xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_gpu.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
1 /*
2  * Copyright (C) 2015 Etnaviv Project
3  *
4  * This program is free software; you can redistribute it and/or modify it
5  * under the terms of the GNU General Public License version 2 as published by
6  * the Free Software Foundation.
7  *
8  * This program is distributed in the hope that it will be useful, but WITHOUT
9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
11  * more details.
12  *
13  * You should have received a copy of the GNU General Public License along with
14  * this program.  If not, see <http://www.gnu.org/licenses/>.
15  */
16 
17 #include <linux/component.h>
18 #include <linux/dma-fence.h>
19 #include <linux/moduleparam.h>
20 #include <linux/of_device.h>
21 #include "etnaviv_dump.h"
22 #include "etnaviv_gpu.h"
23 #include "etnaviv_gem.h"
24 #include "etnaviv_mmu.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 static bool etnaviv_dump_core = true;
36 module_param_named(dump_core, etnaviv_dump_core, bool, 0600);
37 
38 /*
39  * Driver functions:
40  */
41 
42 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
43 {
44 	switch (param) {
45 	case ETNAVIV_PARAM_GPU_MODEL:
46 		*value = gpu->identity.model;
47 		break;
48 
49 	case ETNAVIV_PARAM_GPU_REVISION:
50 		*value = gpu->identity.revision;
51 		break;
52 
53 	case ETNAVIV_PARAM_GPU_FEATURES_0:
54 		*value = gpu->identity.features;
55 		break;
56 
57 	case ETNAVIV_PARAM_GPU_FEATURES_1:
58 		*value = gpu->identity.minor_features0;
59 		break;
60 
61 	case ETNAVIV_PARAM_GPU_FEATURES_2:
62 		*value = gpu->identity.minor_features1;
63 		break;
64 
65 	case ETNAVIV_PARAM_GPU_FEATURES_3:
66 		*value = gpu->identity.minor_features2;
67 		break;
68 
69 	case ETNAVIV_PARAM_GPU_FEATURES_4:
70 		*value = gpu->identity.minor_features3;
71 		break;
72 
73 	case ETNAVIV_PARAM_GPU_FEATURES_5:
74 		*value = gpu->identity.minor_features4;
75 		break;
76 
77 	case ETNAVIV_PARAM_GPU_FEATURES_6:
78 		*value = gpu->identity.minor_features5;
79 		break;
80 
81 	case ETNAVIV_PARAM_GPU_STREAM_COUNT:
82 		*value = gpu->identity.stream_count;
83 		break;
84 
85 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
86 		*value = gpu->identity.register_max;
87 		break;
88 
89 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
90 		*value = gpu->identity.thread_count;
91 		break;
92 
93 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
94 		*value = gpu->identity.vertex_cache_size;
95 		break;
96 
97 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
98 		*value = gpu->identity.shader_core_count;
99 		break;
100 
101 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
102 		*value = gpu->identity.pixel_pipes;
103 		break;
104 
105 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
106 		*value = gpu->identity.vertex_output_buffer_size;
107 		break;
108 
109 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
110 		*value = gpu->identity.buffer_size;
111 		break;
112 
113 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
114 		*value = gpu->identity.instruction_count;
115 		break;
116 
117 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
118 		*value = gpu->identity.num_constants;
119 		break;
120 
121 	case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
122 		*value = gpu->identity.varyings_count;
123 		break;
124 
125 	default:
126 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
127 		return -EINVAL;
128 	}
129 
130 	return 0;
131 }
132 
133 
134 #define etnaviv_is_model_rev(gpu, mod, rev) \
135 	((gpu)->identity.model == chipModel_##mod && \
136 	 (gpu)->identity.revision == rev)
137 #define etnaviv_field(val, field) \
138 	(((val) & field##__MASK) >> field##__SHIFT)
139 
140 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
141 {
142 	if (gpu->identity.minor_features0 &
143 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
144 		u32 specs[4];
145 		unsigned int streams;
146 
147 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
148 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
149 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
150 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
151 
152 		gpu->identity.stream_count = etnaviv_field(specs[0],
153 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
154 		gpu->identity.register_max = etnaviv_field(specs[0],
155 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
156 		gpu->identity.thread_count = etnaviv_field(specs[0],
157 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
158 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
159 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
160 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
161 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
162 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
163 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
164 		gpu->identity.vertex_output_buffer_size =
165 			etnaviv_field(specs[0],
166 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
167 
168 		gpu->identity.buffer_size = etnaviv_field(specs[1],
169 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
170 		gpu->identity.instruction_count = etnaviv_field(specs[1],
171 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
172 		gpu->identity.num_constants = etnaviv_field(specs[1],
173 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
174 
175 		gpu->identity.varyings_count = etnaviv_field(specs[2],
176 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
177 
178 		/* This overrides the value from older register if non-zero */
179 		streams = etnaviv_field(specs[3],
180 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
181 		if (streams)
182 			gpu->identity.stream_count = streams;
183 	}
184 
185 	/* Fill in the stream count if not specified */
186 	if (gpu->identity.stream_count == 0) {
187 		if (gpu->identity.model >= 0x1000)
188 			gpu->identity.stream_count = 4;
189 		else
190 			gpu->identity.stream_count = 1;
191 	}
192 
193 	/* Convert the register max value */
194 	if (gpu->identity.register_max)
195 		gpu->identity.register_max = 1 << gpu->identity.register_max;
196 	else if (gpu->identity.model == chipModel_GC400)
197 		gpu->identity.register_max = 32;
198 	else
199 		gpu->identity.register_max = 64;
200 
201 	/* Convert thread count */
202 	if (gpu->identity.thread_count)
203 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
204 	else if (gpu->identity.model == chipModel_GC400)
205 		gpu->identity.thread_count = 64;
206 	else if (gpu->identity.model == chipModel_GC500 ||
207 		 gpu->identity.model == chipModel_GC530)
208 		gpu->identity.thread_count = 128;
209 	else
210 		gpu->identity.thread_count = 256;
211 
212 	if (gpu->identity.vertex_cache_size == 0)
213 		gpu->identity.vertex_cache_size = 8;
214 
215 	if (gpu->identity.shader_core_count == 0) {
216 		if (gpu->identity.model >= 0x1000)
217 			gpu->identity.shader_core_count = 2;
218 		else
219 			gpu->identity.shader_core_count = 1;
220 	}
221 
222 	if (gpu->identity.pixel_pipes == 0)
223 		gpu->identity.pixel_pipes = 1;
224 
225 	/* Convert virtex buffer size */
226 	if (gpu->identity.vertex_output_buffer_size) {
227 		gpu->identity.vertex_output_buffer_size =
228 			1 << gpu->identity.vertex_output_buffer_size;
229 	} else if (gpu->identity.model == chipModel_GC400) {
230 		if (gpu->identity.revision < 0x4000)
231 			gpu->identity.vertex_output_buffer_size = 512;
232 		else if (gpu->identity.revision < 0x4200)
233 			gpu->identity.vertex_output_buffer_size = 256;
234 		else
235 			gpu->identity.vertex_output_buffer_size = 128;
236 	} else {
237 		gpu->identity.vertex_output_buffer_size = 512;
238 	}
239 
240 	switch (gpu->identity.instruction_count) {
241 	case 0:
242 		if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
243 		    gpu->identity.model == chipModel_GC880)
244 			gpu->identity.instruction_count = 512;
245 		else
246 			gpu->identity.instruction_count = 256;
247 		break;
248 
249 	case 1:
250 		gpu->identity.instruction_count = 1024;
251 		break;
252 
253 	case 2:
254 		gpu->identity.instruction_count = 2048;
255 		break;
256 
257 	default:
258 		gpu->identity.instruction_count = 256;
259 		break;
260 	}
261 
262 	if (gpu->identity.num_constants == 0)
263 		gpu->identity.num_constants = 168;
264 
265 	if (gpu->identity.varyings_count == 0) {
266 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
267 			gpu->identity.varyings_count = 12;
268 		else
269 			gpu->identity.varyings_count = 8;
270 	}
271 
272 	/*
273 	 * For some cores, two varyings are consumed for position, so the
274 	 * maximum varying count needs to be reduced by one.
275 	 */
276 	if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
277 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
278 	    etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
279 	    etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
280 	    etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
281 	    etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
282 	    etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
283 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
284 	    etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
285 	    etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
286 	    etnaviv_is_model_rev(gpu, GC880, 0x5106))
287 		gpu->identity.varyings_count -= 1;
288 }
289 
290 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
291 {
292 	u32 chipIdentity;
293 
294 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
295 
296 	/* Special case for older graphic cores. */
297 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
298 		gpu->identity.model    = chipModel_GC500;
299 		gpu->identity.revision = etnaviv_field(chipIdentity,
300 					 VIVS_HI_CHIP_IDENTITY_REVISION);
301 	} else {
302 
303 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
304 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
305 
306 		/*
307 		 * !!!! HACK ALERT !!!!
308 		 * Because people change device IDs without letting software
309 		 * know about it - here is the hack to make it all look the
310 		 * same.  Only for GC400 family.
311 		 */
312 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
313 		    gpu->identity.model != chipModel_GC420) {
314 			gpu->identity.model = gpu->identity.model & 0x0400;
315 		}
316 
317 		/* Another special case */
318 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
319 			u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
320 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
321 
322 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
323 				/*
324 				 * This IP has an ECO; put the correct
325 				 * revision in it.
326 				 */
327 				gpu->identity.revision = 0x1051;
328 			}
329 		}
330 
331 		/*
332 		 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
333 		 * reality it's just a re-branded GC3000. We can identify this
334 		 * core by the upper half of the revision register being all 1.
335 		 * Fix model/rev here, so all other places can refer to this
336 		 * core by its real identity.
337 		 */
338 		if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
339 			gpu->identity.model = chipModel_GC3000;
340 			gpu->identity.revision &= 0xffff;
341 		}
342 	}
343 
344 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
345 		 gpu->identity.model, gpu->identity.revision);
346 
347 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
348 
349 	/* Disable fast clear on GC700. */
350 	if (gpu->identity.model == chipModel_GC700)
351 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
352 
353 	if ((gpu->identity.model == chipModel_GC500 &&
354 	     gpu->identity.revision < 2) ||
355 	    (gpu->identity.model == chipModel_GC300 &&
356 	     gpu->identity.revision < 0x2000)) {
357 
358 		/*
359 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
360 		 * registers.
361 		 */
362 		gpu->identity.minor_features0 = 0;
363 		gpu->identity.minor_features1 = 0;
364 		gpu->identity.minor_features2 = 0;
365 		gpu->identity.minor_features3 = 0;
366 		gpu->identity.minor_features4 = 0;
367 		gpu->identity.minor_features5 = 0;
368 	} else
369 		gpu->identity.minor_features0 =
370 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
371 
372 	if (gpu->identity.minor_features0 &
373 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
374 		gpu->identity.minor_features1 =
375 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
376 		gpu->identity.minor_features2 =
377 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
378 		gpu->identity.minor_features3 =
379 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
380 		gpu->identity.minor_features4 =
381 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
382 		gpu->identity.minor_features5 =
383 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
384 	}
385 
386 	/* GC600 idle register reports zero bits where modules aren't present */
387 	if (gpu->identity.model == chipModel_GC600) {
388 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
389 				 VIVS_HI_IDLE_STATE_RA |
390 				 VIVS_HI_IDLE_STATE_SE |
391 				 VIVS_HI_IDLE_STATE_PA |
392 				 VIVS_HI_IDLE_STATE_SH |
393 				 VIVS_HI_IDLE_STATE_PE |
394 				 VIVS_HI_IDLE_STATE_DE |
395 				 VIVS_HI_IDLE_STATE_FE;
396 	} else {
397 		gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
398 	}
399 
400 	etnaviv_hw_specs(gpu);
401 }
402 
403 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
404 {
405 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
406 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
407 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
408 }
409 
410 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
411 {
412 	u32 control, idle;
413 	unsigned long timeout;
414 	bool failed = true;
415 
416 	/* TODO
417 	 *
418 	 * - clock gating
419 	 * - puls eater
420 	 * - what about VG?
421 	 */
422 
423 	/* We hope that the GPU resets in under one second */
424 	timeout = jiffies + msecs_to_jiffies(1000);
425 
426 	while (time_is_after_jiffies(timeout)) {
427 		control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
428 			  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
429 
430 		/* enable clock */
431 		etnaviv_gpu_load_clock(gpu, control);
432 
433 		/* Wait for stable clock.  Vivante's code waited for 1ms */
434 		usleep_range(1000, 10000);
435 
436 		/* isolate the GPU. */
437 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
438 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
439 
440 		/* set soft reset. */
441 		control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
442 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
443 
444 		/* wait for reset. */
445 		msleep(1);
446 
447 		/* reset soft reset bit. */
448 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
449 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
450 
451 		/* reset GPU isolation. */
452 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
453 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
454 
455 		/* read idle register. */
456 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
457 
458 		/* try reseting again if FE it not idle */
459 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
460 			dev_dbg(gpu->dev, "FE is not idle\n");
461 			continue;
462 		}
463 
464 		/* read reset register. */
465 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
466 
467 		/* is the GPU idle? */
468 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
469 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
470 			dev_dbg(gpu->dev, "GPU is not idle\n");
471 			continue;
472 		}
473 
474 		failed = false;
475 		break;
476 	}
477 
478 	if (failed) {
479 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
480 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
481 
482 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
483 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
484 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
485 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
486 
487 		return -EBUSY;
488 	}
489 
490 	/* We rely on the GPU running, so program the clock */
491 	control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
492 		  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
493 
494 	/* enable clock */
495 	etnaviv_gpu_load_clock(gpu, control);
496 
497 	return 0;
498 }
499 
500 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
501 {
502 	u32 pmc, ppc;
503 
504 	/* enable clock gating */
505 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
506 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
507 
508 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
509 	if (gpu->identity.revision == 0x4301 ||
510 	    gpu->identity.revision == 0x4302)
511 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
512 
513 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
514 
515 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
516 
517 	/* Disable PA clock gating for GC400+ except for GC420 */
518 	if (gpu->identity.model >= chipModel_GC400 &&
519 	    gpu->identity.model != chipModel_GC420)
520 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
521 
522 	/*
523 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
524 	 * present without a bug fix.
525 	 */
526 	if (gpu->identity.revision < 0x5000 &&
527 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
528 	    !(gpu->identity.minor_features1 &
529 	      chipMinorFeatures1_DISABLE_PE_GATING))
530 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
531 
532 	if (gpu->identity.revision < 0x5422)
533 		pmc |= BIT(15); /* Unknown bit */
534 
535 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
536 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
537 
538 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
539 }
540 
541 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
542 {
543 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
544 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
545 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
546 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
547 }
548 
549 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
550 {
551 	u16 prefetch;
552 
553 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
554 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
555 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
556 		u32 mc_memory_debug;
557 
558 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
559 
560 		if (gpu->identity.revision == 0x5007)
561 			mc_memory_debug |= 0x0c;
562 		else
563 			mc_memory_debug |= 0x08;
564 
565 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
566 	}
567 
568 	/* enable module-level clock gating */
569 	etnaviv_gpu_enable_mlcg(gpu);
570 
571 	/*
572 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
573 	 * This is necessary to prevent the iMX6 SoC locking up.
574 	 */
575 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
576 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
577 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
578 
579 	/* GC2000 rev 5108 needs a special bus config */
580 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
581 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
582 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
583 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
584 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
585 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
586 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
587 	}
588 
589 	/* setup the MMU */
590 	etnaviv_iommu_restore(gpu);
591 
592 	/* Start command processor */
593 	prefetch = etnaviv_buffer_init(gpu);
594 
595 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
596 	etnaviv_gpu_start_fe(gpu, etnaviv_iommu_get_cmdbuf_va(gpu, gpu->buffer),
597 			     prefetch);
598 }
599 
600 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
601 {
602 	int ret, i;
603 
604 	ret = pm_runtime_get_sync(gpu->dev);
605 	if (ret < 0) {
606 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
607 		return ret;
608 	}
609 
610 	etnaviv_hw_identify(gpu);
611 
612 	if (gpu->identity.model == 0) {
613 		dev_err(gpu->dev, "Unknown GPU model\n");
614 		ret = -ENXIO;
615 		goto fail;
616 	}
617 
618 	/* Exclude VG cores with FE2.0 */
619 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
620 	    gpu->identity.features & chipFeatures_FE20) {
621 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
622 		ret = -ENXIO;
623 		goto fail;
624 	}
625 
626 	/*
627 	 * Set the GPU linear window to be at the end of the DMA window, where
628 	 * the CMA area is likely to reside. This ensures that we are able to
629 	 * map the command buffers while having the linear window overlap as
630 	 * much RAM as possible, so we can optimize mappings for other buffers.
631 	 *
632 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
633 	 * to different views of the memory on the individual engines.
634 	 */
635 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
636 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
637 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
638 		if (dma_mask < PHYS_OFFSET + SZ_2G)
639 			gpu->memory_base = PHYS_OFFSET;
640 		else
641 			gpu->memory_base = dma_mask - SZ_2G + 1;
642 	} else if (PHYS_OFFSET >= SZ_2G) {
643 		dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
644 		gpu->memory_base = PHYS_OFFSET;
645 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
646 	}
647 
648 	ret = etnaviv_hw_reset(gpu);
649 	if (ret) {
650 		dev_err(gpu->dev, "GPU reset failed\n");
651 		goto fail;
652 	}
653 
654 	gpu->mmu = etnaviv_iommu_new(gpu);
655 	if (IS_ERR(gpu->mmu)) {
656 		dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n");
657 		ret = PTR_ERR(gpu->mmu);
658 		goto fail;
659 	}
660 
661 	/* Create buffer: */
662 	gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
663 	if (!gpu->buffer) {
664 		ret = -ENOMEM;
665 		dev_err(gpu->dev, "could not create command buffer\n");
666 		goto destroy_iommu;
667 	}
668 
669 	if (gpu->mmu->version == ETNAVIV_IOMMU_V1 &&
670 	    gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
671 		ret = -EINVAL;
672 		dev_err(gpu->dev,
673 			"command buffer outside valid memory window\n");
674 		goto free_buffer;
675 	}
676 
677 	/* Setup event management */
678 	spin_lock_init(&gpu->event_spinlock);
679 	init_completion(&gpu->event_free);
680 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
681 		gpu->event[i].used = false;
682 		complete(&gpu->event_free);
683 	}
684 
685 	/* Now program the hardware */
686 	mutex_lock(&gpu->lock);
687 	etnaviv_gpu_hw_init(gpu);
688 	gpu->exec_state = -1;
689 	mutex_unlock(&gpu->lock);
690 
691 	pm_runtime_mark_last_busy(gpu->dev);
692 	pm_runtime_put_autosuspend(gpu->dev);
693 
694 	return 0;
695 
696 free_buffer:
697 	etnaviv_gpu_cmdbuf_free(gpu->buffer);
698 	gpu->buffer = NULL;
699 destroy_iommu:
700 	etnaviv_iommu_destroy(gpu->mmu);
701 	gpu->mmu = NULL;
702 fail:
703 	pm_runtime_mark_last_busy(gpu->dev);
704 	pm_runtime_put_autosuspend(gpu->dev);
705 
706 	return ret;
707 }
708 
709 #ifdef CONFIG_DEBUG_FS
710 struct dma_debug {
711 	u32 address[2];
712 	u32 state[2];
713 };
714 
715 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
716 {
717 	u32 i;
718 
719 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
720 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
721 
722 	for (i = 0; i < 500; i++) {
723 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
724 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
725 
726 		if (debug->address[0] != debug->address[1])
727 			break;
728 
729 		if (debug->state[0] != debug->state[1])
730 			break;
731 	}
732 }
733 
734 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
735 {
736 	struct dma_debug debug;
737 	u32 dma_lo, dma_hi, axi, idle;
738 	int ret;
739 
740 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
741 
742 	ret = pm_runtime_get_sync(gpu->dev);
743 	if (ret < 0)
744 		return ret;
745 
746 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
747 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
748 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
749 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
750 
751 	verify_dma(gpu, &debug);
752 
753 	seq_puts(m, "\tfeatures\n");
754 	seq_printf(m, "\t minor_features0: 0x%08x\n",
755 		   gpu->identity.minor_features0);
756 	seq_printf(m, "\t minor_features1: 0x%08x\n",
757 		   gpu->identity.minor_features1);
758 	seq_printf(m, "\t minor_features2: 0x%08x\n",
759 		   gpu->identity.minor_features2);
760 	seq_printf(m, "\t minor_features3: 0x%08x\n",
761 		   gpu->identity.minor_features3);
762 	seq_printf(m, "\t minor_features4: 0x%08x\n",
763 		   gpu->identity.minor_features4);
764 	seq_printf(m, "\t minor_features5: 0x%08x\n",
765 		   gpu->identity.minor_features5);
766 
767 	seq_puts(m, "\tspecs\n");
768 	seq_printf(m, "\t stream_count:  %d\n",
769 			gpu->identity.stream_count);
770 	seq_printf(m, "\t register_max: %d\n",
771 			gpu->identity.register_max);
772 	seq_printf(m, "\t thread_count: %d\n",
773 			gpu->identity.thread_count);
774 	seq_printf(m, "\t vertex_cache_size: %d\n",
775 			gpu->identity.vertex_cache_size);
776 	seq_printf(m, "\t shader_core_count: %d\n",
777 			gpu->identity.shader_core_count);
778 	seq_printf(m, "\t pixel_pipes: %d\n",
779 			gpu->identity.pixel_pipes);
780 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
781 			gpu->identity.vertex_output_buffer_size);
782 	seq_printf(m, "\t buffer_size: %d\n",
783 			gpu->identity.buffer_size);
784 	seq_printf(m, "\t instruction_count: %d\n",
785 			gpu->identity.instruction_count);
786 	seq_printf(m, "\t num_constants: %d\n",
787 			gpu->identity.num_constants);
788 	seq_printf(m, "\t varyings_count: %d\n",
789 			gpu->identity.varyings_count);
790 
791 	seq_printf(m, "\taxi: 0x%08x\n", axi);
792 	seq_printf(m, "\tidle: 0x%08x\n", idle);
793 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
794 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
795 		seq_puts(m, "\t FE is not idle\n");
796 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
797 		seq_puts(m, "\t DE is not idle\n");
798 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
799 		seq_puts(m, "\t PE is not idle\n");
800 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
801 		seq_puts(m, "\t SH is not idle\n");
802 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
803 		seq_puts(m, "\t PA is not idle\n");
804 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
805 		seq_puts(m, "\t SE is not idle\n");
806 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
807 		seq_puts(m, "\t RA is not idle\n");
808 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
809 		seq_puts(m, "\t TX is not idle\n");
810 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
811 		seq_puts(m, "\t VG is not idle\n");
812 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
813 		seq_puts(m, "\t IM is not idle\n");
814 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
815 		seq_puts(m, "\t FP is not idle\n");
816 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
817 		seq_puts(m, "\t TS is not idle\n");
818 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
819 		seq_puts(m, "\t AXI low power mode\n");
820 
821 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
822 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
823 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
824 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
825 
826 		seq_puts(m, "\tMC\n");
827 		seq_printf(m, "\t read0: 0x%08x\n", read0);
828 		seq_printf(m, "\t read1: 0x%08x\n", read1);
829 		seq_printf(m, "\t write: 0x%08x\n", write);
830 	}
831 
832 	seq_puts(m, "\tDMA ");
833 
834 	if (debug.address[0] == debug.address[1] &&
835 	    debug.state[0] == debug.state[1]) {
836 		seq_puts(m, "seems to be stuck\n");
837 	} else if (debug.address[0] == debug.address[1]) {
838 		seq_puts(m, "address is constant\n");
839 	} else {
840 		seq_puts(m, "is running\n");
841 	}
842 
843 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
844 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
845 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
846 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
847 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
848 		   dma_lo, dma_hi);
849 
850 	ret = 0;
851 
852 	pm_runtime_mark_last_busy(gpu->dev);
853 	pm_runtime_put_autosuspend(gpu->dev);
854 
855 	return ret;
856 }
857 #endif
858 
859 /*
860  * Hangcheck detection for locked gpu:
861  */
862 static void recover_worker(struct work_struct *work)
863 {
864 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
865 					       recover_work);
866 	unsigned long flags;
867 	unsigned int i;
868 
869 	dev_err(gpu->dev, "hangcheck recover!\n");
870 
871 	if (pm_runtime_get_sync(gpu->dev) < 0)
872 		return;
873 
874 	mutex_lock(&gpu->lock);
875 
876 	/* Only catch the first event, or when manually re-armed */
877 	if (etnaviv_dump_core) {
878 		etnaviv_core_dump(gpu);
879 		etnaviv_dump_core = false;
880 	}
881 
882 	etnaviv_hw_reset(gpu);
883 
884 	/* complete all events, the GPU won't do it after the reset */
885 	spin_lock_irqsave(&gpu->event_spinlock, flags);
886 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
887 		if (!gpu->event[i].used)
888 			continue;
889 		dma_fence_signal(gpu->event[i].fence);
890 		gpu->event[i].fence = NULL;
891 		gpu->event[i].used = false;
892 		complete(&gpu->event_free);
893 	}
894 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
895 	gpu->completed_fence = gpu->active_fence;
896 
897 	etnaviv_gpu_hw_init(gpu);
898 	gpu->lastctx = NULL;
899 	gpu->exec_state = -1;
900 
901 	mutex_unlock(&gpu->lock);
902 	pm_runtime_mark_last_busy(gpu->dev);
903 	pm_runtime_put_autosuspend(gpu->dev);
904 
905 	/* Retire the buffer objects in a work */
906 	etnaviv_queue_work(gpu->drm, &gpu->retire_work);
907 }
908 
909 static void hangcheck_timer_reset(struct etnaviv_gpu *gpu)
910 {
911 	DBG("%s", dev_name(gpu->dev));
912 	mod_timer(&gpu->hangcheck_timer,
913 		  round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES));
914 }
915 
916 static void hangcheck_handler(unsigned long data)
917 {
918 	struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data;
919 	u32 fence = gpu->completed_fence;
920 	bool progress = false;
921 
922 	if (fence != gpu->hangcheck_fence) {
923 		gpu->hangcheck_fence = fence;
924 		progress = true;
925 	}
926 
927 	if (!progress) {
928 		u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
929 		int change = dma_addr - gpu->hangcheck_dma_addr;
930 
931 		if (change < 0 || change > 16) {
932 			gpu->hangcheck_dma_addr = dma_addr;
933 			progress = true;
934 		}
935 	}
936 
937 	if (!progress && fence_after(gpu->active_fence, fence)) {
938 		dev_err(gpu->dev, "hangcheck detected gpu lockup!\n");
939 		dev_err(gpu->dev, "     completed fence: %u\n", fence);
940 		dev_err(gpu->dev, "     active fence: %u\n",
941 			gpu->active_fence);
942 		etnaviv_queue_work(gpu->drm, &gpu->recover_work);
943 	}
944 
945 	/* if still more pending work, reset the hangcheck timer: */
946 	if (fence_after(gpu->active_fence, gpu->hangcheck_fence))
947 		hangcheck_timer_reset(gpu);
948 }
949 
950 static void hangcheck_disable(struct etnaviv_gpu *gpu)
951 {
952 	del_timer_sync(&gpu->hangcheck_timer);
953 	cancel_work_sync(&gpu->recover_work);
954 }
955 
956 /* fence object management */
957 struct etnaviv_fence {
958 	struct etnaviv_gpu *gpu;
959 	struct dma_fence base;
960 };
961 
962 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
963 {
964 	return container_of(fence, struct etnaviv_fence, base);
965 }
966 
967 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
968 {
969 	return "etnaviv";
970 }
971 
972 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
973 {
974 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
975 
976 	return dev_name(f->gpu->dev);
977 }
978 
979 static bool etnaviv_fence_enable_signaling(struct dma_fence *fence)
980 {
981 	return true;
982 }
983 
984 static bool etnaviv_fence_signaled(struct dma_fence *fence)
985 {
986 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
987 
988 	return fence_completed(f->gpu, f->base.seqno);
989 }
990 
991 static void etnaviv_fence_release(struct dma_fence *fence)
992 {
993 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
994 
995 	kfree_rcu(f, base.rcu);
996 }
997 
998 static const struct dma_fence_ops etnaviv_fence_ops = {
999 	.get_driver_name = etnaviv_fence_get_driver_name,
1000 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1001 	.enable_signaling = etnaviv_fence_enable_signaling,
1002 	.signaled = etnaviv_fence_signaled,
1003 	.wait = dma_fence_default_wait,
1004 	.release = etnaviv_fence_release,
1005 };
1006 
1007 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1008 {
1009 	struct etnaviv_fence *f;
1010 
1011 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1012 	if (!f)
1013 		return NULL;
1014 
1015 	f->gpu = gpu;
1016 
1017 	dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1018 		       gpu->fence_context, ++gpu->next_fence);
1019 
1020 	return &f->base;
1021 }
1022 
1023 int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj,
1024 	unsigned int context, bool exclusive)
1025 {
1026 	struct reservation_object *robj = etnaviv_obj->resv;
1027 	struct reservation_object_list *fobj;
1028 	struct dma_fence *fence;
1029 	int i, ret;
1030 
1031 	if (!exclusive) {
1032 		ret = reservation_object_reserve_shared(robj);
1033 		if (ret)
1034 			return ret;
1035 	}
1036 
1037 	/*
1038 	 * If we have any shared fences, then the exclusive fence
1039 	 * should be ignored as it will already have been signalled.
1040 	 */
1041 	fobj = reservation_object_get_list(robj);
1042 	if (!fobj || fobj->shared_count == 0) {
1043 		/* Wait on any existing exclusive fence which isn't our own */
1044 		fence = reservation_object_get_excl(robj);
1045 		if (fence && fence->context != context) {
1046 			ret = dma_fence_wait(fence, true);
1047 			if (ret)
1048 				return ret;
1049 		}
1050 	}
1051 
1052 	if (!exclusive || !fobj)
1053 		return 0;
1054 
1055 	for (i = 0; i < fobj->shared_count; i++) {
1056 		fence = rcu_dereference_protected(fobj->shared[i],
1057 						reservation_object_held(robj));
1058 		if (fence->context != context) {
1059 			ret = dma_fence_wait(fence, true);
1060 			if (ret)
1061 				return ret;
1062 		}
1063 	}
1064 
1065 	return 0;
1066 }
1067 
1068 /*
1069  * event management:
1070  */
1071 
1072 static unsigned int event_alloc(struct etnaviv_gpu *gpu)
1073 {
1074 	unsigned long ret, flags;
1075 	unsigned int i, event = ~0U;
1076 
1077 	ret = wait_for_completion_timeout(&gpu->event_free,
1078 					  msecs_to_jiffies(10 * 10000));
1079 	if (!ret)
1080 		dev_err(gpu->dev, "wait_for_completion_timeout failed");
1081 
1082 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1083 
1084 	/* find first free event */
1085 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
1086 		if (gpu->event[i].used == false) {
1087 			gpu->event[i].used = true;
1088 			event = i;
1089 			break;
1090 		}
1091 	}
1092 
1093 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1094 
1095 	return event;
1096 }
1097 
1098 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1099 {
1100 	unsigned long flags;
1101 
1102 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1103 
1104 	if (gpu->event[event].used == false) {
1105 		dev_warn(gpu->dev, "event %u is already marked as free",
1106 			 event);
1107 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1108 	} else {
1109 		gpu->event[event].used = false;
1110 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1111 
1112 		complete(&gpu->event_free);
1113 	}
1114 }
1115 
1116 /*
1117  * Cmdstream submission/retirement:
1118  */
1119 
1120 struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
1121 	size_t nr_bos)
1122 {
1123 	struct etnaviv_cmdbuf *cmdbuf;
1124 	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
1125 				 sizeof(*cmdbuf));
1126 
1127 	cmdbuf = kzalloc(sz, GFP_KERNEL);
1128 	if (!cmdbuf)
1129 		return NULL;
1130 
1131 	if (gpu->mmu->version == ETNAVIV_IOMMU_V2)
1132 		size = ALIGN(size, SZ_4K);
1133 
1134 	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
1135 				     GFP_KERNEL);
1136 	if (!cmdbuf->vaddr) {
1137 		kfree(cmdbuf);
1138 		return NULL;
1139 	}
1140 
1141 	cmdbuf->gpu = gpu;
1142 	cmdbuf->size = size;
1143 
1144 	return cmdbuf;
1145 }
1146 
1147 void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1148 {
1149 	etnaviv_iommu_put_cmdbuf_va(cmdbuf->gpu, cmdbuf);
1150 	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
1151 		    cmdbuf->paddr);
1152 	kfree(cmdbuf);
1153 }
1154 
1155 static void retire_worker(struct work_struct *work)
1156 {
1157 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1158 					       retire_work);
1159 	u32 fence = gpu->completed_fence;
1160 	struct etnaviv_cmdbuf *cmdbuf, *tmp;
1161 	unsigned int i;
1162 
1163 	mutex_lock(&gpu->lock);
1164 	list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1165 		if (!dma_fence_is_signaled(cmdbuf->fence))
1166 			break;
1167 
1168 		list_del(&cmdbuf->node);
1169 		dma_fence_put(cmdbuf->fence);
1170 
1171 		for (i = 0; i < cmdbuf->nr_bos; i++) {
1172 			struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
1173 			struct etnaviv_gem_object *etnaviv_obj = mapping->object;
1174 
1175 			atomic_dec(&etnaviv_obj->gpu_active);
1176 			/* drop the refcount taken in etnaviv_gpu_submit */
1177 			etnaviv_gem_mapping_unreference(mapping);
1178 		}
1179 
1180 		etnaviv_gpu_cmdbuf_free(cmdbuf);
1181 		/*
1182 		 * We need to balance the runtime PM count caused by
1183 		 * each submission.  Upon submission, we increment
1184 		 * the runtime PM counter, and allocate one event.
1185 		 * So here, we put the runtime PM count for each
1186 		 * completed event.
1187 		 */
1188 		pm_runtime_put_autosuspend(gpu->dev);
1189 	}
1190 
1191 	gpu->retired_fence = fence;
1192 
1193 	mutex_unlock(&gpu->lock);
1194 
1195 	wake_up_all(&gpu->fence_event);
1196 }
1197 
1198 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1199 	u32 fence, struct timespec *timeout)
1200 {
1201 	int ret;
1202 
1203 	if (fence_after(fence, gpu->next_fence)) {
1204 		DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1205 				fence, gpu->next_fence);
1206 		return -EINVAL;
1207 	}
1208 
1209 	if (!timeout) {
1210 		/* No timeout was requested: just test for completion */
1211 		ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1212 	} else {
1213 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1214 
1215 		ret = wait_event_interruptible_timeout(gpu->fence_event,
1216 						fence_completed(gpu, fence),
1217 						remaining);
1218 		if (ret == 0) {
1219 			DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1220 				fence, gpu->retired_fence,
1221 				gpu->completed_fence);
1222 			ret = -ETIMEDOUT;
1223 		} else if (ret != -ERESTARTSYS) {
1224 			ret = 0;
1225 		}
1226 	}
1227 
1228 	return ret;
1229 }
1230 
1231 /*
1232  * Wait for an object to become inactive.  This, on it's own, is not race
1233  * free: the object is moved by the retire worker off the active list, and
1234  * then the iova is put.  Moreover, the object could be re-submitted just
1235  * after we notice that it's become inactive.
1236  *
1237  * Although the retirement happens under the gpu lock, we don't want to hold
1238  * that lock in this function while waiting.
1239  */
1240 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1241 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1242 {
1243 	unsigned long remaining;
1244 	long ret;
1245 
1246 	if (!timeout)
1247 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1248 
1249 	remaining = etnaviv_timeout_to_jiffies(timeout);
1250 
1251 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1252 					       !is_active(etnaviv_obj),
1253 					       remaining);
1254 	if (ret > 0) {
1255 		struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1256 
1257 		/* Synchronise with the retire worker */
1258 		flush_workqueue(priv->wq);
1259 		return 0;
1260 	} else if (ret == -ERESTARTSYS) {
1261 		return -ERESTARTSYS;
1262 	} else {
1263 		return -ETIMEDOUT;
1264 	}
1265 }
1266 
1267 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1268 {
1269 	return pm_runtime_get_sync(gpu->dev);
1270 }
1271 
1272 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1273 {
1274 	pm_runtime_mark_last_busy(gpu->dev);
1275 	pm_runtime_put_autosuspend(gpu->dev);
1276 }
1277 
1278 /* add bo's to gpu's ring, and kick gpu: */
1279 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1280 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1281 {
1282 	struct dma_fence *fence;
1283 	unsigned int event, i;
1284 	int ret;
1285 
1286 	ret = etnaviv_gpu_pm_get_sync(gpu);
1287 	if (ret < 0)
1288 		return ret;
1289 
1290 	/*
1291 	 * TODO
1292 	 *
1293 	 * - flush
1294 	 * - data endian
1295 	 * - prefetch
1296 	 *
1297 	 */
1298 
1299 	event = event_alloc(gpu);
1300 	if (unlikely(event == ~0U)) {
1301 		DRM_ERROR("no free event\n");
1302 		ret = -EBUSY;
1303 		goto out_pm_put;
1304 	}
1305 
1306 	fence = etnaviv_gpu_fence_alloc(gpu);
1307 	if (!fence) {
1308 		event_free(gpu, event);
1309 		ret = -ENOMEM;
1310 		goto out_pm_put;
1311 	}
1312 
1313 	mutex_lock(&gpu->lock);
1314 
1315 	gpu->event[event].fence = fence;
1316 	submit->fence = fence->seqno;
1317 	gpu->active_fence = submit->fence;
1318 
1319 	if (gpu->lastctx != cmdbuf->ctx) {
1320 		gpu->mmu->need_flush = true;
1321 		gpu->switch_context = true;
1322 		gpu->lastctx = cmdbuf->ctx;
1323 	}
1324 
1325 	etnaviv_buffer_queue(gpu, event, cmdbuf);
1326 
1327 	cmdbuf->fence = fence;
1328 	list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1329 
1330 	/* We're committed to adding this command buffer, hold a PM reference */
1331 	pm_runtime_get_noresume(gpu->dev);
1332 
1333 	for (i = 0; i < submit->nr_bos; i++) {
1334 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1335 
1336 		/* Each cmdbuf takes a refcount on the mapping */
1337 		etnaviv_gem_mapping_reference(submit->bos[i].mapping);
1338 		cmdbuf->bo_map[i] = submit->bos[i].mapping;
1339 		atomic_inc(&etnaviv_obj->gpu_active);
1340 
1341 		if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1342 			reservation_object_add_excl_fence(etnaviv_obj->resv,
1343 							  fence);
1344 		else
1345 			reservation_object_add_shared_fence(etnaviv_obj->resv,
1346 							    fence);
1347 	}
1348 	cmdbuf->nr_bos = submit->nr_bos;
1349 	hangcheck_timer_reset(gpu);
1350 	ret = 0;
1351 
1352 	mutex_unlock(&gpu->lock);
1353 
1354 out_pm_put:
1355 	etnaviv_gpu_pm_put(gpu);
1356 
1357 	return ret;
1358 }
1359 
1360 /*
1361  * Init/Cleanup:
1362  */
1363 static irqreturn_t irq_handler(int irq, void *data)
1364 {
1365 	struct etnaviv_gpu *gpu = data;
1366 	irqreturn_t ret = IRQ_NONE;
1367 
1368 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1369 
1370 	if (intr != 0) {
1371 		int event;
1372 
1373 		pm_runtime_mark_last_busy(gpu->dev);
1374 
1375 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1376 
1377 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1378 			dev_err(gpu->dev, "AXI bus error\n");
1379 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1380 		}
1381 
1382 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1383 			int i;
1384 
1385 			dev_err_ratelimited(gpu->dev,
1386 				"MMU fault status 0x%08x\n",
1387 				gpu_read(gpu, VIVS_MMUv2_STATUS));
1388 			for (i = 0; i < 4; i++) {
1389 				dev_err_ratelimited(gpu->dev,
1390 					"MMU %d fault addr 0x%08x\n",
1391 					i, gpu_read(gpu,
1392 					VIVS_MMUv2_EXCEPTION_ADDR(i)));
1393 			}
1394 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1395 		}
1396 
1397 		while ((event = ffs(intr)) != 0) {
1398 			struct dma_fence *fence;
1399 
1400 			event -= 1;
1401 
1402 			intr &= ~(1 << event);
1403 
1404 			dev_dbg(gpu->dev, "event %u\n", event);
1405 
1406 			fence = gpu->event[event].fence;
1407 			gpu->event[event].fence = NULL;
1408 			dma_fence_signal(fence);
1409 
1410 			/*
1411 			 * Events can be processed out of order.  Eg,
1412 			 * - allocate and queue event 0
1413 			 * - allocate event 1
1414 			 * - event 0 completes, we process it
1415 			 * - allocate and queue event 0
1416 			 * - event 1 and event 0 complete
1417 			 * we can end up processing event 0 first, then 1.
1418 			 */
1419 			if (fence_after(fence->seqno, gpu->completed_fence))
1420 				gpu->completed_fence = fence->seqno;
1421 
1422 			event_free(gpu, event);
1423 		}
1424 
1425 		/* Retire the buffer objects in a work */
1426 		etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1427 
1428 		ret = IRQ_HANDLED;
1429 	}
1430 
1431 	return ret;
1432 }
1433 
1434 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1435 {
1436 	int ret;
1437 
1438 	if (gpu->clk_bus) {
1439 		ret = clk_prepare_enable(gpu->clk_bus);
1440 		if (ret)
1441 			return ret;
1442 	}
1443 
1444 	if (gpu->clk_core) {
1445 		ret = clk_prepare_enable(gpu->clk_core);
1446 		if (ret)
1447 			goto disable_clk_bus;
1448 	}
1449 
1450 	if (gpu->clk_shader) {
1451 		ret = clk_prepare_enable(gpu->clk_shader);
1452 		if (ret)
1453 			goto disable_clk_core;
1454 	}
1455 
1456 	return 0;
1457 
1458 disable_clk_core:
1459 	if (gpu->clk_core)
1460 		clk_disable_unprepare(gpu->clk_core);
1461 disable_clk_bus:
1462 	if (gpu->clk_bus)
1463 		clk_disable_unprepare(gpu->clk_bus);
1464 
1465 	return ret;
1466 }
1467 
1468 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1469 {
1470 	if (gpu->clk_shader)
1471 		clk_disable_unprepare(gpu->clk_shader);
1472 	if (gpu->clk_core)
1473 		clk_disable_unprepare(gpu->clk_core);
1474 	if (gpu->clk_bus)
1475 		clk_disable_unprepare(gpu->clk_bus);
1476 
1477 	return 0;
1478 }
1479 
1480 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1481 {
1482 	unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1483 
1484 	do {
1485 		u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1486 
1487 		if ((idle & gpu->idle_mask) == gpu->idle_mask)
1488 			return 0;
1489 
1490 		if (time_is_before_jiffies(timeout)) {
1491 			dev_warn(gpu->dev,
1492 				 "timed out waiting for idle: idle=0x%x\n",
1493 				 idle);
1494 			return -ETIMEDOUT;
1495 		}
1496 
1497 		udelay(5);
1498 	} while (1);
1499 }
1500 
1501 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1502 {
1503 	if (gpu->buffer) {
1504 		/* Replace the last WAIT with END */
1505 		etnaviv_buffer_end(gpu);
1506 
1507 		/*
1508 		 * We know that only the FE is busy here, this should
1509 		 * happen quickly (as the WAIT is only 200 cycles).  If
1510 		 * we fail, just warn and continue.
1511 		 */
1512 		etnaviv_gpu_wait_idle(gpu, 100);
1513 	}
1514 
1515 	return etnaviv_gpu_clk_disable(gpu);
1516 }
1517 
1518 #ifdef CONFIG_PM
1519 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1520 {
1521 	u32 clock;
1522 	int ret;
1523 
1524 	ret = mutex_lock_killable(&gpu->lock);
1525 	if (ret)
1526 		return ret;
1527 
1528 	clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1529 		VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1530 
1531 	etnaviv_gpu_load_clock(gpu, clock);
1532 	etnaviv_gpu_hw_init(gpu);
1533 
1534 	gpu->switch_context = true;
1535 	gpu->exec_state = -1;
1536 
1537 	mutex_unlock(&gpu->lock);
1538 
1539 	return 0;
1540 }
1541 #endif
1542 
1543 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1544 	void *data)
1545 {
1546 	struct drm_device *drm = data;
1547 	struct etnaviv_drm_private *priv = drm->dev_private;
1548 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1549 	int ret;
1550 
1551 #ifdef CONFIG_PM
1552 	ret = pm_runtime_get_sync(gpu->dev);
1553 #else
1554 	ret = etnaviv_gpu_clk_enable(gpu);
1555 #endif
1556 	if (ret < 0)
1557 		return ret;
1558 
1559 	gpu->drm = drm;
1560 	gpu->fence_context = dma_fence_context_alloc(1);
1561 	spin_lock_init(&gpu->fence_spinlock);
1562 
1563 	INIT_LIST_HEAD(&gpu->active_cmd_list);
1564 	INIT_WORK(&gpu->retire_work, retire_worker);
1565 	INIT_WORK(&gpu->recover_work, recover_worker);
1566 	init_waitqueue_head(&gpu->fence_event);
1567 
1568 	setup_deferrable_timer(&gpu->hangcheck_timer, hangcheck_handler,
1569 			       (unsigned long)gpu);
1570 
1571 	priv->gpu[priv->num_gpus++] = gpu;
1572 
1573 	pm_runtime_mark_last_busy(gpu->dev);
1574 	pm_runtime_put_autosuspend(gpu->dev);
1575 
1576 	return 0;
1577 }
1578 
1579 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1580 	void *data)
1581 {
1582 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1583 
1584 	DBG("%s", dev_name(gpu->dev));
1585 
1586 	hangcheck_disable(gpu);
1587 
1588 #ifdef CONFIG_PM
1589 	pm_runtime_get_sync(gpu->dev);
1590 	pm_runtime_put_sync_suspend(gpu->dev);
1591 #else
1592 	etnaviv_gpu_hw_suspend(gpu);
1593 #endif
1594 
1595 	if (gpu->buffer) {
1596 		etnaviv_gpu_cmdbuf_free(gpu->buffer);
1597 		gpu->buffer = NULL;
1598 	}
1599 
1600 	if (gpu->mmu) {
1601 		etnaviv_iommu_destroy(gpu->mmu);
1602 		gpu->mmu = NULL;
1603 	}
1604 
1605 	gpu->drm = NULL;
1606 }
1607 
1608 static const struct component_ops gpu_ops = {
1609 	.bind = etnaviv_gpu_bind,
1610 	.unbind = etnaviv_gpu_unbind,
1611 };
1612 
1613 static const struct of_device_id etnaviv_gpu_match[] = {
1614 	{
1615 		.compatible = "vivante,gc"
1616 	},
1617 	{ /* sentinel */ }
1618 };
1619 
1620 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1621 {
1622 	struct device *dev = &pdev->dev;
1623 	struct etnaviv_gpu *gpu;
1624 	int err;
1625 
1626 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1627 	if (!gpu)
1628 		return -ENOMEM;
1629 
1630 	gpu->dev = &pdev->dev;
1631 	mutex_init(&gpu->lock);
1632 
1633 	/* Map registers: */
1634 	gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1635 	if (IS_ERR(gpu->mmio))
1636 		return PTR_ERR(gpu->mmio);
1637 
1638 	/* Get Interrupt: */
1639 	gpu->irq = platform_get_irq(pdev, 0);
1640 	if (gpu->irq < 0) {
1641 		dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1642 		return gpu->irq;
1643 	}
1644 
1645 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1646 			       dev_name(gpu->dev), gpu);
1647 	if (err) {
1648 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1649 		return err;
1650 	}
1651 
1652 	/* Get Clocks: */
1653 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1654 	DBG("clk_bus: %p", gpu->clk_bus);
1655 	if (IS_ERR(gpu->clk_bus))
1656 		gpu->clk_bus = NULL;
1657 
1658 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1659 	DBG("clk_core: %p", gpu->clk_core);
1660 	if (IS_ERR(gpu->clk_core))
1661 		gpu->clk_core = NULL;
1662 
1663 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1664 	DBG("clk_shader: %p", gpu->clk_shader);
1665 	if (IS_ERR(gpu->clk_shader))
1666 		gpu->clk_shader = NULL;
1667 
1668 	/* TODO: figure out max mapped size */
1669 	dev_set_drvdata(dev, gpu);
1670 
1671 	/*
1672 	 * We treat the device as initially suspended.  The runtime PM
1673 	 * autosuspend delay is rather arbitary: no measurements have
1674 	 * yet been performed to determine an appropriate value.
1675 	 */
1676 	pm_runtime_use_autosuspend(gpu->dev);
1677 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1678 	pm_runtime_enable(gpu->dev);
1679 
1680 	err = component_add(&pdev->dev, &gpu_ops);
1681 	if (err < 0) {
1682 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1683 		return err;
1684 	}
1685 
1686 	return 0;
1687 }
1688 
1689 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1690 {
1691 	component_del(&pdev->dev, &gpu_ops);
1692 	pm_runtime_disable(&pdev->dev);
1693 	return 0;
1694 }
1695 
1696 #ifdef CONFIG_PM
1697 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1698 {
1699 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1700 	u32 idle, mask;
1701 
1702 	/* If we have outstanding fences, we're not idle */
1703 	if (gpu->completed_fence != gpu->active_fence)
1704 		return -EBUSY;
1705 
1706 	/* Check whether the hardware (except FE) is idle */
1707 	mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1708 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1709 	if (idle != mask)
1710 		return -EBUSY;
1711 
1712 	return etnaviv_gpu_hw_suspend(gpu);
1713 }
1714 
1715 static int etnaviv_gpu_rpm_resume(struct device *dev)
1716 {
1717 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1718 	int ret;
1719 
1720 	ret = etnaviv_gpu_clk_enable(gpu);
1721 	if (ret)
1722 		return ret;
1723 
1724 	/* Re-initialise the basic hardware state */
1725 	if (gpu->drm && gpu->buffer) {
1726 		ret = etnaviv_gpu_hw_resume(gpu);
1727 		if (ret) {
1728 			etnaviv_gpu_clk_disable(gpu);
1729 			return ret;
1730 		}
1731 	}
1732 
1733 	return 0;
1734 }
1735 #endif
1736 
1737 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1738 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1739 			   NULL)
1740 };
1741 
1742 struct platform_driver etnaviv_gpu_driver = {
1743 	.driver = {
1744 		.name = "etnaviv-gpu",
1745 		.owner = THIS_MODULE,
1746 		.pm = &etnaviv_gpu_pm_ops,
1747 		.of_match_table = etnaviv_gpu_match,
1748 	},
1749 	.probe = etnaviv_gpu_platform_probe,
1750 	.remove = etnaviv_gpu_platform_remove,
1751 	.id_table = gpu_ids,
1752 };
1753