Hi Matwey, On Saturday, 4 August 2018 11:00:05 EEST Matwey V. Kornilov wrote: > 2018-07-30 18:35 GMT+03:00 Laurent Pinchart: > > On Tuesday, 24 July 2018 21:56:09 EEST Matwey V. Kornilov wrote: > >> 2018-07-23 21:57 GMT+03:00 Alan Stern: > >>> On Mon, 23 Jul 2018, Matwey V. Kornilov wrote: > >>>> I've tried to strategies: > >>>> > >>>> 1) Use dma_unmap and dma_map inside the handler (I suppose this is > >>>> similar to how USB core does when there is no URB_NO_TRANSFER_DMA_MAP) > >>> > >>> Yes. > >>> > >>>> 2) Use sync_cpu and sync_device inside the handler (and dma_map only > >>>> once at memory allocation) > >>>> > >>>> It is interesting that dma_unmap/dma_map pair leads to the lower > >>>> overhead (+1us) than sync_cpu/sync_device (+2us) at x86_64 platform. > >>>> At armv7l platform using dma_unmap/dma_map leads to ~50 usec in the > >>>> handler, and sync_cpu/sync_device - ~65 usec. > >>>> > >>>> However, I am not sure is it mandatory to call > >>>> dma_sync_single_for_device for FROM_DEVICE direction? > >>> > >>> According to Documentation/DMA-API-HOWTO.txt, the CPU should not write > >>> to a DMA_FROM_DEVICE-mapped area, so dma_sync_single_for_device() is > >>> not needed. > >> > >> Well, I measured the following at armv7l. The handler execution time > >> (URB_NO_TRANSFER_DMA_MAP is used for all cases): > >> > >> 1) coherent DMA: ~3000 usec (pwc is not functional) > >> 2) explicit dma_unmap and dma_map in the handler: ~52 usec > >> 3) explicit dma_sync_single_for_cpu (no dma_sync_single_for_device): ~56 > >> usec > > > > I really don't understand why the sync option is slower. Could you please > > investigate ? Before doing anything we need to make sure we have a full > > understanding of the problem. > > Hi, > > I've found one drawback in my measurements. I forgot to fix CPU > frequency at lowest state 300MHz. Now, I remeasured > > 2) dma_unmap and dma_map in the handler: > 2A) dma_unmap_single call: 28.8 +- 1.5 usec > 2B) memcpy and the rest: 58 +- 6 usec > 2C) dma_map_single call: 22 +- 2 usec > Total: 110 +- 7 usec > > 3) dma_sync_single_for_cpu > 3A) dma_sync_single_for_cpu call: 29.4 +- 1.7 usec > 3B) memcpy and the rest: 59 +- 6 usec > 3C) noop (trace events overhead): 5 +- 2 usec > Total: 93 +- 7 usec > > So, now we see that 2A and 3A (as well as 2B and 3B) agree good within > error ranges. Thank you for the time you've spent on these measurements, the information is useful and your work very appreciated. > >> So, I suppose that unfortunately Tomasz suggestion doesn't work. There > >> is no performance improvement when dma_sync_single is used. > >> > >> At x86_64 the following happens: > >> > >> 1) coherent DMA: ~2 usec > > > > What do you mean by coherent DMA for x86_64 ? Is that usb_alloc_coherent() > > ? Could you trace it to see how memory is allocated exactly, and how it's > > mapped to the CPU ? I suspect that it will end up in dma_direct_alloc() > > but I'd like a confirmation. > > usb_alloc_coherents() ends up inside hcd_buffer_alloc() where > dma_alloc_coherent() is called. Keep in mind, that requested size is > 9560 in our case and pool is not used. > > >> 2) explicit dma_unmap and dma_map in the handler: ~3.5 usec > >> 3) explicit dma_sync_single_for_cpu (no dma_sync_single_for_device): ~4 > >> usec > >> > >> So, whats to do next? Personally, I think that DMA streaming API > >> introduces not so great overhead. > > > > It might not be very large, but with USB3 cameras at high resolutions and > > framerates, it might still become noticeable. I wouldn't degrade > > performances on x86, especially if we can decide which option to use > > based on the platform (or perhaps even better based on Kconfig options > > such as DMA_NONCOHERENT). > > PWC is discontinued chip, so there will not be any new USB3 cameras. You're right. I had in mind other USB cameras that would benefit from the same change, and in particular the uvcvideo driver, which is used by USB3 cameras. > Kconfig won't work here, as I said before, DMA config is stored inside > device tree blob on ARM architecture. But couldn't we skip it at least on x86 ? > >> Does anybody happy with turning to streaming DMA or I'll introduce > >> module-level switch as Ezequiel suggested? > > > > A module-level switch isn't a good idea, it will just confuse users. We > > need to establish a strategy and come up with a good heuristic that can > > be applied at compile and/or runtime to automatically decide how to > > allocate buffers. > > I am agree in general, but I cannot understand why webcam driver > should think about memory allocation heuristics. I fully agree with you, this should be handled by either the USB core or the media core (possibly with a few static hints from the driver, such as buffer sizes, to help with heuristics, if needed at all). -- Regards, Laurent Pinchart
