On 09/23/2014 04:24 AM, Frans Klaver wrote: > On Wed, Sep 17, 2014 at 02:13:03PM +0200, Frans Klaver wrote: >> On Wed, Sep 17, 2014 at 08:01:08AM -0400, Peter Hurley wrote: >>> On 09/16/2014 04:50 AM, Frans Klaver wrote: >>>> On Mon, Sep 15, 2014 at 01:31:56PM -0400, Peter Hurley wrote: >>>>> On 09/15/2014 11:39 AM, Peter Hurley wrote: >>>>>> On 09/15/2014 10:00 AM, Frans Klaver wrote: >>>>>>> At 3.6Mbaud, with slightly over 2Mbit/s data coming in, we see 1600 uart >>>>>>> rx buffer overflows within 30 seconds. Threading the interrupt handling reduces >>>>>>> this to about 170 overflows in 10 minutes. >>>>>> >>>>>> Why is the threadirqs kernel boot option not sufficient? >>>>>> Or conversely, shouldn't this be selectable? >>>>> >>>> >>>> I wasn't aware of the threadirqs boot option. I also wouldn't know if >>>> this should be selectable. What would be a reason to favor the >>>> non-threaded irq over the threaded irq? >>> >>> Not everyone cares enough about serial to dedicate kthreads to it :) >> >> Fair enough. In that light, we might not care enough about other >> subsystems to dedicate kthreads to it :). Selectable seems reasonable in >> that case. >> >> >>>>> Also, do you see the same performance differential when you implement this >>>>> in the 8250 driver (that is, on top of Sebastian's omap->8250 conversion)? >>>>> >>>> >>>> I haven't gotten Sebastian's driver to work properly yet on the console. >>>> There was no reason for me yet to throw my omap changes on top of >>>> Sebastian's queue. > > Doing the threaded interrupt change on the 8250 driver doesn't seem as > trivial. Unless I'm mistaken, that version of this patch would mess with > all other 8250 based serial drivers, if it's done properly. Incidentally > I did try using threadirqs, but that didn't give my any significant > results. I mostly noticed a difference in the console. > > >>>> >>>>>> PS - To overflow the 64 byte RX FIFO at those data rates means interrupt >>>>>> latency in excess of 250us? >>>> >>>> At 3686400 baud it should take about 174 us to fill a 64 byte buffer. I >>>> haven't done any measurements on the interrupt latency though. If you >>>> consider that we're sending about 1kB of data, 240 times a second, we're >>>> spending a lot of time reading data from the uart. I can imagine the >>>> system has other work to do as well. >>> >>> System work should not keep irqs from being serviced. Even 174us is a long >>> time not to service an interrupt. Maybe console writes are keeping the isr >>> from running? >> >> That's quite possible. I'll have to redo the test setup I had for this to >> give you a decent answer. I'll have to do that anyway as Sebastian's >> 8250 conversion improves. > > I haven't had time yet to look into this any further. I'll accept that > this patch may fix a case most people aren't the least interested in. > I'll also happily accept that I probably need a better argumentation > than "this works better for us".Would it make sense to drop this patch > and resubmit the other three? As I mentioned in the previous run, I > think these are useful in any case. I would've thought the first 2 patches had already been picked up because they fix div-by-zero faults. I don't really have a problem with the patch (except for it should be selectable, even if that's just a CONFIG_ setting). At the same time, the performance results don't really make sense; so if there's actually an underlying problem, I'd rather that get addressed (and the long interrupt latency may be the underlying problem). As far as the 8250 driver and threaded irqs go, I just was hoping for another data point with a simple hard-coded test jig, not a full-blown patch series for all of them. Sorry for the misunderstanding. Regards, Peter Hurley -- To unsubscribe from this list: send the line "unsubscribe linux-serial" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
