> > in __write_lock_failed; there, EDI is c02d1f4c, which lies in irq_stat. > > EAX is more useful, it points to lock. Hmm, EAX is 0 for all register dumps I see. > This problem look different of described earlier. Now it is plain > deadlock, which is very easy to find. ... > Try to apply the patch appended. It is BH watchdog, it does the > same thing as nmi watchdog, but catches softirq deadlocks. Thanks, this gave me some nice oopses. See the attached archive, which contains undecoded and decoded dumps of three independant trials. In each case, I had two server processes serving the usual individually-small HTTP connections, and did "wc -l /proc/net/tcp" during the run (before or after the run, I can do that without problems, and without reading /proc/net/tcp, the test runs with lots more connections, also without problems). The second oops (*.2.procnettcp) is pretty garbled, but may also contain useful information - there I see tcp_get_info on the call stack. regards Patrick
Attachment:
oops.tar.gz
Description: application/tar-gz
