xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision b737eecd)
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;
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 	},
264 
265 	[GMU_OOB_PERFCOUNTER_SET] = {
266 		.name = "PERFCOUNTER",
267 		.set = 17,
268 		.ack = 25,
269 		.set_new = 28,
270 		.ack_new = 30,
271 	},
272 
273 	[GMU_OOB_BOOT_SLUMBER] = {
274 		.name = "BOOT_SLUMBER",
275 		.set = 22,
276 		.ack = 30,
277 	},
278 
279 	[GMU_OOB_DCVS_SET] = {
280 		.name = "GPU_DCVS",
281 		.set = 23,
282 		.ack = 31,
283 	},
284 };
285 
286 /* Trigger a OOB (out of band) request to the GMU */
287 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
288 {
289 	int ret;
290 	u32 val;
291 	int request, ack;
292 
293 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
294 		return -EINVAL;
295 
296 	if (gmu->legacy) {
297 		request = a6xx_gmu_oob_bits[state].set;
298 		ack = a6xx_gmu_oob_bits[state].ack;
299 	} else {
300 		request = a6xx_gmu_oob_bits[state].set_new;
301 		ack = a6xx_gmu_oob_bits[state].ack_new;
302 		if (!request || !ack) {
303 			DRM_DEV_ERROR(gmu->dev,
304 				      "Invalid non-legacy GMU request %s\n",
305 				      a6xx_gmu_oob_bits[state].name);
306 			return -EINVAL;
307 		}
308 	}
309 
310 	/* Trigger the equested OOB operation */
311 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
312 
313 	/* Wait for the acknowledge interrupt */
314 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
315 		val & (1 << ack), 100, 10000);
316 
317 	if (ret)
318 		DRM_DEV_ERROR(gmu->dev,
319 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
320 				a6xx_gmu_oob_bits[state].name,
321 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
322 
323 	/* Clear the acknowledge interrupt */
324 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
325 
326 	return ret;
327 }
328 
329 /* Clear a pending OOB state in the GMU */
330 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
331 {
332 	int bit;
333 
334 	if (state >= ARRAY_SIZE(a6xx_gmu_oob_bits))
335 		return;
336 
337 	if (gmu->legacy)
338 		bit = a6xx_gmu_oob_bits[state].ack;
339 	else
340 		bit = a6xx_gmu_oob_bits[state].ack_new;
341 
342 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, bit);
343 }
344 
345 /* Enable CPU control of SPTP power power collapse */
346 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
347 {
348 	int ret;
349 	u32 val;
350 
351 	if (!gmu->legacy)
352 		return 0;
353 
354 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
355 
356 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
357 		(val & 0x38) == 0x28, 1, 100);
358 
359 	if (ret) {
360 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
361 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
362 	}
363 
364 	return 0;
365 }
366 
367 /* Disable CPU control of SPTP power power collapse */
368 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
369 {
370 	u32 val;
371 	int ret;
372 
373 	if (!gmu->legacy)
374 		return;
375 
376 	/* Make sure retention is on */
377 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
378 
379 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
380 
381 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
382 		(val & 0x04), 100, 10000);
383 
384 	if (ret)
385 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
386 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
387 }
388 
389 /* Let the GMU know we are starting a boot sequence */
390 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
391 {
392 	u32 vote;
393 
394 	/* Let the GMU know we are getting ready for boot */
395 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
396 
397 	/* Choose the "default" power level as the highest available */
398 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
399 
400 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
401 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
402 
403 	/* Let the GMU know the boot sequence has started */
404 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
405 }
406 
407 /* Let the GMU know that we are about to go into slumber */
408 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
409 {
410 	int ret;
411 
412 	/* Disable the power counter so the GMU isn't busy */
413 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
414 
415 	/* Disable SPTP_PC if the CPU is responsible for it */
416 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
417 		a6xx_sptprac_disable(gmu);
418 
419 	if (!gmu->legacy) {
420 		ret = a6xx_hfi_send_prep_slumber(gmu);
421 		goto out;
422 	}
423 
424 	/* Tell the GMU to get ready to slumber */
425 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
426 
427 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
428 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
429 
430 	if (!ret) {
431 		/* Check to see if the GMU really did slumber */
432 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
433 			!= 0x0f) {
434 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
435 			ret = -ETIMEDOUT;
436 		}
437 	}
438 
439 out:
440 	/* Put fence into allow mode */
441 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
442 	return ret;
443 }
444 
445 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
446 {
447 	int ret;
448 	u32 val;
449 
450 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
451 	/* Wait for the register to finish posting */
452 	wmb();
453 
454 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
455 		val & (1 << 1), 100, 10000);
456 	if (ret) {
457 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
458 		return ret;
459 	}
460 
461 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
462 		!val, 100, 10000);
463 
464 	if (ret) {
465 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
466 		return ret;
467 	}
468 
469 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
470 
471 	/* Set up CX GMU counter 0 to count busy ticks */
472 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
473 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
474 
475 	/* Enable the power counter */
476 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
477 	return 0;
478 }
479 
480 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
481 {
482 	int ret;
483 	u32 val;
484 
485 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
486 
487 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
488 		val, val & (1 << 16), 100, 10000);
489 	if (ret)
490 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
491 
492 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
493 }
494 
495 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
496 {
497 	return msm_writel(value, ptr + (offset << 2));
498 }
499 
500 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
501 		const char *name);
502 
503 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
504 {
505 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
506 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
507 	struct platform_device *pdev = to_platform_device(gmu->dev);
508 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
509 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
510 	uint32_t pdc_address_offset;
511 
512 	if (!pdcptr || !seqptr)
513 		goto err;
514 
515 	if (adreno_is_a618(adreno_gpu) || adreno_is_a640(adreno_gpu))
516 		pdc_address_offset = 0x30090;
517 	else if (adreno_is_a650(adreno_gpu))
518 		pdc_address_offset = 0x300a0;
519 	else
520 		pdc_address_offset = 0x30080;
521 
522 	/* Disable SDE clock gating */
523 	gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
524 
525 	/* Setup RSC PDC handshake for sleep and wakeup */
526 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
527 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
528 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
529 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
530 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
531 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
532 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
533 	gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
534 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
535 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
536 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
537 
538 	/* Load RSC sequencer uCode for sleep and wakeup */
539 	if (adreno_is_a650(adreno_gpu)) {
540 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
541 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
542 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
543 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
544 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
545 	} else {
546 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
547 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
548 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
549 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
550 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
551 	}
552 
553 	/* Load PDC sequencer uCode for power up and power down sequence */
554 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
555 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
556 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
557 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
558 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
559 
560 	/* Set TCS commands used by PDC sequence for low power modes */
561 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
562 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
563 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
564 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
565 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
566 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
567 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
568 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
569 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
570 
571 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
572 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
573 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
574 
575 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
576 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
577 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
578 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
579 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
580 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
581 
582 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
583 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
584 	if (adreno_is_a618(adreno_gpu) || adreno_is_a650(adreno_gpu))
585 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
586 	else
587 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
588 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
589 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
590 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
591 
592 	/* Setup GPU PDC */
593 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
594 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
595 
596 	/* ensure no writes happen before the uCode is fully written */
597 	wmb();
598 
599 err:
600 	if (!IS_ERR_OR_NULL(pdcptr))
601 		iounmap(pdcptr);
602 	if (!IS_ERR_OR_NULL(seqptr))
603 		iounmap(seqptr);
604 }
605 
606 /*
607  * The lowest 16 bits of this value are the number of XO clock cycles for main
608  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
609  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
610  */
611 
612 #define GMU_PWR_COL_HYST 0x000a1680
613 
614 /* Set up the idle state for the GMU */
615 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
616 {
617 	/* Disable GMU WB/RB buffer */
618 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
619 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
620 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
621 
622 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
623 
624 	switch (gmu->idle_level) {
625 	case GMU_IDLE_STATE_IFPC:
626 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
627 			GMU_PWR_COL_HYST);
628 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
629 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
630 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
631 		fallthrough;
632 	case GMU_IDLE_STATE_SPTP:
633 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
634 			GMU_PWR_COL_HYST);
635 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
636 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
637 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
638 	}
639 
640 	/* Enable RPMh GPU client */
641 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
642 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
643 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
644 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
645 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
646 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
647 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
648 }
649 
650 struct block_header {
651 	u32 addr;
652 	u32 size;
653 	u32 type;
654 	u32 value;
655 	u32 data[];
656 };
657 
658 /* this should be a general kernel helper */
659 static int in_range(u32 addr, u32 start, u32 size)
660 {
661 	return addr >= start && addr < start + size;
662 }
663 
664 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
665 {
666 	if (!in_range(blk->addr, bo->iova, bo->size))
667 		return false;
668 
669 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
670 	return true;
671 }
672 
673 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
674 {
675 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
676 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
677 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
678 	const struct block_header *blk;
679 	u32 reg_offset;
680 
681 	u32 itcm_base = 0x00000000;
682 	u32 dtcm_base = 0x00040000;
683 
684 	if (adreno_is_a650(adreno_gpu))
685 		dtcm_base = 0x10004000;
686 
687 	if (gmu->legacy) {
688 		/* Sanity check the size of the firmware that was loaded */
689 		if (fw_image->size > 0x8000) {
690 			DRM_DEV_ERROR(gmu->dev,
691 				"GMU firmware is bigger than the available region\n");
692 			return -EINVAL;
693 		}
694 
695 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
696 			       (u32*) fw_image->data, fw_image->size);
697 		return 0;
698 	}
699 
700 
701 	for (blk = (const struct block_header *) fw_image->data;
702 	     (const u8*) blk < fw_image->data + fw_image->size;
703 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
704 		if (blk->size == 0)
705 			continue;
706 
707 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
708 			reg_offset = (blk->addr - itcm_base) >> 2;
709 			gmu_write_bulk(gmu,
710 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
711 				blk->data, blk->size);
712 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
713 			reg_offset = (blk->addr - dtcm_base) >> 2;
714 			gmu_write_bulk(gmu,
715 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
716 				blk->data, blk->size);
717 		} else if (!fw_block_mem(&gmu->icache, blk) &&
718 			   !fw_block_mem(&gmu->dcache, blk) &&
719 			   !fw_block_mem(&gmu->dummy, blk)) {
720 			DRM_DEV_ERROR(gmu->dev,
721 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
722 				blk->addr, blk->size, blk->data[0]);
723 		}
724 	}
725 
726 	return 0;
727 }
728 
729 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
730 {
731 	static bool rpmh_init;
732 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
733 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
734 	int ret;
735 	u32 chipid;
736 
737 	if (adreno_is_a650(adreno_gpu))
738 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
739 
740 	if (state == GMU_WARM_BOOT) {
741 		ret = a6xx_rpmh_start(gmu);
742 		if (ret)
743 			return ret;
744 	} else {
745 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
746 			"GMU firmware is not loaded\n"))
747 			return -ENOENT;
748 
749 		/* Turn on register retention */
750 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
751 
752 		/* We only need to load the RPMh microcode once */
753 		if (!rpmh_init) {
754 			a6xx_gmu_rpmh_init(gmu);
755 			rpmh_init = true;
756 		} else {
757 			ret = a6xx_rpmh_start(gmu);
758 			if (ret)
759 				return ret;
760 		}
761 
762 		ret = a6xx_gmu_fw_load(gmu);
763 		if (ret)
764 			return ret;
765 	}
766 
767 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
768 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
769 
770 	/* Write the iova of the HFI table */
771 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
772 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
773 
774 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
775 		(1 << 31) | (0xa << 18) | (0xa0));
776 
777 	chipid = adreno_gpu->rev.core << 24;
778 	chipid |= adreno_gpu->rev.major << 16;
779 	chipid |= adreno_gpu->rev.minor << 12;
780 	chipid |= adreno_gpu->rev.patchid << 8;
781 
782 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
783 
784 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
785 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
786 
787 	/* Set up the lowest idle level on the GMU */
788 	a6xx_gmu_power_config(gmu);
789 
790 	ret = a6xx_gmu_start(gmu);
791 	if (ret)
792 		return ret;
793 
794 	if (gmu->legacy) {
795 		ret = a6xx_gmu_gfx_rail_on(gmu);
796 		if (ret)
797 			return ret;
798 	}
799 
800 	/* Enable SPTP_PC if the CPU is responsible for it */
801 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
802 		ret = a6xx_sptprac_enable(gmu);
803 		if (ret)
804 			return ret;
805 	}
806 
807 	ret = a6xx_gmu_hfi_start(gmu);
808 	if (ret)
809 		return ret;
810 
811 	/* FIXME: Do we need this wmb() here? */
812 	wmb();
813 
814 	return 0;
815 }
816 
817 #define A6XX_HFI_IRQ_MASK \
818 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
819 
820 #define A6XX_GMU_IRQ_MASK \
821 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
822 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
823 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
824 
825 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
826 {
827 	disable_irq(gmu->gmu_irq);
828 	disable_irq(gmu->hfi_irq);
829 
830 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
831 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
832 }
833 
834 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
835 {
836 	u32 val;
837 
838 	/* Make sure there are no outstanding RPMh votes */
839 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
840 		(val & 1), 100, 10000);
841 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
842 		(val & 1), 100, 10000);
843 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
844 		(val & 1), 100, 10000);
845 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
846 		(val & 1), 100, 1000);
847 }
848 
849 /* Force the GMU off in case it isn't responsive */
850 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
851 {
852 	/* Flush all the queues */
853 	a6xx_hfi_stop(gmu);
854 
855 	/* Stop the interrupts */
856 	a6xx_gmu_irq_disable(gmu);
857 
858 	/* Force off SPTP in case the GMU is managing it */
859 	a6xx_sptprac_disable(gmu);
860 
861 	/* Make sure there are no outstanding RPMh votes */
862 	a6xx_gmu_rpmh_off(gmu);
863 }
864 
865 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
866 {
867 	struct dev_pm_opp *gpu_opp;
868 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
869 
870 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
871 	if (IS_ERR_OR_NULL(gpu_opp))
872 		return;
873 
874 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
875 	a6xx_gmu_set_freq(gpu, gpu_opp);
876 	dev_pm_opp_put(gpu_opp);
877 }
878 
879 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
880 {
881 	struct dev_pm_opp *gpu_opp;
882 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
883 
884 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
885 	if (IS_ERR_OR_NULL(gpu_opp))
886 		return;
887 
888 	dev_pm_opp_set_opp(&gpu->pdev->dev, gpu_opp);
889 	dev_pm_opp_put(gpu_opp);
890 }
891 
892 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
893 {
894 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
895 	struct msm_gpu *gpu = &adreno_gpu->base;
896 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
897 	int status, ret;
898 
899 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
900 		return 0;
901 
902 	gmu->hung = false;
903 
904 	/* Turn on the resources */
905 	pm_runtime_get_sync(gmu->dev);
906 
907 	/*
908 	 * "enable" the GX power domain which won't actually do anything but it
909 	 * will make sure that the refcounting is correct in case we need to
910 	 * bring down the GX after a GMU failure
911 	 */
912 	if (!IS_ERR_OR_NULL(gmu->gxpd))
913 		pm_runtime_get_sync(gmu->gxpd);
914 
915 	/* Use a known rate to bring up the GMU */
916 	clk_set_rate(gmu->core_clk, 200000000);
917 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
918 	if (ret) {
919 		pm_runtime_put(gmu->gxpd);
920 		pm_runtime_put(gmu->dev);
921 		return ret;
922 	}
923 
924 	/* Set the bus quota to a reasonable value for boot */
925 	a6xx_gmu_set_initial_bw(gpu, gmu);
926 
927 	/* Enable the GMU interrupt */
928 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
929 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
930 	enable_irq(gmu->gmu_irq);
931 
932 	/* Check to see if we are doing a cold or warm boot */
933 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
934 		GMU_WARM_BOOT : GMU_COLD_BOOT;
935 
936 	/*
937 	 * Warm boot path does not work on newer GPUs
938 	 * Presumably this is because icache/dcache regions must be restored
939 	 */
940 	if (!gmu->legacy)
941 		status = GMU_COLD_BOOT;
942 
943 	ret = a6xx_gmu_fw_start(gmu, status);
944 	if (ret)
945 		goto out;
946 
947 	ret = a6xx_hfi_start(gmu, status);
948 	if (ret)
949 		goto out;
950 
951 	/*
952 	 * Turn on the GMU firmware fault interrupt after we know the boot
953 	 * sequence is successful
954 	 */
955 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
956 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
957 	enable_irq(gmu->hfi_irq);
958 
959 	/* Set the GPU to the current freq */
960 	a6xx_gmu_set_initial_freq(gpu, gmu);
961 
962 out:
963 	/* On failure, shut down the GMU to leave it in a good state */
964 	if (ret) {
965 		disable_irq(gmu->gmu_irq);
966 		a6xx_rpmh_stop(gmu);
967 		pm_runtime_put(gmu->gxpd);
968 		pm_runtime_put(gmu->dev);
969 	}
970 
971 	return ret;
972 }
973 
974 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
975 {
976 	u32 reg;
977 
978 	if (!gmu->initialized)
979 		return true;
980 
981 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
982 
983 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
984 		return false;
985 
986 	return true;
987 }
988 
989 #define GBIF_CLIENT_HALT_MASK             BIT(0)
990 #define GBIF_ARB_HALT_MASK                BIT(1)
991 
992 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
993 {
994 	struct msm_gpu *gpu = &adreno_gpu->base;
995 
996 	if (!a6xx_has_gbif(adreno_gpu)) {
997 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
998 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
999 								0xf) == 0xf);
1000 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
1001 
1002 		return;
1003 	}
1004 
1005 	/* Halt new client requests on GBIF */
1006 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
1007 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
1008 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
1009 
1010 	/* Halt all AXI requests on GBIF */
1011 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
1012 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
1013 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
1014 
1015 	/* The GBIF halt needs to be explicitly cleared */
1016 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
1017 }
1018 
1019 /* Gracefully try to shut down the GMU and by extension the GPU */
1020 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1021 {
1022 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1023 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1024 	u32 val;
1025 
1026 	/*
1027 	 * The GMU may still be in slumber unless the GPU started so check and
1028 	 * skip putting it back into slumber if so
1029 	 */
1030 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1031 
1032 	if (val != 0xf) {
1033 		int ret = a6xx_gmu_wait_for_idle(gmu);
1034 
1035 		/* If the GMU isn't responding assume it is hung */
1036 		if (ret) {
1037 			a6xx_gmu_force_off(gmu);
1038 			return;
1039 		}
1040 
1041 		a6xx_bus_clear_pending_transactions(adreno_gpu);
1042 
1043 		/* tell the GMU we want to slumber */
1044 		a6xx_gmu_notify_slumber(gmu);
1045 
1046 		ret = gmu_poll_timeout(gmu,
1047 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1048 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1049 			100, 10000);
1050 
1051 		/*
1052 		 * Let the user know we failed to slumber but don't worry too
1053 		 * much because we are powering down anyway
1054 		 */
1055 
1056 		if (ret)
1057 			DRM_DEV_ERROR(gmu->dev,
1058 				"Unable to slumber GMU: status = 0%x/0%x\n",
1059 				gmu_read(gmu,
1060 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1061 				gmu_read(gmu,
1062 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1063 	}
1064 
1065 	/* Turn off HFI */
1066 	a6xx_hfi_stop(gmu);
1067 
1068 	/* Stop the interrupts and mask the hardware */
1069 	a6xx_gmu_irq_disable(gmu);
1070 
1071 	/* Tell RPMh to power off the GPU */
1072 	a6xx_rpmh_stop(gmu);
1073 }
1074 
1075 
1076 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1077 {
1078 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1079 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1080 
1081 	if (!pm_runtime_active(gmu->dev))
1082 		return 0;
1083 
1084 	/*
1085 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1086 	 * down gracefully
1087 	 */
1088 	if (gmu->hung)
1089 		a6xx_gmu_force_off(gmu);
1090 	else
1091 		a6xx_gmu_shutdown(gmu);
1092 
1093 	/* Remove the bus vote */
1094 	dev_pm_opp_set_opp(&gpu->pdev->dev, NULL);
1095 
1096 	/*
1097 	 * Make sure the GX domain is off before turning off the GMU (CX)
1098 	 * domain. Usually the GMU does this but only if the shutdown sequence
1099 	 * was successful
1100 	 */
1101 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1102 		pm_runtime_put_sync(gmu->gxpd);
1103 
1104 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1105 
1106 	pm_runtime_put_sync(gmu->dev);
1107 
1108 	return 0;
1109 }
1110 
1111 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1112 {
1113 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace, false);
1114 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace, false);
1115 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace, false);
1116 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace, false);
1117 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace, false);
1118 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace, false);
1119 
1120 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1121 	msm_gem_address_space_put(gmu->aspace);
1122 }
1123 
1124 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1125 		size_t size, u64 iova)
1126 {
1127 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1128 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1129 	uint32_t flags = MSM_BO_WC;
1130 	u64 range_start, range_end;
1131 	int ret;
1132 
1133 	size = PAGE_ALIGN(size);
1134 	if (!iova) {
1135 		/* no fixed address - use GMU's uncached range */
1136 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1137 		range_end = 0x80000000;
1138 	} else {
1139 		/* range for fixed address */
1140 		range_start = iova;
1141 		range_end = iova + size;
1142 		/* use IOMMU_PRIV for icache/dcache */
1143 		flags |= MSM_BO_MAP_PRIV;
1144 	}
1145 
1146 	bo->obj = msm_gem_new(dev, size, flags);
1147 	if (IS_ERR(bo->obj))
1148 		return PTR_ERR(bo->obj);
1149 
1150 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1151 		range_start >> PAGE_SHIFT, range_end >> PAGE_SHIFT);
1152 	if (ret) {
1153 		drm_gem_object_put(bo->obj);
1154 		return ret;
1155 	}
1156 
1157 	bo->virt = msm_gem_get_vaddr(bo->obj);
1158 	bo->size = size;
1159 
1160 	return 0;
1161 }
1162 
1163 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1164 {
1165 	struct iommu_domain *domain;
1166 	struct msm_mmu *mmu;
1167 
1168 	domain = iommu_domain_alloc(&platform_bus_type);
1169 	if (!domain)
1170 		return -ENODEV;
1171 
1172 	mmu = msm_iommu_new(gmu->dev, domain);
1173 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1174 	if (IS_ERR(gmu->aspace)) {
1175 		iommu_domain_free(domain);
1176 		return PTR_ERR(gmu->aspace);
1177 	}
1178 
1179 	return 0;
1180 }
1181 
1182 /* Return the 'arc-level' for the given frequency */
1183 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1184 					   unsigned long freq)
1185 {
1186 	struct dev_pm_opp *opp;
1187 	unsigned int val;
1188 
1189 	if (!freq)
1190 		return 0;
1191 
1192 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1193 	if (IS_ERR(opp))
1194 		return 0;
1195 
1196 	val = dev_pm_opp_get_level(opp);
1197 
1198 	dev_pm_opp_put(opp);
1199 
1200 	return val;
1201 }
1202 
1203 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1204 		unsigned long *freqs, int freqs_count, const char *id)
1205 {
1206 	int i, j;
1207 	const u16 *pri, *sec;
1208 	size_t pri_count, sec_count;
1209 
1210 	pri = cmd_db_read_aux_data(id, &pri_count);
1211 	if (IS_ERR(pri))
1212 		return PTR_ERR(pri);
1213 	/*
1214 	 * The data comes back as an array of unsigned shorts so adjust the
1215 	 * count accordingly
1216 	 */
1217 	pri_count >>= 1;
1218 	if (!pri_count)
1219 		return -EINVAL;
1220 
1221 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1222 	if (IS_ERR(sec))
1223 		return PTR_ERR(sec);
1224 
1225 	sec_count >>= 1;
1226 	if (!sec_count)
1227 		return -EINVAL;
1228 
1229 	/* Construct a vote for each frequency */
1230 	for (i = 0; i < freqs_count; i++) {
1231 		u8 pindex = 0, sindex = 0;
1232 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1233 
1234 		/* Get the primary index that matches the arc level */
1235 		for (j = 0; j < pri_count; j++) {
1236 			if (pri[j] >= level) {
1237 				pindex = j;
1238 				break;
1239 			}
1240 		}
1241 
1242 		if (j == pri_count) {
1243 			DRM_DEV_ERROR(dev,
1244 				      "Level %u not found in the RPMh list\n",
1245 				      level);
1246 			DRM_DEV_ERROR(dev, "Available levels:\n");
1247 			for (j = 0; j < pri_count; j++)
1248 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1249 
1250 			return -EINVAL;
1251 		}
1252 
1253 		/*
1254 		 * Look for a level in in the secondary list that matches. If
1255 		 * nothing fits, use the maximum non zero vote
1256 		 */
1257 
1258 		for (j = 0; j < sec_count; j++) {
1259 			if (sec[j] >= level) {
1260 				sindex = j;
1261 				break;
1262 			} else if (sec[j]) {
1263 				sindex = j;
1264 			}
1265 		}
1266 
1267 		/* Construct the vote */
1268 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1269 			(sindex << 8) | pindex;
1270 	}
1271 
1272 	return 0;
1273 }
1274 
1275 /*
1276  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1277  * to construct the list of votes on the CPU and send it over. Query the RPMh
1278  * voltage levels and build the votes
1279  */
1280 
1281 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1282 {
1283 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1284 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1285 	struct msm_gpu *gpu = &adreno_gpu->base;
1286 	int ret;
1287 
1288 	/* Build the GX votes */
1289 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1290 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1291 
1292 	/* Build the CX votes */
1293 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1294 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1295 
1296 	return ret;
1297 }
1298 
1299 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1300 		u32 size)
1301 {
1302 	int count = dev_pm_opp_get_opp_count(dev);
1303 	struct dev_pm_opp *opp;
1304 	int i, index = 0;
1305 	unsigned long freq = 1;
1306 
1307 	/*
1308 	 * The OPP table doesn't contain the "off" frequency level so we need to
1309 	 * add 1 to the table size to account for it
1310 	 */
1311 
1312 	if (WARN(count + 1 > size,
1313 		"The GMU frequency table is being truncated\n"))
1314 		count = size - 1;
1315 
1316 	/* Set the "off" frequency */
1317 	freqs[index++] = 0;
1318 
1319 	for (i = 0; i < count; i++) {
1320 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1321 		if (IS_ERR(opp))
1322 			break;
1323 
1324 		dev_pm_opp_put(opp);
1325 		freqs[index++] = freq++;
1326 	}
1327 
1328 	return index;
1329 }
1330 
1331 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1332 {
1333 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1334 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1335 	struct msm_gpu *gpu = &adreno_gpu->base;
1336 
1337 	int ret = 0;
1338 
1339 	/*
1340 	 * The GMU handles its own frequency switching so build a list of
1341 	 * available frequencies to send during initialization
1342 	 */
1343 	ret = dev_pm_opp_of_add_table(gmu->dev);
1344 	if (ret) {
1345 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1346 		return ret;
1347 	}
1348 
1349 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1350 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1351 
1352 	/*
1353 	 * The GMU also handles GPU frequency switching so build a list
1354 	 * from the GPU OPP table
1355 	 */
1356 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1357 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1358 
1359 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1360 
1361 	/* Build the list of RPMh votes that we'll send to the GMU */
1362 	return a6xx_gmu_rpmh_votes_init(gmu);
1363 }
1364 
1365 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1366 {
1367 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1368 
1369 	if (ret < 1)
1370 		return ret;
1371 
1372 	gmu->nr_clocks = ret;
1373 
1374 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1375 		gmu->nr_clocks, "gmu");
1376 
1377 	return 0;
1378 }
1379 
1380 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1381 		const char *name)
1382 {
1383 	void __iomem *ret;
1384 	struct resource *res = platform_get_resource_byname(pdev,
1385 			IORESOURCE_MEM, name);
1386 
1387 	if (!res) {
1388 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1389 		return ERR_PTR(-EINVAL);
1390 	}
1391 
1392 	ret = ioremap(res->start, resource_size(res));
1393 	if (!ret) {
1394 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1395 		return ERR_PTR(-EINVAL);
1396 	}
1397 
1398 	return ret;
1399 }
1400 
1401 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1402 		const char *name, irq_handler_t handler)
1403 {
1404 	int irq, ret;
1405 
1406 	irq = platform_get_irq_byname(pdev, name);
1407 
1408 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1409 	if (ret) {
1410 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1411 			      name, ret);
1412 		return ret;
1413 	}
1414 
1415 	disable_irq(irq);
1416 
1417 	return irq;
1418 }
1419 
1420 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1421 {
1422 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1423 	struct platform_device *pdev = to_platform_device(gmu->dev);
1424 
1425 	if (!gmu->initialized)
1426 		return;
1427 
1428 	pm_runtime_force_suspend(gmu->dev);
1429 
1430 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1431 		pm_runtime_disable(gmu->gxpd);
1432 		dev_pm_domain_detach(gmu->gxpd, false);
1433 	}
1434 
1435 	iounmap(gmu->mmio);
1436 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1437 		iounmap(gmu->rscc);
1438 	gmu->mmio = NULL;
1439 	gmu->rscc = NULL;
1440 
1441 	a6xx_gmu_memory_free(gmu);
1442 
1443 	free_irq(gmu->gmu_irq, gmu);
1444 	free_irq(gmu->hfi_irq, gmu);
1445 
1446 	/* Drop reference taken in of_find_device_by_node */
1447 	put_device(gmu->dev);
1448 
1449 	gmu->initialized = false;
1450 }
1451 
1452 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1453 {
1454 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1455 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1456 	struct platform_device *pdev = of_find_device_by_node(node);
1457 	int ret;
1458 
1459 	if (!pdev)
1460 		return -ENODEV;
1461 
1462 	gmu->dev = &pdev->dev;
1463 
1464 	of_dma_configure(gmu->dev, node, true);
1465 
1466 	/* Fow now, don't do anything fancy until we get our feet under us */
1467 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1468 
1469 	pm_runtime_enable(gmu->dev);
1470 
1471 	/* Get the list of clocks */
1472 	ret = a6xx_gmu_clocks_probe(gmu);
1473 	if (ret)
1474 		goto err_put_device;
1475 
1476 	ret = a6xx_gmu_memory_probe(gmu);
1477 	if (ret)
1478 		goto err_put_device;
1479 
1480 	/* Allocate memory for the GMU dummy page */
1481 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, SZ_4K, 0x60000000);
1482 	if (ret)
1483 		goto err_memory;
1484 
1485 	if (adreno_is_a650(adreno_gpu)) {
1486 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1487 			SZ_16M - SZ_16K, 0x04000);
1488 		if (ret)
1489 			goto err_memory;
1490 	} else if (adreno_is_a640(adreno_gpu)) {
1491 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1492 			SZ_256K - SZ_16K, 0x04000);
1493 		if (ret)
1494 			goto err_memory;
1495 
1496 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1497 			SZ_256K - SZ_16K, 0x44000);
1498 		if (ret)
1499 			goto err_memory;
1500 	} else {
1501 		/* HFI v1, has sptprac */
1502 		gmu->legacy = true;
1503 
1504 		/* Allocate memory for the GMU debug region */
1505 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0);
1506 		if (ret)
1507 			goto err_memory;
1508 	}
1509 
1510 	/* Allocate memory for for the HFI queues */
1511 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0);
1512 	if (ret)
1513 		goto err_memory;
1514 
1515 	/* Allocate memory for the GMU log region */
1516 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0);
1517 	if (ret)
1518 		goto err_memory;
1519 
1520 	/* Map the GMU registers */
1521 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1522 	if (IS_ERR(gmu->mmio)) {
1523 		ret = PTR_ERR(gmu->mmio);
1524 		goto err_memory;
1525 	}
1526 
1527 	if (adreno_is_a650(adreno_gpu)) {
1528 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1529 		if (IS_ERR(gmu->rscc))
1530 			goto err_mmio;
1531 	} else {
1532 		gmu->rscc = gmu->mmio + 0x23000;
1533 	}
1534 
1535 	/* Get the HFI and GMU interrupts */
1536 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1537 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1538 
1539 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1540 		goto err_mmio;
1541 
1542 	/*
1543 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1544 	 * crash
1545 	 */
1546 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1547 
1548 	/* Get the power levels for the GMU and GPU */
1549 	a6xx_gmu_pwrlevels_probe(gmu);
1550 
1551 	/* Set up the HFI queues */
1552 	a6xx_hfi_init(gmu);
1553 
1554 	gmu->initialized = true;
1555 
1556 	return 0;
1557 
1558 err_mmio:
1559 	iounmap(gmu->mmio);
1560 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1561 		iounmap(gmu->rscc);
1562 	free_irq(gmu->gmu_irq, gmu);
1563 	free_irq(gmu->hfi_irq, gmu);
1564 
1565 	ret = -ENODEV;
1566 
1567 err_memory:
1568 	a6xx_gmu_memory_free(gmu);
1569 err_put_device:
1570 	/* Drop reference taken in of_find_device_by_node */
1571 	put_device(gmu->dev);
1572 
1573 	return ret;
1574 }
1575