On 10/28/2015 07:04 AM, Oliver Neukum wrote: > On Tue, 2015-10-27 at 16:45 +0100, Sven Brauch wrote: >> Hey Oliver, >> >> On 27/10/15 16:07, Oliver Neukum wrote: >>> the theory that the lack of support for put_char() is a >>> major contributor to character loss in cdc-acm can be tested. >>> Sven, could you test the attached patch? It implements >>> the support. >> Thanks a lot for caring about this. I applied the patch to a plain linux >> 4.2 tree, but unfortunately I couldn't see any improvements. Applying > > That is unfortunate. So your problem is with not enough buffers. Well, not necessarily. As I wrote in the original thread, I changed the way input work is scheduled which should impact Sven's use-case. Those patches are on Greg's tty-next branch; afaik Sven has not tested with those. Sven, please test Oliver's patch on that tree. Also, please attach your target .config here. Lastly, please confirm your test method/termios settings (iow, are you using a reproducer or just 'cat big_file > /dev/ttyACM1') >> the patch I originally sent to the list for comparison still fixed the >> issue. > > Peter, what to do about this? 1. Re-evaluate the input kworker latency with my patches in tty-next. If the max input kworker latency can be cut in 1/2, then Sven's use case will be comfortably in only 1/2 the tty buffer space. I have a quick howto doc in progress for measuring that scheduling lag with ftrace, and a simple awk script for processing that ftrace output to aid in determining why. This is the fundamental problem. 2. Fix unthrottle Ever since my work on firewire tty, I've known throttle/unthrottle needed rework, but the i/o path and locking problems were more pressing. This is near top of my TODO list. > I think there are issues with Sven's > original patch. In particular I can't see how it can guarantee > progress. Is there a way to just really increase tty buffers, > e.g. quadruple them? Yes, this can be done per tty port (typically by the port driver install() method). Code fragment below doubles the tty buffer ceiling: retval = tty_buffer_set_limit(port, 128 * 1024); if (retval) goto error....; > Sven, could we run a few more tests before we look at your patch > again? It really has issues, but it does help and nothing else > does. If necessary I will merge an improved version of your patch. I would much rather rework URB flow + unthrottle, as I previously outlined in the original thread instead of introducing another buffering layer. Regards, Peter Hurley -- 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
