Re: [PATCH 1/1] mmc: sdhci-bcm2835: added quirk and removed udelay in write ops

[Stephen Warren <swarren@wwwdotorg.org>]

On 10/15/2014 10:43 AM, Scott Branden wrote:
> Added quirk SDHCI_QUIRK_MULTIBLOCK_READ_ACMD12 present in controller.
> Removed udelay in write ops by using shadow registers for 16 bit
> accesses to 32-bit registers (where necessary).
> Optimized 32-bit operations when doing 8/16 register accesses.

That's 2 or 3 unrelated changes. They'd be better as separate patches,
so that any issues that arise can be bisected down to the smaller changes.

> diff --git a/drivers/mmc/host/sdhci-bcm2835.c b/drivers/mmc/host/sdhci-bcm2835.c

>  /*
>   * The Arasan has a bugette whereby it may lose the content of successive
> + * writes to the same register that are within two SD-card clock cycles of
> + * each other (a clock domain crossing problem).  Problem does not happen with
                                                    ^ The?
See right >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>> ^

> + * data.

Blank line to separate the paragraphs here, to be consistent with the
other paragraph break below?

> + * This wouldn't be a problem with the code except that we can only write the
> + * controller with 32-bit writes.  So two different 16-bit registers in the
> + * written back to back creates the problem.
>   *
> + * In reality, this only happens when a SDHCI_BLOCK_SIZE and SDHCI_BLOCK_COUNT
> + * are written followed by SDHCI_TRANSFER_MODE and SDHCI_COMMAND.

That seems like a rather risky assertion. Even if it's perfectly true
with the MMC core code right now, does the MMC core document a guarantee
that this will always be true? Even if we optimize the WAR for the issue
as you've done, I think we should still have code that validates that
the same register is never written back-to-back to detect this likely
very hard-to-debug problem.

> + * The BLOCK_SIZE and BLOCK_COUNT are meaningless until a command issued so
> + * the work around can be further optimized. We can keep shadow values of
> + * BLOCK_SIZE, BLOCK_COUNT, and TRANSFER_MODE until a COMMAND is issued.
> + * Then, write the BLOCK_SIZE+BLOCK_COUNT in a single 32-bit write followed
> + * by the TRANSFER+COMMAND in another 32-bit write.
>   */

After this patch, the entire WAR for this issue is contained within
bcm2835_sdhci_writew(). It might be a good idea to move the comment next
to that function so it's more at hand to explain the code that's there.
Or at least add a comment to that function the to mention the location
of the explanation for the complex code.

>  static inline u32 bcm2835_sdhci_readl(struct sdhci_host *host, int reg)
>  {
>  	u32 val = readl(host->ioaddr + reg);
> @@ -71,76 +57,83 @@ static inline u32 bcm2835_sdhci_readl(struct sdhci_host *host, int reg)
>  	return val;
>  }
>  
> -static void bcm2835_sdhci_writew(struct sdhci_host *host, u16 val, int reg)
> -{
... (entire function deleted)
> -}

This patch could be a lot smaller if it didn't re-order the functions at
the same time. It makes the patch harder to understand. If you must
re-order the functions, perhaps make that a separate patch that does
nothing else, so that the actual code changes are easier to see?

>  static u16 bcm2835_sdhci_readw(struct sdhci_host *host, int reg)
>  {
> -	u32 val = bcm2835_sdhci_readl(host, (reg & ~3));
> -	u32 word_num = (reg >> 1) & 1;
> -	u32 word_shift = word_num * 16;
> -	u32 word = (val >> word_shift) & 0xffff;
> -
> +	u32 val = bcm2835_sdhci_readl(host->ioaddr, (reg & ~3));

The change from host to host->ioaddr ends up passing the wrong value to
bcm2835_sdhci_readl(). This causes the kernel to crash during boot.

The compiler doesn't warn about this because host->ioaddr is void, so
can be automatically converted to struct sdhci_host *.

> +	u16 word = val >> (reg << 3 & 0x18) & 0xffff;
>  	return word;
>  }

To be honest, I think the existing code is a bit clearer, since it uses
variables with names to explain all the intermediate values. Assuming
the compiler is competent (which admittedly I haven't checked) I would
expect the same code to be generated either way, or at least something
pretty similar. Did you measure the benefit of the optimization?

> +static void bcm2835_sdhci_writew(struct sdhci_host *host, u16 val, int reg)
>  {
> +	struct sdhci_pltfm_host *pltfm_host = sdhci_priv(host);
> +	struct bcm2835_sdhci_host *bcm2835_host = pltfm_host->priv;
> +	u32 word_shift = reg << 3 & 0x18;
> +	u32 mask = 0xffff << word_shift;
> +	u32 oldval;
> +	u32 newval;
> +
> +	if (reg == SDHCI_COMMAND) {
> +		if (bcm2835_host->shadow_blk != 0) {
> +			writel(bcm2835_host->shadow_blk,
> +			       host->ioaddr + SDHCI_BLOCK_SIZE);
> +			bcm2835_host->shadow_blk = 0;
> +		}

Is it absolutely guaranteed that there's never a need to write 0 to that
register? I can see that no data transfer command is likely to transfer
0 blocks. I assume no other type of command uses that register as a
parameter?