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