On 13/11/2019 08:53, Linus Walleij wrote:
> The Ux500 (at least) can only deal with DMA transactions
> starting and ending on an even 4-byte aligned address.
>
> The problem isn't in the DMA engine of the system as such:
> the problem is in the state machine of the MMCI block that
> has some features to handle single bytes but it seems like
> it doesn't quite work.
>
> This problem is probably caused by most of the testing
> being done on mass storage, which will be 512-bytes aligned
> blocks placed neatly in pages and practically never run into
> this situation.
>
> On SDIO (for example in WiFi adapters) this situation is
> common.
>
> By avoiding any such transfers with a special vendor flag,
> we can bail out to PIO when an odd transfer is detected
> while keeping DMA for large transfers of evenly aligned
> packages also for SDIO.
>
> Cc: Ludovic Barre <ludovic.barre@xxxxxx>
> Cc: Brian Masney <masneyb@xxxxxxxxxxxxx>
> Cc: Stephan Gerhold <stephan@xxxxxxxxxxx>
> Cc: Niklas Cassel <niklas.cassel@xxxxxxxxxx>
> Cc: Russell King <rmk+kernel@xxxxxxxxxxxxxxx>
> Signed-off-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> ---
> ChangeLog v1->v3:
> - New patch in v3 after discussion with Ulf
> ---
> drivers/mmc/host/mmci.c | 21 +++++++++++++++++++++
> drivers/mmc/host/mmci.h | 10 ++++++++++
> 2 files changed, 31 insertions(+)
>
> diff --git a/drivers/mmc/host/mmci.c b/drivers/mmc/host/mmci.c
> index 3ffcdf78a428..a08cd845dddc 100644
> --- a/drivers/mmc/host/mmci.c
> +++ b/drivers/mmc/host/mmci.c
> @@ -185,6 +185,7 @@ static struct variant_data variant_ux500 = {
> .irq_pio_mask = MCI_IRQ_PIO_MASK,
> .start_err = MCI_STARTBITERR,
> .opendrain = MCI_OD,
> + .only_long_aligned_dma = true,
> .init = mmci_variant_init,
> };
>
> @@ -219,6 +220,7 @@ static struct variant_data variant_ux500v2 = {
> .irq_pio_mask = MCI_IRQ_PIO_MASK,
> .start_err = MCI_STARTBITERR,
> .opendrain = MCI_OD,
> + .only_long_aligned_dma = true,
> .init = ux500v2_variant_init,
> };
>
> @@ -829,6 +831,25 @@ static int _mmci_dmae_prep_data(struct mmci_host *host, struct mmc_data *data,
> if (data->blksz * data->blocks <= variant->fifosize)
> return -EINVAL;
>
> + /*
> + * Handle the variants with DMA that is broken such that start and
> + * end address must be aligned on a long (32bit) boundary for the DMA
> + * to work. If this occurs, fall back to PIO.
> + */
Nit: why use 'long' as a synonym for "32 bits" ?
Why not name the field "only_32b_aligned_dma" ?
(The size of C's long int is implementation-defined; most 64-bit platforms
have a 64-bit long int.)
Perhaps the ship has already sailed -- what with readl/writel...
Regards.
