xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision f4356947)
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 	a6xx_rpmh_stop(gmu);
625 
626 err:
627 	if (!IS_ERR_OR_NULL(pdcptr))
628 		iounmap(pdcptr);
629 	if (!IS_ERR_OR_NULL(seqptr))
630 		iounmap(seqptr);
631 }
632 
633 /*
634  * The lowest 16 bits of this value are the number of XO clock cycles for main
635  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
636  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
637  */
638 
639 #define GMU_PWR_COL_HYST 0x000a1680
640 
641 /* Set up the idle state for the GMU */
642 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
643 {
644 	/* Disable GMU WB/RB buffer */
645 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
646 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
647 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
648 
649 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
650 
651 	switch (gmu->idle_level) {
652 	case GMU_IDLE_STATE_IFPC:
653 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
654 			GMU_PWR_COL_HYST);
655 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
656 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
657 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
658 		fallthrough;
659 	case GMU_IDLE_STATE_SPTP:
660 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
661 			GMU_PWR_COL_HYST);
662 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
663 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
664 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
665 	}
666 
667 	/* Enable RPMh GPU client */
668 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
669 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
670 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
671 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
672 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
673 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
674 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
675 }
676 
677 struct block_header {
678 	u32 addr;
679 	u32 size;
680 	u32 type;
681 	u32 value;
682 	u32 data[];
683 };
684 
685 /* this should be a general kernel helper */
686 static int in_range(u32 addr, u32 start, u32 size)
687 {
688 	return addr >= start && addr < start + size;
689 }
690 
691 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
692 {
693 	if (!in_range(blk->addr, bo->iova, bo->size))
694 		return false;
695 
696 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
697 	return true;
698 }
699 
700 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
701 {
702 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
703 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
704 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
705 	const struct block_header *blk;
706 	u32 reg_offset;
707 
708 	u32 itcm_base = 0x00000000;
709 	u32 dtcm_base = 0x00040000;
710 
711 	if (adreno_is_a650_family(adreno_gpu))
712 		dtcm_base = 0x10004000;
713 
714 	if (gmu->legacy) {
715 		/* Sanity check the size of the firmware that was loaded */
716 		if (fw_image->size > 0x8000) {
717 			DRM_DEV_ERROR(gmu->dev,
718 				"GMU firmware is bigger than the available region\n");
719 			return -EINVAL;
720 		}
721 
722 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
723 			       (u32*) fw_image->data, fw_image->size);
724 		return 0;
725 	}
726 
727 
728 	for (blk = (const struct block_header *) fw_image->data;
729 	     (const u8*) blk < fw_image->data + fw_image->size;
730 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
731 		if (blk->size == 0)
732 			continue;
733 
734 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
735 			reg_offset = (blk->addr - itcm_base) >> 2;
736 			gmu_write_bulk(gmu,
737 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
738 				blk->data, blk->size);
739 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
740 			reg_offset = (blk->addr - dtcm_base) >> 2;
741 			gmu_write_bulk(gmu,
742 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
743 				blk->data, blk->size);
744 		} else if (!fw_block_mem(&gmu->icache, blk) &&
745 			   !fw_block_mem(&gmu->dcache, blk) &&
746 			   !fw_block_mem(&gmu->dummy, blk)) {
747 			DRM_DEV_ERROR(gmu->dev,
748 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
749 				blk->addr, blk->size, blk->data[0]);
750 		}
751 	}
752 
753 	return 0;
754 }
755 
756 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
757 {
758 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
759 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
760 	int ret;
761 	u32 chipid;
762 
763 	if (adreno_is_a650_family(adreno_gpu)) {
764 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FALNEXT_INTF, 1);
765 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
766 	}
767 
768 	if (state == GMU_WARM_BOOT) {
769 		ret = a6xx_rpmh_start(gmu);
770 		if (ret)
771 			return ret;
772 	} else {
773 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
774 			"GMU firmware is not loaded\n"))
775 			return -ENOENT;
776 
777 		/* Turn on register retention */
778 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
779 
780 		ret = a6xx_rpmh_start(gmu);
781 		if (ret)
782 			return ret;
783 
784 		ret = a6xx_gmu_fw_load(gmu);
785 		if (ret)
786 			return ret;
787 	}
788 
789 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
790 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
791 
792 	/* Write the iova of the HFI table */
793 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
794 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
795 
796 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
797 		(1 << 31) | (0xa << 18) | (0xa0));
798 
799 	chipid = adreno_gpu->rev.core << 24;
800 	chipid |= adreno_gpu->rev.major << 16;
801 	chipid |= adreno_gpu->rev.minor << 12;
802 	chipid |= adreno_gpu->rev.patchid << 8;
803 
804 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
805 
806 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
807 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
808 
809 	/* Set up the lowest idle level on the GMU */
810 	a6xx_gmu_power_config(gmu);
811 
812 	ret = a6xx_gmu_start(gmu);
813 	if (ret)
814 		return ret;
815 
816 	if (gmu->legacy) {
817 		ret = a6xx_gmu_gfx_rail_on(gmu);
818 		if (ret)
819 			return ret;
820 	}
821 
822 	/* Enable SPTP_PC if the CPU is responsible for it */
823 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
824 		ret = a6xx_sptprac_enable(gmu);
825 		if (ret)
826 			return ret;
827 	}
828 
829 	ret = a6xx_gmu_hfi_start(gmu);
830 	if (ret)
831 		return ret;
832 
833 	/* FIXME: Do we need this wmb() here? */
834 	wmb();
835 
836 	return 0;
837 }
838 
839 #define A6XX_HFI_IRQ_MASK \
840 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
841 
842 #define A6XX_GMU_IRQ_MASK \
843 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
844 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
845 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
846 
847 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
848 {
849 	disable_irq(gmu->gmu_irq);
850 	disable_irq(gmu->hfi_irq);
851 
852 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
853 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
854 }
855 
856 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
857 {
858 	u32 val;
859 
860 	/* Make sure there are no outstanding RPMh votes */
861 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
862 		(val & 1), 100, 10000);
863 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
864 		(val & 1), 100, 10000);
865 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
866 		(val & 1), 100, 10000);
867 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
868 		(val & 1), 100, 1000);
869 }
870 
871 #define GBIF_CLIENT_HALT_MASK             BIT(0)
872 #define GBIF_ARB_HALT_MASK                BIT(1)
873 
874 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu,
875 		bool gx_off)
876 {
877 	struct msm_gpu *gpu = &adreno_gpu->base;
878 
879 	if (!a6xx_has_gbif(adreno_gpu)) {
880 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
881 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
882 								0xf) == 0xf);
883 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
884 
885 		return;
886 	}
887 
888 	if (gx_off) {
889 		/* Halt the gx side of GBIF */
890 		gpu_write(gpu, REG_A6XX_RBBM_GBIF_HALT, 1);
891 		spin_until(gpu_read(gpu, REG_A6XX_RBBM_GBIF_HALT_ACK) & 1);
892 	}
893 
894 	/* Halt new client requests on GBIF */
895 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
896 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
897 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
898 
899 	/* Halt all AXI requests on GBIF */
900 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
901 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
902 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
903 
904 	/* The GBIF halt needs to be explicitly cleared */
905 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
906 }
907 
908 /* Force the GMU off in case it isn't responsive */
909 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
910 {
911 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
912 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
913 	struct msm_gpu *gpu = &adreno_gpu->base;
914 
915 	/* Flush all the queues */
916 	a6xx_hfi_stop(gmu);
917 
918 	/* Stop the interrupts */
919 	a6xx_gmu_irq_disable(gmu);
920 
921 	/* Force off SPTP in case the GMU is managing it */
922 	a6xx_sptprac_disable(gmu);
923 
924 	/* Make sure there are no outstanding RPMh votes */
925 	a6xx_gmu_rpmh_off(gmu);
926 
927 	/* Halt the gmu cm3 core */
928 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
929 
930 	a6xx_bus_clear_pending_transactions(adreno_gpu, true);
931 
932 	/* Reset GPU core blocks */
933 	gpu_write(gpu, REG_A6XX_RBBM_SW_RESET_CMD, 1);
934 	udelay(100);
935 }
936 
937 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
938 {
939 	struct dev_pm_opp *gpu_opp;
940 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
941 
942 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
943 	if (IS_ERR(gpu_opp))
944 		return;
945 
946 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
947 	a6xx_gmu_set_freq(gpu, gpu_opp, false);
948 	dev_pm_opp_put(gpu_opp);
949 }
950 
951 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
952 {
953 	struct dev_pm_opp *gpu_opp;
954 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
955 
956 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
957 	if (IS_ERR(gpu_opp))
958 		return;
959 
960 	dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp);
961 	dev_pm_opp_put(gpu_opp);
962 }
963 
964 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
965 {
966 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
967 	struct msm_gpu *gpu = &adreno_gpu->base;
968 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
969 	int status, ret;
970 
971 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
972 		return -EINVAL;
973 
974 	gmu->hung = false;
975 
976 	/* Turn on the resources */
977 	pm_runtime_get_sync(gmu->dev);
978 
979 	/*
980 	 * "enable" the GX power domain which won't actually do anything but it
981 	 * will make sure that the refcounting is correct in case we need to
982 	 * bring down the GX after a GMU failure
983 	 */
984 	if (!IS_ERR_OR_NULL(gmu->gxpd))
985 		pm_runtime_get_sync(gmu->gxpd);
986 
987 	/* Use a known rate to bring up the GMU */
988 	clk_set_rate(gmu->core_clk, 200000000);
989 	clk_set_rate(gmu->hub_clk, 150000000);
990 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
991 	if (ret) {
992 		pm_runtime_put(gmu->gxpd);
993 		pm_runtime_put(gmu->dev);
994 		return ret;
995 	}
996 
997 	/* Set the bus quota to a reasonable value for boot */
998 	a6xx_gmu_set_initial_bw(gpu, gmu);
999 
1000 	/* Enable the GMU interrupt */
1001 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
1002 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
1003 	enable_irq(gmu->gmu_irq);
1004 
1005 	/* Check to see if we are doing a cold or warm boot */
1006 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
1007 		GMU_WARM_BOOT : GMU_COLD_BOOT;
1008 
1009 	/*
1010 	 * Warm boot path does not work on newer GPUs
1011 	 * Presumably this is because icache/dcache regions must be restored
1012 	 */
1013 	if (!gmu->legacy)
1014 		status = GMU_COLD_BOOT;
1015 
1016 	ret = a6xx_gmu_fw_start(gmu, status);
1017 	if (ret)
1018 		goto out;
1019 
1020 	ret = a6xx_hfi_start(gmu, status);
1021 	if (ret)
1022 		goto out;
1023 
1024 	/*
1025 	 * Turn on the GMU firmware fault interrupt after we know the boot
1026 	 * sequence is successful
1027 	 */
1028 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
1029 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
1030 	enable_irq(gmu->hfi_irq);
1031 
1032 	/* Set the GPU to the current freq */
1033 	a6xx_gmu_set_initial_freq(gpu, gmu);
1034 
1035 out:
1036 	/* On failure, shut down the GMU to leave it in a good state */
1037 	if (ret) {
1038 		disable_irq(gmu->gmu_irq);
1039 		a6xx_rpmh_stop(gmu);
1040 		pm_runtime_put(gmu->gxpd);
1041 		pm_runtime_put(gmu->dev);
1042 	}
1043 
1044 	return ret;
1045 }
1046 
1047 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
1048 {
1049 	u32 reg;
1050 
1051 	if (!gmu->initialized)
1052 		return true;
1053 
1054 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
1055 
1056 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
1057 		return false;
1058 
1059 	return true;
1060 }
1061 
1062 /* Gracefully try to shut down the GMU and by extension the GPU */
1063 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1064 {
1065 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1066 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1067 	u32 val;
1068 
1069 	/*
1070 	 * The GMU may still be in slumber unless the GPU started so check and
1071 	 * skip putting it back into slumber if so
1072 	 */
1073 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1074 
1075 	if (val != 0xf) {
1076 		int ret = a6xx_gmu_wait_for_idle(gmu);
1077 
1078 		/* If the GMU isn't responding assume it is hung */
1079 		if (ret) {
1080 			a6xx_gmu_force_off(gmu);
1081 			return;
1082 		}
1083 
1084 		a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung);
1085 
1086 		/* tell the GMU we want to slumber */
1087 		ret = a6xx_gmu_notify_slumber(gmu);
1088 		if (ret) {
1089 			a6xx_gmu_force_off(gmu);
1090 			return;
1091 		}
1092 
1093 		ret = gmu_poll_timeout(gmu,
1094 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1095 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1096 			100, 10000);
1097 
1098 		/*
1099 		 * Let the user know we failed to slumber but don't worry too
1100 		 * much because we are powering down anyway
1101 		 */
1102 
1103 		if (ret)
1104 			DRM_DEV_ERROR(gmu->dev,
1105 				"Unable to slumber GMU: status = 0%x/0%x\n",
1106 				gmu_read(gmu,
1107 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1108 				gmu_read(gmu,
1109 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1110 	}
1111 
1112 	/* Turn off HFI */
1113 	a6xx_hfi_stop(gmu);
1114 
1115 	/* Stop the interrupts and mask the hardware */
1116 	a6xx_gmu_irq_disable(gmu);
1117 
1118 	/* Tell RPMh to power off the GPU */
1119 	a6xx_rpmh_stop(gmu);
1120 }
1121 
1122 
1123 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1124 {
1125 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1126 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1127 
1128 	if (!pm_runtime_active(gmu->dev))
1129 		return 0;
1130 
1131 	/*
1132 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1133 	 * down gracefully
1134 	 */
1135 	if (gmu->hung)
1136 		a6xx_gmu_force_off(gmu);
1137 	else
1138 		a6xx_gmu_shutdown(gmu);
1139 
1140 	/* Remove the bus vote */
1141 	dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1142 
1143 	/*
1144 	 * Make sure the GX domain is off before turning off the GMU (CX)
1145 	 * domain. Usually the GMU does this but only if the shutdown sequence
1146 	 * was successful
1147 	 */
1148 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1149 		pm_runtime_put_sync(gmu->gxpd);
1150 
1151 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1152 
1153 	pm_runtime_put_sync(gmu->dev);
1154 
1155 	return 0;
1156 }
1157 
1158 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1159 {
1160 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace);
1161 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace);
1162 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace);
1163 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace);
1164 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace);
1165 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace);
1166 
1167 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1168 	msm_gem_address_space_put(gmu->aspace);
1169 }
1170 
1171 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1172 		size_t size, u64 iova, const char *name)
1173 {
1174 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1175 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1176 	uint32_t flags = MSM_BO_WC;
1177 	u64 range_start, range_end;
1178 	int ret;
1179 
1180 	size = PAGE_ALIGN(size);
1181 	if (!iova) {
1182 		/* no fixed address - use GMU's uncached range */
1183 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1184 		range_end = 0x80000000;
1185 	} else {
1186 		/* range for fixed address */
1187 		range_start = iova;
1188 		range_end = iova + size;
1189 		/* use IOMMU_PRIV for icache/dcache */
1190 		flags |= MSM_BO_MAP_PRIV;
1191 	}
1192 
1193 	bo->obj = msm_gem_new(dev, size, flags);
1194 	if (IS_ERR(bo->obj))
1195 		return PTR_ERR(bo->obj);
1196 
1197 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1198 					     range_start, range_end);
1199 	if (ret) {
1200 		drm_gem_object_put(bo->obj);
1201 		return ret;
1202 	}
1203 
1204 	bo->virt = msm_gem_get_vaddr(bo->obj);
1205 	bo->size = size;
1206 
1207 	msm_gem_object_set_name(bo->obj, name);
1208 
1209 	return 0;
1210 }
1211 
1212 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1213 {
1214 	struct msm_mmu *mmu;
1215 
1216 	mmu = msm_iommu_new(gmu->dev, 0);
1217 	if (!mmu)
1218 		return -ENODEV;
1219 	if (IS_ERR(mmu))
1220 		return PTR_ERR(mmu);
1221 
1222 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1223 	if (IS_ERR(gmu->aspace))
1224 		return PTR_ERR(gmu->aspace);
1225 
1226 	return 0;
1227 }
1228 
1229 /* Return the 'arc-level' for the given frequency */
1230 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1231 					   unsigned long freq)
1232 {
1233 	struct dev_pm_opp *opp;
1234 	unsigned int val;
1235 
1236 	if (!freq)
1237 		return 0;
1238 
1239 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1240 	if (IS_ERR(opp))
1241 		return 0;
1242 
1243 	val = dev_pm_opp_get_level(opp);
1244 
1245 	dev_pm_opp_put(opp);
1246 
1247 	return val;
1248 }
1249 
1250 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1251 		unsigned long *freqs, int freqs_count, const char *id)
1252 {
1253 	int i, j;
1254 	const u16 *pri, *sec;
1255 	size_t pri_count, sec_count;
1256 
1257 	pri = cmd_db_read_aux_data(id, &pri_count);
1258 	if (IS_ERR(pri))
1259 		return PTR_ERR(pri);
1260 	/*
1261 	 * The data comes back as an array of unsigned shorts so adjust the
1262 	 * count accordingly
1263 	 */
1264 	pri_count >>= 1;
1265 	if (!pri_count)
1266 		return -EINVAL;
1267 
1268 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1269 	if (IS_ERR(sec))
1270 		return PTR_ERR(sec);
1271 
1272 	sec_count >>= 1;
1273 	if (!sec_count)
1274 		return -EINVAL;
1275 
1276 	/* Construct a vote for each frequency */
1277 	for (i = 0; i < freqs_count; i++) {
1278 		u8 pindex = 0, sindex = 0;
1279 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1280 
1281 		/* Get the primary index that matches the arc level */
1282 		for (j = 0; j < pri_count; j++) {
1283 			if (pri[j] >= level) {
1284 				pindex = j;
1285 				break;
1286 			}
1287 		}
1288 
1289 		if (j == pri_count) {
1290 			DRM_DEV_ERROR(dev,
1291 				      "Level %u not found in the RPMh list\n",
1292 				      level);
1293 			DRM_DEV_ERROR(dev, "Available levels:\n");
1294 			for (j = 0; j < pri_count; j++)
1295 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1296 
1297 			return -EINVAL;
1298 		}
1299 
1300 		/*
1301 		 * Look for a level in in the secondary list that matches. If
1302 		 * nothing fits, use the maximum non zero vote
1303 		 */
1304 
1305 		for (j = 0; j < sec_count; j++) {
1306 			if (sec[j] >= level) {
1307 				sindex = j;
1308 				break;
1309 			} else if (sec[j]) {
1310 				sindex = j;
1311 			}
1312 		}
1313 
1314 		/* Construct the vote */
1315 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1316 			(sindex << 8) | pindex;
1317 	}
1318 
1319 	return 0;
1320 }
1321 
1322 /*
1323  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1324  * to construct the list of votes on the CPU and send it over. Query the RPMh
1325  * voltage levels and build the votes
1326  */
1327 
1328 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1329 {
1330 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1331 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1332 	struct msm_gpu *gpu = &adreno_gpu->base;
1333 	int ret;
1334 
1335 	/* Build the GX votes */
1336 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1337 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1338 
1339 	/* Build the CX votes */
1340 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1341 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1342 
1343 	return ret;
1344 }
1345 
1346 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1347 		u32 size)
1348 {
1349 	int count = dev_pm_opp_get_opp_count(dev);
1350 	struct dev_pm_opp *opp;
1351 	int i, index = 0;
1352 	unsigned long freq = 1;
1353 
1354 	/*
1355 	 * The OPP table doesn't contain the "off" frequency level so we need to
1356 	 * add 1 to the table size to account for it
1357 	 */
1358 
1359 	if (WARN(count + 1 > size,
1360 		"The GMU frequency table is being truncated\n"))
1361 		count = size - 1;
1362 
1363 	/* Set the "off" frequency */
1364 	freqs[index++] = 0;
1365 
1366 	for (i = 0; i < count; i++) {
1367 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1368 		if (IS_ERR(opp))
1369 			break;
1370 
1371 		dev_pm_opp_put(opp);
1372 		freqs[index++] = freq++;
1373 	}
1374 
1375 	return index;
1376 }
1377 
1378 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1379 {
1380 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1381 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1382 	struct msm_gpu *gpu = &adreno_gpu->base;
1383 
1384 	int ret = 0;
1385 
1386 	/*
1387 	 * The GMU handles its own frequency switching so build a list of
1388 	 * available frequencies to send during initialization
1389 	 */
1390 	ret = devm_pm_opp_of_add_table(gmu->dev);
1391 	if (ret) {
1392 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1393 		return ret;
1394 	}
1395 
1396 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1397 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1398 
1399 	/*
1400 	 * The GMU also handles GPU frequency switching so build a list
1401 	 * from the GPU OPP table
1402 	 */
1403 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1404 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1405 
1406 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1407 
1408 	/* Build the list of RPMh votes that we'll send to the GMU */
1409 	return a6xx_gmu_rpmh_votes_init(gmu);
1410 }
1411 
1412 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1413 {
1414 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1415 
1416 	if (ret < 1)
1417 		return ret;
1418 
1419 	gmu->nr_clocks = ret;
1420 
1421 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1422 		gmu->nr_clocks, "gmu");
1423 
1424 	gmu->hub_clk = msm_clk_bulk_get_clock(gmu->clocks,
1425 		gmu->nr_clocks, "hub");
1426 
1427 	return 0;
1428 }
1429 
1430 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1431 		const char *name)
1432 {
1433 	void __iomem *ret;
1434 	struct resource *res = platform_get_resource_byname(pdev,
1435 			IORESOURCE_MEM, name);
1436 
1437 	if (!res) {
1438 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1439 		return ERR_PTR(-EINVAL);
1440 	}
1441 
1442 	ret = ioremap(res->start, resource_size(res));
1443 	if (!ret) {
1444 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1445 		return ERR_PTR(-EINVAL);
1446 	}
1447 
1448 	return ret;
1449 }
1450 
1451 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1452 		const char *name, irq_handler_t handler)
1453 {
1454 	int irq, ret;
1455 
1456 	irq = platform_get_irq_byname(pdev, name);
1457 
1458 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1459 	if (ret) {
1460 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1461 			      name, ret);
1462 		return ret;
1463 	}
1464 
1465 	disable_irq(irq);
1466 
1467 	return irq;
1468 }
1469 
1470 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1471 {
1472 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1473 	struct platform_device *pdev = to_platform_device(gmu->dev);
1474 
1475 	if (!gmu->initialized)
1476 		return;
1477 
1478 	pm_runtime_force_suspend(gmu->dev);
1479 
1480 	/*
1481 	 * Since cxpd is a virt device, the devlink with gmu-dev will be removed
1482 	 * automatically when we do detach
1483 	 */
1484 	dev_pm_domain_detach(gmu->cxpd, false);
1485 
1486 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1487 		pm_runtime_disable(gmu->gxpd);
1488 		dev_pm_domain_detach(gmu->gxpd, false);
1489 	}
1490 
1491 	iounmap(gmu->mmio);
1492 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1493 		iounmap(gmu->rscc);
1494 	gmu->mmio = NULL;
1495 	gmu->rscc = NULL;
1496 
1497 	a6xx_gmu_memory_free(gmu);
1498 
1499 	free_irq(gmu->gmu_irq, gmu);
1500 	free_irq(gmu->hfi_irq, gmu);
1501 
1502 	/* Drop reference taken in of_find_device_by_node */
1503 	put_device(gmu->dev);
1504 
1505 	gmu->initialized = false;
1506 }
1507 
1508 static int cxpd_notifier_cb(struct notifier_block *nb,
1509 			unsigned long action, void *data)
1510 {
1511 	struct a6xx_gmu *gmu = container_of(nb, struct a6xx_gmu, pd_nb);
1512 
1513 	if (action == GENPD_NOTIFY_OFF)
1514 		complete_all(&gmu->pd_gate);
1515 
1516 	return 0;
1517 }
1518 
1519 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1520 {
1521 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1522 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1523 	struct platform_device *pdev = of_find_device_by_node(node);
1524 	int ret;
1525 
1526 	if (!pdev)
1527 		return -ENODEV;
1528 
1529 	mutex_init(&gmu->lock);
1530 
1531 	gmu->dev = &pdev->dev;
1532 
1533 	of_dma_configure(gmu->dev, node, true);
1534 
1535 	/* Fow now, don't do anything fancy until we get our feet under us */
1536 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1537 
1538 	pm_runtime_enable(gmu->dev);
1539 
1540 	/* Get the list of clocks */
1541 	ret = a6xx_gmu_clocks_probe(gmu);
1542 	if (ret)
1543 		goto err_put_device;
1544 
1545 	ret = a6xx_gmu_memory_probe(gmu);
1546 	if (ret)
1547 		goto err_put_device;
1548 
1549 
1550 	/* A660 now requires handling "prealloc requests" in GMU firmware
1551 	 * For now just hardcode allocations based on the known firmware.
1552 	 * note: there is no indication that these correspond to "dummy" or
1553 	 * "debug" regions, but this "guess" allows reusing these BOs which
1554 	 * are otherwise unused by a660.
1555 	 */
1556 	gmu->dummy.size = SZ_4K;
1557 	if (adreno_is_a660_family(adreno_gpu)) {
1558 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_4K * 7,
1559 					    0x60400000, "debug");
1560 		if (ret)
1561 			goto err_memory;
1562 
1563 		gmu->dummy.size = SZ_8K;
1564 	}
1565 
1566 	/* Allocate memory for the GMU dummy page */
1567 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, gmu->dummy.size,
1568 				    0x60000000, "dummy");
1569 	if (ret)
1570 		goto err_memory;
1571 
1572 	/* Note that a650 family also includes a660 family: */
1573 	if (adreno_is_a650_family(adreno_gpu)) {
1574 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1575 			SZ_16M - SZ_16K, 0x04000, "icache");
1576 		if (ret)
1577 			goto err_memory;
1578 	/*
1579 	 * NOTE: when porting legacy ("pre-650-family") GPUs you may be tempted to add a condition
1580 	 * to allocate icache/dcache here, as per downstream code flow, but it may not actually be
1581 	 * necessary. If you omit this step and you don't get random pagefaults, you are likely
1582 	 * good to go without this!
1583 	 */
1584 	} else if (adreno_is_a640_family(adreno_gpu)) {
1585 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1586 			SZ_256K - SZ_16K, 0x04000, "icache");
1587 		if (ret)
1588 			goto err_memory;
1589 
1590 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1591 			SZ_256K - SZ_16K, 0x44000, "dcache");
1592 		if (ret)
1593 			goto err_memory;
1594 	} else if (adreno_is_a630(adreno_gpu) || adreno_is_a615_family(adreno_gpu)) {
1595 		/* HFI v1, has sptprac */
1596 		gmu->legacy = true;
1597 
1598 		/* Allocate memory for the GMU debug region */
1599 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0, "debug");
1600 		if (ret)
1601 			goto err_memory;
1602 	}
1603 
1604 	/* Allocate memory for for the HFI queues */
1605 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0, "hfi");
1606 	if (ret)
1607 		goto err_memory;
1608 
1609 	/* Allocate memory for the GMU log region */
1610 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0, "log");
1611 	if (ret)
1612 		goto err_memory;
1613 
1614 	/* Map the GMU registers */
1615 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1616 	if (IS_ERR(gmu->mmio)) {
1617 		ret = PTR_ERR(gmu->mmio);
1618 		goto err_memory;
1619 	}
1620 
1621 	if (adreno_is_a650_family(adreno_gpu)) {
1622 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1623 		if (IS_ERR(gmu->rscc)) {
1624 			ret = -ENODEV;
1625 			goto err_mmio;
1626 		}
1627 	} else {
1628 		gmu->rscc = gmu->mmio + 0x23000;
1629 	}
1630 
1631 	/* Get the HFI and GMU interrupts */
1632 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1633 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1634 
1635 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) {
1636 		ret = -ENODEV;
1637 		goto err_mmio;
1638 	}
1639 
1640 	gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
1641 	if (IS_ERR(gmu->cxpd)) {
1642 		ret = PTR_ERR(gmu->cxpd);
1643 		goto err_mmio;
1644 	}
1645 
1646 	if (!device_link_add(gmu->dev, gmu->cxpd,
1647 					DL_FLAG_PM_RUNTIME)) {
1648 		ret = -ENODEV;
1649 		goto detach_cxpd;
1650 	}
1651 
1652 	init_completion(&gmu->pd_gate);
1653 	complete_all(&gmu->pd_gate);
1654 	gmu->pd_nb.notifier_call = cxpd_notifier_cb;
1655 
1656 	/*
1657 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1658 	 * crash
1659 	 */
1660 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1661 
1662 	/* Get the power levels for the GMU and GPU */
1663 	a6xx_gmu_pwrlevels_probe(gmu);
1664 
1665 	/* Set up the HFI queues */
1666 	a6xx_hfi_init(gmu);
1667 
1668 	/* Initialize RPMh */
1669 	a6xx_gmu_rpmh_init(gmu);
1670 
1671 	gmu->initialized = true;
1672 
1673 	return 0;
1674 
1675 detach_cxpd:
1676 	dev_pm_domain_detach(gmu->cxpd, false);
1677 
1678 err_mmio:
1679 	iounmap(gmu->mmio);
1680 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1681 		iounmap(gmu->rscc);
1682 	free_irq(gmu->gmu_irq, gmu);
1683 	free_irq(gmu->hfi_irq, gmu);
1684 
1685 err_memory:
1686 	a6xx_gmu_memory_free(gmu);
1687 err_put_device:
1688 	/* Drop reference taken in of_find_device_by_node */
1689 	put_device(gmu->dev);
1690 
1691 	return ret;
1692 }
1693