On 30/01/2014 12:07, Bart Van Assche wrote:
On 01/30/14 09:19, Or Gerlitz wrote:
Thanks for narrowing this down, I see your point, however the solution I
propose if to remove this copy altogether... for those rare cases where
fast-registration can't be done -- in SRP I think the code goes to
indirect mode under which there no copy (right?). As for iSER, we will
add to our TODO enhancing the code to avoid using bounce-buffer, few
designs might be possible, one way would be to create set of
FMRS/Fast-reg element which are capable to map chunks which are <
PAGE_SIZE, e.g in the order of single block, and hence can map what ever
arbitrary set of blocks provided by the upper layer.
I'm not sure how important this issue is on x86 systems since the page
size on these systems is 4 KB and since many applications use an I/O
size >= 4 KB. So if the I/O buffers are aligned on page boundaries the
buffer memory can be registered as a single memory region without having
to copy any data. However, on PPC systems, which have a page size of 64
KB, if e.g. a database issues a readv() or writev() call or if an I/O
scheduler coalesces several small I/O requests into a single I/O request
the data buffer of these I/O requests is discontiguous. I think it is
important that data copying can be avoided for such I/O requests.
We know that... and hence iser uses fast registration in chucks of 4KB
even when the page size is 64KB
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
