[openib-general] Re: [PATCH 0 of 20] [RFC] ipath - PathScale InfiniPath driver

Roland Dreier rdreier at cisco.com
Wed Jan 4 13:26:55 PST 2006


    Eric> Given Linus's comments and looking at where you are getting
    Eric> stuck I would recommend you split out support for the
    Eric> nonstandard ipath protocol from the rest of the driver.  If
    Eric> the standard infiniband interfaces for kernel bypass are not
    Eric> sufficient for flinging packets then we need to re-examine
    Eric> them.

Yes, this might be a good idea.  The "core" driver looks like it is
suffering from really being several things stuck together.  It would
probably make things a lot cleaner and easier to maintain if the core
driver just handled synchronizing access to the low-level hardware,
with other stuff split into its own driver.  It seems there might even
be enough stuff to split "core" into three drivers: the real core, the
ultra-high-performance MPI transport, and the management/diagnostitcs
stuff.

Also, there are APIs in the "core" driver that are only exported for a
single user outside the driver -- it would probably make sense to move
that logic directly to where it's used.  I'm thinking of things like
ipath_verbs_send() and the whole ipath_copy.c file.

 - R.



More information about the general mailing list