[linux-dvb] High CPU load in "top" due to budget_av slot polling

hermann pitton hermann-pitton at arcor.de
Fri Apr 18 04:44:23 CEST 2008


Am Freitag, den 18.04.2008, 03:21 +0200 schrieb Oliver Endriss:
> Robert Schedel wrote:
> > Oliver Endriss wrote:
> > > Robert Schedel wrote:
> > >> Hello,
> > >>
> > >> on 2.6.24 I run into a small issue with the budget_av driver:
> > >>
> > >> Hardware: Athlon 64X2 3800+, Satelco Easywatch DVB-C (1894:002c), no CI/CAM
> > >> Software: Linux kernel 2.6.24 (gentoo-r4) SMP x86_64, budget_av module
> > >>
> > >> Error description:
> > >> After the budget_av driver module is loaded (even without any DVB 
> > >> application), the CPU load average in 'top' rises to ~1, but in top no 
> > >> active tasks are found. After unloading the driver, the load decreases 
> > >> again to ~0.
> > >>
> > >> Displaying the blocked tasks during high load with Alt-SysRq-W shows 
> > >> that the kdvb-ca kernel thread seems to be accounted as blocked when it 
> > >> polls for the CI slot status:
> > >> ---------------------------------------------------
> > >> [...]
> > >> ---------------------------------------------------
> > >>
> > >> Enabling debug traces shows that polling for the PSR in function 
> > >> 'saa7146_wait_for_debi_done_sleep' constantly times out after 250x1ms 
> > >> sleeps:
> > >>
> > >>  > saa7146: saa7146_wait_for_debi_done_sleep(): saa7146 (0): 
> > >> saa7146_wait_for_debi_done_sleep timed out while waiting for transfer 
> > >> completion
> > >>
> > >> Increasing the 250ms did not avoid the timeout. And as I understood from 
> > >> previous list mails, this timeout is normal when no CI/CAM is connected 
> > >> to the DEBI. However, for me the high frequency polling does not make 
> > >> sense if someone does not plan to plug in a CI/CAM.
> > >>
> > >> When commenting out two lines in 'dvb_ca_en50221_thread_update_delay' to 
> > >> increase the polling timer for slotstate NONE from 100ms (!) to 60s, the 
> > >> CPU load went down to 0. So this is some kind of workaround for me.
> > > 
> > > Afaics the polling interval could be increased to something like 5s or
> > > 10s if (and only if) the slot is empty. Could you provide a patch?
> > 
> > Attached a patch for 2.6.24.4. Opinions?
> 
> Basically it should work but it has to be tested with CI/CAM, too.
> Furthermore it is not sufficient to test with budget-av because many
> other drivers will be affected.
> 
> So I would prefer a patch which does not touch behaviour for other card
> drivers (if possible).
> 
> Please note for 'final' patches:
> Always run 'checkpatch.pl' and fix the errors.
> Sorry for that. :-(
> 
> > Unfortunately, a 5s poll interval for empty slots still results in a
> > load average of about 0,06 (~ 250ms/5s).
> > 
> > Increasing to 10s would decrease the load and be fine for people without
> > CAM, but increase the delay for those users inserting CAMs during
> > runtime. 5s sounds like a reasonable tradeoff.
> > 
> > Is the 250ms timeout an approved limit? Decreasing it would push the
> > load further down. Probably it still has to cover slow CAMs as well as a
> > stressed PCI bus. Unfortunately, without CAM/CI I cannot make any
> > statements myself.
> > 
> > >> Finally, my questions:
> > >> 1. Did I understand this right, that the behaviour above is expected 
> > >> when no CI is connected?
> > > 
> > > Yes, but afaics the polling interval is way too short.
> > > 
> > >> 2. Are all budget_av cards unable to check CAM slot state via interrupt 
> > >>   for HW reasons (as budget_ci does)?
> > > 
> > > I don't have any budget-av hardware, so I don't know.
> > > But I think that Andrew(?) had a good reason to implement it this way.
> > > (In contrast the budget-ci driver was always interrupt-driven.)
> > > 
> > > If someone finds out that a given card can operate in interrupt mode,
> > > it should be changed for this card. Patches are welcome. ;-)
> > 
> > My impression is, due to the different GPIO layout there is no way to
> > get a CAM change IRQ. But it seems difficult to get information about
> > the HW architecture of cards at all.
> 
> For most cards we have no hw info. :-(
> 
> > Regarding DEBI_E: Just found a av7110 code comment which also reflects
> > what my recent tests showed:
> >           /* Note: Don't try to handle the DEBI error irq (MASK_18), in
> >            * intel mode the timeout is asserted all the time...
> >            */
> > 
> > So really only DEBI_S would be left, see below.
> 
> Did you check whether DEBI_S and/or DEBI_E are ever asserted with your
> setup? If not, an interrupt would never occur anyway...
> 
> > >> 3. Would it be possible to substitute the current PSR DEBI_S polling 
> > >> with an interrupt based solution via IER/ISR? (driver av7110 alreadys 
> > >> seems to do this for its DEBI DMA)? Or was this not considered worthy, 
> > >> due to the expected short waiting period and the tricky IER handling? 
> > >> The code does not state further thoughts about this.
> > > 
> > > The av7110 driver uses interrupt mode for buffer transfers in dma mode.
> > > It does not make much sense to use interrupt mode for single-word
> > > transfers, because the single-word transfers are very fast.
> > > But I understand that the timeout causes a problem in this case.
> > 
> > OK, interrupts of course decrease performance in the "sunny day" cases
> > (=communication with inserted card in slot state READY). Having both
> > approaches (interrupt when slotstate empty, later polling) would combine
> > all benefits but also be somewhat crazy.
> > 
> > >> 4. Are the high timeout periods in debi_done (50ms/250ms) in relation to 
> > >> the 100ms poll timer intended? (I found the recent patch to this code in 
> > >> the mailing list end of last year)
> > > 
> > > That patch was applied to reduce the load on the pci bus in busy-wait
> > > mode. Basically it did not change anything for cam polling. (In fact I
> > > was not aware that the CAM was polled every 100ms. Imho this should be
> > > fixed.)
> > 
> > Only wondered whether the 250ms might have been smaller in former driver
> > versions.
> 
> Iirc it should be even worse with older drivers.
> 
> Basically the 250ms timeout is just a last resort to escape from the
> loop, if the debi transfer hangs for some reason. We might try to reduce
> the timeout but I don't know how far we can go. (Touching 'magic' values
> might be dangeous.)
> 
> CU
> Oliver
> 

Hi,

what are magic values?

Simply something retrieved under NDA with even that hardware not in
established/stable conditions yet, most likely.

If the commercial research facilities are not established for that,
because too expensive, it likely goes out to some university or major
research project around the globe, which recently collected some
external money for something related, to stay alive, and then please
could solve that minor issue on the run for free.

We had lots of high quality contributions in the past, seemingly coming
out of nothing ...

If it really counts, Open Source is a must and everybody knows it.

Cheers,
Hermann






 




More information about the linux-dvb mailing list