treewide: remove redundant IS_ERR() before error code check
[linux/fpc-iii.git] / drivers / gpu / drm / etnaviv / etnaviv_gpu.c
blob799ec20b267dc991f71571596acab277d7307d0c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2015-2018 Etnaviv Project
4 */
6 #include <linux/clk.h>
7 #include <linux/component.h>
8 #include <linux/delay.h>
9 #include <linux/dma-fence.h>
10 #include <linux/dma-mapping.h>
11 #include <linux/module.h>
12 #include <linux/of_device.h>
13 #include <linux/platform_device.h>
14 #include <linux/pm_runtime.h>
15 #include <linux/regulator/consumer.h>
16 #include <linux/thermal.h>
18 #include "etnaviv_cmdbuf.h"
19 #include "etnaviv_dump.h"
20 #include "etnaviv_gpu.h"
21 #include "etnaviv_gem.h"
22 #include "etnaviv_mmu.h"
23 #include "etnaviv_perfmon.h"
24 #include "etnaviv_sched.h"
25 #include "common.xml.h"
26 #include "state.xml.h"
27 #include "state_hi.xml.h"
28 #include "cmdstream.xml.h"
30 #ifndef PHYS_OFFSET
31 #define PHYS_OFFSET 0
32 #endif
34 static const struct platform_device_id gpu_ids[] = {
35 { .name = "etnaviv-gpu,2d" },
36 { },
40 * Driver functions:
43 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
45 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
47 switch (param) {
48 case ETNAVIV_PARAM_GPU_MODEL:
49 *value = gpu->identity.model;
50 break;
52 case ETNAVIV_PARAM_GPU_REVISION:
53 *value = gpu->identity.revision;
54 break;
56 case ETNAVIV_PARAM_GPU_FEATURES_0:
57 *value = gpu->identity.features;
58 break;
60 case ETNAVIV_PARAM_GPU_FEATURES_1:
61 *value = gpu->identity.minor_features0;
62 break;
64 case ETNAVIV_PARAM_GPU_FEATURES_2:
65 *value = gpu->identity.minor_features1;
66 break;
68 case ETNAVIV_PARAM_GPU_FEATURES_3:
69 *value = gpu->identity.minor_features2;
70 break;
72 case ETNAVIV_PARAM_GPU_FEATURES_4:
73 *value = gpu->identity.minor_features3;
74 break;
76 case ETNAVIV_PARAM_GPU_FEATURES_5:
77 *value = gpu->identity.minor_features4;
78 break;
80 case ETNAVIV_PARAM_GPU_FEATURES_6:
81 *value = gpu->identity.minor_features5;
82 break;
84 case ETNAVIV_PARAM_GPU_FEATURES_7:
85 *value = gpu->identity.minor_features6;
86 break;
88 case ETNAVIV_PARAM_GPU_FEATURES_8:
89 *value = gpu->identity.minor_features7;
90 break;
92 case ETNAVIV_PARAM_GPU_FEATURES_9:
93 *value = gpu->identity.minor_features8;
94 break;
96 case ETNAVIV_PARAM_GPU_FEATURES_10:
97 *value = gpu->identity.minor_features9;
98 break;
100 case ETNAVIV_PARAM_GPU_FEATURES_11:
101 *value = gpu->identity.minor_features10;
102 break;
104 case ETNAVIV_PARAM_GPU_FEATURES_12:
105 *value = gpu->identity.minor_features11;
106 break;
108 case ETNAVIV_PARAM_GPU_STREAM_COUNT:
109 *value = gpu->identity.stream_count;
110 break;
112 case ETNAVIV_PARAM_GPU_REGISTER_MAX:
113 *value = gpu->identity.register_max;
114 break;
116 case ETNAVIV_PARAM_GPU_THREAD_COUNT:
117 *value = gpu->identity.thread_count;
118 break;
120 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
121 *value = gpu->identity.vertex_cache_size;
122 break;
124 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
125 *value = gpu->identity.shader_core_count;
126 break;
128 case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
129 *value = gpu->identity.pixel_pipes;
130 break;
132 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
133 *value = gpu->identity.vertex_output_buffer_size;
134 break;
136 case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
137 *value = gpu->identity.buffer_size;
138 break;
140 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
141 *value = gpu->identity.instruction_count;
142 break;
144 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
145 *value = gpu->identity.num_constants;
146 break;
148 case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
149 *value = gpu->identity.varyings_count;
150 break;
152 case ETNAVIV_PARAM_SOFTPIN_START_ADDR:
153 if (priv->mmu_global->version == ETNAVIV_IOMMU_V2)
154 *value = ETNAVIV_SOFTPIN_START_ADDRESS;
155 else
156 *value = ~0ULL;
157 break;
159 default:
160 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
161 return -EINVAL;
164 return 0;
168 #define etnaviv_is_model_rev(gpu, mod, rev) \
169 ((gpu)->identity.model == chipModel_##mod && \
170 (gpu)->identity.revision == rev)
171 #define etnaviv_field(val, field) \
172 (((val) & field##__MASK) >> field##__SHIFT)
174 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
176 if (gpu->identity.minor_features0 &
177 chipMinorFeatures0_MORE_MINOR_FEATURES) {
178 u32 specs[4];
179 unsigned int streams;
181 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
182 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
183 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
184 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
186 gpu->identity.stream_count = etnaviv_field(specs[0],
187 VIVS_HI_CHIP_SPECS_STREAM_COUNT);
188 gpu->identity.register_max = etnaviv_field(specs[0],
189 VIVS_HI_CHIP_SPECS_REGISTER_MAX);
190 gpu->identity.thread_count = etnaviv_field(specs[0],
191 VIVS_HI_CHIP_SPECS_THREAD_COUNT);
192 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
193 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
194 gpu->identity.shader_core_count = etnaviv_field(specs[0],
195 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
196 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
197 VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
198 gpu->identity.vertex_output_buffer_size =
199 etnaviv_field(specs[0],
200 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
202 gpu->identity.buffer_size = etnaviv_field(specs[1],
203 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
204 gpu->identity.instruction_count = etnaviv_field(specs[1],
205 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
206 gpu->identity.num_constants = etnaviv_field(specs[1],
207 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
209 gpu->identity.varyings_count = etnaviv_field(specs[2],
210 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
212 /* This overrides the value from older register if non-zero */
213 streams = etnaviv_field(specs[3],
214 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
215 if (streams)
216 gpu->identity.stream_count = streams;
219 /* Fill in the stream count if not specified */
220 if (gpu->identity.stream_count == 0) {
221 if (gpu->identity.model >= 0x1000)
222 gpu->identity.stream_count = 4;
223 else
224 gpu->identity.stream_count = 1;
227 /* Convert the register max value */
228 if (gpu->identity.register_max)
229 gpu->identity.register_max = 1 << gpu->identity.register_max;
230 else if (gpu->identity.model == chipModel_GC400)
231 gpu->identity.register_max = 32;
232 else
233 gpu->identity.register_max = 64;
235 /* Convert thread count */
236 if (gpu->identity.thread_count)
237 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
238 else if (gpu->identity.model == chipModel_GC400)
239 gpu->identity.thread_count = 64;
240 else if (gpu->identity.model == chipModel_GC500 ||
241 gpu->identity.model == chipModel_GC530)
242 gpu->identity.thread_count = 128;
243 else
244 gpu->identity.thread_count = 256;
246 if (gpu->identity.vertex_cache_size == 0)
247 gpu->identity.vertex_cache_size = 8;
249 if (gpu->identity.shader_core_count == 0) {
250 if (gpu->identity.model >= 0x1000)
251 gpu->identity.shader_core_count = 2;
252 else
253 gpu->identity.shader_core_count = 1;
256 if (gpu->identity.pixel_pipes == 0)
257 gpu->identity.pixel_pipes = 1;
259 /* Convert virtex buffer size */
260 if (gpu->identity.vertex_output_buffer_size) {
261 gpu->identity.vertex_output_buffer_size =
262 1 << gpu->identity.vertex_output_buffer_size;
263 } else if (gpu->identity.model == chipModel_GC400) {
264 if (gpu->identity.revision < 0x4000)
265 gpu->identity.vertex_output_buffer_size = 512;
266 else if (gpu->identity.revision < 0x4200)
267 gpu->identity.vertex_output_buffer_size = 256;
268 else
269 gpu->identity.vertex_output_buffer_size = 128;
270 } else {
271 gpu->identity.vertex_output_buffer_size = 512;
274 switch (gpu->identity.instruction_count) {
275 case 0:
276 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
277 gpu->identity.model == chipModel_GC880)
278 gpu->identity.instruction_count = 512;
279 else
280 gpu->identity.instruction_count = 256;
281 break;
283 case 1:
284 gpu->identity.instruction_count = 1024;
285 break;
287 case 2:
288 gpu->identity.instruction_count = 2048;
289 break;
291 default:
292 gpu->identity.instruction_count = 256;
293 break;
296 if (gpu->identity.num_constants == 0)
297 gpu->identity.num_constants = 168;
299 if (gpu->identity.varyings_count == 0) {
300 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
301 gpu->identity.varyings_count = 12;
302 else
303 gpu->identity.varyings_count = 8;
307 * For some cores, two varyings are consumed for position, so the
308 * maximum varying count needs to be reduced by one.
310 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
311 etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
312 etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
313 etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
314 etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
315 etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
316 etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
317 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
318 etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
319 etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
320 etnaviv_is_model_rev(gpu, GC880, 0x5106))
321 gpu->identity.varyings_count -= 1;
324 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
326 u32 chipIdentity;
328 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
330 /* Special case for older graphic cores. */
331 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
332 gpu->identity.model = chipModel_GC500;
333 gpu->identity.revision = etnaviv_field(chipIdentity,
334 VIVS_HI_CHIP_IDENTITY_REVISION);
335 } else {
337 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
338 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
341 * !!!! HACK ALERT !!!!
342 * Because people change device IDs without letting software
343 * know about it - here is the hack to make it all look the
344 * same. Only for GC400 family.
346 if ((gpu->identity.model & 0xff00) == 0x0400 &&
347 gpu->identity.model != chipModel_GC420) {
348 gpu->identity.model = gpu->identity.model & 0x0400;
351 /* Another special case */
352 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
353 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
354 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
356 if (chipDate == 0x20080814 && chipTime == 0x12051100) {
358 * This IP has an ECO; put the correct
359 * revision in it.
361 gpu->identity.revision = 0x1051;
366 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
367 * reality it's just a re-branded GC3000. We can identify this
368 * core by the upper half of the revision register being all 1.
369 * Fix model/rev here, so all other places can refer to this
370 * core by its real identity.
372 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
373 gpu->identity.model = chipModel_GC3000;
374 gpu->identity.revision &= 0xffff;
378 dev_info(gpu->dev, "model: GC%x, revision: %x\n",
379 gpu->identity.model, gpu->identity.revision);
381 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
383 * If there is a match in the HWDB, we aren't interested in the
384 * remaining register values, as they might be wrong.
386 if (etnaviv_fill_identity_from_hwdb(gpu))
387 return;
389 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
391 /* Disable fast clear on GC700. */
392 if (gpu->identity.model == chipModel_GC700)
393 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
395 if ((gpu->identity.model == chipModel_GC500 &&
396 gpu->identity.revision < 2) ||
397 (gpu->identity.model == chipModel_GC300 &&
398 gpu->identity.revision < 0x2000)) {
401 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
402 * registers.
404 gpu->identity.minor_features0 = 0;
405 gpu->identity.minor_features1 = 0;
406 gpu->identity.minor_features2 = 0;
407 gpu->identity.minor_features3 = 0;
408 gpu->identity.minor_features4 = 0;
409 gpu->identity.minor_features5 = 0;
410 } else
411 gpu->identity.minor_features0 =
412 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
414 if (gpu->identity.minor_features0 &
415 chipMinorFeatures0_MORE_MINOR_FEATURES) {
416 gpu->identity.minor_features1 =
417 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
418 gpu->identity.minor_features2 =
419 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
420 gpu->identity.minor_features3 =
421 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
422 gpu->identity.minor_features4 =
423 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
424 gpu->identity.minor_features5 =
425 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
428 /* GC600 idle register reports zero bits where modules aren't present */
429 if (gpu->identity.model == chipModel_GC600)
430 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
431 VIVS_HI_IDLE_STATE_RA |
432 VIVS_HI_IDLE_STATE_SE |
433 VIVS_HI_IDLE_STATE_PA |
434 VIVS_HI_IDLE_STATE_SH |
435 VIVS_HI_IDLE_STATE_PE |
436 VIVS_HI_IDLE_STATE_DE |
437 VIVS_HI_IDLE_STATE_FE;
439 etnaviv_hw_specs(gpu);
442 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
444 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
445 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
446 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
449 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
451 if (gpu->identity.minor_features2 &
452 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
453 clk_set_rate(gpu->clk_core,
454 gpu->base_rate_core >> gpu->freq_scale);
455 clk_set_rate(gpu->clk_shader,
456 gpu->base_rate_shader >> gpu->freq_scale);
457 } else {
458 unsigned int fscale = 1 << (6 - gpu->freq_scale);
459 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
461 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
462 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
463 etnaviv_gpu_load_clock(gpu, clock);
467 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
469 u32 control, idle;
470 unsigned long timeout;
471 bool failed = true;
473 /* We hope that the GPU resets in under one second */
474 timeout = jiffies + msecs_to_jiffies(1000);
476 while (time_is_after_jiffies(timeout)) {
477 /* enable clock */
478 unsigned int fscale = 1 << (6 - gpu->freq_scale);
479 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
480 etnaviv_gpu_load_clock(gpu, control);
482 /* isolate the GPU. */
483 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
484 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
486 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
487 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
488 VIVS_MMUv2_AHB_CONTROL_RESET);
489 } else {
490 /* set soft reset. */
491 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
492 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
495 /* wait for reset. */
496 usleep_range(10, 20);
498 /* reset soft reset bit. */
499 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
500 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
502 /* reset GPU isolation. */
503 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
504 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
506 /* read idle register. */
507 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
509 /* try reseting again if FE it not idle */
510 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
511 dev_dbg(gpu->dev, "FE is not idle\n");
512 continue;
515 /* read reset register. */
516 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
518 /* is the GPU idle? */
519 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
520 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
521 dev_dbg(gpu->dev, "GPU is not idle\n");
522 continue;
525 /* disable debug registers, as they are not normally needed */
526 control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
527 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
529 failed = false;
530 break;
533 if (failed) {
534 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
535 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
537 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
538 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
539 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
540 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
542 return -EBUSY;
545 /* We rely on the GPU running, so program the clock */
546 etnaviv_gpu_update_clock(gpu);
548 return 0;
551 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
553 u32 pmc, ppc;
555 /* enable clock gating */
556 ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
557 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
559 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
560 if (gpu->identity.revision == 0x4301 ||
561 gpu->identity.revision == 0x4302)
562 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
564 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
566 pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
568 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
569 if (gpu->identity.model >= chipModel_GC400 &&
570 gpu->identity.model != chipModel_GC420 &&
571 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
572 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
575 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
576 * present without a bug fix.
578 if (gpu->identity.revision < 0x5000 &&
579 gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
580 !(gpu->identity.minor_features1 &
581 chipMinorFeatures1_DISABLE_PE_GATING))
582 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
584 if (gpu->identity.revision < 0x5422)
585 pmc |= BIT(15); /* Unknown bit */
587 /* Disable TX clock gating on affected core revisions. */
588 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
589 etnaviv_is_model_rev(gpu, GC2000, 0x5108))
590 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
592 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
593 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
595 gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
598 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
600 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
601 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
602 VIVS_FE_COMMAND_CONTROL_ENABLE |
603 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
605 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
606 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
607 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
608 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
612 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu)
614 u32 address = etnaviv_cmdbuf_get_va(&gpu->buffer,
615 &gpu->mmu_context->cmdbuf_mapping);
616 u16 prefetch;
618 /* setup the MMU */
619 etnaviv_iommu_restore(gpu, gpu->mmu_context);
621 /* Start command processor */
622 prefetch = etnaviv_buffer_init(gpu);
624 etnaviv_gpu_start_fe(gpu, address, prefetch);
627 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
630 * Base value for VIVS_PM_PULSE_EATER register on models where it
631 * cannot be read, extracted from vivante kernel driver.
633 u32 pulse_eater = 0x01590880;
635 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
636 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
637 pulse_eater |= BIT(23);
641 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
642 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
643 pulse_eater &= ~BIT(16);
644 pulse_eater |= BIT(17);
647 if ((gpu->identity.revision > 0x5420) &&
648 (gpu->identity.features & chipFeatures_PIPE_3D))
650 /* Performance fix: disable internal DFS */
651 pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
652 pulse_eater |= BIT(18);
655 gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
658 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
660 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
661 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
662 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
663 u32 mc_memory_debug;
665 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
667 if (gpu->identity.revision == 0x5007)
668 mc_memory_debug |= 0x0c;
669 else
670 mc_memory_debug |= 0x08;
672 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
675 /* enable module-level clock gating */
676 etnaviv_gpu_enable_mlcg(gpu);
679 * Update GPU AXI cache atttribute to "cacheable, no allocate".
680 * This is necessary to prevent the iMX6 SoC locking up.
682 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
683 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
684 VIVS_HI_AXI_CONFIG_ARCACHE(2));
686 /* GC2000 rev 5108 needs a special bus config */
687 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
688 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
689 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
690 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
691 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
692 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
693 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
696 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
697 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
698 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
699 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
702 /* setup the pulse eater */
703 etnaviv_gpu_setup_pulse_eater(gpu);
705 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
708 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
710 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
711 int ret, i;
713 ret = pm_runtime_get_sync(gpu->dev);
714 if (ret < 0) {
715 dev_err(gpu->dev, "Failed to enable GPU power domain\n");
716 return ret;
719 etnaviv_hw_identify(gpu);
721 if (gpu->identity.model == 0) {
722 dev_err(gpu->dev, "Unknown GPU model\n");
723 ret = -ENXIO;
724 goto fail;
727 /* Exclude VG cores with FE2.0 */
728 if (gpu->identity.features & chipFeatures_PIPE_VG &&
729 gpu->identity.features & chipFeatures_FE20) {
730 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
731 ret = -ENXIO;
732 goto fail;
736 * On cores with security features supported, we claim control over the
737 * security states.
739 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
740 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
741 gpu->sec_mode = ETNA_SEC_KERNEL;
743 ret = etnaviv_hw_reset(gpu);
744 if (ret) {
745 dev_err(gpu->dev, "GPU reset failed\n");
746 goto fail;
749 ret = etnaviv_iommu_global_init(gpu);
750 if (ret)
751 goto fail;
754 * Set the GPU linear window to be at the end of the DMA window, where
755 * the CMA area is likely to reside. This ensures that we are able to
756 * map the command buffers while having the linear window overlap as
757 * much RAM as possible, so we can optimize mappings for other buffers.
759 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
760 * to different views of the memory on the individual engines.
762 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
763 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
764 u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
765 if (dma_mask < PHYS_OFFSET + SZ_2G)
766 priv->mmu_global->memory_base = PHYS_OFFSET;
767 else
768 priv->mmu_global->memory_base = dma_mask - SZ_2G + 1;
769 } else if (PHYS_OFFSET >= SZ_2G) {
770 dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
771 priv->mmu_global->memory_base = PHYS_OFFSET;
772 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
775 /* Create buffer: */
776 ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
777 PAGE_SIZE);
778 if (ret) {
779 dev_err(gpu->dev, "could not create command buffer\n");
780 goto fail;
783 /* Setup event management */
784 spin_lock_init(&gpu->event_spinlock);
785 init_completion(&gpu->event_free);
786 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
787 for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
788 complete(&gpu->event_free);
790 /* Now program the hardware */
791 mutex_lock(&gpu->lock);
792 etnaviv_gpu_hw_init(gpu);
793 gpu->exec_state = -1;
794 mutex_unlock(&gpu->lock);
796 pm_runtime_mark_last_busy(gpu->dev);
797 pm_runtime_put_autosuspend(gpu->dev);
799 gpu->initialized = true;
801 return 0;
803 fail:
804 pm_runtime_mark_last_busy(gpu->dev);
805 pm_runtime_put_autosuspend(gpu->dev);
807 return ret;
810 #ifdef CONFIG_DEBUG_FS
811 struct dma_debug {
812 u32 address[2];
813 u32 state[2];
816 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
818 u32 i;
820 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
821 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
823 for (i = 0; i < 500; i++) {
824 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
825 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
827 if (debug->address[0] != debug->address[1])
828 break;
830 if (debug->state[0] != debug->state[1])
831 break;
835 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
837 struct dma_debug debug;
838 u32 dma_lo, dma_hi, axi, idle;
839 int ret;
841 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
843 ret = pm_runtime_get_sync(gpu->dev);
844 if (ret < 0)
845 return ret;
847 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
848 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
849 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
850 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
852 verify_dma(gpu, &debug);
854 seq_puts(m, "\tfeatures\n");
855 seq_printf(m, "\t major_features: 0x%08x\n",
856 gpu->identity.features);
857 seq_printf(m, "\t minor_features0: 0x%08x\n",
858 gpu->identity.minor_features0);
859 seq_printf(m, "\t minor_features1: 0x%08x\n",
860 gpu->identity.minor_features1);
861 seq_printf(m, "\t minor_features2: 0x%08x\n",
862 gpu->identity.minor_features2);
863 seq_printf(m, "\t minor_features3: 0x%08x\n",
864 gpu->identity.minor_features3);
865 seq_printf(m, "\t minor_features4: 0x%08x\n",
866 gpu->identity.minor_features4);
867 seq_printf(m, "\t minor_features5: 0x%08x\n",
868 gpu->identity.minor_features5);
869 seq_printf(m, "\t minor_features6: 0x%08x\n",
870 gpu->identity.minor_features6);
871 seq_printf(m, "\t minor_features7: 0x%08x\n",
872 gpu->identity.minor_features7);
873 seq_printf(m, "\t minor_features8: 0x%08x\n",
874 gpu->identity.minor_features8);
875 seq_printf(m, "\t minor_features9: 0x%08x\n",
876 gpu->identity.minor_features9);
877 seq_printf(m, "\t minor_features10: 0x%08x\n",
878 gpu->identity.minor_features10);
879 seq_printf(m, "\t minor_features11: 0x%08x\n",
880 gpu->identity.minor_features11);
882 seq_puts(m, "\tspecs\n");
883 seq_printf(m, "\t stream_count: %d\n",
884 gpu->identity.stream_count);
885 seq_printf(m, "\t register_max: %d\n",
886 gpu->identity.register_max);
887 seq_printf(m, "\t thread_count: %d\n",
888 gpu->identity.thread_count);
889 seq_printf(m, "\t vertex_cache_size: %d\n",
890 gpu->identity.vertex_cache_size);
891 seq_printf(m, "\t shader_core_count: %d\n",
892 gpu->identity.shader_core_count);
893 seq_printf(m, "\t pixel_pipes: %d\n",
894 gpu->identity.pixel_pipes);
895 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
896 gpu->identity.vertex_output_buffer_size);
897 seq_printf(m, "\t buffer_size: %d\n",
898 gpu->identity.buffer_size);
899 seq_printf(m, "\t instruction_count: %d\n",
900 gpu->identity.instruction_count);
901 seq_printf(m, "\t num_constants: %d\n",
902 gpu->identity.num_constants);
903 seq_printf(m, "\t varyings_count: %d\n",
904 gpu->identity.varyings_count);
906 seq_printf(m, "\taxi: 0x%08x\n", axi);
907 seq_printf(m, "\tidle: 0x%08x\n", idle);
908 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
909 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
910 seq_puts(m, "\t FE is not idle\n");
911 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
912 seq_puts(m, "\t DE is not idle\n");
913 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
914 seq_puts(m, "\t PE is not idle\n");
915 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
916 seq_puts(m, "\t SH is not idle\n");
917 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
918 seq_puts(m, "\t PA is not idle\n");
919 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
920 seq_puts(m, "\t SE is not idle\n");
921 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
922 seq_puts(m, "\t RA is not idle\n");
923 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
924 seq_puts(m, "\t TX is not idle\n");
925 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
926 seq_puts(m, "\t VG is not idle\n");
927 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
928 seq_puts(m, "\t IM is not idle\n");
929 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
930 seq_puts(m, "\t FP is not idle\n");
931 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
932 seq_puts(m, "\t TS is not idle\n");
933 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
934 seq_puts(m, "\t AXI low power mode\n");
936 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
937 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
938 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
939 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
941 seq_puts(m, "\tMC\n");
942 seq_printf(m, "\t read0: 0x%08x\n", read0);
943 seq_printf(m, "\t read1: 0x%08x\n", read1);
944 seq_printf(m, "\t write: 0x%08x\n", write);
947 seq_puts(m, "\tDMA ");
949 if (debug.address[0] == debug.address[1] &&
950 debug.state[0] == debug.state[1]) {
951 seq_puts(m, "seems to be stuck\n");
952 } else if (debug.address[0] == debug.address[1]) {
953 seq_puts(m, "address is constant\n");
954 } else {
955 seq_puts(m, "is running\n");
958 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
959 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
960 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
961 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
962 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
963 dma_lo, dma_hi);
965 ret = 0;
967 pm_runtime_mark_last_busy(gpu->dev);
968 pm_runtime_put_autosuspend(gpu->dev);
970 return ret;
972 #endif
974 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
976 unsigned int i = 0;
978 dev_err(gpu->dev, "recover hung GPU!\n");
980 if (pm_runtime_get_sync(gpu->dev) < 0)
981 return;
983 mutex_lock(&gpu->lock);
985 etnaviv_hw_reset(gpu);
987 /* complete all events, the GPU won't do it after the reset */
988 spin_lock(&gpu->event_spinlock);
989 for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
990 complete(&gpu->event_free);
991 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
992 spin_unlock(&gpu->event_spinlock);
994 etnaviv_gpu_hw_init(gpu);
995 gpu->exec_state = -1;
996 gpu->mmu_context = NULL;
998 mutex_unlock(&gpu->lock);
999 pm_runtime_mark_last_busy(gpu->dev);
1000 pm_runtime_put_autosuspend(gpu->dev);
1003 /* fence object management */
1004 struct etnaviv_fence {
1005 struct etnaviv_gpu *gpu;
1006 struct dma_fence base;
1009 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1011 return container_of(fence, struct etnaviv_fence, base);
1014 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1016 return "etnaviv";
1019 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1021 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1023 return dev_name(f->gpu->dev);
1026 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1028 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1030 return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1033 static void etnaviv_fence_release(struct dma_fence *fence)
1035 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1037 kfree_rcu(f, base.rcu);
1040 static const struct dma_fence_ops etnaviv_fence_ops = {
1041 .get_driver_name = etnaviv_fence_get_driver_name,
1042 .get_timeline_name = etnaviv_fence_get_timeline_name,
1043 .signaled = etnaviv_fence_signaled,
1044 .release = etnaviv_fence_release,
1047 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1049 struct etnaviv_fence *f;
1052 * GPU lock must already be held, otherwise fence completion order might
1053 * not match the seqno order assigned here.
1055 lockdep_assert_held(&gpu->lock);
1057 f = kzalloc(sizeof(*f), GFP_KERNEL);
1058 if (!f)
1059 return NULL;
1061 f->gpu = gpu;
1063 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1064 gpu->fence_context, ++gpu->next_fence);
1066 return &f->base;
1069 /* returns true if fence a comes after fence b */
1070 static inline bool fence_after(u32 a, u32 b)
1072 return (s32)(a - b) > 0;
1076 * event management:
1079 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1080 unsigned int *events)
1082 unsigned long timeout = msecs_to_jiffies(10 * 10000);
1083 unsigned i, acquired = 0;
1085 for (i = 0; i < nr_events; i++) {
1086 unsigned long ret;
1088 ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1090 if (!ret) {
1091 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1092 goto out;
1095 acquired++;
1096 timeout = ret;
1099 spin_lock(&gpu->event_spinlock);
1101 for (i = 0; i < nr_events; i++) {
1102 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1104 events[i] = event;
1105 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1106 set_bit(event, gpu->event_bitmap);
1109 spin_unlock(&gpu->event_spinlock);
1111 return 0;
1113 out:
1114 for (i = 0; i < acquired; i++)
1115 complete(&gpu->event_free);
1117 return -EBUSY;
1120 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1122 if (!test_bit(event, gpu->event_bitmap)) {
1123 dev_warn(gpu->dev, "event %u is already marked as free",
1124 event);
1125 } else {
1126 clear_bit(event, gpu->event_bitmap);
1127 complete(&gpu->event_free);
1132 * Cmdstream submission/retirement:
1134 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1135 u32 id, struct drm_etnaviv_timespec *timeout)
1137 struct dma_fence *fence;
1138 int ret;
1141 * Look up the fence and take a reference. We might still find a fence
1142 * whose refcount has already dropped to zero. dma_fence_get_rcu
1143 * pretends we didn't find a fence in that case.
1145 rcu_read_lock();
1146 fence = idr_find(&gpu->fence_idr, id);
1147 if (fence)
1148 fence = dma_fence_get_rcu(fence);
1149 rcu_read_unlock();
1151 if (!fence)
1152 return 0;
1154 if (!timeout) {
1155 /* No timeout was requested: just test for completion */
1156 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1157 } else {
1158 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1160 ret = dma_fence_wait_timeout(fence, true, remaining);
1161 if (ret == 0)
1162 ret = -ETIMEDOUT;
1163 else if (ret != -ERESTARTSYS)
1164 ret = 0;
1168 dma_fence_put(fence);
1169 return ret;
1173 * Wait for an object to become inactive. This, on it's own, is not race
1174 * free: the object is moved by the scheduler off the active list, and
1175 * then the iova is put. Moreover, the object could be re-submitted just
1176 * after we notice that it's become inactive.
1178 * Although the retirement happens under the gpu lock, we don't want to hold
1179 * that lock in this function while waiting.
1181 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1182 struct etnaviv_gem_object *etnaviv_obj,
1183 struct drm_etnaviv_timespec *timeout)
1185 unsigned long remaining;
1186 long ret;
1188 if (!timeout)
1189 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1191 remaining = etnaviv_timeout_to_jiffies(timeout);
1193 ret = wait_event_interruptible_timeout(gpu->fence_event,
1194 !is_active(etnaviv_obj),
1195 remaining);
1196 if (ret > 0)
1197 return 0;
1198 else if (ret == -ERESTARTSYS)
1199 return -ERESTARTSYS;
1200 else
1201 return -ETIMEDOUT;
1204 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1205 struct etnaviv_event *event, unsigned int flags)
1207 const struct etnaviv_gem_submit *submit = event->submit;
1208 unsigned int i;
1210 for (i = 0; i < submit->nr_pmrs; i++) {
1211 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1213 if (pmr->flags == flags)
1214 etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1218 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1219 struct etnaviv_event *event)
1221 u32 val;
1223 /* disable clock gating */
1224 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1225 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1226 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1228 /* enable debug register */
1229 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1230 val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1231 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1233 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1236 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1237 struct etnaviv_event *event)
1239 const struct etnaviv_gem_submit *submit = event->submit;
1240 unsigned int i;
1241 u32 val;
1243 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1245 for (i = 0; i < submit->nr_pmrs; i++) {
1246 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1248 *pmr->bo_vma = pmr->sequence;
1251 /* disable debug register */
1252 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1253 val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1254 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1256 /* enable clock gating */
1257 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1258 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1259 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1263 /* add bo's to gpu's ring, and kick gpu: */
1264 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1266 struct etnaviv_gpu *gpu = submit->gpu;
1267 struct dma_fence *gpu_fence;
1268 unsigned int i, nr_events = 1, event[3];
1269 int ret;
1271 if (!submit->runtime_resumed) {
1272 ret = pm_runtime_get_sync(gpu->dev);
1273 if (ret < 0)
1274 return NULL;
1275 submit->runtime_resumed = true;
1279 * if there are performance monitor requests we need to have
1280 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1281 * requests.
1282 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1283 * and update the sequence number for userspace.
1285 if (submit->nr_pmrs)
1286 nr_events = 3;
1288 ret = event_alloc(gpu, nr_events, event);
1289 if (ret) {
1290 DRM_ERROR("no free events\n");
1291 return NULL;
1294 mutex_lock(&gpu->lock);
1296 gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1297 if (!gpu_fence) {
1298 for (i = 0; i < nr_events; i++)
1299 event_free(gpu, event[i]);
1301 goto out_unlock;
1304 if (!gpu->mmu_context) {
1305 etnaviv_iommu_context_get(submit->mmu_context);
1306 gpu->mmu_context = submit->mmu_context;
1307 etnaviv_gpu_start_fe_idleloop(gpu);
1308 } else {
1309 etnaviv_iommu_context_get(gpu->mmu_context);
1310 submit->prev_mmu_context = gpu->mmu_context;
1313 if (submit->nr_pmrs) {
1314 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1315 kref_get(&submit->refcount);
1316 gpu->event[event[1]].submit = submit;
1317 etnaviv_sync_point_queue(gpu, event[1]);
1320 gpu->event[event[0]].fence = gpu_fence;
1321 submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1322 etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1323 event[0], &submit->cmdbuf);
1325 if (submit->nr_pmrs) {
1326 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1327 kref_get(&submit->refcount);
1328 gpu->event[event[2]].submit = submit;
1329 etnaviv_sync_point_queue(gpu, event[2]);
1332 out_unlock:
1333 mutex_unlock(&gpu->lock);
1335 return gpu_fence;
1338 static void sync_point_worker(struct work_struct *work)
1340 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1341 sync_point_work);
1342 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1343 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1345 event->sync_point(gpu, event);
1346 etnaviv_submit_put(event->submit);
1347 event_free(gpu, gpu->sync_point_event);
1349 /* restart FE last to avoid GPU and IRQ racing against this worker */
1350 etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1353 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1355 u32 status_reg, status;
1356 int i;
1358 if (gpu->sec_mode == ETNA_SEC_NONE)
1359 status_reg = VIVS_MMUv2_STATUS;
1360 else
1361 status_reg = VIVS_MMUv2_SEC_STATUS;
1363 status = gpu_read(gpu, status_reg);
1364 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1366 for (i = 0; i < 4; i++) {
1367 u32 address_reg;
1369 if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1370 continue;
1372 if (gpu->sec_mode == ETNA_SEC_NONE)
1373 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1374 else
1375 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1377 dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1378 gpu_read(gpu, address_reg));
1382 static irqreturn_t irq_handler(int irq, void *data)
1384 struct etnaviv_gpu *gpu = data;
1385 irqreturn_t ret = IRQ_NONE;
1387 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1389 if (intr != 0) {
1390 int event;
1392 pm_runtime_mark_last_busy(gpu->dev);
1394 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1396 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1397 dev_err(gpu->dev, "AXI bus error\n");
1398 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1401 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1402 dump_mmu_fault(gpu);
1403 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1406 while ((event = ffs(intr)) != 0) {
1407 struct dma_fence *fence;
1409 event -= 1;
1411 intr &= ~(1 << event);
1413 dev_dbg(gpu->dev, "event %u\n", event);
1415 if (gpu->event[event].sync_point) {
1416 gpu->sync_point_event = event;
1417 queue_work(gpu->wq, &gpu->sync_point_work);
1420 fence = gpu->event[event].fence;
1421 if (!fence)
1422 continue;
1424 gpu->event[event].fence = NULL;
1427 * Events can be processed out of order. Eg,
1428 * - allocate and queue event 0
1429 * - allocate event 1
1430 * - event 0 completes, we process it
1431 * - allocate and queue event 0
1432 * - event 1 and event 0 complete
1433 * we can end up processing event 0 first, then 1.
1435 if (fence_after(fence->seqno, gpu->completed_fence))
1436 gpu->completed_fence = fence->seqno;
1437 dma_fence_signal(fence);
1439 event_free(gpu, event);
1442 ret = IRQ_HANDLED;
1445 return ret;
1448 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1450 int ret;
1452 if (gpu->clk_reg) {
1453 ret = clk_prepare_enable(gpu->clk_reg);
1454 if (ret)
1455 return ret;
1458 if (gpu->clk_bus) {
1459 ret = clk_prepare_enable(gpu->clk_bus);
1460 if (ret)
1461 return ret;
1464 if (gpu->clk_core) {
1465 ret = clk_prepare_enable(gpu->clk_core);
1466 if (ret)
1467 goto disable_clk_bus;
1470 if (gpu->clk_shader) {
1471 ret = clk_prepare_enable(gpu->clk_shader);
1472 if (ret)
1473 goto disable_clk_core;
1476 return 0;
1478 disable_clk_core:
1479 if (gpu->clk_core)
1480 clk_disable_unprepare(gpu->clk_core);
1481 disable_clk_bus:
1482 if (gpu->clk_bus)
1483 clk_disable_unprepare(gpu->clk_bus);
1485 return ret;
1488 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1490 if (gpu->clk_shader)
1491 clk_disable_unprepare(gpu->clk_shader);
1492 if (gpu->clk_core)
1493 clk_disable_unprepare(gpu->clk_core);
1494 if (gpu->clk_bus)
1495 clk_disable_unprepare(gpu->clk_bus);
1496 if (gpu->clk_reg)
1497 clk_disable_unprepare(gpu->clk_reg);
1499 return 0;
1502 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1504 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1506 do {
1507 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1509 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1510 return 0;
1512 if (time_is_before_jiffies(timeout)) {
1513 dev_warn(gpu->dev,
1514 "timed out waiting for idle: idle=0x%x\n",
1515 idle);
1516 return -ETIMEDOUT;
1519 udelay(5);
1520 } while (1);
1523 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1525 if (gpu->initialized && gpu->mmu_context) {
1526 /* Replace the last WAIT with END */
1527 mutex_lock(&gpu->lock);
1528 etnaviv_buffer_end(gpu);
1529 mutex_unlock(&gpu->lock);
1532 * We know that only the FE is busy here, this should
1533 * happen quickly (as the WAIT is only 200 cycles). If
1534 * we fail, just warn and continue.
1536 etnaviv_gpu_wait_idle(gpu, 100);
1538 etnaviv_iommu_context_put(gpu->mmu_context);
1539 gpu->mmu_context = NULL;
1542 gpu->exec_state = -1;
1544 return etnaviv_gpu_clk_disable(gpu);
1547 #ifdef CONFIG_PM
1548 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1550 int ret;
1552 ret = mutex_lock_killable(&gpu->lock);
1553 if (ret)
1554 return ret;
1556 etnaviv_gpu_update_clock(gpu);
1557 etnaviv_gpu_hw_init(gpu);
1559 mutex_unlock(&gpu->lock);
1561 return 0;
1563 #endif
1565 static int
1566 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1567 unsigned long *state)
1569 *state = 6;
1571 return 0;
1574 static int
1575 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1576 unsigned long *state)
1578 struct etnaviv_gpu *gpu = cdev->devdata;
1580 *state = gpu->freq_scale;
1582 return 0;
1585 static int
1586 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1587 unsigned long state)
1589 struct etnaviv_gpu *gpu = cdev->devdata;
1591 mutex_lock(&gpu->lock);
1592 gpu->freq_scale = state;
1593 if (!pm_runtime_suspended(gpu->dev))
1594 etnaviv_gpu_update_clock(gpu);
1595 mutex_unlock(&gpu->lock);
1597 return 0;
1600 static struct thermal_cooling_device_ops cooling_ops = {
1601 .get_max_state = etnaviv_gpu_cooling_get_max_state,
1602 .get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1603 .set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1606 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1607 void *data)
1609 struct drm_device *drm = data;
1610 struct etnaviv_drm_private *priv = drm->dev_private;
1611 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1612 int ret;
1614 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1615 gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1616 (char *)dev_name(dev), gpu, &cooling_ops);
1617 if (IS_ERR(gpu->cooling))
1618 return PTR_ERR(gpu->cooling);
1621 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1622 if (!gpu->wq) {
1623 ret = -ENOMEM;
1624 goto out_thermal;
1627 ret = etnaviv_sched_init(gpu);
1628 if (ret)
1629 goto out_workqueue;
1631 #ifdef CONFIG_PM
1632 ret = pm_runtime_get_sync(gpu->dev);
1633 #else
1634 ret = etnaviv_gpu_clk_enable(gpu);
1635 #endif
1636 if (ret < 0)
1637 goto out_sched;
1640 gpu->drm = drm;
1641 gpu->fence_context = dma_fence_context_alloc(1);
1642 idr_init(&gpu->fence_idr);
1643 spin_lock_init(&gpu->fence_spinlock);
1645 INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1646 init_waitqueue_head(&gpu->fence_event);
1648 priv->gpu[priv->num_gpus++] = gpu;
1650 pm_runtime_mark_last_busy(gpu->dev);
1651 pm_runtime_put_autosuspend(gpu->dev);
1653 return 0;
1655 out_sched:
1656 etnaviv_sched_fini(gpu);
1658 out_workqueue:
1659 destroy_workqueue(gpu->wq);
1661 out_thermal:
1662 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1663 thermal_cooling_device_unregister(gpu->cooling);
1665 return ret;
1668 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1669 void *data)
1671 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1673 DBG("%s", dev_name(gpu->dev));
1675 flush_workqueue(gpu->wq);
1676 destroy_workqueue(gpu->wq);
1678 etnaviv_sched_fini(gpu);
1680 #ifdef CONFIG_PM
1681 pm_runtime_get_sync(gpu->dev);
1682 pm_runtime_put_sync_suspend(gpu->dev);
1683 #else
1684 etnaviv_gpu_hw_suspend(gpu);
1685 #endif
1687 if (gpu->initialized) {
1688 etnaviv_cmdbuf_free(&gpu->buffer);
1689 etnaviv_iommu_global_fini(gpu);
1690 gpu->initialized = false;
1693 gpu->drm = NULL;
1694 idr_destroy(&gpu->fence_idr);
1696 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1697 thermal_cooling_device_unregister(gpu->cooling);
1698 gpu->cooling = NULL;
1701 static const struct component_ops gpu_ops = {
1702 .bind = etnaviv_gpu_bind,
1703 .unbind = etnaviv_gpu_unbind,
1706 static const struct of_device_id etnaviv_gpu_match[] = {
1708 .compatible = "vivante,gc"
1710 { /* sentinel */ }
1712 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1714 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1716 struct device *dev = &pdev->dev;
1717 struct etnaviv_gpu *gpu;
1718 int err;
1720 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1721 if (!gpu)
1722 return -ENOMEM;
1724 gpu->dev = &pdev->dev;
1725 mutex_init(&gpu->lock);
1726 mutex_init(&gpu->fence_lock);
1728 /* Map registers: */
1729 gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1730 if (IS_ERR(gpu->mmio))
1731 return PTR_ERR(gpu->mmio);
1733 /* Get Interrupt: */
1734 gpu->irq = platform_get_irq(pdev, 0);
1735 if (gpu->irq < 0) {
1736 dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1737 return gpu->irq;
1740 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1741 dev_name(gpu->dev), gpu);
1742 if (err) {
1743 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1744 return err;
1747 /* Get Clocks: */
1748 gpu->clk_reg = devm_clk_get(&pdev->dev, "reg");
1749 DBG("clk_reg: %p", gpu->clk_reg);
1750 if (IS_ERR(gpu->clk_reg))
1751 gpu->clk_reg = NULL;
1753 gpu->clk_bus = devm_clk_get(&pdev->dev, "bus");
1754 DBG("clk_bus: %p", gpu->clk_bus);
1755 if (IS_ERR(gpu->clk_bus))
1756 gpu->clk_bus = NULL;
1758 gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1759 DBG("clk_core: %p", gpu->clk_core);
1760 if (IS_ERR(gpu->clk_core))
1761 gpu->clk_core = NULL;
1762 gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1764 gpu->clk_shader = devm_clk_get(&pdev->dev, "shader");
1765 DBG("clk_shader: %p", gpu->clk_shader);
1766 if (IS_ERR(gpu->clk_shader))
1767 gpu->clk_shader = NULL;
1768 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1770 /* TODO: figure out max mapped size */
1771 dev_set_drvdata(dev, gpu);
1774 * We treat the device as initially suspended. The runtime PM
1775 * autosuspend delay is rather arbitary: no measurements have
1776 * yet been performed to determine an appropriate value.
1778 pm_runtime_use_autosuspend(gpu->dev);
1779 pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1780 pm_runtime_enable(gpu->dev);
1782 err = component_add(&pdev->dev, &gpu_ops);
1783 if (err < 0) {
1784 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1785 return err;
1788 return 0;
1791 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1793 component_del(&pdev->dev, &gpu_ops);
1794 pm_runtime_disable(&pdev->dev);
1795 return 0;
1798 #ifdef CONFIG_PM
1799 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1801 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1802 u32 idle, mask;
1804 /* If there are any jobs in the HW queue, we're not idle */
1805 if (atomic_read(&gpu->sched.hw_rq_count))
1806 return -EBUSY;
1808 /* Check whether the hardware (except FE) is idle */
1809 mask = gpu->idle_mask & ~VIVS_HI_IDLE_STATE_FE;
1810 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1811 if (idle != mask)
1812 return -EBUSY;
1814 return etnaviv_gpu_hw_suspend(gpu);
1817 static int etnaviv_gpu_rpm_resume(struct device *dev)
1819 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1820 int ret;
1822 ret = etnaviv_gpu_clk_enable(gpu);
1823 if (ret)
1824 return ret;
1826 /* Re-initialise the basic hardware state */
1827 if (gpu->drm && gpu->initialized) {
1828 ret = etnaviv_gpu_hw_resume(gpu);
1829 if (ret) {
1830 etnaviv_gpu_clk_disable(gpu);
1831 return ret;
1835 return 0;
1837 #endif
1839 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1840 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1841 NULL)
1844 struct platform_driver etnaviv_gpu_driver = {
1845 .driver = {
1846 .name = "etnaviv-gpu",
1847 .owner = THIS_MODULE,
1848 .pm = &etnaviv_gpu_pm_ops,
1849 .of_match_table = etnaviv_gpu_match,
1851 .probe = etnaviv_gpu_platform_probe,
1852 .remove = etnaviv_gpu_platform_remove,
1853 .id_table = gpu_ids,