xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision 15e3ae36)
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/dma-mapping.h>
6 #include <linux/interconnect.h>
7 #include <linux/pm_domain.h>
8 #include <linux/pm_opp.h>
9 #include <soc/qcom/cmd-db.h>
10 
11 #include "a6xx_gpu.h"
12 #include "a6xx_gmu.xml.h"
13 
14 static void a6xx_gmu_fault(struct a6xx_gmu *gmu)
15 {
16 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
17 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
18 	struct msm_gpu *gpu = &adreno_gpu->base;
19 	struct drm_device *dev = gpu->dev;
20 	struct msm_drm_private *priv = dev->dev_private;
21 
22 	/* FIXME: add a banner here */
23 	gmu->hung = true;
24 
25 	/* Turn off the hangcheck timer while we are resetting */
26 	del_timer(&gpu->hangcheck_timer);
27 
28 	/* Queue the GPU handler because we need to treat this as a recovery */
29 	queue_work(priv->wq, &gpu->recover_work);
30 }
31 
32 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
33 {
34 	struct a6xx_gmu *gmu = data;
35 	u32 status;
36 
37 	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
38 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
39 
40 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
41 		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
42 
43 		a6xx_gmu_fault(gmu);
44 	}
45 
46 	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
47 		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
48 
49 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
50 		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
51 			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
52 
53 	return IRQ_HANDLED;
54 }
55 
56 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
57 {
58 	struct a6xx_gmu *gmu = data;
59 	u32 status;
60 
61 	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
62 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
63 
64 	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
65 		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
66 
67 		a6xx_gmu_fault(gmu);
68 	}
69 
70 	return IRQ_HANDLED;
71 }
72 
73 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
74 {
75 	u32 val;
76 
77 	/* This can be called from gpu state code so make sure GMU is valid */
78 	if (!gmu->initialized)
79 		return false;
80 
81 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
82 
83 	return !(val &
84 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
85 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
86 }
87 
88 /* Check to see if the GX rail is still powered */
89 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
90 {
91 	u32 val;
92 
93 	/* This can be called from gpu state code so make sure GMU is valid */
94 	if (!gmu->initialized)
95 		return false;
96 
97 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
98 
99 	return !(val &
100 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
101 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
102 }
103 
104 static void __a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index)
105 {
106 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
107 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
108 	struct msm_gpu *gpu = &adreno_gpu->base;
109 	int ret;
110 
111 	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
112 
113 	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
114 		((3 & 0xf) << 28) | index);
115 
116 	/*
117 	 * Send an invalid index as a vote for the bus bandwidth and let the
118 	 * firmware decide on the right vote
119 	 */
120 	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
121 
122 	/* Set and clear the OOB for DCVS to trigger the GMU */
123 	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
124 	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
125 
126 	ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
127 	if (ret)
128 		dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
129 
130 	gmu->freq = gmu->gpu_freqs[index];
131 
132 	/*
133 	 * Eventually we will want to scale the path vote with the frequency but
134 	 * for now leave it at max so that the performance is nominal.
135 	 */
136 	icc_set_bw(gpu->icc_path, 0, MBps_to_icc(7216));
137 }
138 
139 void a6xx_gmu_set_freq(struct msm_gpu *gpu, unsigned long freq)
140 {
141 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
142 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
143 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
144 	u32 perf_index = 0;
145 
146 	if (freq == gmu->freq)
147 		return;
148 
149 	for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
150 		if (freq == gmu->gpu_freqs[perf_index])
151 			break;
152 
153 	gmu->current_perf_index = perf_index;
154 
155 	__a6xx_gmu_set_freq(gmu, perf_index);
156 }
157 
158 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
159 {
160 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
161 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
162 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
163 
164 	return  gmu->freq;
165 }
166 
167 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
168 {
169 	u32 val;
170 	int local = gmu->idle_level;
171 
172 	/* SPTP and IFPC both report as IFPC */
173 	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
174 		local = GMU_IDLE_STATE_IFPC;
175 
176 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
177 
178 	if (val == local) {
179 		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
180 			!a6xx_gmu_gx_is_on(gmu))
181 			return true;
182 	}
183 
184 	return false;
185 }
186 
187 /* Wait for the GMU to get to its most idle state */
188 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
189 {
190 	return spin_until(a6xx_gmu_check_idle_level(gmu));
191 }
192 
193 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
194 {
195 	int ret;
196 	u32 val;
197 
198 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
199 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
200 
201 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
202 		val == 0xbabeface, 100, 10000);
203 
204 	if (ret)
205 		DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
206 
207 	return ret;
208 }
209 
210 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
211 {
212 	u32 val;
213 	int ret;
214 
215 	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
216 
217 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
218 		val & 1, 100, 10000);
219 	if (ret)
220 		DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
221 
222 	return ret;
223 }
224 
225 /* Trigger a OOB (out of band) request to the GMU */
226 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
227 {
228 	int ret;
229 	u32 val;
230 	int request, ack;
231 	const char *name;
232 
233 	switch (state) {
234 	case GMU_OOB_GPU_SET:
235 		request = GMU_OOB_GPU_SET_REQUEST;
236 		ack = GMU_OOB_GPU_SET_ACK;
237 		name = "GPU_SET";
238 		break;
239 	case GMU_OOB_BOOT_SLUMBER:
240 		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
241 		ack = GMU_OOB_BOOT_SLUMBER_ACK;
242 		name = "BOOT_SLUMBER";
243 		break;
244 	case GMU_OOB_DCVS_SET:
245 		request = GMU_OOB_DCVS_REQUEST;
246 		ack = GMU_OOB_DCVS_ACK;
247 		name = "GPU_DCVS";
248 		break;
249 	default:
250 		return -EINVAL;
251 	}
252 
253 	/* Trigger the equested OOB operation */
254 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
255 
256 	/* Wait for the acknowledge interrupt */
257 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
258 		val & (1 << ack), 100, 10000);
259 
260 	if (ret)
261 		DRM_DEV_ERROR(gmu->dev,
262 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
263 				name,
264 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
265 
266 	/* Clear the acknowledge interrupt */
267 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
268 
269 	return ret;
270 }
271 
272 /* Clear a pending OOB state in the GMU */
273 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
274 {
275 	switch (state) {
276 	case GMU_OOB_GPU_SET:
277 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
278 			1 << GMU_OOB_GPU_SET_CLEAR);
279 		break;
280 	case GMU_OOB_BOOT_SLUMBER:
281 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
282 			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
283 		break;
284 	case GMU_OOB_DCVS_SET:
285 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
286 			1 << GMU_OOB_DCVS_CLEAR);
287 		break;
288 	}
289 }
290 
291 /* Enable CPU control of SPTP power power collapse */
292 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
293 {
294 	int ret;
295 	u32 val;
296 
297 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
298 
299 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
300 		(val & 0x38) == 0x28, 1, 100);
301 
302 	if (ret) {
303 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
304 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
305 	}
306 
307 	return 0;
308 }
309 
310 /* Disable CPU control of SPTP power power collapse */
311 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
312 {
313 	u32 val;
314 	int ret;
315 
316 	/* Make sure retention is on */
317 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
318 
319 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
320 
321 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
322 		(val & 0x04), 100, 10000);
323 
324 	if (ret)
325 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
326 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
327 }
328 
329 /* Let the GMU know we are starting a boot sequence */
330 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
331 {
332 	u32 vote;
333 
334 	/* Let the GMU know we are getting ready for boot */
335 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
336 
337 	/* Choose the "default" power level as the highest available */
338 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
339 
340 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
341 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
342 
343 	/* Let the GMU know the boot sequence has started */
344 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
345 }
346 
347 /* Let the GMU know that we are about to go into slumber */
348 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
349 {
350 	int ret;
351 
352 	/* Disable the power counter so the GMU isn't busy */
353 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
354 
355 	/* Disable SPTP_PC if the CPU is responsible for it */
356 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
357 		a6xx_sptprac_disable(gmu);
358 
359 	/* Tell the GMU to get ready to slumber */
360 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
361 
362 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
363 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
364 
365 	if (!ret) {
366 		/* Check to see if the GMU really did slumber */
367 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
368 			!= 0x0f) {
369 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
370 			ret = -ETIMEDOUT;
371 		}
372 	}
373 
374 	/* Put fence into allow mode */
375 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
376 	return ret;
377 }
378 
379 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
380 {
381 	int ret;
382 	u32 val;
383 
384 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
385 	/* Wait for the register to finish posting */
386 	wmb();
387 
388 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
389 		val & (1 << 1), 100, 10000);
390 	if (ret) {
391 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
392 		return ret;
393 	}
394 
395 	ret = gmu_poll_timeout(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
396 		!val, 100, 10000);
397 
398 	if (ret) {
399 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
400 		return ret;
401 	}
402 
403 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
404 
405 	/* Set up CX GMU counter 0 to count busy ticks */
406 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
407 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
408 
409 	/* Enable the power counter */
410 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
411 	return 0;
412 }
413 
414 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
415 {
416 	int ret;
417 	u32 val;
418 
419 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
420 
421 	ret = gmu_poll_timeout(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
422 		val, val & (1 << 16), 100, 10000);
423 	if (ret)
424 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
425 
426 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
427 }
428 
429 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
430 {
431 	return msm_writel(value, ptr + (offset << 2));
432 }
433 
434 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
435 		const char *name);
436 
437 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
438 {
439 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
440 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
441 	struct platform_device *pdev = to_platform_device(gmu->dev);
442 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
443 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
444 
445 	if (!pdcptr || !seqptr)
446 		goto err;
447 
448 	/* Disable SDE clock gating */
449 	gmu_write(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
450 
451 	/* Setup RSC PDC handshake for sleep and wakeup */
452 	gmu_write(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
453 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
454 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
455 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
456 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
457 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
458 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
459 	gmu_write(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
460 	gmu_write(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
461 	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
462 	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
463 
464 	/* Load RSC sequencer uCode for sleep and wakeup */
465 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
466 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
467 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
468 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
469 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
470 
471 	/* Load PDC sequencer uCode for power up and power down sequence */
472 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
473 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
474 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
475 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
476 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
477 
478 	/* Set TCS commands used by PDC sequence for low power modes */
479 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
480 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
481 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
482 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
483 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
484 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
485 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
486 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
487 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
488 
489 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
490 	if (adreno_is_a618(adreno_gpu))
491 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30090);
492 	else
493 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30080);
494 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
495 
496 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
497 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
498 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
499 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
500 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
501 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
502 
503 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
504 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
505 	if (adreno_is_a618(adreno_gpu))
506 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
507 	else
508 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
509 
510 
511 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
512 	if (adreno_is_a618(adreno_gpu))
513 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30090);
514 	else
515 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30080);
516 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
517 
518 	/* Setup GPU PDC */
519 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
520 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
521 
522 	/* ensure no writes happen before the uCode is fully written */
523 	wmb();
524 
525 err:
526 	if (!IS_ERR_OR_NULL(pdcptr))
527 		iounmap(pdcptr);
528 	if (!IS_ERR_OR_NULL(seqptr))
529 		iounmap(seqptr);
530 }
531 
532 /*
533  * The lowest 16 bits of this value are the number of XO clock cycles for main
534  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
535  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
536  */
537 
538 #define GMU_PWR_COL_HYST 0x000a1680
539 
540 /* Set up the idle state for the GMU */
541 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
542 {
543 	/* Disable GMU WB/RB buffer */
544 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
545 
546 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
547 
548 	switch (gmu->idle_level) {
549 	case GMU_IDLE_STATE_IFPC:
550 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
551 			GMU_PWR_COL_HYST);
552 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
553 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
554 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
555 		/* Fall through */
556 	case GMU_IDLE_STATE_SPTP:
557 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
558 			GMU_PWR_COL_HYST);
559 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
560 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
561 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
562 	}
563 
564 	/* Enable RPMh GPU client */
565 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
566 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
567 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
568 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
569 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
570 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
571 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
572 }
573 
574 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
575 {
576 	static bool rpmh_init;
577 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
578 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
579 	int i, ret;
580 	u32 chipid;
581 	u32 *image;
582 
583 	if (state == GMU_WARM_BOOT) {
584 		ret = a6xx_rpmh_start(gmu);
585 		if (ret)
586 			return ret;
587 	} else {
588 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
589 			"GMU firmware is not loaded\n"))
590 			return -ENOENT;
591 
592 		/* Sanity check the size of the firmware that was loaded */
593 		if (adreno_gpu->fw[ADRENO_FW_GMU]->size > 0x8000) {
594 			DRM_DEV_ERROR(gmu->dev,
595 				"GMU firmware is bigger than the available region\n");
596 			return -EINVAL;
597 		}
598 
599 		/* Turn on register retention */
600 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
601 
602 		/* We only need to load the RPMh microcode once */
603 		if (!rpmh_init) {
604 			a6xx_gmu_rpmh_init(gmu);
605 			rpmh_init = true;
606 		} else {
607 			ret = a6xx_rpmh_start(gmu);
608 			if (ret)
609 				return ret;
610 		}
611 
612 		image = (u32 *) adreno_gpu->fw[ADRENO_FW_GMU]->data;
613 
614 		for (i = 0; i < adreno_gpu->fw[ADRENO_FW_GMU]->size >> 2; i++)
615 			gmu_write(gmu, REG_A6XX_GMU_CM3_ITCM_START + i,
616 				image[i]);
617 	}
618 
619 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
620 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
621 
622 	/* Write the iova of the HFI table */
623 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi->iova);
624 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
625 
626 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
627 		(1 << 31) | (0xa << 18) | (0xa0));
628 
629 	chipid = adreno_gpu->rev.core << 24;
630 	chipid |= adreno_gpu->rev.major << 16;
631 	chipid |= adreno_gpu->rev.minor << 12;
632 	chipid |= adreno_gpu->rev.patchid << 8;
633 
634 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
635 
636 	/* Set up the lowest idle level on the GMU */
637 	a6xx_gmu_power_config(gmu);
638 
639 	ret = a6xx_gmu_start(gmu);
640 	if (ret)
641 		return ret;
642 
643 	ret = a6xx_gmu_gfx_rail_on(gmu);
644 	if (ret)
645 		return ret;
646 
647 	/* Enable SPTP_PC if the CPU is responsible for it */
648 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
649 		ret = a6xx_sptprac_enable(gmu);
650 		if (ret)
651 			return ret;
652 	}
653 
654 	ret = a6xx_gmu_hfi_start(gmu);
655 	if (ret)
656 		return ret;
657 
658 	/* FIXME: Do we need this wmb() here? */
659 	wmb();
660 
661 	return 0;
662 }
663 
664 #define A6XX_HFI_IRQ_MASK \
665 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
666 
667 #define A6XX_GMU_IRQ_MASK \
668 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
669 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
670 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
671 
672 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
673 {
674 	disable_irq(gmu->gmu_irq);
675 	disable_irq(gmu->hfi_irq);
676 
677 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
678 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
679 }
680 
681 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
682 {
683 	u32 val;
684 
685 	/* Make sure there are no outstanding RPMh votes */
686 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
687 		(val & 1), 100, 10000);
688 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
689 		(val & 1), 100, 10000);
690 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
691 		(val & 1), 100, 10000);
692 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
693 		(val & 1), 100, 1000);
694 }
695 
696 /* Force the GMU off in case it isn't responsive */
697 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
698 {
699 	/* Flush all the queues */
700 	a6xx_hfi_stop(gmu);
701 
702 	/* Stop the interrupts */
703 	a6xx_gmu_irq_disable(gmu);
704 
705 	/* Force off SPTP in case the GMU is managing it */
706 	a6xx_sptprac_disable(gmu);
707 
708 	/* Make sure there are no outstanding RPMh votes */
709 	a6xx_gmu_rpmh_off(gmu);
710 }
711 
712 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
713 {
714 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
715 	struct msm_gpu *gpu = &adreno_gpu->base;
716 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
717 	int status, ret;
718 
719 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
720 		return 0;
721 
722 	gmu->hung = false;
723 
724 	/* Turn on the resources */
725 	pm_runtime_get_sync(gmu->dev);
726 
727 	/* Use a known rate to bring up the GMU */
728 	clk_set_rate(gmu->core_clk, 200000000);
729 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
730 	if (ret) {
731 		pm_runtime_put(gmu->dev);
732 		return ret;
733 	}
734 
735 	/* Set the bus quota to a reasonable value for boot */
736 	icc_set_bw(gpu->icc_path, 0, MBps_to_icc(3072));
737 
738 	/* Enable the GMU interrupt */
739 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
740 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
741 	enable_irq(gmu->gmu_irq);
742 
743 	/* Check to see if we are doing a cold or warm boot */
744 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
745 		GMU_WARM_BOOT : GMU_COLD_BOOT;
746 
747 	ret = a6xx_gmu_fw_start(gmu, status);
748 	if (ret)
749 		goto out;
750 
751 	ret = a6xx_hfi_start(gmu, status);
752 	if (ret)
753 		goto out;
754 
755 	/*
756 	 * Turn on the GMU firmware fault interrupt after we know the boot
757 	 * sequence is successful
758 	 */
759 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
760 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
761 	enable_irq(gmu->hfi_irq);
762 
763 	/* Set the GPU to the current freq */
764 	__a6xx_gmu_set_freq(gmu, gmu->current_perf_index);
765 
766 	/*
767 	 * "enable" the GX power domain which won't actually do anything but it
768 	 * will make sure that the refcounting is correct in case we need to
769 	 * bring down the GX after a GMU failure
770 	 */
771 	if (!IS_ERR_OR_NULL(gmu->gxpd))
772 		pm_runtime_get(gmu->gxpd);
773 
774 out:
775 	/* On failure, shut down the GMU to leave it in a good state */
776 	if (ret) {
777 		disable_irq(gmu->gmu_irq);
778 		a6xx_rpmh_stop(gmu);
779 		pm_runtime_put(gmu->dev);
780 	}
781 
782 	return ret;
783 }
784 
785 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
786 {
787 	u32 reg;
788 
789 	if (!gmu->initialized)
790 		return true;
791 
792 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
793 
794 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
795 		return false;
796 
797 	return true;
798 }
799 
800 #define GBIF_CLIENT_HALT_MASK             BIT(0)
801 #define GBIF_ARB_HALT_MASK                BIT(1)
802 
803 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
804 {
805 	struct msm_gpu *gpu = &adreno_gpu->base;
806 
807 	if (!a6xx_has_gbif(adreno_gpu)) {
808 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
809 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
810 								0xf) == 0xf);
811 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
812 
813 		return;
814 	}
815 
816 	/* Halt new client requests on GBIF */
817 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
818 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
819 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
820 
821 	/* Halt all AXI requests on GBIF */
822 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
823 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
824 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
825 
826 	/* The GBIF halt needs to be explicitly cleared */
827 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
828 }
829 
830 /* Gracefully try to shut down the GMU and by extension the GPU */
831 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
832 {
833 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
834 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
835 	u32 val;
836 
837 	/*
838 	 * The GMU may still be in slumber unless the GPU started so check and
839 	 * skip putting it back into slumber if so
840 	 */
841 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
842 
843 	if (val != 0xf) {
844 		int ret = a6xx_gmu_wait_for_idle(gmu);
845 
846 		/* If the GMU isn't responding assume it is hung */
847 		if (ret) {
848 			a6xx_gmu_force_off(gmu);
849 			return;
850 		}
851 
852 		a6xx_bus_clear_pending_transactions(adreno_gpu);
853 
854 		/* tell the GMU we want to slumber */
855 		a6xx_gmu_notify_slumber(gmu);
856 
857 		ret = gmu_poll_timeout(gmu,
858 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
859 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
860 			100, 10000);
861 
862 		/*
863 		 * Let the user know we failed to slumber but don't worry too
864 		 * much because we are powering down anyway
865 		 */
866 
867 		if (ret)
868 			DRM_DEV_ERROR(gmu->dev,
869 				"Unable to slumber GMU: status = 0%x/0%x\n",
870 				gmu_read(gmu,
871 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
872 				gmu_read(gmu,
873 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
874 	}
875 
876 	/* Turn off HFI */
877 	a6xx_hfi_stop(gmu);
878 
879 	/* Stop the interrupts and mask the hardware */
880 	a6xx_gmu_irq_disable(gmu);
881 
882 	/* Tell RPMh to power off the GPU */
883 	a6xx_rpmh_stop(gmu);
884 }
885 
886 
887 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
888 {
889 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
890 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
891 
892 	if (!pm_runtime_active(gmu->dev))
893 		return 0;
894 
895 	/*
896 	 * Force the GMU off if we detected a hang, otherwise try to shut it
897 	 * down gracefully
898 	 */
899 	if (gmu->hung)
900 		a6xx_gmu_force_off(gmu);
901 	else
902 		a6xx_gmu_shutdown(gmu);
903 
904 	/* Remove the bus vote */
905 	icc_set_bw(gpu->icc_path, 0, 0);
906 
907 	/*
908 	 * Make sure the GX domain is off before turning off the GMU (CX)
909 	 * domain. Usually the GMU does this but only if the shutdown sequence
910 	 * was successful
911 	 */
912 	if (!IS_ERR_OR_NULL(gmu->gxpd))
913 		pm_runtime_put_sync(gmu->gxpd);
914 
915 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
916 
917 	pm_runtime_put_sync(gmu->dev);
918 
919 	return 0;
920 }
921 
922 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo)
923 {
924 	if (IS_ERR_OR_NULL(bo))
925 		return;
926 
927 	dma_free_wc(gmu->dev, bo->size, bo->virt, bo->iova);
928 	kfree(bo);
929 }
930 
931 static struct a6xx_gmu_bo *a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu,
932 		size_t size)
933 {
934 	struct a6xx_gmu_bo *bo;
935 
936 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
937 	if (!bo)
938 		return ERR_PTR(-ENOMEM);
939 
940 	bo->size = PAGE_ALIGN(size);
941 
942 	bo->virt = dma_alloc_wc(gmu->dev, bo->size, &bo->iova, GFP_KERNEL);
943 
944 	if (!bo->virt) {
945 		kfree(bo);
946 		return ERR_PTR(-ENOMEM);
947 	}
948 
949 	return bo;
950 }
951 
952 /* Return the 'arc-level' for the given frequency */
953 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
954 					   unsigned long freq)
955 {
956 	struct dev_pm_opp *opp;
957 	unsigned int val;
958 
959 	if (!freq)
960 		return 0;
961 
962 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
963 	if (IS_ERR(opp))
964 		return 0;
965 
966 	val = dev_pm_opp_get_level(opp);
967 
968 	dev_pm_opp_put(opp);
969 
970 	return val;
971 }
972 
973 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
974 		unsigned long *freqs, int freqs_count, const char *id)
975 {
976 	int i, j;
977 	const u16 *pri, *sec;
978 	size_t pri_count, sec_count;
979 
980 	pri = cmd_db_read_aux_data(id, &pri_count);
981 	if (IS_ERR(pri))
982 		return PTR_ERR(pri);
983 	/*
984 	 * The data comes back as an array of unsigned shorts so adjust the
985 	 * count accordingly
986 	 */
987 	pri_count >>= 1;
988 	if (!pri_count)
989 		return -EINVAL;
990 
991 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
992 	if (IS_ERR(sec))
993 		return PTR_ERR(sec);
994 
995 	sec_count >>= 1;
996 	if (!sec_count)
997 		return -EINVAL;
998 
999 	/* Construct a vote for each frequency */
1000 	for (i = 0; i < freqs_count; i++) {
1001 		u8 pindex = 0, sindex = 0;
1002 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1003 
1004 		/* Get the primary index that matches the arc level */
1005 		for (j = 0; j < pri_count; j++) {
1006 			if (pri[j] >= level) {
1007 				pindex = j;
1008 				break;
1009 			}
1010 		}
1011 
1012 		if (j == pri_count) {
1013 			DRM_DEV_ERROR(dev,
1014 				"Level %u not found in in the RPMh list\n",
1015 					level);
1016 			DRM_DEV_ERROR(dev, "Available levels:\n");
1017 			for (j = 0; j < pri_count; j++)
1018 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1019 
1020 			return -EINVAL;
1021 		}
1022 
1023 		/*
1024 		 * Look for a level in in the secondary list that matches. If
1025 		 * nothing fits, use the maximum non zero vote
1026 		 */
1027 
1028 		for (j = 0; j < sec_count; j++) {
1029 			if (sec[j] >= level) {
1030 				sindex = j;
1031 				break;
1032 			} else if (sec[j]) {
1033 				sindex = j;
1034 			}
1035 		}
1036 
1037 		/* Construct the vote */
1038 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1039 			(sindex << 8) | pindex;
1040 	}
1041 
1042 	return 0;
1043 }
1044 
1045 /*
1046  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1047  * to construct the list of votes on the CPU and send it over. Query the RPMh
1048  * voltage levels and build the votes
1049  */
1050 
1051 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1052 {
1053 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1054 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1055 	struct msm_gpu *gpu = &adreno_gpu->base;
1056 	int ret;
1057 
1058 	/* Build the GX votes */
1059 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1060 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1061 
1062 	/* Build the CX votes */
1063 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1064 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1065 
1066 	return ret;
1067 }
1068 
1069 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1070 		u32 size)
1071 {
1072 	int count = dev_pm_opp_get_opp_count(dev);
1073 	struct dev_pm_opp *opp;
1074 	int i, index = 0;
1075 	unsigned long freq = 1;
1076 
1077 	/*
1078 	 * The OPP table doesn't contain the "off" frequency level so we need to
1079 	 * add 1 to the table size to account for it
1080 	 */
1081 
1082 	if (WARN(count + 1 > size,
1083 		"The GMU frequency table is being truncated\n"))
1084 		count = size - 1;
1085 
1086 	/* Set the "off" frequency */
1087 	freqs[index++] = 0;
1088 
1089 	for (i = 0; i < count; i++) {
1090 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1091 		if (IS_ERR(opp))
1092 			break;
1093 
1094 		dev_pm_opp_put(opp);
1095 		freqs[index++] = freq++;
1096 	}
1097 
1098 	return index;
1099 }
1100 
1101 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1102 {
1103 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1104 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1105 	struct msm_gpu *gpu = &adreno_gpu->base;
1106 
1107 	int ret = 0;
1108 
1109 	/*
1110 	 * The GMU handles its own frequency switching so build a list of
1111 	 * available frequencies to send during initialization
1112 	 */
1113 	ret = dev_pm_opp_of_add_table(gmu->dev);
1114 	if (ret) {
1115 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1116 		return ret;
1117 	}
1118 
1119 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1120 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1121 
1122 	/*
1123 	 * The GMU also handles GPU frequency switching so build a list
1124 	 * from the GPU OPP table
1125 	 */
1126 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1127 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1128 
1129 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1130 
1131 	/* Build the list of RPMh votes that we'll send to the GMU */
1132 	return a6xx_gmu_rpmh_votes_init(gmu);
1133 }
1134 
1135 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1136 {
1137 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1138 
1139 	if (ret < 1)
1140 		return ret;
1141 
1142 	gmu->nr_clocks = ret;
1143 
1144 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1145 		gmu->nr_clocks, "gmu");
1146 
1147 	return 0;
1148 }
1149 
1150 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1151 		const char *name)
1152 {
1153 	void __iomem *ret;
1154 	struct resource *res = platform_get_resource_byname(pdev,
1155 			IORESOURCE_MEM, name);
1156 
1157 	if (!res) {
1158 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1159 		return ERR_PTR(-EINVAL);
1160 	}
1161 
1162 	ret = ioremap(res->start, resource_size(res));
1163 	if (!ret) {
1164 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1165 		return ERR_PTR(-EINVAL);
1166 	}
1167 
1168 	return ret;
1169 }
1170 
1171 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1172 		const char *name, irq_handler_t handler)
1173 {
1174 	int irq, ret;
1175 
1176 	irq = platform_get_irq_byname(pdev, name);
1177 
1178 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1179 	if (ret) {
1180 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1181 			      name, ret);
1182 		return ret;
1183 	}
1184 
1185 	disable_irq(irq);
1186 
1187 	return irq;
1188 }
1189 
1190 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1191 {
1192 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1193 
1194 	if (!gmu->initialized)
1195 		return;
1196 
1197 	pm_runtime_force_suspend(gmu->dev);
1198 
1199 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1200 		pm_runtime_disable(gmu->gxpd);
1201 		dev_pm_domain_detach(gmu->gxpd, false);
1202 	}
1203 
1204 	iounmap(gmu->mmio);
1205 	gmu->mmio = NULL;
1206 
1207 	a6xx_gmu_memory_free(gmu, gmu->hfi);
1208 
1209 	free_irq(gmu->gmu_irq, gmu);
1210 	free_irq(gmu->hfi_irq, gmu);
1211 
1212 	/* Drop reference taken in of_find_device_by_node */
1213 	put_device(gmu->dev);
1214 
1215 	gmu->initialized = false;
1216 }
1217 
1218 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1219 {
1220 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1221 	struct platform_device *pdev = of_find_device_by_node(node);
1222 	int ret;
1223 
1224 	if (!pdev)
1225 		return -ENODEV;
1226 
1227 	gmu->dev = &pdev->dev;
1228 
1229 	/* Pass force_dma false to require the DT to set the dma region */
1230 	ret = of_dma_configure(gmu->dev, node, false);
1231 	if (ret)
1232 		return ret;
1233 
1234 	/* Set the mask after the of_dma_configure() */
1235 	ret = dma_set_mask_and_coherent(&pdev->dev, DMA_BIT_MASK(31));
1236 	if (ret)
1237 		return ret;
1238 
1239 	/* Fow now, don't do anything fancy until we get our feet under us */
1240 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1241 
1242 	pm_runtime_enable(gmu->dev);
1243 
1244 	/* Get the list of clocks */
1245 	ret = a6xx_gmu_clocks_probe(gmu);
1246 	if (ret)
1247 		goto err_put_device;
1248 
1249 	/* Allocate memory for for the HFI queues */
1250 	gmu->hfi = a6xx_gmu_memory_alloc(gmu, SZ_16K);
1251 	if (IS_ERR(gmu->hfi))
1252 		goto err_memory;
1253 
1254 	/* Allocate memory for the GMU debug region */
1255 	gmu->debug = a6xx_gmu_memory_alloc(gmu, SZ_16K);
1256 	if (IS_ERR(gmu->debug))
1257 		goto err_memory;
1258 
1259 	/* Map the GMU registers */
1260 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1261 	if (IS_ERR(gmu->mmio))
1262 		goto err_memory;
1263 
1264 	/* Get the HFI and GMU interrupts */
1265 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1266 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1267 
1268 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1269 		goto err_mmio;
1270 
1271 	/*
1272 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1273 	 * crash
1274 	 */
1275 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1276 
1277 	/* Get the power levels for the GMU and GPU */
1278 	a6xx_gmu_pwrlevels_probe(gmu);
1279 
1280 	/* Set up the HFI queues */
1281 	a6xx_hfi_init(gmu);
1282 
1283 	gmu->initialized = true;
1284 
1285 	return 0;
1286 
1287 err_mmio:
1288 	iounmap(gmu->mmio);
1289 	free_irq(gmu->gmu_irq, gmu);
1290 	free_irq(gmu->hfi_irq, gmu);
1291 err_memory:
1292 	a6xx_gmu_memory_free(gmu, gmu->hfi);
1293 
1294 	ret = -ENODEV;
1295 
1296 err_put_device:
1297 	/* Drop reference taken in of_find_device_by_node */
1298 	put_device(gmu->dev);
1299 
1300 	return ret;
1301 }
1302