[openib-general] [PATCH 1/2] (fixed) mthca: max_inline_data support

[Roland Dreier]

    Todd> Regarding inline data in posted wqes, this decision can be
    Todd> automatically made for user QPs and for some kernel QPs.
    Todd> Hence the API need not be exposed in user space.

    Todd> If all MRs in the Protection domain have CPU Virtual == IO
    Todd> Virtual, then the post_wqe routine can automatically inline
    Todd> the data as appropriate.  Given this situation, the IO
    Todd> Virtual address in the work request scatter gather list can
    Todd> be used as a direct CPU address to copy data into the WQE.

I don't think this sort of silent semantic change is a good idea at
all.  For example, automatically inlining data will fail in a very
subtle way if an application relies on ordering between work requests
on a send queue to post an RDMA read followed by a send of the same
buffer.  Inlining should only happen if the verbs consumer asks for it.

 - R.