@@ -4771,6 +4771,34 @@ the event is disallowed.
Unless set to -1 (meaning all event), id must be a event ID
(e.g. KVMI_EVENT_UNHOOK, KVMI_EVENT_CR, etc.)
+4.129 KVM_INTROSPECTION_PREUNHOOK
+---------------------------------
+
+:Capability: KVM_CAP_INTROSPECTION
+:Architectures: x86
+:Type: vm ioctl
+:Parameters: none
+:Returns: 0 on success, a negative value on error
+
+Errors:
+
+ ====== ============================================================
+ EFAULT the introspection is not enabled
+ EFAULT the socket (passed with KVM_INTROSPECTION_HOOK) had an error
+ ENOENT the introspection tool didn't subscribed
+ to this type of introspection event (unhook)
+ ====== ============================================================
+
+This ioctl is used to inform that the current VM is
+paused/suspended/migrated/etc.
+
+KVM should send an 'unhook' introspection event to the introspection tool.
+
+If this ioctl is successful, the userspace should give the
+introspection tool a chance to unhook the VM and then it should use
+KVM_INTROSPECTION_UNHOOK to make sure all the introspection structures
+are freed.
+
5. The kvm_run structure
========================
@@ -194,9 +194,10 @@ becomes necessary to remove them before the guest is suspended, moved
(migrated) or a snapshot with memory is created.
The actions are normally performed by the device manager. In the case
-of QEMU, it will use another ioctl to notify the introspection tool and
-wait for a limited amount of time (a few seconds) for a confirmation that
-is OK to proceed.
+of QEMU, it will use the *KVM_INTROSPECTION_PREUNHOOK* ioctl to trigger
+the *KVMI_EVENT_UNHOOK* event and wait for a limited amount of time
+(a few seconds) for a confirmation from the introspection tool
+that is OK to proceed.
Live migrations
---------------
@@ -338,3 +339,65 @@ This command is always allowed.
};
Returns the number of online vCPUs.
+
+Events
+======
+
+All introspection events (VM or vCPU related) are sent
+using the *KVMI_EVENT* message id.
+
+The *KVMI_EVENT_UNHOOK* event doesn't have a reply and share the kvmi_event
+structure, for consistency with the vCPU events.
+
+The message data begins with a common structure, having the size of the
+structure, the vCPU index and the event id::
+
+ struct kvmi_event {
+ __u16 size;
+ __u16 vcpu;
+ __u8 event;
+ __u8 padding[3];
+ struct kvmi_event_arch arch;
+ }
+
+On x86 the structure looks like this::
+
+ struct kvmi_event_arch {
+ __u8 mode;
+ __u8 padding[7];
+ struct kvm_regs regs;
+ struct kvm_sregs sregs;
+ struct {
+ __u64 sysenter_cs;
+ __u64 sysenter_esp;
+ __u64 sysenter_eip;
+ __u64 efer;
+ __u64 star;
+ __u64 lstar;
+ __u64 cstar;
+ __u64 pat;
+ __u64 shadow_gs;
+ } msrs;
+ };
+
+It contains information about the vCPU state at the time of the event.
+
+Specific data can follow these common structures.
+
+1. KVMI_EVENT_UNHOOK
+--------------------
+
+:Architecture: all
+:Versions: >= 1
+:Actions: none
+:Parameters:
+
+::
+
+ struct kvmi_event;
+
+:Returns: none
+
+This event is sent when the device manager has to pause/stop/migrate the
+guest (see **Unhooking**). The introspection tool has a chance to unhook
+and close the KVMI channel (signaling that the operation can proceed).
new file mode 100644
@@ -0,0 +1,29 @@
+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
+#ifndef _UAPI_ASM_X86_KVMI_H
+#define _UAPI_ASM_X86_KVMI_H
+
+/*
+ * KVM introspection - x86 specific structures and definitions
+ */
+
+#include <asm/kvm.h>
+
+struct kvmi_event_arch {
+ __u8 mode; /* 2, 4 or 8 */
+ __u8 padding[7];
+ struct kvm_regs regs;
+ struct kvm_sregs sregs;
+ struct {
+ __u64 sysenter_cs;
+ __u64 sysenter_esp;
+ __u64 sysenter_eip;
+ __u64 efer;
+ __u64 star;
+ __u64 lstar;
+ __u64 cstar;
+ __u64 pat;
+ __u64 shadow_gs;
+ } msrs;
+};
+
+#endif /* _UAPI_ASM_X86_KVMI_H */
@@ -21,6 +21,8 @@ struct kvm_introspection {
DECLARE_BITMAP(cmd_allow_mask, KVMI_NUM_COMMANDS);
DECLARE_BITMAP(event_allow_mask, KVMI_NUM_EVENTS);
+
+ atomic_t ev_seq;
};
#ifdef CONFIG_KVM_INTROSPECTION
@@ -34,6 +36,7 @@ int kvmi_ioctl_hook(struct kvm *kvm, void __user *argp);
int kvmi_ioctl_unhook(struct kvm *kvm);
int kvmi_ioctl_command(struct kvm *kvm, void __user *argp);
int kvmi_ioctl_event(struct kvm *kvm, void __user *argp);
+int kvmi_ioctl_preunhook(struct kvm *kvm);
#else
@@ -1576,6 +1576,8 @@ struct kvm_introspection_feature {
#define KVM_INTROSPECTION_COMMAND _IOW(KVMIO, 0xc5, struct kvm_introspection_feature)
#define KVM_INTROSPECTION_EVENT _IOW(KVMIO, 0xc6, struct kvm_introspection_feature)
+#define KVM_INTROSPECTION_PREUNHOOK _IO(KVMIO, 0xc7)
+
#define KVM_DEV_ASSIGN_ENABLE_IOMMU (1 << 0)
#define KVM_DEV_ASSIGN_PCI_2_3 (1 << 1)
#define KVM_DEV_ASSIGN_MASK_INTX (1 << 2)
@@ -8,12 +8,15 @@
#include <linux/kernel.h>
#include <linux/types.h>
+#include <asm/kvmi.h>
enum {
KVMI_VERSION = 0x00000001
};
enum {
+ KVMI_EVENT = 1,
+
KVMI_GET_VERSION = 2,
KVMI_VM_CHECK_COMMAND = 3,
KVMI_VM_CHECK_EVENT = 4,
@@ -23,6 +26,8 @@ enum {
};
enum {
+ KVMI_EVENT_UNHOOK = 0,
+
KVMI_NUM_EVENTS
};
@@ -70,4 +75,12 @@ struct kvmi_vm_get_info_reply {
__u32 padding[3];
};
+struct kvmi_event {
+ __u16 size;
+ __u16 vcpu;
+ __u8 event;
+ __u8 padding[3];
+ struct kvmi_event_arch arch;
+};
+
#endif /* _UAPI__LINUX_KVMI_H */
@@ -268,6 +268,47 @@ static void test_cmd_get_vm_info(void)
DEBUG("vcpu count: %u\n", rpl.vcpu_count);
}
+static void trigger_event_unhook_notification(struct kvm_vm *vm)
+{
+ int r;
+
+ r = ioctl(vm->fd, KVM_INTROSPECTION_PREUNHOOK, NULL);
+ TEST_ASSERT(r == 0,
+ "KVM_INTROSPECTION_PREUNHOOK failed, errno %d (%s)\n",
+ errno, strerror(errno));
+}
+
+static void receive_event(struct kvmi_msg_hdr *hdr, struct kvmi_event *ev,
+ size_t ev_size, int event_id)
+{
+ receive_data(hdr, sizeof(*hdr));
+
+ TEST_ASSERT(hdr->id == KVMI_EVENT,
+ "Unexpected messages id %d, expected %d\n",
+ hdr->id, KVMI_EVENT);
+
+ TEST_ASSERT(hdr->size == ev_size,
+ "Invalid event size %d, expected %d bytes\n",
+ hdr->size, ev_size);
+
+ receive_data(ev, ev_size);
+
+ TEST_ASSERT(ev->event == event_id,
+ "Unexpected event %d, expected %d\n",
+ ev->event, event_id);
+}
+
+static void test_event_unhook(struct kvm_vm *vm)
+{
+ __u16 id = KVMI_EVENT_UNHOOK;
+ struct kvmi_msg_hdr hdr;
+ struct kvmi_event ev;
+
+ trigger_event_unhook_notification(vm);
+
+ receive_event(&hdr, &ev, sizeof(ev), id);
+}
+
static void test_introspection(struct kvm_vm *vm)
{
setup_socket();
@@ -278,6 +319,7 @@ static void test_introspection(struct kvm_vm *vm)
test_cmd_check_command();
test_cmd_check_event();
test_cmd_get_vm_info();
+ test_event_unhook(vm);
unhook_introspection(vm);
}
@@ -12,6 +12,8 @@
static DECLARE_BITMAP(Kvmi_always_allowed_commands, KVMI_NUM_COMMANDS);
DECLARE_BITMAP(Kvmi_known_events, KVMI_NUM_EVENTS);
+static DECLARE_BITMAP(Kvmi_known_vm_events, KVMI_NUM_EVENTS);
+static DECLARE_BITMAP(Kvmi_known_vcpu_events, KVMI_NUM_EVENTS);
static struct kmem_cache *msg_cache;
@@ -56,7 +58,13 @@ static void setup_always_allowed_commands(void)
static void setup_known_events(void)
{
- bitmap_zero(Kvmi_known_events, KVMI_NUM_EVENTS);
+ bitmap_zero(Kvmi_known_vm_events, KVMI_NUM_EVENTS);
+ set_bit(KVMI_EVENT_UNHOOK, Kvmi_known_vm_events);
+
+ bitmap_zero(Kvmi_known_vcpu_events, KVMI_NUM_EVENTS);
+
+ bitmap_or(Kvmi_known_events, Kvmi_known_vm_events,
+ Kvmi_known_vcpu_events, KVMI_NUM_EVENTS);
}
int kvmi_init(void)
@@ -93,6 +101,8 @@ alloc_kvmi(struct kvm *kvm, const struct kvm_introspection_hook *hook)
bitmap_copy(kvmi->cmd_allow_mask, Kvmi_always_allowed_commands,
KVMI_NUM_COMMANDS);
+ atomic_set(&kvmi->ev_seq, 0);
+
kvmi->kvm = kvm;
return kvmi;
@@ -349,3 +359,33 @@ int kvmi_ioctl_command(struct kvm *kvm, void __user *argp)
mutex_unlock(&kvm->kvmi_lock);
return err;
}
+
+static bool kvmi_unhook_event(struct kvm_introspection *kvmi)
+{
+ int err;
+
+ err = kvmi_msg_send_unhook(kvmi);
+
+ return !err;
+}
+
+int kvmi_ioctl_preunhook(struct kvm *kvm)
+{
+ struct kvm_introspection *kvmi;
+ int err = 0;
+
+ mutex_lock(&kvm->kvmi_lock);
+
+ kvmi = KVMI(kvm);
+ if (!kvmi) {
+ err = -EFAULT;
+ goto out;
+ }
+
+ if (!kvmi_unhook_event(kvmi))
+ err = -ENOENT;
+
+out:
+ mutex_unlock(&kvm->kvmi_lock);
+ return err;
+}
@@ -25,6 +25,7 @@ bool kvmi_sock_get(struct kvm_introspection *kvmi, int fd);
void kvmi_sock_shutdown(struct kvm_introspection *kvmi);
void kvmi_sock_put(struct kvm_introspection *kvmi);
bool kvmi_msg_process(struct kvm_introspection *kvmi);
+int kvmi_msg_send_unhook(struct kvm_introspection *kvmi);
/* kvmi.c */
void *kvmi_msg_alloc(void);
@@ -256,3 +256,41 @@ bool kvmi_msg_process(struct kvm_introspection *kvmi)
out:
return err == 0;
}
+
+static void kvmi_setup_event_msg_hdr(struct kvm_introspection *kvmi,
+ struct kvmi_msg_hdr *hdr,
+ size_t msg_size)
+{
+ memset(hdr, 0, sizeof(*hdr));
+
+ hdr->id = KVMI_EVENT;
+ hdr->seq = atomic_inc_return(&kvmi->ev_seq);
+ hdr->size = msg_size - sizeof(*hdr);
+}
+
+static void kvmi_setup_event_common(struct kvmi_event *ev, u32 ev_id,
+ u16 vcpu_idx)
+{
+ memset(ev, 0, sizeof(*ev));
+
+ ev->vcpu = vcpu_idx;
+ ev->event = ev_id;
+ ev->size = sizeof(*ev);
+}
+
+int kvmi_msg_send_unhook(struct kvm_introspection *kvmi)
+{
+ struct kvmi_msg_hdr hdr;
+ struct kvmi_event common;
+ struct kvec vec[] = {
+ {.iov_base = &hdr, .iov_len = sizeof(hdr) },
+ {.iov_base = &common, .iov_len = sizeof(common)},
+ };
+ size_t msg_size = sizeof(hdr) + sizeof(common);
+ size_t n = ARRAY_SIZE(vec);
+
+ kvmi_setup_event_msg_hdr(kvmi, &hdr, msg_size);
+ kvmi_setup_event_common(&common, KVMI_EVENT_UNHOOK, 0);
+
+ return kvmi_sock_write(kvmi, vec, n, msg_size);
+}
@@ -3596,6 +3596,9 @@ static long kvm_vm_ioctl(struct file *filp,
case KVM_INTROSPECTION_EVENT:
r = kvmi_ioctl_event(kvm, argp);
break;
+ case KVM_INTROSPECTION_PREUNHOOK:
+ r = kvmi_ioctl_preunhook(kvm);
+ break;
#endif /* CONFIG_KVM_INTROSPECTION */
default:
r = kvm_arch_vm_ioctl(filp, ioctl, arg);
In certain situations (when the guest has to be paused, suspended, migrated, etc.), userspace will use the KVM_INTROSPECTION_PREUNHOOK ioctl in order to trigger the KVMI_EVENT_UNHOOK event. If the event is sent successfully (the VM has an active introspection channel), userspace should delay the action (pause/suspend/...) to give the introspection tool the chance to remove its hooks (eg. breakpoints) while the guest is still running. Once a timeout is reached or the introspection tool has closed the socket, userspace should resume the action. Signed-off-by: Adalbert Lazăr <alazar@bitdefender.com> --- Documentation/virt/kvm/api.rst | 28 ++++++++ Documentation/virt/kvm/kvmi.rst | 69 ++++++++++++++++++- arch/x86/include/uapi/asm/kvmi.h | 29 ++++++++ include/linux/kvmi_host.h | 3 + include/uapi/linux/kvm.h | 2 + include/uapi/linux/kvmi.h | 13 ++++ .../testing/selftests/kvm/x86_64/kvmi_test.c | 42 +++++++++++ virt/kvm/introspection/kvmi.c | 42 ++++++++++- virt/kvm/introspection/kvmi_int.h | 1 + virt/kvm/introspection/kvmi_msg.c | 38 ++++++++++ virt/kvm/kvm_main.c | 3 + 11 files changed, 266 insertions(+), 4 deletions(-) create mode 100644 arch/x86/include/uapi/asm/kvmi.h