xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_gpu.c (revision 0883c2c06fb5bcf5b9e008270827e63c09a88c1e)
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/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 "etnaviv_iommu.h"
26 #include "etnaviv_iommu_v2.h"
27 #include "common.xml.h"
28 #include "state.xml.h"
29 #include "state_hi.xml.h"
30 #include "cmdstream.xml.h"
31 
32 static const struct platform_device_id gpu_ids[] = {
33 	{ .name = "etnaviv-gpu,2d" },
34 	{ },
35 };
36 
37 static bool etnaviv_dump_core = true;
38 module_param_named(dump_core, etnaviv_dump_core, bool, 0600);
39 
40 /*
41  * Driver functions:
42  */
43 
44 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
45 {
46 	switch (param) {
47 	case ETNAVIV_PARAM_GPU_MODEL:
48 		*value = gpu->identity.model;
49 		break;
50 
51 	case ETNAVIV_PARAM_GPU_REVISION:
52 		*value = gpu->identity.revision;
53 		break;
54 
55 	case ETNAVIV_PARAM_GPU_FEATURES_0:
56 		*value = gpu->identity.features;
57 		break;
58 
59 	case ETNAVIV_PARAM_GPU_FEATURES_1:
60 		*value = gpu->identity.minor_features0;
61 		break;
62 
63 	case ETNAVIV_PARAM_GPU_FEATURES_2:
64 		*value = gpu->identity.minor_features1;
65 		break;
66 
67 	case ETNAVIV_PARAM_GPU_FEATURES_3:
68 		*value = gpu->identity.minor_features2;
69 		break;
70 
71 	case ETNAVIV_PARAM_GPU_FEATURES_4:
72 		*value = gpu->identity.minor_features3;
73 		break;
74 
75 	case ETNAVIV_PARAM_GPU_FEATURES_5:
76 		*value = gpu->identity.minor_features4;
77 		break;
78 
79 	case ETNAVIV_PARAM_GPU_FEATURES_6:
80 		*value = gpu->identity.minor_features5;
81 		break;
82 
83 	case ETNAVIV_PARAM_GPU_STREAM_COUNT:
84 		*value = gpu->identity.stream_count;
85 		break;
86 
87 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
88 		*value = gpu->identity.register_max;
89 		break;
90 
91 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
92 		*value = gpu->identity.thread_count;
93 		break;
94 
95 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
96 		*value = gpu->identity.vertex_cache_size;
97 		break;
98 
99 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
100 		*value = gpu->identity.shader_core_count;
101 		break;
102 
103 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
104 		*value = gpu->identity.pixel_pipes;
105 		break;
106 
107 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
108 		*value = gpu->identity.vertex_output_buffer_size;
109 		break;
110 
111 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
112 		*value = gpu->identity.buffer_size;
113 		break;
114 
115 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
116 		*value = gpu->identity.instruction_count;
117 		break;
118 
119 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
120 		*value = gpu->identity.num_constants;
121 		break;
122 
123 	case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
124 		*value = gpu->identity.varyings_count;
125 		break;
126 
127 	default:
128 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
129 		return -EINVAL;
130 	}
131 
132 	return 0;
133 }
134 
135 
136 #define etnaviv_is_model_rev(gpu, mod, rev) \
137 	((gpu)->identity.model == chipModel_##mod && \
138 	 (gpu)->identity.revision == rev)
139 #define etnaviv_field(val, field) \
140 	(((val) & field##__MASK) >> field##__SHIFT)
141 
142 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
143 {
144 	if (gpu->identity.minor_features0 &
145 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
146 		u32 specs[4];
147 		unsigned int streams;
148 
149 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
150 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
151 		specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
152 		specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
153 
154 		gpu->identity.stream_count = etnaviv_field(specs[0],
155 					VIVS_HI_CHIP_SPECS_STREAM_COUNT);
156 		gpu->identity.register_max = etnaviv_field(specs[0],
157 					VIVS_HI_CHIP_SPECS_REGISTER_MAX);
158 		gpu->identity.thread_count = etnaviv_field(specs[0],
159 					VIVS_HI_CHIP_SPECS_THREAD_COUNT);
160 		gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
161 					VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
162 		gpu->identity.shader_core_count = etnaviv_field(specs[0],
163 					VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
164 		gpu->identity.pixel_pipes = etnaviv_field(specs[0],
165 					VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
166 		gpu->identity.vertex_output_buffer_size =
167 			etnaviv_field(specs[0],
168 				VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
169 
170 		gpu->identity.buffer_size = etnaviv_field(specs[1],
171 					VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
172 		gpu->identity.instruction_count = etnaviv_field(specs[1],
173 					VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
174 		gpu->identity.num_constants = etnaviv_field(specs[1],
175 					VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
176 
177 		gpu->identity.varyings_count = etnaviv_field(specs[2],
178 					VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
179 
180 		/* This overrides the value from older register if non-zero */
181 		streams = etnaviv_field(specs[3],
182 					VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
183 		if (streams)
184 			gpu->identity.stream_count = streams;
185 	}
186 
187 	/* Fill in the stream count if not specified */
188 	if (gpu->identity.stream_count == 0) {
189 		if (gpu->identity.model >= 0x1000)
190 			gpu->identity.stream_count = 4;
191 		else
192 			gpu->identity.stream_count = 1;
193 	}
194 
195 	/* Convert the register max value */
196 	if (gpu->identity.register_max)
197 		gpu->identity.register_max = 1 << gpu->identity.register_max;
198 	else if (gpu->identity.model == chipModel_GC400)
199 		gpu->identity.register_max = 32;
200 	else
201 		gpu->identity.register_max = 64;
202 
203 	/* Convert thread count */
204 	if (gpu->identity.thread_count)
205 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
206 	else if (gpu->identity.model == chipModel_GC400)
207 		gpu->identity.thread_count = 64;
208 	else if (gpu->identity.model == chipModel_GC500 ||
209 		 gpu->identity.model == chipModel_GC530)
210 		gpu->identity.thread_count = 128;
211 	else
212 		gpu->identity.thread_count = 256;
213 
214 	if (gpu->identity.vertex_cache_size == 0)
215 		gpu->identity.vertex_cache_size = 8;
216 
217 	if (gpu->identity.shader_core_count == 0) {
218 		if (gpu->identity.model >= 0x1000)
219 			gpu->identity.shader_core_count = 2;
220 		else
221 			gpu->identity.shader_core_count = 1;
222 	}
223 
224 	if (gpu->identity.pixel_pipes == 0)
225 		gpu->identity.pixel_pipes = 1;
226 
227 	/* Convert virtex buffer size */
228 	if (gpu->identity.vertex_output_buffer_size) {
229 		gpu->identity.vertex_output_buffer_size =
230 			1 << gpu->identity.vertex_output_buffer_size;
231 	} else if (gpu->identity.model == chipModel_GC400) {
232 		if (gpu->identity.revision < 0x4000)
233 			gpu->identity.vertex_output_buffer_size = 512;
234 		else if (gpu->identity.revision < 0x4200)
235 			gpu->identity.vertex_output_buffer_size = 256;
236 		else
237 			gpu->identity.vertex_output_buffer_size = 128;
238 	} else {
239 		gpu->identity.vertex_output_buffer_size = 512;
240 	}
241 
242 	switch (gpu->identity.instruction_count) {
243 	case 0:
244 		if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
245 		    gpu->identity.model == chipModel_GC880)
246 			gpu->identity.instruction_count = 512;
247 		else
248 			gpu->identity.instruction_count = 256;
249 		break;
250 
251 	case 1:
252 		gpu->identity.instruction_count = 1024;
253 		break;
254 
255 	case 2:
256 		gpu->identity.instruction_count = 2048;
257 		break;
258 
259 	default:
260 		gpu->identity.instruction_count = 256;
261 		break;
262 	}
263 
264 	if (gpu->identity.num_constants == 0)
265 		gpu->identity.num_constants = 168;
266 
267 	if (gpu->identity.varyings_count == 0) {
268 		if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
269 			gpu->identity.varyings_count = 12;
270 		else
271 			gpu->identity.varyings_count = 8;
272 	}
273 
274 	/*
275 	 * For some cores, two varyings are consumed for position, so the
276 	 * maximum varying count needs to be reduced by one.
277 	 */
278 	if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
279 	    etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
280 	    etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
281 	    etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
282 	    etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
283 	    etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
284 	    etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
285 	    etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
286 	    etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
287 	    etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
288 	    etnaviv_is_model_rev(gpu, GC880, 0x5106))
289 		gpu->identity.varyings_count -= 1;
290 }
291 
292 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
293 {
294 	u32 chipIdentity;
295 
296 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
297 
298 	/* Special case for older graphic cores. */
299 	if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
300 		gpu->identity.model    = chipModel_GC500;
301 		gpu->identity.revision = etnaviv_field(chipIdentity,
302 					 VIVS_HI_CHIP_IDENTITY_REVISION);
303 	} else {
304 
305 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
306 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
307 
308 		/*
309 		 * !!!! HACK ALERT !!!!
310 		 * Because people change device IDs without letting software
311 		 * know about it - here is the hack to make it all look the
312 		 * same.  Only for GC400 family.
313 		 */
314 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
315 		    gpu->identity.model != chipModel_GC420) {
316 			gpu->identity.model = gpu->identity.model & 0x0400;
317 		}
318 
319 		/* Another special case */
320 		if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
321 			u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
322 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
323 
324 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
325 				/*
326 				 * This IP has an ECO; put the correct
327 				 * revision in it.
328 				 */
329 				gpu->identity.revision = 0x1051;
330 			}
331 		}
332 	}
333 
334 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
335 		 gpu->identity.model, gpu->identity.revision);
336 
337 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
338 
339 	/* Disable fast clear on GC700. */
340 	if (gpu->identity.model == chipModel_GC700)
341 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
342 
343 	if ((gpu->identity.model == chipModel_GC500 &&
344 	     gpu->identity.revision < 2) ||
345 	    (gpu->identity.model == chipModel_GC300 &&
346 	     gpu->identity.revision < 0x2000)) {
347 
348 		/*
349 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
350 		 * registers.
351 		 */
352 		gpu->identity.minor_features0 = 0;
353 		gpu->identity.minor_features1 = 0;
354 		gpu->identity.minor_features2 = 0;
355 		gpu->identity.minor_features3 = 0;
356 		gpu->identity.minor_features4 = 0;
357 		gpu->identity.minor_features5 = 0;
358 	} else
359 		gpu->identity.minor_features0 =
360 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
361 
362 	if (gpu->identity.minor_features0 &
363 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
364 		gpu->identity.minor_features1 =
365 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
366 		gpu->identity.minor_features2 =
367 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
368 		gpu->identity.minor_features3 =
369 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
370 		gpu->identity.minor_features4 =
371 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
372 		gpu->identity.minor_features5 =
373 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
374 	}
375 
376 	/* GC600 idle register reports zero bits where modules aren't present */
377 	if (gpu->identity.model == chipModel_GC600) {
378 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
379 				 VIVS_HI_IDLE_STATE_RA |
380 				 VIVS_HI_IDLE_STATE_SE |
381 				 VIVS_HI_IDLE_STATE_PA |
382 				 VIVS_HI_IDLE_STATE_SH |
383 				 VIVS_HI_IDLE_STATE_PE |
384 				 VIVS_HI_IDLE_STATE_DE |
385 				 VIVS_HI_IDLE_STATE_FE;
386 	} else {
387 		gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
388 	}
389 
390 	etnaviv_hw_specs(gpu);
391 }
392 
393 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
394 {
395 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
396 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
397 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
398 }
399 
400 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
401 {
402 	u32 control, idle;
403 	unsigned long timeout;
404 	bool failed = true;
405 
406 	/* TODO
407 	 *
408 	 * - clock gating
409 	 * - puls eater
410 	 * - what about VG?
411 	 */
412 
413 	/* We hope that the GPU resets in under one second */
414 	timeout = jiffies + msecs_to_jiffies(1000);
415 
416 	while (time_is_after_jiffies(timeout)) {
417 		control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
418 			  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
419 
420 		/* enable clock */
421 		etnaviv_gpu_load_clock(gpu, control);
422 
423 		/* Wait for stable clock.  Vivante's code waited for 1ms */
424 		usleep_range(1000, 10000);
425 
426 		/* isolate the GPU. */
427 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
428 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
429 
430 		/* set soft reset. */
431 		control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
432 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
433 
434 		/* wait for reset. */
435 		msleep(1);
436 
437 		/* reset soft reset bit. */
438 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
439 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
440 
441 		/* reset GPU isolation. */
442 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
443 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
444 
445 		/* read idle register. */
446 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
447 
448 		/* try reseting again if FE it not idle */
449 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
450 			dev_dbg(gpu->dev, "FE is not idle\n");
451 			continue;
452 		}
453 
454 		/* read reset register. */
455 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
456 
457 		/* is the GPU idle? */
458 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
459 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
460 			dev_dbg(gpu->dev, "GPU is not idle\n");
461 			continue;
462 		}
463 
464 		failed = false;
465 		break;
466 	}
467 
468 	if (failed) {
469 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
470 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
471 
472 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
473 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
474 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
475 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
476 
477 		return -EBUSY;
478 	}
479 
480 	/* We rely on the GPU running, so program the clock */
481 	control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
482 		  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
483 
484 	/* enable clock */
485 	etnaviv_gpu_load_clock(gpu, control);
486 
487 	return 0;
488 }
489 
490 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
491 {
492 	u16 prefetch;
493 
494 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
495 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
496 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
497 		u32 mc_memory_debug;
498 
499 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
500 
501 		if (gpu->identity.revision == 0x5007)
502 			mc_memory_debug |= 0x0c;
503 		else
504 			mc_memory_debug |= 0x08;
505 
506 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
507 	}
508 
509 	/*
510 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
511 	 * This is necessary to prevent the iMX6 SoC locking up.
512 	 */
513 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
514 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
515 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
516 
517 	/* GC2000 rev 5108 needs a special bus config */
518 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
519 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
520 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
521 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
522 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
523 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
524 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
525 	}
526 
527 	/* set base addresses */
528 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_RA, gpu->memory_base);
529 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_FE, gpu->memory_base);
530 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_TX, gpu->memory_base);
531 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PEZ, gpu->memory_base);
532 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PE, gpu->memory_base);
533 
534 	/* setup the MMU page table pointers */
535 	etnaviv_iommu_domain_restore(gpu, gpu->mmu->domain);
536 
537 	/* Start command processor */
538 	prefetch = etnaviv_buffer_init(gpu);
539 
540 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
541 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS,
542 		  gpu->buffer->paddr - gpu->memory_base);
543 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
544 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
545 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
546 }
547 
548 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
549 {
550 	int ret, i;
551 	struct iommu_domain *iommu;
552 	enum etnaviv_iommu_version version;
553 	bool mmuv2;
554 
555 	ret = pm_runtime_get_sync(gpu->dev);
556 	if (ret < 0)
557 		return ret;
558 
559 	etnaviv_hw_identify(gpu);
560 
561 	if (gpu->identity.model == 0) {
562 		dev_err(gpu->dev, "Unknown GPU model\n");
563 		ret = -ENXIO;
564 		goto fail;
565 	}
566 
567 	/* Exclude VG cores with FE2.0 */
568 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
569 	    gpu->identity.features & chipFeatures_FE20) {
570 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
571 		ret = -ENXIO;
572 		goto fail;
573 	}
574 
575 	/*
576 	 * Set the GPU linear window to be at the end of the DMA window, where
577 	 * the CMA area is likely to reside. This ensures that we are able to
578 	 * map the command buffers while having the linear window overlap as
579 	 * much RAM as possible, so we can optimize mappings for other buffers.
580 	 *
581 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
582 	 * to different views of the memory on the individual engines.
583 	 */
584 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
585 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
586 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
587 		if (dma_mask < PHYS_OFFSET + SZ_2G)
588 			gpu->memory_base = PHYS_OFFSET;
589 		else
590 			gpu->memory_base = dma_mask - SZ_2G + 1;
591 	}
592 
593 	ret = etnaviv_hw_reset(gpu);
594 	if (ret)
595 		goto fail;
596 
597 	/* Setup IOMMU.. eventually we will (I think) do this once per context
598 	 * and have separate page tables per context.  For now, to keep things
599 	 * simple and to get something working, just use a single address space:
600 	 */
601 	mmuv2 = gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION;
602 	dev_dbg(gpu->dev, "mmuv2: %d\n", mmuv2);
603 
604 	if (!mmuv2) {
605 		iommu = etnaviv_iommu_domain_alloc(gpu);
606 		version = ETNAVIV_IOMMU_V1;
607 	} else {
608 		iommu = etnaviv_iommu_v2_domain_alloc(gpu);
609 		version = ETNAVIV_IOMMU_V2;
610 	}
611 
612 	if (!iommu) {
613 		ret = -ENOMEM;
614 		goto fail;
615 	}
616 
617 	gpu->mmu = etnaviv_iommu_new(gpu, iommu, version);
618 	if (!gpu->mmu) {
619 		iommu_domain_free(iommu);
620 		ret = -ENOMEM;
621 		goto fail;
622 	}
623 
624 	/* Create buffer: */
625 	gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
626 	if (!gpu->buffer) {
627 		ret = -ENOMEM;
628 		dev_err(gpu->dev, "could not create command buffer\n");
629 		goto destroy_iommu;
630 	}
631 	if (gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
632 		ret = -EINVAL;
633 		dev_err(gpu->dev,
634 			"command buffer outside valid memory window\n");
635 		goto free_buffer;
636 	}
637 
638 	/* Setup event management */
639 	spin_lock_init(&gpu->event_spinlock);
640 	init_completion(&gpu->event_free);
641 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
642 		gpu->event[i].used = false;
643 		complete(&gpu->event_free);
644 	}
645 
646 	/* Now program the hardware */
647 	mutex_lock(&gpu->lock);
648 	etnaviv_gpu_hw_init(gpu);
649 	gpu->exec_state = -1;
650 	mutex_unlock(&gpu->lock);
651 
652 	pm_runtime_mark_last_busy(gpu->dev);
653 	pm_runtime_put_autosuspend(gpu->dev);
654 
655 	return 0;
656 
657 free_buffer:
658 	etnaviv_gpu_cmdbuf_free(gpu->buffer);
659 	gpu->buffer = NULL;
660 destroy_iommu:
661 	etnaviv_iommu_destroy(gpu->mmu);
662 	gpu->mmu = NULL;
663 fail:
664 	pm_runtime_mark_last_busy(gpu->dev);
665 	pm_runtime_put_autosuspend(gpu->dev);
666 
667 	return ret;
668 }
669 
670 #ifdef CONFIG_DEBUG_FS
671 struct dma_debug {
672 	u32 address[2];
673 	u32 state[2];
674 };
675 
676 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
677 {
678 	u32 i;
679 
680 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
681 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
682 
683 	for (i = 0; i < 500; i++) {
684 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
685 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
686 
687 		if (debug->address[0] != debug->address[1])
688 			break;
689 
690 		if (debug->state[0] != debug->state[1])
691 			break;
692 	}
693 }
694 
695 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
696 {
697 	struct dma_debug debug;
698 	u32 dma_lo, dma_hi, axi, idle;
699 	int ret;
700 
701 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
702 
703 	ret = pm_runtime_get_sync(gpu->dev);
704 	if (ret < 0)
705 		return ret;
706 
707 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
708 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
709 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
710 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
711 
712 	verify_dma(gpu, &debug);
713 
714 	seq_puts(m, "\tfeatures\n");
715 	seq_printf(m, "\t minor_features0: 0x%08x\n",
716 		   gpu->identity.minor_features0);
717 	seq_printf(m, "\t minor_features1: 0x%08x\n",
718 		   gpu->identity.minor_features1);
719 	seq_printf(m, "\t minor_features2: 0x%08x\n",
720 		   gpu->identity.minor_features2);
721 	seq_printf(m, "\t minor_features3: 0x%08x\n",
722 		   gpu->identity.minor_features3);
723 	seq_printf(m, "\t minor_features4: 0x%08x\n",
724 		   gpu->identity.minor_features4);
725 	seq_printf(m, "\t minor_features5: 0x%08x\n",
726 		   gpu->identity.minor_features5);
727 
728 	seq_puts(m, "\tspecs\n");
729 	seq_printf(m, "\t stream_count:  %d\n",
730 			gpu->identity.stream_count);
731 	seq_printf(m, "\t register_max: %d\n",
732 			gpu->identity.register_max);
733 	seq_printf(m, "\t thread_count: %d\n",
734 			gpu->identity.thread_count);
735 	seq_printf(m, "\t vertex_cache_size: %d\n",
736 			gpu->identity.vertex_cache_size);
737 	seq_printf(m, "\t shader_core_count: %d\n",
738 			gpu->identity.shader_core_count);
739 	seq_printf(m, "\t pixel_pipes: %d\n",
740 			gpu->identity.pixel_pipes);
741 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
742 			gpu->identity.vertex_output_buffer_size);
743 	seq_printf(m, "\t buffer_size: %d\n",
744 			gpu->identity.buffer_size);
745 	seq_printf(m, "\t instruction_count: %d\n",
746 			gpu->identity.instruction_count);
747 	seq_printf(m, "\t num_constants: %d\n",
748 			gpu->identity.num_constants);
749 	seq_printf(m, "\t varyings_count: %d\n",
750 			gpu->identity.varyings_count);
751 
752 	seq_printf(m, "\taxi: 0x%08x\n", axi);
753 	seq_printf(m, "\tidle: 0x%08x\n", idle);
754 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
755 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
756 		seq_puts(m, "\t FE is not idle\n");
757 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
758 		seq_puts(m, "\t DE is not idle\n");
759 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
760 		seq_puts(m, "\t PE is not idle\n");
761 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
762 		seq_puts(m, "\t SH is not idle\n");
763 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
764 		seq_puts(m, "\t PA is not idle\n");
765 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
766 		seq_puts(m, "\t SE is not idle\n");
767 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
768 		seq_puts(m, "\t RA is not idle\n");
769 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
770 		seq_puts(m, "\t TX is not idle\n");
771 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
772 		seq_puts(m, "\t VG is not idle\n");
773 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
774 		seq_puts(m, "\t IM is not idle\n");
775 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
776 		seq_puts(m, "\t FP is not idle\n");
777 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
778 		seq_puts(m, "\t TS is not idle\n");
779 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
780 		seq_puts(m, "\t AXI low power mode\n");
781 
782 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
783 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
784 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
785 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
786 
787 		seq_puts(m, "\tMC\n");
788 		seq_printf(m, "\t read0: 0x%08x\n", read0);
789 		seq_printf(m, "\t read1: 0x%08x\n", read1);
790 		seq_printf(m, "\t write: 0x%08x\n", write);
791 	}
792 
793 	seq_puts(m, "\tDMA ");
794 
795 	if (debug.address[0] == debug.address[1] &&
796 	    debug.state[0] == debug.state[1]) {
797 		seq_puts(m, "seems to be stuck\n");
798 	} else if (debug.address[0] == debug.address[1]) {
799 		seq_puts(m, "address is constant\n");
800 	} else {
801 		seq_puts(m, "is running\n");
802 	}
803 
804 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
805 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
806 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
807 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
808 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
809 		   dma_lo, dma_hi);
810 
811 	ret = 0;
812 
813 	pm_runtime_mark_last_busy(gpu->dev);
814 	pm_runtime_put_autosuspend(gpu->dev);
815 
816 	return ret;
817 }
818 #endif
819 
820 /*
821  * Power Management:
822  */
823 static int enable_clk(struct etnaviv_gpu *gpu)
824 {
825 	if (gpu->clk_core)
826 		clk_prepare_enable(gpu->clk_core);
827 	if (gpu->clk_shader)
828 		clk_prepare_enable(gpu->clk_shader);
829 
830 	return 0;
831 }
832 
833 static int disable_clk(struct etnaviv_gpu *gpu)
834 {
835 	if (gpu->clk_core)
836 		clk_disable_unprepare(gpu->clk_core);
837 	if (gpu->clk_shader)
838 		clk_disable_unprepare(gpu->clk_shader);
839 
840 	return 0;
841 }
842 
843 static int enable_axi(struct etnaviv_gpu *gpu)
844 {
845 	if (gpu->clk_bus)
846 		clk_prepare_enable(gpu->clk_bus);
847 
848 	return 0;
849 }
850 
851 static int disable_axi(struct etnaviv_gpu *gpu)
852 {
853 	if (gpu->clk_bus)
854 		clk_disable_unprepare(gpu->clk_bus);
855 
856 	return 0;
857 }
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 		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->switch_context = true;
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 fence base;
960 };
961 
962 static inline struct etnaviv_fence *to_etnaviv_fence(struct fence *fence)
963 {
964 	return container_of(fence, struct etnaviv_fence, base);
965 }
966 
967 static const char *etnaviv_fence_get_driver_name(struct fence *fence)
968 {
969 	return "etnaviv";
970 }
971 
972 static const char *etnaviv_fence_get_timeline_name(struct 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 fence *fence)
980 {
981 	return true;
982 }
983 
984 static bool etnaviv_fence_signaled(struct 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 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 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 = fence_default_wait,
1004 	.release = etnaviv_fence_release,
1005 };
1006 
1007 static struct 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 	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 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 = 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 = 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 	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
1132 				     GFP_KERNEL);
1133 	if (!cmdbuf->vaddr) {
1134 		kfree(cmdbuf);
1135 		return NULL;
1136 	}
1137 
1138 	cmdbuf->gpu = gpu;
1139 	cmdbuf->size = size;
1140 
1141 	return cmdbuf;
1142 }
1143 
1144 void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1145 {
1146 	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
1147 		    cmdbuf->paddr);
1148 	kfree(cmdbuf);
1149 }
1150 
1151 static void retire_worker(struct work_struct *work)
1152 {
1153 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1154 					       retire_work);
1155 	u32 fence = gpu->completed_fence;
1156 	struct etnaviv_cmdbuf *cmdbuf, *tmp;
1157 	unsigned int i;
1158 
1159 	mutex_lock(&gpu->lock);
1160 	list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1161 		if (!fence_is_signaled(cmdbuf->fence))
1162 			break;
1163 
1164 		list_del(&cmdbuf->node);
1165 		fence_put(cmdbuf->fence);
1166 
1167 		for (i = 0; i < cmdbuf->nr_bos; i++) {
1168 			struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
1169 			struct etnaviv_gem_object *etnaviv_obj = mapping->object;
1170 
1171 			atomic_dec(&etnaviv_obj->gpu_active);
1172 			/* drop the refcount taken in etnaviv_gpu_submit */
1173 			etnaviv_gem_mapping_unreference(mapping);
1174 		}
1175 
1176 		etnaviv_gpu_cmdbuf_free(cmdbuf);
1177 		/*
1178 		 * We need to balance the runtime PM count caused by
1179 		 * each submission.  Upon submission, we increment
1180 		 * the runtime PM counter, and allocate one event.
1181 		 * So here, we put the runtime PM count for each
1182 		 * completed event.
1183 		 */
1184 		pm_runtime_put_autosuspend(gpu->dev);
1185 	}
1186 
1187 	gpu->retired_fence = fence;
1188 
1189 	mutex_unlock(&gpu->lock);
1190 
1191 	wake_up_all(&gpu->fence_event);
1192 }
1193 
1194 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1195 	u32 fence, struct timespec *timeout)
1196 {
1197 	int ret;
1198 
1199 	if (fence_after(fence, gpu->next_fence)) {
1200 		DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1201 				fence, gpu->next_fence);
1202 		return -EINVAL;
1203 	}
1204 
1205 	if (!timeout) {
1206 		/* No timeout was requested: just test for completion */
1207 		ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1208 	} else {
1209 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1210 
1211 		ret = wait_event_interruptible_timeout(gpu->fence_event,
1212 						fence_completed(gpu, fence),
1213 						remaining);
1214 		if (ret == 0) {
1215 			DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1216 				fence, gpu->retired_fence,
1217 				gpu->completed_fence);
1218 			ret = -ETIMEDOUT;
1219 		} else if (ret != -ERESTARTSYS) {
1220 			ret = 0;
1221 		}
1222 	}
1223 
1224 	return ret;
1225 }
1226 
1227 /*
1228  * Wait for an object to become inactive.  This, on it's own, is not race
1229  * free: the object is moved by the retire worker off the active list, and
1230  * then the iova is put.  Moreover, the object could be re-submitted just
1231  * after we notice that it's become inactive.
1232  *
1233  * Although the retirement happens under the gpu lock, we don't want to hold
1234  * that lock in this function while waiting.
1235  */
1236 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1237 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1238 {
1239 	unsigned long remaining;
1240 	long ret;
1241 
1242 	if (!timeout)
1243 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1244 
1245 	remaining = etnaviv_timeout_to_jiffies(timeout);
1246 
1247 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1248 					       !is_active(etnaviv_obj),
1249 					       remaining);
1250 	if (ret > 0) {
1251 		struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1252 
1253 		/* Synchronise with the retire worker */
1254 		flush_workqueue(priv->wq);
1255 		return 0;
1256 	} else if (ret == -ERESTARTSYS) {
1257 		return -ERESTARTSYS;
1258 	} else {
1259 		return -ETIMEDOUT;
1260 	}
1261 }
1262 
1263 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1264 {
1265 	return pm_runtime_get_sync(gpu->dev);
1266 }
1267 
1268 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1269 {
1270 	pm_runtime_mark_last_busy(gpu->dev);
1271 	pm_runtime_put_autosuspend(gpu->dev);
1272 }
1273 
1274 /* add bo's to gpu's ring, and kick gpu: */
1275 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1276 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1277 {
1278 	struct fence *fence;
1279 	unsigned int event, i;
1280 	int ret;
1281 
1282 	ret = etnaviv_gpu_pm_get_sync(gpu);
1283 	if (ret < 0)
1284 		return ret;
1285 
1286 	mutex_lock(&gpu->lock);
1287 
1288 	/*
1289 	 * TODO
1290 	 *
1291 	 * - flush
1292 	 * - data endian
1293 	 * - prefetch
1294 	 *
1295 	 */
1296 
1297 	event = event_alloc(gpu);
1298 	if (unlikely(event == ~0U)) {
1299 		DRM_ERROR("no free event\n");
1300 		ret = -EBUSY;
1301 		goto out_unlock;
1302 	}
1303 
1304 	fence = etnaviv_gpu_fence_alloc(gpu);
1305 	if (!fence) {
1306 		event_free(gpu, event);
1307 		ret = -ENOMEM;
1308 		goto out_unlock;
1309 	}
1310 
1311 	gpu->event[event].fence = fence;
1312 	submit->fence = fence->seqno;
1313 	gpu->active_fence = submit->fence;
1314 
1315 	if (gpu->lastctx != cmdbuf->ctx) {
1316 		gpu->mmu->need_flush = true;
1317 		gpu->switch_context = true;
1318 		gpu->lastctx = cmdbuf->ctx;
1319 	}
1320 
1321 	etnaviv_buffer_queue(gpu, event, cmdbuf);
1322 
1323 	cmdbuf->fence = fence;
1324 	list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1325 
1326 	/* We're committed to adding this command buffer, hold a PM reference */
1327 	pm_runtime_get_noresume(gpu->dev);
1328 
1329 	for (i = 0; i < submit->nr_bos; i++) {
1330 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1331 
1332 		/* Each cmdbuf takes a refcount on the mapping */
1333 		etnaviv_gem_mapping_reference(submit->bos[i].mapping);
1334 		cmdbuf->bo_map[i] = submit->bos[i].mapping;
1335 		atomic_inc(&etnaviv_obj->gpu_active);
1336 
1337 		if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1338 			reservation_object_add_excl_fence(etnaviv_obj->resv,
1339 							  fence);
1340 		else
1341 			reservation_object_add_shared_fence(etnaviv_obj->resv,
1342 							    fence);
1343 	}
1344 	cmdbuf->nr_bos = submit->nr_bos;
1345 	hangcheck_timer_reset(gpu);
1346 	ret = 0;
1347 
1348 out_unlock:
1349 	mutex_unlock(&gpu->lock);
1350 
1351 	etnaviv_gpu_pm_put(gpu);
1352 
1353 	return ret;
1354 }
1355 
1356 /*
1357  * Init/Cleanup:
1358  */
1359 static irqreturn_t irq_handler(int irq, void *data)
1360 {
1361 	struct etnaviv_gpu *gpu = data;
1362 	irqreturn_t ret = IRQ_NONE;
1363 
1364 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1365 
1366 	if (intr != 0) {
1367 		int event;
1368 
1369 		pm_runtime_mark_last_busy(gpu->dev);
1370 
1371 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1372 
1373 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1374 			dev_err(gpu->dev, "AXI bus error\n");
1375 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1376 		}
1377 
1378 		while ((event = ffs(intr)) != 0) {
1379 			struct fence *fence;
1380 
1381 			event -= 1;
1382 
1383 			intr &= ~(1 << event);
1384 
1385 			dev_dbg(gpu->dev, "event %u\n", event);
1386 
1387 			fence = gpu->event[event].fence;
1388 			gpu->event[event].fence = NULL;
1389 			fence_signal(fence);
1390 
1391 			/*
1392 			 * Events can be processed out of order.  Eg,
1393 			 * - allocate and queue event 0
1394 			 * - allocate event 1
1395 			 * - event 0 completes, we process it
1396 			 * - allocate and queue event 0
1397 			 * - event 1 and event 0 complete
1398 			 * we can end up processing event 0 first, then 1.
1399 			 */
1400 			if (fence_after(fence->seqno, gpu->completed_fence))
1401 				gpu->completed_fence = fence->seqno;
1402 
1403 			event_free(gpu, event);
1404 		}
1405 
1406 		/* Retire the buffer objects in a work */
1407 		etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1408 
1409 		ret = IRQ_HANDLED;
1410 	}
1411 
1412 	return ret;
1413 }
1414 
1415 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1416 {
1417 	int ret;
1418 
1419 	ret = enable_clk(gpu);
1420 	if (ret)
1421 		return ret;
1422 
1423 	ret = enable_axi(gpu);
1424 	if (ret) {
1425 		disable_clk(gpu);
1426 		return ret;
1427 	}
1428 
1429 	return 0;
1430 }
1431 
1432 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1433 {
1434 	int ret;
1435 
1436 	ret = disable_axi(gpu);
1437 	if (ret)
1438 		return ret;
1439 
1440 	ret = disable_clk(gpu);
1441 	if (ret)
1442 		return ret;
1443 
1444 	return 0;
1445 }
1446 
1447 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1448 {
1449 	if (gpu->buffer) {
1450 		unsigned long timeout;
1451 
1452 		/* Replace the last WAIT with END */
1453 		etnaviv_buffer_end(gpu);
1454 
1455 		/*
1456 		 * We know that only the FE is busy here, this should
1457 		 * happen quickly (as the WAIT is only 200 cycles).  If
1458 		 * we fail, just warn and continue.
1459 		 */
1460 		timeout = jiffies + msecs_to_jiffies(100);
1461 		do {
1462 			u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1463 
1464 			if ((idle & gpu->idle_mask) == gpu->idle_mask)
1465 				break;
1466 
1467 			if (time_is_before_jiffies(timeout)) {
1468 				dev_warn(gpu->dev,
1469 					 "timed out waiting for idle: idle=0x%x\n",
1470 					 idle);
1471 				break;
1472 			}
1473 
1474 			udelay(5);
1475 		} while (1);
1476 	}
1477 
1478 	return etnaviv_gpu_clk_disable(gpu);
1479 }
1480 
1481 #ifdef CONFIG_PM
1482 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1483 {
1484 	u32 clock;
1485 	int ret;
1486 
1487 	ret = mutex_lock_killable(&gpu->lock);
1488 	if (ret)
1489 		return ret;
1490 
1491 	clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1492 		VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1493 
1494 	etnaviv_gpu_load_clock(gpu, clock);
1495 	etnaviv_gpu_hw_init(gpu);
1496 
1497 	gpu->switch_context = true;
1498 	gpu->exec_state = -1;
1499 
1500 	mutex_unlock(&gpu->lock);
1501 
1502 	return 0;
1503 }
1504 #endif
1505 
1506 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1507 	void *data)
1508 {
1509 	struct drm_device *drm = data;
1510 	struct etnaviv_drm_private *priv = drm->dev_private;
1511 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1512 	int ret;
1513 
1514 #ifdef CONFIG_PM
1515 	ret = pm_runtime_get_sync(gpu->dev);
1516 #else
1517 	ret = etnaviv_gpu_clk_enable(gpu);
1518 #endif
1519 	if (ret < 0)
1520 		return ret;
1521 
1522 	gpu->drm = drm;
1523 	gpu->fence_context = fence_context_alloc(1);
1524 	spin_lock_init(&gpu->fence_spinlock);
1525 
1526 	INIT_LIST_HEAD(&gpu->active_cmd_list);
1527 	INIT_WORK(&gpu->retire_work, retire_worker);
1528 	INIT_WORK(&gpu->recover_work, recover_worker);
1529 	init_waitqueue_head(&gpu->fence_event);
1530 
1531 	setup_deferrable_timer(&gpu->hangcheck_timer, hangcheck_handler,
1532 			       (unsigned long)gpu);
1533 
1534 	priv->gpu[priv->num_gpus++] = gpu;
1535 
1536 	pm_runtime_mark_last_busy(gpu->dev);
1537 	pm_runtime_put_autosuspend(gpu->dev);
1538 
1539 	return 0;
1540 }
1541 
1542 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1543 	void *data)
1544 {
1545 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1546 
1547 	DBG("%s", dev_name(gpu->dev));
1548 
1549 	hangcheck_disable(gpu);
1550 
1551 #ifdef CONFIG_PM
1552 	pm_runtime_get_sync(gpu->dev);
1553 	pm_runtime_put_sync_suspend(gpu->dev);
1554 #else
1555 	etnaviv_gpu_hw_suspend(gpu);
1556 #endif
1557 
1558 	if (gpu->buffer) {
1559 		etnaviv_gpu_cmdbuf_free(gpu->buffer);
1560 		gpu->buffer = NULL;
1561 	}
1562 
1563 	if (gpu->mmu) {
1564 		etnaviv_iommu_destroy(gpu->mmu);
1565 		gpu->mmu = NULL;
1566 	}
1567 
1568 	gpu->drm = NULL;
1569 }
1570 
1571 static const struct component_ops gpu_ops = {
1572 	.bind = etnaviv_gpu_bind,
1573 	.unbind = etnaviv_gpu_unbind,
1574 };
1575 
1576 static const struct of_device_id etnaviv_gpu_match[] = {
1577 	{
1578 		.compatible = "vivante,gc"
1579 	},
1580 	{ /* sentinel */ }
1581 };
1582 
1583 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1584 {
1585 	struct device *dev = &pdev->dev;
1586 	struct etnaviv_gpu *gpu;
1587 	int err = 0;
1588 
1589 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1590 	if (!gpu)
1591 		return -ENOMEM;
1592 
1593 	gpu->dev = &pdev->dev;
1594 	mutex_init(&gpu->lock);
1595 
1596 	/* Map registers: */
1597 	gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1598 	if (IS_ERR(gpu->mmio))
1599 		return PTR_ERR(gpu->mmio);
1600 
1601 	/* Get Interrupt: */
1602 	gpu->irq = platform_get_irq(pdev, 0);
1603 	if (gpu->irq < 0) {
1604 		err = gpu->irq;
1605 		dev_err(dev, "failed to get irq: %d\n", err);
1606 		goto fail;
1607 	}
1608 
1609 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1610 			       dev_name(gpu->dev), gpu);
1611 	if (err) {
1612 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1613 		goto fail;
1614 	}
1615 
1616 	/* Get Clocks: */
1617 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1618 	DBG("clk_bus: %p", gpu->clk_bus);
1619 	if (IS_ERR(gpu->clk_bus))
1620 		gpu->clk_bus = NULL;
1621 
1622 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1623 	DBG("clk_core: %p", gpu->clk_core);
1624 	if (IS_ERR(gpu->clk_core))
1625 		gpu->clk_core = NULL;
1626 
1627 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1628 	DBG("clk_shader: %p", gpu->clk_shader);
1629 	if (IS_ERR(gpu->clk_shader))
1630 		gpu->clk_shader = NULL;
1631 
1632 	/* TODO: figure out max mapped size */
1633 	dev_set_drvdata(dev, gpu);
1634 
1635 	/*
1636 	 * We treat the device as initially suspended.  The runtime PM
1637 	 * autosuspend delay is rather arbitary: no measurements have
1638 	 * yet been performed to determine an appropriate value.
1639 	 */
1640 	pm_runtime_use_autosuspend(gpu->dev);
1641 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1642 	pm_runtime_enable(gpu->dev);
1643 
1644 	err = component_add(&pdev->dev, &gpu_ops);
1645 	if (err < 0) {
1646 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1647 		goto fail;
1648 	}
1649 
1650 	return 0;
1651 
1652 fail:
1653 	return err;
1654 }
1655 
1656 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1657 {
1658 	component_del(&pdev->dev, &gpu_ops);
1659 	pm_runtime_disable(&pdev->dev);
1660 	return 0;
1661 }
1662 
1663 #ifdef CONFIG_PM
1664 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1665 {
1666 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1667 	u32 idle, mask;
1668 
1669 	/* If we have outstanding fences, we're not idle */
1670 	if (gpu->completed_fence != gpu->active_fence)
1671 		return -EBUSY;
1672 
1673 	/* Check whether the hardware (except FE) is idle */
1674 	mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1675 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1676 	if (idle != mask)
1677 		return -EBUSY;
1678 
1679 	return etnaviv_gpu_hw_suspend(gpu);
1680 }
1681 
1682 static int etnaviv_gpu_rpm_resume(struct device *dev)
1683 {
1684 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1685 	int ret;
1686 
1687 	ret = etnaviv_gpu_clk_enable(gpu);
1688 	if (ret)
1689 		return ret;
1690 
1691 	/* Re-initialise the basic hardware state */
1692 	if (gpu->drm && gpu->buffer) {
1693 		ret = etnaviv_gpu_hw_resume(gpu);
1694 		if (ret) {
1695 			etnaviv_gpu_clk_disable(gpu);
1696 			return ret;
1697 		}
1698 	}
1699 
1700 	return 0;
1701 }
1702 #endif
1703 
1704 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1705 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1706 			   NULL)
1707 };
1708 
1709 struct platform_driver etnaviv_gpu_driver = {
1710 	.driver = {
1711 		.name = "etnaviv-gpu",
1712 		.owner = THIS_MODULE,
1713 		.pm = &etnaviv_gpu_pm_ops,
1714 		.of_match_table = etnaviv_gpu_match,
1715 	},
1716 	.probe = etnaviv_gpu_platform_probe,
1717 	.remove = etnaviv_gpu_platform_remove,
1718 	.id_table = gpu_ids,
1719 };
1720