Re: Maximum data size in a single transfer for MS driver

["Martin K. Petersen" <martin.petersen@oracle.com>]

>>>>> "Alan" == Alan Stern <stern@rowland.harvard.edu> writes:

Alan> James or Jens, is there a reason why this routine limits
Alan> max_sectors to 1024?

Alan> And is there a reason why it sets both max_sectors and
Alan> max_hw_sectors, leaving no way to change one without the other?

The rationale being that we don't want the default value to be too large
because it makes I/O choppy on single-spindle setups.  

If the device driver supports bigger requests we set the hard limit to
whatever they support.  But filesystem I/O is capped at the default
limit which is 1024 for modern devices (and 255 for crappy ones).

You can change the current soft limit on a per-device basis in
/sys/block/<dev>/queue/max_sectors_kb.

I agree it would be nice to have a system-wide default that's not a
#define.  But given that a lot of queues are created early that's not so
easy to do.


Alan> Are drivers supposed to assign values directly to
Alan> q-> limits.max_sectors?  Shouldn't there be a cleaner API for all
Alan> this?

Drivers are supposed to set the hard limit.  The soft limit is a
fs/block layer value and none of the device driver's business.

-- 
Martin K. Petersen	Oracle Linux Engineering