s390/spinlock: Implement SPINLOCK_LOCKVAL with inline assembly

Implement SPINLOCK_LOCKVAL with an inline assembly, which makes use of the
ALTERNATIVE macro, to read spinlock_lockval from lowcore. Provide an
alternative instruction with a different offset in case lowcore is
relocated.

This replaces sequences of two instructions with one instruction.

Before:
  10602a:       a7 78 00 00             lhi     %r7,0
  10602e:       a5 8e 00 00             llilh   %r8,0
  106032:       58 d0 83 ac             l       %r13,940(%r8)
  106036:       ba 7d b5 80             cs      %r7,%r13,1408(%r11)

After:
  10602a:       a7 88 00 00             lhi     %r8,0
  10602e:       e3 70 03 ac 00 58       ly      %r7,940
  106034:       ba 87 b5 80             cs      %r8,%r7,1408(%r11)

Kernel image size change:
add/remove: 756/750 grow/shrink: 646/3435 up/down: 30778/-46326 (-15548)

Acked-by: Vasily Gorbik <gor@linux.ibm.com>
Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>

Author: hcahca

arch/s390/include/asm/spinlock.h

@@ -16,7 +16,23 @@
 #include <asm/processor.h>
 #include <asm/alternative.h>
 
-#define SPINLOCK_LOCKVAL (get_lowcore()->spinlock_lockval)
+static __always_inline unsigned int spinlock_lockval(void)
+{
+	unsigned long lc_lockval;
+	unsigned int lockval;
+
+	BUILD_BUG_ON(sizeof_field(struct lowcore, spinlock_lockval) != sizeof(lockval));
+	lc_lockval = offsetof(struct lowcore, spinlock_lockval);
+	asm_inline(
+		ALTERNATIVE("   ly      %[lockval],%[offzero](%%r0)\n",
+			    "   ly      %[lockval],%[offalt](%%r0)\n",
+			    ALT_FEATURE(MFEATURE_LOWCORE))
+		: [lockval] "=d" (lockval)
+		: [offzero] "i" (lc_lockval),
+		  [offalt] "i" (lc_lockval + LOWCORE_ALT_ADDRESS),
+		  "m" (((struct lowcore *)0)->spinlock_lockval));
+	return lockval;
+}
 
 extern int spin_retry;
 
@@ -60,7 +76,7 @@ static inline int arch_spin_trylock_once(arch_spinlock_t *lp)
 	int old = 0;
 
 	barrier();
-	return likely(arch_try_cmpxchg(&lp->lock, &old, SPINLOCK_LOCKVAL));
+	return likely(arch_try_cmpxchg(&lp->lock, &old, spinlock_lockval()));
 }
 
 static inline void arch_spin_lock(arch_spinlock_t *lp)

arch/s390/lib/spinlock.c

@@ -160,7 +160,7 @@ static inline void arch_spin_lock_queued(arch_spinlock_t *lp)
 
 	ix = get_lowcore()->spinlock_index++;
 	barrier();
-	lockval = SPINLOCK_LOCKVAL;	/* cpu + 1 */
+	lockval = spinlock_lockval();	/* cpu + 1 */
 	node = this_cpu_ptr(&spin_wait[ix]);
 	node->prev = node->next = NULL;
 	node_id = node->node_id;
@@ -251,7 +251,7 @@ static inline void arch_spin_lock_classic(arch_spinlock_t *lp)
 {
 	int lockval, old, new, owner, count;
 
-	lockval = SPINLOCK_LOCKVAL;	/* cpu + 1 */
+	lockval = spinlock_lockval();	/* cpu + 1 */
 
 	/* Pass the virtual CPU to the lock holder if it is not running */
 	owner = arch_spin_yield_target(READ_ONCE(lp->lock), NULL);
@@ -290,7 +290,7 @@ EXPORT_SYMBOL(arch_spin_lock_wait);
 
 int arch_spin_trylock_retry(arch_spinlock_t *lp)
 {
-	int cpu = SPINLOCK_LOCKVAL;
+	int cpu = spinlock_lockval();
 	int owner, count;
 
 	for (count = spin_retry; count > 0; count--) {