Hi Russell, On Monday 03 November 2014 17:04:08 Russell King - ARM Linux wrote: > On Mon, Nov 03, 2014 at 09:57:28PM +0530, Vinod Koul wrote: > > On Sat, Nov 01, 2014 at 02:29:42AM +0200, Laurent Pinchart wrote: > >> Many other drivers suffer from the same problem. While I won't reject > >> your proposed fix, I would prefer a more generic approach. > >> > >> One option that has been discussed previously was to use a work queue to > >> delay starting the DMA transfer to an interruptible context where > >> pm_runtime_get_sync() could be called. However, as Russell pointed out > >> [1], > >> even that won't work in all cases as the DMA slave might need the > >> transfer to be started before enabling part of its hardware (OMAP audio > >> seem to be such a case). > >> > >> I've heard a rumor of a possible DMA engine rework to forbid calling the > >> descriptor preparation API from atomic context. This could be used as a > >> base to implement runtime PM, as DMA slave drivers should not prepare > >> descriptors if they don't need to use them. However that's a long term > >> plan, and we need a solution sooner than that. > > > > Well it is not a rumour :) > > > > I have been contemplating that now that async_tx will be killed so we dont > > have to worry about that usage. For the slave dma usage, we can change the > > prepare API to be non atomic. I think the users will be okay with > > approach. This way drivers can use runtime pm calls in prepare. > > Except we /do/ have a fair number of places where the prep calls are made > from atomic contexts, particularly in serial drivers. You'd need to > introduce a tasklet into almost every serial driver which doesn't > already have one to restart RX DMA after an error or pause. Eg, > > drivers/tty/serial/amba-pl011.c > drivers/tty/serial/pch_uart.c > drivers/tty/serial/imx.c I wonder whether it would be possible to decouple descriptor allocation and descriptor initialization/preparation. If drivers could allocate descriptors and reuse them after they complete, not only would it lower the memory management pressure by getting rid of alloc/free during transmission, but it would also make it possible to easily allocate the transaction descriptors beforehand in non-atomic context. > Probably also: > > drivers/net/ethernet/micrel/ks8842.c > > There could well be other places as well, I've not gone through and > checked exhaustively. -- Regards, Laurent Pinchart -- 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
