On Sat, 2018-01-13 at 19:51 +0100, Manfred Spraul wrote:
Hi Arnd,
On 01/03/2018 12:15 AM, Arnd Bergmann wrote:
2 ipc/sem.c:377:6: warning: '___p1' may be used uninitialized in this function [-Wmaybe-uninitialized]
This code was last touched in 3.16 by the backport of commit 5864a2fd3088 ("ipc/sem.c: fix complex_count vs. simple op race")
The warning is in "smp_load_acquire(&sma->complex_mode))", and I suspect that commit 27d7be1801a4 ("ipc/sem.c: avoid using spin_unlock_wait()") avoided the warning upstream by removing the smp_mb() before it.
The smp_mb() pairs with spin_unlock_wait() in complexmode_enter() It is removed by commit 27d7be1801a4 ("ipc/sem.c: avoid using spin_unlock_wait()").
From what I see, it doesn't exist in any of the stable kernels (intentionally, the above commit is a rewrite for better performance).
___p1 is from smp_load_acquire()
typeof(*p) ___p1 = READ_ONCE(*p); \I don't see how ___p1 could be used uninitialized. Perhaps a compiler issue?
On arm64 smp_load_acquire() was implemented in assembly that only supported 4-byte and 8-byte words. And complex_mode is a bool (1-byte) field.
So I believe the fix is:
878a84d5a8a1 arm64: add missing data types in smp_load_acquire/smp_store_release
Ben.