The branch main has been updated by markj:
URL:
https://cgit.FreeBSD.org/src/commit/?id=568f552b0410f4f72db9ce710f7803acca23f79f
commit 568f552b0410f4f72db9ce710f7803acca23f79f
Author: Mark Johnston <ma...@freebsd.org>
AuthorDate: 2023-01-13 15:01:00 +0000
Commit: Mark Johnston <ma...@freebsd.org>
CommitDate: 2023-01-13 15:37:37 +0000
kvmclock: Fix initialization when EARLY_AP_STARTUP is not defined
To attach to the hypervisor, kvmclock needs to write a per-CPU MSR.
When EARLY_AP_STARTUP is not defined, device attach happens too early:
APs are not yet spun up, so smp_rendezvous only runs the callback on the
local CPU. As a result, the timecounter only gets initialized on the
BSP, and then timekeeping is broken on SMP systems.
Implement handling for !EARLY_AP_STARTUP kernels: keep track of the CPU
on which device attach ran, and then use a SI_SUB_SMP SYSINIT to
register the rest of the CPUs with the hypervisor.
Reported by: Shrikanth R Kamath <kshrika...@juniper.net>
Reviewed by: kib, jhb (earlier versions)
Sponsored by: Klara, Inc.
Sponsored by: Juniper Networks, Inc.
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D37705
---
sys/dev/kvm_clock/kvm_clock.c | 46 ++++++++++++++++++++++++++++++++++++++-----
1 file changed, 41 insertions(+), 5 deletions(-)
diff --git a/sys/dev/kvm_clock/kvm_clock.c b/sys/dev/kvm_clock/kvm_clock.c
index 81c617755385..e6ae03ca01c0 100644
--- a/sys/dev/kvm_clock/kvm_clock.c
+++ b/sys/dev/kvm_clock/kvm_clock.c
@@ -71,10 +71,14 @@ struct kvm_clock_softc {
struct pvclock_vcpu_time_info *timeinfos;
u_int msr_tc;
u_int msr_wc;
+#ifndef EARLY_AP_STARTUP
+ int firstcpu;
+#endif
};
static struct pvclock_wall_clock *kvm_clock_get_wallclock(void *arg);
-static void kvm_clock_system_time_enable(struct kvm_clock_softc *sc);
+static void kvm_clock_system_time_enable(struct kvm_clock_softc *sc,
+ const cpuset_t *cpus);
static void kvm_clock_system_time_enable_pcpu(void *arg);
static void kvm_clock_setup_sysctl(device_t);
@@ -88,9 +92,10 @@ kvm_clock_get_wallclock(void *arg)
}
static void
-kvm_clock_system_time_enable(struct kvm_clock_softc *sc)
+kvm_clock_system_time_enable(struct kvm_clock_softc *sc, const cpuset_t *cpus)
{
- smp_rendezvous(NULL, kvm_clock_system_time_enable_pcpu, NULL, sc);
+ smp_rendezvous_cpus(*cpus, NULL, kvm_clock_system_time_enable_pcpu,
+ NULL, sc);
}
static void
@@ -104,6 +109,32 @@ kvm_clock_system_time_enable_pcpu(void *arg)
wrmsr(sc->msr_tc, vtophys(&(sc->timeinfos)[curcpu]) | 1);
}
+#ifndef EARLY_AP_STARTUP
+static void
+kvm_clock_init_smp(void *arg __unused)
+{
+ devclass_t kvm_clock_devclass;
+ cpuset_t cpus;
+ struct kvm_clock_softc *sc;
+
+ kvm_clock_devclass = devclass_find(KVM_CLOCK_DEVNAME);
+ sc = devclass_get_softc(kvm_clock_devclass, 0);
+ if (sc == NULL || mp_ncpus == 1)
+ return;
+
+ /*
+ * Register with the hypervisor on all CPUs except the one that
+ * registered in kvm_clock_attach().
+ */
+ cpus = all_cpus;
+ KASSERT(CPU_ISSET(sc->firstcpu, &cpus),
+ ("%s: invalid first CPU %d", __func__, sc->firstcpu));
+ CPU_CLR(sc->firstcpu, &cpus);
+ kvm_clock_system_time_enable(sc, &cpus);
+}
+SYSINIT(kvm_clock, SI_SUB_SMP, SI_ORDER_ANY, kvm_clock_init_smp, NULL);
+#endif
+
static void
kvm_clock_identify(driver_t *driver, device_t parent)
{
@@ -150,7 +181,12 @@ kvm_clock_attach(device_t dev)
/* Set up 'struct pvclock_vcpu_time_info' page(s): */
sc->timeinfos = kmem_malloc(mp_ncpus *
sizeof(struct pvclock_vcpu_time_info), M_WAITOK | M_ZERO);
- kvm_clock_system_time_enable(sc);
+#ifdef EARLY_AP_STARTUP
+ kvm_clock_system_time_enable(sc, &all_cpus);
+#else
+ sc->firstcpu = curcpu;
+ kvm_clock_system_time_enable_pcpu(sc);
+#endif
/*
* Init pvclock; register KVM clock wall clock, register KVM clock
@@ -192,7 +228,7 @@ kvm_clock_resume(device_t dev)
* conservatively assume that the system time must be re-inited in
* suspend/resume scenarios.
*/
- kvm_clock_system_time_enable(device_get_softc(dev));
+ kvm_clock_system_time_enable(device_get_softc(dev), &all_cpus);
pvclock_resume();
inittodr(time_second);
return (0);
