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 "cpu.h"
  18 #include "qemu/host-utils.h"
  19 #include "qemu/module.h"
  20 #include "sysemu/kvm.h"
  21 #include "sysemu/runstate.h"
  22 #include "sysemu/hw_accel.h"
  23 #include "kvm_i386.h"
  24 #include "migration/vmstate.h"
  25 #include "hw/sysbus.h"
  26 #include "hw/kvm/clock.h"
  27 #include "hw/qdev-properties.h"
  28 #include "qapi/error.h"
  29
  30 #include <linux/kvm.h>
  31 #include "standard-headers/asm-x86/kvm_para.h"
  32 #include "qom/object.h"
  33
  34 #define TYPE_KVM_CLOCK "kvmclock"
  35 typedef struct KVMClockState KVMClockState;
  36 DECLARE_INSTANCE_CHECKER(KVMClockState, KVM_CLOCK,
  37                          TYPE_KVM_CLOCK)
  38
  39 struct KVMClockState {
  40     /*< private >*/
  41     SysBusDevice busdev;
  42     /*< public >*/
  43
  44     uint64_t clock;
  45     bool clock_valid;
  46
  47     /* whether the 'clock' value was obtained in the 'paused' state */
  48     bool runstate_paused;
  49
  50     /* whether machine type supports reliable KVM_GET_CLOCK */
  51     bool mach_use_reliable_get_clock;
  52
  53     /* whether the 'clock' value was obtained in a host with
  54      * reliable KVM_GET_CLOCK */
  55     bool clock_is_reliable;
  56 };
  57
  58 struct pvclock_vcpu_time_info {
  59     uint32_t   version;
  60     uint32_t   pad0;
  61     uint64_t   tsc_timestamp;
  62     uint64_t   system_time;
  63     uint32_t   tsc_to_system_mul;
  64     int8_t     tsc_shift;
  65     uint8_t    flags;
  66     uint8_t    pad[2];
  67 } __attribute__((__packed__)); /* 32 bytes */
  68
  69 static uint64_t kvmclock_current_nsec(KVMClockState *s)
  70 {
  71     CPUState *cpu = first_cpu;
  72     CPUX86State *env = cpu->env_ptr;
  73     hwaddr kvmclock_struct_pa;
  74     uint64_t migration_tsc = env->tsc;
  75     struct pvclock_vcpu_time_info time;
  76     uint64_t delta;
  77     uint64_t nsec_lo;
  78     uint64_t nsec_hi;
  79     uint64_t nsec;
  80
  81     cpu_synchronize_state(cpu);
  82
  83     if (!(env->system_time_msr & 1ULL)) {
  84         /* KVM clock not active */
  85         return 0;
  86     }
  87
  88     kvmclock_struct_pa = env->system_time_msr & ~1ULL;
  89     cpu_physical_memory_read(kvmclock_struct_pa, &time, sizeof(time));
  90
  91     assert(time.tsc_timestamp <= migration_tsc);
  92     delta = migration_tsc - time.tsc_timestamp;
  93     if (time.tsc_shift < 0) {
  94         delta >>= -time.tsc_shift;
  95     } else {
  96         delta <<= time.tsc_shift;
  97     }
  98
  99     mulu64(&nsec_lo, &nsec_hi, delta, time.tsc_to_system_mul);
 100     nsec = (nsec_lo >> 32) | (nsec_hi << 32);
 101     return nsec + time.system_time;
 102 }
 103
 104 static void kvm_update_clock(KVMClockState *s)
 105 {
 106     struct kvm_clock_data data;
 107     int ret;
 108
 109     ret = kvm_vm_ioctl(kvm_state, KVM_GET_CLOCK, &data);
 110     if (ret < 0) {
 111         fprintf(stderr, "KVM_GET_CLOCK failed: %s\n", strerror(ret));
 112                 abort();
 113     }
 114     s->clock = data.clock;
 115
 116     /* If kvm_has_adjust_clock_stable() is false, KVM_GET_CLOCK returns
 117      * essentially CLOCK_MONOTONIC plus a guest-specific adjustment.  This
 118      * can drift from the TSC-based value that is computed by the guest,
 119      * so we need to go through kvmclock_current_nsec().  If
 120      * kvm_has_adjust_clock_stable() is true, and the flags contain
 121      * KVM_CLOCK_TSC_STABLE, then KVM_GET_CLOCK returns a TSC-based value
 122      * and kvmclock_current_nsec() is not necessary.
 123      *
 124      * Here, however, we need not check KVM_CLOCK_TSC_STABLE.  This is because:
 125      *
 126      * - if the host has disabled the kvmclock master clock, the guest already
 127      *   has protection against time going backwards.  This "safety net" is only
 128      *   absent when kvmclock is stable;
 129      *
 130      * - therefore, we can replace a check like
 131      *
 132      *       if last KVM_GET_CLOCK was not reliable then
 133      *               read from memory
 134      *
 135      *   with
 136      *
 137      *       if last KVM_GET_CLOCK was not reliable && masterclock is enabled
 138      *               read from memory
 139      *
 140      * However:
 141      *
 142      * - if kvm_has_adjust_clock_stable() returns false, the left side is
 143      *   always true (KVM_GET_CLOCK is never reliable), and the right side is
 144      *   unknown (because we don't have data.flags).  We must assume it's true
 145      *   and read from memory.
 146      *
 147      * - if kvm_has_adjust_clock_stable() returns true, the result of the &&
 148      *   is always false (masterclock is enabled iff KVM_GET_CLOCK is reliable)
 149      *
 150      * So we can just use this instead:
 151      *
 152      *       if !kvm_has_adjust_clock_stable() then
 153      *               read from memory
 154      */
 155     s->clock_is_reliable = kvm_has_adjust_clock_stable();
 156 }
 157
 158 static void do_kvmclock_ctrl(CPUState *cpu, run_on_cpu_data data)
 159 {
 160     int ret = kvm_vcpu_ioctl(cpu, KVM_KVMCLOCK_CTRL, 0);
 161
 162     if (ret && ret != -EINVAL) {
 163         fprintf(stderr, "%s: %s\n", __func__, strerror(-ret));
 164     }
 165 }
 166
 167 static void kvmclock_vm_state_change(void *opaque, int running,
 168                                      RunState state)
 169 {
 170     KVMClockState *s = opaque;
 171     CPUState *cpu;
 172     int cap_clock_ctrl = kvm_check_extension(kvm_state, KVM_CAP_KVMCLOCK_CTRL);
 173     int ret;
 174
 175     if (running) {
 176         struct kvm_clock_data data = {};
 177
 178         /*
 179          * If the host where s->clock was read did not support reliable
 180          * KVM_GET_CLOCK, read kvmclock value from memory.
 181          */
 182         if (!s->clock_is_reliable) {
 183             uint64_t pvclock_via_mem = kvmclock_current_nsec(s);
 184             /* We can't rely on the saved clock value, just discard it */
 185             if (pvclock_via_mem) {
 186                 s->clock = pvclock_via_mem;
 187             }
 188         }
 189
 190         s->clock_valid = false;
 191
 192         data.clock = s->clock;
 193         ret = kvm_vm_ioctl(kvm_state, KVM_SET_CLOCK, &data);
 194         if (ret < 0) {
 195             fprintf(stderr, "KVM_SET_CLOCK failed: %s\n", strerror(ret));
 196             abort();
 197         }
 198
 199         if (!cap_clock_ctrl) {
 200             return;
 201         }
 202         CPU_FOREACH(cpu) {
 203             run_on_cpu(cpu, do_kvmclock_ctrl, RUN_ON_CPU_NULL);
 204         }
 205     } else {
 206
 207         if (s->clock_valid) {
 208             return;
 209         }
 210
 211         s->runstate_paused = runstate_check(RUN_STATE_PAUSED);
 212
 213         kvm_synchronize_all_tsc();
 214
 215         kvm_update_clock(s);
 216         /*
 217          * If the VM is stopped, declare the clock state valid to
 218          * avoid re-reading it on next vmsave (which would return
 219          * a different value). Will be reset when the VM is continued.
 220          */
 221         s->clock_valid = true;
 222     }
 223 }
 224
 225 static void kvmclock_realize(DeviceState *dev, Error **errp)
 226 {
 227     KVMClockState *s = KVM_CLOCK(dev);
 228
 229     if (!kvm_enabled()) {
 230         error_setg(errp, "kvmclock device requires KVM");
 231         return;
 232     }
 233
 234     kvm_update_clock(s);
 235
 236     qemu_add_vm_change_state_handler(kvmclock_vm_state_change, s);
 237 }
 238
 239 static bool kvmclock_clock_is_reliable_needed(void *opaque)
 240 {
 241     KVMClockState *s = opaque;
 242
 243     return s->mach_use_reliable_get_clock;
 244 }
 245
 246 static const VMStateDescription kvmclock_reliable_get_clock = {
 247     .name = "kvmclock/clock_is_reliable",
 248     .version_id = 1,
 249     .minimum_version_id = 1,
 250     .needed = kvmclock_clock_is_reliable_needed,
 251     .fields = (VMStateField[]) {
 252         VMSTATE_BOOL(clock_is_reliable, KVMClockState),
 253         VMSTATE_END_OF_LIST()
 254     }
 255 };
 256
 257 /*
 258  * When migrating, assume the source has an unreliable
 259  * KVM_GET_CLOCK unless told otherwise.
 260  */
 261 static int kvmclock_pre_load(void *opaque)
 262 {
 263     KVMClockState *s = opaque;
 264
 265     s->clock_is_reliable = false;
 266
 267     return 0;
 268 }
 269
 270 /*
 271  * When migrating a running guest, read the clock just
 272  * before migration, so that the guest clock counts
 273  * during the events between:
 274  *
 275  *  * vm_stop()
 276  *  *
 277  *  * pre_save()
 278  *
 279  *  This reduces kvmclock difference on migration from 5s
 280  *  to 0.1s (when max_downtime == 5s), because sending the
 281  *  final pages of memory (which happens between vm_stop()
 282  *  and pre_save()) takes max_downtime.
 283  */
 284 static int kvmclock_pre_save(void *opaque)
 285 {
 286     KVMClockState *s = opaque;
 287
 288     if (!s->runstate_paused) {
 289         kvm_update_clock(s);
 290     }
 291
 292     return 0;
 293 }
 294
 295 static const VMStateDescription kvmclock_vmsd = {
 296     .name = "kvmclock",
 297     .version_id = 1,
 298     .minimum_version_id = 1,
 299     .pre_load = kvmclock_pre_load,
 300     .pre_save = kvmclock_pre_save,
 301     .fields = (VMStateField[]) {
 302         VMSTATE_UINT64(clock, KVMClockState),
 303         VMSTATE_END_OF_LIST()
 304     },
 305     .subsections = (const VMStateDescription * []) {
 306         &kvmclock_reliable_get_clock,
 307         NULL
 308     }
 309 };
 310
 311 static Property kvmclock_properties[] = {
 312     DEFINE_PROP_BOOL("x-mach-use-reliable-get-clock", KVMClockState,
 313                       mach_use_reliable_get_clock, true),
 314     DEFINE_PROP_END_OF_LIST(),
 315 };
 316
 317 static void kvmclock_class_init(ObjectClass *klass, void *data)
 318 {
 319     DeviceClass *dc = DEVICE_CLASS(klass);
 320
 321     dc->realize = kvmclock_realize;
 322     dc->vmsd = &kvmclock_vmsd;
 323     device_class_set_props(dc, kvmclock_properties);
 324 }
 325
 326 static const TypeInfo kvmclock_info = {
 327     .name          = TYPE_KVM_CLOCK,
 328     .parent        = TYPE_SYS_BUS_DEVICE,
 329     .instance_size = sizeof(KVMClockState),
 330     .class_init    = kvmclock_class_init,
 331 };
 332
 333 /* Note: Must be called after VCPU initialization. */
 334 void kvmclock_create(void)
 335 {
 336     X86CPU *cpu = X86_CPU(first_cpu);
 337
 338     if (kvm_enabled() &&
 339         cpu->env.features[FEAT_KVM] & ((1ULL << KVM_FEATURE_CLOCKSOURCE) |
 340                                        (1ULL << KVM_FEATURE_CLOCKSOURCE2))) {
 341         sysbus_create_simple(TYPE_KVM_CLOCK, -1, NULL);
 342     }
 343 }
 344
 345 static void kvmclock_register_types(void)
 346 {
 347     type_register_static(&kvmclock_info);
 348 }
 349
 350 type_init(kvmclock_register_types)