hw/i386/kvm/clock.c

   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-common.h"
  18 #include "cpu.h"
  19 #include "qemu/host-utils.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 <linux/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 kvmclock_vm_state_change(void *opaque, int running,
 151                                      RunState state)
 152 {
 153     KVMClockState *s = opaque;
 154     CPUState *cpu;
 155     int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
 156     int ret;
 157
 158     if (running) {
 159         struct kvm_clock_data data = {};
 160
 161         /*
 162          * If the host where s->clock was read did not support reliable
 163          * KVM_GET_CLOCK, read kvmclock value from memory.
 164          */
 165         if (!s->clock_is_reliable) {
 166             uint64_t pvclock_via_mem = kvmclock_current_nsec(s);
 167             /* We can't rely on the saved clock value, just discard it */
 168             if (pvclock_via_mem) {
 169                 s->clock = pvclock_via_mem;
 170             }
 171         }
 172
 173         s->clock_valid = false;
 174
 175         data.clock = s->clock;
 176         ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
 177         if (ret < 0) {
 178             fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret));
 179             abort();
 180         }
 181
 182         if (!cap_clock_ctrl) {
 183             return;
 184         }
 185         CPU_FOREACH(cpu) {
 186             ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0);
 187             if (ret) {
 188                 if (ret != -EINVAL) {
 189                     fprintf(stderr, "%s: %s\n", __func__, strerror(-ret));
 190                 }
 191                 return;
 192             }
 193         }
 194     } else {
 195
 196         if (s->clock_valid) {
 197             return;
 198         }
 199
 200         kvm_synchronize_all_tsc();
 201
 202         kvm_update_clock(s);
 203         /*
 204          * If the VM is stopped, declare the clock state valid to
 205          * avoid re-reading it on next vmsave (which would return
 206          * a different value). Will be reset when the VM is continued.
 207          */
 208         s->clock_valid = true;
 209     }
 210 }
 211
 212 static void kvmclock_realize(DeviceState *dev, Error **errp)
 213 {
 214     KVMClockState *s = KVM_CLOCK(dev);
 215
 216     if (!kvm_enabled()) {
 217         error_setg(errp, "kvmclock device requires KVM");
 218         return;
 219     }
 220
 221     kvm_update_clock(s);
 222
 223     qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
 224 }
 225
 226 static bool kvmclock_clock_is_reliable_needed(void *opaque)
 227 {
 228     KVMClockState *s = opaque;
 229
 230     return s->mach_use_reliable_get_clock;
 231 }
 232
 233 static const VMStateDescription kvmclock_reliable_get_clock = {
 234     .name = "kvmclock/clock_is_reliable",
 235     .version_id = 1,
 236     .minimum_version_id = 1,
 237     .needed = kvmclock_clock_is_reliable_needed,
 238     .fields = (VMStateField[]) {
 239         VMSTATE_BOOL(clock_is_reliable, KVMClockState),
 240         VMSTATE_END_OF_LIST()
 241     }
 242 };
 243
 244 /*
 245  * When migrating, assume the source has an unreliable
 246  * KVM_GET_CLOCK unless told otherwise.
 247  */
 248 static int kvmclock_pre_load(void *opaque)
 249 {
 250     KVMClockState *s = opaque;
 251
 252     s->clock_is_reliable = false;
 253
 254     return 0;
 255 }
 256
 257 /*
 258  * When migrating, read the clock just before migration,
 259  * so that the guest clock counts during the events
 260  * between:
 261  *
 262  *  * vm_stop()
 263  *  *
 264  *  * pre_save()
 265  *
 266  *  This reduces kvmclock difference on migration from 5s
 267  *  to 0.1s (when max_downtime == 5s), because sending the
 268  *  final pages of memory (which happens between vm_stop()
 269  *  and pre_save()) takes max_downtime.
 270  */
 271 static int kvmclock_pre_save(void *opaque)
 272 {
 273     KVMClockState *s = opaque;
 274
 275     kvm_update_clock(s);
 276
 277     return 0;
 278 }
 279
 280 static const VMStateDescription kvmclock_vmsd = {
 281     .name = "kvmclock",
 282     .version_id = 1,
 283     .minimum_version_id = 1,
 284     .pre_load = kvmclock_pre_load,
 285     .pre_save = kvmclock_pre_save,
 286     .fields = (VMStateField[]) {
 287         VMSTATE_UINT64(clock, KVMClockState),
 288         VMSTATE_END_OF_LIST()
 289     },
 290     .subsections = (const VMStateDescription * []) {
 291         &kvmclock_reliable_get_clock,
 292         NULL
 293     }
 294 };
 295
 296 static Property kvmclock_properties[] = {
 297     DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState,
 298                       mach_use_reliable_get_clock, true),
 299     DEFINE_PROP_END_OF_LIST(),
 300 };
 301
 302 static void kvmclock_class_init(ObjectClass *klass, void *data)
 303 {
 304     DeviceClass *dc = DEVICE_CLASS(klass);
 305
 306     dc->realize = kvmclock_realize;
 307     dc->vmsd = &kvmclock_vmsd;
 308     dc->props = kvmclock_properties;
 309 }
 310
 311 static const TypeInfo kvmclock_info = {
 312     .name          = TYPE_KVM_CLOCK,
 313     .parent        = TYPE_SYS_BUS_DEVICE,
 314     .instance_size = sizeof(KVMClockState),
 315     .class_init    = kvmclock_class_init,
 316 };
 317
 318 /* Note: Must be called after VCPU initialization. */
 319 void kvmclock_create(void)
 320 {
 321     X86CPU *cpu = X86_CPU(first_cpu);
 322
 323     if (kvm_enabled() &&
 324         cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) |
 325                                        (1ULL << KVM_FEATURE_CLOCKSOURCE2))) {
 326         sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL);
 327     }
 328 }
 329
 330 static void kvmclock_register_types(void)
 331 {
 332     type_register_static(&kvmclock_info);
 333 }
 334
 335 type_init(kvmclock_register_types)