drm/v3d: Add DRM_IOCTL_V3D_PERFMON_SET_GLOBAL

Add a new ioctl, DRM_IOCTL_V3D_PERFMON_SET_GLOBAL, to allow configuration of a global performance monitor (perfmon). Use the global perfmon for all jobs to ensure consistent performance tracking across submissions. This feature is needed to implement a Perfetto datasources in user-space.

Christian Gmeiner committed 1 year ago

Commit c6eabba1 parent f3c03be6 files · +85 −3Browse files

Unified · Split

6 files

drivers/gpu/drm/v3d/v3d_drv.c +1 −0
drivers/gpu/drm/v3d/v3d_drv.h +8 −0
drivers/gpu/drm/v3d/v3d_perfmon.c +37 −0
drivers/gpu/drm/v3d/v3d_sched.c +11 −3
drivers/gpu/drm/v3d/v3d_submit.c +10 −0
include/uapi/drm/v3d_drm.h +18 −0

drivers/gpu/drm/v3d/v3d_drv.c+1−0

@@ -224,6 +224,7 @@ static const struct drm_ioctl_desc v3d_drm_ioctls[] = { · 223 unmodified lines
224	224	DRM_IOCTL_DEF_DRV(V3D_PERFMON_GET_VALUES, v3d_perfmon_get_values_ioctl, DRM_RENDER_ALLOW),
225	225	DRM_IOCTL_DEF_DRV(V3D_SUBMIT_CPU, v3d_submit_cpu_ioctl, DRM_RENDER_ALLOW \| DRM_AUTH),
226	226	DRM_IOCTL_DEF_DRV(V3D_PERFMON_GET_COUNTER, v3d_perfmon_get_counter_ioctl, DRM_RENDER_ALLOW),
	227	+ DRM_IOCTL_DEF_DRV(V3D_PERFMON_SET_GLOBAL, v3d_perfmon_set_global_ioctl, DRM_RENDER_ALLOW),
227	228	};
228	229
229	230	static const struct drm_driver v3d_drm_driver = {

drivers/gpu/drm/v3d/v3d_drv.h+8−0

@@ -183,6 +183,12 @@ struct v3d_dev { · 182 unmodified lines
183	183	u32 num_allocated;
184	184	u32 pages_allocated;
185	185	} bo_stats;
	186	+
	187	+ /* To support a performance analysis tool in user space, we require
	188	+ * a single, globally configured performance monitor (perfmon) for
	189	+ * all jobs.
	190	+ */
	191	+ struct v3d_perfmon *global_perfmon;
186	192	};
187	193
188	194	static inline struct v3d_dev *
@@ -594,6 +600,8 @@ int v3d_perfmon_get_values_ioctl(struct drm_device dev, void data, · 405 unmodified lines
594	600	struct drm_file *file_priv);
595	601	int v3d_perfmon_get_counter_ioctl(struct drm_device dev, void data,
596	602	struct drm_file *file_priv);
	603	+int v3d_perfmon_set_global_ioctl(struct drm_device dev, void data,
	604	+ struct drm_file *file_priv);
597	605
598	606	/* v3d_sysfs.c */
599	607	int v3d_sysfs_init(struct device *dev);

drivers/gpu/drm/v3d/v3d_perfmon.c+37−0

@@ -313,6 +313,9 @@ static int v3d_perfmon_idr_del(int id, void elem, void data) · 312 unmodified lines
313	313	if (perfmon == v3d->active_perfmon)
314	314	v3d_perfmon_stop(v3d, perfmon, false);
315	315
	316	+ /* If the global perfmon is being destroyed, set it to NULL */
	317	+ cmpxchg(&v3d->global_perfmon, perfmon, NULL);
	318	+
316	319	v3d_perfmon_put(perfmon);
317	320
318	321	return 0;
@@ -398,6 +401,9 @@ int v3d_perfmon_destroy_ioctl(struct drm_device dev, void data, · 79 unmodified lines
398	401	if (perfmon == v3d->active_perfmon)
399	402	v3d_perfmon_stop(v3d, perfmon, false);
400	403
	404	+ /* If the global perfmon is being destroyed, set it to NULL */
	405	+ cmpxchg(&v3d->global_perfmon, perfmon, NULL);
	406	+
401	407	v3d_perfmon_put(perfmon);
402	408
403	409	return 0;
@@ -457,3 +463,34 @@ int v3d_perfmon_get_counter_ioctl(struct drm_device dev, void data, · 53 unmodified lines
457	463
458	464	return 0;
459	465	}
	466	+
	467	+int v3d_perfmon_set_global_ioctl(struct drm_device dev, void data,
	468	+ struct drm_file *file_priv)
	469	+{
	470	+ struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
	471	+ struct drm_v3d_perfmon_set_global *req = data;
	472	+ struct v3d_dev *v3d = to_v3d_dev(dev);
	473	+ struct v3d_perfmon *perfmon;
	474	+
	475	+ if (req->flags & ~DRM_V3D_PERFMON_CLEAR_GLOBAL)
	476	+ return -EINVAL;
	477	+
	478	+ perfmon = v3d_perfmon_find(v3d_priv, req->id);
	479	+ if (!perfmon)
	480	+ return -EINVAL;
	481	+
	482	+ /* If the request is to clear the global performance monitor */
	483	+ if (req->flags & DRM_V3D_PERFMON_CLEAR_GLOBAL) {
	484	+ if (!v3d->global_perfmon)
	485	+ return -EINVAL;
	486	+
	487	+ xchg(&v3d->global_perfmon, NULL);
	488	+
	489	+ return 0;
	490	+ }
	491	+
	492	+ if (cmpxchg(&v3d->global_perfmon, NULL, perfmon))
	493	+ return -EBUSY;
	494	+
	495	+ return 0;
	496	+}

drivers/gpu/drm/v3d/v3d_sched.c+11−3

@@ -120,11 +120,19 @@ v3d_cpu_job_free(struct drm_sched_job *sched_job) · 119 unmodified lines
120	120	static void
121	121	v3d_switch_perfmon(struct v3d_dev v3d, struct v3d_job job)
122	122	{
123		− if (job->perfmon != v3d->active_perfmon)
	123	+ struct v3d_perfmon *perfmon = v3d->global_perfmon;
	124	+
	125	+ if (!perfmon)
	126	+ perfmon = job->perfmon;
	127	+
	128	+ if (perfmon == v3d->active_perfmon)
	129	+ return;
	130	+
	131	+ if (perfmon != v3d->active_perfmon)
124	132	v3d_perfmon_stop(v3d, v3d->active_perfmon, true);
125	133
126		− if (job->perfmon && v3d->active_perfmon != job->perfmon)
127		− v3d_perfmon_start(v3d, job->perfmon);
	134	+ if (perfmon && v3d->active_perfmon != perfmon)
	135	+ v3d_perfmon_start(v3d, perfmon);
128	136	}
129	137
130	138	static void

drivers/gpu/drm/v3d/v3d_submit.c+10−0

@@ -981,6 +981,11 @@ v3d_submit_cl_ioctl(struct drm_device dev, void data, · 980 unmodified lines
981	981	goto fail;
982	982
983	983	if (args->perfmon_id) {
	984	+ if (v3d->global_perfmon) {
	985	+ ret = -EAGAIN;
	986	+ goto fail_perfmon;
	987	+ }
	988	+
984	989	render->base.perfmon = v3d_perfmon_find(v3d_priv,
985	990	args->perfmon_id);
986	991
@@ -1196,6 +1201,11 @@ v3d_submit_csd_ioctl(struct drm_device dev, void data, · 209 unmodified lines
1196	1201	goto fail;
1197	1202
1198	1203	if (args->perfmon_id) {
	1204	+ if (v3d->global_perfmon) {
	1205	+ ret = -EAGAIN;
	1206	+ goto fail_perfmon;
	1207	+ }
	1208	+
1199	1209	job->base.perfmon = v3d_perfmon_find(v3d_priv,
1200	1210	args->perfmon_id);
1201	1211	if (!job->base.perfmon) {

include/uapi/drm/v3d_drm.h+18−0

@@ -43,6 +43,7 @@ extern "C" { · 42 unmodified lines
43	43	#define DRM_V3D_PERFMON_GET_VALUES 0x0a
44	44	#define DRM_V3D_SUBMIT_CPU 0x0b
45	45	#define DRM_V3D_PERFMON_GET_COUNTER 0x0c
	46	+#define DRM_V3D_PERFMON_SET_GLOBAL 0x0d
46	47
47	48	#define DRM_IOCTL_V3D_SUBMIT_CL DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CL, struct drm_v3d_submit_cl)
48	49	#define DRM_IOCTL_V3D_WAIT_BO DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_WAIT_BO, struct drm_v3d_wait_bo)
@@ -61,6 +62,8 @@ extern "C" { · 12 unmodified lines
61	62	#define DRM_IOCTL_V3D_SUBMIT_CPU DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_SUBMIT_CPU, struct drm_v3d_submit_cpu)
62	63	#define DRM_IOCTL_V3D_PERFMON_GET_COUNTER DRM_IOWR(DRM_COMMAND_BASE + DRM_V3D_PERFMON_GET_COUNTER, \
63	64	struct drm_v3d_perfmon_get_counter)
	65	+#define DRM_IOCTL_V3D_PERFMON_SET_GLOBAL DRM_IOW(DRM_COMMAND_BASE + DRM_V3D_PERFMON_SET_GLOBAL, \
	66	+ struct drm_v3d_perfmon_set_global)
64	67
65	68	#define DRM_V3D_SUBMIT_CL_FLUSH_CACHE 0x01
66	69	#define DRM_V3D_SUBMIT_EXTENSION 0x02
@@ -766,6 +769,21 @@ struct drm_v3d_perfmon_get_counter { · 699 unmodified lines
766	769	__u8 reserved[7];
767	770	};
768	771
	772	+#define DRM_V3D_PERFMON_CLEAR_GLOBAL 0x0001
	773	+
	774	+/**
	775	+ * struct drm_v3d_perfmon_set_global - ioctl to define a global performance
	776	+ * monitor
	777	+ *
	778	+ * The global performance monitor will be used for all jobs. If a global
	779	+ * performance monitor is defined, jobs with a self-defined performance
	780	+ * monitor won't be allowed.
	781	+ */
	782	+struct drm_v3d_perfmon_set_global {
	783	+ __u32 flags;
	784	+ __u32 id;
	785	+};
	786	+
769	787	#if defined(__cplusplus)
770	788	}
771	789	#endif