1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Copyright (C) 2018 Broadcom 4 */ 5 6 /** 7 * DOC: VC4 V3D performance monitor module 8 * 9 * The V3D block provides 16 hardware counters which can count various events. 10 */ 11 12 #include "vc4_drv.h" 13 #include "vc4_regs.h" 14 15 #define VC4_PERFMONID_MIN 1 16 #define VC4_PERFMONID_MAX U32_MAX 17 18 void vc4_perfmon_get(struct vc4_perfmon *perfmon) 19 { 20 struct vc4_dev *vc4; 21 22 if (!perfmon) 23 return; 24 25 vc4 = perfmon->dev; 26 if (WARN_ON_ONCE(vc4->is_vc5)) 27 return; 28 29 refcount_inc(&perfmon->refcnt); 30 } 31 32 void vc4_perfmon_put(struct vc4_perfmon *perfmon) 33 { 34 struct vc4_dev *vc4; 35 36 if (!perfmon) 37 return; 38 39 vc4 = perfmon->dev; 40 if (WARN_ON_ONCE(vc4->is_vc5)) 41 return; 42 43 if (refcount_dec_and_test(&perfmon->refcnt)) 44 kfree(perfmon); 45 } 46 47 void vc4_perfmon_start(struct vc4_dev *vc4, struct vc4_perfmon *perfmon) 48 { 49 unsigned int i; 50 u32 mask; 51 52 if (WARN_ON_ONCE(vc4->is_vc5)) 53 return; 54 55 if (WARN_ON_ONCE(!perfmon || vc4->active_perfmon)) 56 return; 57 58 for (i = 0; i < perfmon->ncounters; i++) 59 V3D_WRITE(V3D_PCTRS(i), perfmon->events[i]); 60 61 mask = GENMASK(perfmon->ncounters - 1, 0); 62 V3D_WRITE(V3D_PCTRC, mask); 63 V3D_WRITE(V3D_PCTRE, V3D_PCTRE_EN | mask); 64 vc4->active_perfmon = perfmon; 65 } 66 67 void vc4_perfmon_stop(struct vc4_dev *vc4, struct vc4_perfmon *perfmon, 68 bool capture) 69 { 70 unsigned int i; 71 72 if (WARN_ON_ONCE(vc4->is_vc5)) 73 return; 74 75 if (WARN_ON_ONCE(!vc4->active_perfmon || 76 perfmon != vc4->active_perfmon)) 77 return; 78 79 if (capture) { 80 for (i = 0; i < perfmon->ncounters; i++) 81 perfmon->counters[i] += V3D_READ(V3D_PCTR(i)); 82 } 83 84 V3D_WRITE(V3D_PCTRE, 0); 85 vc4->active_perfmon = NULL; 86 } 87 88 struct vc4_perfmon *vc4_perfmon_find(struct vc4_file *vc4file, int id) 89 { 90 struct vc4_dev *vc4 = vc4file->dev; 91 struct vc4_perfmon *perfmon; 92 93 if (WARN_ON_ONCE(vc4->is_vc5)) 94 return NULL; 95 96 mutex_lock(&vc4file->perfmon.lock); 97 perfmon = idr_find(&vc4file->perfmon.idr, id); 98 vc4_perfmon_get(perfmon); 99 mutex_unlock(&vc4file->perfmon.lock); 100 101 return perfmon; 102 } 103 104 void vc4_perfmon_open_file(struct vc4_file *vc4file) 105 { 106 struct vc4_dev *vc4 = vc4file->dev; 107 108 if (WARN_ON_ONCE(vc4->is_vc5)) 109 return; 110 111 mutex_init(&vc4file->perfmon.lock); 112 idr_init_base(&vc4file->perfmon.idr, VC4_PERFMONID_MIN); 113 vc4file->dev = vc4; 114 } 115 116 static int vc4_perfmon_idr_del(int id, void *elem, void *data) 117 { 118 struct vc4_perfmon *perfmon = elem; 119 struct vc4_dev *vc4 = (struct vc4_dev *)data; 120 121 /* If the active perfmon is being destroyed, stop it first */ 122 if (perfmon == vc4->active_perfmon) 123 vc4_perfmon_stop(vc4, perfmon, false); 124 125 vc4_perfmon_put(perfmon); 126 127 return 0; 128 } 129 130 void vc4_perfmon_close_file(struct vc4_file *vc4file) 131 { 132 struct vc4_dev *vc4 = vc4file->dev; 133 134 if (WARN_ON_ONCE(vc4->is_vc5)) 135 return; 136 137 mutex_lock(&vc4file->perfmon.lock); 138 idr_for_each(&vc4file->perfmon.idr, vc4_perfmon_idr_del, vc4); 139 idr_destroy(&vc4file->perfmon.idr); 140 mutex_unlock(&vc4file->perfmon.lock); 141 mutex_destroy(&vc4file->perfmon.lock); 142 } 143 144 int vc4_perfmon_create_ioctl(struct drm_device *dev, void *data, 145 struct drm_file *file_priv) 146 { 147 struct vc4_dev *vc4 = to_vc4_dev(dev); 148 struct vc4_file *vc4file = file_priv->driver_priv; 149 struct drm_vc4_perfmon_create *req = data; 150 struct vc4_perfmon *perfmon; 151 unsigned int i; 152 int ret; 153 154 if (WARN_ON_ONCE(vc4->is_vc5)) 155 return -ENODEV; 156 157 if (!vc4->v3d) { 158 DRM_DEBUG("Creating perfmon no VC4 V3D probed\n"); 159 return -ENODEV; 160 } 161 162 /* Number of monitored counters cannot exceed HW limits. */ 163 if (req->ncounters > DRM_VC4_MAX_PERF_COUNTERS || 164 !req->ncounters) 165 return -EINVAL; 166 167 /* Make sure all events are valid. */ 168 for (i = 0; i < req->ncounters; i++) { 169 if (req->events[i] >= VC4_PERFCNT_NUM_EVENTS) 170 return -EINVAL; 171 } 172 173 perfmon = kzalloc(struct_size(perfmon, counters, req->ncounters), 174 GFP_KERNEL); 175 if (!perfmon) 176 return -ENOMEM; 177 perfmon->dev = vc4; 178 179 for (i = 0; i < req->ncounters; i++) 180 perfmon->events[i] = req->events[i]; 181 182 perfmon->ncounters = req->ncounters; 183 184 refcount_set(&perfmon->refcnt, 1); 185 186 mutex_lock(&vc4file->perfmon.lock); 187 ret = idr_alloc(&vc4file->perfmon.idr, perfmon, VC4_PERFMONID_MIN, 188 VC4_PERFMONID_MAX, GFP_KERNEL); 189 mutex_unlock(&vc4file->perfmon.lock); 190 191 if (ret < 0) { 192 kfree(perfmon); 193 return ret; 194 } 195 196 req->id = ret; 197 return 0; 198 } 199 200 int vc4_perfmon_destroy_ioctl(struct drm_device *dev, void *data, 201 struct drm_file *file_priv) 202 { 203 struct vc4_dev *vc4 = to_vc4_dev(dev); 204 struct vc4_file *vc4file = file_priv->driver_priv; 205 struct drm_vc4_perfmon_destroy *req = data; 206 struct vc4_perfmon *perfmon; 207 208 if (WARN_ON_ONCE(vc4->is_vc5)) 209 return -ENODEV; 210 211 if (!vc4->v3d) { 212 DRM_DEBUG("Destroying perfmon no VC4 V3D probed\n"); 213 return -ENODEV; 214 } 215 216 mutex_lock(&vc4file->perfmon.lock); 217 perfmon = idr_remove(&vc4file->perfmon.idr, req->id); 218 mutex_unlock(&vc4file->perfmon.lock); 219 220 if (!perfmon) 221 return -EINVAL; 222 223 vc4_perfmon_put(perfmon); 224 return 0; 225 } 226 227 int vc4_perfmon_get_values_ioctl(struct drm_device *dev, void *data, 228 struct drm_file *file_priv) 229 { 230 struct vc4_dev *vc4 = to_vc4_dev(dev); 231 struct vc4_file *vc4file = file_priv->driver_priv; 232 struct drm_vc4_perfmon_get_values *req = data; 233 struct vc4_perfmon *perfmon; 234 int ret; 235 236 if (WARN_ON_ONCE(vc4->is_vc5)) 237 return -ENODEV; 238 239 if (!vc4->v3d) { 240 DRM_DEBUG("Getting perfmon no VC4 V3D probed\n"); 241 return -ENODEV; 242 } 243 244 mutex_lock(&vc4file->perfmon.lock); 245 perfmon = idr_find(&vc4file->perfmon.idr, req->id); 246 vc4_perfmon_get(perfmon); 247 mutex_unlock(&vc4file->perfmon.lock); 248 249 if (!perfmon) 250 return -EINVAL; 251 252 if (copy_to_user(u64_to_user_ptr(req->values_ptr), perfmon->counters, 253 perfmon->ncounters * sizeof(u64))) 254 ret = -EFAULT; 255 else 256 ret = 0; 257 258 vc4_perfmon_put(perfmon); 259 return ret; 260 } 261