Re: [RFC 1/4] sdhci: allow for multiple delays when sending commands

[Luis R. Rodriguez]

On Fri, Mar 18, 2011 at 4:10 PM, Chris Ball <cjb@xxxxxxxxxx> wrote:
> Hi Luis,
>
> On Fri, Mar 18 2011, Luis R. Rodriguez wrote:
>> This is at least known to be required for the ENE 714.
>>
>> Cc: Chris Ball <cjb@xxxxxxxxxx>
>> Cc: Kalle Valo <kvalo@xxxxxxxxxx>
>> Cc: Naveen Singh <nsingh@xxxxxxxxxxx>
>> Cc: Vipin Mehta <Vipin.Mehta@xxxxxxxxxxx>
>> Signed-off-by: Luis R. Rodriguez <lrodriguez@xxxxxxxxxxx>
>
> I think this wins cjb's "I've never been so confused about what a patch
> author thought they were doing before" award.

I warned you :)

>> ---
>> Âdrivers/mmc/host/sdhci.c | Â Â7 ++++---
>> Â1 files changed, 4 insertions(+), 3 deletions(-)
>>
>> diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
>> index 9e15f41..c95dfc2 100644
>> --- a/drivers/mmc/host/sdhci.c
>> +++ b/drivers/mmc/host/sdhci.c
>> @@ -891,8 +891,8 @@ static void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>>
>> Â Â Â WARN_ON(host->cmd);
>>
>> - Â Â /* Wait max 10 ms */
>> - Â Â timeout = 10;
>> + Â Â /* Wait max this amount of ms */
>> + Â Â timeout = (10*256) + 255;
>>
>> Â Â Â mask = SDHCI_CMD_INHIBIT;
>> Â Â Â if ((cmd->data != NULL) || (cmd->flags & MMC_RSP_BUSY))
>
> Okay, so our original plan is to go through the while loop ten times,
> which equals ten mdelay(1)s == 10ms, waiting for the inhibit bit to
> become unset.
>
> After this hunk of your patch, we're set to go through the loop 2815
> times, which would make for 2.8 seconds. ÂThat seems excessive.
>
>> @@ -913,7 +913,8 @@ static void sdhci_send_command(struct sdhci_host *host, struct mmc_command *cmd)
>> Â Â Â Â Â Â Â Â Â Â Â return;
>> Â Â Â Â Â Â Â }
>> Â Â Â Â Â Â Â timeout--;
>> - Â Â Â Â Â Â mdelay(1);
>> + Â Â Â Â Â Â if (!(timeout & 0xFF))
>> + Â Â Â Â Â Â Â Â Â Â mdelay(1);
>> Â Â Â }
>>
>> Â Â Â mod_timer(&host->timer, jiffies + 10 * HZ);
>
> But wait, here you decide *not* to call mdelay(1) every time through the
> loop, and instead call it only on iterations where the bottom eight bits
> are unset. ÂThis disqualifies most of the 2815 values that timeout will
> be set to, and leaves the following values triggering the mdelay(1):
>
> 0x100 0x200 0x300 0x400 0x500 0x600 0x700 0x800 0x900 0xa00
> Â256 Â 512 Â 768 Â1024 Â1280 Â1536 Â1792 Â2048 Â2304 Â2560
>
> The astute observer will notice that there are ten such values.
> So you're calling mdelay(1) ten times. ÂBut that's what we were
> doing before! ÂThe only difference is that now we spin through
> the while loop 2815 times instead of 10, and don't perform any
> explicit delay on 2805 of them. ÂOr am I missing something?
>
> I think you should try:
>
> (a) Reverting the patch and checking that it's actually needed
> (b) Leaving the while loop body alone, but increasing the max
> Â Âtimeout until you bisect to the amount of ms that your controller
> Â Âactually takes to release the inhibit bit.
> (c) Yelling loudly in the direction that this code came from. Â:)

Heh thanks for the review, once I even get ath6kl to even load
properly on x86_64 I'll try reverting this POS hunk and see if it
still works without it. Who knows where the hell this patch came from,
as I noted I was given the entire blob as a huge patch and if you look
at it, the original patch even removes some quirk for tegra.

 Luis
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/