On Thu, Jun 06, 2013 at 11:21:54AM -0700, Greg KH wrote: > On Thu, Jun 06, 2013 at 11:58:56AM +0200, Johan Hovold wrote: > > On Thu, Jun 06, 2013 at 10:50:36AM +0200, Tomaž Šolc wrote: > > > Hi > > > > > > I have noticed that the ftdi_sio serial driver in recent kernel versions > > > has very bad performance when used through the Python's serial library. > > > > > > As a test case I have a custom device that will send a continuous block > > > of 5k characters once every few seconds over a RS-232 line (115200 baud) > > > to an Olimex programmer (based on FT2232C, also tried one with FT2232H). > > > > > > Programmer is connected to a Linux system where a simple Python script > > > reads the device: > > > > > > import serial > > > comm = serial.Serial("/dev/ttyUSB0", 115200) > > > while True: > > > line = comm.readline() > > > > > > With kernels before 3.7.0 the script reads uncorrupted data while using > > > newer kernels (including 3.9.4) the Python script sees heavy byte loss. > > > "top" shows an 95% "idle" CPU. Only very slow transmissions (on the > > > order of tens of bytes per second) will come through uncorrupted. > > > > > > Using git-bisect, I have found the commit that introduced this problem: > > > > > > 6f602912c9d0c84c2edbd446dd9f72660b701605 > > > usb: serial: ftdi_sio: Add missing chars_in_buffer function > > > > > > This might also be related with the unusual way Python serial library > > > reads the device. It uses select() with no timeout and single byte > > > read()s in a loop. strace output: > > > > > > select(4, [3], [], [], NULL) = 1 (in [3]) > > > read(3, "D", 1) = 1 > > > select(4, [3], [], [], NULL) = 1 (in [3]) > > > read(3, "E", 1) = 1 > > > ... > > > > > > With sufficiently large read()s the byte loss can be eliminated. > > > > > > With the commit above, each select() now causes an additional round trip > > > over USB to read the state of the hardware buffer. It's possible that > > > constant status querying triggers some bug in the hardware or the query > > > is simply too slow and causes overflows in the hardware buffer. > > > > You're absolutely right. This is a known issue (the select overhead) > > that was just recently fixed by commit a37025b5c7 ("USB: ftdi_sio: fix > > chars_in_buffer overhead") in v3.10-rc3. Care to give v3.10-rc4 a try? > > > > Greg, perhaps we should consider backporting the wait-until-sent > > patches (i.e. 0693196fe..4746b6c6e)? > > Yes, that's a good idea, I'll do that for the next round of stable > updates, after this next release tomorrow. I applied these, plus a few others in order to get them all to apply and build properly. But the last patch, 4746b6c6e, didn't apply, and I don't think we really need it for the 3.9 kernel, do we? thanks, greg k-h -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html