Re: [PATCH v4 01/24] input: Add ff-memless-next module

["Michal Malý" <madcatxster@devoid-pointer.net>]

On Wednesday 14 of May 2014 11:05:58 Dmitry Torokhov wrote:
> On Wed, May 14, 2014 at 10:35:25AM +0200, Michal Malý wrote:
> > Hi Dmitry,
> > 
> > thank you for reviewing this.
> > 
> > On Tuesday 13 of May 2014 23:38:06 Dmitry Torokhov wrote:
> > > On Sat, Apr 26, 2014 at 05:02:00PM +0200, Michal Malý wrote:
> > > > +
> > > > +/** DEFINITION OF TERMS
> > > > + *
> > > > + * Combined effect - An effect whose force is a superposition of
> > > > forces
> > > > + *                   generated by all effects that can be added
> > > > together.
> > > > + *                   Only one combined effect can be playing at a
> > > > time.
> > > > + *                   Effects that can be added together to create a
> > > > combined + *                   effect are FF_CONSTANT, FF_PERIODIC and
> > > > FF_RAMP. + * Uncombinable effect - An effect that cannot be combined
> > > > with
> > > > another effect. + *                       All conditional effects -
> > > > FF_DAMPER, FF_FRICTION, + *                       FF_INERTIA and
> > > > FF_SPRING are uncombinable. + *                       Number of
> > > > uncombinable effects playing simultaneously + *
> > > > depends on the capabilities of the hardware. + * Rumble effect - An
> > > > effect generated by device's rumble motors instead of + *
> > > > force feedback actuators.
> > > > + *
> > > > + *
> > > > + * HANDLING OF UNCOMBINABLE EFFECTS
> > > > + *
> > > > + * Uncombinable effects cannot be combined together into just one
> > > > effect,
> > > > at + * least not in a clear and obvious manner. Therefore these
> > > > effects
> > > > have to + * be handled individually by ff-memless-next. Handling of
> > > > these
> > > > effects is + * left entirely to the hardware-specific driver,
> > > > ff-memless-next merely + * passes these effects to the
> > > > hardware-specific
> > > > driver at appropriate time. + * ff-memless-next provides the UPLOAD
> > > > command to notify the hardware-specific + * driver that the userspace
> > > > is
> > > > about to request playback of an uncombinable + * effect. The
> > > > hardware-specific driver shall take all steps needed to make + * the
> > > > device ready to play the effect when it receives the UPLOAD command. +
> > > > *
> > > > The actual playback shall commence when START_UNCOMB command is
> > > > received.
> > > > + * Opposite to the UPLOAD command is the ERASE command which tells +
> > > > *
> > > > the hardware-specific driver that the playback has finished and that +
> > > > *
> > > > the effect will not be restarted. STOP_UNCOMB command tells
> > > > + * the hardware-specific driver that the playback shall stop but the
> > > > device + * shall still be ready to resume the playback immediately.
> > > > + *
> > > > + * In case it is not possible to make the device ready to play an
> > > > uncombinable + * effect (all hardware effect slots are occupied), the
> > > > hardware-specific + * driver may return an error when it receives an
> > > > UPLOAD command. If the
> > > 
> > > This part concerns me. It seems to me that devices supporting
> > > "uncombinable" effects are in fact not memoryless devices and we should
> > > not be introducing this term here. If the goal is to work around limited
> > > number of effect slots in the devices by combining certain effects then
> > > it needs to be done at ff-core level as it will be potentially useful
> > > for all devices.
> > 
> > Force generated by a conditional effect (referred to as "uncombinable"
> > within ff-memless-next to make the distinction clear) depends on a
> > position of the device. For instance the more a device is deflected from
> > a neutral position the greater force FF_SPRING generates. A truly
> > memoryless device would have to report its position to the driver, have
> > it calculate the appropriate force and send it back to the device. IMHO
> > such a loop would require a very high USB polling rate to play
> > conditional effects with acceptable quality.
> > 
> > We know for a fact that at least many (all?) Logitech devices that support
> > conditional effects use this "semi-memoryless" approach where FF_CONSTANT
> > and FF_PERIODIC are handled in the memoryless fashion and conditional
> > effects are uploaded to the device (in a somewhat simplified form). The
> > amount of effects that can be uploaded to a device is limited which is
> > why ff-memless-next uses two steps (UPLOAD/ERASE and START/STOP) to
> > handle these effects.
> > 
> > Conditional effects - even if they are of the same type - cannot be
> > effectively combined into one because superposition doesn't seem to work
> > here so they have to be processed one by one.
> > 
> > If we ever come across a really memoryless device it should not be
> > particularly difficult to add another callback to ff-memless-next which
> > would emulate conditional effects with constant force.
> 
> Thank you for the explanation. This further solidifies for me the idea
> that handling of such effects that are in fact uploaded to and managed
> by the device should not be handled by the memoryless core but rather by
> the driver itself. I.e. such drivers should implement their own play(),
> upload(), erase(), etc, and decide whether to use a hardware slot for
> the effect or handle effect in memoryless fashion (if possible). We can
> open ff-memless to allow such drivers to use parts of memoryless
> handling.

Well, these effects are not exactly managed by the device. The only thing that 
is uploaded to the device are parameters of the force to be generated. Other 
parameters - such as timing - still have to be managed by the driver. Any 
driver supporting conditional effects would then have to reimplement at least 
the timing logic.

Another thing of concern is rate limiting. During our testing we have 
discovered that some games can fire off FF commands at a very fast rate - much 
faster than USB polling rate of a device. This eventually overfills the USB 
submit queue and messes everything up. A proper way to fix this would be to 
limit the rate at which the driver sends HW requests. We already have a few 
ideas as to how to implement this in ff-memless-next. If we had the driver use 
ff-memless-next to manage one category of effects and use its own logic to 
manage the rest it'd be next to impossible to do this properly.

It should also be noted that ff-memless-next almost passes the conditional 
effects through to the HW-specific driver that then takes care of everything. A 
practical example of how this works can be found in an experimental port of 
"hid-lg4ff" driver to ff-memless-next by Edwin Velds (https://github.com/edwin-v/linux-hid-lg4ff-next). The only thing that ff-memless-next does is that it 
tells the HW-specific driver that such an effect such start or stop playing.

The reasoning above made me implement the support for conditional effects in 
the way I did. As much as I agree that we're not dealing with purely 
memoryless devices here, I believe that most more advanced FFB wheels/sticks 
are in fact "semi-memoryless" and will therefore benefit from the approach used 
in ff-memless-next.

Michal