1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015-2018 Etnaviv Project
7 #include <linux/component.h>
8 #include <linux/dma-fence.h>
9 #include <linux/moduleparam.h>
10 #include <linux/of_device.h>
11 #include <linux/regulator/consumer.h>
12 #include <linux/thermal.h>
14 #include "etnaviv_cmdbuf.h"
15 #include "etnaviv_dump.h"
16 #include "etnaviv_gpu.h"
17 #include "etnaviv_gem.h"
18 #include "etnaviv_mmu.h"
19 #include "etnaviv_perfmon.h"
20 #include "etnaviv_sched.h"
21 #include "common.xml.h"
22 #include "state.xml.h"
23 #include "state_hi.xml.h"
24 #include "cmdstream.xml.h"
30 static const struct platform_device_id gpu_ids[] = {
31 { .name = "etnaviv-gpu,2d" },
39 int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value)
42 case ETNAVIV_PARAM_GPU_MODEL:
43 *value = gpu->identity.model;
46 case ETNAVIV_PARAM_GPU_REVISION:
47 *value = gpu->identity.revision;
50 case ETNAVIV_PARAM_GPU_FEATURES_0:
51 *value = gpu->identity.features;
54 case ETNAVIV_PARAM_GPU_FEATURES_1:
55 *value = gpu->identity.minor_features0;
58 case ETNAVIV_PARAM_GPU_FEATURES_2:
59 *value = gpu->identity.minor_features1;
62 case ETNAVIV_PARAM_GPU_FEATURES_3:
63 *value = gpu->identity.minor_features2;
66 case ETNAVIV_PARAM_GPU_FEATURES_4:
67 *value = gpu->identity.minor_features3;
70 case ETNAVIV_PARAM_GPU_FEATURES_5:
71 *value = gpu->identity.minor_features4;
74 case ETNAVIV_PARAM_GPU_FEATURES_6:
75 *value = gpu->identity.minor_features5;
78 case ETNAVIV_PARAM_GPU_FEATURES_7:
79 *value = gpu->identity.minor_features6;
82 case ETNAVIV_PARAM_GPU_FEATURES_8:
83 *value = gpu->identity.minor_features7;
86 case ETNAVIV_PARAM_GPU_FEATURES_9:
87 *value = gpu->identity.minor_features8;
90 case ETNAVIV_PARAM_GPU_FEATURES_10:
91 *value = gpu->identity.minor_features9;
94 case ETNAVIV_PARAM_GPU_FEATURES_11:
95 *value = gpu->identity.minor_features10;
98 case ETNAVIV_PARAM_GPU_FEATURES_12:
99 *value = gpu->identity.minor_features11;
102 case ETNAVIV_PARAM_GPU_STREAM_COUNT:
103 *value = gpu->identity.stream_count;
106 case ETNAVIV_PARAM_GPU_REGISTER_MAX:
107 *value = gpu->identity.register_max;
110 case ETNAVIV_PARAM_GPU_THREAD_COUNT:
111 *value = gpu->identity.thread_count;
114 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
115 *value = gpu->identity.vertex_cache_size;
118 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
119 *value = gpu->identity.shader_core_count;
122 case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
123 *value = gpu->identity.pixel_pipes;
126 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
127 *value = gpu->identity.vertex_output_buffer_size;
130 case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
131 *value = gpu->identity.buffer_size;
134 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
135 *value = gpu->identity.instruction_count;
138 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
139 *value = gpu->identity.num_constants;
142 case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
143 *value = gpu->identity.varyings_count;
147 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
155 #define etnaviv_is_model_rev(gpu, mod, rev) \
156 ((gpu)->identity.model == chipModel_##mod && \
157 (gpu)->identity.revision == rev)
158 #define etnaviv_field(val, field) \
159 (((val) & field##__MASK) >> field##__SHIFT)
161 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
163 if (gpu->identity.minor_features0 &
164 chipMinorFeatures0_MORE_MINOR_FEATURES) {
166 unsigned int streams;
168 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
169 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
170 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
171 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
173 gpu->identity.stream_count = etnaviv_field(specs[0],
174 VIVS_HI_CHIP_SPECS_STREAM_COUNT);
175 gpu->identity.register_max = etnaviv_field(specs[0],
176 VIVS_HI_CHIP_SPECS_REGISTER_MAX);
177 gpu->identity.thread_count = etnaviv_field(specs[0],
178 VIVS_HI_CHIP_SPECS_THREAD_COUNT);
179 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
180 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
181 gpu->identity.shader_core_count = etnaviv_field(specs[0],
182 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
183 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
184 VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
185 gpu->identity.vertex_output_buffer_size =
186 etnaviv_field(specs[0],
187 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
189 gpu->identity.buffer_size = etnaviv_field(specs[1],
190 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
191 gpu->identity.instruction_count = etnaviv_field(specs[1],
192 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
193 gpu->identity.num_constants = etnaviv_field(specs[1],
194 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
196 gpu->identity.varyings_count = etnaviv_field(specs[2],
197 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
199 /* This overrides the value from older register if non-zero */
200 streams = etnaviv_field(specs[3],
201 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
203 gpu->identity.stream_count = streams;
206 /* Fill in the stream count if not specified */
207 if (gpu->identity.stream_count == 0) {
208 if (gpu->identity.model >= 0x1000)
209 gpu->identity.stream_count = 4;
211 gpu->identity.stream_count = 1;
214 /* Convert the register max value */
215 if (gpu->identity.register_max)
216 gpu->identity.register_max = 1 << gpu->identity.register_max;
217 else if (gpu->identity.model == chipModel_GC400)
218 gpu->identity.register_max = 32;
220 gpu->identity.register_max = 64;
222 /* Convert thread count */
223 if (gpu->identity.thread_count)
224 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
225 else if (gpu->identity.model == chipModel_GC400)
226 gpu->identity.thread_count = 64;
227 else if (gpu->identity.model == chipModel_GC500 ||
228 gpu->identity.model == chipModel_GC530)
229 gpu->identity.thread_count = 128;
231 gpu->identity.thread_count = 256;
233 if (gpu->identity.vertex_cache_size == 0)
234 gpu->identity.vertex_cache_size = 8;
236 if (gpu->identity.shader_core_count == 0) {
237 if (gpu->identity.model >= 0x1000)
238 gpu->identity.shader_core_count = 2;
240 gpu->identity.shader_core_count = 1;
243 if (gpu->identity.pixel_pipes == 0)
244 gpu->identity.pixel_pipes = 1;
246 /* Convert virtex buffer size */
247 if (gpu->identity.vertex_output_buffer_size) {
248 gpu->identity.vertex_output_buffer_size =
249 1 << gpu->identity.vertex_output_buffer_size;
250 } else if (gpu->identity.model == chipModel_GC400) {
251 if (gpu->identity.revision < 0x4000)
252 gpu->identity.vertex_output_buffer_size = 512;
253 else if (gpu->identity.revision < 0x4200)
254 gpu->identity.vertex_output_buffer_size = 256;
256 gpu->identity.vertex_output_buffer_size = 128;
258 gpu->identity.vertex_output_buffer_size = 512;
261 switch (gpu->identity.instruction_count) {
263 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
264 gpu->identity.model == chipModel_GC880)
265 gpu->identity.instruction_count = 512;
267 gpu->identity.instruction_count = 256;
271 gpu->identity.instruction_count = 1024;
275 gpu->identity.instruction_count = 2048;
279 gpu->identity.instruction_count = 256;
283 if (gpu->identity.num_constants == 0)
284 gpu->identity.num_constants = 168;
286 if (gpu->identity.varyings_count == 0) {
287 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
288 gpu->identity.varyings_count = 12;
290 gpu->identity.varyings_count = 8;
294 * For some cores, two varyings are consumed for position, so the
295 * maximum varying count needs to be reduced by one.
297 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
298 etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
299 etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
300 etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
301 etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
302 etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
303 etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
304 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
305 etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
306 etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
307 etnaviv_is_model_rev(gpu, GC880, 0x5106))
308 gpu->identity.varyings_count -= 1;
311 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
315 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
317 /* Special case for older graphic cores. */
318 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
319 gpu->identity.model = chipModel_GC500;
320 gpu->identity.revision = etnaviv_field(chipIdentity,
321 VIVS_HI_CHIP_IDENTITY_REVISION);
324 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
325 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
328 * !!!! HACK ALERT !!!!
329 * Because people change device IDs without letting software
330 * know about it - here is the hack to make it all look the
331 * same. Only for GC400 family.
333 if ((gpu->identity.model & 0xff00) == 0x0400 &&
334 gpu->identity.model != chipModel_GC420) {
335 gpu->identity.model = gpu->identity.model & 0x0400;
338 /* Another special case */
339 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
340 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
341 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
343 if (chipDate == 0x20080814 && chipTime == 0x12051100) {
345 * This IP has an ECO; put the correct
348 gpu->identity.revision = 0x1051;
353 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
354 * reality it's just a re-branded GC3000. We can identify this
355 * core by the upper half of the revision register being all 1.
356 * Fix model/rev here, so all other places can refer to this
357 * core by its real identity.
359 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
360 gpu->identity.model = chipModel_GC3000;
361 gpu->identity.revision &= 0xffff;
365 dev_info(gpu->dev, "model: GC%x, revision: %x\n",
366 gpu->identity.model, gpu->identity.revision);
369 * If there is a match in the HWDB, we aren't interested in the
370 * remaining register values, as they might be wrong.
372 if (etnaviv_fill_identity_from_hwdb(gpu))
375 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
377 /* Disable fast clear on GC700. */
378 if (gpu->identity.model == chipModel_GC700)
379 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
381 if ((gpu->identity.model == chipModel_GC500 &&
382 gpu->identity.revision < 2) ||
383 (gpu->identity.model == chipModel_GC300 &&
384 gpu->identity.revision < 0x2000)) {
387 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
390 gpu->identity.minor_features0 = 0;
391 gpu->identity.minor_features1 = 0;
392 gpu->identity.minor_features2 = 0;
393 gpu->identity.minor_features3 = 0;
394 gpu->identity.minor_features4 = 0;
395 gpu->identity.minor_features5 = 0;
397 gpu->identity.minor_features0 =
398 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
400 if (gpu->identity.minor_features0 &
401 chipMinorFeatures0_MORE_MINOR_FEATURES) {
402 gpu->identity.minor_features1 =
403 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
404 gpu->identity.minor_features2 =
405 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
406 gpu->identity.minor_features3 =
407 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
408 gpu->identity.minor_features4 =
409 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
410 gpu->identity.minor_features5 =
411 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
414 /* GC600 idle register reports zero bits where modules aren't present */
415 if (gpu->identity.model == chipModel_GC600) {
416 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
417 VIVS_HI_IDLE_STATE_RA |
418 VIVS_HI_IDLE_STATE_SE |
419 VIVS_HI_IDLE_STATE_PA |
420 VIVS_HI_IDLE_STATE_SH |
421 VIVS_HI_IDLE_STATE_PE |
422 VIVS_HI_IDLE_STATE_DE |
423 VIVS_HI_IDLE_STATE_FE;
425 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
428 etnaviv_hw_specs(gpu);
431 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
433 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
434 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
435 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
438 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
440 if (gpu->identity.minor_features2 &
441 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
442 clk_set_rate(gpu->clk_core,
443 gpu->base_rate_core >> gpu->freq_scale);
444 clk_set_rate(gpu->clk_shader,
445 gpu->base_rate_shader >> gpu->freq_scale);
447 unsigned int fscale = 1 << (6 - gpu->freq_scale);
448 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
450 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
451 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
452 etnaviv_gpu_load_clock(gpu, clock);
456 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
459 unsigned long timeout;
462 /* We hope that the GPU resets in under one second */
463 timeout = jiffies + msecs_to_jiffies(1000);
465 while (time_is_after_jiffies(timeout)) {
467 unsigned int fscale = 1 << (6 - gpu->freq_scale);
468 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
469 etnaviv_gpu_load_clock(gpu, control);
471 /* isolate the GPU. */
472 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
473 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
475 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
476 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
477 VIVS_MMUv2_AHB_CONTROL_RESET);
479 /* set soft reset. */
480 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
481 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
484 /* wait for reset. */
485 usleep_range(10, 20);
487 /* reset soft reset bit. */
488 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
489 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
491 /* reset GPU isolation. */
492 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
493 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
495 /* read idle register. */
496 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
498 /* try reseting again if FE it not idle */
499 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
500 dev_dbg(gpu->dev, "FE is not idle\n");
504 /* read reset register. */
505 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
507 /* is the GPU idle? */
508 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
509 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
510 dev_dbg(gpu->dev, "GPU is not idle\n");
514 /* disable debug registers, as they are not normally needed */
515 control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
516 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
523 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
524 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
526 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
527 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
528 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
529 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
534 /* We rely on the GPU running, so program the clock */
535 etnaviv_gpu_update_clock(gpu);
540 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
544 /* enable clock gating */
545 ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
546 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
548 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
549 if (gpu->identity.revision == 0x4301 ||
550 gpu->identity.revision == 0x4302)
551 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
553 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
555 pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
557 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
558 if (gpu->identity.model >= chipModel_GC400 &&
559 gpu->identity.model != chipModel_GC420 &&
560 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
561 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
564 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
565 * present without a bug fix.
567 if (gpu->identity.revision < 0x5000 &&
568 gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
569 !(gpu->identity.minor_features1 &
570 chipMinorFeatures1_DISABLE_PE_GATING))
571 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
573 if (gpu->identity.revision < 0x5422)
574 pmc |= BIT(15); /* Unknown bit */
576 /* Disable TX clock gating on affected core revisions. */
577 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
578 etnaviv_is_model_rev(gpu, GC2000, 0x5108))
579 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
581 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
582 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
584 gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
587 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
589 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
590 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
591 VIVS_FE_COMMAND_CONTROL_ENABLE |
592 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
594 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
595 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
596 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
597 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
601 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
604 * Base value for VIVS_PM_PULSE_EATER register on models where it
605 * cannot be read, extracted from vivante kernel driver.
607 u32 pulse_eater = 0x01590880;
609 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
610 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
611 pulse_eater |= BIT(23);
615 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
616 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
617 pulse_eater &= ~BIT(16);
618 pulse_eater |= BIT(17);
621 if ((gpu->identity.revision > 0x5420) &&
622 (gpu->identity.features & chipFeatures_PIPE_3D))
624 /* Performance fix: disable internal DFS */
625 pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
626 pulse_eater |= BIT(18);
629 gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
632 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
636 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
637 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
638 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
641 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
643 if (gpu->identity.revision == 0x5007)
644 mc_memory_debug |= 0x0c;
646 mc_memory_debug |= 0x08;
648 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
651 /* enable module-level clock gating */
652 etnaviv_gpu_enable_mlcg(gpu);
655 * Update GPU AXI cache atttribute to "cacheable, no allocate".
656 * This is necessary to prevent the iMX6 SoC locking up.
658 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
659 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
660 VIVS_HI_AXI_CONFIG_ARCACHE(2));
662 /* GC2000 rev 5108 needs a special bus config */
663 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
664 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
665 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
666 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
667 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
668 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
669 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
672 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
673 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
674 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
675 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
678 /* setup the pulse eater */
679 etnaviv_gpu_setup_pulse_eater(gpu);
682 etnaviv_iommu_restore(gpu);
684 /* Start command processor */
685 prefetch = etnaviv_buffer_init(gpu);
687 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
688 etnaviv_gpu_start_fe(gpu, etnaviv_cmdbuf_get_va(&gpu->buffer),
692 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
696 ret = pm_runtime_get_sync(gpu->dev);
698 dev_err(gpu->dev, "Failed to enable GPU power domain\n");
702 etnaviv_hw_identify(gpu);
704 if (gpu->identity.model == 0) {
705 dev_err(gpu->dev, "Unknown GPU model\n");
710 /* Exclude VG cores with FE2.0 */
711 if (gpu->identity.features & chipFeatures_PIPE_VG &&
712 gpu->identity.features & chipFeatures_FE20) {
713 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
719 * Set the GPU linear window to be at the end of the DMA window, where
720 * the CMA area is likely to reside. This ensures that we are able to
721 * map the command buffers while having the linear window overlap as
722 * much RAM as possible, so we can optimize mappings for other buffers.
724 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
725 * to different views of the memory on the individual engines.
727 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
728 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
729 u32 dma_mask = (u32)dma_get_required_mask(gpu->dev);
730 if (dma_mask < PHYS_OFFSET + SZ_2G)
731 gpu->memory_base = PHYS_OFFSET;
733 gpu->memory_base = dma_mask - SZ_2G + 1;
734 } else if (PHYS_OFFSET >= SZ_2G) {
735 dev_info(gpu->dev, "Need to move linear window on MC1.0, disabling TS\n");
736 gpu->memory_base = PHYS_OFFSET;
737 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
741 * On cores with security features supported, we claim control over the
744 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
745 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
746 gpu->sec_mode = ETNA_SEC_KERNEL;
748 ret = etnaviv_hw_reset(gpu);
750 dev_err(gpu->dev, "GPU reset failed\n");
754 gpu->mmu = etnaviv_iommu_new(gpu);
755 if (IS_ERR(gpu->mmu)) {
756 dev_err(gpu->dev, "Failed to instantiate GPU IOMMU\n");
757 ret = PTR_ERR(gpu->mmu);
761 gpu->cmdbuf_suballoc = etnaviv_cmdbuf_suballoc_new(gpu);
762 if (IS_ERR(gpu->cmdbuf_suballoc)) {
763 dev_err(gpu->dev, "Failed to create cmdbuf suballocator\n");
764 ret = PTR_ERR(gpu->cmdbuf_suballoc);
769 ret = etnaviv_cmdbuf_init(gpu->cmdbuf_suballoc, &gpu->buffer,
772 dev_err(gpu->dev, "could not create command buffer\n");
776 if (gpu->mmu->version == ETNAVIV_IOMMU_V1 &&
777 etnaviv_cmdbuf_get_va(&gpu->buffer) > 0x80000000) {
780 "command buffer outside valid memory window\n");
784 /* Setup event management */
785 spin_lock_init(&gpu->event_spinlock);
786 init_completion(&gpu->event_free);
787 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
788 for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
789 complete(&gpu->event_free);
791 /* Now program the hardware */
792 mutex_lock(&gpu->lock);
793 etnaviv_gpu_hw_init(gpu);
794 gpu->exec_state = -1;
795 mutex_unlock(&gpu->lock);
797 pm_runtime_mark_last_busy(gpu->dev);
798 pm_runtime_put_autosuspend(gpu->dev);
803 etnaviv_cmdbuf_free(&gpu->buffer);
804 gpu->buffer.suballoc = NULL;
806 etnaviv_iommu_destroy(gpu->mmu);
809 pm_runtime_mark_last_busy(gpu->dev);
810 pm_runtime_put_autosuspend(gpu->dev);
815 #ifdef CONFIG_DEBUG_FS
821 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
825 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
826 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
828 for (i = 0; i < 500; i++) {
829 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
830 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
832 if (debug->address[0] != debug->address[1])
835 if (debug->state[0] != debug->state[1])
840 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
842 struct dma_debug debug;
843 u32 dma_lo, dma_hi, axi, idle;
846 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
848 ret = pm_runtime_get_sync(gpu->dev);
852 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
853 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
854 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
855 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
857 verify_dma(gpu, &debug);
859 seq_puts(m, "\tfeatures\n");
860 seq_printf(m, "\t major_features: 0x%08x\n",
861 gpu->identity.features);
862 seq_printf(m, "\t minor_features0: 0x%08x\n",
863 gpu->identity.minor_features0);
864 seq_printf(m, "\t minor_features1: 0x%08x\n",
865 gpu->identity.minor_features1);
866 seq_printf(m, "\t minor_features2: 0x%08x\n",
867 gpu->identity.minor_features2);
868 seq_printf(m, "\t minor_features3: 0x%08x\n",
869 gpu->identity.minor_features3);
870 seq_printf(m, "\t minor_features4: 0x%08x\n",
871 gpu->identity.minor_features4);
872 seq_printf(m, "\t minor_features5: 0x%08x\n",
873 gpu->identity.minor_features5);
874 seq_printf(m, "\t minor_features6: 0x%08x\n",
875 gpu->identity.minor_features6);
876 seq_printf(m, "\t minor_features7: 0x%08x\n",
877 gpu->identity.minor_features7);
878 seq_printf(m, "\t minor_features8: 0x%08x\n",
879 gpu->identity.minor_features8);
880 seq_printf(m, "\t minor_features9: 0x%08x\n",
881 gpu->identity.minor_features9);
882 seq_printf(m, "\t minor_features10: 0x%08x\n",
883 gpu->identity.minor_features10);
884 seq_printf(m, "\t minor_features11: 0x%08x\n",
885 gpu->identity.minor_features11);
887 seq_puts(m, "\tspecs\n");
888 seq_printf(m, "\t stream_count: %d\n",
889 gpu->identity.stream_count);
890 seq_printf(m, "\t register_max: %d\n",
891 gpu->identity.register_max);
892 seq_printf(m, "\t thread_count: %d\n",
893 gpu->identity.thread_count);
894 seq_printf(m, "\t vertex_cache_size: %d\n",
895 gpu->identity.vertex_cache_size);
896 seq_printf(m, "\t shader_core_count: %d\n",
897 gpu->identity.shader_core_count);
898 seq_printf(m, "\t pixel_pipes: %d\n",
899 gpu->identity.pixel_pipes);
900 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
901 gpu->identity.vertex_output_buffer_size);
902 seq_printf(m, "\t buffer_size: %d\n",
903 gpu->identity.buffer_size);
904 seq_printf(m, "\t instruction_count: %d\n",
905 gpu->identity.instruction_count);
906 seq_printf(m, "\t num_constants: %d\n",
907 gpu->identity.num_constants);
908 seq_printf(m, "\t varyings_count: %d\n",
909 gpu->identity.varyings_count);
911 seq_printf(m, "\taxi: 0x%08x\n", axi);
912 seq_printf(m, "\tidle: 0x%08x\n", idle);
913 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
914 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
915 seq_puts(m, "\t FE is not idle\n");
916 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
917 seq_puts(m, "\t DE is not idle\n");
918 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
919 seq_puts(m, "\t PE is not idle\n");
920 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
921 seq_puts(m, "\t SH is not idle\n");
922 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
923 seq_puts(m, "\t PA is not idle\n");
924 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
925 seq_puts(m, "\t SE is not idle\n");
926 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
927 seq_puts(m, "\t RA is not idle\n");
928 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
929 seq_puts(m, "\t TX is not idle\n");
930 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
931 seq_puts(m, "\t VG is not idle\n");
932 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
933 seq_puts(m, "\t IM is not idle\n");
934 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
935 seq_puts(m, "\t FP is not idle\n");
936 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
937 seq_puts(m, "\t TS is not idle\n");
938 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
939 seq_puts(m, "\t AXI low power mode\n");
941 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
942 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
943 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
944 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
946 seq_puts(m, "\tMC\n");
947 seq_printf(m, "\t read0: 0x%08x\n", read0);
948 seq_printf(m, "\t read1: 0x%08x\n", read1);
949 seq_printf(m, "\t write: 0x%08x\n", write);
952 seq_puts(m, "\tDMA ");
954 if (debug.address[0] == debug.address[1] &&
955 debug.state[0] == debug.state[1]) {
956 seq_puts(m, "seems to be stuck\n");
957 } else if (debug.address[0] == debug.address[1]) {
958 seq_puts(m, "address is constant\n");
960 seq_puts(m, "is running\n");
963 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
964 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
965 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
966 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
967 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
972 pm_runtime_mark_last_busy(gpu->dev);
973 pm_runtime_put_autosuspend(gpu->dev);
979 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
983 dev_err(gpu->dev, "recover hung GPU!\n");
985 if (pm_runtime_get_sync(gpu->dev) < 0)
988 mutex_lock(&gpu->lock);
990 etnaviv_hw_reset(gpu);
992 /* complete all events, the GPU won't do it after the reset */
993 spin_lock(&gpu->event_spinlock);
994 for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
995 complete(&gpu->event_free);
996 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
997 spin_unlock(&gpu->event_spinlock);
999 etnaviv_gpu_hw_init(gpu);
1000 gpu->exec_state = -1;
1002 mutex_unlock(&gpu->lock);
1003 pm_runtime_mark_last_busy(gpu->dev);
1004 pm_runtime_put_autosuspend(gpu->dev);
1007 /* fence object management */
1008 struct etnaviv_fence {
1009 struct etnaviv_gpu *gpu;
1010 struct dma_fence base;
1013 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1015 return container_of(fence, struct etnaviv_fence, base);
1018 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1023 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1025 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1027 return dev_name(f->gpu->dev);
1030 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1032 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1034 return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1037 static void etnaviv_fence_release(struct dma_fence *fence)
1039 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1041 kfree_rcu(f, base.rcu);
1044 static const struct dma_fence_ops etnaviv_fence_ops = {
1045 .get_driver_name = etnaviv_fence_get_driver_name,
1046 .get_timeline_name = etnaviv_fence_get_timeline_name,
1047 .signaled = etnaviv_fence_signaled,
1048 .release = etnaviv_fence_release,
1051 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1053 struct etnaviv_fence *f;
1056 * GPU lock must already be held, otherwise fence completion order might
1057 * not match the seqno order assigned here.
1059 lockdep_assert_held(&gpu->lock);
1061 f = kzalloc(sizeof(*f), GFP_KERNEL);
1067 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1068 gpu->fence_context, ++gpu->next_fence);
1073 /* returns true if fence a comes after fence b */
1074 static inline bool fence_after(u32 a, u32 b)
1076 return (s32)(a - b) > 0;
1083 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1084 unsigned int *events)
1086 unsigned long timeout = msecs_to_jiffies(10 * 10000);
1087 unsigned i, acquired = 0;
1089 for (i = 0; i < nr_events; i++) {
1092 ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1095 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1103 spin_lock(&gpu->event_spinlock);
1105 for (i = 0; i < nr_events; i++) {
1106 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1109 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1110 set_bit(event, gpu->event_bitmap);
1113 spin_unlock(&gpu->event_spinlock);
1118 for (i = 0; i < acquired; i++)
1119 complete(&gpu->event_free);
1124 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1126 if (!test_bit(event, gpu->event_bitmap)) {
1127 dev_warn(gpu->dev, "event %u is already marked as free",
1130 clear_bit(event, gpu->event_bitmap);
1131 complete(&gpu->event_free);
1136 * Cmdstream submission/retirement:
1138 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1139 u32 id, struct timespec *timeout)
1141 struct dma_fence *fence;
1145 * Look up the fence and take a reference. We might still find a fence
1146 * whose refcount has already dropped to zero. dma_fence_get_rcu
1147 * pretends we didn't find a fence in that case.
1150 fence = idr_find(&gpu->fence_idr, id);
1152 fence = dma_fence_get_rcu(fence);
1159 /* No timeout was requested: just test for completion */
1160 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1162 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1164 ret = dma_fence_wait_timeout(fence, true, remaining);
1167 else if (ret != -ERESTARTSYS)
1172 dma_fence_put(fence);
1177 * Wait for an object to become inactive. This, on it's own, is not race
1178 * free: the object is moved by the scheduler off the active list, and
1179 * then the iova is put. Moreover, the object could be re-submitted just
1180 * after we notice that it's become inactive.
1182 * Although the retirement happens under the gpu lock, we don't want to hold
1183 * that lock in this function while waiting.
1185 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1186 struct etnaviv_gem_object *etnaviv_obj, struct timespec *timeout)
1188 unsigned long remaining;
1192 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1194 remaining = etnaviv_timeout_to_jiffies(timeout);
1196 ret = wait_event_interruptible_timeout(gpu->fence_event,
1197 !is_active(etnaviv_obj),
1201 else if (ret == -ERESTARTSYS)
1202 return -ERESTARTSYS;
1207 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1208 struct etnaviv_event *event, unsigned int flags)
1210 const struct etnaviv_gem_submit *submit = event->submit;
1213 for (i = 0; i < submit->nr_pmrs; i++) {
1214 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1216 if (pmr->flags == flags)
1217 etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1221 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1222 struct etnaviv_event *event)
1226 /* disable clock gating */
1227 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1228 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1229 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1231 /* enable debug register */
1232 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1233 val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1234 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1236 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1239 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1240 struct etnaviv_event *event)
1242 const struct etnaviv_gem_submit *submit = event->submit;
1246 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1248 for (i = 0; i < submit->nr_pmrs; i++) {
1249 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1251 *pmr->bo_vma = pmr->sequence;
1254 /* disable debug register */
1255 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1256 val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1257 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1259 /* enable clock gating */
1260 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1261 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1262 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1266 /* add bo's to gpu's ring, and kick gpu: */
1267 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1269 struct etnaviv_gpu *gpu = submit->gpu;
1270 struct dma_fence *gpu_fence;
1271 unsigned int i, nr_events = 1, event[3];
1274 if (!submit->runtime_resumed) {
1275 ret = pm_runtime_get_sync(gpu->dev);
1278 submit->runtime_resumed = true;
1282 * if there are performance monitor requests we need to have
1283 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1285 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1286 * and update the sequence number for userspace.
1288 if (submit->nr_pmrs)
1291 ret = event_alloc(gpu, nr_events, event);
1293 DRM_ERROR("no free events\n");
1297 mutex_lock(&gpu->lock);
1299 gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1301 for (i = 0; i < nr_events; i++)
1302 event_free(gpu, event[i]);
1307 if (submit->nr_pmrs) {
1308 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1309 kref_get(&submit->refcount);
1310 gpu->event[event[1]].submit = submit;
1311 etnaviv_sync_point_queue(gpu, event[1]);
1314 gpu->event[event[0]].fence = gpu_fence;
1315 submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1316 etnaviv_buffer_queue(gpu, submit->exec_state, event[0],
1319 if (submit->nr_pmrs) {
1320 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1321 kref_get(&submit->refcount);
1322 gpu->event[event[2]].submit = submit;
1323 etnaviv_sync_point_queue(gpu, event[2]);
1327 mutex_unlock(&gpu->lock);
1332 static void sync_point_worker(struct work_struct *work)
1334 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1336 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1337 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1339 event->sync_point(gpu, event);
1340 etnaviv_submit_put(event->submit);
1341 event_free(gpu, gpu->sync_point_event);
1343 /* restart FE last to avoid GPU and IRQ racing against this worker */
1344 etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1347 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1349 u32 status_reg, status;
1352 if (gpu->sec_mode == ETNA_SEC_NONE)
1353 status_reg = VIVS_MMUv2_STATUS;
1355 status_reg = VIVS_MMUv2_SEC_STATUS;
1357 status = gpu_read(gpu, status_reg);
1358 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1360 for (i = 0; i < 4; i++) {
1363 if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1366 if (gpu->sec_mode == ETNA_SEC_NONE)
1367 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1369 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1371 dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1372 gpu_read(gpu, address_reg));
1376 static irqreturn_t irq_handler(int irq, void *data)
1378 struct etnaviv_gpu *gpu = data;
1379 irqreturn_t ret = IRQ_NONE;
1381 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1386 pm_runtime_mark_last_busy(gpu->dev);
1388 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1390 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1391 dev_err(gpu->dev, "AXI bus error\n");
1392 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1395 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1396 dump_mmu_fault(gpu);
1397 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1400 while ((event = ffs(intr)) != 0) {
1401 struct dma_fence *fence;
1405 intr &= ~(1 << event);
1407 dev_dbg(gpu->dev, "event %u\n", event);
1409 if (gpu->event[event].sync_point) {
1410 gpu->sync_point_event = event;
1411 queue_work(gpu->wq, &gpu->sync_point_work);
1414 fence = gpu->event[event].fence;
1418 gpu->event[event].fence = NULL;
1421 * Events can be processed out of order. Eg,
1422 * - allocate and queue event 0
1423 * - allocate event 1
1424 * - event 0 completes, we process it
1425 * - allocate and queue event 0
1426 * - event 1 and event 0 complete
1427 * we can end up processing event 0 first, then 1.
1429 if (fence_after(fence->seqno, gpu->completed_fence))
1430 gpu->completed_fence = fence->seqno;
1431 dma_fence_signal(fence);
1433 event_free(gpu, event);
1442 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1447 ret = clk_prepare_enable(gpu->clk_reg);
1453 ret = clk_prepare_enable(gpu->clk_bus);
1458 if (gpu->clk_core) {
1459 ret = clk_prepare_enable(gpu->clk_core);
1461 goto disable_clk_bus;
1464 if (gpu->clk_shader) {
1465 ret = clk_prepare_enable(gpu->clk_shader);
1467 goto disable_clk_core;
1474 clk_disable_unprepare(gpu->clk_core);
1477 clk_disable_unprepare(gpu->clk_bus);
1482 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1484 if (gpu->clk_shader)
1485 clk_disable_unprepare(gpu->clk_shader);
1487 clk_disable_unprepare(gpu->clk_core);
1489 clk_disable_unprepare(gpu->clk_bus);
1491 clk_disable_unprepare(gpu->clk_reg);
1496 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1498 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1501 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1503 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1506 if (time_is_before_jiffies(timeout)) {
1508 "timed out waiting for idle: idle=0x%x\n",
1517 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1519 if (gpu->buffer.suballoc) {
1520 /* Replace the last WAIT with END */
1521 mutex_lock(&gpu->lock);
1522 etnaviv_buffer_end(gpu);
1523 mutex_unlock(&gpu->lock);
1526 * We know that only the FE is busy here, this should
1527 * happen quickly (as the WAIT is only 200 cycles). If
1528 * we fail, just warn and continue.
1530 etnaviv_gpu_wait_idle(gpu, 100);
1533 return etnaviv_gpu_clk_disable(gpu);
1537 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1541 ret = mutex_lock_killable(&gpu->lock);
1545 etnaviv_gpu_update_clock(gpu);
1546 etnaviv_gpu_hw_init(gpu);
1548 gpu->exec_state = -1;
1550 mutex_unlock(&gpu->lock);
1557 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1558 unsigned long *state)
1566 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1567 unsigned long *state)
1569 struct etnaviv_gpu *gpu = cdev->devdata;
1571 *state = gpu->freq_scale;
1577 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1578 unsigned long state)
1580 struct etnaviv_gpu *gpu = cdev->devdata;
1582 mutex_lock(&gpu->lock);
1583 gpu->freq_scale = state;
1584 if (!pm_runtime_suspended(gpu->dev))
1585 etnaviv_gpu_update_clock(gpu);
1586 mutex_unlock(&gpu->lock);
1591 static struct thermal_cooling_device_ops cooling_ops = {
1592 .get_max_state = etnaviv_gpu_cooling_get_max_state,
1593 .get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1594 .set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1597 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1600 struct drm_device *drm = data;
1601 struct etnaviv_drm_private *priv = drm->dev_private;
1602 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1605 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1606 gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1607 (char *)dev_name(dev), gpu, &cooling_ops);
1608 if (IS_ERR(gpu->cooling))
1609 return PTR_ERR(gpu->cooling);
1612 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1618 ret = etnaviv_sched_init(gpu);
1623 ret = pm_runtime_get_sync(gpu->dev);
1625 ret = etnaviv_gpu_clk_enable(gpu);
1632 gpu->fence_context = dma_fence_context_alloc(1);
1633 idr_init(&gpu->fence_idr);
1634 spin_lock_init(&gpu->fence_spinlock);
1636 INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1637 init_waitqueue_head(&gpu->fence_event);
1639 priv->gpu[priv->num_gpus++] = gpu;
1641 pm_runtime_mark_last_busy(gpu->dev);
1642 pm_runtime_put_autosuspend(gpu->dev);
1647 etnaviv_sched_fini(gpu);
1650 destroy_workqueue(gpu->wq);
1653 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1654 thermal_cooling_device_unregister(gpu->cooling);
1659 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1662 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1664 DBG("%s", dev_name(gpu->dev));
1666 flush_workqueue(gpu->wq);
1667 destroy_workqueue(gpu->wq);
1669 etnaviv_sched_fini(gpu);
1672 pm_runtime_get_sync(gpu->dev);
1673 pm_runtime_put_sync_suspend(gpu->dev);
1675 etnaviv_gpu_hw_suspend(gpu);
1678 if (gpu->buffer.suballoc)
1679 etnaviv_cmdbuf_free(&gpu->buffer);
1681 if (gpu->cmdbuf_suballoc) {
1682 etnaviv_cmdbuf_suballoc_destroy(gpu->cmdbuf_suballoc);
1683 gpu->cmdbuf_suballoc = NULL;
1687 etnaviv_iommu_destroy(gpu->mmu);
1692 idr_destroy(&gpu->fence_idr);
1694 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1695 thermal_cooling_device_unregister(gpu->cooling);
1696 gpu->cooling = NULL;
1699 static const struct component_ops gpu_ops = {
1700 .bind = etnaviv_gpu_bind,
1701 .unbind = etnaviv_gpu_unbind,
1704 static const struct of_device_id etnaviv_gpu_match[] = {
1706 .compatible = "vivante,gc"
1710 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1712 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1714 struct device *dev = &pdev->dev;
1715 struct etnaviv_gpu *gpu;
1716 struct resource *res;
1719 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1723 gpu->dev = &pdev->dev;
1724 mutex_init(&gpu->lock);
1725 mutex_init(&gpu->fence_lock);
1727 /* Map registers: */
1728 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1729 gpu->mmio = devm_ioremap_resource(&pdev->dev, res);
1730 if (IS_ERR(gpu->mmio))
1731 return PTR_ERR(gpu->mmio);
1733 /* Get Interrupt: */
1734 gpu->irq = platform_get_irq(pdev, 0);
1736 dev_err(dev, "failed to get irq: %d\n", gpu->irq);
1740 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1741 dev_name(gpu->dev), gpu);
1743 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
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);
1784 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1791 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1793 component_del(&pdev->dev, &gpu_ops);
1794 pm_runtime_disable(&pdev->dev);
1799 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1801 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1804 /* If there are any jobs in the HW queue, we're not idle */
1805 if (atomic_read(&gpu->sched.hw_rq_count))
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;
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);
1822 ret = etnaviv_gpu_clk_enable(gpu);
1826 /* Re-initialise the basic hardware state */
1827 if (gpu->drm && gpu->buffer.suballoc) {
1828 ret = etnaviv_gpu_hw_resume(gpu);
1830 etnaviv_gpu_clk_disable(gpu);
1839 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1840 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1844 struct platform_driver etnaviv_gpu_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,