Re: [PATCH v2 11/40] mmc: sdhci: Add a quirk to skip clearing the transfer mode register on tuning

[Aapo Vienamo]

On Mon, 27 Aug 2018 14:01:52 +0300
Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:

> On 10/08/18 21:08, Aapo Vienamo wrote:
> > Add SDHCI_QUIRK2_TUNE_SKIP_XFERRMODE_REG_PROG to skip programming the
> > SDHCI_TRANSFER_MODE in sdhci_set_transfer_mode() if tuning command is
> > being sent.
> > 
> > On Tegra210 and Tegra186 the tuning sequence hangs if the SDHCI
> > transfer mode register is touched.
> > 
> > Signed-off-by: Aapo Vienamo <avienamo@xxxxxxxxxx>
> > ---
> >  drivers/mmc/host/sdhci.c | 6 ++++++
> >  drivers/mmc/host/sdhci.h | 2 ++
> >  2 files changed, 8 insertions(+)
> > 
> > diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> > index a7b5602..04dc443 100644
> > --- a/drivers/mmc/host/sdhci.c
> > +++ b/drivers/mmc/host/sdhci.c
> > @@ -1028,6 +1028,12 @@ static void sdhci_set_transfer_mode(struct sdhci_host *host,
> >  
> >  	if (data == NULL) {
> >  		if (host->quirks2 &
> > +			SDHCI_QUIRK2_TUNE_SKIP_XFERMODE_REG_PROG &&
> > +				(cmd->opcode == MMC_SEND_TUNING_BLOCK ||
> > +				 cmd->opcode == MMC_SEND_TUNING_BLOCK_HS200)) {
> > +			return;
> > +		}  
> 
> Rather than introduce a quirk, can't we just dodge the unnecessary write e.g.
> 
> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> index 1b3fbd9bd5c5..68af6a67e397 100644
> --- a/drivers/mmc/host/sdhci.c
> +++ b/drivers/mmc/host/sdhci.c
> @@ -1033,10 +1033,19 @@ static void sdhci_set_transfer_mode(struct sdhci_host *host,
>  			if (cmd->opcode != MMC_SEND_TUNING_BLOCK_HS200)
>  				sdhci_writew(host, 0x0, SDHCI_TRANSFER_MODE);
>  		} else {
> -		/* clear Auto CMD settings for no data CMDs */
> +			u16 orig;
> +
> +			/* clear Auto CMD settings for no data CMDs */
>  			mode = sdhci_readw(host, SDHCI_TRANSFER_MODE);
> -			sdhci_writew(host, mode & ~(SDHCI_TRNS_AUTO_CMD12 |
> -				SDHCI_TRNS_AUTO_CMD23), SDHCI_TRANSFER_MODE);
> +			orig = mode;
> +			mode &= ~(SDHCI_TRNS_AUTO_CMD12 | SDHCI_TRNS_AUTO_CMD23);
> +			/*
> +			 * Do not write transfer mode unnecessarily because it
> +			 * can upset some host controllers (e.g. sdhci-tregra)
> +			 * during tuning.
> +			 */
> +			if (mode != orig)
> +				sdhci_writew(host, new_mode, SDHCI_TRANSFER_MODE);
>  		}
>  		return;
>  	}
> 
> 
> Alternatively 
> 
> > +		if (host->quirks2 &
> >  			SDHCI_QUIRK2_CLEAR_TRANSFERMODE_REG_BEFORE_CMD) {
> >  			sdhci_writew(host, 0x0, SDHCI_TRANSFER_MODE);
> >  		} else {
> > diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> > index 23966f8..0a99008 100644
> > --- a/drivers/mmc/host/sdhci.h
> > +++ b/drivers/mmc/host/sdhci.h
> > @@ -450,6 +450,8 @@ struct sdhci_host {
> >   * obtainable timeout.
> >   */
> >  #define SDHCI_QUIRK2_DISABLE_HW_TIMEOUT			(1<<17)
> > +/* Don't clear the SDHCI_TRANSFER_MODE register on tuning commands */
> > +#define SDHCI_QUIRK2_TUNE_SKIP_XFERMODE_REG_PROG	(1<<18)
> >  
> >  	int irq;		/* Device IRQ */
> >  	void __iomem *ioaddr;	/* Mapped address */
> >   
> 

Looks like writes to the transfer mode register are deferred until the
command register is written by tegra_sdhci_writew(). This code was
introduced due to a quirk in Tegra114/124. The Tegra114 ops struct was
probably just copied over when support for Tegra210 was added, so we
end up with this code also getting run on Tegra210.

The issue with tegra_sdhci_writew() is that it breaks successive
read-modify-write operations. The value written to the transfer mode
register isn't returned on read until the command register is also
touched. Removing the write_w accessor from tegra210_sdhci_ops seems to
make the issue originally addressed by this patch to go away.

 -Aapo