Index: mmu.h
===================================================================
--- mmu.h	(revision 6190)
+++ mmu.h	(working copy)
@@ -58,6 +58,10 @@
   uint64_t tf_r10;
   uint64_t tf_r11;
 
+  /* debug registers */
+  uint64_t tf_lbr_to;
+  uint64_t tf_lbr_from;
+
   /* caller-saved registers */
   uint64_t tf_rbx;
   uint64_t tf_rbp;
Index: utrap.h
===================================================================
--- utrap.h	(revision 6190)
+++ utrap.h	(working copy)
@@ -43,6 +43,9 @@
     uint32_t utf_trap_src;
     uint32_t utf_trap_num;
     uint64_t utf_trap_arg;
+
+    uint64_t utf_lbr_from;
+    uint64_t utf_lbr_to;
 };
 
 #endif
Index: locore.S
===================================================================
--- locore.S	(revision 6190)
+++ locore.S	(working copy)
@@ -61,6 +61,13 @@
 trap_ec_entry_stub:
 	// %rsp is pointing at tf_err
 	pushq	%rax		// rax into tf_rax
+
+	pushq	%rcx
+	movq	$MSR_DEBUG_CTL, %rcx
+	xorq	%rax, %rax
+	wrmsr
+	popq	%rcx
+
 	movq	$trap_entry, %rax
 	call	*%rax		// rip into tf__trapentry_rip
 .globl	trap_entry_stub_end
@@ -94,6 +101,10 @@
 trapframe_pop:
 	movq	%rdi,%rsp
 
+	movq	$MSR_DEBUG_CTL, %rcx
+	movq	$1, %rax
+	wrmsr
+
 	movq	tf_rax(%rsp),%rax
 	movq	tf_rbx(%rsp),%rbx
 	movq	tf_rcx(%rsp),%rcx
Index: trap.c
===================================================================
--- trap.c	(revision 6190)
+++ trap.c	(working copy)
@@ -19,7 +19,7 @@
 static int trap_thread_syscall_writeback;
 
 static struct {
-    char trap_entry_code[16] __attribute__ ((aligned (16)));
+    char trap_entry_code[32] __attribute__ ((aligned (32)));
 } trap_entry_stubs[256];
 
 void
@@ -82,6 +82,8 @@
 	    tf->tf_rip, tf->tf_rsp, tf->tf_cs, tf->tf_ss);
     cprintf("rflags %016lx  err %08x\n",
 	    tf->tf_rflags, tf->tf_err);
+    cprintf("LBR FROM %016lx TO %016lx\n",
+	    tf->tf_lbr_from, tf->tf_lbr_to);
 }
 
 static void
@@ -121,6 +123,12 @@
 static void
 trap_dispatch(int trapno, const struct Trapframe *tf)
 {
+    if (trapno == T_DEBUG) {
+	cprintf("thread debug exception: %s\n",
+		trap_thread ? trap_thread->th_ko.ko_name : "null-thread");
+	trapframe_print(tf);
+    }
+
     int64_t r;
 
     if (trapno == T_NMI) {
@@ -169,9 +177,28 @@
     default:
 	r = thread_utrap(trap_thread, UTRAP_SRC_HW, trapno, 0);
 	if (r != 0 && r != -E_RESTART) {
-	    cprintf("Unknown trap %d, cannot utrap: %s.  Trapframe:\n",
-		    trapno, e2s(r));
+	    cprintf("Unknown trap %d, cannot utrap: %s.\n", trapno, e2s(r));
+	    cprintf("Thread %ld (%s), trapframe:\n",
+		    trap_thread->th_ko.ko_id, trap_thread->th_ko.ko_name);
 	    trapframe_print(tf);
+
+	    cprintf("Backtrace:\n");
+	    uint64_t ip = tf->tf_rip;
+	    uint64_t bp = tf->tf_rbp;
+	    for (uint32_t i = 0; i < 8; i++) {
+		cprintf("ip=%lx bp=%lx\n", ip, bp);
+		if (!ip || bp < 0x10000) break;
+		uint64_t *bpp = (uint64_t *) bp;
+		bp = bpp[0];
+		ip = bpp[1];
+	    }
+
+	    if (trapno == T_ILLOP) {
+		uint8_t *ptr = (uint8_t *) tf->tf_rip;
+		for (uint32_t i = 0; i < 16; i++)
+		    cprintf(" %08x", ptr[i]);
+		cprintf("\n");
+	    }
 	    thread_halt(trap_thread);
 	}
     }
@@ -184,6 +211,9 @@
     uint32_t trapno = (trampoline_rip - trap0rip) /
 		      sizeof(trap_entry_stubs[0].trap_entry_code);
 
+    tf->tf_lbr_from = read_msr(MSR_LBR_FROM_IP);
+    tf->tf_lbr_to = read_msr(MSR_LBR_TO_IP);
+
     tf->tf_ds = read_ds();
     tf->tf_es = read_es();
     tf->tf_fs = read_fs();
@@ -276,6 +306,11 @@
     LOAD_SEGMENT_REG(t, gs);
 #undef LOAD_SEGMENT_REG
 
+    write_dr1(0x8700e);
+    write_dr7(DR7_G(1) | DR7_GE |
+	      (DR7_RW_WRITE << DR7_RW_SHIFT(1)) |
+	      (DR7_LEN_8 << DR7_LEN_SHIFT(1)));
+
     trapframe_pop(&t->th_tf);
 }
 
@@ -317,6 +352,7 @@
     UTF_COPY(r8);   UTF_COPY(r9);   UTF_COPY(r10);  UTF_COPY(r11);
     UTF_COPY(r12);  UTF_COPY(r13);  UTF_COPY(r14);  UTF_COPY(r15);
     UTF_COPY(rip);  UTF_COPY(rflags);
+    UTF_COPY(lbr_from); UTF_COPY(lbr_to);
 #undef UTF_COPY
 
     struct UTrapframe *utf = stacktop - sizeof(*utf);