xref: /openbmc/qemu/hw/i386/kvm/clock.c (revision 64552b6b)
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 "cpu.h"
18 #include "qemu/host-utils.h"
19 #include "qemu/module.h"
20 #include "sysemu/sysemu.h"
21 #include "sysemu/kvm.h"
22 #include "sysemu/hw_accel.h"
23 #include "kvm_i386.h"
24 #include "hw/sysbus.h"
25 #include "hw/kvm/clock.h"
26 #include "qapi/error.h"
27 
28 #include <linux/kvm.h>
29 #include "standard-headers/asm-x86/kvm_para.h"
30 
31 #define TYPE_KVM_CLOCK "kvmclock"
32 #define KVM_CLOCK(obj) OBJECT_CHECK(KVMClockState, (obj), TYPE_KVM_CLOCK)
33 
34 typedef struct KVMClockState {
35     /*< private >*/
36     SysBusDevice busdev;
37     /*< public >*/
38 
39     uint64_t clock;
40     bool clock_valid;
41 
42     /* whether machine type supports reliable KVM_GET_CLOCK */
43     bool mach_use_reliable_get_clock;
44 
45     /* whether the 'clock' value was obtained in a host with
46      * reliable KVM_GET_CLOCK */
47     bool clock_is_reliable;
48 } KVMClockState;
49 
50 struct pvclock_vcpu_time_info {
51     uint32_t   version;
52     uint32_t   pad0;
53     uint64_t   tsc_timestamp;
54     uint64_t   system_time;
55     uint32_t   tsc_to_system_mul;
56     int8_t     tsc_shift;
57     uint8_t    flags;
58     uint8_t    pad[2];
59 } __attribute__((__packed__)); /* 32 bytes */
60 
61 static uint64_t kvmclock_current_nsec(KVMClockState *s)
62 {
63     CPUState *cpu = first_cpu;
64     CPUX86State *env = cpu->env_ptr;
65     hwaddr kvmclock_struct_pa;
66     uint64_t migration_tsc = env->tsc;
67     struct pvclock_vcpu_time_info time;
68     uint64_t delta;
69     uint64_t nsec_lo;
70     uint64_t nsec_hi;
71     uint64_t nsec;
72 
73     cpu_synchronize_state(cpu);
74 
75     if (!(env->system_time_msr & 1ULL)) {
76         /* KVM clock not active */
77         return 0;
78     }
79 
80     kvmclock_struct_pa = env->system_time_msr & ~1ULL;
81     cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time));
82 
83     assert(time.tsc_timestamp <= migration_tsc);
84     delta = migration_tsc - time.tsc_timestamp;
85     if (time.tsc_shift < 0) {
86         delta >>= -time.tsc_shift;
87     } else {
88         delta <<= time.tsc_shift;
89     }
90 
91     mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul);
92     nsec = (nsec_lo >> 32) | (nsec_hi << 32);
93     return nsec + time.system_time;
94 }
95 
96 static void kvm_update_clock(KVMClockState *s)
97 {
98     struct kvm_clock_data data;
99     int ret;
100 
101     ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data);
102     if (ret < 0) {
103         fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret));
104                 abort();
105     }
106     s->clock = data.clock;
107 
108     /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns
109      * essentially CLOCK_MONOTONIC plus a guest-specific adjustment.  This
110      * can drift from the TSC-based value that is computed by the guest,
111      * so we need to go through kvmclock_current_nsec().  If
112      * kvm_has_adjust_clock_stable() is true, and the flags contain
113      * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value
114      * and kvmclock_current_nsec() is not necessary.
115      *
116      * Here, however, we need not check KVM_CLOCK_TSC_STABLE.  This is because:
117      *
118      * - if the host has disabled the kvmclock master clock, the guest already
119      *   has protection against time going backwards.  This "safety net" is only
120      *   absent when kvmclock is stable;
121      *
122      * - therefore, we can replace a check like
123      *
124      *       if last KVM_GET_CLOCK was not reliable then
125      *               read from memory
126      *
127      *   with
128      *
129      *       if last KVM_GET_CLOCK was not reliable && masterclock is enabled
130      *               read from memory
131      *
132      * However:
133      *
134      * - if kvm_has_adjust_clock_stable() returns false, the left side is
135      *   always true (KVM_GET_CLOCK is never reliable), and the right side is
136      *   unknown (because we don't have data.flags).  We must assume it's true
137      *   and read from memory.
138      *
139      * - if kvm_has_adjust_clock_stable() returns true, the result of the &&
140      *   is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable)
141      *
142      * So we can just use this instead:
143      *
144      *       if !kvm_has_adjust_clock_stable() then
145      *               read from memory
146      */
147     s->clock_is_reliable = kvm_has_adjust_clock_stable();
148 }
149 
150 static void do_kvmclock_ctrl(CPUState *cpu, run_on_cpu_data data)
151 {
152     int ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0);
153 
154     if (ret && ret != -EINVAL) {
155         fprintf(stderr, "%s: %s\n", __func__, strerror(-ret));
156     }
157 }
158 
159 static void kvmclock_vm_state_change(void *opaque, int running,
160                                      RunState state)
161 {
162     KVMClockState *s = opaque;
163     CPUState *cpu;
164     int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
165     int ret;
166 
167     if (running) {
168         struct kvm_clock_data data = {};
169 
170         /*
171          * If the host where s->clock was read did not support reliable
172          * KVM_GET_CLOCK, read kvmclock value from memory.
173          */
174         if (!s->clock_is_reliable) {
175             uint64_t pvclock_via_mem = kvmclock_current_nsec(s);
176             /* We can't rely on the saved clock value, just discard it */
177             if (pvclock_via_mem) {
178                 s->clock = pvclock_via_mem;
179             }
180         }
181 
182         s->clock_valid = false;
183 
184         data.clock = s->clock;
185         ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
186         if (ret < 0) {
187             fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret));
188             abort();
189         }
190 
191         if (!cap_clock_ctrl) {
192             return;
193         }
194         CPU_FOREACH(cpu) {
195             run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL);
196         }
197     } else {
198 
199         if (s->clock_valid) {
200             return;
201         }
202 
203         kvm_synchronize_all_tsc();
204 
205         kvm_update_clock(s);
206         /*
207          * If the VM is stopped, declare the clock state valid to
208          * avoid re-reading it on next vmsave (which would return
209          * a different value). Will be reset when the VM is continued.
210          */
211         s->clock_valid = true;
212     }
213 }
214 
215 static void kvmclock_realize(DeviceState *dev, Error **errp)
216 {
217     KVMClockState *s = KVM_CLOCK(dev);
218 
219     if (!kvm_enabled()) {
220         error_setg(errp, "kvmclock device requires KVM");
221         return;
222     }
223 
224     kvm_update_clock(s);
225 
226     qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
227 }
228 
229 static bool kvmclock_clock_is_reliable_needed(void *opaque)
230 {
231     KVMClockState *s = opaque;
232 
233     return s->mach_use_reliable_get_clock;
234 }
235 
236 static const VMStateDescription kvmclock_reliable_get_clock = {
237     .name = "kvmclock/clock_is_reliable",
238     .version_id = 1,
239     .minimum_version_id = 1,
240     .needed = kvmclock_clock_is_reliable_needed,
241     .fields = (VMStateField[]) {
242         VMSTATE_BOOL(clock_is_reliable, KVMClockState),
243         VMSTATE_END_OF_LIST()
244     }
245 };
246 
247 /*
248  * When migrating, assume the source has an unreliable
249  * KVM_GET_CLOCK unless told otherwise.
250  */
251 static int kvmclock_pre_load(void *opaque)
252 {
253     KVMClockState *s = opaque;
254 
255     s->clock_is_reliable = false;
256 
257     return 0;
258 }
259 
260 /*
261  * When migrating, read the clock just before migration,
262  * so that the guest clock counts during the events
263  * between:
264  *
265  *  * vm_stop()
266  *  *
267  *  * pre_save()
268  *
269  *  This reduces kvmclock difference on migration from 5s
270  *  to 0.1s (when max_downtime == 5s), because sending the
271  *  final pages of memory (which happens between vm_stop()
272  *  and pre_save()) takes max_downtime.
273  */
274 static int kvmclock_pre_save(void *opaque)
275 {
276     KVMClockState *s = opaque;
277 
278     kvm_update_clock(s);
279 
280     return 0;
281 }
282 
283 static const VMStateDescription kvmclock_vmsd = {
284     .name = "kvmclock",
285     .version_id = 1,
286     .minimum_version_id = 1,
287     .pre_load = kvmclock_pre_load,
288     .pre_save = kvmclock_pre_save,
289     .fields = (VMStateField[]) {
290         VMSTATE_UINT64(clock, KVMClockState),
291         VMSTATE_END_OF_LIST()
292     },
293     .subsections = (const VMStateDescription * []) {
294         &kvmclock_reliable_get_clock,
295         NULL
296     }
297 };
298 
299 static Property kvmclock_properties[] = {
300     DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState,
301                       mach_use_reliable_get_clock, true),
302     DEFINE_PROP_END_OF_LIST(),
303 };
304 
305 static void kvmclock_class_init(ObjectClass *klass, void *data)
306 {
307     DeviceClass *dc = DEVICE_CLASS(klass);
308 
309     dc->realize = kvmclock_realize;
310     dc->vmsd = &kvmclock_vmsd;
311     dc->props = kvmclock_properties;
312 }
313 
314 static const TypeInfo kvmclock_info = {
315     .name          = TYPE_KVM_CLOCK,
316     .parent        = TYPE_SYS_BUS_DEVICE,
317     .instance_size = sizeof(KVMClockState),
318     .class_init    = kvmclock_class_init,
319 };
320 
321 /* Note: Must be called after VCPU initialization. */
322 void kvmclock_create(void)
323 {
324     X86CPU *cpu = X86_CPU(first_cpu);
325 
326     if (kvm_enabled() &&
327         cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) |
328                                        (1ULL << KVM_FEATURE_CLOCKSOURCE2))) {
329         sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL);
330     }
331 }
332 
333 static void kvmclock_register_types(void)
334 {
335     type_register_static(&kvmclock_info);
336 }
337 
338 type_init(kvmclock_register_types)
339