demmt: make pb_pointer_found be per-device

Fixes decoding on CUDA + GL traces

demmt/buffer_decode.c

@@ -30,8 +30,6 @@
 #include "config.h"
 #include "log.h"
 
-static int pb_pointer_found = 0;//TODO: per device
-
 void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, uint32_t len)
 {
 	char pushbuf_desc[1024];
@@ -47,13 +45,14 @@ void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, u
 	if (gpu_addr)
 		gpu_addr -= start;
 
-	if (mapping->fdtype == FDNVIDIA && !pb_pointer_found)
+	if (mapping->fdtype == FDNVIDIA)
 	{
 		struct gpu_object *dev = nvrm_get_device(mapping->object);
 		int chipset = nvrm_get_chipset(dev);
-		int ib_supported = chipset == 0x50 || chipset >= 0x80;
+		bool ib_supported = chipset == 0x50 || chipset >= 0x80;
+		bool pb_pointer_found = nvrm_get_pb_pointer_found(dev);
 
-		if (ib_supported)
+		if (ib_supported && !pb_pointer_found)
 		{
 			uint32_t idx = start / 4;
 			uint32_t *data = (uint32_t *)mapping->data;
@@ -64,12 +63,12 @@ void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, u
 					struct gpu_object *fifo = nvrm_get_fifo(mapping->object, gpu_addr + addr, 1);
 					if (fifo && is_fifo_and_addr_belongs(fifo, gpu_addr + addr))
 					{
-						pb_pointer_found = 1;
+						nvrm_device_set_pb_pointer_found(dev, true);
 						mapping->ib.entries = get_fifo_state(fifo)->ib.entries;
 					}
 				}
 
-				if (!pb_pointer_found)
+				if (!nvrm_get_pb_pointer_found(dev))
 				{
 					uint64_t pb_gpu_addr = (((uint64_t)(data[idx] & 0xff)) << 32) | (data[idx - 1] & 0xfffffffc);
 					struct gpu_object *obj;
@@ -81,17 +80,17 @@ void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, u
 							if (gpu_mapping->address == pb_gpu_addr &&
 								gpu_mapping->length >=  4 * ((data[idx] & 0x7fffffff) >> 10))
 							{
-								pb_pointer_found = 1;
+								nvrm_device_set_pb_pointer_found(dev, true);
 								break;
 							}
 						}
 
-						if (pb_pointer_found)
+						if (nvrm_get_pb_pointer_found(dev))
 							break;
 					}
 				}
 
-				if (pb_pointer_found)
+				if (nvrm_get_pb_pointer_found(dev))
 				{
 					if (info && decode_pb)
 						mmt_printf("IB buffer: %d\n", mapping->id);
@@ -101,7 +100,7 @@ void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, u
 				}
 			}
 		}
-		else
+		else if (!pb_pointer_found)
 		{
 			if (start == 0x40)
 			{
@@ -110,16 +109,16 @@ void buffer_decode_register_write(struct cpu_mapping *mapping, uint32_t start, u
 				{
 					struct gpu_object *fifo = nvrm_get_fifo(mapping->object, gpu_addr + addr, 1);
 					if (fifo && is_fifo_and_addr_belongs(fifo, gpu_addr + addr))
-						pb_pointer_found = 1;
+						nvrm_device_set_pb_pointer_found(dev, true);
 
-					if (!pb_pointer_found)
+					if (!nvrm_get_pb_pointer_found(dev))
 					{
 						struct gpu_mapping *gpu_mapping = gpu_mapping_find(pb_gpu_addr, dev);
 						if (gpu_mapping)
-							pb_pointer_found = 1;
+							nvrm_device_set_pb_pointer_found(dev, true);
 					}
 
-					if (pb_pointer_found)
+					if (nvrm_get_pb_pointer_found(dev))
 					{
 						if (info && decode_pb)
 							mmt_printf("USER buffer: %d\n", mapping->id);

demmt/nvrm.c

@@ -22,6 +22,7 @@
  * OTHER DEALINGS IN THE SOFTWARE.
  */
 
+#include <assert.h>
 #include <inttypes.h>
 #include <stdlib.h>
 #include <string.h>
@@ -39,6 +40,7 @@ struct nvrm_device
 {
 	int chipset;
 	int fifos;
+	bool pb_pointer_found;
 };
 
 static inline struct nvrm_device *nvrm_dev(struct gpu_object *dev)
@@ -249,6 +251,24 @@ int nvrm_get_chipset(struct gpu_object *obj)
 	demmt_abort();
 }
 
+bool nvrm_get_pb_pointer_found(struct gpu_object *obj)
+{
+	struct gpu_object *dev = nvrm_get_device(obj);
+
+	assert(dev);
+	assert(dev->class_data);
+	return nvrm_dev(dev)->pb_pointer_found;
+}
+
+void nvrm_device_set_pb_pointer_found(struct gpu_object *obj, bool pb_pointer_found)
+{
+	struct gpu_object *dev = nvrm_get_device(obj);
+
+	assert(dev);
+	assert(dev->class_data);
+	nvrm_dev(dev)->pb_pointer_found = pb_pointer_found;
+}
+
 static struct gpu_object *nvrm_add_object(uint32_t fd, uint32_t cid, uint32_t parent, uint32_t handle, uint32_t class_)
 {
 	struct gpu_object *obj = gpu_object_add(fd, cid, parent, handle, class_);

demmt/nvrm.h

@@ -1,13 +1,16 @@
 #ifndef DEMMT_NVRM_H
 #define DEMMT_NVRM_H
 
+#include <stdbool.h>
 #include <stdint.h>
 #include "buffer.h"
 #include "mmt_bin_decode_nvidia.h"
 
 struct gpu_object *nvrm_get_device(struct gpu_object *obj);
 int nvrm_get_chipset(struct gpu_object *obj);
 void nvrm_device_set_chipset(struct gpu_object *dev, int chipset);
+bool nvrm_get_pb_pointer_found(struct gpu_object *obj);
+void nvrm_device_set_pb_pointer_found(struct gpu_object *dev, bool found);
 struct gpu_object *nvrm_get_fifo(struct gpu_object *obj, uint64_t gpu_addr, int strict);
 struct gpu_object *nvrm_get_parent_fifo(struct gpu_object *obj);
 int is_fifo_and_addr_belongs(struct gpu_object *obj, uint64_t ctx);