Re: default IRQ affinity change in v2.6.27 (breaking several SMP PPC based systems)

Kumar Gala wrote:

On Oct 24, 2008, at 11:09 AM, Chris Snook wrote:

quoted

Kumar Gala wrote:

quoted

On Oct 24, 2008, at 10:17 AM, Chris Snook wrote:

quoted

Kumar Gala wrote:

quoted

It appears the default IRQ affinity changes from being just cpu 0 
to all cpu's.  This breaks several PPC SMP systems in which only a 
single processor is allowed to be selected as the destination of 
the IRQ.
What is the right answer in fixing this?  Should we:
  cpumask_t irq_default_affinity = 1;
instead of
  cpumask_t irq_default_affinity = CPU_MASK_ALL?

On those systems, perhaps, but not universally.  There's plenty of 
hardware where the physical topology of the machine is abstracted 
away from the OS, and you need to leave the mask wide open and let 
the APIC figure out where to map the IRQs.  Ideally, we should 
probably make this decision based on the APIC, but if there's no PPC 
hardware that uses this technique, then it would suffice to make 
this arch-specific.

What did those systems do before this patch?  Its one thing to expose 
a mask in the ability to change the default mask in 
/proc/irq/default_smp_affinity.  Its another (and a regression in my 
opinion) to change the mask value itself.

Before the patch they took an extremely long time to boot if they had 
storage attached to each node of a multi-chassis system, performed 
poorly unless special irqbalance hackery or manual assignment was 
used, and imposed artificial restrictions on the granularity of 
hardware partitioning to ensure that CPU 0 would always be a CPU that 
could service all interrupts necessary to boot the OS.

quoted

As for making it ARCH specific, that doesn't really help since not 
all PPC hw has the limitation I spoke of.  Not even all MPIC (in our 
cases) have the limitation.

What did those systems do before this patch? :)

Making it arch-specific is an extremely simple way to solve your 
problem without making trouble for the people who wanted this patch in 
the first place.  If PPC needs further refinement to handle particular 
*PICs, you can implement that without touching any arch-generic code.

So why not just have x86 startup code set irq_default_affinity = 
CPU_MASK_ALL than?

It's an issue on Itanium as well, and potentially any SMP architecture with a 
non-trivial interconnect.

-- Chris