locking/qspinlock/x86: Micro-optimize virt_spin_lock()

Optimize virt_spin_lock() to use simpler and faster:

atomic_try_cmpxchg(*ptr, &val, new)

instead of:

atomic_cmpxchg(*ptr, val, new) == val

The x86 CMPXCHG instruction returns success in the ZF flag, so
this change saves a compare after the CMPXCHG.

Also optimize retry loop a bit. atomic_try_cmpxchg() fails iff
&lock->val != 0, so there is no need to load and compare the
lock value again - cpu_relax() can be unconditinally called in
this case. This allows us to generate optimized:

1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 63 jne 8d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 75 5d jne 8d <...>
...
8d: f3 90 pause
8f: eb 93 jmp 24 <...>

instead of:

1f: ba 01 00 00 00 mov $0x1,%edx
24: 8b 03 mov (%rbx),%eax
26: 85 c0 test %eax,%eax
28: 75 13 jne 3d <...>
2a: f0 0f b1 13 lock cmpxchg %edx,(%rbx)
2e: 85 c0 test %eax,%eax
30: 75 f2 jne 24 <...>
...
3d: f3 90 pause
3f: eb e3 jmp 24 <...>

Signed-off-by: Uros Bizjak <[e-mail]>
Signed-off-by: Ingo Molnar <[e-mail]>
Cc: Waiman Long <[e-mail]>
Cc: Linus Torvalds <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]

arch/x86/include/asm/qspinlock.h

···
 #define virt_spin_lock virt_spin_lock
 static inline bool virt_spin_lock(struct qspinlock *lock)
 {
+	int val;
+
 	if (!static_branch_likely(&virt_spin_lock_key))
 		return false;
 
···
 	 * horrible lock 'holder' preemption issues.
 	 */
 
-	do {
-		while (atomic_read(&lock->val) != 0)
-			cpu_relax();
-	} while (atomic_cmpxchg(&lock->val, 0, _Q_LOCKED_VAL) != 0);
+ __retry:
+	val = atomic_read(&lock->val);
+
+	if (val || !atomic_try_cmpxchg(&lock->val, &val, _Q_LOCKED_VAL)) {
+		cpu_relax();
+		goto __retry;
+	}
 
 	return true;
 }