Description of problem: shutdown of autofs hangs; the current pid is 'pidof', stuck in 'R' state. Show Regs sez: Pid/TGid: 1181911819, comm: pidof EIP: 0060:[<c016b4f6>] CPU:0 EIP is at .text.lock.inode [kernel] 0x167 (2.4.21.20.1.2024.2.1.nptlsmp) Call trace is: proc_pid_make_inode proc_pid_lookup real_lookup link_path_walk path_lookup __user_walk sys_stat64 sys_munmap system_call Seems to happen pretty much every shutdown.
Our only changes to autofs from the base kernel are in the O(1)/nptl patch, and they involve signal handling. Have you seen this in the upstream 2.4.21 kernel by any chance?
It seems to work in taroon, which is also NPTL.
I'm still seeing this in 2.4.22-1.2087.nptlsmp on an thlon system.
Ingo, any thoughts?
I use autofs very extensively for NFS mounts, and I've not been seeing this. All my severn workstations are single CPU Athlon or mobile Athlon, though
hm, it seems to have spinlock deadlocked in some inode.c function called by proc_pid_make_inode() - possibly new_inode()? This is not something NPTL has changed or should affect directly.
what could bring us forward would be to know what happens on the other CPU which seems to cause the deadlock of this CPU.
Hmm, I just noticed that I've got an SMP machine running autofs and I can't reproduce this with recent kernels. Maybe fixed...