Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
From:   Thomas Gleixner <tglx@linutronix.de>
To:     Paolo Bonzini <pbonzini@redhat.com>,
        Miklos Szeredi <miklos@szeredi.hu>
Cc:     kvm@vger.kernel.org, Vincenzo Frascino <vincenzo.frascino@arm.com>,
        Juergen Gross <jgross@suse.com>, linux-kernel@vger.kernel.org
Subject: Re: system time goes weird in kvm guest after host suspend/resume
In-Reply-To: <1a1c32fe-d124-0e47-c9e4-695be7ea7567@redhat.com>
Date:   Fri, 05 Jun 2020 12:11:59 +0200
Message-ID: <87k10l7rf4.fsf@nanos.tec.linutronix.de>
MIME-Version: 1.0
Content-Type: text/plain
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Paolo Bonzini <pbonzini@redhat.com> writes:
> On 04/06/20 21:28, Miklos Szeredi wrote:
>> time(2) returns good time, while clock_gettime(2) returns bad time.
>> Here's an example:
>> 
>> time=1591298725 RT=1591300383 MONO=39582 MONO_RAW=39582 BOOT=39582
>> time=1591298726 RT=1591300383 MONO=39582 MONO_RAW=39582 BOOT=39582
>> time=1591298727 RT=1591300383 MONO=39582 MONO_RAW=39582 BOOT=39582
>> time=1591298728 RT=1591300383 MONO=39582 MONO_RAW=39582 BOOT=39582
>> time=1591298729 RT=1591300383 MONO=39582 MONO_RAW=39582 BOOT=39582
>> 
>> As you can see, only time(2) is updated, the others remain the same.
>> date(1) uses clock_gettime(CLOCK_REALTIME) so that shows the bad date.
>> 
>> When the correct time reaches the value returned by CLOCK_REALTIME,
>> the value jumps exactly 2199 seconds.

Which value jumps?

> clockid_to_kclock(CLOCK_REALTIME) is &clock_realtime, so clock_gettime
> calls ktime_get_real_ts64, which is:
>
>
>         do {
>                 seq = read_seqcount_begin(&tk_core.seq);
>
>                 ts->tv_sec = tk->xtime_sec;
>                 nsecs = timekeeping_get_ns(&tk->tkr_mono);
>
>         } while (read_seqcount_retry(&tk_core.seq, seq));
>
>         ts->tv_nsec = 0;
>         timespec64_add_ns(ts, nsecs);
>
> time(2) instead should actually be gettimeofday(2), which just returns
> tk->xtime_sec.

time(2) is either handled in the VDSO or it is handled via syscall and
yes, it's only looking at the xtime_sec value.

gettimeofday(2) returns seconds and microseconds. It's using the same
mechanism as clock_gettime(2) and divides the nanoseconds part by 1000.

> So the problem is the nanosecond part which is off by
> 2199*10^9 nanoseconds, and that is suspiciously close to 2^31...

Not really. It's 2^41.

I can actually now reproduce, but I won't be able to investigate that
before monday.

Thanks,

        tglx