KVM: x86: enable dirty log gradually in small chunks
It could take kvm->mmu_lock for an extended period of time when
enabling dirty log for the first time. The main cost is to clear
all the D-bits of last level SPTEs. This situation can benefit from
manual dirty log protect as well, which can reduce the mmu_lock
time taken. The sequence is like this:
1. Initialize all the bits of the dirty bitmap to 1 when enabling
dirty log for the first time
2. Only write protect the huge pages
3. KVM_GET_DIRTY_LOG returns the dirty bitmap info
4. KVM_CLEAR_DIRTY_LOG will clear D-bit for each of the leaf level
SPTEs gradually in small chunks
Under the Intel(R) Xeon(R) Gold 6152 CPU @ 2.10GHz environment,
I did some tests with a 128G windows VM and counted the time taken
of memory_global_dirty_log_start, here is the numbers:
VM Size Before After optimization
128G 460ms 10ms
Signed-off-by: Jay Zhou <jianjay.zhou@huawei.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
diff --git a/Documentation/virt/kvm/api.rst b/Documentation/virt/kvm/api.rst
index ebd383f..0adef66 100644
--- a/Documentation/virt/kvm/api.rst
+++ b/Documentation/virt/kvm/api.rst
@@ -5707,8 +5707,13 @@
:Architectures: x86, arm, arm64, mips
:Parameters: args[0] whether feature should be enabled or not
-With this capability enabled, KVM_GET_DIRTY_LOG will not automatically
-clear and write-protect all pages that are returned as dirty.
+Valid flags are::
+
+ #define KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (1 << 0)
+ #define KVM_DIRTY_LOG_INITIALLY_SET (1 << 1)
+
+With KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE is set, KVM_GET_DIRTY_LOG will not
+automatically clear and write-protect all pages that are returned as dirty.
Rather, userspace will have to do this operation separately using
KVM_CLEAR_DIRTY_LOG.
@@ -5719,12 +5724,19 @@
take spinlocks for an extended period of time. Second, in some cases a
large amount of time can pass between a call to KVM_GET_DIRTY_LOG and
userspace actually using the data in the page. Pages can be modified
-during this time, which is inefficint for both the guest and userspace:
+during this time, which is inefficient for both the guest and userspace:
the guest will incur a higher penalty due to write protection faults,
while userspace can see false reports of dirty pages. Manual reprotection
helps reducing this time, improving guest performance and reducing the
number of dirty log false positives.
+With KVM_DIRTY_LOG_INITIALLY_SET set, all the bits of the dirty bitmap
+will be initialized to 1 when created. This also improves performance because
+dirty logging can be enabled gradually in small chunks on the first call
+to KVM_CLEAR_DIRTY_LOG. KVM_DIRTY_LOG_INITIALLY_SET depends on
+KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (it is also only available on
+x86 for now).
+
KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 was previously available under the name
KVM_CAP_MANUAL_DIRTY_LOG_PROTECT, but the implementation had bugs that make
it hard or impossible to use it correctly. The availability of
diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
index f58861e..681e230 100644
--- a/arch/x86/include/asm/kvm_host.h
+++ b/arch/x86/include/asm/kvm_host.h
@@ -49,6 +49,9 @@
#define KVM_IRQCHIP_NUM_PINS KVM_IOAPIC_NUM_PINS
+#define KVM_DIRTY_LOG_MANUAL_CAPS (KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE | \
+ KVM_DIRTY_LOG_INITIALLY_SET)
+
/* x86-specific vcpu->requests bit members */
#define KVM_REQ_MIGRATE_TIMER KVM_ARCH_REQ(0)
#define KVM_REQ_REPORT_TPR_ACCESS KVM_ARCH_REQ(1)
@@ -1306,7 +1309,8 @@ void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
void kvm_mmu_reset_context(struct kvm_vcpu *vcpu);
void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
- struct kvm_memory_slot *memslot);
+ struct kvm_memory_slot *memslot,
+ int start_level);
void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
const struct kvm_memory_slot *memslot);
void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 53b776d..db0b66e 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -5864,13 +5864,14 @@ static bool slot_rmap_write_protect(struct kvm *kvm,
}
void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
- struct kvm_memory_slot *memslot)
+ struct kvm_memory_slot *memslot,
+ int start_level)
{
bool flush;
spin_lock(&kvm->mmu_lock);
- flush = slot_handle_all_level(kvm, memslot, slot_rmap_write_protect,
- false);
+ flush = slot_handle_level(kvm, memslot, slot_rmap_write_protect,
+ start_level, PT_MAX_HUGEPAGE_LEVEL, false);
spin_unlock(&kvm->mmu_lock);
/*
diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
index a04017b..2bb4c4e 100644
--- a/arch/x86/kvm/vmx/vmx.c
+++ b/arch/x86/kvm/vmx/vmx.c
@@ -7280,7 +7280,8 @@ static void vmx_sched_in(struct kvm_vcpu *vcpu, int cpu)
static void vmx_slot_enable_log_dirty(struct kvm *kvm,
struct kvm_memory_slot *slot)
{
- kvm_mmu_slot_leaf_clear_dirty(kvm, slot);
+ if (!kvm_dirty_log_manual_protect_and_init_set(kvm))
+ kvm_mmu_slot_leaf_clear_dirty(kvm, slot);
kvm_mmu_slot_largepage_remove_write_access(kvm, slot);
}
diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index ddd1d29..864d0ad 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -9916,7 +9916,7 @@ static void kvm_mmu_slot_apply_flags(struct kvm *kvm,
{
/* Still write protect RO slot */
if (new->flags & KVM_MEM_READONLY) {
- kvm_mmu_slot_remove_write_access(kvm, new);
+ kvm_mmu_slot_remove_write_access(kvm, new, PT_PAGE_TABLE_LEVEL);
return;
}
@@ -9951,10 +9951,23 @@ static void kvm_mmu_slot_apply_flags(struct kvm *kvm,
* See the comments in fast_page_fault().
*/
if (new->flags & KVM_MEM_LOG_DIRTY_PAGES) {
- if (kvm_x86_ops->slot_enable_log_dirty)
+ if (kvm_x86_ops->slot_enable_log_dirty) {
kvm_x86_ops->slot_enable_log_dirty(kvm, new);
- else
- kvm_mmu_slot_remove_write_access(kvm, new);
+ } else {
+ int level =
+ kvm_dirty_log_manual_protect_and_init_set(kvm) ?
+ PT_DIRECTORY_LEVEL : PT_PAGE_TABLE_LEVEL;
+
+ /*
+ * If we're with initial-all-set, we don't need
+ * to write protect any small page because
+ * they're reported as dirty already. However
+ * we still need to write-protect huge pages
+ * so that the page split can happen lazily on
+ * the first write to the huge page.
+ */
+ kvm_mmu_slot_remove_write_access(kvm, new, level);
+ }
} else {
if (kvm_x86_ops->slot_disable_log_dirty)
kvm_x86_ops->slot_disable_log_dirty(kvm, new);
diff --git a/include/linux/kvm_host.h b/include/linux/kvm_host.h
index 4bd5251..127cb08 100644
--- a/include/linux/kvm_host.h
+++ b/include/linux/kvm_host.h
@@ -360,6 +360,10 @@ static inline unsigned long *kvm_second_dirty_bitmap(struct kvm_memory_slot *mem
return memslot->dirty_bitmap + len / sizeof(*memslot->dirty_bitmap);
}
+#ifndef KVM_DIRTY_LOG_MANUAL_CAPS
+#define KVM_DIRTY_LOG_MANUAL_CAPS KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE
+#endif
+
struct kvm_s390_adapter_int {
u64 ind_addr;
u64 summary_addr;
@@ -493,7 +497,7 @@ struct kvm {
#endif
long tlbs_dirty;
struct list_head devices;
- bool manual_dirty_log_protect;
+ u64 manual_dirty_log_protect;
struct dentry *debugfs_dentry;
struct kvm_stat_data **debugfs_stat_data;
struct srcu_struct srcu;
@@ -527,6 +531,11 @@ struct kvm {
#define vcpu_err(vcpu, fmt, ...) \
kvm_err("vcpu%i " fmt, (vcpu)->vcpu_id, ## __VA_ARGS__)
+static inline bool kvm_dirty_log_manual_protect_and_init_set(struct kvm *kvm)
+{
+ return !!(kvm->manual_dirty_log_protect & KVM_DIRTY_LOG_INITIALLY_SET);
+}
+
static inline struct kvm_io_bus *kvm_get_bus(struct kvm *kvm, enum kvm_bus idx)
{
return srcu_dereference_check(kvm->buses[idx], &kvm->srcu,
diff --git a/include/uapi/linux/kvm.h b/include/uapi/linux/kvm.h
index 4b95f9a..d8499d9 100644
--- a/include/uapi/linux/kvm.h
+++ b/include/uapi/linux/kvm.h
@@ -1628,4 +1628,7 @@ struct kvm_hyperv_eventfd {
#define KVM_HYPERV_CONN_ID_MASK 0x00ffffff
#define KVM_HYPERV_EVENTFD_DEASSIGN (1 << 0)
+#define KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE (1 << 0)
+#define KVM_DIRTY_LOG_INITIALLY_SET (1 << 1)
+
#endif /* __LINUX_KVM_H */
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 26ccb6c..699ff9b 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -858,7 +858,7 @@ static int kvm_vm_release(struct inode *inode, struct file *filp)
* Allocation size is twice as large as the actual dirty bitmap size.
* See kvm_vm_ioctl_get_dirty_log() why this is needed.
*/
-static int kvm_create_dirty_bitmap(struct kvm_memory_slot *memslot)
+static int kvm_alloc_dirty_bitmap(struct kvm_memory_slot *memslot)
{
unsigned long dirty_bytes = 2 * kvm_dirty_bitmap_bytes(memslot);
@@ -1288,9 +1288,12 @@ int __kvm_set_memory_region(struct kvm *kvm,
if (!(new.flags & KVM_MEM_LOG_DIRTY_PAGES))
new.dirty_bitmap = NULL;
else if (!new.dirty_bitmap) {
- r = kvm_create_dirty_bitmap(&new);
+ r = kvm_alloc_dirty_bitmap(&new);
if (r)
return r;
+
+ if (kvm_dirty_log_manual_protect_and_init_set(kvm))
+ bitmap_set(new.dirty_bitmap, 0, new.npages);
}
r = kvm_set_memslot(kvm, mem, &old, &new, as_id, change);
@@ -3529,9 +3532,6 @@ static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
case KVM_CAP_IOEVENTFD_ANY_LENGTH:
case KVM_CAP_CHECK_EXTENSION_VM:
case KVM_CAP_ENABLE_CAP_VM:
-#ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
- case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
-#endif
return 1;
#ifdef CONFIG_KVM_MMIO
case KVM_CAP_COALESCED_MMIO:
@@ -3539,6 +3539,10 @@ static long kvm_vm_ioctl_check_extension_generic(struct kvm *kvm, long arg)
case KVM_CAP_COALESCED_PIO:
return 1;
#endif
+#ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
+ case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
+ return KVM_DIRTY_LOG_MANUAL_CAPS;
+#endif
#ifdef CONFIG_HAVE_KVM_IRQ_ROUTING
case KVM_CAP_IRQ_ROUTING:
return KVM_MAX_IRQ_ROUTES;
@@ -3566,11 +3570,17 @@ static int kvm_vm_ioctl_enable_cap_generic(struct kvm *kvm,
{
switch (cap->cap) {
#ifdef CONFIG_KVM_GENERIC_DIRTYLOG_READ_PROTECT
- case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2:
- if (cap->flags || (cap->args[0] & ~1))
+ case KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2: {
+ u64 allowed_options = KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE;
+
+ if (cap->args[0] & KVM_DIRTY_LOG_MANUAL_PROTECT_ENABLE)
+ allowed_options = KVM_DIRTY_LOG_MANUAL_CAPS;
+
+ if (cap->flags || (cap->args[0] & ~allowed_options))
return -EINVAL;
kvm->manual_dirty_log_protect = cap->args[0];
return 0;
+ }
#endif
default:
return kvm_vm_ioctl_enable_cap(kvm, cap);