diff --git a/sys/arch/aarch64/aarch64/cpufunc_asm_armv8.S b/sys/arch/aarch64/aarch64/cpufunc_asm_armv8.S
index 40efba6dcf46..24be1e4584a2 100644
--- a/sys/arch/aarch64/aarch64/cpufunc_asm_armv8.S
+++ b/sys/arch/aarch64/aarch64/cpufunc_asm_armv8.S
@@ -42,9 +42,14 @@
 
 /*
  * Macro to handle the cache. This takes the start address in x0, length
- * in x1. It will corrupt x2-x5.
+ * in x1. It will corrupt x2-x7.
  */
 .macro cache_handle_range dcop = "", icop = ""
+#ifdef __HAVE_PREEMPTION
+0:	mrs	x6, tpidr_el1		/* x6 := curlwp */
+	ldr	x7, [x6, #L_NCSW]	/* x7 := curlwp->l_ncsw */
+#endif
+
 	mrs	x3, ctr_el0
 	mov	x4, #4			/* size of word */
 .ifnb \dcop
@@ -80,6 +85,16 @@
 	dsb	ish
 	isb
 .endif
+
+#ifdef __HAVE_PREEMPTION
+	ldr	x6, [x6, #L_NCSW]	/* x6 := curlwp->l_ncsw again */
+	cmp	x6, x7			/* retry if l_ncsw changed */
+	b.ne	2f			/* forward branch predict not-taken */
+	ret
+2:	b	0b
+#else
+	ret
+#endif
 .endm
 
 
@@ -97,7 +112,6 @@ END(aarch64_cpuid)
  */
 ENTRY(aarch64_dcache_wb_range)
 	cache_handle_range	dcop = cvac
-	ret
 END(aarch64_dcache_wb_range)
 
 /*
@@ -105,7 +119,6 @@ END(aarch64_dcache_wb_range)
  */
 ENTRY(aarch64_dcache_wbinv_range)
 	cache_handle_range	dcop = civac
-	ret
 END(aarch64_dcache_wbinv_range)
 
 /*
@@ -116,7 +129,6 @@ END(aarch64_dcache_wbinv_range)
  */
 ENTRY(aarch64_dcache_inv_range)
 	cache_handle_range	dcop = ivac
-	ret
 END(aarch64_dcache_inv_range)
 
 /*
@@ -124,7 +136,6 @@ END(aarch64_dcache_inv_range)
  */
 ENTRY(aarch64_idcache_wbinv_range)
 	cache_handle_range	dcop = civac, icop = ivau
-	ret
 END(aarch64_idcache_wbinv_range)
 
 /*
@@ -132,7 +143,6 @@ END(aarch64_idcache_wbinv_range)
  */
 ENTRY(aarch64_icache_sync_range)
 	cache_handle_range	dcop = cvau, icop = ivau
-	ret
 END(aarch64_icache_sync_range)
 
 /*
@@ -140,7 +150,6 @@ END(aarch64_icache_sync_range)
  */
 ENTRY(aarch64_icache_inv_range)
 	cache_handle_range	icop = ivau
-	ret
 END(aarch64_icache_inv_range)
 
 /*