On Sun, 18 May 2014, Matthias Pitzl wrote:
> On 14.05.2014 - 15:48:27, Philip Guenther wrote:
> > On Wed, 14 May 2014, Matthias Pitzl wrote:
> > > On our server running with OpenBSD 5.5 calling netstat sometimes lead to
> > > an uvm_fault in sysctl_file(). I can trigger the problem by running some
> > > of our tests which create/delete a lot of files, open/close sockets etc.
> > > and call netstat -anv.
> > >
> > > Attached file contains dmesg, trace and ps output.
> > ...
> > > ddb{3}> trace
> > > sysctl_file() at sysctl_file+0x180
> > > kern_sysctl() at kern_sysctl+0x686
> >
> > You don't happen to have the output of "show reg", do you?
>
> This time with show registers/trace output per CPU.
> Hope this will help to nail down the bug.
Thanks. That shows the "0xdeafbead" value in %rbx that I was expecting,
confirming that it's the "insufficient locking" issue I described. Now I
just need to finish reducing the complexity of exit1() so that we can
bring back sleeping locks to protect allprocess and filehead...
Philip Guenther
