Re: 2.6.24-rc1-54866f032307063776b4eff7eadb131d47f9f9b4 fails to boot: kernel BUG at include/linux/scatterlist.h:49!

[Jens Axboe <jens.axboe@oracle.com>]

On Sun, Nov 04 2007, Torsten Kaiser wrote:
> On 11/4/07, Jens Axboe <jens.axboe@oracle.com> wrote:
> > On Sat, Nov 03 2007, Jens Axboe wrote:
> > > It's probably enough. Only if you use chaining do you need to convert to
> > > using for_each_sg() and so on.
> >
> > Did a grep over ieee1394/, this seems to be all you need.
> 
> What *I* need. For eth1394.
> 
> Stefan Richter wrote:
> > (The other place where the 1394 stack deals with s/g lists is sbp2 where
> > we have to deal with what comes from the SCSI stack or the block layer.
> > But sbp2's s/g list handling is entirely unrelated to ieee1394's own
> > code for isochronous I/O and async streams, which eth1394 needs and
> > caused the bug.)
> 
> I see nothing preventing chained sg lists to be feed into
> sbp2scsi_queuecommand() which then gives these lists to
> sbp2_prep_command_orb_sg().
> 
> And there I find this:
> static void sbp2_prep_command_orb_sg(struct sbp2_command_orb *orb,
>                                      struct sbp2_fwhost_info *hi,
>                                      struct sbp2_command_info *cmd,
>                                      unsigned int scsi_use_sg,
>                                      struct scatterlist *sgpnt,
>                                      u32 orb_direction,
>                                      enum dma_data_direction dma_dir)
> {
> [snip]
> for (i = 0, sg_count = 0 ; i < count; i++, sgpnt++) {

Chained sg lists will only be feed to a scsi host controller that
enables chaining in its host template.

The fix looks fine though, it's just not a requirement or bug fix :-)

> As yesterday my md1_raid5-thread oopsed with the same bug from the
> thread "kernel NULL pointer dereference in blk_rq_map_sg with
> v2.6.23-6815-g0895e91" I'm rather suspicious of anything sg related
> right now. (At least I think its the same bug, as 2.6.23-mm1 does not
> contain the fix from that thread)

Can you post that oops please?

-- 
Jens Axboe