[PATCH v2 6/9] refcount: Fix refcount_dec_not_one()

[Peter Zijlstra]

refcount_dec_not_one() misbehaves when racing against the other
refcount ops, notably it directly compares against REFCOUNT_SATURATED,
even though, through aforementioned races, the value might be slightly
off.

Additionally, when refcount_dec_not_one() is used to decrement zero;
which is valid, because 0 is not one, then it behaves differently from
the other refcount_dec*() functions in that it doesn't call
refcount_warn_saturate().

In order to allow some fuzz around the corners of the positive range,
test for saturation slightly differently.

Signed-off-by: Peter Zijlstra (Intel) <peterz@xxxxxxxxxxxxx>
---
 lib/refcount.c |    5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

--- a/lib/refcount.c
+++ b/lib/refcount.c
@@ -81,7 +81,8 @@ bool refcount_dec_not_one(refcount_t *r)
 	unsigned int new, val = atomic_read(&r->refs);
 
 	do {
-		if (unlikely(val == REFCOUNT_SATURATED))
+		/* SAT+SAT/2 < val < SAT-SAT/2 */
+		if (unlikely(val - (REFCOUNT_SATURATED + REFCOUNT_SATURATED/2) < -REFCOUNT_SATURATED))
 			return true;
 
 		if (val == 1)
@@ -89,7 +90,7 @@ bool refcount_dec_not_one(refcount_t *r)
 
 		new = val - 1;
 		if (new > val) {
-			WARN_ONCE(new > val, "refcount_t: underflow; use-after-free.\n");
+			refcount_warn_saturate(r, REFCOUNT_SUB_UAF);
 			return true;
 		}