diff --git a/sys/i386/i386/machdep.c b/sys/i386/i386/machdep.c
index 89e67a54e5b..ccaeb5e361a 100644
--- a/sys/i386/i386/machdep.c
+++ b/sys/i386/i386/machdep.c
@@ -2636,6 +2636,16 @@ init386(first)
 	dblfault_tss.tss_cs = GSEL(GCODE_SEL, SEL_KPL);
 	dblfault_tss.tss_ldt = GSEL(GLDT_SEL, SEL_KPL);
 
+	/* Initialize the tss (except for the final esp0) early for vm86. */
+	PCPU_SET(common_tss.tss_esp0, thread0.td_kstack +
+	    thread0.td_kstack_pages * PAGE_SIZE - 16);
+	PCPU_SET(common_tss.tss_ss0, GSEL(GDATA_SEL, SEL_KPL));
+	gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
+	PCPU_SET(tss_gdt, &gdt[GPROC0_SEL].sd);
+	PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
+	PCPU_SET(common_tss.tss_ioopt, (sizeof (struct i386tss)) << 16);
+	ltr(gsel_tss);
+
 	vm86_initialize();
 	getmemsize(first);
 	init_param2(physmem);
@@ -2701,14 +2711,10 @@ init386(first)
 	}
 #endif
 	PCPU_SET(curpcb, thread0.td_pcb);
-	/* make an initial tss so cpu can get interrupt stack on syscall! */
+	/* Move esp0 in the tss to its final place. */
 	/* Note: -16 is so we can grow the trapframe if we came from vm86 */
 	PCPU_SET(common_tss.tss_esp0, (vm_offset_t)thread0.td_pcb - 16);
-	PCPU_SET(common_tss.tss_ss0, GSEL(GDATA_SEL, SEL_KPL));
-	gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
-	PCPU_SET(tss_gdt, &gdt[GPROC0_SEL].sd);
-	PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
-	PCPU_SET(common_tss.tss_ioopt, (sizeof (struct i386tss)) << 16);
+	gdt[GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;	/* clear busy bit */
 	ltr(gsel_tss);
 
 	/* make a call gate to reenter kernel with */