Hi Michal,

On 11/03/16 23:07, Michal Nazarewicz wrote:
> On Wed, Mar 09 2016, Felipe F. Tonello wrote:
>> buflen by default (256) is smaller than wMaxPacketSize (512) in high-speed
>> devices.
>>
>> That caused the OUT endpoint to freeze if the host send any data packet of
>> length greater than 256 bytes.
>>
>> This is an example dump of what happended on that enpoint:
>> HOST:   [DATA][Length=260][...]
>> DEVICE: [NAK]
>> HOST:   [PING]
>> DEVICE: [NAK]
>> HOST:   [PING]
>> DEVICE: [NAK]
>> ...
>> HOST:   [PING]
>> DEVICE: [NAK]
>>
>> This patch fixes this problem by setting the minimum usb_request's buffer size
>> for the OUT endpoint as its wMaxPacketSize.
>>
>> Signed-off-by: Felipe F. Tonello <eu@felipetonello.com>
> 
> Acked-by: Michal Nazarewicz <mina86@mina86.com>
> 
> But see comment below:
> 
>> ---
>>  drivers/usb/gadget/function/f_midi.c | 4 +++-
>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/usb/gadget/function/f_midi.c b/drivers/usb/gadget/function/f_midi.c
>> index 8475e3dc82d4..826ba641f29d 100644
>> --- a/drivers/usb/gadget/function/f_midi.c
>> +++ b/drivers/usb/gadget/function/f_midi.c
>> @@ -366,7 +366,9 @@ static int f_midi_set_alt(struct usb_function *f, unsigned intf, unsigned alt)
>>  	/* allocate a bunch of read buffers and queue them all at once. */
>>  	for (i = 0; i < midi->qlen && err == 0; i++) {
>>  		struct usb_request *req =
>> -			midi_alloc_ep_req(midi->out_ep, midi->buflen);
>> +			midi_alloc_ep_req(midi->out_ep,
>> +				max_t(unsigned, midi->buflen,
>> +					bulk_out_desc.wMaxPacketSize));
> 
> Or, just:
> 
> +			midi_alloc_ep_req(midi->out_ep,
> +					  bulk_out_desc.wMaxPacketSize);
> 
> Packet cannot be greater than wMaxPacketSize so there is no need to
> allocate more (if buflen > wMaxPacketSize).

I didn't know that was a constraint. If so, I agree with you.

> 
>>  		if (req == NULL)
>>  			return -ENOMEM;
>>  
> 
> I’m also wondering whether it would be beneficial to get rid of buflen
> all together and use wMaxPacketSie for in endpoints as well?  Is that
> feasible?

Yes, we could just remove the buflen parameter.

The only scenario where I can see buflen been "useful" is if the user
wants to have a smaller buffer size for the OUT endpoint. Should we
support this case or not?

I can rework this patch for any case we agree on.

-- 
Felipe