Em Mon, 29 Oct 2012 14:11 +0000 Simon Farnsworth <simon.farnsworth@xxxxxxxxxxxx> escreveu: > On Monday 29 October 2012 09:32:27 Andy Walls wrote: > > On Mon, 2012-10-29 at 13:02 +0000, Simon Farnsworth wrote: > > > It will affect other drivers as well; the basic cause is that modern chips > > > can enter a package deep sleep state that affects both CPU speed and latency > > > to start of DMA. On older systems, this couldn't happen - the Northbridge > > > kept running at all times, and DMA latencies were low. > > > > > > However, on the Intel Sandybridge system I'm testing with, the maximum wake > > > latency from deep sleep is 109 microseconds; the SAA7134's internal FIFO can't > > > hold onto data for that long, and overflows, resulting in the corruption I'm > > > seeing. The pm QoS request fixes this for me, by telling the PM subsystem > > > that the SAA7134 cannot cope with a long latency on the first write of a DMA > > > transfer. > > > > > > Now, some media bridges (like the ones driven by the cx18 driver) can cope > > > with very high latency before the beginning of a DMA burst. Andy Walls has > > > worked on the cx18 driver to cope in this situation, so it doesn't fail even > > > with the 109 microsecond DMA latency we have on Sandybridge. > > > > Well if brdige wake-up DMA latency is the problem, it is alos the case > > that the CX23418 has a *lot* of on board memory with which to collect > > video and compress it. (And lets face it, the CX23418 is an SoC with > > two ARM cores and a lot of dedicated external memory, as opposed to the > > SAA7134 with 1 kB of FIFO.) That hardware helps quite a bit, if the > > PCI bus is slow to wake up. > > > > I found a SAA7134 sheet for you: > > > > http://www.nxp.com/documents/data_sheet/SAA7134HL.pdf > > > > Section 6.4.3 has a short description of the behaviour when the FIFO > > overflows. > > That's a good description of what I'm seeing, so fits with the idea that it's > FIFO underflow. > > > > But this sheet (close enough): > > > > http://www.nxp.com/documents/data_sheet/SAA7133HL.pdf > > > > Has much nicer examples of the programmed levels of the FIFO in section > > 6.4.3. That 1 kB is for everything: raw VBI, Y, U, V, MPEG, and Audio. > > So you're lucky if one full line of video fits in the FIFO. > > > And that datasheet suggests that my 31 usec request is too high; in the > "fastidious" configuration, I'd need a latency of 22 usec, not 31. > > Does anyone have register-level documentation for the SAA7134, to confirm the > maximum tolerated latency with the FIFO configuration Linux uses? > > > > Others, like the SAA7134, just don't have that much buffering, and we need > > > to ask the pm core to cope with it. I suspect that most old drivers will need > > > updating if anyone wants to use them with modern systems; either they'll have > > > an architecture like the cx18 series, and the type of work Andy has done will > > > fix the problem, or they'll behave like the saa7134, and need a pm_qos request > > > to ensure that the CPU package (and thus memory controller) stay awake. > > > > Unless the chip has a lot of internal memory and processing resources, I > > suspect a pm_qos solution is needed to ensure the PCI bus responds in > > time. > > > > This is a system level issue though. Having the drivers decide what QoS > > they need in the absences of total system requirements, is the right > > thing for the home user. It might not be very friendly for a > > professional solution where someone is trying to tune the use of the > > system IO bandwidth and CPU resources. > > > > The ivtv driver and cx18 driver unconditionally bumping up their PCI > > latency timer to 64 cycles minimum always bugged me: drivers shouldn't > > be deciding QoS in a vaccuum. But, then again, user complaints went > > away and the 64 PCI cycles seemed to be a minimum QoS that everyone > > needed. Also both drivers have a ivtv/cx18_pci_latency module option to > > override the behaviour anyway. > > > So, one trick that the pm_qos request infrastructure gives us is that I only > request reduced latency when we are actually streaming. It's up to the pm core > to determine what that means - e.g. keep CPU awake, program chipset registers, > ignore the request completely. > > The other part of this is that I'm flagging my QoS requirements; if the wakeup > latency is higher than I'm requesting, I'll fail - that is clearly wrong. On > the other hand, if you have reasons to want lower wakeup latencies, the core > will combine all requests (both userspace and kernelspace), and choose the > minimum. If you're sophisticated enough to accept the problems involved in not > waking up in time to service DMA requests, you're probably also up to the task > of changing the kernel slightly to not request lower latencies. Andy/Simon, Thanks for digging into it and getting more data. Do you know if this change it also needed with USB devices that do DMA (isoc and/or bulk)? Or the USB core already handles that? Cheers, Mauro -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
