On Wed, Apr 21, 2021 at 08:31:27PM -0300, Jonas Malaco wrote: > On Wed, Apr 21, 2021 at 10:21:36AM -0700, Guenter Roeck wrote: > > On Wed, Apr 21, 2021 at 01:48:03PM -0300, Jonas Malaco wrote: > > > On Tue, Apr 13, 2021 at 09:45:29AM -0300, Jonas Malaco wrote: > > > > Guenter (and others on this list), > > > > > > Very gentle ping. > > > > > > I also thought posting these questions first would be less disruptive > > > than a RFC patch, but please let me know if you prefer the latter. > > > > > > > It is a difficult subject, and I am struggling myself with the situations > > you are presenting. > > I am somewhat relieved that these issues are not so silly. And I really > appropriate your comments. > > Please take a look at a few more comments bellow. > > > > > > Thanks again, > > > Jonas > > > > > > > > > > > I am getting ready to submit a driver for NZXT Grid+ V3 and Smart Device > > > > (V1) fan controllers, but I am having trouble deciding how to expose > > > > their PWM control due to some device limitations. > > > > > > > > Before getting into those, let me first give some very basic context... > > > > > > > > These devices are USB HIDs, and asynchronously send "status" reports > > > > every 200 ms to communicate speed, current, voltage and control mode for > > > > their channels (one channel per report). > > > > > > > > Fans can be controlled by sending HID output reports to the device, and > > > > both DC and PWM modes are supported. The device features a special > > > > initialization routine (that must be requested during probe) which > > > > automatically detects the appropriate control mode for each channel. > > > > > > > > Back to the device limitations... > > > > > > > > The first is that PWM values can be set, but not read back. And neither > > > > hwmon[1] nor lm-sensors' pwmconfig/fancontrol expect pmw* attributes to > > > > be WO. One solution is to have the driver track the PWM values that are > > > > set through it and return those, but is this acceptable? > > > > I have seen a couple of those recently. I think returning -ENODATA > > if the value isn't known (yet) is the best possible solution. I thought > > about adding that to the ABI, actually. > > We can never read the pwm[1-*] attributes, not even for detected and > controllable fans after the initialization procedure. > > And returning -ENODATA for pwm[1-*] reads makes pwmconfig/fancontrol > unhappy: > Seems to me that pwmconfig is then maybe not appropriate to use, and maybe there should be no driver for this device in the kernel in the first place. Returning a random value after setting the pwm value to 255, removing, and re-inserting the driver is, in my opinion, even worse than returning -ENODATA. After all, the driver doesn't know the pwm value, and it is really a bad idea to report data which doesn't reflect reality. Guenter > # pwmconfig > [...] > Found the following PWM controls: > cat: hwmon0/pwm1: No data available > hwmon0/pwm1 current value: > cat: hwmon0/pwm1: No data available > /bin/pwmconfig: line 201: [: : integer expression expected > cat: hwmon0/pwm1: No data available > hwmon0/pwm1 stuck to > Manual control mode not supported, skipping hwmon0/pwm1. > [...] > > # fancontrol > [...] > Enabling PWM on fans... > cat: hwmon0/pwm1: No data available > Starting automatic fan control... > /bin/fancontrol: line 551: read: read error: 0: No data available > Error reading PWM value from /sys/class/hwmon/hwmon0/pwm1 > Aborting, restoring fans... > cat: hwmon0/pwm1: No data available > /bin/fancontrol: line 458: [: : integer expression expected > hwmon0/pwm1_enable stuck to 1 > Verify fans have returned to full speed > > > > > > > > > > > The other starts with PWM control being disabled for channels that the > > > > device identifies as unconnected. This is not in itself a problem, but > > > > the initialization routine (where the detection happens) is > > > > asynchronous, takes somewhere around 5 seconds, and we do not have any > > > > way of directly querying its result. We only know the control mode of > > > > each channel (be it DC, PWM or disabled) from the regular status > > > > reports. > > > > Again, I think the best solution is to return -ENODATA until the value > > is known. > > Ok. > > > > > > > > > > > These limitations make it complicated to simply use is_visible() to hide > > > > pwm attributes of unconnected channels. We would need to register with > > > > the hwmon subsystem only after getting enough post-initialization status > > > > reports for all channels, and this would essentially mean to sleep for > > > > 6+ seconds. We would also need to unregister and re-register when going > > > > through a suspend-reset_resume cycle, because the device may have its > > > > state wiped, requiring reinitialization.[2] > > > > > > I think the above should resolve that. > > Yes, as well as your suggestion bellow. > > > > > > > A different approach to handle this, which I have preferred _so far,_ is > > > > to use pwm*_enable = 0 to report the unconnected channels to user-space, > > > > while keeping the other pwm attributes visible. But this comes with > > > > other problems. > > > > > > > > First, lm-sensors' pwmconfig expects to be able to write to a > > > > pwm*_enable attribute if it exists, but the device does not support that > > > > operation. The hwmon documentation states that RW values may be RO, but > > > > pwmconfig is out there and in use. So far I simply return 0 to attempts > > > > at those writes, silently ignoring them; functional, but certainly a > > > > hack. > > > > It is a bad idea to return 0 if the value is not accepted. You could check > > if the written value matches the current value and return 0 if it does, > > and an error such as -EOPNOTSUPP or -EINVAL otherwise. > > It worked really well, thanks! > > > > > > > > > > > Second, if PWM control is disabled for a channel, but its pwm* and > > > > pwm*_mode attributes are still visible, what should we return when > > > > user-space attempts to write to them? The practical answer may simply > > > > be to return -EOPNOTSUPP, but this makes me wonder if the whole approach > > > > (of handling these cases with pwm*_enable instead of is_visible()) is > > > > not doomed. > > > > > > Mode isn't really writeable either, isn't it ? If so, use the same trick as > > with the _enable attribute. > > You're right, but setting its visibility to 0444 didn't cause issues for > pwmconfig or fancontrol, so I don't think the trick is necessary here. > > > > > The same is effectively true for the pwm value itself: Since both _enable > > and _mode are effectively read-only, you can accept a write only if > > fan control is enabled, and return an error if it isn't. > > Ok. > > > > > > > A final minor problem is that channels detected as unconnected run at > > > > 40% PWM, but the documentation for pwm*_enable == 0 is a bit too > > > > specific: "no fan speed control (i.e. fan at *full* speed)" (emphasis > > > > mine). > > > > Just document the difference. Reality doesn't always match our expectations. > > Ok. > > Thanks, > Jonas > > > > > Thanks, > > Guenter > > > > > > > > > > Do you have any suggestions and/or recommendations? > > > > > > > > If it helps, a pre-RFC (but functional and mostly clean) version of the > > > > driver can be found at: > > > > > > > > https://github.com/jonasmalacofilho/linux/blob/p-hwmon-add-nzxt-smartdevice-gridplus3/drivers/hwmon/nzxt-smartdevice.c > > > > > > > > Thanks, > > > > Jonas > > > > > > > > [1] According to Documentation/hwmon/sysfs-interface.rst. > > > > [2] The device also does not respond to HID Get_Report, so it is not > > > > trivial to check whether it really needs to be reinitialized, since > > > > the only symptom of that being necessary is the absence of the > > > > asynchronous status reports.
