1 // SPDX-License-Identifier: GPL-2.0 2 /* Copyright (c) 2017-2019 The Linux Foundation. All rights reserved. */ 3 4 #include <linux/clk.h> 5 #include <linux/interconnect.h> 6 #include <linux/pm_domain.h> 7 #include <linux/pm_opp.h> 8 #include <soc/qcom/cmd-db.h> 9 10 #include "a6xx_gpu.h" 11 #include "a6xx_gmu.xml.h" 12 13 static void a6xx_gmu_fault(struct a6xx_gmu *gmu) 14 { 15 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 16 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 17 struct msm_gpu *gpu = &adreno_gpu->base; 18 struct drm_device *dev = gpu->dev; 19 struct msm_drm_private *priv = dev->dev_private; 20 21 /* FIXME: add a banner here */ 22 gmu->hung = true; 23 24 /* Turn off the hangcheck timer while we are resetting */ 25 del_timer(&gpu->hangcheck_timer); 26 27 /* Queue the GPU handler because we need to treat this as a recovery */ 28 queue_work(priv->wq, &gpu->recover_work); 29 } 30 31 static irqreturn_t a6xx_gmu_irq(int irq, void *data) 32 { 33 struct a6xx_gmu *gmu = data; 34 u32 status; 35 36 status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS); 37 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status); 38 39 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) { 40 dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n"); 41 42 a6xx_gmu_fault(gmu); 43 } 44 45 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR) 46 dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n"); 47 48 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR) 49 dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n", 50 gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS)); 51 52 return IRQ_HANDLED; 53 } 54 55 static irqreturn_t a6xx_hfi_irq(int irq, void *data) 56 { 57 struct a6xx_gmu *gmu = data; 58 u32 status; 59 60 status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO); 61 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status); 62 63 if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) { 64 dev_err_ratelimited(gmu->dev, "GMU firmware fault\n"); 65 66 a6xx_gmu_fault(gmu); 67 } 68 69 return IRQ_HANDLED; 70 } 71 72 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu) 73 { 74 u32 val; 75 76 /* This can be called from gpu state code so make sure GMU is valid */ 77 if (!gmu->initialized) 78 return false; 79 80 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS); 81 82 return !(val & 83 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF | 84 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF)); 85 } 86 87 /* Check to see if the GX rail is still powered */ 88 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu) 89 { 90 u32 val; 91 92 /* This can be called from gpu state code so make sure GMU is valid */ 93 if (!gmu->initialized) 94 return false; 95 96 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS); 97 98 return !(val & 99 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF | 100 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF)); 101 } 102 103 static void __a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index) 104 { 105 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 106 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 107 struct msm_gpu *gpu = &adreno_gpu->base; 108 int ret; 109 110 gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0); 111 112 gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING, 113 ((3 & 0xf) << 28) | index); 114 115 /* 116 * Send an invalid index as a vote for the bus bandwidth and let the 117 * firmware decide on the right vote 118 */ 119 gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff); 120 121 /* Set and clear the OOB for DCVS to trigger the GMU */ 122 a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET); 123 a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET); 124 125 ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN); 126 if (ret) 127 dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret); 128 129 gmu->freq = gmu->gpu_freqs[index]; 130 131 /* 132 * Eventually we will want to scale the path vote with the frequency but 133 * for now leave it at max so that the performance is nominal. 134 */ 135 icc_set_bw(gpu->icc_path, 0, MBps_to_icc(7216)); 136 } 137 138 void a6xx_gmu_set_freq(struct msm_gpu *gpu, unsigned long freq) 139 { 140 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 141 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu); 142 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 143 u32 perf_index = 0; 144 145 if (freq == gmu->freq) 146 return; 147 148 for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++) 149 if (freq == gmu->gpu_freqs[perf_index]) 150 break; 151 152 gmu->current_perf_index = perf_index; 153 154 __a6xx_gmu_set_freq(gmu, perf_index); 155 } 156 157 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu) 158 { 159 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 160 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu); 161 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 162 163 return gmu->freq; 164 } 165 166 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu) 167 { 168 u32 val; 169 int local = gmu->idle_level; 170 171 /* SPTP and IFPC both report as IFPC */ 172 if (gmu->idle_level == GMU_IDLE_STATE_SPTP) 173 local = GMU_IDLE_STATE_IFPC; 174 175 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); 176 177 if (val == local) { 178 if (gmu->idle_level != GMU_IDLE_STATE_IFPC || 179 !a6xx_gmu_gx_is_on(gmu)) 180 return true; 181 } 182 183 return false; 184 } 185 186 /* Wait for the GMU to get to its most idle state */ 187 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu) 188 { 189 return spin_until(a6xx_gmu_check_idle_level(gmu)); 190 } 191 192 static int a6xx_gmu_start(struct a6xx_gmu *gmu) 193 { 194 int ret; 195 u32 val; 196 197 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1); 198 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0); 199 200 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val, 201 val == 0xbabeface, 100, 10000); 202 203 if (ret) 204 DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n"); 205 206 return ret; 207 } 208 209 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu) 210 { 211 u32 val; 212 int ret; 213 214 gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1); 215 216 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val, 217 val & 1, 100, 10000); 218 if (ret) 219 DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n"); 220 221 return ret; 222 } 223 224 /* Trigger a OOB (out of band) request to the GMU */ 225 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state) 226 { 227 int ret; 228 u32 val; 229 int request, ack; 230 const char *name; 231 232 switch (state) { 233 case GMU_OOB_GPU_SET: 234 request = GMU_OOB_GPU_SET_REQUEST; 235 ack = GMU_OOB_GPU_SET_ACK; 236 name = "GPU_SET"; 237 break; 238 case GMU_OOB_BOOT_SLUMBER: 239 request = GMU_OOB_BOOT_SLUMBER_REQUEST; 240 ack = GMU_OOB_BOOT_SLUMBER_ACK; 241 name = "BOOT_SLUMBER"; 242 break; 243 case GMU_OOB_DCVS_SET: 244 request = GMU_OOB_DCVS_REQUEST; 245 ack = GMU_OOB_DCVS_ACK; 246 name = "GPU_DCVS"; 247 break; 248 default: 249 return -EINVAL; 250 } 251 252 /* Trigger the equested OOB operation */ 253 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request); 254 255 /* Wait for the acknowledge interrupt */ 256 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val, 257 val & (1 << ack), 100, 10000); 258 259 if (ret) 260 DRM_DEV_ERROR(gmu->dev, 261 "Timeout waiting for GMU OOB set %s: 0x%x\n", 262 name, 263 gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO)); 264 265 /* Clear the acknowledge interrupt */ 266 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack); 267 268 return ret; 269 } 270 271 /* Clear a pending OOB state in the GMU */ 272 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state) 273 { 274 switch (state) { 275 case GMU_OOB_GPU_SET: 276 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 277 1 << GMU_OOB_GPU_SET_CLEAR); 278 break; 279 case GMU_OOB_BOOT_SLUMBER: 280 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 281 1 << GMU_OOB_BOOT_SLUMBER_CLEAR); 282 break; 283 case GMU_OOB_DCVS_SET: 284 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 285 1 << GMU_OOB_DCVS_CLEAR); 286 break; 287 } 288 } 289 290 /* Enable CPU control of SPTP power power collapse */ 291 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu) 292 { 293 int ret; 294 u32 val; 295 296 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000); 297 298 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val, 299 (val & 0x38) == 0x28, 1, 100); 300 301 if (ret) { 302 DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n", 303 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS)); 304 } 305 306 return 0; 307 } 308 309 /* Disable CPU control of SPTP power power collapse */ 310 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu) 311 { 312 u32 val; 313 int ret; 314 315 /* Make sure retention is on */ 316 gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11)); 317 318 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001); 319 320 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val, 321 (val & 0x04), 100, 10000); 322 323 if (ret) 324 DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n", 325 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS)); 326 } 327 328 /* Let the GMU know we are starting a boot sequence */ 329 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu) 330 { 331 u32 vote; 332 333 /* Let the GMU know we are getting ready for boot */ 334 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0); 335 336 /* Choose the "default" power level as the highest available */ 337 vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1]; 338 339 gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff); 340 gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff); 341 342 /* Let the GMU know the boot sequence has started */ 343 return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER); 344 } 345 346 /* Let the GMU know that we are about to go into slumber */ 347 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu) 348 { 349 int ret; 350 351 /* Disable the power counter so the GMU isn't busy */ 352 gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0); 353 354 /* Disable SPTP_PC if the CPU is responsible for it */ 355 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) 356 a6xx_sptprac_disable(gmu); 357 358 /* Tell the GMU to get ready to slumber */ 359 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1); 360 361 ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER); 362 a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER); 363 364 if (!ret) { 365 /* Check to see if the GMU really did slumber */ 366 if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE) 367 != 0x0f) { 368 DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n"); 369 ret = -ETIMEDOUT; 370 } 371 } 372 373 /* Put fence into allow mode */ 374 gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0); 375 return ret; 376 } 377 378 static int a6xx_rpmh_start(struct a6xx_gmu *gmu) 379 { 380 int ret; 381 u32 val; 382 383 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1); 384 /* Wait for the register to finish posting */ 385 wmb(); 386 387 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val, 388 val & (1 << 1), 100, 10000); 389 if (ret) { 390 DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n"); 391 return ret; 392 } 393 394 ret = gmu_poll_timeout(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val, 395 !val, 100, 10000); 396 397 if (ret) { 398 DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n"); 399 return ret; 400 } 401 402 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0); 403 404 /* Set up CX GMU counter 0 to count busy ticks */ 405 gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000); 406 gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20); 407 408 /* Enable the power counter */ 409 gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1); 410 return 0; 411 } 412 413 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu) 414 { 415 int ret; 416 u32 val; 417 418 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1); 419 420 ret = gmu_poll_timeout(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, 421 val, val & (1 << 16), 100, 10000); 422 if (ret) 423 DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n"); 424 425 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0); 426 } 427 428 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value) 429 { 430 return msm_writel(value, ptr + (offset << 2)); 431 } 432 433 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 434 const char *name); 435 436 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu) 437 { 438 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 439 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 440 struct platform_device *pdev = to_platform_device(gmu->dev); 441 void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc"); 442 void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq"); 443 444 if (!pdcptr || !seqptr) 445 goto err; 446 447 /* Disable SDE clock gating */ 448 gmu_write(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24)); 449 450 /* Setup RSC PDC handshake for sleep and wakeup */ 451 gmu_write(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1); 452 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0); 453 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0); 454 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0); 455 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0); 456 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000); 457 gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0); 458 gmu_write(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0); 459 gmu_write(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520); 460 gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510); 461 gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514); 462 463 /* Load RSC sequencer uCode for sleep and wakeup */ 464 gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0); 465 gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7); 466 gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1); 467 gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2); 468 gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8); 469 470 /* Load PDC sequencer uCode for power up and power down sequence */ 471 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1); 472 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2); 473 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0); 474 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284); 475 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc); 476 477 /* Set TCS commands used by PDC sequence for low power modes */ 478 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7); 479 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0); 480 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0); 481 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108); 482 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010); 483 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1); 484 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108); 485 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000); 486 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0); 487 488 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108); 489 if (adreno_is_a618(adreno_gpu)) 490 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30090); 491 else 492 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30080); 493 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0); 494 495 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7); 496 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0); 497 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0); 498 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108); 499 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010); 500 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2); 501 502 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108); 503 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000); 504 if (adreno_is_a618(adreno_gpu)) 505 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2); 506 else 507 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3); 508 509 510 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108); 511 if (adreno_is_a618(adreno_gpu)) 512 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30090); 513 else 514 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30080); 515 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3); 516 517 /* Setup GPU PDC */ 518 pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0); 519 pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001); 520 521 /* ensure no writes happen before the uCode is fully written */ 522 wmb(); 523 524 err: 525 if (!IS_ERR_OR_NULL(pdcptr)) 526 iounmap(pdcptr); 527 if (!IS_ERR_OR_NULL(seqptr)) 528 iounmap(seqptr); 529 } 530 531 /* 532 * The lowest 16 bits of this value are the number of XO clock cycles for main 533 * hysteresis which is set at 0x1680 cycles (300 us). The higher 16 bits are 534 * for the shorter hysteresis that happens after main - this is 0xa (.5 us) 535 */ 536 537 #define GMU_PWR_COL_HYST 0x000a1680 538 539 /* Set up the idle state for the GMU */ 540 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu) 541 { 542 /* Disable GMU WB/RB buffer */ 543 gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1); 544 545 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400); 546 547 switch (gmu->idle_level) { 548 case GMU_IDLE_STATE_IFPC: 549 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST, 550 GMU_PWR_COL_HYST); 551 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 552 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 553 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE); 554 /* Fall through */ 555 case GMU_IDLE_STATE_SPTP: 556 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST, 557 GMU_PWR_COL_HYST); 558 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 559 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 560 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE); 561 } 562 563 /* Enable RPMh GPU client */ 564 gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0, 565 A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE | 566 A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE | 567 A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE | 568 A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE | 569 A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE | 570 A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE); 571 } 572 573 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state) 574 { 575 static bool rpmh_init; 576 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 577 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 578 int i, ret; 579 u32 chipid; 580 u32 *image; 581 582 if (state == GMU_WARM_BOOT) { 583 ret = a6xx_rpmh_start(gmu); 584 if (ret) 585 return ret; 586 } else { 587 if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU], 588 "GMU firmware is not loaded\n")) 589 return -ENOENT; 590 591 /* Sanity check the size of the firmware that was loaded */ 592 if (adreno_gpu->fw[ADRENO_FW_GMU]->size > 0x8000) { 593 DRM_DEV_ERROR(gmu->dev, 594 "GMU firmware is bigger than the available region\n"); 595 return -EINVAL; 596 } 597 598 /* Turn on register retention */ 599 gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1); 600 601 /* We only need to load the RPMh microcode once */ 602 if (!rpmh_init) { 603 a6xx_gmu_rpmh_init(gmu); 604 rpmh_init = true; 605 } else { 606 ret = a6xx_rpmh_start(gmu); 607 if (ret) 608 return ret; 609 } 610 611 image = (u32 *) adreno_gpu->fw[ADRENO_FW_GMU]->data; 612 613 for (i = 0; i < adreno_gpu->fw[ADRENO_FW_GMU]->size >> 2; i++) 614 gmu_write(gmu, REG_A6XX_GMU_CM3_ITCM_START + i, 615 image[i]); 616 } 617 618 gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0); 619 gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02); 620 621 /* Write the iova of the HFI table */ 622 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi->iova); 623 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1); 624 625 gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0, 626 (1 << 31) | (0xa << 18) | (0xa0)); 627 628 chipid = adreno_gpu->rev.core << 24; 629 chipid |= adreno_gpu->rev.major << 16; 630 chipid |= adreno_gpu->rev.minor << 12; 631 chipid |= adreno_gpu->rev.patchid << 8; 632 633 gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid); 634 635 /* Set up the lowest idle level on the GMU */ 636 a6xx_gmu_power_config(gmu); 637 638 ret = a6xx_gmu_start(gmu); 639 if (ret) 640 return ret; 641 642 ret = a6xx_gmu_gfx_rail_on(gmu); 643 if (ret) 644 return ret; 645 646 /* Enable SPTP_PC if the CPU is responsible for it */ 647 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) { 648 ret = a6xx_sptprac_enable(gmu); 649 if (ret) 650 return ret; 651 } 652 653 ret = a6xx_gmu_hfi_start(gmu); 654 if (ret) 655 return ret; 656 657 /* FIXME: Do we need this wmb() here? */ 658 wmb(); 659 660 return 0; 661 } 662 663 #define A6XX_HFI_IRQ_MASK \ 664 (A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) 665 666 #define A6XX_GMU_IRQ_MASK \ 667 (A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \ 668 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \ 669 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR) 670 671 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu) 672 { 673 disable_irq(gmu->gmu_irq); 674 disable_irq(gmu->hfi_irq); 675 676 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0); 677 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0); 678 } 679 680 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu) 681 { 682 u32 val; 683 684 /* Make sure there are no outstanding RPMh votes */ 685 gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val, 686 (val & 1), 100, 10000); 687 gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val, 688 (val & 1), 100, 10000); 689 gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val, 690 (val & 1), 100, 10000); 691 gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val, 692 (val & 1), 100, 1000); 693 } 694 695 /* Force the GMU off in case it isn't responsive */ 696 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu) 697 { 698 /* Flush all the queues */ 699 a6xx_hfi_stop(gmu); 700 701 /* Stop the interrupts */ 702 a6xx_gmu_irq_disable(gmu); 703 704 /* Force off SPTP in case the GMU is managing it */ 705 a6xx_sptprac_disable(gmu); 706 707 /* Make sure there are no outstanding RPMh votes */ 708 a6xx_gmu_rpmh_off(gmu); 709 } 710 711 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu) 712 { 713 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 714 struct msm_gpu *gpu = &adreno_gpu->base; 715 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 716 int status, ret; 717 718 if (WARN(!gmu->initialized, "The GMU is not set up yet\n")) 719 return 0; 720 721 gmu->hung = false; 722 723 /* Turn on the resources */ 724 pm_runtime_get_sync(gmu->dev); 725 726 /* Use a known rate to bring up the GMU */ 727 clk_set_rate(gmu->core_clk, 200000000); 728 ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks); 729 if (ret) { 730 pm_runtime_put(gmu->dev); 731 return ret; 732 } 733 734 /* Set the bus quota to a reasonable value for boot */ 735 icc_set_bw(gpu->icc_path, 0, MBps_to_icc(3072)); 736 737 /* Enable the GMU interrupt */ 738 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0); 739 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK); 740 enable_irq(gmu->gmu_irq); 741 742 /* Check to see if we are doing a cold or warm boot */ 743 status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ? 744 GMU_WARM_BOOT : GMU_COLD_BOOT; 745 746 ret = a6xx_gmu_fw_start(gmu, status); 747 if (ret) 748 goto out; 749 750 ret = a6xx_hfi_start(gmu, status); 751 if (ret) 752 goto out; 753 754 /* 755 * Turn on the GMU firmware fault interrupt after we know the boot 756 * sequence is successful 757 */ 758 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0); 759 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK); 760 enable_irq(gmu->hfi_irq); 761 762 /* Set the GPU to the current freq */ 763 __a6xx_gmu_set_freq(gmu, gmu->current_perf_index); 764 765 /* 766 * "enable" the GX power domain which won't actually do anything but it 767 * will make sure that the refcounting is correct in case we need to 768 * bring down the GX after a GMU failure 769 */ 770 if (!IS_ERR_OR_NULL(gmu->gxpd)) 771 pm_runtime_get(gmu->gxpd); 772 773 out: 774 /* On failure, shut down the GMU to leave it in a good state */ 775 if (ret) { 776 disable_irq(gmu->gmu_irq); 777 a6xx_rpmh_stop(gmu); 778 pm_runtime_put(gmu->dev); 779 } 780 781 return ret; 782 } 783 784 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu) 785 { 786 u32 reg; 787 788 if (!gmu->initialized) 789 return true; 790 791 reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS); 792 793 if (reg & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB) 794 return false; 795 796 return true; 797 } 798 799 #define GBIF_CLIENT_HALT_MASK BIT(0) 800 #define GBIF_ARB_HALT_MASK BIT(1) 801 802 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) 803 { 804 struct msm_gpu *gpu = &adreno_gpu->base; 805 806 if (!a6xx_has_gbif(adreno_gpu)) { 807 gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf); 808 spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) & 809 0xf) == 0xf); 810 gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0); 811 812 return; 813 } 814 815 /* Halt new client requests on GBIF */ 816 gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK); 817 spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) & 818 (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK); 819 820 /* Halt all AXI requests on GBIF */ 821 gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK); 822 spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) & 823 (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK); 824 825 /* The GBIF halt needs to be explicitly cleared */ 826 gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0); 827 } 828 829 /* Gracefully try to shut down the GMU and by extension the GPU */ 830 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu) 831 { 832 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 833 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 834 u32 val; 835 836 /* 837 * The GMU may still be in slumber unless the GPU started so check and 838 * skip putting it back into slumber if so 839 */ 840 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); 841 842 if (val != 0xf) { 843 int ret = a6xx_gmu_wait_for_idle(gmu); 844 845 /* If the GMU isn't responding assume it is hung */ 846 if (ret) { 847 a6xx_gmu_force_off(gmu); 848 return; 849 } 850 851 a6xx_bus_clear_pending_transactions(adreno_gpu); 852 853 /* tell the GMU we want to slumber */ 854 a6xx_gmu_notify_slumber(gmu); 855 856 ret = gmu_poll_timeout(gmu, 857 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val, 858 !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB), 859 100, 10000); 860 861 /* 862 * Let the user know we failed to slumber but don't worry too 863 * much because we are powering down anyway 864 */ 865 866 if (ret) 867 DRM_DEV_ERROR(gmu->dev, 868 "Unable to slumber GMU: status = 0%x/0%x\n", 869 gmu_read(gmu, 870 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS), 871 gmu_read(gmu, 872 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2)); 873 } 874 875 /* Turn off HFI */ 876 a6xx_hfi_stop(gmu); 877 878 /* Stop the interrupts and mask the hardware */ 879 a6xx_gmu_irq_disable(gmu); 880 881 /* Tell RPMh to power off the GPU */ 882 a6xx_rpmh_stop(gmu); 883 } 884 885 886 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu) 887 { 888 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 889 struct msm_gpu *gpu = &a6xx_gpu->base.base; 890 891 if (!pm_runtime_active(gmu->dev)) 892 return 0; 893 894 /* 895 * Force the GMU off if we detected a hang, otherwise try to shut it 896 * down gracefully 897 */ 898 if (gmu->hung) 899 a6xx_gmu_force_off(gmu); 900 else 901 a6xx_gmu_shutdown(gmu); 902 903 /* Remove the bus vote */ 904 icc_set_bw(gpu->icc_path, 0, 0); 905 906 /* 907 * Make sure the GX domain is off before turning off the GMU (CX) 908 * domain. Usually the GMU does this but only if the shutdown sequence 909 * was successful 910 */ 911 if (!IS_ERR_OR_NULL(gmu->gxpd)) 912 pm_runtime_put_sync(gmu->gxpd); 913 914 clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks); 915 916 pm_runtime_put_sync(gmu->dev); 917 918 return 0; 919 } 920 921 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo) 922 { 923 int count, i; 924 u64 iova; 925 926 if (IS_ERR_OR_NULL(bo)) 927 return; 928 929 count = bo->size >> PAGE_SHIFT; 930 iova = bo->iova; 931 932 for (i = 0; i < count; i++, iova += PAGE_SIZE) { 933 iommu_unmap(gmu->domain, iova, PAGE_SIZE); 934 __free_pages(bo->pages[i], 0); 935 } 936 937 kfree(bo->pages); 938 kfree(bo); 939 } 940 941 static struct a6xx_gmu_bo *a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, 942 size_t size) 943 { 944 struct a6xx_gmu_bo *bo; 945 int ret, count, i; 946 947 bo = kzalloc(sizeof(*bo), GFP_KERNEL); 948 if (!bo) 949 return ERR_PTR(-ENOMEM); 950 951 bo->size = PAGE_ALIGN(size); 952 953 count = bo->size >> PAGE_SHIFT; 954 955 bo->pages = kcalloc(count, sizeof(struct page *), GFP_KERNEL); 956 if (!bo->pages) { 957 kfree(bo); 958 return ERR_PTR(-ENOMEM); 959 } 960 961 for (i = 0; i < count; i++) { 962 bo->pages[i] = alloc_page(GFP_KERNEL); 963 if (!bo->pages[i]) 964 goto err; 965 } 966 967 bo->iova = gmu->uncached_iova_base; 968 969 for (i = 0; i < count; i++) { 970 ret = iommu_map(gmu->domain, 971 bo->iova + (PAGE_SIZE * i), 972 page_to_phys(bo->pages[i]), PAGE_SIZE, 973 IOMMU_READ | IOMMU_WRITE); 974 975 if (ret) { 976 DRM_DEV_ERROR(gmu->dev, "Unable to map GMU buffer object\n"); 977 978 for (i = i - 1 ; i >= 0; i--) 979 iommu_unmap(gmu->domain, 980 bo->iova + (PAGE_SIZE * i), 981 PAGE_SIZE); 982 983 goto err; 984 } 985 } 986 987 bo->virt = vmap(bo->pages, count, VM_IOREMAP, 988 pgprot_writecombine(PAGE_KERNEL)); 989 if (!bo->virt) 990 goto err; 991 992 /* Align future IOVA addresses on 1MB boundaries */ 993 gmu->uncached_iova_base += ALIGN(size, SZ_1M); 994 995 return bo; 996 997 err: 998 for (i = 0; i < count; i++) { 999 if (bo->pages[i]) 1000 __free_pages(bo->pages[i], 0); 1001 } 1002 1003 kfree(bo->pages); 1004 kfree(bo); 1005 1006 return ERR_PTR(-ENOMEM); 1007 } 1008 1009 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu) 1010 { 1011 int ret; 1012 1013 /* 1014 * The GMU address space is hardcoded to treat the range 1015 * 0x60000000 - 0x80000000 as un-cached memory. All buffers shared 1016 * between the GMU and the CPU will live in this space 1017 */ 1018 gmu->uncached_iova_base = 0x60000000; 1019 1020 1021 gmu->domain = iommu_domain_alloc(&platform_bus_type); 1022 if (!gmu->domain) 1023 return -ENODEV; 1024 1025 ret = iommu_attach_device(gmu->domain, gmu->dev); 1026 1027 if (ret) { 1028 iommu_domain_free(gmu->domain); 1029 gmu->domain = NULL; 1030 } 1031 1032 return ret; 1033 } 1034 1035 /* Return the 'arc-level' for the given frequency */ 1036 static unsigned int a6xx_gmu_get_arc_level(struct device *dev, 1037 unsigned long freq) 1038 { 1039 struct dev_pm_opp *opp; 1040 unsigned int val; 1041 1042 if (!freq) 1043 return 0; 1044 1045 opp = dev_pm_opp_find_freq_exact(dev, freq, true); 1046 if (IS_ERR(opp)) 1047 return 0; 1048 1049 val = dev_pm_opp_get_level(opp); 1050 1051 dev_pm_opp_put(opp); 1052 1053 return val; 1054 } 1055 1056 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes, 1057 unsigned long *freqs, int freqs_count, const char *id) 1058 { 1059 int i, j; 1060 const u16 *pri, *sec; 1061 size_t pri_count, sec_count; 1062 1063 pri = cmd_db_read_aux_data(id, &pri_count); 1064 if (IS_ERR(pri)) 1065 return PTR_ERR(pri); 1066 /* 1067 * The data comes back as an array of unsigned shorts so adjust the 1068 * count accordingly 1069 */ 1070 pri_count >>= 1; 1071 if (!pri_count) 1072 return -EINVAL; 1073 1074 sec = cmd_db_read_aux_data("mx.lvl", &sec_count); 1075 if (IS_ERR(sec)) 1076 return PTR_ERR(sec); 1077 1078 sec_count >>= 1; 1079 if (!sec_count) 1080 return -EINVAL; 1081 1082 /* Construct a vote for each frequency */ 1083 for (i = 0; i < freqs_count; i++) { 1084 u8 pindex = 0, sindex = 0; 1085 unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]); 1086 1087 /* Get the primary index that matches the arc level */ 1088 for (j = 0; j < pri_count; j++) { 1089 if (pri[j] >= level) { 1090 pindex = j; 1091 break; 1092 } 1093 } 1094 1095 if (j == pri_count) { 1096 DRM_DEV_ERROR(dev, 1097 "Level %u not found in in the RPMh list\n", 1098 level); 1099 DRM_DEV_ERROR(dev, "Available levels:\n"); 1100 for (j = 0; j < pri_count; j++) 1101 DRM_DEV_ERROR(dev, " %u\n", pri[j]); 1102 1103 return -EINVAL; 1104 } 1105 1106 /* 1107 * Look for a level in in the secondary list that matches. If 1108 * nothing fits, use the maximum non zero vote 1109 */ 1110 1111 for (j = 0; j < sec_count; j++) { 1112 if (sec[j] >= level) { 1113 sindex = j; 1114 break; 1115 } else if (sec[j]) { 1116 sindex = j; 1117 } 1118 } 1119 1120 /* Construct the vote */ 1121 votes[i] = ((pri[pindex] & 0xffff) << 16) | 1122 (sindex << 8) | pindex; 1123 } 1124 1125 return 0; 1126 } 1127 1128 /* 1129 * The GMU votes with the RPMh for itself and on behalf of the GPU but we need 1130 * to construct the list of votes on the CPU and send it over. Query the RPMh 1131 * voltage levels and build the votes 1132 */ 1133 1134 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu) 1135 { 1136 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1137 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1138 struct msm_gpu *gpu = &adreno_gpu->base; 1139 int ret; 1140 1141 /* Build the GX votes */ 1142 ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes, 1143 gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl"); 1144 1145 /* Build the CX votes */ 1146 ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes, 1147 gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl"); 1148 1149 return ret; 1150 } 1151 1152 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs, 1153 u32 size) 1154 { 1155 int count = dev_pm_opp_get_opp_count(dev); 1156 struct dev_pm_opp *opp; 1157 int i, index = 0; 1158 unsigned long freq = 1; 1159 1160 /* 1161 * The OPP table doesn't contain the "off" frequency level so we need to 1162 * add 1 to the table size to account for it 1163 */ 1164 1165 if (WARN(count + 1 > size, 1166 "The GMU frequency table is being truncated\n")) 1167 count = size - 1; 1168 1169 /* Set the "off" frequency */ 1170 freqs[index++] = 0; 1171 1172 for (i = 0; i < count; i++) { 1173 opp = dev_pm_opp_find_freq_ceil(dev, &freq); 1174 if (IS_ERR(opp)) 1175 break; 1176 1177 dev_pm_opp_put(opp); 1178 freqs[index++] = freq++; 1179 } 1180 1181 return index; 1182 } 1183 1184 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu) 1185 { 1186 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1187 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1188 struct msm_gpu *gpu = &adreno_gpu->base; 1189 1190 int ret = 0; 1191 1192 /* 1193 * The GMU handles its own frequency switching so build a list of 1194 * available frequencies to send during initialization 1195 */ 1196 ret = dev_pm_opp_of_add_table(gmu->dev); 1197 if (ret) { 1198 DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n"); 1199 return ret; 1200 } 1201 1202 gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev, 1203 gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs)); 1204 1205 /* 1206 * The GMU also handles GPU frequency switching so build a list 1207 * from the GPU OPP table 1208 */ 1209 gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev, 1210 gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs)); 1211 1212 gmu->current_perf_index = gmu->nr_gpu_freqs - 1; 1213 1214 /* Build the list of RPMh votes that we'll send to the GMU */ 1215 return a6xx_gmu_rpmh_votes_init(gmu); 1216 } 1217 1218 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu) 1219 { 1220 int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks); 1221 1222 if (ret < 1) 1223 return ret; 1224 1225 gmu->nr_clocks = ret; 1226 1227 gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks, 1228 gmu->nr_clocks, "gmu"); 1229 1230 return 0; 1231 } 1232 1233 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 1234 const char *name) 1235 { 1236 void __iomem *ret; 1237 struct resource *res = platform_get_resource_byname(pdev, 1238 IORESOURCE_MEM, name); 1239 1240 if (!res) { 1241 DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name); 1242 return ERR_PTR(-EINVAL); 1243 } 1244 1245 ret = ioremap(res->start, resource_size(res)); 1246 if (!ret) { 1247 DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name); 1248 return ERR_PTR(-EINVAL); 1249 } 1250 1251 return ret; 1252 } 1253 1254 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev, 1255 const char *name, irq_handler_t handler) 1256 { 1257 int irq, ret; 1258 1259 irq = platform_get_irq_byname(pdev, name); 1260 1261 ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu); 1262 if (ret) { 1263 DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n", 1264 name, ret); 1265 return ret; 1266 } 1267 1268 disable_irq(irq); 1269 1270 return irq; 1271 } 1272 1273 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu) 1274 { 1275 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1276 1277 if (!gmu->initialized) 1278 return; 1279 1280 pm_runtime_force_suspend(gmu->dev); 1281 1282 if (!IS_ERR_OR_NULL(gmu->gxpd)) { 1283 pm_runtime_disable(gmu->gxpd); 1284 dev_pm_domain_detach(gmu->gxpd, false); 1285 } 1286 1287 iounmap(gmu->mmio); 1288 gmu->mmio = NULL; 1289 1290 a6xx_gmu_memory_free(gmu, gmu->hfi); 1291 1292 iommu_detach_device(gmu->domain, gmu->dev); 1293 1294 iommu_domain_free(gmu->domain); 1295 1296 free_irq(gmu->gmu_irq, gmu); 1297 free_irq(gmu->hfi_irq, gmu); 1298 1299 /* Drop reference taken in of_find_device_by_node */ 1300 put_device(gmu->dev); 1301 1302 gmu->initialized = false; 1303 } 1304 1305 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node) 1306 { 1307 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1308 struct platform_device *pdev = of_find_device_by_node(node); 1309 int ret; 1310 1311 if (!pdev) 1312 return -ENODEV; 1313 1314 gmu->dev = &pdev->dev; 1315 1316 of_dma_configure(gmu->dev, node, true); 1317 1318 /* Fow now, don't do anything fancy until we get our feet under us */ 1319 gmu->idle_level = GMU_IDLE_STATE_ACTIVE; 1320 1321 pm_runtime_enable(gmu->dev); 1322 1323 /* Get the list of clocks */ 1324 ret = a6xx_gmu_clocks_probe(gmu); 1325 if (ret) 1326 goto err_put_device; 1327 1328 /* Set up the IOMMU context bank */ 1329 ret = a6xx_gmu_memory_probe(gmu); 1330 if (ret) 1331 goto err_put_device; 1332 1333 /* Allocate memory for for the HFI queues */ 1334 gmu->hfi = a6xx_gmu_memory_alloc(gmu, SZ_16K); 1335 if (IS_ERR(gmu->hfi)) 1336 goto err_memory; 1337 1338 /* Allocate memory for the GMU debug region */ 1339 gmu->debug = a6xx_gmu_memory_alloc(gmu, SZ_16K); 1340 if (IS_ERR(gmu->debug)) 1341 goto err_memory; 1342 1343 /* Map the GMU registers */ 1344 gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu"); 1345 if (IS_ERR(gmu->mmio)) 1346 goto err_memory; 1347 1348 /* Get the HFI and GMU interrupts */ 1349 gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq); 1350 gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq); 1351 1352 if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) 1353 goto err_mmio; 1354 1355 /* 1356 * Get a link to the GX power domain to reset the GPU in case of GMU 1357 * crash 1358 */ 1359 gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx"); 1360 1361 /* Get the power levels for the GMU and GPU */ 1362 a6xx_gmu_pwrlevels_probe(gmu); 1363 1364 /* Set up the HFI queues */ 1365 a6xx_hfi_init(gmu); 1366 1367 gmu->initialized = true; 1368 1369 return 0; 1370 1371 err_mmio: 1372 iounmap(gmu->mmio); 1373 free_irq(gmu->gmu_irq, gmu); 1374 free_irq(gmu->hfi_irq, gmu); 1375 err_memory: 1376 a6xx_gmu_memory_free(gmu, gmu->hfi); 1377 1378 if (gmu->domain) { 1379 iommu_detach_device(gmu->domain, gmu->dev); 1380 1381 iommu_domain_free(gmu->domain); 1382 } 1383 ret = -ENODEV; 1384 1385 err_put_device: 1386 /* Drop reference taken in of_find_device_by_node */ 1387 put_device(gmu->dev); 1388 1389 return ret; 1390 } 1391