1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015-2018 Etnaviv Project
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/mod_devicetable.h>
12 #include <linux/module.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 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)
41 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
44 case ETNAVIV_PARAM_GPU_MODEL:
45 *value = gpu->identity.model;
48 case ETNAVIV_PARAM_GPU_REVISION:
49 *value = gpu->identity.revision;
52 case ETNAVIV_PARAM_GPU_FEATURES_0:
53 *value = gpu->identity.features;
56 case ETNAVIV_PARAM_GPU_FEATURES_1:
57 *value = gpu->identity.minor_features0;
60 case ETNAVIV_PARAM_GPU_FEATURES_2:
61 *value = gpu->identity.minor_features1;
64 case ETNAVIV_PARAM_GPU_FEATURES_3:
65 *value = gpu->identity.minor_features2;
68 case ETNAVIV_PARAM_GPU_FEATURES_4:
69 *value = gpu->identity.minor_features3;
72 case ETNAVIV_PARAM_GPU_FEATURES_5:
73 *value = gpu->identity.minor_features4;
76 case ETNAVIV_PARAM_GPU_FEATURES_6:
77 *value = gpu->identity.minor_features5;
80 case ETNAVIV_PARAM_GPU_FEATURES_7:
81 *value = gpu->identity.minor_features6;
84 case ETNAVIV_PARAM_GPU_FEATURES_8:
85 *value = gpu->identity.minor_features7;
88 case ETNAVIV_PARAM_GPU_FEATURES_9:
89 *value = gpu->identity.minor_features8;
92 case ETNAVIV_PARAM_GPU_FEATURES_10:
93 *value = gpu->identity.minor_features9;
96 case ETNAVIV_PARAM_GPU_FEATURES_11:
97 *value = gpu->identity.minor_features10;
100 case ETNAVIV_PARAM_GPU_FEATURES_12:
101 *value = gpu->identity.minor_features11;
104 case ETNAVIV_PARAM_GPU_STREAM_COUNT:
105 *value = gpu->identity.stream_count;
108 case ETNAVIV_PARAM_GPU_REGISTER_MAX:
109 *value = gpu->identity.register_max;
112 case ETNAVIV_PARAM_GPU_THREAD_COUNT:
113 *value = gpu->identity.thread_count;
116 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE:
117 *value = gpu->identity.vertex_cache_size;
120 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT:
121 *value = gpu->identity.shader_core_count;
124 case ETNAVIV_PARAM_GPU_PIXEL_PIPES:
125 *value = gpu->identity.pixel_pipes;
128 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE:
129 *value = gpu->identity.vertex_output_buffer_size;
132 case ETNAVIV_PARAM_GPU_BUFFER_SIZE:
133 *value = gpu->identity.buffer_size;
136 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT:
137 *value = gpu->identity.instruction_count;
140 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS:
141 *value = gpu->identity.num_constants;
144 case ETNAVIV_PARAM_GPU_NUM_VARYINGS:
145 *value = gpu->identity.varyings_count;
148 case ETNAVIV_PARAM_SOFTPIN_START_ADDR:
149 if (priv->mmu_global->version == ETNAVIV_IOMMU_V2)
150 *value = ETNAVIV_SOFTPIN_START_ADDRESS;
155 case ETNAVIV_PARAM_GPU_PRODUCT_ID:
156 *value = gpu->identity.product_id;
159 case ETNAVIV_PARAM_GPU_CUSTOMER_ID:
160 *value = gpu->identity.customer_id;
163 case ETNAVIV_PARAM_GPU_ECO_ID:
164 *value = gpu->identity.eco_id;
167 case ETNAVIV_PARAM_GPU_NN_CORE_COUNT:
168 *value = gpu->identity.nn_core_count;
171 case ETNAVIV_PARAM_GPU_NN_MAD_PER_CORE:
172 *value = gpu->identity.nn_mad_per_core;
175 case ETNAVIV_PARAM_GPU_TP_CORE_COUNT:
176 *value = gpu->identity.tp_core_count;
179 case ETNAVIV_PARAM_GPU_ON_CHIP_SRAM_SIZE:
180 *value = gpu->identity.on_chip_sram_size;
183 case ETNAVIV_PARAM_GPU_AXI_SRAM_SIZE:
184 *value = gpu->identity.axi_sram_size;
188 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
196 #define etnaviv_is_model_rev(gpu, mod, rev) \
197 ((gpu)->identity.model == chipModel_##mod && \
198 (gpu)->identity.revision == rev)
199 #define etnaviv_field(val, field) \
200 (((val) & field##__MASK) >> field##__SHIFT)
202 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
204 if (gpu->identity.minor_features0 &
205 chipMinorFeatures0_MORE_MINOR_FEATURES) {
207 unsigned int streams;
209 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
210 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
211 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
212 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
214 gpu->identity.stream_count = etnaviv_field(specs[0],
215 VIVS_HI_CHIP_SPECS_STREAM_COUNT);
216 gpu->identity.register_max = etnaviv_field(specs[0],
217 VIVS_HI_CHIP_SPECS_REGISTER_MAX);
218 gpu->identity.thread_count = etnaviv_field(specs[0],
219 VIVS_HI_CHIP_SPECS_THREAD_COUNT);
220 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
221 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
222 gpu->identity.shader_core_count = etnaviv_field(specs[0],
223 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
224 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
225 VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
226 gpu->identity.vertex_output_buffer_size =
227 etnaviv_field(specs[0],
228 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
230 gpu->identity.buffer_size = etnaviv_field(specs[1],
231 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
232 gpu->identity.instruction_count = etnaviv_field(specs[1],
233 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
234 gpu->identity.num_constants = etnaviv_field(specs[1],
235 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
237 gpu->identity.varyings_count = etnaviv_field(specs[2],
238 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
240 /* This overrides the value from older register if non-zero */
241 streams = etnaviv_field(specs[3],
242 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
244 gpu->identity.stream_count = streams;
247 /* Fill in the stream count if not specified */
248 if (gpu->identity.stream_count == 0) {
249 if (gpu->identity.model >= 0x1000)
250 gpu->identity.stream_count = 4;
252 gpu->identity.stream_count = 1;
255 /* Convert the register max value */
256 if (gpu->identity.register_max)
257 gpu->identity.register_max = 1 << gpu->identity.register_max;
258 else if (gpu->identity.model == chipModel_GC400)
259 gpu->identity.register_max = 32;
261 gpu->identity.register_max = 64;
263 /* Convert thread count */
264 if (gpu->identity.thread_count)
265 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
266 else if (gpu->identity.model == chipModel_GC400)
267 gpu->identity.thread_count = 64;
268 else if (gpu->identity.model == chipModel_GC500 ||
269 gpu->identity.model == chipModel_GC530)
270 gpu->identity.thread_count = 128;
272 gpu->identity.thread_count = 256;
274 if (gpu->identity.vertex_cache_size == 0)
275 gpu->identity.vertex_cache_size = 8;
277 if (gpu->identity.shader_core_count == 0) {
278 if (gpu->identity.model >= 0x1000)
279 gpu->identity.shader_core_count = 2;
281 gpu->identity.shader_core_count = 1;
284 if (gpu->identity.pixel_pipes == 0)
285 gpu->identity.pixel_pipes = 1;
287 /* Convert virtex buffer size */
288 if (gpu->identity.vertex_output_buffer_size) {
289 gpu->identity.vertex_output_buffer_size =
290 1 << gpu->identity.vertex_output_buffer_size;
291 } else if (gpu->identity.model == chipModel_GC400) {
292 if (gpu->identity.revision < 0x4000)
293 gpu->identity.vertex_output_buffer_size = 512;
294 else if (gpu->identity.revision < 0x4200)
295 gpu->identity.vertex_output_buffer_size = 256;
297 gpu->identity.vertex_output_buffer_size = 128;
299 gpu->identity.vertex_output_buffer_size = 512;
302 switch (gpu->identity.instruction_count) {
304 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
305 gpu->identity.model == chipModel_GC880)
306 gpu->identity.instruction_count = 512;
308 gpu->identity.instruction_count = 256;
312 gpu->identity.instruction_count = 1024;
316 gpu->identity.instruction_count = 2048;
320 gpu->identity.instruction_count = 256;
324 if (gpu->identity.num_constants == 0)
325 gpu->identity.num_constants = 168;
327 if (gpu->identity.varyings_count == 0) {
328 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
329 gpu->identity.varyings_count = 12;
331 gpu->identity.varyings_count = 8;
335 * For some cores, two varyings are consumed for position, so the
336 * maximum varying count needs to be reduced by one.
338 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
339 etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
340 etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
341 etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
342 etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
343 etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
344 etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
345 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
346 etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
347 etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
348 etnaviv_is_model_rev(gpu, GC880, 0x5106))
349 gpu->identity.varyings_count -= 1;
352 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
356 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
358 /* Special case for older graphic cores. */
359 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
360 gpu->identity.model = chipModel_GC500;
361 gpu->identity.revision = etnaviv_field(chipIdentity,
362 VIVS_HI_CHIP_IDENTITY_REVISION);
364 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
366 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
367 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
368 gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
371 * Reading these two registers on GC600 rev 0x19 result in a
372 * unhandled fault: external abort on non-linefetch
374 if (!etnaviv_is_model_rev(gpu, GC600, 0x19)) {
375 gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
376 gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
380 * !!!! HACK ALERT !!!!
381 * Because people change device IDs without letting software
382 * know about it - here is the hack to make it all look the
383 * same. Only for GC400 family.
385 if ((gpu->identity.model & 0xff00) == 0x0400 &&
386 gpu->identity.model != chipModel_GC420) {
387 gpu->identity.model = gpu->identity.model & 0x0400;
390 /* Another special case */
391 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
392 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
394 if (chipDate == 0x20080814 && chipTime == 0x12051100) {
396 * This IP has an ECO; put the correct
399 gpu->identity.revision = 0x1051;
404 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
405 * reality it's just a re-branded GC3000. We can identify this
406 * core by the upper half of the revision register being all 1.
407 * Fix model/rev here, so all other places can refer to this
408 * core by its real identity.
410 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
411 gpu->identity.model = chipModel_GC3000;
412 gpu->identity.revision &= 0xffff;
415 if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
416 gpu->identity.eco_id = 1;
418 if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
419 gpu->identity.eco_id = 1;
422 dev_info(gpu->dev, "model: GC%x, revision: %x\n",
423 gpu->identity.model, gpu->identity.revision);
425 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
427 * If there is a match in the HWDB, we aren't interested in the
428 * remaining register values, as they might be wrong.
430 if (etnaviv_fill_identity_from_hwdb(gpu))
433 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
435 /* Disable fast clear on GC700. */
436 if (gpu->identity.model == chipModel_GC700)
437 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
439 /* These models/revisions don't have the 2D pipe bit */
440 if ((gpu->identity.model == chipModel_GC500 &&
441 gpu->identity.revision <= 2) ||
442 gpu->identity.model == chipModel_GC300)
443 gpu->identity.features |= chipFeatures_PIPE_2D;
445 if ((gpu->identity.model == chipModel_GC500 &&
446 gpu->identity.revision < 2) ||
447 (gpu->identity.model == chipModel_GC300 &&
448 gpu->identity.revision < 0x2000)) {
451 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
454 gpu->identity.minor_features0 = 0;
455 gpu->identity.minor_features1 = 0;
456 gpu->identity.minor_features2 = 0;
457 gpu->identity.minor_features3 = 0;
458 gpu->identity.minor_features4 = 0;
459 gpu->identity.minor_features5 = 0;
461 gpu->identity.minor_features0 =
462 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
464 if (gpu->identity.minor_features0 &
465 chipMinorFeatures0_MORE_MINOR_FEATURES) {
466 gpu->identity.minor_features1 =
467 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
468 gpu->identity.minor_features2 =
469 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
470 gpu->identity.minor_features3 =
471 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
472 gpu->identity.minor_features4 =
473 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
474 gpu->identity.minor_features5 =
475 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
478 /* GC600/300 idle register reports zero bits where modules aren't present */
479 if (gpu->identity.model == chipModel_GC600 ||
480 gpu->identity.model == chipModel_GC300)
481 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
482 VIVS_HI_IDLE_STATE_RA |
483 VIVS_HI_IDLE_STATE_SE |
484 VIVS_HI_IDLE_STATE_PA |
485 VIVS_HI_IDLE_STATE_SH |
486 VIVS_HI_IDLE_STATE_PE |
487 VIVS_HI_IDLE_STATE_DE |
488 VIVS_HI_IDLE_STATE_FE;
490 etnaviv_hw_specs(gpu);
493 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
495 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
496 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
497 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
500 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
502 if (gpu->identity.minor_features2 &
503 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
504 clk_set_rate(gpu->clk_core,
505 gpu->base_rate_core >> gpu->freq_scale);
506 clk_set_rate(gpu->clk_shader,
507 gpu->base_rate_shader >> gpu->freq_scale);
509 unsigned int fscale = 1 << (6 - gpu->freq_scale);
510 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
512 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
513 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
514 etnaviv_gpu_load_clock(gpu, clock);
518 * Choose number of wait cycles to target a ~30us (1/32768) max latency
519 * until new work is picked up by the FE when it polls in the idle loop.
520 * If the GPU base frequency is unknown use 200 wait cycles.
522 gpu->fe_waitcycles = clamp(gpu->base_rate_core >> (15 - gpu->freq_scale),
526 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
529 unsigned long timeout;
532 /* We hope that the GPU resets in under one second */
533 timeout = jiffies + msecs_to_jiffies(1000);
535 while (time_is_after_jiffies(timeout)) {
536 unsigned int fscale = 1 << (6 - gpu->freq_scale);
537 u32 pulse_eater = 0x01590880;
539 /* disable clock gating */
540 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, 0x0);
542 /* disable pulse eater */
543 pulse_eater |= BIT(17);
544 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
545 pulse_eater |= BIT(0);
546 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
549 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
550 etnaviv_gpu_load_clock(gpu, control);
552 /* isolate the GPU. */
553 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
554 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
556 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
557 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
558 VIVS_MMUv2_AHB_CONTROL_RESET);
560 /* set soft reset. */
561 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
562 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
565 /* wait for reset. */
566 usleep_range(10, 20);
568 /* reset soft reset bit. */
569 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
570 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
572 /* reset GPU isolation. */
573 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
574 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
576 /* read idle register. */
577 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
579 /* try resetting again if FE is not idle */
580 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
581 dev_dbg(gpu->dev, "FE is not idle\n");
585 /* read reset register. */
586 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
588 /* is the GPU idle? */
589 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
590 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
591 dev_dbg(gpu->dev, "GPU is not idle\n");
595 /* disable debug registers, as they are not normally needed */
596 control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
597 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
604 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
605 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
607 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
608 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
609 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
610 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
615 /* We rely on the GPU running, so program the clock */
616 etnaviv_gpu_update_clock(gpu);
618 gpu->state = ETNA_GPU_STATE_RESET;
619 gpu->exec_state = -1;
620 if (gpu->mmu_context)
621 etnaviv_iommu_context_put(gpu->mmu_context);
622 gpu->mmu_context = NULL;
627 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
631 /* enable clock gating */
632 ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
633 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
635 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
636 if (gpu->identity.revision == 0x4301 ||
637 gpu->identity.revision == 0x4302)
638 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
640 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, ppc);
642 pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS);
644 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
645 if (gpu->identity.model >= chipModel_GC400 &&
646 gpu->identity.model != chipModel_GC420 &&
647 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
648 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
651 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
652 * present without a bug fix.
654 if (gpu->identity.revision < 0x5000 &&
655 gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
656 !(gpu->identity.minor_features1 &
657 chipMinorFeatures1_DISABLE_PE_GATING))
658 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
660 if (gpu->identity.revision < 0x5422)
661 pmc |= BIT(15); /* Unknown bit */
663 /* Disable TX clock gating on affected core revisions. */
664 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
665 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
666 etnaviv_is_model_rev(gpu, GC2000, 0x6202) ||
667 etnaviv_is_model_rev(gpu, GC2000, 0x6203))
668 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
670 /* Disable SE and RA clock gating on affected core revisions. */
671 if (etnaviv_is_model_rev(gpu, GC7000, 0x6202))
672 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE |
673 VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA;
675 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
676 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
678 gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
681 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
683 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
684 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
685 VIVS_FE_COMMAND_CONTROL_ENABLE |
686 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
688 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
689 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
690 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
691 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
695 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu,
696 struct etnaviv_iommu_context *context)
701 WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED);
704 etnaviv_iommu_restore(gpu, context);
706 /* Start command processor */
707 prefetch = etnaviv_buffer_init(gpu);
708 address = etnaviv_cmdbuf_get_va(&gpu->buffer,
709 &gpu->mmu_context->cmdbuf_mapping);
711 etnaviv_gpu_start_fe(gpu, address, prefetch);
713 gpu->state = ETNA_GPU_STATE_RUNNING;
716 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
719 * Base value for VIVS_PM_PULSE_EATER register on models where it
720 * cannot be read, extracted from vivante kernel driver.
722 u32 pulse_eater = 0x01590880;
724 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
725 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
726 pulse_eater |= BIT(23);
730 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
731 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
732 pulse_eater &= ~BIT(16);
733 pulse_eater |= BIT(17);
736 if ((gpu->identity.revision > 0x5420) &&
737 (gpu->identity.features & chipFeatures_PIPE_3D))
739 /* Performance fix: disable internal DFS */
740 pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER);
741 pulse_eater |= BIT(18);
744 gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
747 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
749 WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED ||
750 gpu->state == ETNA_GPU_STATE_RESET));
752 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
753 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
754 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
757 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
759 if (gpu->identity.revision == 0x5007)
760 mc_memory_debug |= 0x0c;
762 mc_memory_debug |= 0x08;
764 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
767 /* enable module-level clock gating */
768 etnaviv_gpu_enable_mlcg(gpu);
771 * Update GPU AXI cache atttribute to "cacheable, no allocate".
772 * This is necessary to prevent the iMX6 SoC locking up.
774 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
775 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
776 VIVS_HI_AXI_CONFIG_ARCACHE(2));
778 /* GC2000 rev 5108 needs a special bus config */
779 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
780 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
781 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
782 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
783 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
784 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
785 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
788 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
789 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
790 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
791 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
794 /* setup the pulse eater */
795 etnaviv_gpu_setup_pulse_eater(gpu);
797 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
799 gpu->state = ETNA_GPU_STATE_INITIALIZED;
802 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
804 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
805 dma_addr_t cmdbuf_paddr;
808 ret = pm_runtime_get_sync(gpu->dev);
810 dev_err(gpu->dev, "Failed to enable GPU power domain\n");
814 etnaviv_hw_identify(gpu);
816 if (gpu->identity.model == 0) {
817 dev_err(gpu->dev, "Unknown GPU model\n");
822 if (gpu->identity.nn_core_count > 0)
823 dev_warn(gpu->dev, "etnaviv has been instantiated on a NPU, "
824 "for which the UAPI is still experimental\n");
826 /* Exclude VG cores with FE2.0 */
827 if (gpu->identity.features & chipFeatures_PIPE_VG &&
828 gpu->identity.features & chipFeatures_FE20) {
829 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
835 * On cores with security features supported, we claim control over the
838 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
839 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
840 gpu->sec_mode = ETNA_SEC_KERNEL;
842 gpu->state = ETNA_GPU_STATE_IDENTIFIED;
844 ret = etnaviv_hw_reset(gpu);
846 dev_err(gpu->dev, "GPU reset failed\n");
850 ret = etnaviv_iommu_global_init(gpu);
855 * If the GPU is part of a system with DMA addressing limitations,
856 * request pages for our SHM backend buffers from the DMA32 zone to
857 * hopefully avoid performance killing SWIOTLB bounce buffering.
859 if (dma_addressing_limited(gpu->dev))
860 priv->shm_gfp_mask |= GFP_DMA32;
863 ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
866 dev_err(gpu->dev, "could not create command buffer\n");
871 * Set the GPU linear window to cover the cmdbuf region, as the GPU
872 * won't be able to start execution otherwise. The alignment to 128M is
873 * chosen arbitrarily but helps in debugging, as the MMU offset
874 * calculations are much more straight forward this way.
876 * On MC1.0 cores the linear window offset is ignored by the TS engine,
877 * leading to inconsistent memory views. Avoid using the offset on those
878 * cores if possible, otherwise disable the TS feature.
880 cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M);
882 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
883 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
884 if (cmdbuf_paddr >= SZ_2G)
885 priv->mmu_global->memory_base = SZ_2G;
887 priv->mmu_global->memory_base = cmdbuf_paddr;
888 } else if (cmdbuf_paddr + SZ_128M >= SZ_2G) {
890 "Need to move linear window on MC1.0, disabling TS\n");
891 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
892 priv->mmu_global->memory_base = SZ_2G;
895 /* Setup event management */
896 spin_lock_init(&gpu->event_spinlock);
897 init_completion(&gpu->event_free);
898 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
899 for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
900 complete(&gpu->event_free);
902 /* Now program the hardware */
903 mutex_lock(&gpu->lock);
904 etnaviv_gpu_hw_init(gpu);
905 mutex_unlock(&gpu->lock);
907 pm_runtime_mark_last_busy(gpu->dev);
908 pm_runtime_put_autosuspend(gpu->dev);
913 pm_runtime_mark_last_busy(gpu->dev);
915 pm_runtime_put_autosuspend(gpu->dev);
920 #ifdef CONFIG_DEBUG_FS
926 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
930 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
931 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
933 for (i = 0; i < 500; i++) {
934 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
935 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
937 if (debug->address[0] != debug->address[1])
940 if (debug->state[0] != debug->state[1])
945 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
947 struct dma_debug debug;
948 u32 dma_lo, dma_hi, axi, idle;
951 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
953 ret = pm_runtime_get_sync(gpu->dev);
957 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
958 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
959 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
960 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
962 verify_dma(gpu, &debug);
964 seq_puts(m, "\tidentity\n");
965 seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
966 seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
967 seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
968 seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
969 seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
971 seq_puts(m, "\tfeatures\n");
972 seq_printf(m, "\t major_features: 0x%08x\n",
973 gpu->identity.features);
974 seq_printf(m, "\t minor_features0: 0x%08x\n",
975 gpu->identity.minor_features0);
976 seq_printf(m, "\t minor_features1: 0x%08x\n",
977 gpu->identity.minor_features1);
978 seq_printf(m, "\t minor_features2: 0x%08x\n",
979 gpu->identity.minor_features2);
980 seq_printf(m, "\t minor_features3: 0x%08x\n",
981 gpu->identity.minor_features3);
982 seq_printf(m, "\t minor_features4: 0x%08x\n",
983 gpu->identity.minor_features4);
984 seq_printf(m, "\t minor_features5: 0x%08x\n",
985 gpu->identity.minor_features5);
986 seq_printf(m, "\t minor_features6: 0x%08x\n",
987 gpu->identity.minor_features6);
988 seq_printf(m, "\t minor_features7: 0x%08x\n",
989 gpu->identity.minor_features7);
990 seq_printf(m, "\t minor_features8: 0x%08x\n",
991 gpu->identity.minor_features8);
992 seq_printf(m, "\t minor_features9: 0x%08x\n",
993 gpu->identity.minor_features9);
994 seq_printf(m, "\t minor_features10: 0x%08x\n",
995 gpu->identity.minor_features10);
996 seq_printf(m, "\t minor_features11: 0x%08x\n",
997 gpu->identity.minor_features11);
999 seq_puts(m, "\tspecs\n");
1000 seq_printf(m, "\t stream_count: %d\n",
1001 gpu->identity.stream_count);
1002 seq_printf(m, "\t register_max: %d\n",
1003 gpu->identity.register_max);
1004 seq_printf(m, "\t thread_count: %d\n",
1005 gpu->identity.thread_count);
1006 seq_printf(m, "\t vertex_cache_size: %d\n",
1007 gpu->identity.vertex_cache_size);
1008 seq_printf(m, "\t shader_core_count: %d\n",
1009 gpu->identity.shader_core_count);
1010 seq_printf(m, "\t nn_core_count: %d\n",
1011 gpu->identity.nn_core_count);
1012 seq_printf(m, "\t pixel_pipes: %d\n",
1013 gpu->identity.pixel_pipes);
1014 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
1015 gpu->identity.vertex_output_buffer_size);
1016 seq_printf(m, "\t buffer_size: %d\n",
1017 gpu->identity.buffer_size);
1018 seq_printf(m, "\t instruction_count: %d\n",
1019 gpu->identity.instruction_count);
1020 seq_printf(m, "\t num_constants: %d\n",
1021 gpu->identity.num_constants);
1022 seq_printf(m, "\t varyings_count: %d\n",
1023 gpu->identity.varyings_count);
1025 seq_printf(m, "\taxi: 0x%08x\n", axi);
1026 seq_printf(m, "\tidle: 0x%08x\n", idle);
1027 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
1028 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
1029 seq_puts(m, "\t FE is not idle\n");
1030 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
1031 seq_puts(m, "\t DE is not idle\n");
1032 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
1033 seq_puts(m, "\t PE is not idle\n");
1034 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
1035 seq_puts(m, "\t SH is not idle\n");
1036 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
1037 seq_puts(m, "\t PA is not idle\n");
1038 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
1039 seq_puts(m, "\t SE is not idle\n");
1040 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
1041 seq_puts(m, "\t RA is not idle\n");
1042 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
1043 seq_puts(m, "\t TX is not idle\n");
1044 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
1045 seq_puts(m, "\t VG is not idle\n");
1046 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
1047 seq_puts(m, "\t IM is not idle\n");
1048 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
1049 seq_puts(m, "\t FP is not idle\n");
1050 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
1051 seq_puts(m, "\t TS is not idle\n");
1052 if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
1053 seq_puts(m, "\t BL is not idle\n");
1054 if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
1055 seq_puts(m, "\t ASYNCFE is not idle\n");
1056 if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
1057 seq_puts(m, "\t MC is not idle\n");
1058 if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
1059 seq_puts(m, "\t PPA is not idle\n");
1060 if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
1061 seq_puts(m, "\t WD is not idle\n");
1062 if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
1063 seq_puts(m, "\t NN is not idle\n");
1064 if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
1065 seq_puts(m, "\t TP is not idle\n");
1066 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
1067 seq_puts(m, "\t AXI low power mode\n");
1069 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
1070 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1071 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1072 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1074 seq_puts(m, "\tMC\n");
1075 seq_printf(m, "\t read0: 0x%08x\n", read0);
1076 seq_printf(m, "\t read1: 0x%08x\n", read1);
1077 seq_printf(m, "\t write: 0x%08x\n", write);
1080 seq_puts(m, "\tDMA ");
1082 if (debug.address[0] == debug.address[1] &&
1083 debug.state[0] == debug.state[1]) {
1084 seq_puts(m, "seems to be stuck\n");
1085 } else if (debug.address[0] == debug.address[1]) {
1086 seq_puts(m, "address is constant\n");
1088 seq_puts(m, "is running\n");
1091 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1092 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1093 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1094 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1095 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1100 pm_runtime_mark_last_busy(gpu->dev);
1102 pm_runtime_put_autosuspend(gpu->dev);
1108 /* fence object management */
1109 struct etnaviv_fence {
1110 struct etnaviv_gpu *gpu;
1111 struct dma_fence base;
1114 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1116 return container_of(fence, struct etnaviv_fence, base);
1119 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1124 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1126 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1128 return dev_name(f->gpu->dev);
1131 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1133 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1135 return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1138 static void etnaviv_fence_release(struct dma_fence *fence)
1140 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1142 kfree_rcu(f, base.rcu);
1145 static const struct dma_fence_ops etnaviv_fence_ops = {
1146 .get_driver_name = etnaviv_fence_get_driver_name,
1147 .get_timeline_name = etnaviv_fence_get_timeline_name,
1148 .signaled = etnaviv_fence_signaled,
1149 .release = etnaviv_fence_release,
1152 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1154 struct etnaviv_fence *f;
1157 * GPU lock must already be held, otherwise fence completion order might
1158 * not match the seqno order assigned here.
1160 lockdep_assert_held(&gpu->lock);
1162 f = kzalloc(sizeof(*f), GFP_KERNEL);
1168 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1169 gpu->fence_context, ++gpu->next_fence);
1174 /* returns true if fence a comes after fence b */
1175 static inline bool fence_after(u32 a, u32 b)
1177 return (s32)(a - b) > 0;
1184 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1185 unsigned int *events)
1187 unsigned long timeout = msecs_to_jiffies(10 * 10000);
1188 unsigned i, acquired = 0, rpm_count = 0;
1191 for (i = 0; i < nr_events; i++) {
1192 unsigned long remaining;
1194 remaining = wait_for_completion_timeout(&gpu->event_free, timeout);
1197 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1203 timeout = remaining;
1206 spin_lock(&gpu->event_spinlock);
1208 for (i = 0; i < nr_events; i++) {
1209 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1212 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1213 set_bit(event, gpu->event_bitmap);
1216 spin_unlock(&gpu->event_spinlock);
1218 for (i = 0; i < nr_events; i++) {
1219 ret = pm_runtime_resume_and_get(gpu->dev);
1228 for (i = 0; i < rpm_count; i++)
1229 pm_runtime_put_autosuspend(gpu->dev);
1231 for (i = 0; i < acquired; i++)
1232 complete(&gpu->event_free);
1237 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1239 if (!test_bit(event, gpu->event_bitmap)) {
1240 dev_warn(gpu->dev, "event %u is already marked as free",
1243 clear_bit(event, gpu->event_bitmap);
1244 complete(&gpu->event_free);
1247 pm_runtime_put_autosuspend(gpu->dev);
1251 * Cmdstream submission/retirement:
1253 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1254 u32 id, struct drm_etnaviv_timespec *timeout)
1256 struct dma_fence *fence;
1260 * Look up the fence and take a reference. We might still find a fence
1261 * whose refcount has already dropped to zero. dma_fence_get_rcu
1262 * pretends we didn't find a fence in that case.
1265 fence = xa_load(&gpu->user_fences, id);
1267 fence = dma_fence_get_rcu(fence);
1274 /* No timeout was requested: just test for completion */
1275 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1277 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1279 ret = dma_fence_wait_timeout(fence, true, remaining);
1282 else if (ret != -ERESTARTSYS)
1287 dma_fence_put(fence);
1292 * Wait for an object to become inactive. This, on it's own, is not race
1293 * free: the object is moved by the scheduler off the active list, and
1294 * then the iova is put. Moreover, the object could be re-submitted just
1295 * after we notice that it's become inactive.
1297 * Although the retirement happens under the gpu lock, we don't want to hold
1298 * that lock in this function while waiting.
1300 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1301 struct etnaviv_gem_object *etnaviv_obj,
1302 struct drm_etnaviv_timespec *timeout)
1304 unsigned long remaining;
1308 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1310 remaining = etnaviv_timeout_to_jiffies(timeout);
1312 ret = wait_event_interruptible_timeout(gpu->fence_event,
1313 !is_active(etnaviv_obj),
1317 else if (ret == -ERESTARTSYS)
1318 return -ERESTARTSYS;
1323 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1324 struct etnaviv_event *event, unsigned int flags)
1326 const struct etnaviv_gem_submit *submit = event->submit;
1329 for (i = 0; i < submit->nr_pmrs; i++) {
1330 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1332 if (pmr->flags == flags)
1333 etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1337 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1338 struct etnaviv_event *event)
1342 /* disable clock gating */
1343 val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1344 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1345 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1347 /* enable debug register */
1348 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1349 val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1350 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1352 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1355 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1356 struct etnaviv_event *event)
1358 const struct etnaviv_gem_submit *submit = event->submit;
1362 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1364 for (i = 0; i < submit->nr_pmrs; i++) {
1365 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1367 *pmr->bo_vma = pmr->sequence;
1370 /* disable debug register */
1371 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1372 val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1373 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1375 /* enable clock gating */
1376 val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS);
1377 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1378 gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val);
1382 /* add bo's to gpu's ring, and kick gpu: */
1383 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1385 struct etnaviv_gpu *gpu = submit->gpu;
1386 struct dma_fence *gpu_fence;
1387 unsigned int i, nr_events = 1, event[3];
1391 * if there are performance monitor requests we need to have
1392 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1394 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1395 * and update the sequence number for userspace.
1397 if (submit->nr_pmrs)
1400 ret = event_alloc(gpu, nr_events, event);
1402 DRM_ERROR("no free events\n");
1403 pm_runtime_put_noidle(gpu->dev);
1407 mutex_lock(&gpu->lock);
1409 gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1411 for (i = 0; i < nr_events; i++)
1412 event_free(gpu, event[i]);
1417 if (gpu->state == ETNA_GPU_STATE_INITIALIZED)
1418 etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context);
1420 if (submit->prev_mmu_context)
1421 etnaviv_iommu_context_put(submit->prev_mmu_context);
1422 submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context);
1424 if (submit->nr_pmrs) {
1425 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1426 kref_get(&submit->refcount);
1427 gpu->event[event[1]].submit = submit;
1428 etnaviv_sync_point_queue(gpu, event[1]);
1431 gpu->event[event[0]].fence = gpu_fence;
1432 submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1433 etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1434 event[0], &submit->cmdbuf);
1436 if (submit->nr_pmrs) {
1437 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1438 kref_get(&submit->refcount);
1439 gpu->event[event[2]].submit = submit;
1440 etnaviv_sync_point_queue(gpu, event[2]);
1444 mutex_unlock(&gpu->lock);
1449 static void sync_point_worker(struct work_struct *work)
1451 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1453 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1454 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1456 event->sync_point(gpu, event);
1457 etnaviv_submit_put(event->submit);
1458 event_free(gpu, gpu->sync_point_event);
1460 /* restart FE last to avoid GPU and IRQ racing against this worker */
1461 etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1464 void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit)
1466 struct etnaviv_gpu *gpu = submit->gpu;
1467 char *comm = NULL, *cmd = NULL;
1468 struct task_struct *task;
1471 dev_err(gpu->dev, "recover hung GPU!\n");
1473 task = get_pid_task(submit->pid, PIDTYPE_PID);
1475 comm = kstrdup(task->comm, GFP_KERNEL);
1476 cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL);
1477 put_task_struct(task);
1481 dev_err(gpu->dev, "offending task: %s (%s)\n", comm, cmd);
1486 if (pm_runtime_get_sync(gpu->dev) < 0)
1489 mutex_lock(&gpu->lock);
1491 etnaviv_hw_reset(gpu);
1493 /* complete all events, the GPU won't do it after the reset */
1494 spin_lock(&gpu->event_spinlock);
1495 for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1497 spin_unlock(&gpu->event_spinlock);
1499 etnaviv_gpu_hw_init(gpu);
1501 mutex_unlock(&gpu->lock);
1502 pm_runtime_mark_last_busy(gpu->dev);
1504 pm_runtime_put_autosuspend(gpu->dev);
1507 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1509 static const char *fault_reasons[] = {
1510 "slave not present",
1514 "read security violation",
1515 "write security violation",
1518 u32 status_reg, status;
1521 if (gpu->sec_mode == ETNA_SEC_NONE)
1522 status_reg = VIVS_MMUv2_STATUS;
1524 status_reg = VIVS_MMUv2_SEC_STATUS;
1526 status = gpu_read(gpu, status_reg);
1527 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1529 for (i = 0; i < 4; i++) {
1530 const char *reason = "unknown";
1534 mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK;
1538 if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons))
1539 reason = fault_reasons[mmu_status - 1];
1541 if (gpu->sec_mode == ETNA_SEC_NONE)
1542 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1544 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1546 dev_err_ratelimited(gpu->dev,
1547 "MMU %d fault (%s) addr 0x%08x\n",
1548 i, reason, gpu_read(gpu, address_reg));
1552 static irqreturn_t irq_handler(int irq, void *data)
1554 struct etnaviv_gpu *gpu = data;
1555 irqreturn_t ret = IRQ_NONE;
1557 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1562 pm_runtime_mark_last_busy(gpu->dev);
1564 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1566 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1567 dev_err(gpu->dev, "AXI bus error\n");
1568 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1571 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1572 dump_mmu_fault(gpu);
1573 gpu->state = ETNA_GPU_STATE_FAULT;
1574 drm_sched_fault(&gpu->sched);
1575 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1578 while ((event = ffs(intr)) != 0) {
1579 struct dma_fence *fence;
1583 intr &= ~(1 << event);
1585 dev_dbg(gpu->dev, "event %u\n", event);
1587 if (gpu->event[event].sync_point) {
1588 gpu->sync_point_event = event;
1589 queue_work(gpu->wq, &gpu->sync_point_work);
1592 fence = gpu->event[event].fence;
1596 gpu->event[event].fence = NULL;
1599 * Events can be processed out of order. Eg,
1600 * - allocate and queue event 0
1601 * - allocate event 1
1602 * - event 0 completes, we process it
1603 * - allocate and queue event 0
1604 * - event 1 and event 0 complete
1605 * we can end up processing event 0 first, then 1.
1607 if (fence_after(fence->seqno, gpu->completed_fence))
1608 gpu->completed_fence = fence->seqno;
1609 dma_fence_signal(fence);
1611 event_free(gpu, event);
1620 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1624 ret = clk_prepare_enable(gpu->clk_reg);
1628 ret = clk_prepare_enable(gpu->clk_bus);
1630 goto disable_clk_reg;
1632 ret = clk_prepare_enable(gpu->clk_core);
1634 goto disable_clk_bus;
1636 ret = clk_prepare_enable(gpu->clk_shader);
1638 goto disable_clk_core;
1643 clk_disable_unprepare(gpu->clk_core);
1645 clk_disable_unprepare(gpu->clk_bus);
1647 clk_disable_unprepare(gpu->clk_reg);
1652 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1654 clk_disable_unprepare(gpu->clk_shader);
1655 clk_disable_unprepare(gpu->clk_core);
1656 clk_disable_unprepare(gpu->clk_bus);
1657 clk_disable_unprepare(gpu->clk_reg);
1662 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1664 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1667 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1669 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1672 if (time_is_before_jiffies(timeout)) {
1674 "timed out waiting for idle: idle=0x%x\n",
1683 static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1685 if (gpu->state == ETNA_GPU_STATE_RUNNING) {
1686 /* Replace the last WAIT with END */
1687 mutex_lock(&gpu->lock);
1688 etnaviv_buffer_end(gpu);
1689 mutex_unlock(&gpu->lock);
1692 * We know that only the FE is busy here, this should
1693 * happen quickly (as the WAIT is only 200 cycles). If
1694 * we fail, just warn and continue.
1696 etnaviv_gpu_wait_idle(gpu, 100);
1698 gpu->state = ETNA_GPU_STATE_INITIALIZED;
1701 gpu->exec_state = -1;
1704 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1708 ret = mutex_lock_killable(&gpu->lock);
1712 etnaviv_gpu_update_clock(gpu);
1713 etnaviv_gpu_hw_init(gpu);
1715 mutex_unlock(&gpu->lock);
1721 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1722 unsigned long *state)
1730 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1731 unsigned long *state)
1733 struct etnaviv_gpu *gpu = cdev->devdata;
1735 *state = gpu->freq_scale;
1741 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1742 unsigned long state)
1744 struct etnaviv_gpu *gpu = cdev->devdata;
1746 mutex_lock(&gpu->lock);
1747 gpu->freq_scale = state;
1748 if (!pm_runtime_suspended(gpu->dev))
1749 etnaviv_gpu_update_clock(gpu);
1750 mutex_unlock(&gpu->lock);
1755 static const struct thermal_cooling_device_ops cooling_ops = {
1756 .get_max_state = etnaviv_gpu_cooling_get_max_state,
1757 .get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1758 .set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1761 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1764 struct drm_device *drm = data;
1765 struct etnaviv_drm_private *priv = drm->dev_private;
1766 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1769 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1770 gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1771 (char *)dev_name(dev), gpu, &cooling_ops);
1772 if (IS_ERR(gpu->cooling))
1773 return PTR_ERR(gpu->cooling);
1776 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1782 ret = etnaviv_sched_init(gpu);
1786 if (!IS_ENABLED(CONFIG_PM)) {
1787 ret = etnaviv_gpu_clk_enable(gpu);
1793 gpu->fence_context = dma_fence_context_alloc(1);
1794 xa_init_flags(&gpu->user_fences, XA_FLAGS_ALLOC);
1795 spin_lock_init(&gpu->fence_spinlock);
1797 INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1798 init_waitqueue_head(&gpu->fence_event);
1800 priv->gpu[priv->num_gpus++] = gpu;
1805 etnaviv_sched_fini(gpu);
1808 destroy_workqueue(gpu->wq);
1811 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1812 thermal_cooling_device_unregister(gpu->cooling);
1817 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1820 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1822 DBG("%s", dev_name(gpu->dev));
1824 destroy_workqueue(gpu->wq);
1826 etnaviv_sched_fini(gpu);
1828 if (IS_ENABLED(CONFIG_PM)) {
1829 pm_runtime_get_sync(gpu->dev);
1830 pm_runtime_put_sync_suspend(gpu->dev);
1832 etnaviv_gpu_hw_suspend(gpu);
1833 etnaviv_gpu_clk_disable(gpu);
1836 if (gpu->mmu_context)
1837 etnaviv_iommu_context_put(gpu->mmu_context);
1839 etnaviv_cmdbuf_free(&gpu->buffer);
1840 etnaviv_iommu_global_fini(gpu);
1843 xa_destroy(&gpu->user_fences);
1845 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1846 thermal_cooling_device_unregister(gpu->cooling);
1847 gpu->cooling = NULL;
1850 static const struct component_ops gpu_ops = {
1851 .bind = etnaviv_gpu_bind,
1852 .unbind = etnaviv_gpu_unbind,
1855 static const struct of_device_id etnaviv_gpu_match[] = {
1857 .compatible = "vivante,gc"
1861 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1863 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1865 struct device *dev = &pdev->dev;
1866 struct etnaviv_gpu *gpu;
1869 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1873 gpu->dev = &pdev->dev;
1874 mutex_init(&gpu->lock);
1875 mutex_init(&gpu->sched_lock);
1877 /* Map registers: */
1878 gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1879 if (IS_ERR(gpu->mmio))
1880 return PTR_ERR(gpu->mmio);
1882 /* Get Interrupt: */
1883 gpu->irq = platform_get_irq(pdev, 0);
1887 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1888 dev_name(gpu->dev), gpu);
1890 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1895 gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1896 DBG("clk_reg: %p", gpu->clk_reg);
1897 if (IS_ERR(gpu->clk_reg))
1898 return PTR_ERR(gpu->clk_reg);
1900 gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1901 DBG("clk_bus: %p", gpu->clk_bus);
1902 if (IS_ERR(gpu->clk_bus))
1903 return PTR_ERR(gpu->clk_bus);
1905 gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1906 DBG("clk_core: %p", gpu->clk_core);
1907 if (IS_ERR(gpu->clk_core))
1908 return PTR_ERR(gpu->clk_core);
1909 gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1911 gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1912 DBG("clk_shader: %p", gpu->clk_shader);
1913 if (IS_ERR(gpu->clk_shader))
1914 return PTR_ERR(gpu->clk_shader);
1915 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1917 /* TODO: figure out max mapped size */
1918 dev_set_drvdata(dev, gpu);
1921 * We treat the device as initially suspended. The runtime PM
1922 * autosuspend delay is rather arbitary: no measurements have
1923 * yet been performed to determine an appropriate value.
1925 pm_runtime_use_autosuspend(gpu->dev);
1926 pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1927 pm_runtime_enable(gpu->dev);
1929 err = component_add(&pdev->dev, &gpu_ops);
1931 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1938 static void etnaviv_gpu_platform_remove(struct platform_device *pdev)
1940 component_del(&pdev->dev, &gpu_ops);
1941 pm_runtime_disable(&pdev->dev);
1944 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1946 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1949 /* If there are any jobs in the HW queue, we're not idle */
1950 if (atomic_read(&gpu->sched.credit_count))
1953 /* Check whether the hardware (except FE and MC) is idle */
1954 mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1955 VIVS_HI_IDLE_STATE_MC);
1956 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1958 dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1963 etnaviv_gpu_hw_suspend(gpu);
1965 gpu->state = ETNA_GPU_STATE_IDENTIFIED;
1967 return etnaviv_gpu_clk_disable(gpu);
1970 static int etnaviv_gpu_rpm_resume(struct device *dev)
1972 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1975 ret = etnaviv_gpu_clk_enable(gpu);
1979 /* Re-initialise the basic hardware state */
1980 if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) {
1981 ret = etnaviv_gpu_hw_resume(gpu);
1983 etnaviv_gpu_clk_disable(gpu);
1991 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1992 RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL)
1995 struct platform_driver etnaviv_gpu_driver = {
1997 .name = "etnaviv-gpu",
1998 .owner = THIS_MODULE,
1999 .pm = pm_ptr(&etnaviv_gpu_pm_ops),
2000 .of_match_table = etnaviv_gpu_match,
2002 .probe = etnaviv_gpu_platform_probe,
2003 .remove_new = etnaviv_gpu_platform_remove,
2004 .id_table = gpu_ids,