Re: [PATCH -tip 3/3] Add get_signal tracepoint

[Ingo Molnar <mingo@elte.hu>]

* Roland McGrath <roland@redhat.com> wrote:

> This is orthogonal to the core-dump tracepoint, I don't see why you 
> call them a unified patch series.
> 
> The proper name for this event is "signal delivery".  But since the 
> proper name for "send_signal" is "signal generation", I suppose "get" 
> is analogously improper to the existing "send" tracepoint.  ;-)

I'd suggest to add include/trace/events/signal.h and put these 
tracepoints there.

> Especially if you call this "get" rather than "deliver", there is 
> another place that should invoke this tracepoint (or perhaps a third 
> one).  sys_rt_sigtimedwait "gets" a signal without delivering it.  In 
> POSIX terminology this is called "accepting" the signal: the three 
> things that can happen in the life of a signal are "generate", 
> "deliver", and "accept".  If you are trying to match up what happened 
> to a signal generated by kill() or whatnot, then you want to notice 
> both delivery and acceptance as the complementary event.
> 
> (And again I have no clue why this signal stuff should be called 
> "sched" at all.)

it shouldnt be called 'sched' - it should go into 'events/signal.h'.

But we also need fuller coverage than this. Coredumps and signal 
delivery events are just a small part of all things signals, we also 
want:

 - signal generation events (send_sig*() variants)

 - signal IPI/wakeup events

 - signal loss events (queue overflow)

 - [ optional: signal blocking/unblocking events ]

 - [ optional: specific signal handler installation/deinstallation ]

That's what we generally require of new events: they should form a 
coherent whole, a logical set of events that 'make sense' and explain 
the workings of a subsystem on a given level of detail.

How finegrained or coarse the level of details is is an open question, 
but if a given level of detail has been picked, we want completeness on 
that level.

So for example in the list above, the '[ optional ]' events are 
finegrained ones that could be left out of the initial version.

We've done this consistently for all subsystems that added tracepoints: 
scheduling, locking, timers, workqueues, block IO, SLAB, IRQs, etc., and 
we want a similar approach for newly covered subsystems (such as 
signals) as well.

Thanks,

	Ingo