xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision 6abeae2a)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright (c) 2017-2019 The Linux Foundation. All rights reserved. */
3 
4 #include <linux/clk.h>
5 #include <linux/interconnect.h>
6 #include <linux/pm_domain.h>
7 #include <linux/pm_opp.h>
8 #include <soc/qcom/cmd-db.h>
9 #include <drm/drm_gem.h>
10 
11 #include "a6xx_gpu.h"
12 #include "a6xx_gmu.xml.h"
13 #include "msm_gem.h"
14 #include "msm_gpu_trace.h"
15 #include "msm_mmu.h"
16 
17 static void a6xx_gmu_fault(struct a6xx_gmu *gmu)
18 {
19 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
20 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
21 	struct msm_gpu *gpu = &adreno_gpu->base;
22 
23 	/* FIXME: add a banner here */
24 	gmu->hung = true;
25 
26 	/* Turn off the hangcheck timer while we are resetting */
27 	del_timer(&gpu->hangcheck_timer);
28 
29 	/* Queue the GPU handler because we need to treat this as a recovery */
30 	kthread_queue_work(gpu->worker, &gpu->recover_work);
31 }
32 
33 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
34 {
35 	struct a6xx_gmu *gmu = data;
36 	u32 status;
37 
38 	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
39 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
40 
41 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
42 		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
43 
44 		a6xx_gmu_fault(gmu);
45 	}
46 
47 	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
48 		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
49 
50 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
51 		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
52 			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
53 
54 	return IRQ_HANDLED;
55 }
56 
57 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
58 {
59 	struct a6xx_gmu *gmu = data;
60 	u32 status;
61 
62 	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
63 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
64 
65 	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
66 		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
67 
68 		a6xx_gmu_fault(gmu);
69 	}
70 
71 	return IRQ_HANDLED;
72 }
73 
74 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
75 {
76 	u32 val;
77 
78 	/* This can be called from gpu state code so make sure GMU is valid */
79 	if (!gmu->initialized)
80 		return false;
81 
82 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
83 
84 	return !(val &
85 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
86 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
87 }
88 
89 /* Check to see if the GX rail is still powered */
90 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
91 {
92 	u32 val;
93 
94 	/* This can be called from gpu state code so make sure GMU is valid */
95 	if (!gmu->initialized)
96 		return false;
97 
98 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
99 
100 	return !(val &
101 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
102 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
103 }
104 
105 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
106 {
107 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
108 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
109 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
110 	u32 perf_index;
111 	unsigned long gpu_freq;
112 	int ret = 0;
113 
114 	gpu_freq = dev_pm_opp_get_freq(opp);
115 
116 	if (gpu_freq == gmu->freq)
117 		return;
118 
119 	for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
120 		if (gpu_freq == gmu->gpu_freqs[perf_index])
121 			break;
122 
123 	gmu->current_perf_index = perf_index;
124 	gmu->freq = gmu->gpu_freqs[perf_index];
125 
126 	trace_msm_gmu_freq_change(gmu->freq, perf_index);
127 
128 	/*
129 	 * This can get called from devfreq while the hardware is idle. Don't
130 	 * bring up the power if it isn't already active
131 	 */
132 	if (pm_runtime_get_if_in_use(gmu->dev) == 0)
133 		return;
134 
135 	if (!gmu->legacy) {
136 		a6xx_hfi_set_freq(gmu, perf_index);
137 		dev_pm_opp_set_bw(&gpu->pdev->dev, opp);
138 		pm_runtime_put(gmu->dev);
139 		return;
140 	}
141 
142 	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
143 
144 	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
145 			((3 & 0xf) << 28) | perf_index);
146 
147 	/*
148 	 * Send an invalid index as a vote for the bus bandwidth and let the
149 	 * firmware decide on the right vote
150 	 */
151 	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
152 
153 	/* Set and clear the OOB for DCVS to trigger the GMU */
154 	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
155 	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
156 
157 	ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
158 	if (ret)
159 		dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
160 
161 	dev_pm_opp_set_bw(&gpu->pdev->dev, opp);
162 	pm_runtime_put(gmu->dev);
163 }
164 
165 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
166 {
167 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
168 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
169 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
170 
171 	return  gmu->freq;
172 }
173 
174 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
175 {
176 	u32 val;
177 	int local = gmu->idle_level;
178 
179 	/* SPTP and IFPC both report as IFPC */
180 	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
181 		local = GMU_IDLE_STATE_IFPC;
182 
183 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
184 
185 	if (val == local) {
186 		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
187 			!a6xx_gmu_gx_is_on(gmu))
188 			return true;
189 	}
190 
191 	return false;
192 }
193 
194 /* Wait for the GMU to get to its most idle state */
195 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
196 {
197 	return spin_until(a6xx_gmu_check_idle_level(gmu));
198 }
199 
200 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
201 {
202 	int ret;
203 	u32 val;
204 	u32 mask, reset_val;
205 
206 	val = gmu_read(gmu, REG_A6XX_GMU_CM3_DTCM_START + 0xff8);
207 	if (val <= 0x20010004) {
208 		mask = 0xffffffff;
209 		reset_val = 0xbabeface;
210 	} else {
211 		mask = 0x1ff;
212 		reset_val = 0x100;
213 	}
214 
215 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
216 
217 	/* Set the log wptr index
218 	 * note: downstream saves the value in poweroff and restores it here
219 	 */
220 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_RESP, 0);
221 
222 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
223 
224 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
225 		(val & mask) == reset_val, 100, 10000);
226 
227 	if (ret)
228 		DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
229 
230 	return ret;
231 }
232 
233 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
234 {
235 	u32 val;
236 	int ret;
237 
238 	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
239 
240 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
241 		val & 1, 100, 10000);
242 	if (ret)
243 		DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
244 
245 	return ret;
246 }
247 
248 /* Trigger a OOB (out of band) request to the GMU */
249 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
250 {
251 	int ret;
252 	u32 val;
253 	int request, ack;
254 	const char *name;
255 
256 	switch (state) {
257 	case GMU_OOB_GPU_SET:
258 		if (gmu->legacy) {
259 			request = GMU_OOB_GPU_SET_REQUEST;
260 			ack = GMU_OOB_GPU_SET_ACK;
261 		} else {
262 			request = GMU_OOB_GPU_SET_REQUEST_NEW;
263 			ack = GMU_OOB_GPU_SET_ACK_NEW;
264 		}
265 		name = "GPU_SET";
266 		break;
267 	case GMU_OOB_BOOT_SLUMBER:
268 		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
269 		ack = GMU_OOB_BOOT_SLUMBER_ACK;
270 		name = "BOOT_SLUMBER";
271 		break;
272 	case GMU_OOB_DCVS_SET:
273 		request = GMU_OOB_DCVS_REQUEST;
274 		ack = GMU_OOB_DCVS_ACK;
275 		name = "GPU_DCVS";
276 		break;
277 	default:
278 		return -EINVAL;
279 	}
280 
281 	/* Trigger the equested OOB operation */
282 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
283 
284 	/* Wait for the acknowledge interrupt */
285 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
286 		val & (1 << ack), 100, 10000);
287 
288 	if (ret)
289 		DRM_DEV_ERROR(gmu->dev,
290 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
291 				name,
292 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
293 
294 	/* Clear the acknowledge interrupt */
295 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
296 
297 	return ret;
298 }
299 
300 /* Clear a pending OOB state in the GMU */
301 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
302 {
303 	if (!gmu->legacy) {
304 		WARN_ON(state != GMU_OOB_GPU_SET);
305 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
306 			1 << GMU_OOB_GPU_SET_CLEAR_NEW);
307 		return;
308 	}
309 
310 	switch (state) {
311 	case GMU_OOB_GPU_SET:
312 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
313 			1 << GMU_OOB_GPU_SET_CLEAR);
314 		break;
315 	case GMU_OOB_BOOT_SLUMBER:
316 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
317 			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
318 		break;
319 	case GMU_OOB_DCVS_SET:
320 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
321 			1 << GMU_OOB_DCVS_CLEAR);
322 		break;
323 	}
324 }
325 
326 /* Enable CPU control of SPTP power power collapse */
327 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
328 {
329 	int ret;
330 	u32 val;
331 
332 	if (!gmu->legacy)
333 		return 0;
334 
335 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
336 
337 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
338 		(val & 0x38) == 0x28, 1, 100);
339 
340 	if (ret) {
341 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
342 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
343 	}
344 
345 	return 0;
346 }
347 
348 /* Disable CPU control of SPTP power power collapse */
349 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
350 {
351 	u32 val;
352 	int ret;
353 
354 	if (!gmu->legacy)
355 		return;
356 
357 	/* Make sure retention is on */
358 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
359 
360 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
361 
362 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
363 		(val & 0x04), 100, 10000);
364 
365 	if (ret)
366 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
367 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
368 }
369 
370 /* Let the GMU know we are starting a boot sequence */
371 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
372 {
373 	u32 vote;
374 
375 	/* Let the GMU know we are getting ready for boot */
376 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
377 
378 	/* Choose the "default" power level as the highest available */
379 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
380 
381 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
382 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
383 
384 	/* Let the GMU know the boot sequence has started */
385 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
386 }
387 
388 /* Let the GMU know that we are about to go into slumber */
389 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
390 {
391 	int ret;
392 
393 	/* Disable the power counter so the GMU isn't busy */
394 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
395 
396 	/* Disable SPTP_PC if the CPU is responsible for it */
397 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
398 		a6xx_sptprac_disable(gmu);
399 
400 	if (!gmu->legacy) {
401 		ret = a6xx_hfi_send_prep_slumber(gmu);
402 		goto out;
403 	}
404 
405 	/* Tell the GMU to get ready to slumber */
406 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
407 
408 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
409 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
410 
411 	if (!ret) {
412 		/* Check to see if the GMU really did slumber */
413 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
414 			!= 0x0f) {
415 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
416 			ret = -ETIMEDOUT;
417 		}
418 	}
419 
420 out:
421 	/* Put fence into allow mode */
422 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
423 	return ret;
424 }
425 
426 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
427 {
428 	int ret;
429 	u32 val;
430 
431 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
432 	/* Wait for the register to finish posting */
433 	wmb();
434 
435 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
436 		val & (1 << 1), 100, 10000);
437 	if (ret) {
438 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
439 		return ret;
440 	}
441 
442 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
443 		!val, 100, 10000);
444 
445 	if (ret) {
446 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
447 		return ret;
448 	}
449 
450 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
451 
452 	/* Set up CX GMU counter 0 to count busy ticks */
453 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
454 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
455 
456 	/* Enable the power counter */
457 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
458 	return 0;
459 }
460 
461 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
462 {
463 	int ret;
464 	u32 val;
465 
466 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
467 
468 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
469 		val, val & (1 << 16), 100, 10000);
470 	if (ret)
471 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
472 
473 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
474 }
475 
476 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
477 {
478 	return msm_writel(value, ptr + (offset << 2));
479 }
480 
481 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
482 		const char *name);
483 
484 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
485 {
486 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
487 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
488 	struct platform_device *pdev = to_platform_device(gmu->dev);
489 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
490 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
491 	uint32_t pdc_address_offset;
492 
493 	if (!pdcptr || !seqptr)
494 		goto err;
495 
496 	if (adreno_is_a618(adreno_gpu) || adreno_is_a640(adreno_gpu))
497 		pdc_address_offset = 0x30090;
498 	else if (adreno_is_a650(adreno_gpu))
499 		pdc_address_offset = 0x300a0;
500 	else
501 		pdc_address_offset = 0x30080;
502 
503 	/* Disable SDE clock gating */
504 	gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
505 
506 	/* Setup RSC PDC handshake for sleep and wakeup */
507 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
508 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
509 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
510 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
511 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
512 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
513 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
514 	gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
515 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
516 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
517 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
518 
519 	/* Load RSC sequencer uCode for sleep and wakeup */
520 	if (adreno_is_a650(adreno_gpu)) {
521 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
522 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
523 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
524 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
525 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
526 	} else {
527 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
528 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
529 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
530 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
531 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
532 	}
533 
534 	/* Load PDC sequencer uCode for power up and power down sequence */
535 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
536 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
537 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
538 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
539 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
540 
541 	/* Set TCS commands used by PDC sequence for low power modes */
542 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
543 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
544 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
545 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
546 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
547 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
548 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
549 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
550 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
551 
552 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
553 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
554 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
555 
556 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
557 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
558 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
559 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
560 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
561 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
562 
563 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
564 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
565 	if (adreno_is_a618(adreno_gpu) || adreno_is_a650(adreno_gpu))
566 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
567 	else
568 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
569 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
570 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
571 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
572 
573 	/* Setup GPU PDC */
574 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
575 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
576 
577 	/* ensure no writes happen before the uCode is fully written */
578 	wmb();
579 
580 err:
581 	if (!IS_ERR_OR_NULL(pdcptr))
582 		iounmap(pdcptr);
583 	if (!IS_ERR_OR_NULL(seqptr))
584 		iounmap(seqptr);
585 }
586 
587 /*
588  * The lowest 16 bits of this value are the number of XO clock cycles for main
589  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
590  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
591  */
592 
593 #define GMU_PWR_COL_HYST 0x000a1680
594 
595 /* Set up the idle state for the GMU */
596 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
597 {
598 	/* Disable GMU WB/RB buffer */
599 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
600 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
601 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
602 
603 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
604 
605 	switch (gmu->idle_level) {
606 	case GMU_IDLE_STATE_IFPC:
607 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
608 			GMU_PWR_COL_HYST);
609 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
610 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
611 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
612 		fallthrough;
613 	case GMU_IDLE_STATE_SPTP:
614 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
615 			GMU_PWR_COL_HYST);
616 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
617 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
618 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
619 	}
620 
621 	/* Enable RPMh GPU client */
622 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
623 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
624 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
625 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
626 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
627 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
628 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
629 }
630 
631 struct block_header {
632 	u32 addr;
633 	u32 size;
634 	u32 type;
635 	u32 value;
636 	u32 data[];
637 };
638 
639 /* this should be a general kernel helper */
640 static int in_range(u32 addr, u32 start, u32 size)
641 {
642 	return addr >= start && addr < start + size;
643 }
644 
645 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
646 {
647 	if (!in_range(blk->addr, bo->iova, bo->size))
648 		return false;
649 
650 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
651 	return true;
652 }
653 
654 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
655 {
656 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
657 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
658 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
659 	const struct block_header *blk;
660 	u32 reg_offset;
661 
662 	u32 itcm_base = 0x00000000;
663 	u32 dtcm_base = 0x00040000;
664 
665 	if (adreno_is_a650(adreno_gpu))
666 		dtcm_base = 0x10004000;
667 
668 	if (gmu->legacy) {
669 		/* Sanity check the size of the firmware that was loaded */
670 		if (fw_image->size > 0x8000) {
671 			DRM_DEV_ERROR(gmu->dev,
672 				"GMU firmware is bigger than the available region\n");
673 			return -EINVAL;
674 		}
675 
676 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
677 			       (u32*) fw_image->data, fw_image->size);
678 		return 0;
679 	}
680 
681 
682 	for (blk = (const struct block_header *) fw_image->data;
683 	     (const u8*) blk < fw_image->data + fw_image->size;
684 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
685 		if (blk->size == 0)
686 			continue;
687 
688 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
689 			reg_offset = (blk->addr - itcm_base) >> 2;
690 			gmu_write_bulk(gmu,
691 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
692 				blk->data, blk->size);
693 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
694 			reg_offset = (blk->addr - dtcm_base) >> 2;
695 			gmu_write_bulk(gmu,
696 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
697 				blk->data, blk->size);
698 		} else if (!fw_block_mem(&gmu->icache, blk) &&
699 			   !fw_block_mem(&gmu->dcache, blk) &&
700 			   !fw_block_mem(&gmu->dummy, blk)) {
701 			DRM_DEV_ERROR(gmu->dev,
702 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
703 				blk->addr, blk->size, blk->data[0]);
704 		}
705 	}
706 
707 	return 0;
708 }
709 
710 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
711 {
712 	static bool rpmh_init;
713 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
714 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
715 	int ret;
716 	u32 chipid;
717 
718 	if (adreno_is_a650(adreno_gpu))
719 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
720 
721 	if (state == GMU_WARM_BOOT) {
722 		ret = a6xx_rpmh_start(gmu);
723 		if (ret)
724 			return ret;
725 	} else {
726 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
727 			"GMU firmware is not loaded\n"))
728 			return -ENOENT;
729 
730 		/* Turn on register retention */
731 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
732 
733 		/* We only need to load the RPMh microcode once */
734 		if (!rpmh_init) {
735 			a6xx_gmu_rpmh_init(gmu);
736 			rpmh_init = true;
737 		} else {
738 			ret = a6xx_rpmh_start(gmu);
739 			if (ret)
740 				return ret;
741 		}
742 
743 		ret = a6xx_gmu_fw_load(gmu);
744 		if (ret)
745 			return ret;
746 	}
747 
748 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
749 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
750 
751 	/* Write the iova of the HFI table */
752 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
753 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
754 
755 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
756 		(1 << 31) | (0xa << 18) | (0xa0));
757 
758 	chipid = adreno_gpu->rev.core << 24;
759 	chipid |= adreno_gpu->rev.major << 16;
760 	chipid |= adreno_gpu->rev.minor << 12;
761 	chipid |= adreno_gpu->rev.patchid << 8;
762 
763 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
764 
765 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
766 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
767 
768 	/* Set up the lowest idle level on the GMU */
769 	a6xx_gmu_power_config(gmu);
770 
771 	ret = a6xx_gmu_start(gmu);
772 	if (ret)
773 		return ret;
774 
775 	if (gmu->legacy) {
776 		ret = a6xx_gmu_gfx_rail_on(gmu);
777 		if (ret)
778 			return ret;
779 	}
780 
781 	/* Enable SPTP_PC if the CPU is responsible for it */
782 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
783 		ret = a6xx_sptprac_enable(gmu);
784 		if (ret)
785 			return ret;
786 	}
787 
788 	ret = a6xx_gmu_hfi_start(gmu);
789 	if (ret)
790 		return ret;
791 
792 	/* FIXME: Do we need this wmb() here? */
793 	wmb();
794 
795 	return 0;
796 }
797 
798 #define A6XX_HFI_IRQ_MASK \
799 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
800 
801 #define A6XX_GMU_IRQ_MASK \
802 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
803 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
804 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
805 
806 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
807 {
808 	disable_irq(gmu->gmu_irq);
809 	disable_irq(gmu->hfi_irq);
810 
811 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
812 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
813 }
814 
815 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
816 {
817 	u32 val;
818 
819 	/* Make sure there are no outstanding RPMh votes */
820 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
821 		(val & 1), 100, 10000);
822 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
823 		(val & 1), 100, 10000);
824 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
825 		(val & 1), 100, 10000);
826 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
827 		(val & 1), 100, 1000);
828 }
829 
830 /* Force the GMU off in case it isn't responsive */
831 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
832 {
833 	/* Flush all the queues */
834 	a6xx_hfi_stop(gmu);
835 
836 	/* Stop the interrupts */
837 	a6xx_gmu_irq_disable(gmu);
838 
839 	/* Force off SPTP in case the GMU is managing it */
840 	a6xx_sptprac_disable(gmu);
841 
842 	/* Make sure there are no outstanding RPMh votes */
843 	a6xx_gmu_rpmh_off(gmu);
844 }
845 
846 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
847 {
848 	struct dev_pm_opp *gpu_opp;
849 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
850 
851 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
852 	if (IS_ERR_OR_NULL(gpu_opp))
853 		return;
854 
855 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
856 	a6xx_gmu_set_freq(gpu, gpu_opp);
857 	dev_pm_opp_put(gpu_opp);
858 }
859 
860 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
861 {
862 	struct dev_pm_opp *gpu_opp;
863 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
864 
865 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
866 	if (IS_ERR_OR_NULL(gpu_opp))
867 		return;
868 
869 	dev_pm_opp_set_bw(&gpu->pdev->dev, gpu_opp);
870 	dev_pm_opp_put(gpu_opp);
871 }
872 
873 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
874 {
875 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
876 	struct msm_gpu *gpu = &adreno_gpu->base;
877 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
878 	int status, ret;
879 
880 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
881 		return 0;
882 
883 	gmu->hung = false;
884 
885 	/* Turn on the resources */
886 	pm_runtime_get_sync(gmu->dev);
887 
888 	/*
889 	 * "enable" the GX power domain which won't actually do anything but it
890 	 * will make sure that the refcounting is correct in case we need to
891 	 * bring down the GX after a GMU failure
892 	 */
893 	if (!IS_ERR_OR_NULL(gmu->gxpd))
894 		pm_runtime_get_sync(gmu->gxpd);
895 
896 	/* Use a known rate to bring up the GMU */
897 	clk_set_rate(gmu->core_clk, 200000000);
898 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
899 	if (ret) {
900 		pm_runtime_put(gmu->gxpd);
901 		pm_runtime_put(gmu->dev);
902 		return ret;
903 	}
904 
905 	/* Set the bus quota to a reasonable value for boot */
906 	a6xx_gmu_set_initial_bw(gpu, gmu);
907 
908 	/* Enable the GMU interrupt */
909 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
910 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
911 	enable_irq(gmu->gmu_irq);
912 
913 	/* Check to see if we are doing a cold or warm boot */
914 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
915 		GMU_WARM_BOOT : GMU_COLD_BOOT;
916 
917 	/*
918 	 * Warm boot path does not work on newer GPUs
919 	 * Presumably this is because icache/dcache regions must be restored
920 	 */
921 	if (!gmu->legacy)
922 		status = GMU_COLD_BOOT;
923 
924 	ret = a6xx_gmu_fw_start(gmu, status);
925 	if (ret)
926 		goto out;
927 
928 	ret = a6xx_hfi_start(gmu, status);
929 	if (ret)
930 		goto out;
931 
932 	/*
933 	 * Turn on the GMU firmware fault interrupt after we know the boot
934 	 * sequence is successful
935 	 */
936 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
937 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
938 	enable_irq(gmu->hfi_irq);
939 
940 	/* Set the GPU to the current freq */
941 	a6xx_gmu_set_initial_freq(gpu, gmu);
942 
943 out:
944 	/* On failure, shut down the GMU to leave it in a good state */
945 	if (ret) {
946 		disable_irq(gmu->gmu_irq);
947 		a6xx_rpmh_stop(gmu);
948 		pm_runtime_put(gmu->gxpd);
949 		pm_runtime_put(gmu->dev);
950 	}
951 
952 	return ret;
953 }
954 
955 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
956 {
957 	u32 reg;
958 
959 	if (!gmu->initialized)
960 		return true;
961 
962 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
963 
964 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
965 		return false;
966 
967 	return true;
968 }
969 
970 #define GBIF_CLIENT_HALT_MASK             BIT(0)
971 #define GBIF_ARB_HALT_MASK                BIT(1)
972 
973 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
974 {
975 	struct msm_gpu *gpu = &adreno_gpu->base;
976 
977 	if (!a6xx_has_gbif(adreno_gpu)) {
978 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
979 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
980 								0xf) == 0xf);
981 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
982 
983 		return;
984 	}
985 
986 	/* Halt new client requests on GBIF */
987 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
988 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
989 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
990 
991 	/* Halt all AXI requests on GBIF */
992 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
993 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
994 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
995 
996 	/* The GBIF halt needs to be explicitly cleared */
997 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
998 }
999 
1000 /* Gracefully try to shut down the GMU and by extension the GPU */
1001 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1002 {
1003 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1004 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1005 	u32 val;
1006 
1007 	/*
1008 	 * The GMU may still be in slumber unless the GPU started so check and
1009 	 * skip putting it back into slumber if so
1010 	 */
1011 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1012 
1013 	if (val != 0xf) {
1014 		int ret = a6xx_gmu_wait_for_idle(gmu);
1015 
1016 		/* If the GMU isn't responding assume it is hung */
1017 		if (ret) {
1018 			a6xx_gmu_force_off(gmu);
1019 			return;
1020 		}
1021 
1022 		a6xx_bus_clear_pending_transactions(adreno_gpu);
1023 
1024 		/* tell the GMU we want to slumber */
1025 		a6xx_gmu_notify_slumber(gmu);
1026 
1027 		ret = gmu_poll_timeout(gmu,
1028 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1029 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1030 			100, 10000);
1031 
1032 		/*
1033 		 * Let the user know we failed to slumber but don't worry too
1034 		 * much because we are powering down anyway
1035 		 */
1036 
1037 		if (ret)
1038 			DRM_DEV_ERROR(gmu->dev,
1039 				"Unable to slumber GMU: status = 0%x/0%x\n",
1040 				gmu_read(gmu,
1041 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1042 				gmu_read(gmu,
1043 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1044 	}
1045 
1046 	/* Turn off HFI */
1047 	a6xx_hfi_stop(gmu);
1048 
1049 	/* Stop the interrupts and mask the hardware */
1050 	a6xx_gmu_irq_disable(gmu);
1051 
1052 	/* Tell RPMh to power off the GPU */
1053 	a6xx_rpmh_stop(gmu);
1054 }
1055 
1056 
1057 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1058 {
1059 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1060 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1061 
1062 	if (!pm_runtime_active(gmu->dev))
1063 		return 0;
1064 
1065 	/*
1066 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1067 	 * down gracefully
1068 	 */
1069 	if (gmu->hung)
1070 		a6xx_gmu_force_off(gmu);
1071 	else
1072 		a6xx_gmu_shutdown(gmu);
1073 
1074 	/* Remove the bus vote */
1075 	dev_pm_opp_set_bw(&gpu->pdev->dev, NULL);
1076 
1077 	/*
1078 	 * Make sure the GX domain is off before turning off the GMU (CX)
1079 	 * domain. Usually the GMU does this but only if the shutdown sequence
1080 	 * was successful
1081 	 */
1082 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1083 		pm_runtime_put_sync(gmu->gxpd);
1084 
1085 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1086 
1087 	pm_runtime_put_sync(gmu->dev);
1088 
1089 	return 0;
1090 }
1091 
1092 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1093 {
1094 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace, false);
1095 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace, false);
1096 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace, false);
1097 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace, false);
1098 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace, false);
1099 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace, false);
1100 
1101 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1102 	msm_gem_address_space_put(gmu->aspace);
1103 }
1104 
1105 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1106 		size_t size, u64 iova)
1107 {
1108 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1109 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1110 	uint32_t flags = MSM_BO_WC;
1111 	u64 range_start, range_end;
1112 	int ret;
1113 
1114 	size = PAGE_ALIGN(size);
1115 	if (!iova) {
1116 		/* no fixed address - use GMU's uncached range */
1117 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1118 		range_end = 0x80000000;
1119 	} else {
1120 		/* range for fixed address */
1121 		range_start = iova;
1122 		range_end = iova + size;
1123 		/* use IOMMU_PRIV for icache/dcache */
1124 		flags |= MSM_BO_MAP_PRIV;
1125 	}
1126 
1127 	bo->obj = msm_gem_new(dev, size, flags);
1128 	if (IS_ERR(bo->obj))
1129 		return PTR_ERR(bo->obj);
1130 
1131 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1132 		range_start >> PAGE_SHIFT, range_end >> PAGE_SHIFT);
1133 	if (ret) {
1134 		drm_gem_object_put(bo->obj);
1135 		return ret;
1136 	}
1137 
1138 	bo->virt = msm_gem_get_vaddr(bo->obj);
1139 	bo->size = size;
1140 
1141 	return 0;
1142 }
1143 
1144 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1145 {
1146 	struct iommu_domain *domain;
1147 	struct msm_mmu *mmu;
1148 
1149 	domain = iommu_domain_alloc(&platform_bus_type);
1150 	if (!domain)
1151 		return -ENODEV;
1152 
1153 	mmu = msm_iommu_new(gmu->dev, domain);
1154 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1155 	if (IS_ERR(gmu->aspace)) {
1156 		iommu_domain_free(domain);
1157 		return PTR_ERR(gmu->aspace);
1158 	}
1159 
1160 	return 0;
1161 }
1162 
1163 /* Return the 'arc-level' for the given frequency */
1164 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1165 					   unsigned long freq)
1166 {
1167 	struct dev_pm_opp *opp;
1168 	unsigned int val;
1169 
1170 	if (!freq)
1171 		return 0;
1172 
1173 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1174 	if (IS_ERR(opp))
1175 		return 0;
1176 
1177 	val = dev_pm_opp_get_level(opp);
1178 
1179 	dev_pm_opp_put(opp);
1180 
1181 	return val;
1182 }
1183 
1184 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1185 		unsigned long *freqs, int freqs_count, const char *id)
1186 {
1187 	int i, j;
1188 	const u16 *pri, *sec;
1189 	size_t pri_count, sec_count;
1190 
1191 	pri = cmd_db_read_aux_data(id, &pri_count);
1192 	if (IS_ERR(pri))
1193 		return PTR_ERR(pri);
1194 	/*
1195 	 * The data comes back as an array of unsigned shorts so adjust the
1196 	 * count accordingly
1197 	 */
1198 	pri_count >>= 1;
1199 	if (!pri_count)
1200 		return -EINVAL;
1201 
1202 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1203 	if (IS_ERR(sec))
1204 		return PTR_ERR(sec);
1205 
1206 	sec_count >>= 1;
1207 	if (!sec_count)
1208 		return -EINVAL;
1209 
1210 	/* Construct a vote for each frequency */
1211 	for (i = 0; i < freqs_count; i++) {
1212 		u8 pindex = 0, sindex = 0;
1213 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1214 
1215 		/* Get the primary index that matches the arc level */
1216 		for (j = 0; j < pri_count; j++) {
1217 			if (pri[j] >= level) {
1218 				pindex = j;
1219 				break;
1220 			}
1221 		}
1222 
1223 		if (j == pri_count) {
1224 			DRM_DEV_ERROR(dev,
1225 				      "Level %u not found in the RPMh list\n",
1226 				      level);
1227 			DRM_DEV_ERROR(dev, "Available levels:\n");
1228 			for (j = 0; j < pri_count; j++)
1229 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1230 
1231 			return -EINVAL;
1232 		}
1233 
1234 		/*
1235 		 * Look for a level in in the secondary list that matches. If
1236 		 * nothing fits, use the maximum non zero vote
1237 		 */
1238 
1239 		for (j = 0; j < sec_count; j++) {
1240 			if (sec[j] >= level) {
1241 				sindex = j;
1242 				break;
1243 			} else if (sec[j]) {
1244 				sindex = j;
1245 			}
1246 		}
1247 
1248 		/* Construct the vote */
1249 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1250 			(sindex << 8) | pindex;
1251 	}
1252 
1253 	return 0;
1254 }
1255 
1256 /*
1257  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1258  * to construct the list of votes on the CPU and send it over. Query the RPMh
1259  * voltage levels and build the votes
1260  */
1261 
1262 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1263 {
1264 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1265 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1266 	struct msm_gpu *gpu = &adreno_gpu->base;
1267 	int ret;
1268 
1269 	/* Build the GX votes */
1270 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1271 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1272 
1273 	/* Build the CX votes */
1274 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1275 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1276 
1277 	return ret;
1278 }
1279 
1280 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1281 		u32 size)
1282 {
1283 	int count = dev_pm_opp_get_opp_count(dev);
1284 	struct dev_pm_opp *opp;
1285 	int i, index = 0;
1286 	unsigned long freq = 1;
1287 
1288 	/*
1289 	 * The OPP table doesn't contain the "off" frequency level so we need to
1290 	 * add 1 to the table size to account for it
1291 	 */
1292 
1293 	if (WARN(count + 1 > size,
1294 		"The GMU frequency table is being truncated\n"))
1295 		count = size - 1;
1296 
1297 	/* Set the "off" frequency */
1298 	freqs[index++] = 0;
1299 
1300 	for (i = 0; i < count; i++) {
1301 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1302 		if (IS_ERR(opp))
1303 			break;
1304 
1305 		dev_pm_opp_put(opp);
1306 		freqs[index++] = freq++;
1307 	}
1308 
1309 	return index;
1310 }
1311 
1312 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1313 {
1314 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1315 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1316 	struct msm_gpu *gpu = &adreno_gpu->base;
1317 
1318 	int ret = 0;
1319 
1320 	/*
1321 	 * The GMU handles its own frequency switching so build a list of
1322 	 * available frequencies to send during initialization
1323 	 */
1324 	ret = dev_pm_opp_of_add_table(gmu->dev);
1325 	if (ret) {
1326 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1327 		return ret;
1328 	}
1329 
1330 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1331 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1332 
1333 	/*
1334 	 * The GMU also handles GPU frequency switching so build a list
1335 	 * from the GPU OPP table
1336 	 */
1337 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1338 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1339 
1340 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1341 
1342 	/* Build the list of RPMh votes that we'll send to the GMU */
1343 	return a6xx_gmu_rpmh_votes_init(gmu);
1344 }
1345 
1346 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1347 {
1348 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1349 
1350 	if (ret < 1)
1351 		return ret;
1352 
1353 	gmu->nr_clocks = ret;
1354 
1355 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1356 		gmu->nr_clocks, "gmu");
1357 
1358 	return 0;
1359 }
1360 
1361 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1362 		const char *name)
1363 {
1364 	void __iomem *ret;
1365 	struct resource *res = platform_get_resource_byname(pdev,
1366 			IORESOURCE_MEM, name);
1367 
1368 	if (!res) {
1369 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1370 		return ERR_PTR(-EINVAL);
1371 	}
1372 
1373 	ret = ioremap(res->start, resource_size(res));
1374 	if (!ret) {
1375 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1376 		return ERR_PTR(-EINVAL);
1377 	}
1378 
1379 	return ret;
1380 }
1381 
1382 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1383 		const char *name, irq_handler_t handler)
1384 {
1385 	int irq, ret;
1386 
1387 	irq = platform_get_irq_byname(pdev, name);
1388 
1389 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1390 	if (ret) {
1391 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1392 			      name, ret);
1393 		return ret;
1394 	}
1395 
1396 	disable_irq(irq);
1397 
1398 	return irq;
1399 }
1400 
1401 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1402 {
1403 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1404 	struct platform_device *pdev = to_platform_device(gmu->dev);
1405 
1406 	if (!gmu->initialized)
1407 		return;
1408 
1409 	pm_runtime_force_suspend(gmu->dev);
1410 
1411 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1412 		pm_runtime_disable(gmu->gxpd);
1413 		dev_pm_domain_detach(gmu->gxpd, false);
1414 	}
1415 
1416 	iounmap(gmu->mmio);
1417 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1418 		iounmap(gmu->rscc);
1419 	gmu->mmio = NULL;
1420 	gmu->rscc = NULL;
1421 
1422 	a6xx_gmu_memory_free(gmu);
1423 
1424 	free_irq(gmu->gmu_irq, gmu);
1425 	free_irq(gmu->hfi_irq, gmu);
1426 
1427 	/* Drop reference taken in of_find_device_by_node */
1428 	put_device(gmu->dev);
1429 
1430 	gmu->initialized = false;
1431 }
1432 
1433 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1434 {
1435 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1436 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1437 	struct platform_device *pdev = of_find_device_by_node(node);
1438 	int ret;
1439 
1440 	if (!pdev)
1441 		return -ENODEV;
1442 
1443 	gmu->dev = &pdev->dev;
1444 
1445 	of_dma_configure(gmu->dev, node, true);
1446 
1447 	/* Fow now, don't do anything fancy until we get our feet under us */
1448 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1449 
1450 	pm_runtime_enable(gmu->dev);
1451 
1452 	/* Get the list of clocks */
1453 	ret = a6xx_gmu_clocks_probe(gmu);
1454 	if (ret)
1455 		goto err_put_device;
1456 
1457 	ret = a6xx_gmu_memory_probe(gmu);
1458 	if (ret)
1459 		goto err_put_device;
1460 
1461 	/* Allocate memory for the GMU dummy page */
1462 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, SZ_4K, 0x60000000);
1463 	if (ret)
1464 		goto err_memory;
1465 
1466 	if (adreno_is_a650(adreno_gpu)) {
1467 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1468 			SZ_16M - SZ_16K, 0x04000);
1469 		if (ret)
1470 			goto err_memory;
1471 	} else if (adreno_is_a640(adreno_gpu)) {
1472 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1473 			SZ_256K - SZ_16K, 0x04000);
1474 		if (ret)
1475 			goto err_memory;
1476 
1477 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1478 			SZ_256K - SZ_16K, 0x44000);
1479 		if (ret)
1480 			goto err_memory;
1481 	} else {
1482 		/* HFI v1, has sptprac */
1483 		gmu->legacy = true;
1484 
1485 		/* Allocate memory for the GMU debug region */
1486 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0);
1487 		if (ret)
1488 			goto err_memory;
1489 	}
1490 
1491 	/* Allocate memory for for the HFI queues */
1492 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0);
1493 	if (ret)
1494 		goto err_memory;
1495 
1496 	/* Allocate memory for the GMU log region */
1497 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0);
1498 	if (ret)
1499 		goto err_memory;
1500 
1501 	/* Map the GMU registers */
1502 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1503 	if (IS_ERR(gmu->mmio)) {
1504 		ret = PTR_ERR(gmu->mmio);
1505 		goto err_memory;
1506 	}
1507 
1508 	if (adreno_is_a650(adreno_gpu)) {
1509 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1510 		if (IS_ERR(gmu->rscc))
1511 			goto err_mmio;
1512 	} else {
1513 		gmu->rscc = gmu->mmio + 0x23000;
1514 	}
1515 
1516 	/* Get the HFI and GMU interrupts */
1517 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1518 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1519 
1520 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1521 		goto err_mmio;
1522 
1523 	/*
1524 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1525 	 * crash
1526 	 */
1527 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1528 
1529 	/* Get the power levels for the GMU and GPU */
1530 	a6xx_gmu_pwrlevels_probe(gmu);
1531 
1532 	/* Set up the HFI queues */
1533 	a6xx_hfi_init(gmu);
1534 
1535 	gmu->initialized = true;
1536 
1537 	return 0;
1538 
1539 err_mmio:
1540 	iounmap(gmu->mmio);
1541 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1542 		iounmap(gmu->rscc);
1543 	free_irq(gmu->gmu_irq, gmu);
1544 	free_irq(gmu->hfi_irq, gmu);
1545 
1546 	ret = -ENODEV;
1547 
1548 err_memory:
1549 	a6xx_gmu_memory_free(gmu);
1550 err_put_device:
1551 	/* Drop reference taken in of_find_device_by_node */
1552 	put_device(gmu->dev);
1553 
1554 	return ret;
1555 }
1556