Re: [PATCH] driver: adc: ltc2497: return directly after reading the adc conversion value

[Uwe Kleine-König]

On Fri, Jun 04, 2021 at 09:20:43AM +0100, Jonathan Cameron wrote:
> On Fri, 4 Jun 2021 06:43:20 +0000
> "Li, Meng" <Meng.Li@xxxxxxxxxxxxx> wrote:
> 
> > > -----Original Message-----
> > > From: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx>
> > > Sent: Friday, June 4, 2021 2:13 PM
> > > To: Li, Meng <Meng.Li@xxxxxxxxxxxxx>
> > > Cc: Jonathan Cameron <jic23@xxxxxxxxxx>; lars@xxxxxxxxxx;
> > > Michael.Hennerich@xxxxxxxxxx; pmeerw@xxxxxxxxxx; linux-
> > > kernel@xxxxxxxxxxxxxxx; linux-iio@xxxxxxxxxxxxxxx
> > > Subject: Re: [PATCH] driver: adc: ltc2497: return directly after reading the adc
> > > conversion value
> > > 
> > > Hello,
> > > 
> > > On Fri, Jun 04, 2021 at 02:16:39AM +0000, Li, Meng wrote:  
> > > >
> > > >  
> > > > > -----Original Message-----
> > > > > From: Jonathan Cameron <jic23@xxxxxxxxxx>
> > > > > Sent: Friday, June 4, 2021 12:20 AM
> > > > > To: Li, Meng <Meng.Li@xxxxxxxxxxxxx>
> > > > > Cc: lars@xxxxxxxxxx; Michael.Hennerich@xxxxxxxxxx;
> > > > > pmeerw@xxxxxxxxxx; u.kleine-koenig@xxxxxxxxxxxxxx; linux-
> > > > > kernel@xxxxxxxxxxxxxxx; linux-iio@xxxxxxxxxxxxxxx
> > > > > Subject: Re: [PATCH] driver: adc: ltc2497: return directly after
> > > > > reading the adc conversion value
> > > > >
> > > > > [Please note: This e-mail is from an EXTERNAL e-mail address]
> > > > >
> > > > > On Tue,  1 Jun 2021 17:28:05 +0800
> > > > > Meng.Li@xxxxxxxxxxxxx wrote:
> > > > >  
> > > > > > From: Meng Li <Meng.Li@xxxxxxxxxxxxx>
> > > > > >
> > > > > > When read adc conversion value with below command:
> > > > > > cat /sys/.../iio:device0/in_voltage0-voltage1_raw
> > > > > > There is an error reported as below:
> > > > > > ltc2497 0-0014: i2c transfer failed: -EREMOTEIO This i2c transfer
> > > > > > issue is introduced by commit 69548b7c2c4f ("iio:
> > > > > > adc: ltc2497: split protocol independent part in a separate module").
> > > > > > When extract the common code into ltc2497-core.c, it change the
> > > > > > code logic of function ltc2497core_read(). With wrong reading
> > > > > > sequence, the action of enable adc channel is sent to chip again
> > > > > > during adc channel is in conversion status. In this way, there is
> > > > > > no ack from chip, and then cause i2c transfer failed.
> > > > > > In order to keep the code logic is the same with original ideal,
> > > > > > it is need to return direct after reading the adc conversion value.  
> > > 
> > > As background about the choice of the .result_and_measure callback:
> > > A difference between the ltc2497 (i2c) and ltc2496 (spi) is that for the latter
> > > reading the result of the last conversion and starting a new is a single bus
> > > operation and the one cannot be done without the other.
> > >   
> > > > > > v2:
> > > > > > According to ltc2497 datasheet, the max value of conversion time
> > > > > > is
> > > > > > 149.9 ms. So, add 20% to the 150msecs so that there is enough time
> > > > > > for data conversion.  
> > > > >
> > > > > Version change logs go below the --- as we don't want to preserve
> > > > > them forever in the git history.
> > > > >
> > > > > I may have lost track of the discussion, but I thought the idea was
> > > > > that perhaps the longer time period would remove the need for the early  
> > > return?  
> > > > >  
> > > >
> > > > No!
> > > > I think the ret is essential.  
> > > 
> > > I'd like to understand why. Currently ltc2497core_read() looks as follows
> > > (simplified by dropping error handling, and unrolling the result_and_measure
> > > callback for the i2c case):
> > > 
> > > 	ltc2497core_wait_conv()
> > > 
> > > 	// result_and_measure(address, NULL)
> > > 	i2c_smbus_write_byte(client, LTC2497_ENABLE | address);
> > > 
> > > 	msleep_interruptible(LTC2497_CONVERSION_TIME_MS)
> > > 
> > > 	// result_and_measure(address, val);
> > > 	i2c_master_recv(client, &buf, 3);
> > > 	i2c_smbus_write_byte(client, LTC2497_ENABLE | address);
> > > 
> > > 
> > > With the early return you suggest to introduce with your patch you save the
> > > last i2c_smbus_write_byte call. The data sheet indeed claims to start a new
> > > conversion at the stop condition.
> > > 
> > > So either the reading of the conversion result and programming of the
> > > (maybe) new address has to be done in a single i2c transfer, or we have to do
> > > something like your patch.
> > > 
> > > What I don't like about your approach is that it changes the semantic of the
> > > callback to result_*or*_measure which is something the spi variant cannot
> > > implement. With the current use of the function this is fine, however if at
> > > some time in the future we implement a bulk conversion shortcut this hurts.
> > > 
> > > So I suggest to do:
> > >   
> > > ---->8----  
> > > From: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx>
> > > Date: Fri, 4 Jun 2021 08:02:44 +0200
> > > Subject: [PATCH] iio: ltc2497: Fix reading conversion results
> > > 
> > > After the result of the previous conversion is read the chip automatically
> > > starts a new conversion and doesn't accept new i2c transfers until this
> > > conversion is completed which makes the function return failure.
> > > 
> > > So add an early return iff the programming of the new address isn't needed.
> > > Note this will not fix the problem in general, but all cases that are currently
> > > used. Once this changes we get the failure back, but this can be addressed
> > > when the need arises.
> > > 
> > > Fixes: 69548b7c2c4f ("iio: adc: ltc2497: split protocol independent part in a
> > > separate module ")
> > > Reported-by: Meng Li <Meng.Li@xxxxxxxxxxxxx>
> > > Signed-off-by: Uwe Kleine-König <u.kleine-koenig@xxxxxxxxxxxxxx>
> > > ---
> > >  drivers/iio/adc/ltc2497.c | 13 +++++++++++++
> > >  1 file changed, 13 insertions(+)
> > > 
> > > diff --git a/drivers/iio/adc/ltc2497.c b/drivers/iio/adc/ltc2497.c
> > > --- a/drivers/iio/adc/ltc2497.c
> > > +++ b/drivers/iio/adc/ltc2497.c
> > > @@ -41,6 +41,19 @@ static int ltc2497_result_and_measure(struct
> > > ltc2497core_driverdata *ddata,
> > >  		}
> > > 
> > >  		*val = (be32_to_cpu(st->buf) >> 14) - (1 << 17);
> > > +
> > > +		/*
> > > +		 * The part started a new conversion at the end of the above i2c
> > > +		 * transfer, so if the address didn't change since the last call
> > > +		 * everything is fine and we can return early.
> > > +		 * If not (which should only happen when some sort of bulk
> > > +		 * conversion is implemented) we have to program the new
> > > +		 * address. Note that this probably fails as the conversion that
> > > +		 * was triggered above is like not complete yet and the two
> > > +		 * operations have to be done in a single transfer.
> > > +		 */
> 
> I'm a little confused by this comment.  It seems to say it will fail if we
> ever do have a different address?  That doesn't seem very helpful...

It's not a real problem in the sense that it triggers today. If you want
to read out (say) the channels 1, 5, 6 and 7, you could do:

	start conversion for channel 1
	wait for the conversion to complete
	read out conversion for channel 1 and start channel 5
	wait for the conversion to complete
	read out conversion for channel 5 and start channel 6
	wait for the conversion to complete
	read out conversion for channel 6 and start channel 7
	wait for the conversion to complete
	read out conversion for channel 7

With this procedure the step "read out conversion for channel X and
start channel Y" has to (and can) be done in a single transfer. But the
status quo is, that when these channels are to be read the following
happens:

	start conversion for channel 1
	wait for the conversion to complete
	read out conversion for channel 1 and (implicitly) start another conversion for channel 1

	wait for the conversion to complete

	start conversion for channel 5
	wait for the conversion to complete
	read out conversion for channel 5 and (implicitly) start another conversion for channel 5

	wait for the conversion to complete

	...

and ltc2497_result_and_measure is well suited to handle this.

So maybe reword the comment to:

	The part started a new conversion at the end of the above i2c
	transfer. With the current implementation of how reading is
	implemented in ltc2497core it never happens that this new
	conversion should be done for a different channel which would
	require writing a new channel address. (Actually writing such a
	new address requires more effort, either another delay must be
	added or the now two transfers must be combined into a single
	one.)

	So check the assumption that the channel really didn't change
	and then return early which does the right thing today.

?

Best regards
Uwe

-- 
Pengutronix e.K.                           | Uwe Kleine-König            |
Industrial Linux Solutions                 | https://www.pengutronix.de/ |

Attachment: signature.asc
Description: PGP signature