alinux: arm64: adjust tk_core memory layout

to #29722367

On some specific hardware with 128 bytes LLC cacheline, tk_core may
cause false sharing problem. We can align it to 128 bytes so that
it won't be affected by other global variables.

This change will make a bit waste on cache utilization but get good
number of performance improvement. So for both 64 and 128 bytes aligned
LLC cacheline, we adjust tk_core memory layout to avoid potential cacheline
contention.

Signed-off-by: Peng Wang <rocking@linux.alibaba.com>
Acked-by: Shanpei Chen <shanpeic@linux.alibaba.com>
Reviewed-by: Shile Zhang <shile.zhang@linux.alibaba.com>

Author: Peng Wang

arch/arm64/Kconfig

@@ -865,6 +865,16 @@ config ARCH_WANT_HUGE_PMD_SHARE
 config ARCH_HAS_CACHE_LINE_SIZE
 	def_bool y
 
+config ARCH_LLC_128_WORKAROUND
+	bool "Workaround for 128 bytes LLC cacheline"
+	depends on ARM64
+	default n
+	help
+	  LLC cacheline size may be up to 128 bytes, and this
+	  is useful if you want to do workaround on such
+	  case. It can be used to align memory address to get
+	  good cache utilization et al.
+
 config SECCOMP
 	bool "Enable seccomp to safely compute untrusted bytecode"
 	---help---

configs/config-4.19.y-aarch64

@@ -435,6 +435,7 @@ CONFIG_HW_PERF_EVENTS=y
 CONFIG_SYS_SUPPORTS_HUGETLBFS=y
 CONFIG_ARCH_WANT_HUGE_PMD_SHARE=y
 CONFIG_ARCH_HAS_CACHE_LINE_SIZE=y
+CONFIG_ARCH_LLC_128_WORKAROUND=y
 CONFIG_SECCOMP=y
 CONFIG_PARAVIRT=y
 CONFIG_PARAVIRT_TIME_ACCOUNTING=y

configs/config-4.19.y-aarch64-debug

@@ -436,6 +436,7 @@ CONFIG_HW_PERF_EVENTS=y
 CONFIG_SYS_SUPPORTS_HUGETLBFS=y
 CONFIG_ARCH_WANT_HUGE_PMD_SHARE=y
 CONFIG_ARCH_HAS_CACHE_LINE_SIZE=y
+CONFIG_ARCH_LLC_128_WORKAROUND=y
 CONFIG_SECCOMP=y
 CONFIG_PARAVIRT=y
 CONFIG_PARAVIRT_TIME_ACCOUNTING=y

kernel/time/timekeeping.c

@@ -48,9 +48,24 @@ enum timekeeping_adv_mode {
  * cache line.
  */
 static struct {
+#ifdef CONFIG_ARCH_LLC_128_WORKAROUND
+	/* Start seq on the middle of 128 bytes aligned address to
+	 * keep some members of tk_core in the same 64 bytes for
+	 * principle of locality while pushing others to another LLC
+	 * cacheline to avoid false sharing.
+	 */
+	u8 padding1[64];
+	seqcount_t		seq;
+	/* Push some timekeeper memebers to another LLC cacheline */
+	u8 padding2[16];
+	struct timekeeper	timekeeper;
+	/* For 128 bytes LLC cacheline */
+} tk_core __aligned(128) = {
+#else
 	seqcount_t		seq;
 	struct timekeeper	timekeeper;
 } tk_core ____cacheline_aligned = {
+#endif
 	.seq = SEQCNT_ZERO(tk_core.seq),
 };