[ofa-general] cxgb3: MW Support

Felix Marti felix at chelsio.com
Wed Aug 13 02:28:52 PDT 2008


 

 

From: general-bounces at lists.openfabrics.org
[mailto:general-bounces at lists.openfabrics.org] On Behalf Of Philip Frey1
Sent: Wednesday, August 13, 2008 1:24 AM
To: Steve Wise
Cc: general at lists.openfabrics.org
Subject: Re: [ofa-general] cxgb3: MW Support

 


> Philip Frey1 wrote:
> >
> > Are memory windows supported by the Chelsio T3 adapter? If so, what
is 
> > the verb to bind/unbind them?
> >
> 
> MWs are supported in ofed-1.4, which is about the beta.   The support
in 
> linux/ofed is only for kernel-mode MWs. 
> 
> Are you doing user mode or kernel mode?

User mode :) 

> For kernel mode, you allocate a mw with ib_alloc_mw().  And you bind
it 
> with ib_bind_mw().  You can invalidate it by posting a IB_WR_LOCAL_INV

> work request to the send queue.  Then you can do a subsequent bind.  
> Deallocate with ib_dealloc_mw().

Ok, thanks. I will keep that in mind for later use (as soon as it is
available). 

> For user mode, I have a patch series that adds support for MWs, but I 
> hasn't been submitted yet for group review.  Its on my (very long)
todo 
> list. :)

No worries, it is not urgent. I was just wondering how expensive it is
to allocate 
a memory window as opposed to allocating a memory region - way cheaper I
guess.





[felix] Yes. Note that QP operations are considered 'fast path' (and are
fully pipelined in T3) with respect to memory/STag management they
include

-          Fast Register Non Shared MR (priv QP only)

-          Bind MW

-          Local Invalidate STag

 

You can issue e.g. a Bind MW and then immediately use the MW STag in
e.g. a following RDMA Write. Ordering guarantees that MW is ready by the
time the RDMA Write is executed. 


Thanks, 
 Philip

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20080813/53e5d37f/attachment.html>


More information about the general mailing list