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.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
341 
342 		/*
343 		 * Reading these two registers on GC600 rev 0x19 result in a
344 		 * unhandled fault: external abort on non-linefetch
345 		 */
346 		if (!etnaviv_is_model_rev(gpu, GC600, 0x19)) {
347 			gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
348 			gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
349 		}
350 
351 		/*
352 		 * !!!! HACK ALERT !!!!
353 		 * Because people change device IDs without letting software
354 		 * know about it - here is the hack to make it all look the
355 		 * same.  Only for GC400 family.
356 		 */
357 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
358 		    gpu->identity.model != chipModel_GC420) {
359 			gpu->identity.model = gpu->identity.model & 0x0400;
360 		}
361 
362 		/* Another special case */
363 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
364 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
365 
366 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
367 				/*
368 				 * This IP has an ECO; put the correct
369 				 * revision in it.
370 				 */
371 				gpu->identity.revision = 0x1051;
372 			}
373 		}
374 
375 		/*
376 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
377 		 * reality it's just a re-branded GC3000. We can identify this
378 		 * core by the upper half of the revision register being all 1.
379 		 * Fix model/rev here, so all other places can refer to this
380 		 * core by its real identity.
381 		 */
382 		if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
383 			gpu->identity.model = chipModel_GC3000;
384 			gpu->identity.revision &= 0xffff;
385 		}
386 
387 		if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
388 			gpu->identity.eco_id = 1;
389 
390 		if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
391 			gpu->identity.eco_id = 1;
392 	}
393 
394 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
395 		 gpu->identity.model, gpu->identity.revision);
396 
397 	gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
398 	/*
399 	 * If there is a match in the HWDB, we aren't interested in the
400 	 * remaining register values, as they might be wrong.
401 	 */
402 	if (etnaviv_fill_identity_from_hwdb(gpu))
403 		return;
404 
405 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
406 
407 	/* Disable fast clear on GC700. */
408 	if (gpu->identity.model == chipModel_GC700)
409 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
410 
411 	if ((gpu->identity.model == chipModel_GC500 &&
412 	     gpu->identity.revision < 2) ||
413 	    (gpu->identity.model == chipModel_GC300 &&
414 	     gpu->identity.revision < 0x2000)) {
415 
416 		/*
417 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
418 		 * registers.
419 		 */
420 		gpu->identity.minor_features0 = 0;
421 		gpu->identity.minor_features1 = 0;
422 		gpu->identity.minor_features2 = 0;
423 		gpu->identity.minor_features3 = 0;
424 		gpu->identity.minor_features4 = 0;
425 		gpu->identity.minor_features5 = 0;
426 	} else
427 		gpu->identity.minor_features0 =
428 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
429 
430 	if (gpu->identity.minor_features0 &
431 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
432 		gpu->identity.minor_features1 =
433 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
434 		gpu->identity.minor_features2 =
435 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
436 		gpu->identity.minor_features3 =
437 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
438 		gpu->identity.minor_features4 =
439 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
440 		gpu->identity.minor_features5 =
441 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
442 	}
443 
444 	/* GC600 idle register reports zero bits where modules aren't present */
445 	if (gpu->identity.model == chipModel_GC600)
446 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
447 				 VIVS_HI_IDLE_STATE_RA |
448 				 VIVS_HI_IDLE_STATE_SE |
449 				 VIVS_HI_IDLE_STATE_PA |
450 				 VIVS_HI_IDLE_STATE_SH |
451 				 VIVS_HI_IDLE_STATE_PE |
452 				 VIVS_HI_IDLE_STATE_DE |
453 				 VIVS_HI_IDLE_STATE_FE;
454 
455 	etnaviv_hw_specs(gpu);
456 }
457 
458 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
459 {
460 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
461 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
462 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
463 }
464 
465 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
466 {
467 	if (gpu->identity.minor_features2 &
468 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
469 		clk_set_rate(gpu->clk_core,
470 			     gpu->base_rate_core >> gpu->freq_scale);
471 		clk_set_rate(gpu->clk_shader,
472 			     gpu->base_rate_shader >> gpu->freq_scale);
473 	} else {
474 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
475 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
476 
477 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
478 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
479 		etnaviv_gpu_load_clock(gpu, clock);
480 	}
481 }
482 
483 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
484 {
485 	u32 control, idle;
486 	unsigned long timeout;
487 	bool failed = true;
488 
489 	/* We hope that the GPU resets in under one second */
490 	timeout = jiffies + msecs_to_jiffies(1000);
491 
492 	while (time_is_after_jiffies(timeout)) {
493 		/* enable clock */
494 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
495 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
496 		etnaviv_gpu_load_clock(gpu, control);
497 
498 		/* isolate the GPU. */
499 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
500 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
501 
502 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
503 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
504 			          VIVS_MMUv2_AHB_CONTROL_RESET);
505 		} else {
506 			/* set soft reset. */
507 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
508 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
509 		}
510 
511 		/* wait for reset. */
512 		usleep_range(10, 20);
513 
514 		/* reset soft reset bit. */
515 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
516 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
517 
518 		/* reset GPU isolation. */
519 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
520 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
521 
522 		/* read idle register. */
523 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
524 
525 		/* try resetting again if FE is not idle */
526 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
527 			dev_dbg(gpu->dev, "FE is not idle\n");
528 			continue;
529 		}
530 
531 		/* read reset register. */
532 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
533 
534 		/* is the GPU idle? */
535 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
536 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
537 			dev_dbg(gpu->dev, "GPU is not idle\n");
538 			continue;
539 		}
540 
541 		/* disable debug registers, as they are not normally needed */
542 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
543 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
544 
545 		failed = false;
546 		break;
547 	}
548 
549 	if (failed) {
550 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
551 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
552 
553 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
554 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
555 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
556 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
557 
558 		return -EBUSY;
559 	}
560 
561 	/* We rely on the GPU running, so program the clock */
562 	etnaviv_gpu_update_clock(gpu);
563 
564 	return 0;
565 }
566 
567 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
568 {
569 	u32 pmc, ppc;
570 
571 	/* enable clock gating */
572 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
573 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
574 
575 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
576 	if (gpu->identity.revision == 0x4301 ||
577 	    gpu->identity.revision == 0x4302)
578 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
579 
580 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
581 
582 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
583 
584 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
585 	if (gpu->identity.model >= chipModel_GC400 &&
586 	    gpu->identity.model != chipModel_GC420 &&
587 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
588 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
589 
590 	/*
591 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
592 	 * present without a bug fix.
593 	 */
594 	if (gpu->identity.revision < 0x5000 &&
595 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
596 	    !(gpu->identity.minor_features1 &
597 	      chipMinorFeatures1_DISABLE_PE_GATING))
598 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
599 
600 	if (gpu->identity.revision < 0x5422)
601 		pmc |= BIT(15); /* Unknown bit */
602 
603 	/* Disable TX clock gating on affected core revisions. */
604 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
605 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108))
606 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
607 
608 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
609 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
610 
611 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
612 }
613 
614 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
615 {
616 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
617 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
618 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
619 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
620 
621 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
622 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
623 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
624 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
625 	}
626 }
627 
628 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu)
629 {
630 	u32 address = etnaviv_cmdbuf_get_va(&gpu->buffer,
631 				&gpu->mmu_context->cmdbuf_mapping);
632 	u16 prefetch;
633 
634 	/* setup the MMU */
635 	etnaviv_iommu_restore(gpu, gpu->mmu_context);
636 
637 	/* Start command processor */
638 	prefetch = etnaviv_buffer_init(gpu);
639 
640 	etnaviv_gpu_start_fe(gpu, address, prefetch);
641 }
642 
643 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
644 {
645 	/*
646 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
647 	 * cannot be read, extracted from vivante kernel driver.
648 	 */
649 	u32 pulse_eater = 0x01590880;
650 
651 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
652 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
653 		pulse_eater |= BIT(23);
654 
655 	}
656 
657 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
658 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
659 		pulse_eater &= ~BIT(16);
660 		pulse_eater |= BIT(17);
661 	}
662 
663 	if ((gpu->identity.revision > 0x5420) &&
664 	    (gpu->identity.features & chipFeatures_PIPE_3D))
665 	{
666 		/* Performance fix: disable internal DFS */
667 		pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
668 		pulse_eater |= BIT(18);
669 	}
670 
671 	gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
672 }
673 
674 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
675 {
676 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
677 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
678 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
679 		u32 mc_memory_debug;
680 
681 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
682 
683 		if (gpu->identity.revision == 0x5007)
684 			mc_memory_debug |= 0x0c;
685 		else
686 			mc_memory_debug |= 0x08;
687 
688 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
689 	}
690 
691 	/* enable module-level clock gating */
692 	etnaviv_gpu_enable_mlcg(gpu);
693 
694 	/*
695 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
696 	 * This is necessary to prevent the iMX6 SoC locking up.
697 	 */
698 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
699 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
700 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
701 
702 	/* GC2000 rev 5108 needs a special bus config */
703 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
704 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
705 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
706 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
707 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
708 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
709 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
710 	}
711 
712 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
713 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
714 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
715 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
716 	}
717 
718 	/* setup the pulse eater */
719 	etnaviv_gpu_setup_pulse_eater(gpu);
720 
721 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
722 }
723 
724 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
725 {
726 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
727 	int ret, i;
728 
729 	ret = pm_runtime_get_sync(gpu->dev);
730 	if (ret < 0) {
731 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
732 		goto pm_put;
733 	}
734 
735 	etnaviv_hw_identify(gpu);
736 
737 	if (gpu->identity.model == 0) {
738 		dev_err(gpu->dev, "Unknown GPU model\n");
739 		ret = -ENXIO;
740 		goto fail;
741 	}
742 
743 	/* Exclude VG cores with FE2.0 */
744 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
745 	    gpu->identity.features & chipFeatures_FE20) {
746 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
747 		ret = -ENXIO;
748 		goto fail;
749 	}
750 
751 	/*
752 	 * On cores with security features supported, we claim control over the
753 	 * security states.
754 	 */
755 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
756 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
757 		gpu->sec_mode = ETNA_SEC_KERNEL;
758 
759 	ret = etnaviv_hw_reset(gpu);
760 	if (ret) {
761 		dev_err(gpu->dev, "GPU reset failed\n");
762 		goto fail;
763 	}
764 
765 	ret = etnaviv_iommu_global_init(gpu);
766 	if (ret)
767 		goto fail;
768 
769 	/*
770 	 * Set the GPU linear window to be at the end of the DMA window, where
771 	 * the CMA area is likely to reside. This ensures that we are able to
772 	 * map the command buffers while having the linear window overlap as
773 	 * much RAM as possible, so we can optimize mappings for other buffers.
774 	 *
775 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
776 	 * to different views of the memory on the individual engines.
777 	 */
778 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
779 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
780 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
781 		if (dma_mask < PHYS_OFFSET + SZ_2G)
782 			priv->mmu_global->memory_base = PHYS_OFFSET;
783 		else
784 			priv->mmu_global->memory_base = dma_mask - SZ_2G + 1;
785 	} else if (PHYS_OFFSET >= SZ_2G) {
786 		dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
787 		priv->mmu_global->memory_base = PHYS_OFFSET;
788 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
789 	}
790 
791 	/*
792 	 * If the GPU is part of a system with DMA addressing limitations,
793 	 * request pages for our SHM backend buffers from the DMA32 zone to
794 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
795 	 */
796 	if (dma_addressing_limited(gpu->dev))
797 		priv->shm_gfp_mask |= GFP_DMA32;
798 
799 	/* Create buffer: */
800 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
801 				  PAGE_SIZE);
802 	if (ret) {
803 		dev_err(gpu->dev, "could not create command buffer\n");
804 		goto fail;
805 	}
806 
807 	/* Setup event management */
808 	spin_lock_init(&gpu->event_spinlock);
809 	init_completion(&gpu->event_free);
810 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
811 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
812 		complete(&gpu->event_free);
813 
814 	/* Now program the hardware */
815 	mutex_lock(&gpu->lock);
816 	etnaviv_gpu_hw_init(gpu);
817 	gpu->exec_state = -1;
818 	mutex_unlock(&gpu->lock);
819 
820 	pm_runtime_mark_last_busy(gpu->dev);
821 	pm_runtime_put_autosuspend(gpu->dev);
822 
823 	gpu->initialized = true;
824 
825 	return 0;
826 
827 fail:
828 	pm_runtime_mark_last_busy(gpu->dev);
829 pm_put:
830 	pm_runtime_put_autosuspend(gpu->dev);
831 
832 	return ret;
833 }
834 
835 #ifdef CONFIG_DEBUG_FS
836 struct dma_debug {
837 	u32 address[2];
838 	u32 state[2];
839 };
840 
841 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
842 {
843 	u32 i;
844 
845 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
846 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
847 
848 	for (i = 0; i < 500; i++) {
849 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
850 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
851 
852 		if (debug->address[0] != debug->address[1])
853 			break;
854 
855 		if (debug->state[0] != debug->state[1])
856 			break;
857 	}
858 }
859 
860 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
861 {
862 	struct dma_debug debug;
863 	u32 dma_lo, dma_hi, axi, idle;
864 	int ret;
865 
866 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
867 
868 	ret = pm_runtime_get_sync(gpu->dev);
869 	if (ret < 0)
870 		goto pm_put;
871 
872 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
873 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
874 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
875 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
876 
877 	verify_dma(gpu, &debug);
878 
879 	seq_puts(m, "\tidentity\n");
880 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
881 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
882 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
883 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
884 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
885 
886 	seq_puts(m, "\tfeatures\n");
887 	seq_printf(m, "\t major_features: 0x%08x\n",
888 		   gpu->identity.features);
889 	seq_printf(m, "\t minor_features0: 0x%08x\n",
890 		   gpu->identity.minor_features0);
891 	seq_printf(m, "\t minor_features1: 0x%08x\n",
892 		   gpu->identity.minor_features1);
893 	seq_printf(m, "\t minor_features2: 0x%08x\n",
894 		   gpu->identity.minor_features2);
895 	seq_printf(m, "\t minor_features3: 0x%08x\n",
896 		   gpu->identity.minor_features3);
897 	seq_printf(m, "\t minor_features4: 0x%08x\n",
898 		   gpu->identity.minor_features4);
899 	seq_printf(m, "\t minor_features5: 0x%08x\n",
900 		   gpu->identity.minor_features5);
901 	seq_printf(m, "\t minor_features6: 0x%08x\n",
902 		   gpu->identity.minor_features6);
903 	seq_printf(m, "\t minor_features7: 0x%08x\n",
904 		   gpu->identity.minor_features7);
905 	seq_printf(m, "\t minor_features8: 0x%08x\n",
906 		   gpu->identity.minor_features8);
907 	seq_printf(m, "\t minor_features9: 0x%08x\n",
908 		   gpu->identity.minor_features9);
909 	seq_printf(m, "\t minor_features10: 0x%08x\n",
910 		   gpu->identity.minor_features10);
911 	seq_printf(m, "\t minor_features11: 0x%08x\n",
912 		   gpu->identity.minor_features11);
913 
914 	seq_puts(m, "\tspecs\n");
915 	seq_printf(m, "\t stream_count:  %d\n",
916 			gpu->identity.stream_count);
917 	seq_printf(m, "\t register_max: %d\n",
918 			gpu->identity.register_max);
919 	seq_printf(m, "\t thread_count: %d\n",
920 			gpu->identity.thread_count);
921 	seq_printf(m, "\t vertex_cache_size: %d\n",
922 			gpu->identity.vertex_cache_size);
923 	seq_printf(m, "\t shader_core_count: %d\n",
924 			gpu->identity.shader_core_count);
925 	seq_printf(m, "\t pixel_pipes: %d\n",
926 			gpu->identity.pixel_pipes);
927 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
928 			gpu->identity.vertex_output_buffer_size);
929 	seq_printf(m, "\t buffer_size: %d\n",
930 			gpu->identity.buffer_size);
931 	seq_printf(m, "\t instruction_count: %d\n",
932 			gpu->identity.instruction_count);
933 	seq_printf(m, "\t num_constants: %d\n",
934 			gpu->identity.num_constants);
935 	seq_printf(m, "\t varyings_count: %d\n",
936 			gpu->identity.varyings_count);
937 
938 	seq_printf(m, "\taxi: 0x%08x\n", axi);
939 	seq_printf(m, "\tidle: 0x%08x\n", idle);
940 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
941 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
942 		seq_puts(m, "\t FE is not idle\n");
943 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
944 		seq_puts(m, "\t DE is not idle\n");
945 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
946 		seq_puts(m, "\t PE is not idle\n");
947 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
948 		seq_puts(m, "\t SH is not idle\n");
949 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
950 		seq_puts(m, "\t PA is not idle\n");
951 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
952 		seq_puts(m, "\t SE is not idle\n");
953 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
954 		seq_puts(m, "\t RA is not idle\n");
955 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
956 		seq_puts(m, "\t TX is not idle\n");
957 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
958 		seq_puts(m, "\t VG is not idle\n");
959 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
960 		seq_puts(m, "\t IM is not idle\n");
961 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
962 		seq_puts(m, "\t FP is not idle\n");
963 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
964 		seq_puts(m, "\t TS is not idle\n");
965 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
966 		seq_puts(m, "\t BL is not idle\n");
967 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
968 		seq_puts(m, "\t ASYNCFE is not idle\n");
969 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
970 		seq_puts(m, "\t MC is not idle\n");
971 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
972 		seq_puts(m, "\t PPA is not idle\n");
973 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
974 		seq_puts(m, "\t WD is not idle\n");
975 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
976 		seq_puts(m, "\t NN is not idle\n");
977 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
978 		seq_puts(m, "\t TP is not idle\n");
979 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
980 		seq_puts(m, "\t AXI low power mode\n");
981 
982 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
983 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
984 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
985 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
986 
987 		seq_puts(m, "\tMC\n");
988 		seq_printf(m, "\t read0: 0x%08x\n", read0);
989 		seq_printf(m, "\t read1: 0x%08x\n", read1);
990 		seq_printf(m, "\t write: 0x%08x\n", write);
991 	}
992 
993 	seq_puts(m, "\tDMA ");
994 
995 	if (debug.address[0] == debug.address[1] &&
996 	    debug.state[0] == debug.state[1]) {
997 		seq_puts(m, "seems to be stuck\n");
998 	} else if (debug.address[0] == debug.address[1]) {
999 		seq_puts(m, "address is constant\n");
1000 	} else {
1001 		seq_puts(m, "is running\n");
1002 	}
1003 
1004 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1005 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1006 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1007 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1008 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1009 		   dma_lo, dma_hi);
1010 
1011 	ret = 0;
1012 
1013 	pm_runtime_mark_last_busy(gpu->dev);
1014 pm_put:
1015 	pm_runtime_put_autosuspend(gpu->dev);
1016 
1017 	return ret;
1018 }
1019 #endif
1020 
1021 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
1022 {
1023 	unsigned int i = 0;
1024 
1025 	dev_err(gpu->dev, "recover hung GPU!\n");
1026 
1027 	if (pm_runtime_get_sync(gpu->dev) < 0)
1028 		goto pm_put;
1029 
1030 	mutex_lock(&gpu->lock);
1031 
1032 	etnaviv_hw_reset(gpu);
1033 
1034 	/* complete all events, the GPU won't do it after the reset */
1035 	spin_lock(&gpu->event_spinlock);
1036 	for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1037 		complete(&gpu->event_free);
1038 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1039 	spin_unlock(&gpu->event_spinlock);
1040 
1041 	etnaviv_gpu_hw_init(gpu);
1042 	gpu->exec_state = -1;
1043 	gpu->mmu_context = NULL;
1044 
1045 	mutex_unlock(&gpu->lock);
1046 	pm_runtime_mark_last_busy(gpu->dev);
1047 pm_put:
1048 	pm_runtime_put_autosuspend(gpu->dev);
1049 }
1050 
1051 /* fence object management */
1052 struct etnaviv_fence {
1053 	struct etnaviv_gpu *gpu;
1054 	struct dma_fence base;
1055 };
1056 
1057 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1058 {
1059 	return container_of(fence, struct etnaviv_fence, base);
1060 }
1061 
1062 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1063 {
1064 	return "etnaviv";
1065 }
1066 
1067 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1068 {
1069 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1070 
1071 	return dev_name(f->gpu->dev);
1072 }
1073 
1074 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1075 {
1076 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1077 
1078 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1079 }
1080 
1081 static void etnaviv_fence_release(struct dma_fence *fence)
1082 {
1083 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1084 
1085 	kfree_rcu(f, base.rcu);
1086 }
1087 
1088 static const struct dma_fence_ops etnaviv_fence_ops = {
1089 	.get_driver_name = etnaviv_fence_get_driver_name,
1090 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1091 	.signaled = etnaviv_fence_signaled,
1092 	.release = etnaviv_fence_release,
1093 };
1094 
1095 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1096 {
1097 	struct etnaviv_fence *f;
1098 
1099 	/*
1100 	 * GPU lock must already be held, otherwise fence completion order might
1101 	 * not match the seqno order assigned here.
1102 	 */
1103 	lockdep_assert_held(&gpu->lock);
1104 
1105 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1106 	if (!f)
1107 		return NULL;
1108 
1109 	f->gpu = gpu;
1110 
1111 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1112 		       gpu->fence_context, ++gpu->next_fence);
1113 
1114 	return &f->base;
1115 }
1116 
1117 /* returns true if fence a comes after fence b */
1118 static inline bool fence_after(u32 a, u32 b)
1119 {
1120 	return (s32)(a - b) > 0;
1121 }
1122 
1123 /*
1124  * event management:
1125  */
1126 
1127 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1128 	unsigned int *events)
1129 {
1130 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1131 	unsigned i, acquired = 0;
1132 
1133 	for (i = 0; i < nr_events; i++) {
1134 		unsigned long ret;
1135 
1136 		ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1137 
1138 		if (!ret) {
1139 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1140 			goto out;
1141 		}
1142 
1143 		acquired++;
1144 		timeout = ret;
1145 	}
1146 
1147 	spin_lock(&gpu->event_spinlock);
1148 
1149 	for (i = 0; i < nr_events; i++) {
1150 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1151 
1152 		events[i] = event;
1153 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1154 		set_bit(event, gpu->event_bitmap);
1155 	}
1156 
1157 	spin_unlock(&gpu->event_spinlock);
1158 
1159 	return 0;
1160 
1161 out:
1162 	for (i = 0; i < acquired; i++)
1163 		complete(&gpu->event_free);
1164 
1165 	return -EBUSY;
1166 }
1167 
1168 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1169 {
1170 	if (!test_bit(event, gpu->event_bitmap)) {
1171 		dev_warn(gpu->dev, "event %u is already marked as free",
1172 			 event);
1173 	} else {
1174 		clear_bit(event, gpu->event_bitmap);
1175 		complete(&gpu->event_free);
1176 	}
1177 }
1178 
1179 /*
1180  * Cmdstream submission/retirement:
1181  */
1182 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1183 	u32 id, struct drm_etnaviv_timespec *timeout)
1184 {
1185 	struct dma_fence *fence;
1186 	int ret;
1187 
1188 	/*
1189 	 * Look up the fence and take a reference. We might still find a fence
1190 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1191 	 * pretends we didn't find a fence in that case.
1192 	 */
1193 	rcu_read_lock();
1194 	fence = idr_find(&gpu->fence_idr, id);
1195 	if (fence)
1196 		fence = dma_fence_get_rcu(fence);
1197 	rcu_read_unlock();
1198 
1199 	if (!fence)
1200 		return 0;
1201 
1202 	if (!timeout) {
1203 		/* No timeout was requested: just test for completion */
1204 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1205 	} else {
1206 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1207 
1208 		ret = dma_fence_wait_timeout(fence, true, remaining);
1209 		if (ret == 0)
1210 			ret = -ETIMEDOUT;
1211 		else if (ret != -ERESTARTSYS)
1212 			ret = 0;
1213 
1214 	}
1215 
1216 	dma_fence_put(fence);
1217 	return ret;
1218 }
1219 
1220 /*
1221  * Wait for an object to become inactive.  This, on it's own, is not race
1222  * free: the object is moved by the scheduler off the active list, and
1223  * then the iova is put.  Moreover, the object could be re-submitted just
1224  * after we notice that it's become inactive.
1225  *
1226  * Although the retirement happens under the gpu lock, we don't want to hold
1227  * that lock in this function while waiting.
1228  */
1229 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1230 	struct etnaviv_gem_object *etnaviv_obj,
1231 	struct drm_etnaviv_timespec *timeout)
1232 {
1233 	unsigned long remaining;
1234 	long ret;
1235 
1236 	if (!timeout)
1237 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1238 
1239 	remaining = etnaviv_timeout_to_jiffies(timeout);
1240 
1241 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1242 					       !is_active(etnaviv_obj),
1243 					       remaining);
1244 	if (ret > 0)
1245 		return 0;
1246 	else if (ret == -ERESTARTSYS)
1247 		return -ERESTARTSYS;
1248 	else
1249 		return -ETIMEDOUT;
1250 }
1251 
1252 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1253 	struct etnaviv_event *event, unsigned int flags)
1254 {
1255 	const struct etnaviv_gem_submit *submit = event->submit;
1256 	unsigned int i;
1257 
1258 	for (i = 0; i < submit->nr_pmrs; i++) {
1259 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1260 
1261 		if (pmr->flags == flags)
1262 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1263 	}
1264 }
1265 
1266 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1267 	struct etnaviv_event *event)
1268 {
1269 	u32 val;
1270 
1271 	/* disable clock gating */
1272 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1273 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1274 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1275 
1276 	/* enable debug register */
1277 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1278 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1279 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1280 
1281 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1282 }
1283 
1284 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1285 	struct etnaviv_event *event)
1286 {
1287 	const struct etnaviv_gem_submit *submit = event->submit;
1288 	unsigned int i;
1289 	u32 val;
1290 
1291 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1292 
1293 	for (i = 0; i < submit->nr_pmrs; i++) {
1294 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1295 
1296 		*pmr->bo_vma = pmr->sequence;
1297 	}
1298 
1299 	/* disable debug register */
1300 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1301 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1302 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1303 
1304 	/* enable clock gating */
1305 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1306 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1307 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1308 }
1309 
1310 
1311 /* add bo's to gpu's ring, and kick gpu: */
1312 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1313 {
1314 	struct etnaviv_gpu *gpu = submit->gpu;
1315 	struct dma_fence *gpu_fence;
1316 	unsigned int i, nr_events = 1, event[3];
1317 	int ret;
1318 
1319 	if (!submit->runtime_resumed) {
1320 		ret = pm_runtime_get_sync(gpu->dev);
1321 		if (ret < 0) {
1322 			pm_runtime_put_noidle(gpu->dev);
1323 			return NULL;
1324 		}
1325 		submit->runtime_resumed = true;
1326 	}
1327 
1328 	/*
1329 	 * if there are performance monitor requests we need to have
1330 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1331 	 *   requests.
1332 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1333 	 *   and update the sequence number for userspace.
1334 	 */
1335 	if (submit->nr_pmrs)
1336 		nr_events = 3;
1337 
1338 	ret = event_alloc(gpu, nr_events, event);
1339 	if (ret) {
1340 		DRM_ERROR("no free events\n");
1341 		pm_runtime_put_noidle(gpu->dev);
1342 		return NULL;
1343 	}
1344 
1345 	mutex_lock(&gpu->lock);
1346 
1347 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1348 	if (!gpu_fence) {
1349 		for (i = 0; i < nr_events; i++)
1350 			event_free(gpu, event[i]);
1351 
1352 		goto out_unlock;
1353 	}
1354 
1355 	if (!gpu->mmu_context) {
1356 		etnaviv_iommu_context_get(submit->mmu_context);
1357 		gpu->mmu_context = submit->mmu_context;
1358 		etnaviv_gpu_start_fe_idleloop(gpu);
1359 	} else {
1360 		etnaviv_iommu_context_get(gpu->mmu_context);
1361 		submit->prev_mmu_context = gpu->mmu_context;
1362 	}
1363 
1364 	if (submit->nr_pmrs) {
1365 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1366 		kref_get(&submit->refcount);
1367 		gpu->event[event[1]].submit = submit;
1368 		etnaviv_sync_point_queue(gpu, event[1]);
1369 	}
1370 
1371 	gpu->event[event[0]].fence = gpu_fence;
1372 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1373 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1374 			     event[0], &submit->cmdbuf);
1375 
1376 	if (submit->nr_pmrs) {
1377 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1378 		kref_get(&submit->refcount);
1379 		gpu->event[event[2]].submit = submit;
1380 		etnaviv_sync_point_queue(gpu, event[2]);
1381 	}
1382 
1383 out_unlock:
1384 	mutex_unlock(&gpu->lock);
1385 
1386 	return gpu_fence;
1387 }
1388 
1389 static void sync_point_worker(struct work_struct *work)
1390 {
1391 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1392 					       sync_point_work);
1393 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1394 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1395 
1396 	event->sync_point(gpu, event);
1397 	etnaviv_submit_put(event->submit);
1398 	event_free(gpu, gpu->sync_point_event);
1399 
1400 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1401 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1402 }
1403 
1404 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1405 {
1406 	u32 status_reg, status;
1407 	int i;
1408 
1409 	if (gpu->sec_mode == ETNA_SEC_NONE)
1410 		status_reg = VIVS_MMUv2_STATUS;
1411 	else
1412 		status_reg = VIVS_MMUv2_SEC_STATUS;
1413 
1414 	status = gpu_read(gpu, status_reg);
1415 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1416 
1417 	for (i = 0; i < 4; i++) {
1418 		u32 address_reg;
1419 
1420 		if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1421 			continue;
1422 
1423 		if (gpu->sec_mode == ETNA_SEC_NONE)
1424 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1425 		else
1426 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1427 
1428 		dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1429 				    gpu_read(gpu, address_reg));
1430 	}
1431 }
1432 
1433 static irqreturn_t irq_handler(int irq, void *data)
1434 {
1435 	struct etnaviv_gpu *gpu = data;
1436 	irqreturn_t ret = IRQ_NONE;
1437 
1438 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1439 
1440 	if (intr != 0) {
1441 		int event;
1442 
1443 		pm_runtime_mark_last_busy(gpu->dev);
1444 
1445 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1446 
1447 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1448 			dev_err(gpu->dev, "AXI bus error\n");
1449 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1450 		}
1451 
1452 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1453 			dump_mmu_fault(gpu);
1454 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1455 		}
1456 
1457 		while ((event = ffs(intr)) != 0) {
1458 			struct dma_fence *fence;
1459 
1460 			event -= 1;
1461 
1462 			intr &= ~(1 << event);
1463 
1464 			dev_dbg(gpu->dev, "event %u\n", event);
1465 
1466 			if (gpu->event[event].sync_point) {
1467 				gpu->sync_point_event = event;
1468 				queue_work(gpu->wq, &gpu->sync_point_work);
1469 			}
1470 
1471 			fence = gpu->event[event].fence;
1472 			if (!fence)
1473 				continue;
1474 
1475 			gpu->event[event].fence = NULL;
1476 
1477 			/*
1478 			 * Events can be processed out of order.  Eg,
1479 			 * - allocate and queue event 0
1480 			 * - allocate event 1
1481 			 * - event 0 completes, we process it
1482 			 * - allocate and queue event 0
1483 			 * - event 1 and event 0 complete
1484 			 * we can end up processing event 0 first, then 1.
1485 			 */
1486 			if (fence_after(fence->seqno, gpu->completed_fence))
1487 				gpu->completed_fence = fence->seqno;
1488 			dma_fence_signal(fence);
1489 
1490 			event_free(gpu, event);
1491 		}
1492 
1493 		ret = IRQ_HANDLED;
1494 	}
1495 
1496 	return ret;
1497 }
1498 
1499 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1500 {
1501 	int ret;
1502 
1503 	ret = clk_prepare_enable(gpu->clk_reg);
1504 	if (ret)
1505 		return ret;
1506 
1507 	ret = clk_prepare_enable(gpu->clk_bus);
1508 	if (ret)
1509 		goto disable_clk_reg;
1510 
1511 	ret = clk_prepare_enable(gpu->clk_core);
1512 	if (ret)
1513 		goto disable_clk_bus;
1514 
1515 	ret = clk_prepare_enable(gpu->clk_shader);
1516 	if (ret)
1517 		goto disable_clk_core;
1518 
1519 	return 0;
1520 
1521 disable_clk_core:
1522 	clk_disable_unprepare(gpu->clk_core);
1523 disable_clk_bus:
1524 	clk_disable_unprepare(gpu->clk_bus);
1525 disable_clk_reg:
1526 	clk_disable_unprepare(gpu->clk_reg);
1527 
1528 	return ret;
1529 }
1530 
1531 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1532 {
1533 	clk_disable_unprepare(gpu->clk_shader);
1534 	clk_disable_unprepare(gpu->clk_core);
1535 	clk_disable_unprepare(gpu->clk_bus);
1536 	clk_disable_unprepare(gpu->clk_reg);
1537 
1538 	return 0;
1539 }
1540 
1541 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1542 {
1543 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1544 
1545 	do {
1546 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1547 
1548 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1549 			return 0;
1550 
1551 		if (time_is_before_jiffies(timeout)) {
1552 			dev_warn(gpu->dev,
1553 				 "timed out waiting for idle: idle=0x%x\n",
1554 				 idle);
1555 			return -ETIMEDOUT;
1556 		}
1557 
1558 		udelay(5);
1559 	} while (1);
1560 }
1561 
1562 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1563 {
1564 	if (gpu->initialized && gpu->mmu_context) {
1565 		/* Replace the last WAIT with END */
1566 		mutex_lock(&gpu->lock);
1567 		etnaviv_buffer_end(gpu);
1568 		mutex_unlock(&gpu->lock);
1569 
1570 		/*
1571 		 * We know that only the FE is busy here, this should
1572 		 * happen quickly (as the WAIT is only 200 cycles).  If
1573 		 * we fail, just warn and continue.
1574 		 */
1575 		etnaviv_gpu_wait_idle(gpu, 100);
1576 
1577 		etnaviv_iommu_context_put(gpu->mmu_context);
1578 		gpu->mmu_context = NULL;
1579 	}
1580 
1581 	gpu->exec_state = -1;
1582 
1583 	return etnaviv_gpu_clk_disable(gpu);
1584 }
1585 
1586 #ifdef CONFIG_PM
1587 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1588 {
1589 	int ret;
1590 
1591 	ret = mutex_lock_killable(&gpu->lock);
1592 	if (ret)
1593 		return ret;
1594 
1595 	etnaviv_gpu_update_clock(gpu);
1596 	etnaviv_gpu_hw_init(gpu);
1597 
1598 	mutex_unlock(&gpu->lock);
1599 
1600 	return 0;
1601 }
1602 #endif
1603 
1604 static int
1605 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1606 				  unsigned long *state)
1607 {
1608 	*state = 6;
1609 
1610 	return 0;
1611 }
1612 
1613 static int
1614 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1615 				  unsigned long *state)
1616 {
1617 	struct etnaviv_gpu *gpu = cdev->devdata;
1618 
1619 	*state = gpu->freq_scale;
1620 
1621 	return 0;
1622 }
1623 
1624 static int
1625 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1626 				  unsigned long state)
1627 {
1628 	struct etnaviv_gpu *gpu = cdev->devdata;
1629 
1630 	mutex_lock(&gpu->lock);
1631 	gpu->freq_scale = state;
1632 	if (!pm_runtime_suspended(gpu->dev))
1633 		etnaviv_gpu_update_clock(gpu);
1634 	mutex_unlock(&gpu->lock);
1635 
1636 	return 0;
1637 }
1638 
1639 static struct thermal_cooling_device_ops cooling_ops = {
1640 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1641 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1642 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1643 };
1644 
1645 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1646 	void *data)
1647 {
1648 	struct drm_device *drm = data;
1649 	struct etnaviv_drm_private *priv = drm->dev_private;
1650 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1651 	int ret;
1652 
1653 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1654 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1655 				(char *)dev_name(dev), gpu, &cooling_ops);
1656 		if (IS_ERR(gpu->cooling))
1657 			return PTR_ERR(gpu->cooling);
1658 	}
1659 
1660 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1661 	if (!gpu->wq) {
1662 		ret = -ENOMEM;
1663 		goto out_thermal;
1664 	}
1665 
1666 	ret = etnaviv_sched_init(gpu);
1667 	if (ret)
1668 		goto out_workqueue;
1669 
1670 #ifdef CONFIG_PM
1671 	ret = pm_runtime_get_sync(gpu->dev);
1672 #else
1673 	ret = etnaviv_gpu_clk_enable(gpu);
1674 #endif
1675 	if (ret < 0)
1676 		goto out_sched;
1677 
1678 
1679 	gpu->drm = drm;
1680 	gpu->fence_context = dma_fence_context_alloc(1);
1681 	idr_init(&gpu->fence_idr);
1682 	spin_lock_init(&gpu->fence_spinlock);
1683 
1684 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1685 	init_waitqueue_head(&gpu->fence_event);
1686 
1687 	priv->gpu[priv->num_gpus++] = gpu;
1688 
1689 	pm_runtime_mark_last_busy(gpu->dev);
1690 	pm_runtime_put_autosuspend(gpu->dev);
1691 
1692 	return 0;
1693 
1694 out_sched:
1695 	etnaviv_sched_fini(gpu);
1696 
1697 out_workqueue:
1698 	destroy_workqueue(gpu->wq);
1699 
1700 out_thermal:
1701 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1702 		thermal_cooling_device_unregister(gpu->cooling);
1703 
1704 	return ret;
1705 }
1706 
1707 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1708 	void *data)
1709 {
1710 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1711 
1712 	DBG("%s", dev_name(gpu->dev));
1713 
1714 	flush_workqueue(gpu->wq);
1715 	destroy_workqueue(gpu->wq);
1716 
1717 	etnaviv_sched_fini(gpu);
1718 
1719 #ifdef CONFIG_PM
1720 	pm_runtime_get_sync(gpu->dev);
1721 	pm_runtime_put_sync_suspend(gpu->dev);
1722 #else
1723 	etnaviv_gpu_hw_suspend(gpu);
1724 #endif
1725 
1726 	if (gpu->initialized) {
1727 		etnaviv_cmdbuf_free(&gpu->buffer);
1728 		etnaviv_iommu_global_fini(gpu);
1729 		gpu->initialized = false;
1730 	}
1731 
1732 	gpu->drm = NULL;
1733 	idr_destroy(&gpu->fence_idr);
1734 
1735 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1736 		thermal_cooling_device_unregister(gpu->cooling);
1737 	gpu->cooling = NULL;
1738 }
1739 
1740 static const struct component_ops gpu_ops = {
1741 	.bind = etnaviv_gpu_bind,
1742 	.unbind = etnaviv_gpu_unbind,
1743 };
1744 
1745 static const struct of_device_id etnaviv_gpu_match[] = {
1746 	{
1747 		.compatible = "vivante,gc"
1748 	},
1749 	{ /* sentinel */ }
1750 };
1751 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1752 
1753 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1754 {
1755 	struct device *dev = &pdev->dev;
1756 	struct etnaviv_gpu *gpu;
1757 	int err;
1758 
1759 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1760 	if (!gpu)
1761 		return -ENOMEM;
1762 
1763 	gpu->dev = &pdev->dev;
1764 	mutex_init(&gpu->lock);
1765 	mutex_init(&gpu->fence_lock);
1766 
1767 	/* Map registers: */
1768 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1769 	if (IS_ERR(gpu->mmio))
1770 		return PTR_ERR(gpu->mmio);
1771 
1772 	/* Get Interrupt: */
1773 	gpu->irq = platform_get_irq(pdev, 0);
1774 	if (gpu->irq < 0) {
1775 		dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1776 		return gpu->irq;
1777 	}
1778 
1779 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1780 			       dev_name(gpu->dev), gpu);
1781 	if (err) {
1782 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1783 		return err;
1784 	}
1785 
1786 	/* Get Clocks: */
1787 	gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1788 	DBG("clk_reg: %p", gpu->clk_reg);
1789 	if (IS_ERR(gpu->clk_reg))
1790 		return PTR_ERR(gpu->clk_reg);
1791 
1792 	gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1793 	DBG("clk_bus: %p", gpu->clk_bus);
1794 	if (IS_ERR(gpu->clk_bus))
1795 		return PTR_ERR(gpu->clk_bus);
1796 
1797 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1798 	DBG("clk_core: %p", gpu->clk_core);
1799 	if (IS_ERR(gpu->clk_core))
1800 		return PTR_ERR(gpu->clk_core);
1801 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1802 
1803 	gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1804 	DBG("clk_shader: %p", gpu->clk_shader);
1805 	if (IS_ERR(gpu->clk_shader))
1806 		return PTR_ERR(gpu->clk_shader);
1807 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1808 
1809 	/* TODO: figure out max mapped size */
1810 	dev_set_drvdata(dev, gpu);
1811 
1812 	/*
1813 	 * We treat the device as initially suspended.  The runtime PM
1814 	 * autosuspend delay is rather arbitary: no measurements have
1815 	 * yet been performed to determine an appropriate value.
1816 	 */
1817 	pm_runtime_use_autosuspend(gpu->dev);
1818 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1819 	pm_runtime_enable(gpu->dev);
1820 
1821 	err = component_add(&pdev->dev, &gpu_ops);
1822 	if (err < 0) {
1823 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1824 		return err;
1825 	}
1826 
1827 	return 0;
1828 }
1829 
1830 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1831 {
1832 	component_del(&pdev->dev, &gpu_ops);
1833 	pm_runtime_disable(&pdev->dev);
1834 	return 0;
1835 }
1836 
1837 #ifdef CONFIG_PM
1838 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1839 {
1840 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1841 	u32 idle, mask;
1842 
1843 	/* If there are any jobs in the HW queue, we're not idle */
1844 	if (atomic_read(&gpu->sched.hw_rq_count))
1845 		return -EBUSY;
1846 
1847 	/* Check whether the hardware (except FE and MC) is idle */
1848 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1849 				  VIVS_HI_IDLE_STATE_MC);
1850 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1851 	if (idle != mask) {
1852 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1853 				     idle);
1854 		return -EBUSY;
1855 	}
1856 
1857 	return etnaviv_gpu_hw_suspend(gpu);
1858 }
1859 
1860 static int etnaviv_gpu_rpm_resume(struct device *dev)
1861 {
1862 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1863 	int ret;
1864 
1865 	ret = etnaviv_gpu_clk_enable(gpu);
1866 	if (ret)
1867 		return ret;
1868 
1869 	/* Re-initialise the basic hardware state */
1870 	if (gpu->drm && gpu->initialized) {
1871 		ret = etnaviv_gpu_hw_resume(gpu);
1872 		if (ret) {
1873 			etnaviv_gpu_clk_disable(gpu);
1874 			return ret;
1875 		}
1876 	}
1877 
1878 	return 0;
1879 }
1880 #endif
1881 
1882 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1883 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1884 			   NULL)
1885 };
1886 
1887 struct platform_driver etnaviv_gpu_driver = {
1888 	.driver = {
1889 		.name = "etnaviv-gpu",
1890 		.owner = THIS_MODULE,
1891 		.pm = &etnaviv_gpu_pm_ops,
1892 		.of_match_table = etnaviv_gpu_match,
1893 	},
1894 	.probe = etnaviv_gpu_platform_probe,
1895 	.remove = etnaviv_gpu_platform_remove,
1896 	.id_table = gpu_ids,
1897 };
1898