On 2/12/2019 6:08 PM, Christoph Hellwig wrote:
On Tue, Feb 12, 2019 at 05:00:45PM +0100, Christoph Hellwig wrote:
On Thu, Feb 07, 2019 at 07:33:18PM +0200, Max Gurtovoy wrote:
From: Israel Rukshin <israelr@xxxxxxxxxxxx>
Instead of adding many arguments to RDMA mapping functions with PI,
create a new data structure 'ib_scatterlist' which contains the needed
information on a mapped scatterlist. In the future, this will also be
used by ULPs that define these attributes explicitly.
Signed-off-by: Israel Rukshin <israelr@xxxxxxxxxxxx>
Reviewed-by: Max Gurtovoy <maxg@xxxxxxxxxxxx>
I can't say I'm a fan of an argument structure in the fast path.
But I'll look how this develops later in the series.
So it seems like this can be embedded into containing strutures at least
in a few places. Can we also do it for the other callers?
Also the name of the structure seems a bit misleading. Maybe
ib_data_buf, although that seems way to generic, but I don't really
want scatterlist in it.
Maybe instead pass a struct struct sg_table instead of it, that takes
care of the orig vs mapped nents. But that leaves the offset as a
separate argument, which might not be too bad.
This is interesting idea, but it's a big change to be done for each ULP
to use sg_table (only NVMeoF use it AFAIK). I think this transition
should be done separately (and well tested !!). We can do it before or
we can do it after this patchset.
Otherwise, it will very hard to merge all these changes in one-shoot.
What is preferred ?
