xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_gpu.c (revision 001821b0e79716c4e17c71d8e053a23599a7a508)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015-2018 Etnaviv Project
4  */
5 
6 #include <linux/clk.h>
7 #include <linux/component.h>
8 #include <linux/delay.h>
9 #include <linux/dma-fence.h>
10 #include <linux/dma-mapping.h>
11 #include <linux/mod_devicetable.h>
12 #include <linux/module.h>
13 #include <linux/platform_device.h>
14 #include <linux/pm_runtime.h>
15 #include <linux/regulator/consumer.h>
16 #include <linux/thermal.h>
17 
18 #include "etnaviv_cmdbuf.h"
19 #include "etnaviv_dump.h"
20 #include "etnaviv_gpu.h"
21 #include "etnaviv_gem.h"
22 #include "etnaviv_mmu.h"
23 #include "etnaviv_perfmon.h"
24 #include "etnaviv_sched.h"
25 #include "common.xml.h"
26 #include "state.xml.h"
27 #include "state_hi.xml.h"
28 #include "cmdstream.xml.h"
29 
30 static const struct platform_device_id gpu_ids[] = {
31 	{ .name = "etnaviv-gpu,2d" },
32 	{ },
33 };
34 
35 /*
36  * Driver functions:
37  */
38 
39 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
40 {
41 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
42 
43 	switch (param) {
44 	case ETNAVIV_PARAM_GPU_MODEL:
45 		*value = gpu->identity.model;
46 		break;
47 
48 	case ETNAVIV_PARAM_GPU_REVISION:
49 		*value = gpu->identity.revision;
50 		break;
51 
52 	case ETNAVIV_PARAM_GPU_FEATURES_0:
53 		*value = gpu->identity.features;
54 		break;
55 
56 	case ETNAVIV_PARAM_GPU_FEATURES_1:
57 		*value = gpu->identity.minor_features0;
58 		break;
59 
60 	case ETNAVIV_PARAM_GPU_FEATURES_2:
61 		*value = gpu->identity.minor_features1;
62 		break;
63 
64 	case ETNAVIV_PARAM_GPU_FEATURES_3:
65 		*value = gpu->identity.minor_features2;
66 		break;
67 
68 	case ETNAVIV_PARAM_GPU_FEATURES_4:
69 		*value = gpu->identity.minor_features3;
70 		break;
71 
72 	case ETNAVIV_PARAM_GPU_FEATURES_5:
73 		*value = gpu->identity.minor_features4;
74 		break;
75 
76 	case ETNAVIV_PARAM_GPU_FEATURES_6:
77 		*value = gpu->identity.minor_features5;
78 		break;
79 
80 	case ETNAVIV_PARAM_GPU_FEATURES_7:
81 		*value = gpu->identity.minor_features6;
82 		break;
83 
84 	case ETNAVIV_PARAM_GPU_FEATURES_8:
85 		*value = gpu->identity.minor_features7;
86 		break;
87 
88 	case ETNAVIV_PARAM_GPU_FEATURES_9:
89 		*value = gpu->identity.minor_features8;
90 		break;
91 
92 	case ETNAVIV_PARAM_GPU_FEATURES_10:
93 		*value = gpu->identity.minor_features9;
94 		break;
95 
96 	case ETNAVIV_PARAM_GPU_FEATURES_11:
97 		*value = gpu->identity.minor_features10;
98 		break;
99 
100 	case ETNAVIV_PARAM_GPU_FEATURES_12:
101 		*value = gpu->identity.minor_features11;
102 		break;
103 
104 	case ETNAVIV_PARAM_GPU_STREAM_COUNT:
105 		*value = gpu->identity.stream_count;
106 		break;
107 
108 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
109 		*value = gpu->identity.register_max;
110 		break;
111 
112 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
113 		*value = gpu->identity.thread_count;
114 		break;
115 
116 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
117 		*value = gpu->identity.vertex_cache_size;
118 		break;
119 
120 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
121 		*value = gpu->identity.shader_core_count;
122 		break;
123 
124 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
125 		*value = gpu->identity.pixel_pipes;
126 		break;
127 
128 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
129 		*value = gpu->identity.vertex_output_buffer_size;
130 		break;
131 
132 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
133 		*value = gpu->identity.buffer_size;
134 		break;
135 
136 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
137 		*value = gpu->identity.instruction_count;
138 		break;
139 
140 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
141 		*value = gpu->identity.num_constants;
142 		break;
143 
144 	case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
145 		*value = gpu->identity.varyings_count;
146 		break;
147 
148 	case ETNAVIV_PARAM_SOFTPIN_START_ADDR:
149 		if (priv->mmu_global->version == ETNAVIV_IOMMU_V2)
150 			*value = ETNAVIV_SOFTPIN_START_ADDRESS;
151 		else
152 			*value = ~0ULL;
153 		break;
154 
155 	case ETNAVIV_PARAM_GPU_PRODUCT_ID:
156 		*value = gpu->identity.product_id;
157 		break;
158 
159 	case ETNAVIV_PARAM_GPU_CUSTOMER_ID:
160 		*value = gpu->identity.customer_id;
161 		break;
162 
163 	case ETNAVIV_PARAM_GPU_ECO_ID:
164 		*value = gpu->identity.eco_id;
165 		break;
166 
167 	default:
168 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
169 		return -EINVAL;
170 	}
171 
172 	return 0;
173 }
174 
175 
176 #define etnaviv_is_model_rev(gpu, mod, rev) \
177 	((gpu)->identity.model == chipModel_##mod && \
178 	 (gpu)->identity.revision == rev)
179 #define etnaviv_field(val, field) \
180 	(((val) & field##__MASK) >> field##__SHIFT)
181 
182 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
183 {
184 	if (gpu->identity.minor_features0 &
185 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
186 		u32 specs[4];
187 		unsigned int streams;
188 
189 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
190 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
191 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
192 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
193 
194 		gpu->identity.stream_count = etnaviv_field(specs[0],
195 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
196 		gpu->identity.register_max = etnaviv_field(specs[0],
197 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
198 		gpu->identity.thread_count = etnaviv_field(specs[0],
199 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
200 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
201 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
202 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
203 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
204 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
205 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
206 		gpu->identity.vertex_output_buffer_size =
207 			etnaviv_field(specs[0],
208 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
209 
210 		gpu->identity.buffer_size = etnaviv_field(specs[1],
211 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
212 		gpu->identity.instruction_count = etnaviv_field(specs[1],
213 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
214 		gpu->identity.num_constants = etnaviv_field(specs[1],
215 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
216 
217 		gpu->identity.varyings_count = etnaviv_field(specs[2],
218 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
219 
220 		/* This overrides the value from older register if non-zero */
221 		streams = etnaviv_field(specs[3],
222 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
223 		if (streams)
224 			gpu->identity.stream_count = streams;
225 	}
226 
227 	/* Fill in the stream count if not specified */
228 	if (gpu->identity.stream_count == 0) {
229 		if (gpu->identity.model >= 0x1000)
230 			gpu->identity.stream_count = 4;
231 		else
232 			gpu->identity.stream_count = 1;
233 	}
234 
235 	/* Convert the register max value */
236 	if (gpu->identity.register_max)
237 		gpu->identity.register_max = 1 << gpu->identity.register_max;
238 	else if (gpu->identity.model == chipModel_GC400)
239 		gpu->identity.register_max = 32;
240 	else
241 		gpu->identity.register_max = 64;
242 
243 	/* Convert thread count */
244 	if (gpu->identity.thread_count)
245 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
246 	else if (gpu->identity.model == chipModel_GC400)
247 		gpu->identity.thread_count = 64;
248 	else if (gpu->identity.model == chipModel_GC500 ||
249 		 gpu->identity.model == chipModel_GC530)
250 		gpu->identity.thread_count = 128;
251 	else
252 		gpu->identity.thread_count = 256;
253 
254 	if (gpu->identity.vertex_cache_size == 0)
255 		gpu->identity.vertex_cache_size = 8;
256 
257 	if (gpu->identity.shader_core_count == 0) {
258 		if (gpu->identity.model >= 0x1000)
259 			gpu->identity.shader_core_count = 2;
260 		else
261 			gpu->identity.shader_core_count = 1;
262 	}
263 
264 	if (gpu->identity.pixel_pipes == 0)
265 		gpu->identity.pixel_pipes = 1;
266 
267 	/* Convert virtex buffer size */
268 	if (gpu->identity.vertex_output_buffer_size) {
269 		gpu->identity.vertex_output_buffer_size =
270 			1 << gpu->identity.vertex_output_buffer_size;
271 	} else if (gpu->identity.model == chipModel_GC400) {
272 		if (gpu->identity.revision < 0x4000)
273 			gpu->identity.vertex_output_buffer_size = 512;
274 		else if (gpu->identity.revision < 0x4200)
275 			gpu->identity.vertex_output_buffer_size = 256;
276 		else
277 			gpu->identity.vertex_output_buffer_size = 128;
278 	} else {
279 		gpu->identity.vertex_output_buffer_size = 512;
280 	}
281 
282 	switch (gpu->identity.instruction_count) {
283 	case 0:
284 		if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
285 		    gpu->identity.model == chipModel_GC880)
286 			gpu->identity.instruction_count = 512;
287 		else
288 			gpu->identity.instruction_count = 256;
289 		break;
290 
291 	case 1:
292 		gpu->identity.instruction_count = 1024;
293 		break;
294 
295 	case 2:
296 		gpu->identity.instruction_count = 2048;
297 		break;
298 
299 	default:
300 		gpu->identity.instruction_count = 256;
301 		break;
302 	}
303 
304 	if (gpu->identity.num_constants == 0)
305 		gpu->identity.num_constants = 168;
306 
307 	if (gpu->identity.varyings_count == 0) {
308 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
309 			gpu->identity.varyings_count = 12;
310 		else
311 			gpu->identity.varyings_count = 8;
312 	}
313 
314 	/*
315 	 * For some cores, two varyings are consumed for position, so the
316 	 * maximum varying count needs to be reduced by one.
317 	 */
318 	if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
319 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
320 	    etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
321 	    etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
322 	    etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
323 	    etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
324 	    etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
325 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
326 	    etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
327 	    etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
328 	    etnaviv_is_model_rev(gpu, GC880, 0x5106))
329 		gpu->identity.varyings_count -= 1;
330 }
331 
332 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
333 {
334 	u32 chipIdentity;
335 
336 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
337 
338 	/* Special case for older graphic cores. */
339 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
340 		gpu->identity.model    = chipModel_GC500;
341 		gpu->identity.revision = etnaviv_field(chipIdentity,
342 					 VIVS_HI_CHIP_IDENTITY_REVISION);
343 	} else {
344 		u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
345 
346 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
347 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
348 		gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
349 
350 		/*
351 		 * Reading these two registers on GC600 rev 0x19 result in a
352 		 * unhandled fault: external abort on non-linefetch
353 		 */
354 		if (!etnaviv_is_model_rev(gpu, GC600, 0x19)) {
355 			gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
356 			gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
357 		}
358 
359 		/*
360 		 * !!!! HACK ALERT !!!!
361 		 * Because people change device IDs without letting software
362 		 * know about it - here is the hack to make it all look the
363 		 * same.  Only for GC400 family.
364 		 */
365 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
366 		    gpu->identity.model != chipModel_GC420) {
367 			gpu->identity.model = gpu->identity.model & 0x0400;
368 		}
369 
370 		/* Another special case */
371 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
372 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
373 
374 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
375 				/*
376 				 * This IP has an ECO; put the correct
377 				 * revision in it.
378 				 */
379 				gpu->identity.revision = 0x1051;
380 			}
381 		}
382 
383 		/*
384 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
385 		 * reality it's just a re-branded GC3000. We can identify this
386 		 * core by the upper half of the revision register being all 1.
387 		 * Fix model/rev here, so all other places can refer to this
388 		 * core by its real identity.
389 		 */
390 		if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
391 			gpu->identity.model = chipModel_GC3000;
392 			gpu->identity.revision &= 0xffff;
393 		}
394 
395 		if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
396 			gpu->identity.eco_id = 1;
397 
398 		if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
399 			gpu->identity.eco_id = 1;
400 	}
401 
402 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
403 		 gpu->identity.model, gpu->identity.revision);
404 
405 	gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
406 	/*
407 	 * If there is a match in the HWDB, we aren't interested in the
408 	 * remaining register values, as they might be wrong.
409 	 */
410 	if (etnaviv_fill_identity_from_hwdb(gpu))
411 		return;
412 
413 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
414 
415 	/* Disable fast clear on GC700. */
416 	if (gpu->identity.model == chipModel_GC700)
417 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
418 
419 	/* These models/revisions don't have the 2D pipe bit */
420 	if ((gpu->identity.model == chipModel_GC500 &&
421 	     gpu->identity.revision <= 2) ||
422 	    gpu->identity.model == chipModel_GC300)
423 		gpu->identity.features |= chipFeatures_PIPE_2D;
424 
425 	if ((gpu->identity.model == chipModel_GC500 &&
426 	     gpu->identity.revision < 2) ||
427 	    (gpu->identity.model == chipModel_GC300 &&
428 	     gpu->identity.revision < 0x2000)) {
429 
430 		/*
431 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
432 		 * registers.
433 		 */
434 		gpu->identity.minor_features0 = 0;
435 		gpu->identity.minor_features1 = 0;
436 		gpu->identity.minor_features2 = 0;
437 		gpu->identity.minor_features3 = 0;
438 		gpu->identity.minor_features4 = 0;
439 		gpu->identity.minor_features5 = 0;
440 	} else
441 		gpu->identity.minor_features0 =
442 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
443 
444 	if (gpu->identity.minor_features0 &
445 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
446 		gpu->identity.minor_features1 =
447 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
448 		gpu->identity.minor_features2 =
449 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
450 		gpu->identity.minor_features3 =
451 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
452 		gpu->identity.minor_features4 =
453 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
454 		gpu->identity.minor_features5 =
455 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
456 	}
457 
458 	/* GC600/300 idle register reports zero bits where modules aren't present */
459 	if (gpu->identity.model == chipModel_GC600 ||
460 	    gpu->identity.model == chipModel_GC300)
461 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
462 				 VIVS_HI_IDLE_STATE_RA |
463 				 VIVS_HI_IDLE_STATE_SE |
464 				 VIVS_HI_IDLE_STATE_PA |
465 				 VIVS_HI_IDLE_STATE_SH |
466 				 VIVS_HI_IDLE_STATE_PE |
467 				 VIVS_HI_IDLE_STATE_DE |
468 				 VIVS_HI_IDLE_STATE_FE;
469 
470 	etnaviv_hw_specs(gpu);
471 }
472 
473 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
474 {
475 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
476 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
477 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
478 }
479 
480 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
481 {
482 	if (gpu->identity.minor_features2 &
483 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
484 		clk_set_rate(gpu->clk_core,
485 			     gpu->base_rate_core >> gpu->freq_scale);
486 		clk_set_rate(gpu->clk_shader,
487 			     gpu->base_rate_shader >> gpu->freq_scale);
488 	} else {
489 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
490 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
491 
492 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
493 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
494 		etnaviv_gpu_load_clock(gpu, clock);
495 	}
496 
497 	/*
498 	 * Choose number of wait cycles to target a ~30us (1/32768) max latency
499 	 * until new work is picked up by the FE when it polls in the idle loop.
500 	 * If the GPU base frequency is unknown use 200 wait cycles.
501 	 */
502 	gpu->fe_waitcycles = clamp(gpu->base_rate_core >> (15 - gpu->freq_scale),
503 				   200UL, 0xffffUL);
504 }
505 
506 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
507 {
508 	u32 control, idle;
509 	unsigned long timeout;
510 	bool failed = true;
511 
512 	/* We hope that the GPU resets in under one second */
513 	timeout = jiffies + msecs_to_jiffies(1000);
514 
515 	while (time_is_after_jiffies(timeout)) {
516 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
517 		u32 pulse_eater = 0x01590880;
518 
519 		/* disable clock gating */
520 		gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, 0x0);
521 
522 		/* disable pulse eater */
523 		pulse_eater |= BIT(17);
524 		gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
525 		pulse_eater |= BIT(0);
526 		gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
527 
528 		/* enable clock */
529 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
530 		etnaviv_gpu_load_clock(gpu, control);
531 
532 		/* isolate the GPU. */
533 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
534 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
535 
536 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
537 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
538 			          VIVS_MMUv2_AHB_CONTROL_RESET);
539 		} else {
540 			/* set soft reset. */
541 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
542 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
543 		}
544 
545 		/* wait for reset. */
546 		usleep_range(10, 20);
547 
548 		/* reset soft reset bit. */
549 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
550 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
551 
552 		/* reset GPU isolation. */
553 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
554 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
555 
556 		/* read idle register. */
557 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
558 
559 		/* try resetting again if FE is not idle */
560 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
561 			dev_dbg(gpu->dev, "FE is not idle\n");
562 			continue;
563 		}
564 
565 		/* read reset register. */
566 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
567 
568 		/* is the GPU idle? */
569 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
570 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
571 			dev_dbg(gpu->dev, "GPU is not idle\n");
572 			continue;
573 		}
574 
575 		/* disable debug registers, as they are not normally needed */
576 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
577 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
578 
579 		failed = false;
580 		break;
581 	}
582 
583 	if (failed) {
584 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
585 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
586 
587 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
588 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
589 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
590 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
591 
592 		return -EBUSY;
593 	}
594 
595 	/* We rely on the GPU running, so program the clock */
596 	etnaviv_gpu_update_clock(gpu);
597 
598 	gpu->state = ETNA_GPU_STATE_RESET;
599 	gpu->exec_state = -1;
600 	if (gpu->mmu_context)
601 		etnaviv_iommu_context_put(gpu->mmu_context);
602 	gpu->mmu_context = NULL;
603 
604 	return 0;
605 }
606 
607 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
608 {
609 	u32 pmc, ppc;
610 
611 	/* enable clock gating */
612 	ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
613 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
614 
615 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
616 	if (gpu->identity.revision == 0x4301 ||
617 	    gpu->identity.revision == 0x4302)
618 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
619 
620 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, ppc);
621 
622 	pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS);
623 
624 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
625 	if (gpu->identity.model >= chipModel_GC400 &&
626 	    gpu->identity.model != chipModel_GC420 &&
627 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
628 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
629 
630 	/*
631 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
632 	 * present without a bug fix.
633 	 */
634 	if (gpu->identity.revision < 0x5000 &&
635 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
636 	    !(gpu->identity.minor_features1 &
637 	      chipMinorFeatures1_DISABLE_PE_GATING))
638 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
639 
640 	if (gpu->identity.revision < 0x5422)
641 		pmc |= BIT(15); /* Unknown bit */
642 
643 	/* Disable TX clock gating on affected core revisions. */
644 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
645 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
646 	    etnaviv_is_model_rev(gpu, GC7000, 0x6202) ||
647 	    etnaviv_is_model_rev(gpu, GC7000, 0x6203))
648 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
649 
650 	/* Disable SE and RA clock gating on affected core revisions. */
651 	if (etnaviv_is_model_rev(gpu, GC7000, 0x6202))
652 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE |
653 		       VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA;
654 
655 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
656 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
657 
658 	gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
659 }
660 
661 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
662 {
663 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
664 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
665 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
666 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
667 
668 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
669 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
670 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
671 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
672 	}
673 }
674 
675 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu,
676 					  struct etnaviv_iommu_context *context)
677 {
678 	u16 prefetch;
679 	u32 address;
680 
681 	WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED);
682 
683 	/* setup the MMU */
684 	etnaviv_iommu_restore(gpu, context);
685 
686 	/* Start command processor */
687 	prefetch = etnaviv_buffer_init(gpu);
688 	address = etnaviv_cmdbuf_get_va(&gpu->buffer,
689 					&gpu->mmu_context->cmdbuf_mapping);
690 
691 	etnaviv_gpu_start_fe(gpu, address, prefetch);
692 
693 	gpu->state = ETNA_GPU_STATE_RUNNING;
694 }
695 
696 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
697 {
698 	/*
699 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
700 	 * cannot be read, extracted from vivante kernel driver.
701 	 */
702 	u32 pulse_eater = 0x01590880;
703 
704 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
705 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
706 		pulse_eater |= BIT(23);
707 
708 	}
709 
710 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
711 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
712 		pulse_eater &= ~BIT(16);
713 		pulse_eater |= BIT(17);
714 	}
715 
716 	if ((gpu->identity.revision > 0x5420) &&
717 	    (gpu->identity.features & chipFeatures_PIPE_3D))
718 	{
719 		/* Performance fix: disable internal DFS */
720 		pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER);
721 		pulse_eater |= BIT(18);
722 	}
723 
724 	gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
725 }
726 
727 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
728 {
729 	WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED ||
730 		  gpu->state == ETNA_GPU_STATE_RESET));
731 
732 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
733 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
734 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
735 		u32 mc_memory_debug;
736 
737 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
738 
739 		if (gpu->identity.revision == 0x5007)
740 			mc_memory_debug |= 0x0c;
741 		else
742 			mc_memory_debug |= 0x08;
743 
744 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
745 	}
746 
747 	/* enable module-level clock gating */
748 	etnaviv_gpu_enable_mlcg(gpu);
749 
750 	/*
751 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
752 	 * This is necessary to prevent the iMX6 SoC locking up.
753 	 */
754 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
755 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
756 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
757 
758 	/* GC2000 rev 5108 needs a special bus config */
759 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
760 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
761 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
762 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
763 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
764 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
765 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
766 	}
767 
768 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
769 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
770 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
771 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
772 	}
773 
774 	/* setup the pulse eater */
775 	etnaviv_gpu_setup_pulse_eater(gpu);
776 
777 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
778 
779 	gpu->state = ETNA_GPU_STATE_INITIALIZED;
780 }
781 
782 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
783 {
784 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
785 	dma_addr_t cmdbuf_paddr;
786 	int ret, i;
787 
788 	ret = pm_runtime_get_sync(gpu->dev);
789 	if (ret < 0) {
790 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
791 		goto pm_put;
792 	}
793 
794 	etnaviv_hw_identify(gpu);
795 
796 	if (gpu->identity.model == 0) {
797 		dev_err(gpu->dev, "Unknown GPU model\n");
798 		ret = -ENXIO;
799 		goto fail;
800 	}
801 
802 	if (gpu->identity.nn_core_count > 0)
803 		dev_warn(gpu->dev, "etnaviv has been instantiated on a NPU, "
804                                    "for which the UAPI is still experimental\n");
805 
806 	/* Exclude VG cores with FE2.0 */
807 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
808 	    gpu->identity.features & chipFeatures_FE20) {
809 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
810 		ret = -ENXIO;
811 		goto fail;
812 	}
813 
814 	/*
815 	 * On cores with security features supported, we claim control over the
816 	 * security states.
817 	 */
818 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
819 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
820 		gpu->sec_mode = ETNA_SEC_KERNEL;
821 
822 	gpu->state = ETNA_GPU_STATE_IDENTIFIED;
823 
824 	ret = etnaviv_hw_reset(gpu);
825 	if (ret) {
826 		dev_err(gpu->dev, "GPU reset failed\n");
827 		goto fail;
828 	}
829 
830 	ret = etnaviv_iommu_global_init(gpu);
831 	if (ret)
832 		goto fail;
833 
834 	/*
835 	 * If the GPU is part of a system with DMA addressing limitations,
836 	 * request pages for our SHM backend buffers from the DMA32 zone to
837 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
838 	 */
839 	if (dma_addressing_limited(gpu->dev))
840 		priv->shm_gfp_mask |= GFP_DMA32;
841 
842 	/* Create buffer: */
843 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
844 				  PAGE_SIZE);
845 	if (ret) {
846 		dev_err(gpu->dev, "could not create command buffer\n");
847 		goto fail;
848 	}
849 
850 	/*
851 	 * Set the GPU linear window to cover the cmdbuf region, as the GPU
852 	 * won't be able to start execution otherwise. The alignment to 128M is
853 	 * chosen arbitrarily but helps in debugging, as the MMU offset
854 	 * calculations are much more straight forward this way.
855 	 *
856 	 * On MC1.0 cores the linear window offset is ignored by the TS engine,
857 	 * leading to inconsistent memory views. Avoid using the offset on those
858 	 * cores if possible, otherwise disable the TS feature.
859 	 */
860 	cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M);
861 
862 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
863 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
864 		if (cmdbuf_paddr >= SZ_2G)
865 			priv->mmu_global->memory_base = SZ_2G;
866 		else
867 			priv->mmu_global->memory_base = cmdbuf_paddr;
868 	} else if (cmdbuf_paddr + SZ_128M >= SZ_2G) {
869 		dev_info(gpu->dev,
870 			 "Need to move linear window on MC1.0, disabling TS\n");
871 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
872 		priv->mmu_global->memory_base = SZ_2G;
873 	}
874 
875 	/* Setup event management */
876 	spin_lock_init(&gpu->event_spinlock);
877 	init_completion(&gpu->event_free);
878 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
879 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
880 		complete(&gpu->event_free);
881 
882 	/* Now program the hardware */
883 	mutex_lock(&gpu->lock);
884 	etnaviv_gpu_hw_init(gpu);
885 	mutex_unlock(&gpu->lock);
886 
887 	pm_runtime_mark_last_busy(gpu->dev);
888 	pm_runtime_put_autosuspend(gpu->dev);
889 
890 	return 0;
891 
892 fail:
893 	pm_runtime_mark_last_busy(gpu->dev);
894 pm_put:
895 	pm_runtime_put_autosuspend(gpu->dev);
896 
897 	return ret;
898 }
899 
900 #ifdef CONFIG_DEBUG_FS
901 struct dma_debug {
902 	u32 address[2];
903 	u32 state[2];
904 };
905 
906 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
907 {
908 	u32 i;
909 
910 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
911 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
912 
913 	for (i = 0; i < 500; i++) {
914 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
915 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
916 
917 		if (debug->address[0] != debug->address[1])
918 			break;
919 
920 		if (debug->state[0] != debug->state[1])
921 			break;
922 	}
923 }
924 
925 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
926 {
927 	struct dma_debug debug;
928 	u32 dma_lo, dma_hi, axi, idle;
929 	int ret;
930 
931 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
932 
933 	ret = pm_runtime_get_sync(gpu->dev);
934 	if (ret < 0)
935 		goto pm_put;
936 
937 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
938 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
939 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
940 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
941 
942 	verify_dma(gpu, &debug);
943 
944 	seq_puts(m, "\tidentity\n");
945 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
946 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
947 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
948 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
949 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
950 
951 	seq_puts(m, "\tfeatures\n");
952 	seq_printf(m, "\t major_features: 0x%08x\n",
953 		   gpu->identity.features);
954 	seq_printf(m, "\t minor_features0: 0x%08x\n",
955 		   gpu->identity.minor_features0);
956 	seq_printf(m, "\t minor_features1: 0x%08x\n",
957 		   gpu->identity.minor_features1);
958 	seq_printf(m, "\t minor_features2: 0x%08x\n",
959 		   gpu->identity.minor_features2);
960 	seq_printf(m, "\t minor_features3: 0x%08x\n",
961 		   gpu->identity.minor_features3);
962 	seq_printf(m, "\t minor_features4: 0x%08x\n",
963 		   gpu->identity.minor_features4);
964 	seq_printf(m, "\t minor_features5: 0x%08x\n",
965 		   gpu->identity.minor_features5);
966 	seq_printf(m, "\t minor_features6: 0x%08x\n",
967 		   gpu->identity.minor_features6);
968 	seq_printf(m, "\t minor_features7: 0x%08x\n",
969 		   gpu->identity.minor_features7);
970 	seq_printf(m, "\t minor_features8: 0x%08x\n",
971 		   gpu->identity.minor_features8);
972 	seq_printf(m, "\t minor_features9: 0x%08x\n",
973 		   gpu->identity.minor_features9);
974 	seq_printf(m, "\t minor_features10: 0x%08x\n",
975 		   gpu->identity.minor_features10);
976 	seq_printf(m, "\t minor_features11: 0x%08x\n",
977 		   gpu->identity.minor_features11);
978 
979 	seq_puts(m, "\tspecs\n");
980 	seq_printf(m, "\t stream_count:  %d\n",
981 			gpu->identity.stream_count);
982 	seq_printf(m, "\t register_max: %d\n",
983 			gpu->identity.register_max);
984 	seq_printf(m, "\t thread_count: %d\n",
985 			gpu->identity.thread_count);
986 	seq_printf(m, "\t vertex_cache_size: %d\n",
987 			gpu->identity.vertex_cache_size);
988 	seq_printf(m, "\t shader_core_count: %d\n",
989 			gpu->identity.shader_core_count);
990 	seq_printf(m, "\t nn_core_count: %d\n",
991 			gpu->identity.nn_core_count);
992 	seq_printf(m, "\t pixel_pipes: %d\n",
993 			gpu->identity.pixel_pipes);
994 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
995 			gpu->identity.vertex_output_buffer_size);
996 	seq_printf(m, "\t buffer_size: %d\n",
997 			gpu->identity.buffer_size);
998 	seq_printf(m, "\t instruction_count: %d\n",
999 			gpu->identity.instruction_count);
1000 	seq_printf(m, "\t num_constants: %d\n",
1001 			gpu->identity.num_constants);
1002 	seq_printf(m, "\t varyings_count: %d\n",
1003 			gpu->identity.varyings_count);
1004 
1005 	seq_printf(m, "\taxi: 0x%08x\n", axi);
1006 	seq_printf(m, "\tidle: 0x%08x\n", idle);
1007 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
1008 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
1009 		seq_puts(m, "\t FE is not idle\n");
1010 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
1011 		seq_puts(m, "\t DE is not idle\n");
1012 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
1013 		seq_puts(m, "\t PE is not idle\n");
1014 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
1015 		seq_puts(m, "\t SH is not idle\n");
1016 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
1017 		seq_puts(m, "\t PA is not idle\n");
1018 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
1019 		seq_puts(m, "\t SE is not idle\n");
1020 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
1021 		seq_puts(m, "\t RA is not idle\n");
1022 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
1023 		seq_puts(m, "\t TX is not idle\n");
1024 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
1025 		seq_puts(m, "\t VG is not idle\n");
1026 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
1027 		seq_puts(m, "\t IM is not idle\n");
1028 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
1029 		seq_puts(m, "\t FP is not idle\n");
1030 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
1031 		seq_puts(m, "\t TS is not idle\n");
1032 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
1033 		seq_puts(m, "\t BL is not idle\n");
1034 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
1035 		seq_puts(m, "\t ASYNCFE is not idle\n");
1036 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
1037 		seq_puts(m, "\t MC is not idle\n");
1038 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
1039 		seq_puts(m, "\t PPA is not idle\n");
1040 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
1041 		seq_puts(m, "\t WD is not idle\n");
1042 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
1043 		seq_puts(m, "\t NN is not idle\n");
1044 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
1045 		seq_puts(m, "\t TP is not idle\n");
1046 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
1047 		seq_puts(m, "\t AXI low power mode\n");
1048 
1049 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
1050 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1051 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1052 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1053 
1054 		seq_puts(m, "\tMC\n");
1055 		seq_printf(m, "\t read0: 0x%08x\n", read0);
1056 		seq_printf(m, "\t read1: 0x%08x\n", read1);
1057 		seq_printf(m, "\t write: 0x%08x\n", write);
1058 	}
1059 
1060 	seq_puts(m, "\tDMA ");
1061 
1062 	if (debug.address[0] == debug.address[1] &&
1063 	    debug.state[0] == debug.state[1]) {
1064 		seq_puts(m, "seems to be stuck\n");
1065 	} else if (debug.address[0] == debug.address[1]) {
1066 		seq_puts(m, "address is constant\n");
1067 	} else {
1068 		seq_puts(m, "is running\n");
1069 	}
1070 
1071 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1072 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1073 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1074 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1075 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1076 		   dma_lo, dma_hi);
1077 
1078 	ret = 0;
1079 
1080 	pm_runtime_mark_last_busy(gpu->dev);
1081 pm_put:
1082 	pm_runtime_put_autosuspend(gpu->dev);
1083 
1084 	return ret;
1085 }
1086 #endif
1087 
1088 /* fence object management */
1089 struct etnaviv_fence {
1090 	struct etnaviv_gpu *gpu;
1091 	struct dma_fence base;
1092 };
1093 
1094 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1095 {
1096 	return container_of(fence, struct etnaviv_fence, base);
1097 }
1098 
1099 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1100 {
1101 	return "etnaviv";
1102 }
1103 
1104 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1105 {
1106 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1107 
1108 	return dev_name(f->gpu->dev);
1109 }
1110 
1111 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1112 {
1113 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1114 
1115 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1116 }
1117 
1118 static void etnaviv_fence_release(struct dma_fence *fence)
1119 {
1120 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1121 
1122 	kfree_rcu(f, base.rcu);
1123 }
1124 
1125 static const struct dma_fence_ops etnaviv_fence_ops = {
1126 	.get_driver_name = etnaviv_fence_get_driver_name,
1127 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1128 	.signaled = etnaviv_fence_signaled,
1129 	.release = etnaviv_fence_release,
1130 };
1131 
1132 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1133 {
1134 	struct etnaviv_fence *f;
1135 
1136 	/*
1137 	 * GPU lock must already be held, otherwise fence completion order might
1138 	 * not match the seqno order assigned here.
1139 	 */
1140 	lockdep_assert_held(&gpu->lock);
1141 
1142 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1143 	if (!f)
1144 		return NULL;
1145 
1146 	f->gpu = gpu;
1147 
1148 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1149 		       gpu->fence_context, ++gpu->next_fence);
1150 
1151 	return &f->base;
1152 }
1153 
1154 /* returns true if fence a comes after fence b */
1155 static inline bool fence_after(u32 a, u32 b)
1156 {
1157 	return (s32)(a - b) > 0;
1158 }
1159 
1160 /*
1161  * event management:
1162  */
1163 
1164 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1165 	unsigned int *events)
1166 {
1167 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1168 	unsigned i, acquired = 0, rpm_count = 0;
1169 	int ret;
1170 
1171 	for (i = 0; i < nr_events; i++) {
1172 		unsigned long remaining;
1173 
1174 		remaining = wait_for_completion_timeout(&gpu->event_free, timeout);
1175 
1176 		if (!remaining) {
1177 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1178 			ret = -EBUSY;
1179 			goto out;
1180 		}
1181 
1182 		acquired++;
1183 		timeout = remaining;
1184 	}
1185 
1186 	spin_lock(&gpu->event_spinlock);
1187 
1188 	for (i = 0; i < nr_events; i++) {
1189 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1190 
1191 		events[i] = event;
1192 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1193 		set_bit(event, gpu->event_bitmap);
1194 	}
1195 
1196 	spin_unlock(&gpu->event_spinlock);
1197 
1198 	for (i = 0; i < nr_events; i++) {
1199 		ret = pm_runtime_resume_and_get(gpu->dev);
1200 		if (ret)
1201 			goto out_rpm;
1202 		rpm_count++;
1203 	}
1204 
1205 	return 0;
1206 
1207 out_rpm:
1208 	for (i = 0; i < rpm_count; i++)
1209 		pm_runtime_put_autosuspend(gpu->dev);
1210 out:
1211 	for (i = 0; i < acquired; i++)
1212 		complete(&gpu->event_free);
1213 
1214 	return ret;
1215 }
1216 
1217 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1218 {
1219 	if (!test_bit(event, gpu->event_bitmap)) {
1220 		dev_warn(gpu->dev, "event %u is already marked as free",
1221 			 event);
1222 	} else {
1223 		clear_bit(event, gpu->event_bitmap);
1224 		complete(&gpu->event_free);
1225 	}
1226 
1227 	pm_runtime_put_autosuspend(gpu->dev);
1228 }
1229 
1230 /*
1231  * Cmdstream submission/retirement:
1232  */
1233 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1234 	u32 id, struct drm_etnaviv_timespec *timeout)
1235 {
1236 	struct dma_fence *fence;
1237 	int ret;
1238 
1239 	/*
1240 	 * Look up the fence and take a reference. We might still find a fence
1241 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1242 	 * pretends we didn't find a fence in that case.
1243 	 */
1244 	rcu_read_lock();
1245 	fence = xa_load(&gpu->user_fences, id);
1246 	if (fence)
1247 		fence = dma_fence_get_rcu(fence);
1248 	rcu_read_unlock();
1249 
1250 	if (!fence)
1251 		return 0;
1252 
1253 	if (!timeout) {
1254 		/* No timeout was requested: just test for completion */
1255 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1256 	} else {
1257 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1258 
1259 		ret = dma_fence_wait_timeout(fence, true, remaining);
1260 		if (ret == 0)
1261 			ret = -ETIMEDOUT;
1262 		else if (ret != -ERESTARTSYS)
1263 			ret = 0;
1264 
1265 	}
1266 
1267 	dma_fence_put(fence);
1268 	return ret;
1269 }
1270 
1271 /*
1272  * Wait for an object to become inactive.  This, on it's own, is not race
1273  * free: the object is moved by the scheduler off the active list, and
1274  * then the iova is put.  Moreover, the object could be re-submitted just
1275  * after we notice that it's become inactive.
1276  *
1277  * Although the retirement happens under the gpu lock, we don't want to hold
1278  * that lock in this function while waiting.
1279  */
1280 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1281 	struct etnaviv_gem_object *etnaviv_obj,
1282 	struct drm_etnaviv_timespec *timeout)
1283 {
1284 	unsigned long remaining;
1285 	long ret;
1286 
1287 	if (!timeout)
1288 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1289 
1290 	remaining = etnaviv_timeout_to_jiffies(timeout);
1291 
1292 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1293 					       !is_active(etnaviv_obj),
1294 					       remaining);
1295 	if (ret > 0)
1296 		return 0;
1297 	else if (ret == -ERESTARTSYS)
1298 		return -ERESTARTSYS;
1299 	else
1300 		return -ETIMEDOUT;
1301 }
1302 
1303 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1304 	struct etnaviv_event *event, unsigned int flags)
1305 {
1306 	const struct etnaviv_gem_submit *submit = event->submit;
1307 	unsigned int i;
1308 
1309 	for (i = 0; i < submit->nr_pmrs; i++) {
1310 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1311 
1312 		if (pmr->flags == flags)
1313 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1314 	}
1315 }
1316 
1317 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1318 	struct etnaviv_event *event)
1319 {
1320 	u32 val;
1321 
1322 	/* disable clock gating */
1323 	val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1324 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1325 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1326 
1327 	/* enable debug register */
1328 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1329 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1330 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1331 
1332 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1333 }
1334 
1335 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1336 	struct etnaviv_event *event)
1337 {
1338 	const struct etnaviv_gem_submit *submit = event->submit;
1339 	unsigned int i;
1340 	u32 val;
1341 
1342 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1343 
1344 	for (i = 0; i < submit->nr_pmrs; i++) {
1345 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1346 
1347 		*pmr->bo_vma = pmr->sequence;
1348 	}
1349 
1350 	/* disable debug register */
1351 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1352 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1353 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1354 
1355 	/* enable clock gating */
1356 	val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1357 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1358 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1359 }
1360 
1361 
1362 /* add bo's to gpu's ring, and kick gpu: */
1363 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1364 {
1365 	struct etnaviv_gpu *gpu = submit->gpu;
1366 	struct dma_fence *gpu_fence;
1367 	unsigned int i, nr_events = 1, event[3];
1368 	int ret;
1369 
1370 	/*
1371 	 * if there are performance monitor requests we need to have
1372 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1373 	 *   requests.
1374 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1375 	 *   and update the sequence number for userspace.
1376 	 */
1377 	if (submit->nr_pmrs)
1378 		nr_events = 3;
1379 
1380 	ret = event_alloc(gpu, nr_events, event);
1381 	if (ret) {
1382 		DRM_ERROR("no free events\n");
1383 		pm_runtime_put_noidle(gpu->dev);
1384 		return NULL;
1385 	}
1386 
1387 	mutex_lock(&gpu->lock);
1388 
1389 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1390 	if (!gpu_fence) {
1391 		for (i = 0; i < nr_events; i++)
1392 			event_free(gpu, event[i]);
1393 
1394 		goto out_unlock;
1395 	}
1396 
1397 	if (gpu->state == ETNA_GPU_STATE_INITIALIZED)
1398 		etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context);
1399 
1400 	if (submit->prev_mmu_context)
1401 		etnaviv_iommu_context_put(submit->prev_mmu_context);
1402 	submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context);
1403 
1404 	if (submit->nr_pmrs) {
1405 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1406 		kref_get(&submit->refcount);
1407 		gpu->event[event[1]].submit = submit;
1408 		etnaviv_sync_point_queue(gpu, event[1]);
1409 	}
1410 
1411 	gpu->event[event[0]].fence = gpu_fence;
1412 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1413 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1414 			     event[0], &submit->cmdbuf);
1415 
1416 	if (submit->nr_pmrs) {
1417 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1418 		kref_get(&submit->refcount);
1419 		gpu->event[event[2]].submit = submit;
1420 		etnaviv_sync_point_queue(gpu, event[2]);
1421 	}
1422 
1423 out_unlock:
1424 	mutex_unlock(&gpu->lock);
1425 
1426 	return gpu_fence;
1427 }
1428 
1429 static void sync_point_worker(struct work_struct *work)
1430 {
1431 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1432 					       sync_point_work);
1433 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1434 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1435 
1436 	event->sync_point(gpu, event);
1437 	etnaviv_submit_put(event->submit);
1438 	event_free(gpu, gpu->sync_point_event);
1439 
1440 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1441 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1442 }
1443 
1444 void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit)
1445 {
1446 	struct etnaviv_gpu *gpu = submit->gpu;
1447 	char *comm = NULL, *cmd = NULL;
1448 	struct task_struct *task;
1449 	unsigned int i;
1450 
1451 	dev_err(gpu->dev, "recover hung GPU!\n");
1452 
1453 	task = get_pid_task(submit->pid, PIDTYPE_PID);
1454 	if (task) {
1455 		comm = kstrdup(task->comm, GFP_KERNEL);
1456 		cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
1457 		put_task_struct(task);
1458 	}
1459 
1460 	if (comm && cmd)
1461 		dev_err(gpu->dev, "offending task: %s (%s)\n", comm, cmd);
1462 
1463 	kfree(cmd);
1464 	kfree(comm);
1465 
1466 	if (pm_runtime_get_sync(gpu->dev) < 0)
1467 		goto pm_put;
1468 
1469 	mutex_lock(&gpu->lock);
1470 
1471 	etnaviv_hw_reset(gpu);
1472 
1473 	/* complete all events, the GPU won't do it after the reset */
1474 	spin_lock(&gpu->event_spinlock);
1475 	for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1476 		event_free(gpu, i);
1477 	spin_unlock(&gpu->event_spinlock);
1478 
1479 	etnaviv_gpu_hw_init(gpu);
1480 
1481 	mutex_unlock(&gpu->lock);
1482 	pm_runtime_mark_last_busy(gpu->dev);
1483 pm_put:
1484 	pm_runtime_put_autosuspend(gpu->dev);
1485 }
1486 
1487 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1488 {
1489 	static const char *fault_reasons[] = {
1490 		"slave not present",
1491 		"page not present",
1492 		"write violation",
1493 		"out of bounds",
1494 		"read security violation",
1495 		"write security violation",
1496 	};
1497 
1498 	u32 status_reg, status;
1499 	int i;
1500 
1501 	if (gpu->sec_mode == ETNA_SEC_NONE)
1502 		status_reg = VIVS_MMUv2_STATUS;
1503 	else
1504 		status_reg = VIVS_MMUv2_SEC_STATUS;
1505 
1506 	status = gpu_read(gpu, status_reg);
1507 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1508 
1509 	for (i = 0; i < 4; i++) {
1510 		const char *reason = "unknown";
1511 		u32 address_reg;
1512 		u32 mmu_status;
1513 
1514 		mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK;
1515 		if (!mmu_status)
1516 			continue;
1517 
1518 		if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons))
1519 			reason = fault_reasons[mmu_status - 1];
1520 
1521 		if (gpu->sec_mode == ETNA_SEC_NONE)
1522 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1523 		else
1524 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1525 
1526 		dev_err_ratelimited(gpu->dev,
1527 				    "MMU %d fault (%s) addr 0x%08x\n",
1528 				    i, reason, gpu_read(gpu, address_reg));
1529 	}
1530 }
1531 
1532 static irqreturn_t irq_handler(int irq, void *data)
1533 {
1534 	struct etnaviv_gpu *gpu = data;
1535 	irqreturn_t ret = IRQ_NONE;
1536 
1537 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1538 
1539 	if (intr != 0) {
1540 		int event;
1541 
1542 		pm_runtime_mark_last_busy(gpu->dev);
1543 
1544 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1545 
1546 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1547 			dev_err(gpu->dev, "AXI bus error\n");
1548 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1549 		}
1550 
1551 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1552 			dump_mmu_fault(gpu);
1553 			gpu->state = ETNA_GPU_STATE_FAULT;
1554 			drm_sched_fault(&gpu->sched);
1555 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1556 		}
1557 
1558 		while ((event = ffs(intr)) != 0) {
1559 			struct dma_fence *fence;
1560 
1561 			event -= 1;
1562 
1563 			intr &= ~(1 << event);
1564 
1565 			dev_dbg(gpu->dev, "event %u\n", event);
1566 
1567 			if (gpu->event[event].sync_point) {
1568 				gpu->sync_point_event = event;
1569 				queue_work(gpu->wq, &gpu->sync_point_work);
1570 			}
1571 
1572 			fence = gpu->event[event].fence;
1573 			if (!fence)
1574 				continue;
1575 
1576 			gpu->event[event].fence = NULL;
1577 
1578 			/*
1579 			 * Events can be processed out of order.  Eg,
1580 			 * - allocate and queue event 0
1581 			 * - allocate event 1
1582 			 * - event 0 completes, we process it
1583 			 * - allocate and queue event 0
1584 			 * - event 1 and event 0 complete
1585 			 * we can end up processing event 0 first, then 1.
1586 			 */
1587 			if (fence_after(fence->seqno, gpu->completed_fence))
1588 				gpu->completed_fence = fence->seqno;
1589 			dma_fence_signal(fence);
1590 
1591 			event_free(gpu, event);
1592 		}
1593 
1594 		ret = IRQ_HANDLED;
1595 	}
1596 
1597 	return ret;
1598 }
1599 
1600 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1601 {
1602 	int ret;
1603 
1604 	ret = clk_prepare_enable(gpu->clk_reg);
1605 	if (ret)
1606 		return ret;
1607 
1608 	ret = clk_prepare_enable(gpu->clk_bus);
1609 	if (ret)
1610 		goto disable_clk_reg;
1611 
1612 	ret = clk_prepare_enable(gpu->clk_core);
1613 	if (ret)
1614 		goto disable_clk_bus;
1615 
1616 	ret = clk_prepare_enable(gpu->clk_shader);
1617 	if (ret)
1618 		goto disable_clk_core;
1619 
1620 	return 0;
1621 
1622 disable_clk_core:
1623 	clk_disable_unprepare(gpu->clk_core);
1624 disable_clk_bus:
1625 	clk_disable_unprepare(gpu->clk_bus);
1626 disable_clk_reg:
1627 	clk_disable_unprepare(gpu->clk_reg);
1628 
1629 	return ret;
1630 }
1631 
1632 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1633 {
1634 	clk_disable_unprepare(gpu->clk_shader);
1635 	clk_disable_unprepare(gpu->clk_core);
1636 	clk_disable_unprepare(gpu->clk_bus);
1637 	clk_disable_unprepare(gpu->clk_reg);
1638 
1639 	return 0;
1640 }
1641 
1642 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1643 {
1644 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1645 
1646 	do {
1647 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1648 
1649 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1650 			return 0;
1651 
1652 		if (time_is_before_jiffies(timeout)) {
1653 			dev_warn(gpu->dev,
1654 				 "timed out waiting for idle: idle=0x%x\n",
1655 				 idle);
1656 			return -ETIMEDOUT;
1657 		}
1658 
1659 		udelay(5);
1660 	} while (1);
1661 }
1662 
1663 static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1664 {
1665 	if (gpu->state == ETNA_GPU_STATE_RUNNING) {
1666 		/* Replace the last WAIT with END */
1667 		mutex_lock(&gpu->lock);
1668 		etnaviv_buffer_end(gpu);
1669 		mutex_unlock(&gpu->lock);
1670 
1671 		/*
1672 		 * We know that only the FE is busy here, this should
1673 		 * happen quickly (as the WAIT is only 200 cycles).  If
1674 		 * we fail, just warn and continue.
1675 		 */
1676 		etnaviv_gpu_wait_idle(gpu, 100);
1677 
1678 		gpu->state = ETNA_GPU_STATE_INITIALIZED;
1679 	}
1680 
1681 	gpu->exec_state = -1;
1682 }
1683 
1684 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1685 {
1686 	int ret;
1687 
1688 	ret = mutex_lock_killable(&gpu->lock);
1689 	if (ret)
1690 		return ret;
1691 
1692 	etnaviv_gpu_update_clock(gpu);
1693 	etnaviv_gpu_hw_init(gpu);
1694 
1695 	mutex_unlock(&gpu->lock);
1696 
1697 	return 0;
1698 }
1699 
1700 static int
1701 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1702 				  unsigned long *state)
1703 {
1704 	*state = 6;
1705 
1706 	return 0;
1707 }
1708 
1709 static int
1710 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1711 				  unsigned long *state)
1712 {
1713 	struct etnaviv_gpu *gpu = cdev->devdata;
1714 
1715 	*state = gpu->freq_scale;
1716 
1717 	return 0;
1718 }
1719 
1720 static int
1721 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1722 				  unsigned long state)
1723 {
1724 	struct etnaviv_gpu *gpu = cdev->devdata;
1725 
1726 	mutex_lock(&gpu->lock);
1727 	gpu->freq_scale = state;
1728 	if (!pm_runtime_suspended(gpu->dev))
1729 		etnaviv_gpu_update_clock(gpu);
1730 	mutex_unlock(&gpu->lock);
1731 
1732 	return 0;
1733 }
1734 
1735 static const struct thermal_cooling_device_ops cooling_ops = {
1736 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1737 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1738 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1739 };
1740 
1741 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1742 	void *data)
1743 {
1744 	struct drm_device *drm = data;
1745 	struct etnaviv_drm_private *priv = drm->dev_private;
1746 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1747 	int ret;
1748 
1749 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1750 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1751 				(char *)dev_name(dev), gpu, &cooling_ops);
1752 		if (IS_ERR(gpu->cooling))
1753 			return PTR_ERR(gpu->cooling);
1754 	}
1755 
1756 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1757 	if (!gpu->wq) {
1758 		ret = -ENOMEM;
1759 		goto out_thermal;
1760 	}
1761 
1762 	ret = etnaviv_sched_init(gpu);
1763 	if (ret)
1764 		goto out_workqueue;
1765 
1766 	if (!IS_ENABLED(CONFIG_PM)) {
1767 		ret = etnaviv_gpu_clk_enable(gpu);
1768 		if (ret < 0)
1769 			goto out_sched;
1770 	}
1771 
1772 	gpu->drm = drm;
1773 	gpu->fence_context = dma_fence_context_alloc(1);
1774 	xa_init_flags(&gpu->user_fences, XA_FLAGS_ALLOC);
1775 	spin_lock_init(&gpu->fence_spinlock);
1776 
1777 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1778 	init_waitqueue_head(&gpu->fence_event);
1779 
1780 	priv->gpu[priv->num_gpus++] = gpu;
1781 
1782 	return 0;
1783 
1784 out_sched:
1785 	etnaviv_sched_fini(gpu);
1786 
1787 out_workqueue:
1788 	destroy_workqueue(gpu->wq);
1789 
1790 out_thermal:
1791 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1792 		thermal_cooling_device_unregister(gpu->cooling);
1793 
1794 	return ret;
1795 }
1796 
1797 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1798 	void *data)
1799 {
1800 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1801 
1802 	DBG("%s", dev_name(gpu->dev));
1803 
1804 	destroy_workqueue(gpu->wq);
1805 
1806 	etnaviv_sched_fini(gpu);
1807 
1808 	if (IS_ENABLED(CONFIG_PM)) {
1809 		pm_runtime_get_sync(gpu->dev);
1810 		pm_runtime_put_sync_suspend(gpu->dev);
1811 	} else {
1812 		etnaviv_gpu_hw_suspend(gpu);
1813 		etnaviv_gpu_clk_disable(gpu);
1814 	}
1815 
1816 	if (gpu->mmu_context)
1817 		etnaviv_iommu_context_put(gpu->mmu_context);
1818 
1819 	etnaviv_cmdbuf_free(&gpu->buffer);
1820 	etnaviv_iommu_global_fini(gpu);
1821 
1822 	gpu->drm = NULL;
1823 	xa_destroy(&gpu->user_fences);
1824 
1825 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1826 		thermal_cooling_device_unregister(gpu->cooling);
1827 	gpu->cooling = NULL;
1828 }
1829 
1830 static const struct component_ops gpu_ops = {
1831 	.bind = etnaviv_gpu_bind,
1832 	.unbind = etnaviv_gpu_unbind,
1833 };
1834 
1835 static const struct of_device_id etnaviv_gpu_match[] = {
1836 	{
1837 		.compatible = "vivante,gc"
1838 	},
1839 	{ /* sentinel */ }
1840 };
1841 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1842 
1843 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1844 {
1845 	struct device *dev = &pdev->dev;
1846 	struct etnaviv_gpu *gpu;
1847 	int err;
1848 
1849 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1850 	if (!gpu)
1851 		return -ENOMEM;
1852 
1853 	gpu->dev = &pdev->dev;
1854 	mutex_init(&gpu->lock);
1855 	mutex_init(&gpu->sched_lock);
1856 
1857 	/* Map registers: */
1858 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1859 	if (IS_ERR(gpu->mmio))
1860 		return PTR_ERR(gpu->mmio);
1861 
1862 	/* Get Interrupt: */
1863 	gpu->irq = platform_get_irq(pdev, 0);
1864 	if (gpu->irq < 0)
1865 		return gpu->irq;
1866 
1867 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1868 			       dev_name(gpu->dev), gpu);
1869 	if (err) {
1870 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1871 		return err;
1872 	}
1873 
1874 	/* Get Clocks: */
1875 	gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1876 	DBG("clk_reg: %p", gpu->clk_reg);
1877 	if (IS_ERR(gpu->clk_reg))
1878 		return PTR_ERR(gpu->clk_reg);
1879 
1880 	gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1881 	DBG("clk_bus: %p", gpu->clk_bus);
1882 	if (IS_ERR(gpu->clk_bus))
1883 		return PTR_ERR(gpu->clk_bus);
1884 
1885 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1886 	DBG("clk_core: %p", gpu->clk_core);
1887 	if (IS_ERR(gpu->clk_core))
1888 		return PTR_ERR(gpu->clk_core);
1889 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1890 
1891 	gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1892 	DBG("clk_shader: %p", gpu->clk_shader);
1893 	if (IS_ERR(gpu->clk_shader))
1894 		return PTR_ERR(gpu->clk_shader);
1895 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1896 
1897 	/* TODO: figure out max mapped size */
1898 	dev_set_drvdata(dev, gpu);
1899 
1900 	/*
1901 	 * We treat the device as initially suspended.  The runtime PM
1902 	 * autosuspend delay is rather arbitary: no measurements have
1903 	 * yet been performed to determine an appropriate value.
1904 	 */
1905 	pm_runtime_use_autosuspend(gpu->dev);
1906 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1907 	pm_runtime_enable(gpu->dev);
1908 
1909 	err = component_add(&pdev->dev, &gpu_ops);
1910 	if (err < 0) {
1911 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1912 		return err;
1913 	}
1914 
1915 	return 0;
1916 }
1917 
1918 static void etnaviv_gpu_platform_remove(struct platform_device *pdev)
1919 {
1920 	component_del(&pdev->dev, &gpu_ops);
1921 	pm_runtime_disable(&pdev->dev);
1922 }
1923 
1924 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1925 {
1926 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1927 	u32 idle, mask;
1928 
1929 	/* If there are any jobs in the HW queue, we're not idle */
1930 	if (atomic_read(&gpu->sched.credit_count))
1931 		return -EBUSY;
1932 
1933 	/* Check whether the hardware (except FE and MC) is idle */
1934 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1935 				  VIVS_HI_IDLE_STATE_MC);
1936 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1937 	if (idle != mask) {
1938 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1939 				     idle);
1940 		return -EBUSY;
1941 	}
1942 
1943 	etnaviv_gpu_hw_suspend(gpu);
1944 
1945 	gpu->state = ETNA_GPU_STATE_IDENTIFIED;
1946 
1947 	return etnaviv_gpu_clk_disable(gpu);
1948 }
1949 
1950 static int etnaviv_gpu_rpm_resume(struct device *dev)
1951 {
1952 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1953 	int ret;
1954 
1955 	ret = etnaviv_gpu_clk_enable(gpu);
1956 	if (ret)
1957 		return ret;
1958 
1959 	/* Re-initialise the basic hardware state */
1960 	if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) {
1961 		ret = etnaviv_gpu_hw_resume(gpu);
1962 		if (ret) {
1963 			etnaviv_gpu_clk_disable(gpu);
1964 			return ret;
1965 		}
1966 	}
1967 
1968 	return 0;
1969 }
1970 
1971 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1972 	RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL)
1973 };
1974 
1975 struct platform_driver etnaviv_gpu_driver = {
1976 	.driver = {
1977 		.name = "etnaviv-gpu",
1978 		.owner = THIS_MODULE,
1979 		.pm = pm_ptr(&etnaviv_gpu_pm_ops),
1980 		.of_match_table = etnaviv_gpu_match,
1981 	},
1982 	.probe = etnaviv_gpu_platform_probe,
1983 	.remove_new = etnaviv_gpu_platform_remove,
1984 	.id_table = gpu_ids,
1985 };
1986