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 
etnaviv_gpu_get_param(struct etnaviv_gpu * gpu,u32 param,u64 * value)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 
etnaviv_hw_specs(struct etnaviv_gpu * gpu)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 
etnaviv_hw_identify(struct etnaviv_gpu * gpu)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 	/* These models/revisions don't have the 2D pipe bit */
420 	if ((gpu->identity.model == chipModel_GC500 &&
421 	     gpu->identity.revision <= 2) ||
422 	    gpu->identity.model == chipModel_GC300)
423 		gpu->identity.features |= chipFeatures_PIPE_2D;
424 
425 	if ((gpu->identity.model == chipModel_GC500 &&
426 	     gpu->identity.revision < 2) ||
427 	    (gpu->identity.model == chipModel_GC300 &&
428 	     gpu->identity.revision < 0x2000)) {
429 
430 		/*
431 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
432 		 * registers.
433 		 */
434 		gpu->identity.minor_features0 = 0;
435 		gpu->identity.minor_features1 = 0;
436 		gpu->identity.minor_features2 = 0;
437 		gpu->identity.minor_features3 = 0;
438 		gpu->identity.minor_features4 = 0;
439 		gpu->identity.minor_features5 = 0;
440 	} else
441 		gpu->identity.minor_features0 =
442 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
443 
444 	if (gpu->identity.minor_features0 &
445 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
446 		gpu->identity.minor_features1 =
447 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
448 		gpu->identity.minor_features2 =
449 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
450 		gpu->identity.minor_features3 =
451 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
452 		gpu->identity.minor_features4 =
453 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
454 		gpu->identity.minor_features5 =
455 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
456 	}
457 
458 	/* GC600/300 idle register reports zero bits where modules aren't present */
459 	if (gpu->identity.model == chipModel_GC600 ||
460 	    gpu->identity.model == chipModel_GC300)
461 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
462 				 VIVS_HI_IDLE_STATE_RA |
463 				 VIVS_HI_IDLE_STATE_SE |
464 				 VIVS_HI_IDLE_STATE_PA |
465 				 VIVS_HI_IDLE_STATE_SH |
466 				 VIVS_HI_IDLE_STATE_PE |
467 				 VIVS_HI_IDLE_STATE_DE |
468 				 VIVS_HI_IDLE_STATE_FE;
469 
470 	etnaviv_hw_specs(gpu);
471 }
472 
etnaviv_gpu_load_clock(struct etnaviv_gpu * gpu,u32 clock)473 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
474 {
475 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
476 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
477 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
478 }
479 
etnaviv_gpu_update_clock(struct etnaviv_gpu * gpu)480 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
481 {
482 	if (gpu->identity.minor_features2 &
483 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
484 		clk_set_rate(gpu->clk_core,
485 			     gpu->base_rate_core >> gpu->freq_scale);
486 		clk_set_rate(gpu->clk_shader,
487 			     gpu->base_rate_shader >> gpu->freq_scale);
488 	} else {
489 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
490 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
491 
492 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
493 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
494 		etnaviv_gpu_load_clock(gpu, clock);
495 	}
496 }
497 
etnaviv_hw_reset(struct etnaviv_gpu * gpu)498 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
499 {
500 	u32 control, idle;
501 	unsigned long timeout;
502 	bool failed = true;
503 
504 	/* We hope that the GPU resets in under one second */
505 	timeout = jiffies + msecs_to_jiffies(1000);
506 
507 	while (time_is_after_jiffies(timeout)) {
508 		/* enable clock */
509 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
510 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
511 		etnaviv_gpu_load_clock(gpu, control);
512 
513 		/* isolate the GPU. */
514 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
515 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
516 
517 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
518 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
519 			          VIVS_MMUv2_AHB_CONTROL_RESET);
520 		} else {
521 			/* set soft reset. */
522 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
523 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
524 		}
525 
526 		/* wait for reset. */
527 		usleep_range(10, 20);
528 
529 		/* reset soft reset bit. */
530 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
531 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
532 
533 		/* reset GPU isolation. */
534 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
535 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
536 
537 		/* read idle register. */
538 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
539 
540 		/* try resetting again if FE is not idle */
541 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
542 			dev_dbg(gpu->dev, "FE is not idle\n");
543 			continue;
544 		}
545 
546 		/* read reset register. */
547 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
548 
549 		/* is the GPU idle? */
550 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
551 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
552 			dev_dbg(gpu->dev, "GPU is not idle\n");
553 			continue;
554 		}
555 
556 		/* disable debug registers, as they are not normally needed */
557 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
558 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
559 
560 		failed = false;
561 		break;
562 	}
563 
564 	if (failed) {
565 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
566 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
567 
568 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
569 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
570 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
571 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
572 
573 		return -EBUSY;
574 	}
575 
576 	/* We rely on the GPU running, so program the clock */
577 	etnaviv_gpu_update_clock(gpu);
578 
579 	gpu->fe_running = false;
580 	gpu->exec_state = -1;
581 	if (gpu->mmu_context)
582 		etnaviv_iommu_context_put(gpu->mmu_context);
583 	gpu->mmu_context = NULL;
584 
585 	return 0;
586 }
587 
etnaviv_gpu_enable_mlcg(struct etnaviv_gpu * gpu)588 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
589 {
590 	u32 pmc, ppc;
591 
592 	/* enable clock gating */
593 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
594 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
595 
596 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
597 	if (gpu->identity.revision == 0x4301 ||
598 	    gpu->identity.revision == 0x4302)
599 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
600 
601 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
602 
603 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
604 
605 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
606 	if (gpu->identity.model >= chipModel_GC400 &&
607 	    gpu->identity.model != chipModel_GC420 &&
608 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
609 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
610 
611 	/*
612 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
613 	 * present without a bug fix.
614 	 */
615 	if (gpu->identity.revision < 0x5000 &&
616 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
617 	    !(gpu->identity.minor_features1 &
618 	      chipMinorFeatures1_DISABLE_PE_GATING))
619 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
620 
621 	if (gpu->identity.revision < 0x5422)
622 		pmc |= BIT(15); /* Unknown bit */
623 
624 	/* Disable TX clock gating on affected core revisions. */
625 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
626 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108))
627 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
628 
629 	/* Disable SE, RA and TX clock gating on affected core revisions. */
630 	if (etnaviv_is_model_rev(gpu, GC7000, 0x6202))
631 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE |
632 		       VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA |
633 		       VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
634 
635 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
636 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
637 
638 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
639 }
640 
etnaviv_gpu_start_fe(struct etnaviv_gpu * gpu,u32 address,u16 prefetch)641 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
642 {
643 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
644 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
645 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
646 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
647 
648 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
649 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
650 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
651 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
652 	}
653 
654 	gpu->fe_running = true;
655 }
656 
etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu * gpu,struct etnaviv_iommu_context * context)657 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu,
658 					  struct etnaviv_iommu_context *context)
659 {
660 	u16 prefetch;
661 	u32 address;
662 
663 	/* setup the MMU */
664 	etnaviv_iommu_restore(gpu, context);
665 
666 	/* Start command processor */
667 	prefetch = etnaviv_buffer_init(gpu);
668 	address = etnaviv_cmdbuf_get_va(&gpu->buffer,
669 					&gpu->mmu_context->cmdbuf_mapping);
670 
671 	etnaviv_gpu_start_fe(gpu, address, prefetch);
672 }
673 
etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu * gpu)674 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
675 {
676 	/*
677 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
678 	 * cannot be read, extracted from vivante kernel driver.
679 	 */
680 	u32 pulse_eater = 0x01590880;
681 
682 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
683 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
684 		pulse_eater |= BIT(23);
685 
686 	}
687 
688 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
689 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
690 		pulse_eater &= ~BIT(16);
691 		pulse_eater |= BIT(17);
692 	}
693 
694 	if ((gpu->identity.revision > 0x5420) &&
695 	    (gpu->identity.features & chipFeatures_PIPE_3D))
696 	{
697 		/* Performance fix: disable internal DFS */
698 		pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
699 		pulse_eater |= BIT(18);
700 	}
701 
702 	gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
703 }
704 
etnaviv_gpu_hw_init(struct etnaviv_gpu * gpu)705 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
706 {
707 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
708 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
709 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
710 		u32 mc_memory_debug;
711 
712 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
713 
714 		if (gpu->identity.revision == 0x5007)
715 			mc_memory_debug |= 0x0c;
716 		else
717 			mc_memory_debug |= 0x08;
718 
719 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
720 	}
721 
722 	/* enable module-level clock gating */
723 	etnaviv_gpu_enable_mlcg(gpu);
724 
725 	/*
726 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
727 	 * This is necessary to prevent the iMX6 SoC locking up.
728 	 */
729 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
730 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
731 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
732 
733 	/* GC2000 rev 5108 needs a special bus config */
734 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
735 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
736 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
737 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
738 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
739 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
740 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
741 	}
742 
743 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
744 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
745 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
746 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
747 	}
748 
749 	/* setup the pulse eater */
750 	etnaviv_gpu_setup_pulse_eater(gpu);
751 
752 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
753 }
754 
etnaviv_gpu_init(struct etnaviv_gpu * gpu)755 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
756 {
757 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
758 	dma_addr_t cmdbuf_paddr;
759 	int ret, i;
760 
761 	ret = pm_runtime_get_sync(gpu->dev);
762 	if (ret < 0) {
763 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
764 		goto pm_put;
765 	}
766 
767 	etnaviv_hw_identify(gpu);
768 
769 	if (gpu->identity.model == 0) {
770 		dev_err(gpu->dev, "Unknown GPU model\n");
771 		ret = -ENXIO;
772 		goto fail;
773 	}
774 
775 	/* Exclude VG cores with FE2.0 */
776 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
777 	    gpu->identity.features & chipFeatures_FE20) {
778 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
779 		ret = -ENXIO;
780 		goto fail;
781 	}
782 
783 	/*
784 	 * On cores with security features supported, we claim control over the
785 	 * security states.
786 	 */
787 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
788 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
789 		gpu->sec_mode = ETNA_SEC_KERNEL;
790 
791 	ret = etnaviv_hw_reset(gpu);
792 	if (ret) {
793 		dev_err(gpu->dev, "GPU reset failed\n");
794 		goto fail;
795 	}
796 
797 	ret = etnaviv_iommu_global_init(gpu);
798 	if (ret)
799 		goto fail;
800 
801 	/*
802 	 * If the GPU is part of a system with DMA addressing limitations,
803 	 * request pages for our SHM backend buffers from the DMA32 zone to
804 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
805 	 */
806 	if (dma_addressing_limited(gpu->dev))
807 		priv->shm_gfp_mask |= GFP_DMA32;
808 
809 	/* Create buffer: */
810 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
811 				  PAGE_SIZE);
812 	if (ret) {
813 		dev_err(gpu->dev, "could not create command buffer\n");
814 		goto fail;
815 	}
816 
817 	/*
818 	 * Set the GPU linear window to cover the cmdbuf region, as the GPU
819 	 * won't be able to start execution otherwise. The alignment to 128M is
820 	 * chosen arbitrarily but helps in debugging, as the MMU offset
821 	 * calculations are much more straight forward this way.
822 	 *
823 	 * On MC1.0 cores the linear window offset is ignored by the TS engine,
824 	 * leading to inconsistent memory views. Avoid using the offset on those
825 	 * cores if possible, otherwise disable the TS feature.
826 	 */
827 	cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M);
828 
829 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
830 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
831 		if (cmdbuf_paddr >= SZ_2G)
832 			priv->mmu_global->memory_base = SZ_2G;
833 		else
834 			priv->mmu_global->memory_base = cmdbuf_paddr;
835 	} else if (cmdbuf_paddr + SZ_128M >= SZ_2G) {
836 		dev_info(gpu->dev,
837 			 "Need to move linear window on MC1.0, disabling TS\n");
838 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
839 		priv->mmu_global->memory_base = SZ_2G;
840 	}
841 
842 	/* Setup event management */
843 	spin_lock_init(&gpu->event_spinlock);
844 	init_completion(&gpu->event_free);
845 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
846 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
847 		complete(&gpu->event_free);
848 
849 	/* Now program the hardware */
850 	mutex_lock(&gpu->lock);
851 	etnaviv_gpu_hw_init(gpu);
852 	mutex_unlock(&gpu->lock);
853 
854 	pm_runtime_mark_last_busy(gpu->dev);
855 	pm_runtime_put_autosuspend(gpu->dev);
856 
857 	gpu->initialized = true;
858 
859 	return 0;
860 
861 fail:
862 	pm_runtime_mark_last_busy(gpu->dev);
863 pm_put:
864 	pm_runtime_put_autosuspend(gpu->dev);
865 
866 	return ret;
867 }
868 
869 #ifdef CONFIG_DEBUG_FS
870 struct dma_debug {
871 	u32 address[2];
872 	u32 state[2];
873 };
874 
verify_dma(struct etnaviv_gpu * gpu,struct dma_debug * debug)875 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
876 {
877 	u32 i;
878 
879 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
880 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
881 
882 	for (i = 0; i < 500; i++) {
883 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
884 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
885 
886 		if (debug->address[0] != debug->address[1])
887 			break;
888 
889 		if (debug->state[0] != debug->state[1])
890 			break;
891 	}
892 }
893 
etnaviv_gpu_debugfs(struct etnaviv_gpu * gpu,struct seq_file * m)894 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
895 {
896 	struct dma_debug debug;
897 	u32 dma_lo, dma_hi, axi, idle;
898 	int ret;
899 
900 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
901 
902 	ret = pm_runtime_get_sync(gpu->dev);
903 	if (ret < 0)
904 		goto pm_put;
905 
906 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
907 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
908 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
909 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
910 
911 	verify_dma(gpu, &debug);
912 
913 	seq_puts(m, "\tidentity\n");
914 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
915 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
916 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
917 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
918 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
919 
920 	seq_puts(m, "\tfeatures\n");
921 	seq_printf(m, "\t major_features: 0x%08x\n",
922 		   gpu->identity.features);
923 	seq_printf(m, "\t minor_features0: 0x%08x\n",
924 		   gpu->identity.minor_features0);
925 	seq_printf(m, "\t minor_features1: 0x%08x\n",
926 		   gpu->identity.minor_features1);
927 	seq_printf(m, "\t minor_features2: 0x%08x\n",
928 		   gpu->identity.minor_features2);
929 	seq_printf(m, "\t minor_features3: 0x%08x\n",
930 		   gpu->identity.minor_features3);
931 	seq_printf(m, "\t minor_features4: 0x%08x\n",
932 		   gpu->identity.minor_features4);
933 	seq_printf(m, "\t minor_features5: 0x%08x\n",
934 		   gpu->identity.minor_features5);
935 	seq_printf(m, "\t minor_features6: 0x%08x\n",
936 		   gpu->identity.minor_features6);
937 	seq_printf(m, "\t minor_features7: 0x%08x\n",
938 		   gpu->identity.minor_features7);
939 	seq_printf(m, "\t minor_features8: 0x%08x\n",
940 		   gpu->identity.minor_features8);
941 	seq_printf(m, "\t minor_features9: 0x%08x\n",
942 		   gpu->identity.minor_features9);
943 	seq_printf(m, "\t minor_features10: 0x%08x\n",
944 		   gpu->identity.minor_features10);
945 	seq_printf(m, "\t minor_features11: 0x%08x\n",
946 		   gpu->identity.minor_features11);
947 
948 	seq_puts(m, "\tspecs\n");
949 	seq_printf(m, "\t stream_count:  %d\n",
950 			gpu->identity.stream_count);
951 	seq_printf(m, "\t register_max: %d\n",
952 			gpu->identity.register_max);
953 	seq_printf(m, "\t thread_count: %d\n",
954 			gpu->identity.thread_count);
955 	seq_printf(m, "\t vertex_cache_size: %d\n",
956 			gpu->identity.vertex_cache_size);
957 	seq_printf(m, "\t shader_core_count: %d\n",
958 			gpu->identity.shader_core_count);
959 	seq_printf(m, "\t pixel_pipes: %d\n",
960 			gpu->identity.pixel_pipes);
961 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
962 			gpu->identity.vertex_output_buffer_size);
963 	seq_printf(m, "\t buffer_size: %d\n",
964 			gpu->identity.buffer_size);
965 	seq_printf(m, "\t instruction_count: %d\n",
966 			gpu->identity.instruction_count);
967 	seq_printf(m, "\t num_constants: %d\n",
968 			gpu->identity.num_constants);
969 	seq_printf(m, "\t varyings_count: %d\n",
970 			gpu->identity.varyings_count);
971 
972 	seq_printf(m, "\taxi: 0x%08x\n", axi);
973 	seq_printf(m, "\tidle: 0x%08x\n", idle);
974 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
975 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
976 		seq_puts(m, "\t FE is not idle\n");
977 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
978 		seq_puts(m, "\t DE is not idle\n");
979 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
980 		seq_puts(m, "\t PE is not idle\n");
981 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
982 		seq_puts(m, "\t SH is not idle\n");
983 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
984 		seq_puts(m, "\t PA is not idle\n");
985 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
986 		seq_puts(m, "\t SE is not idle\n");
987 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
988 		seq_puts(m, "\t RA is not idle\n");
989 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
990 		seq_puts(m, "\t TX is not idle\n");
991 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
992 		seq_puts(m, "\t VG is not idle\n");
993 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
994 		seq_puts(m, "\t IM is not idle\n");
995 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
996 		seq_puts(m, "\t FP is not idle\n");
997 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
998 		seq_puts(m, "\t TS is not idle\n");
999 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
1000 		seq_puts(m, "\t BL is not idle\n");
1001 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
1002 		seq_puts(m, "\t ASYNCFE is not idle\n");
1003 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
1004 		seq_puts(m, "\t MC is not idle\n");
1005 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
1006 		seq_puts(m, "\t PPA is not idle\n");
1007 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
1008 		seq_puts(m, "\t WD is not idle\n");
1009 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
1010 		seq_puts(m, "\t NN is not idle\n");
1011 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
1012 		seq_puts(m, "\t TP is not idle\n");
1013 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
1014 		seq_puts(m, "\t AXI low power mode\n");
1015 
1016 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
1017 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1018 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1019 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1020 
1021 		seq_puts(m, "\tMC\n");
1022 		seq_printf(m, "\t read0: 0x%08x\n", read0);
1023 		seq_printf(m, "\t read1: 0x%08x\n", read1);
1024 		seq_printf(m, "\t write: 0x%08x\n", write);
1025 	}
1026 
1027 	seq_puts(m, "\tDMA ");
1028 
1029 	if (debug.address[0] == debug.address[1] &&
1030 	    debug.state[0] == debug.state[1]) {
1031 		seq_puts(m, "seems to be stuck\n");
1032 	} else if (debug.address[0] == debug.address[1]) {
1033 		seq_puts(m, "address is constant\n");
1034 	} else {
1035 		seq_puts(m, "is running\n");
1036 	}
1037 
1038 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1039 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1040 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1041 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1042 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1043 		   dma_lo, dma_hi);
1044 
1045 	ret = 0;
1046 
1047 	pm_runtime_mark_last_busy(gpu->dev);
1048 pm_put:
1049 	pm_runtime_put_autosuspend(gpu->dev);
1050 
1051 	return ret;
1052 }
1053 #endif
1054 
etnaviv_gpu_recover_hang(struct etnaviv_gpu * gpu)1055 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
1056 {
1057 	unsigned int i;
1058 
1059 	dev_err(gpu->dev, "recover hung GPU!\n");
1060 
1061 	if (pm_runtime_get_sync(gpu->dev) < 0)
1062 		goto pm_put;
1063 
1064 	mutex_lock(&gpu->lock);
1065 
1066 	etnaviv_hw_reset(gpu);
1067 
1068 	/* complete all events, the GPU won't do it after the reset */
1069 	spin_lock(&gpu->event_spinlock);
1070 	for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1071 		complete(&gpu->event_free);
1072 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1073 	spin_unlock(&gpu->event_spinlock);
1074 
1075 	etnaviv_gpu_hw_init(gpu);
1076 
1077 	mutex_unlock(&gpu->lock);
1078 	pm_runtime_mark_last_busy(gpu->dev);
1079 pm_put:
1080 	pm_runtime_put_autosuspend(gpu->dev);
1081 }
1082 
1083 /* fence object management */
1084 struct etnaviv_fence {
1085 	struct etnaviv_gpu *gpu;
1086 	struct dma_fence base;
1087 };
1088 
to_etnaviv_fence(struct dma_fence * fence)1089 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1090 {
1091 	return container_of(fence, struct etnaviv_fence, base);
1092 }
1093 
etnaviv_fence_get_driver_name(struct dma_fence * fence)1094 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1095 {
1096 	return "etnaviv";
1097 }
1098 
etnaviv_fence_get_timeline_name(struct dma_fence * fence)1099 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1100 {
1101 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1102 
1103 	return dev_name(f->gpu->dev);
1104 }
1105 
etnaviv_fence_signaled(struct dma_fence * fence)1106 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1107 {
1108 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1109 
1110 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1111 }
1112 
etnaviv_fence_release(struct dma_fence * fence)1113 static void etnaviv_fence_release(struct dma_fence *fence)
1114 {
1115 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1116 
1117 	kfree_rcu(f, base.rcu);
1118 }
1119 
1120 static const struct dma_fence_ops etnaviv_fence_ops = {
1121 	.get_driver_name = etnaviv_fence_get_driver_name,
1122 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1123 	.signaled = etnaviv_fence_signaled,
1124 	.release = etnaviv_fence_release,
1125 };
1126 
etnaviv_gpu_fence_alloc(struct etnaviv_gpu * gpu)1127 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1128 {
1129 	struct etnaviv_fence *f;
1130 
1131 	/*
1132 	 * GPU lock must already be held, otherwise fence completion order might
1133 	 * not match the seqno order assigned here.
1134 	 */
1135 	lockdep_assert_held(&gpu->lock);
1136 
1137 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1138 	if (!f)
1139 		return NULL;
1140 
1141 	f->gpu = gpu;
1142 
1143 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1144 		       gpu->fence_context, ++gpu->next_fence);
1145 
1146 	return &f->base;
1147 }
1148 
1149 /* returns true if fence a comes after fence b */
fence_after(u32 a,u32 b)1150 static inline bool fence_after(u32 a, u32 b)
1151 {
1152 	return (s32)(a - b) > 0;
1153 }
1154 
1155 /*
1156  * event management:
1157  */
1158 
event_alloc(struct etnaviv_gpu * gpu,unsigned nr_events,unsigned int * events)1159 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1160 	unsigned int *events)
1161 {
1162 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1163 	unsigned i, acquired = 0;
1164 
1165 	for (i = 0; i < nr_events; i++) {
1166 		unsigned long ret;
1167 
1168 		ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1169 
1170 		if (!ret) {
1171 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1172 			goto out;
1173 		}
1174 
1175 		acquired++;
1176 		timeout = ret;
1177 	}
1178 
1179 	spin_lock(&gpu->event_spinlock);
1180 
1181 	for (i = 0; i < nr_events; i++) {
1182 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1183 
1184 		events[i] = event;
1185 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1186 		set_bit(event, gpu->event_bitmap);
1187 	}
1188 
1189 	spin_unlock(&gpu->event_spinlock);
1190 
1191 	return 0;
1192 
1193 out:
1194 	for (i = 0; i < acquired; i++)
1195 		complete(&gpu->event_free);
1196 
1197 	return -EBUSY;
1198 }
1199 
event_free(struct etnaviv_gpu * gpu,unsigned int event)1200 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1201 {
1202 	if (!test_bit(event, gpu->event_bitmap)) {
1203 		dev_warn(gpu->dev, "event %u is already marked as free",
1204 			 event);
1205 	} else {
1206 		clear_bit(event, gpu->event_bitmap);
1207 		complete(&gpu->event_free);
1208 	}
1209 }
1210 
1211 /*
1212  * Cmdstream submission/retirement:
1213  */
etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu * gpu,u32 id,struct drm_etnaviv_timespec * timeout)1214 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1215 	u32 id, struct drm_etnaviv_timespec *timeout)
1216 {
1217 	struct dma_fence *fence;
1218 	int ret;
1219 
1220 	/*
1221 	 * Look up the fence and take a reference. We might still find a fence
1222 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1223 	 * pretends we didn't find a fence in that case.
1224 	 */
1225 	rcu_read_lock();
1226 	fence = idr_find(&gpu->fence_idr, id);
1227 	if (fence)
1228 		fence = dma_fence_get_rcu(fence);
1229 	rcu_read_unlock();
1230 
1231 	if (!fence)
1232 		return 0;
1233 
1234 	if (!timeout) {
1235 		/* No timeout was requested: just test for completion */
1236 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1237 	} else {
1238 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1239 
1240 		ret = dma_fence_wait_timeout(fence, true, remaining);
1241 		if (ret == 0)
1242 			ret = -ETIMEDOUT;
1243 		else if (ret != -ERESTARTSYS)
1244 			ret = 0;
1245 
1246 	}
1247 
1248 	dma_fence_put(fence);
1249 	return ret;
1250 }
1251 
1252 /*
1253  * Wait for an object to become inactive.  This, on it's own, is not race
1254  * free: the object is moved by the scheduler off the active list, and
1255  * then the iova is put.  Moreover, the object could be re-submitted just
1256  * after we notice that it's become inactive.
1257  *
1258  * Although the retirement happens under the gpu lock, we don't want to hold
1259  * that lock in this function while waiting.
1260  */
etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu * gpu,struct etnaviv_gem_object * etnaviv_obj,struct drm_etnaviv_timespec * timeout)1261 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1262 	struct etnaviv_gem_object *etnaviv_obj,
1263 	struct drm_etnaviv_timespec *timeout)
1264 {
1265 	unsigned long remaining;
1266 	long ret;
1267 
1268 	if (!timeout)
1269 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1270 
1271 	remaining = etnaviv_timeout_to_jiffies(timeout);
1272 
1273 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1274 					       !is_active(etnaviv_obj),
1275 					       remaining);
1276 	if (ret > 0)
1277 		return 0;
1278 	else if (ret == -ERESTARTSYS)
1279 		return -ERESTARTSYS;
1280 	else
1281 		return -ETIMEDOUT;
1282 }
1283 
sync_point_perfmon_sample(struct etnaviv_gpu * gpu,struct etnaviv_event * event,unsigned int flags)1284 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1285 	struct etnaviv_event *event, unsigned int flags)
1286 {
1287 	const struct etnaviv_gem_submit *submit = event->submit;
1288 	unsigned int i;
1289 
1290 	for (i = 0; i < submit->nr_pmrs; i++) {
1291 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1292 
1293 		if (pmr->flags == flags)
1294 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1295 	}
1296 }
1297 
sync_point_perfmon_sample_pre(struct etnaviv_gpu * gpu,struct etnaviv_event * event)1298 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1299 	struct etnaviv_event *event)
1300 {
1301 	u32 val;
1302 
1303 	/* disable clock gating */
1304 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1305 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1306 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1307 
1308 	/* enable debug register */
1309 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1310 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1311 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1312 
1313 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1314 }
1315 
sync_point_perfmon_sample_post(struct etnaviv_gpu * gpu,struct etnaviv_event * event)1316 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1317 	struct etnaviv_event *event)
1318 {
1319 	const struct etnaviv_gem_submit *submit = event->submit;
1320 	unsigned int i;
1321 	u32 val;
1322 
1323 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1324 
1325 	for (i = 0; i < submit->nr_pmrs; i++) {
1326 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1327 
1328 		*pmr->bo_vma = pmr->sequence;
1329 	}
1330 
1331 	/* disable debug register */
1332 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1333 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1334 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1335 
1336 	/* enable clock gating */
1337 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1338 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1339 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1340 }
1341 
1342 
1343 /* add bo's to gpu's ring, and kick gpu: */
etnaviv_gpu_submit(struct etnaviv_gem_submit * submit)1344 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1345 {
1346 	struct etnaviv_gpu *gpu = submit->gpu;
1347 	struct dma_fence *gpu_fence;
1348 	unsigned int i, nr_events = 1, event[3];
1349 	int ret;
1350 
1351 	if (!submit->runtime_resumed) {
1352 		ret = pm_runtime_get_sync(gpu->dev);
1353 		if (ret < 0) {
1354 			pm_runtime_put_noidle(gpu->dev);
1355 			return NULL;
1356 		}
1357 		submit->runtime_resumed = true;
1358 	}
1359 
1360 	/*
1361 	 * if there are performance monitor requests we need to have
1362 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1363 	 *   requests.
1364 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1365 	 *   and update the sequence number for userspace.
1366 	 */
1367 	if (submit->nr_pmrs)
1368 		nr_events = 3;
1369 
1370 	ret = event_alloc(gpu, nr_events, event);
1371 	if (ret) {
1372 		DRM_ERROR("no free events\n");
1373 		pm_runtime_put_noidle(gpu->dev);
1374 		return NULL;
1375 	}
1376 
1377 	mutex_lock(&gpu->lock);
1378 
1379 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1380 	if (!gpu_fence) {
1381 		for (i = 0; i < nr_events; i++)
1382 			event_free(gpu, event[i]);
1383 
1384 		goto out_unlock;
1385 	}
1386 
1387 	if (!gpu->fe_running)
1388 		etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context);
1389 
1390 	if (submit->prev_mmu_context)
1391 		etnaviv_iommu_context_put(submit->prev_mmu_context);
1392 	submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context);
1393 
1394 	if (submit->nr_pmrs) {
1395 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1396 		kref_get(&submit->refcount);
1397 		gpu->event[event[1]].submit = submit;
1398 		etnaviv_sync_point_queue(gpu, event[1]);
1399 	}
1400 
1401 	gpu->event[event[0]].fence = gpu_fence;
1402 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1403 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1404 			     event[0], &submit->cmdbuf);
1405 
1406 	if (submit->nr_pmrs) {
1407 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1408 		kref_get(&submit->refcount);
1409 		gpu->event[event[2]].submit = submit;
1410 		etnaviv_sync_point_queue(gpu, event[2]);
1411 	}
1412 
1413 out_unlock:
1414 	mutex_unlock(&gpu->lock);
1415 
1416 	return gpu_fence;
1417 }
1418 
sync_point_worker(struct work_struct * work)1419 static void sync_point_worker(struct work_struct *work)
1420 {
1421 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1422 					       sync_point_work);
1423 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1424 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1425 
1426 	event->sync_point(gpu, event);
1427 	etnaviv_submit_put(event->submit);
1428 	event_free(gpu, gpu->sync_point_event);
1429 
1430 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1431 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1432 }
1433 
dump_mmu_fault(struct etnaviv_gpu * gpu)1434 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1435 {
1436 	u32 status_reg, status;
1437 	int i;
1438 
1439 	if (gpu->sec_mode == ETNA_SEC_NONE)
1440 		status_reg = VIVS_MMUv2_STATUS;
1441 	else
1442 		status_reg = VIVS_MMUv2_SEC_STATUS;
1443 
1444 	status = gpu_read(gpu, status_reg);
1445 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1446 
1447 	for (i = 0; i < 4; i++) {
1448 		u32 address_reg;
1449 
1450 		if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1451 			continue;
1452 
1453 		if (gpu->sec_mode == ETNA_SEC_NONE)
1454 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1455 		else
1456 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1457 
1458 		dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1459 				    gpu_read(gpu, address_reg));
1460 	}
1461 }
1462 
irq_handler(int irq,void * data)1463 static irqreturn_t irq_handler(int irq, void *data)
1464 {
1465 	struct etnaviv_gpu *gpu = data;
1466 	irqreturn_t ret = IRQ_NONE;
1467 
1468 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1469 
1470 	if (intr != 0) {
1471 		int event;
1472 
1473 		pm_runtime_mark_last_busy(gpu->dev);
1474 
1475 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1476 
1477 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1478 			dev_err(gpu->dev, "AXI bus error\n");
1479 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1480 		}
1481 
1482 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1483 			dump_mmu_fault(gpu);
1484 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1485 		}
1486 
1487 		while ((event = ffs(intr)) != 0) {
1488 			struct dma_fence *fence;
1489 
1490 			event -= 1;
1491 
1492 			intr &= ~(1 << event);
1493 
1494 			dev_dbg(gpu->dev, "event %u\n", event);
1495 
1496 			if (gpu->event[event].sync_point) {
1497 				gpu->sync_point_event = event;
1498 				queue_work(gpu->wq, &gpu->sync_point_work);
1499 			}
1500 
1501 			fence = gpu->event[event].fence;
1502 			if (!fence)
1503 				continue;
1504 
1505 			gpu->event[event].fence = NULL;
1506 
1507 			/*
1508 			 * Events can be processed out of order.  Eg,
1509 			 * - allocate and queue event 0
1510 			 * - allocate event 1
1511 			 * - event 0 completes, we process it
1512 			 * - allocate and queue event 0
1513 			 * - event 1 and event 0 complete
1514 			 * we can end up processing event 0 first, then 1.
1515 			 */
1516 			if (fence_after(fence->seqno, gpu->completed_fence))
1517 				gpu->completed_fence = fence->seqno;
1518 			dma_fence_signal(fence);
1519 
1520 			event_free(gpu, event);
1521 		}
1522 
1523 		ret = IRQ_HANDLED;
1524 	}
1525 
1526 	return ret;
1527 }
1528 
etnaviv_gpu_clk_enable(struct etnaviv_gpu * gpu)1529 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1530 {
1531 	int ret;
1532 
1533 	ret = clk_prepare_enable(gpu->clk_reg);
1534 	if (ret)
1535 		return ret;
1536 
1537 	ret = clk_prepare_enable(gpu->clk_bus);
1538 	if (ret)
1539 		goto disable_clk_reg;
1540 
1541 	ret = clk_prepare_enable(gpu->clk_core);
1542 	if (ret)
1543 		goto disable_clk_bus;
1544 
1545 	ret = clk_prepare_enable(gpu->clk_shader);
1546 	if (ret)
1547 		goto disable_clk_core;
1548 
1549 	return 0;
1550 
1551 disable_clk_core:
1552 	clk_disable_unprepare(gpu->clk_core);
1553 disable_clk_bus:
1554 	clk_disable_unprepare(gpu->clk_bus);
1555 disable_clk_reg:
1556 	clk_disable_unprepare(gpu->clk_reg);
1557 
1558 	return ret;
1559 }
1560 
etnaviv_gpu_clk_disable(struct etnaviv_gpu * gpu)1561 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1562 {
1563 	clk_disable_unprepare(gpu->clk_shader);
1564 	clk_disable_unprepare(gpu->clk_core);
1565 	clk_disable_unprepare(gpu->clk_bus);
1566 	clk_disable_unprepare(gpu->clk_reg);
1567 
1568 	return 0;
1569 }
1570 
etnaviv_gpu_wait_idle(struct etnaviv_gpu * gpu,unsigned int timeout_ms)1571 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1572 {
1573 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1574 
1575 	do {
1576 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1577 
1578 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1579 			return 0;
1580 
1581 		if (time_is_before_jiffies(timeout)) {
1582 			dev_warn(gpu->dev,
1583 				 "timed out waiting for idle: idle=0x%x\n",
1584 				 idle);
1585 			return -ETIMEDOUT;
1586 		}
1587 
1588 		udelay(5);
1589 	} while (1);
1590 }
1591 
etnaviv_gpu_hw_suspend(struct etnaviv_gpu * gpu)1592 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1593 {
1594 	if (gpu->initialized && gpu->fe_running) {
1595 		/* Replace the last WAIT with END */
1596 		mutex_lock(&gpu->lock);
1597 		etnaviv_buffer_end(gpu);
1598 		mutex_unlock(&gpu->lock);
1599 
1600 		/*
1601 		 * We know that only the FE is busy here, this should
1602 		 * happen quickly (as the WAIT is only 200 cycles).  If
1603 		 * we fail, just warn and continue.
1604 		 */
1605 		etnaviv_gpu_wait_idle(gpu, 100);
1606 
1607 		gpu->fe_running = false;
1608 	}
1609 
1610 	gpu->exec_state = -1;
1611 
1612 	return etnaviv_gpu_clk_disable(gpu);
1613 }
1614 
1615 #ifdef CONFIG_PM
etnaviv_gpu_hw_resume(struct etnaviv_gpu * gpu)1616 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1617 {
1618 	int ret;
1619 
1620 	ret = mutex_lock_killable(&gpu->lock);
1621 	if (ret)
1622 		return ret;
1623 
1624 	etnaviv_gpu_update_clock(gpu);
1625 	etnaviv_gpu_hw_init(gpu);
1626 
1627 	mutex_unlock(&gpu->lock);
1628 
1629 	return 0;
1630 }
1631 #endif
1632 
1633 static int
etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device * cdev,unsigned long * state)1634 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1635 				  unsigned long *state)
1636 {
1637 	*state = 6;
1638 
1639 	return 0;
1640 }
1641 
1642 static int
etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device * cdev,unsigned long * state)1643 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1644 				  unsigned long *state)
1645 {
1646 	struct etnaviv_gpu *gpu = cdev->devdata;
1647 
1648 	*state = gpu->freq_scale;
1649 
1650 	return 0;
1651 }
1652 
1653 static int
etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device * cdev,unsigned long state)1654 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1655 				  unsigned long state)
1656 {
1657 	struct etnaviv_gpu *gpu = cdev->devdata;
1658 
1659 	mutex_lock(&gpu->lock);
1660 	gpu->freq_scale = state;
1661 	if (!pm_runtime_suspended(gpu->dev))
1662 		etnaviv_gpu_update_clock(gpu);
1663 	mutex_unlock(&gpu->lock);
1664 
1665 	return 0;
1666 }
1667 
1668 static const struct thermal_cooling_device_ops cooling_ops = {
1669 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1670 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1671 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1672 };
1673 
etnaviv_gpu_bind(struct device * dev,struct device * master,void * data)1674 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1675 	void *data)
1676 {
1677 	struct drm_device *drm = data;
1678 	struct etnaviv_drm_private *priv = drm->dev_private;
1679 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1680 	int ret;
1681 
1682 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1683 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1684 				(char *)dev_name(dev), gpu, &cooling_ops);
1685 		if (IS_ERR(gpu->cooling))
1686 			return PTR_ERR(gpu->cooling);
1687 	}
1688 
1689 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1690 	if (!gpu->wq) {
1691 		ret = -ENOMEM;
1692 		goto out_thermal;
1693 	}
1694 
1695 	ret = etnaviv_sched_init(gpu);
1696 	if (ret)
1697 		goto out_workqueue;
1698 
1699 #ifdef CONFIG_PM
1700 	ret = pm_runtime_get_sync(gpu->dev);
1701 #else
1702 	ret = etnaviv_gpu_clk_enable(gpu);
1703 #endif
1704 	if (ret < 0)
1705 		goto out_sched;
1706 
1707 
1708 	gpu->drm = drm;
1709 	gpu->fence_context = dma_fence_context_alloc(1);
1710 	idr_init(&gpu->fence_idr);
1711 	spin_lock_init(&gpu->fence_spinlock);
1712 
1713 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1714 	init_waitqueue_head(&gpu->fence_event);
1715 
1716 	priv->gpu[priv->num_gpus++] = gpu;
1717 
1718 	pm_runtime_mark_last_busy(gpu->dev);
1719 	pm_runtime_put_autosuspend(gpu->dev);
1720 
1721 	return 0;
1722 
1723 out_sched:
1724 	etnaviv_sched_fini(gpu);
1725 
1726 out_workqueue:
1727 	destroy_workqueue(gpu->wq);
1728 
1729 out_thermal:
1730 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1731 		thermal_cooling_device_unregister(gpu->cooling);
1732 
1733 	return ret;
1734 }
1735 
etnaviv_gpu_unbind(struct device * dev,struct device * master,void * data)1736 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1737 	void *data)
1738 {
1739 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1740 
1741 	DBG("%s", dev_name(gpu->dev));
1742 
1743 	destroy_workqueue(gpu->wq);
1744 
1745 	etnaviv_sched_fini(gpu);
1746 
1747 #ifdef CONFIG_PM
1748 	pm_runtime_get_sync(gpu->dev);
1749 	pm_runtime_put_sync_suspend(gpu->dev);
1750 #else
1751 	etnaviv_gpu_hw_suspend(gpu);
1752 #endif
1753 
1754 	if (gpu->mmu_context)
1755 		etnaviv_iommu_context_put(gpu->mmu_context);
1756 
1757 	if (gpu->initialized) {
1758 		etnaviv_cmdbuf_free(&gpu->buffer);
1759 		etnaviv_iommu_global_fini(gpu);
1760 		gpu->initialized = false;
1761 	}
1762 
1763 	gpu->drm = NULL;
1764 	idr_destroy(&gpu->fence_idr);
1765 
1766 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1767 		thermal_cooling_device_unregister(gpu->cooling);
1768 	gpu->cooling = NULL;
1769 }
1770 
1771 static const struct component_ops gpu_ops = {
1772 	.bind = etnaviv_gpu_bind,
1773 	.unbind = etnaviv_gpu_unbind,
1774 };
1775 
1776 static const struct of_device_id etnaviv_gpu_match[] = {
1777 	{
1778 		.compatible = "vivante,gc"
1779 	},
1780 	{ /* sentinel */ }
1781 };
1782 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1783 
etnaviv_gpu_platform_probe(struct platform_device * pdev)1784 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1785 {
1786 	struct device *dev = &pdev->dev;
1787 	struct etnaviv_gpu *gpu;
1788 	int err;
1789 
1790 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1791 	if (!gpu)
1792 		return -ENOMEM;
1793 
1794 	gpu->dev = &pdev->dev;
1795 	mutex_init(&gpu->lock);
1796 	mutex_init(&gpu->fence_lock);
1797 
1798 	/* Map registers: */
1799 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1800 	if (IS_ERR(gpu->mmio))
1801 		return PTR_ERR(gpu->mmio);
1802 
1803 	/* Get Interrupt: */
1804 	gpu->irq = platform_get_irq(pdev, 0);
1805 	if (gpu->irq < 0)
1806 		return gpu->irq;
1807 
1808 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1809 			       dev_name(gpu->dev), gpu);
1810 	if (err) {
1811 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1812 		return err;
1813 	}
1814 
1815 	/* Get Clocks: */
1816 	gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1817 	DBG("clk_reg: %p", gpu->clk_reg);
1818 	if (IS_ERR(gpu->clk_reg))
1819 		return PTR_ERR(gpu->clk_reg);
1820 
1821 	gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1822 	DBG("clk_bus: %p", gpu->clk_bus);
1823 	if (IS_ERR(gpu->clk_bus))
1824 		return PTR_ERR(gpu->clk_bus);
1825 
1826 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1827 	DBG("clk_core: %p", gpu->clk_core);
1828 	if (IS_ERR(gpu->clk_core))
1829 		return PTR_ERR(gpu->clk_core);
1830 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1831 
1832 	gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1833 	DBG("clk_shader: %p", gpu->clk_shader);
1834 	if (IS_ERR(gpu->clk_shader))
1835 		return PTR_ERR(gpu->clk_shader);
1836 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1837 
1838 	/* TODO: figure out max mapped size */
1839 	dev_set_drvdata(dev, gpu);
1840 
1841 	/*
1842 	 * We treat the device as initially suspended.  The runtime PM
1843 	 * autosuspend delay is rather arbitary: no measurements have
1844 	 * yet been performed to determine an appropriate value.
1845 	 */
1846 	pm_runtime_use_autosuspend(gpu->dev);
1847 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1848 	pm_runtime_enable(gpu->dev);
1849 
1850 	err = component_add(&pdev->dev, &gpu_ops);
1851 	if (err < 0) {
1852 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1853 		return err;
1854 	}
1855 
1856 	return 0;
1857 }
1858 
etnaviv_gpu_platform_remove(struct platform_device * pdev)1859 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1860 {
1861 	component_del(&pdev->dev, &gpu_ops);
1862 	pm_runtime_disable(&pdev->dev);
1863 	return 0;
1864 }
1865 
1866 #ifdef CONFIG_PM
etnaviv_gpu_rpm_suspend(struct device * dev)1867 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1868 {
1869 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1870 	u32 idle, mask;
1871 
1872 	/* If there are any jobs in the HW queue, we're not idle */
1873 	if (atomic_read(&gpu->sched.hw_rq_count))
1874 		return -EBUSY;
1875 
1876 	/* Check whether the hardware (except FE and MC) is idle */
1877 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1878 				  VIVS_HI_IDLE_STATE_MC);
1879 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1880 	if (idle != mask) {
1881 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1882 				     idle);
1883 		return -EBUSY;
1884 	}
1885 
1886 	return etnaviv_gpu_hw_suspend(gpu);
1887 }
1888 
etnaviv_gpu_rpm_resume(struct device * dev)1889 static int etnaviv_gpu_rpm_resume(struct device *dev)
1890 {
1891 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1892 	int ret;
1893 
1894 	ret = etnaviv_gpu_clk_enable(gpu);
1895 	if (ret)
1896 		return ret;
1897 
1898 	/* Re-initialise the basic hardware state */
1899 	if (gpu->drm && gpu->initialized) {
1900 		ret = etnaviv_gpu_hw_resume(gpu);
1901 		if (ret) {
1902 			etnaviv_gpu_clk_disable(gpu);
1903 			return ret;
1904 		}
1905 	}
1906 
1907 	return 0;
1908 }
1909 #endif
1910 
1911 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1912 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1913 			   NULL)
1914 };
1915 
1916 struct platform_driver etnaviv_gpu_driver = {
1917 	.driver = {
1918 		.name = "etnaviv-gpu",
1919 		.owner = THIS_MODULE,
1920 		.pm = &etnaviv_gpu_pm_ops,
1921 		.of_match_table = etnaviv_gpu_match,
1922 	},
1923 	.probe = etnaviv_gpu_platform_probe,
1924 	.remove = etnaviv_gpu_platform_remove,
1925 	.id_table = gpu_ids,
1926 };
1927