On Fri, 2 Jul 2021 at 09:03, Arnd Bergmann <arnd@xxxxxxxxxx> wrote: > > On Fri, Jul 2, 2021 at 3:02 AM Shawn Lin <shawn.lin@xxxxxxxxxxxxxx> wrote: > > On 2021/6/30 20:20, Arnd Bergmann wrote: > > > From: Arnd Bergmann <arnd@xxxxxxxx> > > > > > > Jernej Skrabec reported a problem with the cw1200 driver failing on > > > arm64 systems with CONFIG_VMAP_STACK=y. > > > > > > The driver in this case passes a pointer to a stack variable (in vmalloc > > > space) into the sdio layer, which gets translated into an invalid DMA > > > address. > > > > > > Even without CONFIG_VMAP_STACK, the driver is still unreliable, as > > > cache invalidations on the DMA buffer may cause random data corruption > > > in adjacent stack slots. > > > > > > This could be worked around in the SDIO core, but in the discussion we > > > decided that passing a stack variable into SDIO should always be considered > > > a bug, as it is for USB drivers. > > > > > > Change the sdio core to produce a one-time warning for any on-stack > > > (both with and without CONFIG_VMAP_STACK) as well as any vmalloc > > > or module-local address that would have the same translation problem. > > > > This was the previous comment about the same topic. > > Should we check for mmc_io_rw_direct? > > > > https://www.spinics.net/lists/linux-mmc/msg41794.html > > Hi Shawn, > > thank you for remembering that previous discussion, that is a > good question. Looking at the code though, I don't actually > see any part of mmc_io_rw_direct() doing DMA on a caller-provided > buffer. The only thing I see in the code is a 'u8 *out' argument, but > that is just a pointer to a single byte that is set by this function. > > Do you see any other issue with that function, or does that mean > we don't have to change it? I was wrong when I earlier said that we needed to care about mmc_io_rw_direct(). mmc_io_rw_direct() transfer "data" over the CMD line. MMC host drivers can't do DMA on that. I think the $subject patch looks reasonable to me. Kind regards Uffe
