svn commit: r347063 - head/sys/kern

[Ian Lepore]

On Mon, 2019-05-06 at 14:47 -0600, Warner Losh wrote:
> On Mon, May 6, 2019 at 2:40 PM John Baldwin <jhb at freebsd.org> wrote:
> 
> > On 5/6/19 11:45 AM, Mark Johnston wrote:
> > > On Mon, May 06, 2019 at 11:07:18AM -0700, John Baldwin wrote:
> > > > On 5/3/19 2:26 PM, Mark Johnston wrote:
> > > > > Author: markj
> > > > > Date: Fri May  3 21:26:44 2019
> > > > > New Revision: 347063
> > > > > URL: https://svnweb.freebsd.org/changeset/base/347063
> > > > > 
> > > > > Log:
> > > > >   Disallow excessively small times of day in clock_settime(2).
> > > > > 
> > > > >   Reported by:      syzkaller
> > > > >   Reviewed by:      cem, kib
> > > > >   MFC after:        1 week
> > > > >   Sponsored by:     The FreeBSD Foundation
> > > > >   Differential Revision:    https://reviews.freebsd.org/D20151
> > > > > 
> > > > > Modified:
> > > > >   head/sys/kern/kern_time.c
> > > > > 
> > > > > Modified: head/sys/kern/kern_time.c
> > > > > 
> > 
> > ==============================================================================
> > > > > --- head/sys/kern/kern_time.c       Fri May  3 21:13:09 2019
> > 
> > (r347062)
> > > > > +++ head/sys/kern/kern_time.c       Fri May  3 21:26:44 2019
> > 
> > (r347063)
> > > > > @@ -412,7 +412,9 @@ kern_clock_settime(struct thread *td, clockid_t
> > 
> > clock_
> > > > >     if (ats->tv_nsec < 0 || ats->tv_nsec >= 1000000000 ||
> > > > >         ats->tv_sec < 0)
> > > > >             return (EINVAL);
> > > > > -   if (!allow_insane_settime && ats->tv_sec > 8000ULL * 365 * 24 * 60
> > 
> > * 60)
> > > > > +   if (!allow_insane_settime &&
> > > > > +       (ats->tv_sec > 8000ULL * 365 * 24 * 60 * 60 ||
> > > > > +       ats->tv_sec < utc_offset()))
> > > > >             return (EINVAL);
> > > > >     /* XXX Don't convert nsec->usec and back */
> > > > >     TIMESPEC_TO_TIMEVAL(&atv, ats);
> > > > 
> > > > Pardon my ignorance, but I can't see why you are checking against
> > 
> > utc_offset()
> > > > vs some small constant?  None of the discussion in the review mentioned
> > 
> > the
> > > > reason for using this particular value, and I didn't see any comparisons
> > > > against utc_offset or kernadjtz in kern_clock_setttime() or settime()
> > 
> > that
> > > > would have underflowed or panicked.  Can you give a bit more detail on
> > 
> > why
> > > > utc_offset() is the lower bound?  Thanks.
> > > 
> > > I chose it because we subtract utc_offset() from the time passed in to
> > > clock_settime(); see settime_task_func().  That subtraction caused the
> > > underflow that later caused the observed panics.
> > 
> > Ok, thanks.  A few things I didn't see anyone else note in the review then:
> > 
> > 1) This subtraction is actually not done for all rtc drivers, so it seems
> >    like we might block small times for RTC clocks that set
> >    CLOCKF_GETTIME_NO_ADJ.
> > 2) utc_offset can be negative for machines using local time in timezones
> >    "before" UTC.
> > 
> > I suppose we don't think any FreeBSD machines actually need to set the
> > running clock to 0 anyway so fixing it here rather than rejecting invalid
> > values only for RTCs that can't handle it is probably ok, but the
> > connection doesn't feel obvious that we are rejecting times that might
> > be non-representable in RTCs.
> > 
> 
> I'd be tempted to change utc_offset() to be "< 2 * 86400" personally, since
> that's the worst-case UTC offset and will cover all cases for people that
> want to run in 1970. Though honestly, we could also substitute in 946684800
> (First second of 2000) as well, since who needs to run before 2000, right?
> 
> Warner

I do.  The first GPS epoch rollover was August 1999, and I've been
known to use a gps simulator to drive the system through that event and
make sure software handles it correctly.  Now I'd probably be more
likely to use the rollover that happened last month, but still... some
people like to live in the past, at least in simulations. :)

I wasn't even happy about the change that prevents setting time near
the unix epoch.  But, given that there is a tunable/sysctl to disable
the sanity checks, it's probably not all that important; we can set
debug.allow_insane_settime=1.

-- Ian