xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision b830f94f)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2018 The Linux Foundation. All rights reserved. */
3 
4 #include <linux/clk.h>
5 #include <linux/interconnect.h>
6 #include <linux/pm_domain.h>
7 #include <linux/pm_opp.h>
8 #include <soc/qcom/cmd-db.h>
9 
10 #include "a6xx_gpu.h"
11 #include "a6xx_gmu.xml.h"
12 
13 static void a6xx_gmu_fault(struct a6xx_gmu *gmu)
14 {
15 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
16 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
17 	struct msm_gpu *gpu = &adreno_gpu->base;
18 	struct drm_device *dev = gpu->dev;
19 	struct msm_drm_private *priv = dev->dev_private;
20 
21 	/* FIXME: add a banner here */
22 	gmu->hung = true;
23 
24 	/* Turn off the hangcheck timer while we are resetting */
25 	del_timer(&gpu->hangcheck_timer);
26 
27 	/* Queue the GPU handler because we need to treat this as a recovery */
28 	queue_work(priv->wq, &gpu->recover_work);
29 }
30 
31 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
32 {
33 	struct a6xx_gmu *gmu = data;
34 	u32 status;
35 
36 	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
37 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
38 
39 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
40 		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
41 
42 		a6xx_gmu_fault(gmu);
43 	}
44 
45 	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
46 		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
47 
48 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
49 		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
50 			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
51 
52 	return IRQ_HANDLED;
53 }
54 
55 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
56 {
57 	struct a6xx_gmu *gmu = data;
58 	u32 status;
59 
60 	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
61 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
62 
63 	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
64 		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
65 
66 		a6xx_gmu_fault(gmu);
67 	}
68 
69 	return IRQ_HANDLED;
70 }
71 
72 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
73 {
74 	u32 val;
75 
76 	/* This can be called from gpu state code so make sure GMU is valid */
77 	if (!gmu->initialized)
78 		return false;
79 
80 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
81 
82 	return !(val &
83 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
84 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
85 }
86 
87 /* Check to see if the GX rail is still powered */
88 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
89 {
90 	u32 val;
91 
92 	/* This can be called from gpu state code so make sure GMU is valid */
93 	if (!gmu->initialized)
94 		return false;
95 
96 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
97 
98 	return !(val &
99 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
100 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
101 }
102 
103 static void __a6xx_gmu_set_freq(struct a6xx_gmu *gmu, int index)
104 {
105 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
106 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
107 	struct msm_gpu *gpu = &adreno_gpu->base;
108 	int ret;
109 
110 	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
111 
112 	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
113 		((3 & 0xf) << 28) | index);
114 
115 	/*
116 	 * Send an invalid index as a vote for the bus bandwidth and let the
117 	 * firmware decide on the right vote
118 	 */
119 	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
120 
121 	/* Set and clear the OOB for DCVS to trigger the GMU */
122 	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
123 	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
124 
125 	ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
126 	if (ret)
127 		dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
128 
129 	gmu->freq = gmu->gpu_freqs[index];
130 
131 	/*
132 	 * Eventually we will want to scale the path vote with the frequency but
133 	 * for now leave it at max so that the performance is nominal.
134 	 */
135 	icc_set_bw(gpu->icc_path, 0, MBps_to_icc(7216));
136 }
137 
138 void a6xx_gmu_set_freq(struct msm_gpu *gpu, unsigned long freq)
139 {
140 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
141 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
142 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
143 	u32 perf_index = 0;
144 
145 	if (freq == gmu->freq)
146 		return;
147 
148 	for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
149 		if (freq == gmu->gpu_freqs[perf_index])
150 			break;
151 
152 	__a6xx_gmu_set_freq(gmu, perf_index);
153 }
154 
155 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
156 {
157 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
158 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
159 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
160 
161 	return  gmu->freq;
162 }
163 
164 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
165 {
166 	u32 val;
167 	int local = gmu->idle_level;
168 
169 	/* SPTP and IFPC both report as IFPC */
170 	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
171 		local = GMU_IDLE_STATE_IFPC;
172 
173 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
174 
175 	if (val == local) {
176 		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
177 			!a6xx_gmu_gx_is_on(gmu))
178 			return true;
179 	}
180 
181 	return false;
182 }
183 
184 /* Wait for the GMU to get to its most idle state */
185 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
186 {
187 	return spin_until(a6xx_gmu_check_idle_level(gmu));
188 }
189 
190 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
191 {
192 	int ret;
193 	u32 val;
194 
195 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
196 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
197 
198 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
199 		val == 0xbabeface, 100, 10000);
200 
201 	if (ret)
202 		DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
203 
204 	return ret;
205 }
206 
207 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
208 {
209 	u32 val;
210 	int ret;
211 
212 	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
213 
214 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
215 		val & 1, 100, 10000);
216 	if (ret)
217 		DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
218 
219 	return ret;
220 }
221 
222 /* Trigger a OOB (out of band) request to the GMU */
223 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
224 {
225 	int ret;
226 	u32 val;
227 	int request, ack;
228 	const char *name;
229 
230 	switch (state) {
231 	case GMU_OOB_GPU_SET:
232 		request = GMU_OOB_GPU_SET_REQUEST;
233 		ack = GMU_OOB_GPU_SET_ACK;
234 		name = "GPU_SET";
235 		break;
236 	case GMU_OOB_BOOT_SLUMBER:
237 		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
238 		ack = GMU_OOB_BOOT_SLUMBER_ACK;
239 		name = "BOOT_SLUMBER";
240 		break;
241 	case GMU_OOB_DCVS_SET:
242 		request = GMU_OOB_DCVS_REQUEST;
243 		ack = GMU_OOB_DCVS_ACK;
244 		name = "GPU_DCVS";
245 		break;
246 	default:
247 		return -EINVAL;
248 	}
249 
250 	/* Trigger the equested OOB operation */
251 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
252 
253 	/* Wait for the acknowledge interrupt */
254 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
255 		val & (1 << ack), 100, 10000);
256 
257 	if (ret)
258 		DRM_DEV_ERROR(gmu->dev,
259 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
260 				name,
261 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
262 
263 	/* Clear the acknowledge interrupt */
264 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
265 
266 	return ret;
267 }
268 
269 /* Clear a pending OOB state in the GMU */
270 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
271 {
272 	switch (state) {
273 	case GMU_OOB_GPU_SET:
274 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
275 			1 << GMU_OOB_GPU_SET_CLEAR);
276 		break;
277 	case GMU_OOB_BOOT_SLUMBER:
278 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
279 			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
280 		break;
281 	case GMU_OOB_DCVS_SET:
282 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
283 			1 << GMU_OOB_DCVS_CLEAR);
284 		break;
285 	}
286 }
287 
288 /* Enable CPU control of SPTP power power collapse */
289 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
290 {
291 	int ret;
292 	u32 val;
293 
294 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
295 
296 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
297 		(val & 0x38) == 0x28, 1, 100);
298 
299 	if (ret) {
300 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
301 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
302 	}
303 
304 	return 0;
305 }
306 
307 /* Disable CPU control of SPTP power power collapse */
308 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
309 {
310 	u32 val;
311 	int ret;
312 
313 	/* Make sure retention is on */
314 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
315 
316 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
317 
318 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
319 		(val & 0x04), 100, 10000);
320 
321 	if (ret)
322 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
323 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
324 }
325 
326 /* Let the GMU know we are starting a boot sequence */
327 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
328 {
329 	u32 vote;
330 
331 	/* Let the GMU know we are getting ready for boot */
332 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
333 
334 	/* Choose the "default" power level as the highest available */
335 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
336 
337 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
338 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
339 
340 	/* Let the GMU know the boot sequence has started */
341 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
342 }
343 
344 /* Let the GMU know that we are about to go into slumber */
345 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
346 {
347 	int ret;
348 
349 	/* Disable the power counter so the GMU isn't busy */
350 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
351 
352 	/* Disable SPTP_PC if the CPU is responsible for it */
353 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
354 		a6xx_sptprac_disable(gmu);
355 
356 	/* Tell the GMU to get ready to slumber */
357 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
358 
359 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
360 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
361 
362 	if (!ret) {
363 		/* Check to see if the GMU really did slumber */
364 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
365 			!= 0x0f) {
366 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
367 			ret = -ETIMEDOUT;
368 		}
369 	}
370 
371 	/* Put fence into allow mode */
372 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
373 	return ret;
374 }
375 
376 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
377 {
378 	int ret;
379 	u32 val;
380 
381 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
382 	/* Wait for the register to finish posting */
383 	wmb();
384 
385 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
386 		val & (1 << 1), 100, 10000);
387 	if (ret) {
388 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
389 		return ret;
390 	}
391 
392 	ret = gmu_poll_timeout(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
393 		!val, 100, 10000);
394 
395 	if (ret) {
396 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
397 		return ret;
398 	}
399 
400 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
401 
402 	/* Set up CX GMU counter 0 to count busy ticks */
403 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
404 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
405 
406 	/* Enable the power counter */
407 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
408 	return 0;
409 }
410 
411 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
412 {
413 	int ret;
414 	u32 val;
415 
416 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
417 
418 	ret = gmu_poll_timeout(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
419 		val, val & (1 << 16), 100, 10000);
420 	if (ret)
421 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
422 
423 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
424 }
425 
426 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
427 {
428 	return msm_writel(value, ptr + (offset << 2));
429 }
430 
431 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
432 		const char *name);
433 
434 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
435 {
436 	struct platform_device *pdev = to_platform_device(gmu->dev);
437 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
438 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
439 
440 	if (!pdcptr || !seqptr)
441 		goto err;
442 
443 	/* Disable SDE clock gating */
444 	gmu_write(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
445 
446 	/* Setup RSC PDC handshake for sleep and wakeup */
447 	gmu_write(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
448 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
449 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
450 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
451 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
452 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
453 	gmu_write(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
454 	gmu_write(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
455 	gmu_write(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
456 	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
457 	gmu_write(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
458 
459 	/* Load RSC sequencer uCode for sleep and wakeup */
460 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
461 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
462 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
463 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
464 	gmu_write(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
465 
466 	/* Load PDC sequencer uCode for power up and power down sequence */
467 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
468 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
469 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
470 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
471 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
472 
473 	/* Set TCS commands used by PDC sequence for low power modes */
474 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
475 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
476 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
477 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
478 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
479 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
480 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
481 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
482 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
483 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
484 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, 0x30080);
485 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
486 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
487 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
488 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
489 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
490 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
491 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
492 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
493 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
494 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
495 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
496 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, 0x30080);
497 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
498 
499 	/* Setup GPU PDC */
500 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
501 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
502 
503 	/* ensure no writes happen before the uCode is fully written */
504 	wmb();
505 
506 err:
507 	if (!IS_ERR_OR_NULL(pdcptr))
508 		iounmap(pdcptr);
509 	if (!IS_ERR_OR_NULL(seqptr))
510 		iounmap(seqptr);
511 }
512 
513 /*
514  * The lowest 16 bits of this value are the number of XO clock cycles for main
515  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
516  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
517  */
518 
519 #define GMU_PWR_COL_HYST 0x000a1680
520 
521 /* Set up the idle state for the GMU */
522 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
523 {
524 	/* Disable GMU WB/RB buffer */
525 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
526 
527 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
528 
529 	switch (gmu->idle_level) {
530 	case GMU_IDLE_STATE_IFPC:
531 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
532 			GMU_PWR_COL_HYST);
533 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
534 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
535 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
536 		/* Fall through */
537 	case GMU_IDLE_STATE_SPTP:
538 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
539 			GMU_PWR_COL_HYST);
540 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
541 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
542 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
543 	}
544 
545 	/* Enable RPMh GPU client */
546 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
547 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
548 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
549 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
550 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
551 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
552 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
553 }
554 
555 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
556 {
557 	static bool rpmh_init;
558 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
559 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
560 	int i, ret;
561 	u32 chipid;
562 	u32 *image;
563 
564 	if (state == GMU_WARM_BOOT) {
565 		ret = a6xx_rpmh_start(gmu);
566 		if (ret)
567 			return ret;
568 	} else {
569 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
570 			"GMU firmware is not loaded\n"))
571 			return -ENOENT;
572 
573 		/* Sanity check the size of the firmware that was loaded */
574 		if (adreno_gpu->fw[ADRENO_FW_GMU]->size > 0x8000) {
575 			DRM_DEV_ERROR(gmu->dev,
576 				"GMU firmware is bigger than the available region\n");
577 			return -EINVAL;
578 		}
579 
580 		/* Turn on register retention */
581 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
582 
583 		/* We only need to load the RPMh microcode once */
584 		if (!rpmh_init) {
585 			a6xx_gmu_rpmh_init(gmu);
586 			rpmh_init = true;
587 		} else {
588 			ret = a6xx_rpmh_start(gmu);
589 			if (ret)
590 				return ret;
591 		}
592 
593 		image = (u32 *) adreno_gpu->fw[ADRENO_FW_GMU]->data;
594 
595 		for (i = 0; i < adreno_gpu->fw[ADRENO_FW_GMU]->size >> 2; i++)
596 			gmu_write(gmu, REG_A6XX_GMU_CM3_ITCM_START + i,
597 				image[i]);
598 	}
599 
600 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
601 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
602 
603 	/* Write the iova of the HFI table */
604 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi->iova);
605 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
606 
607 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
608 		(1 << 31) | (0xa << 18) | (0xa0));
609 
610 	chipid = adreno_gpu->rev.core << 24;
611 	chipid |= adreno_gpu->rev.major << 16;
612 	chipid |= adreno_gpu->rev.minor << 12;
613 	chipid |= adreno_gpu->rev.patchid << 8;
614 
615 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
616 
617 	/* Set up the lowest idle level on the GMU */
618 	a6xx_gmu_power_config(gmu);
619 
620 	ret = a6xx_gmu_start(gmu);
621 	if (ret)
622 		return ret;
623 
624 	ret = a6xx_gmu_gfx_rail_on(gmu);
625 	if (ret)
626 		return ret;
627 
628 	/* Enable SPTP_PC if the CPU is responsible for it */
629 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
630 		ret = a6xx_sptprac_enable(gmu);
631 		if (ret)
632 			return ret;
633 	}
634 
635 	ret = a6xx_gmu_hfi_start(gmu);
636 	if (ret)
637 		return ret;
638 
639 	/* FIXME: Do we need this wmb() here? */
640 	wmb();
641 
642 	return 0;
643 }
644 
645 #define A6XX_HFI_IRQ_MASK \
646 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
647 
648 #define A6XX_GMU_IRQ_MASK \
649 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
650 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
651 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
652 
653 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
654 {
655 	disable_irq(gmu->gmu_irq);
656 	disable_irq(gmu->hfi_irq);
657 
658 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
659 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
660 }
661 
662 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
663 {
664 	u32 val;
665 
666 	/* Make sure there are no outstanding RPMh votes */
667 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
668 		(val & 1), 100, 10000);
669 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
670 		(val & 1), 100, 10000);
671 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
672 		(val & 1), 100, 10000);
673 	gmu_poll_timeout(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
674 		(val & 1), 100, 1000);
675 }
676 
677 /* Force the GMU off in case it isn't responsive */
678 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
679 {
680 	/* Flush all the queues */
681 	a6xx_hfi_stop(gmu);
682 
683 	/* Stop the interrupts */
684 	a6xx_gmu_irq_disable(gmu);
685 
686 	/* Force off SPTP in case the GMU is managing it */
687 	a6xx_sptprac_disable(gmu);
688 
689 	/* Make sure there are no outstanding RPMh votes */
690 	a6xx_gmu_rpmh_off(gmu);
691 }
692 
693 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
694 {
695 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
696 	struct msm_gpu *gpu = &adreno_gpu->base;
697 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
698 	int status, ret;
699 
700 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
701 		return 0;
702 
703 	gmu->hung = false;
704 
705 	/* Turn on the resources */
706 	pm_runtime_get_sync(gmu->dev);
707 
708 	/* Use a known rate to bring up the GMU */
709 	clk_set_rate(gmu->core_clk, 200000000);
710 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
711 	if (ret) {
712 		pm_runtime_put(gmu->dev);
713 		return ret;
714 	}
715 
716 	/* Set the bus quota to a reasonable value for boot */
717 	icc_set_bw(gpu->icc_path, 0, MBps_to_icc(3072));
718 
719 	/* Enable the GMU interrupt */
720 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
721 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
722 	enable_irq(gmu->gmu_irq);
723 
724 	/* Check to see if we are doing a cold or warm boot */
725 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
726 		GMU_WARM_BOOT : GMU_COLD_BOOT;
727 
728 	ret = a6xx_gmu_fw_start(gmu, status);
729 	if (ret)
730 		goto out;
731 
732 	ret = a6xx_hfi_start(gmu, status);
733 	if (ret)
734 		goto out;
735 
736 	/*
737 	 * Turn on the GMU firmware fault interrupt after we know the boot
738 	 * sequence is successful
739 	 */
740 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
741 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
742 	enable_irq(gmu->hfi_irq);
743 
744 	/* Set the GPU to the highest power frequency */
745 	__a6xx_gmu_set_freq(gmu, gmu->nr_gpu_freqs - 1);
746 
747 	/*
748 	 * "enable" the GX power domain which won't actually do anything but it
749 	 * will make sure that the refcounting is correct in case we need to
750 	 * bring down the GX after a GMU failure
751 	 */
752 	if (!IS_ERR_OR_NULL(gmu->gxpd))
753 		pm_runtime_get(gmu->gxpd);
754 
755 out:
756 	/* On failure, shut down the GMU to leave it in a good state */
757 	if (ret) {
758 		disable_irq(gmu->gmu_irq);
759 		a6xx_rpmh_stop(gmu);
760 		pm_runtime_put(gmu->dev);
761 	}
762 
763 	return ret;
764 }
765 
766 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
767 {
768 	u32 reg;
769 
770 	if (!gmu->initialized)
771 		return true;
772 
773 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
774 
775 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
776 		return false;
777 
778 	return true;
779 }
780 
781 /* Gracefully try to shut down the GMU and by extension the GPU */
782 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
783 {
784 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
785 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
786 	struct msm_gpu *gpu = &adreno_gpu->base;
787 	u32 val;
788 
789 	/*
790 	 * The GMU may still be in slumber unless the GPU started so check and
791 	 * skip putting it back into slumber if so
792 	 */
793 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
794 
795 	if (val != 0xf) {
796 		int ret = a6xx_gmu_wait_for_idle(gmu);
797 
798 		/* If the GMU isn't responding assume it is hung */
799 		if (ret) {
800 			a6xx_gmu_force_off(gmu);
801 			return;
802 		}
803 
804 		/* Clear the VBIF pipe before shutting down */
805 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
806 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) & 0xf)
807 			== 0xf);
808 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
809 
810 		/* tell the GMU we want to slumber */
811 		a6xx_gmu_notify_slumber(gmu);
812 
813 		ret = gmu_poll_timeout(gmu,
814 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
815 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
816 			100, 10000);
817 
818 		/*
819 		 * Let the user know we failed to slumber but don't worry too
820 		 * much because we are powering down anyway
821 		 */
822 
823 		if (ret)
824 			DRM_DEV_ERROR(gmu->dev,
825 				"Unable to slumber GMU: status = 0%x/0%x\n",
826 				gmu_read(gmu,
827 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
828 				gmu_read(gmu,
829 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
830 	}
831 
832 	/* Turn off HFI */
833 	a6xx_hfi_stop(gmu);
834 
835 	/* Stop the interrupts and mask the hardware */
836 	a6xx_gmu_irq_disable(gmu);
837 
838 	/* Tell RPMh to power off the GPU */
839 	a6xx_rpmh_stop(gmu);
840 }
841 
842 
843 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
844 {
845 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
846 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
847 
848 	if (!pm_runtime_active(gmu->dev))
849 		return 0;
850 
851 	/*
852 	 * Force the GMU off if we detected a hang, otherwise try to shut it
853 	 * down gracefully
854 	 */
855 	if (gmu->hung)
856 		a6xx_gmu_force_off(gmu);
857 	else
858 		a6xx_gmu_shutdown(gmu);
859 
860 	/* Remove the bus vote */
861 	icc_set_bw(gpu->icc_path, 0, 0);
862 
863 	/*
864 	 * Make sure the GX domain is off before turning off the GMU (CX)
865 	 * domain. Usually the GMU does this but only if the shutdown sequence
866 	 * was successful
867 	 */
868 	if (!IS_ERR_OR_NULL(gmu->gxpd))
869 		pm_runtime_put_sync(gmu->gxpd);
870 
871 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
872 
873 	pm_runtime_put_sync(gmu->dev);
874 
875 	return 0;
876 }
877 
878 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo)
879 {
880 	int count, i;
881 	u64 iova;
882 
883 	if (IS_ERR_OR_NULL(bo))
884 		return;
885 
886 	count = bo->size >> PAGE_SHIFT;
887 	iova = bo->iova;
888 
889 	for (i = 0; i < count; i++, iova += PAGE_SIZE) {
890 		iommu_unmap(gmu->domain, iova, PAGE_SIZE);
891 		__free_pages(bo->pages[i], 0);
892 	}
893 
894 	kfree(bo->pages);
895 	kfree(bo);
896 }
897 
898 static struct a6xx_gmu_bo *a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu,
899 		size_t size)
900 {
901 	struct a6xx_gmu_bo *bo;
902 	int ret, count, i;
903 
904 	bo = kzalloc(sizeof(*bo), GFP_KERNEL);
905 	if (!bo)
906 		return ERR_PTR(-ENOMEM);
907 
908 	bo->size = PAGE_ALIGN(size);
909 
910 	count = bo->size >> PAGE_SHIFT;
911 
912 	bo->pages = kcalloc(count, sizeof(struct page *), GFP_KERNEL);
913 	if (!bo->pages) {
914 		kfree(bo);
915 		return ERR_PTR(-ENOMEM);
916 	}
917 
918 	for (i = 0; i < count; i++) {
919 		bo->pages[i] = alloc_page(GFP_KERNEL);
920 		if (!bo->pages[i])
921 			goto err;
922 	}
923 
924 	bo->iova = gmu->uncached_iova_base;
925 
926 	for (i = 0; i < count; i++) {
927 		ret = iommu_map(gmu->domain,
928 			bo->iova + (PAGE_SIZE * i),
929 			page_to_phys(bo->pages[i]), PAGE_SIZE,
930 			IOMMU_READ | IOMMU_WRITE);
931 
932 		if (ret) {
933 			DRM_DEV_ERROR(gmu->dev, "Unable to map GMU buffer object\n");
934 
935 			for (i = i - 1 ; i >= 0; i--)
936 				iommu_unmap(gmu->domain,
937 					bo->iova + (PAGE_SIZE * i),
938 					PAGE_SIZE);
939 
940 			goto err;
941 		}
942 	}
943 
944 	bo->virt = vmap(bo->pages, count, VM_IOREMAP,
945 		pgprot_writecombine(PAGE_KERNEL));
946 	if (!bo->virt)
947 		goto err;
948 
949 	/* Align future IOVA addresses on 1MB boundaries */
950 	gmu->uncached_iova_base += ALIGN(size, SZ_1M);
951 
952 	return bo;
953 
954 err:
955 	for (i = 0; i < count; i++) {
956 		if (bo->pages[i])
957 			__free_pages(bo->pages[i], 0);
958 	}
959 
960 	kfree(bo->pages);
961 	kfree(bo);
962 
963 	return ERR_PTR(-ENOMEM);
964 }
965 
966 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
967 {
968 	int ret;
969 
970 	/*
971 	 * The GMU address space is hardcoded to treat the range
972 	 * 0x60000000 - 0x80000000 as un-cached memory. All buffers shared
973 	 * between the GMU and the CPU will live in this space
974 	 */
975 	gmu->uncached_iova_base = 0x60000000;
976 
977 
978 	gmu->domain = iommu_domain_alloc(&platform_bus_type);
979 	if (!gmu->domain)
980 		return -ENODEV;
981 
982 	ret = iommu_attach_device(gmu->domain, gmu->dev);
983 
984 	if (ret) {
985 		iommu_domain_free(gmu->domain);
986 		gmu->domain = NULL;
987 	}
988 
989 	return ret;
990 }
991 
992 /* Return the 'arc-level' for the given frequency */
993 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
994 					   unsigned long freq)
995 {
996 	struct dev_pm_opp *opp;
997 	unsigned int val;
998 
999 	if (!freq)
1000 		return 0;
1001 
1002 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1003 	if (IS_ERR(opp))
1004 		return 0;
1005 
1006 	val = dev_pm_opp_get_level(opp);
1007 
1008 	dev_pm_opp_put(opp);
1009 
1010 	return val;
1011 }
1012 
1013 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1014 		unsigned long *freqs, int freqs_count, const char *id)
1015 {
1016 	int i, j;
1017 	const u16 *pri, *sec;
1018 	size_t pri_count, sec_count;
1019 
1020 	pri = cmd_db_read_aux_data(id, &pri_count);
1021 	if (IS_ERR(pri))
1022 		return PTR_ERR(pri);
1023 	/*
1024 	 * The data comes back as an array of unsigned shorts so adjust the
1025 	 * count accordingly
1026 	 */
1027 	pri_count >>= 1;
1028 	if (!pri_count)
1029 		return -EINVAL;
1030 
1031 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1032 	if (IS_ERR(sec))
1033 		return PTR_ERR(sec);
1034 
1035 	sec_count >>= 1;
1036 	if (!sec_count)
1037 		return -EINVAL;
1038 
1039 	/* Construct a vote for each frequency */
1040 	for (i = 0; i < freqs_count; i++) {
1041 		u8 pindex = 0, sindex = 0;
1042 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1043 
1044 		/* Get the primary index that matches the arc level */
1045 		for (j = 0; j < pri_count; j++) {
1046 			if (pri[j] >= level) {
1047 				pindex = j;
1048 				break;
1049 			}
1050 		}
1051 
1052 		if (j == pri_count) {
1053 			DRM_DEV_ERROR(dev,
1054 				"Level %u not found in in the RPMh list\n",
1055 					level);
1056 			DRM_DEV_ERROR(dev, "Available levels:\n");
1057 			for (j = 0; j < pri_count; j++)
1058 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1059 
1060 			return -EINVAL;
1061 		}
1062 
1063 		/*
1064 		 * Look for a level in in the secondary list that matches. If
1065 		 * nothing fits, use the maximum non zero vote
1066 		 */
1067 
1068 		for (j = 0; j < sec_count; j++) {
1069 			if (sec[j] >= level) {
1070 				sindex = j;
1071 				break;
1072 			} else if (sec[j]) {
1073 				sindex = j;
1074 			}
1075 		}
1076 
1077 		/* Construct the vote */
1078 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1079 			(sindex << 8) | pindex;
1080 	}
1081 
1082 	return 0;
1083 }
1084 
1085 /*
1086  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1087  * to construct the list of votes on the CPU and send it over. Query the RPMh
1088  * voltage levels and build the votes
1089  */
1090 
1091 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1092 {
1093 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1094 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1095 	struct msm_gpu *gpu = &adreno_gpu->base;
1096 	int ret;
1097 
1098 	/* Build the GX votes */
1099 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1100 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1101 
1102 	/* Build the CX votes */
1103 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1104 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1105 
1106 	return ret;
1107 }
1108 
1109 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1110 		u32 size)
1111 {
1112 	int count = dev_pm_opp_get_opp_count(dev);
1113 	struct dev_pm_opp *opp;
1114 	int i, index = 0;
1115 	unsigned long freq = 1;
1116 
1117 	/*
1118 	 * The OPP table doesn't contain the "off" frequency level so we need to
1119 	 * add 1 to the table size to account for it
1120 	 */
1121 
1122 	if (WARN(count + 1 > size,
1123 		"The GMU frequency table is being truncated\n"))
1124 		count = size - 1;
1125 
1126 	/* Set the "off" frequency */
1127 	freqs[index++] = 0;
1128 
1129 	for (i = 0; i < count; i++) {
1130 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1131 		if (IS_ERR(opp))
1132 			break;
1133 
1134 		dev_pm_opp_put(opp);
1135 		freqs[index++] = freq++;
1136 	}
1137 
1138 	return index;
1139 }
1140 
1141 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1142 {
1143 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1144 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1145 	struct msm_gpu *gpu = &adreno_gpu->base;
1146 
1147 	int ret = 0;
1148 
1149 	/*
1150 	 * The GMU handles its own frequency switching so build a list of
1151 	 * available frequencies to send during initialization
1152 	 */
1153 	ret = dev_pm_opp_of_add_table(gmu->dev);
1154 	if (ret) {
1155 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1156 		return ret;
1157 	}
1158 
1159 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1160 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1161 
1162 	/*
1163 	 * The GMU also handles GPU frequency switching so build a list
1164 	 * from the GPU OPP table
1165 	 */
1166 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1167 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1168 
1169 	/* Build the list of RPMh votes that we'll send to the GMU */
1170 	return a6xx_gmu_rpmh_votes_init(gmu);
1171 }
1172 
1173 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1174 {
1175 	int ret = msm_clk_bulk_get(gmu->dev, &gmu->clocks);
1176 
1177 	if (ret < 1)
1178 		return ret;
1179 
1180 	gmu->nr_clocks = ret;
1181 
1182 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1183 		gmu->nr_clocks, "gmu");
1184 
1185 	return 0;
1186 }
1187 
1188 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1189 		const char *name)
1190 {
1191 	void __iomem *ret;
1192 	struct resource *res = platform_get_resource_byname(pdev,
1193 			IORESOURCE_MEM, name);
1194 
1195 	if (!res) {
1196 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1197 		return ERR_PTR(-EINVAL);
1198 	}
1199 
1200 	ret = ioremap(res->start, resource_size(res));
1201 	if (!ret) {
1202 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1203 		return ERR_PTR(-EINVAL);
1204 	}
1205 
1206 	return ret;
1207 }
1208 
1209 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1210 		const char *name, irq_handler_t handler)
1211 {
1212 	int irq, ret;
1213 
1214 	irq = platform_get_irq_byname(pdev, name);
1215 
1216 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1217 	if (ret) {
1218 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1219 			      name, ret);
1220 		return ret;
1221 	}
1222 
1223 	disable_irq(irq);
1224 
1225 	return irq;
1226 }
1227 
1228 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1229 {
1230 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1231 
1232 	if (!gmu->initialized)
1233 		return;
1234 
1235 	pm_runtime_force_suspend(gmu->dev);
1236 
1237 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1238 		pm_runtime_disable(gmu->gxpd);
1239 		dev_pm_domain_detach(gmu->gxpd, false);
1240 	}
1241 
1242 	iounmap(gmu->mmio);
1243 	gmu->mmio = NULL;
1244 
1245 	a6xx_gmu_memory_free(gmu, gmu->hfi);
1246 
1247 	iommu_detach_device(gmu->domain, gmu->dev);
1248 
1249 	iommu_domain_free(gmu->domain);
1250 
1251 	free_irq(gmu->gmu_irq, gmu);
1252 	free_irq(gmu->hfi_irq, gmu);
1253 
1254 	/* Drop reference taken in of_find_device_by_node */
1255 	put_device(gmu->dev);
1256 
1257 	gmu->initialized = false;
1258 }
1259 
1260 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1261 {
1262 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1263 	struct platform_device *pdev = of_find_device_by_node(node);
1264 	int ret;
1265 
1266 	if (!pdev)
1267 		return -ENODEV;
1268 
1269 	gmu->dev = &pdev->dev;
1270 
1271 	of_dma_configure(gmu->dev, node, true);
1272 
1273 	/* Fow now, don't do anything fancy until we get our feet under us */
1274 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1275 
1276 	pm_runtime_enable(gmu->dev);
1277 
1278 	/* Get the list of clocks */
1279 	ret = a6xx_gmu_clocks_probe(gmu);
1280 	if (ret)
1281 		goto err_put_device;
1282 
1283 	/* Set up the IOMMU context bank */
1284 	ret = a6xx_gmu_memory_probe(gmu);
1285 	if (ret)
1286 		goto err_put_device;
1287 
1288 	/* Allocate memory for for the HFI queues */
1289 	gmu->hfi = a6xx_gmu_memory_alloc(gmu, SZ_16K);
1290 	if (IS_ERR(gmu->hfi))
1291 		goto err_memory;
1292 
1293 	/* Allocate memory for the GMU debug region */
1294 	gmu->debug = a6xx_gmu_memory_alloc(gmu, SZ_16K);
1295 	if (IS_ERR(gmu->debug))
1296 		goto err_memory;
1297 
1298 	/* Map the GMU registers */
1299 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1300 	if (IS_ERR(gmu->mmio))
1301 		goto err_memory;
1302 
1303 	/* Get the HFI and GMU interrupts */
1304 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1305 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1306 
1307 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1308 		goto err_mmio;
1309 
1310 	/*
1311 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1312 	 * crash
1313 	 */
1314 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1315 
1316 	/* Get the power levels for the GMU and GPU */
1317 	a6xx_gmu_pwrlevels_probe(gmu);
1318 
1319 	/* Set up the HFI queues */
1320 	a6xx_hfi_init(gmu);
1321 
1322 	gmu->initialized = true;
1323 
1324 	return 0;
1325 
1326 err_mmio:
1327 	iounmap(gmu->mmio);
1328 	free_irq(gmu->gmu_irq, gmu);
1329 	free_irq(gmu->hfi_irq, gmu);
1330 err_memory:
1331 	a6xx_gmu_memory_free(gmu, gmu->hfi);
1332 
1333 	if (gmu->domain) {
1334 		iommu_detach_device(gmu->domain, gmu->dev);
1335 
1336 		iommu_domain_free(gmu->domain);
1337 	}
1338 	ret = -ENODEV;
1339 
1340 err_put_device:
1341 	/* Drop reference taken in of_find_device_by_node */
1342 	put_device(gmu->dev);
1343 
1344 	return ret;
1345 }
1346