On Thu, 2015-11-26 at 23:11 +0530, Vinod Koul wrote: > On Thu, Nov 26, 2015 at 07:24:48PM +0200, Andy Shevchenko wrote: > > On Thu, 2015-11-26 at 22:31 +0530, Vinod Koul wrote: > > > On Thu, Nov 26, 2015 at 05:19:11PM +0200, Andy Shevchenko wrote: > > > > We have to call dw_dma_disable() to stop any ongoing transfer. > > > > > > Ok > > > > > > > On some platforms we can't do that since DMA device is powered > > > > off. > > > > > > Umm, you said we have ongoing transfer which means DMA should be > > > on..!! > > > > Yes, that's true for HSW/BDW and non-affected BYT/CHT. > > Can you please explain even when DMA is in use how can device be > powered > off? That does not sound right to me. It can't, but the problem is we can't distinguish that in this routine! We simple do *not* know the actual power state of DMA. These calls *ensures* that DMA is powered on. Yes, the call to dw_dma_off() when it used to be powered off sounds silly, I agree, though I see no upstreamable (non-hackish) solution for that. Previously I proposed to remove .shutdown hook completely, you were objecting. > Is this on GP DMA on BYT/CHT or > something else? Correct. Affected platforms are BYT-T and some or all of BSW/CHT depending on firmware in use. > > > Like I mentioned here is no possibility to know which platform we > > run > > on. > > > > Would you like to test this on a real device? We can provide you a > > login. > > > > > > > > > Moreover we have no > > > > possibility at that point to check if the platform is affected > > > > or > > > > not. That's > > > > why we call pm_runtime_get_sync() / pm_runtime_put() > > > > unconditionally. On the > > > > other hand we can't use pm_runtime_suspended() because runtime > > > > PM > > > > framework is > > > > not fully used by the driver. > > > > > > Shouldn't that be fixed? > > > > Do you have any solution how? > > > > Rough approach is to turn on it on channel allocation and turn off > > on > > freeing resources. The obvious downside of this solution is power > > consumption of idling device. > > But in that case, the clients should not hold ref of dma chan when > idle and > allocate only when required which is a resonable expectation There is not the case for few drivers. At least for us it's spi-pxa2xx one. It requires channels on its ->probe() stage. Jarkko is Cc'ed here, you may ask him as he is our maintainer for the SPI. -- Andy Shevchenko <andriy.shevchenko@xxxxxxxxxxxxxxx> Intel Finland Oy -- To unsubscribe from this list: send the line "unsubscribe dmaengine" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
