Re: [RFC PATCH 00/11] sched: CFS low-latency features

[Mathieu Desnoyers <mathieu.desnoyers@efficios.com>]

* Peter Zijlstra (peterz@infradead.org) wrote:
> On Fri, 2010-08-27 at 11:21 -0400, Mathieu Desnoyers wrote:
> >        SIGEV_THREAD
> >               Upon  timer  expiration,  invoke  sigev_notify_function as if it
> >               were the start function of a new thread.  (Among the implementa‐
> >               tion  possibilities  here are that each timer notification could
> >               result in the creation of a new thread, or that a single  thread
> >               is  created  to  receive  all  notifications.)   The function is
> >               invoked   with   sigev_value   as   its   sole   argument.    If
> >               sigev_notify_attributes  is  not  NULL,  it  should  point  to a
> >               pthread_attr_t structure that defines  attributes  for  the  new
> >               thread (see pthread_attr_init(3).
> > 
> > So basically, it's the glibc implementation that is broken, not the standard.
> 
> The standard is broken too, what context will the new thread inherit?

Besides pthread_attr_t, thinking of the scheduler/cgroups/etc stuff, I'd think
it might be expected to inherit the state of the thread which calls
timer_create(). But this is not what glibc does right now, and it is not spelled
out clearly by the standard.

> The pthread_attr_t stuff tries to cover some of that, but pthread_attr_t
> doesn't cover all inherited task attributes, and allows for some very
> 'interesting' bugs [1].

(see below)

> 
> The specification also doesn't cover the case where the handler takes
> more time to execute than the timer interval.

Why should it ? It seems valid for a workload to result in spawning many threads
bound to more than a single core on a multi-core system. So concurrency
management should be performed by the application.

> 
> [1] - consider the case where pthread_attr_t includes the stack and we
> use a spawn thread on expire policy and then run into the situation
> where the handler is delayed past the next expiration.

Setting a thread stack and generating the signal more than once is taken into
account in the standard. It leads to unspecified result (IOW: don't do this):

http://www.opengroup.org/onlinepubs/009695399/functions/timer_create.html

"If evp->sigev_sigev_notify is SIGEV_THREAD and sev->sigev_notify_attributes is
not NULL, if the attribute pointed to by sev->sigev_notify_attributes has a
thread stack address specified by a call to pthread_attr_setstack() or
pthread_attr_setstackaddr(), the results are unspecified if the signal is
generated more than once."

Thanks,

Mathieu

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com