1 /* 2 * Copyright (C) 2013 Red Hat 3 * Author: Rob Clark <robdclark@gmail.com> 4 * 5 * Copyright (c) 2014 The Linux Foundation. All rights reserved. 6 * 7 * This program is free software; you can redistribute it and/or modify it 8 * under the terms of the GNU General Public License version 2 as published by 9 * the Free Software Foundation. 10 * 11 * This program is distributed in the hope that it will be useful, but WITHOUT 12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 14 * more details. 15 * 16 * You should have received a copy of the GNU General Public License along with 17 * this program. If not, see <http://www.gnu.org/licenses/>. 18 */ 19 20 #include "adreno_gpu.h" 21 #include "msm_gem.h" 22 #include "msm_mmu.h" 23 24 #define RB_SIZE SZ_32K 25 #define RB_BLKSIZE 16 26 27 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value) 28 { 29 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 30 31 switch (param) { 32 case MSM_PARAM_GPU_ID: 33 *value = adreno_gpu->info->revn; 34 return 0; 35 case MSM_PARAM_GMEM_SIZE: 36 *value = adreno_gpu->gmem; 37 return 0; 38 case MSM_PARAM_CHIP_ID: 39 *value = adreno_gpu->rev.patchid | 40 (adreno_gpu->rev.minor << 8) | 41 (adreno_gpu->rev.major << 16) | 42 (adreno_gpu->rev.core << 24); 43 return 0; 44 default: 45 DBG("%s: invalid param: %u", gpu->name, param); 46 return -EINVAL; 47 } 48 } 49 50 #define rbmemptr(adreno_gpu, member) \ 51 ((adreno_gpu)->memptrs_iova + offsetof(struct adreno_rbmemptrs, member)) 52 53 int adreno_hw_init(struct msm_gpu *gpu) 54 { 55 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 56 int ret; 57 58 DBG("%s", gpu->name); 59 60 ret = msm_gem_get_iova(gpu->rb->bo, gpu->id, &gpu->rb_iova); 61 if (ret) { 62 gpu->rb_iova = 0; 63 dev_err(gpu->dev->dev, "could not map ringbuffer: %d\n", ret); 64 return ret; 65 } 66 67 /* Setup REG_CP_RB_CNTL: */ 68 adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL, 69 /* size is log2(quad-words): */ 70 AXXX_CP_RB_CNTL_BUFSZ(ilog2(gpu->rb->size / 8)) | 71 AXXX_CP_RB_CNTL_BLKSZ(ilog2(RB_BLKSIZE / 8))); 72 73 /* Setup ringbuffer address: */ 74 adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_BASE, gpu->rb_iova); 75 adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR, 76 rbmemptr(adreno_gpu, rptr)); 77 78 /* Setup scratch/timestamp: */ 79 adreno_gpu_write(adreno_gpu, REG_ADRENO_SCRATCH_ADDR, 80 rbmemptr(adreno_gpu, fence)); 81 82 adreno_gpu_write(adreno_gpu, REG_ADRENO_SCRATCH_UMSK, 0x1); 83 84 return 0; 85 } 86 87 static uint32_t get_wptr(struct msm_ringbuffer *ring) 88 { 89 return ring->cur - ring->start; 90 } 91 92 uint32_t adreno_last_fence(struct msm_gpu *gpu) 93 { 94 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 95 return adreno_gpu->memptrs->fence; 96 } 97 98 void adreno_recover(struct msm_gpu *gpu) 99 { 100 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 101 struct drm_device *dev = gpu->dev; 102 int ret; 103 104 gpu->funcs->pm_suspend(gpu); 105 106 /* reset ringbuffer: */ 107 gpu->rb->cur = gpu->rb->start; 108 109 /* reset completed fence seqno, just discard anything pending: */ 110 adreno_gpu->memptrs->fence = gpu->submitted_fence; 111 adreno_gpu->memptrs->rptr = 0; 112 adreno_gpu->memptrs->wptr = 0; 113 114 gpu->funcs->pm_resume(gpu); 115 ret = gpu->funcs->hw_init(gpu); 116 if (ret) { 117 dev_err(dev->dev, "gpu hw init failed: %d\n", ret); 118 /* hmm, oh well? */ 119 } 120 } 121 122 int adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit, 123 struct msm_file_private *ctx) 124 { 125 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 126 struct msm_drm_private *priv = gpu->dev->dev_private; 127 struct msm_ringbuffer *ring = gpu->rb; 128 unsigned i, ibs = 0; 129 130 for (i = 0; i < submit->nr_cmds; i++) { 131 switch (submit->cmd[i].type) { 132 case MSM_SUBMIT_CMD_IB_TARGET_BUF: 133 /* ignore IB-targets */ 134 break; 135 case MSM_SUBMIT_CMD_CTX_RESTORE_BUF: 136 /* ignore if there has not been a ctx switch: */ 137 if (priv->lastctx == ctx) 138 break; 139 case MSM_SUBMIT_CMD_BUF: 140 OUT_PKT3(ring, CP_INDIRECT_BUFFER_PFD, 2); 141 OUT_RING(ring, submit->cmd[i].iova); 142 OUT_RING(ring, submit->cmd[i].size); 143 ibs++; 144 break; 145 } 146 } 147 148 /* on a320, at least, we seem to need to pad things out to an 149 * even number of qwords to avoid issue w/ CP hanging on wrap- 150 * around: 151 */ 152 if (ibs % 2) 153 OUT_PKT2(ring); 154 155 OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1); 156 OUT_RING(ring, submit->fence); 157 158 if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) { 159 /* Flush HLSQ lazy updates to make sure there is nothing 160 * pending for indirect loads after the timestamp has 161 * passed: 162 */ 163 OUT_PKT3(ring, CP_EVENT_WRITE, 1); 164 OUT_RING(ring, HLSQ_FLUSH); 165 166 OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1); 167 OUT_RING(ring, 0x00000000); 168 } 169 170 OUT_PKT3(ring, CP_EVENT_WRITE, 3); 171 OUT_RING(ring, CACHE_FLUSH_TS); 172 OUT_RING(ring, rbmemptr(adreno_gpu, fence)); 173 OUT_RING(ring, submit->fence); 174 175 /* we could maybe be clever and only CP_COND_EXEC the interrupt: */ 176 OUT_PKT3(ring, CP_INTERRUPT, 1); 177 OUT_RING(ring, 0x80000000); 178 179 #if 0 180 if (adreno_is_a3xx(adreno_gpu)) { 181 /* Dummy set-constant to trigger context rollover */ 182 OUT_PKT3(ring, CP_SET_CONSTANT, 2); 183 OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG)); 184 OUT_RING(ring, 0x00000000); 185 } 186 #endif 187 188 gpu->funcs->flush(gpu); 189 190 return 0; 191 } 192 193 void adreno_flush(struct msm_gpu *gpu) 194 { 195 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 196 uint32_t wptr = get_wptr(gpu->rb); 197 198 /* ensure writes to ringbuffer have hit system memory: */ 199 mb(); 200 201 adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr); 202 } 203 204 void adreno_idle(struct msm_gpu *gpu) 205 { 206 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 207 uint32_t wptr = get_wptr(gpu->rb); 208 209 /* wait for CP to drain ringbuffer: */ 210 if (spin_until(adreno_gpu->memptrs->rptr == wptr)) 211 DRM_ERROR("%s: timeout waiting to drain ringbuffer!\n", gpu->name); 212 213 /* TODO maybe we need to reset GPU here to recover from hang? */ 214 } 215 216 #ifdef CONFIG_DEBUG_FS 217 void adreno_show(struct msm_gpu *gpu, struct seq_file *m) 218 { 219 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 220 int i; 221 222 seq_printf(m, "revision: %d (%d.%d.%d.%d)\n", 223 adreno_gpu->info->revn, adreno_gpu->rev.core, 224 adreno_gpu->rev.major, adreno_gpu->rev.minor, 225 adreno_gpu->rev.patchid); 226 227 seq_printf(m, "fence: %d/%d\n", adreno_gpu->memptrs->fence, 228 gpu->submitted_fence); 229 seq_printf(m, "rptr: %d\n", adreno_gpu->memptrs->rptr); 230 seq_printf(m, "wptr: %d\n", adreno_gpu->memptrs->wptr); 231 seq_printf(m, "rb wptr: %d\n", get_wptr(gpu->rb)); 232 233 gpu->funcs->pm_resume(gpu); 234 235 /* dump these out in a form that can be parsed by demsm: */ 236 seq_printf(m, "IO:region %s 00000000 00020000\n", gpu->name); 237 for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) { 238 uint32_t start = adreno_gpu->registers[i]; 239 uint32_t end = adreno_gpu->registers[i+1]; 240 uint32_t addr; 241 242 for (addr = start; addr <= end; addr++) { 243 uint32_t val = gpu_read(gpu, addr); 244 seq_printf(m, "IO:R %08x %08x\n", addr<<2, val); 245 } 246 } 247 248 gpu->funcs->pm_suspend(gpu); 249 } 250 #endif 251 252 /* would be nice to not have to duplicate the _show() stuff with printk(): */ 253 void adreno_dump(struct msm_gpu *gpu) 254 { 255 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 256 int i; 257 258 printk("revision: %d (%d.%d.%d.%d)\n", 259 adreno_gpu->info->revn, adreno_gpu->rev.core, 260 adreno_gpu->rev.major, adreno_gpu->rev.minor, 261 adreno_gpu->rev.patchid); 262 263 printk("fence: %d/%d\n", adreno_gpu->memptrs->fence, 264 gpu->submitted_fence); 265 printk("rptr: %d\n", adreno_gpu->memptrs->rptr); 266 printk("wptr: %d\n", adreno_gpu->memptrs->wptr); 267 printk("rb wptr: %d\n", get_wptr(gpu->rb)); 268 269 /* dump these out in a form that can be parsed by demsm: */ 270 printk("IO:region %s 00000000 00020000\n", gpu->name); 271 for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) { 272 uint32_t start = adreno_gpu->registers[i]; 273 uint32_t end = adreno_gpu->registers[i+1]; 274 uint32_t addr; 275 276 for (addr = start; addr <= end; addr++) { 277 uint32_t val = gpu_read(gpu, addr); 278 printk("IO:R %08x %08x\n", addr<<2, val); 279 } 280 } 281 } 282 283 static uint32_t ring_freewords(struct msm_gpu *gpu) 284 { 285 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 286 uint32_t size = gpu->rb->size / 4; 287 uint32_t wptr = get_wptr(gpu->rb); 288 uint32_t rptr = adreno_gpu->memptrs->rptr; 289 return (rptr + (size - 1) - wptr) % size; 290 } 291 292 void adreno_wait_ring(struct msm_gpu *gpu, uint32_t ndwords) 293 { 294 if (spin_until(ring_freewords(gpu) >= ndwords)) 295 DRM_ERROR("%s: timeout waiting for ringbuffer space\n", gpu->name); 296 } 297 298 static const char *iommu_ports[] = { 299 "gfx3d_user", "gfx3d_priv", 300 "gfx3d1_user", "gfx3d1_priv", 301 }; 302 303 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev, 304 struct adreno_gpu *adreno_gpu, const struct adreno_gpu_funcs *funcs) 305 { 306 struct adreno_platform_config *config = pdev->dev.platform_data; 307 struct msm_gpu *gpu = &adreno_gpu->base; 308 struct msm_mmu *mmu; 309 int ret; 310 311 adreno_gpu->funcs = funcs; 312 adreno_gpu->info = adreno_info(config->rev); 313 adreno_gpu->gmem = adreno_gpu->info->gmem; 314 adreno_gpu->revn = adreno_gpu->info->revn; 315 adreno_gpu->rev = config->rev; 316 317 gpu->fast_rate = config->fast_rate; 318 gpu->slow_rate = config->slow_rate; 319 gpu->bus_freq = config->bus_freq; 320 #ifdef CONFIG_MSM_BUS_SCALING 321 gpu->bus_scale_table = config->bus_scale_table; 322 #endif 323 324 DBG("fast_rate=%u, slow_rate=%u, bus_freq=%u", 325 gpu->fast_rate, gpu->slow_rate, gpu->bus_freq); 326 327 ret = msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base, 328 adreno_gpu->info->name, "kgsl_3d0_reg_memory", "kgsl_3d0_irq", 329 RB_SIZE); 330 if (ret) 331 return ret; 332 333 ret = request_firmware(&adreno_gpu->pm4, adreno_gpu->info->pm4fw, drm->dev); 334 if (ret) { 335 dev_err(drm->dev, "failed to load %s PM4 firmware: %d\n", 336 adreno_gpu->info->pm4fw, ret); 337 return ret; 338 } 339 340 ret = request_firmware(&adreno_gpu->pfp, adreno_gpu->info->pfpfw, drm->dev); 341 if (ret) { 342 dev_err(drm->dev, "failed to load %s PFP firmware: %d\n", 343 adreno_gpu->info->pfpfw, ret); 344 return ret; 345 } 346 347 mmu = gpu->mmu; 348 if (mmu) { 349 ret = mmu->funcs->attach(mmu, iommu_ports, 350 ARRAY_SIZE(iommu_ports)); 351 if (ret) 352 return ret; 353 } 354 355 mutex_lock(&drm->struct_mutex); 356 adreno_gpu->memptrs_bo = msm_gem_new(drm, sizeof(*adreno_gpu->memptrs), 357 MSM_BO_UNCACHED); 358 mutex_unlock(&drm->struct_mutex); 359 if (IS_ERR(adreno_gpu->memptrs_bo)) { 360 ret = PTR_ERR(adreno_gpu->memptrs_bo); 361 adreno_gpu->memptrs_bo = NULL; 362 dev_err(drm->dev, "could not allocate memptrs: %d\n", ret); 363 return ret; 364 } 365 366 adreno_gpu->memptrs = msm_gem_vaddr(adreno_gpu->memptrs_bo); 367 if (!adreno_gpu->memptrs) { 368 dev_err(drm->dev, "could not vmap memptrs\n"); 369 return -ENOMEM; 370 } 371 372 ret = msm_gem_get_iova(adreno_gpu->memptrs_bo, gpu->id, 373 &adreno_gpu->memptrs_iova); 374 if (ret) { 375 dev_err(drm->dev, "could not map memptrs: %d\n", ret); 376 return ret; 377 } 378 379 return 0; 380 } 381 382 void adreno_gpu_cleanup(struct adreno_gpu *gpu) 383 { 384 if (gpu->memptrs_bo) { 385 if (gpu->memptrs_iova) 386 msm_gem_put_iova(gpu->memptrs_bo, gpu->base.id); 387 drm_gem_object_unreference(gpu->memptrs_bo); 388 } 389 release_firmware(gpu->pm4); 390 release_firmware(gpu->pfp); 391 msm_gpu_cleanup(&gpu->base); 392 } 393