1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2021 Raspberry Pi
4 */
5
6 #include "v3d_drv.h"
7 #include "v3d_regs.h"
8
9 #define V3D_PERFMONID_MIN 1
10 #define V3D_PERFMONID_MAX U32_MAX
11
v3d_perfmon_get(struct v3d_perfmon * perfmon)12 void v3d_perfmon_get(struct v3d_perfmon *perfmon)
13 {
14 if (perfmon)
15 refcount_inc(&perfmon->refcnt);
16 }
17
v3d_perfmon_put(struct v3d_perfmon * perfmon)18 void v3d_perfmon_put(struct v3d_perfmon *perfmon)
19 {
20 if (perfmon && refcount_dec_and_test(&perfmon->refcnt)) {
21 mutex_destroy(&perfmon->lock);
22 kfree(perfmon);
23 }
24 }
25
v3d_perfmon_start(struct v3d_dev * v3d,struct v3d_perfmon * perfmon)26 void v3d_perfmon_start(struct v3d_dev *v3d, struct v3d_perfmon *perfmon)
27 {
28 unsigned int i;
29 u32 mask;
30 u8 ncounters;
31
32 if (WARN_ON_ONCE(!perfmon || v3d->active_perfmon))
33 return;
34
35 ncounters = perfmon->ncounters;
36 mask = GENMASK(ncounters - 1, 0);
37
38 for (i = 0; i < ncounters; i++) {
39 u32 source = i / 4;
40 u32 channel = V3D_SET_FIELD(perfmon->counters[i], V3D_PCTR_S0);
41
42 i++;
43 channel |= V3D_SET_FIELD(i < ncounters ? perfmon->counters[i] : 0,
44 V3D_PCTR_S1);
45 i++;
46 channel |= V3D_SET_FIELD(i < ncounters ? perfmon->counters[i] : 0,
47 V3D_PCTR_S2);
48 i++;
49 channel |= V3D_SET_FIELD(i < ncounters ? perfmon->counters[i] : 0,
50 V3D_PCTR_S3);
51 V3D_CORE_WRITE(0, V3D_V4_PCTR_0_SRC_X(source), channel);
52 }
53
54 V3D_CORE_WRITE(0, V3D_V4_PCTR_0_CLR, mask);
55 V3D_CORE_WRITE(0, V3D_PCTR_0_OVERFLOW, mask);
56 V3D_CORE_WRITE(0, V3D_V4_PCTR_0_EN, mask);
57
58 v3d->active_perfmon = perfmon;
59 }
60
v3d_perfmon_stop(struct v3d_dev * v3d,struct v3d_perfmon * perfmon,bool capture)61 void v3d_perfmon_stop(struct v3d_dev *v3d, struct v3d_perfmon *perfmon,
62 bool capture)
63 {
64 unsigned int i;
65
66 if (!perfmon || !v3d->active_perfmon)
67 return;
68
69 mutex_lock(&perfmon->lock);
70 if (perfmon != v3d->active_perfmon) {
71 mutex_unlock(&perfmon->lock);
72 return;
73 }
74
75 if (capture)
76 for (i = 0; i < perfmon->ncounters; i++)
77 perfmon->values[i] += V3D_CORE_READ(0, V3D_PCTR_0_PCTRX(i));
78
79 V3D_CORE_WRITE(0, V3D_V4_PCTR_0_EN, 0);
80
81 v3d->active_perfmon = NULL;
82 mutex_unlock(&perfmon->lock);
83 }
84
v3d_perfmon_find(struct v3d_file_priv * v3d_priv,int id)85 struct v3d_perfmon *v3d_perfmon_find(struct v3d_file_priv *v3d_priv, int id)
86 {
87 struct v3d_perfmon *perfmon;
88
89 mutex_lock(&v3d_priv->perfmon.lock);
90 perfmon = idr_find(&v3d_priv->perfmon.idr, id);
91 v3d_perfmon_get(perfmon);
92 mutex_unlock(&v3d_priv->perfmon.lock);
93
94 return perfmon;
95 }
96
v3d_perfmon_open_file(struct v3d_file_priv * v3d_priv)97 void v3d_perfmon_open_file(struct v3d_file_priv *v3d_priv)
98 {
99 mutex_init(&v3d_priv->perfmon.lock);
100 idr_init_base(&v3d_priv->perfmon.idr, 1);
101 }
102
v3d_perfmon_idr_del(int id,void * elem,void * data)103 static int v3d_perfmon_idr_del(int id, void *elem, void *data)
104 {
105 struct v3d_perfmon *perfmon = elem;
106
107 v3d_perfmon_put(perfmon);
108
109 return 0;
110 }
111
v3d_perfmon_close_file(struct v3d_file_priv * v3d_priv)112 void v3d_perfmon_close_file(struct v3d_file_priv *v3d_priv)
113 {
114 mutex_lock(&v3d_priv->perfmon.lock);
115 idr_for_each(&v3d_priv->perfmon.idr, v3d_perfmon_idr_del, NULL);
116 idr_destroy(&v3d_priv->perfmon.idr);
117 mutex_unlock(&v3d_priv->perfmon.lock);
118 mutex_destroy(&v3d_priv->perfmon.lock);
119 }
120
v3d_perfmon_create_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)121 int v3d_perfmon_create_ioctl(struct drm_device *dev, void *data,
122 struct drm_file *file_priv)
123 {
124 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
125 struct drm_v3d_perfmon_create *req = data;
126 struct v3d_perfmon *perfmon;
127 unsigned int i;
128 int ret;
129
130 /* Number of monitored counters cannot exceed HW limits. */
131 if (req->ncounters > DRM_V3D_MAX_PERF_COUNTERS ||
132 !req->ncounters)
133 return -EINVAL;
134
135 /* Make sure all counters are valid. */
136 for (i = 0; i < req->ncounters; i++) {
137 if (req->counters[i] >= V3D_PERFCNT_NUM)
138 return -EINVAL;
139 }
140
141 perfmon = kzalloc(struct_size(perfmon, values, req->ncounters),
142 GFP_KERNEL);
143 if (!perfmon)
144 return -ENOMEM;
145
146 for (i = 0; i < req->ncounters; i++)
147 perfmon->counters[i] = req->counters[i];
148
149 perfmon->ncounters = req->ncounters;
150
151 refcount_set(&perfmon->refcnt, 1);
152 mutex_init(&perfmon->lock);
153
154 mutex_lock(&v3d_priv->perfmon.lock);
155 ret = idr_alloc(&v3d_priv->perfmon.idr, perfmon, V3D_PERFMONID_MIN,
156 V3D_PERFMONID_MAX, GFP_KERNEL);
157 mutex_unlock(&v3d_priv->perfmon.lock);
158
159 if (ret < 0) {
160 mutex_destroy(&perfmon->lock);
161 kfree(perfmon);
162 return ret;
163 }
164
165 req->id = ret;
166
167 return 0;
168 }
169
v3d_perfmon_destroy_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)170 int v3d_perfmon_destroy_ioctl(struct drm_device *dev, void *data,
171 struct drm_file *file_priv)
172 {
173 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
174 struct drm_v3d_perfmon_destroy *req = data;
175 struct v3d_perfmon *perfmon;
176
177 mutex_lock(&v3d_priv->perfmon.lock);
178 perfmon = idr_remove(&v3d_priv->perfmon.idr, req->id);
179 mutex_unlock(&v3d_priv->perfmon.lock);
180
181 if (!perfmon)
182 return -EINVAL;
183
184 v3d_perfmon_put(perfmon);
185
186 return 0;
187 }
188
v3d_perfmon_get_values_ioctl(struct drm_device * dev,void * data,struct drm_file * file_priv)189 int v3d_perfmon_get_values_ioctl(struct drm_device *dev, void *data,
190 struct drm_file *file_priv)
191 {
192 struct v3d_dev *v3d = to_v3d_dev(dev);
193 struct v3d_file_priv *v3d_priv = file_priv->driver_priv;
194 struct drm_v3d_perfmon_get_values *req = data;
195 struct v3d_perfmon *perfmon;
196 int ret = 0;
197
198 if (req->pad != 0)
199 return -EINVAL;
200
201 mutex_lock(&v3d_priv->perfmon.lock);
202 perfmon = idr_find(&v3d_priv->perfmon.idr, req->id);
203 v3d_perfmon_get(perfmon);
204 mutex_unlock(&v3d_priv->perfmon.lock);
205
206 if (!perfmon)
207 return -EINVAL;
208
209 v3d_perfmon_stop(v3d, perfmon, true);
210
211 if (copy_to_user(u64_to_user_ptr(req->values_ptr), perfmon->values,
212 perfmon->ncounters * sizeof(u64)))
213 ret = -EFAULT;
214
215 v3d_perfmon_put(perfmon);
216
217 return ret;
218 }
219