= FOLL_WRITE;
+
+       return __get_user_pages(tsk, mm, start, 1, flags, page, NULL, NULL);
+}
+
+static inline int check_user_page_hwpoison(unsigned long addr)
+{
+       int rc, flags = FOLL_TOUCH | FOLL_HWPOISON | FOLL_WRITE;
+
+       rc = __get_user_pages(current, current->mm, addr, 1,
+                             flags, NULL, NULL, NULL);
+       return rc == -EHWPOISON;
+}
+
 static pfn_t hva_to_pfn(struct kvm *kvm, unsigned long addr, bool atomic,
                        bool *async, bool write_fault, bool *writable)
 {
@@ -1053,7 +1147,14 @@ static pfn_t hva_to_pfn(struct kvm *kvm, unsigned long addr, bool atomic,
                if (writable)
                        *writable = write_fault;
 
-               npages = get_user_pages_fast(addr, 1, write_fault, page);
+               if (async) {
+                       down_read(&current->mm->mmap_sem);
+                       npages = get_user_page_nowait(current, current->mm,
+                                                    addr, write_fault, page);
+                       up_read(&current->mm->mmap_sem);
+               } else
+                       npages = get_user_pages_fast(addr, 1, write_fault,
+                                                    page);
 
                /* map read fault as writable if possible */
                if (unlikely(!write_fault) && npages == 1) {
@@ -1076,7 +1177,8 @@ static pfn_t hva_to_pfn(struct kvm *kvm, unsigned long addr, bool atomic,
                        return get_fault_pfn();
 
                down_read(&current->mm->mmap_sem);
-               if (is_hwpoison_address(addr)) {
+               if (npages == -EHWPOISON ||
+                       (!async && check_user_page_hwpoison(addr))) {
                        up_read(&current->mm->mmap_sem);
                        get_page(hwpoison_page);
                        return page_to_pfn(hwpoison_page);
@@ -1264,7 +1366,7 @@ int kvm_read_guest_page(struct kvm *kvm, gfn_t gfn, void *data, int offset,
        addr = gfn_to_hva(kvm, gfn);
        if (kvm_is_error_hva(addr))
                return -EFAULT;
-       r = copy_from_user(data, (void __user *)addr + offset, len);
+       r = __copy_from_user(data, (void __user *)addr + offset, len);
        if (r)
                return -EFAULT;
        return 0;
@@ -1320,7 +1422,7 @@ int kvm_write_guest_page(struct kvm *kvm, gfn_t gfn, const void *data,
        addr = gfn_to_hva(kvm, gfn);
        if (kvm_is_error_hva(addr))
                return -EFAULT;
-       r = copy_to_user((void __user *)addr + offset, data, len);
+       r = __copy_to_user((void __user *)addr + offset, data, len);
        if (r)
                return -EFAULT;
        mark_page_dirty(kvm, gfn);
@@ -1380,7 +1482,7 @@ int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
        if (kvm_is_error_hva(ghc->hva))
                return -EFAULT;
 
-       r = copy_to_user((void __user *)ghc->hva, data, len);
+       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);
@@ -1389,6 +1491,26 @@ int kvm_write_guest_cached(struct kvm *kvm, struct gfn_to_hva_cache *ghc,
 }
 EXPORT_SYMBOL_GPL(kvm_write_guest_cached);
 
+int kvm_read_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_from_user(data, (void __user *)ghc->hva, len);
+       if (r)
+               return -EFAULT;
+
+       return 0;
+}
+EXPORT_SYMBOL_GPL(kvm_read_guest_cached);
+
 int kvm_clear_guest_page(struct kvm *kvm, gfn_t gfn, int offset, int len)
 {
        return kvm_write_guest_page(kvm, gfn, (const void *) empty_zero_page,
@@ -1421,7 +1543,8 @@ void mark_page_dirty_in_slot(struct kvm *kvm, struct kvm_memory_slot *memslot,
        if (memslot && memslot->dirty_bitmap) {
                unsigned long rel_gfn = gfn - memslot->base_gfn;
 
-               generic___set_le_bit(rel_gfn, memslot->dirty_bitmap);
+               if (!__test_and_set_bit_le(rel_gfn, memslot->dirty_bitmap))
+                       memslot->nr_dirty_pages++;
        }
 }
 
@@ -1466,18 +1589,55 @@ void kvm_resched(struct kvm_vcpu *vcpu)
 }
 EXPORT_SYMBOL_GPL(kvm_resched);
 
-void kvm_vcpu_on_spin(struct kvm_vcpu *vcpu)
+void kvm_vcpu_on_spin(struct kvm_vcpu *me)
 {
-       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);
+       struct kvm *kvm = me->kvm;
+       struct kvm_vcpu *vcpu;
+       int last_boosted_vcpu = me->kvm->last_boosted_vcpu;
+       int yielded = 0;
+       int pass;
+       int i;
 
-       finish_wait(&vcpu->wq, &wait);
+       /*
+        * We boost the priority of a VCPU that is runnable but not
+        * currently running, because it got preempted by something
+        * else and called schedule in __vcpu_run.  Hopefully that
+        * VCPU is holding the lock that we need and will release it.
+        * We approximate round-robin by starting at the last boosted VCPU.
+        */
+       for (pass = 0; pass < 2 && !yielded; pass++) {
+               kvm_for_each_vcpu(i, vcpu, kvm) {
+                       struct task_struct *task = NULL;
+                       struct pid *pid;
+                       if (!pass && i < last_boosted_vcpu) {
+                               i = last_boosted_vcpu;
+                               continue;
+                       } else if (pass && i > last_boosted_vcpu)
+                               break;
+                       if (vcpu == me)
+                               continue;
+                       if (waitqueue_active(&vcpu->wq))
+                               continue;
+                       rcu_read_lock();
+                       pid = rcu_dereference(vcpu->pid);
+                       if (pid)
+                               task = get_pid_task(vcpu->pid, PIDTYPE_PID);
+                       rcu_read_unlock();
+                       if (!task)
+                               continue;
+                       if (task->flags & PF_VCPU) {
+                               put_task_struct(task);
+                               continue;
+                       }
+                       if (yield_to(task, 1)) {
+                               put_task_struct(task);
+                               kvm->last_boosted_vcpu = i;
+                               yielded = 1;
+                               break;
+                       }
+                       put_task_struct(task);
+               }
+       }
 }
 EXPORT_SYMBOL_GPL(kvm_vcpu_on_spin);
 
@@ -1524,7 +1684,9 @@ static int kvm_vcpu_release(struct inode *inode, struct file *filp)
 static struct file_operations kvm_vcpu_fops = {
        .release        = kvm_vcpu_release,
        .unlocked_ioctl = kvm_vcpu_ioctl,
-       .compat_ioctl   = kvm_vcpu_ioctl,
+#ifdef CONFIG_COMPAT
+       .compat_ioctl   = kvm_vcpu_compat_ioctl,
+#endif
        .mmap           = kvm_vcpu_mmap,
        .llseek         = noop_llseek,
 };
@@ -1553,18 +1715,18 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
 
        r = kvm_arch_vcpu_setup(vcpu);
        if (r)
-               return r;
+               goto vcpu_destroy;
 
        mutex_lock(&kvm->lock);
        if (atomic_read(&kvm->online_vcpus) == KVM_MAX_VCPUS) {
                r = -EINVAL;
-               goto vcpu_destroy;
+               goto unlock_vcpu_destroy;
        }
 
        kvm_for_each_vcpu(r, v, kvm)
                if (v->vcpu_id == id) {
                        r = -EEXIST;
-                       goto vcpu_destroy;
+                       goto unlock_vcpu_destroy;
                }
 
        BUG_ON(kvm->vcpus[atomic_read(&kvm->online_vcpus)]);
@@ -1574,7 +1736,7 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
        r = create_vcpu_fd(vcpu);
        if (r < 0) {
                kvm_put_kvm(kvm);
-               goto vcpu_destroy;
+               goto unlock_vcpu_destroy;
        }
 
        kvm->vcpus[atomic_read(&kvm->online_vcpus)] = vcpu;
@@ -1588,8 +1750,9 @@ static int kvm_vm_ioctl_create_vcpu(struct kvm *kvm, u32 id)
        mutex_unlock(&kvm->lock);
        return r;
 
-vcpu_destroy:
+unlock_vcpu_destroy:
        mutex_unlock(&kvm->lock);
+vcpu_destroy:
        kvm_arch_vcpu_destroy(vcpu);
        return r;
 }
@@ -1812,6 +1975,50 @@ out:
        return r;
 }
 
+#ifdef CONFIG_COMPAT
+static long kvm_vcpu_compat_ioctl(struct file *filp,
+                                 unsigned int ioctl, unsigned long arg)
+{
+       struct kvm_vcpu *vcpu = filp->private_data;
+       void __user *argp = compat_ptr(arg);
+       int r;
+
+       if (vcpu->kvm->mm != current->mm)
+               return -EIO;
+
+       switch (ioctl) {
+       case KVM_SET_SIGNAL_MASK: {
+               struct kvm_signal_mask __user *sigmask_arg = argp;
+               struct kvm_signal_mask kvm_sigmask;
+               compat_sigset_t csigset;
+               sigset_t sigset;
+
+               if (argp) {
+                       r = -EFAULT;
+                       if (copy_from_user(&kvm_sigmask, argp,
+                                          sizeof kvm_sigmask))
+                               goto out;
+                       r = -EINVAL;
+                       if (kvm_sigmask.len != sizeof csigset)
+                               goto out;
+                       r = -EFAULT;
+                       if (copy_from_user(&csigset, sigmask_arg->sigset,
+                                          sizeof csigset))
+                               goto out;
+               }
+               sigset_from_compat(&sigset, &csigset);
+               r = kvm_vcpu_ioctl_set_sigmask(vcpu, &sigset);
+               break;
+       }
+       default:
+               r = kvm_vcpu_ioctl(filp, ioctl, arg);
+       }
+
+out:
+       return r;
+}
+#endif
+
 static long kvm_vm_ioctl(struct file *filp,
                           unsigned int ioctl, unsigned long arg)
 {
@@ -2122,9 +2329,9 @@ static void hardware_enable_nolock(void *junk)
 
 static void hardware_enable(void *junk)
 {
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        hardware_enable_nolock(junk);
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
 }
 
 static void hardware_disable_nolock(void *junk)
@@ -2139,9 +2346,9 @@ static void hardware_disable_nolock(void *junk)
 
 static void hardware_disable(void *junk)
 {
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        hardware_disable_nolock(junk);
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
 }
 
 static void hardware_disable_all_nolock(void)
@@ -2155,16 +2362,16 @@ static void hardware_disable_all_nolock(void)
 
 static void hardware_disable_all(void)
 {
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        hardware_disable_all_nolock();
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
 }
 
 static int hardware_enable_all(void)
 {
        int r = 0;
 
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
 
        kvm_usage_count++;
        if (kvm_usage_count == 1) {
@@ -2177,7 +2384,7 @@ static int hardware_enable_all(void)
                }
        }
 
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
 
        return r;
 }
@@ -2239,24 +2446,92 @@ static void kvm_io_bus_destroy(struct kvm_io_bus *bus)
        int i;
 
        for (i = 0; i < bus->dev_count; i++) {
-               struct kvm_io_device *pos = bus->devs[i];
+               struct kvm_io_device *pos = bus->range[i].dev;
 
                kvm_iodevice_destructor(pos);
        }
        kfree(bus);
 }
 
+int kvm_io_bus_sort_cmp(const void *p1, const void *p2)
+{
+       const struct kvm_io_range *r1 = p1;
+       const struct kvm_io_range *r2 = p2;
+
+       if (r1->addr < r2->addr)
+               return -1;
+       if (r1->addr + r1->len > r2->addr + r2->len)
+               return 1;
+       return 0;
+}
+
+int kvm_io_bus_insert_dev(struct kvm_io_bus *bus, struct kvm_io_device *dev,
+                         gpa_t addr, int len)
+{
+       if (bus->dev_count == NR_IOBUS_DEVS)
+               return -ENOSPC;
+
+       bus->range[bus->dev_count++] = (struct kvm_io_range) {
+               .addr = addr,
+               .len = len,
+               .dev = dev,
+       };
+
+       sort(bus->range, bus->dev_count, sizeof(struct kvm_io_range),
+               kvm_io_bus_sort_cmp, NULL);
+
+       return 0;
+}
+
+int kvm_io_bus_get_first_dev(struct kvm_io_bus *bus,
+                            gpa_t addr, int len)
+{
+       struct kvm_io_range *range, key;
+       int off;
+
+       key = (struct kvm_io_range) {
+               .addr = addr,
+               .len = len,
+       };
+
+       range = bsearch(&key, bus->range, bus->dev_count,
+                       sizeof(struct kvm_io_range), kvm_io_bus_sort_cmp);
+       if (range == NULL)
+               return -ENOENT;
+
+       off = range - bus->range;
+
+       while (off > 0 && kvm_io_bus_sort_cmp(&key, &bus->range[off-1]) == 0)
+               off--;
+
+       return off;
+}
+
 /* kvm_io_bus_write - called under kvm->slots_lock */
 int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
                     int len, const void *val)
 {
-       int i;
+       int idx;
        struct kvm_io_bus *bus;
+       struct kvm_io_range range;
+
+       range = (struct kvm_io_range) {
+               .addr = addr,
+               .len = len,
+       };
 
        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))
+       idx = kvm_io_bus_get_first_dev(bus, addr, len);
+       if (idx < 0)
+               return -EOPNOTSUPP;
+
+       while (idx < bus->dev_count &&
+               kvm_io_bus_sort_cmp(&range, &bus->range[idx]) == 0) {
+               if (!kvm_iodevice_write(bus->range[idx].dev, addr, len, val))
                        return 0;
+               idx++;
+       }
+
        return -EOPNOTSUPP;
 }
 
@@ -2264,19 +2539,33 @@ int kvm_io_bus_write(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
 int kvm_io_bus_read(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
                    int len, void *val)
 {
-       int i;
+       int idx;
        struct kvm_io_bus *bus;
+       struct kvm_io_range range;
+
+       range = (struct kvm_io_range) {
+               .addr = addr,
+               .len = len,
+       };
 
        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))
+       idx = kvm_io_bus_get_first_dev(bus, addr, len);
+       if (idx < 0)
+               return -EOPNOTSUPP;
+
+       while (idx < bus->dev_count &&
+               kvm_io_bus_sort_cmp(&range, &bus->range[idx]) == 0) {
+               if (!kvm_iodevice_read(bus->range[idx].dev, addr, len, val))
                        return 0;
+               idx++;
+       }
+
        return -EOPNOTSUPP;
 }
 
 /* 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 kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx, gpa_t addr,
+                           int len, struct kvm_io_device *dev)
 {
        struct kvm_io_bus *new_bus, *bus;
 
@@ -2284,11 +2573,10 @@ int kvm_io_bus_register_dev(struct kvm *kvm, enum kvm_bus bus_idx,
        if (bus->dev_count > NR_IOBUS_DEVS-1)
                return -ENOSPC;
 
-       new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL);
+       new_bus = kmemdup(bus, 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;
+       kvm_io_bus_insert_dev(new_bus, dev, addr, len);
        rcu_assign_pointer(kvm->buses[bus_idx], new_bus);
        synchronize_srcu_expedited(&kvm->srcu);
        kfree(bus);
@@ -2303,18 +2591,21 @@ int kvm_io_bus_unregister_dev(struct kvm *kvm, enum kvm_bus bus_idx,
        int i, r;
        struct kvm_io_bus *new_bus, *bus;
 
-       new_bus = kzalloc(sizeof(struct kvm_io_bus), GFP_KERNEL);
+       bus = kvm->buses[bus_idx];
+
+       new_bus = kmemdup(bus, sizeof(*bus), GFP_KERNEL);
        if (!new_bus)
                return -ENOMEM;
 
-       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) {
+               if (new_bus->range[i].dev == dev) {
                        r = 0;
-                       new_bus->devs[i] = new_bus->devs[--new_bus->dev_count];
+                       new_bus->dev_count--;
+                       new_bus->range[i] = new_bus->range[new_bus->dev_count];
+                       sort(new_bus->range, new_bus->dev_count,
+                            sizeof(struct kvm_io_range),
+                            kvm_io_bus_sort_cmp, NULL);
                        break;
                }
 
@@ -2339,10 +2630,10 @@ static int vm_stat_get(void *_offset, u64 *val)
        struct kvm *kvm;
 
        *val = 0;
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        list_for_each_entry(kvm, &vm_list, vm_list)
                *val += *(u32 *)((void *)kvm + offset);
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
        return 0;
 }
 
@@ -2356,12 +2647,12 @@ static int vcpu_stat_get(void *_offset, u64 *val)
        int i;
 
        *val = 0;
-       spin_lock(&kvm_lock);
+       raw_spin_lock(&kvm_lock);
        list_for_each_entry(kvm, &vm_list, vm_list)
                kvm_for_each_vcpu(i, vcpu, kvm)
                        *val += *(u32 *)((void *)vcpu + offset);
 
-       spin_unlock(&kvm_lock);
+       raw_spin_unlock(&kvm_lock);
        return 0;
 }
 
@@ -2392,33 +2683,26 @@ static void kvm_exit_debug(void)
        debugfs_remove(kvm_debugfs_dir);
 }
 
-static int kvm_suspend(struct sys_device *dev, pm_message_t state)
+static int kvm_suspend(void)
 {
        if (kvm_usage_count)
                hardware_disable_nolock(NULL);
        return 0;
 }
 
-static int kvm_resume(struct sys_device *dev)
+static void kvm_resume(void)
 {
        if (kvm_usage_count) {
-               WARN_ON(spin_is_locked(&kvm_lock));
+               WARN_ON(raw_spin_is_locked(&kvm_lock));
                hardware_enable_nolock(NULL);
        }
-       return 0;
 }
 
-static struct sysdev_class kvm_sysdev_class = {
-       .name = "kvm",
+static struct syscore_ops kvm_syscore_ops = {
        .suspend = kvm_suspend,
        .resume = kvm_resume,
 };
 
-static struct sys_device kvm_sysdev = {
-       .id = 0,
-       .cls = &kvm_sysdev_class,
-};
-
 struct page *bad_page;
 pfn_t bad_pfn;
 
@@ -2502,14 +2786,6 @@ int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
                goto out_free_2;
        register_reboot_notifier(&kvm_reboot_notifier);
 
-       r = sysdev_class_register(&kvm_sysdev_class);
-       if (r)
-               goto out_free_3;
-
-       r = sysdev_register(&kvm_sysdev);
-       if (r)
-               goto out_free_4;
-
        /* A kmem cache lets us meet the alignment requirements of fx_save. */
        if (!vcpu_align)
                vcpu_align = __alignof__(struct kvm_vcpu);
@@ -2517,7 +2793,7 @@ int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
                                           0, NULL);
        if (!kvm_vcpu_cache) {
                r = -ENOMEM;
-               goto out_free_5;
+               goto out_free_3;
        }
 
        r = kvm_async_pf_init();
@@ -2534,6 +2810,8 @@ int kvm_init(void *opaque, unsigned vcpu_size, unsigned vcpu_align,
                goto out_unreg;
        }
 
+       register_syscore_ops(&kvm_syscore_ops);
+
        kvm_preempt_ops.sched_in = kvm_sched_in;
        kvm_preempt_ops.sched_out = kvm_sched_out;
 
@@ -2545,10 +2823,6 @@ out_unreg:
        kvm_async_pf_deinit();
 out_free:
        kmem_cache_destroy(kvm_vcpu_cache);
-out_free_5:
-       sysdev_unregister(&kvm_sysdev);
-out_free_4:
-       sysdev_class_unregister(&kvm_sysdev_class);
 out_free_3:
        unregister_reboot_notifier(&kvm_reboot_notifier);
        unregister_cpu_notifier(&kvm_cpu_notifier);
@@ -2576,8 +2850,7 @@ void kvm_exit(void)
        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_syscore_ops(&kvm_syscore_ops);
        unregister_reboot_notifier(&kvm_reboot_notifier);
        unregister_cpu_notifier(&kvm_cpu_notifier);
        on_each_cpu(hardware_disable_nolock, NULL, 1);