• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 
etnaviv_gpu_get_param(struct etnaviv_gpu * gpu,u32 param,u64 * value)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 
etnaviv_hw_specs(struct etnaviv_gpu * gpu)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 
etnaviv_hw_identify(struct etnaviv_gpu * gpu)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 	/* These models/revisions don't have the 2D pipe bit */
412 	if ((gpu->identity.model == chipModel_GC500 &&
413 	     gpu->identity.revision <= 2) ||
414 	    gpu->identity.model == chipModel_GC300)
415 		gpu->identity.features |= chipFeatures_PIPE_2D;
416 
417 	if ((gpu->identity.model == chipModel_GC500 &&
418 	     gpu->identity.revision < 2) ||
419 	    (gpu->identity.model == chipModel_GC300 &&
420 	     gpu->identity.revision < 0x2000)) {
421 
422 		/*
423 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
424 		 * registers.
425 		 */
426 		gpu->identity.minor_features0 = 0;
427 		gpu->identity.minor_features1 = 0;
428 		gpu->identity.minor_features2 = 0;
429 		gpu->identity.minor_features3 = 0;
430 		gpu->identity.minor_features4 = 0;
431 		gpu->identity.minor_features5 = 0;
432 	} else
433 		gpu->identity.minor_features0 =
434 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
435 
436 	if (gpu->identity.minor_features0 &
437 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
438 		gpu->identity.minor_features1 =
439 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
440 		gpu->identity.minor_features2 =
441 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
442 		gpu->identity.minor_features3 =
443 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
444 		gpu->identity.minor_features4 =
445 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
446 		gpu->identity.minor_features5 =
447 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
448 	}
449 
450 	/* GC600/300 idle register reports zero bits where modules aren't present */
451 	if (gpu->identity.model == chipModel_GC600 ||
452 	    gpu->identity.model == chipModel_GC300)
453 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
454 				 VIVS_HI_IDLE_STATE_RA |
455 				 VIVS_HI_IDLE_STATE_SE |
456 				 VIVS_HI_IDLE_STATE_PA |
457 				 VIVS_HI_IDLE_STATE_SH |
458 				 VIVS_HI_IDLE_STATE_PE |
459 				 VIVS_HI_IDLE_STATE_DE |
460 				 VIVS_HI_IDLE_STATE_FE;
461 
462 	etnaviv_hw_specs(gpu);
463 }
464 
etnaviv_gpu_load_clock(struct etnaviv_gpu * gpu,u32 clock)465 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
466 {
467 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
468 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
469 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
470 }
471 
etnaviv_gpu_update_clock(struct etnaviv_gpu * gpu)472 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
473 {
474 	if (gpu->identity.minor_features2 &
475 	    chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
476 		clk_set_rate(gpu->clk_core,
477 			     gpu->base_rate_core >> gpu->freq_scale);
478 		clk_set_rate(gpu->clk_shader,
479 			     gpu->base_rate_shader >> gpu->freq_scale);
480 	} else {
481 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
482 		u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
483 
484 		clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
485 		clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
486 		etnaviv_gpu_load_clock(gpu, clock);
487 	}
488 }
489 
etnaviv_hw_reset(struct etnaviv_gpu * gpu)490 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
491 {
492 	u32 control, idle;
493 	unsigned long timeout;
494 	bool failed = true;
495 
496 	/* We hope that the GPU resets in under one second */
497 	timeout = jiffies + msecs_to_jiffies(1000);
498 
499 	while (time_is_after_jiffies(timeout)) {
500 		/* enable clock */
501 		unsigned int fscale = 1 << (6 - gpu->freq_scale);
502 		control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
503 		etnaviv_gpu_load_clock(gpu, control);
504 
505 		/* isolate the GPU. */
506 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
507 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
508 
509 		if (gpu->sec_mode == ETNA_SEC_KERNEL) {
510 			gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
511 			          VIVS_MMUv2_AHB_CONTROL_RESET);
512 		} else {
513 			/* set soft reset. */
514 			control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
515 			gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
516 		}
517 
518 		/* wait for reset. */
519 		usleep_range(10, 20);
520 
521 		/* reset soft reset bit. */
522 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
523 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
524 
525 		/* reset GPU isolation. */
526 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
527 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
528 
529 		/* read idle register. */
530 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
531 
532 		/* try resetting again if FE is not idle */
533 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
534 			dev_dbg(gpu->dev, "FE is not idle\n");
535 			continue;
536 		}
537 
538 		/* read reset register. */
539 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
540 
541 		/* is the GPU idle? */
542 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
543 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
544 			dev_dbg(gpu->dev, "GPU is not idle\n");
545 			continue;
546 		}
547 
548 		/* disable debug registers, as they are not normally needed */
549 		control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
550 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
551 
552 		failed = false;
553 		break;
554 	}
555 
556 	if (failed) {
557 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
558 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
559 
560 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
561 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
562 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
563 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
564 
565 		return -EBUSY;
566 	}
567 
568 	/* We rely on the GPU running, so program the clock */
569 	etnaviv_gpu_update_clock(gpu);
570 
571 	gpu->fe_running = false;
572 	gpu->exec_state = -1;
573 	if (gpu->mmu_context)
574 		etnaviv_iommu_context_put(gpu->mmu_context);
575 	gpu->mmu_context = NULL;
576 
577 	return 0;
578 }
579 
etnaviv_gpu_enable_mlcg(struct etnaviv_gpu * gpu)580 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
581 {
582 	u32 pmc, ppc;
583 
584 	/* enable clock gating */
585 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
586 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
587 
588 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
589 	if (gpu->identity.revision == 0x4301 ||
590 	    gpu->identity.revision == 0x4302)
591 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
592 
593 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
594 
595 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
596 
597 	/* Disable PA clock gating for GC400+ without bugfix except for GC420 */
598 	if (gpu->identity.model >= chipModel_GC400 &&
599 	    gpu->identity.model != chipModel_GC420 &&
600 	    !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
601 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
602 
603 	/*
604 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
605 	 * present without a bug fix.
606 	 */
607 	if (gpu->identity.revision < 0x5000 &&
608 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
609 	    !(gpu->identity.minor_features1 &
610 	      chipMinorFeatures1_DISABLE_PE_GATING))
611 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
612 
613 	if (gpu->identity.revision < 0x5422)
614 		pmc |= BIT(15); /* Unknown bit */
615 
616 	/* Disable TX clock gating on affected core revisions. */
617 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
618 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108))
619 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
620 
621 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
622 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
623 
624 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
625 }
626 
etnaviv_gpu_start_fe(struct etnaviv_gpu * gpu,u32 address,u16 prefetch)627 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
628 {
629 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
630 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
631 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
632 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
633 
634 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
635 		gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
636 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
637 			  VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
638 	}
639 
640 	gpu->fe_running = true;
641 }
642 
etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu * gpu,struct etnaviv_iommu_context * context)643 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu,
644 					  struct etnaviv_iommu_context *context)
645 {
646 	u16 prefetch;
647 	u32 address;
648 
649 	/* setup the MMU */
650 	etnaviv_iommu_restore(gpu, context);
651 
652 	/* Start command processor */
653 	prefetch = etnaviv_buffer_init(gpu);
654 	address = etnaviv_cmdbuf_get_va(&gpu->buffer,
655 					&gpu->mmu_context->cmdbuf_mapping);
656 
657 	etnaviv_gpu_start_fe(gpu, address, prefetch);
658 }
659 
etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu * gpu)660 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
661 {
662 	/*
663 	 * Base value for VIVS_PM_PULSE_EATER register on models where it
664 	 * cannot be read, extracted from vivante kernel driver.
665 	 */
666 	u32 pulse_eater = 0x01590880;
667 
668 	if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
669 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
670 		pulse_eater |= BIT(23);
671 
672 	}
673 
674 	if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
675 	    etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
676 		pulse_eater &= ~BIT(16);
677 		pulse_eater |= BIT(17);
678 	}
679 
680 	if ((gpu->identity.revision > 0x5420) &&
681 	    (gpu->identity.features & chipFeatures_PIPE_3D))
682 	{
683 		/* Performance fix: disable internal DFS */
684 		pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
685 		pulse_eater |= BIT(18);
686 	}
687 
688 	gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
689 }
690 
etnaviv_gpu_hw_init(struct etnaviv_gpu * gpu)691 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
692 {
693 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
694 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
695 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
696 		u32 mc_memory_debug;
697 
698 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
699 
700 		if (gpu->identity.revision == 0x5007)
701 			mc_memory_debug |= 0x0c;
702 		else
703 			mc_memory_debug |= 0x08;
704 
705 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
706 	}
707 
708 	/* enable module-level clock gating */
709 	etnaviv_gpu_enable_mlcg(gpu);
710 
711 	/*
712 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
713 	 * This is necessary to prevent the iMX6 SoC locking up.
714 	 */
715 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
716 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
717 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
718 
719 	/* GC2000 rev 5108 needs a special bus config */
720 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
721 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
722 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
723 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
724 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
725 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
726 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
727 	}
728 
729 	if (gpu->sec_mode == ETNA_SEC_KERNEL) {
730 		u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
731 		val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
732 		gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
733 	}
734 
735 	/* setup the pulse eater */
736 	etnaviv_gpu_setup_pulse_eater(gpu);
737 
738 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
739 }
740 
etnaviv_gpu_init(struct etnaviv_gpu * gpu)741 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
742 {
743 	struct etnaviv_drm_private *priv = gpu->drm->dev_private;
744 	int ret, i;
745 
746 	ret = pm_runtime_get_sync(gpu->dev);
747 	if (ret < 0) {
748 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
749 		goto pm_put;
750 	}
751 
752 	etnaviv_hw_identify(gpu);
753 
754 	if (gpu->identity.model == 0) {
755 		dev_err(gpu->dev, "Unknown GPU model\n");
756 		ret = -ENXIO;
757 		goto fail;
758 	}
759 
760 	/* Exclude VG cores with FE2.0 */
761 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
762 	    gpu->identity.features & chipFeatures_FE20) {
763 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
764 		ret = -ENXIO;
765 		goto fail;
766 	}
767 
768 	/*
769 	 * On cores with security features supported, we claim control over the
770 	 * security states.
771 	 */
772 	if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
773 	    (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
774 		gpu->sec_mode = ETNA_SEC_KERNEL;
775 
776 	ret = etnaviv_hw_reset(gpu);
777 	if (ret) {
778 		dev_err(gpu->dev, "GPU reset failed\n");
779 		goto fail;
780 	}
781 
782 	ret = etnaviv_iommu_global_init(gpu);
783 	if (ret)
784 		goto fail;
785 
786 	/*
787 	 * Set the GPU linear window to be at the end of the DMA window, where
788 	 * the CMA area is likely to reside. This ensures that we are able to
789 	 * map the command buffers while having the linear window overlap as
790 	 * much RAM as possible, so we can optimize mappings for other buffers.
791 	 *
792 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
793 	 * to different views of the memory on the individual engines.
794 	 */
795 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
796 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
797 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
798 		if (dma_mask < PHYS_OFFSET + SZ_2G)
799 			priv->mmu_global->memory_base = PHYS_OFFSET;
800 		else
801 			priv->mmu_global->memory_base = dma_mask - SZ_2G + 1;
802 	} else if (PHYS_OFFSET >= SZ_2G) {
803 		dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
804 		priv->mmu_global->memory_base = PHYS_OFFSET;
805 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
806 	}
807 
808 	/*
809 	 * If the GPU is part of a system with DMA addressing limitations,
810 	 * request pages for our SHM backend buffers from the DMA32 zone to
811 	 * hopefully avoid performance killing SWIOTLB bounce buffering.
812 	 */
813 	if (dma_addressing_limited(gpu->dev))
814 		priv->shm_gfp_mask |= GFP_DMA32;
815 
816 	/* Create buffer: */
817 	ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
818 				  PAGE_SIZE);
819 	if (ret) {
820 		dev_err(gpu->dev, "could not create command buffer\n");
821 		goto fail;
822 	}
823 
824 	/* Setup event management */
825 	spin_lock_init(&gpu->event_spinlock);
826 	init_completion(&gpu->event_free);
827 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
828 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
829 		complete(&gpu->event_free);
830 
831 	/* Now program the hardware */
832 	mutex_lock(&gpu->lock);
833 	etnaviv_gpu_hw_init(gpu);
834 	mutex_unlock(&gpu->lock);
835 
836 	pm_runtime_mark_last_busy(gpu->dev);
837 	pm_runtime_put_autosuspend(gpu->dev);
838 
839 	gpu->initialized = true;
840 
841 	return 0;
842 
843 fail:
844 	pm_runtime_mark_last_busy(gpu->dev);
845 pm_put:
846 	pm_runtime_put_autosuspend(gpu->dev);
847 
848 	return ret;
849 }
850 
851 #ifdef CONFIG_DEBUG_FS
852 struct dma_debug {
853 	u32 address[2];
854 	u32 state[2];
855 };
856 
verify_dma(struct etnaviv_gpu * gpu,struct dma_debug * debug)857 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
858 {
859 	u32 i;
860 
861 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
862 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
863 
864 	for (i = 0; i < 500; i++) {
865 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
866 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
867 
868 		if (debug->address[0] != debug->address[1])
869 			break;
870 
871 		if (debug->state[0] != debug->state[1])
872 			break;
873 	}
874 }
875 
etnaviv_gpu_debugfs(struct etnaviv_gpu * gpu,struct seq_file * m)876 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
877 {
878 	struct dma_debug debug;
879 	u32 dma_lo, dma_hi, axi, idle;
880 	int ret;
881 
882 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
883 
884 	ret = pm_runtime_get_sync(gpu->dev);
885 	if (ret < 0)
886 		goto pm_put;
887 
888 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
889 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
890 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
891 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
892 
893 	verify_dma(gpu, &debug);
894 
895 	seq_puts(m, "\tidentity\n");
896 	seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
897 	seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
898 	seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
899 	seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
900 	seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
901 
902 	seq_puts(m, "\tfeatures\n");
903 	seq_printf(m, "\t major_features: 0x%08x\n",
904 		   gpu->identity.features);
905 	seq_printf(m, "\t minor_features0: 0x%08x\n",
906 		   gpu->identity.minor_features0);
907 	seq_printf(m, "\t minor_features1: 0x%08x\n",
908 		   gpu->identity.minor_features1);
909 	seq_printf(m, "\t minor_features2: 0x%08x\n",
910 		   gpu->identity.minor_features2);
911 	seq_printf(m, "\t minor_features3: 0x%08x\n",
912 		   gpu->identity.minor_features3);
913 	seq_printf(m, "\t minor_features4: 0x%08x\n",
914 		   gpu->identity.minor_features4);
915 	seq_printf(m, "\t minor_features5: 0x%08x\n",
916 		   gpu->identity.minor_features5);
917 	seq_printf(m, "\t minor_features6: 0x%08x\n",
918 		   gpu->identity.minor_features6);
919 	seq_printf(m, "\t minor_features7: 0x%08x\n",
920 		   gpu->identity.minor_features7);
921 	seq_printf(m, "\t minor_features8: 0x%08x\n",
922 		   gpu->identity.minor_features8);
923 	seq_printf(m, "\t minor_features9: 0x%08x\n",
924 		   gpu->identity.minor_features9);
925 	seq_printf(m, "\t minor_features10: 0x%08x\n",
926 		   gpu->identity.minor_features10);
927 	seq_printf(m, "\t minor_features11: 0x%08x\n",
928 		   gpu->identity.minor_features11);
929 
930 	seq_puts(m, "\tspecs\n");
931 	seq_printf(m, "\t stream_count:  %d\n",
932 			gpu->identity.stream_count);
933 	seq_printf(m, "\t register_max: %d\n",
934 			gpu->identity.register_max);
935 	seq_printf(m, "\t thread_count: %d\n",
936 			gpu->identity.thread_count);
937 	seq_printf(m, "\t vertex_cache_size: %d\n",
938 			gpu->identity.vertex_cache_size);
939 	seq_printf(m, "\t shader_core_count: %d\n",
940 			gpu->identity.shader_core_count);
941 	seq_printf(m, "\t pixel_pipes: %d\n",
942 			gpu->identity.pixel_pipes);
943 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
944 			gpu->identity.vertex_output_buffer_size);
945 	seq_printf(m, "\t buffer_size: %d\n",
946 			gpu->identity.buffer_size);
947 	seq_printf(m, "\t instruction_count: %d\n",
948 			gpu->identity.instruction_count);
949 	seq_printf(m, "\t num_constants: %d\n",
950 			gpu->identity.num_constants);
951 	seq_printf(m, "\t varyings_count: %d\n",
952 			gpu->identity.varyings_count);
953 
954 	seq_printf(m, "\taxi: 0x%08x\n", axi);
955 	seq_printf(m, "\tidle: 0x%08x\n", idle);
956 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
957 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
958 		seq_puts(m, "\t FE is not idle\n");
959 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
960 		seq_puts(m, "\t DE is not idle\n");
961 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
962 		seq_puts(m, "\t PE is not idle\n");
963 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
964 		seq_puts(m, "\t SH is not idle\n");
965 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
966 		seq_puts(m, "\t PA is not idle\n");
967 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
968 		seq_puts(m, "\t SE is not idle\n");
969 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
970 		seq_puts(m, "\t RA is not idle\n");
971 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
972 		seq_puts(m, "\t TX is not idle\n");
973 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
974 		seq_puts(m, "\t VG is not idle\n");
975 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
976 		seq_puts(m, "\t IM is not idle\n");
977 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
978 		seq_puts(m, "\t FP is not idle\n");
979 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
980 		seq_puts(m, "\t TS is not idle\n");
981 	if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
982 		seq_puts(m, "\t BL is not idle\n");
983 	if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
984 		seq_puts(m, "\t ASYNCFE is not idle\n");
985 	if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
986 		seq_puts(m, "\t MC is not idle\n");
987 	if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
988 		seq_puts(m, "\t PPA is not idle\n");
989 	if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
990 		seq_puts(m, "\t WD is not idle\n");
991 	if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
992 		seq_puts(m, "\t NN is not idle\n");
993 	if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
994 		seq_puts(m, "\t TP is not idle\n");
995 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
996 		seq_puts(m, "\t AXI low power mode\n");
997 
998 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
999 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1000 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1001 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1002 
1003 		seq_puts(m, "\tMC\n");
1004 		seq_printf(m, "\t read0: 0x%08x\n", read0);
1005 		seq_printf(m, "\t read1: 0x%08x\n", read1);
1006 		seq_printf(m, "\t write: 0x%08x\n", write);
1007 	}
1008 
1009 	seq_puts(m, "\tDMA ");
1010 
1011 	if (debug.address[0] == debug.address[1] &&
1012 	    debug.state[0] == debug.state[1]) {
1013 		seq_puts(m, "seems to be stuck\n");
1014 	} else if (debug.address[0] == debug.address[1]) {
1015 		seq_puts(m, "address is constant\n");
1016 	} else {
1017 		seq_puts(m, "is running\n");
1018 	}
1019 
1020 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1021 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1022 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1023 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1024 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1025 		   dma_lo, dma_hi);
1026 
1027 	ret = 0;
1028 
1029 	pm_runtime_mark_last_busy(gpu->dev);
1030 pm_put:
1031 	pm_runtime_put_autosuspend(gpu->dev);
1032 
1033 	return ret;
1034 }
1035 #endif
1036 
etnaviv_gpu_recover_hang(struct etnaviv_gpu * gpu)1037 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
1038 {
1039 	unsigned int i = 0;
1040 
1041 	dev_err(gpu->dev, "recover hung GPU!\n");
1042 
1043 	if (pm_runtime_get_sync(gpu->dev) < 0)
1044 		goto pm_put;
1045 
1046 	mutex_lock(&gpu->lock);
1047 
1048 	etnaviv_hw_reset(gpu);
1049 
1050 	/* complete all events, the GPU won't do it after the reset */
1051 	spin_lock(&gpu->event_spinlock);
1052 	for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1053 		complete(&gpu->event_free);
1054 	bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1055 	spin_unlock(&gpu->event_spinlock);
1056 
1057 	etnaviv_gpu_hw_init(gpu);
1058 
1059 	mutex_unlock(&gpu->lock);
1060 	pm_runtime_mark_last_busy(gpu->dev);
1061 pm_put:
1062 	pm_runtime_put_autosuspend(gpu->dev);
1063 }
1064 
1065 /* fence object management */
1066 struct etnaviv_fence {
1067 	struct etnaviv_gpu *gpu;
1068 	struct dma_fence base;
1069 };
1070 
to_etnaviv_fence(struct dma_fence * fence)1071 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1072 {
1073 	return container_of(fence, struct etnaviv_fence, base);
1074 }
1075 
etnaviv_fence_get_driver_name(struct dma_fence * fence)1076 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1077 {
1078 	return "etnaviv";
1079 }
1080 
etnaviv_fence_get_timeline_name(struct dma_fence * fence)1081 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1082 {
1083 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1084 
1085 	return dev_name(f->gpu->dev);
1086 }
1087 
etnaviv_fence_signaled(struct dma_fence * fence)1088 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1089 {
1090 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1091 
1092 	return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1093 }
1094 
etnaviv_fence_release(struct dma_fence * fence)1095 static void etnaviv_fence_release(struct dma_fence *fence)
1096 {
1097 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1098 
1099 	kfree_rcu(f, base.rcu);
1100 }
1101 
1102 static const struct dma_fence_ops etnaviv_fence_ops = {
1103 	.get_driver_name = etnaviv_fence_get_driver_name,
1104 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1105 	.signaled = etnaviv_fence_signaled,
1106 	.release = etnaviv_fence_release,
1107 };
1108 
etnaviv_gpu_fence_alloc(struct etnaviv_gpu * gpu)1109 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1110 {
1111 	struct etnaviv_fence *f;
1112 
1113 	/*
1114 	 * GPU lock must already be held, otherwise fence completion order might
1115 	 * not match the seqno order assigned here.
1116 	 */
1117 	lockdep_assert_held(&gpu->lock);
1118 
1119 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1120 	if (!f)
1121 		return NULL;
1122 
1123 	f->gpu = gpu;
1124 
1125 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1126 		       gpu->fence_context, ++gpu->next_fence);
1127 
1128 	return &f->base;
1129 }
1130 
1131 /* returns true if fence a comes after fence b */
fence_after(u32 a,u32 b)1132 static inline bool fence_after(u32 a, u32 b)
1133 {
1134 	return (s32)(a - b) > 0;
1135 }
1136 
1137 /*
1138  * event management:
1139  */
1140 
event_alloc(struct etnaviv_gpu * gpu,unsigned nr_events,unsigned int * events)1141 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1142 	unsigned int *events)
1143 {
1144 	unsigned long timeout = msecs_to_jiffies(10 * 10000);
1145 	unsigned i, acquired = 0;
1146 
1147 	for (i = 0; i < nr_events; i++) {
1148 		unsigned long ret;
1149 
1150 		ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1151 
1152 		if (!ret) {
1153 			dev_err(gpu->dev, "wait_for_completion_timeout failed");
1154 			goto out;
1155 		}
1156 
1157 		acquired++;
1158 		timeout = ret;
1159 	}
1160 
1161 	spin_lock(&gpu->event_spinlock);
1162 
1163 	for (i = 0; i < nr_events; i++) {
1164 		int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1165 
1166 		events[i] = event;
1167 		memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1168 		set_bit(event, gpu->event_bitmap);
1169 	}
1170 
1171 	spin_unlock(&gpu->event_spinlock);
1172 
1173 	return 0;
1174 
1175 out:
1176 	for (i = 0; i < acquired; i++)
1177 		complete(&gpu->event_free);
1178 
1179 	return -EBUSY;
1180 }
1181 
event_free(struct etnaviv_gpu * gpu,unsigned int event)1182 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1183 {
1184 	if (!test_bit(event, gpu->event_bitmap)) {
1185 		dev_warn(gpu->dev, "event %u is already marked as free",
1186 			 event);
1187 	} else {
1188 		clear_bit(event, gpu->event_bitmap);
1189 		complete(&gpu->event_free);
1190 	}
1191 }
1192 
1193 /*
1194  * Cmdstream submission/retirement:
1195  */
etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu * gpu,u32 id,struct drm_etnaviv_timespec * timeout)1196 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1197 	u32 id, struct drm_etnaviv_timespec *timeout)
1198 {
1199 	struct dma_fence *fence;
1200 	int ret;
1201 
1202 	/*
1203 	 * Look up the fence and take a reference. We might still find a fence
1204 	 * whose refcount has already dropped to zero. dma_fence_get_rcu
1205 	 * pretends we didn't find a fence in that case.
1206 	 */
1207 	rcu_read_lock();
1208 	fence = idr_find(&gpu->fence_idr, id);
1209 	if (fence)
1210 		fence = dma_fence_get_rcu(fence);
1211 	rcu_read_unlock();
1212 
1213 	if (!fence)
1214 		return 0;
1215 
1216 	if (!timeout) {
1217 		/* No timeout was requested: just test for completion */
1218 		ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1219 	} else {
1220 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1221 
1222 		ret = dma_fence_wait_timeout(fence, true, remaining);
1223 		if (ret == 0)
1224 			ret = -ETIMEDOUT;
1225 		else if (ret != -ERESTARTSYS)
1226 			ret = 0;
1227 
1228 	}
1229 
1230 	dma_fence_put(fence);
1231 	return ret;
1232 }
1233 
1234 /*
1235  * Wait for an object to become inactive.  This, on it's own, is not race
1236  * free: the object is moved by the scheduler off the active list, and
1237  * then the iova is put.  Moreover, the object could be re-submitted just
1238  * after we notice that it's become inactive.
1239  *
1240  * Although the retirement happens under the gpu lock, we don't want to hold
1241  * that lock in this function while waiting.
1242  */
etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu * gpu,struct etnaviv_gem_object * etnaviv_obj,struct drm_etnaviv_timespec * timeout)1243 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1244 	struct etnaviv_gem_object *etnaviv_obj,
1245 	struct drm_etnaviv_timespec *timeout)
1246 {
1247 	unsigned long remaining;
1248 	long ret;
1249 
1250 	if (!timeout)
1251 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1252 
1253 	remaining = etnaviv_timeout_to_jiffies(timeout);
1254 
1255 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1256 					       !is_active(etnaviv_obj),
1257 					       remaining);
1258 	if (ret > 0)
1259 		return 0;
1260 	else if (ret == -ERESTARTSYS)
1261 		return -ERESTARTSYS;
1262 	else
1263 		return -ETIMEDOUT;
1264 }
1265 
sync_point_perfmon_sample(struct etnaviv_gpu * gpu,struct etnaviv_event * event,unsigned int flags)1266 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1267 	struct etnaviv_event *event, unsigned int flags)
1268 {
1269 	const struct etnaviv_gem_submit *submit = event->submit;
1270 	unsigned int i;
1271 
1272 	for (i = 0; i < submit->nr_pmrs; i++) {
1273 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1274 
1275 		if (pmr->flags == flags)
1276 			etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1277 	}
1278 }
1279 
sync_point_perfmon_sample_pre(struct etnaviv_gpu * gpu,struct etnaviv_event * event)1280 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1281 	struct etnaviv_event *event)
1282 {
1283 	u32 val;
1284 
1285 	/* disable clock gating */
1286 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1287 	val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1288 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1289 
1290 	/* enable debug register */
1291 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1292 	val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1293 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1294 
1295 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1296 }
1297 
sync_point_perfmon_sample_post(struct etnaviv_gpu * gpu,struct etnaviv_event * event)1298 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1299 	struct etnaviv_event *event)
1300 {
1301 	const struct etnaviv_gem_submit *submit = event->submit;
1302 	unsigned int i;
1303 	u32 val;
1304 
1305 	sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1306 
1307 	for (i = 0; i < submit->nr_pmrs; i++) {
1308 		const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1309 
1310 		*pmr->bo_vma = pmr->sequence;
1311 	}
1312 
1313 	/* disable debug register */
1314 	val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1315 	val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1316 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1317 
1318 	/* enable clock gating */
1319 	val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1320 	val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1321 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1322 }
1323 
1324 
1325 /* add bo's to gpu's ring, and kick gpu: */
etnaviv_gpu_submit(struct etnaviv_gem_submit * submit)1326 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1327 {
1328 	struct etnaviv_gpu *gpu = submit->gpu;
1329 	struct dma_fence *gpu_fence;
1330 	unsigned int i, nr_events = 1, event[3];
1331 	int ret;
1332 
1333 	if (!submit->runtime_resumed) {
1334 		ret = pm_runtime_get_sync(gpu->dev);
1335 		if (ret < 0) {
1336 			pm_runtime_put_noidle(gpu->dev);
1337 			return NULL;
1338 		}
1339 		submit->runtime_resumed = true;
1340 	}
1341 
1342 	/*
1343 	 * if there are performance monitor requests we need to have
1344 	 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1345 	 *   requests.
1346 	 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1347 	 *   and update the sequence number for userspace.
1348 	 */
1349 	if (submit->nr_pmrs)
1350 		nr_events = 3;
1351 
1352 	ret = event_alloc(gpu, nr_events, event);
1353 	if (ret) {
1354 		DRM_ERROR("no free events\n");
1355 		pm_runtime_put_noidle(gpu->dev);
1356 		return NULL;
1357 	}
1358 
1359 	mutex_lock(&gpu->lock);
1360 
1361 	gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1362 	if (!gpu_fence) {
1363 		for (i = 0; i < nr_events; i++)
1364 			event_free(gpu, event[i]);
1365 
1366 		goto out_unlock;
1367 	}
1368 
1369 	if (!gpu->fe_running)
1370 		etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context);
1371 
1372 	if (submit->prev_mmu_context)
1373 		etnaviv_iommu_context_put(submit->prev_mmu_context);
1374 	submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context);
1375 
1376 	if (submit->nr_pmrs) {
1377 		gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1378 		kref_get(&submit->refcount);
1379 		gpu->event[event[1]].submit = submit;
1380 		etnaviv_sync_point_queue(gpu, event[1]);
1381 	}
1382 
1383 	gpu->event[event[0]].fence = gpu_fence;
1384 	submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1385 	etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1386 			     event[0], &submit->cmdbuf);
1387 
1388 	if (submit->nr_pmrs) {
1389 		gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1390 		kref_get(&submit->refcount);
1391 		gpu->event[event[2]].submit = submit;
1392 		etnaviv_sync_point_queue(gpu, event[2]);
1393 	}
1394 
1395 out_unlock:
1396 	mutex_unlock(&gpu->lock);
1397 
1398 	return gpu_fence;
1399 }
1400 
sync_point_worker(struct work_struct * work)1401 static void sync_point_worker(struct work_struct *work)
1402 {
1403 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1404 					       sync_point_work);
1405 	struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1406 	u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1407 
1408 	event->sync_point(gpu, event);
1409 	etnaviv_submit_put(event->submit);
1410 	event_free(gpu, gpu->sync_point_event);
1411 
1412 	/* restart FE last to avoid GPU and IRQ racing against this worker */
1413 	etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1414 }
1415 
dump_mmu_fault(struct etnaviv_gpu * gpu)1416 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1417 {
1418 	u32 status_reg, status;
1419 	int i;
1420 
1421 	if (gpu->sec_mode == ETNA_SEC_NONE)
1422 		status_reg = VIVS_MMUv2_STATUS;
1423 	else
1424 		status_reg = VIVS_MMUv2_SEC_STATUS;
1425 
1426 	status = gpu_read(gpu, status_reg);
1427 	dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1428 
1429 	for (i = 0; i < 4; i++) {
1430 		u32 address_reg;
1431 
1432 		if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1433 			continue;
1434 
1435 		if (gpu->sec_mode == ETNA_SEC_NONE)
1436 			address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1437 		else
1438 			address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1439 
1440 		dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1441 				    gpu_read(gpu, address_reg));
1442 	}
1443 }
1444 
irq_handler(int irq,void * data)1445 static irqreturn_t irq_handler(int irq, void *data)
1446 {
1447 	struct etnaviv_gpu *gpu = data;
1448 	irqreturn_t ret = IRQ_NONE;
1449 
1450 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1451 
1452 	if (intr != 0) {
1453 		int event;
1454 
1455 		pm_runtime_mark_last_busy(gpu->dev);
1456 
1457 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1458 
1459 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1460 			dev_err(gpu->dev, "AXI bus error\n");
1461 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1462 		}
1463 
1464 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1465 			dump_mmu_fault(gpu);
1466 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1467 		}
1468 
1469 		while ((event = ffs(intr)) != 0) {
1470 			struct dma_fence *fence;
1471 
1472 			event -= 1;
1473 
1474 			intr &= ~(1 << event);
1475 
1476 			dev_dbg(gpu->dev, "event %u\n", event);
1477 
1478 			if (gpu->event[event].sync_point) {
1479 				gpu->sync_point_event = event;
1480 				queue_work(gpu->wq, &gpu->sync_point_work);
1481 			}
1482 
1483 			fence = gpu->event[event].fence;
1484 			if (!fence)
1485 				continue;
1486 
1487 			gpu->event[event].fence = NULL;
1488 
1489 			/*
1490 			 * Events can be processed out of order.  Eg,
1491 			 * - allocate and queue event 0
1492 			 * - allocate event 1
1493 			 * - event 0 completes, we process it
1494 			 * - allocate and queue event 0
1495 			 * - event 1 and event 0 complete
1496 			 * we can end up processing event 0 first, then 1.
1497 			 */
1498 			if (fence_after(fence->seqno, gpu->completed_fence))
1499 				gpu->completed_fence = fence->seqno;
1500 			dma_fence_signal(fence);
1501 
1502 			event_free(gpu, event);
1503 		}
1504 
1505 		ret = IRQ_HANDLED;
1506 	}
1507 
1508 	return ret;
1509 }
1510 
etnaviv_gpu_clk_enable(struct etnaviv_gpu * gpu)1511 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1512 {
1513 	int ret;
1514 
1515 	ret = clk_prepare_enable(gpu->clk_reg);
1516 	if (ret)
1517 		return ret;
1518 
1519 	ret = clk_prepare_enable(gpu->clk_bus);
1520 	if (ret)
1521 		goto disable_clk_reg;
1522 
1523 	ret = clk_prepare_enable(gpu->clk_core);
1524 	if (ret)
1525 		goto disable_clk_bus;
1526 
1527 	ret = clk_prepare_enable(gpu->clk_shader);
1528 	if (ret)
1529 		goto disable_clk_core;
1530 
1531 	return 0;
1532 
1533 disable_clk_core:
1534 	clk_disable_unprepare(gpu->clk_core);
1535 disable_clk_bus:
1536 	clk_disable_unprepare(gpu->clk_bus);
1537 disable_clk_reg:
1538 	clk_disable_unprepare(gpu->clk_reg);
1539 
1540 	return ret;
1541 }
1542 
etnaviv_gpu_clk_disable(struct etnaviv_gpu * gpu)1543 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1544 {
1545 	clk_disable_unprepare(gpu->clk_shader);
1546 	clk_disable_unprepare(gpu->clk_core);
1547 	clk_disable_unprepare(gpu->clk_bus);
1548 	clk_disable_unprepare(gpu->clk_reg);
1549 
1550 	return 0;
1551 }
1552 
etnaviv_gpu_wait_idle(struct etnaviv_gpu * gpu,unsigned int timeout_ms)1553 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1554 {
1555 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1556 
1557 	do {
1558 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1559 
1560 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1561 			return 0;
1562 
1563 		if (time_is_before_jiffies(timeout)) {
1564 			dev_warn(gpu->dev,
1565 				 "timed out waiting for idle: idle=0x%x\n",
1566 				 idle);
1567 			return -ETIMEDOUT;
1568 		}
1569 
1570 		udelay(5);
1571 	} while (1);
1572 }
1573 
etnaviv_gpu_hw_suspend(struct etnaviv_gpu * gpu)1574 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1575 {
1576 	if (gpu->initialized && gpu->fe_running) {
1577 		/* Replace the last WAIT with END */
1578 		mutex_lock(&gpu->lock);
1579 		etnaviv_buffer_end(gpu);
1580 		mutex_unlock(&gpu->lock);
1581 
1582 		/*
1583 		 * We know that only the FE is busy here, this should
1584 		 * happen quickly (as the WAIT is only 200 cycles).  If
1585 		 * we fail, just warn and continue.
1586 		 */
1587 		etnaviv_gpu_wait_idle(gpu, 100);
1588 
1589 		gpu->fe_running = false;
1590 	}
1591 
1592 	gpu->exec_state = -1;
1593 
1594 	return etnaviv_gpu_clk_disable(gpu);
1595 }
1596 
1597 #ifdef CONFIG_PM
etnaviv_gpu_hw_resume(struct etnaviv_gpu * gpu)1598 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1599 {
1600 	int ret;
1601 
1602 	ret = mutex_lock_killable(&gpu->lock);
1603 	if (ret)
1604 		return ret;
1605 
1606 	etnaviv_gpu_update_clock(gpu);
1607 	etnaviv_gpu_hw_init(gpu);
1608 
1609 	mutex_unlock(&gpu->lock);
1610 
1611 	return 0;
1612 }
1613 #endif
1614 
1615 static int
etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device * cdev,unsigned long * state)1616 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1617 				  unsigned long *state)
1618 {
1619 	*state = 6;
1620 
1621 	return 0;
1622 }
1623 
1624 static int
etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device * cdev,unsigned long * state)1625 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1626 				  unsigned long *state)
1627 {
1628 	struct etnaviv_gpu *gpu = cdev->devdata;
1629 
1630 	*state = gpu->freq_scale;
1631 
1632 	return 0;
1633 }
1634 
1635 static int
etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device * cdev,unsigned long state)1636 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1637 				  unsigned long state)
1638 {
1639 	struct etnaviv_gpu *gpu = cdev->devdata;
1640 
1641 	mutex_lock(&gpu->lock);
1642 	gpu->freq_scale = state;
1643 	if (!pm_runtime_suspended(gpu->dev))
1644 		etnaviv_gpu_update_clock(gpu);
1645 	mutex_unlock(&gpu->lock);
1646 
1647 	return 0;
1648 }
1649 
1650 static struct thermal_cooling_device_ops cooling_ops = {
1651 	.get_max_state = etnaviv_gpu_cooling_get_max_state,
1652 	.get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1653 	.set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1654 };
1655 
etnaviv_gpu_bind(struct device * dev,struct device * master,void * data)1656 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1657 	void *data)
1658 {
1659 	struct drm_device *drm = data;
1660 	struct etnaviv_drm_private *priv = drm->dev_private;
1661 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1662 	int ret;
1663 
1664 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1665 		gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1666 				(char *)dev_name(dev), gpu, &cooling_ops);
1667 		if (IS_ERR(gpu->cooling))
1668 			return PTR_ERR(gpu->cooling);
1669 	}
1670 
1671 	gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1672 	if (!gpu->wq) {
1673 		ret = -ENOMEM;
1674 		goto out_thermal;
1675 	}
1676 
1677 	ret = etnaviv_sched_init(gpu);
1678 	if (ret)
1679 		goto out_workqueue;
1680 
1681 #ifdef CONFIG_PM
1682 	ret = pm_runtime_get_sync(gpu->dev);
1683 #else
1684 	ret = etnaviv_gpu_clk_enable(gpu);
1685 #endif
1686 	if (ret < 0)
1687 		goto out_sched;
1688 
1689 
1690 	gpu->drm = drm;
1691 	gpu->fence_context = dma_fence_context_alloc(1);
1692 	idr_init(&gpu->fence_idr);
1693 	spin_lock_init(&gpu->fence_spinlock);
1694 
1695 	INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1696 	init_waitqueue_head(&gpu->fence_event);
1697 
1698 	priv->gpu[priv->num_gpus++] = gpu;
1699 
1700 	pm_runtime_mark_last_busy(gpu->dev);
1701 	pm_runtime_put_autosuspend(gpu->dev);
1702 
1703 	return 0;
1704 
1705 out_sched:
1706 	etnaviv_sched_fini(gpu);
1707 
1708 out_workqueue:
1709 	destroy_workqueue(gpu->wq);
1710 
1711 out_thermal:
1712 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1713 		thermal_cooling_device_unregister(gpu->cooling);
1714 
1715 	return ret;
1716 }
1717 
etnaviv_gpu_unbind(struct device * dev,struct device * master,void * data)1718 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1719 	void *data)
1720 {
1721 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1722 
1723 	DBG("%s", dev_name(gpu->dev));
1724 
1725 	flush_workqueue(gpu->wq);
1726 	destroy_workqueue(gpu->wq);
1727 
1728 	etnaviv_sched_fini(gpu);
1729 
1730 #ifdef CONFIG_PM
1731 	pm_runtime_get_sync(gpu->dev);
1732 	pm_runtime_put_sync_suspend(gpu->dev);
1733 #else
1734 	etnaviv_gpu_hw_suspend(gpu);
1735 #endif
1736 
1737 	if (gpu->mmu_context)
1738 		etnaviv_iommu_context_put(gpu->mmu_context);
1739 
1740 	if (gpu->initialized) {
1741 		etnaviv_cmdbuf_free(&gpu->buffer);
1742 		etnaviv_iommu_global_fini(gpu);
1743 		gpu->initialized = false;
1744 	}
1745 
1746 	gpu->drm = NULL;
1747 	idr_destroy(&gpu->fence_idr);
1748 
1749 	if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1750 		thermal_cooling_device_unregister(gpu->cooling);
1751 	gpu->cooling = NULL;
1752 }
1753 
1754 static const struct component_ops gpu_ops = {
1755 	.bind = etnaviv_gpu_bind,
1756 	.unbind = etnaviv_gpu_unbind,
1757 };
1758 
1759 static const struct of_device_id etnaviv_gpu_match[] = {
1760 	{
1761 		.compatible = "vivante,gc"
1762 	},
1763 	{ /* sentinel */ }
1764 };
1765 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1766 
etnaviv_gpu_platform_probe(struct platform_device * pdev)1767 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1768 {
1769 	struct device *dev = &pdev->dev;
1770 	struct etnaviv_gpu *gpu;
1771 	int err;
1772 
1773 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1774 	if (!gpu)
1775 		return -ENOMEM;
1776 
1777 	gpu->dev = &pdev->dev;
1778 	mutex_init(&gpu->lock);
1779 	mutex_init(&gpu->fence_lock);
1780 
1781 	/* Map registers: */
1782 	gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1783 	if (IS_ERR(gpu->mmio))
1784 		return PTR_ERR(gpu->mmio);
1785 
1786 	/* Get Interrupt: */
1787 	gpu->irq = platform_get_irq(pdev, 0);
1788 	if (gpu->irq < 0) {
1789 		dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1790 		return gpu->irq;
1791 	}
1792 
1793 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1794 			       dev_name(gpu->dev), gpu);
1795 	if (err) {
1796 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1797 		return err;
1798 	}
1799 
1800 	/* Get Clocks: */
1801 	gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1802 	DBG("clk_reg: %p", gpu->clk_reg);
1803 	if (IS_ERR(gpu->clk_reg))
1804 		return PTR_ERR(gpu->clk_reg);
1805 
1806 	gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1807 	DBG("clk_bus: %p", gpu->clk_bus);
1808 	if (IS_ERR(gpu->clk_bus))
1809 		return PTR_ERR(gpu->clk_bus);
1810 
1811 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1812 	DBG("clk_core: %p", gpu->clk_core);
1813 	if (IS_ERR(gpu->clk_core))
1814 		return PTR_ERR(gpu->clk_core);
1815 	gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1816 
1817 	gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1818 	DBG("clk_shader: %p", gpu->clk_shader);
1819 	if (IS_ERR(gpu->clk_shader))
1820 		return PTR_ERR(gpu->clk_shader);
1821 	gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1822 
1823 	/* TODO: figure out max mapped size */
1824 	dev_set_drvdata(dev, gpu);
1825 
1826 	/*
1827 	 * We treat the device as initially suspended.  The runtime PM
1828 	 * autosuspend delay is rather arbitary: no measurements have
1829 	 * yet been performed to determine an appropriate value.
1830 	 */
1831 	pm_runtime_use_autosuspend(gpu->dev);
1832 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1833 	pm_runtime_enable(gpu->dev);
1834 
1835 	err = component_add(&pdev->dev, &gpu_ops);
1836 	if (err < 0) {
1837 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1838 		return err;
1839 	}
1840 
1841 	return 0;
1842 }
1843 
etnaviv_gpu_platform_remove(struct platform_device * pdev)1844 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1845 {
1846 	component_del(&pdev->dev, &gpu_ops);
1847 	pm_runtime_disable(&pdev->dev);
1848 	return 0;
1849 }
1850 
1851 #ifdef CONFIG_PM
etnaviv_gpu_rpm_suspend(struct device * dev)1852 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1853 {
1854 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1855 	u32 idle, mask;
1856 
1857 	/* If there are any jobs in the HW queue, we're not idle */
1858 	if (atomic_read(&gpu->sched.hw_rq_count))
1859 		return -EBUSY;
1860 
1861 	/* Check whether the hardware (except FE and MC) is idle */
1862 	mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1863 				  VIVS_HI_IDLE_STATE_MC);
1864 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1865 	if (idle != mask) {
1866 		dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1867 				     idle);
1868 		return -EBUSY;
1869 	}
1870 
1871 	return etnaviv_gpu_hw_suspend(gpu);
1872 }
1873 
etnaviv_gpu_rpm_resume(struct device * dev)1874 static int etnaviv_gpu_rpm_resume(struct device *dev)
1875 {
1876 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1877 	int ret;
1878 
1879 	ret = etnaviv_gpu_clk_enable(gpu);
1880 	if (ret)
1881 		return ret;
1882 
1883 	/* Re-initialise the basic hardware state */
1884 	if (gpu->drm && gpu->initialized) {
1885 		ret = etnaviv_gpu_hw_resume(gpu);
1886 		if (ret) {
1887 			etnaviv_gpu_clk_disable(gpu);
1888 			return ret;
1889 		}
1890 	}
1891 
1892 	return 0;
1893 }
1894 #endif
1895 
1896 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1897 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1898 			   NULL)
1899 };
1900 
1901 struct platform_driver etnaviv_gpu_driver = {
1902 	.driver = {
1903 		.name = "etnaviv-gpu",
1904 		.owner = THIS_MODULE,
1905 		.pm = &etnaviv_gpu_pm_ops,
1906 		.of_match_table = etnaviv_gpu_match,
1907 	},
1908 	.probe = etnaviv_gpu_platform_probe,
1909 	.remove = etnaviv_gpu_platform_remove,
1910 	.id_table = gpu_ids,
1911 };
1912