Re: [ANNOUNCE] New utility: 'trace'

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

* Tom Zanussi (tzanussi@gmail.com) wrote:
> On Wed, 2010-11-17 at 13:36 -0500, Mathieu Desnoyers wrote:
> > * Tom Zanussi (tzanussi@gmail.com) wrote:
> > [...]
> > > IIRC, I think the conclusion we came to was that it could be done
> > > mechanically if for example the right-hand-side of an assignment in
> > > TP_fast_assign() only involved a simple variable assignment, but as
> > > Steve pointed out, some assignments are more complicated than that.
> > 
> > Yep, we came up to the same conclusions in UST.
> > 
> > > For example, in the sched_switch tracepoint assignments:
> > > 
> > > 	__entry->prev_prio      = prev->prio;
> > >         __entry->prev_state     = __trace_sched_switch_state(prev);
> > > 
> > > so the prev_prio should be able to be tested 'in-line' but the
> > > prev_state would require a temporary buffer to write the value into
> > > before doing the test as mentioned by Steve.  At which point you're no
> > > further ahead (in that case) than the current situation...
> > 
> > if we change all assignments to, e.g.:
> > 
> > _tp_assign(__entry->prev_prio, prev->prio)
> > _tp_assign(__entry->prev_state, __trace_sched_switch_state(prev))
> > 
> > then we can redefine the macros for filtering much more easily than with the
> > " = " assignment operator.
> > 
> > About your comment above, what is the problem with evaluating
> > "__trace_sched_switch_state(prev)" twice ? It will typically be cache-hot after
> > the first evaluation, so I wonder if, in practice, we really save a significant
> > amount of cycles by saving its result between filtering and writing into trace
> > buffers. As I pointed out earlier, for my customers, having a very, very fast
> > filter "out" case is more important that trying to squeeze a few cycles out of
> > the filter passed case.
> > 
> 
> But the idea is to avoid allocating the trace buffer in the first place,
> until we've decided we want the event.  So how do you check the result
> of __trace_sched_switch_state(prev) with the filter value if you don't
> have it temporarily stored somewhere (not in the trace buffer, which
> doesn't exist yet as far as this event is concerned)?

It seems I might be missing something important, but what's wrong with using
registers or the stack to hold the value for comparison ? In this case, it's a
"long", so a register seems perfectly reasonable. But again, I feel I'm missing
a key point -- what is it ?

Thanks,

Mathieu

> 
> Tom
> 
> > Also, how many of these "__trace_sched_switch_state(prev)" are static inlines vs
> > actual function calls ? If it's mostly static inlines to dereference a few
> > pointers, doing it the second time when the filter passed won't hurt much.
> > 
> > Thanks,
> > 
> > Mathieu
> > 
> 
> 

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