1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright 2019 Collabora Ltd */ 3 4 #include <drm/drm_file.h> 5 #include <drm/drm_gem_shmem_helper.h> 6 #include <drm/panfrost_drm.h> 7 #include <linux/completion.h> 8 #include <linux/iopoll.h> 9 #include <linux/pm_runtime.h> 10 #include <linux/slab.h> 11 #include <linux/uaccess.h> 12 13 #include "panfrost_device.h" 14 #include "panfrost_features.h" 15 #include "panfrost_gem.h" 16 #include "panfrost_issues.h" 17 #include "panfrost_job.h" 18 #include "panfrost_mmu.h" 19 #include "panfrost_perfcnt.h" 20 #include "panfrost_regs.h" 21 22 #define COUNTERS_PER_BLOCK 64 23 #define BYTES_PER_COUNTER 4 24 #define BLOCKS_PER_COREGROUP 8 25 #define V4_SHADERS_PER_COREGROUP 4 26 27 struct panfrost_perfcnt { 28 struct panfrost_gem_mapping *mapping; 29 size_t bosize; 30 void *buf; 31 struct panfrost_file_priv *user; 32 struct mutex lock; 33 struct completion dump_comp; 34 }; 35 36 void panfrost_perfcnt_clean_cache_done(struct panfrost_device *pfdev) 37 { 38 complete(&pfdev->perfcnt->dump_comp); 39 } 40 41 void panfrost_perfcnt_sample_done(struct panfrost_device *pfdev) 42 { 43 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_CACHES); 44 } 45 46 static int panfrost_perfcnt_dump_locked(struct panfrost_device *pfdev) 47 { 48 u64 gpuva; 49 int ret; 50 51 reinit_completion(&pfdev->perfcnt->dump_comp); 52 gpuva = pfdev->perfcnt->mapping->mmnode.start << PAGE_SHIFT; 53 gpu_write(pfdev, GPU_PERFCNT_BASE_LO, gpuva); 54 gpu_write(pfdev, GPU_PERFCNT_BASE_HI, gpuva >> 32); 55 gpu_write(pfdev, GPU_INT_CLEAR, 56 GPU_IRQ_CLEAN_CACHES_COMPLETED | 57 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED); 58 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_SAMPLE); 59 ret = wait_for_completion_interruptible_timeout(&pfdev->perfcnt->dump_comp, 60 msecs_to_jiffies(1000)); 61 if (!ret) 62 ret = -ETIMEDOUT; 63 else if (ret > 0) 64 ret = 0; 65 66 return ret; 67 } 68 69 static int panfrost_perfcnt_enable_locked(struct panfrost_device *pfdev, 70 struct drm_file *file_priv, 71 unsigned int counterset) 72 { 73 struct panfrost_file_priv *user = file_priv->driver_priv; 74 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt; 75 struct drm_gem_shmem_object *bo; 76 u32 cfg, as; 77 int ret; 78 79 if (user == perfcnt->user) 80 return 0; 81 else if (perfcnt->user) 82 return -EBUSY; 83 84 ret = pm_runtime_get_sync(pfdev->dev); 85 if (ret < 0) 86 return ret; 87 88 bo = drm_gem_shmem_create(pfdev->ddev, perfcnt->bosize); 89 if (IS_ERR(bo)) 90 return PTR_ERR(bo); 91 92 /* Map the perfcnt buf in the address space attached to file_priv. */ 93 ret = panfrost_gem_open(&bo->base, file_priv); 94 if (ret) 95 goto err_put_bo; 96 97 perfcnt->mapping = panfrost_gem_mapping_get(to_panfrost_bo(&bo->base), 98 user); 99 if (!perfcnt->mapping) { 100 ret = -EINVAL; 101 goto err_close_bo; 102 } 103 104 perfcnt->buf = drm_gem_shmem_vmap(&bo->base); 105 if (IS_ERR(perfcnt->buf)) { 106 ret = PTR_ERR(perfcnt->buf); 107 goto err_put_mapping; 108 } 109 110 /* 111 * Invalidate the cache and clear the counters to start from a fresh 112 * state. 113 */ 114 reinit_completion(&pfdev->perfcnt->dump_comp); 115 gpu_write(pfdev, GPU_INT_CLEAR, 116 GPU_IRQ_CLEAN_CACHES_COMPLETED | 117 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED); 118 gpu_write(pfdev, GPU_CMD, GPU_CMD_PERFCNT_CLEAR); 119 gpu_write(pfdev, GPU_CMD, GPU_CMD_CLEAN_INV_CACHES); 120 ret = wait_for_completion_timeout(&pfdev->perfcnt->dump_comp, 121 msecs_to_jiffies(1000)); 122 if (!ret) { 123 ret = -ETIMEDOUT; 124 goto err_vunmap; 125 } 126 127 perfcnt->user = user; 128 129 as = panfrost_mmu_as_get(pfdev, perfcnt->mapping->mmu); 130 cfg = GPU_PERFCNT_CFG_AS(as) | 131 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_MANUAL); 132 133 /* 134 * Bifrost GPUs have 2 set of counters, but we're only interested by 135 * the first one for now. 136 */ 137 if (panfrost_model_is_bifrost(pfdev)) 138 cfg |= GPU_PERFCNT_CFG_SETSEL(counterset); 139 140 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0xffffffff); 141 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0xffffffff); 142 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0xffffffff); 143 144 /* 145 * Due to PRLAM-8186 we need to disable the Tiler before we enable HW 146 * counters. 147 */ 148 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186)) 149 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0); 150 else 151 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff); 152 153 gpu_write(pfdev, GPU_PERFCNT_CFG, cfg); 154 155 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8186)) 156 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0xffffffff); 157 158 /* The BO ref is retained by the mapping. */ 159 drm_gem_object_put_unlocked(&bo->base); 160 161 return 0; 162 163 err_vunmap: 164 drm_gem_shmem_vunmap(&bo->base, perfcnt->buf); 165 err_put_mapping: 166 panfrost_gem_mapping_put(perfcnt->mapping); 167 err_close_bo: 168 panfrost_gem_close(&bo->base, file_priv); 169 err_put_bo: 170 drm_gem_object_put_unlocked(&bo->base); 171 return ret; 172 } 173 174 static int panfrost_perfcnt_disable_locked(struct panfrost_device *pfdev, 175 struct drm_file *file_priv) 176 { 177 struct panfrost_file_priv *user = file_priv->driver_priv; 178 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt; 179 180 if (user != perfcnt->user) 181 return -EINVAL; 182 183 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0x0); 184 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0x0); 185 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0x0); 186 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0); 187 gpu_write(pfdev, GPU_PERFCNT_CFG, 188 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF)); 189 190 perfcnt->user = NULL; 191 drm_gem_shmem_vunmap(&perfcnt->mapping->obj->base.base, perfcnt->buf); 192 perfcnt->buf = NULL; 193 panfrost_gem_close(&perfcnt->mapping->obj->base.base, file_priv); 194 panfrost_mmu_as_put(pfdev, perfcnt->mapping->mmu); 195 panfrost_gem_mapping_put(perfcnt->mapping); 196 perfcnt->mapping = NULL; 197 pm_runtime_mark_last_busy(pfdev->dev); 198 pm_runtime_put_autosuspend(pfdev->dev); 199 200 return 0; 201 } 202 203 int panfrost_ioctl_perfcnt_enable(struct drm_device *dev, void *data, 204 struct drm_file *file_priv) 205 { 206 struct panfrost_device *pfdev = dev->dev_private; 207 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt; 208 struct drm_panfrost_perfcnt_enable *req = data; 209 int ret; 210 211 ret = panfrost_unstable_ioctl_check(); 212 if (ret) 213 return ret; 214 215 /* Only Bifrost GPUs have 2 set of counters. */ 216 if (req->counterset > (panfrost_model_is_bifrost(pfdev) ? 1 : 0)) 217 return -EINVAL; 218 219 mutex_lock(&perfcnt->lock); 220 if (req->enable) 221 ret = panfrost_perfcnt_enable_locked(pfdev, file_priv, 222 req->counterset); 223 else 224 ret = panfrost_perfcnt_disable_locked(pfdev, file_priv); 225 mutex_unlock(&perfcnt->lock); 226 227 return ret; 228 } 229 230 int panfrost_ioctl_perfcnt_dump(struct drm_device *dev, void *data, 231 struct drm_file *file_priv) 232 { 233 struct panfrost_device *pfdev = dev->dev_private; 234 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt; 235 struct drm_panfrost_perfcnt_dump *req = data; 236 void __user *user_ptr = (void __user *)(uintptr_t)req->buf_ptr; 237 int ret; 238 239 ret = panfrost_unstable_ioctl_check(); 240 if (ret) 241 return ret; 242 243 mutex_lock(&perfcnt->lock); 244 if (perfcnt->user != file_priv->driver_priv) { 245 ret = -EINVAL; 246 goto out; 247 } 248 249 ret = panfrost_perfcnt_dump_locked(pfdev); 250 if (ret) 251 goto out; 252 253 if (copy_to_user(user_ptr, perfcnt->buf, perfcnt->bosize)) 254 ret = -EFAULT; 255 256 out: 257 mutex_unlock(&perfcnt->lock); 258 259 return ret; 260 } 261 262 void panfrost_perfcnt_close(struct drm_file *file_priv) 263 { 264 struct panfrost_file_priv *pfile = file_priv->driver_priv; 265 struct panfrost_device *pfdev = pfile->pfdev; 266 struct panfrost_perfcnt *perfcnt = pfdev->perfcnt; 267 268 pm_runtime_get_sync(pfdev->dev); 269 mutex_lock(&perfcnt->lock); 270 if (perfcnt->user == pfile) 271 panfrost_perfcnt_disable_locked(pfdev, file_priv); 272 mutex_unlock(&perfcnt->lock); 273 pm_runtime_mark_last_busy(pfdev->dev); 274 pm_runtime_put_autosuspend(pfdev->dev); 275 } 276 277 int panfrost_perfcnt_init(struct panfrost_device *pfdev) 278 { 279 struct panfrost_perfcnt *perfcnt; 280 size_t size; 281 282 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_V4)) { 283 unsigned int ncoregroups; 284 285 ncoregroups = hweight64(pfdev->features.l2_present); 286 size = ncoregroups * BLOCKS_PER_COREGROUP * 287 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER; 288 } else { 289 unsigned int nl2c, ncores; 290 291 /* 292 * TODO: define a macro to extract the number of l2 caches from 293 * mem_features. 294 */ 295 nl2c = ((pfdev->features.mem_features >> 8) & GENMASK(3, 0)) + 1; 296 297 /* 298 * shader_present might be sparse, but the counters layout 299 * forces to dump unused regions too, hence the fls64() call 300 * instead of hweight64(). 301 */ 302 ncores = fls64(pfdev->features.shader_present); 303 304 /* 305 * There's always one JM and one Tiler block, hence the '+ 2' 306 * here. 307 */ 308 size = (nl2c + ncores + 2) * 309 COUNTERS_PER_BLOCK * BYTES_PER_COUNTER; 310 } 311 312 perfcnt = devm_kzalloc(pfdev->dev, sizeof(*perfcnt), GFP_KERNEL); 313 if (!perfcnt) 314 return -ENOMEM; 315 316 perfcnt->bosize = size; 317 318 /* Start with everything disabled. */ 319 gpu_write(pfdev, GPU_PERFCNT_CFG, 320 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF)); 321 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0); 322 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0); 323 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0); 324 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0); 325 326 init_completion(&perfcnt->dump_comp); 327 mutex_init(&perfcnt->lock); 328 pfdev->perfcnt = perfcnt; 329 330 return 0; 331 } 332 333 void panfrost_perfcnt_fini(struct panfrost_device *pfdev) 334 { 335 /* Disable everything before leaving. */ 336 gpu_write(pfdev, GPU_PERFCNT_CFG, 337 GPU_PERFCNT_CFG_MODE(GPU_PERFCNT_CFG_MODE_OFF)); 338 gpu_write(pfdev, GPU_PRFCNT_JM_EN, 0); 339 gpu_write(pfdev, GPU_PRFCNT_SHADER_EN, 0); 340 gpu_write(pfdev, GPU_PRFCNT_MMU_L2_EN, 0); 341 gpu_write(pfdev, GPU_PRFCNT_TILER_EN, 0); 342 } 343