Re: A new ide warning message

[Jens Axboe]

On Fri, Aug 02 2002, Marcin Dalecki wrote:
> U?ytkownik Stephen Lord napisa?:
> >On Fri, 2002-08-02 at 06:48, Marcin Dalecki wrote:
> >
> >>Uz.ytkownik Jens Axboe napisa?:
> >>
> >>>On Fri, Aug 02 2002, Stephen Lord wrote:
> >>>
> >>>
> >>>>In 2.5.30 I started getting these warning messages out ide during
> >>>>the mount of an XFS filesystem:
> >>>>
> >>>>ide-dma: received 1 phys segments, build 2
> >>>>
> >>>>Can anyone translate that into English please.
> >>>
> >>>
> >>>Well I added that message when switching to the 2.5 style request
> >>>mapping functions, and I think the message is perfectly clear :-). Never
> >>>the less, it means that a segment that came into the ide layer with an
> >>>advertised size of 1 segment was returned from blk_rq_map_sg() as having
> >>>_two_. This can be a problem with dynamically allocated sg table (not
> >>>that ide uses those, but still).
> >>>
> >>>It's a bug and usually a critical one when this happens. I'd be inclined
> >>>to think that Adam's changes in this path are to blame for this error.
> >>
> >>Carefull carefull. it can be that the generic BIO code doesn't honour
> >>the limits Adam was setting properly. And it can be of course
> >>as well the XFS doesn't cooperate properly with those limits as well,
> >>since ther kernel appears to be patched to support them.
> >>
> >
> >
> >Well, this is happening when reading the log up from disk during
> >mount, we will be asking for somewhere around 32K of data at a
> >time, but it might not be well aligned. I will instrument it and
> >report back - will be a few hours, the box is at work and I just
> >tripped it up in some other code, I cannot reset it from here.
> >
> >
> >>It would be helpfull as well to know on which brand of host controller 
> >>chip this was found. In esp. trm290 maybe?
> >
> >
> >Since it is down I cannot give you the ide boot messages right now,
> >but it is a Tyan Tiger BX motherboard using the built in IDE chipset,
> >so pretty generic stuff.
> OK. Could you then deliberately change the following in ide/main.c
> 
> +	/* Most controllers cannot do transfers across 64kB boundaries.
> +	   trm290 can do transfers within a 4GB boundary, so it changes
> +	   this mask accordingly. */
> +	ch->seg_boundary_mask = 0xffff;
> +
> +	/* Some chipsets (cs5530, any others?) think a 64kB transfer
> +	   is 0 byte transfer, so set the limit one sector smaller.
> +	   In the future, we may default to 64kB transfers and let
> +	   invidual chipsets with this problem change ch->max_segment_size. 
> */
> +	ch->max_segment_size = (1<<16) - 512;
> 
> 
> I would in esp. like to see the result of setting  ch->max_segment_size 
> = (1 << 15).

This might not be such a good idea, since the limit-bio-size etc stuff
isn't in yet, depending on _exactly_ how big the bio's xfs are building
are. If they are max 8 pages (I seem to recall so), then yeah the above
test would be nice to see. If they are bigger than 8 pages, then the
above would be a meaningless test.

I'll hack up a rq_dump() function to slap in pcidma.c as well.

-- 
Jens Axboe