svn commit: r238828 - head/sys/sys

[Gleb Smirnoff]

On Fri, Jul 27, 2012 at 10:32:55PM +1000, Bruce Evans wrote:
B> > On Fri, Jul 27, 2012 at 02:12:37PM +0300, Konstantin Belousov wrote:
B> > K> On Fri, Jul 27, 2012 at 09:16:48AM +0000, Gleb Smirnoff wrote:
B> > K> > ...
B> > K> > Log:
B> > K> >   Add assertion for refcount overflow.
B> > K> >
B> > K> >   Submitted by:	Andrey Zonov <andrey zonov.org>
B> > K> >   Reviewed by:	kib
B> > K> It was discussed rather then reviewed.
B> > K>
B> > K> I suggest that the assert may be expressed as a check after the increment,
B> > K> which verifies that counter is != 0. This allows to avoid namespace
B> > K> pollution due to limits.h.
B> >
B> > Hmm, overflowing unsigned is a defined behavior in C. If Bruce agrees,
B> > then I'm happy with KASSERT after increment.
B> 
B> Comparing with (uint_t)-1 before is equivalent.  You can even omit the
B> cast (depend on the default promotion).
B> 
B> I just noticed that there is a technical problem -- the count is read
B> unlocked in the KASSERT.  And since the comparision is for equality,
B> if you lose the race reading the count when it reaches the overflow
B> threshold, then you won't see it overflow unless it wraps again and
B> you win the race next time (or later).  atomic_cmpset could be used
B> to clamp the value at the max, but that is too much for an assertion.

We have discussed that. As alternative I proposed:

@@ -50,8 +51,14 @@
 static __inline void
 refcount_acquire(volatile u_int *count)
 {
+#ifdef INVARIANTS
+       u_int old;
+       old = atomic_fetchadd_int(count, 1);
+       KASSERT(old < UINT_MAX, ("refcount %p overflowed", count));
+#else
        atomic_add_acq_int(count, 1);
+#endif
 }

Konstantin didn't like that production code differs from INVARIANTS.

So we ended with what I committed, advocating to the fact that although
assertion is racy and bad panics still can occur, the "good panics"
would occur much more often, and a single "good panic" is enough to
show what's going on.

-- 
Totus tuus, Glebius.