Re: [PATCH 3/3] timer: Reduce unnecessary sighand lock contention

[Frederic Weisbecker <fweisbec@gmail.com>]

On Wed, Aug 26, 2015 at 04:32:34PM -0700, Jason Low wrote:
> On Thu, 2015-08-27 at 00:56 +0200, Frederic Weisbecker wrote:
> > On Tue, Aug 25, 2015 at 08:17:48PM -0700, Jason Low wrote:
> > > It was found while running a database workload on large systems that
> > > significant time was spent trying to acquire the sighand lock.
> > > 
> > > The issue was that whenever an itimer expired, many threads ended up
> > > simultaneously trying to send the signal. Most of the time, nothing
> > > happened after acquiring the sighand lock because another thread
> > > had already sent the signal and updated the "next expire" time. The
> > > fastpath_timer_check() didn't help much since the "next expire" time
> > > was updated later.
> > >  
> > > This patch addresses this by having the thread_group_cputimer structure
> > > maintain a boolean to signify when a thread in the group is already
> > > checking for process wide timers, and adds extra logic in the fastpath
> > > to check the boolean.
> > > 
> > > Signed-off-by: Jason Low <jason.low2@hp.com>
> > > ---
> > >  include/linux/init_task.h      |    1 +
> > >  include/linux/sched.h          |    3 +++
> > >  kernel/time/posix-cpu-timers.c |   19 +++++++++++++++++--
> > >  3 files changed, 21 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/include/linux/init_task.h b/include/linux/init_task.h
> > > index d0b380e..3350c77 100644
> > > --- a/include/linux/init_task.h
> > > +++ b/include/linux/init_task.h
> > > @@ -53,6 +53,7 @@ extern struct fs_struct init_fs;
> > >  	.cputimer	= { 						\
> > >  		.cputime_atomic	= INIT_CPUTIME_ATOMIC,			\
> > >  		.running	= 0,					\
> > > +		.checking_timer	= 0,					\
> > >  	},								\
> > >  	INIT_PREV_CPUTIME(sig)						\
> > >  	.cred_guard_mutex =						\
> > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > index 119823d..a6c8334 100644
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -619,6 +619,8 @@ struct task_cputime_atomic {
> > >   * @cputime_atomic:	atomic thread group interval timers.
> > >   * @running:		non-zero when there are timers running and
> > >   * 			@cputime receives updates.
> > > + * @checking_timer:	non-zero when a thread is in the process of
> > > + *			checking for thread group timers.
> > >   *
> > >   * This structure contains the version of task_cputime, above, that is
> > >   * used for thread group CPU timer calculations.
> > > @@ -626,6 +628,7 @@ struct task_cputime_atomic {
> > >  struct thread_group_cputimer {
> > >  	struct task_cputime_atomic cputime_atomic;
> > >  	int running;
> > > +	int checking_timer;
> > 
> > How about a flag in the "running" field instead?
> > 
> > 1) Space in signal_struct is not as important as in task_strut but it
> >    still matters.
> 
> George Spelvin suggested that we convert them to booleans which would
> make them take up 2 bytes.
> 
> > 2) We already read the "running" field locklessly. Adding a new field like
> >    checking_timer gets even more complicated. Ideally there should be at
> >    least a paired memory barrier between both. Let's just simplify that
> >    with a single field.
> 
> hmmm, so having 1 "flag" where we access bits for the "running" and
> "checking_timer"?

Sure, like:

#define CPUTIMER_RUNNING 0x1
#define CPUTIMER_CHECKING 0x2

struct thread_group_cputimer {
    struct task_cputime_atomic cputime_atomic;
    int status;
}

So from cputimer_running() you just need to check:

     if (cputimer->status & CPUTIMER_RUNNING)

And from run_posix_cpu_timer() fast-path:

     if (cputimer->status == CPUTIMER_RUNNING)

so that ignores CPUTIMER_CHECKING case.

> 
> > Now concerning the solution for your problem, I'm a bit uncomfortable with
> > lockless magics like this. When the thread sets checking_timer to 1, there is
> > no guarantee that the other threads in the process will see it "fast" enough
> > to avoid the slow path checks. Then there is also the risk that the threads
> > don't see "fast" enough that checking_timers has toggled to 0 and as a result
> > a timer may expire late. Now the lockless access of "running" already induces
> > such race. So if it really solves issues in practice, why not.
> 
> Perhaps to be safer, we use something like load_acquire() and
> store_release() for accessing both the ->running and ->checking_timer
> fields?

Well it depends against what we want to order them. If it's a single field
we don't need to order them together at least.