Loading...
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
32static const struct platform_device_id gpu_ids[] = {
33 { .name = "etnaviv-gpu,2d" },
34 { },
35};
36
37static bool etnaviv_dump_core = true;
38module_param_named(dump_core, etnaviv_dump_core, bool, 0600);
39
40/*
41 * Driver functions:
42 */
43
44int 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
142static 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
292static 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
393static 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
400static 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
490static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
491{
492 u16 prefetch;
493
494 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
495 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
496 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
497 u32 mc_memory_debug;
498
499 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
500
501 if (gpu->identity.revision == 0x5007)
502 mc_memory_debug |= 0x0c;
503 else
504 mc_memory_debug |= 0x08;
505
506 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
507 }
508
509 /*
510 * Update GPU AXI cache atttribute to "cacheable, no allocate".
511 * This is necessary to prevent the iMX6 SoC locking up.
512 */
513 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
514 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
515 VIVS_HI_AXI_CONFIG_ARCACHE(2));
516
517 /* GC2000 rev 5108 needs a special bus config */
518 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
519 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
520 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
521 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
522 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
523 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
524 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
525 }
526
527 /* set base addresses */
528 gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_RA, gpu->memory_base);
529 gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_FE, gpu->memory_base);
530 gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_TX, gpu->memory_base);
531 gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PEZ, gpu->memory_base);
532 gpu_write(gpu, VIVS_MC_MEMORY_BASE_ADDR_PE, gpu->memory_base);
533
534 /* setup the MMU page table pointers */
535 etnaviv_iommu_domain_restore(gpu, gpu->mmu->domain);
536
537 /* Start command processor */
538 prefetch = etnaviv_buffer_init(gpu);
539
540 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
541 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS,
542 gpu->buffer->paddr - gpu->memory_base);
543 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
544 VIVS_FE_COMMAND_CONTROL_ENABLE |
545 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
546}
547
548int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
549{
550 int ret, i;
551 struct iommu_domain *iommu;
552 enum etnaviv_iommu_version version;
553 bool mmuv2;
554
555 ret = pm_runtime_get_sync(gpu->dev);
556 if (ret < 0)
557 return ret;
558
559 etnaviv_hw_identify(gpu);
560
561 if (gpu->identity.model == 0) {
562 dev_err(gpu->dev, "Unknown GPU model\n");
563 ret = -ENXIO;
564 goto fail;
565 }
566
567 /* Exclude VG cores with FE2.0 */
568 if (gpu->identity.features & chipFeatures_PIPE_VG &&
569 gpu->identity.features & chipFeatures_FE20) {
570 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
571 ret = -ENXIO;
572 goto fail;
573 }
574
575 /*
576 * Set the GPU linear window to be at the end of the DMA window, where
577 * the CMA area is likely to reside. This ensures that we are able to
578 * map the command buffers while having the linear window overlap as
579 * much RAM as possible, so we can optimize mappings for other buffers.
580 *
581 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
582 * to different views of the memory on the individual engines.
583 */
584 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
585 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
586 u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
587 if (dma_mask < PHYS_OFFSET + SZ_2G)
588 gpu->memory_base = PHYS_OFFSET;
589 else
590 gpu->memory_base = dma_mask - SZ_2G + 1;
591 }
592
593 ret = etnaviv_hw_reset(gpu);
594 if (ret)
595 goto fail;
596
597 /* Setup IOMMU.. eventually we will (I think) do this once per context
598 * and have separate page tables per context. For now, to keep things
599 * simple and to get something working, just use a single address space:
600 */
601 mmuv2 = gpu->identity.minor_features1 & chipMinorFeatures1_MMU_VERSION;
602 dev_dbg(gpu->dev, "mmuv2: %d\n", mmuv2);
603
604 if (!mmuv2) {
605 iommu = etnaviv_iommu_domain_alloc(gpu);
606 version = ETNAVIV_IOMMU_V1;
607 } else {
608 iommu = etnaviv_iommu_v2_domain_alloc(gpu);
609 version = ETNAVIV_IOMMU_V2;
610 }
611
612 if (!iommu) {
613 ret = -ENOMEM;
614 goto fail;
615 }
616
617 gpu->mmu = etnaviv_iommu_new(gpu, iommu, version);
618 if (!gpu->mmu) {
619 iommu_domain_free(iommu);
620 ret = -ENOMEM;
621 goto fail;
622 }
623
624 /* Create buffer: */
625 gpu->buffer = etnaviv_gpu_cmdbuf_new(gpu, PAGE_SIZE, 0);
626 if (!gpu->buffer) {
627 ret = -ENOMEM;
628 dev_err(gpu->dev, "could not create command buffer\n");
629 goto destroy_iommu;
630 }
631 if (gpu->buffer->paddr - gpu->memory_base > 0x80000000) {
632 ret = -EINVAL;
633 dev_err(gpu->dev,
634 "command buffer outside valid memory window\n");
635 goto free_buffer;
636 }
637
638 /* Setup event management */
639 spin_lock_init(&gpu->event_spinlock);
640 init_completion(&gpu->event_free);
641 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
642 gpu->event[i].used = false;
643 complete(&gpu->event_free);
644 }
645
646 /* Now program the hardware */
647 mutex_lock(&gpu->lock);
648 etnaviv_gpu_hw_init(gpu);
649 gpu->exec_state = -1;
650 mutex_unlock(&gpu->lock);
651
652 pm_runtime_mark_last_busy(gpu->dev);
653 pm_runtime_put_autosuspend(gpu->dev);
654
655 return 0;
656
657free_buffer:
658 etnaviv_gpu_cmdbuf_free(gpu->buffer);
659 gpu->buffer = NULL;
660destroy_iommu:
661 etnaviv_iommu_destroy(gpu->mmu);
662 gpu->mmu = NULL;
663fail:
664 pm_runtime_mark_last_busy(gpu->dev);
665 pm_runtime_put_autosuspend(gpu->dev);
666
667 return ret;
668}
669
670#ifdef CONFIG_DEBUG_FS
671struct dma_debug {
672 u32 address[2];
673 u32 state[2];
674};
675
676static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
677{
678 u32 i;
679
680 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
681 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
682
683 for (i = 0; i < 500; i++) {
684 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
685 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
686
687 if (debug->address[0] != debug->address[1])
688 break;
689
690 if (debug->state[0] != debug->state[1])
691 break;
692 }
693}
694
695int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
696{
697 struct dma_debug debug;
698 u32 dma_lo, dma_hi, axi, idle;
699 int ret;
700
701 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
702
703 ret = pm_runtime_get_sync(gpu->dev);
704 if (ret < 0)
705 return ret;
706
707 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
708 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
709 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
710 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
711
712 verify_dma(gpu, &debug);
713
714 seq_puts(m, "\tfeatures\n");
715 seq_printf(m, "\t minor_features0: 0x%08x\n",
716 gpu->identity.minor_features0);
717 seq_printf(m, "\t minor_features1: 0x%08x\n",
718 gpu->identity.minor_features1);
719 seq_printf(m, "\t minor_features2: 0x%08x\n",
720 gpu->identity.minor_features2);
721 seq_printf(m, "\t minor_features3: 0x%08x\n",
722 gpu->identity.minor_features3);
723 seq_printf(m, "\t minor_features4: 0x%08x\n",
724 gpu->identity.minor_features4);
725 seq_printf(m, "\t minor_features5: 0x%08x\n",
726 gpu->identity.minor_features5);
727
728 seq_puts(m, "\tspecs\n");
729 seq_printf(m, "\t stream_count: %d\n",
730 gpu->identity.stream_count);
731 seq_printf(m, "\t register_max: %d\n",
732 gpu->identity.register_max);
733 seq_printf(m, "\t thread_count: %d\n",
734 gpu->identity.thread_count);
735 seq_printf(m, "\t vertex_cache_size: %d\n",
736 gpu->identity.vertex_cache_size);
737 seq_printf(m, "\t shader_core_count: %d\n",
738 gpu->identity.shader_core_count);
739 seq_printf(m, "\t pixel_pipes: %d\n",
740 gpu->identity.pixel_pipes);
741 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
742 gpu->identity.vertex_output_buffer_size);
743 seq_printf(m, "\t buffer_size: %d\n",
744 gpu->identity.buffer_size);
745 seq_printf(m, "\t instruction_count: %d\n",
746 gpu->identity.instruction_count);
747 seq_printf(m, "\t num_constants: %d\n",
748 gpu->identity.num_constants);
749 seq_printf(m, "\t varyings_count: %d\n",
750 gpu->identity.varyings_count);
751
752 seq_printf(m, "\taxi: 0x%08x\n", axi);
753 seq_printf(m, "\tidle: 0x%08x\n", idle);
754 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
755 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
756 seq_puts(m, "\t FE is not idle\n");
757 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
758 seq_puts(m, "\t DE is not idle\n");
759 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
760 seq_puts(m, "\t PE is not idle\n");
761 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
762 seq_puts(m, "\t SH is not idle\n");
763 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
764 seq_puts(m, "\t PA is not idle\n");
765 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
766 seq_puts(m, "\t SE is not idle\n");
767 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
768 seq_puts(m, "\t RA is not idle\n");
769 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
770 seq_puts(m, "\t TX is not idle\n");
771 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
772 seq_puts(m, "\t VG is not idle\n");
773 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
774 seq_puts(m, "\t IM is not idle\n");
775 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
776 seq_puts(m, "\t FP is not idle\n");
777 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
778 seq_puts(m, "\t TS is not idle\n");
779 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
780 seq_puts(m, "\t AXI low power mode\n");
781
782 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
783 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
784 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
785 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
786
787 seq_puts(m, "\tMC\n");
788 seq_printf(m, "\t read0: 0x%08x\n", read0);
789 seq_printf(m, "\t read1: 0x%08x\n", read1);
790 seq_printf(m, "\t write: 0x%08x\n", write);
791 }
792
793 seq_puts(m, "\tDMA ");
794
795 if (debug.address[0] == debug.address[1] &&
796 debug.state[0] == debug.state[1]) {
797 seq_puts(m, "seems to be stuck\n");
798 } else if (debug.address[0] == debug.address[1]) {
799 seq_puts(m, "adress is constant\n");
800 } else {
801 seq_puts(m, "is runing\n");
802 }
803
804 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
805 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
806 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
807 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
808 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
809 dma_lo, dma_hi);
810
811 ret = 0;
812
813 pm_runtime_mark_last_busy(gpu->dev);
814 pm_runtime_put_autosuspend(gpu->dev);
815
816 return ret;
817}
818#endif
819
820/*
821 * Power Management:
822 */
823static int enable_clk(struct etnaviv_gpu *gpu)
824{
825 if (gpu->clk_core)
826 clk_prepare_enable(gpu->clk_core);
827 if (gpu->clk_shader)
828 clk_prepare_enable(gpu->clk_shader);
829
830 return 0;
831}
832
833static int disable_clk(struct etnaviv_gpu *gpu)
834{
835 if (gpu->clk_core)
836 clk_disable_unprepare(gpu->clk_core);
837 if (gpu->clk_shader)
838 clk_disable_unprepare(gpu->clk_shader);
839
840 return 0;
841}
842
843static int enable_axi(struct etnaviv_gpu *gpu)
844{
845 if (gpu->clk_bus)
846 clk_prepare_enable(gpu->clk_bus);
847
848 return 0;
849}
850
851static int disable_axi(struct etnaviv_gpu *gpu)
852{
853 if (gpu->clk_bus)
854 clk_disable_unprepare(gpu->clk_bus);
855
856 return 0;
857}
858
859/*
860 * Hangcheck detection for locked gpu:
861 */
862static void recover_worker(struct work_struct *work)
863{
864 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
865 recover_work);
866 unsigned long flags;
867 unsigned int i;
868
869 dev_err(gpu->dev, "hangcheck recover!\n");
870
871 if (pm_runtime_get_sync(gpu->dev) < 0)
872 return;
873
874 mutex_lock(&gpu->lock);
875
876 /* Only catch the first event, or when manually re-armed */
877 if (etnaviv_dump_core) {
878 etnaviv_core_dump(gpu);
879 etnaviv_dump_core = false;
880 }
881
882 etnaviv_hw_reset(gpu);
883
884 /* complete all events, the GPU won't do it after the reset */
885 spin_lock_irqsave(&gpu->event_spinlock, flags);
886 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
887 if (!gpu->event[i].used)
888 continue;
889 fence_signal(gpu->event[i].fence);
890 gpu->event[i].fence = NULL;
891 gpu->event[i].used = false;
892 complete(&gpu->event_free);
893 }
894 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
895 gpu->completed_fence = gpu->active_fence;
896
897 etnaviv_gpu_hw_init(gpu);
898 gpu->switch_context = true;
899 gpu->exec_state = -1;
900
901 mutex_unlock(&gpu->lock);
902 pm_runtime_mark_last_busy(gpu->dev);
903 pm_runtime_put_autosuspend(gpu->dev);
904
905 /* Retire the buffer objects in a work */
906 etnaviv_queue_work(gpu->drm, &gpu->retire_work);
907}
908
909static void hangcheck_timer_reset(struct etnaviv_gpu *gpu)
910{
911 DBG("%s", dev_name(gpu->dev));
912 mod_timer(&gpu->hangcheck_timer,
913 round_jiffies_up(jiffies + DRM_ETNAVIV_HANGCHECK_JIFFIES));
914}
915
916static void hangcheck_handler(unsigned long data)
917{
918 struct etnaviv_gpu *gpu = (struct etnaviv_gpu *)data;
919 u32 fence = gpu->completed_fence;
920 bool progress = false;
921
922 if (fence != gpu->hangcheck_fence) {
923 gpu->hangcheck_fence = fence;
924 progress = true;
925 }
926
927 if (!progress) {
928 u32 dma_addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
929 int change = dma_addr - gpu->hangcheck_dma_addr;
930
931 if (change < 0 || change > 16) {
932 gpu->hangcheck_dma_addr = dma_addr;
933 progress = true;
934 }
935 }
936
937 if (!progress && fence_after(gpu->active_fence, fence)) {
938 dev_err(gpu->dev, "hangcheck detected gpu lockup!\n");
939 dev_err(gpu->dev, " completed fence: %u\n", fence);
940 dev_err(gpu->dev, " active fence: %u\n",
941 gpu->active_fence);
942 etnaviv_queue_work(gpu->drm, &gpu->recover_work);
943 }
944
945 /* if still more pending work, reset the hangcheck timer: */
946 if (fence_after(gpu->active_fence, gpu->hangcheck_fence))
947 hangcheck_timer_reset(gpu);
948}
949
950static void hangcheck_disable(struct etnaviv_gpu *gpu)
951{
952 del_timer_sync(&gpu->hangcheck_timer);
953 cancel_work_sync(&gpu->recover_work);
954}
955
956/* fence object management */
957struct etnaviv_fence {
958 struct etnaviv_gpu *gpu;
959 struct fence base;
960};
961
962static inline struct etnaviv_fence *to_etnaviv_fence(struct fence *fence)
963{
964 return container_of(fence, struct etnaviv_fence, base);
965}
966
967static const char *etnaviv_fence_get_driver_name(struct fence *fence)
968{
969 return "etnaviv";
970}
971
972static const char *etnaviv_fence_get_timeline_name(struct fence *fence)
973{
974 struct etnaviv_fence *f = to_etnaviv_fence(fence);
975
976 return dev_name(f->gpu->dev);
977}
978
979static bool etnaviv_fence_enable_signaling(struct fence *fence)
980{
981 return true;
982}
983
984static bool etnaviv_fence_signaled(struct fence *fence)
985{
986 struct etnaviv_fence *f = to_etnaviv_fence(fence);
987
988 return fence_completed(f->gpu, f->base.seqno);
989}
990
991static void etnaviv_fence_release(struct fence *fence)
992{
993 struct etnaviv_fence *f = to_etnaviv_fence(fence);
994
995 kfree_rcu(f, base.rcu);
996}
997
998static const struct fence_ops etnaviv_fence_ops = {
999 .get_driver_name = etnaviv_fence_get_driver_name,
1000 .get_timeline_name = etnaviv_fence_get_timeline_name,
1001 .enable_signaling = etnaviv_fence_enable_signaling,
1002 .signaled = etnaviv_fence_signaled,
1003 .wait = fence_default_wait,
1004 .release = etnaviv_fence_release,
1005};
1006
1007static struct fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1008{
1009 struct etnaviv_fence *f;
1010
1011 f = kzalloc(sizeof(*f), GFP_KERNEL);
1012 if (!f)
1013 return NULL;
1014
1015 f->gpu = gpu;
1016
1017 fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1018 gpu->fence_context, ++gpu->next_fence);
1019
1020 return &f->base;
1021}
1022
1023int etnaviv_gpu_fence_sync_obj(struct etnaviv_gem_object *etnaviv_obj,
1024 unsigned int context, bool exclusive)
1025{
1026 struct reservation_object *robj = etnaviv_obj->resv;
1027 struct reservation_object_list *fobj;
1028 struct fence *fence;
1029 int i, ret;
1030
1031 if (!exclusive) {
1032 ret = reservation_object_reserve_shared(robj);
1033 if (ret)
1034 return ret;
1035 }
1036
1037 /*
1038 * If we have any shared fences, then the exclusive fence
1039 * should be ignored as it will already have been signalled.
1040 */
1041 fobj = reservation_object_get_list(robj);
1042 if (!fobj || fobj->shared_count == 0) {
1043 /* Wait on any existing exclusive fence which isn't our own */
1044 fence = reservation_object_get_excl(robj);
1045 if (fence && fence->context != context) {
1046 ret = fence_wait(fence, true);
1047 if (ret)
1048 return ret;
1049 }
1050 }
1051
1052 if (!exclusive || !fobj)
1053 return 0;
1054
1055 for (i = 0; i < fobj->shared_count; i++) {
1056 fence = rcu_dereference_protected(fobj->shared[i],
1057 reservation_object_held(robj));
1058 if (fence->context != context) {
1059 ret = fence_wait(fence, true);
1060 if (ret)
1061 return ret;
1062 }
1063 }
1064
1065 return 0;
1066}
1067
1068/*
1069 * event management:
1070 */
1071
1072static unsigned int event_alloc(struct etnaviv_gpu *gpu)
1073{
1074 unsigned long ret, flags;
1075 unsigned int i, event = ~0U;
1076
1077 ret = wait_for_completion_timeout(&gpu->event_free,
1078 msecs_to_jiffies(10 * 10000));
1079 if (!ret)
1080 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1081
1082 spin_lock_irqsave(&gpu->event_spinlock, flags);
1083
1084 /* find first free event */
1085 for (i = 0; i < ARRAY_SIZE(gpu->event); i++) {
1086 if (gpu->event[i].used == false) {
1087 gpu->event[i].used = true;
1088 event = i;
1089 break;
1090 }
1091 }
1092
1093 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1094
1095 return event;
1096}
1097
1098static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1099{
1100 unsigned long flags;
1101
1102 spin_lock_irqsave(&gpu->event_spinlock, flags);
1103
1104 if (gpu->event[event].used == false) {
1105 dev_warn(gpu->dev, "event %u is already marked as free",
1106 event);
1107 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1108 } else {
1109 gpu->event[event].used = false;
1110 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1111
1112 complete(&gpu->event_free);
1113 }
1114}
1115
1116/*
1117 * Cmdstream submission/retirement:
1118 */
1119
1120struct etnaviv_cmdbuf *etnaviv_gpu_cmdbuf_new(struct etnaviv_gpu *gpu, u32 size,
1121 size_t nr_bos)
1122{
1123 struct etnaviv_cmdbuf *cmdbuf;
1124 size_t sz = size_vstruct(nr_bos, sizeof(cmdbuf->bo_map[0]),
1125 sizeof(*cmdbuf));
1126
1127 cmdbuf = kzalloc(sz, GFP_KERNEL);
1128 if (!cmdbuf)
1129 return NULL;
1130
1131 cmdbuf->vaddr = dma_alloc_wc(gpu->dev, size, &cmdbuf->paddr,
1132 GFP_KERNEL);
1133 if (!cmdbuf->vaddr) {
1134 kfree(cmdbuf);
1135 return NULL;
1136 }
1137
1138 cmdbuf->gpu = gpu;
1139 cmdbuf->size = size;
1140
1141 return cmdbuf;
1142}
1143
1144void etnaviv_gpu_cmdbuf_free(struct etnaviv_cmdbuf *cmdbuf)
1145{
1146 dma_free_wc(cmdbuf->gpu->dev, cmdbuf->size, cmdbuf->vaddr,
1147 cmdbuf->paddr);
1148 kfree(cmdbuf);
1149}
1150
1151static void retire_worker(struct work_struct *work)
1152{
1153 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1154 retire_work);
1155 u32 fence = gpu->completed_fence;
1156 struct etnaviv_cmdbuf *cmdbuf, *tmp;
1157 unsigned int i;
1158
1159 mutex_lock(&gpu->lock);
1160 list_for_each_entry_safe(cmdbuf, tmp, &gpu->active_cmd_list, node) {
1161 if (!fence_is_signaled(cmdbuf->fence))
1162 break;
1163
1164 list_del(&cmdbuf->node);
1165 fence_put(cmdbuf->fence);
1166
1167 for (i = 0; i < cmdbuf->nr_bos; i++) {
1168 struct etnaviv_vram_mapping *mapping = cmdbuf->bo_map[i];
1169 struct etnaviv_gem_object *etnaviv_obj = mapping->object;
1170
1171 atomic_dec(&etnaviv_obj->gpu_active);
1172 /* drop the refcount taken in etnaviv_gpu_submit */
1173 etnaviv_gem_mapping_unreference(mapping);
1174 }
1175
1176 etnaviv_gpu_cmdbuf_free(cmdbuf);
1177 /*
1178 * We need to balance the runtime PM count caused by
1179 * each submission. Upon submission, we increment
1180 * the runtime PM counter, and allocate one event.
1181 * So here, we put the runtime PM count for each
1182 * completed event.
1183 */
1184 pm_runtime_put_autosuspend(gpu->dev);
1185 }
1186
1187 gpu->retired_fence = fence;
1188
1189 mutex_unlock(&gpu->lock);
1190
1191 wake_up_all(&gpu->fence_event);
1192}
1193
1194int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1195 u32 fence, struct timespec *timeout)
1196{
1197 int ret;
1198
1199 if (fence_after(fence, gpu->next_fence)) {
1200 DRM_ERROR("waiting on invalid fence: %u (of %u)\n",
1201 fence, gpu->next_fence);
1202 return -EINVAL;
1203 }
1204
1205 if (!timeout) {
1206 /* No timeout was requested: just test for completion */
1207 ret = fence_completed(gpu, fence) ? 0 : -EBUSY;
1208 } else {
1209 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1210
1211 ret = wait_event_interruptible_timeout(gpu->fence_event,
1212 fence_completed(gpu, fence),
1213 remaining);
1214 if (ret == 0) {
1215 DBG("timeout waiting for fence: %u (retired: %u completed: %u)",
1216 fence, gpu->retired_fence,
1217 gpu->completed_fence);
1218 ret = -ETIMEDOUT;
1219 } else if (ret != -ERESTARTSYS) {
1220 ret = 0;
1221 }
1222 }
1223
1224 return ret;
1225}
1226
1227/*
1228 * Wait for an object to become inactive. This, on it's own, is not race
1229 * free: the object is moved by the retire worker off the active list, and
1230 * then the iova is put. Moreover, the object could be re-submitted just
1231 * after we notice that it's become inactive.
1232 *
1233 * Although the retirement happens under the gpu lock, we don't want to hold
1234 * that lock in this function while waiting.
1235 */
1236int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1237 struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1238{
1239 unsigned long remaining;
1240 long ret;
1241
1242 if (!timeout)
1243 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1244
1245 remaining = etnaviv_timeout_to_jiffies(timeout);
1246
1247 ret = wait_event_interruptible_timeout(gpu->fence_event,
1248 !is_active(etnaviv_obj),
1249 remaining);
1250 if (ret > 0) {
1251 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
1252
1253 /* Synchronise with the retire worker */
1254 flush_workqueue(priv->wq);
1255 return 0;
1256 } else if (ret == -ERESTARTSYS) {
1257 return -ERESTARTSYS;
1258 } else {
1259 return -ETIMEDOUT;
1260 }
1261}
1262
1263int etnaviv_gpu_pm_get_sync(struct etnaviv_gpu *gpu)
1264{
1265 return pm_runtime_get_sync(gpu->dev);
1266}
1267
1268void etnaviv_gpu_pm_put(struct etnaviv_gpu *gpu)
1269{
1270 pm_runtime_mark_last_busy(gpu->dev);
1271 pm_runtime_put_autosuspend(gpu->dev);
1272}
1273
1274/* add bo's to gpu's ring, and kick gpu: */
1275int etnaviv_gpu_submit(struct etnaviv_gpu *gpu,
1276 struct etnaviv_gem_submit *submit, struct etnaviv_cmdbuf *cmdbuf)
1277{
1278 struct fence *fence;
1279 unsigned int event, i;
1280 int ret;
1281
1282 ret = etnaviv_gpu_pm_get_sync(gpu);
1283 if (ret < 0)
1284 return ret;
1285
1286 mutex_lock(&gpu->lock);
1287
1288 /*
1289 * TODO
1290 *
1291 * - flush
1292 * - data endian
1293 * - prefetch
1294 *
1295 */
1296
1297 event = event_alloc(gpu);
1298 if (unlikely(event == ~0U)) {
1299 DRM_ERROR("no free event\n");
1300 ret = -EBUSY;
1301 goto out_unlock;
1302 }
1303
1304 fence = etnaviv_gpu_fence_alloc(gpu);
1305 if (!fence) {
1306 event_free(gpu, event);
1307 ret = -ENOMEM;
1308 goto out_unlock;
1309 }
1310
1311 gpu->event[event].fence = fence;
1312 submit->fence = fence->seqno;
1313 gpu->active_fence = submit->fence;
1314
1315 if (gpu->lastctx != cmdbuf->ctx) {
1316 gpu->mmu->need_flush = true;
1317 gpu->switch_context = true;
1318 gpu->lastctx = cmdbuf->ctx;
1319 }
1320
1321 etnaviv_buffer_queue(gpu, event, cmdbuf);
1322
1323 cmdbuf->fence = fence;
1324 list_add_tail(&cmdbuf->node, &gpu->active_cmd_list);
1325
1326 /* We're committed to adding this command buffer, hold a PM reference */
1327 pm_runtime_get_noresume(gpu->dev);
1328
1329 for (i = 0; i < submit->nr_bos; i++) {
1330 struct etnaviv_gem_object *etnaviv_obj = submit->bos[i].obj;
1331
1332 /* Each cmdbuf takes a refcount on the mapping */
1333 etnaviv_gem_mapping_reference(submit->bos[i].mapping);
1334 cmdbuf->bo_map[i] = submit->bos[i].mapping;
1335 atomic_inc(&etnaviv_obj->gpu_active);
1336
1337 if (submit->bos[i].flags & ETNA_SUBMIT_BO_WRITE)
1338 reservation_object_add_excl_fence(etnaviv_obj->resv,
1339 fence);
1340 else
1341 reservation_object_add_shared_fence(etnaviv_obj->resv,
1342 fence);
1343 }
1344 cmdbuf->nr_bos = submit->nr_bos;
1345 hangcheck_timer_reset(gpu);
1346 ret = 0;
1347
1348out_unlock:
1349 mutex_unlock(&gpu->lock);
1350
1351 etnaviv_gpu_pm_put(gpu);
1352
1353 return ret;
1354}
1355
1356/*
1357 * Init/Cleanup:
1358 */
1359static irqreturn_t irq_handler(int irq, void *data)
1360{
1361 struct etnaviv_gpu *gpu = data;
1362 irqreturn_t ret = IRQ_NONE;
1363
1364 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1365
1366 if (intr != 0) {
1367 int event;
1368
1369 pm_runtime_mark_last_busy(gpu->dev);
1370
1371 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1372
1373 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1374 dev_err(gpu->dev, "AXI bus error\n");
1375 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1376 }
1377
1378 while ((event = ffs(intr)) != 0) {
1379 struct fence *fence;
1380
1381 event -= 1;
1382
1383 intr &= ~(1 << event);
1384
1385 dev_dbg(gpu->dev, "event %u\n", event);
1386
1387 fence = gpu->event[event].fence;
1388 gpu->event[event].fence = NULL;
1389 fence_signal(fence);
1390
1391 /*
1392 * Events can be processed out of order. Eg,
1393 * - allocate and queue event 0
1394 * - allocate event 1
1395 * - event 0 completes, we process it
1396 * - allocate and queue event 0
1397 * - event 1 and event 0 complete
1398 * we can end up processing event 0 first, then 1.
1399 */
1400 if (fence_after(fence->seqno, gpu->completed_fence))
1401 gpu->completed_fence = fence->seqno;
1402
1403 event_free(gpu, event);
1404 }
1405
1406 /* Retire the buffer objects in a work */
1407 etnaviv_queue_work(gpu->drm, &gpu->retire_work);
1408
1409 ret = IRQ_HANDLED;
1410 }
1411
1412 return ret;
1413}
1414
1415static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1416{
1417 int ret;
1418
1419 ret = enable_clk(gpu);
1420 if (ret)
1421 return ret;
1422
1423 ret = enable_axi(gpu);
1424 if (ret) {
1425 disable_clk(gpu);
1426 return ret;
1427 }
1428
1429 return 0;
1430}
1431
1432static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1433{
1434 int ret;
1435
1436 ret = disable_axi(gpu);
1437 if (ret)
1438 return ret;
1439
1440 ret = disable_clk(gpu);
1441 if (ret)
1442 return ret;
1443
1444 return 0;
1445}
1446
1447static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1448{
1449 if (gpu->buffer) {
1450 unsigned long timeout;
1451
1452 /* Replace the last WAIT with END */
1453 etnaviv_buffer_end(gpu);
1454
1455 /*
1456 * We know that only the FE is busy here, this should
1457 * happen quickly (as the WAIT is only 200 cycles). If
1458 * we fail, just warn and continue.
1459 */
1460 timeout = jiffies + msecs_to_jiffies(100);
1461 do {
1462 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1463
1464 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1465 break;
1466
1467 if (time_is_before_jiffies(timeout)) {
1468 dev_warn(gpu->dev,
1469 "timed out waiting for idle: idle=0x%x\n",
1470 idle);
1471 break;
1472 }
1473
1474 udelay(5);
1475 } while (1);
1476 }
1477
1478 return etnaviv_gpu_clk_disable(gpu);
1479}
1480
1481#ifdef CONFIG_PM
1482static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1483{
1484 u32 clock;
1485 int ret;
1486
1487 ret = mutex_lock_killable(&gpu->lock);
1488 if (ret)
1489 return ret;
1490
1491 clock = VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS |
1492 VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(0x40);
1493
1494 etnaviv_gpu_load_clock(gpu, clock);
1495 etnaviv_gpu_hw_init(gpu);
1496
1497 gpu->switch_context = true;
1498 gpu->exec_state = -1;
1499
1500 mutex_unlock(&gpu->lock);
1501
1502 return 0;
1503}
1504#endif
1505
1506static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1507 void *data)
1508{
1509 struct drm_device *drm = data;
1510 struct etnaviv_drm_private *priv = drm->dev_private;
1511 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1512 int ret;
1513
1514#ifdef CONFIG_PM
1515 ret = pm_runtime_get_sync(gpu->dev);
1516#else
1517 ret = etnaviv_gpu_clk_enable(gpu);
1518#endif
1519 if (ret < 0)
1520 return ret;
1521
1522 gpu->drm = drm;
1523 gpu->fence_context = fence_context_alloc(1);
1524 spin_lock_init(&gpu->fence_spinlock);
1525
1526 INIT_LIST_HEAD(&gpu->active_cmd_list);
1527 INIT_WORK(&gpu->retire_work, retire_worker);
1528 INIT_WORK(&gpu->recover_work, recover_worker);
1529 init_waitqueue_head(&gpu->fence_event);
1530
1531 setup_timer(&gpu->hangcheck_timer, hangcheck_handler,
1532 (unsigned long)gpu);
1533
1534 priv->gpu[priv->num_gpus++] = gpu;
1535
1536 pm_runtime_mark_last_busy(gpu->dev);
1537 pm_runtime_put_autosuspend(gpu->dev);
1538
1539 return 0;
1540}
1541
1542static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1543 void *data)
1544{
1545 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1546
1547 DBG("%s", dev_name(gpu->dev));
1548
1549 hangcheck_disable(gpu);
1550
1551#ifdef CONFIG_PM
1552 pm_runtime_get_sync(gpu->dev);
1553 pm_runtime_put_sync_suspend(gpu->dev);
1554#else
1555 etnaviv_gpu_hw_suspend(gpu);
1556#endif
1557
1558 if (gpu->buffer) {
1559 etnaviv_gpu_cmdbuf_free(gpu->buffer);
1560 gpu->buffer = NULL;
1561 }
1562
1563 if (gpu->mmu) {
1564 etnaviv_iommu_destroy(gpu->mmu);
1565 gpu->mmu = NULL;
1566 }
1567
1568 gpu->drm = NULL;
1569}
1570
1571static const struct component_ops gpu_ops = {
1572 .bind = etnaviv_gpu_bind,
1573 .unbind = etnaviv_gpu_unbind,
1574};
1575
1576static const struct of_device_id etnaviv_gpu_match[] = {
1577 {
1578 .compatible = "vivante,gc"
1579 },
1580 { /* sentinel */ }
1581};
1582
1583static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1584{
1585 struct device *dev = &pdev->dev;
1586 struct etnaviv_gpu *gpu;
1587 int err = 0;
1588
1589 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1590 if (!gpu)
1591 return -ENOMEM;
1592
1593 gpu->dev = &pdev->dev;
1594 mutex_init(&gpu->lock);
1595
1596 /* Map registers: */
1597 gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1598 if (IS_ERR(gpu->mmio))
1599 return PTR_ERR(gpu->mmio);
1600
1601 /* Get Interrupt: */
1602 gpu->irq = platform_get_irq(pdev, 0);
1603 if (gpu->irq < 0) {
1604 err = gpu->irq;
1605 dev_err(dev, "failed to get irq: %d\n", err);
1606 goto fail;
1607 }
1608
1609 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1610 dev_name(gpu->dev), gpu);
1611 if (err) {
1612 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1613 goto fail;
1614 }
1615
1616 /* Get Clocks: */
1617 gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1618 DBG("clk_bus: %p", gpu->clk_bus);
1619 if (IS_ERR(gpu->clk_bus))
1620 gpu->clk_bus = NULL;
1621
1622 gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1623 DBG("clk_core: %p", gpu->clk_core);
1624 if (IS_ERR(gpu->clk_core))
1625 gpu->clk_core = NULL;
1626
1627 gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1628 DBG("clk_shader: %p", gpu->clk_shader);
1629 if (IS_ERR(gpu->clk_shader))
1630 gpu->clk_shader = NULL;
1631
1632 /* TODO: figure out max mapped size */
1633 dev_set_drvdata(dev, gpu);
1634
1635 /*
1636 * We treat the device as initially suspended. The runtime PM
1637 * autosuspend delay is rather arbitary: no measurements have
1638 * yet been performed to determine an appropriate value.
1639 */
1640 pm_runtime_use_autosuspend(gpu->dev);
1641 pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1642 pm_runtime_enable(gpu->dev);
1643
1644 err = component_add(&pdev->dev, &gpu_ops);
1645 if (err < 0) {
1646 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1647 goto fail;
1648 }
1649
1650 return 0;
1651
1652fail:
1653 return err;
1654}
1655
1656static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1657{
1658 component_del(&pdev->dev, &gpu_ops);
1659 pm_runtime_disable(&pdev->dev);
1660 return 0;
1661}
1662
1663#ifdef CONFIG_PM
1664static int etnaviv_gpu_rpm_suspend(struct device *dev)
1665{
1666 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1667 u32 idle, mask;
1668
1669 /* If we have outstanding fences, we're not idle */
1670 if (gpu->completed_fence != gpu->active_fence)
1671 return -EBUSY;
1672
1673 /* Check whether the hardware (except FE) is idle */
1674 mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1675 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1676 if (idle != mask)
1677 return -EBUSY;
1678
1679 return etnaviv_gpu_hw_suspend(gpu);
1680}
1681
1682static int etnaviv_gpu_rpm_resume(struct device *dev)
1683{
1684 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1685 int ret;
1686
1687 ret = etnaviv_gpu_clk_enable(gpu);
1688 if (ret)
1689 return ret;
1690
1691 /* Re-initialise the basic hardware state */
1692 if (gpu->drm && gpu->buffer) {
1693 ret = etnaviv_gpu_hw_resume(gpu);
1694 if (ret) {
1695 etnaviv_gpu_clk_disable(gpu);
1696 return ret;
1697 }
1698 }
1699
1700 return 0;
1701}
1702#endif
1703
1704static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1705 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1706 NULL)
1707};
1708
1709struct platform_driver etnaviv_gpu_driver = {
1710 .driver = {
1711 .name = "etnaviv-gpu",
1712 .owner = THIS_MODULE,
1713 .pm = &etnaviv_gpu_pm_ops,
1714 .of_match_table = etnaviv_gpu_match,
1715 },
1716 .probe = etnaviv_gpu_platform_probe,
1717 .remove = etnaviv_gpu_platform_remove,
1718 .id_table = gpu_ids,
1719};
1/*
2 * Copyright (C) 2015 Etnaviv Project
3 *
4 * This program is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 as published by
6 * the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
11 * more details.
12 *
13 * You should have received a copy of the GNU General Public License along with
14 * this program. If not, see <http://www.gnu.org/licenses/>.
15 */
16
17#include <linux/component.h>
18#include <linux/dma-fence.h>
19#include <linux/moduleparam.h>
20#include <linux/of_device.h>
21#include <linux/thermal.h>
22
23#include "etnaviv_cmdbuf.h"
24#include "etnaviv_dump.h"
25#include "etnaviv_gpu.h"
26#include "etnaviv_gem.h"
27#include "etnaviv_mmu.h"
28#include "etnaviv_perfmon.h"
29#include "etnaviv_sched.h"
30#include "common.xml.h"
31#include "state.xml.h"
32#include "state_hi.xml.h"
33#include "cmdstream.xml.h"
34
35#ifndef PHYS_OFFSET
36#define PHYS_OFFSET 0
37#endif
38
39static const struct platform_device_id gpu_ids[] = {
40 { .name = "etnaviv-gpu,2d" },
41 { },
42};
43
44/*
45 * Driver functions:
46 */
47
48int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
49{
50 switch (param) {
51 case ETNAVIV_PARAM_GPU_MODEL:
52 *value = gpu->identity.model;
53 break;
54
55 case ETNAVIV_PARAM_GPU_REVISION:
56 *value = gpu->identity.revision;
57 break;
58
59 case ETNAVIV_PARAM_GPU_FEATURES_0:
60 *value = gpu->identity.features;
61 break;
62
63 case ETNAVIV_PARAM_GPU_FEATURES_1:
64 *value = gpu->identity.minor_features0;
65 break;
66
67 case ETNAVIV_PARAM_GPU_FEATURES_2:
68 *value = gpu->identity.minor_features1;
69 break;
70
71 case ETNAVIV_PARAM_GPU_FEATURES_3:
72 *value = gpu->identity.minor_features2;
73 break;
74
75 case ETNAVIV_PARAM_GPU_FEATURES_4:
76 *value = gpu->identity.minor_features3;
77 break;
78
79 case ETNAVIV_PARAM_GPU_FEATURES_5:
80 *value = gpu->identity.minor_features4;
81 break;
82
83 case ETNAVIV_PARAM_GPU_FEATURES_6:
84 *value = gpu->identity.minor_features5;
85 break;
86
87 case ETNAVIV_PARAM_GPU_FEATURES_7:
88 *value = gpu->identity.minor_features6;
89 break;
90
91 case ETNAVIV_PARAM_GPU_FEATURES_8:
92 *value = gpu->identity.minor_features7;
93 break;
94
95 case ETNAVIV_PARAM_GPU_FEATURES_9:
96 *value = gpu->identity.minor_features8;
97 break;
98
99 case ETNAVIV_PARAM_GPU_FEATURES_10:
100 *value = gpu->identity.minor_features9;
101 break;
102
103 case ETNAVIV_PARAM_GPU_FEATURES_11:
104 *value = gpu->identity.minor_features10;
105 break;
106
107 case ETNAVIV_PARAM_GPU_FEATURES_12:
108 *value = gpu->identity.minor_features11;
109 break;
110
111 case ETNAVIV_PARAM_GPU_STREAM_COUNT:
112 *value = gpu->identity.stream_count;
113 break;
114
115 case ETNAVIV_PARAM_GPU_REGISTER_MAX:
116 *value = gpu->identity.register_max;
117 break;
118
119 case ETNAVIV_PARAM_GPU_THREAD_COUNT:
120 *value = gpu->identity.thread_count;
121 break;
122
123 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
124 *value = gpu->identity.vertex_cache_size;
125 break;
126
127 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
128 *value = gpu->identity.shader_core_count;
129 break;
130
131 case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
132 *value = gpu->identity.pixel_pipes;
133 break;
134
135 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
136 *value = gpu->identity.vertex_output_buffer_size;
137 break;
138
139 case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
140 *value = gpu->identity.buffer_size;
141 break;
142
143 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
144 *value = gpu->identity.instruction_count;
145 break;
146
147 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
148 *value = gpu->identity.num_constants;
149 break;
150
151 case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
152 *value = gpu->identity.varyings_count;
153 break;
154
155 default:
156 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
157 return -EINVAL;
158 }
159
160 return 0;
161}
162
163
164#define etnaviv_is_model_rev(gpu, mod, rev) \
165 ((gpu)->identity.model == chipModel_##mod && \
166 (gpu)->identity.revision == rev)
167#define etnaviv_field(val, field) \
168 (((val) & field##__MASK) >> field##__SHIFT)
169
170static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
171{
172 if (gpu->identity.minor_features0 &
173 chipMinorFeatures0_MORE_MINOR_FEATURES) {
174 u32 specs[4];
175 unsigned int streams;
176
177 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
178 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
179 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
180 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
181
182 gpu->identity.stream_count = etnaviv_field(specs[0],
183 VIVS_HI_CHIP_SPECS_STREAM_COUNT);
184 gpu->identity.register_max = etnaviv_field(specs[0],
185 VIVS_HI_CHIP_SPECS_REGISTER_MAX);
186 gpu->identity.thread_count = etnaviv_field(specs[0],
187 VIVS_HI_CHIP_SPECS_THREAD_COUNT);
188 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
189 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
190 gpu->identity.shader_core_count = etnaviv_field(specs[0],
191 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
192 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
193 VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
194 gpu->identity.vertex_output_buffer_size =
195 etnaviv_field(specs[0],
196 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
197
198 gpu->identity.buffer_size = etnaviv_field(specs[1],
199 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
200 gpu->identity.instruction_count = etnaviv_field(specs[1],
201 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
202 gpu->identity.num_constants = etnaviv_field(specs[1],
203 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
204
205 gpu->identity.varyings_count = etnaviv_field(specs[2],
206 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
207
208 /* This overrides the value from older register if non-zero */
209 streams = etnaviv_field(specs[3],
210 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
211 if (streams)
212 gpu->identity.stream_count = streams;
213 }
214
215 /* Fill in the stream count if not specified */
216 if (gpu->identity.stream_count == 0) {
217 if (gpu->identity.model >= 0x1000)
218 gpu->identity.stream_count = 4;
219 else
220 gpu->identity.stream_count = 1;
221 }
222
223 /* Convert the register max value */
224 if (gpu->identity.register_max)
225 gpu->identity.register_max = 1 << gpu->identity.register_max;
226 else if (gpu->identity.model == chipModel_GC400)
227 gpu->identity.register_max = 32;
228 else
229 gpu->identity.register_max = 64;
230
231 /* Convert thread count */
232 if (gpu->identity.thread_count)
233 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
234 else if (gpu->identity.model == chipModel_GC400)
235 gpu->identity.thread_count = 64;
236 else if (gpu->identity.model == chipModel_GC500 ||
237 gpu->identity.model == chipModel_GC530)
238 gpu->identity.thread_count = 128;
239 else
240 gpu->identity.thread_count = 256;
241
242 if (gpu->identity.vertex_cache_size == 0)
243 gpu->identity.vertex_cache_size = 8;
244
245 if (gpu->identity.shader_core_count == 0) {
246 if (gpu->identity.model >= 0x1000)
247 gpu->identity.shader_core_count = 2;
248 else
249 gpu->identity.shader_core_count = 1;
250 }
251
252 if (gpu->identity.pixel_pipes == 0)
253 gpu->identity.pixel_pipes = 1;
254
255 /* Convert virtex buffer size */
256 if (gpu->identity.vertex_output_buffer_size) {
257 gpu->identity.vertex_output_buffer_size =
258 1 << gpu->identity.vertex_output_buffer_size;
259 } else if (gpu->identity.model == chipModel_GC400) {
260 if (gpu->identity.revision < 0x4000)
261 gpu->identity.vertex_output_buffer_size = 512;
262 else if (gpu->identity.revision < 0x4200)
263 gpu->identity.vertex_output_buffer_size = 256;
264 else
265 gpu->identity.vertex_output_buffer_size = 128;
266 } else {
267 gpu->identity.vertex_output_buffer_size = 512;
268 }
269
270 switch (gpu->identity.instruction_count) {
271 case 0:
272 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
273 gpu->identity.model == chipModel_GC880)
274 gpu->identity.instruction_count = 512;
275 else
276 gpu->identity.instruction_count = 256;
277 break;
278
279 case 1:
280 gpu->identity.instruction_count = 1024;
281 break;
282
283 case 2:
284 gpu->identity.instruction_count = 2048;
285 break;
286
287 default:
288 gpu->identity.instruction_count = 256;
289 break;
290 }
291
292 if (gpu->identity.num_constants == 0)
293 gpu->identity.num_constants = 168;
294
295 if (gpu->identity.varyings_count == 0) {
296 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
297 gpu->identity.varyings_count = 12;
298 else
299 gpu->identity.varyings_count = 8;
300 }
301
302 /*
303 * For some cores, two varyings are consumed for position, so the
304 * maximum varying count needs to be reduced by one.
305 */
306 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
307 etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
308 etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
309 etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
310 etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
311 etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
312 etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
313 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
314 etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
315 etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
316 etnaviv_is_model_rev(gpu, GC880, 0x5106))
317 gpu->identity.varyings_count -= 1;
318}
319
320static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
321{
322 u32 chipIdentity;
323
324 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
325
326 /* Special case for older graphic cores. */
327 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
328 gpu->identity.model = chipModel_GC500;
329 gpu->identity.revision = etnaviv_field(chipIdentity,
330 VIVS_HI_CHIP_IDENTITY_REVISION);
331 } else {
332
333 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
334 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
335
336 /*
337 * !!!! HACK ALERT !!!!
338 * Because people change device IDs without letting software
339 * know about it - here is the hack to make it all look the
340 * same. Only for GC400 family.
341 */
342 if ((gpu->identity.model & 0xff00) == 0x0400 &&
343 gpu->identity.model != chipModel_GC420) {
344 gpu->identity.model = gpu->identity.model & 0x0400;
345 }
346
347 /* Another special case */
348 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
349 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
350 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
351
352 if (chipDate == 0x20080814 && chipTime == 0x12051100) {
353 /*
354 * This IP has an ECO; put the correct
355 * revision in it.
356 */
357 gpu->identity.revision = 0x1051;
358 }
359 }
360
361 /*
362 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
363 * reality it's just a re-branded GC3000. We can identify this
364 * core by the upper half of the revision register being all 1.
365 * Fix model/rev here, so all other places can refer to this
366 * core by its real identity.
367 */
368 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
369 gpu->identity.model = chipModel_GC3000;
370 gpu->identity.revision &= 0xffff;
371 }
372 }
373
374 dev_info(gpu->dev, "model: GC%x, revision: %x\n",
375 gpu->identity.model, gpu->identity.revision);
376
377 /*
378 * If there is a match in the HWDB, we aren't interested in the
379 * remaining register values, as they might be wrong.
380 */
381 if (etnaviv_fill_identity_from_hwdb(gpu))
382 return;
383
384 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
385
386 /* Disable fast clear on GC700. */
387 if (gpu->identity.model == chipModel_GC700)
388 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
389
390 if ((gpu->identity.model == chipModel_GC500 &&
391 gpu->identity.revision < 2) ||
392 (gpu->identity.model == chipModel_GC300 &&
393 gpu->identity.revision < 0x2000)) {
394
395 /*
396 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
397 * registers.
398 */
399 gpu->identity.minor_features0 = 0;
400 gpu->identity.minor_features1 = 0;
401 gpu->identity.minor_features2 = 0;
402 gpu->identity.minor_features3 = 0;
403 gpu->identity.minor_features4 = 0;
404 gpu->identity.minor_features5 = 0;
405 } else
406 gpu->identity.minor_features0 =
407 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
408
409 if (gpu->identity.minor_features0 &
410 chipMinorFeatures0_MORE_MINOR_FEATURES) {
411 gpu->identity.minor_features1 =
412 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
413 gpu->identity.minor_features2 =
414 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
415 gpu->identity.minor_features3 =
416 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
417 gpu->identity.minor_features4 =
418 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
419 gpu->identity.minor_features5 =
420 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
421 }
422
423 /* GC600 idle register reports zero bits where modules aren't present */
424 if (gpu->identity.model == chipModel_GC600) {
425 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
426 VIVS_HI_IDLE_STATE_RA |
427 VIVS_HI_IDLE_STATE_SE |
428 VIVS_HI_IDLE_STATE_PA |
429 VIVS_HI_IDLE_STATE_SH |
430 VIVS_HI_IDLE_STATE_PE |
431 VIVS_HI_IDLE_STATE_DE |
432 VIVS_HI_IDLE_STATE_FE;
433 } else {
434 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
435 }
436
437 etnaviv_hw_specs(gpu);
438}
439
440static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
441{
442 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
443 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
444 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
445}
446
447static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
448{
449 if (gpu->identity.minor_features2 &
450 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
451 clk_set_rate(gpu->clk_core,
452 gpu->base_rate_core >> gpu->freq_scale);
453 clk_set_rate(gpu->clk_shader,
454 gpu->base_rate_shader >> gpu->freq_scale);
455 } else {
456 unsigned int fscale = 1 << (6 - gpu->freq_scale);
457 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
458
459 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
460 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
461 etnaviv_gpu_load_clock(gpu, clock);
462 }
463}
464
465static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
466{
467 u32 control, idle;
468 unsigned long timeout;
469 bool failed = true;
470
471 /* We hope that the GPU resets in under one second */
472 timeout = jiffies + msecs_to_jiffies(1000);
473
474 while (time_is_after_jiffies(timeout)) {
475 /* enable clock */
476 unsigned int fscale = 1 << (6 - gpu->freq_scale);
477 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
478 etnaviv_gpu_load_clock(gpu, control);
479
480 /* isolate the GPU. */
481 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
482 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
483
484 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
485 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
486 VIVS_MMUv2_AHB_CONTROL_RESET);
487 } else {
488 /* set soft reset. */
489 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
490 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
491 }
492
493 /* wait for reset. */
494 usleep_range(10, 20);
495
496 /* reset soft reset bit. */
497 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
498 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
499
500 /* reset GPU isolation. */
501 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
502 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
503
504 /* read idle register. */
505 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
506
507 /* try reseting again if FE it not idle */
508 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
509 dev_dbg(gpu->dev, "FE is not idle\n");
510 continue;
511 }
512
513 /* read reset register. */
514 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
515
516 /* is the GPU idle? */
517 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
518 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
519 dev_dbg(gpu->dev, "GPU is not idle\n");
520 continue;
521 }
522
523 /* disable debug registers, as they are not normally needed */
524 control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
525 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
526
527 failed = false;
528 break;
529 }
530
531 if (failed) {
532 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
533 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
534
535 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
536 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
537 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
538 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
539
540 return -EBUSY;
541 }
542
543 /* We rely on the GPU running, so program the clock */
544 etnaviv_gpu_update_clock(gpu);
545
546 return 0;
547}
548
549static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
550{
551 u32 pmc, ppc;
552
553 /* enable clock gating */
554 ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
555 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
556
557 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
558 if (gpu->identity.revision == 0x4301 ||
559 gpu->identity.revision == 0x4302)
560 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
561
562 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
563
564 pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
565
566 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
567 if (gpu->identity.model >= chipModel_GC400 &&
568 gpu->identity.model != chipModel_GC420 &&
569 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
570 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
571
572 /*
573 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
574 * present without a bug fix.
575 */
576 if (gpu->identity.revision < 0x5000 &&
577 gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
578 !(gpu->identity.minor_features1 &
579 chipMinorFeatures1_DISABLE_PE_GATING))
580 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
581
582 if (gpu->identity.revision < 0x5422)
583 pmc |= BIT(15); /* Unknown bit */
584
585 /* Disable TX clock gating on affected core revisions. */
586 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
587 etnaviv_is_model_rev(gpu, GC2000, 0x5108))
588 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
589
590 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
591 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
592
593 gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
594}
595
596void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
597{
598 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
599 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
600 VIVS_FE_COMMAND_CONTROL_ENABLE |
601 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
602
603 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
604 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
605 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
606 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
607 }
608}
609
610static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
611{
612 /*
613 * Base value for VIVS_PM_PULSE_EATER register on models where it
614 * cannot be read, extracted from vivante kernel driver.
615 */
616 u32 pulse_eater = 0x01590880;
617
618 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
619 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
620 pulse_eater |= BIT(23);
621
622 }
623
624 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
625 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
626 pulse_eater &= ~BIT(16);
627 pulse_eater |= BIT(17);
628 }
629
630 if ((gpu->identity.revision > 0x5420) &&
631 (gpu->identity.features & chipFeatures_PIPE_3D))
632 {
633 /* Performance fix: disable internal DFS */
634 pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
635 pulse_eater |= BIT(18);
636 }
637
638 gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
639}
640
641static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
642{
643 u16 prefetch;
644
645 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
646 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
647 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
648 u32 mc_memory_debug;
649
650 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
651
652 if (gpu->identity.revision == 0x5007)
653 mc_memory_debug |= 0x0c;
654 else
655 mc_memory_debug |= 0x08;
656
657 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
658 }
659
660 /* enable module-level clock gating */
661 etnaviv_gpu_enable_mlcg(gpu);
662
663 /*
664 * Update GPU AXI cache atttribute to "cacheable, no allocate".
665 * This is necessary to prevent the iMX6 SoC locking up.
666 */
667 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
668 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
669 VIVS_HI_AXI_CONFIG_ARCACHE(2));
670
671 /* GC2000 rev 5108 needs a special bus config */
672 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
673 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
674 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
675 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
676 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
677 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
678 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
679 }
680
681 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
682 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
683 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
684 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
685 }
686
687 /* setup the pulse eater */
688 etnaviv_gpu_setup_pulse_eater(gpu);
689
690 /* setup the MMU */
691 etnaviv_iommu_restore(gpu);
692
693 /* Start command processor */
694 prefetch = etnaviv_buffer_init(gpu);
695
696 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
697 etnaviv_gpu_start_fe(gpu, etnaviv_cmdbuf_get_va(&gpu->buffer),
698 prefetch);
699}
700
701int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
702{
703 int ret, i;
704
705 ret = pm_runtime_get_sync(gpu->dev);
706 if (ret < 0) {
707 dev_err(gpu->dev, "Failed to enable GPU power domain\n");
708 return ret;
709 }
710
711 etnaviv_hw_identify(gpu);
712
713 if (gpu->identity.model == 0) {
714 dev_err(gpu->dev, "Unknown GPU model\n");
715 ret = -ENXIO;
716 goto fail;
717 }
718
719 /* Exclude VG cores with FE2.0 */
720 if (gpu->identity.features & chipFeatures_PIPE_VG &&
721 gpu->identity.features & chipFeatures_FE20) {
722 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
723 ret = -ENXIO;
724 goto fail;
725 }
726
727 /*
728 * Set the GPU linear window to be at the end of the DMA window, where
729 * the CMA area is likely to reside. This ensures that we are able to
730 * map the command buffers while having the linear window overlap as
731 * much RAM as possible, so we can optimize mappings for other buffers.
732 *
733 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
734 * to different views of the memory on the individual engines.
735 */
736 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
737 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
738 u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
739 if (dma_mask < PHYS_OFFSET + SZ_2G)
740 gpu->memory_base = PHYS_OFFSET;
741 else
742 gpu->memory_base = dma_mask - SZ_2G + 1;
743 } else if (PHYS_OFFSET >= SZ_2G) {
744 dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
745 gpu->memory_base = PHYS_OFFSET;
746 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
747 }
748
749 /*
750 * On cores with security features supported, we claim control over the
751 * security states.
752 */
753 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
754 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
755 gpu->sec_mode = ETNA_SEC_KERNEL;
756
757 ret = etnaviv_hw_reset(gpu);
758 if (ret) {
759 dev_err(gpu->dev, "GPU reset failed\n");
760 goto fail;
761 }
762
763 gpu->mmu = etnaviv_iommu_new(gpu);
764 if (IS_ERR(gpu->mmu)) {
765 dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n");
766 ret = PTR_ERR(gpu->mmu);
767 goto fail;
768 }
769
770 gpu->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(gpu);
771 if (IS_ERR(gpu->cmdbuf_suballoc)) {
772 dev_err(gpu->dev, "Failed to create cmdbuf suballocator\n");
773 ret = PTR_ERR(gpu->cmdbuf_suballoc);
774 goto fail;
775 }
776
777 /* Create buffer: */
778 ret = etnaviv_cmdbuf_init(gpu->cmdbuf_suballoc, &gpu->buffer,
779 PAGE_SIZE);
780 if (ret) {
781 dev_err(gpu->dev, "could not create command buffer\n");
782 goto destroy_iommu;
783 }
784
785 if (gpu->mmu->version == ETNAVIV_IOMMU_V1 &&
786 etnaviv_cmdbuf_get_va(&gpu->buffer) > 0x80000000) {
787 ret = -EINVAL;
788 dev_err(gpu->dev,
789 "command buffer outside valid memory window\n");
790 goto free_buffer;
791 }
792
793 /* Setup event management */
794 spin_lock_init(&gpu->event_spinlock);
795 init_completion(&gpu->event_free);
796 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
797 for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
798 complete(&gpu->event_free);
799
800 /* Now program the hardware */
801 mutex_lock(&gpu->lock);
802 etnaviv_gpu_hw_init(gpu);
803 gpu->exec_state = -1;
804 mutex_unlock(&gpu->lock);
805
806 pm_runtime_mark_last_busy(gpu->dev);
807 pm_runtime_put_autosuspend(gpu->dev);
808
809 return 0;
810
811free_buffer:
812 etnaviv_cmdbuf_free(&gpu->buffer);
813destroy_iommu:
814 etnaviv_iommu_destroy(gpu->mmu);
815 gpu->mmu = NULL;
816fail:
817 pm_runtime_mark_last_busy(gpu->dev);
818 pm_runtime_put_autosuspend(gpu->dev);
819
820 return ret;
821}
822
823#ifdef CONFIG_DEBUG_FS
824struct dma_debug {
825 u32 address[2];
826 u32 state[2];
827};
828
829static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
830{
831 u32 i;
832
833 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
834 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
835
836 for (i = 0; i < 500; i++) {
837 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
838 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
839
840 if (debug->address[0] != debug->address[1])
841 break;
842
843 if (debug->state[0] != debug->state[1])
844 break;
845 }
846}
847
848int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
849{
850 struct dma_debug debug;
851 u32 dma_lo, dma_hi, axi, idle;
852 int ret;
853
854 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
855
856 ret = pm_runtime_get_sync(gpu->dev);
857 if (ret < 0)
858 return ret;
859
860 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
861 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
862 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
863 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
864
865 verify_dma(gpu, &debug);
866
867 seq_puts(m, "\tfeatures\n");
868 seq_printf(m, "\t major_features: 0x%08x\n",
869 gpu->identity.features);
870 seq_printf(m, "\t minor_features0: 0x%08x\n",
871 gpu->identity.minor_features0);
872 seq_printf(m, "\t minor_features1: 0x%08x\n",
873 gpu->identity.minor_features1);
874 seq_printf(m, "\t minor_features2: 0x%08x\n",
875 gpu->identity.minor_features2);
876 seq_printf(m, "\t minor_features3: 0x%08x\n",
877 gpu->identity.minor_features3);
878 seq_printf(m, "\t minor_features4: 0x%08x\n",
879 gpu->identity.minor_features4);
880 seq_printf(m, "\t minor_features5: 0x%08x\n",
881 gpu->identity.minor_features5);
882 seq_printf(m, "\t minor_features6: 0x%08x\n",
883 gpu->identity.minor_features6);
884 seq_printf(m, "\t minor_features7: 0x%08x\n",
885 gpu->identity.minor_features7);
886 seq_printf(m, "\t minor_features8: 0x%08x\n",
887 gpu->identity.minor_features8);
888 seq_printf(m, "\t minor_features9: 0x%08x\n",
889 gpu->identity.minor_features9);
890 seq_printf(m, "\t minor_features10: 0x%08x\n",
891 gpu->identity.minor_features10);
892 seq_printf(m, "\t minor_features11: 0x%08x\n",
893 gpu->identity.minor_features11);
894
895 seq_puts(m, "\tspecs\n");
896 seq_printf(m, "\t stream_count: %d\n",
897 gpu->identity.stream_count);
898 seq_printf(m, "\t register_max: %d\n",
899 gpu->identity.register_max);
900 seq_printf(m, "\t thread_count: %d\n",
901 gpu->identity.thread_count);
902 seq_printf(m, "\t vertex_cache_size: %d\n",
903 gpu->identity.vertex_cache_size);
904 seq_printf(m, "\t shader_core_count: %d\n",
905 gpu->identity.shader_core_count);
906 seq_printf(m, "\t pixel_pipes: %d\n",
907 gpu->identity.pixel_pipes);
908 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
909 gpu->identity.vertex_output_buffer_size);
910 seq_printf(m, "\t buffer_size: %d\n",
911 gpu->identity.buffer_size);
912 seq_printf(m, "\t instruction_count: %d\n",
913 gpu->identity.instruction_count);
914 seq_printf(m, "\t num_constants: %d\n",
915 gpu->identity.num_constants);
916 seq_printf(m, "\t varyings_count: %d\n",
917 gpu->identity.varyings_count);
918
919 seq_printf(m, "\taxi: 0x%08x\n", axi);
920 seq_printf(m, "\tidle: 0x%08x\n", idle);
921 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
922 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
923 seq_puts(m, "\t FE is not idle\n");
924 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
925 seq_puts(m, "\t DE is not idle\n");
926 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
927 seq_puts(m, "\t PE is not idle\n");
928 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
929 seq_puts(m, "\t SH is not idle\n");
930 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
931 seq_puts(m, "\t PA is not idle\n");
932 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
933 seq_puts(m, "\t SE is not idle\n");
934 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
935 seq_puts(m, "\t RA is not idle\n");
936 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
937 seq_puts(m, "\t TX is not idle\n");
938 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
939 seq_puts(m, "\t VG is not idle\n");
940 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
941 seq_puts(m, "\t IM is not idle\n");
942 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
943 seq_puts(m, "\t FP is not idle\n");
944 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
945 seq_puts(m, "\t TS is not idle\n");
946 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
947 seq_puts(m, "\t AXI low power mode\n");
948
949 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
950 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
951 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
952 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
953
954 seq_puts(m, "\tMC\n");
955 seq_printf(m, "\t read0: 0x%08x\n", read0);
956 seq_printf(m, "\t read1: 0x%08x\n", read1);
957 seq_printf(m, "\t write: 0x%08x\n", write);
958 }
959
960 seq_puts(m, "\tDMA ");
961
962 if (debug.address[0] == debug.address[1] &&
963 debug.state[0] == debug.state[1]) {
964 seq_puts(m, "seems to be stuck\n");
965 } else if (debug.address[0] == debug.address[1]) {
966 seq_puts(m, "address is constant\n");
967 } else {
968 seq_puts(m, "is running\n");
969 }
970
971 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
972 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
973 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
974 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
975 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
976 dma_lo, dma_hi);
977
978 ret = 0;
979
980 pm_runtime_mark_last_busy(gpu->dev);
981 pm_runtime_put_autosuspend(gpu->dev);
982
983 return ret;
984}
985#endif
986
987void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
988{
989 unsigned long flags;
990 unsigned int i = 0;
991
992 dev_err(gpu->dev, "recover hung GPU!\n");
993
994 if (pm_runtime_get_sync(gpu->dev) < 0)
995 return;
996
997 mutex_lock(&gpu->lock);
998
999 etnaviv_hw_reset(gpu);
1000
1001 /* complete all events, the GPU won't do it after the reset */
1002 spin_lock_irqsave(&gpu->event_spinlock, flags);
1003 for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1004 complete(&gpu->event_free);
1005 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1006 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1007 gpu->completed_fence = gpu->active_fence;
1008
1009 etnaviv_gpu_hw_init(gpu);
1010 gpu->lastctx = NULL;
1011 gpu->exec_state = -1;
1012
1013 mutex_unlock(&gpu->lock);
1014 pm_runtime_mark_last_busy(gpu->dev);
1015 pm_runtime_put_autosuspend(gpu->dev);
1016}
1017
1018/* fence object management */
1019struct etnaviv_fence {
1020 struct etnaviv_gpu *gpu;
1021 struct dma_fence base;
1022};
1023
1024static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1025{
1026 return container_of(fence, struct etnaviv_fence, base);
1027}
1028
1029static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1030{
1031 return "etnaviv";
1032}
1033
1034static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1035{
1036 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1037
1038 return dev_name(f->gpu->dev);
1039}
1040
1041static bool etnaviv_fence_enable_signaling(struct dma_fence *fence)
1042{
1043 return true;
1044}
1045
1046static bool etnaviv_fence_signaled(struct dma_fence *fence)
1047{
1048 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1049
1050 return fence_completed(f->gpu, f->base.seqno);
1051}
1052
1053static void etnaviv_fence_release(struct dma_fence *fence)
1054{
1055 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1056
1057 kfree_rcu(f, base.rcu);
1058}
1059
1060static const struct dma_fence_ops etnaviv_fence_ops = {
1061 .get_driver_name = etnaviv_fence_get_driver_name,
1062 .get_timeline_name = etnaviv_fence_get_timeline_name,
1063 .enable_signaling = etnaviv_fence_enable_signaling,
1064 .signaled = etnaviv_fence_signaled,
1065 .wait = dma_fence_default_wait,
1066 .release = etnaviv_fence_release,
1067};
1068
1069static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1070{
1071 struct etnaviv_fence *f;
1072
1073 /*
1074 * GPU lock must already be held, otherwise fence completion order might
1075 * not match the seqno order assigned here.
1076 */
1077 lockdep_assert_held(&gpu->lock);
1078
1079 f = kzalloc(sizeof(*f), GFP_KERNEL);
1080 if (!f)
1081 return NULL;
1082
1083 f->gpu = gpu;
1084
1085 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1086 gpu->fence_context, ++gpu->next_fence);
1087
1088 return &f->base;
1089}
1090
1091/*
1092 * event management:
1093 */
1094
1095static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1096 unsigned int *events)
1097{
1098 unsigned long flags, timeout = msecs_to_jiffies(10 * 10000);
1099 unsigned i, acquired = 0;
1100
1101 for (i = 0; i < nr_events; i++) {
1102 unsigned long ret;
1103
1104 ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1105
1106 if (!ret) {
1107 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1108 goto out;
1109 }
1110
1111 acquired++;
1112 timeout = ret;
1113 }
1114
1115 spin_lock_irqsave(&gpu->event_spinlock, flags);
1116
1117 for (i = 0; i < nr_events; i++) {
1118 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1119
1120 events[i] = event;
1121 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1122 set_bit(event, gpu->event_bitmap);
1123 }
1124
1125 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1126
1127 return 0;
1128
1129out:
1130 for (i = 0; i < acquired; i++)
1131 complete(&gpu->event_free);
1132
1133 return -EBUSY;
1134}
1135
1136static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1137{
1138 unsigned long flags;
1139
1140 spin_lock_irqsave(&gpu->event_spinlock, flags);
1141
1142 if (!test_bit(event, gpu->event_bitmap)) {
1143 dev_warn(gpu->dev, "event %u is already marked as free",
1144 event);
1145 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1146 } else {
1147 clear_bit(event, gpu->event_bitmap);
1148 spin_unlock_irqrestore(&gpu->event_spinlock, flags);
1149
1150 complete(&gpu->event_free);
1151 }
1152}
1153
1154/*
1155 * Cmdstream submission/retirement:
1156 */
1157int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1158 u32 id, struct timespec *timeout)
1159{
1160 struct dma_fence *fence;
1161 int ret;
1162
1163 /*
1164 * Look up the fence and take a reference. We might still find a fence
1165 * whose refcount has already dropped to zero. dma_fence_get_rcu
1166 * pretends we didn't find a fence in that case.
1167 */
1168 rcu_read_lock();
1169 fence = idr_find(&gpu->fence_idr, id);
1170 if (fence)
1171 fence = dma_fence_get_rcu(fence);
1172 rcu_read_unlock();
1173
1174 if (!fence)
1175 return 0;
1176
1177 if (!timeout) {
1178 /* No timeout was requested: just test for completion */
1179 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1180 } else {
1181 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1182
1183 ret = dma_fence_wait_timeout(fence, true, remaining);
1184 if (ret == 0)
1185 ret = -ETIMEDOUT;
1186 else if (ret != -ERESTARTSYS)
1187 ret = 0;
1188
1189 }
1190
1191 dma_fence_put(fence);
1192 return ret;
1193}
1194
1195/*
1196 * Wait for an object to become inactive. This, on it's own, is not race
1197 * free: the object is moved by the scheduler off the active list, and
1198 * then the iova is put. Moreover, the object could be re-submitted just
1199 * after we notice that it's become inactive.
1200 *
1201 * Although the retirement happens under the gpu lock, we don't want to hold
1202 * that lock in this function while waiting.
1203 */
1204int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1205 struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1206{
1207 unsigned long remaining;
1208 long ret;
1209
1210 if (!timeout)
1211 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1212
1213 remaining = etnaviv_timeout_to_jiffies(timeout);
1214
1215 ret = wait_event_interruptible_timeout(gpu->fence_event,
1216 !is_active(etnaviv_obj),
1217 remaining);
1218 if (ret > 0)
1219 return 0;
1220 else if (ret == -ERESTARTSYS)
1221 return -ERESTARTSYS;
1222 else
1223 return -ETIMEDOUT;
1224}
1225
1226static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1227 struct etnaviv_event *event, unsigned int flags)
1228{
1229 const struct etnaviv_gem_submit *submit = event->submit;
1230 unsigned int i;
1231
1232 for (i = 0; i < submit->nr_pmrs; i++) {
1233 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1234
1235 if (pmr->flags == flags)
1236 etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1237 }
1238}
1239
1240static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1241 struct etnaviv_event *event)
1242{
1243 u32 val;
1244
1245 /* disable clock gating */
1246 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1247 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1248 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1249
1250 /* enable debug register */
1251 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1252 val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1253 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1254
1255 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1256}
1257
1258static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1259 struct etnaviv_event *event)
1260{
1261 const struct etnaviv_gem_submit *submit = event->submit;
1262 unsigned int i;
1263 u32 val;
1264
1265 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1266
1267 for (i = 0; i < submit->nr_pmrs; i++) {
1268 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1269
1270 *pmr->bo_vma = pmr->sequence;
1271 }
1272
1273 /* disable debug register */
1274 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1275 val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1276 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1277
1278 /* enable clock gating */
1279 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1280 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1281 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1282}
1283
1284
1285/* add bo's to gpu's ring, and kick gpu: */
1286struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1287{
1288 struct etnaviv_gpu *gpu = submit->gpu;
1289 struct dma_fence *gpu_fence;
1290 unsigned int i, nr_events = 1, event[3];
1291 int ret;
1292
1293 if (!submit->runtime_resumed) {
1294 ret = pm_runtime_get_sync(gpu->dev);
1295 if (ret < 0)
1296 return NULL;
1297 submit->runtime_resumed = true;
1298 }
1299
1300 /*
1301 * if there are performance monitor requests we need to have
1302 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1303 * requests.
1304 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1305 * and update the sequence number for userspace.
1306 */
1307 if (submit->nr_pmrs)
1308 nr_events = 3;
1309
1310 ret = event_alloc(gpu, nr_events, event);
1311 if (ret) {
1312 DRM_ERROR("no free events\n");
1313 return NULL;
1314 }
1315
1316 mutex_lock(&gpu->lock);
1317
1318 gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1319 if (!gpu_fence) {
1320 for (i = 0; i < nr_events; i++)
1321 event_free(gpu, event[i]);
1322
1323 goto out_unlock;
1324 }
1325
1326 gpu->active_fence = gpu_fence->seqno;
1327
1328 if (submit->nr_pmrs) {
1329 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1330 kref_get(&submit->refcount);
1331 gpu->event[event[1]].submit = submit;
1332 etnaviv_sync_point_queue(gpu, event[1]);
1333 }
1334
1335 gpu->event[event[0]].fence = gpu_fence;
1336 submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1337 etnaviv_buffer_queue(gpu, submit->exec_state, event[0],
1338 &submit->cmdbuf);
1339
1340 if (submit->nr_pmrs) {
1341 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1342 kref_get(&submit->refcount);
1343 gpu->event[event[2]].submit = submit;
1344 etnaviv_sync_point_queue(gpu, event[2]);
1345 }
1346
1347out_unlock:
1348 mutex_unlock(&gpu->lock);
1349
1350 return gpu_fence;
1351}
1352
1353static void sync_point_worker(struct work_struct *work)
1354{
1355 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1356 sync_point_work);
1357 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1358 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1359
1360 event->sync_point(gpu, event);
1361 etnaviv_submit_put(event->submit);
1362 event_free(gpu, gpu->sync_point_event);
1363
1364 /* restart FE last to avoid GPU and IRQ racing against this worker */
1365 etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1366}
1367
1368static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1369{
1370 u32 status_reg, status;
1371 int i;
1372
1373 if (gpu->sec_mode == ETNA_SEC_NONE)
1374 status_reg = VIVS_MMUv2_STATUS;
1375 else
1376 status_reg = VIVS_MMUv2_SEC_STATUS;
1377
1378 status = gpu_read(gpu, status_reg);
1379 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1380
1381 for (i = 0; i < 4; i++) {
1382 u32 address_reg;
1383
1384 if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1385 continue;
1386
1387 if (gpu->sec_mode == ETNA_SEC_NONE)
1388 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1389 else
1390 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1391
1392 dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1393 gpu_read(gpu, address_reg));
1394 }
1395}
1396
1397static irqreturn_t irq_handler(int irq, void *data)
1398{
1399 struct etnaviv_gpu *gpu = data;
1400 irqreturn_t ret = IRQ_NONE;
1401
1402 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1403
1404 if (intr != 0) {
1405 int event;
1406
1407 pm_runtime_mark_last_busy(gpu->dev);
1408
1409 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1410
1411 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1412 dev_err(gpu->dev, "AXI bus error\n");
1413 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1414 }
1415
1416 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1417 dump_mmu_fault(gpu);
1418 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1419 }
1420
1421 while ((event = ffs(intr)) != 0) {
1422 struct dma_fence *fence;
1423
1424 event -= 1;
1425
1426 intr &= ~(1 << event);
1427
1428 dev_dbg(gpu->dev, "event %u\n", event);
1429
1430 if (gpu->event[event].sync_point) {
1431 gpu->sync_point_event = event;
1432 queue_work(gpu->wq, &gpu->sync_point_work);
1433 }
1434
1435 fence = gpu->event[event].fence;
1436 if (!fence)
1437 continue;
1438
1439 gpu->event[event].fence = NULL;
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 dma_fence_signal(fence);
1453
1454 event_free(gpu, event);
1455 }
1456
1457 ret = IRQ_HANDLED;
1458 }
1459
1460 return ret;
1461}
1462
1463static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1464{
1465 int ret;
1466
1467 if (gpu->clk_reg) {
1468 ret = clk_prepare_enable(gpu->clk_reg);
1469 if (ret)
1470 return ret;
1471 }
1472
1473 if (gpu->clk_bus) {
1474 ret = clk_prepare_enable(gpu->clk_bus);
1475 if (ret)
1476 return ret;
1477 }
1478
1479 if (gpu->clk_core) {
1480 ret = clk_prepare_enable(gpu->clk_core);
1481 if (ret)
1482 goto disable_clk_bus;
1483 }
1484
1485 if (gpu->clk_shader) {
1486 ret = clk_prepare_enable(gpu->clk_shader);
1487 if (ret)
1488 goto disable_clk_core;
1489 }
1490
1491 return 0;
1492
1493disable_clk_core:
1494 if (gpu->clk_core)
1495 clk_disable_unprepare(gpu->clk_core);
1496disable_clk_bus:
1497 if (gpu->clk_bus)
1498 clk_disable_unprepare(gpu->clk_bus);
1499
1500 return ret;
1501}
1502
1503static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1504{
1505 if (gpu->clk_shader)
1506 clk_disable_unprepare(gpu->clk_shader);
1507 if (gpu->clk_core)
1508 clk_disable_unprepare(gpu->clk_core);
1509 if (gpu->clk_bus)
1510 clk_disable_unprepare(gpu->clk_bus);
1511 if (gpu->clk_reg)
1512 clk_disable_unprepare(gpu->clk_reg);
1513
1514 return 0;
1515}
1516
1517int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1518{
1519 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1520
1521 do {
1522 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1523
1524 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1525 return 0;
1526
1527 if (time_is_before_jiffies(timeout)) {
1528 dev_warn(gpu->dev,
1529 "timed out waiting for idle: idle=0x%x\n",
1530 idle);
1531 return -ETIMEDOUT;
1532 }
1533
1534 udelay(5);
1535 } while (1);
1536}
1537
1538static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1539{
1540 if (gpu->buffer.suballoc) {
1541 /* Replace the last WAIT with END */
1542 mutex_lock(&gpu->lock);
1543 etnaviv_buffer_end(gpu);
1544 mutex_unlock(&gpu->lock);
1545
1546 /*
1547 * We know that only the FE is busy here, this should
1548 * happen quickly (as the WAIT is only 200 cycles). If
1549 * we fail, just warn and continue.
1550 */
1551 etnaviv_gpu_wait_idle(gpu, 100);
1552 }
1553
1554 return etnaviv_gpu_clk_disable(gpu);
1555}
1556
1557#ifdef CONFIG_PM
1558static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1559{
1560 int ret;
1561
1562 ret = mutex_lock_killable(&gpu->lock);
1563 if (ret)
1564 return ret;
1565
1566 etnaviv_gpu_update_clock(gpu);
1567 etnaviv_gpu_hw_init(gpu);
1568
1569 gpu->lastctx = NULL;
1570 gpu->exec_state = -1;
1571
1572 mutex_unlock(&gpu->lock);
1573
1574 return 0;
1575}
1576#endif
1577
1578static int
1579etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1580 unsigned long *state)
1581{
1582 *state = 6;
1583
1584 return 0;
1585}
1586
1587static int
1588etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1589 unsigned long *state)
1590{
1591 struct etnaviv_gpu *gpu = cdev->devdata;
1592
1593 *state = gpu->freq_scale;
1594
1595 return 0;
1596}
1597
1598static int
1599etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1600 unsigned long state)
1601{
1602 struct etnaviv_gpu *gpu = cdev->devdata;
1603
1604 mutex_lock(&gpu->lock);
1605 gpu->freq_scale = state;
1606 if (!pm_runtime_suspended(gpu->dev))
1607 etnaviv_gpu_update_clock(gpu);
1608 mutex_unlock(&gpu->lock);
1609
1610 return 0;
1611}
1612
1613static struct thermal_cooling_device_ops cooling_ops = {
1614 .get_max_state = etnaviv_gpu_cooling_get_max_state,
1615 .get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1616 .set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1617};
1618
1619static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1620 void *data)
1621{
1622 struct drm_device *drm = data;
1623 struct etnaviv_drm_private *priv = drm->dev_private;
1624 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1625 int ret;
1626
1627 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1628 gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1629 (char *)dev_name(dev), gpu, &cooling_ops);
1630 if (IS_ERR(gpu->cooling))
1631 return PTR_ERR(gpu->cooling);
1632 }
1633
1634 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1635 if (!gpu->wq) {
1636 ret = -ENOMEM;
1637 goto out_thermal;
1638 }
1639
1640 ret = etnaviv_sched_init(gpu);
1641 if (ret)
1642 goto out_workqueue;
1643
1644#ifdef CONFIG_PM
1645 ret = pm_runtime_get_sync(gpu->dev);
1646#else
1647 ret = etnaviv_gpu_clk_enable(gpu);
1648#endif
1649 if (ret < 0)
1650 goto out_sched;
1651
1652
1653 gpu->drm = drm;
1654 gpu->fence_context = dma_fence_context_alloc(1);
1655 idr_init(&gpu->fence_idr);
1656 spin_lock_init(&gpu->fence_spinlock);
1657
1658 INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1659 init_waitqueue_head(&gpu->fence_event);
1660
1661 priv->gpu[priv->num_gpus++] = gpu;
1662
1663 pm_runtime_mark_last_busy(gpu->dev);
1664 pm_runtime_put_autosuspend(gpu->dev);
1665
1666 return 0;
1667
1668out_sched:
1669 etnaviv_sched_fini(gpu);
1670
1671out_workqueue:
1672 destroy_workqueue(gpu->wq);
1673
1674out_thermal:
1675 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1676 thermal_cooling_device_unregister(gpu->cooling);
1677
1678 return ret;
1679}
1680
1681static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1682 void *data)
1683{
1684 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1685
1686 DBG("%s", dev_name(gpu->dev));
1687
1688 flush_workqueue(gpu->wq);
1689 destroy_workqueue(gpu->wq);
1690
1691 etnaviv_sched_fini(gpu);
1692
1693#ifdef CONFIG_PM
1694 pm_runtime_get_sync(gpu->dev);
1695 pm_runtime_put_sync_suspend(gpu->dev);
1696#else
1697 etnaviv_gpu_hw_suspend(gpu);
1698#endif
1699
1700 if (gpu->buffer.suballoc)
1701 etnaviv_cmdbuf_free(&gpu->buffer);
1702
1703 if (gpu->cmdbuf_suballoc) {
1704 etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc);
1705 gpu->cmdbuf_suballoc = NULL;
1706 }
1707
1708 if (gpu->mmu) {
1709 etnaviv_iommu_destroy(gpu->mmu);
1710 gpu->mmu = NULL;
1711 }
1712
1713 gpu->drm = NULL;
1714 idr_destroy(&gpu->fence_idr);
1715
1716 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1717 thermal_cooling_device_unregister(gpu->cooling);
1718 gpu->cooling = NULL;
1719}
1720
1721static const struct component_ops gpu_ops = {
1722 .bind = etnaviv_gpu_bind,
1723 .unbind = etnaviv_gpu_unbind,
1724};
1725
1726static const struct of_device_id etnaviv_gpu_match[] = {
1727 {
1728 .compatible = "vivante,gc"
1729 },
1730 { /* sentinel */ }
1731};
1732MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1733
1734static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1735{
1736 struct device *dev = &pdev->dev;
1737 struct etnaviv_gpu *gpu;
1738 int err;
1739
1740 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1741 if (!gpu)
1742 return -ENOMEM;
1743
1744 gpu->dev = &pdev->dev;
1745 mutex_init(&gpu->lock);
1746 mutex_init(&gpu->fence_idr_lock);
1747
1748 /* Map registers: */
1749 gpu->mmio = etnaviv_ioremap(pdev, NULL, dev_name(gpu->dev));
1750 if (IS_ERR(gpu->mmio))
1751 return PTR_ERR(gpu->mmio);
1752
1753 /* Get Interrupt: */
1754 gpu->irq = platform_get_irq(pdev, 0);
1755 if (gpu->irq < 0) {
1756 dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1757 return gpu->irq;
1758 }
1759
1760 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1761 dev_name(gpu->dev), gpu);
1762 if (err) {
1763 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1764 return err;
1765 }
1766
1767 /* Get Clocks: */
1768 gpu->clk_reg = devm_clk_get(&pdev->dev, "reg");
1769 DBG("clk_reg: %p", gpu->clk_reg);
1770 if (IS_ERR(gpu->clk_reg))
1771 gpu->clk_reg = NULL;
1772
1773 gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1774 DBG("clk_bus: %p", gpu->clk_bus);
1775 if (IS_ERR(gpu->clk_bus))
1776 gpu->clk_bus = NULL;
1777
1778 gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1779 DBG("clk_core: %p", gpu->clk_core);
1780 if (IS_ERR(gpu->clk_core))
1781 gpu->clk_core = NULL;
1782 gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1783
1784 gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1785 DBG("clk_shader: %p", gpu->clk_shader);
1786 if (IS_ERR(gpu->clk_shader))
1787 gpu->clk_shader = NULL;
1788 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1789
1790 /* TODO: figure out max mapped size */
1791 dev_set_drvdata(dev, gpu);
1792
1793 /*
1794 * We treat the device as initially suspended. The runtime PM
1795 * autosuspend delay is rather arbitary: no measurements have
1796 * yet been performed to determine an appropriate value.
1797 */
1798 pm_runtime_use_autosuspend(gpu->dev);
1799 pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1800 pm_runtime_enable(gpu->dev);
1801
1802 err = component_add(&pdev->dev, &gpu_ops);
1803 if (err < 0) {
1804 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1805 return err;
1806 }
1807
1808 return 0;
1809}
1810
1811static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1812{
1813 component_del(&pdev->dev, &gpu_ops);
1814 pm_runtime_disable(&pdev->dev);
1815 return 0;
1816}
1817
1818#ifdef CONFIG_PM
1819static int etnaviv_gpu_rpm_suspend(struct device *dev)
1820{
1821 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1822 u32 idle, mask;
1823
1824 /* If we have outstanding fences, we're not idle */
1825 if (gpu->completed_fence != gpu->active_fence)
1826 return -EBUSY;
1827
1828 /* Check whether the hardware (except FE) is idle */
1829 mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1830 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1831 if (idle != mask)
1832 return -EBUSY;
1833
1834 return etnaviv_gpu_hw_suspend(gpu);
1835}
1836
1837static int etnaviv_gpu_rpm_resume(struct device *dev)
1838{
1839 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1840 int ret;
1841
1842 ret = etnaviv_gpu_clk_enable(gpu);
1843 if (ret)
1844 return ret;
1845
1846 /* Re-initialise the basic hardware state */
1847 if (gpu->drm && gpu->buffer.suballoc) {
1848 ret = etnaviv_gpu_hw_resume(gpu);
1849 if (ret) {
1850 etnaviv_gpu_clk_disable(gpu);
1851 return ret;
1852 }
1853 }
1854
1855 return 0;
1856}
1857#endif
1858
1859static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1860 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1861 NULL)
1862};
1863
1864struct platform_driver etnaviv_gpu_driver = {
1865 .driver = {
1866 .name = "etnaviv-gpu",
1867 .owner = THIS_MODULE,
1868 .pm = &etnaviv_gpu_pm_ops,
1869 .of_match_table = etnaviv_gpu_match,
1870 },
1871 .probe = etnaviv_gpu_platform_probe,
1872 .remove = etnaviv_gpu_platform_remove,
1873 .id_table = gpu_ids,
1874};