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_enable_mlcg(struct etnaviv_gpu *gpu)
491 {
492 	u32 pmc, ppc;
493 
494 	/* enable clock gating */
495 	ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
496 	ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
497 
498 	/* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
499 	if (gpu->identity.revision == 0x4301 ||
500 	    gpu->identity.revision == 0x4302)
501 		ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
502 
503 	gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
504 
505 	pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
506 
507 	/* Disable PA clock gating for GC400+ except for GC420 */
508 	if (gpu->identity.model >= chipModel_GC400 &&
509 	    gpu->identity.model != chipModel_GC420)
510 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
511 
512 	/*
513 	 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
514 	 * present without a bug fix.
515 	 */
516 	if (gpu->identity.revision < 0x5000 &&
517 	    gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
518 	    !(gpu->identity.minor_features1 &
519 	      chipMinorFeatures1_DISABLE_PE_GATING))
520 		pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
521 
522 	if (gpu->identity.revision < 0x5422)
523 		pmc |= BIT(15); /* Unknown bit */
524 
525 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
526 	pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
527 
528 	gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
529 }
530 
531 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
532 {
533 	u16 prefetch;
534 
535 	if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
536 	     etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
537 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
538 		u32 mc_memory_debug;
539 
540 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
541 
542 		if (gpu->identity.revision == 0x5007)
543 			mc_memory_debug |= 0x0c;
544 		else
545 			mc_memory_debug |= 0x08;
546 
547 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
548 	}
549 
550 	/* enable module-level clock gating */
551 	etnaviv_gpu_enable_mlcg(gpu);
552 
553 	/*
554 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
555 	 * This is necessary to prevent the iMX6 SoC locking up.
556 	 */
557 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
558 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
559 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
560 
561 	/* GC2000 rev 5108 needs a special bus config */
562 	if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
563 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
564 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
565 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
566 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
567 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
568 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
569 	}
570 
571 	/* set base addresses */
572 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_RA, gpu->memory_base);
573 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_FE, gpu->memory_base);
574 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_TX, gpu->memory_base);
575 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PEZ, gpu->memory_base);
576 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PE, gpu->memory_base);
577 
578 	/* setup the MMU page table pointers */
579 	etnaviv_iommu_domain_restore(gpu, gpu->mmu->domain);
580 
581 	/* Start command processor */
582 	prefetch = etnaviv_buffer_init(gpu);
583 
584 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
585 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS,
586 		  gpu->buffer->paddr - gpu->memory_base);
587 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
588 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
589 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
590 }
591 
592 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
593 {
594 	int ret, i;
595 	struct iommu_domain *iommu;
596 	enum etnaviv_iommu_version version;
597 	bool mmuv2;
598 
599 	ret = pm_runtime_get_sync(gpu->dev);
600 	if (ret < 0) {
601 		dev_err(gpu->dev, "Failed to enable GPU power domain\n");
602 		return ret;
603 	}
604 
605 	etnaviv_hw_identify(gpu);
606 
607 	if (gpu->identity.model == 0) {
608 		dev_err(gpu->dev, "Unknown GPU model\n");
609 		ret = -ENXIO;
610 		goto fail;
611 	}
612 
613 	/* Exclude VG cores with FE2.0 */
614 	if (gpu->identity.features & chipFeatures_PIPE_VG &&
615 	    gpu->identity.features & chipFeatures_FE20) {
616 		dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
617 		ret = -ENXIO;
618 		goto fail;
619 	}
620 
621 	/*
622 	 * Set the GPU linear window to be at the end of the DMA window, where
623 	 * the CMA area is likely to reside. This ensures that we are able to
624 	 * map the command buffers while having the linear window overlap as
625 	 * much RAM as possible, so we can optimize mappings for other buffers.
626 	 *
627 	 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
628 	 * to different views of the memory on the individual engines.
629 	 */
630 	if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
631 	    (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
632 		u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
633 		if (dma_mask < PHYS_OFFSET + SZ_2G)
634 			gpu->memory_base = PHYS_OFFSET;
635 		else
636 			gpu->memory_base = dma_mask - SZ_2G + 1;
637 	}
638 
639 	ret = etnaviv_hw_reset(gpu);
640 	if (ret) {
641 		dev_err(gpu->dev, "GPU reset failed\n");
642 		goto fail;
643 	}
644 
645 	/* Setup IOMMU.. eventually we will (I think) do this once per context
646 	 * and have separate page tables per context.  For now, to keep things
647 	 * simple and to get something working, just use a single address space:
648 	 */
649 	mmuv2 = gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION;
650 	dev_dbg(gpu->dev, "mmuv2: %d\n", mmuv2);
651 
652 	if (!mmuv2) {
653 		iommu = etnaviv_iommu_domain_alloc(gpu);
654 		version = ETNAVIV_IOMMU_V1;
655 	} else {
656 		iommu = etnaviv_iommu_v2_domain_alloc(gpu);
657 		version = ETNAVIV_IOMMU_V2;
658 	}
659 
660 	if (!iommu) {
661 		dev_err(gpu->dev, "Failed to allocate GPU IOMMU domain\n");
662 		ret = -ENOMEM;
663 		goto fail;
664 	}
665 
666 	gpu->mmu = etnaviv_iommu_new(gpu, iommu, version);
667 	if (!gpu->mmu) {
668 		dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n");
669 		iommu_domain_free(iommu);
670 		ret = -ENOMEM;
671 		goto fail;
672 	}
673 
674 	/* Create buffer: */
675 	gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
676 	if (!gpu->buffer) {
677 		ret = -ENOMEM;
678 		dev_err(gpu->dev, "could not create command buffer\n");
679 		goto destroy_iommu;
680 	}
681 	if (gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
682 		ret = -EINVAL;
683 		dev_err(gpu->dev,
684 			"command buffer outside valid memory window\n");
685 		goto free_buffer;
686 	}
687 
688 	/* Setup event management */
689 	spin_lock_init(&gpu->event_spinlock);
690 	init_completion(&gpu->event_free);
691 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
692 		gpu->event[i].used = false;
693 		complete(&gpu->event_free);
694 	}
695 
696 	/* Now program the hardware */
697 	mutex_lock(&gpu->lock);
698 	etnaviv_gpu_hw_init(gpu);
699 	gpu->exec_state = -1;
700 	mutex_unlock(&gpu->lock);
701 
702 	pm_runtime_mark_last_busy(gpu->dev);
703 	pm_runtime_put_autosuspend(gpu->dev);
704 
705 	return 0;
706 
707 free_buffer:
708 	etnaviv_gpu_cmdbuf_free(gpu->buffer);
709 	gpu->buffer = NULL;
710 destroy_iommu:
711 	etnaviv_iommu_destroy(gpu->mmu);
712 	gpu->mmu = NULL;
713 fail:
714 	pm_runtime_mark_last_busy(gpu->dev);
715 	pm_runtime_put_autosuspend(gpu->dev);
716 
717 	return ret;
718 }
719 
720 #ifdef CONFIG_DEBUG_FS
721 struct dma_debug {
722 	u32 address[2];
723 	u32 state[2];
724 };
725 
726 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
727 {
728 	u32 i;
729 
730 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
731 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
732 
733 	for (i = 0; i < 500; i++) {
734 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
735 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
736 
737 		if (debug->address[0] != debug->address[1])
738 			break;
739 
740 		if (debug->state[0] != debug->state[1])
741 			break;
742 	}
743 }
744 
745 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
746 {
747 	struct dma_debug debug;
748 	u32 dma_lo, dma_hi, axi, idle;
749 	int ret;
750 
751 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
752 
753 	ret = pm_runtime_get_sync(gpu->dev);
754 	if (ret < 0)
755 		return ret;
756 
757 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
758 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
759 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
760 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
761 
762 	verify_dma(gpu, &debug);
763 
764 	seq_puts(m, "\tfeatures\n");
765 	seq_printf(m, "\t minor_features0: 0x%08x\n",
766 		   gpu->identity.minor_features0);
767 	seq_printf(m, "\t minor_features1: 0x%08x\n",
768 		   gpu->identity.minor_features1);
769 	seq_printf(m, "\t minor_features2: 0x%08x\n",
770 		   gpu->identity.minor_features2);
771 	seq_printf(m, "\t minor_features3: 0x%08x\n",
772 		   gpu->identity.minor_features3);
773 	seq_printf(m, "\t minor_features4: 0x%08x\n",
774 		   gpu->identity.minor_features4);
775 	seq_printf(m, "\t minor_features5: 0x%08x\n",
776 		   gpu->identity.minor_features5);
777 
778 	seq_puts(m, "\tspecs\n");
779 	seq_printf(m, "\t stream_count:  %d\n",
780 			gpu->identity.stream_count);
781 	seq_printf(m, "\t register_max: %d\n",
782 			gpu->identity.register_max);
783 	seq_printf(m, "\t thread_count: %d\n",
784 			gpu->identity.thread_count);
785 	seq_printf(m, "\t vertex_cache_size: %d\n",
786 			gpu->identity.vertex_cache_size);
787 	seq_printf(m, "\t shader_core_count: %d\n",
788 			gpu->identity.shader_core_count);
789 	seq_printf(m, "\t pixel_pipes: %d\n",
790 			gpu->identity.pixel_pipes);
791 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
792 			gpu->identity.vertex_output_buffer_size);
793 	seq_printf(m, "\t buffer_size: %d\n",
794 			gpu->identity.buffer_size);
795 	seq_printf(m, "\t instruction_count: %d\n",
796 			gpu->identity.instruction_count);
797 	seq_printf(m, "\t num_constants: %d\n",
798 			gpu->identity.num_constants);
799 	seq_printf(m, "\t varyings_count: %d\n",
800 			gpu->identity.varyings_count);
801 
802 	seq_printf(m, "\taxi: 0x%08x\n", axi);
803 	seq_printf(m, "\tidle: 0x%08x\n", idle);
804 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
805 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
806 		seq_puts(m, "\t FE is not idle\n");
807 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
808 		seq_puts(m, "\t DE is not idle\n");
809 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
810 		seq_puts(m, "\t PE is not idle\n");
811 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
812 		seq_puts(m, "\t SH is not idle\n");
813 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
814 		seq_puts(m, "\t PA is not idle\n");
815 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
816 		seq_puts(m, "\t SE is not idle\n");
817 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
818 		seq_puts(m, "\t RA is not idle\n");
819 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
820 		seq_puts(m, "\t TX is not idle\n");
821 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
822 		seq_puts(m, "\t VG is not idle\n");
823 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
824 		seq_puts(m, "\t IM is not idle\n");
825 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
826 		seq_puts(m, "\t FP is not idle\n");
827 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
828 		seq_puts(m, "\t TS is not idle\n");
829 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
830 		seq_puts(m, "\t AXI low power mode\n");
831 
832 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
833 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
834 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
835 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
836 
837 		seq_puts(m, "\tMC\n");
838 		seq_printf(m, "\t read0: 0x%08x\n", read0);
839 		seq_printf(m, "\t read1: 0x%08x\n", read1);
840 		seq_printf(m, "\t write: 0x%08x\n", write);
841 	}
842 
843 	seq_puts(m, "\tDMA ");
844 
845 	if (debug.address[0] == debug.address[1] &&
846 	    debug.state[0] == debug.state[1]) {
847 		seq_puts(m, "seems to be stuck\n");
848 	} else if (debug.address[0] == debug.address[1]) {
849 		seq_puts(m, "address is constant\n");
850 	} else {
851 		seq_puts(m, "is running\n");
852 	}
853 
854 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
855 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
856 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
857 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
858 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
859 		   dma_lo, dma_hi);
860 
861 	ret = 0;
862 
863 	pm_runtime_mark_last_busy(gpu->dev);
864 	pm_runtime_put_autosuspend(gpu->dev);
865 
866 	return ret;
867 }
868 #endif
869 
870 /*
871  * Power Management:
872  */
873 static int enable_clk(struct etnaviv_gpu *gpu)
874 {
875 	if (gpu->clk_core)
876 		clk_prepare_enable(gpu->clk_core);
877 	if (gpu->clk_shader)
878 		clk_prepare_enable(gpu->clk_shader);
879 
880 	return 0;
881 }
882 
883 static int disable_clk(struct etnaviv_gpu *gpu)
884 {
885 	if (gpu->clk_core)
886 		clk_disable_unprepare(gpu->clk_core);
887 	if (gpu->clk_shader)
888 		clk_disable_unprepare(gpu->clk_shader);
889 
890 	return 0;
891 }
892 
893 static int enable_axi(struct etnaviv_gpu *gpu)
894 {
895 	if (gpu->clk_bus)
896 		clk_prepare_enable(gpu->clk_bus);
897 
898 	return 0;
899 }
900 
901 static int disable_axi(struct etnaviv_gpu *gpu)
902 {
903 	if (gpu->clk_bus)
904 		clk_disable_unprepare(gpu->clk_bus);
905 
906 	return 0;
907 }
908 
909 /*
910  * Hangcheck detection for locked gpu:
911  */
912 static void recover_worker(struct work_struct *work)
913 {
914 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
915 					       recover_work);
916 	unsigned long flags;
917 	unsigned int i;
918 
919 	dev_err(gpu->dev, "hangcheck recover!\n");
920 
921 	if (pm_runtime_get_sync(gpu->dev) < 0)
922 		return;
923 
924 	mutex_lock(&gpu->lock);
925 
926 	/* Only catch the first event, or when manually re-armed */
927 	if (etnaviv_dump_core) {
928 		etnaviv_core_dump(gpu);
929 		etnaviv_dump_core = false;
930 	}
931 
932 	etnaviv_hw_reset(gpu);
933 
934 	/* complete all events, the GPU won't do it after the reset */
935 	spin_lock_irqsave(&gpu->event_spinlock, flags);
936 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
937 		if (!gpu->event[i].used)
938 			continue;
939 		fence_signal(gpu->event[i].fence);
940 		gpu->event[i].fence = NULL;
941 		gpu->event[i].used = false;
942 		complete(&gpu->event_free);
943 	}
944 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
945 	gpu->completed_fence = gpu->active_fence;
946 
947 	etnaviv_gpu_hw_init(gpu);
948 	gpu->switch_context = true;
949 	gpu->exec_state = -1;
950 
951 	mutex_unlock(&gpu->lock);
952 	pm_runtime_mark_last_busy(gpu->dev);
953 	pm_runtime_put_autosuspend(gpu->dev);
954 
955 	/* Retire the buffer objects in a work */
956 	etnaviv_queue_work(gpu->drm, &gpu->retire_work);
957 }
958 
959 static void hangcheck_timer_reset(struct etnaviv_gpu *gpu)
960 {
961 	DBG("%s", dev_name(gpu->dev));
962 	mod_timer(&gpu->hangcheck_timer,
963 		  round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES));
964 }
965 
966 static void hangcheck_handler(unsigned long data)
967 {
968 	struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data;
969 	u32 fence = gpu->completed_fence;
970 	bool progress = false;
971 
972 	if (fence != gpu->hangcheck_fence) {
973 		gpu->hangcheck_fence = fence;
974 		progress = true;
975 	}
976 
977 	if (!progress) {
978 		u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
979 		int change = dma_addr - gpu->hangcheck_dma_addr;
980 
981 		if (change < 0 || change > 16) {
982 			gpu->hangcheck_dma_addr = dma_addr;
983 			progress = true;
984 		}
985 	}
986 
987 	if (!progress && fence_after(gpu->active_fence, fence)) {
988 		dev_err(gpu->dev, "hangcheck detected gpu lockup!\n");
989 		dev_err(gpu->dev, "     completed fence: %u\n", fence);
990 		dev_err(gpu->dev, "     active fence: %u\n",
991 			gpu->active_fence);
992 		etnaviv_queue_work(gpu->drm, &gpu->recover_work);
993 	}
994 
995 	/* if still more pending work, reset the hangcheck timer: */
996 	if (fence_after(gpu->active_fence, gpu->hangcheck_fence))
997 		hangcheck_timer_reset(gpu);
998 }
999 
1000 static void hangcheck_disable(struct etnaviv_gpu *gpu)
1001 {
1002 	del_timer_sync(&gpu->hangcheck_timer);
1003 	cancel_work_sync(&gpu->recover_work);
1004 }
1005 
1006 /* fence object management */
1007 struct etnaviv_fence {
1008 	struct etnaviv_gpu *gpu;
1009 	struct fence base;
1010 };
1011 
1012 static inline struct etnaviv_fence *to_etnaviv_fence(struct fence *fence)
1013 {
1014 	return container_of(fence, struct etnaviv_fence, base);
1015 }
1016 
1017 static const char *etnaviv_fence_get_driver_name(struct fence *fence)
1018 {
1019 	return "etnaviv";
1020 }
1021 
1022 static const char *etnaviv_fence_get_timeline_name(struct fence *fence)
1023 {
1024 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1025 
1026 	return dev_name(f->gpu->dev);
1027 }
1028 
1029 static bool etnaviv_fence_enable_signaling(struct fence *fence)
1030 {
1031 	return true;
1032 }
1033 
1034 static bool etnaviv_fence_signaled(struct fence *fence)
1035 {
1036 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1037 
1038 	return fence_completed(f->gpu, f->base.seqno);
1039 }
1040 
1041 static void etnaviv_fence_release(struct fence *fence)
1042 {
1043 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
1044 
1045 	kfree_rcu(f, base.rcu);
1046 }
1047 
1048 static const struct fence_ops etnaviv_fence_ops = {
1049 	.get_driver_name = etnaviv_fence_get_driver_name,
1050 	.get_timeline_name = etnaviv_fence_get_timeline_name,
1051 	.enable_signaling = etnaviv_fence_enable_signaling,
1052 	.signaled = etnaviv_fence_signaled,
1053 	.wait = fence_default_wait,
1054 	.release = etnaviv_fence_release,
1055 };
1056 
1057 static struct fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1058 {
1059 	struct etnaviv_fence *f;
1060 
1061 	f = kzalloc(sizeof(*f), GFP_KERNEL);
1062 	if (!f)
1063 		return NULL;
1064 
1065 	f->gpu = gpu;
1066 
1067 	fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1068 		   gpu->fence_context, ++gpu->next_fence);
1069 
1070 	return &f->base;
1071 }
1072 
1073 int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj,
1074 	unsigned int context, bool exclusive)
1075 {
1076 	struct reservation_object *robj = etnaviv_obj->resv;
1077 	struct reservation_object_list *fobj;
1078 	struct fence *fence;
1079 	int i, ret;
1080 
1081 	if (!exclusive) {
1082 		ret = reservation_object_reserve_shared(robj);
1083 		if (ret)
1084 			return ret;
1085 	}
1086 
1087 	/*
1088 	 * If we have any shared fences, then the exclusive fence
1089 	 * should be ignored as it will already have been signalled.
1090 	 */
1091 	fobj = reservation_object_get_list(robj);
1092 	if (!fobj || fobj->shared_count == 0) {
1093 		/* Wait on any existing exclusive fence which isn't our own */
1094 		fence = reservation_object_get_excl(robj);
1095 		if (fence && fence->context != context) {
1096 			ret = fence_wait(fence, true);
1097 			if (ret)
1098 				return ret;
1099 		}
1100 	}
1101 
1102 	if (!exclusive || !fobj)
1103 		return 0;
1104 
1105 	for (i = 0; i < fobj->shared_count; i++) {
1106 		fence = rcu_dereference_protected(fobj->shared[i],
1107 						reservation_object_held(robj));
1108 		if (fence->context != context) {
1109 			ret = fence_wait(fence, true);
1110 			if (ret)
1111 				return ret;
1112 		}
1113 	}
1114 
1115 	return 0;
1116 }
1117 
1118 /*
1119  * event management:
1120  */
1121 
1122 static unsigned int event_alloc(struct etnaviv_gpu *gpu)
1123 {
1124 	unsigned long ret, flags;
1125 	unsigned int i, event = ~0U;
1126 
1127 	ret = wait_for_completion_timeout(&gpu->event_free,
1128 					  msecs_to_jiffies(10 * 10000));
1129 	if (!ret)
1130 		dev_err(gpu->dev, "wait_for_completion_timeout failed");
1131 
1132 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1133 
1134 	/* find first free event */
1135 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
1136 		if (gpu->event[i].used == false) {
1137 			gpu->event[i].used = true;
1138 			event = i;
1139 			break;
1140 		}
1141 	}
1142 
1143 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1144 
1145 	return event;
1146 }
1147 
1148 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1149 {
1150 	unsigned long flags;
1151 
1152 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1153 
1154 	if (gpu->event[event].used == false) {
1155 		dev_warn(gpu->dev, "event %u is already marked as free",
1156 			 event);
1157 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1158 	} else {
1159 		gpu->event[event].used = false;
1160 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1161 
1162 		complete(&gpu->event_free);
1163 	}
1164 }
1165 
1166 /*
1167  * Cmdstream submission/retirement:
1168  */
1169 
1170 struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
1171 	size_t nr_bos)
1172 {
1173 	struct etnaviv_cmdbuf *cmdbuf;
1174 	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
1175 				 sizeof(*cmdbuf));
1176 
1177 	cmdbuf = kzalloc(sz, GFP_KERNEL);
1178 	if (!cmdbuf)
1179 		return NULL;
1180 
1181 	cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
1182 				     GFP_KERNEL);
1183 	if (!cmdbuf->vaddr) {
1184 		kfree(cmdbuf);
1185 		return NULL;
1186 	}
1187 
1188 	cmdbuf->gpu = gpu;
1189 	cmdbuf->size = size;
1190 
1191 	return cmdbuf;
1192 }
1193 
1194 void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1195 {
1196 	dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
1197 		    cmdbuf->paddr);
1198 	kfree(cmdbuf);
1199 }
1200 
1201 static void retire_worker(struct work_struct *work)
1202 {
1203 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1204 					       retire_work);
1205 	u32 fence = gpu->completed_fence;
1206 	struct etnaviv_cmdbuf *cmdbuf, *tmp;
1207 	unsigned int i;
1208 
1209 	mutex_lock(&gpu->lock);
1210 	list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1211 		if (!fence_is_signaled(cmdbuf->fence))
1212 			break;
1213 
1214 		list_del(&cmdbuf->node);
1215 		fence_put(cmdbuf->fence);
1216 
1217 		for (i = 0; i < cmdbuf->nr_bos; i++) {
1218 			struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
1219 			struct etnaviv_gem_object *etnaviv_obj = mapping->object;
1220 
1221 			atomic_dec(&etnaviv_obj->gpu_active);
1222 			/* drop the refcount taken in etnaviv_gpu_submit */
1223 			etnaviv_gem_mapping_unreference(mapping);
1224 		}
1225 
1226 		etnaviv_gpu_cmdbuf_free(cmdbuf);
1227 		/*
1228 		 * We need to balance the runtime PM count caused by
1229 		 * each submission.  Upon submission, we increment
1230 		 * the runtime PM counter, and allocate one event.
1231 		 * So here, we put the runtime PM count for each
1232 		 * completed event.
1233 		 */
1234 		pm_runtime_put_autosuspend(gpu->dev);
1235 	}
1236 
1237 	gpu->retired_fence = fence;
1238 
1239 	mutex_unlock(&gpu->lock);
1240 
1241 	wake_up_all(&gpu->fence_event);
1242 }
1243 
1244 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1245 	u32 fence, struct timespec *timeout)
1246 {
1247 	int ret;
1248 
1249 	if (fence_after(fence, gpu->next_fence)) {
1250 		DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1251 				fence, gpu->next_fence);
1252 		return -EINVAL;
1253 	}
1254 
1255 	if (!timeout) {
1256 		/* No timeout was requested: just test for completion */
1257 		ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1258 	} else {
1259 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1260 
1261 		ret = wait_event_interruptible_timeout(gpu->fence_event,
1262 						fence_completed(gpu, fence),
1263 						remaining);
1264 		if (ret == 0) {
1265 			DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1266 				fence, gpu->retired_fence,
1267 				gpu->completed_fence);
1268 			ret = -ETIMEDOUT;
1269 		} else if (ret != -ERESTARTSYS) {
1270 			ret = 0;
1271 		}
1272 	}
1273 
1274 	return ret;
1275 }
1276 
1277 /*
1278  * Wait for an object to become inactive.  This, on it's own, is not race
1279  * free: the object is moved by the retire worker off the active list, and
1280  * then the iova is put.  Moreover, the object could be re-submitted just
1281  * after we notice that it's become inactive.
1282  *
1283  * Although the retirement happens under the gpu lock, we don't want to hold
1284  * that lock in this function while waiting.
1285  */
1286 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1287 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1288 {
1289 	unsigned long remaining;
1290 	long ret;
1291 
1292 	if (!timeout)
1293 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1294 
1295 	remaining = etnaviv_timeout_to_jiffies(timeout);
1296 
1297 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1298 					       !is_active(etnaviv_obj),
1299 					       remaining);
1300 	if (ret > 0) {
1301 		struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1302 
1303 		/* Synchronise with the retire worker */
1304 		flush_workqueue(priv->wq);
1305 		return 0;
1306 	} else if (ret == -ERESTARTSYS) {
1307 		return -ERESTARTSYS;
1308 	} else {
1309 		return -ETIMEDOUT;
1310 	}
1311 }
1312 
1313 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1314 {
1315 	return pm_runtime_get_sync(gpu->dev);
1316 }
1317 
1318 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1319 {
1320 	pm_runtime_mark_last_busy(gpu->dev);
1321 	pm_runtime_put_autosuspend(gpu->dev);
1322 }
1323 
1324 /* add bo's to gpu's ring, and kick gpu: */
1325 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1326 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1327 {
1328 	struct fence *fence;
1329 	unsigned int event, i;
1330 	int ret;
1331 
1332 	ret = etnaviv_gpu_pm_get_sync(gpu);
1333 	if (ret < 0)
1334 		return ret;
1335 
1336 	mutex_lock(&gpu->lock);
1337 
1338 	/*
1339 	 * TODO
1340 	 *
1341 	 * - flush
1342 	 * - data endian
1343 	 * - prefetch
1344 	 *
1345 	 */
1346 
1347 	event = event_alloc(gpu);
1348 	if (unlikely(event == ~0U)) {
1349 		DRM_ERROR("no free event\n");
1350 		ret = -EBUSY;
1351 		goto out_unlock;
1352 	}
1353 
1354 	fence = etnaviv_gpu_fence_alloc(gpu);
1355 	if (!fence) {
1356 		event_free(gpu, event);
1357 		ret = -ENOMEM;
1358 		goto out_unlock;
1359 	}
1360 
1361 	gpu->event[event].fence = fence;
1362 	submit->fence = fence->seqno;
1363 	gpu->active_fence = submit->fence;
1364 
1365 	if (gpu->lastctx != cmdbuf->ctx) {
1366 		gpu->mmu->need_flush = true;
1367 		gpu->switch_context = true;
1368 		gpu->lastctx = cmdbuf->ctx;
1369 	}
1370 
1371 	etnaviv_buffer_queue(gpu, event, cmdbuf);
1372 
1373 	cmdbuf->fence = fence;
1374 	list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1375 
1376 	/* We're committed to adding this command buffer, hold a PM reference */
1377 	pm_runtime_get_noresume(gpu->dev);
1378 
1379 	for (i = 0; i < submit->nr_bos; i++) {
1380 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1381 
1382 		/* Each cmdbuf takes a refcount on the mapping */
1383 		etnaviv_gem_mapping_reference(submit->bos[i].mapping);
1384 		cmdbuf->bo_map[i] = submit->bos[i].mapping;
1385 		atomic_inc(&etnaviv_obj->gpu_active);
1386 
1387 		if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1388 			reservation_object_add_excl_fence(etnaviv_obj->resv,
1389 							  fence);
1390 		else
1391 			reservation_object_add_shared_fence(etnaviv_obj->resv,
1392 							    fence);
1393 	}
1394 	cmdbuf->nr_bos = submit->nr_bos;
1395 	hangcheck_timer_reset(gpu);
1396 	ret = 0;
1397 
1398 out_unlock:
1399 	mutex_unlock(&gpu->lock);
1400 
1401 	etnaviv_gpu_pm_put(gpu);
1402 
1403 	return ret;
1404 }
1405 
1406 /*
1407  * Init/Cleanup:
1408  */
1409 static irqreturn_t irq_handler(int irq, void *data)
1410 {
1411 	struct etnaviv_gpu *gpu = data;
1412 	irqreturn_t ret = IRQ_NONE;
1413 
1414 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1415 
1416 	if (intr != 0) {
1417 		int event;
1418 
1419 		pm_runtime_mark_last_busy(gpu->dev);
1420 
1421 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1422 
1423 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1424 			dev_err(gpu->dev, "AXI bus error\n");
1425 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1426 		}
1427 
1428 		while ((event = ffs(intr)) != 0) {
1429 			struct fence *fence;
1430 
1431 			event -= 1;
1432 
1433 			intr &= ~(1 << event);
1434 
1435 			dev_dbg(gpu->dev, "event %u\n", event);
1436 
1437 			fence = gpu->event[event].fence;
1438 			gpu->event[event].fence = NULL;
1439 			fence_signal(fence);
1440 
1441 			/*
1442 			 * Events can be processed out of order.  Eg,
1443 			 * - allocate and queue event 0
1444 			 * - allocate event 1
1445 			 * - event 0 completes, we process it
1446 			 * - allocate and queue event 0
1447 			 * - event 1 and event 0 complete
1448 			 * we can end up processing event 0 first, then 1.
1449 			 */
1450 			if (fence_after(fence->seqno, gpu->completed_fence))
1451 				gpu->completed_fence = fence->seqno;
1452 
1453 			event_free(gpu, event);
1454 		}
1455 
1456 		/* Retire the buffer objects in a work */
1457 		etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1458 
1459 		ret = IRQ_HANDLED;
1460 	}
1461 
1462 	return ret;
1463 }
1464 
1465 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1466 {
1467 	int ret;
1468 
1469 	ret = enable_clk(gpu);
1470 	if (ret)
1471 		return ret;
1472 
1473 	ret = enable_axi(gpu);
1474 	if (ret) {
1475 		disable_clk(gpu);
1476 		return ret;
1477 	}
1478 
1479 	return 0;
1480 }
1481 
1482 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1483 {
1484 	int ret;
1485 
1486 	ret = disable_axi(gpu);
1487 	if (ret)
1488 		return ret;
1489 
1490 	ret = disable_clk(gpu);
1491 	if (ret)
1492 		return ret;
1493 
1494 	return 0;
1495 }
1496 
1497 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1498 {
1499 	if (gpu->buffer) {
1500 		unsigned long timeout;
1501 
1502 		/* Replace the last WAIT with END */
1503 		etnaviv_buffer_end(gpu);
1504 
1505 		/*
1506 		 * We know that only the FE is busy here, this should
1507 		 * happen quickly (as the WAIT is only 200 cycles).  If
1508 		 * we fail, just warn and continue.
1509 		 */
1510 		timeout = jiffies + msecs_to_jiffies(100);
1511 		do {
1512 			u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1513 
1514 			if ((idle & gpu->idle_mask) == gpu->idle_mask)
1515 				break;
1516 
1517 			if (time_is_before_jiffies(timeout)) {
1518 				dev_warn(gpu->dev,
1519 					 "timed out waiting for idle: idle=0x%x\n",
1520 					 idle);
1521 				break;
1522 			}
1523 
1524 			udelay(5);
1525 		} while (1);
1526 	}
1527 
1528 	return etnaviv_gpu_clk_disable(gpu);
1529 }
1530 
1531 #ifdef CONFIG_PM
1532 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1533 {
1534 	u32 clock;
1535 	int ret;
1536 
1537 	ret = mutex_lock_killable(&gpu->lock);
1538 	if (ret)
1539 		return ret;
1540 
1541 	clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1542 		VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1543 
1544 	etnaviv_gpu_load_clock(gpu, clock);
1545 	etnaviv_gpu_hw_init(gpu);
1546 
1547 	gpu->switch_context = true;
1548 	gpu->exec_state = -1;
1549 
1550 	mutex_unlock(&gpu->lock);
1551 
1552 	return 0;
1553 }
1554 #endif
1555 
1556 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1557 	void *data)
1558 {
1559 	struct drm_device *drm = data;
1560 	struct etnaviv_drm_private *priv = drm->dev_private;
1561 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1562 	int ret;
1563 
1564 #ifdef CONFIG_PM
1565 	ret = pm_runtime_get_sync(gpu->dev);
1566 #else
1567 	ret = etnaviv_gpu_clk_enable(gpu);
1568 #endif
1569 	if (ret < 0)
1570 		return ret;
1571 
1572 	gpu->drm = drm;
1573 	gpu->fence_context = fence_context_alloc(1);
1574 	spin_lock_init(&gpu->fence_spinlock);
1575 
1576 	INIT_LIST_HEAD(&gpu->active_cmd_list);
1577 	INIT_WORK(&gpu->retire_work, retire_worker);
1578 	INIT_WORK(&gpu->recover_work, recover_worker);
1579 	init_waitqueue_head(&gpu->fence_event);
1580 
1581 	setup_deferrable_timer(&gpu->hangcheck_timer, hangcheck_handler,
1582 			       (unsigned long)gpu);
1583 
1584 	priv->gpu[priv->num_gpus++] = gpu;
1585 
1586 	pm_runtime_mark_last_busy(gpu->dev);
1587 	pm_runtime_put_autosuspend(gpu->dev);
1588 
1589 	return 0;
1590 }
1591 
1592 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1593 	void *data)
1594 {
1595 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1596 
1597 	DBG("%s", dev_name(gpu->dev));
1598 
1599 	hangcheck_disable(gpu);
1600 
1601 #ifdef CONFIG_PM
1602 	pm_runtime_get_sync(gpu->dev);
1603 	pm_runtime_put_sync_suspend(gpu->dev);
1604 #else
1605 	etnaviv_gpu_hw_suspend(gpu);
1606 #endif
1607 
1608 	if (gpu->buffer) {
1609 		etnaviv_gpu_cmdbuf_free(gpu->buffer);
1610 		gpu->buffer = NULL;
1611 	}
1612 
1613 	if (gpu->mmu) {
1614 		etnaviv_iommu_destroy(gpu->mmu);
1615 		gpu->mmu = NULL;
1616 	}
1617 
1618 	gpu->drm = NULL;
1619 }
1620 
1621 static const struct component_ops gpu_ops = {
1622 	.bind = etnaviv_gpu_bind,
1623 	.unbind = etnaviv_gpu_unbind,
1624 };
1625 
1626 static const struct of_device_id etnaviv_gpu_match[] = {
1627 	{
1628 		.compatible = "vivante,gc"
1629 	},
1630 	{ /* sentinel */ }
1631 };
1632 
1633 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1634 {
1635 	struct device *dev = &pdev->dev;
1636 	struct etnaviv_gpu *gpu;
1637 	int err = 0;
1638 
1639 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1640 	if (!gpu)
1641 		return -ENOMEM;
1642 
1643 	gpu->dev = &pdev->dev;
1644 	mutex_init(&gpu->lock);
1645 
1646 	/* Map registers: */
1647 	gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1648 	if (IS_ERR(gpu->mmio))
1649 		return PTR_ERR(gpu->mmio);
1650 
1651 	/* Get Interrupt: */
1652 	gpu->irq = platform_get_irq(pdev, 0);
1653 	if (gpu->irq < 0) {
1654 		err = gpu->irq;
1655 		dev_err(dev, "failed to get irq: %d\n", err);
1656 		goto fail;
1657 	}
1658 
1659 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1660 			       dev_name(gpu->dev), gpu);
1661 	if (err) {
1662 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1663 		goto fail;
1664 	}
1665 
1666 	/* Get Clocks: */
1667 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1668 	DBG("clk_bus: %p", gpu->clk_bus);
1669 	if (IS_ERR(gpu->clk_bus))
1670 		gpu->clk_bus = NULL;
1671 
1672 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1673 	DBG("clk_core: %p", gpu->clk_core);
1674 	if (IS_ERR(gpu->clk_core))
1675 		gpu->clk_core = NULL;
1676 
1677 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1678 	DBG("clk_shader: %p", gpu->clk_shader);
1679 	if (IS_ERR(gpu->clk_shader))
1680 		gpu->clk_shader = NULL;
1681 
1682 	/* TODO: figure out max mapped size */
1683 	dev_set_drvdata(dev, gpu);
1684 
1685 	/*
1686 	 * We treat the device as initially suspended.  The runtime PM
1687 	 * autosuspend delay is rather arbitary: no measurements have
1688 	 * yet been performed to determine an appropriate value.
1689 	 */
1690 	pm_runtime_use_autosuspend(gpu->dev);
1691 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1692 	pm_runtime_enable(gpu->dev);
1693 
1694 	err = component_add(&pdev->dev, &gpu_ops);
1695 	if (err < 0) {
1696 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1697 		goto fail;
1698 	}
1699 
1700 	return 0;
1701 
1702 fail:
1703 	return err;
1704 }
1705 
1706 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1707 {
1708 	component_del(&pdev->dev, &gpu_ops);
1709 	pm_runtime_disable(&pdev->dev);
1710 	return 0;
1711 }
1712 
1713 #ifdef CONFIG_PM
1714 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1715 {
1716 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1717 	u32 idle, mask;
1718 
1719 	/* If we have outstanding fences, we're not idle */
1720 	if (gpu->completed_fence != gpu->active_fence)
1721 		return -EBUSY;
1722 
1723 	/* Check whether the hardware (except FE) is idle */
1724 	mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1725 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1726 	if (idle != mask)
1727 		return -EBUSY;
1728 
1729 	return etnaviv_gpu_hw_suspend(gpu);
1730 }
1731 
1732 static int etnaviv_gpu_rpm_resume(struct device *dev)
1733 {
1734 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1735 	int ret;
1736 
1737 	ret = etnaviv_gpu_clk_enable(gpu);
1738 	if (ret)
1739 		return ret;
1740 
1741 	/* Re-initialise the basic hardware state */
1742 	if (gpu->drm && gpu->buffer) {
1743 		ret = etnaviv_gpu_hw_resume(gpu);
1744 		if (ret) {
1745 			etnaviv_gpu_clk_disable(gpu);
1746 			return ret;
1747 		}
1748 	}
1749 
1750 	return 0;
1751 }
1752 #endif
1753 
1754 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1755 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1756 			   NULL)
1757 };
1758 
1759 struct platform_driver etnaviv_gpu_driver = {
1760 	.driver = {
1761 		.name = "etnaviv-gpu",
1762 		.owner = THIS_MODULE,
1763 		.pm = &etnaviv_gpu_pm_ops,
1764 		.of_match_table = etnaviv_gpu_match,
1765 	},
1766 	.probe = etnaviv_gpu_platform_probe,
1767 	.remove = etnaviv_gpu_platform_remove,
1768 	.id_table = gpu_ids,
1769 };
1770