> 
> So ho-hum.  As a first step, David, could you please take a look into
> what's up with downgrade_write()?

Yes. Patch attached. The critical bit is in the hunk beginning at line 95.

It appears my rwsem testsuite didn't have a downgrade write test. I added one
and the bug became obvious.

> (Then again, we need to have a serious think about the overflow bug.  It's
> fatal.  Should we fix it?  If so, the current rwsem implementation is
> probably unsalvageable).

The optimised rwsem implementation is probably okay, provided you don't have a
32-bit machine with silly amounts of memory, for which the 32767 process limit
is probably reasonable. Maybe make it contingent on CONFIG_HIGHMEM on X86?

If you don't want to use an optimised version, there _is_ a spinlock version
as well.

For 64-bit archs like x86_64 and ppc64 this algorithm can easily made to use a
64-bit counter (lib/rwsem.c wants a signed long counter), so it's just a
matter of using XADDQ, LDARX/STDCX or equivalents (I have patches for those
two). I've also got MIPS64 patches somewhere too, though they're horribly out
of date.

Alpha and S390 already do the 64-bit counter thing. And if sparc64 has a
64-bit CAS, that can be used. IA64's rwsems should be 64-bitable too.

Using a 64-bit counter gives you a limit of 2 billion processes.

David