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/cdefs.h>
  40 #include <sys/param.h>
  41 #include <sys/bus.h>
  42 #include <sys/domainset.h>
  43 #include <sys/kernel.h>
  44 #include <sys/malloc.h>
  45 #include <sys/module.h>
  46 #include <sys/smp.h>
  47 #include <sys/sysctl.h>
  48
  49 #include <vm/vm.h>
  50 #include <vm/pmap.h>
  51 #include <vm/vm_extern.h>
  52
  53 #include <machine/pvclock.h>
  54 #include <x86/kvm.h>
  55
  56 #include "clock_if.h"
  57
  58 #define KVM_CLOCK_DEVNAME               "kvmclock"
  59 /*
  60  * Note: Chosen to be (1) above HPET's value (always 950), (2) above the TSC's
  61  * default value of 800, and (3) below the TSC's value when it supports the
  62  * "Invariant TSC" feature and is believed to be synchronized across all CPUs.
  63  */
  64 #define KVM_CLOCK_TC_QUALITY            975
  65
  66 struct kvm_clock_softc {
  67         struct pvclock                   pvc;
  68         struct pvclock_wall_clock        wc;
  69         struct pvclock_vcpu_time_info   *timeinfos;
  70         u_int                            msr_tc;
  71         u_int                            msr_wc;
  72 #ifndef EARLY_AP_STARTUP
  73         int                              firstcpu;
  74 #endif
  75 };
  76
  77 static struct pvclock_wall_clock *kvm_clock_get_wallclock(void *arg);
  78 static void     kvm_clock_system_time_enable(struct kvm_clock_softc *sc,
  79                     const cpuset_t *cpus);
  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, const cpuset_t *cpus)
  94 {
  95         smp_rendezvous_cpus(*cpus, NULL, kvm_clock_system_time_enable_pcpu,
  96             NULL, sc);
  97 }
  98
  99 static void
 100 kvm_clock_system_time_enable_pcpu(void *arg)
 101 {
 102         struct kvm_clock_softc *sc = arg;
 103
 104         /*
 105          * See [2]; the lsb of this MSR is the system time enable bit.
 106          */
 107         wrmsr(sc->msr_tc, vtophys(&(sc->timeinfos)[curcpu]) | 1);
 108 }
 109
 110 #ifndef EARLY_AP_STARTUP
 111 static void
 112 kvm_clock_init_smp(void *arg __unused)
 113 {
 114         devclass_t kvm_clock_devclass;
 115         cpuset_t cpus;
 116         struct kvm_clock_softc *sc;
 117
 118         kvm_clock_devclass = devclass_find(KVM_CLOCK_DEVNAME);
 119         sc = devclass_get_softc(kvm_clock_devclass, 0);
 120         if (sc == NULL || mp_ncpus == 1)
 121                 return;
 122
 123         /*
 124          * Register with the hypervisor on all CPUs except the one that
 125          * registered in kvm_clock_attach().
 126          */
 127         cpus = all_cpus;
 128         KASSERT(CPU_ISSET(sc->firstcpu, &cpus),
 129             ("%s: invalid first CPU %d", __func__, sc->firstcpu));
 130         CPU_CLR(sc->firstcpu, &cpus);
 131         kvm_clock_system_time_enable(sc, &cpus);
 132 }
 133 SYSINIT(kvm_clock, SI_SUB_SMP, SI_ORDER_ANY, kvm_clock_init_smp, NULL);
 134 #endif
 135
 136 static void
 137 kvm_clock_identify(driver_t *driver, device_t parent)
 138 {
 139         u_int regs[4];
 140
 141         kvm_cpuid_get_features(regs);
 142         if ((regs[0] &
 143             (KVM_FEATURE_CLOCKSOURCE2 | KVM_FEATURE_CLOCKSOURCE)) == 0)
 144                 return;
 145         if (device_find_child(parent, KVM_CLOCK_DEVNAME, -1))
 146                 return;
 147         BUS_ADD_CHILD(parent, 0, KVM_CLOCK_DEVNAME, 0);
 148 }
 149
 150 static int
 151 kvm_clock_probe(device_t dev)
 152 {
 153         device_set_desc(dev, "KVM paravirtual clock");
 154         return (BUS_PROBE_DEFAULT);
 155 }
 156
 157 static int
 158 kvm_clock_attach(device_t dev)
 159 {
 160         u_int regs[4];
 161         struct kvm_clock_softc *sc = device_get_softc(dev);
 162         bool stable_flag_supported;
 163
 164         /* Process KVM "features" CPUID leaf content: */
 165         kvm_cpuid_get_features(regs);
 166         if ((regs[0] & KVM_FEATURE_CLOCKSOURCE2) != 0) {
 167                 sc->msr_tc = KVM_MSR_SYSTEM_TIME_NEW;
 168                 sc->msr_wc = KVM_MSR_WALL_CLOCK_NEW;
 169         } else {
 170                 KASSERT((regs[0] & KVM_FEATURE_CLOCKSOURCE) != 0,
 171                     ("Clocksource feature flags disappeared since "
 172                     "kvm_clock_identify: regs[0] %#0x.", regs[0]));
 173                 sc->msr_tc = KVM_MSR_SYSTEM_TIME;
 174                 sc->msr_wc = KVM_MSR_WALL_CLOCK;
 175         }
 176         stable_flag_supported =
 177             (regs[0] & KVM_FEATURE_CLOCKSOURCE_STABLE_BIT) != 0;
 178
 179         /* Set up 'struct pvclock_vcpu_time_info' page(s): */
 180         sc->timeinfos = kmem_malloc(mp_ncpus *
 181             sizeof(struct pvclock_vcpu_time_info), M_WAITOK | M_ZERO);
 182 #ifdef EARLY_AP_STARTUP
 183         kvm_clock_system_time_enable(sc, &all_cpus);
 184 #else
 185         sc->firstcpu = curcpu;
 186         kvm_clock_system_time_enable_pcpu(sc);
 187 #endif
 188
 189         /*
 190          * Init pvclock; register KVM clock wall clock, register KVM clock
 191          * timecounter, and set up the requisite infrastructure for vDSO access
 192          * to this timecounter.
 193          *     Regarding 'tc_flags': Since the KVM MSR documentation does not
 194          *     specifically discuss suspend/resume scenarios, conservatively
 195          *     leave 'TC_FLAGS_SUSPEND_SAFE' cleared and assume that the system
 196          *     time must be re-inited in such cases.
 197          */
 198         sc->pvc.get_wallclock = kvm_clock_get_wallclock;
 199         sc->pvc.get_wallclock_arg = sc;
 200         sc->pvc.timeinfos = sc->timeinfos;
 201         sc->pvc.stable_flag_supported = stable_flag_supported;
 202         pvclock_init(&sc->pvc, dev, KVM_CLOCK_DEVNAME, KVM_CLOCK_TC_QUALITY, 0);
 203         kvm_clock_setup_sysctl(dev);
 204         return (0);
 205 }
 206
 207 static int
 208 kvm_clock_detach(device_t dev)
 209 {
 210         struct kvm_clock_softc *sc = device_get_softc(dev);
 211
 212         return (pvclock_destroy(&sc->pvc));
 213 }
 214
 215 static int
 216 kvm_clock_suspend(device_t dev)
 217 {
 218         return (0);
 219 }
 220
 221 static int
 222 kvm_clock_resume(device_t dev)
 223 {
 224         /*
 225          * See note in 'kvm_clock_attach()' regarding 'TC_FLAGS_SUSPEND_SAFE';
 226          * conservatively assume that the system time must be re-inited in
 227          * suspend/resume scenarios.
 228          */
 229         kvm_clock_system_time_enable(device_get_softc(dev), &all_cpus);
 230         pvclock_resume();
 231         inittodr(time_second);
 232         return (0);
 233 }
 234
 235 static int
 236 kvm_clock_gettime(device_t dev, struct timespec *ts)
 237 {
 238         struct kvm_clock_softc *sc = device_get_softc(dev);
 239
 240         pvclock_gettime(&sc->pvc, ts);
 241         return (0);
 242 }
 243
 244 static int
 245 kvm_clock_settime(device_t dev, struct timespec *ts)
 246 {
 247         /*
 248          * Even though it is not possible to set the KVM clock's wall clock, to
 249          * avoid the possibility of periodic benign error messages from
 250          * 'settime_task_func()', report success rather than, e.g., 'ENODEV'.
 251          */
 252         return (0);
 253 }
 254
 255 static int
 256 kvm_clock_tsc_freq_sysctl(SYSCTL_HANDLER_ARGS)
 257 {
 258         struct kvm_clock_softc *sc = oidp->oid_arg1;
 259         uint64_t freq = pvclock_tsc_freq(sc->timeinfos);
 260
 261         return (sysctl_handle_64(oidp, &freq, 0, req));
 262 }
 263
 264 static void
 265 kvm_clock_setup_sysctl(device_t dev)
 266 {
 267         struct kvm_clock_softc *sc = device_get_softc(dev);
 268         struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(dev);
 269         struct sysctl_oid *tree = device_get_sysctl_tree(dev);
 270         struct sysctl_oid_list *child = SYSCTL_CHILDREN(tree);
 271
 272         SYSCTL_ADD_PROC(ctx, child, OID_AUTO, "tsc_freq",
 273             CTLTYPE_U64 | CTLFLAG_RD | CTLFLAG_MPSAFE, sc, 0,
 274             kvm_clock_tsc_freq_sysctl, "QU",
 275             "Time Stamp Counter frequency");
 276 }
 277
 278 static device_method_t kvm_clock_methods[] = {
 279         DEVMETHOD(device_identify,      kvm_clock_identify),
 280         DEVMETHOD(device_probe,         kvm_clock_probe),
 281         DEVMETHOD(device_attach,        kvm_clock_attach),
 282         DEVMETHOD(device_detach,        kvm_clock_detach),
 283         DEVMETHOD(device_suspend,       kvm_clock_suspend),
 284         DEVMETHOD(device_resume,        kvm_clock_resume),
 285         /* clock interface */
 286         DEVMETHOD(clock_gettime,        kvm_clock_gettime),
 287         DEVMETHOD(clock_settime,        kvm_clock_settime),
 288
 289         DEVMETHOD_END
 290 };
 291
 292 static driver_t kvm_clock_driver = {
 293         KVM_CLOCK_DEVNAME,
 294         kvm_clock_methods,
 295         sizeof(struct kvm_clock_softc),
 296 };
 297
 298 DRIVER_MODULE(kvm_clock, nexus, kvm_clock_driver, 0, 0);