On 23-05-18 16:49:45, Frank Li wrote:
> When the address of a request was not aligned with an 8-byte boundary, the
> USB DMA was unable to process it, necessitating the use of an internal
> bounce buffer.
>
> In these cases, the request->buf had to be copied to/from this bounce
> buffer. However, if this unaligned address scenario arises, it is
> unnecessary to perform heavy cache maintenance operations like
> usb_gadget_map(unmap)_request_by_dev() on the request->buf, as the DMA
> does not utilize it at all. it can be skipped at this case.
>
> iperf3 tests on the rndis case:
>
> Transmit speed (TX): Improved from 299Mbps to 440Mbps
> Receive speed (RX): Improved from 290Mbps to 500Mbps
>
> Signed-off-by: Frank Li <Frank.Li@xxxxxxx>
> ---
> drivers/usb/cdns3/cdns3-gadget.c | 11 +++++++----
> 1 file changed, 7 insertions(+), 4 deletions(-)
>
> diff --git a/drivers/usb/cdns3/cdns3-gadget.c b/drivers/usb/cdns3/cdns3-gadget.c
> index 1dcadef933e3..09a0882a4e97 100644
> --- a/drivers/usb/cdns3/cdns3-gadget.c
> +++ b/drivers/usb/cdns3/cdns3-gadget.c
> @@ -800,7 +800,8 @@ void cdns3_gadget_giveback(struct cdns3_endpoint *priv_ep,
> if (request->status == -EINPROGRESS)
> request->status = status;
>
> - usb_gadget_unmap_request_by_dev(priv_dev->sysdev, request,
> + if (likely(!(priv_req->flags & REQUEST_UNALIGNED)))
> + usb_gadget_unmap_request_by_dev(priv_dev->sysdev, request,
> priv_ep->dir);
>
> if ((priv_req->flags & REQUEST_UNALIGNED) &&
> @@ -2543,10 +2544,12 @@ static int __cdns3_gadget_ep_queue(struct usb_ep *ep,
> if (ret < 0)
> return ret;
>
> - ret = usb_gadget_map_request_by_dev(priv_dev->sysdev, request,
> + if (likely(!(priv_req->flags & REQUEST_UNALIGNED))) {
> + ret = usb_gadget_map_request_by_dev(priv_dev->sysdev, request,
> usb_endpoint_dir_in(ep->desc));
So, the possible reason for performance drop is do cache coherency
operation twice for unaligned buffers?
Peter
> - if (ret)
> - return ret;
> + if (ret)
> + return ret;
> + }
>
> list_add_tail(&request->list, &priv_ep->deferred_req_list);
>
> --
> 2.34.1
>
--
Thanks,
Peter Chen
