Peter Zijlstra wrote:

On Fri, 2008-08-29 at 14:03 -0400, Gregory Haskins wrote:
  

quoted

*Patch submitted for inclusion in PREEMPT_RT 26-rt4.  Applies to 2.6.26.3-rt3*

Hi Ingo, Steven, Thomas,
  Please consider for -rt4.  This fixes a nasty deadlock on my systems under
  heavy load.

[
Changelog:
	v2: only touch seqlock_t because raw_seqlock_t doesn't require
	    serialization and userspace cannot modify data during a read

	v1: initial release
]

-Greg

----
seqlock: serialize against writers

Seqlocks have always advertised that readers do not "block", but this was
never really true.  Readers have always logically blocked at the head of
the critical section under contention with writers, regardless of whether
they were allowed to run code or not.

Recent changes in this space (88a411c07b6fedcfc97b8dc51ae18540bd2beda0)
have turned this into a more explicit blocking operation in mainline.
However, this change highlights a short-coming in -rt because the
normal seqlock_ts are preemptible.  This means that we can potentially
deadlock should a reader spin waiting for a write critical-section to end
while the writer is preempted.
    

Ah, the point I was missing is higher-priority realtime task, in which
case the write side will never run because it wont preempt.
  

Yep

quoted

This patch changes the internal implementation to use a rwlock and forces
the readers to serialize with the writers under contention.  This will
have the advantage that -rt seqlocks_t will sleep the reader if deadlock
were imminent, and it will pi-boost the writer to prevent inversion.

This fixes a deadlock discovered under testing where all high prioritiy
readers were hogging the cpus and preventing a writer from releasing the
lock.

Since seqlocks are designed to be used as rarely-write locks, this should
not affect the performance in the fast-path
    

Still dont like this patch, once you have a rwlock you might as well go
all the way.

Why?  A full rwlock will still be much slower since the readers will
always need an atomic op.  This construct only uses atomic ops in the
slow path under contention, which should be rare, and is thus still
superior when retries are permissible to the design.

 Esp since this half-arsed construct defeats PI in certain
cases.
  

Ouch.  While I admit that you can still get into inversion scenarios
once the reader leaves the seqbegin, this is the nature of seqlocks. 
The only ways I can think of to get around this involve atomic ops in
the fast path, which I think should be avoided.  What would you suggest
otherwise?

-Greg