Re: [PATCH 1/1] serial: 8250, disable "too much work" messages

[Jiri Slaby <jslaby@suse.cz>]

On 04/01/2014 03:40 PM, One Thousand Gnomes wrote:
> On Tue,  1 Apr 2014 13:37:00 +0200
> Jiri Slaby <jslaby@suse.cz> wrote:
> 
>> The 8250 driver now reports many of these:
>>   serial8250: too much work for irq4
>> These messages turned out to be common these days with a use of
>> virtualization. I tried to increase the limit of processed characters
>> in commit e7328ae1848966181a7ac47e8ae6cddbd2cf55f3 (serial: 8250,
>> increase PASS_LIMIT) in 2011. It was raised from 256 to 512, but it is
>> still not enough, apparently.
> 
> A lot of emulations model the queue completely incorrectly. However
> simply hiding it with a pr_debug is the wrong answer - it wants fixing.
> 
> If we set a large PASS_LIMIT then it's not going to be a big loss on real
> hardware - we'll burp for a second or two and continue, but it ought to
> cure the virtualisation case.
> 
> If it doesn't we've got a bigger problem because it means we are jammed
> in the kernel spinning in an IRQ handler feeding data to a fake serial
> port that never stops being an IRQ and we end up hanging the virtualised
> OS for a long period
> 
> If that is happening then we need to actually workaround whatever
> crapware emulator is triggering it so we don't hang the guest for long
> periods if there is a big I/O.
> 
> If its a real port that is jammed our normal time around the loop on the
> LPC bus is going to be a shade over 24uS (32uS if TX is jammed on)
> 
> So we certainly ought to be able to go a bit higher without major
> crisis. Beyond that if it is still tripping then instead of whining we
> need to set IIR_NO_INT and set a polling timer to turn the IRQ back on
> next timer tick. That way a crappy emulated port can't hang the guest
> with a continual stream of data and a busted real one might actually sort
> out.

So, according to Takashi's measurements, we would need over 15000 loops
on a single port. Of course, this value is highly dependent on a system.
On my system, it is like 7 times lower (2100). And it lasts ~300ms here.

I suppose a limit like 32k loops is way too much and I just should go
and implement the polling. Or what about adding inter-character sleeps
to qemu to correspond to the speed? I can do that too, but I am not sure
if limiting the throughput will be accepted by them.

thanks,
-- 
js
suse labs