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_STREAM_COUNT:
76 		*value = gpu->identity.stream_count;
77 		break;
78 
79 	case ETNAVIV_PARAM_GPU_REGISTER_MAX:
80 		*value = gpu->identity.register_max;
81 		break;
82 
83 	case ETNAVIV_PARAM_GPU_THREAD_COUNT:
84 		*value = gpu->identity.thread_count;
85 		break;
86 
87 	case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
88 		*value = gpu->identity.vertex_cache_size;
89 		break;
90 
91 	case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
92 		*value = gpu->identity.shader_core_count;
93 		break;
94 
95 	case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
96 		*value = gpu->identity.pixel_pipes;
97 		break;
98 
99 	case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
100 		*value = gpu->identity.vertex_output_buffer_size;
101 		break;
102 
103 	case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
104 		*value = gpu->identity.buffer_size;
105 		break;
106 
107 	case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
108 		*value = gpu->identity.instruction_count;
109 		break;
110 
111 	case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
112 		*value = gpu->identity.num_constants;
113 		break;
114 
115 	default:
116 		DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
117 		return -EINVAL;
118 	}
119 
120 	return 0;
121 }
122 
123 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
124 {
125 	if (gpu->identity.minor_features0 &
126 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
127 		u32 specs[2];
128 
129 		specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
130 		specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
131 
132 		gpu->identity.stream_count =
133 			(specs[0] & VIVS_HI_CHIP_SPECS_STREAM_COUNT__MASK)
134 				>> VIVS_HI_CHIP_SPECS_STREAM_COUNT__SHIFT;
135 		gpu->identity.register_max =
136 			(specs[0] & VIVS_HI_CHIP_SPECS_REGISTER_MAX__MASK)
137 				>> VIVS_HI_CHIP_SPECS_REGISTER_MAX__SHIFT;
138 		gpu->identity.thread_count =
139 			(specs[0] & VIVS_HI_CHIP_SPECS_THREAD_COUNT__MASK)
140 				>> VIVS_HI_CHIP_SPECS_THREAD_COUNT__SHIFT;
141 		gpu->identity.vertex_cache_size =
142 			(specs[0] & VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE__MASK)
143 				>> VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE__SHIFT;
144 		gpu->identity.shader_core_count =
145 			(specs[0] & VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT__MASK)
146 				>> VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT__SHIFT;
147 		gpu->identity.pixel_pipes =
148 			(specs[0] & VIVS_HI_CHIP_SPECS_PIXEL_PIPES__MASK)
149 				>> VIVS_HI_CHIP_SPECS_PIXEL_PIPES__SHIFT;
150 		gpu->identity.vertex_output_buffer_size =
151 			(specs[0] & VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE__MASK)
152 				>> VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE__SHIFT;
153 
154 		gpu->identity.buffer_size =
155 			(specs[1] & VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE__MASK)
156 				>> VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE__SHIFT;
157 		gpu->identity.instruction_count =
158 			(specs[1] & VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT__MASK)
159 				>> VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT__SHIFT;
160 		gpu->identity.num_constants =
161 			(specs[1] & VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS__MASK)
162 				>> VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS__SHIFT;
163 	}
164 
165 	/* Fill in the stream count if not specified */
166 	if (gpu->identity.stream_count == 0) {
167 		if (gpu->identity.model >= 0x1000)
168 			gpu->identity.stream_count = 4;
169 		else
170 			gpu->identity.stream_count = 1;
171 	}
172 
173 	/* Convert the register max value */
174 	if (gpu->identity.register_max)
175 		gpu->identity.register_max = 1 << gpu->identity.register_max;
176 	else if (gpu->identity.model == 0x0400)
177 		gpu->identity.register_max = 32;
178 	else
179 		gpu->identity.register_max = 64;
180 
181 	/* Convert thread count */
182 	if (gpu->identity.thread_count)
183 		gpu->identity.thread_count = 1 << gpu->identity.thread_count;
184 	else if (gpu->identity.model == 0x0400)
185 		gpu->identity.thread_count = 64;
186 	else if (gpu->identity.model == 0x0500 ||
187 		 gpu->identity.model == 0x0530)
188 		gpu->identity.thread_count = 128;
189 	else
190 		gpu->identity.thread_count = 256;
191 
192 	if (gpu->identity.vertex_cache_size == 0)
193 		gpu->identity.vertex_cache_size = 8;
194 
195 	if (gpu->identity.shader_core_count == 0) {
196 		if (gpu->identity.model >= 0x1000)
197 			gpu->identity.shader_core_count = 2;
198 		else
199 			gpu->identity.shader_core_count = 1;
200 	}
201 
202 	if (gpu->identity.pixel_pipes == 0)
203 		gpu->identity.pixel_pipes = 1;
204 
205 	/* Convert virtex buffer size */
206 	if (gpu->identity.vertex_output_buffer_size) {
207 		gpu->identity.vertex_output_buffer_size =
208 			1 << gpu->identity.vertex_output_buffer_size;
209 	} else if (gpu->identity.model == 0x0400) {
210 		if (gpu->identity.revision < 0x4000)
211 			gpu->identity.vertex_output_buffer_size = 512;
212 		else if (gpu->identity.revision < 0x4200)
213 			gpu->identity.vertex_output_buffer_size = 256;
214 		else
215 			gpu->identity.vertex_output_buffer_size = 128;
216 	} else {
217 		gpu->identity.vertex_output_buffer_size = 512;
218 	}
219 
220 	switch (gpu->identity.instruction_count) {
221 	case 0:
222 		if ((gpu->identity.model == 0x2000 &&
223 		     gpu->identity.revision == 0x5108) ||
224 		    gpu->identity.model == 0x880)
225 			gpu->identity.instruction_count = 512;
226 		else
227 			gpu->identity.instruction_count = 256;
228 		break;
229 
230 	case 1:
231 		gpu->identity.instruction_count = 1024;
232 		break;
233 
234 	case 2:
235 		gpu->identity.instruction_count = 2048;
236 		break;
237 
238 	default:
239 		gpu->identity.instruction_count = 256;
240 		break;
241 	}
242 
243 	if (gpu->identity.num_constants == 0)
244 		gpu->identity.num_constants = 168;
245 }
246 
247 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
248 {
249 	u32 chipIdentity;
250 
251 	chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
252 
253 	/* Special case for older graphic cores. */
254 	if (((chipIdentity & VIVS_HI_CHIP_IDENTITY_FAMILY__MASK)
255 	     >> VIVS_HI_CHIP_IDENTITY_FAMILY__SHIFT) ==  0x01) {
256 		gpu->identity.model    = 0x500; /* gc500 */
257 		gpu->identity.revision =
258 			(chipIdentity & VIVS_HI_CHIP_IDENTITY_REVISION__MASK)
259 			>> VIVS_HI_CHIP_IDENTITY_REVISION__SHIFT;
260 	} else {
261 
262 		gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
263 		gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
264 
265 		/*
266 		 * !!!! HACK ALERT !!!!
267 		 * Because people change device IDs without letting software
268 		 * know about it - here is the hack to make it all look the
269 		 * same.  Only for GC400 family.
270 		 */
271 		if ((gpu->identity.model & 0xff00) == 0x0400 &&
272 		    gpu->identity.model != 0x0420) {
273 			gpu->identity.model = gpu->identity.model & 0x0400;
274 		}
275 
276 		/* Another special case */
277 		if (gpu->identity.model == 0x300 &&
278 		    gpu->identity.revision == 0x2201) {
279 			u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
280 			u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
281 
282 			if (chipDate == 0x20080814 && chipTime == 0x12051100) {
283 				/*
284 				 * This IP has an ECO; put the correct
285 				 * revision in it.
286 				 */
287 				gpu->identity.revision = 0x1051;
288 			}
289 		}
290 	}
291 
292 	dev_info(gpu->dev, "model: GC%x, revision: %x\n",
293 		 gpu->identity.model, gpu->identity.revision);
294 
295 	gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
296 
297 	/* Disable fast clear on GC700. */
298 	if (gpu->identity.model == 0x700)
299 		gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
300 
301 	if ((gpu->identity.model == 0x500 && gpu->identity.revision < 2) ||
302 	    (gpu->identity.model == 0x300 && gpu->identity.revision < 0x2000)) {
303 
304 		/*
305 		 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
306 		 * registers.
307 		 */
308 		gpu->identity.minor_features0 = 0;
309 		gpu->identity.minor_features1 = 0;
310 		gpu->identity.minor_features2 = 0;
311 		gpu->identity.minor_features3 = 0;
312 	} else
313 		gpu->identity.minor_features0 =
314 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
315 
316 	if (gpu->identity.minor_features0 &
317 	    chipMinorFeatures0_MORE_MINOR_FEATURES) {
318 		gpu->identity.minor_features1 =
319 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
320 		gpu->identity.minor_features2 =
321 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
322 		gpu->identity.minor_features3 =
323 				gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
324 	}
325 
326 	/* GC600 idle register reports zero bits where modules aren't present */
327 	if (gpu->identity.model == chipModel_GC600) {
328 		gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
329 				 VIVS_HI_IDLE_STATE_RA |
330 				 VIVS_HI_IDLE_STATE_SE |
331 				 VIVS_HI_IDLE_STATE_PA |
332 				 VIVS_HI_IDLE_STATE_SH |
333 				 VIVS_HI_IDLE_STATE_PE |
334 				 VIVS_HI_IDLE_STATE_DE |
335 				 VIVS_HI_IDLE_STATE_FE;
336 	} else {
337 		gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
338 	}
339 
340 	etnaviv_hw_specs(gpu);
341 }
342 
343 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
344 {
345 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
346 		  VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
347 	gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
348 }
349 
350 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
351 {
352 	u32 control, idle;
353 	unsigned long timeout;
354 	bool failed = true;
355 
356 	/* TODO
357 	 *
358 	 * - clock gating
359 	 * - puls eater
360 	 * - what about VG?
361 	 */
362 
363 	/* We hope that the GPU resets in under one second */
364 	timeout = jiffies + msecs_to_jiffies(1000);
365 
366 	while (time_is_after_jiffies(timeout)) {
367 		control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
368 			  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
369 
370 		/* enable clock */
371 		etnaviv_gpu_load_clock(gpu, control);
372 
373 		/* Wait for stable clock.  Vivante's code waited for 1ms */
374 		usleep_range(1000, 10000);
375 
376 		/* isolate the GPU. */
377 		control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
378 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
379 
380 		/* set soft reset. */
381 		control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
382 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
383 
384 		/* wait for reset. */
385 		msleep(1);
386 
387 		/* reset soft reset bit. */
388 		control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
389 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
390 
391 		/* reset GPU isolation. */
392 		control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
393 		gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
394 
395 		/* read idle register. */
396 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
397 
398 		/* try reseting again if FE it not idle */
399 		if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
400 			dev_dbg(gpu->dev, "FE is not idle\n");
401 			continue;
402 		}
403 
404 		/* read reset register. */
405 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
406 
407 		/* is the GPU idle? */
408 		if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
409 		    ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
410 			dev_dbg(gpu->dev, "GPU is not idle\n");
411 			continue;
412 		}
413 
414 		failed = false;
415 		break;
416 	}
417 
418 	if (failed) {
419 		idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
420 		control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
421 
422 		dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
423 			idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
424 			control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
425 			control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
426 
427 		return -EBUSY;
428 	}
429 
430 	/* We rely on the GPU running, so program the clock */
431 	control = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
432 		  VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
433 
434 	/* enable clock */
435 	etnaviv_gpu_load_clock(gpu, control);
436 
437 	return 0;
438 }
439 
440 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
441 {
442 	u16 prefetch;
443 
444 	if (gpu->identity.model == chipModel_GC320 &&
445 	    gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400 &&
446 	    (gpu->identity.revision == 0x5007 ||
447 	     gpu->identity.revision == 0x5220)) {
448 		u32 mc_memory_debug;
449 
450 		mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
451 
452 		if (gpu->identity.revision == 0x5007)
453 			mc_memory_debug |= 0x0c;
454 		else
455 			mc_memory_debug |= 0x08;
456 
457 		gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
458 	}
459 
460 	/*
461 	 * Update GPU AXI cache atttribute to "cacheable, no allocate".
462 	 * This is necessary to prevent the iMX6 SoC locking up.
463 	 */
464 	gpu_write(gpu, VIVS_HI_AXI_CONFIG,
465 		  VIVS_HI_AXI_CONFIG_AWCACHE(2) |
466 		  VIVS_HI_AXI_CONFIG_ARCACHE(2));
467 
468 	/* GC2000 rev 5108 needs a special bus config */
469 	if (gpu->identity.model == 0x2000 && gpu->identity.revision == 0x5108) {
470 		u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
471 		bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
472 				VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
473 		bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
474 			      VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
475 		gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
476 	}
477 
478 	/* set base addresses */
479 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_RA, gpu->memory_base);
480 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_FE, gpu->memory_base);
481 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_TX, gpu->memory_base);
482 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PEZ, gpu->memory_base);
483 	gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PE, gpu->memory_base);
484 
485 	/* setup the MMU page table pointers */
486 	etnaviv_iommu_domain_restore(gpu, gpu->mmu->domain);
487 
488 	/* Start command processor */
489 	prefetch = etnaviv_buffer_init(gpu);
490 
491 	gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
492 	gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS,
493 		  gpu->buffer->paddr - gpu->memory_base);
494 	gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
495 		  VIVS_FE_COMMAND_CONTROL_ENABLE |
496 		  VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
497 }
498 
499 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
500 {
501 	int ret, i;
502 	struct iommu_domain *iommu;
503 	enum etnaviv_iommu_version version;
504 	bool mmuv2;
505 
506 	ret = pm_runtime_get_sync(gpu->dev);
507 	if (ret < 0)
508 		return ret;
509 
510 	etnaviv_hw_identify(gpu);
511 
512 	if (gpu->identity.model == 0) {
513 		dev_err(gpu->dev, "Unknown GPU model\n");
514 		pm_runtime_put_autosuspend(gpu->dev);
515 		return -ENXIO;
516 	}
517 
518 	ret = etnaviv_hw_reset(gpu);
519 	if (ret)
520 		goto fail;
521 
522 	/* Setup IOMMU.. eventually we will (I think) do this once per context
523 	 * and have separate page tables per context.  For now, to keep things
524 	 * simple and to get something working, just use a single address space:
525 	 */
526 	mmuv2 = gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION;
527 	dev_dbg(gpu->dev, "mmuv2: %d\n", mmuv2);
528 
529 	if (!mmuv2) {
530 		iommu = etnaviv_iommu_domain_alloc(gpu);
531 		version = ETNAVIV_IOMMU_V1;
532 	} else {
533 		iommu = etnaviv_iommu_v2_domain_alloc(gpu);
534 		version = ETNAVIV_IOMMU_V2;
535 	}
536 
537 	if (!iommu) {
538 		ret = -ENOMEM;
539 		goto fail;
540 	}
541 
542 	/* TODO: we will leak here memory - fix it! */
543 
544 	gpu->mmu = etnaviv_iommu_new(gpu, iommu, version);
545 	if (!gpu->mmu) {
546 		ret = -ENOMEM;
547 		goto fail;
548 	}
549 
550 	/* Create buffer: */
551 	gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
552 	if (!gpu->buffer) {
553 		ret = -ENOMEM;
554 		dev_err(gpu->dev, "could not create command buffer\n");
555 		goto fail;
556 	}
557 	if (gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
558 		ret = -EINVAL;
559 		dev_err(gpu->dev,
560 			"command buffer outside valid memory window\n");
561 		goto free_buffer;
562 	}
563 
564 	/* Setup event management */
565 	spin_lock_init(&gpu->event_spinlock);
566 	init_completion(&gpu->event_free);
567 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
568 		gpu->event[i].used = false;
569 		complete(&gpu->event_free);
570 	}
571 
572 	/* Now program the hardware */
573 	mutex_lock(&gpu->lock);
574 	etnaviv_gpu_hw_init(gpu);
575 	mutex_unlock(&gpu->lock);
576 
577 	pm_runtime_mark_last_busy(gpu->dev);
578 	pm_runtime_put_autosuspend(gpu->dev);
579 
580 	return 0;
581 
582 free_buffer:
583 	etnaviv_gpu_cmdbuf_free(gpu->buffer);
584 	gpu->buffer = NULL;
585 fail:
586 	pm_runtime_mark_last_busy(gpu->dev);
587 	pm_runtime_put_autosuspend(gpu->dev);
588 
589 	return ret;
590 }
591 
592 #ifdef CONFIG_DEBUG_FS
593 struct dma_debug {
594 	u32 address[2];
595 	u32 state[2];
596 };
597 
598 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
599 {
600 	u32 i;
601 
602 	debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
603 	debug->state[0]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
604 
605 	for (i = 0; i < 500; i++) {
606 		debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
607 		debug->state[1]   = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
608 
609 		if (debug->address[0] != debug->address[1])
610 			break;
611 
612 		if (debug->state[0] != debug->state[1])
613 			break;
614 	}
615 }
616 
617 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
618 {
619 	struct dma_debug debug;
620 	u32 dma_lo, dma_hi, axi, idle;
621 	int ret;
622 
623 	seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
624 
625 	ret = pm_runtime_get_sync(gpu->dev);
626 	if (ret < 0)
627 		return ret;
628 
629 	dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
630 	dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
631 	axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
632 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
633 
634 	verify_dma(gpu, &debug);
635 
636 	seq_puts(m, "\tfeatures\n");
637 	seq_printf(m, "\t minor_features0: 0x%08x\n",
638 		   gpu->identity.minor_features0);
639 	seq_printf(m, "\t minor_features1: 0x%08x\n",
640 		   gpu->identity.minor_features1);
641 	seq_printf(m, "\t minor_features2: 0x%08x\n",
642 		   gpu->identity.minor_features2);
643 	seq_printf(m, "\t minor_features3: 0x%08x\n",
644 		   gpu->identity.minor_features3);
645 
646 	seq_puts(m, "\tspecs\n");
647 	seq_printf(m, "\t stream_count:  %d\n",
648 			gpu->identity.stream_count);
649 	seq_printf(m, "\t register_max: %d\n",
650 			gpu->identity.register_max);
651 	seq_printf(m, "\t thread_count: %d\n",
652 			gpu->identity.thread_count);
653 	seq_printf(m, "\t vertex_cache_size: %d\n",
654 			gpu->identity.vertex_cache_size);
655 	seq_printf(m, "\t shader_core_count: %d\n",
656 			gpu->identity.shader_core_count);
657 	seq_printf(m, "\t pixel_pipes: %d\n",
658 			gpu->identity.pixel_pipes);
659 	seq_printf(m, "\t vertex_output_buffer_size: %d\n",
660 			gpu->identity.vertex_output_buffer_size);
661 	seq_printf(m, "\t buffer_size: %d\n",
662 			gpu->identity.buffer_size);
663 	seq_printf(m, "\t instruction_count: %d\n",
664 			gpu->identity.instruction_count);
665 	seq_printf(m, "\t num_constants: %d\n",
666 			gpu->identity.num_constants);
667 
668 	seq_printf(m, "\taxi: 0x%08x\n", axi);
669 	seq_printf(m, "\tidle: 0x%08x\n", idle);
670 	idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
671 	if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
672 		seq_puts(m, "\t FE is not idle\n");
673 	if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
674 		seq_puts(m, "\t DE is not idle\n");
675 	if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
676 		seq_puts(m, "\t PE is not idle\n");
677 	if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
678 		seq_puts(m, "\t SH is not idle\n");
679 	if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
680 		seq_puts(m, "\t PA is not idle\n");
681 	if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
682 		seq_puts(m, "\t SE is not idle\n");
683 	if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
684 		seq_puts(m, "\t RA is not idle\n");
685 	if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
686 		seq_puts(m, "\t TX is not idle\n");
687 	if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
688 		seq_puts(m, "\t VG is not idle\n");
689 	if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
690 		seq_puts(m, "\t IM is not idle\n");
691 	if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
692 		seq_puts(m, "\t FP is not idle\n");
693 	if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
694 		seq_puts(m, "\t TS is not idle\n");
695 	if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
696 		seq_puts(m, "\t AXI low power mode\n");
697 
698 	if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
699 		u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
700 		u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
701 		u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
702 
703 		seq_puts(m, "\tMC\n");
704 		seq_printf(m, "\t read0: 0x%08x\n", read0);
705 		seq_printf(m, "\t read1: 0x%08x\n", read1);
706 		seq_printf(m, "\t write: 0x%08x\n", write);
707 	}
708 
709 	seq_puts(m, "\tDMA ");
710 
711 	if (debug.address[0] == debug.address[1] &&
712 	    debug.state[0] == debug.state[1]) {
713 		seq_puts(m, "seems to be stuck\n");
714 	} else if (debug.address[0] == debug.address[1]) {
715 		seq_puts(m, "adress is constant\n");
716 	} else {
717 		seq_puts(m, "is runing\n");
718 	}
719 
720 	seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
721 	seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
722 	seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
723 	seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
724 	seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
725 		   dma_lo, dma_hi);
726 
727 	ret = 0;
728 
729 	pm_runtime_mark_last_busy(gpu->dev);
730 	pm_runtime_put_autosuspend(gpu->dev);
731 
732 	return ret;
733 }
734 #endif
735 
736 /*
737  * Power Management:
738  */
739 static int enable_clk(struct etnaviv_gpu *gpu)
740 {
741 	if (gpu->clk_core)
742 		clk_prepare_enable(gpu->clk_core);
743 	if (gpu->clk_shader)
744 		clk_prepare_enable(gpu->clk_shader);
745 
746 	return 0;
747 }
748 
749 static int disable_clk(struct etnaviv_gpu *gpu)
750 {
751 	if (gpu->clk_core)
752 		clk_disable_unprepare(gpu->clk_core);
753 	if (gpu->clk_shader)
754 		clk_disable_unprepare(gpu->clk_shader);
755 
756 	return 0;
757 }
758 
759 static int enable_axi(struct etnaviv_gpu *gpu)
760 {
761 	if (gpu->clk_bus)
762 		clk_prepare_enable(gpu->clk_bus);
763 
764 	return 0;
765 }
766 
767 static int disable_axi(struct etnaviv_gpu *gpu)
768 {
769 	if (gpu->clk_bus)
770 		clk_disable_unprepare(gpu->clk_bus);
771 
772 	return 0;
773 }
774 
775 /*
776  * Hangcheck detection for locked gpu:
777  */
778 static void recover_worker(struct work_struct *work)
779 {
780 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
781 					       recover_work);
782 	unsigned long flags;
783 	unsigned int i;
784 
785 	dev_err(gpu->dev, "hangcheck recover!\n");
786 
787 	if (pm_runtime_get_sync(gpu->dev) < 0)
788 		return;
789 
790 	mutex_lock(&gpu->lock);
791 
792 	/* Only catch the first event, or when manually re-armed */
793 	if (etnaviv_dump_core) {
794 		etnaviv_core_dump(gpu);
795 		etnaviv_dump_core = false;
796 	}
797 
798 	etnaviv_hw_reset(gpu);
799 
800 	/* complete all events, the GPU won't do it after the reset */
801 	spin_lock_irqsave(&gpu->event_spinlock, flags);
802 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
803 		if (!gpu->event[i].used)
804 			continue;
805 		fence_signal(gpu->event[i].fence);
806 		gpu->event[i].fence = NULL;
807 		gpu->event[i].used = false;
808 		complete(&gpu->event_free);
809 		/*
810 		 * Decrement the PM count for each stuck event. This is safe
811 		 * even in atomic context as we use ASYNC RPM here.
812 		 */
813 		pm_runtime_put_autosuspend(gpu->dev);
814 	}
815 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
816 	gpu->completed_fence = gpu->active_fence;
817 
818 	etnaviv_gpu_hw_init(gpu);
819 	gpu->switch_context = true;
820 
821 	mutex_unlock(&gpu->lock);
822 	pm_runtime_mark_last_busy(gpu->dev);
823 	pm_runtime_put_autosuspend(gpu->dev);
824 
825 	/* Retire the buffer objects in a work */
826 	etnaviv_queue_work(gpu->drm, &gpu->retire_work);
827 }
828 
829 static void hangcheck_timer_reset(struct etnaviv_gpu *gpu)
830 {
831 	DBG("%s", dev_name(gpu->dev));
832 	mod_timer(&gpu->hangcheck_timer,
833 		  round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES));
834 }
835 
836 static void hangcheck_handler(unsigned long data)
837 {
838 	struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data;
839 	u32 fence = gpu->completed_fence;
840 	bool progress = false;
841 
842 	if (fence != gpu->hangcheck_fence) {
843 		gpu->hangcheck_fence = fence;
844 		progress = true;
845 	}
846 
847 	if (!progress) {
848 		u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
849 		int change = dma_addr - gpu->hangcheck_dma_addr;
850 
851 		if (change < 0 || change > 16) {
852 			gpu->hangcheck_dma_addr = dma_addr;
853 			progress = true;
854 		}
855 	}
856 
857 	if (!progress && fence_after(gpu->active_fence, fence)) {
858 		dev_err(gpu->dev, "hangcheck detected gpu lockup!\n");
859 		dev_err(gpu->dev, "     completed fence: %u\n", fence);
860 		dev_err(gpu->dev, "     active fence: %u\n",
861 			gpu->active_fence);
862 		etnaviv_queue_work(gpu->drm, &gpu->recover_work);
863 	}
864 
865 	/* if still more pending work, reset the hangcheck timer: */
866 	if (fence_after(gpu->active_fence, gpu->hangcheck_fence))
867 		hangcheck_timer_reset(gpu);
868 }
869 
870 static void hangcheck_disable(struct etnaviv_gpu *gpu)
871 {
872 	del_timer_sync(&gpu->hangcheck_timer);
873 	cancel_work_sync(&gpu->recover_work);
874 }
875 
876 /* fence object management */
877 struct etnaviv_fence {
878 	struct etnaviv_gpu *gpu;
879 	struct fence base;
880 };
881 
882 static inline struct etnaviv_fence *to_etnaviv_fence(struct fence *fence)
883 {
884 	return container_of(fence, struct etnaviv_fence, base);
885 }
886 
887 static const char *etnaviv_fence_get_driver_name(struct fence *fence)
888 {
889 	return "etnaviv";
890 }
891 
892 static const char *etnaviv_fence_get_timeline_name(struct fence *fence)
893 {
894 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
895 
896 	return dev_name(f->gpu->dev);
897 }
898 
899 static bool etnaviv_fence_enable_signaling(struct fence *fence)
900 {
901 	return true;
902 }
903 
904 static bool etnaviv_fence_signaled(struct fence *fence)
905 {
906 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
907 
908 	return fence_completed(f->gpu, f->base.seqno);
909 }
910 
911 static void etnaviv_fence_release(struct fence *fence)
912 {
913 	struct etnaviv_fence *f = to_etnaviv_fence(fence);
914 
915 	kfree_rcu(f, base.rcu);
916 }
917 
918 static const struct fence_ops etnaviv_fence_ops = {
919 	.get_driver_name = etnaviv_fence_get_driver_name,
920 	.get_timeline_name = etnaviv_fence_get_timeline_name,
921 	.enable_signaling = etnaviv_fence_enable_signaling,
922 	.signaled = etnaviv_fence_signaled,
923 	.wait = fence_default_wait,
924 	.release = etnaviv_fence_release,
925 };
926 
927 static struct fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
928 {
929 	struct etnaviv_fence *f;
930 
931 	f = kzalloc(sizeof(*f), GFP_KERNEL);
932 	if (!f)
933 		return NULL;
934 
935 	f->gpu = gpu;
936 
937 	fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
938 		   gpu->fence_context, ++gpu->next_fence);
939 
940 	return &f->base;
941 }
942 
943 int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj,
944 	unsigned int context, bool exclusive)
945 {
946 	struct reservation_object *robj = etnaviv_obj->resv;
947 	struct reservation_object_list *fobj;
948 	struct fence *fence;
949 	int i, ret;
950 
951 	if (!exclusive) {
952 		ret = reservation_object_reserve_shared(robj);
953 		if (ret)
954 			return ret;
955 	}
956 
957 	/*
958 	 * If we have any shared fences, then the exclusive fence
959 	 * should be ignored as it will already have been signalled.
960 	 */
961 	fobj = reservation_object_get_list(robj);
962 	if (!fobj || fobj->shared_count == 0) {
963 		/* Wait on any existing exclusive fence which isn't our own */
964 		fence = reservation_object_get_excl(robj);
965 		if (fence && fence->context != context) {
966 			ret = fence_wait(fence, true);
967 			if (ret)
968 				return ret;
969 		}
970 	}
971 
972 	if (!exclusive || !fobj)
973 		return 0;
974 
975 	for (i = 0; i < fobj->shared_count; i++) {
976 		fence = rcu_dereference_protected(fobj->shared[i],
977 						reservation_object_held(robj));
978 		if (fence->context != context) {
979 			ret = fence_wait(fence, true);
980 			if (ret)
981 				return ret;
982 		}
983 	}
984 
985 	return 0;
986 }
987 
988 /*
989  * event management:
990  */
991 
992 static unsigned int event_alloc(struct etnaviv_gpu *gpu)
993 {
994 	unsigned long ret, flags;
995 	unsigned int i, event = ~0U;
996 
997 	ret = wait_for_completion_timeout(&gpu->event_free,
998 					  msecs_to_jiffies(10 * 10000));
999 	if (!ret)
1000 		dev_err(gpu->dev, "wait_for_completion_timeout failed");
1001 
1002 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1003 
1004 	/* find first free event */
1005 	for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
1006 		if (gpu->event[i].used == false) {
1007 			gpu->event[i].used = true;
1008 			event = i;
1009 			break;
1010 		}
1011 	}
1012 
1013 	spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1014 
1015 	return event;
1016 }
1017 
1018 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1019 {
1020 	unsigned long flags;
1021 
1022 	spin_lock_irqsave(&gpu->event_spinlock, flags);
1023 
1024 	if (gpu->event[event].used == false) {
1025 		dev_warn(gpu->dev, "event %u is already marked as free",
1026 			 event);
1027 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1028 	} else {
1029 		gpu->event[event].used = false;
1030 		spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1031 
1032 		complete(&gpu->event_free);
1033 	}
1034 }
1035 
1036 /*
1037  * Cmdstream submission/retirement:
1038  */
1039 
1040 struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
1041 	size_t nr_bos)
1042 {
1043 	struct etnaviv_cmdbuf *cmdbuf;
1044 	size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo[0]),
1045 				 sizeof(*cmdbuf));
1046 
1047 	cmdbuf = kzalloc(sz, GFP_KERNEL);
1048 	if (!cmdbuf)
1049 		return NULL;
1050 
1051 	cmdbuf->vaddr = dma_alloc_writecombine(gpu->dev, size, &cmdbuf->paddr,
1052 					       GFP_KERNEL);
1053 	if (!cmdbuf->vaddr) {
1054 		kfree(cmdbuf);
1055 		return NULL;
1056 	}
1057 
1058 	cmdbuf->gpu = gpu;
1059 	cmdbuf->size = size;
1060 
1061 	return cmdbuf;
1062 }
1063 
1064 void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1065 {
1066 	dma_free_writecombine(cmdbuf->gpu->dev, cmdbuf->size,
1067 			      cmdbuf->vaddr, cmdbuf->paddr);
1068 	kfree(cmdbuf);
1069 }
1070 
1071 static void retire_worker(struct work_struct *work)
1072 {
1073 	struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1074 					       retire_work);
1075 	u32 fence = gpu->completed_fence;
1076 	struct etnaviv_cmdbuf *cmdbuf, *tmp;
1077 	unsigned int i;
1078 
1079 	mutex_lock(&gpu->lock);
1080 	list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1081 		if (!fence_is_signaled(cmdbuf->fence))
1082 			break;
1083 
1084 		list_del(&cmdbuf->node);
1085 		fence_put(cmdbuf->fence);
1086 
1087 		for (i = 0; i < cmdbuf->nr_bos; i++) {
1088 			struct etnaviv_gem_object *etnaviv_obj = cmdbuf->bo[i];
1089 
1090 			atomic_dec(&etnaviv_obj->gpu_active);
1091 			/* drop the refcount taken in etnaviv_gpu_submit */
1092 			etnaviv_gem_put_iova(gpu, &etnaviv_obj->base);
1093 		}
1094 
1095 		etnaviv_gpu_cmdbuf_free(cmdbuf);
1096 	}
1097 
1098 	gpu->retired_fence = fence;
1099 
1100 	mutex_unlock(&gpu->lock);
1101 
1102 	wake_up_all(&gpu->fence_event);
1103 }
1104 
1105 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1106 	u32 fence, struct timespec *timeout)
1107 {
1108 	int ret;
1109 
1110 	if (fence_after(fence, gpu->next_fence)) {
1111 		DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1112 				fence, gpu->next_fence);
1113 		return -EINVAL;
1114 	}
1115 
1116 	if (!timeout) {
1117 		/* No timeout was requested: just test for completion */
1118 		ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1119 	} else {
1120 		unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1121 
1122 		ret = wait_event_interruptible_timeout(gpu->fence_event,
1123 						fence_completed(gpu, fence),
1124 						remaining);
1125 		if (ret == 0) {
1126 			DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1127 				fence, gpu->retired_fence,
1128 				gpu->completed_fence);
1129 			ret = -ETIMEDOUT;
1130 		} else if (ret != -ERESTARTSYS) {
1131 			ret = 0;
1132 		}
1133 	}
1134 
1135 	return ret;
1136 }
1137 
1138 /*
1139  * Wait for an object to become inactive.  This, on it's own, is not race
1140  * free: the object is moved by the retire worker off the active list, and
1141  * then the iova is put.  Moreover, the object could be re-submitted just
1142  * after we notice that it's become inactive.
1143  *
1144  * Although the retirement happens under the gpu lock, we don't want to hold
1145  * that lock in this function while waiting.
1146  */
1147 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1148 	struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1149 {
1150 	unsigned long remaining;
1151 	long ret;
1152 
1153 	if (!timeout)
1154 		return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1155 
1156 	remaining = etnaviv_timeout_to_jiffies(timeout);
1157 
1158 	ret = wait_event_interruptible_timeout(gpu->fence_event,
1159 					       !is_active(etnaviv_obj),
1160 					       remaining);
1161 	if (ret > 0) {
1162 		struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1163 
1164 		/* Synchronise with the retire worker */
1165 		flush_workqueue(priv->wq);
1166 		return 0;
1167 	} else if (ret == -ERESTARTSYS) {
1168 		return -ERESTARTSYS;
1169 	} else {
1170 		return -ETIMEDOUT;
1171 	}
1172 }
1173 
1174 int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1175 {
1176 	return pm_runtime_get_sync(gpu->dev);
1177 }
1178 
1179 void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1180 {
1181 	pm_runtime_mark_last_busy(gpu->dev);
1182 	pm_runtime_put_autosuspend(gpu->dev);
1183 }
1184 
1185 /* add bo's to gpu's ring, and kick gpu: */
1186 int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1187 	struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1188 {
1189 	struct fence *fence;
1190 	unsigned int event, i;
1191 	int ret;
1192 
1193 	ret = etnaviv_gpu_pm_get_sync(gpu);
1194 	if (ret < 0)
1195 		return ret;
1196 
1197 	mutex_lock(&gpu->lock);
1198 
1199 	/*
1200 	 * TODO
1201 	 *
1202 	 * - flush
1203 	 * - data endian
1204 	 * - prefetch
1205 	 *
1206 	 */
1207 
1208 	event = event_alloc(gpu);
1209 	if (unlikely(event == ~0U)) {
1210 		DRM_ERROR("no free event\n");
1211 		ret = -EBUSY;
1212 		goto out_unlock;
1213 	}
1214 
1215 	fence = etnaviv_gpu_fence_alloc(gpu);
1216 	if (!fence) {
1217 		event_free(gpu, event);
1218 		ret = -ENOMEM;
1219 		goto out_unlock;
1220 	}
1221 
1222 	gpu->event[event].fence = fence;
1223 	submit->fence = fence->seqno;
1224 	gpu->active_fence = submit->fence;
1225 
1226 	if (gpu->lastctx != cmdbuf->ctx) {
1227 		gpu->mmu->need_flush = true;
1228 		gpu->switch_context = true;
1229 		gpu->lastctx = cmdbuf->ctx;
1230 	}
1231 
1232 	etnaviv_buffer_queue(gpu, event, cmdbuf);
1233 
1234 	cmdbuf->fence = fence;
1235 	list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1236 
1237 	/* We're committed to adding this command buffer, hold a PM reference */
1238 	pm_runtime_get_noresume(gpu->dev);
1239 
1240 	for (i = 0; i < submit->nr_bos; i++) {
1241 		struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1242 		u32 iova;
1243 
1244 		/* Each cmdbuf takes a refcount on the iova */
1245 		etnaviv_gem_get_iova(gpu, &etnaviv_obj->base, &iova);
1246 		cmdbuf->bo[i] = etnaviv_obj;
1247 		atomic_inc(&etnaviv_obj->gpu_active);
1248 
1249 		if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1250 			reservation_object_add_excl_fence(etnaviv_obj->resv,
1251 							  fence);
1252 		else
1253 			reservation_object_add_shared_fence(etnaviv_obj->resv,
1254 							    fence);
1255 	}
1256 	cmdbuf->nr_bos = submit->nr_bos;
1257 	hangcheck_timer_reset(gpu);
1258 	ret = 0;
1259 
1260 out_unlock:
1261 	mutex_unlock(&gpu->lock);
1262 
1263 	etnaviv_gpu_pm_put(gpu);
1264 
1265 	return ret;
1266 }
1267 
1268 /*
1269  * Init/Cleanup:
1270  */
1271 static irqreturn_t irq_handler(int irq, void *data)
1272 {
1273 	struct etnaviv_gpu *gpu = data;
1274 	irqreturn_t ret = IRQ_NONE;
1275 
1276 	u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1277 
1278 	if (intr != 0) {
1279 		int event;
1280 
1281 		pm_runtime_mark_last_busy(gpu->dev);
1282 
1283 		dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1284 
1285 		if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1286 			dev_err(gpu->dev, "AXI bus error\n");
1287 			intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1288 		}
1289 
1290 		while ((event = ffs(intr)) != 0) {
1291 			struct fence *fence;
1292 
1293 			event -= 1;
1294 
1295 			intr &= ~(1 << event);
1296 
1297 			dev_dbg(gpu->dev, "event %u\n", event);
1298 
1299 			fence = gpu->event[event].fence;
1300 			gpu->event[event].fence = NULL;
1301 			fence_signal(fence);
1302 
1303 			/*
1304 			 * Events can be processed out of order.  Eg,
1305 			 * - allocate and queue event 0
1306 			 * - allocate event 1
1307 			 * - event 0 completes, we process it
1308 			 * - allocate and queue event 0
1309 			 * - event 1 and event 0 complete
1310 			 * we can end up processing event 0 first, then 1.
1311 			 */
1312 			if (fence_after(fence->seqno, gpu->completed_fence))
1313 				gpu->completed_fence = fence->seqno;
1314 
1315 			event_free(gpu, event);
1316 
1317 			/*
1318 			 * We need to balance the runtime PM count caused by
1319 			 * each submission.  Upon submission, we increment
1320 			 * the runtime PM counter, and allocate one event.
1321 			 * So here, we put the runtime PM count for each
1322 			 * completed event.
1323 			 */
1324 			pm_runtime_put_autosuspend(gpu->dev);
1325 		}
1326 
1327 		/* Retire the buffer objects in a work */
1328 		etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1329 
1330 		ret = IRQ_HANDLED;
1331 	}
1332 
1333 	return ret;
1334 }
1335 
1336 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1337 {
1338 	int ret;
1339 
1340 	ret = enable_clk(gpu);
1341 	if (ret)
1342 		return ret;
1343 
1344 	ret = enable_axi(gpu);
1345 	if (ret) {
1346 		disable_clk(gpu);
1347 		return ret;
1348 	}
1349 
1350 	return 0;
1351 }
1352 
1353 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1354 {
1355 	int ret;
1356 
1357 	ret = disable_axi(gpu);
1358 	if (ret)
1359 		return ret;
1360 
1361 	ret = disable_clk(gpu);
1362 	if (ret)
1363 		return ret;
1364 
1365 	return 0;
1366 }
1367 
1368 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1369 {
1370 	if (gpu->buffer) {
1371 		unsigned long timeout;
1372 
1373 		/* Replace the last WAIT with END */
1374 		etnaviv_buffer_end(gpu);
1375 
1376 		/*
1377 		 * We know that only the FE is busy here, this should
1378 		 * happen quickly (as the WAIT is only 200 cycles).  If
1379 		 * we fail, just warn and continue.
1380 		 */
1381 		timeout = jiffies + msecs_to_jiffies(100);
1382 		do {
1383 			u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1384 
1385 			if ((idle & gpu->idle_mask) == gpu->idle_mask)
1386 				break;
1387 
1388 			if (time_is_before_jiffies(timeout)) {
1389 				dev_warn(gpu->dev,
1390 					 "timed out waiting for idle: idle=0x%x\n",
1391 					 idle);
1392 				break;
1393 			}
1394 
1395 			udelay(5);
1396 		} while (1);
1397 	}
1398 
1399 	return etnaviv_gpu_clk_disable(gpu);
1400 }
1401 
1402 #ifdef CONFIG_PM
1403 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1404 {
1405 	u32 clock;
1406 	int ret;
1407 
1408 	ret = mutex_lock_killable(&gpu->lock);
1409 	if (ret)
1410 		return ret;
1411 
1412 	clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1413 		VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1414 
1415 	etnaviv_gpu_load_clock(gpu, clock);
1416 	etnaviv_gpu_hw_init(gpu);
1417 
1418 	gpu->switch_context = true;
1419 
1420 	mutex_unlock(&gpu->lock);
1421 
1422 	return 0;
1423 }
1424 #endif
1425 
1426 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1427 	void *data)
1428 {
1429 	struct drm_device *drm = data;
1430 	struct etnaviv_drm_private *priv = drm->dev_private;
1431 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1432 	int ret;
1433 
1434 #ifdef CONFIG_PM
1435 	ret = pm_runtime_get_sync(gpu->dev);
1436 #else
1437 	ret = etnaviv_gpu_clk_enable(gpu);
1438 #endif
1439 	if (ret < 0)
1440 		return ret;
1441 
1442 	gpu->drm = drm;
1443 	gpu->fence_context = fence_context_alloc(1);
1444 	spin_lock_init(&gpu->fence_spinlock);
1445 
1446 	INIT_LIST_HEAD(&gpu->active_cmd_list);
1447 	INIT_WORK(&gpu->retire_work, retire_worker);
1448 	INIT_WORK(&gpu->recover_work, recover_worker);
1449 	init_waitqueue_head(&gpu->fence_event);
1450 
1451 	setup_timer(&gpu->hangcheck_timer, hangcheck_handler,
1452 			(unsigned long)gpu);
1453 
1454 	priv->gpu[priv->num_gpus++] = gpu;
1455 
1456 	pm_runtime_mark_last_busy(gpu->dev);
1457 	pm_runtime_put_autosuspend(gpu->dev);
1458 
1459 	return 0;
1460 }
1461 
1462 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1463 	void *data)
1464 {
1465 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1466 
1467 	DBG("%s", dev_name(gpu->dev));
1468 
1469 	hangcheck_disable(gpu);
1470 
1471 #ifdef CONFIG_PM
1472 	pm_runtime_get_sync(gpu->dev);
1473 	pm_runtime_put_sync_suspend(gpu->dev);
1474 #else
1475 	etnaviv_gpu_hw_suspend(gpu);
1476 #endif
1477 
1478 	if (gpu->buffer) {
1479 		etnaviv_gpu_cmdbuf_free(gpu->buffer);
1480 		gpu->buffer = NULL;
1481 	}
1482 
1483 	if (gpu->mmu) {
1484 		etnaviv_iommu_destroy(gpu->mmu);
1485 		gpu->mmu = NULL;
1486 	}
1487 
1488 	gpu->drm = NULL;
1489 }
1490 
1491 static const struct component_ops gpu_ops = {
1492 	.bind = etnaviv_gpu_bind,
1493 	.unbind = etnaviv_gpu_unbind,
1494 };
1495 
1496 static const struct of_device_id etnaviv_gpu_match[] = {
1497 	{
1498 		.compatible = "vivante,gc"
1499 	},
1500 	{ /* sentinel */ }
1501 };
1502 
1503 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1504 {
1505 	struct device *dev = &pdev->dev;
1506 	struct etnaviv_gpu *gpu;
1507 	int err = 0;
1508 
1509 	gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1510 	if (!gpu)
1511 		return -ENOMEM;
1512 
1513 	gpu->dev = &pdev->dev;
1514 	mutex_init(&gpu->lock);
1515 
1516 	/*
1517 	 * Set the GPU base address to the start of physical memory.  This
1518 	 * ensures that if we have up to 2GB, the v1 MMU can address the
1519 	 * highest memory.  This is important as command buffers may be
1520 	 * allocated outside of this limit.
1521 	 */
1522 	gpu->memory_base = PHYS_OFFSET;
1523 
1524 	/* Map registers: */
1525 	gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1526 	if (IS_ERR(gpu->mmio))
1527 		return PTR_ERR(gpu->mmio);
1528 
1529 	/* Get Interrupt: */
1530 	gpu->irq = platform_get_irq(pdev, 0);
1531 	if (gpu->irq < 0) {
1532 		err = gpu->irq;
1533 		dev_err(dev, "failed to get irq: %d\n", err);
1534 		goto fail;
1535 	}
1536 
1537 	err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1538 			       dev_name(gpu->dev), gpu);
1539 	if (err) {
1540 		dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1541 		goto fail;
1542 	}
1543 
1544 	/* Get Clocks: */
1545 	gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1546 	DBG("clk_bus: %p", gpu->clk_bus);
1547 	if (IS_ERR(gpu->clk_bus))
1548 		gpu->clk_bus = NULL;
1549 
1550 	gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1551 	DBG("clk_core: %p", gpu->clk_core);
1552 	if (IS_ERR(gpu->clk_core))
1553 		gpu->clk_core = NULL;
1554 
1555 	gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1556 	DBG("clk_shader: %p", gpu->clk_shader);
1557 	if (IS_ERR(gpu->clk_shader))
1558 		gpu->clk_shader = NULL;
1559 
1560 	/* TODO: figure out max mapped size */
1561 	dev_set_drvdata(dev, gpu);
1562 
1563 	/*
1564 	 * We treat the device as initially suspended.  The runtime PM
1565 	 * autosuspend delay is rather arbitary: no measurements have
1566 	 * yet been performed to determine an appropriate value.
1567 	 */
1568 	pm_runtime_use_autosuspend(gpu->dev);
1569 	pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1570 	pm_runtime_enable(gpu->dev);
1571 
1572 	err = component_add(&pdev->dev, &gpu_ops);
1573 	if (err < 0) {
1574 		dev_err(&pdev->dev, "failed to register component: %d\n", err);
1575 		goto fail;
1576 	}
1577 
1578 	return 0;
1579 
1580 fail:
1581 	return err;
1582 }
1583 
1584 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1585 {
1586 	component_del(&pdev->dev, &gpu_ops);
1587 	pm_runtime_disable(&pdev->dev);
1588 	return 0;
1589 }
1590 
1591 #ifdef CONFIG_PM
1592 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1593 {
1594 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1595 	u32 idle, mask;
1596 
1597 	/* If we have outstanding fences, we're not idle */
1598 	if (gpu->completed_fence != gpu->active_fence)
1599 		return -EBUSY;
1600 
1601 	/* Check whether the hardware (except FE) is idle */
1602 	mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1603 	idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1604 	if (idle != mask)
1605 		return -EBUSY;
1606 
1607 	return etnaviv_gpu_hw_suspend(gpu);
1608 }
1609 
1610 static int etnaviv_gpu_rpm_resume(struct device *dev)
1611 {
1612 	struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1613 	int ret;
1614 
1615 	ret = etnaviv_gpu_clk_enable(gpu);
1616 	if (ret)
1617 		return ret;
1618 
1619 	/* Re-initialise the basic hardware state */
1620 	if (gpu->drm && gpu->buffer) {
1621 		ret = etnaviv_gpu_hw_resume(gpu);
1622 		if (ret) {
1623 			etnaviv_gpu_clk_disable(gpu);
1624 			return ret;
1625 		}
1626 	}
1627 
1628 	return 0;
1629 }
1630 #endif
1631 
1632 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1633 	SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1634 			   NULL)
1635 };
1636 
1637 struct platform_driver etnaviv_gpu_driver = {
1638 	.driver = {
1639 		.name = "etnaviv-gpu",
1640 		.owner = THIS_MODULE,
1641 		.pm = &etnaviv_gpu_pm_ops,
1642 		.of_match_table = etnaviv_gpu_match,
1643 	},
1644 	.probe = etnaviv_gpu_platform_probe,
1645 	.remove = etnaviv_gpu_platform_remove,
1646 	.id_table = gpu_ids,
1647 };
1648