On Tue, Sep 27, 2022 at 7:14 AM Jens Axboe [off-list ref] wrote:

Hi,

The passthrough IO path currently doesn't do any request allocation
batching like we do for normal IO. Wire this up through the usual
blk_mq_alloc_request() allocation helper.

Similarly, we don't currently supported batched completions for
passthrough IO. Allow the request->end_io() handler to return back
whether or not it retains ownership of the request. By default all
handlers are converted to returning RQ_END_IO_NONE, which retains
the existing behavior. But with that in place, we can tweak the
nvme uring_cmd end_io handler to pass back ownership, and hence enable
completion batching for passthrough requests as well.

This is good for a 10% improvement for passthrough performance. For
a non-drive limited test case, passthrough IO is now more efficient
than the regular bdev O_DIRECT path.

Changes since v1:
- Remove spurious semicolon
- Cleanup struct nvme_uring_cmd_pdu handling

--
Jens Axboe

I see an improvement of ~12% (2.34 to 2.63 MIOPS) with polling enabled and
an improvement of ~4% (1.84 to 1.92 MIOPS) with polling disabled using the
t/io_uring utility (in fio) in my setup with this patch series!

--
Anuj Gupta