[openib-general] Re: [PATCH] cm refcount race fix

[Michael S. Tsirkin]

Quoting r. Roland Dreier <rdreier at cisco.com>:
> Changing the refcount to an int doesn't seem like that daunting a
> change to make...

My problem is, we have the same bug in:

mthca
mad_rmpp.c
cma.c
mad.c
ucm.c
ucma.c
multicast.c

So changing all refcounts from atomic to int and making sure no one keeps the
lock when calling atomic_dec/atomic_inc is a big task.  As opposed to this, the
wake_up/wait_event idiom is easy to spot and just touching it is easier.

And while the trivially correct variant that I posted would be 2.6.17
material (or even -stable if I start running into this crash a lot),
this bigger change seems too risky.

But - if you are up to it ... Patch?

-- 
MST