202 static int perf_tracepoint_open(struct gpu_perf *gp, int tp_id,  in perf_tracepoint_open()  argument
222 	n = gp->nr_cpus * (gp->nr_events+1);  in perf_tracepoint_open()
223 	fd = realloc(gp->fd, n*sizeof(int));  in perf_tracepoint_open()
224 	sample = realloc(gp->sample, n*sizeof(*gp->sample));  in perf_tracepoint_open()
227 	gp->fd = fd;  in perf_tracepoint_open()
228 	gp->sample = sample;  in perf_tracepoint_open()
230 	fd += gp->nr_events * gp->nr_cpus;  in perf_tracepoint_open()
231 	sample += gp->nr_events * gp->nr_cpus;  in perf_tracepoint_open()
232 	for (n = 0; n < gp->nr_cpus; n++) {  in perf_tracepoint_open()
246 	gp->nr_events++;  in perf_tracepoint_open()
250 static int perf_mmap(struct gpu_perf *gp)  in perf_mmap()  argument
252 	int size = (1 + N_PAGES) * gp->page_size;  in perf_mmap()
255 	gp->map = malloc(sizeof(void *)*gp->nr_cpus);  in perf_mmap()
256 	if (gp->map == NULL)  in perf_mmap()
259 	fd = gp->fd;  in perf_mmap()
260 	for (j = 0; j < gp->nr_cpus; j++) {  in perf_mmap()
261 		gp->map[j] = mmap(NULL, size, PROT_READ | PROT_WRITE, MAP_SHARED, *fd++, 0);  in perf_mmap()
262 		if (gp->map[j] == (void *)-1)  in perf_mmap()
266 	for (i = 1; i < gp->nr_events; i++) {  in perf_mmap()
267 		for (j = 0; j < gp->nr_cpus; j++)  in perf_mmap()
268 			ioctl(*fd++, PERF_EVENT_IOC_SET_OUTPUT, gp->fd[j]);  in perf_mmap()
275 		munmap(gp->map[j], size);  in perf_mmap()
276 	free(gp->map);  in perf_mmap()
277 	gp->map = NULL;  in perf_mmap()
302 lookup_comm(struct gpu_perf *gp, pid_t pid)  in lookup_comm()  argument
309 	for (comm = gp->comm; comm != NULL; comm = comm->next) {  in lookup_comm()
324 		comm->next = gp->comm;  in lookup_comm()
325 		gp->comm = comm;  in lookup_comm()
331 static int request_add(struct gpu_perf *gp, const void *event)  in request_add()  argument
336 	comm = lookup_comm(gp, sample->pid);  in request_add()
344 static int flip_complete(struct gpu_perf *gp, const void *event)  in flip_complete()  argument
348 	gp->flip_complete[READ_TP_FIELD_U32(sample, TP_FLIP_COMPLETE, plane)]++;  in flip_complete()
352 static int ctx_switch(struct gpu_perf *gp, const void *event)  in ctx_switch()  argument
356 	gp->ctx_switch[GET_RING_ID(sample, TP_GEM_RING_SWITCH_CONTEXT)]++;  in ctx_switch()
360 static int ring_sync(struct gpu_perf *gp, const void *event)  in ring_sync()  argument
365 	comm = lookup_comm(gp, sample->pid);  in ring_sync()
373 static int wait_begin(struct gpu_perf *gp, const void *event)  in wait_begin()  argument
379 	comm = lookup_comm(gp, sample->pid);  in wait_begin()
394 	wait->next = gp->wait[GET_RING_ID(sample, TP_GEM_REQUEST_WAIT_BEGIN)];  in wait_begin()
395 	gp->wait[GET_RING_ID(sample, TP_GEM_REQUEST_WAIT_BEGIN)] = wait;  in wait_begin()
400 static int wait_end(struct gpu_perf *gp, const void *event)  in wait_end()  argument
408 	for (prev = &gp->wait[engine]; (wait = *prev) != NULL; prev = &wait->next) {  in wait_end()
423 void gpu_perf_init(struct gpu_perf *gp, unsigned flags)  in gpu_perf_init()  argument
425 	memset(gp, 0, sizeof(*gp));  in gpu_perf_init()
426 	gp->nr_cpus = sysconf(_SC_NPROCESSORS_ONLN);  in gpu_perf_init()
427 	gp->page_size = getpagesize();  in gpu_perf_init()
429 	perf_tracepoint_open(gp, TP_GEM_REQUEST_ADD, request_add);  in gpu_perf_init()
430 	if (perf_tracepoint_open(gp, TP_GEM_REQUEST_WAIT_BEGIN, wait_begin) == 0)  in gpu_perf_init()
431 		perf_tracepoint_open(gp, TP_GEM_REQUEST_WAIT_END, wait_end);  in gpu_perf_init()
432 	perf_tracepoint_open(gp, TP_FLIP_COMPLETE, flip_complete);  in gpu_perf_init()
433 	perf_tracepoint_open(gp, TP_GEM_RING_SYNC_TO, ring_sync);  in gpu_perf_init()
434 	perf_tracepoint_open(gp, TP_GEM_RING_SWITCH_CONTEXT, ctx_switch);  in gpu_perf_init()
436 	if (gp->nr_events == 0) {  in gpu_perf_init()
437 		gp->error = "i915.ko tracepoints not available";  in gpu_perf_init()
441 	if (perf_mmap(gp))  in gpu_perf_init()
445 static int process_sample(struct gpu_perf *gp, int cpu,  in process_sample()  argument
452 	for (n = 0; n < gp->nr_events; n++) {  in process_sample()
453 		int m = n * gp->nr_cpus + cpu;  in process_sample()
454 		if (gp->sample[m].id != sample->id)  in process_sample()
457 		update = gp->sample[m].func(gp, sample);  in process_sample()
464 int gpu_perf_update(struct gpu_perf *gp)  in gpu_perf_update()  argument
466 	const int size = N_PAGES * gp->page_size;  in gpu_perf_update()
472 	if (gp->map == NULL)  in gpu_perf_update()
475 	for (n = 0; n < gp->nr_cpus; n++) {  in gpu_perf_update()
476 		struct perf_event_mmap_page *mmap = gp->map[n];  in gpu_perf_update()
492 		data = (uint8_t *)mmap + gp->page_size;  in gpu_perf_update()
522 				update += process_sample(gp, n, header);  in gpu_perf_update()