LTTng » LTTng-UST

09:37 < Compudj> pwoegere: you enable events in two channels
09:37 < Compudj> this would enabled events both in .so and in the executable twice, right ?
09:38 < Compudj> "As already stated, this effect only occurs when the tracepoint gets emitted from a dlopened shared lib. For all other emitted tracepoints the behavior is as it was until LTTng 2.1." -> ah, it seems not
09:39 < pwoegere> Compudj: My understanding was that if an event was enabled for channel A, it should not go also in channel B (even if B collects all remaining userspace events)
09:40 < Compudj> pwoegere: I think early versions were doing this, but it was more an unfortunate design issue than done on purpose
09:40 < pwoegere> Compudj: otherwhise how can you reasonably control which events go where ?
09:40 < Compudj> you don't enable "all" events
09:41 < Compudj> or eventually, we'd introduce "all except A B and C" 
09:41 < pwoegere> Compudj: lttng enable-event -u -a -c ch2
09:41 < Compudj> I mean: if you enable all events, you will have duplicates
09:41 < Compudj> (if an event ends up matching for 2 channels)
09:41 < Compudj> e.g.:
09:41 < Compudj> lttng enable-event -u -a -c ch1
09:41 < Compudj> lttng enable-event -u -a -c ch2
09:41 < Compudj> you will have all events in each channel
09:42 < pwoegere> Compudj: Hmm
09:42 < pwoegere> Compudj: I would not have expected that
09:42 < pwoegere> Compudj: what sense could it make to get the same event persisted in two channels ?
09:43 < Compudj> each channel can have its own config
09:43 < Compudj> e.g. different buffer size, timers, behavior in general
09:43 < Compudj> so we might want to put important info into one
09:44 < Compudj> eventually, also, we might want to put an event with a specific loglevel or filter associated into a channel
09:44 < Compudj> so, an example of valid use-case:
09:44 < Compudj> enable all events of loglevel debug with filter on TID 123 into channel A
09:44 < Compudj> and enable some events of loglevel critical into channel B
09:45 < Compudj> so the "all events" set is made more specific by the loglevel and filter info
09:45 < Compudj> so someone could very well want to enable a "duplicate" of an event with different loglevel and filter info
09:46 < Compudj> (into another channel, _or_ into the same channel)
09:46 < pwoegere> Compudj: Hmmm, I though of it more like the unix pipe. data flows from one place to another. If you want duplicate use explicit tee
09:47 < Compudj> events can be written into all the sessions dans are interested in them
09:47 < Compudj> and can also be written in all channels that are interested in them
09:47 < Compudj> however, let's say that you enable an event twice for the same channel
09:47 < Compudj> with a wildcard and the specific event name
09:47 < Compudj> it's only going to be saved once
09:47 < Compudj> no duplicate
09:47 < Compudj> but if you do so for multiple channels, how should we know which one to keep ?
09:48 < pwoegere> Compudj: by using the mental model of a pipe
09:49 < pwoegere> Compudj: e.g.
09:49 < pwoegere> Compudj: lttng enable-event 'my_events:*' -u -c ch1 --dontConsume
09:50 < Compudj> what is --dontConsume ?
09:50 < pwoegere> Compudj: I made it up for demonstration purpose
09:51 < Compudj> it puts a ring buffer in overwrite mode, without consumer attached ?
09:51 < pwoegere> lttng enable-event 'my_events:*' -u -c ch1
09:52 < pwoegere> with --dontConsume would mean capute that kind of event in ch1 but do not take it from the pipe so that a later ...
09:52 < pwoegere> lttng enable-event -u -a -c ch2
09:52 < pwoegere> would also hava a chance to capture it
09:53 < pwoegere> Compudj: but I guess thats already too late, since the semantic is already defined differently
09:54 < Compudj> indeed
09:55 < Compudj> if you want close control on which event goes where, you might want to rethink you use of the wildcards (*)
09:55 < Compudj> so an event won't match for two channels
09:56 < Compudj> by the way, it also behaves like this if the event is in the program main executable
09:57 < Compudj> so the behavior seems coherent
09:59 < pwoegere> Compudj: Ok, for now I will adjust my script to not make use of multiple userspace channels anymore. this has to wait until we have event exclusion. Thanks !
10:00 < Compudj> pwoegere: sorry about that.. this slight change of behavior has been required by the addition of loglevel and filter features