xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision 249592bf)
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_opp(&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_opp(&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 struct a6xx_gmu_oob_bits {
249 	int set, ack, set_new, ack_new, clear, clear_new;
250 	const char *name;
251 };
252 
253 /* These are the interrupt / ack bits for each OOB request that are set
254  * in a6xx_gmu_set_oob and a6xx_clear_oob
255  */
256 static const struct a6xx_gmu_oob_bits a6xx_gmu_oob_bits[] = {
257 	[GMU_OOB_GPU_SET] = {
258 		.name = "GPU_SET",
259 		.set = 16,
260 		.ack = 24,
261 		.set_new = 30,
262 		.ack_new = 31,
263 		.clear = 24,
264 		.clear_new = 31,
265 	},
266 
267 	[GMU_OOB_PERFCOUNTER_SET] = {
268 		.name = "PERFCOUNTER",
269 		.set = 17,
270 		.ack = 25,
271 		.set_new = 28,
272 		.ack_new = 30,
273 		.clear = 25,
274 		.clear_new = 29,
275 	},
276 
277 	[GMU_OOB_BOOT_SLUMBER] = {
278 		.name = "BOOT_SLUMBER",
279 		.set = 22,
280 		.ack = 30,
281 		.clear = 30,
282 	},
283 
284 	[GMU_OOB_DCVS_SET] = {
285 		.name = "GPU_DCVS",
286 		.set = 23,
287 		.ack = 31,
288 		.clear = 31,
289 	},
290 };
291 
292 /* Trigger a OOB (out of band) request to the GMU */
293 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
294 {
295 	int ret;
296 	u32 val;
297 	int request, ack;
298 
299 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
300 		return -EINVAL;
301 
302 	if (gmu->legacy) {
303 		request = a6xx_gmu_oob_bits[state].set;
304 		ack = a6xx_gmu_oob_bits[state].ack;
305 	} else {
306 		request = a6xx_gmu_oob_bits[state].set_new;
307 		ack = a6xx_gmu_oob_bits[state].ack_new;
308 		if (!request || !ack) {
309 			DRM_DEV_ERROR(gmu->dev,
310 				      "Invalid non-legacy GMU request %s\n",
311 				      a6xx_gmu_oob_bits[state].name);
312 			return -EINVAL;
313 		}
314 	}
315 
316 	/* Trigger the equested OOB operation */
317 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
318 
319 	/* Wait for the acknowledge interrupt */
320 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
321 		val & (1 << ack), 100, 10000);
322 
323 	if (ret)
324 		DRM_DEV_ERROR(gmu->dev,
325 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
326 				a6xx_gmu_oob_bits[state].name,
327 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
328 
329 	/* Clear the acknowledge interrupt */
330 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
331 
332 	return ret;
333 }
334 
335 /* Clear a pending OOB state in the GMU */
336 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
337 {
338 	int bit;
339 
340 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
341 		return;
342 
343 	if (gmu->legacy)
344 		bit = a6xx_gmu_oob_bits[state].clear;
345 	else
346 		bit = a6xx_gmu_oob_bits[state].clear_new;
347 
348 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << bit);
349 }
350 
351 /* Enable CPU control of SPTP power power collapse */
352 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
353 {
354 	int ret;
355 	u32 val;
356 
357 	if (!gmu->legacy)
358 		return 0;
359 
360 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
361 
362 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
363 		(val & 0x38) == 0x28, 1, 100);
364 
365 	if (ret) {
366 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
367 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
368 	}
369 
370 	return 0;
371 }
372 
373 /* Disable CPU control of SPTP power power collapse */
374 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
375 {
376 	u32 val;
377 	int ret;
378 
379 	if (!gmu->legacy)
380 		return;
381 
382 	/* Make sure retention is on */
383 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
384 
385 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
386 
387 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
388 		(val & 0x04), 100, 10000);
389 
390 	if (ret)
391 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
392 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
393 }
394 
395 /* Let the GMU know we are starting a boot sequence */
396 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
397 {
398 	u32 vote;
399 
400 	/* Let the GMU know we are getting ready for boot */
401 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
402 
403 	/* Choose the "default" power level as the highest available */
404 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
405 
406 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
407 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
408 
409 	/* Let the GMU know the boot sequence has started */
410 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
411 }
412 
413 /* Let the GMU know that we are about to go into slumber */
414 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
415 {
416 	int ret;
417 
418 	/* Disable the power counter so the GMU isn't busy */
419 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
420 
421 	/* Disable SPTP_PC if the CPU is responsible for it */
422 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
423 		a6xx_sptprac_disable(gmu);
424 
425 	if (!gmu->legacy) {
426 		ret = a6xx_hfi_send_prep_slumber(gmu);
427 		goto out;
428 	}
429 
430 	/* Tell the GMU to get ready to slumber */
431 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
432 
433 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
434 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
435 
436 	if (!ret) {
437 		/* Check to see if the GMU really did slumber */
438 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
439 			!= 0x0f) {
440 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
441 			ret = -ETIMEDOUT;
442 		}
443 	}
444 
445 out:
446 	/* Put fence into allow mode */
447 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
448 	return ret;
449 }
450 
451 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
452 {
453 	int ret;
454 	u32 val;
455 
456 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
457 	/* Wait for the register to finish posting */
458 	wmb();
459 
460 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
461 		val & (1 << 1), 100, 10000);
462 	if (ret) {
463 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
464 		return ret;
465 	}
466 
467 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
468 		!val, 100, 10000);
469 
470 	if (ret) {
471 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
472 		return ret;
473 	}
474 
475 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
476 
477 	/* Set up CX GMU counter 0 to count busy ticks */
478 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
479 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
480 
481 	/* Enable the power counter */
482 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
483 	return 0;
484 }
485 
486 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
487 {
488 	int ret;
489 	u32 val;
490 
491 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
492 
493 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
494 		val, val & (1 << 16), 100, 10000);
495 	if (ret)
496 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
497 
498 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
499 }
500 
501 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
502 {
503 	return msm_writel(value, ptr + (offset << 2));
504 }
505 
506 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
507 		const char *name);
508 
509 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
510 {
511 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
512 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
513 	struct platform_device *pdev = to_platform_device(gmu->dev);
514 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
515 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
516 	uint32_t pdc_address_offset;
517 
518 	if (!pdcptr || !seqptr)
519 		goto err;
520 
521 	if (adreno_is_a618(adreno_gpu) || adreno_is_a640(adreno_gpu))
522 		pdc_address_offset = 0x30090;
523 	else if (adreno_is_a650(adreno_gpu))
524 		pdc_address_offset = 0x300a0;
525 	else
526 		pdc_address_offset = 0x30080;
527 
528 	/* Disable SDE clock gating */
529 	gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
530 
531 	/* Setup RSC PDC handshake for sleep and wakeup */
532 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
533 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
534 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
535 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
536 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
537 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
538 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
539 	gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
540 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
541 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
542 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
543 
544 	/* Load RSC sequencer uCode for sleep and wakeup */
545 	if (adreno_is_a650(adreno_gpu)) {
546 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
547 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
548 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
549 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
550 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
551 	} else {
552 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
553 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
554 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
555 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
556 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
557 	}
558 
559 	/* Load PDC sequencer uCode for power up and power down sequence */
560 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
561 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
562 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
563 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
564 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
565 
566 	/* Set TCS commands used by PDC sequence for low power modes */
567 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
568 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
569 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
570 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
571 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
572 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
573 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
574 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
575 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
576 
577 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
578 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
579 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
580 
581 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
582 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
583 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
584 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
585 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
586 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
587 
588 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
589 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
590 	if (adreno_is_a618(adreno_gpu) || adreno_is_a650(adreno_gpu))
591 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
592 	else
593 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
594 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
595 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
596 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
597 
598 	/* Setup GPU PDC */
599 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
600 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
601 
602 	/* ensure no writes happen before the uCode is fully written */
603 	wmb();
604 
605 err:
606 	if (!IS_ERR_OR_NULL(pdcptr))
607 		iounmap(pdcptr);
608 	if (!IS_ERR_OR_NULL(seqptr))
609 		iounmap(seqptr);
610 }
611 
612 /*
613  * The lowest 16 bits of this value are the number of XO clock cycles for main
614  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
615  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
616  */
617 
618 #define GMU_PWR_COL_HYST 0x000a1680
619 
620 /* Set up the idle state for the GMU */
621 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
622 {
623 	/* Disable GMU WB/RB buffer */
624 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
625 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
626 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
627 
628 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
629 
630 	switch (gmu->idle_level) {
631 	case GMU_IDLE_STATE_IFPC:
632 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
633 			GMU_PWR_COL_HYST);
634 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
635 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
636 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
637 		fallthrough;
638 	case GMU_IDLE_STATE_SPTP:
639 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
640 			GMU_PWR_COL_HYST);
641 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
642 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
643 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
644 	}
645 
646 	/* Enable RPMh GPU client */
647 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
648 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
649 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
650 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
651 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
652 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
653 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
654 }
655 
656 struct block_header {
657 	u32 addr;
658 	u32 size;
659 	u32 type;
660 	u32 value;
661 	u32 data[];
662 };
663 
664 /* this should be a general kernel helper */
665 static int in_range(u32 addr, u32 start, u32 size)
666 {
667 	return addr >= start && addr < start + size;
668 }
669 
670 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
671 {
672 	if (!in_range(blk->addr, bo->iova, bo->size))
673 		return false;
674 
675 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
676 	return true;
677 }
678 
679 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
680 {
681 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
682 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
683 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
684 	const struct block_header *blk;
685 	u32 reg_offset;
686 
687 	u32 itcm_base = 0x00000000;
688 	u32 dtcm_base = 0x00040000;
689 
690 	if (adreno_is_a650(adreno_gpu))
691 		dtcm_base = 0x10004000;
692 
693 	if (gmu->legacy) {
694 		/* Sanity check the size of the firmware that was loaded */
695 		if (fw_image->size > 0x8000) {
696 			DRM_DEV_ERROR(gmu->dev,
697 				"GMU firmware is bigger than the available region\n");
698 			return -EINVAL;
699 		}
700 
701 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
702 			       (u32*) fw_image->data, fw_image->size);
703 		return 0;
704 	}
705 
706 
707 	for (blk = (const struct block_header *) fw_image->data;
708 	     (const u8*) blk < fw_image->data + fw_image->size;
709 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
710 		if (blk->size == 0)
711 			continue;
712 
713 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
714 			reg_offset = (blk->addr - itcm_base) >> 2;
715 			gmu_write_bulk(gmu,
716 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
717 				blk->data, blk->size);
718 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
719 			reg_offset = (blk->addr - dtcm_base) >> 2;
720 			gmu_write_bulk(gmu,
721 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
722 				blk->data, blk->size);
723 		} else if (!fw_block_mem(&gmu->icache, blk) &&
724 			   !fw_block_mem(&gmu->dcache, blk) &&
725 			   !fw_block_mem(&gmu->dummy, blk)) {
726 			DRM_DEV_ERROR(gmu->dev,
727 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
728 				blk->addr, blk->size, blk->data[0]);
729 		}
730 	}
731 
732 	return 0;
733 }
734 
735 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
736 {
737 	static bool rpmh_init;
738 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
739 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
740 	int ret;
741 	u32 chipid;
742 
743 	if (adreno_is_a650(adreno_gpu))
744 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
745 
746 	if (state == GMU_WARM_BOOT) {
747 		ret = a6xx_rpmh_start(gmu);
748 		if (ret)
749 			return ret;
750 	} else {
751 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
752 			"GMU firmware is not loaded\n"))
753 			return -ENOENT;
754 
755 		/* Turn on register retention */
756 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
757 
758 		/* We only need to load the RPMh microcode once */
759 		if (!rpmh_init) {
760 			a6xx_gmu_rpmh_init(gmu);
761 			rpmh_init = true;
762 		} else {
763 			ret = a6xx_rpmh_start(gmu);
764 			if (ret)
765 				return ret;
766 		}
767 
768 		ret = a6xx_gmu_fw_load(gmu);
769 		if (ret)
770 			return ret;
771 	}
772 
773 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
774 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
775 
776 	/* Write the iova of the HFI table */
777 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
778 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
779 
780 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
781 		(1 << 31) | (0xa << 18) | (0xa0));
782 
783 	chipid = adreno_gpu->rev.core << 24;
784 	chipid |= adreno_gpu->rev.major << 16;
785 	chipid |= adreno_gpu->rev.minor << 12;
786 	chipid |= adreno_gpu->rev.patchid << 8;
787 
788 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
789 
790 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
791 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
792 
793 	/* Set up the lowest idle level on the GMU */
794 	a6xx_gmu_power_config(gmu);
795 
796 	ret = a6xx_gmu_start(gmu);
797 	if (ret)
798 		return ret;
799 
800 	if (gmu->legacy) {
801 		ret = a6xx_gmu_gfx_rail_on(gmu);
802 		if (ret)
803 			return ret;
804 	}
805 
806 	/* Enable SPTP_PC if the CPU is responsible for it */
807 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
808 		ret = a6xx_sptprac_enable(gmu);
809 		if (ret)
810 			return ret;
811 	}
812 
813 	ret = a6xx_gmu_hfi_start(gmu);
814 	if (ret)
815 		return ret;
816 
817 	/* FIXME: Do we need this wmb() here? */
818 	wmb();
819 
820 	return 0;
821 }
822 
823 #define A6XX_HFI_IRQ_MASK \
824 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
825 
826 #define A6XX_GMU_IRQ_MASK \
827 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
828 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
829 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
830 
831 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
832 {
833 	disable_irq(gmu->gmu_irq);
834 	disable_irq(gmu->hfi_irq);
835 
836 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
837 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
838 }
839 
840 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
841 {
842 	u32 val;
843 
844 	/* Make sure there are no outstanding RPMh votes */
845 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
846 		(val & 1), 100, 10000);
847 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
848 		(val & 1), 100, 10000);
849 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
850 		(val & 1), 100, 10000);
851 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
852 		(val & 1), 100, 1000);
853 }
854 
855 /* Force the GMU off in case it isn't responsive */
856 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
857 {
858 	/* Flush all the queues */
859 	a6xx_hfi_stop(gmu);
860 
861 	/* Stop the interrupts */
862 	a6xx_gmu_irq_disable(gmu);
863 
864 	/* Force off SPTP in case the GMU is managing it */
865 	a6xx_sptprac_disable(gmu);
866 
867 	/* Make sure there are no outstanding RPMh votes */
868 	a6xx_gmu_rpmh_off(gmu);
869 }
870 
871 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
872 {
873 	struct dev_pm_opp *gpu_opp;
874 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
875 
876 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
877 	if (IS_ERR_OR_NULL(gpu_opp))
878 		return;
879 
880 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
881 	a6xx_gmu_set_freq(gpu, gpu_opp);
882 	dev_pm_opp_put(gpu_opp);
883 }
884 
885 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
886 {
887 	struct dev_pm_opp *gpu_opp;
888 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
889 
890 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
891 	if (IS_ERR_OR_NULL(gpu_opp))
892 		return;
893 
894 	dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp);
895 	dev_pm_opp_put(gpu_opp);
896 }
897 
898 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
899 {
900 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
901 	struct msm_gpu *gpu = &adreno_gpu->base;
902 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
903 	int status, ret;
904 
905 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
906 		return 0;
907 
908 	gmu->hung = false;
909 
910 	/* Turn on the resources */
911 	pm_runtime_get_sync(gmu->dev);
912 
913 	/*
914 	 * "enable" the GX power domain which won't actually do anything but it
915 	 * will make sure that the refcounting is correct in case we need to
916 	 * bring down the GX after a GMU failure
917 	 */
918 	if (!IS_ERR_OR_NULL(gmu->gxpd))
919 		pm_runtime_get_sync(gmu->gxpd);
920 
921 	/* Use a known rate to bring up the GMU */
922 	clk_set_rate(gmu->core_clk, 200000000);
923 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
924 	if (ret) {
925 		pm_runtime_put(gmu->gxpd);
926 		pm_runtime_put(gmu->dev);
927 		return ret;
928 	}
929 
930 	/* Set the bus quota to a reasonable value for boot */
931 	a6xx_gmu_set_initial_bw(gpu, gmu);
932 
933 	/* Enable the GMU interrupt */
934 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
935 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
936 	enable_irq(gmu->gmu_irq);
937 
938 	/* Check to see if we are doing a cold or warm boot */
939 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
940 		GMU_WARM_BOOT : GMU_COLD_BOOT;
941 
942 	/*
943 	 * Warm boot path does not work on newer GPUs
944 	 * Presumably this is because icache/dcache regions must be restored
945 	 */
946 	if (!gmu->legacy)
947 		status = GMU_COLD_BOOT;
948 
949 	ret = a6xx_gmu_fw_start(gmu, status);
950 	if (ret)
951 		goto out;
952 
953 	ret = a6xx_hfi_start(gmu, status);
954 	if (ret)
955 		goto out;
956 
957 	/*
958 	 * Turn on the GMU firmware fault interrupt after we know the boot
959 	 * sequence is successful
960 	 */
961 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
962 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
963 	enable_irq(gmu->hfi_irq);
964 
965 	/* Set the GPU to the current freq */
966 	a6xx_gmu_set_initial_freq(gpu, gmu);
967 
968 out:
969 	/* On failure, shut down the GMU to leave it in a good state */
970 	if (ret) {
971 		disable_irq(gmu->gmu_irq);
972 		a6xx_rpmh_stop(gmu);
973 		pm_runtime_put(gmu->gxpd);
974 		pm_runtime_put(gmu->dev);
975 	}
976 
977 	return ret;
978 }
979 
980 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
981 {
982 	u32 reg;
983 
984 	if (!gmu->initialized)
985 		return true;
986 
987 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
988 
989 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
990 		return false;
991 
992 	return true;
993 }
994 
995 #define GBIF_CLIENT_HALT_MASK             BIT(0)
996 #define GBIF_ARB_HALT_MASK                BIT(1)
997 
998 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
999 {
1000 	struct msm_gpu *gpu = &adreno_gpu->base;
1001 
1002 	if (!a6xx_has_gbif(adreno_gpu)) {
1003 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
1004 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
1005 								0xf) == 0xf);
1006 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
1007 
1008 		return;
1009 	}
1010 
1011 	/* Halt new client requests on GBIF */
1012 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
1013 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
1014 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
1015 
1016 	/* Halt all AXI requests on GBIF */
1017 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
1018 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
1019 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
1020 
1021 	/* The GBIF halt needs to be explicitly cleared */
1022 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
1023 }
1024 
1025 /* Gracefully try to shut down the GMU and by extension the GPU */
1026 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1027 {
1028 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1029 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1030 	u32 val;
1031 
1032 	/*
1033 	 * The GMU may still be in slumber unless the GPU started so check and
1034 	 * skip putting it back into slumber if so
1035 	 */
1036 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1037 
1038 	if (val != 0xf) {
1039 		int ret = a6xx_gmu_wait_for_idle(gmu);
1040 
1041 		/* If the GMU isn't responding assume it is hung */
1042 		if (ret) {
1043 			a6xx_gmu_force_off(gmu);
1044 			return;
1045 		}
1046 
1047 		a6xx_bus_clear_pending_transactions(adreno_gpu);
1048 
1049 		/* tell the GMU we want to slumber */
1050 		a6xx_gmu_notify_slumber(gmu);
1051 
1052 		ret = gmu_poll_timeout(gmu,
1053 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1054 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1055 			100, 10000);
1056 
1057 		/*
1058 		 * Let the user know we failed to slumber but don't worry too
1059 		 * much because we are powering down anyway
1060 		 */
1061 
1062 		if (ret)
1063 			DRM_DEV_ERROR(gmu->dev,
1064 				"Unable to slumber GMU: status = 0%x/0%x\n",
1065 				gmu_read(gmu,
1066 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1067 				gmu_read(gmu,
1068 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1069 	}
1070 
1071 	/* Turn off HFI */
1072 	a6xx_hfi_stop(gmu);
1073 
1074 	/* Stop the interrupts and mask the hardware */
1075 	a6xx_gmu_irq_disable(gmu);
1076 
1077 	/* Tell RPMh to power off the GPU */
1078 	a6xx_rpmh_stop(gmu);
1079 }
1080 
1081 
1082 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1083 {
1084 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1085 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1086 
1087 	if (!pm_runtime_active(gmu->dev))
1088 		return 0;
1089 
1090 	/*
1091 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1092 	 * down gracefully
1093 	 */
1094 	if (gmu->hung)
1095 		a6xx_gmu_force_off(gmu);
1096 	else
1097 		a6xx_gmu_shutdown(gmu);
1098 
1099 	/* Remove the bus vote */
1100 	dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1101 
1102 	/*
1103 	 * Make sure the GX domain is off before turning off the GMU (CX)
1104 	 * domain. Usually the GMU does this but only if the shutdown sequence
1105 	 * was successful
1106 	 */
1107 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1108 		pm_runtime_put_sync(gmu->gxpd);
1109 
1110 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1111 
1112 	pm_runtime_put_sync(gmu->dev);
1113 
1114 	return 0;
1115 }
1116 
1117 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1118 {
1119 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace, false);
1120 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace, false);
1121 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace, false);
1122 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace, false);
1123 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace, false);
1124 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace, false);
1125 
1126 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1127 	msm_gem_address_space_put(gmu->aspace);
1128 }
1129 
1130 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1131 		size_t size, u64 iova)
1132 {
1133 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1134 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1135 	uint32_t flags = MSM_BO_WC;
1136 	u64 range_start, range_end;
1137 	int ret;
1138 
1139 	size = PAGE_ALIGN(size);
1140 	if (!iova) {
1141 		/* no fixed address - use GMU's uncached range */
1142 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1143 		range_end = 0x80000000;
1144 	} else {
1145 		/* range for fixed address */
1146 		range_start = iova;
1147 		range_end = iova + size;
1148 		/* use IOMMU_PRIV for icache/dcache */
1149 		flags |= MSM_BO_MAP_PRIV;
1150 	}
1151 
1152 	bo->obj = msm_gem_new(dev, size, flags);
1153 	if (IS_ERR(bo->obj))
1154 		return PTR_ERR(bo->obj);
1155 
1156 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1157 		range_start >> PAGE_SHIFT, range_end >> PAGE_SHIFT);
1158 	if (ret) {
1159 		drm_gem_object_put(bo->obj);
1160 		return ret;
1161 	}
1162 
1163 	bo->virt = msm_gem_get_vaddr(bo->obj);
1164 	bo->size = size;
1165 
1166 	return 0;
1167 }
1168 
1169 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1170 {
1171 	struct iommu_domain *domain;
1172 	struct msm_mmu *mmu;
1173 
1174 	domain = iommu_domain_alloc(&platform_bus_type);
1175 	if (!domain)
1176 		return -ENODEV;
1177 
1178 	mmu = msm_iommu_new(gmu->dev, domain);
1179 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1180 	if (IS_ERR(gmu->aspace)) {
1181 		iommu_domain_free(domain);
1182 		return PTR_ERR(gmu->aspace);
1183 	}
1184 
1185 	return 0;
1186 }
1187 
1188 /* Return the 'arc-level' for the given frequency */
1189 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1190 					   unsigned long freq)
1191 {
1192 	struct dev_pm_opp *opp;
1193 	unsigned int val;
1194 
1195 	if (!freq)
1196 		return 0;
1197 
1198 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1199 	if (IS_ERR(opp))
1200 		return 0;
1201 
1202 	val = dev_pm_opp_get_level(opp);
1203 
1204 	dev_pm_opp_put(opp);
1205 
1206 	return val;
1207 }
1208 
1209 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1210 		unsigned long *freqs, int freqs_count, const char *id)
1211 {
1212 	int i, j;
1213 	const u16 *pri, *sec;
1214 	size_t pri_count, sec_count;
1215 
1216 	pri = cmd_db_read_aux_data(id, &pri_count);
1217 	if (IS_ERR(pri))
1218 		return PTR_ERR(pri);
1219 	/*
1220 	 * The data comes back as an array of unsigned shorts so adjust the
1221 	 * count accordingly
1222 	 */
1223 	pri_count >>= 1;
1224 	if (!pri_count)
1225 		return -EINVAL;
1226 
1227 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1228 	if (IS_ERR(sec))
1229 		return PTR_ERR(sec);
1230 
1231 	sec_count >>= 1;
1232 	if (!sec_count)
1233 		return -EINVAL;
1234 
1235 	/* Construct a vote for each frequency */
1236 	for (i = 0; i < freqs_count; i++) {
1237 		u8 pindex = 0, sindex = 0;
1238 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1239 
1240 		/* Get the primary index that matches the arc level */
1241 		for (j = 0; j < pri_count; j++) {
1242 			if (pri[j] >= level) {
1243 				pindex = j;
1244 				break;
1245 			}
1246 		}
1247 
1248 		if (j == pri_count) {
1249 			DRM_DEV_ERROR(dev,
1250 				      "Level %u not found in the RPMh list\n",
1251 				      level);
1252 			DRM_DEV_ERROR(dev, "Available levels:\n");
1253 			for (j = 0; j < pri_count; j++)
1254 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1255 
1256 			return -EINVAL;
1257 		}
1258 
1259 		/*
1260 		 * Look for a level in in the secondary list that matches. If
1261 		 * nothing fits, use the maximum non zero vote
1262 		 */
1263 
1264 		for (j = 0; j < sec_count; j++) {
1265 			if (sec[j] >= level) {
1266 				sindex = j;
1267 				break;
1268 			} else if (sec[j]) {
1269 				sindex = j;
1270 			}
1271 		}
1272 
1273 		/* Construct the vote */
1274 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1275 			(sindex << 8) | pindex;
1276 	}
1277 
1278 	return 0;
1279 }
1280 
1281 /*
1282  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1283  * to construct the list of votes on the CPU and send it over. Query the RPMh
1284  * voltage levels and build the votes
1285  */
1286 
1287 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1288 {
1289 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1290 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1291 	struct msm_gpu *gpu = &adreno_gpu->base;
1292 	int ret;
1293 
1294 	/* Build the GX votes */
1295 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1296 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1297 
1298 	/* Build the CX votes */
1299 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1300 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1301 
1302 	return ret;
1303 }
1304 
1305 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1306 		u32 size)
1307 {
1308 	int count = dev_pm_opp_get_opp_count(dev);
1309 	struct dev_pm_opp *opp;
1310 	int i, index = 0;
1311 	unsigned long freq = 1;
1312 
1313 	/*
1314 	 * The OPP table doesn't contain the "off" frequency level so we need to
1315 	 * add 1 to the table size to account for it
1316 	 */
1317 
1318 	if (WARN(count + 1 > size,
1319 		"The GMU frequency table is being truncated\n"))
1320 		count = size - 1;
1321 
1322 	/* Set the "off" frequency */
1323 	freqs[index++] = 0;
1324 
1325 	for (i = 0; i < count; i++) {
1326 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1327 		if (IS_ERR(opp))
1328 			break;
1329 
1330 		dev_pm_opp_put(opp);
1331 		freqs[index++] = freq++;
1332 	}
1333 
1334 	return index;
1335 }
1336 
1337 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1338 {
1339 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1340 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1341 	struct msm_gpu *gpu = &adreno_gpu->base;
1342 
1343 	int ret = 0;
1344 
1345 	/*
1346 	 * The GMU handles its own frequency switching so build a list of
1347 	 * available frequencies to send during initialization
1348 	 */
1349 	ret = dev_pm_opp_of_add_table(gmu->dev);
1350 	if (ret) {
1351 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1352 		return ret;
1353 	}
1354 
1355 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1356 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1357 
1358 	/*
1359 	 * The GMU also handles GPU frequency switching so build a list
1360 	 * from the GPU OPP table
1361 	 */
1362 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1363 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1364 
1365 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1366 
1367 	/* Build the list of RPMh votes that we'll send to the GMU */
1368 	return a6xx_gmu_rpmh_votes_init(gmu);
1369 }
1370 
1371 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1372 {
1373 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1374 
1375 	if (ret < 1)
1376 		return ret;
1377 
1378 	gmu->nr_clocks = ret;
1379 
1380 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1381 		gmu->nr_clocks, "gmu");
1382 
1383 	return 0;
1384 }
1385 
1386 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1387 		const char *name)
1388 {
1389 	void __iomem *ret;
1390 	struct resource *res = platform_get_resource_byname(pdev,
1391 			IORESOURCE_MEM, name);
1392 
1393 	if (!res) {
1394 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1395 		return ERR_PTR(-EINVAL);
1396 	}
1397 
1398 	ret = ioremap(res->start, resource_size(res));
1399 	if (!ret) {
1400 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1401 		return ERR_PTR(-EINVAL);
1402 	}
1403 
1404 	return ret;
1405 }
1406 
1407 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1408 		const char *name, irq_handler_t handler)
1409 {
1410 	int irq, ret;
1411 
1412 	irq = platform_get_irq_byname(pdev, name);
1413 
1414 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1415 	if (ret) {
1416 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1417 			      name, ret);
1418 		return ret;
1419 	}
1420 
1421 	disable_irq(irq);
1422 
1423 	return irq;
1424 }
1425 
1426 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1427 {
1428 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1429 	struct platform_device *pdev = to_platform_device(gmu->dev);
1430 
1431 	if (!gmu->initialized)
1432 		return;
1433 
1434 	pm_runtime_force_suspend(gmu->dev);
1435 
1436 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1437 		pm_runtime_disable(gmu->gxpd);
1438 		dev_pm_domain_detach(gmu->gxpd, false);
1439 	}
1440 
1441 	iounmap(gmu->mmio);
1442 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1443 		iounmap(gmu->rscc);
1444 	gmu->mmio = NULL;
1445 	gmu->rscc = NULL;
1446 
1447 	a6xx_gmu_memory_free(gmu);
1448 
1449 	free_irq(gmu->gmu_irq, gmu);
1450 	free_irq(gmu->hfi_irq, gmu);
1451 
1452 	/* Drop reference taken in of_find_device_by_node */
1453 	put_device(gmu->dev);
1454 
1455 	gmu->initialized = false;
1456 }
1457 
1458 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1459 {
1460 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1461 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1462 	struct platform_device *pdev = of_find_device_by_node(node);
1463 	int ret;
1464 
1465 	if (!pdev)
1466 		return -ENODEV;
1467 
1468 	gmu->dev = &pdev->dev;
1469 
1470 	of_dma_configure(gmu->dev, node, true);
1471 
1472 	/* Fow now, don't do anything fancy until we get our feet under us */
1473 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1474 
1475 	pm_runtime_enable(gmu->dev);
1476 
1477 	/* Get the list of clocks */
1478 	ret = a6xx_gmu_clocks_probe(gmu);
1479 	if (ret)
1480 		goto err_put_device;
1481 
1482 	ret = a6xx_gmu_memory_probe(gmu);
1483 	if (ret)
1484 		goto err_put_device;
1485 
1486 	/* Allocate memory for the GMU dummy page */
1487 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, SZ_4K, 0x60000000);
1488 	if (ret)
1489 		goto err_memory;
1490 
1491 	if (adreno_is_a650(adreno_gpu)) {
1492 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1493 			SZ_16M - SZ_16K, 0x04000);
1494 		if (ret)
1495 			goto err_memory;
1496 	} else if (adreno_is_a640(adreno_gpu)) {
1497 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1498 			SZ_256K - SZ_16K, 0x04000);
1499 		if (ret)
1500 			goto err_memory;
1501 
1502 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1503 			SZ_256K - SZ_16K, 0x44000);
1504 		if (ret)
1505 			goto err_memory;
1506 	} else {
1507 		/* HFI v1, has sptprac */
1508 		gmu->legacy = true;
1509 
1510 		/* Allocate memory for the GMU debug region */
1511 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0);
1512 		if (ret)
1513 			goto err_memory;
1514 	}
1515 
1516 	/* Allocate memory for for the HFI queues */
1517 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0);
1518 	if (ret)
1519 		goto err_memory;
1520 
1521 	/* Allocate memory for the GMU log region */
1522 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0);
1523 	if (ret)
1524 		goto err_memory;
1525 
1526 	/* Map the GMU registers */
1527 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1528 	if (IS_ERR(gmu->mmio)) {
1529 		ret = PTR_ERR(gmu->mmio);
1530 		goto err_memory;
1531 	}
1532 
1533 	if (adreno_is_a650(adreno_gpu)) {
1534 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1535 		if (IS_ERR(gmu->rscc))
1536 			goto err_mmio;
1537 	} else {
1538 		gmu->rscc = gmu->mmio + 0x23000;
1539 	}
1540 
1541 	/* Get the HFI and GMU interrupts */
1542 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1543 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1544 
1545 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1546 		goto err_mmio;
1547 
1548 	/*
1549 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1550 	 * crash
1551 	 */
1552 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1553 
1554 	/* Get the power levels for the GMU and GPU */
1555 	a6xx_gmu_pwrlevels_probe(gmu);
1556 
1557 	/* Set up the HFI queues */
1558 	a6xx_hfi_init(gmu);
1559 
1560 	gmu->initialized = true;
1561 
1562 	return 0;
1563 
1564 err_mmio:
1565 	iounmap(gmu->mmio);
1566 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1567 		iounmap(gmu->rscc);
1568 	free_irq(gmu->gmu_irq, gmu);
1569 	free_irq(gmu->hfi_irq, gmu);
1570 
1571 	ret = -ENODEV;
1572 
1573 err_memory:
1574 	a6xx_gmu_memory_free(gmu);
1575 err_put_device:
1576 	/* Drop reference taken in of_find_device_by_node */
1577 	put_device(gmu->dev);
1578 
1579 	return ret;
1580 }
1581