Re: [PATCH 2/2] hwmon: (dell-smm) Unify i8k_ioctl() and i8k_ioctl_unlocked()

[Pali Rohár]

On Monday 22 November 2021 11:28:30 Guenter Roeck wrote:
> On 11/22/21 11:10 AM, Pali Rohár wrote:
> > On Monday 22 November 2021 19:50:14 Armin Wolf wrote:
> > > Am 22.11.21 um 18:55 schrieb Guenter Roeck:
> > > > On 11/22/21 8:01 AM, Pali Rohár wrote:
> > > > > On Saturday 20 November 2021 18:03:19 Armin Wolf wrote:
> > > > > > The only purpose of i8k_ioctl() is to call i8k_ioctl_unlocked()
> > > > > > with i8k_mutex held. Judging from the hwmon code, this mutex
> > > > > > only needs to be held when setting the fan speed/mode.
> > > > > 
> > > > > Really? I think that there is no difference between setting and getting
> > > > > fan speed/mode. At least I do not see why 'set' needs mutex and 'get' do
> > > > > not need it. Some more explanation is needed...
> > > > > 
> > > > I8K_SET_FAN sets the fan speed and returns the current status. Without
> > > > locking, the returned status may not match or be associated with the
> > > > previous
> > > > set operation.
> > > > 
> > > > Maybe that doesn't matter, and the synchronization is not needed. If so,
> > > > you can probably remove the locking entirely.
> > > > 
> > > > Guenter
> > > 
> > > That is the reason i kept the locking code. Since i do not want to break
> > > the ioctl interfacein any way, removing the locking code seems too risky
> > > to me.
> > 
> > I see. That is a good point.
> > 
> > But there is same race condition also when at the same time going to
> > change speed via ioctl and also via hwmon sysfs.
> > 
> 
> I thought the sysfs code does not change the fan speed and report the
> fan status in the same request. Did I miss something ?

No. I mean something different. Let me to write trace call:

CPU 0:                          CPU 1:
                                1. dell_smm_write()
1. ioctl(I8K_SET_FAN)
2. i8k_set_fan()
                                2. i8k_set_fan()
3. i8k_get_fan_status()

So to ensure that i8k_get_fan_status() on CPU 0 returns value which
belongs to i8k_set_fan() from CPU 0 it is needed to still use mutex.

Armin is right here and I think that patch is correct.