drivers: hv: mshv_vtl: Implement restore partition time IOCTL

Hyper-V supports a restore partition time hypercall that resets the TSC to provided values. This call is typically made by a Windows guest when restoring from hibernate. The paravisor needs to support this call in order to allow the guest to be able to successfully perform the operation. Since the reference time is being changed, clocks need to be updated to maintain consistent foward progress.

Author: hargar19

arch/x86/include/asm/mshyperv.h

@@ -241,6 +241,8 @@ void hyperv_setup_mmu_ops(void);
 void set_hv_tscchange_cb(void (*cb)(void));
 void clear_hv_tscchange_cb(void);
 void hyperv_stop_tsc_emulation(void);
+void hv_save_sched_clock_state(void);
+void hv_restore_sched_clock_state(void);
 int hyperv_flush_guest_mapping(u64 as);
 int hyperv_flush_guest_mapping_range(u64 as,
 		hyperv_fill_flush_list_func fill_func, void *data);

arch/x86/kernel/cpu/mshyperv.c

@@ -266,13 +266,13 @@ static void restore_hv_clock_tsc_state(void)
  * suspend-resume and the offset used to measure time needs to be
  * corrected, post resume.
  */
-static void hv_save_sched_clock_state(void)
+void hv_save_sched_clock_state(void)
 {
 	old_save_sched_clock_state();
 	save_hv_clock_tsc_state();
 }
 
-static void hv_restore_sched_clock_state(void)
+void hv_restore_sched_clock_state(void)
 {
 	restore_hv_clock_tsc_state();
 	old_restore_sched_clock_state();

drivers/hv/mshv_vtl_main.c

@@ -32,6 +32,7 @@
 
 #ifdef CONFIG_X86_64
 #include <linux/cleanup.h>
+#include <linux/stop_machine.h>
 
 #include <asm/apic.h>
 #include <uapi/asm/mtrr.h>
@@ -42,6 +43,7 @@
 #include <asm/vmx.h>
 
 #include "../../kernel/fpu/legacy.h"
+#include "../../kernel/time/timekeeping.h"
 
 #endif
 
@@ -865,6 +867,55 @@ static int mshv_vtl_ioctl_add_vtl0_mem(struct mshv_vtl *vtl, void __user *arg)
 	return 0;
 }
 
+#ifdef CONFIG_X86_64
+static int restore_partition_time_with_cpus_stopped(void *data)
+{
+	struct mshv_partition_time *partition_time = data;
+	struct hv_input_restore_partition_time *input;
+	int result = 0;
+	u64 status;
+
+	/* Save current clock state. Other CPUs are waiting in stop code, so no locks are taken. */
+	sched_clock_suspend();
+	timekeeping_suspend();
+	hv_save_sched_clock_state();
+
+	/* Interrupts are disabled. Make the hypercall to update the TSC. */
+	input = *this_cpu_ptr(hyperv_pcpu_input_arg);
+	input->partition_id = HV_PARTITION_ID_SELF;
+	input->tsc_sequence = partition_time->tsc_sequence;
+	input->reserved = 0;
+	input->reference_time_in_100_ns = partition_time->reference_time_in_100_ns;
+	input->tsc = partition_time->tsc;
+	status = hv_do_hypercall(HVCALL_RESTORE_PARTITION_TIME, input, NULL);
+	if (!hv_result_success(status)) {
+		pr_err("HVCALL_RESTORE_PARTITION_TIME failed with %#llx\n", status);
+		result = -EINVAL;
+	}
+
+	/* Restore clock state using current TSC value. */
+	hv_restore_sched_clock_state();
+	timekeeping_resume();
+	sched_clock_resume();
+
+	return result;
+}
+
+static int mshv_restore_partition_time(void __user *arg)
+{
+	struct mshv_partition_time partition_time;
+	int ret;
+
+	if (copy_from_user(&partition_time, arg, sizeof(partition_time)))
+		return -EFAULT;
+
+	/* Stop other CPUs, using the current one to restore partition time. */
+	ret = stop_machine(restore_partition_time_with_cpus_stopped, &partition_time,
+			cpumask_of(raw_smp_processor_id()));
+	return ret;
+}
+#endif
+
 static void mshv_vtl_cancel(int cpu)
 {
 	int here = get_cpu();
@@ -2596,6 +2647,13 @@ mshv_vtl_ioctl(struct file *filp, unsigned int ioctl, unsigned long arg)
 	case MSHV_VTL_ADD_VTL0_MEMORY:
 		ret = mshv_vtl_ioctl_add_vtl0_mem(vtl, (void __user *)arg);
 		break;
+
+#if defined(CONFIG_X86_64)
+	case MSHV_RESTORE_PARTITION_TIME:
+		ret = mshv_restore_partition_time((void __user *)arg);
+		break;
+#endif
+
 #if defined(CONFIG_X86_64) && defined(CONFIG_INTEL_TDX_GUEST)
 	case MSHV_VTL_TDCALL:
 		ret = mshv_vtl_ioctl_tdcall((void __user *)arg);

include/uapi/hyperv/hvgdk_mini.h

@@ -261,6 +261,8 @@ union hv_hypervisor_version_info {
 #define HVCALL_GET_VP_CPUID_VALUES		0x00f4
 #define HVCALL_START_VP				0x0099
 #define HVCALL_GET_VP_INDEX_FROM_APIC_ID		0x009a
+#define HVCALL_RESTORE_PARTITION_TIME		0x0103
+
 
 /*
  * Some macros - i.e. GENMASK_ULL and BIT_ULL - are not currently supported by
@@ -1310,4 +1312,14 @@ struct hv_input_install_intercept {
 	union hv_intercept_parameters intercept_parameter;
 } __packed;
 
+#if defined(__x86_64__)
+struct hv_input_restore_partition_time {
+	__u64 partition_id;
+	__u32 tsc_sequence;
+	__u32 reserved;
+	__u64 reference_time_in_100_ns;
+	__u64 tsc;
+} __packed;
+#endif
+
 #endif /* _UAPI_HV_HVGDK_MINI_H */

include/uapi/linux/mshv.h

@@ -306,6 +306,13 @@ struct mshv_map_device_intr {
 	__u8 padding[7];
 } __packed;
 
+struct mshv_partition_time {
+	__u32 tsc_sequence;
+	__u32 reserved;
+	__u64 reference_time_in_100_ns;
+	__u64 tsc;
+} __packed;
+
 #define MSHV_KICK_CPUS_FLAG_WAIT_FOR_CPUS	(1 << 0)
 #define MSHV_KICK_CPUS_FLAG_CANCEL_CPU_RUN	(1 << 1)
 
@@ -371,6 +378,7 @@ struct mshv_map_device_intr {
 						       struct mshv_map_device_intr)
 
 /* For x86-64 only */
+#define MSHV_RESTORE_PARTITION_TIME	_IOW(MSHV_IOCTL, 0x13, struct mshv_partition_time)
 #define MSHV_VTL_GUEST_VSM_VMSA_PFN	_IOWR(MSHV_IOCTL, 0x34, __u64)
 
 /* For x86-64 SEV-SNP only */