Re: [PATCH 0/11] LTTng-core (basic tracing infrastructure) 0.5.108

[Martin Bligh <mbligh@mbligh.org>]

Ingo Molnar wrote:
> * Martin Bligh <mbligh@mbligh.org> wrote:
> 
>>>i very much agree that they should become as fast as possible. So to 
>>>rephrase the question: can we make dynamic tracepoints as fast (or 
>>>nearly as fast) as static tracepoints? If yes, should we care about 
>>>static tracers at all?
>>
>>Depends how many nops you're willing to add, I guess. Anything, even 
>>the static tracepoints really needs at least a branch to be useful, 
>>IMHO. At least for what I've been doing with it, you need to stop the 
>>data flow after a while (when the event you're interested in happens, 
>>I'm using it like a flight data recorder, so we can go back and do 
>>postmortem on what went wrong). I should imagine branch prediction 
>>makes it very cheap on most modern CPUs, but don't have hard data to 
>>hand.
> 
> only 5 bytes of NOP are needed by default, so that a kprobe can insert a 
> call/callq instruction. The easiest way in practice is to insert a 
> _single_, unconditional function call that is patched out to NOPs upon 
> its first occurance (doing this is not a performance issue at all). That 
> way the only cost is the NOP and the function parameter preparation 
> side-effects. (which might or might not be significant - with register 
> calling conventions and most parameters being readily available it 
> should be small.)
> 
> note that such a limited, minimally invasive 'data extraction point' 
> infrastructure is not actually what the LTT patches are doing. It's not 
> even close, and i think you'll be surprised. Let me quote from the 
> latest LTT patch (patch-2.6.17-lttng-0.5.108, which is the same version 
> submitted to lkml - although no specific tracepoints were submitted):

OK, I grant you that's pretty scary ;-) However, it's not the only way
to do it. Most things we're using write a statically sized 64-bit event
into a relayfs buffer, with a timestamp, a minor and major event type,
and a byte of data payload.

> believe it or not, this is inlined into: kernel/sched.c ...
> 
> 'enuff said. LTT is so far from being even considerable that it's not 
> even funny.

Particularly if we're doing more complex things like that, I'd agree
that the overhead of doing the out of line jump is non-existant by
comparison. Even with the relayfs logging alone, perhaps the jump is
not that heavy ... hmmm.

If we put the NOPs in (at least as an option on some architectures)
from a macro, you don't really need the full kprobes implemented to
to tracing, even ... just overwrite the nops with a jump, so presumably
would be easier to port. However, not sure how local variable data
is specified in that case ... perhaps the kprobes guys know better.
Most of the complexity seemed to be with relocating existing code
because you didn't have nops.

To me, the main thing is to have hooks for the at least some of the
basic needs maintained in-kernel - from the dtrace paper Val pointed
me to, that seems to be exactly what they do too, and it integrates
with the newly added dynamic ones where necessary. Plus I hate the
whole awk thing, and general complexity of systemtap, but we can
probably avoid that easily enough - either the embedded C option
you mentioned, or just a different definiton for the same hook macros
under a config option.

So perhaps it'll all work. Still need a little bit of data maintained
in tree though.

M.