Initial gfx9 debugger node suspend/resume

Change-Id: I2a5dac3d02265c11f5b6985ab457e2d1caa0a033
Signed-off-by: Philip Cox <Philip.Cox@amd.com>


[ROCm/ROCR-Runtime commit: 37858f2311]

projects/rocr-runtime/include/hsakmt.h

@@ -633,14 +633,11 @@ hsaKmtDbgAddressWatch(
     - HSAKMT_STATUS_INVALID_HANDLE if any QueueId is invalid.
 */
 HSAKMT_STATUS
-HSAKMTAPI
-hsaKmtQueueSuspend(
-    HSAuint32   NumQueues,     //IN
-    HSA_QUEUEID QueueId[],     //IN
-    bool        NoGracePeriod, //IN
-    bool        MemFence       //IN
-    );
-
+    HSAKMTAPI
+    hsaKmtNodeSuspend(
+            HSAuint32 Pid,
+            HSAuint32 NodeId,
+            HSAuint32 Flags);
 /**
    Resume the execution of a set of queues. If a queue is not
    suspended by hsaKmtDbgQueueSuspend() then it remains executing.
@@ -658,12 +655,11 @@ hsaKmtQueueSuspend(
     - HSAKMT_STATUS_INVALID_HANDLE if any QueueId is invalid.
 */
 HSAKMT_STATUS
-HSAKMTAPI
-hsaKmtQueueResume(
-    HSAuint32   NumQueues, //IN
-    HSA_QUEUEID QueueId[], //IN
-    bool        MemFence   //IN
-    );
+    HSAKMTAPI
+    hsaKmtNodeResume(
+            HSAuint32 Pid,
+            HSAuint32 NodeId,
+            HSAuint32 Flags);
 
 /**
   Enable debug trap for NodeId. If QueueId is INVALID_QUEUEID then

projects/rocr-runtime/include/hsakmttypes.h

@@ -75,6 +75,12 @@ typedef HSAuint64          HSA_QUEUEID;
 // An HSA_QUEUEID that is never a valid queue ID.
 #define INVALID_QUEUEID 0xFFFFFFFFFFFFFFFFULL
 
+// A PID that is never a valid process ID.
+#define INVALID_PID 0xFFFFFFFF
+
+// // A HSA_NODEID that is never a valid node ID.
+#define INVALID_NODEID 0xFFFFFFFF
+
 // This is included in order to force the alignments to be 4 bytes so that
 // it avoids extra padding added by the compiler when a 64-bit binary is generated.
 #pragma pack(push, hsakmttypes_h, 4)
@@ -785,6 +791,13 @@ typedef enum _HSA_DBG_WAVE_LAUNCH_MODE
     HSA_DBG_WAVE_LAUNCH_MODE_NUM
 } HSA_DBG_WAVE_LAUNCH_MODE;
 
+typedef enum HSA_DBG_NODE_CONTROL {
+      HSA_DBG_NODE_CONTROL_NO_GRACE_PERIOD = 0x01,
+      HSA_DBG_NODE_CONTROL_MEMORY_FENCE = 0x02,
+      HSA_DBG_NODE_CONTROL_UPDATE_CONTEXT = 0x04,
+} HSA_DBG_NODE_CONTROL;
+
+
 //This structure is hardware specific and may change in the future
 typedef struct _HsaDbgWaveMsgAMDGen2
 {

projects/rocr-runtime/include/linux/kfd_ioctl.h

@@ -193,32 +193,62 @@ struct kfd_ioctl_dbg_wave_control_args {
 /* KFD_IOC_DBG_TRAP_ENABLE:
  * data1: 0=disable, 1=enable
  * data2: queue ID (for future use)
+ * data3: unused
+ * data4: unused
  */
 #define KFD_IOC_DBG_TRAP_ENABLE 0
 
 /* KFD_IOC_DBG_TRAP_SET_TRAP_DATA:
  * data1: SPI_GDBG_TRAP_DATA0
  * data2: SPI_GDBG_TRAP_DATA1
+ * data3: unused
+ * data4: unused
  */
 #define KFD_IOC_DBG_TRAP_SET_TRAP_DATA 1
 
 /* KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE:
  * data1: override mode: 0=OR, 1=REPLACE
  * data2: mask
+ * data3: unused
+ * data4: unused
  */
 #define KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE 2
 
 /* KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE:
  * data1: 0=normal, 1=halt, 2=kill, 3=singlestep, 4=disable
  * data2: unused
+ * data3: unused
+ * data4: unused
  */
 #define KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE 3
 
+
+#define KFD__DBG_NODE_SUSPEND_NO_GRACE          0x01
+#define KFD__DBG_NODE_SUSPEND_MEMORY_FENCE      0x02
+#define KFD__DBG_NODE_SUSPEND_UPDATE_CONTEXT    0x04
+/* KFD_IOC_DBG_TRAP_NODE_SUSPEND:
+ * data1: pid
+ * data2: nodeid
+ * data3: flags no_grace=0x01 memory_fence=0x02 update_context=0x04
+ * data4: unused
+ */
+#define KFD_IOC_DBG_TRAP_NODE_SUSPEND 4
+
+/* KFD_IOC_DBG_TRAP_NODE_RESUME:
+ * data1: pid
+ * data2: nodeid
+ * data3: flags no_grace=0x01 memory_fence=0x02 update_context=0x04
+ * data4: unused
+ */
+#define KFD_IOC_DBG_TRAP_NODE_RESUME 5
+
 struct kfd_ioctl_dbg_trap_args {
 	__u32 gpu_id;  /* to KFD */
 	__u32 op;      /* to KFD */
 	__u32 data1;   /* to KFD */
 	__u32 data2;   /* to KFD */
+	__u32 data3;   /* to KFD */
+	__u32 data4;   /* to KFD */
 };
 
 /* Matching HSA_EVENTTYPE */

projects/rocr-runtime/src/debug.c

@@ -27,6 +27,7 @@
 #include "linux/kfd_ioctl.h"
 #include <stdlib.h>
 #include <string.h>
+#include <unistd.h>
 
 static bool *is_device_debugged;
 int debug_get_reg_status(uint32_t node_id, bool *is_debugged);
@@ -271,7 +272,8 @@ int debug_get_reg_status(uint32_t node_id, bool *is_debugged)
 static HSAKMT_STATUS debug_trap(HSAuint32 NodeId,
 				HSAuint32 op,
 				HSAuint32 data1,
-				HSAuint32 data2)
+				HSAuint32 data2,
+				HSAuint32 data3)
 {
 	uint32_t gpu_id;
 	HSAKMT_STATUS result;
@@ -296,6 +298,8 @@ static HSAKMT_STATUS debug_trap(HSAuint32 NodeId,
 	args.op = op;
 	args.data1 = data1;
 	args.data2 = data2;
+	args.data3 = data3;
+	args.data4 = 0;
 
 	long err = kmtIoctl(kfd_fd, AMDKFD_IOC_DBG_TRAP, &args);
 
@@ -313,12 +317,12 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtEnableDebugTrap(HSAuint32   NodeId,
 	if (QueueId != INVALID_QUEUEID)
 		return HSAKMT_STATUS_NOT_SUPPORTED;
 
-	return debug_trap(NodeId, KFD_IOC_DBG_TRAP_ENABLE, 1, QueueId);
+	return debug_trap(NodeId, KFD_IOC_DBG_TRAP_ENABLE, 1, QueueId, 0);
 }
 
 HSAKMT_STATUS HSAKMTAPI hsaKmtDisableDebugTrap(HSAuint32 NodeId)
 {
-	return  debug_trap(NodeId, KFD_IOC_DBG_TRAP_ENABLE, 0, 0);
+	return  debug_trap(NodeId, KFD_IOC_DBG_TRAP_ENABLE, 0, 0, 0);
 }
 
 HSAKMT_STATUS HSAKMTAPI hsaKmtSetDebugTrapData2(HSAuint32 NodeId,
@@ -328,7 +332,8 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtSetDebugTrapData2(HSAuint32 NodeId,
 	return debug_trap(NodeId,
 				KFD_IOC_DBG_TRAP_SET_TRAP_DATA,
 				TrapData0,
-				TrapData1);
+				TrapData1,
+				0);
 }
 
 HSAKMT_STATUS HSAKMTAPI hsaKmtSetWaveLaunchTrapOverride(
@@ -342,7 +347,8 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtSetWaveLaunchTrapOverride(
 	return debug_trap(NodeId,
 				KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_OVERRIDE,
 				TrapOverride,
-				TrapMask);
+				TrapMask,
+				0);
 }
 
 HSAKMT_STATUS HSAKMTAPI hsaKmtSetWaveLaunchMode(
@@ -352,5 +358,52 @@ HSAKMT_STATUS HSAKMTAPI hsaKmtSetWaveLaunchMode(
 	return debug_trap(NodeId,
 				KFD_IOC_DBG_TRAP_SET_WAVE_LAUNCH_MODE,
 				WaveLaunchMode,
+				0,
 				0);
 }
+
+HSAKMT_STATUS HSAKMTAPI hsaKmtNodeSuspend(
+		HSAuint32 Pid,
+		HSAuint32 NodeId,
+		HSAuint32 Flags)
+{
+	pid_t current_pid = INVALID_PID;
+
+	CHECK_KFD_OPEN();
+
+	if (Pid == INVALID_PID) {
+		current_pid = getpid();
+		if (current_pid == 0)
+			return HSAKMT_STATUS_INVALID_HANDLE;
+		Pid = (HSAuint32) current_pid;
+	}
+
+	return debug_trap(NodeId,
+			KFD_IOC_DBG_TRAP_NODE_SUSPEND,
+			Pid,
+			NodeId,
+			Flags);
+}
+
+HSAKMT_STATUS HSAKMTAPI hsaKmtNodeResume(
+		HSAuint32 Pid,
+		HSAuint32 NodeId,
+		HSAuint32 Flags)
+{
+	pid_t current_pid = INVALID_PID;
+
+	CHECK_KFD_OPEN();
+
+	if (Pid == INVALID_PID) {
+		current_pid = getpid();
+		if (current_pid == 0)
+			return HSAKMT_STATUS_INVALID_HANDLE;
+		Pid = (HSAuint32) current_pid;
+	}
+
+	return debug_trap(NodeId,
+			KFD_IOC_DBG_TRAP_NODE_RESUME,
+			Pid,
+			NodeId,
+			Flags);
+}

projects/rocr-runtime/src/libhsakmt.ver

@@ -61,6 +61,8 @@ hsaKmtDisableDebugTrap;
 hsaKmtSetDebugTrapData2;
 hsaKmtSetWaveLaunchTrapOverride;
 hsaKmtSetWaveLaunchMode;
+hsaKmtNodeSuspend;
+hsaKmtNodeResume;
 
 local: *;
 };