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 static 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 static 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 /* Set up CX GMU counter 0 to count busy ticks */ 483 gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000); 484 gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20); 485 486 /* Enable the power counter */ 487 gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1); 488 return 0; 489 } 490 491 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu) 492 { 493 int ret; 494 u32 val; 495 496 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1); 497 498 ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, 499 val, val & (1 << 16), 100, 10000); 500 if (ret) 501 DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n"); 502 503 gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0); 504 } 505 506 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value) 507 { 508 msm_writel(value, ptr + (offset << 2)); 509 } 510 511 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 512 const char *name); 513 514 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu) 515 { 516 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 517 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 518 struct platform_device *pdev = to_platform_device(gmu->dev); 519 void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc"); 520 void __iomem *seqptr = NULL; 521 uint32_t pdc_address_offset; 522 bool pdc_in_aop = false; 523 524 if (IS_ERR(pdcptr)) 525 goto err; 526 527 if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu)) 528 pdc_in_aop = true; 529 else if (adreno_is_a618(adreno_gpu) || adreno_is_a640_family(adreno_gpu)) 530 pdc_address_offset = 0x30090; 531 else if (adreno_is_a619(adreno_gpu)) 532 pdc_address_offset = 0x300a0; 533 else 534 pdc_address_offset = 0x30080; 535 536 if (!pdc_in_aop) { 537 seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq"); 538 if (IS_ERR(seqptr)) 539 goto err; 540 } 541 542 /* Disable SDE clock gating */ 543 gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24)); 544 545 /* Setup RSC PDC handshake for sleep and wakeup */ 546 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1); 547 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0); 548 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0); 549 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0); 550 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0); 551 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000); 552 gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0); 553 gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0); 554 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520); 555 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510); 556 gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514); 557 558 /* Load RSC sequencer uCode for sleep and wakeup */ 559 if (adreno_is_a650_family(adreno_gpu)) { 560 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0); 561 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab); 562 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581); 563 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2); 564 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad); 565 } else { 566 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0); 567 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7); 568 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1); 569 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2); 570 gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8); 571 } 572 573 if (pdc_in_aop) 574 goto setup_pdc; 575 576 /* Load PDC sequencer uCode for power up and power down sequence */ 577 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1); 578 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2); 579 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0); 580 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284); 581 pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc); 582 583 /* Set TCS commands used by PDC sequence for low power modes */ 584 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7); 585 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0); 586 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0); 587 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108); 588 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010); 589 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1); 590 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108); 591 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000); 592 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0); 593 594 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108); 595 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset); 596 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0); 597 598 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7); 599 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0); 600 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0); 601 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108); 602 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010); 603 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2); 604 605 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108); 606 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000); 607 if (adreno_is_a618(adreno_gpu) || adreno_is_a619(adreno_gpu) || 608 adreno_is_a650_family(adreno_gpu)) 609 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2); 610 else 611 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3); 612 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108); 613 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset); 614 pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3); 615 616 /* Setup GPU PDC */ 617 setup_pdc: 618 pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0); 619 pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001); 620 621 /* ensure no writes happen before the uCode is fully written */ 622 wmb(); 623 624 err: 625 if (!IS_ERR_OR_NULL(pdcptr)) 626 iounmap(pdcptr); 627 if (!IS_ERR_OR_NULL(seqptr)) 628 iounmap(seqptr); 629 } 630 631 /* 632 * The lowest 16 bits of this value are the number of XO clock cycles for main 633 * hysteresis which is set at 0x1680 cycles (300 us). The higher 16 bits are 634 * for the shorter hysteresis that happens after main - this is 0xa (.5 us) 635 */ 636 637 #define GMU_PWR_COL_HYST 0x000a1680 638 639 /* Set up the idle state for the GMU */ 640 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu) 641 { 642 /* Disable GMU WB/RB buffer */ 643 gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1); 644 gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1); 645 gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1); 646 647 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400); 648 649 switch (gmu->idle_level) { 650 case GMU_IDLE_STATE_IFPC: 651 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST, 652 GMU_PWR_COL_HYST); 653 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 654 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 655 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE); 656 fallthrough; 657 case GMU_IDLE_STATE_SPTP: 658 gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST, 659 GMU_PWR_COL_HYST); 660 gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0, 661 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE | 662 A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE); 663 } 664 665 /* Enable RPMh GPU client */ 666 gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0, 667 A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE | 668 A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE | 669 A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE | 670 A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE | 671 A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE | 672 A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE); 673 } 674 675 struct block_header { 676 u32 addr; 677 u32 size; 678 u32 type; 679 u32 value; 680 u32 data[]; 681 }; 682 683 /* this should be a general kernel helper */ 684 static int in_range(u32 addr, u32 start, u32 size) 685 { 686 return addr >= start && addr < start + size; 687 } 688 689 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk) 690 { 691 if (!in_range(blk->addr, bo->iova, bo->size)) 692 return false; 693 694 memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size); 695 return true; 696 } 697 698 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu) 699 { 700 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 701 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 702 const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU]; 703 const struct block_header *blk; 704 u32 reg_offset; 705 706 u32 itcm_base = 0x00000000; 707 u32 dtcm_base = 0x00040000; 708 709 if (adreno_is_a650_family(adreno_gpu)) 710 dtcm_base = 0x10004000; 711 712 if (gmu->legacy) { 713 /* Sanity check the size of the firmware that was loaded */ 714 if (fw_image->size > 0x8000) { 715 DRM_DEV_ERROR(gmu->dev, 716 "GMU firmware is bigger than the available region\n"); 717 return -EINVAL; 718 } 719 720 gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START, 721 (u32*) fw_image->data, fw_image->size); 722 return 0; 723 } 724 725 726 for (blk = (const struct block_header *) fw_image->data; 727 (const u8*) blk < fw_image->data + fw_image->size; 728 blk = (const struct block_header *) &blk->data[blk->size >> 2]) { 729 if (blk->size == 0) 730 continue; 731 732 if (in_range(blk->addr, itcm_base, SZ_16K)) { 733 reg_offset = (blk->addr - itcm_base) >> 2; 734 gmu_write_bulk(gmu, 735 REG_A6XX_GMU_CM3_ITCM_START + reg_offset, 736 blk->data, blk->size); 737 } else if (in_range(blk->addr, dtcm_base, SZ_16K)) { 738 reg_offset = (blk->addr - dtcm_base) >> 2; 739 gmu_write_bulk(gmu, 740 REG_A6XX_GMU_CM3_DTCM_START + reg_offset, 741 blk->data, blk->size); 742 } else if (!fw_block_mem(&gmu->icache, blk) && 743 !fw_block_mem(&gmu->dcache, blk) && 744 !fw_block_mem(&gmu->dummy, blk)) { 745 DRM_DEV_ERROR(gmu->dev, 746 "failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n", 747 blk->addr, blk->size, blk->data[0]); 748 } 749 } 750 751 return 0; 752 } 753 754 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state) 755 { 756 static bool rpmh_init; 757 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 758 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 759 int ret; 760 u32 chipid; 761 762 if (adreno_is_a650_family(adreno_gpu)) { 763 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FALNEXT_INTF, 1); 764 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1); 765 } 766 767 if (state == GMU_WARM_BOOT) { 768 ret = a6xx_rpmh_start(gmu); 769 if (ret) 770 return ret; 771 } else { 772 if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU], 773 "GMU firmware is not loaded\n")) 774 return -ENOENT; 775 776 /* Turn on register retention */ 777 gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1); 778 779 /* We only need to load the RPMh microcode once */ 780 if (!rpmh_init) { 781 a6xx_gmu_rpmh_init(gmu); 782 rpmh_init = true; 783 } else { 784 ret = a6xx_rpmh_start(gmu); 785 if (ret) 786 return ret; 787 } 788 789 ret = a6xx_gmu_fw_load(gmu); 790 if (ret) 791 return ret; 792 } 793 794 gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0); 795 gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02); 796 797 /* Write the iova of the HFI table */ 798 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova); 799 gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1); 800 801 gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0, 802 (1 << 31) | (0xa << 18) | (0xa0)); 803 804 chipid = adreno_gpu->rev.core << 24; 805 chipid |= adreno_gpu->rev.major << 16; 806 chipid |= adreno_gpu->rev.minor << 12; 807 chipid |= adreno_gpu->rev.patchid << 8; 808 809 gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid); 810 811 gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG, 812 gmu->log.iova | (gmu->log.size / SZ_4K - 1)); 813 814 /* Set up the lowest idle level on the GMU */ 815 a6xx_gmu_power_config(gmu); 816 817 ret = a6xx_gmu_start(gmu); 818 if (ret) 819 return ret; 820 821 if (gmu->legacy) { 822 ret = a6xx_gmu_gfx_rail_on(gmu); 823 if (ret) 824 return ret; 825 } 826 827 /* Enable SPTP_PC if the CPU is responsible for it */ 828 if (gmu->idle_level < GMU_IDLE_STATE_SPTP) { 829 ret = a6xx_sptprac_enable(gmu); 830 if (ret) 831 return ret; 832 } 833 834 ret = a6xx_gmu_hfi_start(gmu); 835 if (ret) 836 return ret; 837 838 /* FIXME: Do we need this wmb() here? */ 839 wmb(); 840 841 return 0; 842 } 843 844 #define A6XX_HFI_IRQ_MASK \ 845 (A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) 846 847 #define A6XX_GMU_IRQ_MASK \ 848 (A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \ 849 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \ 850 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR) 851 852 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu) 853 { 854 disable_irq(gmu->gmu_irq); 855 disable_irq(gmu->hfi_irq); 856 857 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0); 858 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0); 859 } 860 861 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu) 862 { 863 u32 val; 864 865 /* Make sure there are no outstanding RPMh votes */ 866 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val, 867 (val & 1), 100, 10000); 868 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val, 869 (val & 1), 100, 10000); 870 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val, 871 (val & 1), 100, 10000); 872 gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val, 873 (val & 1), 100, 1000); 874 } 875 876 #define GBIF_CLIENT_HALT_MASK BIT(0) 877 #define GBIF_ARB_HALT_MASK BIT(1) 878 879 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu) 880 { 881 struct msm_gpu *gpu = &adreno_gpu->base; 882 883 if (!a6xx_has_gbif(adreno_gpu)) { 884 gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf); 885 spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) & 886 0xf) == 0xf); 887 gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0); 888 889 return; 890 } 891 892 /* Halt the gx side of GBIF */ 893 gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 1); 894 spin_until(gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT_ACK) & 1); 895 896 /* Halt new client requests on GBIF */ 897 gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK); 898 spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) & 899 (GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK); 900 901 /* Halt all AXI requests on GBIF */ 902 gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK); 903 spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) & 904 (GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK); 905 906 /* The GBIF halt needs to be explicitly cleared */ 907 gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0); 908 } 909 910 /* Force the GMU off in case it isn't responsive */ 911 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu) 912 { 913 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 914 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 915 struct msm_gpu *gpu = &adreno_gpu->base; 916 917 /* Flush all the queues */ 918 a6xx_hfi_stop(gmu); 919 920 /* Stop the interrupts */ 921 a6xx_gmu_irq_disable(gmu); 922 923 /* Force off SPTP in case the GMU is managing it */ 924 a6xx_sptprac_disable(gmu); 925 926 /* Make sure there are no outstanding RPMh votes */ 927 a6xx_gmu_rpmh_off(gmu); 928 929 /* Halt the gmu cm3 core */ 930 gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1); 931 932 a6xx_bus_clear_pending_transactions(adreno_gpu); 933 934 /* Reset GPU core blocks */ 935 gpu_write(gpu, REG_A6XX_RBBM_SW_RESET_CMD, 1); 936 udelay(100); 937 } 938 939 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu) 940 { 941 struct dev_pm_opp *gpu_opp; 942 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index]; 943 944 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true); 945 if (IS_ERR(gpu_opp)) 946 return; 947 948 gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */ 949 a6xx_gmu_set_freq(gpu, gpu_opp, false); 950 dev_pm_opp_put(gpu_opp); 951 } 952 953 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu) 954 { 955 struct dev_pm_opp *gpu_opp; 956 unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index]; 957 958 gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true); 959 if (IS_ERR(gpu_opp)) 960 return; 961 962 dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp); 963 dev_pm_opp_put(gpu_opp); 964 } 965 966 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu) 967 { 968 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 969 struct msm_gpu *gpu = &adreno_gpu->base; 970 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 971 int status, ret; 972 973 if (WARN(!gmu->initialized, "The GMU is not set up yet\n")) 974 return 0; 975 976 gmu->hung = false; 977 978 /* Turn on the resources */ 979 pm_runtime_get_sync(gmu->dev); 980 981 /* 982 * "enable" the GX power domain which won't actually do anything but it 983 * will make sure that the refcounting is correct in case we need to 984 * bring down the GX after a GMU failure 985 */ 986 if (!IS_ERR_OR_NULL(gmu->gxpd)) 987 pm_runtime_get_sync(gmu->gxpd); 988 989 /* Use a known rate to bring up the GMU */ 990 clk_set_rate(gmu->core_clk, 200000000); 991 clk_set_rate(gmu->hub_clk, 150000000); 992 ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks); 993 if (ret) { 994 pm_runtime_put(gmu->gxpd); 995 pm_runtime_put(gmu->dev); 996 return ret; 997 } 998 999 /* Set the bus quota to a reasonable value for boot */ 1000 a6xx_gmu_set_initial_bw(gpu, gmu); 1001 1002 /* Enable the GMU interrupt */ 1003 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0); 1004 gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK); 1005 enable_irq(gmu->gmu_irq); 1006 1007 /* Check to see if we are doing a cold or warm boot */ 1008 status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ? 1009 GMU_WARM_BOOT : GMU_COLD_BOOT; 1010 1011 /* 1012 * Warm boot path does not work on newer GPUs 1013 * Presumably this is because icache/dcache regions must be restored 1014 */ 1015 if (!gmu->legacy) 1016 status = GMU_COLD_BOOT; 1017 1018 ret = a6xx_gmu_fw_start(gmu, status); 1019 if (ret) 1020 goto out; 1021 1022 ret = a6xx_hfi_start(gmu, status); 1023 if (ret) 1024 goto out; 1025 1026 /* 1027 * Turn on the GMU firmware fault interrupt after we know the boot 1028 * sequence is successful 1029 */ 1030 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0); 1031 gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK); 1032 enable_irq(gmu->hfi_irq); 1033 1034 /* Set the GPU to the current freq */ 1035 a6xx_gmu_set_initial_freq(gpu, gmu); 1036 1037 out: 1038 /* On failure, shut down the GMU to leave it in a good state */ 1039 if (ret) { 1040 disable_irq(gmu->gmu_irq); 1041 a6xx_rpmh_stop(gmu); 1042 pm_runtime_put(gmu->gxpd); 1043 pm_runtime_put(gmu->dev); 1044 } 1045 1046 return ret; 1047 } 1048 1049 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu) 1050 { 1051 u32 reg; 1052 1053 if (!gmu->initialized) 1054 return true; 1055 1056 reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS); 1057 1058 if (reg & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB) 1059 return false; 1060 1061 return true; 1062 } 1063 1064 /* Gracefully try to shut down the GMU and by extension the GPU */ 1065 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu) 1066 { 1067 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1068 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1069 u32 val; 1070 1071 /* 1072 * The GMU may still be in slumber unless the GPU started so check and 1073 * skip putting it back into slumber if so 1074 */ 1075 val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE); 1076 1077 if (val != 0xf) { 1078 int ret = a6xx_gmu_wait_for_idle(gmu); 1079 1080 /* If the GMU isn't responding assume it is hung */ 1081 if (ret) { 1082 a6xx_gmu_force_off(gmu); 1083 return; 1084 } 1085 1086 a6xx_bus_clear_pending_transactions(adreno_gpu); 1087 1088 /* tell the GMU we want to slumber */ 1089 ret = a6xx_gmu_notify_slumber(gmu); 1090 if (ret) { 1091 a6xx_gmu_force_off(gmu); 1092 return; 1093 } 1094 1095 ret = gmu_poll_timeout(gmu, 1096 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val, 1097 !(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB), 1098 100, 10000); 1099 1100 /* 1101 * Let the user know we failed to slumber but don't worry too 1102 * much because we are powering down anyway 1103 */ 1104 1105 if (ret) 1106 DRM_DEV_ERROR(gmu->dev, 1107 "Unable to slumber GMU: status = 0%x/0%x\n", 1108 gmu_read(gmu, 1109 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS), 1110 gmu_read(gmu, 1111 REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2)); 1112 } 1113 1114 /* Turn off HFI */ 1115 a6xx_hfi_stop(gmu); 1116 1117 /* Stop the interrupts and mask the hardware */ 1118 a6xx_gmu_irq_disable(gmu); 1119 1120 /* Tell RPMh to power off the GPU */ 1121 a6xx_rpmh_stop(gmu); 1122 } 1123 1124 1125 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu) 1126 { 1127 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1128 struct msm_gpu *gpu = &a6xx_gpu->base.base; 1129 1130 if (!pm_runtime_active(gmu->dev)) 1131 return 0; 1132 1133 /* 1134 * Force the GMU off if we detected a hang, otherwise try to shut it 1135 * down gracefully 1136 */ 1137 if (gmu->hung) 1138 a6xx_gmu_force_off(gmu); 1139 else 1140 a6xx_gmu_shutdown(gmu); 1141 1142 /* Remove the bus vote */ 1143 dev_pm_opp_set_opp(&gpu->pdev->dev, NULL); 1144 1145 /* 1146 * Make sure the GX domain is off before turning off the GMU (CX) 1147 * domain. Usually the GMU does this but only if the shutdown sequence 1148 * was successful 1149 */ 1150 if (!IS_ERR_OR_NULL(gmu->gxpd)) 1151 pm_runtime_put_sync(gmu->gxpd); 1152 1153 clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks); 1154 1155 pm_runtime_put_sync(gmu->dev); 1156 1157 return 0; 1158 } 1159 1160 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu) 1161 { 1162 msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace); 1163 msm_gem_kernel_put(gmu->debug.obj, gmu->aspace); 1164 msm_gem_kernel_put(gmu->icache.obj, gmu->aspace); 1165 msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace); 1166 msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace); 1167 msm_gem_kernel_put(gmu->log.obj, gmu->aspace); 1168 1169 gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu); 1170 msm_gem_address_space_put(gmu->aspace); 1171 } 1172 1173 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo, 1174 size_t size, u64 iova, const char *name) 1175 { 1176 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1177 struct drm_device *dev = a6xx_gpu->base.base.dev; 1178 uint32_t flags = MSM_BO_WC; 1179 u64 range_start, range_end; 1180 int ret; 1181 1182 size = PAGE_ALIGN(size); 1183 if (!iova) { 1184 /* no fixed address - use GMU's uncached range */ 1185 range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */ 1186 range_end = 0x80000000; 1187 } else { 1188 /* range for fixed address */ 1189 range_start = iova; 1190 range_end = iova + size; 1191 /* use IOMMU_PRIV for icache/dcache */ 1192 flags |= MSM_BO_MAP_PRIV; 1193 } 1194 1195 bo->obj = msm_gem_new(dev, size, flags); 1196 if (IS_ERR(bo->obj)) 1197 return PTR_ERR(bo->obj); 1198 1199 ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova, 1200 range_start, range_end); 1201 if (ret) { 1202 drm_gem_object_put(bo->obj); 1203 return ret; 1204 } 1205 1206 bo->virt = msm_gem_get_vaddr(bo->obj); 1207 bo->size = size; 1208 1209 msm_gem_object_set_name(bo->obj, name); 1210 1211 return 0; 1212 } 1213 1214 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu) 1215 { 1216 struct msm_mmu *mmu; 1217 1218 mmu = msm_iommu_new(gmu->dev, 0); 1219 if (!mmu) 1220 return -ENODEV; 1221 if (IS_ERR(mmu)) 1222 return PTR_ERR(mmu); 1223 1224 gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000); 1225 if (IS_ERR(gmu->aspace)) 1226 return PTR_ERR(gmu->aspace); 1227 1228 return 0; 1229 } 1230 1231 /* Return the 'arc-level' for the given frequency */ 1232 static unsigned int a6xx_gmu_get_arc_level(struct device *dev, 1233 unsigned long freq) 1234 { 1235 struct dev_pm_opp *opp; 1236 unsigned int val; 1237 1238 if (!freq) 1239 return 0; 1240 1241 opp = dev_pm_opp_find_freq_exact(dev, freq, true); 1242 if (IS_ERR(opp)) 1243 return 0; 1244 1245 val = dev_pm_opp_get_level(opp); 1246 1247 dev_pm_opp_put(opp); 1248 1249 return val; 1250 } 1251 1252 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes, 1253 unsigned long *freqs, int freqs_count, const char *id) 1254 { 1255 int i, j; 1256 const u16 *pri, *sec; 1257 size_t pri_count, sec_count; 1258 1259 pri = cmd_db_read_aux_data(id, &pri_count); 1260 if (IS_ERR(pri)) 1261 return PTR_ERR(pri); 1262 /* 1263 * The data comes back as an array of unsigned shorts so adjust the 1264 * count accordingly 1265 */ 1266 pri_count >>= 1; 1267 if (!pri_count) 1268 return -EINVAL; 1269 1270 sec = cmd_db_read_aux_data("mx.lvl", &sec_count); 1271 if (IS_ERR(sec)) 1272 return PTR_ERR(sec); 1273 1274 sec_count >>= 1; 1275 if (!sec_count) 1276 return -EINVAL; 1277 1278 /* Construct a vote for each frequency */ 1279 for (i = 0; i < freqs_count; i++) { 1280 u8 pindex = 0, sindex = 0; 1281 unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]); 1282 1283 /* Get the primary index that matches the arc level */ 1284 for (j = 0; j < pri_count; j++) { 1285 if (pri[j] >= level) { 1286 pindex = j; 1287 break; 1288 } 1289 } 1290 1291 if (j == pri_count) { 1292 DRM_DEV_ERROR(dev, 1293 "Level %u not found in the RPMh list\n", 1294 level); 1295 DRM_DEV_ERROR(dev, "Available levels:\n"); 1296 for (j = 0; j < pri_count; j++) 1297 DRM_DEV_ERROR(dev, " %u\n", pri[j]); 1298 1299 return -EINVAL; 1300 } 1301 1302 /* 1303 * Look for a level in in the secondary list that matches. If 1304 * nothing fits, use the maximum non zero vote 1305 */ 1306 1307 for (j = 0; j < sec_count; j++) { 1308 if (sec[j] >= level) { 1309 sindex = j; 1310 break; 1311 } else if (sec[j]) { 1312 sindex = j; 1313 } 1314 } 1315 1316 /* Construct the vote */ 1317 votes[i] = ((pri[pindex] & 0xffff) << 16) | 1318 (sindex << 8) | pindex; 1319 } 1320 1321 return 0; 1322 } 1323 1324 /* 1325 * The GMU votes with the RPMh for itself and on behalf of the GPU but we need 1326 * to construct the list of votes on the CPU and send it over. Query the RPMh 1327 * voltage levels and build the votes 1328 */ 1329 1330 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu) 1331 { 1332 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1333 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1334 struct msm_gpu *gpu = &adreno_gpu->base; 1335 int ret; 1336 1337 /* Build the GX votes */ 1338 ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes, 1339 gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl"); 1340 1341 /* Build the CX votes */ 1342 ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes, 1343 gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl"); 1344 1345 return ret; 1346 } 1347 1348 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs, 1349 u32 size) 1350 { 1351 int count = dev_pm_opp_get_opp_count(dev); 1352 struct dev_pm_opp *opp; 1353 int i, index = 0; 1354 unsigned long freq = 1; 1355 1356 /* 1357 * The OPP table doesn't contain the "off" frequency level so we need to 1358 * add 1 to the table size to account for it 1359 */ 1360 1361 if (WARN(count + 1 > size, 1362 "The GMU frequency table is being truncated\n")) 1363 count = size - 1; 1364 1365 /* Set the "off" frequency */ 1366 freqs[index++] = 0; 1367 1368 for (i = 0; i < count; i++) { 1369 opp = dev_pm_opp_find_freq_ceil(dev, &freq); 1370 if (IS_ERR(opp)) 1371 break; 1372 1373 dev_pm_opp_put(opp); 1374 freqs[index++] = freq++; 1375 } 1376 1377 return index; 1378 } 1379 1380 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu) 1381 { 1382 struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu); 1383 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1384 struct msm_gpu *gpu = &adreno_gpu->base; 1385 1386 int ret = 0; 1387 1388 /* 1389 * The GMU handles its own frequency switching so build a list of 1390 * available frequencies to send during initialization 1391 */ 1392 ret = devm_pm_opp_of_add_table(gmu->dev); 1393 if (ret) { 1394 DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n"); 1395 return ret; 1396 } 1397 1398 gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev, 1399 gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs)); 1400 1401 /* 1402 * The GMU also handles GPU frequency switching so build a list 1403 * from the GPU OPP table 1404 */ 1405 gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev, 1406 gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs)); 1407 1408 gmu->current_perf_index = gmu->nr_gpu_freqs - 1; 1409 1410 /* Build the list of RPMh votes that we'll send to the GMU */ 1411 return a6xx_gmu_rpmh_votes_init(gmu); 1412 } 1413 1414 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu) 1415 { 1416 int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks); 1417 1418 if (ret < 1) 1419 return ret; 1420 1421 gmu->nr_clocks = ret; 1422 1423 gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks, 1424 gmu->nr_clocks, "gmu"); 1425 1426 gmu->hub_clk = msm_clk_bulk_get_clock(gmu->clocks, 1427 gmu->nr_clocks, "hub"); 1428 1429 return 0; 1430 } 1431 1432 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev, 1433 const char *name) 1434 { 1435 void __iomem *ret; 1436 struct resource *res = platform_get_resource_byname(pdev, 1437 IORESOURCE_MEM, name); 1438 1439 if (!res) { 1440 DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name); 1441 return ERR_PTR(-EINVAL); 1442 } 1443 1444 ret = ioremap(res->start, resource_size(res)); 1445 if (!ret) { 1446 DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name); 1447 return ERR_PTR(-EINVAL); 1448 } 1449 1450 return ret; 1451 } 1452 1453 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev, 1454 const char *name, irq_handler_t handler) 1455 { 1456 int irq, ret; 1457 1458 irq = platform_get_irq_byname(pdev, name); 1459 1460 ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu); 1461 if (ret) { 1462 DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n", 1463 name, ret); 1464 return ret; 1465 } 1466 1467 disable_irq(irq); 1468 1469 return irq; 1470 } 1471 1472 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu) 1473 { 1474 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1475 struct platform_device *pdev = to_platform_device(gmu->dev); 1476 1477 if (!gmu->initialized) 1478 return; 1479 1480 pm_runtime_force_suspend(gmu->dev); 1481 1482 if (!IS_ERR_OR_NULL(gmu->gxpd)) { 1483 pm_runtime_disable(gmu->gxpd); 1484 dev_pm_domain_detach(gmu->gxpd, false); 1485 } 1486 1487 iounmap(gmu->mmio); 1488 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc")) 1489 iounmap(gmu->rscc); 1490 gmu->mmio = NULL; 1491 gmu->rscc = NULL; 1492 1493 a6xx_gmu_memory_free(gmu); 1494 1495 free_irq(gmu->gmu_irq, gmu); 1496 free_irq(gmu->hfi_irq, gmu); 1497 1498 /* Drop reference taken in of_find_device_by_node */ 1499 put_device(gmu->dev); 1500 1501 gmu->initialized = false; 1502 } 1503 1504 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node) 1505 { 1506 struct adreno_gpu *adreno_gpu = &a6xx_gpu->base; 1507 struct a6xx_gmu *gmu = &a6xx_gpu->gmu; 1508 struct platform_device *pdev = of_find_device_by_node(node); 1509 int ret; 1510 1511 if (!pdev) 1512 return -ENODEV; 1513 1514 mutex_init(&gmu->lock); 1515 1516 gmu->dev = &pdev->dev; 1517 1518 of_dma_configure(gmu->dev, node, true); 1519 1520 /* Fow now, don't do anything fancy until we get our feet under us */ 1521 gmu->idle_level = GMU_IDLE_STATE_ACTIVE; 1522 1523 pm_runtime_enable(gmu->dev); 1524 1525 /* Get the list of clocks */ 1526 ret = a6xx_gmu_clocks_probe(gmu); 1527 if (ret) 1528 goto err_put_device; 1529 1530 ret = a6xx_gmu_memory_probe(gmu); 1531 if (ret) 1532 goto err_put_device; 1533 1534 1535 /* A660 now requires handling "prealloc requests" in GMU firmware 1536 * For now just hardcode allocations based on the known firmware. 1537 * note: there is no indication that these correspond to "dummy" or 1538 * "debug" regions, but this "guess" allows reusing these BOs which 1539 * are otherwise unused by a660. 1540 */ 1541 gmu->dummy.size = SZ_4K; 1542 if (adreno_is_a660_family(adreno_gpu)) { 1543 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_4K * 7, 1544 0x60400000, "debug"); 1545 if (ret) 1546 goto err_memory; 1547 1548 gmu->dummy.size = SZ_8K; 1549 } 1550 1551 /* Allocate memory for the GMU dummy page */ 1552 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, gmu->dummy.size, 1553 0x60000000, "dummy"); 1554 if (ret) 1555 goto err_memory; 1556 1557 /* Note that a650 family also includes a660 family: */ 1558 if (adreno_is_a650_family(adreno_gpu)) { 1559 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache, 1560 SZ_16M - SZ_16K, 0x04000, "icache"); 1561 if (ret) 1562 goto err_memory; 1563 /* 1564 * NOTE: when porting legacy ("pre-650-family") GPUs you may be tempted to add a condition 1565 * to allocate icache/dcache here, as per downstream code flow, but it may not actually be 1566 * necessary. If you omit this step and you don't get random pagefaults, you are likely 1567 * good to go without this! 1568 */ 1569 } else if (adreno_is_a640_family(adreno_gpu)) { 1570 ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache, 1571 SZ_256K - SZ_16K, 0x04000, "icache"); 1572 if (ret) 1573 goto err_memory; 1574 1575 ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache, 1576 SZ_256K - SZ_16K, 0x44000, "dcache"); 1577 if (ret) 1578 goto err_memory; 1579 } else if (adreno_is_a630(adreno_gpu) || adreno_is_a615_family(adreno_gpu)) { 1580 /* HFI v1, has sptprac */ 1581 gmu->legacy = true; 1582 1583 /* Allocate memory for the GMU debug region */ 1584 ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0, "debug"); 1585 if (ret) 1586 goto err_memory; 1587 } 1588 1589 /* Allocate memory for for the HFI queues */ 1590 ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0, "hfi"); 1591 if (ret) 1592 goto err_memory; 1593 1594 /* Allocate memory for the GMU log region */ 1595 ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0, "log"); 1596 if (ret) 1597 goto err_memory; 1598 1599 /* Map the GMU registers */ 1600 gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu"); 1601 if (IS_ERR(gmu->mmio)) { 1602 ret = PTR_ERR(gmu->mmio); 1603 goto err_memory; 1604 } 1605 1606 if (adreno_is_a650_family(adreno_gpu)) { 1607 gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc"); 1608 if (IS_ERR(gmu->rscc)) 1609 goto err_mmio; 1610 } else { 1611 gmu->rscc = gmu->mmio + 0x23000; 1612 } 1613 1614 /* Get the HFI and GMU interrupts */ 1615 gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq); 1616 gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq); 1617 1618 if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) 1619 goto err_mmio; 1620 1621 /* 1622 * Get a link to the GX power domain to reset the GPU in case of GMU 1623 * crash 1624 */ 1625 gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx"); 1626 1627 /* Get the power levels for the GMU and GPU */ 1628 a6xx_gmu_pwrlevels_probe(gmu); 1629 1630 /* Set up the HFI queues */ 1631 a6xx_hfi_init(gmu); 1632 1633 gmu->initialized = true; 1634 1635 return 0; 1636 1637 err_mmio: 1638 iounmap(gmu->mmio); 1639 if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc")) 1640 iounmap(gmu->rscc); 1641 free_irq(gmu->gmu_irq, gmu); 1642 free_irq(gmu->hfi_irq, gmu); 1643 1644 ret = -ENODEV; 1645 1646 err_memory: 1647 a6xx_gmu_memory_free(gmu); 1648 err_put_device: 1649 /* Drop reference taken in of_find_device_by_node */ 1650 put_device(gmu->dev); 1651 1652 return ret; 1653 } 1654