On Thu, 2008-10-02 at 18:58 +0200, Jens Axboe wrote: > On Thu, Oct 02 2008, James Bottomley wrote: > > The bug would appear to be that we sometimes only look at q->max_sectors > > when deciding on mergability. Either we have to insist on max_sectors > > <= hw_max_sectors, or we have to start using min(q->max_sectors, > > q->max_hw_sectors) for this. > > q->max_sectors MUST always be <= q->max_hw_sectors, otherwise we could > be sending down requests that are too large for the device to handle. So > that condition would be a big bug. The sysfs interface checks for this, > and blk_queue_max_sectors() makes sure that is true as well. Yes, that seems always to be enforced. Perhaps there are other ways of tripping this problem ... I'm still sure if it occurs it's because we do a physical merge where a virtual merge is forbidden. > The fixes proposed still look weird. There is no phys vs hw segment > constraints, the request must adhere to the limits set by both. It's > mostly a moot point anyway, as 2.6.28 will get rid of the hw accounting > anyway. Agree all three proposed fixes look wrong. However, if it's what I think, just getting rid of hw accounting won't fix the problem because we'll still have the case where a physical merge is forbidden by iommu constraints ... this still needs to be accounted for. What we really need is a demonstration of what actually is going wrong ... James -- To unsubscribe from this list: send the line "unsubscribe linux-scsi" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html