Hi Guenter, Quoting Guenter Roeck (2015-10-20 06:30:03) > On 10/20/2015 06:17 AM, Marc Titinger wrote: > > On 20/10/2015 14:54, Guenter Roeck wrote: > >> On 10/20/2015 01:20 AM, Marc Titinger wrote: > >>> With the current implementation, the driver will prevent a readout at a > >>> pace faster than the default conversion time (2ms) times the averaging > >>> setting, min AVG being 1:1. > >>> > >>> Any sysfs "show" read access from the client app faster than 500 Hz > >>> will be > >>> 'cached' by the driver, but actually since do_update reads all 8 > >>> registers, > >>> the best achievable measurement rate is roughly 8*800 us (for the time > >>> spent in i2c-core) i.e. <= 156Hz with Beagle Bone Black. > >>> > >>> This change set uses a register mask to allow for the readout of a single > >>> i2c register at a time. Furthermore, performing subsequent reads on the > >>> same register will make use of the ability of the i2c chip to retain the > >>> last reg offset, hence use a shorter i2c message (roughly 400us > >>> instead of > >>> 800us spent in i2c-core.c). > >>> > >> That doesn't work. There could be accesses from other sources (such as > >> through > >> i2c-dev, or in multi-master systems) between two reads. > > > > Re-setting the register address with each read transaction will not prevent another master to change the configuration in your back, in this case. That sounds like a general issue of concurrent clients for one device, this is beyond just reading one register IMO. > > > That is an invasive change, though, not just a simple read. Sure, it is > a risk as well. But it is a different level of risk than someone using > i2cget or i2cdump while the driver is running. > > >> > >>> The best readout rate for a single measurement is now around 2kHz. And > >>> for > >>> four measurements around (1/(4*800us) = 312 Hz. Since for any readout > >>> rate > >>> faster than 160 Hz the interval is set by the i2c transactions > >>> completion, > >>> the 'last-update' anti-flooding code will not have a limiting effect in > >>> practice. Hence I also remove the elapsed time checking in the hwmon > >>> driver > >>> for ina2xx. > >>> > >>> To summarize, the patch provides a max bandwidth improvement with hwmon > >>> client apps from ~160 Hz to ~320 Hz, and better in single-channel > >>> polling mode. > >>> > >> Overall your patch pretty much re-implements regmap. Since you drop > >> caching, > >> it is also unnecessary to read all registers at a time, so you can just use > >> a function to read _one_ register and returns its value (with retries). > >> Or use regmap. Either case, do_update() and ina2xx_update_device() are no > >> longer needed. > > Agreed. > > > >> > >> If you want to convert the driver to regmap, just look for 'regmap' in > >> drivers/hwmon for examples. > > > > Fair enough, but based on your comments, I may look into an iio driver instead for this device, given our application, rather than 'twisting' the hwmon interface. > > > > Sorry, you lost me there. How are you twisting the hwmon interface ? > Because I am concerned about multiple accesses from multiple sources ? > How is iio going to solve that problem ? By ignoring it ? There is no twisting going on here :-) We're going to investigate IIO with or without the multi-client/regmap issues raised in this thread. The purpose is to investigate if we can get better performance/higher sample rate for our application (the ACME power measurement device). > > Not that I mind if you want to convert the driver to iio. One less driver > to take care of. Just asking. Yup, it's orthogonal to the issues raised here. Regards, Mike > > Thanks, > Guenter > _______________________________________________ lm-sensors mailing list lm-sensors@xxxxxxxxxxxxxx http://lists.lm-sensors.org/mailman/listinfo/lm-sensors
