Re: [PATCH] x86/paravirt: revert exports to restore old behaviour

[Takashi Iwai <tiwai@suse.de>]

At Tue, 13 Nov 2007 16:51:04 -0800,
Zachary Amsden wrote:
> 
> On Tue, 2007-11-13 at 22:22 +0000, Christoph Hellwig wrote:
> > On Tue, Nov 13, 2007 at 12:21:16PM -0800, Jeremy Fitzhardinge wrote:
> > > Subject: x86/paravirt: revert exports to restore old behaviour
> > > 
> > > Subdividing the paravirt_ops structure caused a regression in certain
> > > non-GPL modules which try to use mmu_ops and cpu_ops.  This restores
> > > the old behaviour, and makes it consistent with the
> > > non-CONFIG_PARAVIRT case.
> > 
> > NACK, both of these are internal and graphics drivers should not be
> > using them.
> 
> Some of them are internal, but some are not, they just happened to be
> privileged CPU operations available to anyone.
> 
> Does anyone know what hooks they are actually using?  Something like
> reading MSRs is not internal at all, it is CPU specific, and the
> graphics drivers might have very good reasons to read them to figure out
> how AGP is configured for example.
> 
> The graphics drivers most certainly don't need to be paravirtualized
> however, so they could always work around this with raw asm constructs.
> 
> The mmu_ops hook is more debatable.  But until someone figures out what
> hooks they want, we can't know whether they are internal or not.  In any
> case, this is a regression.

I took at this problem (as I have an nvidia card on one of my
workstations), and found out that the following suffer from
EXPORT_SYMBOL_GPL changes:

* local_disable_irq(), local_irq_save*(), etc.
* MSR-related macros like rdmsr(), wrmsr(), read_cr0(), etc.
  wbinvd(), too.
* pmd_val(), pgd_val(), etc are all involved with pv_mm_ops.
  pmd_large() and pmd_bad() is also indirectly involved.
  __flush_tlb() and friends suffer, too.

The easiest workaround I found was to undefine CONFIG_PARAVIRT before
inclusion of linux kernel headers, but it is really ugly and hacky.
Redefinig with raw_*() and native_*() is another way, but it takes
much more work than defining these primitive functions in assembly.

So, in short, with EXPORT_SYMBOL_GPL change, it's pretty hard to write
a non-GPL driver in a same manner...


Takashi