Re: Re: i686 hang on boot in userspace

[Jens Axboe <axboe@suse.de>]

On Tue, Jul 25 2006, gmu 2k6 wrote:
> On 7/25/06, Jens Axboe <axboe@suse.de> wrote:
> >On Tue, Jul 25 2006, gmu 2k6 wrote:
> >> On 7/25/06, Jens Axboe <axboe@suse.de> wrote:
> >> >On Tue, Jul 25 2006, gmu 2k6 wrote:
> >> >> On 7/25/06, Jens Axboe <axboe@suse.de> wrote:
> >> >> >On Tue, Jul 25 2006, gmu 2k6 wrote:
> >> >> >> On 7/25/06, Jens Axboe <axboe@suse.de> wrote:
> >> >> >> >On Mon, Jul 24 2006, gmu 2k6 wrote:
> >> >> >> >> the problem I have with hangs is related to changes in CFQ and 
> >that
> >> >> >> >> CFQ is now the default. 2.6.17-git12 had the problem but booting
> >> >> >> >> it with elevator=deadline fixes the hang.
> >> >> >> >>
> >> >> >> >> symptoms encountered during git-bisecting between v2.6.17 and
> >> >> >> >v2.6.18-rc1:
> >> >> >> >> A hang while starting network services
> >> >> >> >> B hang while trying to login
> >> >> >> >>   1 on remote console [not SSH] it hang after typing <uid><CR>
> >> >> >> >>   1 via OpenSSH it hang after typing <pwd><CR> when doing slogin
> >> >> >> >root@<IP>
> >> >> >> >>
> >> >> >> >> A is the problem I got in the first place and this seems to be 
> >the
> >> >> >> >> case since 2.6.17-git11 definitely although git-bisect pointed 
> >me
> >> >at
> >> >> >> >> the following
> >> >> >> >> changeset which is included since 2.6.17-git12:
> >> >> >> >>
> >> >> >> >> caaa5f9f0a75d1dc5e812e69afdbb8720e077fd3
> >> >> >> >> by Jens Axboe
> >> >> >> >> titled "[PATCH] cfq-iosched: many performance fixes"
> >> >> >> >>
> >> >> >> >> strange enough it also hangs with 2.6.17-git11 which did not
> >> >include
> >> >> >that
> >> >> >> >> one changeset yet.
> >> >> >> >
> >> >> >> >So perhaps your bisect isn't 100% trust worthy? Can you do a 
> >manual
> >> >> >> >-gitX bisect to see which 2.6.17-gitX introduced the problem?
> >> >> >> >
> >> >> >> >Also please put a serial console or similar on the machine, so you
> >> >can
> >> >> >> >log + store the sysrq+t output.
> >> >> >>
> >> >> >> well I didn't say that caa....fd3 is the exact change which broke 
> >it,
> >> >> >> just that it's related to 1) CFQ changes and 2) CFQ being the 
> >default
> >> >> >> now.
> >> >> >> I have a Remote Serial Console via HP's integrated Lights-Out Java
> >> >> >> Applet but am not sure how to enable serial console via kernel boot
> >> >> >> params (will try to find out).
> >> >> >> I will first try to find the 2.6.17-git* revision working before
> >> >> >> bisecting it against -git11 or git12.
> >> >> >
> >> >> >Thanks, would be much appreciated to try and narrow it down to a
> >> >> >specific fix.
> >> >> >
> >> >> >Are you seeing the hang on cciss?
> >> >>
> >> >> I'm not sure it is in the cciss driver, but the SmartArray is driven 
> >by
> >> >> cciss.
> >> >> starting git<11 boot tests in a minute now.
> >> >
> >> >Ok, thanks for confirming it's cciss. The bug is likely an interaction
> >> >between cciss and cfq I think, so it would be very useful if you can pin
> >> >point which of the cfq patches make it stall.
> >>
> >> is there anything special about cciss or did you just deduce that it
> >> must be cciss in that particular box and are suspecting interaction
> >> problems with that driver and your CFQ changes?
> >
> >Nothing really special about cciss, but a few months ago I had a similar
> >discussion about cciss and a strange hang.
> >
> >If possible, please also try a known bad kernel and apply the below
> >patch and see if it still reproduces:
> >
> >diff --git a/drivers/block/cciss.c b/drivers/block/cciss.c
> >index 1c4df22..2b36e7a 100644
> >--- a/drivers/block/cciss.c
> >+++ b/drivers/block/cciss.c
> >@@ -2362,7 +2362,11 @@ static inline void complete_command(ctlr
> >        cmd->rq->completion_data = cmd;
> >        cmd->rq->errors = status;
> >        blk_add_trace_rq(cmd->rq->q, cmd->rq, BLK_TA_COMPLETE);
> >+#if 1
> >+       cciss_softirq_done(cmd->rq);
> >+#else
> >        blk_complete_request(cmd->rq);
> >+#endif
> > }
> >
> > /*
> 
> manually nailed it down to 2.6.17-git7 being the first broken revision.
> going to try whether Linus' git tree knows the -git revisions and do a 
> bisect
> otherwise interdiff and looking for CFQ or cciss changes as best I can.

Hmm, there are no cfq/cciss changes between git6 and git7. Some SCSI
changes, though. Are you using SCSI for anything?

We really need that sysrq-t dump.

-- 
Jens Axboe