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/dma-fence.h>
9 #include <linux/moduleparam.h>
10 #include <linux/of_device.h>
11 #include <linux/regulator/consumer.h>
12 #include <linux/thermal.h>
13 
14 #include "etnaviv_cmdbuf.h"
15 #include "etnaviv_dump.h"
16 #include "etnaviv_gpu.h"
17 #include "etnaviv_gem.h"
18 #include "etnaviv_mmu.h"
19 #include "etnaviv_perfmon.h"
20 #include "etnaviv_sched.h"
21 #include "common.xml.h"
22 #include "state.xml.h"
23 #include "state_hi.xml.h"
24 #include "cmdstream.xml.h"
25 
26 #ifndef PHYS_OFFSET
27 #define PHYS_OFFSET 0
28 #endif
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 	switch (param) {
42 	case ETNAVIV_PARAM_GPU_MODEL:
43 		*value = gpu->identity.model;
44 		break;
45 
46 	case ETNAVIV_PARAM_GPU_REVISION:
47 		*value = gpu->identity.revision;
48 		break;
49 
50 	case ETNAVIV_PARAM_GPU_FEATURES_0:
51 		*value = gpu->identity.features;
52 		break;
53 
54 	case ETNAVIV_PARAM_GPU_FEATURES_1:
55 		*value = gpu->identity.minor_features0;
56 		break;
57 
58 	case ETNAVIV_PARAM_GPU_FEATURES_2:
59 		*value = gpu->identity.minor_features1;
60 		break;
61 
62 	case ETNAVIV_PARAM_GPU_FEATURES_3:
63 		*value = gpu->identity.minor_features2;
64 		break;
65 
66 	case ETNAVIV_PARAM_GPU_FEATURES_4:
67 		*value = gpu->identity.minor_features3;
68 		break;
69 
70 	case ETNAVIV_PARAM_GPU_FEATURES_5:
71 		*value = gpu->identity.minor_features4;
72 		break;
73 
74 	case ETNAVIV_PARAM_GPU_FEATURES_6:
75 		*value = gpu->identity.minor_features5;
76 		break;
77 
78 	case ETNAVIV_PARAM_GPU_FEATURES_7:
79 		*value = gpu->identity.minor_features6;
80 		break;
81 
82 	case ETNAVIV_PARAM_GPU_FEATURES_8:
83 		*value = gpu->identity.minor_features7;
84 		break;
85 
86 	case ETNAVIV_PARAM_GPU_FEATURES_9:
87 		*value = gpu->identity.minor_features8;
88 		break;
89 
90 	case ETNAVIV_PARAM_GPU_FEATURES_10:
91 		*value = gpu->identity.minor_features9;
92 		break;
93 
94 	case ETNAVIV_PARAM_GPU_FEATURES_11:
95 		*value = gpu->identity.minor_features10;
96 		break;
97 
98 	case ETNAVIV_PARAM_GPU_FEATURES_12:
99 		*value = gpu->identity.minor_features11;
100 		break;
101 
102 	case ETNAVIV_PARAM_GPU_STREAM_COUNT:
103 		*value = gpu->identity.stream_count;
104 		break;
105 
106 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
107 		*value = gpu->identity.register_max;
108 		break;
109 
110 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
111 		*value = gpu->identity.thread_count;
112 		break;
113 
114 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
115 		*value = gpu->identity.vertex_cache_size;
116 		break;
117 
118 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
119 		*value = gpu->identity.shader_core_count;
120 		break;
121 
122 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
123 		*value = gpu->identity.pixel_pipes;
124 		break;
125 
126 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
127 		*value = gpu->identity.vertex_output_buffer_size;
128 		break;
129 
130 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
131 		*value = gpu->identity.buffer_size;
132 		break;
133 
134 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
135 		*value = gpu->identity.instruction_count;
136 		break;
137 
138 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
139 		*value = gpu->identity.num_constants;
140 		break;
141 
142 	case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
143 		*value = gpu->identity.varyings_count;
144 		break;
145 
146 	default:
147 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
148 		return -EINVAL;
149 	}
150 
151 	return 0;
152 }
153 
154 
155 #define etnaviv_is_model_rev(gpu, mod, rev) \
156 	((gpu)->identity.model == chipModel_##mod && \
157 	 (gpu)->identity.revision == rev)
158 #define etnaviv_field(val, field) \
159 	(((val) & field##__MASK) >> field##__SHIFT)
160 
161 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
162 {
163 	if (gpu->identity.minor_features0 &
164 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
165 		u32 specs[4];
166 		unsigned int streams;
167 
168 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
169 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
170 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
171 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
172 
173 		gpu->identity.stream_count = etnaviv_field(specs[0],
174 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
175 		gpu->identity.register_max = etnaviv_field(specs[0],
176 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
177 		gpu->identity.thread_count = etnaviv_field(specs[0],
178 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
179 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
180 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
181 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
182 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
183 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
184 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
185 		gpu->identity.vertex_output_buffer_size =
186 			etnaviv_field(specs[0],
187 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
188 
189 		gpu->identity.buffer_size = etnaviv_field(specs[1],
190 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
191 		gpu->identity.instruction_count = etnaviv_field(specs[1],
192 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
193 		gpu->identity.num_constants = etnaviv_field(specs[1],
194 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
195 
196 		gpu->identity.varyings_count = etnaviv_field(specs[2],
197 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
198 
199 		/* This overrides the value from older register if non-zero */
200 		streams = etnaviv_field(specs[3],
201 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
202 		if (streams)
203 			gpu->identity.stream_count = streams;
204 	}
205 
206 	/* Fill in the stream count if not specified */
207 	if (gpu->identity.stream_count == 0) {
208 		if (gpu->identity.model >= 0x1000)
209 			gpu->identity.stream_count = 4;
210 		else
211 			gpu->identity.stream_count = 1;
212 	}
213 
214 	/* Convert the register max value */
215 	if (gpu->identity.register_max)
216 		gpu->identity.register_max = 1 << gpu->identity.register_max;
217 	else if (gpu->identity.model == chipModel_GC400)
218 		gpu->identity.register_max = 32;
219 	else
220 		gpu->identity.register_max = 64;
221 
222 	/* Convert thread count */
223 	if (gpu->identity.thread_count)
224 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
225 	else if (gpu->identity.model == chipModel_GC400)
226 		gpu->identity.thread_count = 64;
227 	else if (gpu->identity.model == chipModel_GC500 ||
228 		 gpu->identity.model == chipModel_GC530)
229 		gpu->identity.thread_count = 128;
230 	else
231 		gpu->identity.thread_count = 256;
232 
233 	if (gpu->identity.vertex_cache_size == 0)
234 		gpu->identity.vertex_cache_size = 8;
235 
236 	if (gpu->identity.shader_core_count == 0) {
237 		if (gpu->identity.model >= 0x1000)
238 			gpu->identity.shader_core_count = 2;
239 		else
240 			gpu->identity.shader_core_count = 1;
241 	}
242 
243 	if (gpu->identity.pixel_pipes == 0)
244 		gpu->identity.pixel_pipes = 1;
245 
246 	/* Convert virtex buffer size */
247 	if (gpu->identity.vertex_output_buffer_size) {
248 		gpu->identity.vertex_output_buffer_size =
249 			1 << gpu->identity.vertex_output_buffer_size;
250 	} else if (gpu->identity.model == chipModel_GC400) {
251 		if (gpu->identity.revision < 0x4000)
252 			gpu->identity.vertex_output_buffer_size = 512;
253 		else if (gpu->identity.revision < 0x4200)
254 			gpu->identity.vertex_output_buffer_size = 256;
255 		else
256 			gpu->identity.vertex_output_buffer_size = 128;
257 	} else {
258 		gpu->identity.vertex_output_buffer_size = 512;
259 	}
260 
261 	switch (gpu->identity.instruction_count) {
262 	case 0:
263 		if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
264 		    gpu->identity.model == chipModel_GC880)
265 			gpu->identity.instruction_count = 512;
266 		else
267 			gpu->identity.instruction_count = 256;
268 		break;
269 
270 	case 1:
271 		gpu->identity.instruction_count = 1024;
272 		break;
273 
274 	case 2:
275 		gpu->identity.instruction_count = 2048;
276 		break;
277 
278 	default:
279 		gpu->identity.instruction_count = 256;
280 		break;
281 	}
282 
283 	if (gpu->identity.num_constants == 0)
284 		gpu->identity.num_constants = 168;
285 
286 	if (gpu->identity.varyings_count == 0) {
287 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
288 			gpu->identity.varyings_count = 12;
289 		else
290 			gpu->identity.varyings_count = 8;
291 	}
292 
293 	/*
294 	 * For some cores, two varyings are consumed for position, so the
295 	 * maximum varying count needs to be reduced by one.
296 	 */
297 	if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
298 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
299 	    etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
300 	    etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
301 	    etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
302 	    etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
303 	    etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
304 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
305 	    etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
306 	    etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
307 	    etnaviv_is_model_rev(gpu, GC880, 0x5106))
308 		gpu->identity.varyings_count -= 1;
309 }
310 
311 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
312 {
313 	u32 chipIdentity;
314 
315 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
316 
317 	/* Special case for older graphic cores. */
318 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
319 		gpu->identity.model    = chipModel_GC500;
320 		gpu->identity.revision = etnaviv_field(chipIdentity,
321 					 VIVS_HI_CHIP_IDENTITY_REVISION);
322 	} else {
323 
324 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
325 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
326 
327 		/*
328 		 * !!!! HACK ALERT !!!!
329 		 * Because people change device IDs without letting software
330 		 * know about it - here is the hack to make it all look the
331 		 * same.  Only for GC400 family.
332 		 */
333 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
334 		    gpu->identity.model != chipModel_GC420) {
335 			gpu->identity.model = gpu->identity.model & 0x0400;
336 		}
337 
338 		/* Another special case */
339 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
340 			u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
341 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
342 
343 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
344 				/*
345 				 * This IP has an ECO; put the correct
346 				 * revision in it.
347 				 */
348 				gpu->identity.revision = 0x1051;
349 			}
350 		}
351 
352 		/*
353 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
354 		 * reality it's just a re-branded GC3000. We can identify this
355 		 * core by the upper half of the revision register being all 1.
356 		 * Fix model/rev here, so all other places can refer to this
357 		 * core by its real identity.
358 		 */
359 		if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
360 			gpu->identity.model = chipModel_GC3000;
361 			gpu->identity.revision &= 0xffff;
362 		}
363 	}
364 
365 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
366 		 gpu->identity.model, gpu->identity.revision);
367 
368 	gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
369 	/*
370 	 * If there is a match in the HWDB, we aren't interested in the
371 	 * remaining register values, as they might be wrong.
372 	 */
373 	if (etnaviv_fill_identity_from_hwdb(gpu))
374 		return;
375 
376 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
377 
378 	/* Disable fast clear on GC700. */
379 	if (gpu->identity.model == chipModel_GC700)
380 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
381 
382 	if ((gpu->identity.model == chipModel_GC500 &&
383 	     gpu->identity.revision < 2) ||
384 	    (gpu->identity.model == chipModel_GC300 &&
385 	     gpu->identity.revision < 0x2000)) {
386 
387 		/*
388 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
389 		 * registers.
390 		 */
391 		gpu->identity.minor_features0 = 0;
392 		gpu->identity.minor_features1 = 0;
393 		gpu->identity.minor_features2 = 0;
394 		gpu->identity.minor_features3 = 0;
395 		gpu->identity.minor_features4 = 0;
396 		gpu->identity.minor_features5 = 0;
397 	} else
398 		gpu->identity.minor_features0 =
399 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
400 
401 	if (gpu->identity.minor_features0 &
402 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
403 		gpu->identity.minor_features1 =
404 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
405 		gpu->identity.minor_features2 =
406 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
407 		gpu->identity.minor_features3 =
408 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
409 		gpu->identity.minor_features4 =
410 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
411 		gpu->identity.minor_features5 =
412 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
413 	}
414 
415 	/* GC600 idle register reports zero bits where modules aren't present */
416 	if (gpu->identity.model == chipModel_GC600)
417 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
418 				 VIVS_HI_IDLE_STATE_RA |
419 				 VIVS_HI_IDLE_STATE_SE |
420 				 VIVS_HI_IDLE_STATE_PA |
421 				 VIVS_HI_IDLE_STATE_SH |
422 				 VIVS_HI_IDLE_STATE_PE |
423 				 VIVS_HI_IDLE_STATE_DE |
424 				 VIVS_HI_IDLE_STATE_FE;
425 
426 	etnaviv_hw_specs(gpu);
427 }
428 
429 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
430 {
431 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
432 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
433 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
434 }
435 
436 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
437 {
438 	if (gpu->identity.minor_features2 &
439 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
440 		clk_set_rate(gpu->clk_core,
441 			     gpu->base_rate_core >> gpu->freq_scale);
442 		clk_set_rate(gpu->clk_shader,
443 			     gpu->base_rate_shader >> gpu->freq_scale);
444 	} else {
445 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
446 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
447 
448 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
449 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
450 		etnaviv_gpu_load_clock(gpu, clock);
451 	}
452 }
453 
454 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
455 {
456 	u32 control, idle;
457 	unsigned long timeout;
458 	bool failed = true;
459 
460 	/* We hope that the GPU resets in under one second */
461 	timeout = jiffies + msecs_to_jiffies(1000);
462 
463 	while (time_is_after_jiffies(timeout)) {
464 		/* enable clock */
465 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
466 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
467 		etnaviv_gpu_load_clock(gpu, control);
468 
469 		/* isolate the GPU. */
470 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
471 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
472 
473 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
474 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
475 			          VIVS_MMUv2_AHB_CONTROL_RESET);
476 		} else {
477 			/* set soft reset. */
478 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
479 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
480 		}
481 
482 		/* wait for reset. */
483 		usleep_range(10, 20);
484 
485 		/* reset soft reset bit. */
486 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
487 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
488 
489 		/* reset GPU isolation. */
490 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
491 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
492 
493 		/* read idle register. */
494 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
495 
496 		/* try reseting again if FE it not idle */
497 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
498 			dev_dbg(gpu->dev, "FE is not idle\n");
499 			continue;
500 		}
501 
502 		/* read reset register. */
503 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
504 
505 		/* is the GPU idle? */
506 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
507 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
508 			dev_dbg(gpu->dev, "GPU is not idle\n");
509 			continue;
510 		}
511 
512 		/* disable debug registers, as they are not normally needed */
513 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
514 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
515 
516 		failed = false;
517 		break;
518 	}
519 
520 	if (failed) {
521 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
522 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
523 
524 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
525 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
526 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
527 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
528 
529 		return -EBUSY;
530 	}
531 
532 	/* We rely on the GPU running, so program the clock */
533 	etnaviv_gpu_update_clock(gpu);
534 
535 	return 0;
536 }
537 
538 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
539 {
540 	u32 pmc, ppc;
541 
542 	/* enable clock gating */
543 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
544 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
545 
546 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
547 	if (gpu->identity.revision == 0x4301 ||
548 	    gpu->identity.revision == 0x4302)
549 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
550 
551 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
552 
553 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
554 
555 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
556 	if (gpu->identity.model >= chipModel_GC400 &&
557 	    gpu->identity.model != chipModel_GC420 &&
558 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
559 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
560 
561 	/*
562 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
563 	 * present without a bug fix.
564 	 */
565 	if (gpu->identity.revision < 0x5000 &&
566 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
567 	    !(gpu->identity.minor_features1 &
568 	      chipMinorFeatures1_DISABLE_PE_GATING))
569 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
570 
571 	if (gpu->identity.revision < 0x5422)
572 		pmc |= BIT(15); /* Unknown bit */
573 
574 	/* Disable TX clock gating on affected core revisions. */
575 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
576 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108))
577 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
578 
579 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
580 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
581 
582 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
583 }
584 
585 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
586 {
587 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
588 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
589 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
590 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
591 
592 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
593 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
594 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
595 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
596 	}
597 }
598 
599 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
600 {
601 	/*
602 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
603 	 * cannot be read, extracted from vivante kernel driver.
604 	 */
605 	u32 pulse_eater = 0x01590880;
606 
607 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
608 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
609 		pulse_eater |= BIT(23);
610 
611 	}
612 
613 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
614 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
615 		pulse_eater &= ~BIT(16);
616 		pulse_eater |= BIT(17);
617 	}
618 
619 	if ((gpu->identity.revision > 0x5420) &&
620 	    (gpu->identity.features & chipFeatures_PIPE_3D))
621 	{
622 		/* Performance fix: disable internal DFS */
623 		pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
624 		pulse_eater |= BIT(18);
625 	}
626 
627 	gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
628 }
629 
630 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
631 {
632 	u16 prefetch;
633 
634 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
635 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
636 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
637 		u32 mc_memory_debug;
638 
639 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
640 
641 		if (gpu->identity.revision == 0x5007)
642 			mc_memory_debug |= 0x0c;
643 		else
644 			mc_memory_debug |= 0x08;
645 
646 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
647 	}
648 
649 	/* enable module-level clock gating */
650 	etnaviv_gpu_enable_mlcg(gpu);
651 
652 	/*
653 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
654 	 * This is necessary to prevent the iMX6 SoC locking up.
655 	 */
656 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
657 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
658 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
659 
660 	/* GC2000 rev 5108 needs a special bus config */
661 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
662 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
663 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
664 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
665 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
666 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
667 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
668 	}
669 
670 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
671 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
672 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
673 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
674 	}
675 
676 	/* setup the pulse eater */
677 	etnaviv_gpu_setup_pulse_eater(gpu);
678 
679 	/* setup the MMU */
680 	etnaviv_iommu_restore(gpu);
681 
682 	/* Start command processor */
683 	prefetch = etnaviv_buffer_init(gpu);
684 
685 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
686 	etnaviv_gpu_start_fe(gpu, etnaviv_cmdbuf_get_va(&gpu->buffer),
687 			     prefetch);
688 }
689 
690 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
691 {
692 	int ret, i;
693 
694 	ret = pm_runtime_get_sync(gpu->dev);
695 	if (ret < 0) {
696 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
697 		return ret;
698 	}
699 
700 	etnaviv_hw_identify(gpu);
701 
702 	if (gpu->identity.model == 0) {
703 		dev_err(gpu->dev, "Unknown GPU model\n");
704 		ret = -ENXIO;
705 		goto fail;
706 	}
707 
708 	/* Exclude VG cores with FE2.0 */
709 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
710 	    gpu->identity.features & chipFeatures_FE20) {
711 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
712 		ret = -ENXIO;
713 		goto fail;
714 	}
715 
716 	/*
717 	 * Set the GPU linear window to be at the end of the DMA window, where
718 	 * the CMA area is likely to reside. This ensures that we are able to
719 	 * map the command buffers while having the linear window overlap as
720 	 * much RAM as possible, so we can optimize mappings for other buffers.
721 	 *
722 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
723 	 * to different views of the memory on the individual engines.
724 	 */
725 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
726 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
727 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
728 		if (dma_mask < PHYS_OFFSET + SZ_2G)
729 			gpu->memory_base = PHYS_OFFSET;
730 		else
731 			gpu->memory_base = dma_mask - SZ_2G + 1;
732 	} else if (PHYS_OFFSET >= SZ_2G) {
733 		dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
734 		gpu->memory_base = PHYS_OFFSET;
735 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
736 	}
737 
738 	/*
739 	 * On cores with security features supported, we claim control over the
740 	 * security states.
741 	 */
742 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
743 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
744 		gpu->sec_mode = ETNA_SEC_KERNEL;
745 
746 	ret = etnaviv_hw_reset(gpu);
747 	if (ret) {
748 		dev_err(gpu->dev, "GPU reset failed\n");
749 		goto fail;
750 	}
751 
752 	gpu->mmu = etnaviv_iommu_new(gpu);
753 	if (IS_ERR(gpu->mmu)) {
754 		dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n");
755 		ret = PTR_ERR(gpu->mmu);
756 		goto fail;
757 	}
758 
759 	gpu->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(gpu);
760 	if (IS_ERR(gpu->cmdbuf_suballoc)) {
761 		dev_err(gpu->dev, "Failed to create cmdbuf suballocator\n");
762 		ret = PTR_ERR(gpu->cmdbuf_suballoc);
763 		goto destroy_iommu;
764 	}
765 
766 	/* Create buffer: */
767 	ret = etnaviv_cmdbuf_init(gpu->cmdbuf_suballoc, &gpu->buffer,
768 				  PAGE_SIZE);
769 	if (ret) {
770 		dev_err(gpu->dev, "could not create command buffer\n");
771 		goto destroy_suballoc;
772 	}
773 
774 	if (gpu->mmu->version == ETNAVIV_IOMMU_V1 &&
775 	    etnaviv_cmdbuf_get_va(&gpu->buffer) > 0x80000000) {
776 		ret = -EINVAL;
777 		dev_err(gpu->dev,
778 			"command buffer outside valid memory window\n");
779 		goto free_buffer;
780 	}
781 
782 	/* Setup event management */
783 	spin_lock_init(&gpu->event_spinlock);
784 	init_completion(&gpu->event_free);
785 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
786 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
787 		complete(&gpu->event_free);
788 
789 	/* Now program the hardware */
790 	mutex_lock(&gpu->lock);
791 	etnaviv_gpu_hw_init(gpu);
792 	gpu->exec_state = -1;
793 	mutex_unlock(&gpu->lock);
794 
795 	pm_runtime_mark_last_busy(gpu->dev);
796 	pm_runtime_put_autosuspend(gpu->dev);
797 
798 	return 0;
799 
800 free_buffer:
801 	etnaviv_cmdbuf_free(&gpu->buffer);
802 	gpu->buffer.suballoc = NULL;
803 destroy_suballoc:
804 	etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc);
805 	gpu->cmdbuf_suballoc = NULL;
806 destroy_iommu:
807 	etnaviv_iommu_destroy(gpu->mmu);
808 	gpu->mmu = NULL;
809 fail:
810 	pm_runtime_mark_last_busy(gpu->dev);
811 	pm_runtime_put_autosuspend(gpu->dev);
812 
813 	return ret;
814 }
815 
816 #ifdef CONFIG_DEBUG_FS
817 struct dma_debug {
818 	u32 address[2];
819 	u32 state[2];
820 };
821 
822 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
823 {
824 	u32 i;
825 
826 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
827 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
828 
829 	for (i = 0; i < 500; i++) {
830 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
831 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
832 
833 		if (debug->address[0] != debug->address[1])
834 			break;
835 
836 		if (debug->state[0] != debug->state[1])
837 			break;
838 	}
839 }
840 
841 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
842 {
843 	struct dma_debug debug;
844 	u32 dma_lo, dma_hi, axi, idle;
845 	int ret;
846 
847 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
848 
849 	ret = pm_runtime_get_sync(gpu->dev);
850 	if (ret < 0)
851 		return ret;
852 
853 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
854 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
855 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
856 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
857 
858 	verify_dma(gpu, &debug);
859 
860 	seq_puts(m, "\tfeatures\n");
861 	seq_printf(m, "\t major_features: 0x%08x\n",
862 		   gpu->identity.features);
863 	seq_printf(m, "\t minor_features0: 0x%08x\n",
864 		   gpu->identity.minor_features0);
865 	seq_printf(m, "\t minor_features1: 0x%08x\n",
866 		   gpu->identity.minor_features1);
867 	seq_printf(m, "\t minor_features2: 0x%08x\n",
868 		   gpu->identity.minor_features2);
869 	seq_printf(m, "\t minor_features3: 0x%08x\n",
870 		   gpu->identity.minor_features3);
871 	seq_printf(m, "\t minor_features4: 0x%08x\n",
872 		   gpu->identity.minor_features4);
873 	seq_printf(m, "\t minor_features5: 0x%08x\n",
874 		   gpu->identity.minor_features5);
875 	seq_printf(m, "\t minor_features6: 0x%08x\n",
876 		   gpu->identity.minor_features6);
877 	seq_printf(m, "\t minor_features7: 0x%08x\n",
878 		   gpu->identity.minor_features7);
879 	seq_printf(m, "\t minor_features8: 0x%08x\n",
880 		   gpu->identity.minor_features8);
881 	seq_printf(m, "\t minor_features9: 0x%08x\n",
882 		   gpu->identity.minor_features9);
883 	seq_printf(m, "\t minor_features10: 0x%08x\n",
884 		   gpu->identity.minor_features10);
885 	seq_printf(m, "\t minor_features11: 0x%08x\n",
886 		   gpu->identity.minor_features11);
887 
888 	seq_puts(m, "\tspecs\n");
889 	seq_printf(m, "\t stream_count:  %d\n",
890 			gpu->identity.stream_count);
891 	seq_printf(m, "\t register_max: %d\n",
892 			gpu->identity.register_max);
893 	seq_printf(m, "\t thread_count: %d\n",
894 			gpu->identity.thread_count);
895 	seq_printf(m, "\t vertex_cache_size: %d\n",
896 			gpu->identity.vertex_cache_size);
897 	seq_printf(m, "\t shader_core_count: %d\n",
898 			gpu->identity.shader_core_count);
899 	seq_printf(m, "\t pixel_pipes: %d\n",
900 			gpu->identity.pixel_pipes);
901 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
902 			gpu->identity.vertex_output_buffer_size);
903 	seq_printf(m, "\t buffer_size: %d\n",
904 			gpu->identity.buffer_size);
905 	seq_printf(m, "\t instruction_count: %d\n",
906 			gpu->identity.instruction_count);
907 	seq_printf(m, "\t num_constants: %d\n",
908 			gpu->identity.num_constants);
909 	seq_printf(m, "\t varyings_count: %d\n",
910 			gpu->identity.varyings_count);
911 
912 	seq_printf(m, "\taxi: 0x%08x\n", axi);
913 	seq_printf(m, "\tidle: 0x%08x\n", idle);
914 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
915 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
916 		seq_puts(m, "\t FE is not idle\n");
917 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
918 		seq_puts(m, "\t DE is not idle\n");
919 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
920 		seq_puts(m, "\t PE is not idle\n");
921 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
922 		seq_puts(m, "\t SH is not idle\n");
923 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
924 		seq_puts(m, "\t PA is not idle\n");
925 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
926 		seq_puts(m, "\t SE is not idle\n");
927 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
928 		seq_puts(m, "\t RA is not idle\n");
929 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
930 		seq_puts(m, "\t TX is not idle\n");
931 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
932 		seq_puts(m, "\t VG is not idle\n");
933 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
934 		seq_puts(m, "\t IM is not idle\n");
935 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
936 		seq_puts(m, "\t FP is not idle\n");
937 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
938 		seq_puts(m, "\t TS is not idle\n");
939 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
940 		seq_puts(m, "\t AXI low power mode\n");
941 
942 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
943 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
944 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
945 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
946 
947 		seq_puts(m, "\tMC\n");
948 		seq_printf(m, "\t read0: 0x%08x\n", read0);
949 		seq_printf(m, "\t read1: 0x%08x\n", read1);
950 		seq_printf(m, "\t write: 0x%08x\n", write);
951 	}
952 
953 	seq_puts(m, "\tDMA ");
954 
955 	if (debug.address[0] == debug.address[1] &&
956 	    debug.state[0] == debug.state[1]) {
957 		seq_puts(m, "seems to be stuck\n");
958 	} else if (debug.address[0] == debug.address[1]) {
959 		seq_puts(m, "address is constant\n");
960 	} else {
961 		seq_puts(m, "is running\n");
962 	}
963 
964 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
965 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
966 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
967 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
968 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
969 		   dma_lo, dma_hi);
970 
971 	ret = 0;
972 
973 	pm_runtime_mark_last_busy(gpu->dev);
974 	pm_runtime_put_autosuspend(gpu->dev);
975 
976 	return ret;
977 }
978 #endif
979 
980 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
981 {
982 	unsigned int i = 0;
983 
984 	dev_err(gpu->dev, "recover hung GPU!\n");
985 
986 	if (pm_runtime_get_sync(gpu->dev) < 0)
987 		return;
988 
989 	mutex_lock(&gpu->lock);
990 
991 	etnaviv_hw_reset(gpu);
992 
993 	/* complete all events, the GPU won't do it after the reset */
994 	spin_lock(&gpu->event_spinlock);
995 	for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
996 		complete(&gpu->event_free);
997 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
998 	spin_unlock(&gpu->event_spinlock);
999 
1000 	etnaviv_gpu_hw_init(gpu);
1001 	gpu->exec_state = -1;
1002 
1003 	mutex_unlock(&gpu->lock);
1004 	pm_runtime_mark_last_busy(gpu->dev);
1005 	pm_runtime_put_autosuspend(gpu->dev);
1006 }
1007 
1008 /* fence object management */
1009 struct etnaviv_fence {
1010 	struct etnaviv_gpu *gpu;
1011 	struct dma_fence base;
1012 };
1013 
1014 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1015 {
1016 	return container_of(fence, struct etnaviv_fence, base);
1017 }
1018 
1019 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1020 {
1021 	return "etnaviv";
1022 }
1023 
1024 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1025 {
1026 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1027 
1028 	return dev_name(f->gpu->dev);
1029 }
1030 
1031 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1032 {
1033 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1034 
1035 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1036 }
1037 
1038 static void etnaviv_fence_release(struct dma_fence *fence)
1039 {
1040 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1041 
1042 	kfree_rcu(f, base.rcu);
1043 }
1044 
1045 static const struct dma_fence_ops etnaviv_fence_ops = {
1046 	.get_driver_name = etnaviv_fence_get_driver_name,
1047 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1048 	.signaled = etnaviv_fence_signaled,
1049 	.release = etnaviv_fence_release,
1050 };
1051 
1052 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1053 {
1054 	struct etnaviv_fence *f;
1055 
1056 	/*
1057 	 * GPU lock must already be held, otherwise fence completion order might
1058 	 * not match the seqno order assigned here.
1059 	 */
1060 	lockdep_assert_held(&gpu->lock);
1061 
1062 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1063 	if (!f)
1064 		return NULL;
1065 
1066 	f->gpu = gpu;
1067 
1068 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1069 		       gpu->fence_context, ++gpu->next_fence);
1070 
1071 	return &f->base;
1072 }
1073 
1074 /* returns true if fence a comes after fence b */
1075 static inline bool fence_after(u32 a, u32 b)
1076 {
1077 	return (s32)(a - b) > 0;
1078 }
1079 
1080 /*
1081  * event management:
1082  */
1083 
1084 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1085 	unsigned int *events)
1086 {
1087 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1088 	unsigned i, acquired = 0;
1089 
1090 	for (i = 0; i < nr_events; i++) {
1091 		unsigned long ret;
1092 
1093 		ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1094 
1095 		if (!ret) {
1096 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1097 			goto out;
1098 		}
1099 
1100 		acquired++;
1101 		timeout = ret;
1102 	}
1103 
1104 	spin_lock(&gpu->event_spinlock);
1105 
1106 	for (i = 0; i < nr_events; i++) {
1107 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1108 
1109 		events[i] = event;
1110 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1111 		set_bit(event, gpu->event_bitmap);
1112 	}
1113 
1114 	spin_unlock(&gpu->event_spinlock);
1115 
1116 	return 0;
1117 
1118 out:
1119 	for (i = 0; i < acquired; i++)
1120 		complete(&gpu->event_free);
1121 
1122 	return -EBUSY;
1123 }
1124 
1125 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1126 {
1127 	if (!test_bit(event, gpu->event_bitmap)) {
1128 		dev_warn(gpu->dev, "event %u is already marked as free",
1129 			 event);
1130 	} else {
1131 		clear_bit(event, gpu->event_bitmap);
1132 		complete(&gpu->event_free);
1133 	}
1134 }
1135 
1136 /*
1137  * Cmdstream submission/retirement:
1138  */
1139 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1140 	u32 id, struct timespec *timeout)
1141 {
1142 	struct dma_fence *fence;
1143 	int ret;
1144 
1145 	/*
1146 	 * Look up the fence and take a reference. We might still find a fence
1147 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1148 	 * pretends we didn't find a fence in that case.
1149 	 */
1150 	rcu_read_lock();
1151 	fence = idr_find(&gpu->fence_idr, id);
1152 	if (fence)
1153 		fence = dma_fence_get_rcu(fence);
1154 	rcu_read_unlock();
1155 
1156 	if (!fence)
1157 		return 0;
1158 
1159 	if (!timeout) {
1160 		/* No timeout was requested: just test for completion */
1161 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1162 	} else {
1163 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1164 
1165 		ret = dma_fence_wait_timeout(fence, true, remaining);
1166 		if (ret == 0)
1167 			ret = -ETIMEDOUT;
1168 		else if (ret != -ERESTARTSYS)
1169 			ret = 0;
1170 
1171 	}
1172 
1173 	dma_fence_put(fence);
1174 	return ret;
1175 }
1176 
1177 /*
1178  * Wait for an object to become inactive.  This, on it's own, is not race
1179  * free: the object is moved by the scheduler off the active list, and
1180  * then the iova is put.  Moreover, the object could be re-submitted just
1181  * after we notice that it's become inactive.
1182  *
1183  * Although the retirement happens under the gpu lock, we don't want to hold
1184  * that lock in this function while waiting.
1185  */
1186 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1187 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1188 {
1189 	unsigned long remaining;
1190 	long ret;
1191 
1192 	if (!timeout)
1193 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1194 
1195 	remaining = etnaviv_timeout_to_jiffies(timeout);
1196 
1197 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1198 					       !is_active(etnaviv_obj),
1199 					       remaining);
1200 	if (ret > 0)
1201 		return 0;
1202 	else if (ret == -ERESTARTSYS)
1203 		return -ERESTARTSYS;
1204 	else
1205 		return -ETIMEDOUT;
1206 }
1207 
1208 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1209 	struct etnaviv_event *event, unsigned int flags)
1210 {
1211 	const struct etnaviv_gem_submit *submit = event->submit;
1212 	unsigned int i;
1213 
1214 	for (i = 0; i < submit->nr_pmrs; i++) {
1215 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1216 
1217 		if (pmr->flags == flags)
1218 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1219 	}
1220 }
1221 
1222 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1223 	struct etnaviv_event *event)
1224 {
1225 	u32 val;
1226 
1227 	/* disable clock gating */
1228 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1229 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1230 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1231 
1232 	/* enable debug register */
1233 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1234 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1235 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1236 
1237 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1238 }
1239 
1240 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1241 	struct etnaviv_event *event)
1242 {
1243 	const struct etnaviv_gem_submit *submit = event->submit;
1244 	unsigned int i;
1245 	u32 val;
1246 
1247 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1248 
1249 	for (i = 0; i < submit->nr_pmrs; i++) {
1250 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1251 
1252 		*pmr->bo_vma = pmr->sequence;
1253 	}
1254 
1255 	/* disable debug register */
1256 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1257 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1258 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1259 
1260 	/* enable clock gating */
1261 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1262 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1263 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1264 }
1265 
1266 
1267 /* add bo's to gpu's ring, and kick gpu: */
1268 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1269 {
1270 	struct etnaviv_gpu *gpu = submit->gpu;
1271 	struct dma_fence *gpu_fence;
1272 	unsigned int i, nr_events = 1, event[3];
1273 	int ret;
1274 
1275 	if (!submit->runtime_resumed) {
1276 		ret = pm_runtime_get_sync(gpu->dev);
1277 		if (ret < 0)
1278 			return NULL;
1279 		submit->runtime_resumed = true;
1280 	}
1281 
1282 	/*
1283 	 * if there are performance monitor requests we need to have
1284 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1285 	 *   requests.
1286 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1287 	 *   and update the sequence number for userspace.
1288 	 */
1289 	if (submit->nr_pmrs)
1290 		nr_events = 3;
1291 
1292 	ret = event_alloc(gpu, nr_events, event);
1293 	if (ret) {
1294 		DRM_ERROR("no free events\n");
1295 		return NULL;
1296 	}
1297 
1298 	mutex_lock(&gpu->lock);
1299 
1300 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1301 	if (!gpu_fence) {
1302 		for (i = 0; i < nr_events; i++)
1303 			event_free(gpu, event[i]);
1304 
1305 		goto out_unlock;
1306 	}
1307 
1308 	if (submit->nr_pmrs) {
1309 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1310 		kref_get(&submit->refcount);
1311 		gpu->event[event[1]].submit = submit;
1312 		etnaviv_sync_point_queue(gpu, event[1]);
1313 	}
1314 
1315 	gpu->event[event[0]].fence = gpu_fence;
1316 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1317 	etnaviv_buffer_queue(gpu, submit->exec_state, event[0],
1318 			     &submit->cmdbuf);
1319 
1320 	if (submit->nr_pmrs) {
1321 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1322 		kref_get(&submit->refcount);
1323 		gpu->event[event[2]].submit = submit;
1324 		etnaviv_sync_point_queue(gpu, event[2]);
1325 	}
1326 
1327 out_unlock:
1328 	mutex_unlock(&gpu->lock);
1329 
1330 	return gpu_fence;
1331 }
1332 
1333 static void sync_point_worker(struct work_struct *work)
1334 {
1335 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1336 					       sync_point_work);
1337 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1338 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1339 
1340 	event->sync_point(gpu, event);
1341 	etnaviv_submit_put(event->submit);
1342 	event_free(gpu, gpu->sync_point_event);
1343 
1344 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1345 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1346 }
1347 
1348 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1349 {
1350 	u32 status_reg, status;
1351 	int i;
1352 
1353 	if (gpu->sec_mode == ETNA_SEC_NONE)
1354 		status_reg = VIVS_MMUv2_STATUS;
1355 	else
1356 		status_reg = VIVS_MMUv2_SEC_STATUS;
1357 
1358 	status = gpu_read(gpu, status_reg);
1359 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1360 
1361 	for (i = 0; i < 4; i++) {
1362 		u32 address_reg;
1363 
1364 		if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1365 			continue;
1366 
1367 		if (gpu->sec_mode == ETNA_SEC_NONE)
1368 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1369 		else
1370 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1371 
1372 		dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1373 				    gpu_read(gpu, address_reg));
1374 	}
1375 }
1376 
1377 static irqreturn_t irq_handler(int irq, void *data)
1378 {
1379 	struct etnaviv_gpu *gpu = data;
1380 	irqreturn_t ret = IRQ_NONE;
1381 
1382 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1383 
1384 	if (intr != 0) {
1385 		int event;
1386 
1387 		pm_runtime_mark_last_busy(gpu->dev);
1388 
1389 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1390 
1391 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1392 			dev_err(gpu->dev, "AXI bus error\n");
1393 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1394 		}
1395 
1396 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1397 			dump_mmu_fault(gpu);
1398 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1399 		}
1400 
1401 		while ((event = ffs(intr)) != 0) {
1402 			struct dma_fence *fence;
1403 
1404 			event -= 1;
1405 
1406 			intr &= ~(1 << event);
1407 
1408 			dev_dbg(gpu->dev, "event %u\n", event);
1409 
1410 			if (gpu->event[event].sync_point) {
1411 				gpu->sync_point_event = event;
1412 				queue_work(gpu->wq, &gpu->sync_point_work);
1413 			}
1414 
1415 			fence = gpu->event[event].fence;
1416 			if (!fence)
1417 				continue;
1418 
1419 			gpu->event[event].fence = NULL;
1420 
1421 			/*
1422 			 * Events can be processed out of order.  Eg,
1423 			 * - allocate and queue event 0
1424 			 * - allocate event 1
1425 			 * - event 0 completes, we process it
1426 			 * - allocate and queue event 0
1427 			 * - event 1 and event 0 complete
1428 			 * we can end up processing event 0 first, then 1.
1429 			 */
1430 			if (fence_after(fence->seqno, gpu->completed_fence))
1431 				gpu->completed_fence = fence->seqno;
1432 			dma_fence_signal(fence);
1433 
1434 			event_free(gpu, event);
1435 		}
1436 
1437 		ret = IRQ_HANDLED;
1438 	}
1439 
1440 	return ret;
1441 }
1442 
1443 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1444 {
1445 	int ret;
1446 
1447 	if (gpu->clk_reg) {
1448 		ret = clk_prepare_enable(gpu->clk_reg);
1449 		if (ret)
1450 			return ret;
1451 	}
1452 
1453 	if (gpu->clk_bus) {
1454 		ret = clk_prepare_enable(gpu->clk_bus);
1455 		if (ret)
1456 			return ret;
1457 	}
1458 
1459 	if (gpu->clk_core) {
1460 		ret = clk_prepare_enable(gpu->clk_core);
1461 		if (ret)
1462 			goto disable_clk_bus;
1463 	}
1464 
1465 	if (gpu->clk_shader) {
1466 		ret = clk_prepare_enable(gpu->clk_shader);
1467 		if (ret)
1468 			goto disable_clk_core;
1469 	}
1470 
1471 	return 0;
1472 
1473 disable_clk_core:
1474 	if (gpu->clk_core)
1475 		clk_disable_unprepare(gpu->clk_core);
1476 disable_clk_bus:
1477 	if (gpu->clk_bus)
1478 		clk_disable_unprepare(gpu->clk_bus);
1479 
1480 	return ret;
1481 }
1482 
1483 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1484 {
1485 	if (gpu->clk_shader)
1486 		clk_disable_unprepare(gpu->clk_shader);
1487 	if (gpu->clk_core)
1488 		clk_disable_unprepare(gpu->clk_core);
1489 	if (gpu->clk_bus)
1490 		clk_disable_unprepare(gpu->clk_bus);
1491 	if (gpu->clk_reg)
1492 		clk_disable_unprepare(gpu->clk_reg);
1493 
1494 	return 0;
1495 }
1496 
1497 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1498 {
1499 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1500 
1501 	do {
1502 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1503 
1504 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1505 			return 0;
1506 
1507 		if (time_is_before_jiffies(timeout)) {
1508 			dev_warn(gpu->dev,
1509 				 "timed out waiting for idle: idle=0x%x\n",
1510 				 idle);
1511 			return -ETIMEDOUT;
1512 		}
1513 
1514 		udelay(5);
1515 	} while (1);
1516 }
1517 
1518 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1519 {
1520 	if (gpu->buffer.suballoc) {
1521 		/* Replace the last WAIT with END */
1522 		mutex_lock(&gpu->lock);
1523 		etnaviv_buffer_end(gpu);
1524 		mutex_unlock(&gpu->lock);
1525 
1526 		/*
1527 		 * We know that only the FE is busy here, this should
1528 		 * happen quickly (as the WAIT is only 200 cycles).  If
1529 		 * we fail, just warn and continue.
1530 		 */
1531 		etnaviv_gpu_wait_idle(gpu, 100);
1532 	}
1533 
1534 	return etnaviv_gpu_clk_disable(gpu);
1535 }
1536 
1537 #ifdef CONFIG_PM
1538 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1539 {
1540 	int ret;
1541 
1542 	ret = mutex_lock_killable(&gpu->lock);
1543 	if (ret)
1544 		return ret;
1545 
1546 	etnaviv_gpu_update_clock(gpu);
1547 	etnaviv_gpu_hw_init(gpu);
1548 
1549 	gpu->exec_state = -1;
1550 
1551 	mutex_unlock(&gpu->lock);
1552 
1553 	return 0;
1554 }
1555 #endif
1556 
1557 static int
1558 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1559 				  unsigned long *state)
1560 {
1561 	*state = 6;
1562 
1563 	return 0;
1564 }
1565 
1566 static int
1567 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1568 				  unsigned long *state)
1569 {
1570 	struct etnaviv_gpu *gpu = cdev->devdata;
1571 
1572 	*state = gpu->freq_scale;
1573 
1574 	return 0;
1575 }
1576 
1577 static int
1578 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1579 				  unsigned long state)
1580 {
1581 	struct etnaviv_gpu *gpu = cdev->devdata;
1582 
1583 	mutex_lock(&gpu->lock);
1584 	gpu->freq_scale = state;
1585 	if (!pm_runtime_suspended(gpu->dev))
1586 		etnaviv_gpu_update_clock(gpu);
1587 	mutex_unlock(&gpu->lock);
1588 
1589 	return 0;
1590 }
1591 
1592 static struct thermal_cooling_device_ops cooling_ops = {
1593 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1594 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1595 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1596 };
1597 
1598 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1599 	void *data)
1600 {
1601 	struct drm_device *drm = data;
1602 	struct etnaviv_drm_private *priv = drm->dev_private;
1603 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1604 	int ret;
1605 
1606 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1607 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1608 				(char *)dev_name(dev), gpu, &cooling_ops);
1609 		if (IS_ERR(gpu->cooling))
1610 			return PTR_ERR(gpu->cooling);
1611 	}
1612 
1613 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1614 	if (!gpu->wq) {
1615 		ret = -ENOMEM;
1616 		goto out_thermal;
1617 	}
1618 
1619 	ret = etnaviv_sched_init(gpu);
1620 	if (ret)
1621 		goto out_workqueue;
1622 
1623 #ifdef CONFIG_PM
1624 	ret = pm_runtime_get_sync(gpu->dev);
1625 #else
1626 	ret = etnaviv_gpu_clk_enable(gpu);
1627 #endif
1628 	if (ret < 0)
1629 		goto out_sched;
1630 
1631 
1632 	gpu->drm = drm;
1633 	gpu->fence_context = dma_fence_context_alloc(1);
1634 	idr_init(&gpu->fence_idr);
1635 	spin_lock_init(&gpu->fence_spinlock);
1636 
1637 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1638 	init_waitqueue_head(&gpu->fence_event);
1639 
1640 	priv->gpu[priv->num_gpus++] = gpu;
1641 
1642 	pm_runtime_mark_last_busy(gpu->dev);
1643 	pm_runtime_put_autosuspend(gpu->dev);
1644 
1645 	return 0;
1646 
1647 out_sched:
1648 	etnaviv_sched_fini(gpu);
1649 
1650 out_workqueue:
1651 	destroy_workqueue(gpu->wq);
1652 
1653 out_thermal:
1654 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1655 		thermal_cooling_device_unregister(gpu->cooling);
1656 
1657 	return ret;
1658 }
1659 
1660 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1661 	void *data)
1662 {
1663 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1664 
1665 	DBG("%s", dev_name(gpu->dev));
1666 
1667 	flush_workqueue(gpu->wq);
1668 	destroy_workqueue(gpu->wq);
1669 
1670 	etnaviv_sched_fini(gpu);
1671 
1672 #ifdef CONFIG_PM
1673 	pm_runtime_get_sync(gpu->dev);
1674 	pm_runtime_put_sync_suspend(gpu->dev);
1675 #else
1676 	etnaviv_gpu_hw_suspend(gpu);
1677 #endif
1678 
1679 	if (gpu->buffer.suballoc)
1680 		etnaviv_cmdbuf_free(&gpu->buffer);
1681 
1682 	if (gpu->cmdbuf_suballoc) {
1683 		etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc);
1684 		gpu->cmdbuf_suballoc = NULL;
1685 	}
1686 
1687 	if (gpu->mmu) {
1688 		etnaviv_iommu_destroy(gpu->mmu);
1689 		gpu->mmu = NULL;
1690 	}
1691 
1692 	gpu->drm = NULL;
1693 	idr_destroy(&gpu->fence_idr);
1694 
1695 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1696 		thermal_cooling_device_unregister(gpu->cooling);
1697 	gpu->cooling = NULL;
1698 }
1699 
1700 static const struct component_ops gpu_ops = {
1701 	.bind = etnaviv_gpu_bind,
1702 	.unbind = etnaviv_gpu_unbind,
1703 };
1704 
1705 static const struct of_device_id etnaviv_gpu_match[] = {
1706 	{
1707 		.compatible = "vivante,gc"
1708 	},
1709 	{ /* sentinel */ }
1710 };
1711 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1712 
1713 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1714 {
1715 	struct device *dev = &pdev->dev;
1716 	struct etnaviv_gpu *gpu;
1717 	struct resource *res;
1718 	int err;
1719 
1720 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1721 	if (!gpu)
1722 		return -ENOMEM;
1723 
1724 	gpu->dev = &pdev->dev;
1725 	mutex_init(&gpu->lock);
1726 	mutex_init(&gpu->fence_lock);
1727 
1728 	/* Map registers: */
1729 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1730 	gpu->mmio = devm_ioremap_resource(&pdev->dev, res);
1731 	if (IS_ERR(gpu->mmio))
1732 		return PTR_ERR(gpu->mmio);
1733 
1734 	/* Get Interrupt: */
1735 	gpu->irq = platform_get_irq(pdev, 0);
1736 	if (gpu->irq < 0) {
1737 		dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1738 		return gpu->irq;
1739 	}
1740 
1741 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1742 			       dev_name(gpu->dev), gpu);
1743 	if (err) {
1744 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1745 		return err;
1746 	}
1747 
1748 	/* Get Clocks: */
1749 	gpu->clk_reg = devm_clk_get(&pdev->dev, "reg");
1750 	DBG("clk_reg: %p", gpu->clk_reg);
1751 	if (IS_ERR(gpu->clk_reg))
1752 		gpu->clk_reg = NULL;
1753 
1754 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1755 	DBG("clk_bus: %p", gpu->clk_bus);
1756 	if (IS_ERR(gpu->clk_bus))
1757 		gpu->clk_bus = NULL;
1758 
1759 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1760 	DBG("clk_core: %p", gpu->clk_core);
1761 	if (IS_ERR(gpu->clk_core))
1762 		gpu->clk_core = NULL;
1763 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1764 
1765 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1766 	DBG("clk_shader: %p", gpu->clk_shader);
1767 	if (IS_ERR(gpu->clk_shader))
1768 		gpu->clk_shader = NULL;
1769 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1770 
1771 	/* TODO: figure out max mapped size */
1772 	dev_set_drvdata(dev, gpu);
1773 
1774 	/*
1775 	 * We treat the device as initially suspended.  The runtime PM
1776 	 * autosuspend delay is rather arbitary: no measurements have
1777 	 * yet been performed to determine an appropriate value.
1778 	 */
1779 	pm_runtime_use_autosuspend(gpu->dev);
1780 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1781 	pm_runtime_enable(gpu->dev);
1782 
1783 	err = component_add(&pdev->dev, &gpu_ops);
1784 	if (err < 0) {
1785 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1786 		return err;
1787 	}
1788 
1789 	return 0;
1790 }
1791 
1792 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1793 {
1794 	component_del(&pdev->dev, &gpu_ops);
1795 	pm_runtime_disable(&pdev->dev);
1796 	return 0;
1797 }
1798 
1799 #ifdef CONFIG_PM
1800 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1801 {
1802 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1803 	u32 idle, mask;
1804 
1805 	/* If there are any jobs in the HW queue, we're not idle */
1806 	if (atomic_read(&gpu->sched.hw_rq_count))
1807 		return -EBUSY;
1808 
1809 	/* Check whether the hardware (except FE) is idle */
1810 	mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1811 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1812 	if (idle != mask)
1813 		return -EBUSY;
1814 
1815 	return etnaviv_gpu_hw_suspend(gpu);
1816 }
1817 
1818 static int etnaviv_gpu_rpm_resume(struct device *dev)
1819 {
1820 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1821 	int ret;
1822 
1823 	ret = etnaviv_gpu_clk_enable(gpu);
1824 	if (ret)
1825 		return ret;
1826 
1827 	/* Re-initialise the basic hardware state */
1828 	if (gpu->drm && gpu->buffer.suballoc) {
1829 		ret = etnaviv_gpu_hw_resume(gpu);
1830 		if (ret) {
1831 			etnaviv_gpu_clk_disable(gpu);
1832 			return ret;
1833 		}
1834 	}
1835 
1836 	return 0;
1837 }
1838 #endif
1839 
1840 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1841 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1842 			   NULL)
1843 };
1844 
1845 struct platform_driver etnaviv_gpu_driver = {
1846 	.driver = {
1847 		.name = "etnaviv-gpu",
1848 		.owner = THIS_MODULE,
1849 		.pm = &etnaviv_gpu_pm_ops,
1850 		.of_match_table = etnaviv_gpu_match,
1851 	},
1852 	.probe = etnaviv_gpu_platform_probe,
1853 	.remove = etnaviv_gpu_platform_remove,
1854 	.id_table = gpu_ids,
1855 };
1856