]>
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" |
33c11879 | 17 | #include "cpu.h" |
9a48bcd1 | 18 | #include "qemu/host-utils.h" |
0b8fa32f | 19 | #include "qemu/module.h" |
9c17d615 | 20 | #include "sysemu/kvm.h" |
54d31236 | 21 | #include "sysemu/runstate.h" |
e2b6c171 | 22 | #include "sysemu/hw_accel.h" |
0fd7e098 | 23 | #include "kvm_i386.h" |
d6454270 | 24 | #include "migration/vmstate.h" |
3b9a6ee5 JK |
25 | #include "hw/sysbus.h" |
26 | #include "hw/kvm/clock.h" | |
a27bd6c7 | 27 | #include "hw/qdev-properties.h" |
ca2edcd3 | 28 | #include "qapi/error.h" |
0ec329da | 29 | |
0ec329da | 30 | #include <linux/kvm.h> |
1814eab6 | 31 | #include "standard-headers/asm-x86/kvm_para.h" |
0ec329da | 32 | |
98bdc0d7 HT |
33 | #define TYPE_KVM_CLOCK "kvmclock" |
34 | #define KVM_CLOCK(obj) OBJECT_CHECK(KVMClockState, (obj), TYPE_KVM_CLOCK) | |
35 | ||
0ec329da | 36 | typedef 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 MT |
43 | |
44 | /* whether machine type supports reliable KVM_GET_CLOCK */ | |
45 | bool mach_use_reliable_get_clock; | |
46 | ||
47 | /* whether the 'clock' value was obtained in a host with | |
48 | * reliable KVM_GET_CLOCK */ | |
49 | bool clock_is_reliable; | |
0ec329da JK |
50 | } KVMClockState; |
51 | ||
9a48bcd1 AG |
52 | struct pvclock_vcpu_time_info { |
53 | uint32_t version; | |
54 | uint32_t pad0; | |
55 | uint64_t tsc_timestamp; | |
56 | uint64_t system_time; | |
57 | uint32_t tsc_to_system_mul; | |
58 | int8_t tsc_shift; | |
59 | uint8_t flags; | |
60 | uint8_t pad[2]; | |
61 | } __attribute__((__packed__)); /* 32 bytes */ | |
62 | ||
63 | static uint64_t kvmclock_current_nsec(KVMClockState *s) | |
64 | { | |
65 | CPUState *cpu = first_cpu; | |
66 | CPUX86State *env = cpu->env_ptr; | |
346b1215 | 67 | hwaddr kvmclock_struct_pa; |
9a48bcd1 AG |
68 | uint64_t migration_tsc = env->tsc; |
69 | struct pvclock_vcpu_time_info time; | |
70 | uint64_t delta; | |
71 | uint64_t nsec_lo; | |
72 | uint64_t nsec_hi; | |
73 | uint64_t nsec; | |
74 | ||
e2b6c171 DP |
75 | cpu_synchronize_state(cpu); |
76 | ||
9a48bcd1 AG |
77 | if (!(env->system_time_msr & 1ULL)) { |
78 | /* KVM clock not active */ | |
79 | return 0; | |
80 | } | |
81 | ||
346b1215 | 82 | kvmclock_struct_pa = env->system_time_msr & ~1ULL; |
9a48bcd1 AG |
83 | cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time)); |
84 | ||
85 | assert(time.tsc_timestamp <= migration_tsc); | |
86 | delta = migration_tsc - time.tsc_timestamp; | |
87 | if (time.tsc_shift < 0) { | |
88 | delta >>= -time.tsc_shift; | |
89 | } else { | |
90 | delta <<= time.tsc_shift; | |
91 | } | |
92 | ||
93 | mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul); | |
94 | nsec = (nsec_lo >> 32) | (nsec_hi << 32); | |
95 | return nsec + time.system_time; | |
96 | } | |
0ec329da | 97 | |
6053a86f MT |
98 | static void kvm_update_clock(KVMClockState *s) |
99 | { | |
100 | struct kvm_clock_data data; | |
101 | int ret; | |
102 | ||
103 | ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data); | |
104 | if (ret < 0) { | |
105 | fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret)); | |
106 | abort(); | |
107 | } | |
108 | s->clock = data.clock; | |
109 | ||
110 | /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns | |
111 | * essentially CLOCK_MONOTONIC plus a guest-specific adjustment. This | |
112 | * can drift from the TSC-based value that is computed by the guest, | |
113 | * so we need to go through kvmclock_current_nsec(). If | |
114 | * kvm_has_adjust_clock_stable() is true, and the flags contain | |
115 | * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value | |
116 | * and kvmclock_current_nsec() is not necessary. | |
117 | * | |
118 | * Here, however, we need not check KVM_CLOCK_TSC_STABLE. This is because: | |
119 | * | |
120 | * - if the host has disabled the kvmclock master clock, the guest already | |
121 | * has protection against time going backwards. This "safety net" is only | |
122 | * absent when kvmclock is stable; | |
123 | * | |
124 | * - therefore, we can replace a check like | |
125 | * | |
126 | * if last KVM_GET_CLOCK was not reliable then | |
127 | * read from memory | |
128 | * | |
129 | * with | |
130 | * | |
131 | * if last KVM_GET_CLOCK was not reliable && masterclock is enabled | |
132 | * read from memory | |
133 | * | |
134 | * However: | |
135 | * | |
136 | * - if kvm_has_adjust_clock_stable() returns false, the left side is | |
137 | * always true (KVM_GET_CLOCK is never reliable), and the right side is | |
138 | * unknown (because we don't have data.flags). We must assume it's true | |
139 | * and read from memory. | |
140 | * | |
141 | * - if kvm_has_adjust_clock_stable() returns true, the result of the && | |
142 | * is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable) | |
143 | * | |
144 | * So we can just use this instead: | |
145 | * | |
146 | * if !kvm_has_adjust_clock_stable() then | |
147 | * read from memory | |
148 | */ | |
149 | s->clock_is_reliable = kvm_has_adjust_clock_stable(); | |
150 | } | |
151 | ||
4c3e2506 YX |
152 | static void do_kvmclock_ctrl(CPUState *cpu, run_on_cpu_data data) |
153 | { | |
154 | int ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0); | |
155 | ||
156 | if (ret && ret != -EINVAL) { | |
157 | fprintf(stderr, "%s: %s\n", __func__, strerror(-ret)); | |
158 | } | |
159 | } | |
160 | ||
1dfb4dd9 LC |
161 | static void kvmclock_vm_state_change(void *opaque, int running, |
162 | RunState state) | |
0ec329da JK |
163 | { |
164 | KVMClockState *s = opaque; | |
e76d05c2 | 165 | CPUState *cpu; |
f349c12c EM |
166 | int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL); |
167 | int ret; | |
0ec329da JK |
168 | |
169 | if (running) { | |
5e0b7d88 | 170 | struct kvm_clock_data data = {}; |
f349c12c | 171 | |
6053a86f MT |
172 | /* |
173 | * If the host where s->clock was read did not support reliable | |
174 | * KVM_GET_CLOCK, read kvmclock value from memory. | |
175 | */ | |
176 | if (!s->clock_is_reliable) { | |
177 | uint64_t pvclock_via_mem = kvmclock_current_nsec(s); | |
178 | /* We can't rely on the saved clock value, just discard it */ | |
179 | if (pvclock_via_mem) { | |
180 | s->clock = pvclock_via_mem; | |
181 | } | |
9a48bcd1 AG |
182 | } |
183 | ||
6053a86f MT |
184 | s->clock_valid = false; |
185 | ||
00f4d64e | 186 | data.clock = s->clock; |
00f4d64e MT |
187 | ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data); |
188 | if (ret < 0) { | |
189 | fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret)); | |
190 | abort(); | |
191 | } | |
192 | ||
f349c12c EM |
193 | if (!cap_clock_ctrl) { |
194 | return; | |
195 | } | |
bdc44640 | 196 | CPU_FOREACH(cpu) { |
4c3e2506 | 197 | run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL); |
f349c12c | 198 | } |
00f4d64e | 199 | } else { |
00f4d64e MT |
200 | |
201 | if (s->clock_valid) { | |
202 | return; | |
203 | } | |
317b0a6d | 204 | |
0fd7e098 | 205 | kvm_synchronize_all_tsc(); |
1154d84d | 206 | |
6053a86f | 207 | kvm_update_clock(s); |
00f4d64e MT |
208 | /* |
209 | * If the VM is stopped, declare the clock state valid to | |
210 | * avoid re-reading it on next vmsave (which would return | |
211 | * a different value). Will be reset when the VM is continued. | |
212 | */ | |
213 | s->clock_valid = true; | |
0ec329da JK |
214 | } |
215 | } | |
216 | ||
913bc638 | 217 | static void kvmclock_realize(DeviceState *dev, Error **errp) |
0ec329da | 218 | { |
98bdc0d7 | 219 | KVMClockState *s = KVM_CLOCK(dev); |
0ec329da | 220 | |
ca2edcd3 EH |
221 | if (!kvm_enabled()) { |
222 | error_setg(errp, "kvmclock device requires KVM"); | |
223 | return; | |
224 | } | |
225 | ||
6053a86f MT |
226 | kvm_update_clock(s); |
227 | ||
0ec329da | 228 | qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s); |
0ec329da JK |
229 | } |
230 | ||
6053a86f MT |
231 | static bool kvmclock_clock_is_reliable_needed(void *opaque) |
232 | { | |
233 | KVMClockState *s = opaque; | |
234 | ||
235 | return s->mach_use_reliable_get_clock; | |
236 | } | |
237 | ||
238 | static const VMStateDescription kvmclock_reliable_get_clock = { | |
239 | .name = "kvmclock/clock_is_reliable", | |
240 | .version_id = 1, | |
241 | .minimum_version_id = 1, | |
242 | .needed = kvmclock_clock_is_reliable_needed, | |
243 | .fields = (VMStateField[]) { | |
244 | VMSTATE_BOOL(clock_is_reliable, KVMClockState), | |
245 | VMSTATE_END_OF_LIST() | |
246 | } | |
247 | }; | |
248 | ||
c2b01cfe MC |
249 | /* |
250 | * When migrating, assume the source has an unreliable | |
251 | * KVM_GET_CLOCK unless told otherwise. | |
252 | */ | |
253 | static int kvmclock_pre_load(void *opaque) | |
254 | { | |
255 | KVMClockState *s = opaque; | |
256 | ||
257 | s->clock_is_reliable = false; | |
258 | ||
259 | return 0; | |
260 | } | |
261 | ||
6053a86f MT |
262 | /* |
263 | * When migrating, read the clock just before migration, | |
264 | * so that the guest clock counts during the events | |
265 | * between: | |
266 | * | |
267 | * * vm_stop() | |
268 | * * | |
269 | * * pre_save() | |
270 | * | |
271 | * This reduces kvmclock difference on migration from 5s | |
272 | * to 0.1s (when max_downtime == 5s), because sending the | |
273 | * final pages of memory (which happens between vm_stop() | |
274 | * and pre_save()) takes max_downtime. | |
275 | */ | |
44b1ff31 | 276 | static int kvmclock_pre_save(void *opaque) |
6053a86f MT |
277 | { |
278 | KVMClockState *s = opaque; | |
279 | ||
280 | kvm_update_clock(s); | |
44b1ff31 DDAG |
281 | |
282 | return 0; | |
6053a86f MT |
283 | } |
284 | ||
0ec329da JK |
285 | static const VMStateDescription kvmclock_vmsd = { |
286 | .name = "kvmclock", | |
287 | .version_id = 1, | |
288 | .minimum_version_id = 1, | |
c2b01cfe | 289 | .pre_load = kvmclock_pre_load, |
6053a86f | 290 | .pre_save = kvmclock_pre_save, |
0ec329da JK |
291 | .fields = (VMStateField[]) { |
292 | VMSTATE_UINT64(clock, KVMClockState), | |
293 | VMSTATE_END_OF_LIST() | |
6053a86f MT |
294 | }, |
295 | .subsections = (const VMStateDescription * []) { | |
296 | &kvmclock_reliable_get_clock, | |
297 | NULL | |
0ec329da JK |
298 | } |
299 | }; | |
300 | ||
6053a86f MT |
301 | static Property kvmclock_properties[] = { |
302 | DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState, | |
303 | mach_use_reliable_get_clock, true), | |
304 | DEFINE_PROP_END_OF_LIST(), | |
305 | }; | |
306 | ||
999e12bb AL |
307 | static void kvmclock_class_init(ObjectClass *klass, void *data) |
308 | { | |
39bffca2 | 309 | DeviceClass *dc = DEVICE_CLASS(klass); |
999e12bb | 310 | |
913bc638 | 311 | dc->realize = kvmclock_realize; |
39bffca2 | 312 | dc->vmsd = &kvmclock_vmsd; |
6053a86f | 313 | dc->props = kvmclock_properties; |
999e12bb AL |
314 | } |
315 | ||
8c43a6f0 | 316 | static const TypeInfo kvmclock_info = { |
98bdc0d7 | 317 | .name = TYPE_KVM_CLOCK, |
39bffca2 AL |
318 | .parent = TYPE_SYS_BUS_DEVICE, |
319 | .instance_size = sizeof(KVMClockState), | |
320 | .class_init = kvmclock_class_init, | |
0ec329da JK |
321 | }; |
322 | ||
323 | /* Note: Must be called after VCPU initialization. */ | |
324 | void kvmclock_create(void) | |
325 | { | |
182735ef AF |
326 | X86CPU *cpu = X86_CPU(first_cpu); |
327 | ||
0ec329da | 328 | if (kvm_enabled() && |
182735ef AF |
329 | cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) | |
330 | (1ULL << KVM_FEATURE_CLOCKSOURCE2))) { | |
98bdc0d7 | 331 | sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL); |
0ec329da JK |
332 | } |
333 | } | |
334 | ||
83f7d43a | 335 | static void kvmclock_register_types(void) |
0ec329da | 336 | { |
39bffca2 | 337 | type_register_static(&kvmclock_info); |
0ec329da JK |
338 | } |
339 | ||
83f7d43a | 340 | type_init(kvmclock_register_types) |