If it was possible to create a driver model that allowed module compatibility
across different releases/revisions without recompilation and with minimal
performance hit would this be desirable?
If such a model was possible should development be avoided to discourage
binary only modules?
James
* James Clark ([email protected]) wrote:
> If it was possible to create a driver model that allowed module compatibility
> across different releases/revisions without recompilation and with minimal
> performance hit would this be desirable?
>
> If such a model was possible should development be avoided to discourage
> binary only modules?
I would argue that is exactly the situation seen by XFree with quite a
few cards requiring binary modules for full functionality.
Dave
-----Open up your eyes, open up your mind, open up your code -------
/ Dr. David Alan Gilbert | Running GNU/Linux on Alpha,68K| Happy \
\ gro.gilbert @ treblig.org | MIPS,x86,ARM,SPARC,PPC & HPPA | In Hex /
\ _________________________|_____ http://www.treblig.org |_______/
1. If it was possible to create a driver model that allowed module
compatibility across different releases/revisions without recompilation and
with minimal performance hit would this be desirable?
2. If such a model was possible should development be avoided to discourage
binary only modules?
James
On Sun, 14 Sep 2003 21:38:32 BST, James Clark <[email protected]> said:
> 1. If it was possible to create a driver model that allowed module
> compatibility across different releases/revisions without recompilation and
> with minimal performance hit would this be desirable?
>
> 2. If such a model was possible should development be avoided to discourage
> binary only modules?
-ENOTAGAIN
On Sul, 2003-09-14 at 21:38, James Clark wrote:
> 1. If it was possible to create a driver model that allowed module
> compatibility across different releases/revisions without recompilation and
> with minimal performance hit would this be desirable?
We've spent years trying to reduce the number of binary differences.
Since it involves the most basic blocks of code, choice of instructions
it isnt going to be avoidable.
> How does your proposal differ/improve on the demand-load module system?
Modules must be (re)compiled for each specific kernel release. This is because
of the very tight coupling with the kernel. The kernel changes so much that
it is often necessary to amend a module to keep up. My suggestion is to
create an abstract interface between the two. The interface (called a binary
interface) would define the way the two 'talk' to one another. Once the
interface has been released as v1.0 any module could be made compatible with
the v1.0 interface. After a while it would be necessary to release improved
interfaces. This would not break the existing modules as the old interface
could remain due to its abstract nature. This technique is used in everything
from Windows drivers, Netscape pluggins and even normal programs - image
saying you must buy a new version of Word Perfect for every minor release of
the OS - it has 'binary' compatibility. Unfortunately many see this as active
encouragement of 'binary only' modules which have no source code. This idea
would cost in performance. I suggest this loss is worthwhile to resolve
compatibility and usability.
> This seemed to work well in Mandrake, or am I missing the point?!?
Unless you build the module yourself you are the mercy of Mandrake to make
each module - I don't see this as being much different from the situation
with Windows. Poor old Mandrake must remake almost every module almost
everytime. You cannot use a module from Redhat's kernel. In effect everyone
is duplicating effort. Modules must currently be compiled, this requires a
degree of expertise and can be daunting to someone who just wants to use the
OS. My suggestion would not remove need to release modules in source form,
but once compiled they could be plugged into any kernel and distributed in
binary+source form.
> Also, the only reply to your posting that I read (and I've only read a
> fraction of them!) which offered any genuine usefulness was the point
> that if you offer the code then they'll consider the merits. Are you
> considering an actual solution, or just suggesting a policy change?
I understand the way competitor binary interfaces work. Currently I'm not
going to roll up my sleeves and write this system. I don't have the technical
expertise to design such a thing and although I could learn the curve at this
time is too steep. I do think that my experience with similar competitor
systems allows me to speak on the subject. However, I feel that pushing for
this change is a positive thing. It has started a debate on 'misusing' the
GPL to prevent binary only modules and has resulted in some positive
comments. If they debate it rationally and then decide not to bother I will
have achieved a lot. I do feel qualified to make this (obvious) suggestion
and comment on the design of any resulting interface.
James
On Tue, Sep 16, 2003 at 12:13:38AM +0100, James Clark wrote:
> I understand the way competitor binary interfaces work. Currently I'm not
> going to roll up my sleeves and write this system. I don't have the technical
> expertise to design such a thing and although I could learn the curve at this
> time is too steep. I do think that my experience with similar competitor
> systems allows me to speak on the subject. However, I feel that pushing for
> this change is a positive thing. It has started a debate on 'misusing' the
> GPL to prevent binary only modules and has resulted in some positive
> comments. If they debate it rationally and then decide not to bother I will
> have achieved a lot. I do feel qualified to make this (obvious) suggestion
> and comment on the design of any resulting interface.
I and others have repeatedly pointed out the big failings in this, which you
have ignored.
To do this means you must freeze the size of any structure passed across the
kernel-module boundary. You must freeze any functions which are in-lined.
spin_lock() must not have debugging code which is enabled by a CONFIG_
option, or it must become a function call. You can never optimize away
stuff that only matters on SMP systems. You can never have a structure
include members conditionally upon CONFIG_ options, or you have to provide
accessor functions for every field of every struct.
This is what you are not getting. Because module code is a first-class
citizen in Linux, you CAN'T decouple them, without imposing BIG UGLY
restrictions on all modules. Do you understand the examples I am giving
you? I can be more explicit, if you like.
Now, if you, or someone can devise a cleanish way to do it, SHOW US CODE.
It doesn't have to work fully, just enough to convince us it could work AT
ALL.
I assert that any solution will have one or more of the following
attributes, any of which are unacceptable:
* will perform terribly
* will be cumbersome to code
* will impose new restrictions on kernel developers
* will be unmaintainable
I BEG YOU to prove me wrong.
--
Notice that as computers are becoming easier and easier to use,
suddenly there's a big market for "Dummies" books. Cause and effect,
or merely an ironic juxtaposition of unrelated facts?