]>
Commit | Line | Data |
---|---|---|
0ec329da JK |
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 | * | |
6b620ca3 PB |
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. | |
0ec329da JK |
14 | */ |
15 | ||
b6a0aa05 | 16 | #include "qemu/osdep.h" |
9a48bcd1 | 17 | #include "qemu/host-utils.h" |
0b8fa32f | 18 | #include "qemu/module.h" |
9c17d615 | 19 | #include "sysemu/kvm.h" |
54d31236 | 20 | #include "sysemu/runstate.h" |
e2b6c171 | 21 | #include "sysemu/hw_accel.h" |
a9dc68d9 | 22 | #include "kvm/kvm_i386.h" |
d6454270 | 23 | #include "migration/vmstate.h" |
3b9a6ee5 JK |
24 | #include "hw/sysbus.h" |
25 | #include "hw/kvm/clock.h" | |
a27bd6c7 | 26 | #include "hw/qdev-properties.h" |
ca2edcd3 | 27 | #include "qapi/error.h" |
0ec329da | 28 | |
0ec329da | 29 | #include <linux/kvm.h> |
1814eab6 | 30 | #include "standard-headers/asm-x86/kvm_para.h" |
db1015e9 | 31 | #include "qom/object.h" |
0ec329da | 32 | |
98bdc0d7 | 33 | #define TYPE_KVM_CLOCK "kvmclock" |
8063396b | 34 | OBJECT_DECLARE_SIMPLE_TYPE(KVMClockState, KVM_CLOCK) |
98bdc0d7 | 35 | |
db1015e9 | 36 | struct KVMClockState { |
98bdc0d7 | 37 | /*< private >*/ |
0ec329da | 38 | SysBusDevice busdev; |
98bdc0d7 HT |
39 | /*< public >*/ |
40 | ||
0ec329da JK |
41 | uint64_t clock; |
42 | bool clock_valid; | |
6053a86f | 43 | |
41733249 MG |
44 | /* whether the 'clock' value was obtained in the 'paused' state */ |
45 | bool runstate_paused; | |
46 | ||
6053a86f MT |
47 | /* whether machine type supports reliable KVM_GET_CLOCK */ |
48 | bool mach_use_reliable_get_clock; | |
49 | ||
50 | /* whether the 'clock' value was obtained in a host with | |
51 | * reliable KVM_GET_CLOCK */ | |
52 | bool clock_is_reliable; | |
db1015e9 | 53 | }; |
0ec329da | 54 | |
9a48bcd1 AG |
55 | struct pvclock_vcpu_time_info { |
56 | uint32_t version; | |
57 | uint32_t pad0; | |
58 | uint64_t tsc_timestamp; | |
59 | uint64_t system_time; | |
60 | uint32_t tsc_to_system_mul; | |
61 | int8_t tsc_shift; | |
62 | uint8_t flags; | |
63 | uint8_t pad[2]; | |
64 | } __attribute__((__packed__)); /* 32 bytes */ | |
65 | ||
66 | static uint64_t kvmclock_current_nsec(KVMClockState *s) | |
67 | { | |
68 | CPUState *cpu = first_cpu; | |
69 | CPUX86State *env = cpu->env_ptr; | |
346b1215 | 70 | hwaddr kvmclock_struct_pa; |
9a48bcd1 AG |
71 | uint64_t migration_tsc = env->tsc; |
72 | struct pvclock_vcpu_time_info time; | |
73 | uint64_t delta; | |
74 | uint64_t nsec_lo; | |
75 | uint64_t nsec_hi; | |
76 | uint64_t nsec; | |
77 | ||
e2b6c171 DP |
78 | cpu_synchronize_state(cpu); |
79 | ||
9a48bcd1 AG |
80 | if (!(env->system_time_msr & 1ULL)) { |
81 | /* KVM clock not active */ | |
82 | return 0; | |
83 | } | |
84 | ||
346b1215 | 85 | kvmclock_struct_pa = env->system_time_msr & ~1ULL; |
9a48bcd1 AG |
86 | cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time)); |
87 | ||
88 | assert(time.tsc_timestamp <= migration_tsc); | |
89 | delta = migration_tsc - time.tsc_timestamp; | |
90 | if (time.tsc_shift < 0) { | |
91 | delta >>= -time.tsc_shift; | |
92 | } else { | |
93 | delta <<= time.tsc_shift; | |
94 | } | |
95 | ||
96 | mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul); | |
97 | nsec = (nsec_lo >> 32) | (nsec_hi << 32); | |
98 | return nsec + time.system_time; | |
99 | } | |
0ec329da | 100 | |
6053a86f MT |
101 | static void kvm_update_clock(KVMClockState *s) |
102 | { | |
103 | struct kvm_clock_data data; | |
104 | int ret; | |
105 | ||
106 | ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data); | |
107 | if (ret < 0) { | |
108 | fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret)); | |
109 | abort(); | |
110 | } | |
111 | s->clock = data.clock; | |
112 | ||
113 | /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns | |
114 | * essentially CLOCK_MONOTONIC plus a guest-specific adjustment. This | |
115 | * can drift from the TSC-based value that is computed by the guest, | |
116 | * so we need to go through kvmclock_current_nsec(). If | |
117 | * kvm_has_adjust_clock_stable() is true, and the flags contain | |
118 | * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value | |
119 | * and kvmclock_current_nsec() is not necessary. | |
120 | * | |
121 | * Here, however, we need not check KVM_CLOCK_TSC_STABLE. This is because: | |
122 | * | |
123 | * - if the host has disabled the kvmclock master clock, the guest already | |
124 | * has protection against time going backwards. This "safety net" is only | |
125 | * absent when kvmclock is stable; | |
126 | * | |
127 | * - therefore, we can replace a check like | |
128 | * | |
129 | * if last KVM_GET_CLOCK was not reliable then | |
130 | * read from memory | |
131 | * | |
132 | * with | |
133 | * | |
134 | * if last KVM_GET_CLOCK was not reliable && masterclock is enabled | |
135 | * read from memory | |
136 | * | |
137 | * However: | |
138 | * | |
139 | * - if kvm_has_adjust_clock_stable() returns false, the left side is | |
140 | * always true (KVM_GET_CLOCK is never reliable), and the right side is | |
141 | * unknown (because we don't have data.flags). We must assume it's true | |
142 | * and read from memory. | |
143 | * | |
144 | * - if kvm_has_adjust_clock_stable() returns true, the result of the && | |
145 | * is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable) | |
146 | * | |
147 | * So we can just use this instead: | |
148 | * | |
149 | * if !kvm_has_adjust_clock_stable() then | |
150 | * read from memory | |
151 | */ | |
152 | s->clock_is_reliable = kvm_has_adjust_clock_stable(); | |
153 | } | |
154 | ||
4c3e2506 YX |
155 | static void do_kvmclock_ctrl(CPUState *cpu, run_on_cpu_data data) |
156 | { | |
157 | int ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0); | |
158 | ||
159 | if (ret && ret != -EINVAL) { | |
160 | fprintf(stderr, "%s: %s\n", __func__, strerror(-ret)); | |
161 | } | |
162 | } | |
163 | ||
538f0497 | 164 | static void kvmclock_vm_state_change(void *opaque, bool running, |
1dfb4dd9 | 165 | RunState state) |
0ec329da JK |
166 | { |
167 | KVMClockState *s = opaque; | |
e76d05c2 | 168 | CPUState *cpu; |
f349c12c EM |
169 | int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL); |
170 | int ret; | |
0ec329da JK |
171 | |
172 | if (running) { | |
5e0b7d88 | 173 | struct kvm_clock_data data = {}; |
f349c12c | 174 | |
6053a86f MT |
175 | /* |
176 | * If the host where s->clock was read did not support reliable | |
177 | * KVM_GET_CLOCK, read kvmclock value from memory. | |
178 | */ | |
179 | if (!s->clock_is_reliable) { | |
180 | uint64_t pvclock_via_mem = kvmclock_current_nsec(s); | |
181 | /* We can't rely on the saved clock value, just discard it */ | |
182 | if (pvclock_via_mem) { | |
183 | s->clock = pvclock_via_mem; | |
184 | } | |
9a48bcd1 AG |
185 | } |
186 | ||
6053a86f MT |
187 | s->clock_valid = false; |
188 | ||
00f4d64e | 189 | data.clock = s->clock; |
00f4d64e MT |
190 | ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data); |
191 | if (ret < 0) { | |
192 | fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret)); | |
193 | abort(); | |
194 | } | |
195 | ||
f349c12c EM |
196 | if (!cap_clock_ctrl) { |
197 | return; | |
198 | } | |
bdc44640 | 199 | CPU_FOREACH(cpu) { |
4c3e2506 | 200 | run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL); |
f349c12c | 201 | } |
00f4d64e | 202 | } else { |
00f4d64e MT |
203 | |
204 | if (s->clock_valid) { | |
205 | return; | |
206 | } | |
317b0a6d | 207 | |
41733249 MG |
208 | s->runstate_paused = runstate_check(RUN_STATE_PAUSED); |
209 | ||
0fd7e098 | 210 | kvm_synchronize_all_tsc(); |
1154d84d | 211 | |
6053a86f | 212 | kvm_update_clock(s); |
00f4d64e MT |
213 | /* |
214 | * If the VM is stopped, declare the clock state valid to | |
215 | * avoid re-reading it on next vmsave (which would return | |
216 | * a different value). Will be reset when the VM is continued. | |
217 | */ | |
218 | s->clock_valid = true; | |
0ec329da JK |
219 | } |
220 | } | |
221 | ||
913bc638 | 222 | static void kvmclock_realize(DeviceState *dev, Error **errp) |
0ec329da | 223 | { |
98bdc0d7 | 224 | KVMClockState *s = KVM_CLOCK(dev); |
0ec329da | 225 | |
ca2edcd3 EH |
226 | if (!kvm_enabled()) { |
227 | error_setg(errp, "kvmclock device requires KVM"); | |
228 | return; | |
229 | } | |
230 | ||
6053a86f MT |
231 | kvm_update_clock(s); |
232 | ||
0ec329da | 233 | qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s); |
0ec329da JK |
234 | } |
235 | ||
6053a86f MT |
236 | static bool kvmclock_clock_is_reliable_needed(void *opaque) |
237 | { | |
238 | KVMClockState *s = opaque; | |
239 | ||
240 | return s->mach_use_reliable_get_clock; | |
241 | } | |
242 | ||
243 | static const VMStateDescription kvmclock_reliable_get_clock = { | |
244 | .name = "kvmclock/clock_is_reliable", | |
245 | .version_id = 1, | |
246 | .minimum_version_id = 1, | |
247 | .needed = kvmclock_clock_is_reliable_needed, | |
248 | .fields = (VMStateField[]) { | |
249 | VMSTATE_BOOL(clock_is_reliable, KVMClockState), | |
250 | VMSTATE_END_OF_LIST() | |
251 | } | |
252 | }; | |
253 | ||
c2b01cfe MC |
254 | /* |
255 | * When migrating, assume the source has an unreliable | |
256 | * KVM_GET_CLOCK unless told otherwise. | |
257 | */ | |
258 | static int kvmclock_pre_load(void *opaque) | |
259 | { | |
260 | KVMClockState *s = opaque; | |
261 | ||
262 | s->clock_is_reliable = false; | |
263 | ||
264 | return 0; | |
265 | } | |
266 | ||
6053a86f | 267 | /* |
41733249 MG |
268 | * When migrating a running guest, read the clock just |
269 | * before migration, so that the guest clock counts | |
270 | * during the events between: | |
6053a86f MT |
271 | * |
272 | * * vm_stop() | |
273 | * * | |
274 | * * pre_save() | |
275 | * | |
276 | * This reduces kvmclock difference on migration from 5s | |
277 | * to 0.1s (when max_downtime == 5s), because sending the | |
278 | * final pages of memory (which happens between vm_stop() | |
279 | * and pre_save()) takes max_downtime. | |
280 | */ | |
44b1ff31 | 281 | static int kvmclock_pre_save(void *opaque) |
6053a86f MT |
282 | { |
283 | KVMClockState *s = opaque; | |
284 | ||
41733249 MG |
285 | if (!s->runstate_paused) { |
286 | kvm_update_clock(s); | |
287 | } | |
44b1ff31 DDAG |
288 | |
289 | return 0; | |
6053a86f MT |
290 | } |
291 | ||
0ec329da JK |
292 | static const VMStateDescription kvmclock_vmsd = { |
293 | .name = "kvmclock", | |
294 | .version_id = 1, | |
295 | .minimum_version_id = 1, | |
c2b01cfe | 296 | .pre_load = kvmclock_pre_load, |
6053a86f | 297 | .pre_save = kvmclock_pre_save, |
0ec329da JK |
298 | .fields = (VMStateField[]) { |
299 | VMSTATE_UINT64(clock, KVMClockState), | |
300 | VMSTATE_END_OF_LIST() | |
6053a86f MT |
301 | }, |
302 | .subsections = (const VMStateDescription * []) { | |
303 | &kvmclock_reliable_get_clock, | |
304 | NULL | |
0ec329da JK |
305 | } |
306 | }; | |
307 | ||
6053a86f MT |
308 | static Property kvmclock_properties[] = { |
309 | DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState, | |
310 | mach_use_reliable_get_clock, true), | |
311 | DEFINE_PROP_END_OF_LIST(), | |
312 | }; | |
313 | ||
999e12bb AL |
314 | static void kvmclock_class_init(ObjectClass *klass, void *data) |
315 | { | |
39bffca2 | 316 | DeviceClass *dc = DEVICE_CLASS(klass); |
999e12bb | 317 | |
913bc638 | 318 | dc->realize = kvmclock_realize; |
39bffca2 | 319 | dc->vmsd = &kvmclock_vmsd; |
4f67d30b | 320 | device_class_set_props(dc, kvmclock_properties); |
999e12bb AL |
321 | } |
322 | ||
8c43a6f0 | 323 | static const TypeInfo kvmclock_info = { |
98bdc0d7 | 324 | .name = TYPE_KVM_CLOCK, |
39bffca2 AL |
325 | .parent = TYPE_SYS_BUS_DEVICE, |
326 | .instance_size = sizeof(KVMClockState), | |
327 | .class_init = kvmclock_class_init, | |
0ec329da JK |
328 | }; |
329 | ||
330 | /* Note: Must be called after VCPU initialization. */ | |
8700a984 | 331 | void kvmclock_create(bool create_always) |
0ec329da | 332 | { |
182735ef AF |
333 | X86CPU *cpu = X86_CPU(first_cpu); |
334 | ||
8700a984 VK |
335 | if (!kvm_enabled() || !kvm_has_adjust_clock()) |
336 | return; | |
337 | ||
338 | if (create_always || | |
182735ef AF |
339 | cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) | |
340 | (1ULL << KVM_FEATURE_CLOCKSOURCE2))) { | |
98bdc0d7 | 341 | sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL); |
0ec329da JK |
342 | } |
343 | } | |
344 | ||
83f7d43a | 345 | static void kvmclock_register_types(void) |
0ec329da | 346 | { |
39bffca2 | 347 | type_register_static(&kvmclock_info); |
0ec329da JK |
348 | } |
349 | ||
83f7d43a | 350 | type_init(kvmclock_register_types) |