xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_gpu.c (revision 6e7fd890f1d6ac83805409e9c346240de2705584)
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 static inline bool etnaviv_is_model_rev(struct etnaviv_gpu *gpu, u32 model, u32 revision)
176 {
177 	return gpu->identity.model == model &&
178 	       gpu->identity.revision == revision;
179 }
180 
181 #define etnaviv_field(val, field) \
182 	(((val) & field##__MASK) >> field##__SHIFT)
183 
184 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
185 {
186 	if (gpu->identity.minor_features0 &
187 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
188 		u32 specs[4];
189 		unsigned int streams;
190 
191 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
192 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
193 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
194 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
195 
196 		gpu->identity.stream_count = etnaviv_field(specs[0],
197 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
198 		gpu->identity.register_max = etnaviv_field(specs[0],
199 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
200 		gpu->identity.thread_count = etnaviv_field(specs[0],
201 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
202 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
203 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
204 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
205 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
206 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
207 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
208 		gpu->identity.vertex_output_buffer_size =
209 			etnaviv_field(specs[0],
210 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
211 
212 		gpu->identity.buffer_size = etnaviv_field(specs[1],
213 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
214 		gpu->identity.instruction_count = etnaviv_field(specs[1],
215 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
216 		gpu->identity.num_constants = etnaviv_field(specs[1],
217 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
218 
219 		gpu->identity.varyings_count = etnaviv_field(specs[2],
220 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
221 
222 		/* This overrides the value from older register if non-zero */
223 		streams = etnaviv_field(specs[3],
224 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
225 		if (streams)
226 			gpu->identity.stream_count = streams;
227 	}
228 
229 	/* Fill in the stream count if not specified */
230 	if (gpu->identity.stream_count == 0) {
231 		if (gpu->identity.model >= 0x1000)
232 			gpu->identity.stream_count = 4;
233 		else
234 			gpu->identity.stream_count = 1;
235 	}
236 
237 	/* Convert the register max value */
238 	if (gpu->identity.register_max)
239 		gpu->identity.register_max = 1 << gpu->identity.register_max;
240 	else if (gpu->identity.model == chipModel_GC400)
241 		gpu->identity.register_max = 32;
242 	else
243 		gpu->identity.register_max = 64;
244 
245 	/* Convert thread count */
246 	if (gpu->identity.thread_count)
247 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
248 	else if (gpu->identity.model == chipModel_GC400)
249 		gpu->identity.thread_count = 64;
250 	else if (gpu->identity.model == chipModel_GC500 ||
251 		 gpu->identity.model == chipModel_GC530)
252 		gpu->identity.thread_count = 128;
253 	else
254 		gpu->identity.thread_count = 256;
255 
256 	if (gpu->identity.vertex_cache_size == 0)
257 		gpu->identity.vertex_cache_size = 8;
258 
259 	if (gpu->identity.shader_core_count == 0) {
260 		if (gpu->identity.model >= 0x1000)
261 			gpu->identity.shader_core_count = 2;
262 		else
263 			gpu->identity.shader_core_count = 1;
264 	}
265 
266 	if (gpu->identity.pixel_pipes == 0)
267 		gpu->identity.pixel_pipes = 1;
268 
269 	/* Convert virtex buffer size */
270 	if (gpu->identity.vertex_output_buffer_size) {
271 		gpu->identity.vertex_output_buffer_size =
272 			1 << gpu->identity.vertex_output_buffer_size;
273 	} else if (gpu->identity.model == chipModel_GC400) {
274 		if (gpu->identity.revision < 0x4000)
275 			gpu->identity.vertex_output_buffer_size = 512;
276 		else if (gpu->identity.revision < 0x4200)
277 			gpu->identity.vertex_output_buffer_size = 256;
278 		else
279 			gpu->identity.vertex_output_buffer_size = 128;
280 	} else {
281 		gpu->identity.vertex_output_buffer_size = 512;
282 	}
283 
284 	switch (gpu->identity.instruction_count) {
285 	case 0:
286 		if (etnaviv_is_model_rev(gpu, 0x2000, 0x5108) ||
287 		    gpu->identity.model == chipModel_GC880)
288 			gpu->identity.instruction_count = 512;
289 		else
290 			gpu->identity.instruction_count = 256;
291 		break;
292 
293 	case 1:
294 		gpu->identity.instruction_count = 1024;
295 		break;
296 
297 	case 2:
298 		gpu->identity.instruction_count = 2048;
299 		break;
300 
301 	default:
302 		gpu->identity.instruction_count = 256;
303 		break;
304 	}
305 
306 	if (gpu->identity.num_constants == 0)
307 		gpu->identity.num_constants = 168;
308 
309 	if (gpu->identity.varyings_count == 0) {
310 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
311 			gpu->identity.varyings_count = 12;
312 		else
313 			gpu->identity.varyings_count = 8;
314 	}
315 
316 	/*
317 	 * For some cores, two varyings are consumed for position, so the
318 	 * maximum varying count needs to be reduced by one.
319 	 */
320 	if (etnaviv_is_model_rev(gpu, 0x5000, 0x5434) ||
321 	    etnaviv_is_model_rev(gpu, 0x4000, 0x5222) ||
322 	    etnaviv_is_model_rev(gpu, 0x4000, 0x5245) ||
323 	    etnaviv_is_model_rev(gpu, 0x4000, 0x5208) ||
324 	    etnaviv_is_model_rev(gpu, 0x3000, 0x5435) ||
325 	    etnaviv_is_model_rev(gpu, 0x2200, 0x5244) ||
326 	    etnaviv_is_model_rev(gpu, 0x2100, 0x5108) ||
327 	    etnaviv_is_model_rev(gpu, 0x2000, 0x5108) ||
328 	    etnaviv_is_model_rev(gpu, 0x1500, 0x5246) ||
329 	    etnaviv_is_model_rev(gpu, 0x880, 0x5107) ||
330 	    etnaviv_is_model_rev(gpu, 0x880, 0x5106))
331 		gpu->identity.varyings_count -= 1;
332 }
333 
334 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
335 {
336 	u32 chipIdentity;
337 
338 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
339 
340 	/* Special case for older graphic cores. */
341 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
342 		gpu->identity.model    = chipModel_GC500;
343 		gpu->identity.revision = etnaviv_field(chipIdentity,
344 					 VIVS_HI_CHIP_IDENTITY_REVISION);
345 	} else {
346 		u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
347 
348 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
349 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
350 		gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
351 
352 		/*
353 		 * Reading these two registers on GC600 rev 0x19 result in a
354 		 * unhandled fault: external abort on non-linefetch
355 		 */
356 		if (!etnaviv_is_model_rev(gpu, 0x600, 0x19)) {
357 			gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
358 			gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
359 		}
360 
361 		/*
362 		 * !!!! HACK ALERT !!!!
363 		 * Because people change device IDs without letting software
364 		 * know about it - here is the hack to make it all look the
365 		 * same.  Only for GC400 family.
366 		 */
367 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
368 		    gpu->identity.model != chipModel_GC420) {
369 			gpu->identity.model = gpu->identity.model & 0x0400;
370 		}
371 
372 		/* Another special case */
373 		if (etnaviv_is_model_rev(gpu, 0x300, 0x2201)) {
374 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
375 
376 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
377 				/*
378 				 * This IP has an ECO; put the correct
379 				 * revision in it.
380 				 */
381 				gpu->identity.revision = 0x1051;
382 			}
383 		}
384 
385 		/*
386 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
387 		 * reality it's just a re-branded GC3000. We can identify this
388 		 * core by the upper half of the revision register being all 1.
389 		 * Fix model/rev here, so all other places can refer to this
390 		 * core by its real identity.
391 		 */
392 		if (etnaviv_is_model_rev(gpu, 0x2000, 0xffff5450)) {
393 			gpu->identity.model = chipModel_GC3000;
394 			gpu->identity.revision &= 0xffff;
395 		}
396 
397 		if (etnaviv_is_model_rev(gpu, 0x1000, 0x5037) && (chipDate == 0x20120617))
398 			gpu->identity.eco_id = 1;
399 
400 		if (etnaviv_is_model_rev(gpu, 0x320, 0x5303) && (chipDate == 0x20140511))
401 			gpu->identity.eco_id = 1;
402 	}
403 
404 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
405 		 gpu->identity.model, gpu->identity.revision);
406 
407 	gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
408 	/*
409 	 * If there is a match in the HWDB, we aren't interested in the
410 	 * remaining register values, as they might be wrong.
411 	 */
412 	if (etnaviv_fill_identity_from_hwdb(gpu))
413 		return;
414 
415 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
416 
417 	/* Disable fast clear on GC700. */
418 	if (gpu->identity.model == chipModel_GC700)
419 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
420 
421 	/* These models/revisions don't have the 2D pipe bit */
422 	if ((gpu->identity.model == chipModel_GC500 &&
423 	     gpu->identity.revision <= 2) ||
424 	    gpu->identity.model == chipModel_GC300)
425 		gpu->identity.features |= chipFeatures_PIPE_2D;
426 
427 	if ((gpu->identity.model == chipModel_GC500 &&
428 	     gpu->identity.revision < 2) ||
429 	    (gpu->identity.model == chipModel_GC300 &&
430 	     gpu->identity.revision < 0x2000)) {
431 
432 		/*
433 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
434 		 * registers.
435 		 */
436 		gpu->identity.minor_features0 = 0;
437 		gpu->identity.minor_features1 = 0;
438 		gpu->identity.minor_features2 = 0;
439 		gpu->identity.minor_features3 = 0;
440 		gpu->identity.minor_features4 = 0;
441 		gpu->identity.minor_features5 = 0;
442 	} else
443 		gpu->identity.minor_features0 =
444 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
445 
446 	if (gpu->identity.minor_features0 &
447 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
448 		gpu->identity.minor_features1 =
449 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
450 		gpu->identity.minor_features2 =
451 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
452 		gpu->identity.minor_features3 =
453 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
454 		gpu->identity.minor_features4 =
455 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
456 		gpu->identity.minor_features5 =
457 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
458 	}
459 
460 	/* GC600/300 idle register reports zero bits where modules aren't present */
461 	if (gpu->identity.model == chipModel_GC600 ||
462 	    gpu->identity.model == chipModel_GC300)
463 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
464 				 VIVS_HI_IDLE_STATE_RA |
465 				 VIVS_HI_IDLE_STATE_SE |
466 				 VIVS_HI_IDLE_STATE_PA |
467 				 VIVS_HI_IDLE_STATE_SH |
468 				 VIVS_HI_IDLE_STATE_PE |
469 				 VIVS_HI_IDLE_STATE_DE |
470 				 VIVS_HI_IDLE_STATE_FE;
471 
472 	etnaviv_hw_specs(gpu);
473 }
474 
475 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
476 {
477 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
478 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
479 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
480 }
481 
482 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
483 {
484 	if (gpu->identity.minor_features2 &
485 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
486 		clk_set_rate(gpu->clk_core,
487 			     gpu->base_rate_core >> gpu->freq_scale);
488 		clk_set_rate(gpu->clk_shader,
489 			     gpu->base_rate_shader >> gpu->freq_scale);
490 	} else {
491 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
492 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
493 
494 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
495 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
496 		etnaviv_gpu_load_clock(gpu, clock);
497 	}
498 
499 	/*
500 	 * Choose number of wait cycles to target a ~30us (1/32768) max latency
501 	 * until new work is picked up by the FE when it polls in the idle loop.
502 	 * If the GPU base frequency is unknown use 200 wait cycles.
503 	 */
504 	gpu->fe_waitcycles = clamp(gpu->base_rate_core >> (15 - gpu->freq_scale),
505 				   200UL, 0xffffUL);
506 }
507 
508 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
509 {
510 	u32 control, idle;
511 	unsigned long timeout;
512 	bool failed = true;
513 
514 	/* We hope that the GPU resets in under one second */
515 	timeout = jiffies + msecs_to_jiffies(1000);
516 
517 	while (time_is_after_jiffies(timeout)) {
518 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
519 		u32 pulse_eater = 0x01590880;
520 
521 		/* disable clock gating */
522 		gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, 0x0);
523 
524 		/* disable pulse eater */
525 		pulse_eater |= BIT(17);
526 		gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
527 		pulse_eater |= BIT(0);
528 		gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
529 
530 		/* enable clock */
531 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
532 		etnaviv_gpu_load_clock(gpu, control);
533 
534 		/* isolate the GPU. */
535 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
536 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
537 
538 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
539 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
540 			          VIVS_MMUv2_AHB_CONTROL_RESET);
541 		} else {
542 			/* set soft reset. */
543 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
544 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
545 		}
546 
547 		/* wait for reset. */
548 		usleep_range(10, 20);
549 
550 		/* reset soft reset bit. */
551 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
552 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
553 
554 		/* reset GPU isolation. */
555 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
556 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
557 
558 		/* read idle register. */
559 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
560 
561 		/* try resetting again if FE is not idle */
562 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
563 			dev_dbg(gpu->dev, "FE is not idle\n");
564 			continue;
565 		}
566 
567 		/* read reset register. */
568 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
569 
570 		/* is the GPU idle? */
571 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
572 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
573 			dev_dbg(gpu->dev, "GPU is not idle\n");
574 			continue;
575 		}
576 
577 		/* disable debug registers, as they are not normally needed */
578 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
579 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
580 
581 		failed = false;
582 		break;
583 	}
584 
585 	if (failed) {
586 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
587 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
588 
589 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
590 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
591 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
592 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
593 
594 		return -EBUSY;
595 	}
596 
597 	/* We rely on the GPU running, so program the clock */
598 	etnaviv_gpu_update_clock(gpu);
599 
600 	gpu->state = ETNA_GPU_STATE_RESET;
601 	gpu->exec_state = -1;
602 	if (gpu->mmu_context)
603 		etnaviv_iommu_context_put(gpu->mmu_context);
604 	gpu->mmu_context = NULL;
605 
606 	return 0;
607 }
608 
609 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
610 {
611 	u32 pmc, ppc;
612 
613 	/* enable clock gating */
614 	ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
615 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
616 
617 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
618 	if (gpu->identity.revision == 0x4301 ||
619 	    gpu->identity.revision == 0x4302)
620 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
621 
622 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, ppc);
623 
624 	pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS);
625 
626 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
627 	if (gpu->identity.model >= chipModel_GC400 &&
628 	    gpu->identity.model != chipModel_GC420 &&
629 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
630 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
631 
632 	/*
633 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
634 	 * present without a bug fix.
635 	 */
636 	if (gpu->identity.revision < 0x5000 &&
637 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
638 	    !(gpu->identity.minor_features1 &
639 	      chipMinorFeatures1_DISABLE_PE_GATING))
640 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
641 
642 	if (gpu->identity.revision < 0x5422)
643 		pmc |= BIT(15); /* Unknown bit */
644 
645 	/* Disable TX clock gating on affected core revisions. */
646 	if (etnaviv_is_model_rev(gpu, 0x4000, 0x5222) ||
647 	    etnaviv_is_model_rev(gpu, 0x2000, 0x5108) ||
648 	    etnaviv_is_model_rev(gpu, 0x7000, 0x6202) ||
649 	    etnaviv_is_model_rev(gpu, 0x7000, 0x6203))
650 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
651 
652 	/* Disable SE and RA clock gating on affected core revisions. */
653 	if (etnaviv_is_model_rev(gpu, 0x7000, 0x6202))
654 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE |
655 		       VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA;
656 
657 	/* Disable SH_EU clock gating on affected core revisions. */
658 	if (etnaviv_is_model_rev(gpu, 0x8000, 0x7200) ||
659 	    etnaviv_is_model_rev(gpu, 0x8000, 0x8002) ||
660 	    etnaviv_is_model_rev(gpu, 0x9200, 0x6304))
661 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SH_EU;
662 
663 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
664 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
665 
666 	gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
667 }
668 
669 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
670 {
671 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
672 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
673 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
674 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
675 
676 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
677 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
678 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
679 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
680 	}
681 }
682 
683 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu,
684 					  struct etnaviv_iommu_context *context)
685 {
686 	u16 prefetch;
687 	u32 address;
688 
689 	WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED);
690 
691 	/* setup the MMU */
692 	etnaviv_iommu_restore(gpu, context);
693 
694 	/* Start command processor */
695 	prefetch = etnaviv_buffer_init(gpu);
696 	address = etnaviv_cmdbuf_get_va(&gpu->buffer,
697 					&gpu->mmu_context->cmdbuf_mapping);
698 
699 	etnaviv_gpu_start_fe(gpu, address, prefetch);
700 
701 	gpu->state = ETNA_GPU_STATE_RUNNING;
702 }
703 
704 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
705 {
706 	/*
707 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
708 	 * cannot be read, extracted from vivante kernel driver.
709 	 */
710 	u32 pulse_eater = 0x01590880;
711 
712 	if (etnaviv_is_model_rev(gpu, 0x4000, 0x5208) ||
713 	    etnaviv_is_model_rev(gpu, 0x4000, 0x5222)) {
714 		pulse_eater |= BIT(23);
715 
716 	}
717 
718 	if (etnaviv_is_model_rev(gpu, 0x1000, 0x5039) ||
719 	    etnaviv_is_model_rev(gpu, 0x1000, 0x5040)) {
720 		pulse_eater &= ~BIT(16);
721 		pulse_eater |= BIT(17);
722 	}
723 
724 	if ((gpu->identity.revision > 0x5420) &&
725 	    (gpu->identity.features & chipFeatures_PIPE_3D))
726 	{
727 		/* Performance fix: disable internal DFS */
728 		pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER);
729 		pulse_eater |= BIT(18);
730 	}
731 
732 	gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
733 }
734 
735 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
736 {
737 	WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED ||
738 		  gpu->state == ETNA_GPU_STATE_RESET));
739 
740 	if ((etnaviv_is_model_rev(gpu, 0x320, 0x5007) ||
741 	     etnaviv_is_model_rev(gpu, 0x320, 0x5220)) &&
742 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
743 		u32 mc_memory_debug;
744 
745 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
746 
747 		if (gpu->identity.revision == 0x5007)
748 			mc_memory_debug |= 0x0c;
749 		else
750 			mc_memory_debug |= 0x08;
751 
752 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
753 	}
754 
755 	/* enable module-level clock gating */
756 	etnaviv_gpu_enable_mlcg(gpu);
757 
758 	/*
759 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
760 	 * This is necessary to prevent the iMX6 SoC locking up.
761 	 */
762 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
763 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
764 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
765 
766 	/* GC2000 rev 5108 needs a special bus config */
767 	if (etnaviv_is_model_rev(gpu, 0x2000, 0x5108)) {
768 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
769 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
770 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
771 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
772 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
773 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
774 	}
775 
776 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
777 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
778 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
779 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
780 	}
781 
782 	/* setup the pulse eater */
783 	etnaviv_gpu_setup_pulse_eater(gpu);
784 
785 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
786 
787 	gpu->state = ETNA_GPU_STATE_INITIALIZED;
788 }
789 
790 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
791 {
792 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
793 	dma_addr_t cmdbuf_paddr;
794 	int ret, i;
795 
796 	ret = pm_runtime_get_sync(gpu->dev);
797 	if (ret < 0) {
798 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
799 		goto pm_put;
800 	}
801 
802 	etnaviv_hw_identify(gpu);
803 
804 	if (gpu->identity.model == 0) {
805 		dev_err(gpu->dev, "Unknown GPU model\n");
806 		ret = -ENXIO;
807 		goto fail;
808 	}
809 
810 	if (gpu->identity.nn_core_count > 0)
811 		dev_warn(gpu->dev, "etnaviv has been instantiated on a NPU, "
812                                    "for which the UAPI is still experimental\n");
813 
814 	/* Exclude VG cores with FE2.0 */
815 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
816 	    gpu->identity.features & chipFeatures_FE20) {
817 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
818 		ret = -ENXIO;
819 		goto fail;
820 	}
821 
822 	/*
823 	 * On cores with security features supported, we claim control over the
824 	 * security states.
825 	 */
826 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
827 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
828 		gpu->sec_mode = ETNA_SEC_KERNEL;
829 
830 	gpu->state = ETNA_GPU_STATE_IDENTIFIED;
831 
832 	ret = etnaviv_hw_reset(gpu);
833 	if (ret) {
834 		dev_err(gpu->dev, "GPU reset failed\n");
835 		goto fail;
836 	}
837 
838 	ret = etnaviv_iommu_global_init(gpu);
839 	if (ret)
840 		goto fail;
841 
842 	/*
843 	 * If the GPU is part of a system with DMA addressing limitations,
844 	 * request pages for our SHM backend buffers from the DMA32 zone to
845 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
846 	 */
847 	if (dma_addressing_limited(gpu->dev))
848 		priv->shm_gfp_mask |= GFP_DMA32;
849 
850 	/* Create buffer: */
851 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
852 				  PAGE_SIZE);
853 	if (ret) {
854 		dev_err(gpu->dev, "could not create command buffer\n");
855 		goto fail;
856 	}
857 
858 	/*
859 	 * Set the GPU linear window to cover the cmdbuf region, as the GPU
860 	 * won't be able to start execution otherwise. The alignment to 128M is
861 	 * chosen arbitrarily but helps in debugging, as the MMU offset
862 	 * calculations are much more straight forward this way.
863 	 *
864 	 * On MC1.0 cores the linear window offset is ignored by the TS engine,
865 	 * leading to inconsistent memory views. Avoid using the offset on those
866 	 * cores if possible, otherwise disable the TS feature. MMUv2 doesn't
867 	 * expose this issue, as all TS accesses are MMU translated, so the
868 	 * linear window offset won't be used.
869 	 */
870 	cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M);
871 
872 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
873 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20) ||
874 	    (gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION)) {
875 		if (cmdbuf_paddr >= SZ_2G)
876 			priv->mmu_global->memory_base = SZ_2G;
877 		else
878 			priv->mmu_global->memory_base = cmdbuf_paddr;
879 	} else if (cmdbuf_paddr + SZ_128M >= SZ_2G) {
880 		dev_info(gpu->dev,
881 			 "Need to move linear window on MC1.0, disabling TS\n");
882 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
883 		priv->mmu_global->memory_base = SZ_2G;
884 	}
885 
886 	/* Setup event management */
887 	spin_lock_init(&gpu->event_spinlock);
888 	init_completion(&gpu->event_free);
889 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
890 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
891 		complete(&gpu->event_free);
892 
893 	/* Now program the hardware */
894 	mutex_lock(&gpu->lock);
895 	etnaviv_gpu_hw_init(gpu);
896 	mutex_unlock(&gpu->lock);
897 
898 	pm_runtime_mark_last_busy(gpu->dev);
899 	pm_runtime_put_autosuspend(gpu->dev);
900 
901 	return 0;
902 
903 fail:
904 	pm_runtime_mark_last_busy(gpu->dev);
905 pm_put:
906 	pm_runtime_put_autosuspend(gpu->dev);
907 
908 	return ret;
909 }
910 
911 #ifdef CONFIG_DEBUG_FS
912 struct dma_debug {
913 	u32 address[2];
914 	u32 state[2];
915 };
916 
917 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
918 {
919 	u32 i;
920 
921 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
922 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
923 
924 	for (i = 0; i < 500; i++) {
925 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
926 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
927 
928 		if (debug->address[0] != debug->address[1])
929 			break;
930 
931 		if (debug->state[0] != debug->state[1])
932 			break;
933 	}
934 }
935 
936 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
937 {
938 	struct dma_debug debug;
939 	u32 dma_lo, dma_hi, axi, idle;
940 	int ret;
941 
942 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
943 
944 	ret = pm_runtime_get_sync(gpu->dev);
945 	if (ret < 0)
946 		goto pm_put;
947 
948 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
949 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
950 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
951 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
952 
953 	verify_dma(gpu, &debug);
954 
955 	seq_puts(m, "\tidentity\n");
956 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
957 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
958 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
959 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
960 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
961 
962 	seq_puts(m, "\tfeatures\n");
963 	seq_printf(m, "\t major_features: 0x%08x\n",
964 		   gpu->identity.features);
965 	seq_printf(m, "\t minor_features0: 0x%08x\n",
966 		   gpu->identity.minor_features0);
967 	seq_printf(m, "\t minor_features1: 0x%08x\n",
968 		   gpu->identity.minor_features1);
969 	seq_printf(m, "\t minor_features2: 0x%08x\n",
970 		   gpu->identity.minor_features2);
971 	seq_printf(m, "\t minor_features3: 0x%08x\n",
972 		   gpu->identity.minor_features3);
973 	seq_printf(m, "\t minor_features4: 0x%08x\n",
974 		   gpu->identity.minor_features4);
975 	seq_printf(m, "\t minor_features5: 0x%08x\n",
976 		   gpu->identity.minor_features5);
977 	seq_printf(m, "\t minor_features6: 0x%08x\n",
978 		   gpu->identity.minor_features6);
979 	seq_printf(m, "\t minor_features7: 0x%08x\n",
980 		   gpu->identity.minor_features7);
981 	seq_printf(m, "\t minor_features8: 0x%08x\n",
982 		   gpu->identity.minor_features8);
983 	seq_printf(m, "\t minor_features9: 0x%08x\n",
984 		   gpu->identity.minor_features9);
985 	seq_printf(m, "\t minor_features10: 0x%08x\n",
986 		   gpu->identity.minor_features10);
987 	seq_printf(m, "\t minor_features11: 0x%08x\n",
988 		   gpu->identity.minor_features11);
989 
990 	seq_puts(m, "\tspecs\n");
991 	seq_printf(m, "\t stream_count:  %d\n",
992 			gpu->identity.stream_count);
993 	seq_printf(m, "\t register_max: %d\n",
994 			gpu->identity.register_max);
995 	seq_printf(m, "\t thread_count: %d\n",
996 			gpu->identity.thread_count);
997 	seq_printf(m, "\t vertex_cache_size: %d\n",
998 			gpu->identity.vertex_cache_size);
999 	seq_printf(m, "\t shader_core_count: %d\n",
1000 			gpu->identity.shader_core_count);
1001 	seq_printf(m, "\t nn_core_count: %d\n",
1002 			gpu->identity.nn_core_count);
1003 	seq_printf(m, "\t pixel_pipes: %d\n",
1004 			gpu->identity.pixel_pipes);
1005 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
1006 			gpu->identity.vertex_output_buffer_size);
1007 	seq_printf(m, "\t buffer_size: %d\n",
1008 			gpu->identity.buffer_size);
1009 	seq_printf(m, "\t instruction_count: %d\n",
1010 			gpu->identity.instruction_count);
1011 	seq_printf(m, "\t num_constants: %d\n",
1012 			gpu->identity.num_constants);
1013 	seq_printf(m, "\t varyings_count: %d\n",
1014 			gpu->identity.varyings_count);
1015 
1016 	seq_printf(m, "\taxi: 0x%08x\n", axi);
1017 	seq_printf(m, "\tidle: 0x%08x\n", idle);
1018 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
1019 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
1020 		seq_puts(m, "\t FE is not idle\n");
1021 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
1022 		seq_puts(m, "\t DE is not idle\n");
1023 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
1024 		seq_puts(m, "\t PE is not idle\n");
1025 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
1026 		seq_puts(m, "\t SH is not idle\n");
1027 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
1028 		seq_puts(m, "\t PA is not idle\n");
1029 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
1030 		seq_puts(m, "\t SE is not idle\n");
1031 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
1032 		seq_puts(m, "\t RA is not idle\n");
1033 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
1034 		seq_puts(m, "\t TX is not idle\n");
1035 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
1036 		seq_puts(m, "\t VG is not idle\n");
1037 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
1038 		seq_puts(m, "\t IM is not idle\n");
1039 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
1040 		seq_puts(m, "\t FP is not idle\n");
1041 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
1042 		seq_puts(m, "\t TS is not idle\n");
1043 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
1044 		seq_puts(m, "\t BL is not idle\n");
1045 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
1046 		seq_puts(m, "\t ASYNCFE is not idle\n");
1047 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
1048 		seq_puts(m, "\t MC is not idle\n");
1049 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
1050 		seq_puts(m, "\t PPA is not idle\n");
1051 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
1052 		seq_puts(m, "\t WD is not idle\n");
1053 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
1054 		seq_puts(m, "\t NN is not idle\n");
1055 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
1056 		seq_puts(m, "\t TP is not idle\n");
1057 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
1058 		seq_puts(m, "\t AXI low power mode\n");
1059 
1060 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
1061 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1062 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1063 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1064 
1065 		seq_puts(m, "\tMC\n");
1066 		seq_printf(m, "\t read0: 0x%08x\n", read0);
1067 		seq_printf(m, "\t read1: 0x%08x\n", read1);
1068 		seq_printf(m, "\t write: 0x%08x\n", write);
1069 	}
1070 
1071 	seq_puts(m, "\tDMA ");
1072 
1073 	if (debug.address[0] == debug.address[1] &&
1074 	    debug.state[0] == debug.state[1]) {
1075 		seq_puts(m, "seems to be stuck\n");
1076 	} else if (debug.address[0] == debug.address[1]) {
1077 		seq_puts(m, "address is constant\n");
1078 	} else {
1079 		seq_puts(m, "is running\n");
1080 	}
1081 
1082 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1083 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1084 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1085 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1086 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1087 		   dma_lo, dma_hi);
1088 
1089 	ret = 0;
1090 
1091 	pm_runtime_mark_last_busy(gpu->dev);
1092 pm_put:
1093 	pm_runtime_put_autosuspend(gpu->dev);
1094 
1095 	return ret;
1096 }
1097 #endif
1098 
1099 /* fence object management */
1100 struct etnaviv_fence {
1101 	struct etnaviv_gpu *gpu;
1102 	struct dma_fence base;
1103 };
1104 
1105 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1106 {
1107 	return container_of(fence, struct etnaviv_fence, base);
1108 }
1109 
1110 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1111 {
1112 	return "etnaviv";
1113 }
1114 
1115 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1116 {
1117 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1118 
1119 	return dev_name(f->gpu->dev);
1120 }
1121 
1122 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1123 {
1124 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1125 
1126 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1127 }
1128 
1129 static void etnaviv_fence_release(struct dma_fence *fence)
1130 {
1131 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1132 
1133 	kfree_rcu(f, base.rcu);
1134 }
1135 
1136 static const struct dma_fence_ops etnaviv_fence_ops = {
1137 	.get_driver_name = etnaviv_fence_get_driver_name,
1138 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1139 	.signaled = etnaviv_fence_signaled,
1140 	.release = etnaviv_fence_release,
1141 };
1142 
1143 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1144 {
1145 	struct etnaviv_fence *f;
1146 
1147 	/*
1148 	 * GPU lock must already be held, otherwise fence completion order might
1149 	 * not match the seqno order assigned here.
1150 	 */
1151 	lockdep_assert_held(&gpu->lock);
1152 
1153 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1154 	if (!f)
1155 		return NULL;
1156 
1157 	f->gpu = gpu;
1158 
1159 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1160 		       gpu->fence_context, ++gpu->next_fence);
1161 
1162 	return &f->base;
1163 }
1164 
1165 /* returns true if fence a comes after fence b */
1166 static inline bool fence_after(u32 a, u32 b)
1167 {
1168 	return (s32)(a - b) > 0;
1169 }
1170 
1171 /*
1172  * event management:
1173  */
1174 
1175 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1176 	unsigned int *events)
1177 {
1178 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1179 	unsigned i, acquired = 0, rpm_count = 0;
1180 	int ret;
1181 
1182 	for (i = 0; i < nr_events; i++) {
1183 		unsigned long remaining;
1184 
1185 		remaining = wait_for_completion_timeout(&gpu->event_free, timeout);
1186 
1187 		if (!remaining) {
1188 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1189 			ret = -EBUSY;
1190 			goto out;
1191 		}
1192 
1193 		acquired++;
1194 		timeout = remaining;
1195 	}
1196 
1197 	spin_lock(&gpu->event_spinlock);
1198 
1199 	for (i = 0; i < nr_events; i++) {
1200 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1201 
1202 		events[i] = event;
1203 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1204 		set_bit(event, gpu->event_bitmap);
1205 	}
1206 
1207 	spin_unlock(&gpu->event_spinlock);
1208 
1209 	for (i = 0; i < nr_events; i++) {
1210 		ret = pm_runtime_resume_and_get(gpu->dev);
1211 		if (ret)
1212 			goto out_rpm;
1213 		rpm_count++;
1214 	}
1215 
1216 	return 0;
1217 
1218 out_rpm:
1219 	for (i = 0; i < rpm_count; i++)
1220 		pm_runtime_put_autosuspend(gpu->dev);
1221 out:
1222 	for (i = 0; i < acquired; i++)
1223 		complete(&gpu->event_free);
1224 
1225 	return ret;
1226 }
1227 
1228 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1229 {
1230 	if (!test_bit(event, gpu->event_bitmap)) {
1231 		dev_warn(gpu->dev, "event %u is already marked as free",
1232 			 event);
1233 	} else {
1234 		clear_bit(event, gpu->event_bitmap);
1235 		complete(&gpu->event_free);
1236 	}
1237 
1238 	pm_runtime_put_autosuspend(gpu->dev);
1239 }
1240 
1241 /*
1242  * Cmdstream submission/retirement:
1243  */
1244 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1245 	u32 id, struct drm_etnaviv_timespec *timeout)
1246 {
1247 	struct dma_fence *fence;
1248 	int ret;
1249 
1250 	/*
1251 	 * Look up the fence and take a reference. We might still find a fence
1252 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1253 	 * pretends we didn't find a fence in that case.
1254 	 */
1255 	rcu_read_lock();
1256 	fence = xa_load(&gpu->user_fences, id);
1257 	if (fence)
1258 		fence = dma_fence_get_rcu(fence);
1259 	rcu_read_unlock();
1260 
1261 	if (!fence)
1262 		return 0;
1263 
1264 	if (!timeout) {
1265 		/* No timeout was requested: just test for completion */
1266 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1267 	} else {
1268 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1269 
1270 		ret = dma_fence_wait_timeout(fence, true, remaining);
1271 		if (ret == 0)
1272 			ret = -ETIMEDOUT;
1273 		else if (ret != -ERESTARTSYS)
1274 			ret = 0;
1275 
1276 	}
1277 
1278 	dma_fence_put(fence);
1279 	return ret;
1280 }
1281 
1282 /*
1283  * Wait for an object to become inactive.  This, on it's own, is not race
1284  * free: the object is moved by the scheduler off the active list, and
1285  * then the iova is put.  Moreover, the object could be re-submitted just
1286  * after we notice that it's become inactive.
1287  *
1288  * Although the retirement happens under the gpu lock, we don't want to hold
1289  * that lock in this function while waiting.
1290  */
1291 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1292 	struct etnaviv_gem_object *etnaviv_obj,
1293 	struct drm_etnaviv_timespec *timeout)
1294 {
1295 	unsigned long remaining;
1296 	long ret;
1297 
1298 	if (!timeout)
1299 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1300 
1301 	remaining = etnaviv_timeout_to_jiffies(timeout);
1302 
1303 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1304 					       !is_active(etnaviv_obj),
1305 					       remaining);
1306 	if (ret > 0)
1307 		return 0;
1308 	else if (ret == -ERESTARTSYS)
1309 		return -ERESTARTSYS;
1310 	else
1311 		return -ETIMEDOUT;
1312 }
1313 
1314 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1315 	struct etnaviv_event *event, unsigned int flags)
1316 {
1317 	const struct etnaviv_gem_submit *submit = event->submit;
1318 	unsigned int i;
1319 
1320 	for (i = 0; i < submit->nr_pmrs; i++) {
1321 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1322 
1323 		if (pmr->flags == flags)
1324 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1325 	}
1326 }
1327 
1328 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1329 	struct etnaviv_event *event)
1330 {
1331 	u32 val;
1332 
1333 	/* disable clock gating */
1334 	val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1335 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1336 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1337 
1338 	/* enable debug register */
1339 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1340 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1341 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1342 
1343 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1344 }
1345 
1346 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1347 	struct etnaviv_event *event)
1348 {
1349 	const struct etnaviv_gem_submit *submit = event->submit;
1350 	unsigned int i;
1351 	u32 val;
1352 
1353 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1354 
1355 	for (i = 0; i < submit->nr_pmrs; i++) {
1356 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1357 
1358 		*pmr->bo_vma = pmr->sequence;
1359 	}
1360 
1361 	/* disable debug register */
1362 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1363 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1364 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1365 
1366 	/* enable clock gating */
1367 	val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1368 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1369 	gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1370 }
1371 
1372 
1373 /* add bo's to gpu's ring, and kick gpu: */
1374 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1375 {
1376 	struct etnaviv_gpu *gpu = submit->gpu;
1377 	struct dma_fence *gpu_fence;
1378 	unsigned int i, nr_events = 1, event[3];
1379 	int ret;
1380 
1381 	/*
1382 	 * if there are performance monitor requests we need to have
1383 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1384 	 *   requests.
1385 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1386 	 *   and update the sequence number for userspace.
1387 	 */
1388 	if (submit->nr_pmrs)
1389 		nr_events = 3;
1390 
1391 	ret = event_alloc(gpu, nr_events, event);
1392 	if (ret) {
1393 		DRM_ERROR("no free events\n");
1394 		pm_runtime_put_noidle(gpu->dev);
1395 		return NULL;
1396 	}
1397 
1398 	mutex_lock(&gpu->lock);
1399 
1400 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1401 	if (!gpu_fence) {
1402 		for (i = 0; i < nr_events; i++)
1403 			event_free(gpu, event[i]);
1404 
1405 		goto out_unlock;
1406 	}
1407 
1408 	if (gpu->state == ETNA_GPU_STATE_INITIALIZED)
1409 		etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context);
1410 
1411 	if (submit->prev_mmu_context)
1412 		etnaviv_iommu_context_put(submit->prev_mmu_context);
1413 	submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context);
1414 
1415 	if (submit->nr_pmrs) {
1416 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1417 		kref_get(&submit->refcount);
1418 		gpu->event[event[1]].submit = submit;
1419 		etnaviv_sync_point_queue(gpu, event[1]);
1420 	}
1421 
1422 	gpu->event[event[0]].fence = gpu_fence;
1423 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1424 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1425 			     event[0], &submit->cmdbuf);
1426 
1427 	if (submit->nr_pmrs) {
1428 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1429 		kref_get(&submit->refcount);
1430 		gpu->event[event[2]].submit = submit;
1431 		etnaviv_sync_point_queue(gpu, event[2]);
1432 	}
1433 
1434 out_unlock:
1435 	mutex_unlock(&gpu->lock);
1436 
1437 	return gpu_fence;
1438 }
1439 
1440 static void sync_point_worker(struct work_struct *work)
1441 {
1442 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1443 					       sync_point_work);
1444 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1445 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1446 
1447 	event->sync_point(gpu, event);
1448 	etnaviv_submit_put(event->submit);
1449 	event_free(gpu, gpu->sync_point_event);
1450 
1451 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1452 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1453 }
1454 
1455 void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit)
1456 {
1457 	struct etnaviv_gpu *gpu = submit->gpu;
1458 	char *comm = NULL, *cmd = NULL;
1459 	struct task_struct *task;
1460 	unsigned int i;
1461 
1462 	dev_err(gpu->dev, "recover hung GPU!\n");
1463 
1464 	task = get_pid_task(submit->pid, PIDTYPE_PID);
1465 	if (task) {
1466 		comm = kstrdup(task->comm, GFP_KERNEL);
1467 		cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
1468 		put_task_struct(task);
1469 	}
1470 
1471 	if (comm && cmd)
1472 		dev_err(gpu->dev, "offending task: %s (%s)\n", comm, cmd);
1473 
1474 	kfree(cmd);
1475 	kfree(comm);
1476 
1477 	if (pm_runtime_get_sync(gpu->dev) < 0)
1478 		goto pm_put;
1479 
1480 	mutex_lock(&gpu->lock);
1481 
1482 	etnaviv_hw_reset(gpu);
1483 
1484 	/* complete all events, the GPU won't do it after the reset */
1485 	spin_lock(&gpu->event_spinlock);
1486 	for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1487 		event_free(gpu, i);
1488 	spin_unlock(&gpu->event_spinlock);
1489 
1490 	etnaviv_gpu_hw_init(gpu);
1491 
1492 	mutex_unlock(&gpu->lock);
1493 	pm_runtime_mark_last_busy(gpu->dev);
1494 pm_put:
1495 	pm_runtime_put_autosuspend(gpu->dev);
1496 }
1497 
1498 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1499 {
1500 	static const char *fault_reasons[] = {
1501 		"slave not present",
1502 		"page not present",
1503 		"write violation",
1504 		"out of bounds",
1505 		"read security violation",
1506 		"write security violation",
1507 	};
1508 
1509 	u32 status_reg, status;
1510 	int i;
1511 
1512 	if (gpu->sec_mode == ETNA_SEC_NONE)
1513 		status_reg = VIVS_MMUv2_STATUS;
1514 	else
1515 		status_reg = VIVS_MMUv2_SEC_STATUS;
1516 
1517 	status = gpu_read(gpu, status_reg);
1518 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1519 
1520 	for (i = 0; i < 4; i++) {
1521 		const char *reason = "unknown";
1522 		u32 address_reg;
1523 		u32 mmu_status;
1524 
1525 		mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK;
1526 		if (!mmu_status)
1527 			continue;
1528 
1529 		if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons))
1530 			reason = fault_reasons[mmu_status - 1];
1531 
1532 		if (gpu->sec_mode == ETNA_SEC_NONE)
1533 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1534 		else
1535 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1536 
1537 		dev_err_ratelimited(gpu->dev,
1538 				    "MMU %d fault (%s) addr 0x%08x\n",
1539 				    i, reason, gpu_read(gpu, address_reg));
1540 	}
1541 }
1542 
1543 static irqreturn_t irq_handler(int irq, void *data)
1544 {
1545 	struct etnaviv_gpu *gpu = data;
1546 	irqreturn_t ret = IRQ_NONE;
1547 
1548 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1549 
1550 	if (intr != 0) {
1551 		ktime_t now = ktime_get();
1552 		int event;
1553 
1554 		pm_runtime_mark_last_busy(gpu->dev);
1555 
1556 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1557 
1558 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1559 			dev_err(gpu->dev, "AXI bus error\n");
1560 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1561 		}
1562 
1563 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1564 			dump_mmu_fault(gpu);
1565 			gpu->state = ETNA_GPU_STATE_FAULT;
1566 			drm_sched_fault(&gpu->sched);
1567 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1568 		}
1569 
1570 		while ((event = ffs(intr)) != 0) {
1571 			struct dma_fence *fence;
1572 
1573 			event -= 1;
1574 
1575 			intr &= ~(1 << event);
1576 
1577 			dev_dbg(gpu->dev, "event %u\n", event);
1578 
1579 			if (gpu->event[event].sync_point) {
1580 				gpu->sync_point_event = event;
1581 				queue_work(gpu->wq, &gpu->sync_point_work);
1582 			}
1583 
1584 			fence = gpu->event[event].fence;
1585 			if (!fence)
1586 				continue;
1587 
1588 			gpu->event[event].fence = NULL;
1589 
1590 			/*
1591 			 * Events can be processed out of order.  Eg,
1592 			 * - allocate and queue event 0
1593 			 * - allocate event 1
1594 			 * - event 0 completes, we process it
1595 			 * - allocate and queue event 0
1596 			 * - event 1 and event 0 complete
1597 			 * we can end up processing event 0 first, then 1.
1598 			 */
1599 			if (fence_after(fence->seqno, gpu->completed_fence))
1600 				gpu->completed_fence = fence->seqno;
1601 			dma_fence_signal_timestamp(fence, now);
1602 
1603 			event_free(gpu, event);
1604 		}
1605 
1606 		ret = IRQ_HANDLED;
1607 	}
1608 
1609 	return ret;
1610 }
1611 
1612 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1613 {
1614 	int ret;
1615 
1616 	ret = clk_prepare_enable(gpu->clk_reg);
1617 	if (ret)
1618 		return ret;
1619 
1620 	ret = clk_prepare_enable(gpu->clk_bus);
1621 	if (ret)
1622 		goto disable_clk_reg;
1623 
1624 	ret = clk_prepare_enable(gpu->clk_core);
1625 	if (ret)
1626 		goto disable_clk_bus;
1627 
1628 	ret = clk_prepare_enable(gpu->clk_shader);
1629 	if (ret)
1630 		goto disable_clk_core;
1631 
1632 	return 0;
1633 
1634 disable_clk_core:
1635 	clk_disable_unprepare(gpu->clk_core);
1636 disable_clk_bus:
1637 	clk_disable_unprepare(gpu->clk_bus);
1638 disable_clk_reg:
1639 	clk_disable_unprepare(gpu->clk_reg);
1640 
1641 	return ret;
1642 }
1643 
1644 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1645 {
1646 	clk_disable_unprepare(gpu->clk_shader);
1647 	clk_disable_unprepare(gpu->clk_core);
1648 	clk_disable_unprepare(gpu->clk_bus);
1649 	clk_disable_unprepare(gpu->clk_reg);
1650 
1651 	return 0;
1652 }
1653 
1654 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1655 {
1656 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1657 
1658 	do {
1659 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1660 
1661 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1662 			return 0;
1663 
1664 		if (time_is_before_jiffies(timeout)) {
1665 			dev_warn(gpu->dev,
1666 				 "timed out waiting for idle: idle=0x%x\n",
1667 				 idle);
1668 			return -ETIMEDOUT;
1669 		}
1670 
1671 		udelay(5);
1672 	} while (1);
1673 }
1674 
1675 static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1676 {
1677 	if (gpu->state == ETNA_GPU_STATE_RUNNING) {
1678 		/* Replace the last WAIT with END */
1679 		mutex_lock(&gpu->lock);
1680 		etnaviv_buffer_end(gpu);
1681 		mutex_unlock(&gpu->lock);
1682 
1683 		/*
1684 		 * We know that only the FE is busy here, this should
1685 		 * happen quickly (as the WAIT is only 200 cycles).  If
1686 		 * we fail, just warn and continue.
1687 		 */
1688 		etnaviv_gpu_wait_idle(gpu, 100);
1689 
1690 		gpu->state = ETNA_GPU_STATE_INITIALIZED;
1691 	}
1692 
1693 	gpu->exec_state = -1;
1694 }
1695 
1696 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1697 {
1698 	int ret;
1699 
1700 	ret = mutex_lock_killable(&gpu->lock);
1701 	if (ret)
1702 		return ret;
1703 
1704 	etnaviv_gpu_update_clock(gpu);
1705 	etnaviv_gpu_hw_init(gpu);
1706 
1707 	mutex_unlock(&gpu->lock);
1708 
1709 	return 0;
1710 }
1711 
1712 static int
1713 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1714 				  unsigned long *state)
1715 {
1716 	*state = 6;
1717 
1718 	return 0;
1719 }
1720 
1721 static int
1722 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1723 				  unsigned long *state)
1724 {
1725 	struct etnaviv_gpu *gpu = cdev->devdata;
1726 
1727 	*state = gpu->freq_scale;
1728 
1729 	return 0;
1730 }
1731 
1732 static int
1733 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1734 				  unsigned long state)
1735 {
1736 	struct etnaviv_gpu *gpu = cdev->devdata;
1737 
1738 	mutex_lock(&gpu->lock);
1739 	gpu->freq_scale = state;
1740 	if (!pm_runtime_suspended(gpu->dev))
1741 		etnaviv_gpu_update_clock(gpu);
1742 	mutex_unlock(&gpu->lock);
1743 
1744 	return 0;
1745 }
1746 
1747 static const struct thermal_cooling_device_ops cooling_ops = {
1748 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1749 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1750 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1751 };
1752 
1753 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1754 	void *data)
1755 {
1756 	struct drm_device *drm = data;
1757 	struct etnaviv_drm_private *priv = drm->dev_private;
1758 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1759 	int ret;
1760 
1761 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1762 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1763 				(char *)dev_name(dev), gpu, &cooling_ops);
1764 		if (IS_ERR(gpu->cooling))
1765 			return PTR_ERR(gpu->cooling);
1766 	}
1767 
1768 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1769 	if (!gpu->wq) {
1770 		ret = -ENOMEM;
1771 		goto out_thermal;
1772 	}
1773 
1774 	ret = etnaviv_sched_init(gpu);
1775 	if (ret)
1776 		goto out_workqueue;
1777 
1778 	if (!IS_ENABLED(CONFIG_PM)) {
1779 		ret = etnaviv_gpu_clk_enable(gpu);
1780 		if (ret < 0)
1781 			goto out_sched;
1782 	}
1783 
1784 	gpu->drm = drm;
1785 	gpu->fence_context = dma_fence_context_alloc(1);
1786 	xa_init_flags(&gpu->user_fences, XA_FLAGS_ALLOC);
1787 	spin_lock_init(&gpu->fence_spinlock);
1788 
1789 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1790 	init_waitqueue_head(&gpu->fence_event);
1791 
1792 	priv->gpu[priv->num_gpus++] = gpu;
1793 
1794 	return 0;
1795 
1796 out_sched:
1797 	etnaviv_sched_fini(gpu);
1798 
1799 out_workqueue:
1800 	destroy_workqueue(gpu->wq);
1801 
1802 out_thermal:
1803 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1804 		thermal_cooling_device_unregister(gpu->cooling);
1805 
1806 	return ret;
1807 }
1808 
1809 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1810 	void *data)
1811 {
1812 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1813 
1814 	DBG("%s", dev_name(gpu->dev));
1815 
1816 	destroy_workqueue(gpu->wq);
1817 
1818 	etnaviv_sched_fini(gpu);
1819 
1820 	if (IS_ENABLED(CONFIG_PM)) {
1821 		pm_runtime_get_sync(gpu->dev);
1822 		pm_runtime_put_sync_suspend(gpu->dev);
1823 	} else {
1824 		etnaviv_gpu_hw_suspend(gpu);
1825 		etnaviv_gpu_clk_disable(gpu);
1826 	}
1827 
1828 	if (gpu->mmu_context)
1829 		etnaviv_iommu_context_put(gpu->mmu_context);
1830 
1831 	etnaviv_cmdbuf_free(&gpu->buffer);
1832 	etnaviv_iommu_global_fini(gpu);
1833 
1834 	gpu->drm = NULL;
1835 	xa_destroy(&gpu->user_fences);
1836 
1837 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1838 		thermal_cooling_device_unregister(gpu->cooling);
1839 	gpu->cooling = NULL;
1840 }
1841 
1842 static const struct component_ops gpu_ops = {
1843 	.bind = etnaviv_gpu_bind,
1844 	.unbind = etnaviv_gpu_unbind,
1845 };
1846 
1847 static const struct of_device_id etnaviv_gpu_match[] = {
1848 	{
1849 		.compatible = "vivante,gc"
1850 	},
1851 	{ /* sentinel */ }
1852 };
1853 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1854 
1855 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1856 {
1857 	struct device *dev = &pdev->dev;
1858 	struct etnaviv_gpu *gpu;
1859 	int err;
1860 
1861 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1862 	if (!gpu)
1863 		return -ENOMEM;
1864 
1865 	gpu->dev = &pdev->dev;
1866 	mutex_init(&gpu->lock);
1867 	mutex_init(&gpu->sched_lock);
1868 
1869 	/* Map registers: */
1870 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1871 	if (IS_ERR(gpu->mmio))
1872 		return PTR_ERR(gpu->mmio);
1873 
1874 	/* Get Interrupt: */
1875 	gpu->irq = platform_get_irq(pdev, 0);
1876 	if (gpu->irq < 0)
1877 		return gpu->irq;
1878 
1879 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1880 			       dev_name(gpu->dev), gpu);
1881 	if (err) {
1882 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1883 		return err;
1884 	}
1885 
1886 	/* Get Clocks: */
1887 	gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1888 	DBG("clk_reg: %p", gpu->clk_reg);
1889 	if (IS_ERR(gpu->clk_reg))
1890 		return PTR_ERR(gpu->clk_reg);
1891 
1892 	gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1893 	DBG("clk_bus: %p", gpu->clk_bus);
1894 	if (IS_ERR(gpu->clk_bus))
1895 		return PTR_ERR(gpu->clk_bus);
1896 
1897 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1898 	DBG("clk_core: %p", gpu->clk_core);
1899 	if (IS_ERR(gpu->clk_core))
1900 		return PTR_ERR(gpu->clk_core);
1901 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1902 
1903 	gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1904 	DBG("clk_shader: %p", gpu->clk_shader);
1905 	if (IS_ERR(gpu->clk_shader))
1906 		return PTR_ERR(gpu->clk_shader);
1907 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1908 
1909 	/* TODO: figure out max mapped size */
1910 	dev_set_drvdata(dev, gpu);
1911 
1912 	/*
1913 	 * We treat the device as initially suspended.  The runtime PM
1914 	 * autosuspend delay is rather arbitary: no measurements have
1915 	 * yet been performed to determine an appropriate value.
1916 	 */
1917 	pm_runtime_use_autosuspend(gpu->dev);
1918 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1919 	pm_runtime_enable(gpu->dev);
1920 
1921 	err = component_add(&pdev->dev, &gpu_ops);
1922 	if (err < 0) {
1923 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1924 		return err;
1925 	}
1926 
1927 	return 0;
1928 }
1929 
1930 static void etnaviv_gpu_platform_remove(struct platform_device *pdev)
1931 {
1932 	component_del(&pdev->dev, &gpu_ops);
1933 	pm_runtime_disable(&pdev->dev);
1934 }
1935 
1936 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1937 {
1938 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1939 	u32 idle, mask;
1940 
1941 	/* If there are any jobs in the HW queue, we're not idle */
1942 	if (atomic_read(&gpu->sched.credit_count))
1943 		return -EBUSY;
1944 
1945 	/* Check whether the hardware (except FE and MC) is idle */
1946 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1947 				  VIVS_HI_IDLE_STATE_MC);
1948 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1949 	if (idle != mask) {
1950 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1951 				     idle);
1952 		return -EBUSY;
1953 	}
1954 
1955 	etnaviv_gpu_hw_suspend(gpu);
1956 
1957 	gpu->state = ETNA_GPU_STATE_IDENTIFIED;
1958 
1959 	return etnaviv_gpu_clk_disable(gpu);
1960 }
1961 
1962 static int etnaviv_gpu_rpm_resume(struct device *dev)
1963 {
1964 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1965 	int ret;
1966 
1967 	ret = etnaviv_gpu_clk_enable(gpu);
1968 	if (ret)
1969 		return ret;
1970 
1971 	/* Re-initialise the basic hardware state */
1972 	if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) {
1973 		ret = etnaviv_gpu_hw_resume(gpu);
1974 		if (ret) {
1975 			etnaviv_gpu_clk_disable(gpu);
1976 			return ret;
1977 		}
1978 	}
1979 
1980 	return 0;
1981 }
1982 
1983 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1984 	RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL)
1985 };
1986 
1987 struct platform_driver etnaviv_gpu_driver = {
1988 	.driver = {
1989 		.name = "etnaviv-gpu",
1990 		.pm = pm_ptr(&etnaviv_gpu_pm_ops),
1991 		.of_match_table = etnaviv_gpu_match,
1992 	},
1993 	.probe = etnaviv_gpu_platform_probe,
1994 	.remove_new = etnaviv_gpu_platform_remove,
1995 	.id_table = gpu_ids,
1996 };
1997