kvmclock: add a new function to update env->tsc.
[kvmfornfv.git] / qemu / hw / i386 / kvm / clock.c
1 /*
2  * QEMU KVM support, paravirtual clock device
3  *
4  * Copyright (C) 2011 Siemens AG
5  *
6  * Authors:
7  *  Jan Kiszka        <jan.kiszka@siemens.com>
8  *
9  * This work is licensed under the terms of the GNU GPL version 2.
10  * See the COPYING file in the top-level directory.
11  *
12  * Contributions after 2012-01-13 are licensed under the terms of the
13  * GNU GPL, version 2 or (at your option) any later version.
14  */
15
16 #include "qemu-common.h"
17 #include "qemu/host-utils.h"
18 #include "sysemu/sysemu.h"
19 #include "sysemu/kvm.h"
20 #include "kvm_i386.h"
21 #include "hw/sysbus.h"
22 #include "hw/kvm/clock.h"
23
24 #include <linux/kvm.h>
25 #include <linux/kvm_para.h>
26
27 #define TYPE_KVM_CLOCK "kvmclock"
28 #define KVM_CLOCK(obj) OBJECT_CHECK(KVMClockState, (obj), TYPE_KVM_CLOCK)
29
30 typedef struct KVMClockState {
31     /*< private >*/
32     SysBusDevice busdev;
33     /*< public >*/
34
35     uint64_t clock;
36     bool clock_valid;
37 } KVMClockState;
38
39 struct pvclock_vcpu_time_info {
40     uint32_t   version;
41     uint32_t   pad0;
42     uint64_t   tsc_timestamp;
43     uint64_t   system_time;
44     uint32_t   tsc_to_system_mul;
45     int8_t     tsc_shift;
46     uint8_t    flags;
47     uint8_t    pad[2];
48 } __attribute__((__packed__)); /* 32 bytes */
49
50 static uint64_t kvmclock_current_nsec(KVMClockState *s)
51 {
52     CPUState *cpu = first_cpu;
53     CPUX86State *env = cpu->env_ptr;
54     hwaddr kvmclock_struct_pa = env->system_time_msr & ~1ULL;
55     uint64_t migration_tsc = env->tsc;
56     struct pvclock_vcpu_time_info time;
57     uint64_t delta;
58     uint64_t nsec_lo;
59     uint64_t nsec_hi;
60     uint64_t nsec;
61
62     if (!(env->system_time_msr & 1ULL)) {
63         /* KVM clock not active */
64         return 0;
65     }
66
67     cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time));
68
69     assert(time.tsc_timestamp <= migration_tsc);
70     delta = migration_tsc - time.tsc_timestamp;
71     if (time.tsc_shift < 0) {
72         delta >>= -time.tsc_shift;
73     } else {
74         delta <<= time.tsc_shift;
75     }
76
77     mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul);
78     nsec = (nsec_lo >> 32) | (nsec_hi << 32);
79     return nsec + time.system_time;
80 }
81
82 static void kvmclock_vm_state_change(void *opaque, int running,
83                                      RunState state)
84 {
85     KVMClockState *s = opaque;
86     CPUState *cpu;
87     int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
88     int ret;
89
90     if (running) {
91         struct kvm_clock_data data = {};
92         uint64_t time_at_migration = kvmclock_current_nsec(s);
93
94         s->clock_valid = false;
95
96         /* We can't rely on the migrated clock value, just discard it */
97         if (time_at_migration) {
98             s->clock = time_at_migration;
99         }
100
101         data.clock = s->clock;
102         ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
103         if (ret < 0) {
104             fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret));
105             abort();
106         }
107
108         if (!cap_clock_ctrl) {
109             return;
110         }
111         CPU_FOREACH(cpu) {
112             ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0);
113             if (ret) {
114                 if (ret != -EINVAL) {
115                     fprintf(stderr, "%s: %s\n", __func__, strerror(-ret));
116                 }
117                 return;
118             }
119         }
120     } else {
121         struct kvm_clock_data data;
122         int ret;
123
124         if (s->clock_valid) {
125             return;
126         }
127
128         kvm_synchronize_all_tsc();
129
130         ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data);
131         if (ret < 0) {
132             fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret));
133             abort();
134         }
135         s->clock = data.clock;
136
137         /*
138          * If the VM is stopped, declare the clock state valid to
139          * avoid re-reading it on next vmsave (which would return
140          * a different value). Will be reset when the VM is continued.
141          */
142         s->clock_valid = true;
143     }
144 }
145
146 static void kvmclock_realize(DeviceState *dev, Error **errp)
147 {
148     KVMClockState *s = KVM_CLOCK(dev);
149
150     qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
151 }
152
153 static const VMStateDescription kvmclock_vmsd = {
154     .name = "kvmclock",
155     .version_id = 1,
156     .minimum_version_id = 1,
157     .fields = (VMStateField[]) {
158         VMSTATE_UINT64(clock, KVMClockState),
159         VMSTATE_END_OF_LIST()
160     }
161 };
162
163 static void kvmclock_class_init(ObjectClass *klass, void *data)
164 {
165     DeviceClass *dc = DEVICE_CLASS(klass);
166
167     dc->realize = kvmclock_realize;
168     dc->vmsd = &kvmclock_vmsd;
169 }
170
171 static const TypeInfo kvmclock_info = {
172     .name          = TYPE_KVM_CLOCK,
173     .parent        = TYPE_SYS_BUS_DEVICE,
174     .instance_size = sizeof(KVMClockState),
175     .class_init    = kvmclock_class_init,
176 };
177
178 /* Note: Must be called after VCPU initialization. */
179 void kvmclock_create(void)
180 {
181     X86CPU *cpu = X86_CPU(first_cpu);
182
183     if (kvm_enabled() &&
184         cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) |
185                                        (1ULL << KVM_FEATURE_CLOCKSOURCE2))) {
186         sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL);
187     }
188 }
189
190 static void kvmclock_register_types(void)
191 {
192     type_register_static(&kvmclock_info);
193 }
194
195 type_init(kvmclock_register_types)