xref: /openbmc/linux/drivers/gpu/drm/msm/adreno/a6xx_gmu.c (revision e15a5365)
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 	struct drm_device *dev = gpu->dev;
23 	struct msm_drm_private *priv = dev->dev_private;
24 
25 	/* FIXME: add a banner here */
26 	gmu->hung = true;
27 
28 	/* Turn off the hangcheck timer while we are resetting */
29 	del_timer(&gpu->hangcheck_timer);
30 
31 	/* Queue the GPU handler because we need to treat this as a recovery */
32 	queue_work(priv->wq, &gpu->recover_work);
33 }
34 
35 static irqreturn_t a6xx_gmu_irq(int irq, void *data)
36 {
37 	struct a6xx_gmu *gmu = data;
38 	u32 status;
39 
40 	status = gmu_read(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_STATUS);
41 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, status);
42 
43 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE) {
44 		dev_err_ratelimited(gmu->dev, "GMU watchdog expired\n");
45 
46 		a6xx_gmu_fault(gmu);
47 	}
48 
49 	if (status &  A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR)
50 		dev_err_ratelimited(gmu->dev, "GMU AHB bus error\n");
51 
52 	if (status & A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
53 		dev_err_ratelimited(gmu->dev, "GMU fence error: 0x%x\n",
54 			gmu_read(gmu, REG_A6XX_GMU_AHB_FENCE_STATUS));
55 
56 	return IRQ_HANDLED;
57 }
58 
59 static irqreturn_t a6xx_hfi_irq(int irq, void *data)
60 {
61 	struct a6xx_gmu *gmu = data;
62 	u32 status;
63 
64 	status = gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO);
65 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, status);
66 
67 	if (status & A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT) {
68 		dev_err_ratelimited(gmu->dev, "GMU firmware fault\n");
69 
70 		a6xx_gmu_fault(gmu);
71 	}
72 
73 	return IRQ_HANDLED;
74 }
75 
76 bool a6xx_gmu_sptprac_is_on(struct a6xx_gmu *gmu)
77 {
78 	u32 val;
79 
80 	/* This can be called from gpu state code so make sure GMU is valid */
81 	if (!gmu->initialized)
82 		return false;
83 
84 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
85 
86 	return !(val &
87 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SPTPRAC_GDSC_POWER_OFF |
88 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_SP_CLOCK_OFF));
89 }
90 
91 /* Check to see if the GX rail is still powered */
92 bool a6xx_gmu_gx_is_on(struct a6xx_gmu *gmu)
93 {
94 	u32 val;
95 
96 	/* This can be called from gpu state code so make sure GMU is valid */
97 	if (!gmu->initialized)
98 		return false;
99 
100 	val = gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS);
101 
102 	return !(val &
103 		(A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_GDSC_POWER_OFF |
104 		A6XX_GMU_SPTPRAC_PWR_CLK_STATUS_GX_HM_CLK_OFF));
105 }
106 
107 void a6xx_gmu_set_freq(struct msm_gpu *gpu, struct dev_pm_opp *opp)
108 {
109 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
110 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
111 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
112 	u32 perf_index;
113 	unsigned long gpu_freq;
114 	int ret = 0;
115 
116 	gpu_freq = dev_pm_opp_get_freq(opp);
117 
118 	if (gpu_freq == gmu->freq)
119 		return;
120 
121 	for (perf_index = 0; perf_index < gmu->nr_gpu_freqs - 1; perf_index++)
122 		if (gpu_freq == gmu->gpu_freqs[perf_index])
123 			break;
124 
125 	gmu->current_perf_index = perf_index;
126 	gmu->freq = gmu->gpu_freqs[perf_index];
127 
128 	trace_msm_gmu_freq_change(gmu->freq, perf_index);
129 
130 	/*
131 	 * This can get called from devfreq while the hardware is idle. Don't
132 	 * bring up the power if it isn't already active
133 	 */
134 	if (pm_runtime_get_if_in_use(gmu->dev) == 0)
135 		return;
136 
137 	if (!gmu->legacy) {
138 		a6xx_hfi_set_freq(gmu, perf_index);
139 		dev_pm_opp_set_bw(&gpu->pdev->dev, opp);
140 		pm_runtime_put(gmu->dev);
141 		return;
142 	}
143 
144 	gmu_write(gmu, REG_A6XX_GMU_DCVS_ACK_OPTION, 0);
145 
146 	gmu_write(gmu, REG_A6XX_GMU_DCVS_PERF_SETTING,
147 			((3 & 0xf) << 28) | perf_index);
148 
149 	/*
150 	 * Send an invalid index as a vote for the bus bandwidth and let the
151 	 * firmware decide on the right vote
152 	 */
153 	gmu_write(gmu, REG_A6XX_GMU_DCVS_BW_SETTING, 0xff);
154 
155 	/* Set and clear the OOB for DCVS to trigger the GMU */
156 	a6xx_gmu_set_oob(gmu, GMU_OOB_DCVS_SET);
157 	a6xx_gmu_clear_oob(gmu, GMU_OOB_DCVS_SET);
158 
159 	ret = gmu_read(gmu, REG_A6XX_GMU_DCVS_RETURN);
160 	if (ret)
161 		dev_err(gmu->dev, "GMU set GPU frequency error: %d\n", ret);
162 
163 	dev_pm_opp_set_bw(&gpu->pdev->dev, opp);
164 	pm_runtime_put(gmu->dev);
165 }
166 
167 unsigned long a6xx_gmu_get_freq(struct msm_gpu *gpu)
168 {
169 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
170 	struct a6xx_gpu *a6xx_gpu = to_a6xx_gpu(adreno_gpu);
171 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
172 
173 	return  gmu->freq;
174 }
175 
176 static bool a6xx_gmu_check_idle_level(struct a6xx_gmu *gmu)
177 {
178 	u32 val;
179 	int local = gmu->idle_level;
180 
181 	/* SPTP and IFPC both report as IFPC */
182 	if (gmu->idle_level == GMU_IDLE_STATE_SPTP)
183 		local = GMU_IDLE_STATE_IFPC;
184 
185 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
186 
187 	if (val == local) {
188 		if (gmu->idle_level != GMU_IDLE_STATE_IFPC ||
189 			!a6xx_gmu_gx_is_on(gmu))
190 			return true;
191 	}
192 
193 	return false;
194 }
195 
196 /* Wait for the GMU to get to its most idle state */
197 int a6xx_gmu_wait_for_idle(struct a6xx_gmu *gmu)
198 {
199 	return spin_until(a6xx_gmu_check_idle_level(gmu));
200 }
201 
202 static int a6xx_gmu_start(struct a6xx_gmu *gmu)
203 {
204 	int ret;
205 	u32 val;
206 	u32 mask, reset_val;
207 
208 	val = gmu_read(gmu, REG_A6XX_GMU_CM3_DTCM_START + 0xff8);
209 	if (val <= 0x20010004) {
210 		mask = 0xffffffff;
211 		reset_val = 0xbabeface;
212 	} else {
213 		mask = 0x1ff;
214 		reset_val = 0x100;
215 	}
216 
217 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 1);
218 
219 	/* Set the log wptr index
220 	 * note: downstream saves the value in poweroff and restores it here
221 	 */
222 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_RESP, 0);
223 
224 	gmu_write(gmu, REG_A6XX_GMU_CM3_SYSRESET, 0);
225 
226 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, val,
227 		(val & mask) == reset_val, 100, 10000);
228 
229 	if (ret)
230 		DRM_DEV_ERROR(gmu->dev, "GMU firmware initialization timed out\n");
231 
232 	return ret;
233 }
234 
235 static int a6xx_gmu_hfi_start(struct a6xx_gmu *gmu)
236 {
237 	u32 val;
238 	int ret;
239 
240 	gmu_write(gmu, REG_A6XX_GMU_HFI_CTRL_INIT, 1);
241 
242 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_HFI_CTRL_STATUS, val,
243 		val & 1, 100, 10000);
244 	if (ret)
245 		DRM_DEV_ERROR(gmu->dev, "Unable to start the HFI queues\n");
246 
247 	return ret;
248 }
249 
250 /* Trigger a OOB (out of band) request to the GMU */
251 int a6xx_gmu_set_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
252 {
253 	int ret;
254 	u32 val;
255 	int request, ack;
256 	const char *name;
257 
258 	switch (state) {
259 	case GMU_OOB_GPU_SET:
260 		if (gmu->legacy) {
261 			request = GMU_OOB_GPU_SET_REQUEST;
262 			ack = GMU_OOB_GPU_SET_ACK;
263 		} else {
264 			request = GMU_OOB_GPU_SET_REQUEST_NEW;
265 			ack = GMU_OOB_GPU_SET_ACK_NEW;
266 		}
267 		name = "GPU_SET";
268 		break;
269 	case GMU_OOB_BOOT_SLUMBER:
270 		request = GMU_OOB_BOOT_SLUMBER_REQUEST;
271 		ack = GMU_OOB_BOOT_SLUMBER_ACK;
272 		name = "BOOT_SLUMBER";
273 		break;
274 	case GMU_OOB_DCVS_SET:
275 		request = GMU_OOB_DCVS_REQUEST;
276 		ack = GMU_OOB_DCVS_ACK;
277 		name = "GPU_DCVS";
278 		break;
279 	default:
280 		return -EINVAL;
281 	}
282 
283 	/* Trigger the equested OOB operation */
284 	gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET, 1 << request);
285 
286 	/* Wait for the acknowledge interrupt */
287 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO, val,
288 		val & (1 << ack), 100, 10000);
289 
290 	if (ret)
291 		DRM_DEV_ERROR(gmu->dev,
292 			"Timeout waiting for GMU OOB set %s: 0x%x\n",
293 				name,
294 				gmu_read(gmu, REG_A6XX_GMU_GMU2HOST_INTR_INFO));
295 
296 	/* Clear the acknowledge interrupt */
297 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, 1 << ack);
298 
299 	return ret;
300 }
301 
302 /* Clear a pending OOB state in the GMU */
303 void a6xx_gmu_clear_oob(struct a6xx_gmu *gmu, enum a6xx_gmu_oob_state state)
304 {
305 	if (!gmu->legacy) {
306 		WARN_ON(state != GMU_OOB_GPU_SET);
307 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
308 			1 << GMU_OOB_GPU_SET_CLEAR_NEW);
309 		return;
310 	}
311 
312 	switch (state) {
313 	case GMU_OOB_GPU_SET:
314 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
315 			1 << GMU_OOB_GPU_SET_CLEAR);
316 		break;
317 	case GMU_OOB_BOOT_SLUMBER:
318 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
319 			1 << GMU_OOB_BOOT_SLUMBER_CLEAR);
320 		break;
321 	case GMU_OOB_DCVS_SET:
322 		gmu_write(gmu, REG_A6XX_GMU_HOST2GMU_INTR_SET,
323 			1 << GMU_OOB_DCVS_CLEAR);
324 		break;
325 	}
326 }
327 
328 /* Enable CPU control of SPTP power power collapse */
329 static int a6xx_sptprac_enable(struct a6xx_gmu *gmu)
330 {
331 	int ret;
332 	u32 val;
333 
334 	if (!gmu->legacy)
335 		return 0;
336 
337 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778000);
338 
339 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
340 		(val & 0x38) == 0x28, 1, 100);
341 
342 	if (ret) {
343 		DRM_DEV_ERROR(gmu->dev, "Unable to power on SPTPRAC: 0x%x\n",
344 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
345 	}
346 
347 	return 0;
348 }
349 
350 /* Disable CPU control of SPTP power power collapse */
351 static void a6xx_sptprac_disable(struct a6xx_gmu *gmu)
352 {
353 	u32 val;
354 	int ret;
355 
356 	if (!gmu->legacy)
357 		return;
358 
359 	/* Make sure retention is on */
360 	gmu_rmw(gmu, REG_A6XX_GPU_CC_GX_GDSCR, 0, (1 << 11));
361 
362 	gmu_write(gmu, REG_A6XX_GMU_GX_SPTPRAC_POWER_CONTROL, 0x778001);
363 
364 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS, val,
365 		(val & 0x04), 100, 10000);
366 
367 	if (ret)
368 		DRM_DEV_ERROR(gmu->dev, "failed to power off SPTPRAC: 0x%x\n",
369 			gmu_read(gmu, REG_A6XX_GMU_SPTPRAC_PWR_CLK_STATUS));
370 }
371 
372 /* Let the GMU know we are starting a boot sequence */
373 static int a6xx_gmu_gfx_rail_on(struct a6xx_gmu *gmu)
374 {
375 	u32 vote;
376 
377 	/* Let the GMU know we are getting ready for boot */
378 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 0);
379 
380 	/* Choose the "default" power level as the highest available */
381 	vote = gmu->gx_arc_votes[gmu->nr_gpu_freqs - 1];
382 
383 	gmu_write(gmu, REG_A6XX_GMU_GX_VOTE_IDX, vote & 0xff);
384 	gmu_write(gmu, REG_A6XX_GMU_MX_VOTE_IDX, (vote >> 8) & 0xff);
385 
386 	/* Let the GMU know the boot sequence has started */
387 	return a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
388 }
389 
390 /* Let the GMU know that we are about to go into slumber */
391 static int a6xx_gmu_notify_slumber(struct a6xx_gmu *gmu)
392 {
393 	int ret;
394 
395 	/* Disable the power counter so the GMU isn't busy */
396 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 0);
397 
398 	/* Disable SPTP_PC if the CPU is responsible for it */
399 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP)
400 		a6xx_sptprac_disable(gmu);
401 
402 	if (!gmu->legacy) {
403 		ret = a6xx_hfi_send_prep_slumber(gmu);
404 		goto out;
405 	}
406 
407 	/* Tell the GMU to get ready to slumber */
408 	gmu_write(gmu, REG_A6XX_GMU_BOOT_SLUMBER_OPTION, 1);
409 
410 	ret = a6xx_gmu_set_oob(gmu, GMU_OOB_BOOT_SLUMBER);
411 	a6xx_gmu_clear_oob(gmu, GMU_OOB_BOOT_SLUMBER);
412 
413 	if (!ret) {
414 		/* Check to see if the GMU really did slumber */
415 		if (gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE)
416 			!= 0x0f) {
417 			DRM_DEV_ERROR(gmu->dev, "The GMU did not go into slumber\n");
418 			ret = -ETIMEDOUT;
419 		}
420 	}
421 
422 out:
423 	/* Put fence into allow mode */
424 	gmu_write(gmu, REG_A6XX_GMU_AO_AHB_FENCE_CTRL, 0);
425 	return ret;
426 }
427 
428 static int a6xx_rpmh_start(struct a6xx_gmu *gmu)
429 {
430 	int ret;
431 	u32 val;
432 
433 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1 << 1);
434 	/* Wait for the register to finish posting */
435 	wmb();
436 
437 	ret = gmu_poll_timeout(gmu, REG_A6XX_GMU_RSCC_CONTROL_ACK, val,
438 		val & (1 << 1), 100, 10000);
439 	if (ret) {
440 		DRM_DEV_ERROR(gmu->dev, "Unable to power on the GPU RSC\n");
441 		return ret;
442 	}
443 
444 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_SEQ_BUSY_DRV0, val,
445 		!val, 100, 10000);
446 
447 	if (ret) {
448 		DRM_DEV_ERROR(gmu->dev, "GPU RSC sequence stuck while waking up the GPU\n");
449 		return ret;
450 	}
451 
452 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
453 
454 	/* Set up CX GMU counter 0 to count busy ticks */
455 	gmu_write(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_MASK, 0xff000000);
456 	gmu_rmw(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_SELECT_0, 0xff, 0x20);
457 
458 	/* Enable the power counter */
459 	gmu_write(gmu, REG_A6XX_GMU_CX_GMU_POWER_COUNTER_ENABLE, 1);
460 	return 0;
461 }
462 
463 static void a6xx_rpmh_stop(struct a6xx_gmu *gmu)
464 {
465 	int ret;
466 	u32 val;
467 
468 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 1);
469 
470 	ret = gmu_poll_timeout_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0,
471 		val, val & (1 << 16), 100, 10000);
472 	if (ret)
473 		DRM_DEV_ERROR(gmu->dev, "Unable to power off the GPU RSC\n");
474 
475 	gmu_write(gmu, REG_A6XX_GMU_RSCC_CONTROL_REQ, 0);
476 }
477 
478 static inline void pdc_write(void __iomem *ptr, u32 offset, u32 value)
479 {
480 	return msm_writel(value, ptr + (offset << 2));
481 }
482 
483 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
484 		const char *name);
485 
486 static void a6xx_gmu_rpmh_init(struct a6xx_gmu *gmu)
487 {
488 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
489 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
490 	struct platform_device *pdev = to_platform_device(gmu->dev);
491 	void __iomem *pdcptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc");
492 	void __iomem *seqptr = a6xx_gmu_get_mmio(pdev, "gmu_pdc_seq");
493 	uint32_t pdc_address_offset;
494 
495 	if (!pdcptr || !seqptr)
496 		goto err;
497 
498 	if (adreno_is_a618(adreno_gpu) || adreno_is_a640(adreno_gpu))
499 		pdc_address_offset = 0x30090;
500 	else if (adreno_is_a650(adreno_gpu))
501 		pdc_address_offset = 0x300a0;
502 	else
503 		pdc_address_offset = 0x30080;
504 
505 	/* Disable SDE clock gating */
506 	gmu_write_rscc(gmu, REG_A6XX_GPU_RSCC_RSC_STATUS0_DRV0, BIT(24));
507 
508 	/* Setup RSC PDC handshake for sleep and wakeup */
509 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SLAVE_ID_DRV0, 1);
510 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA, 0);
511 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR, 0);
512 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 2, 0);
513 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 2, 0);
514 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_DATA + 4, 0x80000000);
515 	gmu_write_rscc(gmu, REG_A6XX_RSCC_HIDDEN_TCS_CMD0_ADDR + 4, 0);
516 	gmu_write_rscc(gmu, REG_A6XX_RSCC_OVERRIDE_START_ADDR, 0);
517 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_SEQ_START_ADDR, 0x4520);
518 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_LO, 0x4510);
519 	gmu_write_rscc(gmu, REG_A6XX_RSCC_PDC_MATCH_VALUE_HI, 0x4514);
520 
521 	/* Load RSC sequencer uCode for sleep and wakeup */
522 	if (adreno_is_a650(adreno_gpu)) {
523 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xeaaae5a0);
524 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xe1a1ebab);
525 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e0a581);
526 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xecac82e2);
527 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020edad);
528 	} else {
529 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0, 0xa7a506a0);
530 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 1, 0xa1e6a6e7);
531 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 2, 0xa2e081e1);
532 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 3, 0xe9a982e2);
533 		gmu_write_rscc(gmu, REG_A6XX_RSCC_SEQ_MEM_0_DRV0 + 4, 0x0020e8a8);
534 	}
535 
536 	/* Load PDC sequencer uCode for power up and power down sequence */
537 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0, 0xfebea1e1);
538 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 1, 0xa5a4a3a2);
539 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 2, 0x8382a6e0);
540 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 3, 0xbce3e284);
541 	pdc_write(seqptr, REG_A6XX_PDC_GPU_SEQ_MEM_0 + 4, 0x002081fc);
542 
543 	/* Set TCS commands used by PDC sequence for low power modes */
544 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_ENABLE_BANK, 7);
545 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD_WAIT_FOR_CMPL_BANK, 0);
546 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CONTROL, 0);
547 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID, 0x10108);
548 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR, 0x30010);
549 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA, 1);
550 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 4, 0x10108);
551 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 4, 0x30000);
552 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 4, 0x0);
553 
554 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_MSGID + 8, 0x10108);
555 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_ADDR + 8, pdc_address_offset);
556 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS1_CMD0_DATA + 8, 0x0);
557 
558 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_ENABLE_BANK, 7);
559 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD_WAIT_FOR_CMPL_BANK, 0);
560 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CONTROL, 0);
561 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID, 0x10108);
562 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR, 0x30010);
563 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA, 2);
564 
565 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 4, 0x10108);
566 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 4, 0x30000);
567 	if (adreno_is_a618(adreno_gpu) || adreno_is_a650(adreno_gpu))
568 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x2);
569 	else
570 		pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 4, 0x3);
571 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_MSGID + 8, 0x10108);
572 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_ADDR + 8, pdc_address_offset);
573 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_TCS3_CMD0_DATA + 8, 0x3);
574 
575 	/* Setup GPU PDC */
576 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_SEQ_START_ADDR, 0);
577 	pdc_write(pdcptr, REG_A6XX_PDC_GPU_ENABLE_PDC, 0x80000001);
578 
579 	/* ensure no writes happen before the uCode is fully written */
580 	wmb();
581 
582 err:
583 	if (!IS_ERR_OR_NULL(pdcptr))
584 		iounmap(pdcptr);
585 	if (!IS_ERR_OR_NULL(seqptr))
586 		iounmap(seqptr);
587 }
588 
589 /*
590  * The lowest 16 bits of this value are the number of XO clock cycles for main
591  * hysteresis which is set at 0x1680 cycles (300 us).  The higher 16 bits are
592  * for the shorter hysteresis that happens after main - this is 0xa (.5 us)
593  */
594 
595 #define GMU_PWR_COL_HYST 0x000a1680
596 
597 /* Set up the idle state for the GMU */
598 static void a6xx_gmu_power_config(struct a6xx_gmu *gmu)
599 {
600 	/* Disable GMU WB/RB buffer */
601 	gmu_write(gmu, REG_A6XX_GMU_SYS_BUS_CONFIG, 0x1);
602 	gmu_write(gmu, REG_A6XX_GMU_ICACHE_CONFIG, 0x1);
603 	gmu_write(gmu, REG_A6XX_GMU_DCACHE_CONFIG, 0x1);
604 
605 	gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0x9c40400);
606 
607 	switch (gmu->idle_level) {
608 	case GMU_IDLE_STATE_IFPC:
609 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_HYST,
610 			GMU_PWR_COL_HYST);
611 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
612 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
613 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_HM_POWER_COLLAPSE_ENABLE);
614 		fallthrough;
615 	case GMU_IDLE_STATE_SPTP:
616 		gmu_write(gmu, REG_A6XX_GMU_PWR_COL_SPTPRAC_HYST,
617 			GMU_PWR_COL_HYST);
618 		gmu_rmw(gmu, REG_A6XX_GMU_PWR_COL_INTER_FRAME_CTRL, 0,
619 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_IFPC_ENABLE |
620 			A6XX_GMU_PWR_COL_INTER_FRAME_CTRL_SPTPRAC_POWER_CONTROL_ENABLE);
621 	}
622 
623 	/* Enable RPMh GPU client */
624 	gmu_rmw(gmu, REG_A6XX_GMU_RPMH_CTRL, 0,
625 		A6XX_GMU_RPMH_CTRL_RPMH_INTERFACE_ENABLE |
626 		A6XX_GMU_RPMH_CTRL_LLC_VOTE_ENABLE |
627 		A6XX_GMU_RPMH_CTRL_DDR_VOTE_ENABLE |
628 		A6XX_GMU_RPMH_CTRL_MX_VOTE_ENABLE |
629 		A6XX_GMU_RPMH_CTRL_CX_VOTE_ENABLE |
630 		A6XX_GMU_RPMH_CTRL_GFX_VOTE_ENABLE);
631 }
632 
633 struct block_header {
634 	u32 addr;
635 	u32 size;
636 	u32 type;
637 	u32 value;
638 	u32 data[];
639 };
640 
641 /* this should be a general kernel helper */
642 static int in_range(u32 addr, u32 start, u32 size)
643 {
644 	return addr >= start && addr < start + size;
645 }
646 
647 static bool fw_block_mem(struct a6xx_gmu_bo *bo, const struct block_header *blk)
648 {
649 	if (!in_range(blk->addr, bo->iova, bo->size))
650 		return false;
651 
652 	memcpy(bo->virt + blk->addr - bo->iova, blk->data, blk->size);
653 	return true;
654 }
655 
656 static int a6xx_gmu_fw_load(struct a6xx_gmu *gmu)
657 {
658 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
659 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
660 	const struct firmware *fw_image = adreno_gpu->fw[ADRENO_FW_GMU];
661 	const struct block_header *blk;
662 	u32 reg_offset;
663 
664 	u32 itcm_base = 0x00000000;
665 	u32 dtcm_base = 0x00040000;
666 
667 	if (adreno_is_a650(adreno_gpu))
668 		dtcm_base = 0x10004000;
669 
670 	if (gmu->legacy) {
671 		/* Sanity check the size of the firmware that was loaded */
672 		if (fw_image->size > 0x8000) {
673 			DRM_DEV_ERROR(gmu->dev,
674 				"GMU firmware is bigger than the available region\n");
675 			return -EINVAL;
676 		}
677 
678 		gmu_write_bulk(gmu, REG_A6XX_GMU_CM3_ITCM_START,
679 			       (u32*) fw_image->data, fw_image->size);
680 		return 0;
681 	}
682 
683 
684 	for (blk = (const struct block_header *) fw_image->data;
685 	     (const u8*) blk < fw_image->data + fw_image->size;
686 	     blk = (const struct block_header *) &blk->data[blk->size >> 2]) {
687 		if (blk->size == 0)
688 			continue;
689 
690 		if (in_range(blk->addr, itcm_base, SZ_16K)) {
691 			reg_offset = (blk->addr - itcm_base) >> 2;
692 			gmu_write_bulk(gmu,
693 				REG_A6XX_GMU_CM3_ITCM_START + reg_offset,
694 				blk->data, blk->size);
695 		} else if (in_range(blk->addr, dtcm_base, SZ_16K)) {
696 			reg_offset = (blk->addr - dtcm_base) >> 2;
697 			gmu_write_bulk(gmu,
698 				REG_A6XX_GMU_CM3_DTCM_START + reg_offset,
699 				blk->data, blk->size);
700 		} else if (!fw_block_mem(&gmu->icache, blk) &&
701 			   !fw_block_mem(&gmu->dcache, blk) &&
702 			   !fw_block_mem(&gmu->dummy, blk)) {
703 			DRM_DEV_ERROR(gmu->dev,
704 				"failed to match fw block (addr=%.8x size=%d data[0]=%.8x)\n",
705 				blk->addr, blk->size, blk->data[0]);
706 		}
707 	}
708 
709 	return 0;
710 }
711 
712 static int a6xx_gmu_fw_start(struct a6xx_gmu *gmu, unsigned int state)
713 {
714 	static bool rpmh_init;
715 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
716 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
717 	int ret;
718 	u32 chipid;
719 
720 	if (adreno_is_a650(adreno_gpu))
721 		gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_CX_FAL_INTF, 1);
722 
723 	if (state == GMU_WARM_BOOT) {
724 		ret = a6xx_rpmh_start(gmu);
725 		if (ret)
726 			return ret;
727 	} else {
728 		if (WARN(!adreno_gpu->fw[ADRENO_FW_GMU],
729 			"GMU firmware is not loaded\n"))
730 			return -ENOENT;
731 
732 		/* Turn on register retention */
733 		gmu_write(gmu, REG_A6XX_GMU_GENERAL_7, 1);
734 
735 		/* We only need to load the RPMh microcode once */
736 		if (!rpmh_init) {
737 			a6xx_gmu_rpmh_init(gmu);
738 			rpmh_init = true;
739 		} else {
740 			ret = a6xx_rpmh_start(gmu);
741 			if (ret)
742 				return ret;
743 		}
744 
745 		ret = a6xx_gmu_fw_load(gmu);
746 		if (ret)
747 			return ret;
748 	}
749 
750 	gmu_write(gmu, REG_A6XX_GMU_CM3_FW_INIT_RESULT, 0);
751 	gmu_write(gmu, REG_A6XX_GMU_CM3_BOOT_CONFIG, 0x02);
752 
753 	/* Write the iova of the HFI table */
754 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_ADDR, gmu->hfi.iova);
755 	gmu_write(gmu, REG_A6XX_GMU_HFI_QTBL_INFO, 1);
756 
757 	gmu_write(gmu, REG_A6XX_GMU_AHB_FENCE_RANGE_0,
758 		(1 << 31) | (0xa << 18) | (0xa0));
759 
760 	chipid = adreno_gpu->rev.core << 24;
761 	chipid |= adreno_gpu->rev.major << 16;
762 	chipid |= adreno_gpu->rev.minor << 12;
763 	chipid |= adreno_gpu->rev.patchid << 8;
764 
765 	gmu_write(gmu, REG_A6XX_GMU_HFI_SFR_ADDR, chipid);
766 
767 	gmu_write(gmu, REG_A6XX_GPU_GMU_CX_GMU_PWR_COL_CP_MSG,
768 		  gmu->log.iova | (gmu->log.size / SZ_4K - 1));
769 
770 	/* Set up the lowest idle level on the GMU */
771 	a6xx_gmu_power_config(gmu);
772 
773 	ret = a6xx_gmu_start(gmu);
774 	if (ret)
775 		return ret;
776 
777 	if (gmu->legacy) {
778 		ret = a6xx_gmu_gfx_rail_on(gmu);
779 		if (ret)
780 			return ret;
781 	}
782 
783 	/* Enable SPTP_PC if the CPU is responsible for it */
784 	if (gmu->idle_level < GMU_IDLE_STATE_SPTP) {
785 		ret = a6xx_sptprac_enable(gmu);
786 		if (ret)
787 			return ret;
788 	}
789 
790 	ret = a6xx_gmu_hfi_start(gmu);
791 	if (ret)
792 		return ret;
793 
794 	/* FIXME: Do we need this wmb() here? */
795 	wmb();
796 
797 	return 0;
798 }
799 
800 #define A6XX_HFI_IRQ_MASK \
801 	(A6XX_GMU_GMU2HOST_INTR_INFO_CM3_FAULT)
802 
803 #define A6XX_GMU_IRQ_MASK \
804 	(A6XX_GMU_AO_HOST_INTERRUPT_STATUS_WDOG_BITE | \
805 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_HOST_AHB_BUS_ERROR | \
806 	 A6XX_GMU_AO_HOST_INTERRUPT_STATUS_FENCE_ERR)
807 
808 static void a6xx_gmu_irq_disable(struct a6xx_gmu *gmu)
809 {
810 	disable_irq(gmu->gmu_irq);
811 	disable_irq(gmu->hfi_irq);
812 
813 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~0);
814 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~0);
815 }
816 
817 static void a6xx_gmu_rpmh_off(struct a6xx_gmu *gmu)
818 {
819 	u32 val;
820 
821 	/* Make sure there are no outstanding RPMh votes */
822 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS0_DRV0_STATUS, val,
823 		(val & 1), 100, 10000);
824 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS1_DRV0_STATUS, val,
825 		(val & 1), 100, 10000);
826 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS2_DRV0_STATUS, val,
827 		(val & 1), 100, 10000);
828 	gmu_poll_timeout_rscc(gmu, REG_A6XX_RSCC_TCS3_DRV0_STATUS, val,
829 		(val & 1), 100, 1000);
830 }
831 
832 /* Force the GMU off in case it isn't responsive */
833 static void a6xx_gmu_force_off(struct a6xx_gmu *gmu)
834 {
835 	/* Flush all the queues */
836 	a6xx_hfi_stop(gmu);
837 
838 	/* Stop the interrupts */
839 	a6xx_gmu_irq_disable(gmu);
840 
841 	/* Force off SPTP in case the GMU is managing it */
842 	a6xx_sptprac_disable(gmu);
843 
844 	/* Make sure there are no outstanding RPMh votes */
845 	a6xx_gmu_rpmh_off(gmu);
846 }
847 
848 static void a6xx_gmu_set_initial_freq(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
849 {
850 	struct dev_pm_opp *gpu_opp;
851 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
852 
853 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
854 	if (IS_ERR_OR_NULL(gpu_opp))
855 		return;
856 
857 	gmu->freq = 0; /* so a6xx_gmu_set_freq() doesn't exit early */
858 	a6xx_gmu_set_freq(gpu, gpu_opp);
859 	dev_pm_opp_put(gpu_opp);
860 }
861 
862 static void a6xx_gmu_set_initial_bw(struct msm_gpu *gpu, struct a6xx_gmu *gmu)
863 {
864 	struct dev_pm_opp *gpu_opp;
865 	unsigned long gpu_freq = gmu->gpu_freqs[gmu->current_perf_index];
866 
867 	gpu_opp = dev_pm_opp_find_freq_exact(&gpu->pdev->dev, gpu_freq, true);
868 	if (IS_ERR_OR_NULL(gpu_opp))
869 		return;
870 
871 	dev_pm_opp_set_bw(&gpu->pdev->dev, gpu_opp);
872 	dev_pm_opp_put(gpu_opp);
873 }
874 
875 int a6xx_gmu_resume(struct a6xx_gpu *a6xx_gpu)
876 {
877 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
878 	struct msm_gpu *gpu = &adreno_gpu->base;
879 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
880 	int status, ret;
881 
882 	if (WARN(!gmu->initialized, "The GMU is not set up yet\n"))
883 		return 0;
884 
885 	gmu->hung = false;
886 
887 	/* Turn on the resources */
888 	pm_runtime_get_sync(gmu->dev);
889 
890 	/*
891 	 * "enable" the GX power domain which won't actually do anything but it
892 	 * will make sure that the refcounting is correct in case we need to
893 	 * bring down the GX after a GMU failure
894 	 */
895 	if (!IS_ERR_OR_NULL(gmu->gxpd))
896 		pm_runtime_get_sync(gmu->gxpd);
897 
898 	/* Use a known rate to bring up the GMU */
899 	clk_set_rate(gmu->core_clk, 200000000);
900 	ret = clk_bulk_prepare_enable(gmu->nr_clocks, gmu->clocks);
901 	if (ret) {
902 		pm_runtime_put(gmu->gxpd);
903 		pm_runtime_put(gmu->dev);
904 		return ret;
905 	}
906 
907 	/* Set the bus quota to a reasonable value for boot */
908 	a6xx_gmu_set_initial_bw(gpu, gmu);
909 
910 	/* Enable the GMU interrupt */
911 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_CLR, ~0);
912 	gmu_write(gmu, REG_A6XX_GMU_AO_HOST_INTERRUPT_MASK, ~A6XX_GMU_IRQ_MASK);
913 	enable_irq(gmu->gmu_irq);
914 
915 	/* Check to see if we are doing a cold or warm boot */
916 	status = gmu_read(gmu, REG_A6XX_GMU_GENERAL_7) == 1 ?
917 		GMU_WARM_BOOT : GMU_COLD_BOOT;
918 
919 	/*
920 	 * Warm boot path does not work on newer GPUs
921 	 * Presumably this is because icache/dcache regions must be restored
922 	 */
923 	if (!gmu->legacy)
924 		status = GMU_COLD_BOOT;
925 
926 	ret = a6xx_gmu_fw_start(gmu, status);
927 	if (ret)
928 		goto out;
929 
930 	ret = a6xx_hfi_start(gmu, status);
931 	if (ret)
932 		goto out;
933 
934 	/*
935 	 * Turn on the GMU firmware fault interrupt after we know the boot
936 	 * sequence is successful
937 	 */
938 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_CLR, ~0);
939 	gmu_write(gmu, REG_A6XX_GMU_GMU2HOST_INTR_MASK, ~A6XX_HFI_IRQ_MASK);
940 	enable_irq(gmu->hfi_irq);
941 
942 	/* Set the GPU to the current freq */
943 	a6xx_gmu_set_initial_freq(gpu, gmu);
944 
945 out:
946 	/* On failure, shut down the GMU to leave it in a good state */
947 	if (ret) {
948 		disable_irq(gmu->gmu_irq);
949 		a6xx_rpmh_stop(gmu);
950 		pm_runtime_put(gmu->gxpd);
951 		pm_runtime_put(gmu->dev);
952 	}
953 
954 	return ret;
955 }
956 
957 bool a6xx_gmu_isidle(struct a6xx_gmu *gmu)
958 {
959 	u32 reg;
960 
961 	if (!gmu->initialized)
962 		return true;
963 
964 	reg = gmu_read(gmu, REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS);
965 
966 	if (reg &  A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB)
967 		return false;
968 
969 	return true;
970 }
971 
972 #define GBIF_CLIENT_HALT_MASK             BIT(0)
973 #define GBIF_ARB_HALT_MASK                BIT(1)
974 
975 static void a6xx_bus_clear_pending_transactions(struct adreno_gpu *adreno_gpu)
976 {
977 	struct msm_gpu *gpu = &adreno_gpu->base;
978 
979 	if (!a6xx_has_gbif(adreno_gpu)) {
980 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0xf);
981 		spin_until((gpu_read(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL1) &
982 								0xf) == 0xf);
983 		gpu_write(gpu, REG_A6XX_VBIF_XIN_HALT_CTRL0, 0);
984 
985 		return;
986 	}
987 
988 	/* Halt new client requests on GBIF */
989 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_CLIENT_HALT_MASK);
990 	spin_until((gpu_read(gpu, REG_A6XX_GBIF_HALT_ACK) &
991 			(GBIF_CLIENT_HALT_MASK)) == GBIF_CLIENT_HALT_MASK);
992 
993 	/* Halt all AXI requests on GBIF */
994 	gpu_write(gpu, REG_A6XX_GBIF_HALT, GBIF_ARB_HALT_MASK);
995 	spin_until((gpu_read(gpu,  REG_A6XX_GBIF_HALT_ACK) &
996 			(GBIF_ARB_HALT_MASK)) == GBIF_ARB_HALT_MASK);
997 
998 	/* The GBIF halt needs to be explicitly cleared */
999 	gpu_write(gpu, REG_A6XX_GBIF_HALT, 0x0);
1000 }
1001 
1002 /* Gracefully try to shut down the GMU and by extension the GPU */
1003 static void a6xx_gmu_shutdown(struct a6xx_gmu *gmu)
1004 {
1005 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1006 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1007 	u32 val;
1008 
1009 	/*
1010 	 * The GMU may still be in slumber unless the GPU started so check and
1011 	 * skip putting it back into slumber if so
1012 	 */
1013 	val = gmu_read(gmu, REG_A6XX_GPU_GMU_CX_GMU_RPMH_POWER_STATE);
1014 
1015 	if (val != 0xf) {
1016 		int ret = a6xx_gmu_wait_for_idle(gmu);
1017 
1018 		/* If the GMU isn't responding assume it is hung */
1019 		if (ret) {
1020 			a6xx_gmu_force_off(gmu);
1021 			return;
1022 		}
1023 
1024 		a6xx_bus_clear_pending_transactions(adreno_gpu);
1025 
1026 		/* tell the GMU we want to slumber */
1027 		a6xx_gmu_notify_slumber(gmu);
1028 
1029 		ret = gmu_poll_timeout(gmu,
1030 			REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS, val,
1031 			!(val & A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS_GPUBUSYIGNAHB),
1032 			100, 10000);
1033 
1034 		/*
1035 		 * Let the user know we failed to slumber but don't worry too
1036 		 * much because we are powering down anyway
1037 		 */
1038 
1039 		if (ret)
1040 			DRM_DEV_ERROR(gmu->dev,
1041 				"Unable to slumber GMU: status = 0%x/0%x\n",
1042 				gmu_read(gmu,
1043 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS),
1044 				gmu_read(gmu,
1045 					REG_A6XX_GPU_GMU_AO_GPU_CX_BUSY_STATUS2));
1046 	}
1047 
1048 	/* Turn off HFI */
1049 	a6xx_hfi_stop(gmu);
1050 
1051 	/* Stop the interrupts and mask the hardware */
1052 	a6xx_gmu_irq_disable(gmu);
1053 
1054 	/* Tell RPMh to power off the GPU */
1055 	a6xx_rpmh_stop(gmu);
1056 }
1057 
1058 
1059 int a6xx_gmu_stop(struct a6xx_gpu *a6xx_gpu)
1060 {
1061 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1062 	struct msm_gpu *gpu = &a6xx_gpu->base.base;
1063 
1064 	if (!pm_runtime_active(gmu->dev))
1065 		return 0;
1066 
1067 	/*
1068 	 * Force the GMU off if we detected a hang, otherwise try to shut it
1069 	 * down gracefully
1070 	 */
1071 	if (gmu->hung)
1072 		a6xx_gmu_force_off(gmu);
1073 	else
1074 		a6xx_gmu_shutdown(gmu);
1075 
1076 	/* Remove the bus vote */
1077 	dev_pm_opp_set_bw(&gpu->pdev->dev, NULL);
1078 
1079 	/*
1080 	 * Make sure the GX domain is off before turning off the GMU (CX)
1081 	 * domain. Usually the GMU does this but only if the shutdown sequence
1082 	 * was successful
1083 	 */
1084 	if (!IS_ERR_OR_NULL(gmu->gxpd))
1085 		pm_runtime_put_sync(gmu->gxpd);
1086 
1087 	clk_bulk_disable_unprepare(gmu->nr_clocks, gmu->clocks);
1088 
1089 	pm_runtime_put_sync(gmu->dev);
1090 
1091 	return 0;
1092 }
1093 
1094 static void a6xx_gmu_memory_free(struct a6xx_gmu *gmu)
1095 {
1096 	msm_gem_kernel_put(gmu->hfi.obj, gmu->aspace, false);
1097 	msm_gem_kernel_put(gmu->debug.obj, gmu->aspace, false);
1098 	msm_gem_kernel_put(gmu->icache.obj, gmu->aspace, false);
1099 	msm_gem_kernel_put(gmu->dcache.obj, gmu->aspace, false);
1100 	msm_gem_kernel_put(gmu->dummy.obj, gmu->aspace, false);
1101 	msm_gem_kernel_put(gmu->log.obj, gmu->aspace, false);
1102 
1103 	gmu->aspace->mmu->funcs->detach(gmu->aspace->mmu);
1104 	msm_gem_address_space_put(gmu->aspace);
1105 }
1106 
1107 static int a6xx_gmu_memory_alloc(struct a6xx_gmu *gmu, struct a6xx_gmu_bo *bo,
1108 		size_t size, u64 iova)
1109 {
1110 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1111 	struct drm_device *dev = a6xx_gpu->base.base.dev;
1112 	uint32_t flags = MSM_BO_WC;
1113 	u64 range_start, range_end;
1114 	int ret;
1115 
1116 	size = PAGE_ALIGN(size);
1117 	if (!iova) {
1118 		/* no fixed address - use GMU's uncached range */
1119 		range_start = 0x60000000 + PAGE_SIZE; /* skip dummy page */
1120 		range_end = 0x80000000;
1121 	} else {
1122 		/* range for fixed address */
1123 		range_start = iova;
1124 		range_end = iova + size;
1125 		/* use IOMMU_PRIV for icache/dcache */
1126 		flags |= MSM_BO_MAP_PRIV;
1127 	}
1128 
1129 	bo->obj = msm_gem_new(dev, size, flags);
1130 	if (IS_ERR(bo->obj))
1131 		return PTR_ERR(bo->obj);
1132 
1133 	ret = msm_gem_get_and_pin_iova_range(bo->obj, gmu->aspace, &bo->iova,
1134 		range_start >> PAGE_SHIFT, range_end >> PAGE_SHIFT);
1135 	if (ret) {
1136 		drm_gem_object_put(bo->obj);
1137 		return ret;
1138 	}
1139 
1140 	bo->virt = msm_gem_get_vaddr(bo->obj);
1141 	bo->size = size;
1142 
1143 	return 0;
1144 }
1145 
1146 static int a6xx_gmu_memory_probe(struct a6xx_gmu *gmu)
1147 {
1148 	struct iommu_domain *domain;
1149 	struct msm_mmu *mmu;
1150 
1151 	domain = iommu_domain_alloc(&platform_bus_type);
1152 	if (!domain)
1153 		return -ENODEV;
1154 
1155 	mmu = msm_iommu_new(gmu->dev, domain);
1156 	gmu->aspace = msm_gem_address_space_create(mmu, "gmu", 0x0, 0x80000000);
1157 	if (IS_ERR(gmu->aspace)) {
1158 		iommu_domain_free(domain);
1159 		return PTR_ERR(gmu->aspace);
1160 	}
1161 
1162 	return 0;
1163 }
1164 
1165 /* Return the 'arc-level' for the given frequency */
1166 static unsigned int a6xx_gmu_get_arc_level(struct device *dev,
1167 					   unsigned long freq)
1168 {
1169 	struct dev_pm_opp *opp;
1170 	unsigned int val;
1171 
1172 	if (!freq)
1173 		return 0;
1174 
1175 	opp = dev_pm_opp_find_freq_exact(dev, freq, true);
1176 	if (IS_ERR(opp))
1177 		return 0;
1178 
1179 	val = dev_pm_opp_get_level(opp);
1180 
1181 	dev_pm_opp_put(opp);
1182 
1183 	return val;
1184 }
1185 
1186 static int a6xx_gmu_rpmh_arc_votes_init(struct device *dev, u32 *votes,
1187 		unsigned long *freqs, int freqs_count, const char *id)
1188 {
1189 	int i, j;
1190 	const u16 *pri, *sec;
1191 	size_t pri_count, sec_count;
1192 
1193 	pri = cmd_db_read_aux_data(id, &pri_count);
1194 	if (IS_ERR(pri))
1195 		return PTR_ERR(pri);
1196 	/*
1197 	 * The data comes back as an array of unsigned shorts so adjust the
1198 	 * count accordingly
1199 	 */
1200 	pri_count >>= 1;
1201 	if (!pri_count)
1202 		return -EINVAL;
1203 
1204 	sec = cmd_db_read_aux_data("mx.lvl", &sec_count);
1205 	if (IS_ERR(sec))
1206 		return PTR_ERR(sec);
1207 
1208 	sec_count >>= 1;
1209 	if (!sec_count)
1210 		return -EINVAL;
1211 
1212 	/* Construct a vote for each frequency */
1213 	for (i = 0; i < freqs_count; i++) {
1214 		u8 pindex = 0, sindex = 0;
1215 		unsigned int level = a6xx_gmu_get_arc_level(dev, freqs[i]);
1216 
1217 		/* Get the primary index that matches the arc level */
1218 		for (j = 0; j < pri_count; j++) {
1219 			if (pri[j] >= level) {
1220 				pindex = j;
1221 				break;
1222 			}
1223 		}
1224 
1225 		if (j == pri_count) {
1226 			DRM_DEV_ERROR(dev,
1227 				      "Level %u not found in the RPMh list\n",
1228 				      level);
1229 			DRM_DEV_ERROR(dev, "Available levels:\n");
1230 			for (j = 0; j < pri_count; j++)
1231 				DRM_DEV_ERROR(dev, "  %u\n", pri[j]);
1232 
1233 			return -EINVAL;
1234 		}
1235 
1236 		/*
1237 		 * Look for a level in in the secondary list that matches. If
1238 		 * nothing fits, use the maximum non zero vote
1239 		 */
1240 
1241 		for (j = 0; j < sec_count; j++) {
1242 			if (sec[j] >= level) {
1243 				sindex = j;
1244 				break;
1245 			} else if (sec[j]) {
1246 				sindex = j;
1247 			}
1248 		}
1249 
1250 		/* Construct the vote */
1251 		votes[i] = ((pri[pindex] & 0xffff) << 16) |
1252 			(sindex << 8) | pindex;
1253 	}
1254 
1255 	return 0;
1256 }
1257 
1258 /*
1259  * The GMU votes with the RPMh for itself and on behalf of the GPU but we need
1260  * to construct the list of votes on the CPU and send it over. Query the RPMh
1261  * voltage levels and build the votes
1262  */
1263 
1264 static int a6xx_gmu_rpmh_votes_init(struct a6xx_gmu *gmu)
1265 {
1266 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1267 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1268 	struct msm_gpu *gpu = &adreno_gpu->base;
1269 	int ret;
1270 
1271 	/* Build the GX votes */
1272 	ret = a6xx_gmu_rpmh_arc_votes_init(&gpu->pdev->dev, gmu->gx_arc_votes,
1273 		gmu->gpu_freqs, gmu->nr_gpu_freqs, "gfx.lvl");
1274 
1275 	/* Build the CX votes */
1276 	ret |= a6xx_gmu_rpmh_arc_votes_init(gmu->dev, gmu->cx_arc_votes,
1277 		gmu->gmu_freqs, gmu->nr_gmu_freqs, "cx.lvl");
1278 
1279 	return ret;
1280 }
1281 
1282 static int a6xx_gmu_build_freq_table(struct device *dev, unsigned long *freqs,
1283 		u32 size)
1284 {
1285 	int count = dev_pm_opp_get_opp_count(dev);
1286 	struct dev_pm_opp *opp;
1287 	int i, index = 0;
1288 	unsigned long freq = 1;
1289 
1290 	/*
1291 	 * The OPP table doesn't contain the "off" frequency level so we need to
1292 	 * add 1 to the table size to account for it
1293 	 */
1294 
1295 	if (WARN(count + 1 > size,
1296 		"The GMU frequency table is being truncated\n"))
1297 		count = size - 1;
1298 
1299 	/* Set the "off" frequency */
1300 	freqs[index++] = 0;
1301 
1302 	for (i = 0; i < count; i++) {
1303 		opp = dev_pm_opp_find_freq_ceil(dev, &freq);
1304 		if (IS_ERR(opp))
1305 			break;
1306 
1307 		dev_pm_opp_put(opp);
1308 		freqs[index++] = freq++;
1309 	}
1310 
1311 	return index;
1312 }
1313 
1314 static int a6xx_gmu_pwrlevels_probe(struct a6xx_gmu *gmu)
1315 {
1316 	struct a6xx_gpu *a6xx_gpu = container_of(gmu, struct a6xx_gpu, gmu);
1317 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1318 	struct msm_gpu *gpu = &adreno_gpu->base;
1319 
1320 	int ret = 0;
1321 
1322 	/*
1323 	 * The GMU handles its own frequency switching so build a list of
1324 	 * available frequencies to send during initialization
1325 	 */
1326 	ret = dev_pm_opp_of_add_table(gmu->dev);
1327 	if (ret) {
1328 		DRM_DEV_ERROR(gmu->dev, "Unable to set the OPP table for the GMU\n");
1329 		return ret;
1330 	}
1331 
1332 	gmu->nr_gmu_freqs = a6xx_gmu_build_freq_table(gmu->dev,
1333 		gmu->gmu_freqs, ARRAY_SIZE(gmu->gmu_freqs));
1334 
1335 	/*
1336 	 * The GMU also handles GPU frequency switching so build a list
1337 	 * from the GPU OPP table
1338 	 */
1339 	gmu->nr_gpu_freqs = a6xx_gmu_build_freq_table(&gpu->pdev->dev,
1340 		gmu->gpu_freqs, ARRAY_SIZE(gmu->gpu_freqs));
1341 
1342 	gmu->current_perf_index = gmu->nr_gpu_freqs - 1;
1343 
1344 	/* Build the list of RPMh votes that we'll send to the GMU */
1345 	return a6xx_gmu_rpmh_votes_init(gmu);
1346 }
1347 
1348 static int a6xx_gmu_clocks_probe(struct a6xx_gmu *gmu)
1349 {
1350 	int ret = devm_clk_bulk_get_all(gmu->dev, &gmu->clocks);
1351 
1352 	if (ret < 1)
1353 		return ret;
1354 
1355 	gmu->nr_clocks = ret;
1356 
1357 	gmu->core_clk = msm_clk_bulk_get_clock(gmu->clocks,
1358 		gmu->nr_clocks, "gmu");
1359 
1360 	return 0;
1361 }
1362 
1363 static void __iomem *a6xx_gmu_get_mmio(struct platform_device *pdev,
1364 		const char *name)
1365 {
1366 	void __iomem *ret;
1367 	struct resource *res = platform_get_resource_byname(pdev,
1368 			IORESOURCE_MEM, name);
1369 
1370 	if (!res) {
1371 		DRM_DEV_ERROR(&pdev->dev, "Unable to find the %s registers\n", name);
1372 		return ERR_PTR(-EINVAL);
1373 	}
1374 
1375 	ret = ioremap(res->start, resource_size(res));
1376 	if (!ret) {
1377 		DRM_DEV_ERROR(&pdev->dev, "Unable to map the %s registers\n", name);
1378 		return ERR_PTR(-EINVAL);
1379 	}
1380 
1381 	return ret;
1382 }
1383 
1384 static int a6xx_gmu_get_irq(struct a6xx_gmu *gmu, struct platform_device *pdev,
1385 		const char *name, irq_handler_t handler)
1386 {
1387 	int irq, ret;
1388 
1389 	irq = platform_get_irq_byname(pdev, name);
1390 
1391 	ret = request_irq(irq, handler, IRQF_TRIGGER_HIGH, name, gmu);
1392 	if (ret) {
1393 		DRM_DEV_ERROR(&pdev->dev, "Unable to get interrupt %s %d\n",
1394 			      name, ret);
1395 		return ret;
1396 	}
1397 
1398 	disable_irq(irq);
1399 
1400 	return irq;
1401 }
1402 
1403 void a6xx_gmu_remove(struct a6xx_gpu *a6xx_gpu)
1404 {
1405 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1406 	struct platform_device *pdev = to_platform_device(gmu->dev);
1407 
1408 	if (!gmu->initialized)
1409 		return;
1410 
1411 	pm_runtime_force_suspend(gmu->dev);
1412 
1413 	if (!IS_ERR_OR_NULL(gmu->gxpd)) {
1414 		pm_runtime_disable(gmu->gxpd);
1415 		dev_pm_domain_detach(gmu->gxpd, false);
1416 	}
1417 
1418 	iounmap(gmu->mmio);
1419 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1420 		iounmap(gmu->rscc);
1421 	gmu->mmio = NULL;
1422 	gmu->rscc = NULL;
1423 
1424 	a6xx_gmu_memory_free(gmu);
1425 
1426 	free_irq(gmu->gmu_irq, gmu);
1427 	free_irq(gmu->hfi_irq, gmu);
1428 
1429 	/* Drop reference taken in of_find_device_by_node */
1430 	put_device(gmu->dev);
1431 
1432 	gmu->initialized = false;
1433 }
1434 
1435 int a6xx_gmu_init(struct a6xx_gpu *a6xx_gpu, struct device_node *node)
1436 {
1437 	struct adreno_gpu *adreno_gpu = &a6xx_gpu->base;
1438 	struct a6xx_gmu *gmu = &a6xx_gpu->gmu;
1439 	struct platform_device *pdev = of_find_device_by_node(node);
1440 	int ret;
1441 
1442 	if (!pdev)
1443 		return -ENODEV;
1444 
1445 	gmu->dev = &pdev->dev;
1446 
1447 	of_dma_configure(gmu->dev, node, true);
1448 
1449 	/* Fow now, don't do anything fancy until we get our feet under us */
1450 	gmu->idle_level = GMU_IDLE_STATE_ACTIVE;
1451 
1452 	pm_runtime_enable(gmu->dev);
1453 
1454 	/* Get the list of clocks */
1455 	ret = a6xx_gmu_clocks_probe(gmu);
1456 	if (ret)
1457 		goto err_put_device;
1458 
1459 	ret = a6xx_gmu_memory_probe(gmu);
1460 	if (ret)
1461 		goto err_put_device;
1462 
1463 	/* Allocate memory for the GMU dummy page */
1464 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->dummy, SZ_4K, 0x60000000);
1465 	if (ret)
1466 		goto err_memory;
1467 
1468 	if (adreno_is_a650(adreno_gpu)) {
1469 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1470 			SZ_16M - SZ_16K, 0x04000);
1471 		if (ret)
1472 			goto err_memory;
1473 	} else if (adreno_is_a640(adreno_gpu)) {
1474 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->icache,
1475 			SZ_256K - SZ_16K, 0x04000);
1476 		if (ret)
1477 			goto err_memory;
1478 
1479 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->dcache,
1480 			SZ_256K - SZ_16K, 0x44000);
1481 		if (ret)
1482 			goto err_memory;
1483 	} else {
1484 		/* HFI v1, has sptprac */
1485 		gmu->legacy = true;
1486 
1487 		/* Allocate memory for the GMU debug region */
1488 		ret = a6xx_gmu_memory_alloc(gmu, &gmu->debug, SZ_16K, 0);
1489 		if (ret)
1490 			goto err_memory;
1491 	}
1492 
1493 	/* Allocate memory for for the HFI queues */
1494 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->hfi, SZ_16K, 0);
1495 	if (ret)
1496 		goto err_memory;
1497 
1498 	/* Allocate memory for the GMU log region */
1499 	ret = a6xx_gmu_memory_alloc(gmu, &gmu->log, SZ_4K, 0);
1500 	if (ret)
1501 		goto err_memory;
1502 
1503 	/* Map the GMU registers */
1504 	gmu->mmio = a6xx_gmu_get_mmio(pdev, "gmu");
1505 	if (IS_ERR(gmu->mmio)) {
1506 		ret = PTR_ERR(gmu->mmio);
1507 		goto err_memory;
1508 	}
1509 
1510 	if (adreno_is_a650(adreno_gpu)) {
1511 		gmu->rscc = a6xx_gmu_get_mmio(pdev, "rscc");
1512 		if (IS_ERR(gmu->rscc))
1513 			goto err_mmio;
1514 	} else {
1515 		gmu->rscc = gmu->mmio + 0x23000;
1516 	}
1517 
1518 	/* Get the HFI and GMU interrupts */
1519 	gmu->hfi_irq = a6xx_gmu_get_irq(gmu, pdev, "hfi", a6xx_hfi_irq);
1520 	gmu->gmu_irq = a6xx_gmu_get_irq(gmu, pdev, "gmu", a6xx_gmu_irq);
1521 
1522 	if (gmu->hfi_irq < 0 || gmu->gmu_irq < 0)
1523 		goto err_mmio;
1524 
1525 	/*
1526 	 * Get a link to the GX power domain to reset the GPU in case of GMU
1527 	 * crash
1528 	 */
1529 	gmu->gxpd = dev_pm_domain_attach_by_name(gmu->dev, "gx");
1530 
1531 	/* Get the power levels for the GMU and GPU */
1532 	a6xx_gmu_pwrlevels_probe(gmu);
1533 
1534 	/* Set up the HFI queues */
1535 	a6xx_hfi_init(gmu);
1536 
1537 	gmu->initialized = true;
1538 
1539 	return 0;
1540 
1541 err_mmio:
1542 	iounmap(gmu->mmio);
1543 	if (platform_get_resource_byname(pdev, IORESOURCE_MEM, "rscc"))
1544 		iounmap(gmu->rscc);
1545 	free_irq(gmu->gmu_irq, gmu);
1546 	free_irq(gmu->hfi_irq, gmu);
1547 
1548 	ret = -ENODEV;
1549 
1550 err_memory:
1551 	a6xx_gmu_memory_free(gmu);
1552 err_put_device:
1553 	/* Drop reference taken in of_find_device_by_node */
1554 	put_device(gmu->dev);
1555 
1556 	return ret;
1557 }
1558