[U-Boot] [PATCH] Timeout for SDHCI commands

[U-Boot] [PATCH] Timeout for SDHCI commands

[Eli Billauer]

The patch in the following mail is a result of a problem I had using an eMMC
device on a Xilinx Zynq ARM processor. It turned out that the waiting for a
certain response from the SDIO interface was made with a plain loop, with
no absolute time measurement. Since I'm using a relatively fast processor,
the timeout expired before the eMMC chip managed to acknowledge an
mmc_switch() with EXT_CSD_HS_TIMING.

This patch fixes the problem on my board, and eMMC works properly with it.
However I can't say that I really understand what I did, and the 100 ms
timeout was chosen with a finger in the wind. If there's a reason why the
timeout should be longer or shorter, this is a good time to come forward.

Thanks,
  Eli

Eli Billauer (1):
  mmc: sdhci: Fixed timeout for sdhci_send_command()

 drivers/mmc/sdhci.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

-- 
1.7.2.3

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Eli Billauer]

The current wait loop just reads the status 10000 times, which makes the
actual timeout period platform-dependent. The udelay() call within the loop
makes the new timeout ~100 ms.

Signed-off-by: Eli Billauer <eli.billauer@gmail.com>
---
 drivers/mmc/sdhci.c |    1 +
 1 files changed, 1 insertions(+), 0 deletions(-)

diff --git a/drivers/mmc/sdhci.c b/drivers/mmc/sdhci.c
index 3125d13..80f3a91 100644
--- a/drivers/mmc/sdhci.c
+++ b/drivers/mmc/sdhci.c
@@ -226,6 +226,7 @@ int sdhci_send_command(struct mmc *mmc, struct mmc_cmd *cmd,
 			break;
 		if (--retry == 0)
 			break;
+		udelay(10);
 	} while ((stat & mask) != mask);
 
 	if (retry == 0) {
-- 
1.7.2.3

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Andy Fleming]

On Thu, Jun 12, 2014 at 4:41 AM, Eli Billauer <eli.billauer@gmail.com> wrote:
> The current wait loop just reads the status 10000 times, which makes the
> actual timeout period platform-dependent. The udelay() call within the loop
> makes the new timeout ~100 ms.
>
> Signed-off-by: Eli Billauer <eli.billauer@gmail.com>
> ---
>  drivers/mmc/sdhci.c |    1 +
>  1 files changed, 1 insertions(+), 0 deletions(-)
>
> diff --git a/drivers/mmc/sdhci.c b/drivers/mmc/sdhci.c
> index 3125d13..80f3a91 100644
> --- a/drivers/mmc/sdhci.c
> +++ b/drivers/mmc/sdhci.c
> @@ -226,6 +226,7 @@ int sdhci_send_command(struct mmc *mmc, struct mmc_cmd *cmd,
>                         break;
>                 if (--retry == 0)
>                         break;
> +               udelay(10);
>         } while ((stat & mask) != mask);


Hmmm...

Is 100ms part of the spec? I like the idea of making the timeout more
time-based, but it seems to me that this changes the timeout quite
significantly. If it took N ms before, it now takes N + 100 ms.

I think, if we want the timeout to be ~100ms, we should use a udelay
of 100 or 1000, and then reduce "retry" accordingly.

Andy

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Eli Billauer]

On 19/06/14 19:43, Andy Fleming wrote:
> On Thu, Jun 12, 2014 at 4:41 AM, Eli Billauer<eli.billauer@gmail.com>  wrote:
>    
>> The current wait loop just reads the status 10000 times, which makes the
>> actual timeout period platform-dependent. The udelay() call within the loop
>> makes the new timeout ~100 ms.
>>
>> [ snipped patch ]
>>      
>
> Hmmm...
>
> Is 100ms part of the spec? I like the idea of making the timeout more
> time-based, but it seems to me that this changes the timeout quite
> significantly. If it took N ms before, it now takes N + 100 ms.
>
> I think, if we want the timeout to be ~100ms, we should use a udelay
> of 100 or 1000, and then reduce "retry" accordingly.
>    
Hi,

As I said in the mail preceding this patch, I don't know what the 
timeout should be. Maybe someone with a better knowledge on MMC could 
come forward.

Regards,
    Eli
> Andy
>
>    

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Pantelis Antoniou]

Hi Eli,

On Jun 12, 2014, at 12:41 PM, Eli Billauer wrote:

> The current wait loop just reads the status 10000 times, which makes the
> actual timeout period platform-dependent. The udelay() call within the loop
> makes the new timeout ~100 ms.
> 
> Signed-off-by: Eli Billauer <eli.billauer@gmail.com>
> ---
> drivers/mmc/sdhci.c |    1 +
> 1 files changed, 1 insertions(+), 0 deletions(-)
> 
> diff --git a/drivers/mmc/sdhci.c b/drivers/mmc/sdhci.c
> index 3125d13..80f3a91 100644
> --- a/drivers/mmc/sdhci.c
> +++ b/drivers/mmc/sdhci.c
> @@ -226,6 +226,7 @@ int sdhci_send_command(struct mmc *mmc, struct mmc_cmd *cmd,
> 			break;
> 		if (--retry == 0)
> 			break;
> +		udelay(10);
> 	} while ((stat & mask) != mask);
> 
> 	if (retry == 0) {
> -- 
> 1.7.2.3

Looking at the linux sources is no good, cause linux is interrupt driven.
This delay is used because the driver is not interrupt driven, so you have
to wait until the interrupt indication is delivered.

The only reference to interrupt latency I found is related to tuning and is
set to 50ms which I supposed is very pessimistic.
I think a timeout of 100ms would be fine.

Regards

-- Pantelis

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Andy Fleming]

On Fri, Jun 27, 2014 at 4:37 AM, Pantelis Antoniou <
pantelis.antoniou@gmail.com> wrote:

> Hi Eli,
>
> On Jun 12, 2014, at 12:41 PM, Eli Billauer wrote:
>
> > The current wait loop just reads the status 10000 times, which makes the
> > actual timeout period platform-dependent. The udelay() call within the
> loop
> > makes the new timeout ~100 ms.
> >
> > Signed-off-by: Eli Billauer <eli.billauer@gmail.com>
> > ---
> > drivers/mmc/sdhci.c |    1 +
> > 1 files changed, 1 insertions(+), 0 deletions(-)
> >
> > diff --git a/drivers/mmc/sdhci.c b/drivers/mmc/sdhci.c
> > index 3125d13..80f3a91 100644
> > --- a/drivers/mmc/sdhci.c
> > +++ b/drivers/mmc/sdhci.c
> > @@ -226,6 +226,7 @@ int sdhci_send_command(struct mmc *mmc, struct
> mmc_cmd *cmd,
> >                       break;
> >               if (--retry == 0)
> >                       break;
> > +             udelay(10);
> >       } while ((stat & mask) != mask);
> >
> >       if (retry == 0) {
> > --
> > 1.7.2.3
>
> Looking at the linux sources is no good, cause linux is interrupt driven.
> This delay is used because the driver is not interrupt driven, so you have
> to wait until the interrupt indication is delivered.
>
> The only reference to interrupt latency I found is related to tuning and is
> set to 50ms which I supposed is very pessimistic.
> I think a timeout of 100ms would be fine.
>
>
I suspect the timeout of 100ms is fine (though it's always nice when we tie
such numbers to something more concrete than: "it works if I make it wait
longer"). My main point was that this actually *adds* 100ms to the
preexisting timeout, instead of making the timeout ~100ms. If we reduced
the number of checks and increased the delay, the delay would completely
dominate the timeout loop, and total time would become closer to ~100ms.

Andy

[U-Boot] [PATCH] mmc: sdhci: Fixed timeout for sdhci_send_command()

[Steve Rae]

Tested-by: Steve Rae <srae@broadcom.com>

(does resolve the issue on our board!)

On 14-06-27 02:37 AM, Pantelis Antoniou wrote:
> Hi Eli,
>
> On Jun 12, 2014, at 12:41 PM, Eli Billauer wrote:
>
>> The current wait loop just reads the status 10000 times, which makes the
>> actual timeout period platform-dependent. The udelay() call within the loop
>> makes the new timeout ~100 ms.
>>
>> Signed-off-by: Eli Billauer <eli.billauer@gmail.com>
>> ---
>> drivers/mmc/sdhci.c |    1 +
>> 1 files changed, 1 insertions(+), 0 deletions(-)
>>
>> diff --git a/drivers/mmc/sdhci.c b/drivers/mmc/sdhci.c
>> index 3125d13..80f3a91 100644
>> --- a/drivers/mmc/sdhci.c
>> +++ b/drivers/mmc/sdhci.c
>> @@ -226,6 +226,7 @@ int sdhci_send_command(struct mmc *mmc, struct mmc_cmd *cmd,
>> 			break;
>> 		if (--retry == 0)
>> 			break;
>> +		udelay(10);
>> 	} while ((stat & mask) != mask);
>>
>> 	if (retry == 0) {
>> --
>> 1.7.2.3
>
> Looking at the linux sources is no good, cause linux is interrupt driven.
> This delay is used because the driver is not interrupt driven, so you have
> to wait until the interrupt indication is delivered.
>
> The only reference to interrupt latency I found is related to tuning and is
> set to 50ms which I supposed is very pessimistic.
> I think a timeout of 100ms would be fine.
>
> Regards
>
> -- Pantelis
>
> _______________________________________________
> U-Boot mailing list
> U-Boot at lists.denx.de
> http://lists.denx.de/mailman/listinfo/u-boot
>