1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */ 3 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */ 4 /* Copyright 2019 Collabora ltd. */ 5 #include <linux/bitfield.h> 6 #include <linux/bitmap.h> 7 #include <linux/delay.h> 8 #include <linux/dma-mapping.h> 9 #include <linux/interrupt.h> 10 #include <linux/io.h> 11 #include <linux/iopoll.h> 12 #include <linux/platform_device.h> 13 14 #include "panfrost_device.h" 15 #include "panfrost_features.h" 16 #include "panfrost_issues.h" 17 #include "panfrost_gpu.h" 18 #include "panfrost_perfcnt.h" 19 #include "panfrost_regs.h" 20 21 static irqreturn_t panfrost_gpu_irq_handler(int irq, void *data) 22 { 23 struct panfrost_device *pfdev = data; 24 u32 state = gpu_read(pfdev, GPU_INT_STAT); 25 u32 fault_status = gpu_read(pfdev, GPU_FAULT_STATUS); 26 27 if (!state) 28 return IRQ_NONE; 29 30 if (state & GPU_IRQ_MASK_ERROR) { 31 u64 address = (u64) gpu_read(pfdev, GPU_FAULT_ADDRESS_HI) << 32; 32 address |= gpu_read(pfdev, GPU_FAULT_ADDRESS_LO); 33 34 dev_warn(pfdev->dev, "GPU Fault 0x%08x (%s) at 0x%016llx\n", 35 fault_status & 0xFF, panfrost_exception_name(pfdev, fault_status), 36 address); 37 38 if (state & GPU_IRQ_MULTIPLE_FAULT) 39 dev_warn(pfdev->dev, "There were multiple GPU faults - some have not been reported\n"); 40 41 gpu_write(pfdev, GPU_INT_MASK, 0); 42 } 43 44 if (state & GPU_IRQ_PERFCNT_SAMPLE_COMPLETED) 45 panfrost_perfcnt_sample_done(pfdev); 46 47 if (state & GPU_IRQ_CLEAN_CACHES_COMPLETED) 48 panfrost_perfcnt_clean_cache_done(pfdev); 49 50 gpu_write(pfdev, GPU_INT_CLEAR, state); 51 52 return IRQ_HANDLED; 53 } 54 55 int panfrost_gpu_soft_reset(struct panfrost_device *pfdev) 56 { 57 int ret; 58 u32 val; 59 60 gpu_write(pfdev, GPU_INT_MASK, 0); 61 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_RESET_COMPLETED); 62 gpu_write(pfdev, GPU_CMD, GPU_CMD_SOFT_RESET); 63 64 ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT, 65 val, val & GPU_IRQ_RESET_COMPLETED, 100, 10000); 66 67 if (ret) { 68 dev_err(pfdev->dev, "gpu soft reset timed out\n"); 69 return ret; 70 } 71 72 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL); 73 gpu_write(pfdev, GPU_INT_MASK, GPU_IRQ_MASK_ALL); 74 75 return 0; 76 } 77 78 static void panfrost_gpu_init_quirks(struct panfrost_device *pfdev) 79 { 80 u32 quirks = 0; 81 82 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8443) || 83 panfrost_has_hw_issue(pfdev, HW_ISSUE_11035)) 84 quirks |= SC_LS_PAUSEBUFFER_DISABLE; 85 86 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10327)) 87 quirks |= SC_SDC_DISABLE_OQ_DISCARD; 88 89 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10797)) 90 quirks |= SC_ENABLE_TEXGRD_FLAGS; 91 92 if (!panfrost_has_hw_issue(pfdev, GPUCORE_1619)) { 93 if (panfrost_model_cmp(pfdev, 0x750) < 0) /* T60x, T62x, T72x */ 94 quirks |= SC_LS_ATTR_CHECK_DISABLE; 95 else if (panfrost_model_cmp(pfdev, 0x880) <= 0) /* T76x, T8xx */ 96 quirks |= SC_LS_ALLOW_ATTR_TYPES; 97 } 98 99 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_TLS_HASHING)) 100 quirks |= SC_TLS_HASH_ENABLE; 101 102 if (quirks) 103 gpu_write(pfdev, GPU_SHADER_CONFIG, quirks); 104 105 106 quirks = gpu_read(pfdev, GPU_TILER_CONFIG); 107 108 /* Set tiler clock gate override if required */ 109 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_T76X_3953)) 110 quirks |= TC_CLOCK_GATE_OVERRIDE; 111 112 gpu_write(pfdev, GPU_TILER_CONFIG, quirks); 113 114 115 quirks = gpu_read(pfdev, GPU_L2_MMU_CONFIG); 116 117 /* Limit read & write ID width for AXI */ 118 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_3BIT_EXT_RW_L2_MMU_CONFIG)) 119 quirks &= ~(L2_MMU_CONFIG_3BIT_LIMIT_EXTERNAL_READS | 120 L2_MMU_CONFIG_3BIT_LIMIT_EXTERNAL_WRITES); 121 else 122 quirks &= ~(L2_MMU_CONFIG_LIMIT_EXTERNAL_READS | 123 L2_MMU_CONFIG_LIMIT_EXTERNAL_WRITES); 124 125 gpu_write(pfdev, GPU_L2_MMU_CONFIG, quirks); 126 127 quirks = 0; 128 if ((panfrost_model_eq(pfdev, 0x860) || panfrost_model_eq(pfdev, 0x880)) && 129 pfdev->features.revision >= 0x2000) 130 quirks |= JM_MAX_JOB_THROTTLE_LIMIT << JM_JOB_THROTTLE_LIMIT_SHIFT; 131 else if (panfrost_model_eq(pfdev, 0x6000) && 132 pfdev->features.coherency_features == COHERENCY_ACE) 133 quirks |= (COHERENCY_ACE_LITE | COHERENCY_ACE) << 134 JM_FORCE_COHERENCY_FEATURES_SHIFT; 135 136 if (quirks) 137 gpu_write(pfdev, GPU_JM_CONFIG, quirks); 138 } 139 140 #define MAX_HW_REVS 6 141 142 struct panfrost_model { 143 const char *name; 144 u32 id; 145 u32 id_mask; 146 u64 features; 147 u64 issues; 148 struct { 149 u32 revision; 150 u64 issues; 151 } revs[MAX_HW_REVS]; 152 }; 153 154 #define GPU_MODEL(_name, _id, ...) \ 155 {\ 156 .name = __stringify(_name), \ 157 .id = _id, \ 158 .features = hw_features_##_name, \ 159 .issues = hw_issues_##_name, \ 160 .revs = { __VA_ARGS__ }, \ 161 } 162 163 #define GPU_REV_EXT(name, _rev, _p, _s, stat) \ 164 {\ 165 .revision = (_rev) << 12 | (_p) << 4 | (_s), \ 166 .issues = hw_issues_##name##_r##_rev##p##_p##stat, \ 167 } 168 #define GPU_REV(name, r, p) GPU_REV_EXT(name, r, p, 0, ) 169 170 static const struct panfrost_model gpu_models[] = { 171 /* T60x has an oddball version */ 172 GPU_MODEL(t600, 0x600, 173 GPU_REV_EXT(t600, 0, 0, 1, _15dev0)), 174 GPU_MODEL(t620, 0x620, 175 GPU_REV(t620, 0, 1), GPU_REV(t620, 1, 0)), 176 GPU_MODEL(t720, 0x720), 177 GPU_MODEL(t760, 0x750, 178 GPU_REV(t760, 0, 0), GPU_REV(t760, 0, 1), 179 GPU_REV_EXT(t760, 0, 1, 0, _50rel0), 180 GPU_REV(t760, 0, 2), GPU_REV(t760, 0, 3)), 181 GPU_MODEL(t820, 0x820), 182 GPU_MODEL(t830, 0x830), 183 GPU_MODEL(t860, 0x860), 184 GPU_MODEL(t880, 0x880), 185 186 GPU_MODEL(g71, 0x6000, 187 GPU_REV_EXT(g71, 0, 0, 1, _05dev0)), 188 GPU_MODEL(g72, 0x6001), 189 GPU_MODEL(g51, 0x7000), 190 GPU_MODEL(g76, 0x7001), 191 GPU_MODEL(g52, 0x7002), 192 GPU_MODEL(g31, 0x7003, 193 GPU_REV(g31, 1, 0)), 194 }; 195 196 static void panfrost_gpu_init_features(struct panfrost_device *pfdev) 197 { 198 u32 gpu_id, num_js, major, minor, status, rev; 199 const char *name = "unknown"; 200 u64 hw_feat = 0; 201 u64 hw_issues = hw_issues_all; 202 const struct panfrost_model *model; 203 int i; 204 205 pfdev->features.l2_features = gpu_read(pfdev, GPU_L2_FEATURES); 206 pfdev->features.core_features = gpu_read(pfdev, GPU_CORE_FEATURES); 207 pfdev->features.tiler_features = gpu_read(pfdev, GPU_TILER_FEATURES); 208 pfdev->features.mem_features = gpu_read(pfdev, GPU_MEM_FEATURES); 209 pfdev->features.mmu_features = gpu_read(pfdev, GPU_MMU_FEATURES); 210 pfdev->features.thread_features = gpu_read(pfdev, GPU_THREAD_FEATURES); 211 pfdev->features.max_threads = gpu_read(pfdev, GPU_THREAD_MAX_THREADS); 212 pfdev->features.thread_max_workgroup_sz = gpu_read(pfdev, GPU_THREAD_MAX_WORKGROUP_SIZE); 213 pfdev->features.thread_max_barrier_sz = gpu_read(pfdev, GPU_THREAD_MAX_BARRIER_SIZE); 214 pfdev->features.coherency_features = gpu_read(pfdev, GPU_COHERENCY_FEATURES); 215 for (i = 0; i < 4; i++) 216 pfdev->features.texture_features[i] = gpu_read(pfdev, GPU_TEXTURE_FEATURES(i)); 217 218 pfdev->features.as_present = gpu_read(pfdev, GPU_AS_PRESENT); 219 220 pfdev->features.js_present = gpu_read(pfdev, GPU_JS_PRESENT); 221 num_js = hweight32(pfdev->features.js_present); 222 for (i = 0; i < num_js; i++) 223 pfdev->features.js_features[i] = gpu_read(pfdev, GPU_JS_FEATURES(i)); 224 225 pfdev->features.shader_present = gpu_read(pfdev, GPU_SHADER_PRESENT_LO); 226 pfdev->features.shader_present |= (u64)gpu_read(pfdev, GPU_SHADER_PRESENT_HI) << 32; 227 228 pfdev->features.tiler_present = gpu_read(pfdev, GPU_TILER_PRESENT_LO); 229 pfdev->features.tiler_present |= (u64)gpu_read(pfdev, GPU_TILER_PRESENT_HI) << 32; 230 231 pfdev->features.l2_present = gpu_read(pfdev, GPU_L2_PRESENT_LO); 232 pfdev->features.l2_present |= (u64)gpu_read(pfdev, GPU_L2_PRESENT_HI) << 32; 233 pfdev->features.nr_core_groups = hweight64(pfdev->features.l2_present); 234 235 pfdev->features.stack_present = gpu_read(pfdev, GPU_STACK_PRESENT_LO); 236 pfdev->features.stack_present |= (u64)gpu_read(pfdev, GPU_STACK_PRESENT_HI) << 32; 237 238 pfdev->features.thread_tls_alloc = gpu_read(pfdev, GPU_THREAD_TLS_ALLOC); 239 240 gpu_id = gpu_read(pfdev, GPU_ID); 241 pfdev->features.revision = gpu_id & 0xffff; 242 pfdev->features.id = gpu_id >> 16; 243 244 /* The T60x has an oddball ID value. Fix it up to the standard Midgard 245 * format so we (and userspace) don't have to special case it. 246 */ 247 if (pfdev->features.id == 0x6956) 248 pfdev->features.id = 0x0600; 249 250 major = (pfdev->features.revision >> 12) & 0xf; 251 minor = (pfdev->features.revision >> 4) & 0xff; 252 status = pfdev->features.revision & 0xf; 253 rev = pfdev->features.revision; 254 255 gpu_id = pfdev->features.id; 256 257 for (model = gpu_models; model->name; model++) { 258 int best = -1; 259 260 if (!panfrost_model_eq(pfdev, model->id)) 261 continue; 262 263 name = model->name; 264 hw_feat = model->features; 265 hw_issues |= model->issues; 266 for (i = 0; i < MAX_HW_REVS; i++) { 267 if (model->revs[i].revision == rev) { 268 best = i; 269 break; 270 } else if (model->revs[i].revision == (rev & ~0xf)) 271 best = i; 272 } 273 274 if (best >= 0) 275 hw_issues |= model->revs[best].issues; 276 277 break; 278 } 279 280 bitmap_from_u64(pfdev->features.hw_features, hw_feat); 281 bitmap_from_u64(pfdev->features.hw_issues, hw_issues); 282 283 dev_info(pfdev->dev, "mali-%s id 0x%x major 0x%x minor 0x%x status 0x%x", 284 name, gpu_id, major, minor, status); 285 dev_info(pfdev->dev, "features: %64pb, issues: %64pb", 286 pfdev->features.hw_features, 287 pfdev->features.hw_issues); 288 289 dev_info(pfdev->dev, "Features: L2:0x%08x Shader:0x%08x Tiler:0x%08x Mem:0x%0x MMU:0x%08x AS:0x%x JS:0x%x", 290 pfdev->features.l2_features, 291 pfdev->features.core_features, 292 pfdev->features.tiler_features, 293 pfdev->features.mem_features, 294 pfdev->features.mmu_features, 295 pfdev->features.as_present, 296 pfdev->features.js_present); 297 298 dev_info(pfdev->dev, "shader_present=0x%0llx l2_present=0x%0llx", 299 pfdev->features.shader_present, pfdev->features.l2_present); 300 } 301 302 void panfrost_gpu_power_on(struct panfrost_device *pfdev) 303 { 304 int ret; 305 u32 val; 306 307 /* Just turn on everything for now */ 308 gpu_write(pfdev, L2_PWRON_LO, pfdev->features.l2_present); 309 ret = readl_relaxed_poll_timeout(pfdev->iomem + L2_READY_LO, 310 val, val == pfdev->features.l2_present, 100, 1000); 311 312 gpu_write(pfdev, STACK_PWRON_LO, pfdev->features.stack_present); 313 ret |= readl_relaxed_poll_timeout(pfdev->iomem + STACK_READY_LO, 314 val, val == pfdev->features.stack_present, 100, 1000); 315 316 gpu_write(pfdev, SHADER_PWRON_LO, pfdev->features.shader_present); 317 ret |= readl_relaxed_poll_timeout(pfdev->iomem + SHADER_READY_LO, 318 val, val == pfdev->features.shader_present, 100, 1000); 319 320 gpu_write(pfdev, TILER_PWRON_LO, pfdev->features.tiler_present); 321 ret |= readl_relaxed_poll_timeout(pfdev->iomem + TILER_READY_LO, 322 val, val == pfdev->features.tiler_present, 100, 1000); 323 324 if (ret) 325 dev_err(pfdev->dev, "error powering up gpu"); 326 } 327 328 void panfrost_gpu_power_off(struct panfrost_device *pfdev) 329 { 330 gpu_write(pfdev, TILER_PWROFF_LO, 0); 331 gpu_write(pfdev, SHADER_PWROFF_LO, 0); 332 gpu_write(pfdev, STACK_PWROFF_LO, 0); 333 gpu_write(pfdev, L2_PWROFF_LO, 0); 334 } 335 336 int panfrost_gpu_init(struct panfrost_device *pfdev) 337 { 338 int err, irq; 339 340 err = panfrost_gpu_soft_reset(pfdev); 341 if (err) 342 return err; 343 344 panfrost_gpu_init_features(pfdev); 345 346 dma_set_mask_and_coherent(pfdev->dev, 347 DMA_BIT_MASK(FIELD_GET(0xff00, pfdev->features.mmu_features))); 348 349 irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "gpu"); 350 if (irq <= 0) 351 return -ENODEV; 352 353 err = devm_request_irq(pfdev->dev, irq, panfrost_gpu_irq_handler, 354 IRQF_SHARED, "gpu", pfdev); 355 if (err) { 356 dev_err(pfdev->dev, "failed to request gpu irq"); 357 return err; 358 } 359 360 panfrost_gpu_init_quirks(pfdev); 361 panfrost_gpu_power_on(pfdev); 362 363 return 0; 364 } 365 366 void panfrost_gpu_fini(struct panfrost_device *pfdev) 367 { 368 panfrost_gpu_power_off(pfdev); 369 } 370 371 u32 panfrost_gpu_get_latest_flush_id(struct panfrost_device *pfdev) 372 { 373 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) 374 return gpu_read(pfdev, GPU_LATEST_FLUSH_ID); 375 return 0; 376 } 377