sys/dev/kvm_clock/kvm_clock.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   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/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #include <sys/param.h>
  43 #include <sys/bus.h>
  44 #include <sys/domainset.h>
  45 #include <sys/kernel.h>
  46 #include <sys/malloc.h>
  47 #include <sys/module.h>
  48 #include <sys/smp.h>
  49 #include <sys/sysctl.h>
  50
  51 #include <vm/vm.h>
  52 #include <vm/pmap.h>
  53 #include <vm/vm_extern.h>
  54
  55 #include <machine/pvclock.h>
  56 #include <x86/kvm.h>
  57
  58 #include "clock_if.h"
  59
  60 #define KVM_CLOCK_DEVNAME               "kvmclock"
  61 /*
  62  * Note: Chosen to be (1) above HPET's value (always 950), (2) above the TSC's
  63  * default value of 800, and (3) below the TSC's value when it supports the
  64  * "Invariant TSC" feature and is believed to be synchronized across all CPUs.
  65  */
  66 #define KVM_CLOCK_TC_QUALITY            975
  67
  68 struct kvm_clock_softc {
  69         struct pvclock                   pvc;
  70         struct pvclock_wall_clock        wc;
  71         struct pvclock_vcpu_time_info   *timeinfos;
  72         u_int                            msr_tc;
  73         u_int                            msr_wc;
  74 };
  75
  76 static devclass_t       kvm_clock_devclass;
  77
  78 static struct pvclock_wall_clock *kvm_clock_get_wallclock(void *arg);
  79 static void     kvm_clock_system_time_enable(struct kvm_clock_softc *sc);
  80 static void     kvm_clock_system_time_enable_pcpu(void *arg);
  81 static void     kvm_clock_setup_sysctl(device_t);
  82
  83 static struct pvclock_wall_clock *
  84 kvm_clock_get_wallclock(void *arg)
  85 {
  86         struct kvm_clock_softc *sc = arg;
  87
  88         wrmsr(sc->msr_wc, vtophys(&sc->wc));
  89         return (&sc->wc);
  90 }
  91
  92 static void
  93 kvm_clock_system_time_enable(struct kvm_clock_softc *sc)
  94 {
  95         smp_rendezvous(NULL, kvm_clock_system_time_enable_pcpu, 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 static void
 110 kvm_clock_identify(driver_t *driver, device_t parent)
 111 {
 112         u_int regs[4];
 113
 114         kvm_cpuid_get_features(regs);
 115         if ((regs[0] &
 116             (KVM_FEATURE_CLOCKSOURCE2 | KVM_FEATURE_CLOCKSOURCE)) == 0)
 117                 return;
 118         if (device_find_child(parent, KVM_CLOCK_DEVNAME, -1))
 119                 return;
 120         BUS_ADD_CHILD(parent, 0, KVM_CLOCK_DEVNAME, 0);
 121 }
 122
 123 static int
 124 kvm_clock_probe(device_t dev)
 125 {
 126         device_set_desc(dev, "KVM paravirtual clock");
 127         return (BUS_PROBE_DEFAULT);
 128 }
 129
 130 static int
 131 kvm_clock_attach(device_t dev)
 132 {
 133         u_int regs[4];
 134         struct kvm_clock_softc *sc = device_get_softc(dev);
 135         bool stable_flag_supported;
 136
 137         /* Process KVM "features" CPUID leaf content: */
 138         kvm_cpuid_get_features(regs);
 139         if ((regs[0] & KVM_FEATURE_CLOCKSOURCE2) != 0) {
 140                 sc->msr_tc = KVM_MSR_SYSTEM_TIME_NEW;
 141                 sc->msr_wc = KVM_MSR_WALL_CLOCK_NEW;
 142         } else {
 143                 KASSERT((regs[0] & KVM_FEATURE_CLOCKSOURCE) != 0,
 144                     ("Clocksource feature flags disappeared since "
 145                     "kvm_clock_identify: regs[0] %#0x.", regs[0]));
 146                 sc->msr_tc = KVM_MSR_SYSTEM_TIME;
 147                 sc->msr_wc = KVM_MSR_WALL_CLOCK;
 148         }
 149         stable_flag_supported =
 150             (regs[0] & KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) != 0;
 151
 152         /* Set up 'struct pvclock_vcpu_time_info' page(s): */
 153         sc->timeinfos = (struct pvclock_vcpu_time_info *)kmem_malloc(mp_ncpus *
 154             sizeof(struct pvclock_vcpu_time_info), M_WAITOK | M_ZERO);
 155         kvm_clock_system_time_enable(sc);
 156
 157         /*
 158          * Init pvclock; register KVM clock wall clock, register KVM clock
 159          * timecounter, and set up the requisite infrastructure for vDSO access
 160          * to this timecounter.
 161          *     Regarding 'tc_flags': Since the KVM MSR documentation does not
 162          *     specifically discuss suspend/resume scenarios, conservatively
 163          *     leave 'TC_FLAGS_SUSPEND_SAFE' cleared and assume that the system
 164          *     time must be re-inited in such cases.
 165          */
 166         sc->pvc.get_wallclock = kvm_clock_get_wallclock;
 167         sc->pvc.get_wallclock_arg = sc;
 168         sc->pvc.timeinfos = sc->timeinfos;
 169         sc->pvc.stable_flag_supported = stable_flag_supported;
 170         pvclock_init(&sc->pvc, dev, KVM_CLOCK_DEVNAME, KVM_CLOCK_TC_QUALITY, 0);
 171         kvm_clock_setup_sysctl(dev);
 172         return (0);
 173 }
 174
 175 static int
 176 kvm_clock_detach(device_t dev)
 177 {
 178         struct kvm_clock_softc *sc = device_get_softc(dev);
 179
 180         return (pvclock_destroy(&sc->pvc));
 181 }
 182
 183 static int
 184 kvm_clock_suspend(device_t dev)
 185 {
 186         return (0);
 187 }
 188
 189 static int
 190 kvm_clock_resume(device_t dev)
 191 {
 192         /*
 193          * See note in 'kvm_clock_attach()' regarding 'TC_FLAGS_SUSPEND_SAFE';
 194          * conservatively assume that the system time must be re-inited in
 195          * suspend/resume scenarios.
 196          */
 197         kvm_clock_system_time_enable(device_get_softc(dev));
 198         pvclock_resume();
 199         inittodr(time_second);
 200         return (0);
 201 }
 202
 203 static int
 204 kvm_clock_gettime(device_t dev, struct timespec *ts)
 205 {
 206         struct kvm_clock_softc *sc = device_get_softc(dev);
 207
 208         pvclock_gettime(&sc->pvc, ts);
 209         return (0);
 210 }
 211
 212 static int
 213 kvm_clock_settime(device_t dev, struct timespec *ts)
 214 {
 215         /*
 216          * Even though it is not possible to set the KVM clock's wall clock, to
 217          * avoid the possibility of periodic benign error messages from
 218          * 'settime_task_func()', report success rather than, e.g., 'ENODEV'.
 219          */
 220         return (0);
 221 }
 222
 223 static int
 224 kvm_clock_tsc_freq_sysctl(SYSCTL_HANDLER_ARGS)
 225 {
 226         struct kvm_clock_softc *sc = oidp->oid_arg1;
 227         uint64_t freq = pvclock_tsc_freq(sc->timeinfos);
 228
 229         return (sysctl_handle_64(oidp, &freq, 0, req));
 230 }
 231
 232 static void
 233 kvm_clock_setup_sysctl(device_t dev)
 234 {
 235         struct kvm_clock_softc *sc = device_get_softc(dev);
 236         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(dev);
 237         struct sysctl_oid *tree = device_get_sysctl_tree(dev);
 238         struct sysctl_oid_list *child = SYSCTL_CHILDREN(tree);
 239
 240         SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "tsc_freq",
 241             CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
 242             kvm_clock_tsc_freq_sysctl, "QU",
 243             "Time Stamp Counter frequency");
 244 }
 245
 246 static device_method_t kvm_clock_methods[] = {
 247         DEVMETHOD(device_identify,      kvm_clock_identify),
 248         DEVMETHOD(device_probe,         kvm_clock_probe),
 249         DEVMETHOD(device_attach,        kvm_clock_attach),
 250         DEVMETHOD(device_detach,        kvm_clock_detach),
 251         DEVMETHOD(device_suspend,       kvm_clock_suspend),
 252         DEVMETHOD(device_resume,        kvm_clock_resume),
 253         /* clock interface */
 254         DEVMETHOD(clock_gettime,        kvm_clock_gettime),
 255         DEVMETHOD(clock_settime,        kvm_clock_settime),
 256
 257         DEVMETHOD_END
 258 };
 259
 260 static driver_t kvm_clock_driver = {
 261         KVM_CLOCK_DEVNAME,
 262         kvm_clock_methods,
 263         sizeof(struct kvm_clock_softc),
 264 };
 265
 266 DRIVER_MODULE(kvm_clock, nexus, kvm_clock_driver, kvm_clock_devclass, 0, 0);