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 #include <drm/drm_gem.h> 10 11 #include "a6xx_gpu.h" 12 #include "a6xx_gmu.xml.h" 13 #include "msm_gem.h" 14 #include "msm_gpu_trace.h" 15 #include "msm_mmu.h" 16 17 static void a6xx_gmu_fault(struct a6xx_gmu *gmu) 18 { 19 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 20 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 21 struct msm_gpu *gpu = &adreno_gpu->base; 22 23 /* FIXME: add a banner here */ 24 gmu->hung = true; 25 26 /* Turn off the hangcheck timer while we are resetting */ 27 del_timer(&gpu->hangcheck_timer); 28 29 /* Queue the GPU handler because we need to treat this as a recovery */ 30 kthread_queue_work(gpu->worker, &gpu->recover_work); 31 } 32 33 static irqreturn_t a6xx_gmu_irq(int irq, void *data) 34 { 35 struct a6xx_gmu *gmu = data; 36 u32 status; 37 38 status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS); 39 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status); 40 41 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) { 42 dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n"); 43 44 a6xx_gmu_fault(gmu); 45 } 46 47 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR) 48 dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n"); 49 50 if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR) 51 dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n", 52 gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS)); 53 54 return IRQ_HANDLED; 55 } 56 57 static irqreturn_t a6xx_hfi_irq(int irq, void *data) 58 { 59 struct a6xx_gmu *gmu = data; 60 u32 status; 61 62 status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO); 63 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status); 64 65 if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) { 66 dev_err_ratelimited(gmu->dev, "GMU firmware fault\n"); 67 68 a6xx_gmu_fault(gmu); 69 } 70 71 return IRQ_HANDLED; 72 } 73 74 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu) 75 { 76 u32 val; 77 78 /* This can be called from gpu state code so make sure GMU is valid */ 79 if (!gmu->initialized) 80 return false; 81 82 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS); 83 84 return !(val & 85 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF | 86 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF)); 87 } 88 89 /* Check to see if the GX rail is still powered */ 90 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu) 91 { 92 u32 val; 93 94 /* This can be called from gpu state code so make sure GMU is valid */ 95 if (!gmu->initialized) 96 return false; 97 98 val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS); 99 100 return !(val & 101 (A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF | 102 A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF)); 103 } 104 105 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp, 106 bool suspended) 107 { 108 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 109 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu); 110 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 111 u32 perf_index; 112 unsigned long gpu_freq; 113 int ret = 0; 114 115 gpu_freq = dev_pm_opp_get_freq(opp); 116 117 if (gpu_freq == gmu->freq) 118 return; 119 120 for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++) 121 if (gpu_freq == gmu->gpu_freqs[perf_index]) 122 break; 123 124 gmu->current_perf_index = perf_index; 125 gmu->freq = gmu->gpu_freqs[perf_index]; 126 127 trace_msm_gmu_freq_change(gmu->freq, perf_index); 128 129 /* 130 * This can get called from devfreq while the hardware is idle. Don't 131 * bring up the power if it isn't already active. All we're doing here 132 * is updating the frequency so that when we come back online we're at 133 * the right rate. 134 */ 135 if (suspended) 136 return; 137 138 if (!gmu->legacy) { 139 a6xx_hfi_set_freq(gmu, perf_index); 140 dev_pm_opp_set_opp(&gpu->pdev->dev, opp); 141 return; 142 } 143 144 gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0); 145 146 gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING, 147 ((3 & 0xf) << 28) | perf_index); 148 149 /* 150 * Send an invalid index as a vote for the bus bandwidth and let the 151 * firmware decide on the right vote 152 */ 153 gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff); 154 155 /* Set and clear the OOB for DCVS to trigger the GMU */ 156 a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET); 157 a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET); 158 159 ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN); 160 if (ret) 161 dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret); 162 163 dev_pm_opp_set_opp(&gpu->pdev->dev, opp); 164 } 165 166 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu) 167 { 168 struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu); 169 struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu); 170 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 171 172 return gmu->freq; 173 } 174 175 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu) 176 { 177 u32 val; 178 int local = gmu->idle_level; 179 180 /* SPTP and IFPC both report as IFPC */ 181 if (gmu->idle_level == GMU_IDLE_STATE_SPTP) 182 local = GMU_IDLE_STATE_IFPC; 183 184 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); 185 186 if (val == local) { 187 if (gmu->idle_level != GMU_IDLE_STATE_IFPC || 188 !a6xx_gmu_gx_is_on(gmu)) 189 return true; 190 } 191 192 return false; 193 } 194 195 /* Wait for the GMU to get to its most idle state */ 196 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu) 197 { 198 return spin_until(a6xx_gmu_check_idle_level(gmu)); 199 } 200 201 static int a6xx_gmu_start(struct a6xx_gmu *gmu) 202 { 203 int ret; 204 u32 val; 205 u32 mask, reset_val; 206 207 val = gmu_read(gmu, REG_A6XX_GMU_CM3_DTCM_START + 0xff8); 208 if (val <= 0x20010004) { 209 mask = 0xffffffff; 210 reset_val = 0xbabeface; 211 } else { 212 mask = 0x1ff; 213 reset_val = 0x100; 214 } 215 216 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1); 217 218 /* Set the log wptr index 219 * note: downstream saves the value in poweroff and restores it here 220 */ 221 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_RESP, 0); 222 223 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0); 224 225 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val, 226 (val & mask) == reset_val, 100, 10000); 227 228 if (ret) 229 DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n"); 230 231 return ret; 232 } 233 234 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu) 235 { 236 u32 val; 237 int ret; 238 239 gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1); 240 241 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val, 242 val & 1, 100, 10000); 243 if (ret) 244 DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n"); 245 246 return ret; 247 } 248 249 struct a6xx_gmu_oob_bits { 250 int set, ack, set_new, ack_new, clear, clear_new; 251 const char *name; 252 }; 253 254 /* These are the interrupt / ack bits for each OOB request that are set 255 * in a6xx_gmu_set_oob and a6xx_clear_oob 256 */ 257 static const struct a6xx_gmu_oob_bits a6xx_gmu_oob_bits[] = { 258 [GMU_OOB_GPU_SET] = { 259 .name = "GPU_SET", 260 .set = 16, 261 .ack = 24, 262 .set_new = 30, 263 .ack_new = 31, 264 .clear = 24, 265 .clear_new = 31, 266 }, 267 268 [GMU_OOB_PERFCOUNTER_SET] = { 269 .name = "PERFCOUNTER", 270 .set = 17, 271 .ack = 25, 272 .set_new = 28, 273 .ack_new = 30, 274 .clear = 25, 275 .clear_new = 29, 276 }, 277 278 [GMU_OOB_BOOT_SLUMBER] = { 279 .name = "BOOT_SLUMBER", 280 .set = 22, 281 .ack = 30, 282 .clear = 30, 283 }, 284 285 [GMU_OOB_DCVS_SET] = { 286 .name = "GPU_DCVS", 287 .set = 23, 288 .ack = 31, 289 .clear = 31, 290 }, 291 }; 292 293 /* Trigger a OOB (out of band) request to the GMU */ 294 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state) 295 { 296 int ret; 297 u32 val; 298 int request, ack; 299 300 WARN_ON_ONCE(!mutex_is_locked(&gmu->lock)); 301 302 if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits)) 303 return -EINVAL; 304 305 if (gmu->legacy) { 306 request = a6xx_gmu_oob_bits[state].set; 307 ack = a6xx_gmu_oob_bits[state].ack; 308 } else { 309 request = a6xx_gmu_oob_bits[state].set_new; 310 ack = a6xx_gmu_oob_bits[state].ack_new; 311 if (!request || !ack) { 312 DRM_DEV_ERROR(gmu->dev, 313 "Invalid non-legacy GMU request %s\n", 314 a6xx_gmu_oob_bits[state].name); 315 return -EINVAL; 316 } 317 } 318 319 /* Trigger the equested OOB operation */ 320 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request); 321 322 /* Wait for the acknowledge interrupt */ 323 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val, 324 val & (1 << ack), 100, 10000); 325 326 if (ret) 327 DRM_DEV_ERROR(gmu->dev, 328 "Timeout waiting for GMU OOB set %s: 0x%x\n", 329 a6xx_gmu_oob_bits[state].name, 330 gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO)); 331 332 /* Clear the acknowledge interrupt */ 333 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack); 334 335 return ret; 336 } 337 338 /* Clear a pending OOB state in the GMU */ 339 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state) 340 { 341 int bit; 342 343 WARN_ON_ONCE(!mutex_is_locked(&gmu->lock)); 344 345 if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits)) 346 return; 347 348 if (gmu->legacy) 349 bit = a6xx_gmu_oob_bits[state].clear; 350 else 351 bit = a6xx_gmu_oob_bits[state].clear_new; 352 353 gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << bit); 354 } 355 356 /* Enable CPU control of SPTP power power collapse */ 357 int a6xx_sptprac_enable(struct a6xx_gmu *gmu) 358 { 359 int ret; 360 u32 val; 361 362 if (!gmu->legacy) 363 return 0; 364 365 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000); 366 367 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val, 368 (val & 0x38) == 0x28, 1, 100); 369 370 if (ret) { 371 DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n", 372 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS)); 373 } 374 375 return 0; 376 } 377 378 /* Disable CPU control of SPTP power power collapse */ 379 void a6xx_sptprac_disable(struct a6xx_gmu *gmu) 380 { 381 u32 val; 382 int ret; 383 384 if (!gmu->legacy) 385 return; 386 387 /* Make sure retention is on */ 388 gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11)); 389 390 gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001); 391 392 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val, 393 (val & 0x04), 100, 10000); 394 395 if (ret) 396 DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n", 397 gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS)); 398 } 399 400 /* Let the GMU know we are starting a boot sequence */ 401 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu) 402 { 403 u32 vote; 404 405 /* Let the GMU know we are getting ready for boot */ 406 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0); 407 408 /* Choose the "default" power level as the highest available */ 409 vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1]; 410 411 gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff); 412 gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff); 413 414 /* Let the GMU know the boot sequence has started */ 415 return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER); 416 } 417 418 /* Let the GMU know that we are about to go into slumber */ 419 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu) 420 { 421 int ret; 422 423 /* Disable the power counter so the GMU isn't busy */ 424 gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0); 425 426 /* Disable SPTP_PC if the CPU is responsible for it */ 427 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) 428 a6xx_sptprac_disable(gmu); 429 430 if (!gmu->legacy) { 431 ret = a6xx_hfi_send_prep_slumber(gmu); 432 goto out; 433 } 434 435 /* Tell the GMU to get ready to slumber */ 436 gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1); 437 438 ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER); 439 a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER); 440 441 if (!ret) { 442 /* Check to see if the GMU really did slumber */ 443 if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE) 444 != 0x0f) { 445 DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n"); 446 ret = -ETIMEDOUT; 447 } 448 } 449 450 out: 451 /* Put fence into allow mode */ 452 gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0); 453 return ret; 454 } 455 456 static int a6xx_rpmh_start(struct a6xx_gmu *gmu) 457 { 458 int ret; 459 u32 val; 460 461 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1); 462 /* Wait for the register to finish posting */ 463 wmb(); 464 465 ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val, 466 val & (1 << 1), 100, 10000); 467 if (ret) { 468 DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n"); 469 return ret; 470 } 471 472 ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val, 473 !val, 100, 10000); 474 475 if (ret) { 476 DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n"); 477 return ret; 478 } 479 480 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0); 481 482 return 0; 483 } 484 485 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu) 486 { 487 int ret; 488 u32 val; 489 490 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1); 491 492 ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, 493 val, val & (1 << 16), 100, 10000); 494 if (ret) 495 DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n"); 496 497 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0); 498 } 499 500 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value) 501 { 502 msm_writel(value, ptr + (offset << 2)); 503 } 504 505 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 506 const char *name); 507 508 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu) 509 { 510 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 511 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 512 struct platform_device *pdev = to_platform_device(gmu->dev); 513 void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc"); 514 void __iomem *seqptr = NULL; 515 uint32_t pdc_address_offset; 516 bool pdc_in_aop = false; 517 518 if (IS_ERR(pdcptr)) 519 goto err; 520 521 if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu)) 522 pdc_in_aop = true; 523 else if (adreno_is_a618(adreno_gpu) || adreno_is_a640_family(adreno_gpu)) 524 pdc_address_offset = 0x30090; 525 else if (adreno_is_a619(adreno_gpu)) 526 pdc_address_offset = 0x300a0; 527 else 528 pdc_address_offset = 0x30080; 529 530 if (!pdc_in_aop) { 531 seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq"); 532 if (IS_ERR(seqptr)) 533 goto err; 534 } 535 536 /* Disable SDE clock gating */ 537 gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24)); 538 539 /* Setup RSC PDC handshake for sleep and wakeup */ 540 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1); 541 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0); 542 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0); 543 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0); 544 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0); 545 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000); 546 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0); 547 gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0); 548 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520); 549 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510); 550 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514); 551 552 /* Load RSC sequencer uCode for sleep and wakeup */ 553 if (adreno_is_a650_family(adreno_gpu)) { 554 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0); 555 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab); 556 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581); 557 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2); 558 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad); 559 } else { 560 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0); 561 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7); 562 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1); 563 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2); 564 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8); 565 } 566 567 if (pdc_in_aop) 568 goto setup_pdc; 569 570 /* Load PDC sequencer uCode for power up and power down sequence */ 571 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1); 572 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2); 573 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0); 574 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284); 575 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc); 576 577 /* Set TCS commands used by PDC sequence for low power modes */ 578 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7); 579 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0); 580 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0); 581 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108); 582 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010); 583 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1); 584 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108); 585 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000); 586 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0); 587 588 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108); 589 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset); 590 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0); 591 592 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7); 593 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0); 594 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0); 595 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108); 596 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010); 597 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2); 598 599 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108); 600 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000); 601 if (adreno_is_a618(adreno_gpu) || adreno_is_a619(adreno_gpu) || 602 adreno_is_a650_family(adreno_gpu)) 603 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2); 604 else 605 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3); 606 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108); 607 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset); 608 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3); 609 610 /* Setup GPU PDC */ 611 setup_pdc: 612 pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0); 613 pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001); 614 615 /* ensure no writes happen before the uCode is fully written */ 616 wmb(); 617 618 a6xx_rpmh_stop(gmu); 619 620 err: 621 if (!IS_ERR_OR_NULL(pdcptr)) 622 iounmap(pdcptr); 623 if (!IS_ERR_OR_NULL(seqptr)) 624 iounmap(seqptr); 625 } 626 627 /* 628 * The lowest 16 bits of this value are the number of XO clock cycles for main 629 * hysteresis which is set at 0x1680 cycles (300 us). The higher 16 bits are 630 * for the shorter hysteresis that happens after main - this is 0xa (.5 us) 631 */ 632 633 #define GMU_PWR_COL_HYST 0x000a1680 634 635 /* Set up the idle state for the GMU */ 636 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu) 637 { 638 /* Disable GMU WB/RB buffer */ 639 gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1); 640 gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1); 641 gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1); 642 643 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400); 644 645 switch (gmu->idle_level) { 646 case GMU_IDLE_STATE_IFPC: 647 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST, 648 GMU_PWR_COL_HYST); 649 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 650 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 651 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE); 652 fallthrough; 653 case GMU_IDLE_STATE_SPTP: 654 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST, 655 GMU_PWR_COL_HYST); 656 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 657 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 658 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE); 659 } 660 661 /* Enable RPMh GPU client */ 662 gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0, 663 A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE | 664 A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE | 665 A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE | 666 A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE | 667 A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE | 668 A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE); 669 } 670 671 struct block_header { 672 u32 addr; 673 u32 size; 674 u32 type; 675 u32 value; 676 u32 data[]; 677 }; 678 679 /* this should be a general kernel helper */ 680 static int in_range(u32 addr, u32 start, u32 size) 681 { 682 return addr >= start && addr < start + size; 683 } 684 685 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk) 686 { 687 if (!in_range(blk->addr, bo->iova, bo->size)) 688 return false; 689 690 memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size); 691 return true; 692 } 693 694 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu) 695 { 696 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 697 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 698 const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU]; 699 const struct block_header *blk; 700 u32 reg_offset; 701 702 u32 itcm_base = 0x00000000; 703 u32 dtcm_base = 0x00040000; 704 705 if (adreno_is_a650_family(adreno_gpu)) 706 dtcm_base = 0x10004000; 707 708 if (gmu->legacy) { 709 /* Sanity check the size of the firmware that was loaded */ 710 if (fw_image->size > 0x8000) { 711 DRM_DEV_ERROR(gmu->dev, 712 "GMU firmware is bigger than the available region\n"); 713 return -EINVAL; 714 } 715 716 gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START, 717 (u32*) fw_image->data, fw_image->size); 718 return 0; 719 } 720 721 722 for (blk = (const struct block_header *) fw_image->data; 723 (const u8*) blk < fw_image->data + fw_image->size; 724 blk = (const struct block_header *) &blk->data[blk->size >> 2]) { 725 if (blk->size == 0) 726 continue; 727 728 if (in_range(blk->addr, itcm_base, SZ_16K)) { 729 reg_offset = (blk->addr - itcm_base) >> 2; 730 gmu_write_bulk(gmu, 731 REG_A6XX_GMU_CM3_ITCM_START + reg_offset, 732 blk->data, blk->size); 733 } else if (in_range(blk->addr, dtcm_base, SZ_16K)) { 734 reg_offset = (blk->addr - dtcm_base) >> 2; 735 gmu_write_bulk(gmu, 736 REG_A6XX_GMU_CM3_DTCM_START + reg_offset, 737 blk->data, blk->size); 738 } else if (!fw_block_mem(&gmu->icache, blk) && 739 !fw_block_mem(&gmu->dcache, blk) && 740 !fw_block_mem(&gmu->dummy, blk)) { 741 DRM_DEV_ERROR(gmu->dev, 742 "failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n", 743 blk->addr, blk->size, blk->data[0]); 744 } 745 } 746 747 return 0; 748 } 749 750 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state) 751 { 752 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 753 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 754 int ret; 755 u32 chipid; 756 757 if (adreno_is_a650_family(adreno_gpu)) { 758 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FALNEXT_INTF, 1); 759 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1); 760 } 761 762 if (state == GMU_WARM_BOOT) { 763 ret = a6xx_rpmh_start(gmu); 764 if (ret) 765 return ret; 766 } else { 767 if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU], 768 "GMU firmware is not loaded\n")) 769 return -ENOENT; 770 771 /* Turn on register retention */ 772 gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1); 773 774 ret = a6xx_rpmh_start(gmu); 775 if (ret) 776 return ret; 777 778 ret = a6xx_gmu_fw_load(gmu); 779 if (ret) 780 return ret; 781 } 782 783 gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0); 784 gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02); 785 786 /* Write the iova of the HFI table */ 787 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova); 788 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1); 789 790 gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0, 791 (1 << 31) | (0xa << 18) | (0xa0)); 792 793 chipid = adreno_gpu->rev.core << 24; 794 chipid |= adreno_gpu->rev.major << 16; 795 chipid |= adreno_gpu->rev.minor << 12; 796 chipid |= adreno_gpu->rev.patchid << 8; 797 798 gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid); 799 800 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG, 801 gmu->log.iova | (gmu->log.size / SZ_4K - 1)); 802 803 /* Set up the lowest idle level on the GMU */ 804 a6xx_gmu_power_config(gmu); 805 806 ret = a6xx_gmu_start(gmu); 807 if (ret) 808 return ret; 809 810 if (gmu->legacy) { 811 ret = a6xx_gmu_gfx_rail_on(gmu); 812 if (ret) 813 return ret; 814 } 815 816 /* Enable SPTP_PC if the CPU is responsible for it */ 817 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) { 818 ret = a6xx_sptprac_enable(gmu); 819 if (ret) 820 return ret; 821 } 822 823 ret = a6xx_gmu_hfi_start(gmu); 824 if (ret) 825 return ret; 826 827 /* FIXME: Do we need this wmb() here? */ 828 wmb(); 829 830 return 0; 831 } 832 833 #define A6XX_HFI_IRQ_MASK \ 834 (A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) 835 836 #define A6XX_GMU_IRQ_MASK \ 837 (A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \ 838 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \ 839 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR) 840 841 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu) 842 { 843 disable_irq(gmu->gmu_irq); 844 disable_irq(gmu->hfi_irq); 845 846 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0); 847 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0); 848 } 849 850 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu) 851 { 852 u32 val; 853 854 /* Make sure there are no outstanding RPMh votes */ 855 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val, 856 (val & 1), 100, 10000); 857 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val, 858 (val & 1), 100, 10000); 859 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val, 860 (val & 1), 100, 10000); 861 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val, 862 (val & 1), 100, 1000); 863 } 864 865 /* Force the GMU off in case it isn't responsive */ 866 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu) 867 { 868 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 869 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 870 struct msm_gpu *gpu = &adreno_gpu->base; 871 872 /* 873 * Turn off keep alive that might have been enabled by the hang 874 * interrupt 875 */ 876 gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0); 877 878 /* Flush all the queues */ 879 a6xx_hfi_stop(gmu); 880 881 /* Stop the interrupts */ 882 a6xx_gmu_irq_disable(gmu); 883 884 /* Force off SPTP in case the GMU is managing it */ 885 a6xx_sptprac_disable(gmu); 886 887 /* Make sure there are no outstanding RPMh votes */ 888 a6xx_gmu_rpmh_off(gmu); 889 890 /* Halt the gmu cm3 core */ 891 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1); 892 893 a6xx_bus_clear_pending_transactions(adreno_gpu, true); 894 895 /* Reset GPU core blocks */ 896 a6xx_gpu_sw_reset(gpu, true); 897 } 898 899 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu) 900 { 901 struct dev_pm_opp *gpu_opp; 902 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index]; 903 904 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true); 905 if (IS_ERR(gpu_opp)) 906 return; 907 908 gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */ 909 a6xx_gmu_set_freq(gpu, gpu_opp, false); 910 dev_pm_opp_put(gpu_opp); 911 } 912 913 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu) 914 { 915 struct dev_pm_opp *gpu_opp; 916 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index]; 917 918 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true); 919 if (IS_ERR(gpu_opp)) 920 return; 921 922 dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp); 923 dev_pm_opp_put(gpu_opp); 924 } 925 926 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu) 927 { 928 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 929 struct msm_gpu *gpu = &adreno_gpu->base; 930 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 931 int status, ret; 932 933 if (WARN(!gmu->initialized, "The GMU is not set up yet\n")) 934 return -EINVAL; 935 936 gmu->hung = false; 937 938 /* Turn on the resources */ 939 pm_runtime_get_sync(gmu->dev); 940 941 /* 942 * "enable" the GX power domain which won't actually do anything but it 943 * will make sure that the refcounting is correct in case we need to 944 * bring down the GX after a GMU failure 945 */ 946 if (!IS_ERR_OR_NULL(gmu->gxpd)) 947 pm_runtime_get_sync(gmu->gxpd); 948 949 /* Use a known rate to bring up the GMU */ 950 clk_set_rate(gmu->core_clk, 200000000); 951 clk_set_rate(gmu->hub_clk, 150000000); 952 ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks); 953 if (ret) { 954 pm_runtime_put(gmu->gxpd); 955 pm_runtime_put(gmu->dev); 956 return ret; 957 } 958 959 /* Set the bus quota to a reasonable value for boot */ 960 a6xx_gmu_set_initial_bw(gpu, gmu); 961 962 /* Enable the GMU interrupt */ 963 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0); 964 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK); 965 enable_irq(gmu->gmu_irq); 966 967 /* Check to see if we are doing a cold or warm boot */ 968 status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ? 969 GMU_WARM_BOOT : GMU_COLD_BOOT; 970 971 /* 972 * Warm boot path does not work on newer GPUs 973 * Presumably this is because icache/dcache regions must be restored 974 */ 975 if (!gmu->legacy) 976 status = GMU_COLD_BOOT; 977 978 ret = a6xx_gmu_fw_start(gmu, status); 979 if (ret) 980 goto out; 981 982 ret = a6xx_hfi_start(gmu, status); 983 if (ret) 984 goto out; 985 986 /* 987 * Turn on the GMU firmware fault interrupt after we know the boot 988 * sequence is successful 989 */ 990 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0); 991 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK); 992 enable_irq(gmu->hfi_irq); 993 994 /* Set the GPU to the current freq */ 995 a6xx_gmu_set_initial_freq(gpu, gmu); 996 997 out: 998 /* On failure, shut down the GMU to leave it in a good state */ 999 if (ret) { 1000 disable_irq(gmu->gmu_irq); 1001 a6xx_rpmh_stop(gmu); 1002 pm_runtime_put(gmu->gxpd); 1003 pm_runtime_put(gmu->dev); 1004 } 1005 1006 return ret; 1007 } 1008 1009 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu) 1010 { 1011 u32 reg; 1012 1013 if (!gmu->initialized) 1014 return true; 1015 1016 reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS); 1017 1018 if (reg & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB) 1019 return false; 1020 1021 return true; 1022 } 1023 1024 /* Gracefully try to shut down the GMU and by extension the GPU */ 1025 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu) 1026 { 1027 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1028 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1029 u32 val; 1030 1031 /* 1032 * The GMU may still be in slumber unless the GPU started so check and 1033 * skip putting it back into slumber if so 1034 */ 1035 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); 1036 1037 if (val != 0xf) { 1038 int ret = a6xx_gmu_wait_for_idle(gmu); 1039 1040 /* If the GMU isn't responding assume it is hung */ 1041 if (ret) { 1042 a6xx_gmu_force_off(gmu); 1043 return; 1044 } 1045 1046 a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung); 1047 1048 /* tell the GMU we want to slumber */ 1049 ret = a6xx_gmu_notify_slumber(gmu); 1050 if (ret) { 1051 a6xx_gmu_force_off(gmu); 1052 return; 1053 } 1054 1055 ret = gmu_poll_timeout(gmu, 1056 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val, 1057 !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB), 1058 100, 10000); 1059 1060 /* 1061 * Let the user know we failed to slumber but don't worry too 1062 * much because we are powering down anyway 1063 */ 1064 1065 if (ret) 1066 DRM_DEV_ERROR(gmu->dev, 1067 "Unable to slumber GMU: status = 0%x/0%x\n", 1068 gmu_read(gmu, 1069 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS), 1070 gmu_read(gmu, 1071 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2)); 1072 } 1073 1074 /* Turn off HFI */ 1075 a6xx_hfi_stop(gmu); 1076 1077 /* Stop the interrupts and mask the hardware */ 1078 a6xx_gmu_irq_disable(gmu); 1079 1080 /* Tell RPMh to power off the GPU */ 1081 a6xx_rpmh_stop(gmu); 1082 } 1083 1084 1085 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu) 1086 { 1087 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1088 struct msm_gpu *gpu = &a6xx_gpu->base.base; 1089 1090 if (!pm_runtime_active(gmu->dev)) 1091 return 0; 1092 1093 /* 1094 * Force the GMU off if we detected a hang, otherwise try to shut it 1095 * down gracefully 1096 */ 1097 if (gmu->hung) 1098 a6xx_gmu_force_off(gmu); 1099 else 1100 a6xx_gmu_shutdown(gmu); 1101 1102 /* Remove the bus vote */ 1103 dev_pm_opp_set_opp(&gpu->pdev->dev, NULL); 1104 1105 /* 1106 * Make sure the GX domain is off before turning off the GMU (CX) 1107 * domain. Usually the GMU does this but only if the shutdown sequence 1108 * was successful 1109 */ 1110 if (!IS_ERR_OR_NULL(gmu->gxpd)) 1111 pm_runtime_put_sync(gmu->gxpd); 1112 1113 clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks); 1114 1115 pm_runtime_put_sync(gmu->dev); 1116 1117 return 0; 1118 } 1119 1120 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu) 1121 { 1122 msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace); 1123 msm_gem_kernel_put(gmu->debug.obj, gmu->aspace); 1124 msm_gem_kernel_put(gmu->icache.obj, gmu->aspace); 1125 msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace); 1126 msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace); 1127 msm_gem_kernel_put(gmu->log.obj, gmu->aspace); 1128 1129 gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu); 1130 msm_gem_address_space_put(gmu->aspace); 1131 } 1132 1133 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo, 1134 size_t size, u64 iova, const char *name) 1135 { 1136 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1137 struct drm_device *dev = a6xx_gpu->base.base.dev; 1138 uint32_t flags = MSM_BO_WC; 1139 u64 range_start, range_end; 1140 int ret; 1141 1142 size = PAGE_ALIGN(size); 1143 if (!iova) { 1144 /* no fixed address - use GMU's uncached range */ 1145 range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */ 1146 range_end = 0x80000000; 1147 } else { 1148 /* range for fixed address */ 1149 range_start = iova; 1150 range_end = iova + size; 1151 /* use IOMMU_PRIV for icache/dcache */ 1152 flags |= MSM_BO_MAP_PRIV; 1153 } 1154 1155 bo->obj = msm_gem_new(dev, size, flags); 1156 if (IS_ERR(bo->obj)) 1157 return PTR_ERR(bo->obj); 1158 1159 ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova, 1160 range_start, range_end); 1161 if (ret) { 1162 drm_gem_object_put(bo->obj); 1163 return ret; 1164 } 1165 1166 bo->virt = msm_gem_get_vaddr(bo->obj); 1167 bo->size = size; 1168 1169 msm_gem_object_set_name(bo->obj, name); 1170 1171 return 0; 1172 } 1173 1174 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu) 1175 { 1176 struct msm_mmu *mmu; 1177 1178 mmu = msm_iommu_new(gmu->dev, 0); 1179 if (!mmu) 1180 return -ENODEV; 1181 if (IS_ERR(mmu)) 1182 return PTR_ERR(mmu); 1183 1184 gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000); 1185 if (IS_ERR(gmu->aspace)) 1186 return PTR_ERR(gmu->aspace); 1187 1188 return 0; 1189 } 1190 1191 /* Return the 'arc-level' for the given frequency */ 1192 static unsigned int a6xx_gmu_get_arc_level(struct device *dev, 1193 unsigned long freq) 1194 { 1195 struct dev_pm_opp *opp; 1196 unsigned int val; 1197 1198 if (!freq) 1199 return 0; 1200 1201 opp = dev_pm_opp_find_freq_exact(dev, freq, true); 1202 if (IS_ERR(opp)) 1203 return 0; 1204 1205 val = dev_pm_opp_get_level(opp); 1206 1207 dev_pm_opp_put(opp); 1208 1209 return val; 1210 } 1211 1212 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes, 1213 unsigned long *freqs, int freqs_count, const char *id) 1214 { 1215 int i, j; 1216 const u16 *pri, *sec; 1217 size_t pri_count, sec_count; 1218 1219 pri = cmd_db_read_aux_data(id, &pri_count); 1220 if (IS_ERR(pri)) 1221 return PTR_ERR(pri); 1222 /* 1223 * The data comes back as an array of unsigned shorts so adjust the 1224 * count accordingly 1225 */ 1226 pri_count >>= 1; 1227 if (!pri_count) 1228 return -EINVAL; 1229 1230 sec = cmd_db_read_aux_data("mx.lvl", &sec_count); 1231 if (IS_ERR(sec)) 1232 return PTR_ERR(sec); 1233 1234 sec_count >>= 1; 1235 if (!sec_count) 1236 return -EINVAL; 1237 1238 /* Construct a vote for each frequency */ 1239 for (i = 0; i < freqs_count; i++) { 1240 u8 pindex = 0, sindex = 0; 1241 unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]); 1242 1243 /* Get the primary index that matches the arc level */ 1244 for (j = 0; j < pri_count; j++) { 1245 if (pri[j] >= level) { 1246 pindex = j; 1247 break; 1248 } 1249 } 1250 1251 if (j == pri_count) { 1252 DRM_DEV_ERROR(dev, 1253 "Level %u not found in the RPMh list\n", 1254 level); 1255 DRM_DEV_ERROR(dev, "Available levels:\n"); 1256 for (j = 0; j < pri_count; j++) 1257 DRM_DEV_ERROR(dev, " %u\n", pri[j]); 1258 1259 return -EINVAL; 1260 } 1261 1262 /* 1263 * Look for a level in in the secondary list that matches. If 1264 * nothing fits, use the maximum non zero vote 1265 */ 1266 1267 for (j = 0; j < sec_count; j++) { 1268 if (sec[j] >= level) { 1269 sindex = j; 1270 break; 1271 } else if (sec[j]) { 1272 sindex = j; 1273 } 1274 } 1275 1276 /* Construct the vote */ 1277 votes[i] = ((pri[pindex] & 0xffff) << 16) | 1278 (sindex << 8) | pindex; 1279 } 1280 1281 return 0; 1282 } 1283 1284 /* 1285 * The GMU votes with the RPMh for itself and on behalf of the GPU but we need 1286 * to construct the list of votes on the CPU and send it over. Query the RPMh 1287 * voltage levels and build the votes 1288 */ 1289 1290 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu) 1291 { 1292 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1293 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1294 struct msm_gpu *gpu = &adreno_gpu->base; 1295 int ret; 1296 1297 /* Build the GX votes */ 1298 ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes, 1299 gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl"); 1300 1301 /* Build the CX votes */ 1302 ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes, 1303 gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl"); 1304 1305 return ret; 1306 } 1307 1308 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs, 1309 u32 size) 1310 { 1311 int count = dev_pm_opp_get_opp_count(dev); 1312 struct dev_pm_opp *opp; 1313 int i, index = 0; 1314 unsigned long freq = 1; 1315 1316 /* 1317 * The OPP table doesn't contain the "off" frequency level so we need to 1318 * add 1 to the table size to account for it 1319 */ 1320 1321 if (WARN(count + 1 > size, 1322 "The GMU frequency table is being truncated\n")) 1323 count = size - 1; 1324 1325 /* Set the "off" frequency */ 1326 freqs[index++] = 0; 1327 1328 for (i = 0; i < count; i++) { 1329 opp = dev_pm_opp_find_freq_ceil(dev, &freq); 1330 if (IS_ERR(opp)) 1331 break; 1332 1333 dev_pm_opp_put(opp); 1334 freqs[index++] = freq++; 1335 } 1336 1337 return index; 1338 } 1339 1340 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu) 1341 { 1342 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1343 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1344 struct msm_gpu *gpu = &adreno_gpu->base; 1345 1346 int ret = 0; 1347 1348 /* 1349 * The GMU handles its own frequency switching so build a list of 1350 * available frequencies to send during initialization 1351 */ 1352 ret = devm_pm_opp_of_add_table(gmu->dev); 1353 if (ret) { 1354 DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n"); 1355 return ret; 1356 } 1357 1358 gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev, 1359 gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs)); 1360 1361 /* 1362 * The GMU also handles GPU frequency switching so build a list 1363 * from the GPU OPP table 1364 */ 1365 gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev, 1366 gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs)); 1367 1368 gmu->current_perf_index = gmu->nr_gpu_freqs - 1; 1369 1370 /* Build the list of RPMh votes that we'll send to the GMU */ 1371 return a6xx_gmu_rpmh_votes_init(gmu); 1372 } 1373 1374 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu) 1375 { 1376 int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks); 1377 1378 if (ret < 1) 1379 return ret; 1380 1381 gmu->nr_clocks = ret; 1382 1383 gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks, 1384 gmu->nr_clocks, "gmu"); 1385 1386 gmu->hub_clk = msm_clk_bulk_get_clock(gmu->clocks, 1387 gmu->nr_clocks, "hub"); 1388 1389 return 0; 1390 } 1391 1392 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 1393 const char *name) 1394 { 1395 void __iomem *ret; 1396 struct resource *res = platform_get_resource_byname(pdev, 1397 IORESOURCE_MEM, name); 1398 1399 if (!res) { 1400 DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name); 1401 return ERR_PTR(-EINVAL); 1402 } 1403 1404 ret = ioremap(res->start, resource_size(res)); 1405 if (!ret) { 1406 DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name); 1407 return ERR_PTR(-EINVAL); 1408 } 1409 1410 return ret; 1411 } 1412 1413 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev, 1414 const char *name, irq_handler_t handler) 1415 { 1416 int irq, ret; 1417 1418 irq = platform_get_irq_byname(pdev, name); 1419 1420 ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu); 1421 if (ret) { 1422 DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n", 1423 name, ret); 1424 return ret; 1425 } 1426 1427 disable_irq(irq); 1428 1429 return irq; 1430 } 1431 1432 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu) 1433 { 1434 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1435 struct platform_device *pdev = to_platform_device(gmu->dev); 1436 1437 if (!gmu->initialized) 1438 return; 1439 1440 pm_runtime_force_suspend(gmu->dev); 1441 1442 /* 1443 * Since cxpd is a virt device, the devlink with gmu-dev will be removed 1444 * automatically when we do detach 1445 */ 1446 dev_pm_domain_detach(gmu->cxpd, false); 1447 1448 if (!IS_ERR_OR_NULL(gmu->gxpd)) { 1449 pm_runtime_disable(gmu->gxpd); 1450 dev_pm_domain_detach(gmu->gxpd, false); 1451 } 1452 1453 iounmap(gmu->mmio); 1454 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc")) 1455 iounmap(gmu->rscc); 1456 gmu->mmio = NULL; 1457 gmu->rscc = NULL; 1458 1459 a6xx_gmu_memory_free(gmu); 1460 1461 free_irq(gmu->gmu_irq, gmu); 1462 free_irq(gmu->hfi_irq, gmu); 1463 1464 /* Drop reference taken in of_find_device_by_node */ 1465 put_device(gmu->dev); 1466 1467 gmu->initialized = false; 1468 } 1469 1470 static int cxpd_notifier_cb(struct notifier_block *nb, 1471 unsigned long action, void *data) 1472 { 1473 struct a6xx_gmu *gmu = container_of(nb, struct a6xx_gmu, pd_nb); 1474 1475 if (action == GENPD_NOTIFY_OFF) 1476 complete_all(&gmu->pd_gate); 1477 1478 return 0; 1479 } 1480 1481 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node) 1482 { 1483 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1484 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1485 struct platform_device *pdev = of_find_device_by_node(node); 1486 int ret; 1487 1488 if (!pdev) 1489 return -ENODEV; 1490 1491 gmu->dev = &pdev->dev; 1492 1493 of_dma_configure(gmu->dev, node, true); 1494 1495 /* Fow now, don't do anything fancy until we get our feet under us */ 1496 gmu->idle_level = GMU_IDLE_STATE_ACTIVE; 1497 1498 pm_runtime_enable(gmu->dev); 1499 1500 /* Get the list of clocks */ 1501 ret = a6xx_gmu_clocks_probe(gmu); 1502 if (ret) 1503 goto err_put_device; 1504 1505 ret = a6xx_gmu_memory_probe(gmu); 1506 if (ret) 1507 goto err_put_device; 1508 1509 1510 /* A660 now requires handling "prealloc requests" in GMU firmware 1511 * For now just hardcode allocations based on the known firmware. 1512 * note: there is no indication that these correspond to "dummy" or 1513 * "debug" regions, but this "guess" allows reusing these BOs which 1514 * are otherwise unused by a660. 1515 */ 1516 gmu->dummy.size = SZ_4K; 1517 if (adreno_is_a660_family(adreno_gpu)) { 1518 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_4K * 7, 1519 0x60400000, "debug"); 1520 if (ret) 1521 goto err_memory; 1522 1523 gmu->dummy.size = SZ_8K; 1524 } 1525 1526 /* Allocate memory for the GMU dummy page */ 1527 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, gmu->dummy.size, 1528 0x60000000, "dummy"); 1529 if (ret) 1530 goto err_memory; 1531 1532 /* Note that a650 family also includes a660 family: */ 1533 if (adreno_is_a650_family(adreno_gpu)) { 1534 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache, 1535 SZ_16M - SZ_16K, 0x04000, "icache"); 1536 if (ret) 1537 goto err_memory; 1538 /* 1539 * NOTE: when porting legacy ("pre-650-family") GPUs you may be tempted to add a condition 1540 * to allocate icache/dcache here, as per downstream code flow, but it may not actually be 1541 * necessary. If you omit this step and you don't get random pagefaults, you are likely 1542 * good to go without this! 1543 */ 1544 } else if (adreno_is_a640_family(adreno_gpu)) { 1545 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache, 1546 SZ_256K - SZ_16K, 0x04000, "icache"); 1547 if (ret) 1548 goto err_memory; 1549 1550 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache, 1551 SZ_256K - SZ_16K, 0x44000, "dcache"); 1552 if (ret) 1553 goto err_memory; 1554 } else if (adreno_is_a630(adreno_gpu) || adreno_is_a615_family(adreno_gpu)) { 1555 /* HFI v1, has sptprac */ 1556 gmu->legacy = true; 1557 1558 /* Allocate memory for the GMU debug region */ 1559 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0, "debug"); 1560 if (ret) 1561 goto err_memory; 1562 } 1563 1564 /* Allocate memory for for the HFI queues */ 1565 ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0, "hfi"); 1566 if (ret) 1567 goto err_memory; 1568 1569 /* Allocate memory for the GMU log region */ 1570 ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0, "log"); 1571 if (ret) 1572 goto err_memory; 1573 1574 /* Map the GMU registers */ 1575 gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu"); 1576 if (IS_ERR(gmu->mmio)) { 1577 ret = PTR_ERR(gmu->mmio); 1578 goto err_memory; 1579 } 1580 1581 if (adreno_is_a650_family(adreno_gpu)) { 1582 gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc"); 1583 if (IS_ERR(gmu->rscc)) { 1584 ret = -ENODEV; 1585 goto err_mmio; 1586 } 1587 } else { 1588 gmu->rscc = gmu->mmio + 0x23000; 1589 } 1590 1591 /* Get the HFI and GMU interrupts */ 1592 gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq); 1593 gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq); 1594 1595 if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) { 1596 ret = -ENODEV; 1597 goto err_mmio; 1598 } 1599 1600 gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx"); 1601 if (IS_ERR(gmu->cxpd)) { 1602 ret = PTR_ERR(gmu->cxpd); 1603 goto err_mmio; 1604 } 1605 1606 if (!device_link_add(gmu->dev, gmu->cxpd, 1607 DL_FLAG_PM_RUNTIME)) { 1608 ret = -ENODEV; 1609 goto detach_cxpd; 1610 } 1611 1612 init_completion(&gmu->pd_gate); 1613 complete_all(&gmu->pd_gate); 1614 gmu->pd_nb.notifier_call = cxpd_notifier_cb; 1615 1616 /* 1617 * Get a link to the GX power domain to reset the GPU in case of GMU 1618 * crash 1619 */ 1620 gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx"); 1621 1622 /* Get the power levels for the GMU and GPU */ 1623 a6xx_gmu_pwrlevels_probe(gmu); 1624 1625 /* Set up the HFI queues */ 1626 a6xx_hfi_init(gmu); 1627 1628 /* Initialize RPMh */ 1629 a6xx_gmu_rpmh_init(gmu); 1630 1631 gmu->initialized = true; 1632 1633 return 0; 1634 1635 detach_cxpd: 1636 dev_pm_domain_detach(gmu->cxpd, false); 1637 1638 err_mmio: 1639 iounmap(gmu->mmio); 1640 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc")) 1641 iounmap(gmu->rscc); 1642 free_irq(gmu->gmu_irq, gmu); 1643 free_irq(gmu->hfi_irq, gmu); 1644 1645 err_memory: 1646 a6xx_gmu_memory_free(gmu); 1647 err_put_device: 1648 /* Drop reference taken in of_find_device_by_node */ 1649 put_device(gmu->dev); 1650 1651 return ret; 1652 } 1653