[ofw] RE: Possible IPoIB DMA problems

[Anh Duong]

>In any case, I wander if you saw this problems happening on a real
machine, >or is this something that you have reached by reading the
code?

I saw this problem while I was just reading the code.

Anh

-----Original Message-----
From: Fab Tillier [mailto:ftillier at windows.microsoft.com] 
Sent: Thursday, August 21, 2008 12:11 PM
To: Tzachi Dar; Anh Duong; ofw at lists.openfabrics.org
Cc: Stan C. Smith; John Russo; Alex Estrin; Usha Srinivasan
Subject: RE: Possible IPoIB DMA problems



>From: Tzachi Dar [mailto:tzachid at mellanox.co.il]
>Sent: Thursday, August 21, 2008 9:07 AM
>To: Anh Duong; ofw at lists.openfabrics.org
>Cc: Stan C. Smith; John Russo; Fab Tillier; Alex Estrin; Usha
Srinivasan
>Subject: RE: Possible IPoIB DMA problems
>
>As far as I know all low-level drivers on open fabrics don't use DMA
>addresses. Leonid will be able to say more once he is back from
>vacation.

For SRP and 'normal' IPoIB packets the storport and NDIS drivers do the
DMA mappings, respectively.  The solution for IPoIB's special packet
handling is to allocate a common buffer for the receive buffers.  It
should be pretty straight forward.

IBAL has the same problem internally with MAD buffers.  Allocating a
common buffer for this is much more complicated due to how IBAL is
designed (it doesn't know about device objects and Windows DMA).  A
potential work around here would be to size the special QPs large enough
that the MAD payload could be sent as an inline buffer.

-Fab