On Wed, Sep 02, 2020 at 10:31:47AM +0200, Ulf Hansson wrote: > On Tue, 1 Sep 2020 at 17:40, Christoph Hellwig <hch@xxxxxx> wrote: > > > > On Tue, Sep 01, 2020 at 05:36:17PM +0200, Ulf Hansson wrote: > > > > I still don't think this makes sense, as the dma_mask should always > > > > be non-NULL here. > > > > > > If that is the case, I wonder how the driver could even have worked without DMA. > > > > > > Because in the existing code, host->dma_dev gets assigned to > > > spi->master->dev.parent->dma_mask - which seems to turn on the DMA > > > usage in the driver. > > > > > > What am I missing? > > > > Do you know of other non-DMA users? For SH nommu it probably worked > > I don't know of other non-DMA users. As I said, I wish someone could > step in and take better care of mmc_spi - as I know it's being used a > lot. > > > because SH nommu used to provide a DMA implementation that worked > > fine for streaming maps, but was completely broken for coherent > > allocation. And this driver appears to only use the former. > > Alright, so you are saying the DMA support may potentially never have > been optional to this driver. In any case, I can remove the check in > $subject patch, as it shouldn't matter. DMA support was always optional, because even on systems where DMA is present, it doesn't necessarily mean the SPI controller uses DMA. In particular, pure bit-banged SPI via GPIOs doesn't have DMA, but has always worked. See my previous reply to Christoph about host->dma_dev for my current-best understanding of what's going on here. > Anyway, let's see what Rich thinks of this. I am curious to see if the > patch works on his SH boards - as I haven't been able to test it. I'll rebuild and retest just to confirm, but I already tested a functionally equivalent patch that just did the #ifdef inline (rather than moving the logic out to separate functions) and it worked fine. Rich
