Merge branch 'preempt'

Heiko Carstens says:

====================

The option to select PREEMPT_NONE will go away for all architectures which
support PREEMPT_LAZY [1]. Until now all distributions provide kernels built
with PREEMPT_NONE enabled for s390. In particular this means that all
preempt_disable() / preempt_enable() pairs are optimized away during
compile time.

With PREEMPT_LAZY this is not the case. Switching to PREEMPT_LAZY leads
to a kernel image size increase of ~218kb (defconfig, gcc15).

s390 provides optimized preempt primitives, however there is still room for
improvement. Since support for relocatable lowcore was added access to
preempt_count in lowcore requires an extra call of get_lowcore(), which
generates an extra instruction. Also all instructions have to use a base
register which is not zero to access preempt_count.

Address this by adding a couple of inline assemblies with alternatives.

This generates better code and reduces the size of a PREEMPT_LAZY built
kernel image by ~58kb.

[1] https://lore.kernel.org/all/20251219101502.GB1132199@noisy.programming.kicks-ass.net/

====================

Signed-off-by: Heiko Carstens <hca@linux.ibm.com>

2 files changed, 45 insertions, 4 deletions

diff --git a/arch/s390/include/asm/asm.h b/arch/s390/include/asm/asm.h
index e9062b01e2a2..510901c2a5f9 100644
--- a/arch/s390/include/asm/asm.h
+++ b/arch/s390/include/asm/asm.h
@@ -30,7 +30,7 @@
  */
 #if defined(__GCC_ASM_FLAG_OUTPUTS__) && !(IS_ENABLED(CONFIG_CC_ASM_FLAG_OUTPUT_BROKEN))
 
-#define __HAVE_ASM_FLAG_OUTPUTS__
+#define __HAVE_ASM_FLAG_OUTPUTS__ 1
 
 #define CC_IPM(sym)
 #define CC_OUT(sym, var)	"=@cc" (var)
diff --git a/arch/s390/include/asm/preempt.h b/arch/s390/include/asm/preempt.h
index 6ccd033acfe5..6e5821bb047e 100644
--- a/arch/s390/include/asm/preempt.h
+++ b/arch/s390/include/asm/preempt.h
@@ -8,7 +8,10 @@
 #include <asm/cmpxchg.h>
 #include <asm/march.h>
 
-/* We use the MSB mostly because its available */
+/*
+ * Use MSB so it is possible to read preempt_count with LLGT which
+ * reads the least significant 31 bits with a single instruction.
+ */
 #define PREEMPT_NEED_RESCHED	0x80000000
 
 /*
@@ -23,7 +26,20 @@
  */
 static __always_inline int preempt_count(void)
 {
-	return READ_ONCE(get_lowcore()->preempt_count) & ~PREEMPT_NEED_RESCHED;
+	unsigned long lc_preempt, count;
+
+	BUILD_BUG_ON(sizeof_field(struct lowcore, preempt_count) != sizeof(int));
+	lc_preempt = offsetof(struct lowcore, preempt_count);
+	/* READ_ONCE(get_lowcore()->preempt_count) & ~PREEMPT_NEED_RESCHED */
+	asm_inline(
+		ALTERNATIVE("llgt	%[count],%[offzero](%%r0)\n",
+			    "llgt	%[count],%[offalt](%%r0)\n",
+			    ALT_FEATURE(MFEATURE_LOWCORE))
+		: [count] "=d" (count)
+		: [offzero] "i" (lc_preempt),
+		  [offalt] "i" (lc_preempt + LOWCORE_ALT_ADDRESS),
+		  "m" (((struct lowcore *)0)->preempt_count));
+	return count;
 }
 
 static __always_inline void preempt_count_set(int pc)
@@ -68,7 +84,17 @@ static __always_inline void __preempt_count_add(int val)
 	 */
 	if (!IS_ENABLED(CONFIG_PROFILE_ALL_BRANCHES)) {
 		if (__builtin_constant_p(val) && (val >= -128) && (val <= 127)) {
-			__atomic_add_const(val, &get_lowcore()->preempt_count);
+			unsigned long lc_preempt;
+
+			lc_preempt = offsetof(struct lowcore, preempt_count);
+			asm_inline(
+				ALTERNATIVE("asi	%[offzero](%%r0),%[val]\n",
+					    "asi	%[offalt](%%r0),%[val]\n",
+					    ALT_FEATURE(MFEATURE_LOWCORE))
+				: "+m" (((struct lowcore *)0)->preempt_count)
+				: [offzero] "i" (lc_preempt), [val] "i" (val),
+				  [offalt] "i" (lc_preempt + LOWCORE_ALT_ADDRESS)
+				: "cc");
 			return;
 		}
 	}
@@ -87,7 +113,22 @@ static __always_inline void __preempt_count_sub(int val)
  */
 static __always_inline bool __preempt_count_dec_and_test(void)
 {
+#ifdef __HAVE_ASM_FLAG_OUTPUTS__
+	unsigned long lc_preempt;
+	int cc;
+
+	lc_preempt = offsetof(struct lowcore, preempt_count);
+	asm_inline(
+		ALTERNATIVE("alsi	%[offzero](%%r0),%[val]\n",
+			    "alsi	%[offalt](%%r0),%[val]\n",
+			    ALT_FEATURE(MFEATURE_LOWCORE))
+		: "=@cc" (cc), "+m" (((struct lowcore *)0)->preempt_count)
+		: [offzero] "i" (lc_preempt), [val] "i" (-1),
+		[offalt] "i" (lc_preempt + LOWCORE_ALT_ADDRESS));
+	return (cc == 0) || (cc == 2);
+#else
 	return __atomic_add_const_and_test(-1, &get_lowcore()->preempt_count);
+#endif
 }
 
 /*