2005-10-28 08:00:13

by John Bowler

Subject: [PATCH] 2.6.14-rc5-mm1 net/ipv4/route.c: spin_unlock call fails to compile

The spin_unlock call which gets an rt_hash_lock_addr(i)
argument fails to compile on rc5-mm1 uniprocessor with
no CONFIG_DEBUG_SPINLOCK because, in mm1, the spin_unlock
macro expands to a statement of the form:

do { (void)&ptr->raw; } while (0)

and 'ptr' is 'NULL'. Since the expression is address-of
(therefore not evaluated as an r-value - i.e. no attempt
is made to load the non-existent raw, just take its address)
this expression is harmless.

The patch makes 'ptr' ((spinlock_t*)NULL), which the compiler
can compile (sizeof *ptr is 0 in the case - the spinlock_t
struct is empty - so the patch is particularly safe.)

The patch applies to both 2.6.14 and 2.6.14-rc5-mm1 (although
only the latter requires it.)

Signed-off-by: John Bowler <[email protected]>

--- linux-2.6.14-rc5/net/ipv4/route.c 2005-10-26 08:37:20.752285410 -0700
+++ patched/net/ipv4/route.c 2005-10-26 12:17:00.761651111 -0700
@@ -231,7 +231,7 @@ static spinlock_t *rt_hash_locks;
spin_lock_init(&rt_hash_locks[i]); \
}
#else
-# define rt_hash_lock_addr(slot) NULL
+# define rt_hash_lock_addr(slot) ((spinlock_t*)NULL)
# define rt_hash_lock_init()
#endif

2005-10-28 09:00:21

by Ingo Molnar

Subject: Re: [PATCH] 2.6.14-rc5-mm1 net/ipv4/route.c: spin_unlock call fails to compile

* John Bowler <[email protected]> wrote:

> #else
> -# define rt_hash_lock_addr(slot) NULL
> +# define rt_hash_lock_addr(slot) ((spinlock_t*)NULL)
> # define rt_hash_lock_init()

looks good to me - it gives (slightly) more type safety too.

Ingo