Re: [PATCH] mmc: sh_mmcif: unmap DMA buffers only after the transfer has stopped

[Samarth Narang]

Sorry, the sender header on this mail came from a misconfigured local
git identity. I can do a patch resend if required.

Best,
Samarth


On Fri, Oct 30, 2026 at 10:31 AM Codex <samarth.colleges@xxxxxxxxx> wrote:
>
> From: Samarth Narang <samarth.colleges@xxxxxxxxx>
>
> When sh_mmcif_end_cmd() is woken by a data error interrupt or a DMA
> timeout, the DMA descriptor may still be active. However, the function
> unmaps data->sg before clearing BUF_ACC_DMAREN/DMAWEN and calling
> dmaengine_terminate_sync(), potentially releasing the mapping while the
> controller is still accessing it.
>
> Since data errors can be triggered by the card, a faulty or malicious
> card may cause DMA accesses to continue after unmapping, potentially
> affecting reused IOVA or SWIOTLB mappings or causing premature cache
> maintenance on non-coherent systems.
>
> Move the unmap after DMA termination so that the mapping remains valid
> until the transfer has either completed normally or been terminated
> synchronously.
>
> Fixes: eae309836509 ("mmc: sh_mmcif: Terminate DMA transactions when detecting timeout or error")
> Assisted-by: LLM
> Signed-off-by: Samarth Narang <samarth.colleges@xxxxxxxxx>
> ---
> This issue was identified through static analysis of DMA lifecycle
> ordering and has not been validated on hardware.
>
> Since the request pages are not released until mmc_request_done() is
> called after DMA termination, direct-mapped pages remain owned by the
> request. The potential exposure is therefore limited to the DMA mapping
> layer, including IOVA or SWIOTLB slot reuse and premature cache
> maintenance on non-coherent systems.
>
> The fix also relies on dmaengine_terminate_sync() stopping the channel
> correctly on both MMCIF DMA providers. However, shdma clears CHCR.DE
> without confirming that the channel has stopped, while rcar-dmac polls
> the bit for approximately 1 ms but only logs a timeout if it remains
> set.
>
>  drivers/mmc/host/sh_mmcif.c |   23 ++++++++++++++---------
>  1 file changed, 14 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/mmc/host/sh_mmcif.c b/drivers/mmc/host/sh_mmcif.c
> index 8bd9cdeda..7b33f832b 100644
> --- a/drivers/mmc/host/sh_mmcif.c
> +++ b/drivers/mmc/host/sh_mmcif.c
> @@ -1173,15 +1173,6 @@ static bool sh_mmcif_end_cmd(struct sh_mmcif_host *host)
>         time = wait_for_completion_interruptible_timeout(&host->dma_complete,
>                                                          host->timeout);
>
> -       if (data->flags & MMC_DATA_READ)
> -               dma_unmap_sg(host->chan_rx->device->dev,
> -                            data->sg, data->sg_len,
> -                            DMA_FROM_DEVICE);
> -       else
> -               dma_unmap_sg(host->chan_tx->device->dev,
> -                            data->sg, data->sg_len,
> -                            DMA_TO_DEVICE);
> -
>         if (host->sd_error) {
>                 dev_err(host->mmc->parent,
>                         "Error IRQ while waiting for DMA completion!\n");
> @@ -1208,6 +1199,20 @@ static bool sh_mmcif_end_cmd(struct sh_mmcif_host *host)
>                         dmaengine_terminate_sync(host->chan_tx);
>         }
>
> +       /*
> +        * An error IRQ or a timeout wakes us up while the DMA transfer may
> +        * still be running: only unmap once it has completed or been
> +        * terminated above.
> +        */
> +       if (data->flags & MMC_DATA_READ)
> +               dma_unmap_sg(host->chan_rx->device->dev,
> +                            data->sg, data->sg_len,
> +                            DMA_FROM_DEVICE);
> +       else
> +               dma_unmap_sg(host->chan_tx->device->dev,
> +                            data->sg, data->sg_len,
> +                            DMA_TO_DEVICE);
> +
>         return false;
>  }
>
>
> base-commit: 24a19b953e2a3d0ea5a8c4951a0ab8f32720c907