On Mon, Feb 19, 2018 at 02:01:43PM +0000, Will Deacon wrote:
quoted
The non serializing __clear_bit() was getting "lost"
80543b8e: ld_s r2,[r13,0] <--- (A) Finds PG_locked is set
80543b90: or r3,r2,1 <--- (B) other core unlocks right here
80543b94: st_s r3,[r13,0] <--- (C) sets PG_locked (overwrites unlock)
Ah, so it's problematic for the case where atomics are built using locks.
Got it. I'll err on the side of caution here and have the asm-generic header
(which should be bitops/lock.h not bitops/atomic.h) conditionally define
__clear_bit_unlock as clear_bit_lock unless the architecture has provided
its own implementation.
So I think we get it all right if we use atomic_set_release(). If the
atomics are implemented using locks, atomic_set*() should be implemented
like atomic_xchg() and avoid the above problem.