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