[U-Boot] [PATCH V4 10/17] usb: gadget: mv_udc: optimize bounce

[Troy Kisky <troy.kisky@boundarydevices.com>]

On 9/22/2013 5:02 PM, Marek Vasut wrote:
> Dear Troy Kisky,
>
>> On 9/20/2013 3:58 AM, Marek Vasut wrote:
>>> Dear Troy Kisky,
>>>
>>>> Only perform one copy, either in the bounce
>>>> routine for IN transfers, or the debounce
>>>> rtn for OUT transfer.
>>>>
>>>> On out transfers, only copy the number
>>>> of bytes received from the bounce buffer
>>>>
>>>> Signed-off-by: Troy Kisky <troy.kisky@boundarydevices.com>
>>>>
>>>> ---
>>>> v4: no change
>>> Just a question here. Are you sure we never Send AND Reserve the data in
>>> one turn? Because that would need two copyings.
>>    ???   s/Reserve/Receive/
>>
>> As far as I'm aware, a single buffer is only ever used to capture or
>> provide data not both.
>> But, if 2 transfers were queued, an OUT and then an IN using the same
Actually, I should have said "an IN (tx to the host controller) and then 
an OUT(rx from the host controller)"
>> buffer, if it worked before
>> this patch, it should work after as well.
> How come? Before, it was doing flush before and inval after the transfer, right
> ?

The 1st "IN" transfer (tx to the host), will [copy]/flush on mv_bounce 
and [free]/nothing on mv_debounce.
The 2nd "OUT" transfer (rx from the host) will flush on mv_bounce and 
invalidate/[copy/free] on mv_debounce.

>
> btw what does this part of the patch do/mean ? Why is it there?
>
> @@ -387,10 +383,9 @@ static void handle_ep_complete(struct mv_ep *ep)
>                         num, in ? "in" : "out", item->info, item->page0);
>   
>          len = (item->info >> 16) & 0x7fff;
> -
> -       mv_debounce(ep);
> -
>          ep->req.length -= len;
> +       mv_debounce(ep, in);
> +
>
That implements the "On out transfers, only copy the number of bytes 
received from the bounce buffer"
portion of the commit message.


Thanks
Troy