arm64: spinlock: Ensure forward-progress in spin_unlock_wait

Rather than wait until we observe the lock being free (which might never
happen), we can also return from spin_unlock_wait if we observe that the
lock is now held by somebody else, which implies that it was unlocked
but we just missed seeing it in that state.

Furthermore, in such a scenario there is no longer a need to write back
the value that we loaded, since we know that there has been a lock
hand-off, which is sufficient to publish any stores prior to the
unlock_wait because the ARm architecture ensures that a Store-Release
instruction is multi-copy atomic when observed by a Load-Acquire
instruction.

The litmus test is something like:

AArch64
{
0:X1=x; 0:X3=y;
1:X1=y;
2:X1=y; 2:X3=x;
}
 P0          | P1           | P2           ;
 MOV W0,#1   | MOV W0,#1    | LDAR W0,[X1] ;
 STR W0,[X1] | STLR W0,[X1] | LDR W2,[X3]  ;
 DMB SY      |              |              ;
 LDR W2,[X3] |              |              ;
exists
(0:X2=0 /\ 2:X0=1 /\ 2:X2=0)

where P0 is doing spin_unlock_wait, P1 is doing spin_unlock and P2 is
doing spin_lock.

Signed-off-by: Will Deacon <will.deacon@arm.com>

1 files changed, 24 insertions, 3 deletions

diff --git a/arch/arm64/include/asm/spinlock.h b/arch/arm64/include/asm/spinlock.h
index d5c894253e73..e875a5a551d7 100644
--- a/arch/arm64/include/asm/spinlock.h
+++ b/arch/arm64/include/asm/spinlock.h
@@ -30,20 +30,39 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 {
 	unsigned int tmp;
 	arch_spinlock_t lockval;
+	u32 owner;
 
 	/*
 	 * Ensure prior spin_lock operations to other locks have completed
 	 * on this CPU before we test whether "lock" is locked.
 	 */
 	smp_mb();
+	owner = READ_ONCE(lock->owner) << 16;
 
 	asm volatile(
 "	sevl\n"
 "1:	wfe\n"
 "2:	ldaxr	%w0, %2\n"
+	/* Is the lock free? */
 "	eor	%w1, %w0, %w0, ror #16\n"
-"	cbnz	%w1, 1b\n"
-	/* Serialise against any concurrent lockers */
+"	cbz	%w1, 3f\n"
+	/* Lock taken -- has there been a subsequent unlock->lock transition? */
+"	eor	%w1, %w3, %w0, lsl #16\n"
+"	cbz	%w1, 1b\n"
+	/*
+	 * The owner has been updated, so there was an unlock->lock
+	 * transition that we missed. That means we can rely on the
+	 * store-release of the unlock operation paired with the
+	 * load-acquire of the lock operation to publish any of our
+	 * previous stores to the new lock owner and therefore don't
+	 * need to bother with the writeback below.
+	 */
+"	b	4f\n"
+"3:\n"
+	/*
+	 * Serialise against any concurrent lockers by writing back the
+	 * unlocked lock value
+	 */
 	ARM64_LSE_ATOMIC_INSN(
 	/* LL/SC */
 "	stxr	%w1, %w0, %2\n"
@@ -53,9 +72,11 @@ static inline void arch_spin_unlock_wait(arch_spinlock_t *lock)
 "	mov	%w1, %w0\n"
 "	cas	%w0, %w0, %2\n"
 "	eor	%w1, %w1, %w0\n")
+	/* Somebody else wrote to the lock, GOTO 10 and reload the value */
 "	cbnz	%w1, 2b\n"
+"4:"
 	: "=&r" (lockval), "=&r" (tmp), "+Q" (*lock)
-	:
+	: "r" (owner)
 	: "memory");
 }