drm/vc4: Expose performance counters to userspace

The V3D engine has various hardware counters which might be interesting
to userspace performance analysis tools.

Expose new ioctls to create/destroy a performance monitor object and
query the counter values of this perfmance monitor.

Note that a perfomance monitor is given an ID that is only valid on the
file descriptor it has been allocated from. A performance monitor can be
attached to a CL submission and the driver will enable HW counters for
this request and update the performance monitor values at the end of the
job.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Eric Anholt <eric@anholt.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20180112090926.12538-1-boris.brezillon@free-electrons.com

9 files changed, 474 insertions, 72 deletions

diff --git a/drivers/gpu/drm/vc4/Makefile b/drivers/gpu/drm/vc4/Makefile
index f5500df51686..4a3a868235f8 100644
--- a/drivers/gpu/drm/vc4/Makefile
+++ b/drivers/gpu/drm/vc4/Makefile
@@ -15,6 +15,7 @@ vc4-y := \
 	vc4_vec.o \
 	vc4_hvs.o \
 	vc4_irq.o \
+	vc4_perfmon.o \
 	vc4_plane.o \
 	vc4_render_cl.o \
 	vc4_trace_points.o \
diff --git a/drivers/gpu/drm/vc4/vc4_drv.c b/drivers/gpu/drm/vc4/vc4_drv.c
index ceb385fd69c5..94b99c90425a 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.c
+++ b/drivers/gpu/drm/vc4/vc4_drv.c
@@ -101,6 +101,7 @@ static int vc4_get_param_ioctl(struct drm_device *dev, void *data,
 	case DRM_VC4_PARAM_SUPPORTS_THREADED_FS:
 	case DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER:
 	case DRM_VC4_PARAM_SUPPORTS_MADVISE:
+	case DRM_VC4_PARAM_SUPPORTS_PERFMON:
 		args->value = true;
 		break;
 	default:
@@ -111,6 +112,26 @@ static int vc4_get_param_ioctl(struct drm_device *dev, void *data,
 	return 0;
 }
 
+static int vc4_open(struct drm_device *dev, struct drm_file *file)
+{
+	struct vc4_file *vc4file;
+
+	vc4file = kzalloc(sizeof(*vc4file), GFP_KERNEL);
+	if (!vc4file)
+		return -ENOMEM;
+
+	vc4_perfmon_open_file(vc4file);
+	file->driver_priv = vc4file;
+	return 0;
+}
+
+static void vc4_close(struct drm_device *dev, struct drm_file *file)
+{
+	struct vc4_file *vc4file = file->driver_priv;
+
+	vc4_perfmon_close_file(vc4file);
+}
+
 static const struct vm_operations_struct vc4_vm_ops = {
 	.fault = vc4_fault,
 	.open = drm_gem_vm_open,
@@ -143,6 +164,9 @@ static const struct drm_ioctl_desc vc4_drm_ioctls[] = {
 	DRM_IOCTL_DEF_DRV(VC4_GET_TILING, vc4_get_tiling_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(VC4_LABEL_BO, vc4_label_bo_ioctl, DRM_RENDER_ALLOW),
 	DRM_IOCTL_DEF_DRV(VC4_GEM_MADVISE, vc4_gem_madvise_ioctl, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(VC4_PERFMON_CREATE, vc4_perfmon_create_ioctl, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(VC4_PERFMON_DESTROY, vc4_perfmon_destroy_ioctl, DRM_RENDER_ALLOW),
+	DRM_IOCTL_DEF_DRV(VC4_PERFMON_GET_VALUES, vc4_perfmon_get_values_ioctl, DRM_RENDER_ALLOW),
 };
 
 static struct drm_driver vc4_drm_driver = {
@@ -153,6 +177,8 @@ static struct drm_driver vc4_drm_driver = {
 			    DRIVER_RENDER |
 			    DRIVER_PRIME),
 	.lastclose = drm_fb_helper_lastclose,
+	.open = vc4_open,
+	.postclose = vc4_close,
 	.irq_handler = vc4_irq,
 	.irq_preinstall = vc4_irq_preinstall,
 	.irq_postinstall = vc4_irq_postinstall,
diff --git a/drivers/gpu/drm/vc4/vc4_drv.h b/drivers/gpu/drm/vc4/vc4_drv.h
index 3af22936d9b3..fefa1664a9f5 100644
--- a/drivers/gpu/drm/vc4/vc4_drv.h
+++ b/drivers/gpu/drm/vc4/vc4_drv.h
@@ -11,6 +11,8 @@
 #include <drm/drm_encoder.h>
 #include <drm/drm_gem_cma_helper.h>
 
+#include "uapi/drm/vc4_drm.h"
+
 /* Don't forget to update vc4_bo.c: bo_type_names[] when adding to
  * this.
  */
@@ -29,6 +31,36 @@ enum vc4_kernel_bo_type {
 	VC4_BO_TYPE_COUNT
 };
 
+/* Performance monitor object. The perform lifetime is controlled by userspace
+ * using perfmon related ioctls. A perfmon can be attached to a submit_cl
+ * request, and when this is the case, HW perf counters will be activated just
+ * before the submit_cl is submitted to the GPU and disabled when the job is
+ * done. This way, only events related to a specific job will be counted.
+ */
+struct vc4_perfmon {
+	/* Tracks the number of users of the perfmon, when this counter reaches
+	 * zero the perfmon is destroyed.
+	 */
+	refcount_t refcnt;
+
+	/* Number of counters activated in this perfmon instance
+	 * (should be less than DRM_VC4_MAX_PERF_COUNTERS).
+	 */
+	u8 ncounters;
+
+	/* Events counted by the HW perf counters. */
+	u8 events[DRM_VC4_MAX_PERF_COUNTERS];
+
+	/* Storage for counter values. Counters are incremented by the HW
+	 * perf counter values every time the perfmon is attached to a GPU job.
+	 * This way, perfmon users don't have to retrieve the results after
+	 * each job if they want to track events covering several submissions.
+	 * Note that counter values can't be reset, but you can fake a reset by
+	 * destroying the perfmon and creating a new one.
+	 */
+	u64 counters[0];
+};
+
 struct vc4_dev {
 	struct drm_device *dev;
 
@@ -121,6 +153,11 @@ struct vc4_dev {
 	wait_queue_head_t job_wait_queue;
 	struct work_struct job_done_work;
 
+	/* Used to track the active perfmon if any. Access to this field is
+	 * protected by job_lock.
+	 */
+	struct vc4_perfmon *active_perfmon;
+
 	/* List of struct vc4_seqno_cb for callbacks to be made from a
 	 * workqueue when the given seqno is passed.
 	 */
@@ -406,6 +443,21 @@ struct vc4_exec_info {
 	void *uniforms_v;
 	uint32_t uniforms_p;
 	uint32_t uniforms_size;
+
+	/* Pointer to a performance monitor object if the user requested it,
+	 * NULL otherwise.
+	 */
+	struct vc4_perfmon *perfmon;
+};
+
+/* Per-open file private data. Any driver-specific resource that has to be
+ * released when the DRM file is closed should be placed here.
+ */
+struct vc4_file {
+	struct {
+		struct idr idr;
+		struct mutex lock;
+	} perfmon;
 };
 
 static inline struct vc4_exec_info *
@@ -646,3 +698,19 @@ bool vc4_check_tex_size(struct vc4_exec_info *exec,
 /* vc4_validate_shader.c */
 struct vc4_validated_shader_info *
 vc4_validate_shader(struct drm_gem_cma_object *shader_obj);
+
+/* vc4_perfmon.c */
+void vc4_perfmon_get(struct vc4_perfmon *perfmon);
+void vc4_perfmon_put(struct vc4_perfmon *perfmon);
+void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon);
+void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
+		      bool capture);
+struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id);
+void vc4_perfmon_open_file(struct vc4_file *vc4file);
+void vc4_perfmon_close_file(struct vc4_file *vc4file);
+int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
+			     struct drm_file *file_priv);
+int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *file_priv);
+int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
+				 struct drm_file *file_priv);
diff --git a/drivers/gpu/drm/vc4/vc4_gem.c b/drivers/gpu/drm/vc4/vc4_gem.c
index 638540943c61..3ac801b14d4e 100644
--- a/drivers/gpu/drm/vc4/vc4_gem.c
+++ b/drivers/gpu/drm/vc4/vc4_gem.c
@@ -454,14 +454,30 @@ again:
 
 	vc4_flush_caches(dev);
 
+	/* Only start the perfmon if it was not already started by a previous
+	 * job.
+	 */
+	if (exec->perfmon && vc4->active_perfmon != exec->perfmon)
+		vc4_perfmon_start(vc4, exec->perfmon);
+
 	/* Either put the job in the binner if it uses the binner, or
 	 * immediately move it to the to-be-rendered queue.
 	 */
 	if (exec->ct0ca != exec->ct0ea) {
 		submit_cl(dev, 0, exec->ct0ca, exec->ct0ea);
 	} else {
+		struct vc4_exec_info *next;
+
 		vc4_move_job_to_render(dev, exec);
-		goto again;
+		next = vc4_first_bin_job(vc4);
+
+		/* We can't start the next bin job if the previous job had a
+		 * different perfmon instance attached to it. The same goes
+		 * if one of them had a perfmon attached to it and the other
+		 * one doesn't.
+		 */
+		if (next && next->perfmon == exec->perfmon)
+			goto again;
 	}
 }
 
@@ -621,6 +637,7 @@ vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
 		 struct ww_acquire_ctx *acquire_ctx)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_exec_info *renderjob;
 	uint64_t seqno;
 	unsigned long irqflags;
 	struct vc4_fence *fence;
@@ -646,11 +663,14 @@ vc4_queue_submit(struct drm_device *dev, struct vc4_exec_info *exec,
 
 	list_add_tail(&exec->head, &vc4->bin_job_list);
 
-	/* If no job was executing, kick ours off.  Otherwise, it'll
-	 * get started when the previous job's flush done interrupt
-	 * occurs.
+	/* If no bin job was executing and if the render job (if any) has the
+	 * same perfmon as our job attached to it (or if both jobs don't have
+	 * perfmon activated), then kick ours off.  Otherwise, it'll get
+	 * started when the previous job's flush/render done interrupt occurs.
 	 */
-	if (vc4_first_bin_job(vc4) == exec) {
+	renderjob = vc4_first_render_job(vc4);
+	if (vc4_first_bin_job(vc4) == exec &&
+	    (!renderjob || renderjob->perfmon == exec->perfmon)) {
 		vc4_submit_next_bin_job(dev);
 		vc4_queue_hangcheck(dev);
 	}
@@ -915,6 +935,9 @@ vc4_complete_exec(struct drm_device *dev, struct vc4_exec_info *exec)
 	vc4->bin_alloc_used &= ~exec->bin_slots;
 	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 
+	/* Release the reference we had on the perf monitor. */
+	vc4_perfmon_put(exec->perfmon);
+
 	mutex_lock(&vc4->power_lock);
 	if (--vc4->power_refcount == 0) {
 		pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
@@ -1067,6 +1090,7 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 		    struct drm_file *file_priv)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
+	struct vc4_file *vc4file = file_priv->driver_priv;
 	struct drm_vc4_submit_cl *args = data;
 	struct vc4_exec_info *exec;
 	struct ww_acquire_ctx acquire_ctx;
@@ -1080,6 +1104,11 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 		return -EINVAL;
 	}
 
+	if (args->pad2 != 0) {
+		DRM_DEBUG("->pad2 must be set to zero\n");
+		return -EINVAL;
+	}
+
 	exec = kcalloc(1, sizeof(*exec), GFP_KERNEL);
 	if (!exec) {
 		DRM_ERROR("malloc failure on exec struct\n");
@@ -1105,6 +1134,15 @@ vc4_submit_cl_ioctl(struct drm_device *dev, void *data,
 	if (ret)
 		goto fail;
 
+	if (args->perfmonid) {
+		exec->perfmon = vc4_perfmon_find(vc4file,
+						 args->perfmonid);
+		if (!exec->perfmon) {
+			ret = -ENOENT;
+			goto fail;
+		}
+	}
+
 	if (exec->args->bin_cl_size != 0) {
 		ret = vc4_get_bcl(dev, exec);
 		if (ret)
diff --git a/drivers/gpu/drm/vc4/vc4_irq.c b/drivers/gpu/drm/vc4/vc4_irq.c
index 3dd62d75f531..4cd2ccfe15f4 100644
--- a/drivers/gpu/drm/vc4/vc4_irq.c
+++ b/drivers/gpu/drm/vc4/vc4_irq.c
@@ -104,13 +104,20 @@ static void
 vc4_irq_finish_bin_job(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
-	struct vc4_exec_info *exec = vc4_first_bin_job(vc4);
+	struct vc4_exec_info *next, *exec = vc4_first_bin_job(vc4);
 
 	if (!exec)
 		return;
 
 	vc4_move_job_to_render(dev, exec);
-	vc4_submit_next_bin_job(dev);
+	next = vc4_first_bin_job(vc4);
+
+	/* Only submit the next job in the bin list if it matches the perfmon
+	 * attached to the one that just finished (or if both jobs don't have
+	 * perfmon attached to them).
+	 */
+	if (next && next->perfmon == exec->perfmon)
+		vc4_submit_next_bin_job(dev);
 }
 
 static void
@@ -122,6 +129,10 @@ vc4_cancel_bin_job(struct drm_device *dev)
 	if (!exec)
 		return;
 
+	/* Stop the perfmon so that the next bin job can be started. */
+	if (exec->perfmon)
+		vc4_perfmon_stop(vc4, exec->perfmon, false);
+
 	list_move_tail(&exec->head, &vc4->bin_job_list);
 	vc4_submit_next_bin_job(dev);
 }
@@ -131,18 +142,41 @@ vc4_irq_finish_render_job(struct drm_device *dev)
 {
 	struct vc4_dev *vc4 = to_vc4_dev(dev);
 	struct vc4_exec_info *exec = vc4_first_render_job(vc4);
+	struct vc4_exec_info *nextbin, *nextrender;
 
 	if (!exec)
 		return;
 
 	vc4->finished_seqno++;
 	list_move_tail(&exec->head, &vc4->job_done_list);
+
+	nextbin = vc4_first_bin_job(vc4);
+	nextrender = vc4_first_render_job(vc4);
+
+	/* Only stop the perfmon if following jobs in the queue don't expect it
+	 * to be enabled.
+	 */
+	if (exec->perfmon && !nextrender &&
+	    (!nextbin || nextbin->perfmon != exec->perfmon))
+		vc4_perfmon_stop(vc4, exec->perfmon, true);
+
+	/* If there's a render job waiting, start it. If this is not the case
+	 * we may have to unblock the binner if it's been stalled because of
+	 * perfmon (this can be checked by comparing the perfmon attached to
+	 * the finished renderjob to the one attached to the next bin job: if
+	 * they don't match, this means the binner is stalled and should be
+	 * restarted).
+	 */
+	if (nextrender)
+		vc4_submit_next_render_job(dev);
+	else if (nextbin && nextbin->perfmon != exec->perfmon)
+		vc4_submit_next_bin_job(dev);
+
 	if (exec->fence) {
 		dma_fence_signal_locked(exec->fence);
 		dma_fence_put(exec->fence);
 		exec->fence = NULL;
 	}
-	vc4_submit_next_render_job(dev);
 
 	wake_up_all(&vc4->job_wait_queue);
 	schedule_work(&vc4->job_done_work);
diff --git a/drivers/gpu/drm/vc4/vc4_perfmon.c b/drivers/gpu/drm/vc4/vc4_perfmon.c
new file mode 100644
index 000000000000..437e7a27f21d
--- /dev/null
+++ b/drivers/gpu/drm/vc4/vc4_perfmon.c
@@ -0,0 +1,188 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright (C) 2018 Broadcom
+ */
+
+/**
+ * DOC: VC4 V3D performance monitor module
+ *
+ * The V3D block provides 16 hardware counters which can count various events.
+ */
+
+#include "vc4_drv.h"
+#include "vc4_regs.h"
+
+#define VC4_PERFMONID_MIN	1
+#define VC4_PERFMONID_MAX	U32_MAX
+
+void vc4_perfmon_get(struct vc4_perfmon *perfmon)
+{
+	if (perfmon)
+		refcount_inc(&perfmon->refcnt);
+}
+
+void vc4_perfmon_put(struct vc4_perfmon *perfmon)
+{
+	if (perfmon && refcount_dec_and_test(&perfmon->refcnt))
+		kfree(perfmon);
+}
+
+void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon)
+{
+	unsigned int i;
+	u32 mask;
+
+	if (WARN_ON_ONCE(!perfmon || vc4->active_perfmon))
+		return;
+
+	for (i = 0; i < perfmon->ncounters; i++)
+		V3D_WRITE(V3D_PCTRS(i), perfmon->events[i]);
+
+	mask = GENMASK(perfmon->ncounters - 1, 0);
+	V3D_WRITE(V3D_PCTRC, mask);
+	V3D_WRITE(V3D_PCTRE, V3D_PCTRE_EN | mask);
+	vc4->active_perfmon = perfmon;
+}
+
+void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon,
+		      bool capture)
+{
+	unsigned int i;
+
+	if (WARN_ON_ONCE(!vc4->active_perfmon ||
+			 perfmon != vc4->active_perfmon))
+		return;
+
+	if (capture) {
+		for (i = 0; i < perfmon->ncounters; i++)
+			perfmon->counters[i] += V3D_READ(V3D_PCTR(i));
+	}
+
+	V3D_WRITE(V3D_PCTRE, 0);
+	vc4->active_perfmon = NULL;
+}
+
+struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id)
+{
+	struct vc4_perfmon *perfmon;
+
+	mutex_lock(&vc4file->perfmon.lock);
+	perfmon = idr_find(&vc4file->perfmon.idr, id);
+	vc4_perfmon_get(perfmon);
+	mutex_unlock(&vc4file->perfmon.lock);
+
+	return perfmon;
+}
+
+void vc4_perfmon_open_file(struct vc4_file *vc4file)
+{
+	mutex_init(&vc4file->perfmon.lock);
+	idr_init(&vc4file->perfmon.idr);
+}
+
+static int vc4_perfmon_idr_del(int id, void *elem, void *data)
+{
+	struct vc4_perfmon *perfmon = elem;
+
+	vc4_perfmon_put(perfmon);
+
+	return 0;
+}
+
+void vc4_perfmon_close_file(struct vc4_file *vc4file)
+{
+	mutex_lock(&vc4file->perfmon.lock);
+	idr_for_each(&vc4file->perfmon.idr, vc4_perfmon_idr_del, NULL);
+	idr_destroy(&vc4file->perfmon.idr);
+	mutex_unlock(&vc4file->perfmon.lock);
+}
+
+int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data,
+			     struct drm_file *file_priv)
+{
+	struct vc4_file *vc4file = file_priv->driver_priv;
+	struct drm_vc4_perfmon_create *req = data;
+	struct vc4_perfmon *perfmon;
+	unsigned int i;
+	int ret;
+
+	/* Number of monitored counters cannot exceed HW limits. */
+	if (req->ncounters > DRM_VC4_MAX_PERF_COUNTERS ||
+	    !req->ncounters)
+		return -EINVAL;
+
+	/* Make sure all events are valid. */
+	for (i = 0; i < req->ncounters; i++) {
+		if (req->events[i] >= VC4_PERFCNT_NUM_EVENTS)
+			return -EINVAL;
+	}
+
+	perfmon = kzalloc(sizeof(*perfmon) + (req->ncounters * sizeof(u64)),
+			  GFP_KERNEL);
+	if (!perfmon)
+		return -ENOMEM;
+
+	for (i = 0; i < req->ncounters; i++)
+		perfmon->events[i] = req->events[i];
+
+	perfmon->ncounters = req->ncounters;
+
+	refcount_set(&perfmon->refcnt, 1);
+
+	mutex_lock(&vc4file->perfmon.lock);
+	ret = idr_alloc(&vc4file->perfmon.idr, perfmon, VC4_PERFMONID_MIN,
+			VC4_PERFMONID_MAX, GFP_KERNEL);
+	mutex_unlock(&vc4file->perfmon.lock);
+
+	if (ret < 0) {
+		kfree(perfmon);
+		return ret;
+	}
+
+	req->id = ret;
+	return 0;
+}
+
+int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
+			      struct drm_file *file_priv)
+{
+	struct vc4_file *vc4file = file_priv->driver_priv;
+	struct drm_vc4_perfmon_destroy *req = data;
+	struct vc4_perfmon *perfmon;
+
+	mutex_lock(&vc4file->perfmon.lock);
+	perfmon = idr_remove(&vc4file->perfmon.idr, req->id);
+	mutex_unlock(&vc4file->perfmon.lock);
+
+	if (!perfmon)
+		return -EINVAL;
+
+	vc4_perfmon_put(perfmon);
+	return 0;
+}
+
+int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
+				 struct drm_file *file_priv)
+{
+	struct vc4_file *vc4file = file_priv->driver_priv;
+	struct drm_vc4_perfmon_get_values *req = data;
+	struct vc4_perfmon *perfmon;
+	int ret;
+
+	mutex_lock(&vc4file->perfmon.lock);
+	perfmon = idr_find(&vc4file->perfmon.idr, req->id);
+	vc4_perfmon_get(perfmon);
+	mutex_unlock(&vc4file->perfmon.lock);
+
+	if (!perfmon)
+		return -EINVAL;
+
+	if (copy_to_user(u64_to_user_ptr(req->values_ptr), perfmon->counters,
+			 perfmon->ncounters * sizeof(u64)))
+		ret = -EFAULT;
+	else
+		ret = 0;
+
+	vc4_perfmon_put(perfmon);
+	return ret;
+}
diff --git a/drivers/gpu/drm/vc4/vc4_regs.h b/drivers/gpu/drm/vc4/vc4_regs.h
index 55677bd50f66..b9749cb24063 100644
--- a/drivers/gpu/drm/vc4/vc4_regs.h
+++ b/drivers/gpu/drm/vc4/vc4_regs.h
@@ -122,38 +122,9 @@
 #define V3D_VPMBASE  0x00504
 #define V3D_PCTRC    0x00670
 #define V3D_PCTRE    0x00674
-#define V3D_PCTR0    0x00680
-#define V3D_PCTRS0   0x00684
-#define V3D_PCTR1    0x00688
-#define V3D_PCTRS1   0x0068c
-#define V3D_PCTR2    0x00690
-#define V3D_PCTRS2   0x00694
-#define V3D_PCTR3    0x00698
-#define V3D_PCTRS3   0x0069c
-#define V3D_PCTR4    0x006a0
-#define V3D_PCTRS4   0x006a4
-#define V3D_PCTR5    0x006a8
-#define V3D_PCTRS5   0x006ac
-#define V3D_PCTR6    0x006b0
-#define V3D_PCTRS6   0x006b4
-#define V3D_PCTR7    0x006b8
-#define V3D_PCTRS7   0x006bc
-#define V3D_PCTR8    0x006c0
-#define V3D_PCTRS8   0x006c4
-#define V3D_PCTR9    0x006c8
-#define V3D_PCTRS9   0x006cc
-#define V3D_PCTR10   0x006d0
-#define V3D_PCTRS10  0x006d4
-#define V3D_PCTR11   0x006d8
-#define V3D_PCTRS11  0x006dc
-#define V3D_PCTR12   0x006e0
-#define V3D_PCTRS12  0x006e4
-#define V3D_PCTR13   0x006e8
-#define V3D_PCTRS13  0x006ec
-#define V3D_PCTR14   0x006f0
-#define V3D_PCTRS14  0x006f4
-#define V3D_PCTR15   0x006f8
-#define V3D_PCTRS15  0x006fc
+# define V3D_PCTRE_EN	BIT(31)
+#define V3D_PCTR(x)  (0x00680 + ((x) * 8))
+#define V3D_PCTRS(x) (0x00684 + ((x) * 8))
 #define V3D_DBGE     0x00f00
 #define V3D_FDBGO    0x00f04
 #define V3D_FDBGB    0x00f08
diff --git a/drivers/gpu/drm/vc4/vc4_v3d.c b/drivers/gpu/drm/vc4/vc4_v3d.c
index 493f392b3a0a..bfc2fa73d2ae 100644
--- a/drivers/gpu/drm/vc4/vc4_v3d.c
+++ b/drivers/gpu/drm/vc4/vc4_v3d.c
@@ -68,38 +68,38 @@ static const struct {
 	REGDEF(V3D_VPMBASE),
 	REGDEF(V3D_PCTRC),
 	REGDEF(V3D_PCTRE),
-	REGDEF(V3D_PCTR0),
-	REGDEF(V3D_PCTRS0),
-	REGDEF(V3D_PCTR1),
-	REGDEF(V3D_PCTRS1),
-	REGDEF(V3D_PCTR2),
-	REGDEF(V3D_PCTRS2),
-	REGDEF(V3D_PCTR3),
-	REGDEF(V3D_PCTRS3),
-	REGDEF(V3D_PCTR4),
-	REGDEF(V3D_PCTRS4),
-	REGDEF(V3D_PCTR5),
-	REGDEF(V3D_PCTRS5),
-	REGDEF(V3D_PCTR6),
-	REGDEF(V3D_PCTRS6),
-	REGDEF(V3D_PCTR7),
-	REGDEF(V3D_PCTRS7),
-	REGDEF(V3D_PCTR8),
-	REGDEF(V3D_PCTRS8),
-	REGDEF(V3D_PCTR9),
-	REGDEF(V3D_PCTRS9),
-	REGDEF(V3D_PCTR10),
-	REGDEF(V3D_PCTRS10),
-	REGDEF(V3D_PCTR11),
-	REGDEF(V3D_PCTRS11),
-	REGDEF(V3D_PCTR12),
-	REGDEF(V3D_PCTRS12),
-	REGDEF(V3D_PCTR13),
-	REGDEF(V3D_PCTRS13),
-	REGDEF(V3D_PCTR14),
-	REGDEF(V3D_PCTRS14),
-	REGDEF(V3D_PCTR15),
-	REGDEF(V3D_PCTRS15),
+	REGDEF(V3D_PCTR(0)),
+	REGDEF(V3D_PCTRS(0)),
+	REGDEF(V3D_PCTR(1)),
+	REGDEF(V3D_PCTRS(1)),
+	REGDEF(V3D_PCTR(2)),
+	REGDEF(V3D_PCTRS(2)),
+	REGDEF(V3D_PCTR(3)),
+	REGDEF(V3D_PCTRS(3)),
+	REGDEF(V3D_PCTR(4)),
+	REGDEF(V3D_PCTRS(4)),
+	REGDEF(V3D_PCTR(5)),
+	REGDEF(V3D_PCTRS(5)),
+	REGDEF(V3D_PCTR(6)),
+	REGDEF(V3D_PCTRS(6)),
+	REGDEF(V3D_PCTR(7)),
+	REGDEF(V3D_PCTRS(7)),
+	REGDEF(V3D_PCTR(8)),
+	REGDEF(V3D_PCTRS(8)),
+	REGDEF(V3D_PCTR(9)),
+	REGDEF(V3D_PCTRS(9)),
+	REGDEF(V3D_PCTR(10)),
+	REGDEF(V3D_PCTRS(10)),
+	REGDEF(V3D_PCTR(11)),
+	REGDEF(V3D_PCTRS(11)),
+	REGDEF(V3D_PCTR(12)),
+	REGDEF(V3D_PCTRS(12)),
+	REGDEF(V3D_PCTR(13)),
+	REGDEF(V3D_PCTRS(13)),
+	REGDEF(V3D_PCTR(14)),
+	REGDEF(V3D_PCTRS(14)),
+	REGDEF(V3D_PCTR(15)),
+	REGDEF(V3D_PCTRS(15)),
 	REGDEF(V3D_DBGE),
 	REGDEF(V3D_FDBGO),
 	REGDEF(V3D_FDBGB),
diff --git a/include/uapi/drm/vc4_drm.h b/include/uapi/drm/vc4_drm.h
index 52263b575bdc..b95a0e11cb07 100644
--- a/include/uapi/drm/vc4_drm.h
+++ b/include/uapi/drm/vc4_drm.h
@@ -42,6 +42,9 @@ extern "C" {
 #define DRM_VC4_GET_TILING                        0x09
 #define DRM_VC4_LABEL_BO                          0x0a
 #define DRM_VC4_GEM_MADVISE                       0x0b
+#define DRM_VC4_PERFMON_CREATE                    0x0c
+#define DRM_VC4_PERFMON_DESTROY                   0x0d
+#define DRM_VC4_PERFMON_GET_VALUES                0x0e
 
 #define DRM_IOCTL_VC4_SUBMIT_CL           DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_SUBMIT_CL, struct drm_vc4_submit_cl)
 #define DRM_IOCTL_VC4_WAIT_SEQNO          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_WAIT_SEQNO, struct drm_vc4_wait_seqno)
@@ -55,6 +58,9 @@ extern "C" {
 #define DRM_IOCTL_VC4_GET_TILING          DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_GET_TILING, struct drm_vc4_get_tiling)
 #define DRM_IOCTL_VC4_LABEL_BO            DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_LABEL_BO, struct drm_vc4_label_bo)
 #define DRM_IOCTL_VC4_GEM_MADVISE         DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_GEM_MADVISE, struct drm_vc4_gem_madvise)
+#define DRM_IOCTL_VC4_PERFMON_CREATE      DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_PERFMON_CREATE, struct drm_vc4_perfmon_create)
+#define DRM_IOCTL_VC4_PERFMON_DESTROY     DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_PERFMON_DESTROY, struct drm_vc4_perfmon_destroy)
+#define DRM_IOCTL_VC4_PERFMON_GET_VALUES  DRM_IOWR(DRM_COMMAND_BASE + DRM_VC4_PERFMON_GET_VALUES, struct drm_vc4_perfmon_get_values)
 
 struct drm_vc4_submit_rcl_surface {
 	__u32 hindex; /* Handle index, or ~0 if not present. */
@@ -173,6 +179,15 @@ struct drm_vc4_submit_cl {
 	 * wait ioctl).
 	 */
 	__u64 seqno;
+
+	/* ID of the perfmon to attach to this job. 0 means no perfmon. */
+	__u32 perfmonid;
+
+	/* Unused field to align this struct on 64 bits. Must be set to 0.
+	 * If one ever needs to add an u32 field to this struct, this field
+	 * can be used.
+	 */
+	__u32 pad2;
 };
 
 /**
@@ -308,6 +323,7 @@ struct drm_vc4_get_hang_state {
 #define DRM_VC4_PARAM_SUPPORTS_THREADED_FS	5
 #define DRM_VC4_PARAM_SUPPORTS_FIXED_RCL_ORDER	6
 #define DRM_VC4_PARAM_SUPPORTS_MADVISE		7
+#define DRM_VC4_PARAM_SUPPORTS_PERFMON		8
 
 struct drm_vc4_get_param {
 	__u32 param;
@@ -352,6 +368,66 @@ struct drm_vc4_gem_madvise {
 	__u32 pad;
 };
 
+enum {
+	VC4_PERFCNT_FEP_VALID_PRIMS_NO_RENDER,
+	VC4_PERFCNT_FEP_VALID_PRIMS_RENDER,
+	VC4_PERFCNT_FEP_CLIPPED_QUADS,
+	VC4_PERFCNT_FEP_VALID_QUADS,
+	VC4_PERFCNT_TLB_QUADS_NOT_PASSING_STENCIL,
+	VC4_PERFCNT_TLB_QUADS_NOT_PASSING_Z_AND_STENCIL,
+	VC4_PERFCNT_TLB_QUADS_PASSING_Z_AND_STENCIL,
+	VC4_PERFCNT_TLB_QUADS_ZERO_COVERAGE,
+	VC4_PERFCNT_TLB_QUADS_NON_ZERO_COVERAGE,
+	VC4_PERFCNT_TLB_QUADS_WRITTEN_TO_COLOR_BUF,
+	VC4_PERFCNT_PLB_PRIMS_OUTSIDE_VIEWPORT,
+	VC4_PERFCNT_PLB_PRIMS_NEED_CLIPPING,
+	VC4_PERFCNT_PSE_PRIMS_REVERSED,
+	VC4_PERFCNT_QPU_TOTAL_IDLE_CYCLES,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_VERTEX_COORD_SHADING,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_FRAGMENT_SHADING,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_EXEC_VALID_INST,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_WAITING_TMUS,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_WAITING_SCOREBOARD,
+	VC4_PERFCNT_QPU_TOTAL_CLK_CYCLES_WAITING_VARYINGS,
+	VC4_PERFCNT_QPU_TOTAL_INST_CACHE_HIT,
+	VC4_PERFCNT_QPU_TOTAL_INST_CACHE_MISS,
+	VC4_PERFCNT_QPU_TOTAL_UNIFORM_CACHE_HIT,
+	VC4_PERFCNT_QPU_TOTAL_UNIFORM_CACHE_MISS,
+	VC4_PERFCNT_TMU_TOTAL_TEXT_QUADS_PROCESSED,
+	VC4_PERFCNT_TMU_TOTAL_TEXT_CACHE_MISS,
+	VC4_PERFCNT_VPM_TOTAL_CLK_CYCLES_VDW_STALLED,
+	VC4_PERFCNT_VPM_TOTAL_CLK_CYCLES_VCD_STALLED,
+	VC4_PERFCNT_L2C_TOTAL_L2_CACHE_HIT,
+	VC4_PERFCNT_L2C_TOTAL_L2_CACHE_MISS,
+	VC4_PERFCNT_NUM_EVENTS,
+};
+
+#define DRM_VC4_MAX_PERF_COUNTERS	16
+
+struct drm_vc4_perfmon_create {
+	__u32 id;
+	__u32 ncounters;
+	__u8 events[DRM_VC4_MAX_PERF_COUNTERS];
+};
+
+struct drm_vc4_perfmon_destroy {
+	__u32 id;
+};
+
+/*
+ * Returns the values of the performance counters tracked by this
+ * perfmon (as an array of ncounters u64 values).
+ *
+ * No implicit synchronization is performed, so the user has to
+ * guarantee that any jobs using this perfmon have already been
+ * completed  (probably by blocking on the seqno returned by the
+ * last exec that used the perfmon).
+ */
+struct drm_vc4_perfmon_get_values {
+	__u32 id;
+	__u64 values_ptr;
+};
+
 #if defined(__cplusplus)
 }
 #endif