[linux-dvb] [RFC] video-buf redesign
Mauro Carvalho Chehab
mchehab at infradead.org
Wed Sep 26 23:54:30 CEST 2007
Em Qua, 2007-09-26 às 22:58 +0200, hermann pitton escreveu:
> Am Mittwoch, den 26.09.2007, 11:25 -0400 schrieb mkrufky at linuxtv.org:
> > Mauro Carvalho Chehab wrote:
> > >> Unfortunately, I see the same broken behavior in the videobuf tree and
> > >> the tm6000-new tree. Only the master branch v4l-dvb tree is
> > >> functioning properly.
> > >>
> > >
> > > Unfortunately, there's nothing I can do without the logs. So, _please_
> > > send the logs also.
> > >
> > > Every test I did here (and also Hermann reports) didn't produce the
> > > issues you're showing.
> > >
> > > All I get here is clean mpeg streams, with videobuf-dvb and cx88-dvb.
> > >
> > > There is, however, a known issue with IRQ and SMP machines that might be
> > > influencing. So, I need you to do also "cat /proc/cpuinfo". If the
> > > machine you're using to test has more than one CPU core, you may also
> > > need to run irqbalance. Another valuable test would be to disable the
> > > second core, and see what's happening.
> I also conducted the tests only on UP machines on the saa7134 driver.
My tests with cx88-dvb were with dualcore AMD. I noticed some logs,
indicating that irq handling is not fast enough to handle one irq for
every frame buffer. Yet, my machine were capable of not losing any
frames (this happened on both the current videobuf and the newer one).
Probably, if my machine were a little slow, or if the irq where shared
with some other devices, I would lose frames.
I'll re-run the tests disabling the second CPU and with both running,
but without irqbalance. Let's see.
> > I understand that neither you nor Hermann were able to reproduce this
> > problem, but I assure you, this is indeed a regression, and it's a
> > show-stopper for 2.6.24 -- it absolutely must be fixed prior to a merge.
> Yes, I agree then. We should be careful with this small testing base we
> have and Mike obviously noticed issues.
Fixing is important.
> Why this happens on his machine with VDR and an additional saa7146
> device, I still don't know. We have also no reports for the new
> video_buf design from saa7146 users.
It may be due to IRQ sharing. If you have two boards receiving
interrupts at the same IRQ line, you may suffer this kind of troubles,
since both drivers will be called, on a serial way, to handle that IRQ
line. This will delay videobuf handling. In this case, a small time
increment at videobuf processing may produce such troubles.
I'm writing a small patch that will reduce a little bit the videobuf
timings. Hopefully, this will fix the issue.
> Just installed the tm6000-new tree on two machines again and xawtv4
> seems to be fine on both, UP, no DVB-T currently, no HD streams.
More information about the linux-dvb