On Mon, 2016-05-23 at 13:49 +0200, Hans de Goede wrote:
> Commit 198de51dbc34 ("USB: uas: Limit qdepth at the scsi-host level")
> removed the scsi_change_queue_depth() call from uas_slave_configure()
> assuming that the slave would inherit the host's queue_depth, which
> that commit sets to the same value.
>
> This is incorrect, without the scsi_change_queue_depth() call the
> slave's queue_depth defaults to 1, introducing a performance
> regression.
>
> This commit restores the call, fixing the performance regression.
>
> Cc: [email protected]
> Fixes: 198de51dbc34 ("USB: uas: Limit qdepth at the scsi-host level")
> Reported-by: Tom Yan <[email protected]>
> Signed-off-by: Hans de Goede <[email protected]>
> ---
> drivers/usb/storage/uas.c | 1 +
> 1 file changed, 1 insertion(+)
>
> diff --git a/drivers/usb/storage/uas.c b/drivers/usb/storage/uas.c
> index 16bc679..ecc7d4b 100644
> --- a/drivers/usb/storage/uas.c
> +++ b/drivers/usb/storage/uas.c
> @@ -835,6 +835,7 @@ static int uas_slave_configure(struct scsi_device
> *sdev)
> if (devinfo->flags & US_FL_BROKEN_FUA)
> sdev->broken_fua = 1;
>
> + scsi_change_queue_depth(sdev, devinfo->qdepth - 2);
Are you sure about this? For spinning rust, experiments imply that the
optimal queue depth per device is somewhere between 2 and 4. Obviously
that's not true for SSDs, so it depends on your use case. Plus, for
ATA NCQ devices (which I believe most UAS is bridged to) you have a
maximum NCQ depth of 31.
There's a good reason why you don't want a queue deeper than you can
handle: it tends to interfere with writeback because you build up a lot
of pending I/O in the queue which can't be issued (it's very similar to
why bufferbloat is a problem in networks). In theory, as long as your
devices return the correct indicator (QUEUE_FULL status), we'll handle
most of this in the mid-layer by plugging the block queue, but given
what I've seen from UAS devices, that's less than probable.
James
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
