Le ven 07/11/2003 à 10:09, Jens Axboe a écrit :
> On Fri, Nov 07 2003, Nicolas Mailhot wrote:
> > Le ven 07/11/2003 à 09:24, Jens Axboe a écrit :
> > > On Wed, Nov 05 2003, Alan Stern wrote:
> > 
> > > > In any case, it quite likely _does_ point to a driver bug.  But since
> > > > sddr09_read_data() was handed this sg entry and didn't change it, if there
> > > > is such a bug it must lie in a higher-level driver.  Maybe the scsi layer, 
> > > > maybe the block layer, maybe the memory-management system, maybe the file 
> > > > system.  That was my original point.
> > > 
> > > Well, the sg entry looks perfectly valid. And that was my original
> > > point :-). And that is why I said it looks like a driver bug, not in
> > > upper layers. How much memory did the system that crashed have? If the
> > > system has highmem, try testing with scsi_calculate_bounce_limit()
> > > unconditionally returning BLK_BOUNCE_HIGH.
> > 
> > The system has 1 GiB of memory, ie just enough to make stuff like
> > radeonfb fail
> 
> Try with this debug patch then, does it work now?
> 
> ===== drivers/scsi/scsi_lib.c 1.77 vs edited =====
> --- 1.77/drivers/scsi/scsi_lib.c	Tue Oct 14 09:28:06 2003
> +++ edited/drivers/scsi/scsi_lib.c	Fri Nov  7 10:08:52 2003
> @@ -1215,6 +1215,7 @@
>  
>  u64 scsi_calculate_bounce_limit(struct Scsi_Host *shost)
>  {
> +#if 0
>  	struct device *host_dev;
>  
>  	if (shost->unchecked_isa_dma)
> @@ -1229,6 +1230,9 @@
>  	 * hardware have no practical limit.
>  	 */
>  	return BLK_BOUNCE_ANY;
> +#else
> +	return BLK_BOUNCE_HIGH;
> +#endif
>  }
>  
>  struct request_queue *scsi_alloc_queue(struct scsi_device *sdev)

Will try this evening when I have physical access to the system. (It's
difficult to plug a USB device via ssh;)

Cheers,

-- 
Nicolas Mailhot