xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision e5242c5f)
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/of_platform.h>
7 #include <linux/platform_device.h>
8 #include <linux/pm_domain.h>
9 #include <linux/pm_opp.h>
10 #include <soc/qcom/cmd-db.h>
11 #include <drm/drm_gem.h>
12 
13 #include "a6xx_gpu.h"
14 #include "a6xx_gmu.xml.h"
15 #include "msm_gem.h"
16 #include "msm_gpu_trace.h"
17 #include "msm_mmu.h"
18 
19 static void a6xx_gmu_fault(struct a6xx_gmu *gmu)
20 {
21 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
22 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
23 	struct msm_gpu *gpu = &adreno_gpu->base;
24 
25 	/* FIXME: add a banner here */
26 	gmu->hung = true;
27 
28 	/* Turn off the hangcheck timer while we are resetting */
29 	del_timer(&gpu->hangcheck_timer);
30 
31 	/* Queue the GPU handler because we need to treat this as a recovery */
32 	kthread_queue_work(gpu->worker, &gpu->recover_work);
33 }
34 
35 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
36 {
37 	struct a6xx_gmu *gmu = data;
38 	u32 status;
39 
40 	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
41 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
42 
43 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
44 		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
45 
46 		a6xx_gmu_fault(gmu);
47 	}
48 
49 	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
50 		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
51 
52 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
53 		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
54 			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
55 
56 	return IRQ_HANDLED;
57 }
58 
59 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
60 {
61 	struct a6xx_gmu *gmu = data;
62 	u32 status;
63 
64 	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
65 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
66 
67 	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
68 		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
69 
70 		a6xx_gmu_fault(gmu);
71 	}
72 
73 	return IRQ_HANDLED;
74 }
75 
76 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
77 {
78 	u32 val;
79 
80 	/* This can be called from gpu state code so make sure GMU is valid */
81 	if (!gmu->initialized)
82 		return false;
83 
84 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
85 
86 	return !(val &
87 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
88 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
89 }
90 
91 /* Check to see if the GX rail is still powered */
92 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
93 {
94 	u32 val;
95 
96 	/* This can be called from gpu state code so make sure GMU is valid */
97 	if (!gmu->initialized)
98 		return false;
99 
100 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
101 
102 	return !(val &
103 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
104 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
105 }
106 
107 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp,
108 		       bool suspended)
109 {
110 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
111 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
112 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
113 	u32 perf_index;
114 	unsigned long gpu_freq;
115 	int ret = 0;
116 
117 	gpu_freq = dev_pm_opp_get_freq(opp);
118 
119 	if (gpu_freq == gmu->freq)
120 		return;
121 
122 	for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
123 		if (gpu_freq == gmu->gpu_freqs[perf_index])
124 			break;
125 
126 	gmu->current_perf_index = perf_index;
127 	gmu->freq = gmu->gpu_freqs[perf_index];
128 
129 	trace_msm_gmu_freq_change(gmu->freq, perf_index);
130 
131 	/*
132 	 * This can get called from devfreq while the hardware is idle. Don't
133 	 * bring up the power if it isn't already active. All we're doing here
134 	 * is updating the frequency so that when we come back online we're at
135 	 * the right rate.
136 	 */
137 	if (suspended)
138 		return;
139 
140 	if (!gmu->legacy) {
141 		a6xx_hfi_set_freq(gmu, perf_index);
142 		dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
143 		return;
144 	}
145 
146 	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
147 
148 	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
149 			((3 & 0xf) << 28) | perf_index);
150 
151 	/*
152 	 * Send an invalid index as a vote for the bus bandwidth and let the
153 	 * firmware decide on the right vote
154 	 */
155 	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
156 
157 	/* Set and clear the OOB for DCVS to trigger the GMU */
158 	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
159 	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
160 
161 	ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
162 	if (ret)
163 		dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
164 
165 	dev_pm_opp_set_opp(&gpu->pdev->dev, opp);
166 }
167 
168 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
169 {
170 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
171 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
172 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
173 
174 	return  gmu->freq;
175 }
176 
177 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
178 {
179 	u32 val;
180 	int local = gmu->idle_level;
181 
182 	/* SPTP and IFPC both report as IFPC */
183 	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
184 		local = GMU_IDLE_STATE_IFPC;
185 
186 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
187 
188 	if (val == local) {
189 		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
190 			!a6xx_gmu_gx_is_on(gmu))
191 			return true;
192 	}
193 
194 	return false;
195 }
196 
197 /* Wait for the GMU to get to its most idle state */
198 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
199 {
200 	return spin_until(a6xx_gmu_check_idle_level(gmu));
201 }
202 
203 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
204 {
205 	int ret;
206 	u32 val;
207 	u32 mask, reset_val;
208 
209 	val = gmu_read(gmu, REG_A6XX_GMU_CM3_DTCM_START + 0xff8);
210 	if (val <= 0x20010004) {
211 		mask = 0xffffffff;
212 		reset_val = 0xbabeface;
213 	} else {
214 		mask = 0x1ff;
215 		reset_val = 0x100;
216 	}
217 
218 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
219 
220 	/* Set the log wptr index
221 	 * note: downstream saves the value in poweroff and restores it here
222 	 */
223 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_RESP, 0);
224 
225 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
226 
227 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
228 		(val & mask) == reset_val, 100, 10000);
229 
230 	if (ret)
231 		DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
232 
233 	return ret;
234 }
235 
236 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
237 {
238 	u32 val;
239 	int ret;
240 
241 	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
242 
243 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
244 		val & 1, 100, 10000);
245 	if (ret)
246 		DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
247 
248 	return ret;
249 }
250 
251 struct a6xx_gmu_oob_bits {
252 	int set, ack, set_new, ack_new, clear, clear_new;
253 	const char *name;
254 };
255 
256 /* These are the interrupt / ack bits for each OOB request that are set
257  * in a6xx_gmu_set_oob and a6xx_clear_oob
258  */
259 static const struct a6xx_gmu_oob_bits a6xx_gmu_oob_bits[] = {
260 	[GMU_OOB_GPU_SET] = {
261 		.name = "GPU_SET",
262 		.set = 16,
263 		.ack = 24,
264 		.set_new = 30,
265 		.ack_new = 31,
266 		.clear = 24,
267 		.clear_new = 31,
268 	},
269 
270 	[GMU_OOB_PERFCOUNTER_SET] = {
271 		.name = "PERFCOUNTER",
272 		.set = 17,
273 		.ack = 25,
274 		.set_new = 28,
275 		.ack_new = 30,
276 		.clear = 25,
277 		.clear_new = 29,
278 	},
279 
280 	[GMU_OOB_BOOT_SLUMBER] = {
281 		.name = "BOOT_SLUMBER",
282 		.set = 22,
283 		.ack = 30,
284 		.clear = 30,
285 	},
286 
287 	[GMU_OOB_DCVS_SET] = {
288 		.name = "GPU_DCVS",
289 		.set = 23,
290 		.ack = 31,
291 		.clear = 31,
292 	},
293 };
294 
295 /* Trigger a OOB (out of band) request to the GMU */
296 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
297 {
298 	int ret;
299 	u32 val;
300 	int request, ack;
301 
302 	WARN_ON_ONCE(!mutex_is_locked(&gmu->lock));
303 
304 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
305 		return -EINVAL;
306 
307 	if (gmu->legacy) {
308 		request = a6xx_gmu_oob_bits[state].set;
309 		ack = a6xx_gmu_oob_bits[state].ack;
310 	} else {
311 		request = a6xx_gmu_oob_bits[state].set_new;
312 		ack = a6xx_gmu_oob_bits[state].ack_new;
313 		if (!request || !ack) {
314 			DRM_DEV_ERROR(gmu->dev,
315 				      "Invalid non-legacy GMU request %s\n",
316 				      a6xx_gmu_oob_bits[state].name);
317 			return -EINVAL;
318 		}
319 	}
320 
321 	/* Trigger the equested OOB operation */
322 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
323 
324 	/* Wait for the acknowledge interrupt */
325 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
326 		val & (1 << ack), 100, 10000);
327 
328 	if (ret)
329 		DRM_DEV_ERROR(gmu->dev,
330 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
331 				a6xx_gmu_oob_bits[state].name,
332 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
333 
334 	/* Clear the acknowledge interrupt */
335 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
336 
337 	return ret;
338 }
339 
340 /* Clear a pending OOB state in the GMU */
341 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
342 {
343 	int bit;
344 
345 	WARN_ON_ONCE(!mutex_is_locked(&gmu->lock));
346 
347 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
348 		return;
349 
350 	if (gmu->legacy)
351 		bit = a6xx_gmu_oob_bits[state].clear;
352 	else
353 		bit = a6xx_gmu_oob_bits[state].clear_new;
354 
355 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << bit);
356 }
357 
358 /* Enable CPU control of SPTP power power collapse */
359 int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
360 {
361 	int ret;
362 	u32 val;
363 
364 	if (!gmu->legacy)
365 		return 0;
366 
367 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
368 
369 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
370 		(val & 0x38) == 0x28, 1, 100);
371 
372 	if (ret) {
373 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
374 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
375 	}
376 
377 	return 0;
378 }
379 
380 /* Disable CPU control of SPTP power power collapse */
381 void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
382 {
383 	u32 val;
384 	int ret;
385 
386 	if (!gmu->legacy)
387 		return;
388 
389 	/* Make sure retention is on */
390 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
391 
392 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
393 
394 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
395 		(val & 0x04), 100, 10000);
396 
397 	if (ret)
398 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
399 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
400 }
401 
402 /* Let the GMU know we are starting a boot sequence */
403 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
404 {
405 	u32 vote;
406 
407 	/* Let the GMU know we are getting ready for boot */
408 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
409 
410 	/* Choose the "default" power level as the highest available */
411 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
412 
413 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
414 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
415 
416 	/* Let the GMU know the boot sequence has started */
417 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
418 }
419 
420 /* Let the GMU know that we are about to go into slumber */
421 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
422 {
423 	int ret;
424 
425 	/* Disable the power counter so the GMU isn't busy */
426 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
427 
428 	/* Disable SPTP_PC if the CPU is responsible for it */
429 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
430 		a6xx_sptprac_disable(gmu);
431 
432 	if (!gmu->legacy) {
433 		ret = a6xx_hfi_send_prep_slumber(gmu);
434 		goto out;
435 	}
436 
437 	/* Tell the GMU to get ready to slumber */
438 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
439 
440 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
441 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
442 
443 	if (!ret) {
444 		/* Check to see if the GMU really did slumber */
445 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
446 			!= 0x0f) {
447 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
448 			ret = -ETIMEDOUT;
449 		}
450 	}
451 
452 out:
453 	/* Put fence into allow mode */
454 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
455 	return ret;
456 }
457 
458 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
459 {
460 	int ret;
461 	u32 val;
462 
463 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
464 	/* Wait for the register to finish posting */
465 	wmb();
466 
467 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
468 		val & (1 << 1), 100, 10000);
469 	if (ret) {
470 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
471 		return ret;
472 	}
473 
474 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
475 		!val, 100, 10000);
476 
477 	if (ret) {
478 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
479 		return ret;
480 	}
481 
482 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
483 
484 	return 0;
485 }
486 
487 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
488 {
489 	int ret;
490 	u32 val;
491 
492 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
493 
494 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
495 		val, val & (1 << 16), 100, 10000);
496 	if (ret)
497 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
498 
499 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
500 }
501 
502 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
503 {
504 	msm_writel(value, ptr + (offset << 2));
505 }
506 
507 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
508 		const char *name);
509 
510 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
511 {
512 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
513 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
514 	struct platform_device *pdev = to_platform_device(gmu->dev);
515 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
516 	void __iomem *seqptr = NULL;
517 	uint32_t pdc_address_offset;
518 	bool pdc_in_aop = false;
519 
520 	if (IS_ERR(pdcptr))
521 		goto err;
522 
523 	if (adreno_is_a650(adreno_gpu) || adreno_is_a660_family(adreno_gpu))
524 		pdc_in_aop = true;
525 	else if (adreno_is_a618(adreno_gpu) || adreno_is_a640_family(adreno_gpu))
526 		pdc_address_offset = 0x30090;
527 	else if (adreno_is_a619(adreno_gpu))
528 		pdc_address_offset = 0x300a0;
529 	else
530 		pdc_address_offset = 0x30080;
531 
532 	if (!pdc_in_aop) {
533 		seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
534 		if (IS_ERR(seqptr))
535 			goto err;
536 	}
537 
538 	/* Disable SDE clock gating */
539 	gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
540 
541 	/* Setup RSC PDC handshake for sleep and wakeup */
542 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
543 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
544 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
545 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
546 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
547 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
548 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
549 	gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
550 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
551 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
552 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
553 
554 	/* Load RSC sequencer uCode for sleep and wakeup */
555 	if (adreno_is_a650_family(adreno_gpu)) {
556 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
557 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
558 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
559 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
560 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
561 	} else {
562 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
563 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
564 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
565 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
566 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
567 	}
568 
569 	if (pdc_in_aop)
570 		goto setup_pdc;
571 
572 	/* Load PDC sequencer uCode for power up and power down sequence */
573 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
574 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
575 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
576 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
577 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
578 
579 	/* Set TCS commands used by PDC sequence for low power modes */
580 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
581 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
582 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
583 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
584 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
585 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
586 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
587 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
588 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
589 
590 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
591 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
592 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
593 
594 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
595 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
596 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
597 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
598 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
599 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
600 
601 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
602 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
603 	if (adreno_is_a618(adreno_gpu) || adreno_is_a619(adreno_gpu) ||
604 			adreno_is_a650_family(adreno_gpu))
605 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
606 	else
607 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
608 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
609 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
610 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
611 
612 	/* Setup GPU PDC */
613 setup_pdc:
614 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
615 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
616 
617 	/* ensure no writes happen before the uCode is fully written */
618 	wmb();
619 
620 	a6xx_rpmh_stop(gmu);
621 
622 err:
623 	if (!IS_ERR_OR_NULL(pdcptr))
624 		iounmap(pdcptr);
625 	if (!IS_ERR_OR_NULL(seqptr))
626 		iounmap(seqptr);
627 }
628 
629 /*
630  * The lowest 16 bits of this value are the number of XO clock cycles for main
631  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
632  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
633  */
634 
635 #define GMU_PWR_COL_HYST 0x000a1680
636 
637 /* Set up the idle state for the GMU */
638 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
639 {
640 	/* Disable GMU WB/RB buffer */
641 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
642 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
643 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
644 
645 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
646 
647 	switch (gmu->idle_level) {
648 	case GMU_IDLE_STATE_IFPC:
649 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
650 			GMU_PWR_COL_HYST);
651 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
652 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
653 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
654 		fallthrough;
655 	case GMU_IDLE_STATE_SPTP:
656 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
657 			GMU_PWR_COL_HYST);
658 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
659 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
660 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
661 	}
662 
663 	/* Enable RPMh GPU client */
664 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
665 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
666 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
667 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
668 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
669 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
670 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
671 }
672 
673 struct block_header {
674 	u32 addr;
675 	u32 size;
676 	u32 type;
677 	u32 value;
678 	u32 data[];
679 };
680 
681 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
682 {
683 	if (!in_range(blk->addr, bo->iova, bo->size))
684 		return false;
685 
686 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
687 	return true;
688 }
689 
690 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
691 {
692 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
693 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
694 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
695 	const struct block_header *blk;
696 	u32 reg_offset;
697 
698 	u32 itcm_base = 0x00000000;
699 	u32 dtcm_base = 0x00040000;
700 
701 	if (adreno_is_a650_family(adreno_gpu))
702 		dtcm_base = 0x10004000;
703 
704 	if (gmu->legacy) {
705 		/* Sanity check the size of the firmware that was loaded */
706 		if (fw_image->size > 0x8000) {
707 			DRM_DEV_ERROR(gmu->dev,
708 				"GMU firmware is bigger than the available region\n");
709 			return -EINVAL;
710 		}
711 
712 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
713 			       (u32*) fw_image->data, fw_image->size);
714 		return 0;
715 	}
716 
717 
718 	for (blk = (const struct block_header *) fw_image->data;
719 	     (const u8*) blk < fw_image->data + fw_image->size;
720 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
721 		if (blk->size == 0)
722 			continue;
723 
724 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
725 			reg_offset = (blk->addr - itcm_base) >> 2;
726 			gmu_write_bulk(gmu,
727 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
728 				blk->data, blk->size);
729 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
730 			reg_offset = (blk->addr - dtcm_base) >> 2;
731 			gmu_write_bulk(gmu,
732 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
733 				blk->data, blk->size);
734 		} else if (!fw_block_mem(&gmu->icache, blk) &&
735 			   !fw_block_mem(&gmu->dcache, blk) &&
736 			   !fw_block_mem(&gmu->dummy, blk)) {
737 			DRM_DEV_ERROR(gmu->dev,
738 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
739 				blk->addr, blk->size, blk->data[0]);
740 		}
741 	}
742 
743 	return 0;
744 }
745 
746 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
747 {
748 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
749 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
750 	int ret;
751 	u32 chipid;
752 
753 	if (adreno_is_a650_family(adreno_gpu)) {
754 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FALNEXT_INTF, 1);
755 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
756 	}
757 
758 	if (state == GMU_WARM_BOOT) {
759 		ret = a6xx_rpmh_start(gmu);
760 		if (ret)
761 			return ret;
762 	} else {
763 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
764 			"GMU firmware is not loaded\n"))
765 			return -ENOENT;
766 
767 		/* Turn on register retention */
768 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
769 
770 		ret = a6xx_rpmh_start(gmu);
771 		if (ret)
772 			return ret;
773 
774 		ret = a6xx_gmu_fw_load(gmu);
775 		if (ret)
776 			return ret;
777 	}
778 
779 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
780 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
781 
782 	/* Write the iova of the HFI table */
783 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
784 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
785 
786 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
787 		(1 << 31) | (0xa << 18) | (0xa0));
788 
789 	/*
790 	 * Snapshots toggle the NMI bit which will result in a jump to the NMI
791 	 * handler instead of __main. Set the M3 config value to avoid that.
792 	 */
793 	gmu_write(gmu, REG_A6XX_GMU_CM3_CFG, 0x4052);
794 
795 	/*
796 	 * Note that the GMU has a slightly different layout for
797 	 * chip_id, for whatever reason, so a bit of massaging
798 	 * is needed.  The upper 16b are the same, but minor and
799 	 * patchid are packed in four bits each with the lower
800 	 * 8b unused:
801 	 */
802 	chipid  = adreno_gpu->chip_id & 0xffff0000;
803 	chipid |= (adreno_gpu->chip_id << 4) & 0xf000; /* minor */
804 	chipid |= (adreno_gpu->chip_id << 8) & 0x0f00; /* patchid */
805 
806 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
807 
808 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
809 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
810 
811 	/* Set up the lowest idle level on the GMU */
812 	a6xx_gmu_power_config(gmu);
813 
814 	ret = a6xx_gmu_start(gmu);
815 	if (ret)
816 		return ret;
817 
818 	if (gmu->legacy) {
819 		ret = a6xx_gmu_gfx_rail_on(gmu);
820 		if (ret)
821 			return ret;
822 	}
823 
824 	/* Enable SPTP_PC if the CPU is responsible for it */
825 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
826 		ret = a6xx_sptprac_enable(gmu);
827 		if (ret)
828 			return ret;
829 	}
830 
831 	ret = a6xx_gmu_hfi_start(gmu);
832 	if (ret)
833 		return ret;
834 
835 	/* FIXME: Do we need this wmb() here? */
836 	wmb();
837 
838 	return 0;
839 }
840 
841 #define A6XX_HFI_IRQ_MASK \
842 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
843 
844 #define A6XX_GMU_IRQ_MASK \
845 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
846 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
847 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
848 
849 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
850 {
851 	disable_irq(gmu->gmu_irq);
852 	disable_irq(gmu->hfi_irq);
853 
854 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
855 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
856 }
857 
858 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
859 {
860 	u32 val;
861 
862 	/* Make sure there are no outstanding RPMh votes */
863 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
864 		(val & 1), 100, 10000);
865 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
866 		(val & 1), 100, 10000);
867 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
868 		(val & 1), 100, 10000);
869 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
870 		(val & 1), 100, 1000);
871 }
872 
873 /* Force the GMU off in case it isn't responsive */
874 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
875 {
876 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
877 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
878 	struct msm_gpu *gpu = &adreno_gpu->base;
879 
880 	/*
881 	 * Turn off keep alive that might have been enabled by the hang
882 	 * interrupt
883 	 */
884 	gmu_write(&a6xx_gpu->gmu, REG_A6XX_GMU_GMU_PWR_COL_KEEPALIVE, 0);
885 
886 	/* Flush all the queues */
887 	a6xx_hfi_stop(gmu);
888 
889 	/* Stop the interrupts */
890 	a6xx_gmu_irq_disable(gmu);
891 
892 	/* Force off SPTP in case the GMU is managing it */
893 	a6xx_sptprac_disable(gmu);
894 
895 	/* Make sure there are no outstanding RPMh votes */
896 	a6xx_gmu_rpmh_off(gmu);
897 
898 	/* Clear the WRITEDROPPED fields and put fence into allow mode */
899 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS_CLR, 0x7);
900 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
901 
902 	/* Make sure the above writes go through */
903 	wmb();
904 
905 	/* Halt the gmu cm3 core */
906 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
907 
908 	a6xx_bus_clear_pending_transactions(adreno_gpu, true);
909 
910 	/* Reset GPU core blocks */
911 	a6xx_gpu_sw_reset(gpu, true);
912 }
913 
914 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
915 {
916 	struct dev_pm_opp *gpu_opp;
917 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
918 
919 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
920 	if (IS_ERR(gpu_opp))
921 		return;
922 
923 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
924 	a6xx_gmu_set_freq(gpu, gpu_opp, false);
925 	dev_pm_opp_put(gpu_opp);
926 }
927 
928 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
929 {
930 	struct dev_pm_opp *gpu_opp;
931 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
932 
933 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
934 	if (IS_ERR(gpu_opp))
935 		return;
936 
937 	dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp);
938 	dev_pm_opp_put(gpu_opp);
939 }
940 
941 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
942 {
943 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
944 	struct msm_gpu *gpu = &adreno_gpu->base;
945 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
946 	int status, ret;
947 
948 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
949 		return -EINVAL;
950 
951 	gmu->hung = false;
952 
953 	/* Turn on the resources */
954 	pm_runtime_get_sync(gmu->dev);
955 
956 	/*
957 	 * "enable" the GX power domain which won't actually do anything but it
958 	 * will make sure that the refcounting is correct in case we need to
959 	 * bring down the GX after a GMU failure
960 	 */
961 	if (!IS_ERR_OR_NULL(gmu->gxpd))
962 		pm_runtime_get_sync(gmu->gxpd);
963 
964 	/* Use a known rate to bring up the GMU */
965 	clk_set_rate(gmu->core_clk, 200000000);
966 	clk_set_rate(gmu->hub_clk, 150000000);
967 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
968 	if (ret) {
969 		pm_runtime_put(gmu->gxpd);
970 		pm_runtime_put(gmu->dev);
971 		return ret;
972 	}
973 
974 	/* Set the bus quota to a reasonable value for boot */
975 	a6xx_gmu_set_initial_bw(gpu, gmu);
976 
977 	/* Enable the GMU interrupt */
978 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
979 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
980 	enable_irq(gmu->gmu_irq);
981 
982 	/* Check to see if we are doing a cold or warm boot */
983 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
984 		GMU_WARM_BOOT : GMU_COLD_BOOT;
985 
986 	/*
987 	 * Warm boot path does not work on newer GPUs
988 	 * Presumably this is because icache/dcache regions must be restored
989 	 */
990 	if (!gmu->legacy)
991 		status = GMU_COLD_BOOT;
992 
993 	ret = a6xx_gmu_fw_start(gmu, status);
994 	if (ret)
995 		goto out;
996 
997 	ret = a6xx_hfi_start(gmu, status);
998 	if (ret)
999 		goto out;
1000 
1001 	/*
1002 	 * Turn on the GMU firmware fault interrupt after we know the boot
1003 	 * sequence is successful
1004 	 */
1005 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
1006 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
1007 	enable_irq(gmu->hfi_irq);
1008 
1009 	/* Set the GPU to the current freq */
1010 	a6xx_gmu_set_initial_freq(gpu, gmu);
1011 
1012 out:
1013 	/* On failure, shut down the GMU to leave it in a good state */
1014 	if (ret) {
1015 		disable_irq(gmu->gmu_irq);
1016 		a6xx_rpmh_stop(gmu);
1017 		pm_runtime_put(gmu->gxpd);
1018 		pm_runtime_put(gmu->dev);
1019 	}
1020 
1021 	return ret;
1022 }
1023 
1024 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
1025 {
1026 	u32 reg;
1027 
1028 	if (!gmu->initialized)
1029 		return true;
1030 
1031 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
1032 
1033 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
1034 		return false;
1035 
1036 	return true;
1037 }
1038 
1039 /* Gracefully try to shut down the GMU and by extension the GPU */
1040 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1041 {
1042 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1043 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1044 	u32 val;
1045 
1046 	/*
1047 	 * The GMU may still be in slumber unless the GPU started so check and
1048 	 * skip putting it back into slumber if so
1049 	 */
1050 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1051 
1052 	if (val != 0xf) {
1053 		int ret = a6xx_gmu_wait_for_idle(gmu);
1054 
1055 		/* If the GMU isn't responding assume it is hung */
1056 		if (ret) {
1057 			a6xx_gmu_force_off(gmu);
1058 			return;
1059 		}
1060 
1061 		a6xx_bus_clear_pending_transactions(adreno_gpu, a6xx_gpu->hung);
1062 
1063 		/* tell the GMU we want to slumber */
1064 		ret = a6xx_gmu_notify_slumber(gmu);
1065 		if (ret) {
1066 			a6xx_gmu_force_off(gmu);
1067 			return;
1068 		}
1069 
1070 		ret = gmu_poll_timeout(gmu,
1071 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1072 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1073 			100, 10000);
1074 
1075 		/*
1076 		 * Let the user know we failed to slumber but don't worry too
1077 		 * much because we are powering down anyway
1078 		 */
1079 
1080 		if (ret)
1081 			DRM_DEV_ERROR(gmu->dev,
1082 				"Unable to slumber GMU: status = 0%x/0%x\n",
1083 				gmu_read(gmu,
1084 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1085 				gmu_read(gmu,
1086 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1087 	}
1088 
1089 	/* Turn off HFI */
1090 	a6xx_hfi_stop(gmu);
1091 
1092 	/* Stop the interrupts and mask the hardware */
1093 	a6xx_gmu_irq_disable(gmu);
1094 
1095 	/* Tell RPMh to power off the GPU */
1096 	a6xx_rpmh_stop(gmu);
1097 }
1098 
1099 
1100 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1101 {
1102 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1103 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1104 
1105 	if (!pm_runtime_active(gmu->dev))
1106 		return 0;
1107 
1108 	/*
1109 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1110 	 * down gracefully
1111 	 */
1112 	if (gmu->hung)
1113 		a6xx_gmu_force_off(gmu);
1114 	else
1115 		a6xx_gmu_shutdown(gmu);
1116 
1117 	/* Remove the bus vote */
1118 	dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1119 
1120 	/*
1121 	 * Make sure the GX domain is off before turning off the GMU (CX)
1122 	 * domain. Usually the GMU does this but only if the shutdown sequence
1123 	 * was successful
1124 	 */
1125 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1126 		pm_runtime_put_sync(gmu->gxpd);
1127 
1128 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1129 
1130 	pm_runtime_put_sync(gmu->dev);
1131 
1132 	return 0;
1133 }
1134 
1135 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1136 {
1137 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace);
1138 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace);
1139 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace);
1140 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace);
1141 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace);
1142 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace);
1143 
1144 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1145 	msm_gem_address_space_put(gmu->aspace);
1146 }
1147 
1148 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1149 		size_t size, u64 iova, const char *name)
1150 {
1151 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1152 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1153 	uint32_t flags = MSM_BO_WC;
1154 	u64 range_start, range_end;
1155 	int ret;
1156 
1157 	size = PAGE_ALIGN(size);
1158 	if (!iova) {
1159 		/* no fixed address - use GMU's uncached range */
1160 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1161 		range_end = 0x80000000;
1162 	} else {
1163 		/* range for fixed address */
1164 		range_start = iova;
1165 		range_end = iova + size;
1166 		/* use IOMMU_PRIV for icache/dcache */
1167 		flags |= MSM_BO_MAP_PRIV;
1168 	}
1169 
1170 	bo->obj = msm_gem_new(dev, size, flags);
1171 	if (IS_ERR(bo->obj))
1172 		return PTR_ERR(bo->obj);
1173 
1174 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1175 					     range_start, range_end);
1176 	if (ret) {
1177 		drm_gem_object_put(bo->obj);
1178 		return ret;
1179 	}
1180 
1181 	bo->virt = msm_gem_get_vaddr(bo->obj);
1182 	bo->size = size;
1183 
1184 	msm_gem_object_set_name(bo->obj, name);
1185 
1186 	return 0;
1187 }
1188 
1189 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1190 {
1191 	struct msm_mmu *mmu;
1192 
1193 	mmu = msm_iommu_new(gmu->dev, 0);
1194 	if (!mmu)
1195 		return -ENODEV;
1196 	if (IS_ERR(mmu))
1197 		return PTR_ERR(mmu);
1198 
1199 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1200 	if (IS_ERR(gmu->aspace))
1201 		return PTR_ERR(gmu->aspace);
1202 
1203 	return 0;
1204 }
1205 
1206 /* Return the 'arc-level' for the given frequency */
1207 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1208 					   unsigned long freq)
1209 {
1210 	struct dev_pm_opp *opp;
1211 	unsigned int val;
1212 
1213 	if (!freq)
1214 		return 0;
1215 
1216 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1217 	if (IS_ERR(opp))
1218 		return 0;
1219 
1220 	val = dev_pm_opp_get_level(opp);
1221 
1222 	dev_pm_opp_put(opp);
1223 
1224 	return val;
1225 }
1226 
1227 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1228 		unsigned long *freqs, int freqs_count, const char *id)
1229 {
1230 	int i, j;
1231 	const u16 *pri, *sec;
1232 	size_t pri_count, sec_count;
1233 
1234 	pri = cmd_db_read_aux_data(id, &pri_count);
1235 	if (IS_ERR(pri))
1236 		return PTR_ERR(pri);
1237 	/*
1238 	 * The data comes back as an array of unsigned shorts so adjust the
1239 	 * count accordingly
1240 	 */
1241 	pri_count >>= 1;
1242 	if (!pri_count)
1243 		return -EINVAL;
1244 
1245 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1246 	if (IS_ERR(sec))
1247 		return PTR_ERR(sec);
1248 
1249 	sec_count >>= 1;
1250 	if (!sec_count)
1251 		return -EINVAL;
1252 
1253 	/* Construct a vote for each frequency */
1254 	for (i = 0; i < freqs_count; i++) {
1255 		u8 pindex = 0, sindex = 0;
1256 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1257 
1258 		/* Get the primary index that matches the arc level */
1259 		for (j = 0; j < pri_count; j++) {
1260 			if (pri[j] >= level) {
1261 				pindex = j;
1262 				break;
1263 			}
1264 		}
1265 
1266 		if (j == pri_count) {
1267 			DRM_DEV_ERROR(dev,
1268 				      "Level %u not found in the RPMh list\n",
1269 				      level);
1270 			DRM_DEV_ERROR(dev, "Available levels:\n");
1271 			for (j = 0; j < pri_count; j++)
1272 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1273 
1274 			return -EINVAL;
1275 		}
1276 
1277 		/*
1278 		 * Look for a level in in the secondary list that matches. If
1279 		 * nothing fits, use the maximum non zero vote
1280 		 */
1281 
1282 		for (j = 0; j < sec_count; j++) {
1283 			if (sec[j] >= level) {
1284 				sindex = j;
1285 				break;
1286 			} else if (sec[j]) {
1287 				sindex = j;
1288 			}
1289 		}
1290 
1291 		/* Construct the vote */
1292 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1293 			(sindex << 8) | pindex;
1294 	}
1295 
1296 	return 0;
1297 }
1298 
1299 /*
1300  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1301  * to construct the list of votes on the CPU and send it over. Query the RPMh
1302  * voltage levels and build the votes
1303  */
1304 
1305 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1306 {
1307 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1308 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1309 	struct msm_gpu *gpu = &adreno_gpu->base;
1310 	int ret;
1311 
1312 	/* Build the GX votes */
1313 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1314 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1315 
1316 	/* Build the CX votes */
1317 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1318 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1319 
1320 	return ret;
1321 }
1322 
1323 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1324 		u32 size)
1325 {
1326 	int count = dev_pm_opp_get_opp_count(dev);
1327 	struct dev_pm_opp *opp;
1328 	int i, index = 0;
1329 	unsigned long freq = 1;
1330 
1331 	/*
1332 	 * The OPP table doesn't contain the "off" frequency level so we need to
1333 	 * add 1 to the table size to account for it
1334 	 */
1335 
1336 	if (WARN(count + 1 > size,
1337 		"The GMU frequency table is being truncated\n"))
1338 		count = size - 1;
1339 
1340 	/* Set the "off" frequency */
1341 	freqs[index++] = 0;
1342 
1343 	for (i = 0; i < count; i++) {
1344 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1345 		if (IS_ERR(opp))
1346 			break;
1347 
1348 		dev_pm_opp_put(opp);
1349 		freqs[index++] = freq++;
1350 	}
1351 
1352 	return index;
1353 }
1354 
1355 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1356 {
1357 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1358 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1359 	struct msm_gpu *gpu = &adreno_gpu->base;
1360 
1361 	int ret = 0;
1362 
1363 	/*
1364 	 * The GMU handles its own frequency switching so build a list of
1365 	 * available frequencies to send during initialization
1366 	 */
1367 	ret = devm_pm_opp_of_add_table(gmu->dev);
1368 	if (ret) {
1369 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1370 		return ret;
1371 	}
1372 
1373 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1374 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1375 
1376 	/*
1377 	 * The GMU also handles GPU frequency switching so build a list
1378 	 * from the GPU OPP table
1379 	 */
1380 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1381 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1382 
1383 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1384 
1385 	/* Build the list of RPMh votes that we'll send to the GMU */
1386 	return a6xx_gmu_rpmh_votes_init(gmu);
1387 }
1388 
1389 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1390 {
1391 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1392 
1393 	if (ret < 1)
1394 		return ret;
1395 
1396 	gmu->nr_clocks = ret;
1397 
1398 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1399 		gmu->nr_clocks, "gmu");
1400 
1401 	gmu->hub_clk = msm_clk_bulk_get_clock(gmu->clocks,
1402 		gmu->nr_clocks, "hub");
1403 
1404 	return 0;
1405 }
1406 
1407 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1408 		const char *name)
1409 {
1410 	void __iomem *ret;
1411 	struct resource *res = platform_get_resource_byname(pdev,
1412 			IORESOURCE_MEM, name);
1413 
1414 	if (!res) {
1415 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1416 		return ERR_PTR(-EINVAL);
1417 	}
1418 
1419 	ret = ioremap(res->start, resource_size(res));
1420 	if (!ret) {
1421 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1422 		return ERR_PTR(-EINVAL);
1423 	}
1424 
1425 	return ret;
1426 }
1427 
1428 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1429 		const char *name, irq_handler_t handler)
1430 {
1431 	int irq, ret;
1432 
1433 	irq = platform_get_irq_byname(pdev, name);
1434 
1435 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH | IRQF_NO_AUTOEN, name, gmu);
1436 	if (ret) {
1437 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1438 			      name, ret);
1439 		return ret;
1440 	}
1441 
1442 	return irq;
1443 }
1444 
1445 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1446 {
1447 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1448 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1449 	struct platform_device *pdev = to_platform_device(gmu->dev);
1450 
1451 	mutex_lock(&gmu->lock);
1452 	if (!gmu->initialized) {
1453 		mutex_unlock(&gmu->lock);
1454 		return;
1455 	}
1456 
1457 	gmu->initialized = false;
1458 
1459 	mutex_unlock(&gmu->lock);
1460 
1461 	pm_runtime_force_suspend(gmu->dev);
1462 
1463 	/*
1464 	 * Since cxpd is a virt device, the devlink with gmu-dev will be removed
1465 	 * automatically when we do detach
1466 	 */
1467 	dev_pm_domain_detach(gmu->cxpd, false);
1468 
1469 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1470 		pm_runtime_disable(gmu->gxpd);
1471 		dev_pm_domain_detach(gmu->gxpd, false);
1472 	}
1473 
1474 	iounmap(gmu->mmio);
1475 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1476 		iounmap(gmu->rscc);
1477 	gmu->mmio = NULL;
1478 	gmu->rscc = NULL;
1479 
1480 	if (!adreno_has_gmu_wrapper(adreno_gpu)) {
1481 		a6xx_gmu_memory_free(gmu);
1482 
1483 		free_irq(gmu->gmu_irq, gmu);
1484 		free_irq(gmu->hfi_irq, gmu);
1485 	}
1486 
1487 	/* Drop reference taken in of_find_device_by_node */
1488 	put_device(gmu->dev);
1489 }
1490 
1491 static int cxpd_notifier_cb(struct notifier_block *nb,
1492 			unsigned long action, void *data)
1493 {
1494 	struct a6xx_gmu *gmu = container_of(nb, struct a6xx_gmu, pd_nb);
1495 
1496 	if (action == GENPD_NOTIFY_OFF)
1497 		complete_all(&gmu->pd_gate);
1498 
1499 	return 0;
1500 }
1501 
1502 int a6xx_gmu_wrapper_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1503 {
1504 	struct platform_device *pdev = of_find_device_by_node(node);
1505 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1506 	int ret;
1507 
1508 	if (!pdev)
1509 		return -ENODEV;
1510 
1511 	gmu->dev = &pdev->dev;
1512 
1513 	of_dma_configure(gmu->dev, node, true);
1514 
1515 	pm_runtime_enable(gmu->dev);
1516 
1517 	/* Mark legacy for manual SPTPRAC control */
1518 	gmu->legacy = true;
1519 
1520 	/* Map the GMU registers */
1521 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1522 	if (IS_ERR(gmu->mmio)) {
1523 		ret = PTR_ERR(gmu->mmio);
1524 		goto err_mmio;
1525 	}
1526 
1527 	gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
1528 	if (IS_ERR(gmu->cxpd)) {
1529 		ret = PTR_ERR(gmu->cxpd);
1530 		goto err_mmio;
1531 	}
1532 
1533 	if (!device_link_add(gmu->dev, gmu->cxpd, DL_FLAG_PM_RUNTIME)) {
1534 		ret = -ENODEV;
1535 		goto detach_cxpd;
1536 	}
1537 
1538 	init_completion(&gmu->pd_gate);
1539 	complete_all(&gmu->pd_gate);
1540 	gmu->pd_nb.notifier_call = cxpd_notifier_cb;
1541 
1542 	/* Get a link to the GX power domain to reset the GPU */
1543 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1544 	if (IS_ERR(gmu->gxpd)) {
1545 		ret = PTR_ERR(gmu->gxpd);
1546 		goto err_mmio;
1547 	}
1548 
1549 	gmu->initialized = true;
1550 
1551 	return 0;
1552 
1553 detach_cxpd:
1554 	dev_pm_domain_detach(gmu->cxpd, false);
1555 
1556 err_mmio:
1557 	iounmap(gmu->mmio);
1558 
1559 	/* Drop reference taken in of_find_device_by_node */
1560 	put_device(gmu->dev);
1561 
1562 	return ret;
1563 }
1564 
1565 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1566 {
1567 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1568 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1569 	struct platform_device *pdev = of_find_device_by_node(node);
1570 	int ret;
1571 
1572 	if (!pdev)
1573 		return -ENODEV;
1574 
1575 	gmu->dev = &pdev->dev;
1576 
1577 	of_dma_configure(gmu->dev, node, true);
1578 
1579 	/* Fow now, don't do anything fancy until we get our feet under us */
1580 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1581 
1582 	pm_runtime_enable(gmu->dev);
1583 
1584 	/* Get the list of clocks */
1585 	ret = a6xx_gmu_clocks_probe(gmu);
1586 	if (ret)
1587 		goto err_put_device;
1588 
1589 	ret = a6xx_gmu_memory_probe(gmu);
1590 	if (ret)
1591 		goto err_put_device;
1592 
1593 
1594 	/* A660 now requires handling "prealloc requests" in GMU firmware
1595 	 * For now just hardcode allocations based on the known firmware.
1596 	 * note: there is no indication that these correspond to "dummy" or
1597 	 * "debug" regions, but this "guess" allows reusing these BOs which
1598 	 * are otherwise unused by a660.
1599 	 */
1600 	gmu->dummy.size = SZ_4K;
1601 	if (adreno_is_a660_family(adreno_gpu)) {
1602 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_4K * 7,
1603 					    0x60400000, "debug");
1604 		if (ret)
1605 			goto err_memory;
1606 
1607 		gmu->dummy.size = SZ_8K;
1608 	}
1609 
1610 	/* Allocate memory for the GMU dummy page */
1611 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, gmu->dummy.size,
1612 				    0x60000000, "dummy");
1613 	if (ret)
1614 		goto err_memory;
1615 
1616 	/* Note that a650 family also includes a660 family: */
1617 	if (adreno_is_a650_family(adreno_gpu)) {
1618 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1619 			SZ_16M - SZ_16K, 0x04000, "icache");
1620 		if (ret)
1621 			goto err_memory;
1622 	/*
1623 	 * NOTE: when porting legacy ("pre-650-family") GPUs you may be tempted to add a condition
1624 	 * to allocate icache/dcache here, as per downstream code flow, but it may not actually be
1625 	 * necessary. If you omit this step and you don't get random pagefaults, you are likely
1626 	 * good to go without this!
1627 	 */
1628 	} else if (adreno_is_a640_family(adreno_gpu)) {
1629 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1630 			SZ_256K - SZ_16K, 0x04000, "icache");
1631 		if (ret)
1632 			goto err_memory;
1633 
1634 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1635 			SZ_256K - SZ_16K, 0x44000, "dcache");
1636 		if (ret)
1637 			goto err_memory;
1638 	} else if (adreno_is_a630_family(adreno_gpu)) {
1639 		/* HFI v1, has sptprac */
1640 		gmu->legacy = true;
1641 
1642 		/* Allocate memory for the GMU debug region */
1643 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0, "debug");
1644 		if (ret)
1645 			goto err_memory;
1646 	}
1647 
1648 	/* Allocate memory for the GMU log region */
1649 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_16K, 0, "log");
1650 	if (ret)
1651 		goto err_memory;
1652 
1653 	/* Allocate memory for for the HFI queues */
1654 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0, "hfi");
1655 	if (ret)
1656 		goto err_memory;
1657 
1658 	/* Map the GMU registers */
1659 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1660 	if (IS_ERR(gmu->mmio)) {
1661 		ret = PTR_ERR(gmu->mmio);
1662 		goto err_memory;
1663 	}
1664 
1665 	if (adreno_is_a650_family(adreno_gpu)) {
1666 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1667 		if (IS_ERR(gmu->rscc)) {
1668 			ret = -ENODEV;
1669 			goto err_mmio;
1670 		}
1671 	} else {
1672 		gmu->rscc = gmu->mmio + 0x23000;
1673 	}
1674 
1675 	/* Get the HFI and GMU interrupts */
1676 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1677 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1678 
1679 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0) {
1680 		ret = -ENODEV;
1681 		goto err_mmio;
1682 	}
1683 
1684 	gmu->cxpd = dev_pm_domain_attach_by_name(gmu->dev, "cx");
1685 	if (IS_ERR(gmu->cxpd)) {
1686 		ret = PTR_ERR(gmu->cxpd);
1687 		goto err_mmio;
1688 	}
1689 
1690 	if (!device_link_add(gmu->dev, gmu->cxpd,
1691 					DL_FLAG_PM_RUNTIME)) {
1692 		ret = -ENODEV;
1693 		goto detach_cxpd;
1694 	}
1695 
1696 	init_completion(&gmu->pd_gate);
1697 	complete_all(&gmu->pd_gate);
1698 	gmu->pd_nb.notifier_call = cxpd_notifier_cb;
1699 
1700 	/*
1701 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1702 	 * crash
1703 	 */
1704 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1705 
1706 	/* Get the power levels for the GMU and GPU */
1707 	a6xx_gmu_pwrlevels_probe(gmu);
1708 
1709 	/* Set up the HFI queues */
1710 	a6xx_hfi_init(gmu);
1711 
1712 	/* Initialize RPMh */
1713 	a6xx_gmu_rpmh_init(gmu);
1714 
1715 	gmu->initialized = true;
1716 
1717 	return 0;
1718 
1719 detach_cxpd:
1720 	dev_pm_domain_detach(gmu->cxpd, false);
1721 
1722 err_mmio:
1723 	iounmap(gmu->mmio);
1724 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1725 		iounmap(gmu->rscc);
1726 	free_irq(gmu->gmu_irq, gmu);
1727 	free_irq(gmu->hfi_irq, gmu);
1728 
1729 err_memory:
1730 	a6xx_gmu_memory_free(gmu);
1731 err_put_device:
1732 	/* Drop reference taken in of_find_device_by_node */
1733 	put_device(gmu->dev);
1734 
1735 	return ret;
1736 }
1737