[ofa-general] Re: [PATCH 08 of 11] anon-vma-rwsem

[Robin Holt]

On Wed, May 14, 2008 at 06:11:22AM +0200, Nick Piggin wrote:
> On Tue, May 13, 2008 at 10:32:38AM -0500, Robin Holt wrote:
> > On Tue, May 13, 2008 at 10:06:44PM +1000, Nick Piggin wrote:
> > > On Thursday 08 May 2008 10:38, Robin Holt wrote:
> > > > In order to invalidate the remote page table entries, we need to message
> > > > (uses XPC) to the remote side.  The remote side needs to acquire the
> > > > importing process's mmap_sem and call zap_page_range().  Between the
> > > > messaging and the acquiring a sleeping lock, I would argue this will
> > > > require sleeping locks in the path prior to the mmu_notifier invalidate_*
> > > > callouts().
> > > 
> > > Why do you need to take mmap_sem in order to shoot down pagetables of
> > > the process? It would be nice if this can just be done without
> > > sleeping.
> > 
> > We are trying to shoot down page tables of a different process running
> > on a different instance of Linux running on Numa-link connected portions
> > of the same machine.
> 
> Right. You can zap page tables without sleeping, if you're careful. I
> don't know that we quite do that for anonymous pages at the moment, but it
> should be possible with a bit of thought, I believe.
> 
>  
> > The messaging is clearly going to require sleeping.  Are you suggesting
> > we need to rework XPC communications to not require sleeping?  I think
> > that is going to be impossible since the transfer engine requires a
> > sleeping context.
> 
> I guess that you have found a way to perform TLB flushing within coherent
> domains over the numalink interconnect without sleeping. I'm sure it would
> be possible to send similar messages between non coherent domains.

I assume by coherent domains, your are actually talking about system
images.  Our memory coherence domain on the 3700 family is 512 processors
on 128 nodes.  On the 4700 family, it is 16,384 processors on 4096 nodes.
We extend a "Read-Exclusive" mode beyond the coherence domain so any
processor is able to read any cacheline on the system.  We also provide
uncached access for certain types of memory beyond the coherence domain.

For the other partitions, the exporting partition does not know what
virtual address the imported pages are mapped.  The pages are frequently
mapped in a different order by the MPI library to help with MPI collective
operations.

For the exporting side to do those TLB flushes, we would need to replicate
all that importing information back to the exporting side.

Additionally, the hardware that does the TLB flushing is protected
by a spinlock on each system image.  We would need to change that
simple spinlock into a type of hardware lock that would work (on 3700)
outside the processors coherence domain.  The only way to do that is to
use uncached addresses with our Atomic Memory Operations which do the
cmpxchg at the memory controller.  The uncached accesses are an order
of magnitude or more slower.

> So yes, I'd much rather rework such highly specialized system to fit in
> closer with Linux than rework Linux to fit with these machines (and
> apparently slow everyone else down).

But it isn't that we are having a problem adapting to just the hardware.
One of the limiting factors is Linux on the other partition.

> > Additionally, the call to zap_page_range expects to have the mmap_sem
> > held.  I suppose we could use something other than zap_page_range and
> > atomically clear the process page tables.
> 
> zap_page_range does not expect to have mmap_sem held. I think for anon
> pages it is always called with mmap_sem, however try_to_unmap_anon is
> not (although it expects page lock to be held, I think we should be able
> to avoid that).

zap_page_range calls unmap_vmas which walks to vma->next.  Are you saying
that can be walked without grabbing the mmap_sem at least readably?
I feel my understanding of list management and locking completely
shifting.

> >  Doing that will not alleviate
> > the need to sleep for the messaging to the other partitions.
> 
> No, but I'd venture to guess that is not impossible to implement even
> on your current hardware (maybe a firmware update is needed)?

Are you suggesting the sending side would not need to sleep or the
receiving side?  Assuming you meant the sender, it spins waiting for the
remote side to acknowledge the invalidate request?  We place the data
into a previously agreed upon buffer and send an interrupt.  At this
point, we would need to start spinning and waiting for completion.
Let's assume we never run out of buffer space.

The receiving side receives an interrupt.  The interrupt currently wakes
an XPC thread to do the work of transfering and delivering the message
to XPMEM.  The transfer of the data which XPC does uses the BTE engine
which takes up to 28 seconds to timeout (hardware timeout before raising
and error) and the BTE code automatically does a retry for certain
types of failure.  We currently need to grab semaphores which _MAY_
be able to be reworked into other types of locks.


Thanks,
Robin