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/module.h>
12 #include <linux/of_device.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 #ifndef PHYS_OFFSET
31 #define PHYS_OFFSET 0
32 #endif
33 
34 static const struct platform_device_id gpu_ids[] = {
35 	{ .name = "etnaviv-gpu,2d" },
36 	{ },
37 };
38 
39 /*
40  * Driver functions:
41  */
42 
43 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
44 {
45 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
46 
47 	switch (param) {
48 	case ETNAVIV_PARAM_GPU_MODEL:
49 		*value = gpu->identity.model;
50 		break;
51 
52 	case ETNAVIV_PARAM_GPU_REVISION:
53 		*value = gpu->identity.revision;
54 		break;
55 
56 	case ETNAVIV_PARAM_GPU_FEATURES_0:
57 		*value = gpu->identity.features;
58 		break;
59 
60 	case ETNAVIV_PARAM_GPU_FEATURES_1:
61 		*value = gpu->identity.minor_features0;
62 		break;
63 
64 	case ETNAVIV_PARAM_GPU_FEATURES_2:
65 		*value = gpu->identity.minor_features1;
66 		break;
67 
68 	case ETNAVIV_PARAM_GPU_FEATURES_3:
69 		*value = gpu->identity.minor_features2;
70 		break;
71 
72 	case ETNAVIV_PARAM_GPU_FEATURES_4:
73 		*value = gpu->identity.minor_features3;
74 		break;
75 
76 	case ETNAVIV_PARAM_GPU_FEATURES_5:
77 		*value = gpu->identity.minor_features4;
78 		break;
79 
80 	case ETNAVIV_PARAM_GPU_FEATURES_6:
81 		*value = gpu->identity.minor_features5;
82 		break;
83 
84 	case ETNAVIV_PARAM_GPU_FEATURES_7:
85 		*value = gpu->identity.minor_features6;
86 		break;
87 
88 	case ETNAVIV_PARAM_GPU_FEATURES_8:
89 		*value = gpu->identity.minor_features7;
90 		break;
91 
92 	case ETNAVIV_PARAM_GPU_FEATURES_9:
93 		*value = gpu->identity.minor_features8;
94 		break;
95 
96 	case ETNAVIV_PARAM_GPU_FEATURES_10:
97 		*value = gpu->identity.minor_features9;
98 		break;
99 
100 	case ETNAVIV_PARAM_GPU_FEATURES_11:
101 		*value = gpu->identity.minor_features10;
102 		break;
103 
104 	case ETNAVIV_PARAM_GPU_FEATURES_12:
105 		*value = gpu->identity.minor_features11;
106 		break;
107 
108 	case ETNAVIV_PARAM_GPU_STREAM_COUNT:
109 		*value = gpu->identity.stream_count;
110 		break;
111 
112 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
113 		*value = gpu->identity.register_max;
114 		break;
115 
116 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
117 		*value = gpu->identity.thread_count;
118 		break;
119 
120 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
121 		*value = gpu->identity.vertex_cache_size;
122 		break;
123 
124 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
125 		*value = gpu->identity.shader_core_count;
126 		break;
127 
128 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
129 		*value = gpu->identity.pixel_pipes;
130 		break;
131 
132 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
133 		*value = gpu->identity.vertex_output_buffer_size;
134 		break;
135 
136 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
137 		*value = gpu->identity.buffer_size;
138 		break;
139 
140 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
141 		*value = gpu->identity.instruction_count;
142 		break;
143 
144 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
145 		*value = gpu->identity.num_constants;
146 		break;
147 
148 	case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
149 		*value = gpu->identity.varyings_count;
150 		break;
151 
152 	case ETNAVIV_PARAM_SOFTPIN_START_ADDR:
153 		if (priv->mmu_global->version == ETNAVIV_IOMMU_V2)
154 			*value = ETNAVIV_SOFTPIN_START_ADDRESS;
155 		else
156 			*value = ~0ULL;
157 		break;
158 
159 	default:
160 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
161 		return -EINVAL;
162 	}
163 
164 	return 0;
165 }
166 
167 
168 #define etnaviv_is_model_rev(gpu, mod, rev) \
169 	((gpu)->identity.model == chipModel_##mod && \
170 	 (gpu)->identity.revision == rev)
171 #define etnaviv_field(val, field) \
172 	(((val) & field##__MASK) >> field##__SHIFT)
173 
174 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
175 {
176 	if (gpu->identity.minor_features0 &
177 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
178 		u32 specs[4];
179 		unsigned int streams;
180 
181 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
182 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
183 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
184 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
185 
186 		gpu->identity.stream_count = etnaviv_field(specs[0],
187 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
188 		gpu->identity.register_max = etnaviv_field(specs[0],
189 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
190 		gpu->identity.thread_count = etnaviv_field(specs[0],
191 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
192 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
193 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
194 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
195 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
196 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
197 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
198 		gpu->identity.vertex_output_buffer_size =
199 			etnaviv_field(specs[0],
200 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
201 
202 		gpu->identity.buffer_size = etnaviv_field(specs[1],
203 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
204 		gpu->identity.instruction_count = etnaviv_field(specs[1],
205 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
206 		gpu->identity.num_constants = etnaviv_field(specs[1],
207 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
208 
209 		gpu->identity.varyings_count = etnaviv_field(specs[2],
210 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
211 
212 		/* This overrides the value from older register if non-zero */
213 		streams = etnaviv_field(specs[3],
214 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
215 		if (streams)
216 			gpu->identity.stream_count = streams;
217 	}
218 
219 	/* Fill in the stream count if not specified */
220 	if (gpu->identity.stream_count == 0) {
221 		if (gpu->identity.model >= 0x1000)
222 			gpu->identity.stream_count = 4;
223 		else
224 			gpu->identity.stream_count = 1;
225 	}
226 
227 	/* Convert the register max value */
228 	if (gpu->identity.register_max)
229 		gpu->identity.register_max = 1 << gpu->identity.register_max;
230 	else if (gpu->identity.model == chipModel_GC400)
231 		gpu->identity.register_max = 32;
232 	else
233 		gpu->identity.register_max = 64;
234 
235 	/* Convert thread count */
236 	if (gpu->identity.thread_count)
237 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
238 	else if (gpu->identity.model == chipModel_GC400)
239 		gpu->identity.thread_count = 64;
240 	else if (gpu->identity.model == chipModel_GC500 ||
241 		 gpu->identity.model == chipModel_GC530)
242 		gpu->identity.thread_count = 128;
243 	else
244 		gpu->identity.thread_count = 256;
245 
246 	if (gpu->identity.vertex_cache_size == 0)
247 		gpu->identity.vertex_cache_size = 8;
248 
249 	if (gpu->identity.shader_core_count == 0) {
250 		if (gpu->identity.model >= 0x1000)
251 			gpu->identity.shader_core_count = 2;
252 		else
253 			gpu->identity.shader_core_count = 1;
254 	}
255 
256 	if (gpu->identity.pixel_pipes == 0)
257 		gpu->identity.pixel_pipes = 1;
258 
259 	/* Convert virtex buffer size */
260 	if (gpu->identity.vertex_output_buffer_size) {
261 		gpu->identity.vertex_output_buffer_size =
262 			1 << gpu->identity.vertex_output_buffer_size;
263 	} else if (gpu->identity.model == chipModel_GC400) {
264 		if (gpu->identity.revision < 0x4000)
265 			gpu->identity.vertex_output_buffer_size = 512;
266 		else if (gpu->identity.revision < 0x4200)
267 			gpu->identity.vertex_output_buffer_size = 256;
268 		else
269 			gpu->identity.vertex_output_buffer_size = 128;
270 	} else {
271 		gpu->identity.vertex_output_buffer_size = 512;
272 	}
273 
274 	switch (gpu->identity.instruction_count) {
275 	case 0:
276 		if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
277 		    gpu->identity.model == chipModel_GC880)
278 			gpu->identity.instruction_count = 512;
279 		else
280 			gpu->identity.instruction_count = 256;
281 		break;
282 
283 	case 1:
284 		gpu->identity.instruction_count = 1024;
285 		break;
286 
287 	case 2:
288 		gpu->identity.instruction_count = 2048;
289 		break;
290 
291 	default:
292 		gpu->identity.instruction_count = 256;
293 		break;
294 	}
295 
296 	if (gpu->identity.num_constants == 0)
297 		gpu->identity.num_constants = 168;
298 
299 	if (gpu->identity.varyings_count == 0) {
300 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
301 			gpu->identity.varyings_count = 12;
302 		else
303 			gpu->identity.varyings_count = 8;
304 	}
305 
306 	/*
307 	 * For some cores, two varyings are consumed for position, so the
308 	 * maximum varying count needs to be reduced by one.
309 	 */
310 	if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
311 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
312 	    etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
313 	    etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
314 	    etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
315 	    etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
316 	    etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
317 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
318 	    etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
319 	    etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
320 	    etnaviv_is_model_rev(gpu, GC880, 0x5106))
321 		gpu->identity.varyings_count -= 1;
322 }
323 
324 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
325 {
326 	u32 chipIdentity;
327 
328 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
329 
330 	/* Special case for older graphic cores. */
331 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
332 		gpu->identity.model    = chipModel_GC500;
333 		gpu->identity.revision = etnaviv_field(chipIdentity,
334 					 VIVS_HI_CHIP_IDENTITY_REVISION);
335 	} else {
336 		u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
337 
338 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
339 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
340 		gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
341 		gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
342 		gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
343 
344 		/*
345 		 * !!!! HACK ALERT !!!!
346 		 * Because people change device IDs without letting software
347 		 * know about it - here is the hack to make it all look the
348 		 * same.  Only for GC400 family.
349 		 */
350 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
351 		    gpu->identity.model != chipModel_GC420) {
352 			gpu->identity.model = gpu->identity.model & 0x0400;
353 		}
354 
355 		/* Another special case */
356 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
357 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
358 
359 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
360 				/*
361 				 * This IP has an ECO; put the correct
362 				 * revision in it.
363 				 */
364 				gpu->identity.revision = 0x1051;
365 			}
366 		}
367 
368 		/*
369 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
370 		 * reality it's just a re-branded GC3000. We can identify this
371 		 * core by the upper half of the revision register being all 1.
372 		 * Fix model/rev here, so all other places can refer to this
373 		 * core by its real identity.
374 		 */
375 		if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
376 			gpu->identity.model = chipModel_GC3000;
377 			gpu->identity.revision &= 0xffff;
378 		}
379 
380 		if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
381 			gpu->identity.eco_id = 1;
382 
383 		if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
384 			gpu->identity.eco_id = 1;
385 	}
386 
387 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
388 		 gpu->identity.model, gpu->identity.revision);
389 
390 	gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
391 	/*
392 	 * If there is a match in the HWDB, we aren't interested in the
393 	 * remaining register values, as they might be wrong.
394 	 */
395 	if (etnaviv_fill_identity_from_hwdb(gpu))
396 		return;
397 
398 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
399 
400 	/* Disable fast clear on GC700. */
401 	if (gpu->identity.model == chipModel_GC700)
402 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
403 
404 	if ((gpu->identity.model == chipModel_GC500 &&
405 	     gpu->identity.revision < 2) ||
406 	    (gpu->identity.model == chipModel_GC300 &&
407 	     gpu->identity.revision < 0x2000)) {
408 
409 		/*
410 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
411 		 * registers.
412 		 */
413 		gpu->identity.minor_features0 = 0;
414 		gpu->identity.minor_features1 = 0;
415 		gpu->identity.minor_features2 = 0;
416 		gpu->identity.minor_features3 = 0;
417 		gpu->identity.minor_features4 = 0;
418 		gpu->identity.minor_features5 = 0;
419 	} else
420 		gpu->identity.minor_features0 =
421 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
422 
423 	if (gpu->identity.minor_features0 &
424 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
425 		gpu->identity.minor_features1 =
426 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
427 		gpu->identity.minor_features2 =
428 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
429 		gpu->identity.minor_features3 =
430 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
431 		gpu->identity.minor_features4 =
432 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
433 		gpu->identity.minor_features5 =
434 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
435 	}
436 
437 	/* GC600 idle register reports zero bits where modules aren't present */
438 	if (gpu->identity.model == chipModel_GC600)
439 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
440 				 VIVS_HI_IDLE_STATE_RA |
441 				 VIVS_HI_IDLE_STATE_SE |
442 				 VIVS_HI_IDLE_STATE_PA |
443 				 VIVS_HI_IDLE_STATE_SH |
444 				 VIVS_HI_IDLE_STATE_PE |
445 				 VIVS_HI_IDLE_STATE_DE |
446 				 VIVS_HI_IDLE_STATE_FE;
447 
448 	etnaviv_hw_specs(gpu);
449 }
450 
451 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
452 {
453 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
454 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
455 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
456 }
457 
458 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
459 {
460 	if (gpu->identity.minor_features2 &
461 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
462 		clk_set_rate(gpu->clk_core,
463 			     gpu->base_rate_core >> gpu->freq_scale);
464 		clk_set_rate(gpu->clk_shader,
465 			     gpu->base_rate_shader >> gpu->freq_scale);
466 	} else {
467 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
468 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
469 
470 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
471 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
472 		etnaviv_gpu_load_clock(gpu, clock);
473 	}
474 }
475 
476 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
477 {
478 	u32 control, idle;
479 	unsigned long timeout;
480 	bool failed = true;
481 
482 	/* We hope that the GPU resets in under one second */
483 	timeout = jiffies + msecs_to_jiffies(1000);
484 
485 	while (time_is_after_jiffies(timeout)) {
486 		/* enable clock */
487 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
488 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
489 		etnaviv_gpu_load_clock(gpu, control);
490 
491 		/* isolate the GPU. */
492 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
493 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
494 
495 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
496 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
497 			          VIVS_MMUv2_AHB_CONTROL_RESET);
498 		} else {
499 			/* set soft reset. */
500 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
501 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
502 		}
503 
504 		/* wait for reset. */
505 		usleep_range(10, 20);
506 
507 		/* reset soft reset bit. */
508 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
509 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
510 
511 		/* reset GPU isolation. */
512 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
513 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
514 
515 		/* read idle register. */
516 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
517 
518 		/* try resetting again if FE is not idle */
519 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
520 			dev_dbg(gpu->dev, "FE is not idle\n");
521 			continue;
522 		}
523 
524 		/* read reset register. */
525 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
526 
527 		/* is the GPU idle? */
528 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
529 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
530 			dev_dbg(gpu->dev, "GPU is not idle\n");
531 			continue;
532 		}
533 
534 		/* disable debug registers, as they are not normally needed */
535 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
536 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
537 
538 		failed = false;
539 		break;
540 	}
541 
542 	if (failed) {
543 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
544 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
545 
546 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
547 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
548 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
549 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
550 
551 		return -EBUSY;
552 	}
553 
554 	/* We rely on the GPU running, so program the clock */
555 	etnaviv_gpu_update_clock(gpu);
556 
557 	return 0;
558 }
559 
560 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
561 {
562 	u32 pmc, ppc;
563 
564 	/* enable clock gating */
565 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
566 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
567 
568 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
569 	if (gpu->identity.revision == 0x4301 ||
570 	    gpu->identity.revision == 0x4302)
571 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
572 
573 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
574 
575 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
576 
577 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
578 	if (gpu->identity.model >= chipModel_GC400 &&
579 	    gpu->identity.model != chipModel_GC420 &&
580 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
581 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
582 
583 	/*
584 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
585 	 * present without a bug fix.
586 	 */
587 	if (gpu->identity.revision < 0x5000 &&
588 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
589 	    !(gpu->identity.minor_features1 &
590 	      chipMinorFeatures1_DISABLE_PE_GATING))
591 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
592 
593 	if (gpu->identity.revision < 0x5422)
594 		pmc |= BIT(15); /* Unknown bit */
595 
596 	/* Disable TX clock gating on affected core revisions. */
597 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
598 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108))
599 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
600 
601 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
602 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
603 
604 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
605 }
606 
607 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
608 {
609 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
610 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
611 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
612 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
613 
614 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
615 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
616 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
617 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
618 	}
619 }
620 
621 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu)
622 {
623 	u32 address = etnaviv_cmdbuf_get_va(&gpu->buffer,
624 				&gpu->mmu_context->cmdbuf_mapping);
625 	u16 prefetch;
626 
627 	/* setup the MMU */
628 	etnaviv_iommu_restore(gpu, gpu->mmu_context);
629 
630 	/* Start command processor */
631 	prefetch = etnaviv_buffer_init(gpu);
632 
633 	etnaviv_gpu_start_fe(gpu, address, prefetch);
634 }
635 
636 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
637 {
638 	/*
639 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
640 	 * cannot be read, extracted from vivante kernel driver.
641 	 */
642 	u32 pulse_eater = 0x01590880;
643 
644 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
645 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
646 		pulse_eater |= BIT(23);
647 
648 	}
649 
650 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
651 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
652 		pulse_eater &= ~BIT(16);
653 		pulse_eater |= BIT(17);
654 	}
655 
656 	if ((gpu->identity.revision > 0x5420) &&
657 	    (gpu->identity.features & chipFeatures_PIPE_3D))
658 	{
659 		/* Performance fix: disable internal DFS */
660 		pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
661 		pulse_eater |= BIT(18);
662 	}
663 
664 	gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
665 }
666 
667 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
668 {
669 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
670 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
671 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
672 		u32 mc_memory_debug;
673 
674 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
675 
676 		if (gpu->identity.revision == 0x5007)
677 			mc_memory_debug |= 0x0c;
678 		else
679 			mc_memory_debug |= 0x08;
680 
681 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
682 	}
683 
684 	/* enable module-level clock gating */
685 	etnaviv_gpu_enable_mlcg(gpu);
686 
687 	/*
688 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
689 	 * This is necessary to prevent the iMX6 SoC locking up.
690 	 */
691 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
692 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
693 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
694 
695 	/* GC2000 rev 5108 needs a special bus config */
696 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
697 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
698 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
699 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
700 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
701 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
702 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
703 	}
704 
705 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
706 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
707 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
708 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
709 	}
710 
711 	/* setup the pulse eater */
712 	etnaviv_gpu_setup_pulse_eater(gpu);
713 
714 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
715 }
716 
717 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
718 {
719 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
720 	int ret, i;
721 
722 	ret = pm_runtime_get_sync(gpu->dev);
723 	if (ret < 0) {
724 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
725 		return ret;
726 	}
727 
728 	etnaviv_hw_identify(gpu);
729 
730 	if (gpu->identity.model == 0) {
731 		dev_err(gpu->dev, "Unknown GPU model\n");
732 		ret = -ENXIO;
733 		goto fail;
734 	}
735 
736 	/* Exclude VG cores with FE2.0 */
737 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
738 	    gpu->identity.features & chipFeatures_FE20) {
739 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
740 		ret = -ENXIO;
741 		goto fail;
742 	}
743 
744 	/*
745 	 * On cores with security features supported, we claim control over the
746 	 * security states.
747 	 */
748 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
749 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
750 		gpu->sec_mode = ETNA_SEC_KERNEL;
751 
752 	ret = etnaviv_hw_reset(gpu);
753 	if (ret) {
754 		dev_err(gpu->dev, "GPU reset failed\n");
755 		goto fail;
756 	}
757 
758 	ret = etnaviv_iommu_global_init(gpu);
759 	if (ret)
760 		goto fail;
761 
762 	/*
763 	 * Set the GPU linear window to be at the end of the DMA window, where
764 	 * the CMA area is likely to reside. This ensures that we are able to
765 	 * map the command buffers while having the linear window overlap as
766 	 * much RAM as possible, so we can optimize mappings for other buffers.
767 	 *
768 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
769 	 * to different views of the memory on the individual engines.
770 	 */
771 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
772 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
773 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
774 		if (dma_mask < PHYS_OFFSET + SZ_2G)
775 			priv->mmu_global->memory_base = PHYS_OFFSET;
776 		else
777 			priv->mmu_global->memory_base = dma_mask - SZ_2G + 1;
778 	} else if (PHYS_OFFSET >= SZ_2G) {
779 		dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
780 		priv->mmu_global->memory_base = PHYS_OFFSET;
781 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
782 	}
783 
784 	/*
785 	 * If the GPU is part of a system with DMA addressing limitations,
786 	 * request pages for our SHM backend buffers from the DMA32 zone to
787 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
788 	 */
789 	if (dma_addressing_limited(gpu->dev))
790 		priv->shm_gfp_mask |= GFP_DMA32;
791 
792 	/* Create buffer: */
793 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
794 				  PAGE_SIZE);
795 	if (ret) {
796 		dev_err(gpu->dev, "could not create command buffer\n");
797 		goto fail;
798 	}
799 
800 	/* Setup event management */
801 	spin_lock_init(&gpu->event_spinlock);
802 	init_completion(&gpu->event_free);
803 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
804 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
805 		complete(&gpu->event_free);
806 
807 	/* Now program the hardware */
808 	mutex_lock(&gpu->lock);
809 	etnaviv_gpu_hw_init(gpu);
810 	gpu->exec_state = -1;
811 	mutex_unlock(&gpu->lock);
812 
813 	pm_runtime_mark_last_busy(gpu->dev);
814 	pm_runtime_put_autosuspend(gpu->dev);
815 
816 	gpu->initialized = true;
817 
818 	return 0;
819 
820 fail:
821 	pm_runtime_mark_last_busy(gpu->dev);
822 	pm_runtime_put_autosuspend(gpu->dev);
823 
824 	return ret;
825 }
826 
827 #ifdef CONFIG_DEBUG_FS
828 struct dma_debug {
829 	u32 address[2];
830 	u32 state[2];
831 };
832 
833 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
834 {
835 	u32 i;
836 
837 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
838 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
839 
840 	for (i = 0; i < 500; i++) {
841 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
842 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
843 
844 		if (debug->address[0] != debug->address[1])
845 			break;
846 
847 		if (debug->state[0] != debug->state[1])
848 			break;
849 	}
850 }
851 
852 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
853 {
854 	struct dma_debug debug;
855 	u32 dma_lo, dma_hi, axi, idle;
856 	int ret;
857 
858 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
859 
860 	ret = pm_runtime_get_sync(gpu->dev);
861 	if (ret < 0)
862 		return ret;
863 
864 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
865 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
866 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
867 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
868 
869 	verify_dma(gpu, &debug);
870 
871 	seq_puts(m, "\tidentity\n");
872 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
873 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
874 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
875 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
876 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
877 
878 	seq_puts(m, "\tfeatures\n");
879 	seq_printf(m, "\t major_features: 0x%08x\n",
880 		   gpu->identity.features);
881 	seq_printf(m, "\t minor_features0: 0x%08x\n",
882 		   gpu->identity.minor_features0);
883 	seq_printf(m, "\t minor_features1: 0x%08x\n",
884 		   gpu->identity.minor_features1);
885 	seq_printf(m, "\t minor_features2: 0x%08x\n",
886 		   gpu->identity.minor_features2);
887 	seq_printf(m, "\t minor_features3: 0x%08x\n",
888 		   gpu->identity.minor_features3);
889 	seq_printf(m, "\t minor_features4: 0x%08x\n",
890 		   gpu->identity.minor_features4);
891 	seq_printf(m, "\t minor_features5: 0x%08x\n",
892 		   gpu->identity.minor_features5);
893 	seq_printf(m, "\t minor_features6: 0x%08x\n",
894 		   gpu->identity.minor_features6);
895 	seq_printf(m, "\t minor_features7: 0x%08x\n",
896 		   gpu->identity.minor_features7);
897 	seq_printf(m, "\t minor_features8: 0x%08x\n",
898 		   gpu->identity.minor_features8);
899 	seq_printf(m, "\t minor_features9: 0x%08x\n",
900 		   gpu->identity.minor_features9);
901 	seq_printf(m, "\t minor_features10: 0x%08x\n",
902 		   gpu->identity.minor_features10);
903 	seq_printf(m, "\t minor_features11: 0x%08x\n",
904 		   gpu->identity.minor_features11);
905 
906 	seq_puts(m, "\tspecs\n");
907 	seq_printf(m, "\t stream_count:  %d\n",
908 			gpu->identity.stream_count);
909 	seq_printf(m, "\t register_max: %d\n",
910 			gpu->identity.register_max);
911 	seq_printf(m, "\t thread_count: %d\n",
912 			gpu->identity.thread_count);
913 	seq_printf(m, "\t vertex_cache_size: %d\n",
914 			gpu->identity.vertex_cache_size);
915 	seq_printf(m, "\t shader_core_count: %d\n",
916 			gpu->identity.shader_core_count);
917 	seq_printf(m, "\t pixel_pipes: %d\n",
918 			gpu->identity.pixel_pipes);
919 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
920 			gpu->identity.vertex_output_buffer_size);
921 	seq_printf(m, "\t buffer_size: %d\n",
922 			gpu->identity.buffer_size);
923 	seq_printf(m, "\t instruction_count: %d\n",
924 			gpu->identity.instruction_count);
925 	seq_printf(m, "\t num_constants: %d\n",
926 			gpu->identity.num_constants);
927 	seq_printf(m, "\t varyings_count: %d\n",
928 			gpu->identity.varyings_count);
929 
930 	seq_printf(m, "\taxi: 0x%08x\n", axi);
931 	seq_printf(m, "\tidle: 0x%08x\n", idle);
932 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
933 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
934 		seq_puts(m, "\t FE is not idle\n");
935 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
936 		seq_puts(m, "\t DE is not idle\n");
937 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
938 		seq_puts(m, "\t PE is not idle\n");
939 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
940 		seq_puts(m, "\t SH is not idle\n");
941 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
942 		seq_puts(m, "\t PA is not idle\n");
943 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
944 		seq_puts(m, "\t SE is not idle\n");
945 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
946 		seq_puts(m, "\t RA is not idle\n");
947 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
948 		seq_puts(m, "\t TX is not idle\n");
949 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
950 		seq_puts(m, "\t VG is not idle\n");
951 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
952 		seq_puts(m, "\t IM is not idle\n");
953 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
954 		seq_puts(m, "\t FP is not idle\n");
955 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
956 		seq_puts(m, "\t TS is not idle\n");
957 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
958 		seq_puts(m, "\t BL is not idle\n");
959 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
960 		seq_puts(m, "\t ASYNCFE is not idle\n");
961 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
962 		seq_puts(m, "\t MC is not idle\n");
963 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
964 		seq_puts(m, "\t PPA is not idle\n");
965 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
966 		seq_puts(m, "\t WD is not idle\n");
967 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
968 		seq_puts(m, "\t NN is not idle\n");
969 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
970 		seq_puts(m, "\t TP is not idle\n");
971 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
972 		seq_puts(m, "\t AXI low power mode\n");
973 
974 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
975 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
976 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
977 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
978 
979 		seq_puts(m, "\tMC\n");
980 		seq_printf(m, "\t read0: 0x%08x\n", read0);
981 		seq_printf(m, "\t read1: 0x%08x\n", read1);
982 		seq_printf(m, "\t write: 0x%08x\n", write);
983 	}
984 
985 	seq_puts(m, "\tDMA ");
986 
987 	if (debug.address[0] == debug.address[1] &&
988 	    debug.state[0] == debug.state[1]) {
989 		seq_puts(m, "seems to be stuck\n");
990 	} else if (debug.address[0] == debug.address[1]) {
991 		seq_puts(m, "address is constant\n");
992 	} else {
993 		seq_puts(m, "is running\n");
994 	}
995 
996 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
997 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
998 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
999 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1000 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1001 		   dma_lo, dma_hi);
1002 
1003 	ret = 0;
1004 
1005 	pm_runtime_mark_last_busy(gpu->dev);
1006 	pm_runtime_put_autosuspend(gpu->dev);
1007 
1008 	return ret;
1009 }
1010 #endif
1011 
1012 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
1013 {
1014 	unsigned int i = 0;
1015 
1016 	dev_err(gpu->dev, "recover hung GPU!\n");
1017 
1018 	if (pm_runtime_get_sync(gpu->dev) < 0)
1019 		return;
1020 
1021 	mutex_lock(&gpu->lock);
1022 
1023 	etnaviv_hw_reset(gpu);
1024 
1025 	/* complete all events, the GPU won't do it after the reset */
1026 	spin_lock(&gpu->event_spinlock);
1027 	for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1028 		complete(&gpu->event_free);
1029 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1030 	spin_unlock(&gpu->event_spinlock);
1031 
1032 	etnaviv_gpu_hw_init(gpu);
1033 	gpu->exec_state = -1;
1034 	gpu->mmu_context = NULL;
1035 
1036 	mutex_unlock(&gpu->lock);
1037 	pm_runtime_mark_last_busy(gpu->dev);
1038 	pm_runtime_put_autosuspend(gpu->dev);
1039 }
1040 
1041 /* fence object management */
1042 struct etnaviv_fence {
1043 	struct etnaviv_gpu *gpu;
1044 	struct dma_fence base;
1045 };
1046 
1047 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1048 {
1049 	return container_of(fence, struct etnaviv_fence, base);
1050 }
1051 
1052 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1053 {
1054 	return "etnaviv";
1055 }
1056 
1057 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1058 {
1059 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1060 
1061 	return dev_name(f->gpu->dev);
1062 }
1063 
1064 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1065 {
1066 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1067 
1068 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1069 }
1070 
1071 static void etnaviv_fence_release(struct dma_fence *fence)
1072 {
1073 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1074 
1075 	kfree_rcu(f, base.rcu);
1076 }
1077 
1078 static const struct dma_fence_ops etnaviv_fence_ops = {
1079 	.get_driver_name = etnaviv_fence_get_driver_name,
1080 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1081 	.signaled = etnaviv_fence_signaled,
1082 	.release = etnaviv_fence_release,
1083 };
1084 
1085 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1086 {
1087 	struct etnaviv_fence *f;
1088 
1089 	/*
1090 	 * GPU lock must already be held, otherwise fence completion order might
1091 	 * not match the seqno order assigned here.
1092 	 */
1093 	lockdep_assert_held(&gpu->lock);
1094 
1095 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1096 	if (!f)
1097 		return NULL;
1098 
1099 	f->gpu = gpu;
1100 
1101 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1102 		       gpu->fence_context, ++gpu->next_fence);
1103 
1104 	return &f->base;
1105 }
1106 
1107 /* returns true if fence a comes after fence b */
1108 static inline bool fence_after(u32 a, u32 b)
1109 {
1110 	return (s32)(a - b) > 0;
1111 }
1112 
1113 /*
1114  * event management:
1115  */
1116 
1117 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1118 	unsigned int *events)
1119 {
1120 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1121 	unsigned i, acquired = 0;
1122 
1123 	for (i = 0; i < nr_events; i++) {
1124 		unsigned long ret;
1125 
1126 		ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1127 
1128 		if (!ret) {
1129 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1130 			goto out;
1131 		}
1132 
1133 		acquired++;
1134 		timeout = ret;
1135 	}
1136 
1137 	spin_lock(&gpu->event_spinlock);
1138 
1139 	for (i = 0; i < nr_events; i++) {
1140 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1141 
1142 		events[i] = event;
1143 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1144 		set_bit(event, gpu->event_bitmap);
1145 	}
1146 
1147 	spin_unlock(&gpu->event_spinlock);
1148 
1149 	return 0;
1150 
1151 out:
1152 	for (i = 0; i < acquired; i++)
1153 		complete(&gpu->event_free);
1154 
1155 	return -EBUSY;
1156 }
1157 
1158 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1159 {
1160 	if (!test_bit(event, gpu->event_bitmap)) {
1161 		dev_warn(gpu->dev, "event %u is already marked as free",
1162 			 event);
1163 	} else {
1164 		clear_bit(event, gpu->event_bitmap);
1165 		complete(&gpu->event_free);
1166 	}
1167 }
1168 
1169 /*
1170  * Cmdstream submission/retirement:
1171  */
1172 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1173 	u32 id, struct drm_etnaviv_timespec *timeout)
1174 {
1175 	struct dma_fence *fence;
1176 	int ret;
1177 
1178 	/*
1179 	 * Look up the fence and take a reference. We might still find a fence
1180 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1181 	 * pretends we didn't find a fence in that case.
1182 	 */
1183 	rcu_read_lock();
1184 	fence = idr_find(&gpu->fence_idr, id);
1185 	if (fence)
1186 		fence = dma_fence_get_rcu(fence);
1187 	rcu_read_unlock();
1188 
1189 	if (!fence)
1190 		return 0;
1191 
1192 	if (!timeout) {
1193 		/* No timeout was requested: just test for completion */
1194 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1195 	} else {
1196 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1197 
1198 		ret = dma_fence_wait_timeout(fence, true, remaining);
1199 		if (ret == 0)
1200 			ret = -ETIMEDOUT;
1201 		else if (ret != -ERESTARTSYS)
1202 			ret = 0;
1203 
1204 	}
1205 
1206 	dma_fence_put(fence);
1207 	return ret;
1208 }
1209 
1210 /*
1211  * Wait for an object to become inactive.  This, on it's own, is not race
1212  * free: the object is moved by the scheduler off the active list, and
1213  * then the iova is put.  Moreover, the object could be re-submitted just
1214  * after we notice that it's become inactive.
1215  *
1216  * Although the retirement happens under the gpu lock, we don't want to hold
1217  * that lock in this function while waiting.
1218  */
1219 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1220 	struct etnaviv_gem_object *etnaviv_obj,
1221 	struct drm_etnaviv_timespec *timeout)
1222 {
1223 	unsigned long remaining;
1224 	long ret;
1225 
1226 	if (!timeout)
1227 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1228 
1229 	remaining = etnaviv_timeout_to_jiffies(timeout);
1230 
1231 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1232 					       !is_active(etnaviv_obj),
1233 					       remaining);
1234 	if (ret > 0)
1235 		return 0;
1236 	else if (ret == -ERESTARTSYS)
1237 		return -ERESTARTSYS;
1238 	else
1239 		return -ETIMEDOUT;
1240 }
1241 
1242 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1243 	struct etnaviv_event *event, unsigned int flags)
1244 {
1245 	const struct etnaviv_gem_submit *submit = event->submit;
1246 	unsigned int i;
1247 
1248 	for (i = 0; i < submit->nr_pmrs; i++) {
1249 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1250 
1251 		if (pmr->flags == flags)
1252 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1253 	}
1254 }
1255 
1256 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1257 	struct etnaviv_event *event)
1258 {
1259 	u32 val;
1260 
1261 	/* disable clock gating */
1262 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1263 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1264 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1265 
1266 	/* enable debug register */
1267 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1268 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1269 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1270 
1271 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1272 }
1273 
1274 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1275 	struct etnaviv_event *event)
1276 {
1277 	const struct etnaviv_gem_submit *submit = event->submit;
1278 	unsigned int i;
1279 	u32 val;
1280 
1281 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1282 
1283 	for (i = 0; i < submit->nr_pmrs; i++) {
1284 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1285 
1286 		*pmr->bo_vma = pmr->sequence;
1287 	}
1288 
1289 	/* disable debug register */
1290 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1291 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1292 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1293 
1294 	/* enable clock gating */
1295 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1296 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1297 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1298 }
1299 
1300 
1301 /* add bo's to gpu's ring, and kick gpu: */
1302 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1303 {
1304 	struct etnaviv_gpu *gpu = submit->gpu;
1305 	struct dma_fence *gpu_fence;
1306 	unsigned int i, nr_events = 1, event[3];
1307 	int ret;
1308 
1309 	if (!submit->runtime_resumed) {
1310 		ret = pm_runtime_get_sync(gpu->dev);
1311 		if (ret < 0)
1312 			return NULL;
1313 		submit->runtime_resumed = true;
1314 	}
1315 
1316 	/*
1317 	 * if there are performance monitor requests we need to have
1318 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1319 	 *   requests.
1320 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1321 	 *   and update the sequence number for userspace.
1322 	 */
1323 	if (submit->nr_pmrs)
1324 		nr_events = 3;
1325 
1326 	ret = event_alloc(gpu, nr_events, event);
1327 	if (ret) {
1328 		DRM_ERROR("no free events\n");
1329 		return NULL;
1330 	}
1331 
1332 	mutex_lock(&gpu->lock);
1333 
1334 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1335 	if (!gpu_fence) {
1336 		for (i = 0; i < nr_events; i++)
1337 			event_free(gpu, event[i]);
1338 
1339 		goto out_unlock;
1340 	}
1341 
1342 	if (!gpu->mmu_context) {
1343 		etnaviv_iommu_context_get(submit->mmu_context);
1344 		gpu->mmu_context = submit->mmu_context;
1345 		etnaviv_gpu_start_fe_idleloop(gpu);
1346 	} else {
1347 		etnaviv_iommu_context_get(gpu->mmu_context);
1348 		submit->prev_mmu_context = gpu->mmu_context;
1349 	}
1350 
1351 	if (submit->nr_pmrs) {
1352 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1353 		kref_get(&submit->refcount);
1354 		gpu->event[event[1]].submit = submit;
1355 		etnaviv_sync_point_queue(gpu, event[1]);
1356 	}
1357 
1358 	gpu->event[event[0]].fence = gpu_fence;
1359 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1360 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1361 			     event[0], &submit->cmdbuf);
1362 
1363 	if (submit->nr_pmrs) {
1364 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1365 		kref_get(&submit->refcount);
1366 		gpu->event[event[2]].submit = submit;
1367 		etnaviv_sync_point_queue(gpu, event[2]);
1368 	}
1369 
1370 out_unlock:
1371 	mutex_unlock(&gpu->lock);
1372 
1373 	return gpu_fence;
1374 }
1375 
1376 static void sync_point_worker(struct work_struct *work)
1377 {
1378 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1379 					       sync_point_work);
1380 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1381 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1382 
1383 	event->sync_point(gpu, event);
1384 	etnaviv_submit_put(event->submit);
1385 	event_free(gpu, gpu->sync_point_event);
1386 
1387 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1388 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1389 }
1390 
1391 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1392 {
1393 	u32 status_reg, status;
1394 	int i;
1395 
1396 	if (gpu->sec_mode == ETNA_SEC_NONE)
1397 		status_reg = VIVS_MMUv2_STATUS;
1398 	else
1399 		status_reg = VIVS_MMUv2_SEC_STATUS;
1400 
1401 	status = gpu_read(gpu, status_reg);
1402 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1403 
1404 	for (i = 0; i < 4; i++) {
1405 		u32 address_reg;
1406 
1407 		if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1408 			continue;
1409 
1410 		if (gpu->sec_mode == ETNA_SEC_NONE)
1411 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1412 		else
1413 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1414 
1415 		dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1416 				    gpu_read(gpu, address_reg));
1417 	}
1418 }
1419 
1420 static irqreturn_t irq_handler(int irq, void *data)
1421 {
1422 	struct etnaviv_gpu *gpu = data;
1423 	irqreturn_t ret = IRQ_NONE;
1424 
1425 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1426 
1427 	if (intr != 0) {
1428 		int event;
1429 
1430 		pm_runtime_mark_last_busy(gpu->dev);
1431 
1432 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1433 
1434 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1435 			dev_err(gpu->dev, "AXI bus error\n");
1436 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1437 		}
1438 
1439 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1440 			dump_mmu_fault(gpu);
1441 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1442 		}
1443 
1444 		while ((event = ffs(intr)) != 0) {
1445 			struct dma_fence *fence;
1446 
1447 			event -= 1;
1448 
1449 			intr &= ~(1 << event);
1450 
1451 			dev_dbg(gpu->dev, "event %u\n", event);
1452 
1453 			if (gpu->event[event].sync_point) {
1454 				gpu->sync_point_event = event;
1455 				queue_work(gpu->wq, &gpu->sync_point_work);
1456 			}
1457 
1458 			fence = gpu->event[event].fence;
1459 			if (!fence)
1460 				continue;
1461 
1462 			gpu->event[event].fence = NULL;
1463 
1464 			/*
1465 			 * Events can be processed out of order.  Eg,
1466 			 * - allocate and queue event 0
1467 			 * - allocate event 1
1468 			 * - event 0 completes, we process it
1469 			 * - allocate and queue event 0
1470 			 * - event 1 and event 0 complete
1471 			 * we can end up processing event 0 first, then 1.
1472 			 */
1473 			if (fence_after(fence->seqno, gpu->completed_fence))
1474 				gpu->completed_fence = fence->seqno;
1475 			dma_fence_signal(fence);
1476 
1477 			event_free(gpu, event);
1478 		}
1479 
1480 		ret = IRQ_HANDLED;
1481 	}
1482 
1483 	return ret;
1484 }
1485 
1486 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1487 {
1488 	int ret;
1489 
1490 	if (gpu->clk_reg) {
1491 		ret = clk_prepare_enable(gpu->clk_reg);
1492 		if (ret)
1493 			return ret;
1494 	}
1495 
1496 	if (gpu->clk_bus) {
1497 		ret = clk_prepare_enable(gpu->clk_bus);
1498 		if (ret)
1499 			return ret;
1500 	}
1501 
1502 	if (gpu->clk_core) {
1503 		ret = clk_prepare_enable(gpu->clk_core);
1504 		if (ret)
1505 			goto disable_clk_bus;
1506 	}
1507 
1508 	if (gpu->clk_shader) {
1509 		ret = clk_prepare_enable(gpu->clk_shader);
1510 		if (ret)
1511 			goto disable_clk_core;
1512 	}
1513 
1514 	return 0;
1515 
1516 disable_clk_core:
1517 	if (gpu->clk_core)
1518 		clk_disable_unprepare(gpu->clk_core);
1519 disable_clk_bus:
1520 	if (gpu->clk_bus)
1521 		clk_disable_unprepare(gpu->clk_bus);
1522 
1523 	return ret;
1524 }
1525 
1526 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1527 {
1528 	if (gpu->clk_shader)
1529 		clk_disable_unprepare(gpu->clk_shader);
1530 	if (gpu->clk_core)
1531 		clk_disable_unprepare(gpu->clk_core);
1532 	if (gpu->clk_bus)
1533 		clk_disable_unprepare(gpu->clk_bus);
1534 	if (gpu->clk_reg)
1535 		clk_disable_unprepare(gpu->clk_reg);
1536 
1537 	return 0;
1538 }
1539 
1540 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1541 {
1542 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1543 
1544 	do {
1545 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1546 
1547 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1548 			return 0;
1549 
1550 		if (time_is_before_jiffies(timeout)) {
1551 			dev_warn(gpu->dev,
1552 				 "timed out waiting for idle: idle=0x%x\n",
1553 				 idle);
1554 			return -ETIMEDOUT;
1555 		}
1556 
1557 		udelay(5);
1558 	} while (1);
1559 }
1560 
1561 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1562 {
1563 	if (gpu->initialized && gpu->mmu_context) {
1564 		/* Replace the last WAIT with END */
1565 		mutex_lock(&gpu->lock);
1566 		etnaviv_buffer_end(gpu);
1567 		mutex_unlock(&gpu->lock);
1568 
1569 		/*
1570 		 * We know that only the FE is busy here, this should
1571 		 * happen quickly (as the WAIT is only 200 cycles).  If
1572 		 * we fail, just warn and continue.
1573 		 */
1574 		etnaviv_gpu_wait_idle(gpu, 100);
1575 
1576 		etnaviv_iommu_context_put(gpu->mmu_context);
1577 		gpu->mmu_context = NULL;
1578 	}
1579 
1580 	gpu->exec_state = -1;
1581 
1582 	return etnaviv_gpu_clk_disable(gpu);
1583 }
1584 
1585 #ifdef CONFIG_PM
1586 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1587 {
1588 	int ret;
1589 
1590 	ret = mutex_lock_killable(&gpu->lock);
1591 	if (ret)
1592 		return ret;
1593 
1594 	etnaviv_gpu_update_clock(gpu);
1595 	etnaviv_gpu_hw_init(gpu);
1596 
1597 	mutex_unlock(&gpu->lock);
1598 
1599 	return 0;
1600 }
1601 #endif
1602 
1603 static int
1604 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1605 				  unsigned long *state)
1606 {
1607 	*state = 6;
1608 
1609 	return 0;
1610 }
1611 
1612 static int
1613 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1614 				  unsigned long *state)
1615 {
1616 	struct etnaviv_gpu *gpu = cdev->devdata;
1617 
1618 	*state = gpu->freq_scale;
1619 
1620 	return 0;
1621 }
1622 
1623 static int
1624 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1625 				  unsigned long state)
1626 {
1627 	struct etnaviv_gpu *gpu = cdev->devdata;
1628 
1629 	mutex_lock(&gpu->lock);
1630 	gpu->freq_scale = state;
1631 	if (!pm_runtime_suspended(gpu->dev))
1632 		etnaviv_gpu_update_clock(gpu);
1633 	mutex_unlock(&gpu->lock);
1634 
1635 	return 0;
1636 }
1637 
1638 static struct thermal_cooling_device_ops cooling_ops = {
1639 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1640 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1641 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1642 };
1643 
1644 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1645 	void *data)
1646 {
1647 	struct drm_device *drm = data;
1648 	struct etnaviv_drm_private *priv = drm->dev_private;
1649 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1650 	int ret;
1651 
1652 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1653 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1654 				(char *)dev_name(dev), gpu, &cooling_ops);
1655 		if (IS_ERR(gpu->cooling))
1656 			return PTR_ERR(gpu->cooling);
1657 	}
1658 
1659 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1660 	if (!gpu->wq) {
1661 		ret = -ENOMEM;
1662 		goto out_thermal;
1663 	}
1664 
1665 	ret = etnaviv_sched_init(gpu);
1666 	if (ret)
1667 		goto out_workqueue;
1668 
1669 #ifdef CONFIG_PM
1670 	ret = pm_runtime_get_sync(gpu->dev);
1671 #else
1672 	ret = etnaviv_gpu_clk_enable(gpu);
1673 #endif
1674 	if (ret < 0)
1675 		goto out_sched;
1676 
1677 
1678 	gpu->drm = drm;
1679 	gpu->fence_context = dma_fence_context_alloc(1);
1680 	idr_init(&gpu->fence_idr);
1681 	spin_lock_init(&gpu->fence_spinlock);
1682 
1683 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1684 	init_waitqueue_head(&gpu->fence_event);
1685 
1686 	priv->gpu[priv->num_gpus++] = gpu;
1687 
1688 	pm_runtime_mark_last_busy(gpu->dev);
1689 	pm_runtime_put_autosuspend(gpu->dev);
1690 
1691 	return 0;
1692 
1693 out_sched:
1694 	etnaviv_sched_fini(gpu);
1695 
1696 out_workqueue:
1697 	destroy_workqueue(gpu->wq);
1698 
1699 out_thermal:
1700 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1701 		thermal_cooling_device_unregister(gpu->cooling);
1702 
1703 	return ret;
1704 }
1705 
1706 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1707 	void *data)
1708 {
1709 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1710 
1711 	DBG("%s", dev_name(gpu->dev));
1712 
1713 	flush_workqueue(gpu->wq);
1714 	destroy_workqueue(gpu->wq);
1715 
1716 	etnaviv_sched_fini(gpu);
1717 
1718 #ifdef CONFIG_PM
1719 	pm_runtime_get_sync(gpu->dev);
1720 	pm_runtime_put_sync_suspend(gpu->dev);
1721 #else
1722 	etnaviv_gpu_hw_suspend(gpu);
1723 #endif
1724 
1725 	if (gpu->initialized) {
1726 		etnaviv_cmdbuf_free(&gpu->buffer);
1727 		etnaviv_iommu_global_fini(gpu);
1728 		gpu->initialized = false;
1729 	}
1730 
1731 	gpu->drm = NULL;
1732 	idr_destroy(&gpu->fence_idr);
1733 
1734 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1735 		thermal_cooling_device_unregister(gpu->cooling);
1736 	gpu->cooling = NULL;
1737 }
1738 
1739 static const struct component_ops gpu_ops = {
1740 	.bind = etnaviv_gpu_bind,
1741 	.unbind = etnaviv_gpu_unbind,
1742 };
1743 
1744 static const struct of_device_id etnaviv_gpu_match[] = {
1745 	{
1746 		.compatible = "vivante,gc"
1747 	},
1748 	{ /* sentinel */ }
1749 };
1750 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1751 
1752 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1753 {
1754 	struct device *dev = &pdev->dev;
1755 	struct etnaviv_gpu *gpu;
1756 	int err;
1757 
1758 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1759 	if (!gpu)
1760 		return -ENOMEM;
1761 
1762 	gpu->dev = &pdev->dev;
1763 	mutex_init(&gpu->lock);
1764 	mutex_init(&gpu->fence_lock);
1765 
1766 	/* Map registers: */
1767 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1768 	if (IS_ERR(gpu->mmio))
1769 		return PTR_ERR(gpu->mmio);
1770 
1771 	/* Get Interrupt: */
1772 	gpu->irq = platform_get_irq(pdev, 0);
1773 	if (gpu->irq < 0) {
1774 		dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1775 		return gpu->irq;
1776 	}
1777 
1778 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1779 			       dev_name(gpu->dev), gpu);
1780 	if (err) {
1781 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1782 		return err;
1783 	}
1784 
1785 	/* Get Clocks: */
1786 	gpu->clk_reg = devm_clk_get(&pdev->dev, "reg");
1787 	DBG("clk_reg: %p", gpu->clk_reg);
1788 	if (IS_ERR(gpu->clk_reg))
1789 		gpu->clk_reg = NULL;
1790 
1791 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1792 	DBG("clk_bus: %p", gpu->clk_bus);
1793 	if (IS_ERR(gpu->clk_bus))
1794 		gpu->clk_bus = NULL;
1795 
1796 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1797 	DBG("clk_core: %p", gpu->clk_core);
1798 	if (IS_ERR(gpu->clk_core))
1799 		gpu->clk_core = NULL;
1800 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1801 
1802 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1803 	DBG("clk_shader: %p", gpu->clk_shader);
1804 	if (IS_ERR(gpu->clk_shader))
1805 		gpu->clk_shader = NULL;
1806 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1807 
1808 	/* TODO: figure out max mapped size */
1809 	dev_set_drvdata(dev, gpu);
1810 
1811 	/*
1812 	 * We treat the device as initially suspended.  The runtime PM
1813 	 * autosuspend delay is rather arbitary: no measurements have
1814 	 * yet been performed to determine an appropriate value.
1815 	 */
1816 	pm_runtime_use_autosuspend(gpu->dev);
1817 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1818 	pm_runtime_enable(gpu->dev);
1819 
1820 	err = component_add(&pdev->dev, &gpu_ops);
1821 	if (err < 0) {
1822 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1823 		return err;
1824 	}
1825 
1826 	return 0;
1827 }
1828 
1829 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1830 {
1831 	component_del(&pdev->dev, &gpu_ops);
1832 	pm_runtime_disable(&pdev->dev);
1833 	return 0;
1834 }
1835 
1836 #ifdef CONFIG_PM
1837 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1838 {
1839 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1840 	u32 idle, mask;
1841 
1842 	/* If there are any jobs in the HW queue, we're not idle */
1843 	if (atomic_read(&gpu->sched.hw_rq_count))
1844 		return -EBUSY;
1845 
1846 	/* Check whether the hardware (except FE and MC) is idle */
1847 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1848 				  VIVS_HI_IDLE_STATE_MC);
1849 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1850 	if (idle != mask) {
1851 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1852 				     idle);
1853 		return -EBUSY;
1854 	}
1855 
1856 	return etnaviv_gpu_hw_suspend(gpu);
1857 }
1858 
1859 static int etnaviv_gpu_rpm_resume(struct device *dev)
1860 {
1861 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1862 	int ret;
1863 
1864 	ret = etnaviv_gpu_clk_enable(gpu);
1865 	if (ret)
1866 		return ret;
1867 
1868 	/* Re-initialise the basic hardware state */
1869 	if (gpu->drm && gpu->initialized) {
1870 		ret = etnaviv_gpu_hw_resume(gpu);
1871 		if (ret) {
1872 			etnaviv_gpu_clk_disable(gpu);
1873 			return ret;
1874 		}
1875 	}
1876 
1877 	return 0;
1878 }
1879 #endif
1880 
1881 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1882 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1883 			   NULL)
1884 };
1885 
1886 struct platform_driver etnaviv_gpu_driver = {
1887 	.driver = {
1888 		.name = "etnaviv-gpu",
1889 		.owner = THIS_MODULE,
1890 		.pm = &etnaviv_gpu_pm_ops,
1891 		.of_match_table = etnaviv_gpu_match,
1892 	},
1893 	.probe = etnaviv_gpu_platform_probe,
1894 	.remove = etnaviv_gpu_platform_remove,
1895 	.id_table = gpu_ids,
1896 };
1897