On Tue, 25 Oct 2022 at 13:04, Arnd Bergmann <arnd@xxxxxxxx> wrote: > > On Tue, Oct 25, 2022, at 12:51, Ulf Hansson wrote: > > Hi Wolfram, > > > > On Thu, 6 Oct 2022 at 21:05, Wolfram Sang > > <wsa+renesas@xxxxxxxxxxxxxxxxxxxx> wrote: > >> > >> Motivation for this series from patch 5: > >> > >> === > >> So far, we have been relying on access_end interrupts only to mark DMA > >> transfers as done implying that DMA end interrupts have occurred by then > >> anyhow. On some SoCs under some conditions, this turned out to be not > >> enough. So, we enable DMA interrupts as well and make sure that both > >> events, DMA irq and access_end irq, have happened before finishing the > >> DMA transfer. > > > > The tmio/sdhi core still relies on using tasklets. I think we should > > strive to move away from tasklets for all mmc host drivers and to use > > threaded irqs instead. > > > > That said, I am worried that it might be harder to move away from > > tasklets beyond $subject series, for tmio/sdhi, but I might be wrong? > > So, I am wondering if it perhaps would be better to make that > > modernization/conversion as the first step instead? > > Moving away from tasklets is probably a good idea overall, but I'm > not sure that MMC actually needs a custom IRQ deferral mechanism > in addition to the existing BLOCK_SOFTIRQ. I would expect that block > drivers usually operate in the context of the blk_mq caller, and > adding in another thread context can add substantial latency. Well, it's not that simple as it depends on what the MMC controller supports too (and whether the MMC/SD card really conforms to the specs). We can't poll for busy completions in IRQ context, for example. And in some cases, we simply need to poll with a CMD13. Don't get me wrong, I am not promoting a custom deferral mechanism for MMC, but just a regular threaded IRQ handler (per host driver of course) when that is needed. And most mmc host drivers already have that today. For more sophisticated HWs, we have the mmc hsq interface, that host drivers can hook into, to potentially avoid some unnecessary context switchings. > > Arnd Kind regards Uffe
