Re: [Question] Hooks for scheduler tracing (CFS)

* Frank Ch. Eigler (fche@redhat.com) wrote:

Mathieu Desnoyers [off-list ref] writes:

quoted

[...]
The problem is also in _stp_print_flush, not *only* in relay code:
void _stp_print_flush (void)
...
                spin_lock(&_stp_print_lock);
                ...
                spin_unlock(&_stp_print_lock);

Those will turn into mutexes with -rt.

Indeed, plus systemtap-generated locking code uses rwlocks,
local_irq_save/restore or preempt_disable, in various places.  Could
someone point to a place that spells out what would be more
appropriate way of ensuring atomicity while being compatible with -rt?

- FChE

AFAIK, for your needs either:
- Use atomic ops
- Use per-cpu data with preempt disabling/irq disabling
- Use the original "real" spin locks/rwlocks (raw_*).
- Don't play with timers or wakeups, since this kernel code uses the
  "standard" spin locks (sleepable in -rt).

You just don't want to sleep in the tracing code.

Make sure that the sub-buffer switch code respects that too: it is the
most tricky and yet less executed part of the tracing code, so it's easy
for bugs to slip there and yet be undetected for a while.

Since you will likely disable preemption, make sure your tracing code
executes in a deterministic time.

Mathieu

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68