Re: [PATCH 00/17] ublk: avoid accessing ublk_queue to handle ublksrv_io_cmd

[Ming Lei <ming.lei@redhat.com>]

On Wed, Sep 17, 2025 at 07:49:36PM -0600, Caleb Sander Mateos wrote:
> For ublk servers with many ublk queues, accessing the ublk_queue in
> ublk_ch_uring_cmd_local() and the functions it calls is a frequent cache miss.
> The ublk_queue is only accessed for its q_depth and flags, which are also
> available on ublk_device. And ublk_device is already accessed for nr_hw_queues,
> so it will already be cached. Unfortunately, the UBLK_IO_NEED_GET_DATA path
> still needs to access the ublk_queue for io_cmd_buf, so it's not possible to
> avoid accessing the ublk_queue there. (Allocating a single io_cmd_buf for all of
> a ublk_device's I/Os could be done in the future.) At least we can optimize
> UBLK_IO_FETCH_REQ, UBLK_IO_COMMIT_AND_FETCH_REQ, UBLK_IO_REGISTER_IO_BUF, and
> UBLK_IO_UNREGISTER_IO_BUF.
> Using only the ublk_device and not the ublk_queue in ublk_dispatch_req() is also
> possible, but left for a future change.

The idea looks good: avoid to read ublk_queue since querying ublk_device is 
inevitable & enough.

For the series,

Reviewed-by: Ming Lei <ming.lei@redhat.com>

BTW, 'const struct ublk_device *' can be passed for several helpers, and it
can be one follow-up.


Thanks,
Ming