Re: PROPOSAL: /proc/dev
H. Peter Anvin (hpa@transmeta.com)
Wed, 31 Dec 1997 20:31:54 -0800 (PST)
> > > Why not? If you could give us some good reasons why not then we probably
> > > will drop it. But several times, I have heard this come up, and never a
> > > good reason why it shouldn't be done. The two reasons that I have heard are:
> > > 1) Memory usage -- so if you can't spare the memory, don't use it!
> > > 2) Lack of user configurablity (IE different permissions, different names)
> > > -- so we have callbacks through a userspace daemon. Or we let people
> > > write really simple kernelspace modules. (Basicly a big select and a
> > > function call or three with constant arguments for each case.)
> >
> > These are the arguments the proponents shoot down constantly, but they
> > never address any of the REAL (IMO) problems (although the kernel
> > space usage is likely to be prohibitive if done anything remotely like
> > correctly.) YOU REALLY DON'T WANT CALLBACKS -- if you do, then your
> > device performance and reliability is going straight into the toilet.
>
> We already have callbacks for the f_ops. So this scheme will increase
> the callback depth by one. Is that really a big deal? We're only
> talking about a handfull of cycles or so. I'm sure the syscall
> overhead is much greater.
I'm talking about user space callbacks, as were you. They are
expensive and fragile.
> > At some point I'll probably get my act together and write an FAQ on
> > this. I'm sick and tired of rehashing the same argument every four
> > months for about four years now...
>
> Jeez, take it easy. I'm just trying to be constructive.
Sorry, sometimes even my patience wears thin... not with you put the
fact that this gets re-suggested over and over and over and over and
over...
-hpa