Re: [PATCH] hwmon: sht3x: set initial jiffies to last_update

Re: [PATCH] hwmon: sht3x: set initial jiffies to last_update

[Matt Ranostay]

On Wed, Jul 20, 2016 at 11:10 PM, Jean Delvare <jdelvare@suse.de> wrote:
> Hi Matt,
>
> On Fri, 8 Jul 2016 17:22:10 -0700, Matt Ranostay wrote:
>> On Fri, Jul 8, 2016 at 12:56 AM, Jean Delvare <jdelvare@suse.de> wrote:
>> > On jeu., 2016-07-07 at 19:46 -0700, Matt Ranostay wrote:
>> >> Handling the wraparound requires the data->last_update to be set to an
>> >> initial jiffies value. Otherwise you can start in a state where the
>> >> sensor will never request a reading.
>> >
>> > I can't see how. As I read the code, in the worst case, readings can be
>> > blocked for interval_ms (2 seconds maximum.)
>>
>> On 64-bit systems this is never an issue because the jiffies counter
>> will never wrap around.
>>
>> But my system is a 32-bit ARM core, so the the kernel sets the initial
>> value to 0xfffb6c20 so it will wrap around in 5 minutes to find buggy
>> code.
>>
>> So looking at time_after(0xfffb6c20, 0) will return false always till
>> it finally rolls over.
>
> I've always been confused by how jiffies wrapping is handled. So I
> tested it, and you are correct.
>

Yeah but of course on 64-bit systems most just use the lower 32-bits
of the jiffies counter.

But since that could rollover, and the unlikely event no sensor data
was read for ~57 days you could be in the same state that couldn't get
a data sample ....

Not sure if we want to check this edge case.. and also this effects a
few iio drivers (mostly mine :)) if we really care about this.


>> >> Cc: Guenter Roeck <linux@roeck-us.net>
>> >> Cc: David Frey <david.frey@sensirion.com>
>> >> Signed-off-by: Matt Ranostay <mranostay@gmail.com>
>> >> ---
>> >>  drivers/hwmon/sht3x.c | 2 +-
>> >>  1 file changed, 1 insertion(+), 1 deletion(-)
>> >>
>> >> diff --git a/drivers/hwmon/sht3x.c b/drivers/hwmon/sht3x.c
>> >> index 450645b6053d..05a925257938 100644
>> >> --- a/drivers/hwmon/sht3x.c
>> >> +++ b/drivers/hwmon/sht3x.c
>> >> @@ -722,7 +722,7 @@ static int sht3x_probe(struct i2c_client *client,
>> >>       data->setup.blocking_io = false;
>> >>       data->setup.high_precision = true;
>> >>       data->mode = 0;
>> >> -     data->last_update = 0;
>> >> +     data->last_update = jiffies;
>> >>       data->client = client;
>> >>       crc8_populate_msb(sht3x_crc8_table, SHT3X_CRC8_POLYNOMIAL);
>> >>
>> >
>> > Both look equally wrong to me. With your proposal, accessing the sysfs
>> > attributes right after loading the driver will not trigger a reading.
>> >
>> > In order to guarantee that the first access will trigger a reading,
>> > data->last_update should be initialized to jiffies -
>> > msecs_to_jiffies(2000) (the maximum interval value.)
>>
>> Ok that is fine.  Rather do  jiffies + (2 * HZ)
>
> 2 * HZ is fine. But it's minus, not plus. You want to initialize the
> last update time to 2 seconds BEFORE the driver is being loaded, so
> that the current time is at least 2 seconds AFTER the (fake) last
> update at first access. And you should subtract another jiffy to be on
> the safe side.
>
> Can you please send an updated patch?

Sure will do.

>
> --
> Jean Delvare
> SUSE L3 Support

Re: [PATCH] hwmon: sht3x: set initial jiffies to last_update

[Jean Delvare]

Hi Matt,

On Sun, 24 Jul 2016 16:50:39 -0700, Matt Ranostay wrote:
> On Wed, Jul 20, 2016 at 11:10 PM, Jean Delvare <jdelvare@suse.de> wrote:
> > On Fri, 8 Jul 2016 17:22:10 -0700, Matt Ranostay wrote:
> >> On Fri, Jul 8, 2016 at 12:56 AM, Jean Delvare <jdelvare@suse.de> wrote:
> >> > On jeu., 2016-07-07 at 19:46 -0700, Matt Ranostay wrote:
> >> >> Handling the wraparound requires the data->last_update to be set to an
> >> >> initial jiffies value. Otherwise you can start in a state where the
> >> >> sensor will never request a reading.
> >> >
> >> > I can't see how. As I read the code, in the worst case, readings can be
> >> > blocked for interval_ms (2 seconds maximum.)
> >>
> >> On 64-bit systems this is never an issue because the jiffies counter
> >> will never wrap around.
> >>
> >> But my system is a 32-bit ARM core, so the the kernel sets the initial
> >> value to 0xfffb6c20 so it will wrap around in 5 minutes to find buggy
> >> code.
> >>
> >> So looking at time_after(0xfffb6c20, 0) will return false always till
> >> it finally rolls over.
> >
> > I've always been confused by how jiffies wrapping is handled. So I
> > tested it, and you are correct.
> 
> Yeah but of course on 64-bit systems most just use the lower 32-bits
> of the jiffies counter.

Actually it seems the jiffies are initialized the same on 32-bit and
64-bit systems, so on 64-bit systems you'll get 33-bit values after 5
minutes, and 34-bit values after about 100 days (if HZ=1000.) But
you'll never realistically wrap, you are right.

> But since that could rollover, and the unlikely event no sensor data
> was read for ~57 days you could be in the same state that couldn't get
> a data sample ....

At HZ=1000 the jiffies wrap after 49 days, IIRC, not 57. Not that it
matters ;-)

> Not sure if we want to check this edge case.. and also this effects a
> few iio drivers (mostly mine :)) if we really care about this.

All other hwmon drivers have the same problem. I can't think of a cheap
way to solve this problem, nor do I think we need to. If you monitor
your system, you are supposed to poll the device on a regular basis,
not once every 49 days.

-- 
Jean Delvare
SUSE L3 Support