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/osdep.h"
17 #include "qemu/host-utils.h"
18 #include "qemu/module.h"
19 #include "system/kvm.h"
20 #include "system/runstate.h"
21 #include "system/hw_accel.h"
22 #include "kvm/kvm_i386.h"
23 #include "migration/vmstate.h"
24 #include "hw/sysbus.h"
25 #include "hw/i386/kvm/clock.h"
26 #include "hw/qdev-properties.h"
27 #include "qapi/error.h"
28
29 #include <linux/kvm.h>
30 #include "qom/object.h"
31
32 #define TYPE_KVM_CLOCK "kvmclock"
33 OBJECT_DECLARE_SIMPLE_TYPE(KVMClockState, KVM_CLOCK)
34
35 struct KVMClockState {
36 /*< private >*/
37 SysBusDevice busdev;
38 /*< public >*/
39
40 uint64_t clock;
41 bool clock_valid;
42
43 /* whether the 'clock' value was obtained in the 'paused' state */
44 bool runstate_paused;
45
46 /* whether machine type supports reliable KVM_GET_CLOCK */
47 bool mach_use_reliable_get_clock;
48
49 /* whether the 'clock' value was obtained in a host with
50 * reliable KVM_GET_CLOCK */
51 bool clock_is_reliable;
52 };
53
54 struct pvclock_vcpu_time_info {
55 uint32_t version;
56 uint32_t pad0;
57 uint64_t tsc_timestamp;
58 uint64_t system_time;
59 uint32_t tsc_to_system_mul;
60 int8_t tsc_shift;
61 uint8_t flags;
62 uint8_t pad[2];
63 } __attribute__((__packed__)); /* 32 bytes */
64
kvmclock_current_nsec(KVMClockState * s)65 static uint64_t kvmclock_current_nsec(KVMClockState *s)
66 {
67 CPUState *cpu = first_cpu;
68 CPUX86State *env = cpu_env(cpu);
69 hwaddr kvmclock_struct_pa;
70 uint64_t migration_tsc = env->tsc;
71 struct pvclock_vcpu_time_info time;
72 uint64_t delta;
73 uint64_t nsec_lo;
74 uint64_t nsec_hi;
75 uint64_t nsec;
76
77 cpu_synchronize_state(cpu);
78
79 if (!(env->system_time_msr & 1ULL)) {
80 /* KVM clock not active */
81 return 0;
82 }
83
84 kvmclock_struct_pa = env->system_time_msr & ~1ULL;
85 cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time));
86
87 assert(time.tsc_timestamp <= migration_tsc);
88 delta = migration_tsc - time.tsc_timestamp;
89 if (time.tsc_shift < 0) {
90 delta >>= -time.tsc_shift;
91 } else {
92 delta <<= time.tsc_shift;
93 }
94
95 mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul);
96 nsec = (nsec_lo >> 32) | (nsec_hi << 32);
97 return nsec + time.system_time;
98 }
99
kvm_update_clock(KVMClockState * s)100 static void kvm_update_clock(KVMClockState *s)
101 {
102 struct kvm_clock_data data;
103 int ret;
104
105 ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data);
106 if (ret < 0) {
107 fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(-ret));
108 abort();
109 }
110 s->clock = data.clock;
111
112 /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns
113 * essentially CLOCK_MONOTONIC plus a guest-specific adjustment. This
114 * can drift from the TSC-based value that is computed by the guest,
115 * so we need to go through kvmclock_current_nsec(). If
116 * kvm_has_adjust_clock_stable() is true, and the flags contain
117 * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value
118 * and kvmclock_current_nsec() is not necessary.
119 *
120 * Here, however, we need not check KVM_CLOCK_TSC_STABLE. This is because:
121 *
122 * - if the host has disabled the kvmclock master clock, the guest already
123 * has protection against time going backwards. This "safety net" is only
124 * absent when kvmclock is stable;
125 *
126 * - therefore, we can replace a check like
127 *
128 * if last KVM_GET_CLOCK was not reliable then
129 * read from memory
130 *
131 * with
132 *
133 * if last KVM_GET_CLOCK was not reliable && masterclock is enabled
134 * read from memory
135 *
136 * However:
137 *
138 * - if kvm_has_adjust_clock_stable() returns false, the left side is
139 * always true (KVM_GET_CLOCK is never reliable), and the right side is
140 * unknown (because we don't have data.flags). We must assume it's true
141 * and read from memory.
142 *
143 * - if kvm_has_adjust_clock_stable() returns true, the result of the &&
144 * is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable)
145 *
146 * So we can just use this instead:
147 *
148 * if !kvm_has_adjust_clock_stable() then
149 * read from memory
150 */
151 s->clock_is_reliable = kvm_has_adjust_clock_stable();
152 }
153
do_kvmclock_ctrl(CPUState * cpu,run_on_cpu_data data)154 static void do_kvmclock_ctrl(CPUState *cpu, run_on_cpu_data data)
155 {
156 int ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0);
157
158 if (ret && ret != -EINVAL) {
159 fprintf(stderr, "%s: %s\n", __func__, strerror(-ret));
160 }
161 }
162
kvmclock_vm_state_change(void * opaque,bool running,RunState state)163 static void kvmclock_vm_state_change(void *opaque, bool running,
164 RunState state)
165 {
166 KVMClockState *s = opaque;
167 CPUState *cpu;
168 int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
169 int ret;
170
171 if (running) {
172 struct kvm_clock_data data = {};
173
174 /*
175 * If the host where s->clock was read did not support reliable
176 * KVM_GET_CLOCK, read kvmclock value from memory.
177 */
178 if (!s->clock_is_reliable) {
179 uint64_t pvclock_via_mem = kvmclock_current_nsec(s);
180 /* We can't rely on the saved clock value, just discard it */
181 if (pvclock_via_mem) {
182 s->clock = pvclock_via_mem;
183 }
184 }
185
186 s->clock_valid = false;
187
188 data.clock = s->clock;
189 ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
190 if (ret < 0) {
191 fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(-ret));
192 abort();
193 }
194
195 if (!cap_clock_ctrl) {
196 return;
197 }
198 CPU_FOREACH(cpu) {
199 run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL);
200 }
201 } else {
202
203 if (s->clock_valid) {
204 return;
205 }
206
207 s->runstate_paused = runstate_check(RUN_STATE_PAUSED);
208
209 kvm_synchronize_all_tsc();
210
211 kvm_update_clock(s);
212 /*
213 * If the VM is stopped, declare the clock state valid to
214 * avoid re-reading it on next vmsave (which would return
215 * a different value). Will be reset when the VM is continued.
216 */
217 s->clock_valid = true;
218 }
219 }
220
kvmclock_realize(DeviceState * dev,Error ** errp)221 static void kvmclock_realize(DeviceState *dev, Error **errp)
222 {
223 KVMClockState *s = KVM_CLOCK(dev);
224
225 if (!kvm_enabled()) {
226 error_setg(errp, "kvmclock device requires KVM");
227 return;
228 }
229
230 kvm_update_clock(s);
231
232 qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
233 }
234
kvmclock_clock_is_reliable_needed(void * opaque)235 static bool kvmclock_clock_is_reliable_needed(void *opaque)
236 {
237 KVMClockState *s = opaque;
238
239 return s->mach_use_reliable_get_clock;
240 }
241
242 static const VMStateDescription kvmclock_reliable_get_clock = {
243 .name = "kvmclock/clock_is_reliable",
244 .version_id = 1,
245 .minimum_version_id = 1,
246 .needed = kvmclock_clock_is_reliable_needed,
247 .fields = (const VMStateField[]) {
248 VMSTATE_BOOL(clock_is_reliable, KVMClockState),
249 VMSTATE_END_OF_LIST()
250 }
251 };
252
253 /*
254 * When migrating, assume the source has an unreliable
255 * KVM_GET_CLOCK unless told otherwise.
256 */
kvmclock_pre_load(void * opaque)257 static int kvmclock_pre_load(void *opaque)
258 {
259 KVMClockState *s = opaque;
260
261 s->clock_is_reliable = false;
262
263 return 0;
264 }
265
266 /*
267 * When migrating a running guest, read the clock just
268 * before migration, so that the guest clock counts
269 * during the events between:
270 *
271 * * vm_stop()
272 * *
273 * * pre_save()
274 *
275 * This reduces kvmclock difference on migration from 5s
276 * to 0.1s (when max_downtime == 5s), because sending the
277 * final pages of memory (which happens between vm_stop()
278 * and pre_save()) takes max_downtime.
279 */
kvmclock_pre_save(void * opaque)280 static int kvmclock_pre_save(void *opaque)
281 {
282 KVMClockState *s = opaque;
283
284 if (!s->runstate_paused) {
285 kvm_update_clock(s);
286 }
287
288 return 0;
289 }
290
291 static const VMStateDescription kvmclock_vmsd = {
292 .name = "kvmclock",
293 .version_id = 1,
294 .minimum_version_id = 1,
295 .pre_load = kvmclock_pre_load,
296 .pre_save = kvmclock_pre_save,
297 .fields = (const VMStateField[]) {
298 VMSTATE_UINT64(clock, KVMClockState),
299 VMSTATE_END_OF_LIST()
300 },
301 .subsections = (const VMStateDescription * const []) {
302 &kvmclock_reliable_get_clock,
303 NULL
304 }
305 };
306
307 static const Property kvmclock_properties[] = {
308 DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState,
309 mach_use_reliable_get_clock, true),
310 };
311
kvmclock_class_init(ObjectClass * klass,void * data)312 static void kvmclock_class_init(ObjectClass *klass, void *data)
313 {
314 DeviceClass *dc = DEVICE_CLASS(klass);
315
316 dc->realize = kvmclock_realize;
317 dc->vmsd = &kvmclock_vmsd;
318 device_class_set_props(dc, kvmclock_properties);
319 }
320
321 static const TypeInfo kvmclock_info = {
322 .name = TYPE_KVM_CLOCK,
323 .parent = TYPE_SYS_BUS_DEVICE,
324 .instance_size = sizeof(KVMClockState),
325 .class_init = kvmclock_class_init,
326 };
327
328 /* Note: Must be called after VCPU initialization. */
kvmclock_create(bool create_always)329 void kvmclock_create(bool create_always)
330 {
331 X86CPU *cpu = X86_CPU(first_cpu);
332
333 assert(kvm_enabled());
334 if (create_always ||
335 cpu->env.features[FEAT_KVM] & (CPUID_KVM_CLOCK |
336 CPUID_KVM_CLOCK2)) {
337 sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL);
338 }
339 }
340
kvmclock_register_types(void)341 static void kvmclock_register_types(void)
342 {
343 type_register_static(&kvmclock_info);
344 }
345
346 type_init(kvmclock_register_types)
347