PCI init vs. memory init

PCI init vs. memory init

[Linas Vepstas]

Paul,

I'd previously reported a problem where the PCI subsystem 
was getting inited before the memory subsystem. In the unusual
case of an EEH failure, this resulted in a crash in kmalloc,
which I hacked around with a if(!mem_init_done).

You mentioned that the sequence of inits seemed wrong, that the 
PCI init should be done later, after the memory init. I think
I agree; but when I took a very very quick look at the code, there 
was no obvious hook in later init to move the PCI init over to. 

Are you pursuing this further? Should I dig into it?  Any bright
ideas? Am I missing something obvious?  

There are several spots in in the powerpc PCI init code where 
a boot_mem alloc is used instead of kmalloc, and this boot_mem is 
then hacked around in the case of a PCI hotplug remove.  It would 
be nice to fix this...

--linas

Re: PCI init vs. memory init

[Paul Mackerras]

Linas Vepstas writes:

> You mentioned that the sequence of inits seemed wrong, that the 
> PCI init should be done later, after the memory init. I think
> I agree; but when I took a very very quick look at the code, there 
> was no obvious hook in later init to move the PCI init over to. 
> 
> Are you pursuing this further? Should I dig into it?  Any bright
> ideas? Am I missing something obvious?  

I assume you're talking about find_and_init_phbs() and eeh_init(),
which are currently called from pSeries_setup_arch().

Would a core_initcall be early enough for those?  It seems to me that
it probably would be.  What are the actual dependencies?  Clearly it
needs to be before pcibios_init(), which is a subsys_initcall.  Is
there anything else that they need to come before?

> There are several spots in in the powerpc PCI init code where 
> a boot_mem alloc is used instead of kmalloc, and this boot_mem is 
> then hacked around in the case of a PCI hotplug remove.  It would 
> be nice to fix this...

Indeed.

Paul.