diff --git a/include/hsakmt.h b/include/hsakmt.h
index cc96e492b6..237f80bbc7 100644
--- a/include/hsakmt.h
+++ b/include/hsakmt.h
@@ -990,6 +990,7 @@ HSAKMT_STATUS
 HSAKMTAPI
 hsaKmtSetAddressWatch(
     HSAuint32          NodeId,        //IN
+    HSAuint32          Pid,           //IN
     HSA_DBG_WATCH_MODE WatchMode,     //IN
     void*              WatchAddress,  //IN
     HSAuint64          WatchAddrMask, //IN
@@ -1013,6 +1014,7 @@ HSAKMT_STATUS
 HSAKMTAPI
 hsaKmtClearAddressWatch(
     HSAuint32 NodeId, //IN
+    HSAuint32 Pid,    //IN
     HSAuint32 WatchId //IN
     );
 
diff --git a/include/hsakmttypes.h b/include/hsakmttypes.h
index 3aeba807ea..9abf4963c5 100644
--- a/include/hsakmttypes.h
+++ b/include/hsakmttypes.h
@@ -223,10 +223,10 @@ typedef union
     HSAuint64 Value;
     struct
     {
-        HSAuint64 WatchAddrMaskLoBit: 6; // Only bits
+        HSAuint64 WatchAddrMaskLoBit: 4; // Only bits
                                         // WatchAddrMaskLoBit..WatchAddrMaskHiBit
                                         // of the
-        HSAuint64 WatchAddrMaskHiBit: 4; // watch address mask are used.
+        HSAuint64 WatchAddrMaskHiBit: 6; // watch address mask are used.
                                          // 0 is the least significant bit.
         HSAuint64 TrapDataCount: 4;      // Number of 32 bit TrapData
                                          // registers supported.
diff --git a/include/linux/kfd_ioctl.h b/include/linux/kfd_ioctl.h
index aa4292e629..6a7dd1f44a 100644
--- a/include/linux/kfd_ioctl.h
+++ b/include/linux/kfd_ioctl.h
@@ -28,8 +28,25 @@
 
 #define KFD_IOCTL_MAJOR_VERSION 1
 #define KFD_IOCTL_MINOR_VERSION 2
-#define KFD_IOCTL_DBG_MAJOR_VERSION	1
-#define KFD_IOCTL_DBG_MINOR_VERSION	0
+
+/*
+ * Debug revision change log
+ *
+ * 0.1 - Initial revision
+ * 0.2 - Fix to include querying pending event that is both trap and vmfault
+ * 1.0 - Removed function to set debug data (renumbering functions broke ABI)
+ * 1.1 - Allow attaching to processes that have not opened /dev/kfd yet
+ * 1.2 - Allow flag option to clear queue status on queue suspend
+ * 1.3 - Fix race condition between clear on suspend and trap event handling
+ * 1.3 - Fix race condition between clear on suspend and trap event handling
+ * 1.4 - Fix bad kfifo free
+ * 1.5 - Fix ABA issue between queue snapshot and suspend
+ * 2.0 - Return number of queues suspended/resumed and mask invalid/error
+ *       array slots
+ * 2.1 - Add Set Address Watch, and Clear Address Watch support.
+ */
+#define KFD_IOCTL_DBG_MAJOR_VERSION	2
+#define KFD_IOCTL_DBG_MINOR_VERSION	1
 
 struct kfd_ioctl_get_version_args {
 	__u32 major_version;	/* from KFD */
@@ -257,6 +274,7 @@ struct kfd_ioctl_dbg_wave_control_args {
  * data2: flags (IN)
  * data3: suspend[2:2], event type [1:0] (OUT)
  */
+
 #define KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT 5
 
 /* KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT:
@@ -268,13 +286,28 @@ struct kfd_ioctl_dbg_wave_control_args {
 #define KFD_IOC_DBG_TRAP_GET_QUEUE_SNAPSHOT 6
 
 /* KFD_IOC_DBG_TRAP_GET_VERSION:
- * prt: unsused
+ * ptr: unsused
  * data1: major version (OUT)
  * data2: minor version (OUT)
  * data3: unused
  */
 #define KFD_IOC_DBG_TRAP_GET_VERSION	7
 
+/* KFD_IOC_DBG_TRAP_CLEAR_ADDRESS_WATCH:
+ * ptr: unused
+ * data1: watch ID
+ * data2: unused
+ * data3: unused
+ */
+#define KFD_IOC_DBG_TRAP_CLEAR_ADDRESS_WATCH 8
+
+/* KFD_IOC_DBG_TRAP_SET_ADDRESS_WATCH:
+ * ptr:   Watch address
+ * data1: Watch ID (OUT)
+ * data2: watch_mode: 0=read, 1=nonread, 2=atomic, 3=all
+ * data3: watch address mask
+ */
+#define KFD_IOC_DBG_TRAP_SET_ADDRESS_WATCH 9
 
 struct kfd_ioctl_dbg_trap_args {
 	__u64 ptr;     /* to KFD -- used for pointer arguments: queue arrays */
diff --git a/src/debug.c b/src/debug.c
index f7d9cbd22d..caddc79a8e 100644
--- a/src/debug.c
+++ b/src/debug.c
@@ -30,6 +30,8 @@
 #include <unistd.h>
 
 static bool *is_device_debugged;
+
+
 int debug_get_reg_status(uint32_t node_id, bool *is_debugged);
 
 HSAKMT_STATUS init_device_debugging_memory(unsigned int NumNodes)
@@ -669,3 +671,63 @@ hsaKmtGetQueueSnapshot(
 
 	return 0;
 }
+
+HSAKMT_STATUS HSAKMTAPI hsaKmtSetAddressWatch(
+		HSAuint32		NodeId,
+		HSAuint32		Pid,
+		HSA_DBG_WATCH_MODE	WatchMode,
+		void			*WatchAddress,
+		HSAuint64		WatchAddrMask,
+		HSAuint32		*WatchId
+		)
+{
+
+	HSAKMT_STATUS result;
+	HSAuint32     TruncatedWatchAddressMask;
+	struct kfd_ioctl_dbg_trap_args argout = {0};
+
+	/* Right now we only support 32 bit watch address masks, so we need
+	 * to check that we aren't losing data when we truncate the mask
+	 * to be passed to the kernel.
+	 */
+	if (WatchAddrMask > (HSAuint64) UINT_MAX)
+	{
+		return HSAKMT_STATUS_INVALID_PARAMETER;
+	}
+	TruncatedWatchAddressMask = (HSAuint32) WatchAddrMask;
+
+	if (WatchId == NULL)
+		return HSAKMT_STATUS_INVALID_PARAMETER;
+
+	result = debug_trap(NodeId,
+			KFD_IOC_DBG_TRAP_SET_ADDRESS_WATCH, // op
+			*WatchId,
+			WatchMode,
+			TruncatedWatchAddressMask,
+			Pid,
+			(HSAuint64) WatchAddress,
+			&argout);
+	*WatchId = argout.data1;
+
+	return result;
+}
+
+HSAKMT_STATUS HSAKMTAPI hsaKmtClearAddressWatch(
+		HSAuint32 NodeId,
+		HSAuint32 Pid,
+		HSAuint32 WatchId
+		)
+{
+
+	HSAKMT_STATUS result;
+
+	result = debug_trap(NodeId,
+			KFD_IOC_DBG_TRAP_CLEAR_ADDRESS_WATCH, // op
+			WatchId,
+			0,
+			0,
+			Pid,
+			0,
+			NULL);
+	return result;
+}
diff --git a/src/libhsakmt.ver b/src/libhsakmt.ver
index fc9564f3ec..d47f68d2d0 100644
--- a/src/libhsakmt.ver
+++ b/src/libhsakmt.ver
@@ -69,6 +69,8 @@ hsaKmtQueueResume;
 hsaKmtAllocQueueGWS;
 hsaKmtGetKernelDebugTrapVersionInfo;
 hsaKmtGetThunkDebugTrapVersionInfo;
+hsaKmtSetAddressWatch;
+hsaKmtClearAddressWatch;
 
 local: *;
 };