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/module.h>
12 #include <linux/of_device.h>
13 #include <linux/platform_device.h>
14 #include <linux/pm_runtime.h>
15 #include <linux/regulator/consumer.h>
16 #include <linux/thermal.h>
18 #include "etnaviv_cmdbuf.h"
19 #include "etnaviv_dump.h"
20 #include "etnaviv_gpu.h"
21 #include "etnaviv_gem.h"
22 #include "etnaviv_mmu.h"
23 #include "etnaviv_perfmon.h"
24 #include "etnaviv_sched.h"
25 #include "common.xml.h"
26 #include "state.xml.h"
27 #include "state_hi.xml.h"
28 #include "cmdstream.xml.h"
30 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;
168 DBG("%s: invalid param: %u", dev_name(gpu->dev), param);
176 #define etnaviv_is_model_rev(gpu, mod, rev) \
177 ((gpu)->identity.model == chipModel_##mod && \
178 (gpu)->identity.revision == rev)
179 #define etnaviv_field(val, field) \
180 (((val) & field##__MASK) >> field##__SHIFT)
182 static void etnaviv_hw_specs(struct etnaviv_gpu *gpu)
184 if (gpu->identity.minor_features0 &
185 chipMinorFeatures0_MORE_MINOR_FEATURES) {
187 unsigned int streams;
189 specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS);
190 specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2);
191 specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3);
192 specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4);
194 gpu->identity.stream_count = etnaviv_field(specs[0],
195 VIVS_HI_CHIP_SPECS_STREAM_COUNT);
196 gpu->identity.register_max = etnaviv_field(specs[0],
197 VIVS_HI_CHIP_SPECS_REGISTER_MAX);
198 gpu->identity.thread_count = etnaviv_field(specs[0],
199 VIVS_HI_CHIP_SPECS_THREAD_COUNT);
200 gpu->identity.vertex_cache_size = etnaviv_field(specs[0],
201 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE);
202 gpu->identity.shader_core_count = etnaviv_field(specs[0],
203 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT);
204 gpu->identity.pixel_pipes = etnaviv_field(specs[0],
205 VIVS_HI_CHIP_SPECS_PIXEL_PIPES);
206 gpu->identity.vertex_output_buffer_size =
207 etnaviv_field(specs[0],
208 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE);
210 gpu->identity.buffer_size = etnaviv_field(specs[1],
211 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE);
212 gpu->identity.instruction_count = etnaviv_field(specs[1],
213 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT);
214 gpu->identity.num_constants = etnaviv_field(specs[1],
215 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS);
217 gpu->identity.varyings_count = etnaviv_field(specs[2],
218 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT);
220 /* This overrides the value from older register if non-zero */
221 streams = etnaviv_field(specs[3],
222 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT);
224 gpu->identity.stream_count = streams;
227 /* Fill in the stream count if not specified */
228 if (gpu->identity.stream_count == 0) {
229 if (gpu->identity.model >= 0x1000)
230 gpu->identity.stream_count = 4;
232 gpu->identity.stream_count = 1;
235 /* Convert the register max value */
236 if (gpu->identity.register_max)
237 gpu->identity.register_max = 1 << gpu->identity.register_max;
238 else if (gpu->identity.model == chipModel_GC400)
239 gpu->identity.register_max = 32;
241 gpu->identity.register_max = 64;
243 /* Convert thread count */
244 if (gpu->identity.thread_count)
245 gpu->identity.thread_count = 1 << gpu->identity.thread_count;
246 else if (gpu->identity.model == chipModel_GC400)
247 gpu->identity.thread_count = 64;
248 else if (gpu->identity.model == chipModel_GC500 ||
249 gpu->identity.model == chipModel_GC530)
250 gpu->identity.thread_count = 128;
252 gpu->identity.thread_count = 256;
254 if (gpu->identity.vertex_cache_size == 0)
255 gpu->identity.vertex_cache_size = 8;
257 if (gpu->identity.shader_core_count == 0) {
258 if (gpu->identity.model >= 0x1000)
259 gpu->identity.shader_core_count = 2;
261 gpu->identity.shader_core_count = 1;
264 if (gpu->identity.pixel_pipes == 0)
265 gpu->identity.pixel_pipes = 1;
267 /* Convert virtex buffer size */
268 if (gpu->identity.vertex_output_buffer_size) {
269 gpu->identity.vertex_output_buffer_size =
270 1 << gpu->identity.vertex_output_buffer_size;
271 } else if (gpu->identity.model == chipModel_GC400) {
272 if (gpu->identity.revision < 0x4000)
273 gpu->identity.vertex_output_buffer_size = 512;
274 else if (gpu->identity.revision < 0x4200)
275 gpu->identity.vertex_output_buffer_size = 256;
277 gpu->identity.vertex_output_buffer_size = 128;
279 gpu->identity.vertex_output_buffer_size = 512;
282 switch (gpu->identity.instruction_count) {
284 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
285 gpu->identity.model == chipModel_GC880)
286 gpu->identity.instruction_count = 512;
288 gpu->identity.instruction_count = 256;
292 gpu->identity.instruction_count = 1024;
296 gpu->identity.instruction_count = 2048;
300 gpu->identity.instruction_count = 256;
304 if (gpu->identity.num_constants == 0)
305 gpu->identity.num_constants = 168;
307 if (gpu->identity.varyings_count == 0) {
308 if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0)
309 gpu->identity.varyings_count = 12;
311 gpu->identity.varyings_count = 8;
315 * For some cores, two varyings are consumed for position, so the
316 * maximum varying count needs to be reduced by one.
318 if (etnaviv_is_model_rev(gpu, GC5000, 0x5434) ||
319 etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
320 etnaviv_is_model_rev(gpu, GC4000, 0x5245) ||
321 etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
322 etnaviv_is_model_rev(gpu, GC3000, 0x5435) ||
323 etnaviv_is_model_rev(gpu, GC2200, 0x5244) ||
324 etnaviv_is_model_rev(gpu, GC2100, 0x5108) ||
325 etnaviv_is_model_rev(gpu, GC2000, 0x5108) ||
326 etnaviv_is_model_rev(gpu, GC1500, 0x5246) ||
327 etnaviv_is_model_rev(gpu, GC880, 0x5107) ||
328 etnaviv_is_model_rev(gpu, GC880, 0x5106))
329 gpu->identity.varyings_count -= 1;
332 static void etnaviv_hw_identify(struct etnaviv_gpu *gpu)
336 chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY);
338 /* Special case for older graphic cores. */
339 if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) {
340 gpu->identity.model = chipModel_GC500;
341 gpu->identity.revision = etnaviv_field(chipIdentity,
342 VIVS_HI_CHIP_IDENTITY_REVISION);
344 u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE);
346 gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL);
347 gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV);
348 gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID);
351 * Reading these two registers on GC600 rev 0x19 result in a
352 * unhandled fault: external abort on non-linefetch
354 if (!etnaviv_is_model_rev(gpu, GC600, 0x19)) {
355 gpu->identity.product_id = gpu_read(gpu, VIVS_HI_CHIP_PRODUCT_ID);
356 gpu->identity.eco_id = gpu_read(gpu, VIVS_HI_CHIP_ECO_ID);
360 * !!!! HACK ALERT !!!!
361 * Because people change device IDs without letting software
362 * know about it - here is the hack to make it all look the
363 * same. Only for GC400 family.
365 if ((gpu->identity.model & 0xff00) == 0x0400 &&
366 gpu->identity.model != chipModel_GC420) {
367 gpu->identity.model = gpu->identity.model & 0x0400;
370 /* Another special case */
371 if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) {
372 u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME);
374 if (chipDate == 0x20080814 && chipTime == 0x12051100) {
376 * This IP has an ECO; put the correct
379 gpu->identity.revision = 0x1051;
384 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
385 * reality it's just a re-branded GC3000. We can identify this
386 * core by the upper half of the revision register being all 1.
387 * Fix model/rev here, so all other places can refer to this
388 * core by its real identity.
390 if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) {
391 gpu->identity.model = chipModel_GC3000;
392 gpu->identity.revision &= 0xffff;
395 if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617))
396 gpu->identity.eco_id = 1;
398 if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511))
399 gpu->identity.eco_id = 1;
402 dev_info(gpu->dev, "model: GC%x, revision: %x\n",
403 gpu->identity.model, gpu->identity.revision);
405 gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP;
407 * If there is a match in the HWDB, we aren't interested in the
408 * remaining register values, as they might be wrong.
410 if (etnaviv_fill_identity_from_hwdb(gpu))
413 gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE);
415 /* Disable fast clear on GC700. */
416 if (gpu->identity.model == chipModel_GC700)
417 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
419 if ((gpu->identity.model == chipModel_GC500 &&
420 gpu->identity.revision < 2) ||
421 (gpu->identity.model == chipModel_GC300 &&
422 gpu->identity.revision < 0x2000)) {
425 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
428 gpu->identity.minor_features0 = 0;
429 gpu->identity.minor_features1 = 0;
430 gpu->identity.minor_features2 = 0;
431 gpu->identity.minor_features3 = 0;
432 gpu->identity.minor_features4 = 0;
433 gpu->identity.minor_features5 = 0;
435 gpu->identity.minor_features0 =
436 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0);
438 if (gpu->identity.minor_features0 &
439 chipMinorFeatures0_MORE_MINOR_FEATURES) {
440 gpu->identity.minor_features1 =
441 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1);
442 gpu->identity.minor_features2 =
443 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2);
444 gpu->identity.minor_features3 =
445 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3);
446 gpu->identity.minor_features4 =
447 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4);
448 gpu->identity.minor_features5 =
449 gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5);
452 /* GC600 idle register reports zero bits where modules aren't present */
453 if (gpu->identity.model == chipModel_GC600)
454 gpu->idle_mask = VIVS_HI_IDLE_STATE_TX |
455 VIVS_HI_IDLE_STATE_RA |
456 VIVS_HI_IDLE_STATE_SE |
457 VIVS_HI_IDLE_STATE_PA |
458 VIVS_HI_IDLE_STATE_SH |
459 VIVS_HI_IDLE_STATE_PE |
460 VIVS_HI_IDLE_STATE_DE |
461 VIVS_HI_IDLE_STATE_FE;
463 etnaviv_hw_specs(gpu);
466 static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock)
468 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock |
469 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD);
470 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock);
473 static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu)
475 if (gpu->identity.minor_features2 &
476 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) {
477 clk_set_rate(gpu->clk_core,
478 gpu->base_rate_core >> gpu->freq_scale);
479 clk_set_rate(gpu->clk_shader,
480 gpu->base_rate_shader >> gpu->freq_scale);
482 unsigned int fscale = 1 << (6 - gpu->freq_scale);
483 u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
485 clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK;
486 clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
487 etnaviv_gpu_load_clock(gpu, clock);
491 static int etnaviv_hw_reset(struct etnaviv_gpu *gpu)
494 unsigned long timeout;
497 /* We hope that the GPU resets in under one second */
498 timeout = jiffies + msecs_to_jiffies(1000);
500 while (time_is_after_jiffies(timeout)) {
502 unsigned int fscale = 1 << (6 - gpu->freq_scale);
503 control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale);
504 etnaviv_gpu_load_clock(gpu, control);
506 /* isolate the GPU. */
507 control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
508 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
510 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
511 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL,
512 VIVS_MMUv2_AHB_CONTROL_RESET);
514 /* set soft reset. */
515 control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
516 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
519 /* wait for reset. */
520 usleep_range(10, 20);
522 /* reset soft reset bit. */
523 control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET;
524 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
526 /* reset GPU isolation. */
527 control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU;
528 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
530 /* read idle register. */
531 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
533 /* try resetting again if FE is not idle */
534 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) {
535 dev_dbg(gpu->dev, "FE is not idle\n");
539 /* read reset register. */
540 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
542 /* is the GPU idle? */
543 if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) ||
544 ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) {
545 dev_dbg(gpu->dev, "GPU is not idle\n");
549 /* disable debug registers, as they are not normally needed */
550 control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
551 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control);
558 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
559 control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
561 dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
562 idle & VIVS_HI_IDLE_STATE_FE ? "" : "not ",
563 control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not ",
564 control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not ");
569 /* We rely on the GPU running, so program the clock */
570 etnaviv_gpu_update_clock(gpu);
575 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu)
579 /* enable clock gating */
580 ppc = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
581 ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
583 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
584 if (gpu->identity.revision == 0x4301 ||
585 gpu->identity.revision == 0x4302)
586 ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING;
588 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, ppc);
590 pmc = gpu_read(gpu, VIVS_PM_MODULE_CONTROLS);
592 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
593 if (gpu->identity.model >= chipModel_GC400 &&
594 gpu->identity.model != chipModel_GC420 &&
595 !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12))
596 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA;
599 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
600 * present without a bug fix.
602 if (gpu->identity.revision < 0x5000 &&
603 gpu->identity.minor_features0 & chipMinorFeatures0_HZ &&
604 !(gpu->identity.minor_features1 &
605 chipMinorFeatures1_DISABLE_PE_GATING))
606 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE;
608 if (gpu->identity.revision < 0x5422)
609 pmc |= BIT(15); /* Unknown bit */
611 /* Disable TX clock gating on affected core revisions. */
612 if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) ||
613 etnaviv_is_model_rev(gpu, GC2000, 0x5108))
614 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
616 /* Disable SE, RA and TX clock gating on affected core revisions. */
617 if (etnaviv_is_model_rev(gpu, GC7000, 0x6202))
618 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE |
619 VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA |
620 VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX;
622 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ;
623 pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ;
625 gpu_write(gpu, VIVS_PM_MODULE_CONTROLS, pmc);
628 void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch)
630 gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address);
631 gpu_write(gpu, VIVS_FE_COMMAND_CONTROL,
632 VIVS_FE_COMMAND_CONTROL_ENABLE |
633 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch));
635 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
636 gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL,
637 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE |
638 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch));
642 static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu)
644 u32 address = etnaviv_cmdbuf_get_va(&gpu->buffer,
645 &gpu->mmu_context->cmdbuf_mapping);
649 etnaviv_iommu_restore(gpu, gpu->mmu_context);
651 /* Start command processor */
652 prefetch = etnaviv_buffer_init(gpu);
654 etnaviv_gpu_start_fe(gpu, address, prefetch);
657 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu)
660 * Base value for VIVS_PM_PULSE_EATER register on models where it
661 * cannot be read, extracted from vivante kernel driver.
663 u32 pulse_eater = 0x01590880;
665 if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) ||
666 etnaviv_is_model_rev(gpu, GC4000, 0x5222)) {
667 pulse_eater |= BIT(23);
671 if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) ||
672 etnaviv_is_model_rev(gpu, GC1000, 0x5040)) {
673 pulse_eater &= ~BIT(16);
674 pulse_eater |= BIT(17);
677 if ((gpu->identity.revision > 0x5420) &&
678 (gpu->identity.features & chipFeatures_PIPE_3D))
680 /* Performance fix: disable internal DFS */
681 pulse_eater = gpu_read(gpu, VIVS_PM_PULSE_EATER);
682 pulse_eater |= BIT(18);
685 gpu_write(gpu, VIVS_PM_PULSE_EATER, pulse_eater);
688 static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu)
690 if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) ||
691 etnaviv_is_model_rev(gpu, GC320, 0x5220)) &&
692 gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) {
695 mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff;
697 if (gpu->identity.revision == 0x5007)
698 mc_memory_debug |= 0x0c;
700 mc_memory_debug |= 0x08;
702 gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, mc_memory_debug);
705 /* enable module-level clock gating */
706 etnaviv_gpu_enable_mlcg(gpu);
709 * Update GPU AXI cache atttribute to "cacheable, no allocate".
710 * This is necessary to prevent the iMX6 SoC locking up.
712 gpu_write(gpu, VIVS_HI_AXI_CONFIG,
713 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
714 VIVS_HI_AXI_CONFIG_ARCACHE(2));
716 /* GC2000 rev 5108 needs a special bus config */
717 if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) {
718 u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG);
719 bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK |
720 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK);
721 bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
722 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
723 gpu_write(gpu, VIVS_MC_BUS_CONFIG, bus_config);
726 if (gpu->sec_mode == ETNA_SEC_KERNEL) {
727 u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL);
728 val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS;
729 gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, val);
732 /* setup the pulse eater */
733 etnaviv_gpu_setup_pulse_eater(gpu);
735 gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U);
738 int etnaviv_gpu_init(struct etnaviv_gpu *gpu)
740 struct etnaviv_drm_private *priv = gpu->drm->dev_private;
741 dma_addr_t cmdbuf_paddr;
744 ret = pm_runtime_get_sync(gpu->dev);
746 dev_err(gpu->dev, "Failed to enable GPU power domain\n");
750 etnaviv_hw_identify(gpu);
752 if (gpu->identity.model == 0) {
753 dev_err(gpu->dev, "Unknown GPU model\n");
758 /* Exclude VG cores with FE2.0 */
759 if (gpu->identity.features & chipFeatures_PIPE_VG &&
760 gpu->identity.features & chipFeatures_FE20) {
761 dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n");
767 * On cores with security features supported, we claim control over the
770 if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) &&
771 (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB))
772 gpu->sec_mode = ETNA_SEC_KERNEL;
774 ret = etnaviv_hw_reset(gpu);
776 dev_err(gpu->dev, "GPU reset failed\n");
780 ret = etnaviv_iommu_global_init(gpu);
785 * If the GPU is part of a system with DMA addressing limitations,
786 * request pages for our SHM backend buffers from the DMA32 zone to
787 * hopefully avoid performance killing SWIOTLB bounce buffering.
789 if (dma_addressing_limited(gpu->dev))
790 priv->shm_gfp_mask |= GFP_DMA32;
793 ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer,
796 dev_err(gpu->dev, "could not create command buffer\n");
801 * Set the GPU linear window to cover the cmdbuf region, as the GPU
802 * won't be able to start execution otherwise. The alignment to 128M is
803 * chosen arbitrarily but helps in debugging, as the MMU offset
804 * calculations are much more straight forward this way.
806 * On MC1.0 cores the linear window offset is ignored by the TS engine,
807 * leading to inconsistent memory views. Avoid using the offset on those
808 * cores if possible, otherwise disable the TS feature.
810 cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M);
812 if (!(gpu->identity.features & chipFeatures_PIPE_3D) ||
813 (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) {
814 if (cmdbuf_paddr >= SZ_2G)
815 priv->mmu_global->memory_base = SZ_2G;
817 priv->mmu_global->memory_base = cmdbuf_paddr;
818 } else if (cmdbuf_paddr + SZ_128M >= SZ_2G) {
820 "Need to move linear window on MC1.0, disabling TS\n");
821 gpu->identity.features &= ~chipFeatures_FAST_CLEAR;
822 priv->mmu_global->memory_base = SZ_2G;
825 /* Setup event management */
826 spin_lock_init(&gpu->event_spinlock);
827 init_completion(&gpu->event_free);
828 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
829 for (i = 0; i < ARRAY_SIZE(gpu->event); i++)
830 complete(&gpu->event_free);
832 /* Now program the hardware */
833 mutex_lock(&gpu->lock);
834 etnaviv_gpu_hw_init(gpu);
835 gpu->exec_state = -1;
836 mutex_unlock(&gpu->lock);
838 pm_runtime_mark_last_busy(gpu->dev);
839 pm_runtime_put_autosuspend(gpu->dev);
841 gpu->initialized = true;
846 pm_runtime_mark_last_busy(gpu->dev);
848 pm_runtime_put_autosuspend(gpu->dev);
853 #ifdef CONFIG_DEBUG_FS
859 static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug)
863 debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
864 debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
866 for (i = 0; i < 500; i++) {
867 debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
868 debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE);
870 if (debug->address[0] != debug->address[1])
873 if (debug->state[0] != debug->state[1])
878 int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m)
880 struct dma_debug debug;
881 u32 dma_lo, dma_hi, axi, idle;
884 seq_printf(m, "%s Status:\n", dev_name(gpu->dev));
886 ret = pm_runtime_get_sync(gpu->dev);
890 dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW);
891 dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH);
892 axi = gpu_read(gpu, VIVS_HI_AXI_STATUS);
893 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
895 verify_dma(gpu, &debug);
897 seq_puts(m, "\tidentity\n");
898 seq_printf(m, "\t model: 0x%x\n", gpu->identity.model);
899 seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision);
900 seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id);
901 seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id);
902 seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id);
904 seq_puts(m, "\tfeatures\n");
905 seq_printf(m, "\t major_features: 0x%08x\n",
906 gpu->identity.features);
907 seq_printf(m, "\t minor_features0: 0x%08x\n",
908 gpu->identity.minor_features0);
909 seq_printf(m, "\t minor_features1: 0x%08x\n",
910 gpu->identity.minor_features1);
911 seq_printf(m, "\t minor_features2: 0x%08x\n",
912 gpu->identity.minor_features2);
913 seq_printf(m, "\t minor_features3: 0x%08x\n",
914 gpu->identity.minor_features3);
915 seq_printf(m, "\t minor_features4: 0x%08x\n",
916 gpu->identity.minor_features4);
917 seq_printf(m, "\t minor_features5: 0x%08x\n",
918 gpu->identity.minor_features5);
919 seq_printf(m, "\t minor_features6: 0x%08x\n",
920 gpu->identity.minor_features6);
921 seq_printf(m, "\t minor_features7: 0x%08x\n",
922 gpu->identity.minor_features7);
923 seq_printf(m, "\t minor_features8: 0x%08x\n",
924 gpu->identity.minor_features8);
925 seq_printf(m, "\t minor_features9: 0x%08x\n",
926 gpu->identity.minor_features9);
927 seq_printf(m, "\t minor_features10: 0x%08x\n",
928 gpu->identity.minor_features10);
929 seq_printf(m, "\t minor_features11: 0x%08x\n",
930 gpu->identity.minor_features11);
932 seq_puts(m, "\tspecs\n");
933 seq_printf(m, "\t stream_count: %d\n",
934 gpu->identity.stream_count);
935 seq_printf(m, "\t register_max: %d\n",
936 gpu->identity.register_max);
937 seq_printf(m, "\t thread_count: %d\n",
938 gpu->identity.thread_count);
939 seq_printf(m, "\t vertex_cache_size: %d\n",
940 gpu->identity.vertex_cache_size);
941 seq_printf(m, "\t shader_core_count: %d\n",
942 gpu->identity.shader_core_count);
943 seq_printf(m, "\t pixel_pipes: %d\n",
944 gpu->identity.pixel_pipes);
945 seq_printf(m, "\t vertex_output_buffer_size: %d\n",
946 gpu->identity.vertex_output_buffer_size);
947 seq_printf(m, "\t buffer_size: %d\n",
948 gpu->identity.buffer_size);
949 seq_printf(m, "\t instruction_count: %d\n",
950 gpu->identity.instruction_count);
951 seq_printf(m, "\t num_constants: %d\n",
952 gpu->identity.num_constants);
953 seq_printf(m, "\t varyings_count: %d\n",
954 gpu->identity.varyings_count);
956 seq_printf(m, "\taxi: 0x%08x\n", axi);
957 seq_printf(m, "\tidle: 0x%08x\n", idle);
958 idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP;
959 if ((idle & VIVS_HI_IDLE_STATE_FE) == 0)
960 seq_puts(m, "\t FE is not idle\n");
961 if ((idle & VIVS_HI_IDLE_STATE_DE) == 0)
962 seq_puts(m, "\t DE is not idle\n");
963 if ((idle & VIVS_HI_IDLE_STATE_PE) == 0)
964 seq_puts(m, "\t PE is not idle\n");
965 if ((idle & VIVS_HI_IDLE_STATE_SH) == 0)
966 seq_puts(m, "\t SH is not idle\n");
967 if ((idle & VIVS_HI_IDLE_STATE_PA) == 0)
968 seq_puts(m, "\t PA is not idle\n");
969 if ((idle & VIVS_HI_IDLE_STATE_SE) == 0)
970 seq_puts(m, "\t SE is not idle\n");
971 if ((idle & VIVS_HI_IDLE_STATE_RA) == 0)
972 seq_puts(m, "\t RA is not idle\n");
973 if ((idle & VIVS_HI_IDLE_STATE_TX) == 0)
974 seq_puts(m, "\t TX is not idle\n");
975 if ((idle & VIVS_HI_IDLE_STATE_VG) == 0)
976 seq_puts(m, "\t VG is not idle\n");
977 if ((idle & VIVS_HI_IDLE_STATE_IM) == 0)
978 seq_puts(m, "\t IM is not idle\n");
979 if ((idle & VIVS_HI_IDLE_STATE_FP) == 0)
980 seq_puts(m, "\t FP is not idle\n");
981 if ((idle & VIVS_HI_IDLE_STATE_TS) == 0)
982 seq_puts(m, "\t TS is not idle\n");
983 if ((idle & VIVS_HI_IDLE_STATE_BL) == 0)
984 seq_puts(m, "\t BL is not idle\n");
985 if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0)
986 seq_puts(m, "\t ASYNCFE is not idle\n");
987 if ((idle & VIVS_HI_IDLE_STATE_MC) == 0)
988 seq_puts(m, "\t MC is not idle\n");
989 if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0)
990 seq_puts(m, "\t PPA is not idle\n");
991 if ((idle & VIVS_HI_IDLE_STATE_WD) == 0)
992 seq_puts(m, "\t WD is not idle\n");
993 if ((idle & VIVS_HI_IDLE_STATE_NN) == 0)
994 seq_puts(m, "\t NN is not idle\n");
995 if ((idle & VIVS_HI_IDLE_STATE_TP) == 0)
996 seq_puts(m, "\t TP is not idle\n");
997 if (idle & VIVS_HI_IDLE_STATE_AXI_LP)
998 seq_puts(m, "\t AXI low power mode\n");
1000 if (gpu->identity.features & chipFeatures_DEBUG_MODE) {
1001 u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0);
1002 u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1);
1003 u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE);
1005 seq_puts(m, "\tMC\n");
1006 seq_printf(m, "\t read0: 0x%08x\n", read0);
1007 seq_printf(m, "\t read1: 0x%08x\n", read1);
1008 seq_printf(m, "\t write: 0x%08x\n", write);
1011 seq_puts(m, "\tDMA ");
1013 if (debug.address[0] == debug.address[1] &&
1014 debug.state[0] == debug.state[1]) {
1015 seq_puts(m, "seems to be stuck\n");
1016 } else if (debug.address[0] == debug.address[1]) {
1017 seq_puts(m, "address is constant\n");
1019 seq_puts(m, "is running\n");
1022 seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]);
1023 seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]);
1024 seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]);
1025 seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]);
1026 seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
1031 pm_runtime_mark_last_busy(gpu->dev);
1033 pm_runtime_put_autosuspend(gpu->dev);
1039 void etnaviv_gpu_recover_hang(struct etnaviv_gpu *gpu)
1043 dev_err(gpu->dev, "recover hung GPU!\n");
1045 if (pm_runtime_get_sync(gpu->dev) < 0)
1048 mutex_lock(&gpu->lock);
1050 etnaviv_hw_reset(gpu);
1052 /* complete all events, the GPU won't do it after the reset */
1053 spin_lock(&gpu->event_spinlock);
1054 for_each_set_bit_from(i, gpu->event_bitmap, ETNA_NR_EVENTS)
1055 complete(&gpu->event_free);
1056 bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS);
1057 spin_unlock(&gpu->event_spinlock);
1059 etnaviv_gpu_hw_init(gpu);
1060 gpu->exec_state = -1;
1061 gpu->mmu_context = NULL;
1063 mutex_unlock(&gpu->lock);
1064 pm_runtime_mark_last_busy(gpu->dev);
1066 pm_runtime_put_autosuspend(gpu->dev);
1069 /* fence object management */
1070 struct etnaviv_fence {
1071 struct etnaviv_gpu *gpu;
1072 struct dma_fence base;
1075 static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence)
1077 return container_of(fence, struct etnaviv_fence, base);
1080 static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence)
1085 static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence)
1087 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1089 return dev_name(f->gpu->dev);
1092 static bool etnaviv_fence_signaled(struct dma_fence *fence)
1094 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1096 return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0;
1099 static void etnaviv_fence_release(struct dma_fence *fence)
1101 struct etnaviv_fence *f = to_etnaviv_fence(fence);
1103 kfree_rcu(f, base.rcu);
1106 static const struct dma_fence_ops etnaviv_fence_ops = {
1107 .get_driver_name = etnaviv_fence_get_driver_name,
1108 .get_timeline_name = etnaviv_fence_get_timeline_name,
1109 .signaled = etnaviv_fence_signaled,
1110 .release = etnaviv_fence_release,
1113 static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu)
1115 struct etnaviv_fence *f;
1118 * GPU lock must already be held, otherwise fence completion order might
1119 * not match the seqno order assigned here.
1121 lockdep_assert_held(&gpu->lock);
1123 f = kzalloc(sizeof(*f), GFP_KERNEL);
1129 dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock,
1130 gpu->fence_context, ++gpu->next_fence);
1135 /* returns true if fence a comes after fence b */
1136 static inline bool fence_after(u32 a, u32 b)
1138 return (s32)(a - b) > 0;
1145 static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events,
1146 unsigned int *events)
1148 unsigned long timeout = msecs_to_jiffies(10 * 10000);
1149 unsigned i, acquired = 0;
1151 for (i = 0; i < nr_events; i++) {
1154 ret = wait_for_completion_timeout(&gpu->event_free, timeout);
1157 dev_err(gpu->dev, "wait_for_completion_timeout failed");
1165 spin_lock(&gpu->event_spinlock);
1167 for (i = 0; i < nr_events; i++) {
1168 int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS);
1171 memset(&gpu->event[event], 0, sizeof(struct etnaviv_event));
1172 set_bit(event, gpu->event_bitmap);
1175 spin_unlock(&gpu->event_spinlock);
1180 for (i = 0; i < acquired; i++)
1181 complete(&gpu->event_free);
1186 static void event_free(struct etnaviv_gpu *gpu, unsigned int event)
1188 if (!test_bit(event, gpu->event_bitmap)) {
1189 dev_warn(gpu->dev, "event %u is already marked as free",
1192 clear_bit(event, gpu->event_bitmap);
1193 complete(&gpu->event_free);
1198 * Cmdstream submission/retirement:
1200 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu,
1201 u32 id, struct drm_etnaviv_timespec *timeout)
1203 struct dma_fence *fence;
1207 * Look up the fence and take a reference. We might still find a fence
1208 * whose refcount has already dropped to zero. dma_fence_get_rcu
1209 * pretends we didn't find a fence in that case.
1212 fence = idr_find(&gpu->fence_idr, id);
1214 fence = dma_fence_get_rcu(fence);
1221 /* No timeout was requested: just test for completion */
1222 ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY;
1224 unsigned long remaining = etnaviv_timeout_to_jiffies(timeout);
1226 ret = dma_fence_wait_timeout(fence, true, remaining);
1229 else if (ret != -ERESTARTSYS)
1234 dma_fence_put(fence);
1239 * Wait for an object to become inactive. This, on it's own, is not race
1240 * free: the object is moved by the scheduler off the active list, and
1241 * then the iova is put. Moreover, the object could be re-submitted just
1242 * after we notice that it's become inactive.
1244 * Although the retirement happens under the gpu lock, we don't want to hold
1245 * that lock in this function while waiting.
1247 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu,
1248 struct etnaviv_gem_object *etnaviv_obj,
1249 struct drm_etnaviv_timespec *timeout)
1251 unsigned long remaining;
1255 return !is_active(etnaviv_obj) ? 0 : -EBUSY;
1257 remaining = etnaviv_timeout_to_jiffies(timeout);
1259 ret = wait_event_interruptible_timeout(gpu->fence_event,
1260 !is_active(etnaviv_obj),
1264 else if (ret == -ERESTARTSYS)
1265 return -ERESTARTSYS;
1270 static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu,
1271 struct etnaviv_event *event, unsigned int flags)
1273 const struct etnaviv_gem_submit *submit = event->submit;
1276 for (i = 0; i < submit->nr_pmrs; i++) {
1277 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1279 if (pmr->flags == flags)
1280 etnaviv_perfmon_process(gpu, pmr, submit->exec_state);
1284 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu,
1285 struct etnaviv_event *event)
1289 /* disable clock gating */
1290 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1291 val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1292 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1294 /* enable debug register */
1295 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1296 val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1297 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1299 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE);
1302 static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu,
1303 struct etnaviv_event *event)
1305 const struct etnaviv_gem_submit *submit = event->submit;
1309 sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST);
1311 for (i = 0; i < submit->nr_pmrs; i++) {
1312 const struct etnaviv_perfmon_request *pmr = submit->pmrs + i;
1314 *pmr->bo_vma = pmr->sequence;
1317 /* disable debug register */
1318 val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL);
1319 val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS;
1320 gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val);
1322 /* enable clock gating */
1323 val = gpu_read(gpu, VIVS_PM_POWER_CONTROLS);
1324 val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING;
1325 gpu_write(gpu, VIVS_PM_POWER_CONTROLS, val);
1329 /* add bo's to gpu's ring, and kick gpu: */
1330 struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit)
1332 struct etnaviv_gpu *gpu = submit->gpu;
1333 struct dma_fence *gpu_fence;
1334 unsigned int i, nr_events = 1, event[3];
1337 if (!submit->runtime_resumed) {
1338 ret = pm_runtime_get_sync(gpu->dev);
1340 pm_runtime_put_noidle(gpu->dev);
1343 submit->runtime_resumed = true;
1347 * if there are performance monitor requests we need to have
1348 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1350 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1351 * and update the sequence number for userspace.
1353 if (submit->nr_pmrs)
1356 ret = event_alloc(gpu, nr_events, event);
1358 DRM_ERROR("no free events\n");
1359 pm_runtime_put_noidle(gpu->dev);
1363 mutex_lock(&gpu->lock);
1365 gpu_fence = etnaviv_gpu_fence_alloc(gpu);
1367 for (i = 0; i < nr_events; i++)
1368 event_free(gpu, event[i]);
1373 if (!gpu->mmu_context) {
1374 etnaviv_iommu_context_get(submit->mmu_context);
1375 gpu->mmu_context = submit->mmu_context;
1376 etnaviv_gpu_start_fe_idleloop(gpu);
1378 etnaviv_iommu_context_get(gpu->mmu_context);
1379 submit->prev_mmu_context = gpu->mmu_context;
1382 if (submit->nr_pmrs) {
1383 gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre;
1384 kref_get(&submit->refcount);
1385 gpu->event[event[1]].submit = submit;
1386 etnaviv_sync_point_queue(gpu, event[1]);
1389 gpu->event[event[0]].fence = gpu_fence;
1390 submit->cmdbuf.user_size = submit->cmdbuf.size - 8;
1391 etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context,
1392 event[0], &submit->cmdbuf);
1394 if (submit->nr_pmrs) {
1395 gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post;
1396 kref_get(&submit->refcount);
1397 gpu->event[event[2]].submit = submit;
1398 etnaviv_sync_point_queue(gpu, event[2]);
1402 mutex_unlock(&gpu->lock);
1407 static void sync_point_worker(struct work_struct *work)
1409 struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu,
1411 struct etnaviv_event *event = &gpu->event[gpu->sync_point_event];
1412 u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS);
1414 event->sync_point(gpu, event);
1415 etnaviv_submit_put(event->submit);
1416 event_free(gpu, gpu->sync_point_event);
1418 /* restart FE last to avoid GPU and IRQ racing against this worker */
1419 etnaviv_gpu_start_fe(gpu, addr + 2, 2);
1422 static void dump_mmu_fault(struct etnaviv_gpu *gpu)
1424 u32 status_reg, status;
1427 if (gpu->sec_mode == ETNA_SEC_NONE)
1428 status_reg = VIVS_MMUv2_STATUS;
1430 status_reg = VIVS_MMUv2_SEC_STATUS;
1432 status = gpu_read(gpu, status_reg);
1433 dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status);
1435 for (i = 0; i < 4; i++) {
1438 if (!(status & (VIVS_MMUv2_STATUS_EXCEPTION0__MASK << (i * 4))))
1441 if (gpu->sec_mode == ETNA_SEC_NONE)
1442 address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i);
1444 address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR;
1446 dev_err_ratelimited(gpu->dev, "MMU %d fault addr 0x%08x\n", i,
1447 gpu_read(gpu, address_reg));
1451 static irqreturn_t irq_handler(int irq, void *data)
1453 struct etnaviv_gpu *gpu = data;
1454 irqreturn_t ret = IRQ_NONE;
1456 u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE);
1461 pm_runtime_mark_last_busy(gpu->dev);
1463 dev_dbg(gpu->dev, "intr 0x%08x\n", intr);
1465 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) {
1466 dev_err(gpu->dev, "AXI bus error\n");
1467 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR;
1470 if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) {
1471 dump_mmu_fault(gpu);
1472 intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION;
1475 while ((event = ffs(intr)) != 0) {
1476 struct dma_fence *fence;
1480 intr &= ~(1 << event);
1482 dev_dbg(gpu->dev, "event %u\n", event);
1484 if (gpu->event[event].sync_point) {
1485 gpu->sync_point_event = event;
1486 queue_work(gpu->wq, &gpu->sync_point_work);
1489 fence = gpu->event[event].fence;
1493 gpu->event[event].fence = NULL;
1496 * Events can be processed out of order. Eg,
1497 * - allocate and queue event 0
1498 * - allocate event 1
1499 * - event 0 completes, we process it
1500 * - allocate and queue event 0
1501 * - event 1 and event 0 complete
1502 * we can end up processing event 0 first, then 1.
1504 if (fence_after(fence->seqno, gpu->completed_fence))
1505 gpu->completed_fence = fence->seqno;
1506 dma_fence_signal(fence);
1508 event_free(gpu, event);
1517 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu)
1521 ret = clk_prepare_enable(gpu->clk_reg);
1525 ret = clk_prepare_enable(gpu->clk_bus);
1527 goto disable_clk_reg;
1529 ret = clk_prepare_enable(gpu->clk_core);
1531 goto disable_clk_bus;
1533 ret = clk_prepare_enable(gpu->clk_shader);
1535 goto disable_clk_core;
1540 clk_disable_unprepare(gpu->clk_core);
1542 clk_disable_unprepare(gpu->clk_bus);
1544 clk_disable_unprepare(gpu->clk_reg);
1549 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu)
1551 clk_disable_unprepare(gpu->clk_shader);
1552 clk_disable_unprepare(gpu->clk_core);
1553 clk_disable_unprepare(gpu->clk_bus);
1554 clk_disable_unprepare(gpu->clk_reg);
1559 int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms)
1561 unsigned long timeout = jiffies + msecs_to_jiffies(timeout_ms);
1564 u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE);
1566 if ((idle & gpu->idle_mask) == gpu->idle_mask)
1569 if (time_is_before_jiffies(timeout)) {
1571 "timed out waiting for idle: idle=0x%x\n",
1580 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu)
1582 if (gpu->initialized && gpu->mmu_context) {
1583 /* Replace the last WAIT with END */
1584 mutex_lock(&gpu->lock);
1585 etnaviv_buffer_end(gpu);
1586 mutex_unlock(&gpu->lock);
1589 * We know that only the FE is busy here, this should
1590 * happen quickly (as the WAIT is only 200 cycles). If
1591 * we fail, just warn and continue.
1593 etnaviv_gpu_wait_idle(gpu, 100);
1595 etnaviv_iommu_context_put(gpu->mmu_context);
1596 gpu->mmu_context = NULL;
1599 gpu->exec_state = -1;
1601 return etnaviv_gpu_clk_disable(gpu);
1605 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu)
1609 ret = mutex_lock_killable(&gpu->lock);
1613 etnaviv_gpu_update_clock(gpu);
1614 etnaviv_gpu_hw_init(gpu);
1616 mutex_unlock(&gpu->lock);
1623 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev,
1624 unsigned long *state)
1632 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev,
1633 unsigned long *state)
1635 struct etnaviv_gpu *gpu = cdev->devdata;
1637 *state = gpu->freq_scale;
1643 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev,
1644 unsigned long state)
1646 struct etnaviv_gpu *gpu = cdev->devdata;
1648 mutex_lock(&gpu->lock);
1649 gpu->freq_scale = state;
1650 if (!pm_runtime_suspended(gpu->dev))
1651 etnaviv_gpu_update_clock(gpu);
1652 mutex_unlock(&gpu->lock);
1657 static struct thermal_cooling_device_ops cooling_ops = {
1658 .get_max_state = etnaviv_gpu_cooling_get_max_state,
1659 .get_cur_state = etnaviv_gpu_cooling_get_cur_state,
1660 .set_cur_state = etnaviv_gpu_cooling_set_cur_state,
1663 static int etnaviv_gpu_bind(struct device *dev, struct device *master,
1666 struct drm_device *drm = data;
1667 struct etnaviv_drm_private *priv = drm->dev_private;
1668 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1671 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) {
1672 gpu->cooling = thermal_of_cooling_device_register(dev->of_node,
1673 (char *)dev_name(dev), gpu, &cooling_ops);
1674 if (IS_ERR(gpu->cooling))
1675 return PTR_ERR(gpu->cooling);
1678 gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0);
1684 ret = etnaviv_sched_init(gpu);
1689 ret = pm_runtime_get_sync(gpu->dev);
1691 ret = etnaviv_gpu_clk_enable(gpu);
1698 gpu->fence_context = dma_fence_context_alloc(1);
1699 idr_init(&gpu->fence_idr);
1700 spin_lock_init(&gpu->fence_spinlock);
1702 INIT_WORK(&gpu->sync_point_work, sync_point_worker);
1703 init_waitqueue_head(&gpu->fence_event);
1705 priv->gpu[priv->num_gpus++] = gpu;
1707 pm_runtime_mark_last_busy(gpu->dev);
1708 pm_runtime_put_autosuspend(gpu->dev);
1713 etnaviv_sched_fini(gpu);
1716 destroy_workqueue(gpu->wq);
1719 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1720 thermal_cooling_device_unregister(gpu->cooling);
1725 static void etnaviv_gpu_unbind(struct device *dev, struct device *master,
1728 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1730 DBG("%s", dev_name(gpu->dev));
1732 flush_workqueue(gpu->wq);
1733 destroy_workqueue(gpu->wq);
1735 etnaviv_sched_fini(gpu);
1738 pm_runtime_get_sync(gpu->dev);
1739 pm_runtime_put_sync_suspend(gpu->dev);
1741 etnaviv_gpu_hw_suspend(gpu);
1744 if (gpu->initialized) {
1745 etnaviv_cmdbuf_free(&gpu->buffer);
1746 etnaviv_iommu_global_fini(gpu);
1747 gpu->initialized = false;
1751 idr_destroy(&gpu->fence_idr);
1753 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL))
1754 thermal_cooling_device_unregister(gpu->cooling);
1755 gpu->cooling = NULL;
1758 static const struct component_ops gpu_ops = {
1759 .bind = etnaviv_gpu_bind,
1760 .unbind = etnaviv_gpu_unbind,
1763 static const struct of_device_id etnaviv_gpu_match[] = {
1765 .compatible = "vivante,gc"
1769 MODULE_DEVICE_TABLE(of, etnaviv_gpu_match);
1771 static int etnaviv_gpu_platform_probe(struct platform_device *pdev)
1773 struct device *dev = &pdev->dev;
1774 struct etnaviv_gpu *gpu;
1777 gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL);
1781 gpu->dev = &pdev->dev;
1782 mutex_init(&gpu->lock);
1783 mutex_init(&gpu->fence_lock);
1785 /* Map registers: */
1786 gpu->mmio = devm_platform_ioremap_resource(pdev, 0);
1787 if (IS_ERR(gpu->mmio))
1788 return PTR_ERR(gpu->mmio);
1790 /* Get Interrupt: */
1791 gpu->irq = platform_get_irq(pdev, 0);
1795 err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0,
1796 dev_name(gpu->dev), gpu);
1798 dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err);
1803 gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg");
1804 DBG("clk_reg: %p", gpu->clk_reg);
1805 if (IS_ERR(gpu->clk_reg))
1806 return PTR_ERR(gpu->clk_reg);
1808 gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus");
1809 DBG("clk_bus: %p", gpu->clk_bus);
1810 if (IS_ERR(gpu->clk_bus))
1811 return PTR_ERR(gpu->clk_bus);
1813 gpu->clk_core = devm_clk_get(&pdev->dev, "core");
1814 DBG("clk_core: %p", gpu->clk_core);
1815 if (IS_ERR(gpu->clk_core))
1816 return PTR_ERR(gpu->clk_core);
1817 gpu->base_rate_core = clk_get_rate(gpu->clk_core);
1819 gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader");
1820 DBG("clk_shader: %p", gpu->clk_shader);
1821 if (IS_ERR(gpu->clk_shader))
1822 return PTR_ERR(gpu->clk_shader);
1823 gpu->base_rate_shader = clk_get_rate(gpu->clk_shader);
1825 /* TODO: figure out max mapped size */
1826 dev_set_drvdata(dev, gpu);
1829 * We treat the device as initially suspended. The runtime PM
1830 * autosuspend delay is rather arbitary: no measurements have
1831 * yet been performed to determine an appropriate value.
1833 pm_runtime_use_autosuspend(gpu->dev);
1834 pm_runtime_set_autosuspend_delay(gpu->dev, 200);
1835 pm_runtime_enable(gpu->dev);
1837 err = component_add(&pdev->dev, &gpu_ops);
1839 dev_err(&pdev->dev, "failed to register component: %d\n", err);
1846 static int etnaviv_gpu_platform_remove(struct platform_device *pdev)
1848 component_del(&pdev->dev, &gpu_ops);
1849 pm_runtime_disable(&pdev->dev);
1854 static int etnaviv_gpu_rpm_suspend(struct device *dev)
1856 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1859 /* If there are any jobs in the HW queue, we're not idle */
1860 if (atomic_read(&gpu->sched.hw_rq_count))
1863 /* Check whether the hardware (except FE and MC) is idle */
1864 mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE |
1865 VIVS_HI_IDLE_STATE_MC);
1866 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask;
1868 dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n",
1873 return etnaviv_gpu_hw_suspend(gpu);
1876 static int etnaviv_gpu_rpm_resume(struct device *dev)
1878 struct etnaviv_gpu *gpu = dev_get_drvdata(dev);
1881 ret = etnaviv_gpu_clk_enable(gpu);
1885 /* Re-initialise the basic hardware state */
1886 if (gpu->drm && gpu->initialized) {
1887 ret = etnaviv_gpu_hw_resume(gpu);
1889 etnaviv_gpu_clk_disable(gpu);
1898 static const struct dev_pm_ops etnaviv_gpu_pm_ops = {
1899 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume,
1903 struct platform_driver etnaviv_gpu_driver = {
1905 .name = "etnaviv-gpu",
1906 .owner = THIS_MODULE,
1907 .pm = &etnaviv_gpu_pm_ops,
1908 .of_match_table = etnaviv_gpu_match,
1910 .probe = etnaviv_gpu_platform_probe,
1911 .remove = etnaviv_gpu_platform_remove,
1912 .id_table = gpu_ids,