diff --git a/arch/x86_64/src/common/x86_64_syscall.c b/arch/x86_64/src/common/x86_64_syscall.c
index 196453f5d6ae8..a3a7f26f448ac 100644
--- a/arch/x86_64/src/common/x86_64_syscall.c
+++ b/arch/x86_64/src/common/x86_64_syscall.c
@@ -39,6 +39,14 @@
 
 #include "x86_64_internal.h"
 
+/****************************************************************************
+ * Pre-processor Definitions
+ ****************************************************************************/
+
+/* Red zone the System V AMD64 ABI reserves below the user stack pointer */
+
+#define X86_64_ABI_RED_ZONE 128
+
 /****************************************************************************
  * Private Types
  ****************************************************************************/
@@ -247,16 +255,24 @@ uint64_t *x86_64_syscall(uint64_t *regs)
             {
               uint64_t usp;
 
-              /* Copy "info" into user stack */
+              /* Save the kernel stack pointer to restore on handler
+               * return
+               */
 
-              usp = rtcb->xcp.saved_ursp - 8;
+              rtcb->xcp.kstkptr = (uintptr_t *)regs[REG_RSP];
 
-              /* Create a frame for info and copy the kernel info */
+              /* Create a 16 byte aligned frame for info below the red
+               * zone of the interrupted user code
+               */
 
-              usp = usp - sizeof(siginfo_t);
+              usp = (rtcb->xcp.saved_ursp - X86_64_ABI_RED_ZONE -
+                     sizeof(siginfo_t)) & ~0x0f;
               memcpy((void *)usp, (void *)regs[REG_RSI], sizeof(siginfo_t));
 
-              /* Now set the updated SP and user copy of "info" to RSI */
+              /* Set the new SP and the user copy of "info" to RSI.
+               * The naked trampoline is entered with SP 16 byte
+               * aligned; its call provides the return address slot.
+               */
 
               regs[REG_RSP] = usp;
               regs[REG_RSI] = usp;
@@ -290,12 +306,21 @@ uint64_t *x86_64_syscall(uint64_t *regs)
           DEBUGASSERT(rtcb->xcp.sigreturn != 0);
 
           regs[REG_RCX]       = rtcb->xcp.sigreturn;
-          regs[REG_RSP]       = rtcb->xcp.saved_rsp;
           rtcb->xcp.sigreturn = 0;
 
-          /* For kernel mode, we should be already on a correct kernel stack
-           * which was recovered in x86_64_syscall_entry.
-           */
+#ifdef CONFIG_ARCH_KERNEL_STACK
+          if (rtcb->xcp.kstack != NULL)
+            {
+              /* Return to the kernel stack saved at dispatch */
+
+              regs[REG_RSP]     = (uint64_t)rtcb->xcp.kstkptr;
+              rtcb->xcp.kstkptr = rtcb->xcp.ktopstk;
+            }
+          else
+#endif
+            {
+              regs[REG_RSP] = rtcb->xcp.saved_rsp;
+            }
 
           break;
         }
diff --git a/arch/x86_64/src/intel64/intel64_schedulesigaction.c b/arch/x86_64/src/intel64/intel64_schedulesigaction.c
index 2bd5cc52e53f7..7b8fae360b5eb 100644
--- a/arch/x86_64/src/intel64/intel64_schedulesigaction.c
+++ b/arch/x86_64/src/intel64/intel64_schedulesigaction.c
@@ -74,6 +74,8 @@
 
 void up_schedule_sigaction(struct tcb_s *tcb)
 {
+  uint64_t sp = tcb->xcp.regs[REG_RSP];
+
   sinfo("tcb=%p, rtcb=%p current_regs=%p\n", tcb,
         this_task(), this_task()->xcp.regs);
 
@@ -83,7 +85,7 @@ void up_schedule_sigaction(struct tcb_s *tcb)
    */
 
   tcb->xcp.saved_rip        = tcb->xcp.regs[REG_RIP];
-  tcb->xcp.saved_rsp        = tcb->xcp.regs[REG_RSP];
+  tcb->xcp.saved_rsp        = sp;
   tcb->xcp.saved_rflags     = tcb->xcp.regs[REG_RFLAGS];
 
   /* Then set up to vector to the trampoline with interrupts
@@ -91,6 +93,21 @@ void up_schedule_sigaction(struct tcb_s *tcb)
    */
 
   tcb->xcp.regs[REG_RIP]    = (uint64_t)x86_64_sigdeliver;
-  tcb->xcp.regs[REG_RSP]    = tcb->xcp.regs[REG_RSP] - 8;
   tcb->xcp.regs[REG_RFLAGS] = 0;
+
+#ifdef CONFIG_ARCH_KERNEL_STACK
+  /* Run the trampoline on the thread kernel stack when the thread was
+   * interrupted in user mode: the signal handler runs on the user
+   * stack and would overwrite the trampoline frame there.
+   */
+
+  if (tcb->xcp.kstack != NULL &&
+      (sp < (uint64_t)tcb->xcp.kstack || sp > (uint64_t)tcb->xcp.ktopstk))
+    {
+      tcb->xcp.saved_ursp = sp;
+      sp                  = (uint64_t)tcb->xcp.ktopstk;
+    }
+#endif
+
+  tcb->xcp.regs[REG_RSP]    = sp - 8;
 }