[ofa-general] Re: InfiniBand card (mthca) in Linux

[Roland Dreier]

 > Well, using swiotlb=force (which turns on the bounce buffers) I do not get
 > oops any more. On the other hand, I got some oopses in memcpy of the bounce
 > buffers which I try to solve with Xen developpers.

So it seems there is a problem with the normal Xen PCI mapping API
then.  It would be better to avoid bounce buffers for this if
possible, because as I said that would double the memory consumption
and potentially exhaust your swiotlb space (because this hardware
context memory is not used for "in-flight" IOs, it is essentially
given to the hardware permanently).

Also, could you please CC me on any threads with the Xen developers?
It's kind of annoying to only get half of the story about what's going
on with debugging this.

Thanks,
  Roland