Re: [Bug 9734] New: I/O error when inserting a second firewire sata disk

[Stefan Richter <stefanr@s5r6.in-berlin.de>]

I wrote on 2008-01-13:
> James Bottomley wrote:
>> Firewire list cc'd
>>> Jan 12 16:50:49 x3400 kernel: firewire_sbp2: orb reply timed out, rcode=0x11
>>> Jan 12 16:50:49 x3400 kernel: sd 11:0:0:0: [sdc] Result: hostbyte=DID_BUS_BUSY
>>> driverbyte=DRIVER_OK,SUGGEST_OK
>> Best I can tell, this is the source of the problem.  The sbp2 driver is
>> replying DID_BUS_BUSY until that gets sorted out, which seems to be
>> never.
> 
> When something was plugged in or out at the same bus, fw-sbp2 has to
> reconnect == renew the login to each logical unit.  The syslog in the
> report is inconclusive whether that happened or failed.

In any case, there are frequently commands retried or newly enqueued
while fw-sbp2 waits to get the login renewed.  (And fw-sbp2 continues to
complete them with DID_BUS_BUSY until the reconnection didn't succeed.

Whoever caused that I/O, e.g. dd like in the reporter's and my own
tests, will quickly fail.

> As a side note, the old sbp2 driver does not quit commands with
> DID_BUS_BUSY between bus reset and reconnect.  Instead it blocks the
> Scsi_Host in order to not receive commands during that time at all.

I experimented with this yesterday.  First I tried
scsi_internal_device_block() because we
  - want to block logical units individually if possible,
  - need to block from within atomic context (softirq context).
However, this failed miserably with all sorts of lock inversion bug
backtraces (alleged ones or real ones, I don't know) and with occasional
kernel lock-ups (so it were probably real lock inversions).  These
locking issues cannot be solved easily because block layer and scsi_lib
play nauseating games with their locks.

So, I switched over to scsi_block_requests(), i.e. blocking the whole
host like the old sbp2 driver does.  This doesn't seem to have
scsi_internal_device_block()'s locking issues.  However, the sbp2 driver
has one Scsi_Host for each logical unit while the new fw-sbp2 driver
however has one Scsi_Host for each target.  Hence there are difficulties
with targets with multiple logical units, but I probably got them sorted
out now.

There remain frequent problems with reconnection + re-login failures
though.  These failures don't happen with exactly the same bus topology
if I don't run I/O during the bus resets.  I have an idea though what to
try next...
-- 
Stefan Richter
-=====-==--- --=- ---==
http://arcgraph.de/sr/

-------------------------------------------------------------------------
This SF.net email is sponsored by: Microsoft
Defy all challenges. Microsoft(R) Visual Studio 2008.
http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/