svn commit: r300110 - head/sys/kern

[Mark Johnston]

On Tue, May 17, 2016 at 10:22:18PM -0700, NGie Cooper wrote:
> 
> > On May 17, 2016, at 20:55, Mark Johnston <markj at FreeBSD.org> wrote:
> > 
> > Author: markj
> > Date: Wed May 18 03:55:54 2016
> > New Revision: 300110
> > URL: https://svnweb.freebsd.org/changeset/base/300110
> > 
> > Log:
> >  Do not acquire the thread lock in hardclock_cnt() unless needed.
> > 
> >  This function only sets thread flags if a SIGPROF or SIGVTALRM timer
> >  has fired, which is almost never the case.
> > 
> >  MFC after:    2 weeks
> > 
> > Modified:
> >  head/sys/kern/kern_clock.c
> > 
> > Modified: head/sys/kern/kern_clock.c
> > ==============================================================================
> > --- head/sys/kern/kern_clock.c    Wed May 18 03:50:21 2016    (r300109)
> > +++ head/sys/kern/kern_clock.c    Wed May 18 03:55:54 2016    (r300110)
> > @@ -570,9 +570,11 @@ hardclock_cnt(int cnt, int usermode)
> >            flags |= TDF_PROFPEND | TDF_ASTPENDING;
> >        PROC_ITIMUNLOCK(p);
> >    }
> > -    thread_lock(td);
> > -    td->td_flags |= flags;
> > -    thread_unlock(td);
> > +    if (flags != 0) {
> > +        thread_lock(td);
> > +        td->td_flags |= flags;
> > +        thread_unlock(td);
> > +    }
> 
> Use predict_false?

I don't think that's really warranted here. A __predict_false could
become incorrect after a future change elsewhere in this function, and
this isn't enough of a hot path for such a change to make a measurable
difference on its own.

> > 
> > #ifdef    HWPMC_HOOKS
> >    if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid)))
> >