]>
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" |
0ec329da | 17 | #include "qemu-common.h" |
33c11879 | 18 | #include "cpu.h" |
9a48bcd1 | 19 | #include "qemu/host-utils.h" |
9c17d615 PB |
20 | #include "sysemu/sysemu.h" |
21 | #include "sysemu/kvm.h" | |
0fd7e098 | 22 | #include "kvm_i386.h" |
3b9a6ee5 JK |
23 | #include "hw/sysbus.h" |
24 | #include "hw/kvm/clock.h" | |
0ec329da | 25 | |
0ec329da JK |
26 | #include <linux/kvm.h> |
27 | #include <linux/kvm_para.h> | |
28 | ||
98bdc0d7 HT |
29 | #define TYPE_KVM_CLOCK "kvmclock" |
30 | #define KVM_CLOCK(obj) OBJECT_CHECK(KVMClockState, (obj), TYPE_KVM_CLOCK) | |
31 | ||
0ec329da | 32 | typedef struct KVMClockState { |
98bdc0d7 | 33 | /*< private >*/ |
0ec329da | 34 | SysBusDevice busdev; |
98bdc0d7 HT |
35 | /*< public >*/ |
36 | ||
0ec329da JK |
37 | uint64_t clock; |
38 | bool clock_valid; | |
6053a86f MT |
39 | |
40 | /* whether machine type supports reliable KVM_GET_CLOCK */ | |
41 | bool mach_use_reliable_get_clock; | |
42 | ||
43 | /* whether the 'clock' value was obtained in a host with | |
44 | * reliable KVM_GET_CLOCK */ | |
45 | bool clock_is_reliable; | |
0ec329da JK |
46 | } KVMClockState; |
47 | ||
9a48bcd1 AG |
48 | struct pvclock_vcpu_time_info { |
49 | uint32_t version; | |
50 | uint32_t pad0; | |
51 | uint64_t tsc_timestamp; | |
52 | uint64_t system_time; | |
53 | uint32_t tsc_to_system_mul; | |
54 | int8_t tsc_shift; | |
55 | uint8_t flags; | |
56 | uint8_t pad[2]; | |
57 | } __attribute__((__packed__)); /* 32 bytes */ | |
58 | ||
59 | static uint64_t kvmclock_current_nsec(KVMClockState *s) | |
60 | { | |
61 | CPUState *cpu = first_cpu; | |
62 | CPUX86State *env = cpu->env_ptr; | |
63 | hwaddr kvmclock_struct_pa = env->system_time_msr & ~1ULL; | |
64 | uint64_t migration_tsc = env->tsc; | |
65 | struct pvclock_vcpu_time_info time; | |
66 | uint64_t delta; | |
67 | uint64_t nsec_lo; | |
68 | uint64_t nsec_hi; | |
69 | uint64_t nsec; | |
70 | ||
71 | if (!(env->system_time_msr & 1ULL)) { | |
72 | /* KVM clock not active */ | |
73 | return 0; | |
74 | } | |
75 | ||
76 | cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time)); | |
77 | ||
78 | assert(time.tsc_timestamp <= migration_tsc); | |
79 | delta = migration_tsc - time.tsc_timestamp; | |
80 | if (time.tsc_shift < 0) { | |
81 | delta >>= -time.tsc_shift; | |
82 | } else { | |
83 | delta <<= time.tsc_shift; | |
84 | } | |
85 | ||
86 | mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul); | |
87 | nsec = (nsec_lo >> 32) | (nsec_hi << 32); | |
88 | return nsec + time.system_time; | |
89 | } | |
0ec329da | 90 | |
6053a86f MT |
91 | static void kvm_update_clock(KVMClockState *s) |
92 | { | |
93 | struct kvm_clock_data data; | |
94 | int ret; | |
95 | ||
96 | ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data); | |
97 | if (ret < 0) { | |
98 | fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret)); | |
99 | abort(); | |
100 | } | |
101 | s->clock = data.clock; | |
102 | ||
103 | /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns | |
104 | * essentially CLOCK_MONOTONIC plus a guest-specific adjustment. This | |
105 | * can drift from the TSC-based value that is computed by the guest, | |
106 | * so we need to go through kvmclock_current_nsec(). If | |
107 | * kvm_has_adjust_clock_stable() is true, and the flags contain | |
108 | * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value | |
109 | * and kvmclock_current_nsec() is not necessary. | |
110 | * | |
111 | * Here, however, we need not check KVM_CLOCK_TSC_STABLE. This is because: | |
112 | * | |
113 | * - if the host has disabled the kvmclock master clock, the guest already | |
114 | * has protection against time going backwards. This "safety net" is only | |
115 | * absent when kvmclock is stable; | |
116 | * | |
117 | * - therefore, we can replace a check like | |
118 | * | |
119 | * if last KVM_GET_CLOCK was not reliable then | |
120 | * read from memory | |
121 | * | |
122 | * with | |
123 | * | |
124 | * if last KVM_GET_CLOCK was not reliable && masterclock is enabled | |
125 | * read from memory | |
126 | * | |
127 | * However: | |
128 | * | |
129 | * - if kvm_has_adjust_clock_stable() returns false, the left side is | |
130 | * always true (KVM_GET_CLOCK is never reliable), and the right side is | |
131 | * unknown (because we don't have data.flags). We must assume it's true | |
132 | * and read from memory. | |
133 | * | |
134 | * - if kvm_has_adjust_clock_stable() returns true, the result of the && | |
135 | * is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable) | |
136 | * | |
137 | * So we can just use this instead: | |
138 | * | |
139 | * if !kvm_has_adjust_clock_stable() then | |
140 | * read from memory | |
141 | */ | |
142 | s->clock_is_reliable = kvm_has_adjust_clock_stable(); | |
143 | } | |
144 | ||
1dfb4dd9 LC |
145 | static void kvmclock_vm_state_change(void *opaque, int running, |
146 | RunState state) | |
0ec329da JK |
147 | { |
148 | KVMClockState *s = opaque; | |
e76d05c2 | 149 | CPUState *cpu; |
f349c12c EM |
150 | int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL); |
151 | int ret; | |
0ec329da JK |
152 | |
153 | if (running) { | |
5e0b7d88 | 154 | struct kvm_clock_data data = {}; |
f349c12c | 155 | |
6053a86f MT |
156 | /* |
157 | * If the host where s->clock was read did not support reliable | |
158 | * KVM_GET_CLOCK, read kvmclock value from memory. | |
159 | */ | |
160 | if (!s->clock_is_reliable) { | |
161 | uint64_t pvclock_via_mem = kvmclock_current_nsec(s); | |
162 | /* We can't rely on the saved clock value, just discard it */ | |
163 | if (pvclock_via_mem) { | |
164 | s->clock = pvclock_via_mem; | |
165 | } | |
9a48bcd1 AG |
166 | } |
167 | ||
6053a86f MT |
168 | s->clock_valid = false; |
169 | ||
00f4d64e | 170 | data.clock = s->clock; |
00f4d64e MT |
171 | ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data); |
172 | if (ret < 0) { | |
173 | fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret)); | |
174 | abort(); | |
175 | } | |
176 | ||
f349c12c EM |
177 | if (!cap_clock_ctrl) { |
178 | return; | |
179 | } | |
bdc44640 | 180 | CPU_FOREACH(cpu) { |
182735ef | 181 | ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0); |
f349c12c EM |
182 | if (ret) { |
183 | if (ret != -EINVAL) { | |
184 | fprintf(stderr, "%s: %s\n", __func__, strerror(-ret)); | |
185 | } | |
186 | return; | |
187 | } | |
188 | } | |
00f4d64e | 189 | } else { |
00f4d64e MT |
190 | |
191 | if (s->clock_valid) { | |
192 | return; | |
193 | } | |
317b0a6d | 194 | |
0fd7e098 | 195 | kvm_synchronize_all_tsc(); |
1154d84d | 196 | |
6053a86f | 197 | kvm_update_clock(s); |
00f4d64e MT |
198 | /* |
199 | * If the VM is stopped, declare the clock state valid to | |
200 | * avoid re-reading it on next vmsave (which would return | |
201 | * a different value). Will be reset when the VM is continued. | |
202 | */ | |
203 | s->clock_valid = true; | |
0ec329da JK |
204 | } |
205 | } | |
206 | ||
913bc638 | 207 | static void kvmclock_realize(DeviceState *dev, Error **errp) |
0ec329da | 208 | { |
98bdc0d7 | 209 | KVMClockState *s = KVM_CLOCK(dev); |
0ec329da | 210 | |
6053a86f MT |
211 | kvm_update_clock(s); |
212 | ||
0ec329da | 213 | qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s); |
0ec329da JK |
214 | } |
215 | ||
6053a86f MT |
216 | static bool kvmclock_clock_is_reliable_needed(void *opaque) |
217 | { | |
218 | KVMClockState *s = opaque; | |
219 | ||
220 | return s->mach_use_reliable_get_clock; | |
221 | } | |
222 | ||
223 | static const VMStateDescription kvmclock_reliable_get_clock = { | |
224 | .name = "kvmclock/clock_is_reliable", | |
225 | .version_id = 1, | |
226 | .minimum_version_id = 1, | |
227 | .needed = kvmclock_clock_is_reliable_needed, | |
228 | .fields = (VMStateField[]) { | |
229 | VMSTATE_BOOL(clock_is_reliable, KVMClockState), | |
230 | VMSTATE_END_OF_LIST() | |
231 | } | |
232 | }; | |
233 | ||
234 | /* | |
235 | * When migrating, read the clock just before migration, | |
236 | * so that the guest clock counts during the events | |
237 | * between: | |
238 | * | |
239 | * * vm_stop() | |
240 | * * | |
241 | * * pre_save() | |
242 | * | |
243 | * This reduces kvmclock difference on migration from 5s | |
244 | * to 0.1s (when max_downtime == 5s), because sending the | |
245 | * final pages of memory (which happens between vm_stop() | |
246 | * and pre_save()) takes max_downtime. | |
247 | */ | |
248 | static void kvmclock_pre_save(void *opaque) | |
249 | { | |
250 | KVMClockState *s = opaque; | |
251 | ||
252 | kvm_update_clock(s); | |
253 | } | |
254 | ||
0ec329da JK |
255 | static const VMStateDescription kvmclock_vmsd = { |
256 | .name = "kvmclock", | |
257 | .version_id = 1, | |
258 | .minimum_version_id = 1, | |
6053a86f | 259 | .pre_save = kvmclock_pre_save, |
0ec329da JK |
260 | .fields = (VMStateField[]) { |
261 | VMSTATE_UINT64(clock, KVMClockState), | |
262 | VMSTATE_END_OF_LIST() | |
6053a86f MT |
263 | }, |
264 | .subsections = (const VMStateDescription * []) { | |
265 | &kvmclock_reliable_get_clock, | |
266 | NULL | |
0ec329da JK |
267 | } |
268 | }; | |
269 | ||
6053a86f MT |
270 | static Property kvmclock_properties[] = { |
271 | DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState, | |
272 | mach_use_reliable_get_clock, true), | |
273 | DEFINE_PROP_END_OF_LIST(), | |
274 | }; | |
275 | ||
999e12bb AL |
276 | static void kvmclock_class_init(ObjectClass *klass, void *data) |
277 | { | |
39bffca2 | 278 | DeviceClass *dc = DEVICE_CLASS(klass); |
999e12bb | 279 | |
913bc638 | 280 | dc->realize = kvmclock_realize; |
39bffca2 | 281 | dc->vmsd = &kvmclock_vmsd; |
6053a86f | 282 | dc->props = kvmclock_properties; |
999e12bb AL |
283 | } |
284 | ||
8c43a6f0 | 285 | static const TypeInfo kvmclock_info = { |
98bdc0d7 | 286 | .name = TYPE_KVM_CLOCK, |
39bffca2 AL |
287 | .parent = TYPE_SYS_BUS_DEVICE, |
288 | .instance_size = sizeof(KVMClockState), | |
289 | .class_init = kvmclock_class_init, | |
0ec329da JK |
290 | }; |
291 | ||
292 | /* Note: Must be called after VCPU initialization. */ | |
293 | void kvmclock_create(void) | |
294 | { | |
182735ef AF |
295 | X86CPU *cpu = X86_CPU(first_cpu); |
296 | ||
0ec329da | 297 | if (kvm_enabled() && |
182735ef AF |
298 | cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) | |
299 | (1ULL << KVM_FEATURE_CLOCKSOURCE2))) { | |
98bdc0d7 | 300 | sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL); |
0ec329da JK |
301 | } |
302 | } | |
303 | ||
83f7d43a | 304 | static void kvmclock_register_types(void) |
0ec329da | 305 | { |
39bffca2 | 306 | type_register_static(&kvmclock_info); |
0ec329da JK |
307 | } |
308 | ||
83f7d43a | 309 | type_init(kvmclock_register_types) |