Re: ff-core effect id handling in case of a failed effect upload

[Elias Vanderstuyft]

On Wed, Feb 19, 2014 at 5:49 PM, Anssi Hannula  wrote:
> (added Dmitry to CC)
>
> 19.02.2014 13:42, Elias Vanderstuyft kirjoitti:
> > Hi,
>
> Hi,
> > In the process of reviewing the Wine DInput translation layer, I
> > noticed an inconvenience (in the ff-core implementation?) that can
> > possibly lead to confusing problems to application developers (not
> > only for Wine), in short:
> > If a new (id==-1) effect was uploaded (look at
> > ff-core.c::input_ff_upload(...)) that failed (e.g. returning EINVAL),
> > ff-core will have assigned a positive number to the effect id. This
> > can be confusing because the dev->ff->effects[] array will not contain
> > an element at the index of that new effect id.
>
> I agree that this is a bit confusing, and the kernel code should
> probably be modified to not clobber the ioctl-provided effect on failure
> (effect->id is set to an "undefined" value, i.e. next free effect slot).
>
> Dmitry, WDYT?
>
> The downside is that if changed, any new userspace code may unknowingly
> depend on the fixed non-clobbering behavior (though apparently they
> already do, as evidenced by Wine DInput, so might not be much of an
> argument...).

I don't think that's a problem.
Something that can be more problematic, is that a newer application
assumes the fixed non-clobbering behaviour, but runs on a kernel with
the old clobbering behaviour, could experience problems.
> > Here is a more elaborated description/discussion:
> > - This is a bug that is either the responsibility of the Linux kernel,
> >
> > or of Wine (and possibly other applications that do the same thing as
> > described below):
> >     It is caused by the following situation:
> >         When uploading an effect, the specific kernel device driver
> > may return an error,
> >         e.g. EINVAL for example when a periodic's effect period is set to zero.
> >         This error will then be returned by "ioctl(*(This->fd),
> > EVIOCSFF, &This->effect)".
> >         With or without error, one can find out that
> > /drivers/input/ff-core.c:input_ff_upload(...) is called,
> >         which will set effect->id >= 0, if it was set to -1 (=> new
> > effect created) by the user.
> >     But in case of an error:
> >     - Assume effect->id was set to -1 by the user:
> >         The error is reported by ff->upload(...) at
> >
> > /drivers/input/ff-core.c:167,
> >             the effect->id will also be set >= 0 (*).
> >             The offending effect will not be saved in the
> > ff->effects[] array (***).
> >         - Assume effect->id was set >= 0 by the user (and
> > ff->effects[effect->id] is a valid existing effect):
> >             The error is reported by ff->upload(...) at
> > /drivers/input/ff-core.c:167,
> >             the effect->id will remain unchanged (**).
> >             The offending effect will not overwrite the
> > ff->effects[effect->id] element (****).
> >     Is this (see *, **, *** and ****) desired behaviour?
> >     - If yes:
> >         Change the following in Wine's dinput/effect_linuxinput.c:90 :
> >             if (ioctl(*(This->fd), EVIOCSFF, &This->effect) == -1) {
> >         to :
> >             int effectId_old = This->effect.id;
> >             if (ioctl(*(This->fd), EVIOCSFF, &This->effect) == -1) {
> >                 This->effect.id = effectId_old;
> >     - If no for *:
> >         Kernel code /drivers/input/ff-core.c:input_ff_upload(...)
> >
> > has to be patched to revert "effect->id" back to its original value
> > set by the user,
> >             which is only needed when the initial (by user) value of
> > "effect->id" was equal to -1.
> >         - If no for **** (or maybe also ***):
> >             ff->effects[effect->id] could be replaced by an 'empty'
> > effect (however this can get complex because the effect's type has to
> > remain unchanged)
> >             This would be a change in the kernel code
> > /drivers/input/ff-core.c:input_ff_upload(...).
> >         - If no for **:
> >             I don't really know. Discussion is needed.
> >     - In my opinion, **, *** and **** are desired behaviour, while
> > - should leave the effect->id at -1.
>
> Right.
> >         In that case, Wine's dinput implementation does not have
> >
> > to be patched, and the kernel code only should apply a minor patch.
>
> Well, maybe better to change dinput anyway so that it can work properly
> with current kernel versions (assuming this gets changed in the future)?
> Not my call, though.

Yes, agreed.
I'll include the code (with effectId_old variable) and this discussion
when I submit my whole patchset for Wine's DInput libs to Wine-devel
mailing list. Right now, I don't know who I should personally mail to.
> --
> Anssi Hannula

Elias