Bodo Stroesser wrote:
> Hi,
> 
> my dual Xeon machine freezes, if connection between
> FC switch and tape drives is broken while writing to tapes.
> 
> There is one SCSI target with 16 tape LUNs connected to my
> FC controller via FC switch. I can reproduce the problem by
> starting "dd if=/dev/zero of=/dev/st[0-7] bs=256K" on the
> first 8 LUNs. Then I unplug the connection between switch and
> tapes.
> 
> It doesn't matter if using LP9802 or one channel of LP9402DC.
> The problem happens immediately after cfg_nodev_tmo has
> run out. If nodev_tmo is changed, time from breaking connection
> to machine freezing changes accordingly.
> 
> After the problem happened, even NMIs no longer are handled.
> I added nmi_watchdog=1 to cmdline and added some simple code
> to nmi handler, that writes the nmi counter directly to video ram.
> In case of error, nmi no longer counts (but I have no idea, how
> this can happen, maybe there is some HW bug).

Sorry, I was wrong. The machine panics but I didn't note that,
because panic_blink() doesn't work and panic messages were not
written to the fb-console.

Meanwhile I could reproduce the problem with verbose logging in
lpfc-driver and with a modified panic(), that writes out full
log_buf to serial port. I've attached the logging starting from
lpfc insertion. I added some comments to it, to show what I was
doing.

If I understand panic info correctly, the machine crashes in
st_sleep_done[st] because of an invalid pointer STp, but the
real reason for the crash seems to be some unexpected event in
lpfc, after nodev_tmo expired.

I hope, this helps to find the problem. If other traces are
needed, please let me know.

Please CC me, I'm not on the lists.

Regards
		Bodo