On Thu, May 31, 2012 at 03:09:26PM +0530, Rajaram R wrote: > On Thu, May 31, 2012 at 1:41 PM, Felipe Balbi <balbi@xxxxxx> wrote: > > Hi, > > > > (please don't top-post, ever) > > hmm i missed that... thanks for pointing ;-) > > > > > On Thu, May 31, 2012 at 12:06:21PM +0530, Rajaram R wrote: > >> Hi Felipe/Paul > >> > >> Any reason for removing short_not_ok ? I believe it still holds true > >> even after this patch. I will push a patch for the same if there is no > >> objection. > > > > It doesn't hold true at all. We _want_ a short packet in order to end > > the transfer. Let me illustrate: > > > > We want to receive 31-byte CBw, but our UDC can't handle short packets > > on OUT direction, so we start a transfer for wMaxPacketSize. If we said > > short_not_ok, the driver wouldn't like when we received 31-byte and > > ended the transfer. > > IMO any tweak specific to a controller should be in the platform > files.. In case of musb also the controller has some limitation and > short packets are handled appropriately within the musb driver. I have explained why it wouldn't be a good choice to do that for dwc3. USB3 transfers are quite high bandwidth (5Gbps), if we needed to use a bounce buffer for every non-aligned transfer, we would have to do quite a lot of memcpy() and that would degrade performance quite a lot and cause quite some load on the CPU. That was the reason why we chose to change g_mass_storage instead :-s > In case of DWC, if there is a limitation you can always ignore this > field. This field is being used by musb driver for managing DMA > transfers. Removing it will affect musb drivers. Please clarify yeah, musb is quite buggy :-p patches fixing that are very welcome actually. To be fair, the whole EP handling in MUSB is quite messed up and should change. Together with DMA setup, endpoint initialization, FIFO allocation (and re-allocation if necessary after SetInterface) and so on... Just look at rxstate() and txstate(), those two functions are too big and they know too much about the underlying DMA engine. I'm sure many many bugs are hidden on those two functions. If you would like to help out, just splitting txstate() and rxstate() into dma and non-dma functions would already help you see issues. After that, go back to the MUSB programming guide and unify DMA programming (it can be done) so that all those ifdefs are dropped and checks for which DMA engine are dropped. -- balbi
Attachment:
signature.asc
Description: Digital signature
