Re: Inter-module compatibility

On Wed, 10 Aug 2005, David Dawes wrote:
> On Wed, Aug 10, 2005 at 10:49:49AM -0600, Marc Aurele La France wrote:
>> On Tue, 9 Aug 2005, David Dawes wrote:
>>> On Tue, Aug 09, 2005 at 10:45:15AM -0600, Marc Aurele La France wrote:
>>>> Our loader scheme has been, and is, reasonably good at dealing with
>>>> incompatibilities between modules and the core binary.  But it is not so
>>>> good with ensuring modules are compatible with each other.

>>>> As a case in point, I am now faced with strong motivation to rework the
>>>> three Vbe.*InfoBlock structures currently exported by vbe.h.

>>>> Now, it's not my purpose here to get into these motivations, but reworking
>>>> these structures would have consequences, one of which I am unsure how to
>>>> deal with.  At minimum, what I need to do is ...

>>>> a) prevent drivers compiled with the current vbe.h from using a vbe module
>>>>    created with the new vbe.h;  and
>>>> b) prevent drivers compiled with the new vbe.h from using a vbe module
>>>>    created with the current vbe.h.

>>>> The hammer approach would suggest bumping the video driver ABI version
>>>> number all the way up to 1.0.  (It currently sits at 0.8).  This is
>>>> overkill as doing so would prevent the loading of _any_ version 0.*
>>>> module, vendor-provided or not, making this "politically" unwise.

>>>> Bumping this version number to 0.9 (or leaving it alone) is insufficient
>>>> as it prevents neither "a)" nor "b)", so changes would be required.

>>>> I've prototyped a way of dealing with "a)".  Basically, I'd have the vbe
>>>> module's Setup() function call the common layer to associate a
>>>> (potentiall  extended) XF86ModReqInfo occurrence with the vbe module, 
>>>> which the loader would then use to check parent modules against.

>>>> Dealing with "b)" is trickier.  All I can think of right now would require
>>>> changes to the way video drivers load the vbe module, something I'd like
>>>> to stay away from, especially in the presence of vendor-provided source
>>>> and/or binaries.

>>> "b)" could be handled by changing the drivers to specify a vbe module
>>> version requirement when loading the vbe module.  I don't think that is
>>> unreasonable given that child module version checking is already a part
>>> of the loadModule interface.

>> Yes.  One issue with this is that it doesn't do anything for vendor-provided
>> drivers.

> Vendors would need to handle the "b)" case in their drivers.

>> On the other hand, in this case, this would only need to be done to that
>> subset of drivers that actually peek/poke inside the structures to be
>> reworked, rather than passing structure occurrences around.  This
>> requirement would need to be clearly documented.

>> Also, vbe.h could provide a macro to load the vbe module, the idea being to
>> make the module control its own loading.

>>> If we were to start adding module interface version checking for
>>> other shared module interfaces, that would head off similar issues
>>> in the future.

>> Yes, and how a module is loaded could be controlled by its main header, or
>> something like "xxxModule.h".

> The "a)" case can be made to fail cleanly by having the old names being
> functions that print out an informative message and exit.

> The "b)" case could be handled like this:

> #define VBEExtendedInit(a,b,c) \
>  (xf86LoaderCheckSymbol("newVBEExtendedInit") : \
>    newVBEExtendedInit(a,b,c) ?
>    (FatalError(INFORMATIVE_MESSAGE), NULL))

> However, I think it's better to control module loading from the module
> public headers, thus incorporating interface version checking with the
> interface definition and avoiding the need to resort to tricks like this.

The attached (against current CVS) is a nearly finalised implementation of 
the above.  I've tested this with various versions of the vbe and vesa 
modules, and it works as expected.

However, what I have here is _very_ specific to the problem at hand, that 
being dealing with the consequences of changing the vbe InfoBlock structures. 
This means that I only changed those drivers (i810 (i830, actually), savage, 
sis, vesa and xgi) that actually care about the layout of these structures. 
For these drivers, both "a)" and "b)" are enforced.

For other "vbe'ing" drivers, such as my own atimisc, "b)" is not enforced 
(good), but "a)" is, unnessarily.  That's because, as I have it here, 
vbeSetup() doesn't have a reasonable way of detecting its parent module 
doesn't peek/poke into the changed structures.  So, I'm left with an 
assymmetric implementation.

The options, at this point, are ...

1) Leave this as is;
2) Change all vbe'ing drivers.  This would put them all on the same footing
    WRT enforcing "a)" and "b)".
3) Make vbeSetup() smarter.  One possibility here is to check version
    requirements only if the parent references, say, VBEGetModeInfo(),
    although this wouldn't work for the combined i810 module, where only the
    i830 code path looks into these structures.

I don't have a strong preference, except to commit a finalised version of 
this change all in one go, to eliminate exposure.

Comments?

Thanks.

Marc.

+----------------------------------+-----------------------------------+
|  Marc Aurele La France           |  work:   1-780-492-9310           |
|  Academic Information and        |  fax:    1-780-492-1729           |
|    Communications Technologies   |  email:  [email protected]          |
|  352 General Services Building   +-----------------------------------+
|  University of Alberta           |                                   |
|  Edmonton, Alberta               |     Standard disclaimers apply    |
|  T6G 2H1                         |                                   |
|  CANADA                          |                                   |
+----------------------------------+-----------------------------------+
XFree86 developer and VP.  ATI driver and X server internals.