Re: [avrdude-dev] [bug #38831] [PATCH] Fix avrftdi onchip-buffer overrun

[Hermann Kraus]

On Fri, 26 Apr 2013 13:09:11 +0200, Hannes Weisbach  
<address@hidden> wrote:
> this is not an issue, because the FIFO is only of importance in UART  
> mode. avrftdi uses the MPSSE mode, in which packet sizes are handled by  
> the USB spec and subsystem of the operating system, specifically libusb.

I have seen the same problem as in #38659 while trying to verify or read  
an atmega16.
I got the same
> avrftdi.c:536 avrftdi_transmit() ftdi_write_data(pdata->ftdic, buf,  
> buf_size + 4) != buf_size + 4: Connection timed out (110)
error messages. Reading an mega8 worked however. I tracked it down to the  
different "buf_size" values (512 vs. 256). It stops working as soon as  
buf_size is bigger than 380. Together with the 4 bytes protocol overhead I  
get exactly the fifo size of 384 bytes.

Everything you write about not caring about packet sizes you wrote is  
correct. But libusb only handles splitting the data into packets that  
match the USB endpoint size, but it know nothing about buffers internal to  
the chip.
I'm not sure if write-only access would work (I guess it works, as  
programming the flash does work), but writing + reading data back  
definitely doesn't work.

You send an 512 byte OUT request which produces 512 bytes of data. However  
there is no IN request to read the data. So after processing 384 bytes the  
FT2232 stall its endpoint till the results are read, while on the other  
hand the host waits for the full 512 bytes to be sent before it issues an  
IN request. This creates a dead-lock condition. The USB timeout expires  
and the transaction is aborted.

There are three solutions to this problem:
a) Issue an asynchronous IN request before the OUT request.
b) Issue an asynchronous OUT request and immediately afterwards an IN  
request.
c) Split data in packets that fit into the FIFO buffer of the FT2232.

Solution c) looked like the easiest one to implement so I chose it.

Regards,
Hermann