--=-=-=
Content-Transfer-Encoding: quoted-printable


Berndt Josef Wulf <wulf@ping.net.au> writes:

> On Friday 21 July 2006 05:09, Greg Troxel wrote:
>> I can see two reasons not to have these changes compiled in all the
>> time:
>>
>> It's easier to convince oneself that committing the code won't cause
>> trouble with the ifdef, which is why I suggested that route to Joanne.
>>
>> This adds a bit of code to GENERIC.  In general it seems good to be
>> able to exclude functionality one doesn't need.
>>
>> Right now it's only the USRP I know of that can use the changes.  But
>> really anything that sends a lot of data over bulk endpoints to user
>> space would benefit.
>
> Don't you agree that options that make a kernel device/interface to
> comply more closely to the specifications should be enabled by
> default? The intended changes are not to extend but to add missing
> code making this device more compliant.

There are two issues: functionality/semantics and performance.  The
change increases functionality but changes semantics.  Thus, it's not
appropriate to enable RA_WB all the time.

The basic issue is the relationship of system calls and USB
transactions.  Now, with ugen(4), a read system call causes a USB read
transaction.  For some devices, that might be important, where the
read transcation initiates a measurement, or something else.

Write transaction buffering is less problematic, but write-behind
breaks the current notion that the transaction has finished (at some
layer - I'm not sure all the way down).

It's perhaps true for the USRP and USRP-like devices that one would
always want RA_WB, but it's also quite trivial for the USRP library to
enable this feature.

=2D-=20
    Greg Troxel <gdt@ir.bbn.com>

--=-=-=
Content-Type: application/pgp-signature

-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.4 (NetBSD)

iD8DBQFEwRG3+vesoDJhHiURAiqaAJ9ZnyDQLzIxyWWKKb7vl7VHpaObKACfda+h
/U7PXGfdnWsiCAmRpXbmK0k=
=wYNy
-----END PGP SIGNATURE-----
--=-=-=--