i2c algo bit timeout question

i2c algo bit timeout question

[Dave Airlie]

Hi to any i2c people,

So I've been debugging some EDID fetching failures and wanted to ask
about the use of time_after_eq in the i2c bit banging code.

EDID specification recommends 2ms timeout for the ack on the initial
read, so we set the timeout in our code to usecs_to_jiffies(2200) (10%
margin of error). On my systems this ends up as 1, and we seem to fail
to retrieve EDID one in 10-20 times. Changing the value to 2, always
gets me the EDID I want.

So looking at drivers/i2c/algos/i2c-algo-bit.c it appears it uses
time_after_eq on jiffies, start + timeout value. So if we have a 10ms
jiffie resolution and enter this at the 9ms point in the 10ms window,
we will seem to exit the loop after 1ms instead of the minimum which I
asked for which is 2.2ms. Should this code use time_after instead of
time_after_eq?

or am I missing something else.
Dave.

Re: i2c algo bit timeout question

[Jean Delvare]

Hi Dave,

On Thu, 23 Apr 2009 09:19:21 +1000, Dave Airlie wrote:
> Hi to any i2c people,

i2c people tend to live on the linux-i2c list, Cc'd.

> So I've been debugging some EDID fetching failures and wanted to ask
> about the use of time_after_eq in the i2c bit banging code.
> 
> EDID specification recommends 2ms timeout for the ack on the initial
> read, so we set the timeout in our code to usecs_to_jiffies(2200) (10%
> margin of error). On my systems this ends up as 1, and we seem to fail
> to retrieve EDID one in 10-20 times. Changing the value to 2, always
> gets me the EDID I want.
> 
> So looking at drivers/i2c/algos/i2c-algo-bit.c it appears it uses
> time_after_eq on jiffies, start + timeout value. So if we have a 10ms
> jiffie resolution and enter this at the 9ms point in the 10ms window,
> we will seem to exit the loop after 1ms instead of the minimum which I
> asked for which is 2.2ms. Should this code use time_after instead of
> time_after_eq?

Yes, I think it should. This bug has been there pretty much since
forever. I suppose people didn't notice because they usually use a
large timeout value.

Please send a patch fixing this and I'll apply it.

-- 
Jean Delvare

Re: i2c algo bit timeout question

[Jean Delvare]

Hi Dave,

On Thu, 23 Apr 2009 09:19:21 +1000, Dave Airlie wrote:
> Hi to any i2c people,

i2c people tend to live on the linux-i2c list, Cc'd.

> So I've been debugging some EDID fetching failures and wanted to ask
> about the use of time_after_eq in the i2c bit banging code.
> 
> EDID specification recommends 2ms timeout for the ack on the initial
> read, so we set the timeout in our code to usecs_to_jiffies(2200) (10%
> margin of error). On my systems this ends up as 1, and we seem to fail
> to retrieve EDID one in 10-20 times. Changing the value to 2, always
> gets me the EDID I want.
> 
> So looking at drivers/i2c/algos/i2c-algo-bit.c it appears it uses
> time_after_eq on jiffies, start + timeout value. So if we have a 10ms
> jiffie resolution and enter this at the 9ms point in the 10ms window,
> we will seem to exit the loop after 1ms instead of the minimum which I
> asked for which is 2.2ms. Should this code use time_after instead of
> time_after_eq?

Yes, I think it should. This bug has been there pretty much since
forever. I suppose people didn't notice because they usually use a
large timeout value.

Please send a patch fixing this and I'll apply it.

-- 
Jean Delvare