[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    Re: deadlock tracing patch against 2.0.30
From:       "Leonard N. Zubkoff" <lnz () dandelion ! com>
Date:       1997-05-21 10:15:09
[Download RAW message or body]

  Date: Tue, 20 May 1997 13:23:26 +0200 (MET DST)
  From: Ingo Molnar <mingo@pc5829.hil.siemens.at>

  and here is a patch that does 'deadlock tracing' on 2.0 SMP. It's usage is
  much different from earlier deadlock detection patches (also it's much
  simplified). This one catches even uncatchable deadlocks ;) A short FAQ:

  How it works:

  it prints the current trace position onto the screen (it writes directly
  into textmode videomemory). Thus you have to use this in textmode ... 
  whatever X crash you can produce, you've got to do it without going into
  graphics mode. Whenever the machine locks up, the latest and greatest
  trace will be on the screen. Here comes the hard part: write down the
  trace buffer from the screen and send it us. (WARNING: we need your
  System.map too ... that one shouldnt go to mailing lists)

Very clever!  This will complement nicely the method I'm using, which is to
augment my deadlock patch to count the number of iterations waiting for the
kernel lock, and call a routine when a 32bit register loops back to 0 (or
perhaps a lower limit).  This has negligible overhead, so we can leave it in
place at all times.  If deadlock is detected, I print a message and plan to
force an OOPS on the other processor so that we get a stack backtrace.  Then
we'll know there is a deadlock problem (rather than a hardware lockup) and your
deadlock tracer can be used to get further information.

		Leonard

[prev in list] [next in list] [prev in thread] [next in thread]