Re: measuring accept queueing time (revisited)

[fche at redhat dot com (Frank Ch. Eigler)]

Hi, Eric -


normalperson (we'll just see about that :-) wrote:

> Hi all, I'm wondering if there's an updated/better version of
> the [script] posted here:
>
>   http://sourceware.org/ml/systemtap/2006-q1/msg00385.html
>
> I can probably figure it out on my own, but would like to avoid
> doing so if somebody's already done the work :)

Yeah, some changes since then:
- data structures are more easily traversed by pure
  script code -> and @cast type operators
- the kernel has useful tracepoints compiled in for more events,
  **but** this does not currently seem to include any interesting
  tcp/network/socket stuff
- the translator has gotten better placing probes
- the compiler has gotten better at generating debug data to let
  us find data 

So these days, I'd try to go for intercepting the inline functions

   probe kernel.function("sk_acceptq_added") {
      /* record $sk, timestamp */
   }
   probe kernel.function("sk_acceptq_removed") { 
      /* match $sk  */
      /* record elapsed time, execname(), pid() */
   }

or their (source-code-reading-determined) callers.

And perhaps try to raise a flag on lkml to get some tracepoints into
the tcp/ip layer pretty please.


> [...]  I'd also like to be able to get the same accept queueing time
> for Unix sockets, too.

There doesn't seem to be much queuing logic in UNIX sockets, for the
connect/accept path at least.


- FChE