sys/dev/kvm_clock/kvm_clock.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   3  *
   4  * Copyright (c) 2014 Bryan Venteicher <bryanv@FreeBSD.org>
   5  * Copyright (c) 2021 Mathieu Chouquet-Stringer
   6  * Copyright (c) 2021 Juniper Networks, Inc.
   7  * Copyright (c) 2021 Klara, Inc.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  */
  30
  31 /*
  32  * Linux KVM paravirtual clock support
  33  *
  34  * References:
  35  *     - [1] https://www.kernel.org/doc/html/latest/virt/kvm/cpuid.html
  36  *     - [2] https://www.kernel.org/doc/html/latest/virt/kvm/msr.html
  37  */
  38
  39 #include <sys/param.h>
  40 #include <sys/bus.h>
  41 #include <sys/domainset.h>
  42 #include <sys/kernel.h>
  43 #include <sys/malloc.h>
  44 #include <sys/module.h>
  45 #include <sys/smp.h>
  46 #include <sys/sysctl.h>
  47
  48 #include <vm/vm.h>
  49 #include <vm/pmap.h>
  50 #include <vm/vm_extern.h>
  51
  52 #include <machine/pvclock.h>
  53 #include <x86/kvm.h>
  54
  55 #include "clock_if.h"
  56
  57 #define KVM_CLOCK_DEVNAME               "kvmclock"
  58 /*
  59  * Note: Chosen to be (1) above HPET's value (always 950), (2) above the TSC's
  60  * default value of 800, and (3) below the TSC's value when it supports the
  61  * "Invariant TSC" feature and is believed to be synchronized across all CPUs.
  62  */
  63 #define KVM_CLOCK_TC_QUALITY            975
  64
  65 struct kvm_clock_softc {
  66         struct pvclock                   pvc;
  67         struct pvclock_wall_clock        wc;
  68         struct pvclock_vcpu_time_info   *timeinfos;
  69         u_int                            msr_tc;
  70         u_int                            msr_wc;
  71 #ifndef EARLY_AP_STARTUP
  72         int                              firstcpu;
  73 #endif
  74 };
  75
  76 static struct pvclock_wall_clock *kvm_clock_get_wallclock(void *arg);
  77 static void     kvm_clock_system_time_enable(struct kvm_clock_softc *sc,
  78                     const cpuset_t *cpus);
  79 static void     kvm_clock_system_time_enable_pcpu(void *arg);
  80 static void     kvm_clock_setup_sysctl(device_t);
  81
  82 static struct pvclock_wall_clock *
  83 kvm_clock_get_wallclock(void *arg)
  84 {
  85         struct kvm_clock_softc *sc = arg;
  86
  87         wrmsr(sc->msr_wc, vtophys(&sc->wc));
  88         return (&sc->wc);
  89 }
  90
  91 static void
  92 kvm_clock_system_time_enable(struct kvm_clock_softc *sc, const cpuset_t *cpus)
  93 {
  94         smp_rendezvous_cpus(*cpus, NULL, kvm_clock_system_time_enable_pcpu,
  95             NULL, sc);
  96 }
  97
  98 static void
  99 kvm_clock_system_time_enable_pcpu(void *arg)
 100 {
 101         struct kvm_clock_softc *sc = arg;
 102
 103         /*
 104          * See [2]; the lsb of this MSR is the system time enable bit.
 105          */
 106         wrmsr(sc->msr_tc, vtophys(&(sc->timeinfos)[curcpu]) | 1);
 107 }
 108
 109 #ifndef EARLY_AP_STARTUP
 110 static void
 111 kvm_clock_init_smp(void *arg __unused)
 112 {
 113         devclass_t kvm_clock_devclass;
 114         cpuset_t cpus;
 115         struct kvm_clock_softc *sc;
 116
 117         kvm_clock_devclass = devclass_find(KVM_CLOCK_DEVNAME);
 118         sc = devclass_get_softc(kvm_clock_devclass, 0);
 119         if (sc == NULL || mp_ncpus == 1)
 120                 return;
 121
 122         /*
 123          * Register with the hypervisor on all CPUs except the one that
 124          * registered in kvm_clock_attach().
 125          */
 126         cpus = all_cpus;
 127         KASSERT(CPU_ISSET(sc->firstcpu, &cpus),
 128             ("%s: invalid first CPU %d", __func__, sc->firstcpu));
 129         CPU_CLR(sc->firstcpu, &cpus);
 130         kvm_clock_system_time_enable(sc, &cpus);
 131 }
 132 SYSINIT(kvm_clock, SI_SUB_SMP, SI_ORDER_ANY, kvm_clock_init_smp, NULL);
 133 #endif
 134
 135 static void
 136 kvm_clock_identify(driver_t *driver, device_t parent)
 137 {
 138         u_int regs[4];
 139
 140         kvm_cpuid_get_features(regs);
 141         if ((regs[0] &
 142             (KVM_FEATURE_CLOCKSOURCE2 | KVM_FEATURE_CLOCKSOURCE)) == 0)
 143                 return;
 144         if (device_find_child(parent, KVM_CLOCK_DEVNAME, -1))
 145                 return;
 146         BUS_ADD_CHILD(parent, 0, KVM_CLOCK_DEVNAME, 0);
 147 }
 148
 149 static int
 150 kvm_clock_probe(device_t dev)
 151 {
 152         device_set_desc(dev, "KVM paravirtual clock");
 153         return (BUS_PROBE_DEFAULT);
 154 }
 155
 156 static int
 157 kvm_clock_attach(device_t dev)
 158 {
 159         u_int regs[4];
 160         struct kvm_clock_softc *sc = device_get_softc(dev);
 161         bool stable_flag_supported;
 162
 163         /* Process KVM "features" CPUID leaf content: */
 164         kvm_cpuid_get_features(regs);
 165         if ((regs[0] & KVM_FEATURE_CLOCKSOURCE2) != 0) {
 166                 sc->msr_tc = KVM_MSR_SYSTEM_TIME_NEW;
 167                 sc->msr_wc = KVM_MSR_WALL_CLOCK_NEW;
 168         } else {
 169                 KASSERT((regs[0] & KVM_FEATURE_CLOCKSOURCE) != 0,
 170                     ("Clocksource feature flags disappeared since "
 171                     "kvm_clock_identify: regs[0] %#0x.", regs[0]));
 172                 sc->msr_tc = KVM_MSR_SYSTEM_TIME;
 173                 sc->msr_wc = KVM_MSR_WALL_CLOCK;
 174         }
 175         stable_flag_supported =
 176             (regs[0] & KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) != 0;
 177
 178         /* Set up 'struct pvclock_vcpu_time_info' page(s): */
 179         sc->timeinfos = kmem_malloc(mp_ncpus *
 180             sizeof(struct pvclock_vcpu_time_info), M_WAITOK | M_ZERO);
 181 #ifdef EARLY_AP_STARTUP
 182         kvm_clock_system_time_enable(sc, &all_cpus);
 183 #else
 184         sc->firstcpu = curcpu;
 185         kvm_clock_system_time_enable_pcpu(sc);
 186 #endif
 187
 188         /*
 189          * Init pvclock; register KVM clock wall clock, register KVM clock
 190          * timecounter, and set up the requisite infrastructure for vDSO access
 191          * to this timecounter.
 192          *     Regarding 'tc_flags': Since the KVM MSR documentation does not
 193          *     specifically discuss suspend/resume scenarios, conservatively
 194          *     leave 'TC_FLAGS_SUSPEND_SAFE' cleared and assume that the system
 195          *     time must be re-inited in such cases.
 196          */
 197         sc->pvc.get_wallclock = kvm_clock_get_wallclock;
 198         sc->pvc.get_wallclock_arg = sc;
 199         sc->pvc.timeinfos = sc->timeinfos;
 200         sc->pvc.stable_flag_supported = stable_flag_supported;
 201         pvclock_init(&sc->pvc, dev, KVM_CLOCK_DEVNAME, KVM_CLOCK_TC_QUALITY, 0);
 202         kvm_clock_setup_sysctl(dev);
 203         return (0);
 204 }
 205
 206 static int
 207 kvm_clock_detach(device_t dev)
 208 {
 209         struct kvm_clock_softc *sc = device_get_softc(dev);
 210
 211         return (pvclock_destroy(&sc->pvc));
 212 }
 213
 214 static int
 215 kvm_clock_suspend(device_t dev)
 216 {
 217         return (0);
 218 }
 219
 220 static int
 221 kvm_clock_resume(device_t dev)
 222 {
 223         /*
 224          * See note in 'kvm_clock_attach()' regarding 'TC_FLAGS_SUSPEND_SAFE';
 225          * conservatively assume that the system time must be re-inited in
 226          * suspend/resume scenarios.
 227          */
 228         kvm_clock_system_time_enable(device_get_softc(dev), &all_cpus);
 229         pvclock_resume();
 230         inittodr(time_second);
 231         return (0);
 232 }
 233
 234 static int
 235 kvm_clock_gettime(device_t dev, struct timespec *ts)
 236 {
 237         struct kvm_clock_softc *sc = device_get_softc(dev);
 238
 239         pvclock_gettime(&sc->pvc, ts);
 240         return (0);
 241 }
 242
 243 static int
 244 kvm_clock_settime(device_t dev, struct timespec *ts)
 245 {
 246         /*
 247          * Even though it is not possible to set the KVM clock's wall clock, to
 248          * avoid the possibility of periodic benign error messages from
 249          * 'settime_task_func()', report success rather than, e.g., 'ENODEV'.
 250          */
 251         return (0);
 252 }
 253
 254 static int
 255 kvm_clock_tsc_freq_sysctl(SYSCTL_HANDLER_ARGS)
 256 {
 257         struct kvm_clock_softc *sc = oidp->oid_arg1;
 258         uint64_t freq = pvclock_tsc_freq(sc->timeinfos);
 259
 260         return (sysctl_handle_64(oidp, &freq, 0, req));
 261 }
 262
 263 static void
 264 kvm_clock_setup_sysctl(device_t dev)
 265 {
 266         struct kvm_clock_softc *sc = device_get_softc(dev);
 267         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(dev);
 268         struct sysctl_oid *tree = device_get_sysctl_tree(dev);
 269         struct sysctl_oid_list *child = SYSCTL_CHILDREN(tree);
 270
 271         SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "tsc_freq",
 272             CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
 273             kvm_clock_tsc_freq_sysctl, "QU",
 274             "Time Stamp Counter frequency");
 275 }
 276
 277 static device_method_t kvm_clock_methods[] = {
 278         DEVMETHOD(device_identify,      kvm_clock_identify),
 279         DEVMETHOD(device_probe,         kvm_clock_probe),
 280         DEVMETHOD(device_attach,        kvm_clock_attach),
 281         DEVMETHOD(device_detach,        kvm_clock_detach),
 282         DEVMETHOD(device_suspend,       kvm_clock_suspend),
 283         DEVMETHOD(device_resume,        kvm_clock_resume),
 284         /* clock interface */
 285         DEVMETHOD(clock_gettime,        kvm_clock_gettime),
 286         DEVMETHOD(clock_settime,        kvm_clock_settime),
 287
 288         DEVMETHOD_END
 289 };
 290
 291 static driver_t kvm_clock_driver = {
 292         KVM_CLOCK_DEVNAME,
 293         kvm_clock_methods,
 294         sizeof(struct kvm_clock_softc),
 295 };
 296
 297 DRIVER_MODULE(kvm_clock, nexus, kvm_clock_driver, 0, 0);