public inbox for systemtap@sourceware.org
 help / color / mirror / Atom feed
* function("*") probing
@ 2006-11-20 21:12 Frank Ch. Eigler
  0 siblings, 0 replies; only message in thread
From: Frank Ch. Eigler @ 2006-11-20 21:12 UTC (permalink / raw)
  To: systemtap


Hi -

Thanks to a well-working kexec/kdump setup on fc5, fc6, and rhel5,
I've made some progress in working out the reasons for the crashes we
encounter when indiscriminately probing kernel functions.  Often, the
stack traces include multiple nested faults, sometimes bottoming out
on some random error, sometimes on a hung lock.

One problem is an old bugaboo: reentrancy.  It turns out that many of
the locking primitives we sparingly use, which ideally should be
inlined, in fact turn into function calls.  The main bunch of problems
occurs when the locking-related kernel functions (_read_lock and many
pals) are themselves probed.  Putting these into the translator
blacklist makes a big difference.  I'm working on characterizing the
callees of increasingly complex probes, and am considering
blacklisting many of them.  I was under the impression that kprobes
tries to detect & prevent such reentrancy but perhaps that too needs
some work.

Another problem is our lack of self-throttling.  It is related to bug
#2685 ("skip probes on insufficient stack") but is more like the linux
nmi-watchdog looking at /proc/interrupts and what dtrace has: a way of
detecting excessive probing load, and consequential probe skipping or
outright session shutdown.  I just created bug #3545 for this part.

- FChE

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2006-11-20 21:02 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2006-11-20 21:12 function("*") probing Frank Ch. Eigler

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).