1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2015-2018 Etnaviv Project
6 #include <linux/component.h>
7 #include <linux/dma-fence.h>
8 #include <linux/moduleparam.h>
9 #include <linux/of_device.h>
10 #include <linux/thermal.h>
12 #include "etnaviv_cmdbuf.h"
13 #include "etnaviv_dump.h"
14 #include "etnaviv_gpu.h"
15 #include "etnaviv_gem.h"
16 #include "etnaviv_mmu.h"
17 #include "etnaviv_perfmon.h"
18 #include "etnaviv_sched.h"
19 #include "common.xml.h"
20 #include "state.xml.h"
21 #include "state_hi.xml.h"
22 #include "cmdstream.xml.h"
28 static const struct platform_device_id gpu_ids
[] = {
29 { .name
= "etnaviv-gpu,2d" },
37 int etnaviv_gpu_get_param(struct etnaviv_gpu
*gpu
, u32 param
, u64
*value
)
40 case ETNAVIV_PARAM_GPU_MODEL
:
41 *value
= gpu
->identity
.model
;
44 case ETNAVIV_PARAM_GPU_REVISION
:
45 *value
= gpu
->identity
.revision
;
48 case ETNAVIV_PARAM_GPU_FEATURES_0
:
49 *value
= gpu
->identity
.features
;
52 case ETNAVIV_PARAM_GPU_FEATURES_1
:
53 *value
= gpu
->identity
.minor_features0
;
56 case ETNAVIV_PARAM_GPU_FEATURES_2
:
57 *value
= gpu
->identity
.minor_features1
;
60 case ETNAVIV_PARAM_GPU_FEATURES_3
:
61 *value
= gpu
->identity
.minor_features2
;
64 case ETNAVIV_PARAM_GPU_FEATURES_4
:
65 *value
= gpu
->identity
.minor_features3
;
68 case ETNAVIV_PARAM_GPU_FEATURES_5
:
69 *value
= gpu
->identity
.minor_features4
;
72 case ETNAVIV_PARAM_GPU_FEATURES_6
:
73 *value
= gpu
->identity
.minor_features5
;
76 case ETNAVIV_PARAM_GPU_FEATURES_7
:
77 *value
= gpu
->identity
.minor_features6
;
80 case ETNAVIV_PARAM_GPU_FEATURES_8
:
81 *value
= gpu
->identity
.minor_features7
;
84 case ETNAVIV_PARAM_GPU_FEATURES_9
:
85 *value
= gpu
->identity
.minor_features8
;
88 case ETNAVIV_PARAM_GPU_FEATURES_10
:
89 *value
= gpu
->identity
.minor_features9
;
92 case ETNAVIV_PARAM_GPU_FEATURES_11
:
93 *value
= gpu
->identity
.minor_features10
;
96 case ETNAVIV_PARAM_GPU_FEATURES_12
:
97 *value
= gpu
->identity
.minor_features11
;
100 case ETNAVIV_PARAM_GPU_STREAM_COUNT
:
101 *value
= gpu
->identity
.stream_count
;
104 case ETNAVIV_PARAM_GPU_REGISTER_MAX
:
105 *value
= gpu
->identity
.register_max
;
108 case ETNAVIV_PARAM_GPU_THREAD_COUNT
:
109 *value
= gpu
->identity
.thread_count
;
112 case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE
:
113 *value
= gpu
->identity
.vertex_cache_size
;
116 case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT
:
117 *value
= gpu
->identity
.shader_core_count
;
120 case ETNAVIV_PARAM_GPU_PIXEL_PIPES
:
121 *value
= gpu
->identity
.pixel_pipes
;
124 case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE
:
125 *value
= gpu
->identity
.vertex_output_buffer_size
;
128 case ETNAVIV_PARAM_GPU_BUFFER_SIZE
:
129 *value
= gpu
->identity
.buffer_size
;
132 case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT
:
133 *value
= gpu
->identity
.instruction_count
;
136 case ETNAVIV_PARAM_GPU_NUM_CONSTANTS
:
137 *value
= gpu
->identity
.num_constants
;
140 case ETNAVIV_PARAM_GPU_NUM_VARYINGS
:
141 *value
= gpu
->identity
.varyings_count
;
145 DBG("%s: invalid param: %u", dev_name(gpu
->dev
), param
);
153 #define etnaviv_is_model_rev(gpu, mod, rev) \
154 ((gpu)->identity.model == chipModel_##mod && \
155 (gpu)->identity.revision == rev)
156 #define etnaviv_field(val, field) \
157 (((val) & field##__MASK) >> field##__SHIFT)
159 static void etnaviv_hw_specs(struct etnaviv_gpu
*gpu
)
161 if (gpu
->identity
.minor_features0
&
162 chipMinorFeatures0_MORE_MINOR_FEATURES
) {
164 unsigned int streams
;
166 specs
[0] = gpu_read(gpu
, VIVS_HI_CHIP_SPECS
);
167 specs
[1] = gpu_read(gpu
, VIVS_HI_CHIP_SPECS_2
);
168 specs
[2] = gpu_read(gpu
, VIVS_HI_CHIP_SPECS_3
);
169 specs
[3] = gpu_read(gpu
, VIVS_HI_CHIP_SPECS_4
);
171 gpu
->identity
.stream_count
= etnaviv_field(specs
[0],
172 VIVS_HI_CHIP_SPECS_STREAM_COUNT
);
173 gpu
->identity
.register_max
= etnaviv_field(specs
[0],
174 VIVS_HI_CHIP_SPECS_REGISTER_MAX
);
175 gpu
->identity
.thread_count
= etnaviv_field(specs
[0],
176 VIVS_HI_CHIP_SPECS_THREAD_COUNT
);
177 gpu
->identity
.vertex_cache_size
= etnaviv_field(specs
[0],
178 VIVS_HI_CHIP_SPECS_VERTEX_CACHE_SIZE
);
179 gpu
->identity
.shader_core_count
= etnaviv_field(specs
[0],
180 VIVS_HI_CHIP_SPECS_SHADER_CORE_COUNT
);
181 gpu
->identity
.pixel_pipes
= etnaviv_field(specs
[0],
182 VIVS_HI_CHIP_SPECS_PIXEL_PIPES
);
183 gpu
->identity
.vertex_output_buffer_size
=
184 etnaviv_field(specs
[0],
185 VIVS_HI_CHIP_SPECS_VERTEX_OUTPUT_BUFFER_SIZE
);
187 gpu
->identity
.buffer_size
= etnaviv_field(specs
[1],
188 VIVS_HI_CHIP_SPECS_2_BUFFER_SIZE
);
189 gpu
->identity
.instruction_count
= etnaviv_field(specs
[1],
190 VIVS_HI_CHIP_SPECS_2_INSTRUCTION_COUNT
);
191 gpu
->identity
.num_constants
= etnaviv_field(specs
[1],
192 VIVS_HI_CHIP_SPECS_2_NUM_CONSTANTS
);
194 gpu
->identity
.varyings_count
= etnaviv_field(specs
[2],
195 VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT
);
197 /* This overrides the value from older register if non-zero */
198 streams
= etnaviv_field(specs
[3],
199 VIVS_HI_CHIP_SPECS_4_STREAM_COUNT
);
201 gpu
->identity
.stream_count
= streams
;
204 /* Fill in the stream count if not specified */
205 if (gpu
->identity
.stream_count
== 0) {
206 if (gpu
->identity
.model
>= 0x1000)
207 gpu
->identity
.stream_count
= 4;
209 gpu
->identity
.stream_count
= 1;
212 /* Convert the register max value */
213 if (gpu
->identity
.register_max
)
214 gpu
->identity
.register_max
= 1 << gpu
->identity
.register_max
;
215 else if (gpu
->identity
.model
== chipModel_GC400
)
216 gpu
->identity
.register_max
= 32;
218 gpu
->identity
.register_max
= 64;
220 /* Convert thread count */
221 if (gpu
->identity
.thread_count
)
222 gpu
->identity
.thread_count
= 1 << gpu
->identity
.thread_count
;
223 else if (gpu
->identity
.model
== chipModel_GC400
)
224 gpu
->identity
.thread_count
= 64;
225 else if (gpu
->identity
.model
== chipModel_GC500
||
226 gpu
->identity
.model
== chipModel_GC530
)
227 gpu
->identity
.thread_count
= 128;
229 gpu
->identity
.thread_count
= 256;
231 if (gpu
->identity
.vertex_cache_size
== 0)
232 gpu
->identity
.vertex_cache_size
= 8;
234 if (gpu
->identity
.shader_core_count
== 0) {
235 if (gpu
->identity
.model
>= 0x1000)
236 gpu
->identity
.shader_core_count
= 2;
238 gpu
->identity
.shader_core_count
= 1;
241 if (gpu
->identity
.pixel_pipes
== 0)
242 gpu
->identity
.pixel_pipes
= 1;
244 /* Convert virtex buffer size */
245 if (gpu
->identity
.vertex_output_buffer_size
) {
246 gpu
->identity
.vertex_output_buffer_size
=
247 1 << gpu
->identity
.vertex_output_buffer_size
;
248 } else if (gpu
->identity
.model
== chipModel_GC400
) {
249 if (gpu
->identity
.revision
< 0x4000)
250 gpu
->identity
.vertex_output_buffer_size
= 512;
251 else if (gpu
->identity
.revision
< 0x4200)
252 gpu
->identity
.vertex_output_buffer_size
= 256;
254 gpu
->identity
.vertex_output_buffer_size
= 128;
256 gpu
->identity
.vertex_output_buffer_size
= 512;
259 switch (gpu
->identity
.instruction_count
) {
261 if (etnaviv_is_model_rev(gpu
, GC2000
, 0x5108) ||
262 gpu
->identity
.model
== chipModel_GC880
)
263 gpu
->identity
.instruction_count
= 512;
265 gpu
->identity
.instruction_count
= 256;
269 gpu
->identity
.instruction_count
= 1024;
273 gpu
->identity
.instruction_count
= 2048;
277 gpu
->identity
.instruction_count
= 256;
281 if (gpu
->identity
.num_constants
== 0)
282 gpu
->identity
.num_constants
= 168;
284 if (gpu
->identity
.varyings_count
== 0) {
285 if (gpu
->identity
.minor_features1
& chipMinorFeatures1_HALTI0
)
286 gpu
->identity
.varyings_count
= 12;
288 gpu
->identity
.varyings_count
= 8;
292 * For some cores, two varyings are consumed for position, so the
293 * maximum varying count needs to be reduced by one.
295 if (etnaviv_is_model_rev(gpu
, GC5000
, 0x5434) ||
296 etnaviv_is_model_rev(gpu
, GC4000
, 0x5222) ||
297 etnaviv_is_model_rev(gpu
, GC4000
, 0x5245) ||
298 etnaviv_is_model_rev(gpu
, GC4000
, 0x5208) ||
299 etnaviv_is_model_rev(gpu
, GC3000
, 0x5435) ||
300 etnaviv_is_model_rev(gpu
, GC2200
, 0x5244) ||
301 etnaviv_is_model_rev(gpu
, GC2100
, 0x5108) ||
302 etnaviv_is_model_rev(gpu
, GC2000
, 0x5108) ||
303 etnaviv_is_model_rev(gpu
, GC1500
, 0x5246) ||
304 etnaviv_is_model_rev(gpu
, GC880
, 0x5107) ||
305 etnaviv_is_model_rev(gpu
, GC880
, 0x5106))
306 gpu
->identity
.varyings_count
-= 1;
309 static void etnaviv_hw_identify(struct etnaviv_gpu
*gpu
)
313 chipIdentity
= gpu_read(gpu
, VIVS_HI_CHIP_IDENTITY
);
315 /* Special case for older graphic cores. */
316 if (etnaviv_field(chipIdentity
, VIVS_HI_CHIP_IDENTITY_FAMILY
) == 0x01) {
317 gpu
->identity
.model
= chipModel_GC500
;
318 gpu
->identity
.revision
= etnaviv_field(chipIdentity
,
319 VIVS_HI_CHIP_IDENTITY_REVISION
);
322 gpu
->identity
.model
= gpu_read(gpu
, VIVS_HI_CHIP_MODEL
);
323 gpu
->identity
.revision
= gpu_read(gpu
, VIVS_HI_CHIP_REV
);
326 * !!!! HACK ALERT !!!!
327 * Because people change device IDs without letting software
328 * know about it - here is the hack to make it all look the
329 * same. Only for GC400 family.
331 if ((gpu
->identity
.model
& 0xff00) == 0x0400 &&
332 gpu
->identity
.model
!= chipModel_GC420
) {
333 gpu
->identity
.model
= gpu
->identity
.model
& 0x0400;
336 /* Another special case */
337 if (etnaviv_is_model_rev(gpu
, GC300
, 0x2201)) {
338 u32 chipDate
= gpu_read(gpu
, VIVS_HI_CHIP_DATE
);
339 u32 chipTime
= gpu_read(gpu
, VIVS_HI_CHIP_TIME
);
341 if (chipDate
== 0x20080814 && chipTime
== 0x12051100) {
343 * This IP has an ECO; put the correct
346 gpu
->identity
.revision
= 0x1051;
351 * NXP likes to call the GPU on the i.MX6QP GC2000+, but in
352 * reality it's just a re-branded GC3000. We can identify this
353 * core by the upper half of the revision register being all 1.
354 * Fix model/rev here, so all other places can refer to this
355 * core by its real identity.
357 if (etnaviv_is_model_rev(gpu
, GC2000
, 0xffff5450)) {
358 gpu
->identity
.model
= chipModel_GC3000
;
359 gpu
->identity
.revision
&= 0xffff;
363 dev_info(gpu
->dev
, "model: GC%x, revision: %x\n",
364 gpu
->identity
.model
, gpu
->identity
.revision
);
367 * If there is a match in the HWDB, we aren't interested in the
368 * remaining register values, as they might be wrong.
370 if (etnaviv_fill_identity_from_hwdb(gpu
))
373 gpu
->identity
.features
= gpu_read(gpu
, VIVS_HI_CHIP_FEATURE
);
375 /* Disable fast clear on GC700. */
376 if (gpu
->identity
.model
== chipModel_GC700
)
377 gpu
->identity
.features
&= ~chipFeatures_FAST_CLEAR
;
379 if ((gpu
->identity
.model
== chipModel_GC500
&&
380 gpu
->identity
.revision
< 2) ||
381 (gpu
->identity
.model
== chipModel_GC300
&&
382 gpu
->identity
.revision
< 0x2000)) {
385 * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these
388 gpu
->identity
.minor_features0
= 0;
389 gpu
->identity
.minor_features1
= 0;
390 gpu
->identity
.minor_features2
= 0;
391 gpu
->identity
.minor_features3
= 0;
392 gpu
->identity
.minor_features4
= 0;
393 gpu
->identity
.minor_features5
= 0;
395 gpu
->identity
.minor_features0
=
396 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_0
);
398 if (gpu
->identity
.minor_features0
&
399 chipMinorFeatures0_MORE_MINOR_FEATURES
) {
400 gpu
->identity
.minor_features1
=
401 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_1
);
402 gpu
->identity
.minor_features2
=
403 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_2
);
404 gpu
->identity
.minor_features3
=
405 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_3
);
406 gpu
->identity
.minor_features4
=
407 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_4
);
408 gpu
->identity
.minor_features5
=
409 gpu_read(gpu
, VIVS_HI_CHIP_MINOR_FEATURE_5
);
412 /* GC600 idle register reports zero bits where modules aren't present */
413 if (gpu
->identity
.model
== chipModel_GC600
) {
414 gpu
->idle_mask
= VIVS_HI_IDLE_STATE_TX
|
415 VIVS_HI_IDLE_STATE_RA
|
416 VIVS_HI_IDLE_STATE_SE
|
417 VIVS_HI_IDLE_STATE_PA
|
418 VIVS_HI_IDLE_STATE_SH
|
419 VIVS_HI_IDLE_STATE_PE
|
420 VIVS_HI_IDLE_STATE_DE
|
421 VIVS_HI_IDLE_STATE_FE
;
423 gpu
->idle_mask
= ~VIVS_HI_IDLE_STATE_AXI_LP
;
426 etnaviv_hw_specs(gpu
);
429 static void etnaviv_gpu_load_clock(struct etnaviv_gpu
*gpu
, u32 clock
)
431 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, clock
|
432 VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD
);
433 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, clock
);
436 static void etnaviv_gpu_update_clock(struct etnaviv_gpu
*gpu
)
438 if (gpu
->identity
.minor_features2
&
439 chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING
) {
440 clk_set_rate(gpu
->clk_core
,
441 gpu
->base_rate_core
>> gpu
->freq_scale
);
442 clk_set_rate(gpu
->clk_shader
,
443 gpu
->base_rate_shader
>> gpu
->freq_scale
);
445 unsigned int fscale
= 1 << (6 - gpu
->freq_scale
);
446 u32 clock
= gpu_read(gpu
, VIVS_HI_CLOCK_CONTROL
);
448 clock
&= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK
;
449 clock
|= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale
);
450 etnaviv_gpu_load_clock(gpu
, clock
);
454 static int etnaviv_hw_reset(struct etnaviv_gpu
*gpu
)
457 unsigned long timeout
;
460 /* We hope that the GPU resets in under one second */
461 timeout
= jiffies
+ msecs_to_jiffies(1000);
463 while (time_is_after_jiffies(timeout
)) {
465 unsigned int fscale
= 1 << (6 - gpu
->freq_scale
);
466 control
= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale
);
467 etnaviv_gpu_load_clock(gpu
, control
);
469 /* isolate the GPU. */
470 control
|= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU
;
471 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, control
);
473 if (gpu
->sec_mode
== ETNA_SEC_KERNEL
) {
474 gpu_write(gpu
, VIVS_MMUv2_AHB_CONTROL
,
475 VIVS_MMUv2_AHB_CONTROL_RESET
);
477 /* set soft reset. */
478 control
|= VIVS_HI_CLOCK_CONTROL_SOFT_RESET
;
479 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, control
);
482 /* wait for reset. */
483 usleep_range(10, 20);
485 /* reset soft reset bit. */
486 control
&= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET
;
487 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, control
);
489 /* reset GPU isolation. */
490 control
&= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU
;
491 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, control
);
493 /* read idle register. */
494 idle
= gpu_read(gpu
, VIVS_HI_IDLE_STATE
);
496 /* try reseting again if FE it not idle */
497 if ((idle
& VIVS_HI_IDLE_STATE_FE
) == 0) {
498 dev_dbg(gpu
->dev
, "FE is not idle\n");
502 /* read reset register. */
503 control
= gpu_read(gpu
, VIVS_HI_CLOCK_CONTROL
);
505 /* is the GPU idle? */
506 if (((control
& VIVS_HI_CLOCK_CONTROL_IDLE_3D
) == 0) ||
507 ((control
& VIVS_HI_CLOCK_CONTROL_IDLE_2D
) == 0)) {
508 dev_dbg(gpu
->dev
, "GPU is not idle\n");
512 /* disable debug registers, as they are not normally needed */
513 control
|= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS
;
514 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, control
);
521 idle
= gpu_read(gpu
, VIVS_HI_IDLE_STATE
);
522 control
= gpu_read(gpu
, VIVS_HI_CLOCK_CONTROL
);
524 dev_err(gpu
->dev
, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n",
525 idle
& VIVS_HI_IDLE_STATE_FE
? "" : "not ",
526 control
& VIVS_HI_CLOCK_CONTROL_IDLE_3D
? "" : "not ",
527 control
& VIVS_HI_CLOCK_CONTROL_IDLE_2D
? "" : "not ");
532 /* We rely on the GPU running, so program the clock */
533 etnaviv_gpu_update_clock(gpu
);
538 static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu
*gpu
)
542 /* enable clock gating */
543 ppc
= gpu_read(gpu
, VIVS_PM_POWER_CONTROLS
);
544 ppc
|= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING
;
546 /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */
547 if (gpu
->identity
.revision
== 0x4301 ||
548 gpu
->identity
.revision
== 0x4302)
549 ppc
|= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING
;
551 gpu_write(gpu
, VIVS_PM_POWER_CONTROLS
, ppc
);
553 pmc
= gpu_read(gpu
, VIVS_PM_MODULE_CONTROLS
);
555 /* Disable PA clock gating for GC400+ without bugfix except for GC420 */
556 if (gpu
->identity
.model
>= chipModel_GC400
&&
557 gpu
->identity
.model
!= chipModel_GC420
&&
558 !(gpu
->identity
.minor_features3
& chipMinorFeatures3_BUG_FIXES12
))
559 pmc
|= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA
;
562 * Disable PE clock gating on revs < 5.0.0.0 when HZ is
563 * present without a bug fix.
565 if (gpu
->identity
.revision
< 0x5000 &&
566 gpu
->identity
.minor_features0
& chipMinorFeatures0_HZ
&&
567 !(gpu
->identity
.minor_features1
&
568 chipMinorFeatures1_DISABLE_PE_GATING
))
569 pmc
|= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE
;
571 if (gpu
->identity
.revision
< 0x5422)
572 pmc
|= BIT(15); /* Unknown bit */
574 /* Disable TX clock gating on affected core revisions. */
575 if (etnaviv_is_model_rev(gpu
, GC4000
, 0x5222) ||
576 etnaviv_is_model_rev(gpu
, GC2000
, 0x5108))
577 pmc
|= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX
;
579 pmc
|= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ
;
580 pmc
|= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ
;
582 gpu_write(gpu
, VIVS_PM_MODULE_CONTROLS
, pmc
);
585 void etnaviv_gpu_start_fe(struct etnaviv_gpu
*gpu
, u32 address
, u16 prefetch
)
587 gpu_write(gpu
, VIVS_FE_COMMAND_ADDRESS
, address
);
588 gpu_write(gpu
, VIVS_FE_COMMAND_CONTROL
,
589 VIVS_FE_COMMAND_CONTROL_ENABLE
|
590 VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch
));
592 if (gpu
->sec_mode
== ETNA_SEC_KERNEL
) {
593 gpu_write(gpu
, VIVS_MMUv2_SEC_COMMAND_CONTROL
,
594 VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE
|
595 VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch
));
599 static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu
*gpu
)
602 * Base value for VIVS_PM_PULSE_EATER register on models where it
603 * cannot be read, extracted from vivante kernel driver.
605 u32 pulse_eater
= 0x01590880;
607 if (etnaviv_is_model_rev(gpu
, GC4000
, 0x5208) ||
608 etnaviv_is_model_rev(gpu
, GC4000
, 0x5222)) {
609 pulse_eater
|= BIT(23);
613 if (etnaviv_is_model_rev(gpu
, GC1000
, 0x5039) ||
614 etnaviv_is_model_rev(gpu
, GC1000
, 0x5040)) {
615 pulse_eater
&= ~BIT(16);
616 pulse_eater
|= BIT(17);
619 if ((gpu
->identity
.revision
> 0x5420) &&
620 (gpu
->identity
.features
& chipFeatures_PIPE_3D
))
622 /* Performance fix: disable internal DFS */
623 pulse_eater
= gpu_read(gpu
, VIVS_PM_PULSE_EATER
);
624 pulse_eater
|= BIT(18);
627 gpu_write(gpu
, VIVS_PM_PULSE_EATER
, pulse_eater
);
630 static void etnaviv_gpu_hw_init(struct etnaviv_gpu
*gpu
)
634 if ((etnaviv_is_model_rev(gpu
, GC320
, 0x5007) ||
635 etnaviv_is_model_rev(gpu
, GC320
, 0x5220)) &&
636 gpu_read(gpu
, VIVS_HI_CHIP_TIME
) != 0x2062400) {
639 mc_memory_debug
= gpu_read(gpu
, VIVS_MC_DEBUG_MEMORY
) & ~0xff;
641 if (gpu
->identity
.revision
== 0x5007)
642 mc_memory_debug
|= 0x0c;
644 mc_memory_debug
|= 0x08;
646 gpu_write(gpu
, VIVS_MC_DEBUG_MEMORY
, mc_memory_debug
);
649 /* enable module-level clock gating */
650 etnaviv_gpu_enable_mlcg(gpu
);
653 * Update GPU AXI cache atttribute to "cacheable, no allocate".
654 * This is necessary to prevent the iMX6 SoC locking up.
656 gpu_write(gpu
, VIVS_HI_AXI_CONFIG
,
657 VIVS_HI_AXI_CONFIG_AWCACHE(2) |
658 VIVS_HI_AXI_CONFIG_ARCACHE(2));
660 /* GC2000 rev 5108 needs a special bus config */
661 if (etnaviv_is_model_rev(gpu
, GC2000
, 0x5108)) {
662 u32 bus_config
= gpu_read(gpu
, VIVS_MC_BUS_CONFIG
);
663 bus_config
&= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK
|
664 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK
);
665 bus_config
|= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) |
666 VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0);
667 gpu_write(gpu
, VIVS_MC_BUS_CONFIG
, bus_config
);
670 if (gpu
->sec_mode
== ETNA_SEC_KERNEL
) {
671 u32 val
= gpu_read(gpu
, VIVS_MMUv2_AHB_CONTROL
);
672 val
|= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS
;
673 gpu_write(gpu
, VIVS_MMUv2_AHB_CONTROL
, val
);
676 /* setup the pulse eater */
677 etnaviv_gpu_setup_pulse_eater(gpu
);
680 etnaviv_iommu_restore(gpu
);
682 /* Start command processor */
683 prefetch
= etnaviv_buffer_init(gpu
);
685 gpu_write(gpu
, VIVS_HI_INTR_ENBL
, ~0U);
686 etnaviv_gpu_start_fe(gpu
, etnaviv_cmdbuf_get_va(&gpu
->buffer
),
690 int etnaviv_gpu_init(struct etnaviv_gpu
*gpu
)
694 ret
= pm_runtime_get_sync(gpu
->dev
);
696 dev_err(gpu
->dev
, "Failed to enable GPU power domain\n");
700 etnaviv_hw_identify(gpu
);
702 if (gpu
->identity
.model
== 0) {
703 dev_err(gpu
->dev
, "Unknown GPU model\n");
708 /* Exclude VG cores with FE2.0 */
709 if (gpu
->identity
.features
& chipFeatures_PIPE_VG
&&
710 gpu
->identity
.features
& chipFeatures_FE20
) {
711 dev_info(gpu
->dev
, "Ignoring GPU with VG and FE2.0\n");
717 * Set the GPU linear window to be at the end of the DMA window, where
718 * the CMA area is likely to reside. This ensures that we are able to
719 * map the command buffers while having the linear window overlap as
720 * much RAM as possible, so we can optimize mappings for other buffers.
722 * For 3D cores only do this if MC2.0 is present, as with MC1.0 it leads
723 * to different views of the memory on the individual engines.
725 if (!(gpu
->identity
.features
& chipFeatures_PIPE_3D
) ||
726 (gpu
->identity
.minor_features0
& chipMinorFeatures0_MC20
)) {
727 u32 dma_mask
= (u32
)dma_get_required_mask(gpu
->dev
);
728 if (dma_mask
< PHYS_OFFSET
+ SZ_2G
)
729 gpu
->memory_base
= PHYS_OFFSET
;
731 gpu
->memory_base
= dma_mask
- SZ_2G
+ 1;
732 } else if (PHYS_OFFSET
>= SZ_2G
) {
733 dev_info(gpu
->dev
, "Need to move linear window on MC1.0, disabling TS\n");
734 gpu
->memory_base
= PHYS_OFFSET
;
735 gpu
->identity
.features
&= ~chipFeatures_FAST_CLEAR
;
739 * On cores with security features supported, we claim control over the
742 if ((gpu
->identity
.minor_features7
& chipMinorFeatures7_BIT_SECURITY
) &&
743 (gpu
->identity
.minor_features10
& chipMinorFeatures10_SECURITY_AHB
))
744 gpu
->sec_mode
= ETNA_SEC_KERNEL
;
746 ret
= etnaviv_hw_reset(gpu
);
748 dev_err(gpu
->dev
, "GPU reset failed\n");
752 gpu
->mmu
= etnaviv_iommu_new(gpu
);
753 if (IS_ERR(gpu
->mmu
)) {
754 dev_err(gpu
->dev
, "Failed to instantiate GPU IOMMU\n");
755 ret
= PTR_ERR(gpu
->mmu
);
759 gpu
->cmdbuf_suballoc
= etnaviv_cmdbuf_suballoc_new(gpu
);
760 if (IS_ERR(gpu
->cmdbuf_suballoc
)) {
761 dev_err(gpu
->dev
, "Failed to create cmdbuf suballocator\n");
762 ret
= PTR_ERR(gpu
->cmdbuf_suballoc
);
767 ret
= etnaviv_cmdbuf_init(gpu
->cmdbuf_suballoc
, &gpu
->buffer
,
770 dev_err(gpu
->dev
, "could not create command buffer\n");
771 goto destroy_suballoc
;
774 if (gpu
->mmu
->version
== ETNAVIV_IOMMU_V1
&&
775 etnaviv_cmdbuf_get_va(&gpu
->buffer
) > 0x80000000) {
778 "command buffer outside valid memory window\n");
782 /* Setup event management */
783 spin_lock_init(&gpu
->event_spinlock
);
784 init_completion(&gpu
->event_free
);
785 bitmap_zero(gpu
->event_bitmap
, ETNA_NR_EVENTS
);
786 for (i
= 0; i
< ARRAY_SIZE(gpu
->event
); i
++)
787 complete(&gpu
->event_free
);
789 /* Now program the hardware */
790 mutex_lock(&gpu
->lock
);
791 etnaviv_gpu_hw_init(gpu
);
792 gpu
->exec_state
= -1;
793 mutex_unlock(&gpu
->lock
);
795 pm_runtime_mark_last_busy(gpu
->dev
);
796 pm_runtime_put_autosuspend(gpu
->dev
);
801 etnaviv_cmdbuf_free(&gpu
->buffer
);
802 gpu
->buffer
.suballoc
= NULL
;
804 etnaviv_cmdbuf_suballoc_destroy(gpu
->cmdbuf_suballoc
);
805 gpu
->cmdbuf_suballoc
= NULL
;
807 etnaviv_iommu_destroy(gpu
->mmu
);
810 pm_runtime_mark_last_busy(gpu
->dev
);
811 pm_runtime_put_autosuspend(gpu
->dev
);
816 #ifdef CONFIG_DEBUG_FS
822 static void verify_dma(struct etnaviv_gpu
*gpu
, struct dma_debug
*debug
)
826 debug
->address
[0] = gpu_read(gpu
, VIVS_FE_DMA_ADDRESS
);
827 debug
->state
[0] = gpu_read(gpu
, VIVS_FE_DMA_DEBUG_STATE
);
829 for (i
= 0; i
< 500; i
++) {
830 debug
->address
[1] = gpu_read(gpu
, VIVS_FE_DMA_ADDRESS
);
831 debug
->state
[1] = gpu_read(gpu
, VIVS_FE_DMA_DEBUG_STATE
);
833 if (debug
->address
[0] != debug
->address
[1])
836 if (debug
->state
[0] != debug
->state
[1])
841 int etnaviv_gpu_debugfs(struct etnaviv_gpu
*gpu
, struct seq_file
*m
)
843 struct dma_debug debug
;
844 u32 dma_lo
, dma_hi
, axi
, idle
;
847 seq_printf(m
, "%s Status:\n", dev_name(gpu
->dev
));
849 ret
= pm_runtime_get_sync(gpu
->dev
);
853 dma_lo
= gpu_read(gpu
, VIVS_FE_DMA_LOW
);
854 dma_hi
= gpu_read(gpu
, VIVS_FE_DMA_HIGH
);
855 axi
= gpu_read(gpu
, VIVS_HI_AXI_STATUS
);
856 idle
= gpu_read(gpu
, VIVS_HI_IDLE_STATE
);
858 verify_dma(gpu
, &debug
);
860 seq_puts(m
, "\tfeatures\n");
861 seq_printf(m
, "\t major_features: 0x%08x\n",
862 gpu
->identity
.features
);
863 seq_printf(m
, "\t minor_features0: 0x%08x\n",
864 gpu
->identity
.minor_features0
);
865 seq_printf(m
, "\t minor_features1: 0x%08x\n",
866 gpu
->identity
.minor_features1
);
867 seq_printf(m
, "\t minor_features2: 0x%08x\n",
868 gpu
->identity
.minor_features2
);
869 seq_printf(m
, "\t minor_features3: 0x%08x\n",
870 gpu
->identity
.minor_features3
);
871 seq_printf(m
, "\t minor_features4: 0x%08x\n",
872 gpu
->identity
.minor_features4
);
873 seq_printf(m
, "\t minor_features5: 0x%08x\n",
874 gpu
->identity
.minor_features5
);
875 seq_printf(m
, "\t minor_features6: 0x%08x\n",
876 gpu
->identity
.minor_features6
);
877 seq_printf(m
, "\t minor_features7: 0x%08x\n",
878 gpu
->identity
.minor_features7
);
879 seq_printf(m
, "\t minor_features8: 0x%08x\n",
880 gpu
->identity
.minor_features8
);
881 seq_printf(m
, "\t minor_features9: 0x%08x\n",
882 gpu
->identity
.minor_features9
);
883 seq_printf(m
, "\t minor_features10: 0x%08x\n",
884 gpu
->identity
.minor_features10
);
885 seq_printf(m
, "\t minor_features11: 0x%08x\n",
886 gpu
->identity
.minor_features11
);
888 seq_puts(m
, "\tspecs\n");
889 seq_printf(m
, "\t stream_count: %d\n",
890 gpu
->identity
.stream_count
);
891 seq_printf(m
, "\t register_max: %d\n",
892 gpu
->identity
.register_max
);
893 seq_printf(m
, "\t thread_count: %d\n",
894 gpu
->identity
.thread_count
);
895 seq_printf(m
, "\t vertex_cache_size: %d\n",
896 gpu
->identity
.vertex_cache_size
);
897 seq_printf(m
, "\t shader_core_count: %d\n",
898 gpu
->identity
.shader_core_count
);
899 seq_printf(m
, "\t pixel_pipes: %d\n",
900 gpu
->identity
.pixel_pipes
);
901 seq_printf(m
, "\t vertex_output_buffer_size: %d\n",
902 gpu
->identity
.vertex_output_buffer_size
);
903 seq_printf(m
, "\t buffer_size: %d\n",
904 gpu
->identity
.buffer_size
);
905 seq_printf(m
, "\t instruction_count: %d\n",
906 gpu
->identity
.instruction_count
);
907 seq_printf(m
, "\t num_constants: %d\n",
908 gpu
->identity
.num_constants
);
909 seq_printf(m
, "\t varyings_count: %d\n",
910 gpu
->identity
.varyings_count
);
912 seq_printf(m
, "\taxi: 0x%08x\n", axi
);
913 seq_printf(m
, "\tidle: 0x%08x\n", idle
);
914 idle
|= ~gpu
->idle_mask
& ~VIVS_HI_IDLE_STATE_AXI_LP
;
915 if ((idle
& VIVS_HI_IDLE_STATE_FE
) == 0)
916 seq_puts(m
, "\t FE is not idle\n");
917 if ((idle
& VIVS_HI_IDLE_STATE_DE
) == 0)
918 seq_puts(m
, "\t DE is not idle\n");
919 if ((idle
& VIVS_HI_IDLE_STATE_PE
) == 0)
920 seq_puts(m
, "\t PE is not idle\n");
921 if ((idle
& VIVS_HI_IDLE_STATE_SH
) == 0)
922 seq_puts(m
, "\t SH is not idle\n");
923 if ((idle
& VIVS_HI_IDLE_STATE_PA
) == 0)
924 seq_puts(m
, "\t PA is not idle\n");
925 if ((idle
& VIVS_HI_IDLE_STATE_SE
) == 0)
926 seq_puts(m
, "\t SE is not idle\n");
927 if ((idle
& VIVS_HI_IDLE_STATE_RA
) == 0)
928 seq_puts(m
, "\t RA is not idle\n");
929 if ((idle
& VIVS_HI_IDLE_STATE_TX
) == 0)
930 seq_puts(m
, "\t TX is not idle\n");
931 if ((idle
& VIVS_HI_IDLE_STATE_VG
) == 0)
932 seq_puts(m
, "\t VG is not idle\n");
933 if ((idle
& VIVS_HI_IDLE_STATE_IM
) == 0)
934 seq_puts(m
, "\t IM is not idle\n");
935 if ((idle
& VIVS_HI_IDLE_STATE_FP
) == 0)
936 seq_puts(m
, "\t FP is not idle\n");
937 if ((idle
& VIVS_HI_IDLE_STATE_TS
) == 0)
938 seq_puts(m
, "\t TS is not idle\n");
939 if (idle
& VIVS_HI_IDLE_STATE_AXI_LP
)
940 seq_puts(m
, "\t AXI low power mode\n");
942 if (gpu
->identity
.features
& chipFeatures_DEBUG_MODE
) {
943 u32 read0
= gpu_read(gpu
, VIVS_MC_DEBUG_READ0
);
944 u32 read1
= gpu_read(gpu
, VIVS_MC_DEBUG_READ1
);
945 u32 write
= gpu_read(gpu
, VIVS_MC_DEBUG_WRITE
);
947 seq_puts(m
, "\tMC\n");
948 seq_printf(m
, "\t read0: 0x%08x\n", read0
);
949 seq_printf(m
, "\t read1: 0x%08x\n", read1
);
950 seq_printf(m
, "\t write: 0x%08x\n", write
);
953 seq_puts(m
, "\tDMA ");
955 if (debug
.address
[0] == debug
.address
[1] &&
956 debug
.state
[0] == debug
.state
[1]) {
957 seq_puts(m
, "seems to be stuck\n");
958 } else if (debug
.address
[0] == debug
.address
[1]) {
959 seq_puts(m
, "address is constant\n");
961 seq_puts(m
, "is running\n");
964 seq_printf(m
, "\t address 0: 0x%08x\n", debug
.address
[0]);
965 seq_printf(m
, "\t address 1: 0x%08x\n", debug
.address
[1]);
966 seq_printf(m
, "\t state 0: 0x%08x\n", debug
.state
[0]);
967 seq_printf(m
, "\t state 1: 0x%08x\n", debug
.state
[1]);
968 seq_printf(m
, "\t last fetch 64 bit word: 0x%08x 0x%08x\n",
973 pm_runtime_mark_last_busy(gpu
->dev
);
974 pm_runtime_put_autosuspend(gpu
->dev
);
980 void etnaviv_gpu_recover_hang(struct etnaviv_gpu
*gpu
)
985 dev_err(gpu
->dev
, "recover hung GPU!\n");
987 if (pm_runtime_get_sync(gpu
->dev
) < 0)
990 mutex_lock(&gpu
->lock
);
992 etnaviv_hw_reset(gpu
);
994 /* complete all events, the GPU won't do it after the reset */
995 spin_lock_irqsave(&gpu
->event_spinlock
, flags
);
996 for_each_set_bit_from(i
, gpu
->event_bitmap
, ETNA_NR_EVENTS
)
997 complete(&gpu
->event_free
);
998 bitmap_zero(gpu
->event_bitmap
, ETNA_NR_EVENTS
);
999 spin_unlock_irqrestore(&gpu
->event_spinlock
, flags
);
1000 gpu
->completed_fence
= gpu
->active_fence
;
1002 etnaviv_gpu_hw_init(gpu
);
1003 gpu
->lastctx
= NULL
;
1004 gpu
->exec_state
= -1;
1006 mutex_unlock(&gpu
->lock
);
1007 pm_runtime_mark_last_busy(gpu
->dev
);
1008 pm_runtime_put_autosuspend(gpu
->dev
);
1011 /* fence object management */
1012 struct etnaviv_fence
{
1013 struct etnaviv_gpu
*gpu
;
1014 struct dma_fence base
;
1017 static inline struct etnaviv_fence
*to_etnaviv_fence(struct dma_fence
*fence
)
1019 return container_of(fence
, struct etnaviv_fence
, base
);
1022 static const char *etnaviv_fence_get_driver_name(struct dma_fence
*fence
)
1027 static const char *etnaviv_fence_get_timeline_name(struct dma_fence
*fence
)
1029 struct etnaviv_fence
*f
= to_etnaviv_fence(fence
);
1031 return dev_name(f
->gpu
->dev
);
1034 static bool etnaviv_fence_signaled(struct dma_fence
*fence
)
1036 struct etnaviv_fence
*f
= to_etnaviv_fence(fence
);
1038 return fence_completed(f
->gpu
, f
->base
.seqno
);
1041 static void etnaviv_fence_release(struct dma_fence
*fence
)
1043 struct etnaviv_fence
*f
= to_etnaviv_fence(fence
);
1045 kfree_rcu(f
, base
.rcu
);
1048 static const struct dma_fence_ops etnaviv_fence_ops
= {
1049 .get_driver_name
= etnaviv_fence_get_driver_name
,
1050 .get_timeline_name
= etnaviv_fence_get_timeline_name
,
1051 .signaled
= etnaviv_fence_signaled
,
1052 .release
= etnaviv_fence_release
,
1055 static struct dma_fence
*etnaviv_gpu_fence_alloc(struct etnaviv_gpu
*gpu
)
1057 struct etnaviv_fence
*f
;
1060 * GPU lock must already be held, otherwise fence completion order might
1061 * not match the seqno order assigned here.
1063 lockdep_assert_held(&gpu
->lock
);
1065 f
= kzalloc(sizeof(*f
), GFP_KERNEL
);
1071 dma_fence_init(&f
->base
, &etnaviv_fence_ops
, &gpu
->fence_spinlock
,
1072 gpu
->fence_context
, ++gpu
->next_fence
);
1081 static int event_alloc(struct etnaviv_gpu
*gpu
, unsigned nr_events
,
1082 unsigned int *events
)
1084 unsigned long flags
, timeout
= msecs_to_jiffies(10 * 10000);
1085 unsigned i
, acquired
= 0;
1087 for (i
= 0; i
< nr_events
; i
++) {
1090 ret
= wait_for_completion_timeout(&gpu
->event_free
, timeout
);
1093 dev_err(gpu
->dev
, "wait_for_completion_timeout failed");
1101 spin_lock_irqsave(&gpu
->event_spinlock
, flags
);
1103 for (i
= 0; i
< nr_events
; i
++) {
1104 int event
= find_first_zero_bit(gpu
->event_bitmap
, ETNA_NR_EVENTS
);
1107 memset(&gpu
->event
[event
], 0, sizeof(struct etnaviv_event
));
1108 set_bit(event
, gpu
->event_bitmap
);
1111 spin_unlock_irqrestore(&gpu
->event_spinlock
, flags
);
1116 for (i
= 0; i
< acquired
; i
++)
1117 complete(&gpu
->event_free
);
1122 static void event_free(struct etnaviv_gpu
*gpu
, unsigned int event
)
1124 unsigned long flags
;
1126 spin_lock_irqsave(&gpu
->event_spinlock
, flags
);
1128 if (!test_bit(event
, gpu
->event_bitmap
)) {
1129 dev_warn(gpu
->dev
, "event %u is already marked as free",
1131 spin_unlock_irqrestore(&gpu
->event_spinlock
, flags
);
1133 clear_bit(event
, gpu
->event_bitmap
);
1134 spin_unlock_irqrestore(&gpu
->event_spinlock
, flags
);
1136 complete(&gpu
->event_free
);
1141 * Cmdstream submission/retirement:
1143 int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu
*gpu
,
1144 u32 id
, struct timespec
*timeout
)
1146 struct dma_fence
*fence
;
1150 * Look up the fence and take a reference. We might still find a fence
1151 * whose refcount has already dropped to zero. dma_fence_get_rcu
1152 * pretends we didn't find a fence in that case.
1155 fence
= idr_find(&gpu
->fence_idr
, id
);
1157 fence
= dma_fence_get_rcu(fence
);
1164 /* No timeout was requested: just test for completion */
1165 ret
= dma_fence_is_signaled(fence
) ? 0 : -EBUSY
;
1167 unsigned long remaining
= etnaviv_timeout_to_jiffies(timeout
);
1169 ret
= dma_fence_wait_timeout(fence
, true, remaining
);
1172 else if (ret
!= -ERESTARTSYS
)
1177 dma_fence_put(fence
);
1182 * Wait for an object to become inactive. This, on it's own, is not race
1183 * free: the object is moved by the scheduler off the active list, and
1184 * then the iova is put. Moreover, the object could be re-submitted just
1185 * after we notice that it's become inactive.
1187 * Although the retirement happens under the gpu lock, we don't want to hold
1188 * that lock in this function while waiting.
1190 int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu
*gpu
,
1191 struct etnaviv_gem_object
*etnaviv_obj
, struct timespec
*timeout
)
1193 unsigned long remaining
;
1197 return !is_active(etnaviv_obj
) ? 0 : -EBUSY
;
1199 remaining
= etnaviv_timeout_to_jiffies(timeout
);
1201 ret
= wait_event_interruptible_timeout(gpu
->fence_event
,
1202 !is_active(etnaviv_obj
),
1206 else if (ret
== -ERESTARTSYS
)
1207 return -ERESTARTSYS
;
1212 static void sync_point_perfmon_sample(struct etnaviv_gpu
*gpu
,
1213 struct etnaviv_event
*event
, unsigned int flags
)
1215 const struct etnaviv_gem_submit
*submit
= event
->submit
;
1218 for (i
= 0; i
< submit
->nr_pmrs
; i
++) {
1219 const struct etnaviv_perfmon_request
*pmr
= submit
->pmrs
+ i
;
1221 if (pmr
->flags
== flags
)
1222 etnaviv_perfmon_process(gpu
, pmr
, submit
->exec_state
);
1226 static void sync_point_perfmon_sample_pre(struct etnaviv_gpu
*gpu
,
1227 struct etnaviv_event
*event
)
1231 /* disable clock gating */
1232 val
= gpu_read(gpu
, VIVS_PM_POWER_CONTROLS
);
1233 val
&= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING
;
1234 gpu_write(gpu
, VIVS_PM_POWER_CONTROLS
, val
);
1236 /* enable debug register */
1237 val
= gpu_read(gpu
, VIVS_HI_CLOCK_CONTROL
);
1238 val
&= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS
;
1239 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, val
);
1241 sync_point_perfmon_sample(gpu
, event
, ETNA_PM_PROCESS_PRE
);
1244 static void sync_point_perfmon_sample_post(struct etnaviv_gpu
*gpu
,
1245 struct etnaviv_event
*event
)
1247 const struct etnaviv_gem_submit
*submit
= event
->submit
;
1251 sync_point_perfmon_sample(gpu
, event
, ETNA_PM_PROCESS_POST
);
1253 for (i
= 0; i
< submit
->nr_pmrs
; i
++) {
1254 const struct etnaviv_perfmon_request
*pmr
= submit
->pmrs
+ i
;
1256 *pmr
->bo_vma
= pmr
->sequence
;
1259 /* disable debug register */
1260 val
= gpu_read(gpu
, VIVS_HI_CLOCK_CONTROL
);
1261 val
|= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS
;
1262 gpu_write(gpu
, VIVS_HI_CLOCK_CONTROL
, val
);
1264 /* enable clock gating */
1265 val
= gpu_read(gpu
, VIVS_PM_POWER_CONTROLS
);
1266 val
|= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING
;
1267 gpu_write(gpu
, VIVS_PM_POWER_CONTROLS
, val
);
1271 /* add bo's to gpu's ring, and kick gpu: */
1272 struct dma_fence
*etnaviv_gpu_submit(struct etnaviv_gem_submit
*submit
)
1274 struct etnaviv_gpu
*gpu
= submit
->gpu
;
1275 struct dma_fence
*gpu_fence
;
1276 unsigned int i
, nr_events
= 1, event
[3];
1279 if (!submit
->runtime_resumed
) {
1280 ret
= pm_runtime_get_sync(gpu
->dev
);
1283 submit
->runtime_resumed
= true;
1287 * if there are performance monitor requests we need to have
1288 * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE
1290 * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests
1291 * and update the sequence number for userspace.
1293 if (submit
->nr_pmrs
)
1296 ret
= event_alloc(gpu
, nr_events
, event
);
1298 DRM_ERROR("no free events\n");
1302 mutex_lock(&gpu
->lock
);
1304 gpu_fence
= etnaviv_gpu_fence_alloc(gpu
);
1306 for (i
= 0; i
< nr_events
; i
++)
1307 event_free(gpu
, event
[i
]);
1312 gpu
->active_fence
= gpu_fence
->seqno
;
1314 if (submit
->nr_pmrs
) {
1315 gpu
->event
[event
[1]].sync_point
= &sync_point_perfmon_sample_pre
;
1316 kref_get(&submit
->refcount
);
1317 gpu
->event
[event
[1]].submit
= submit
;
1318 etnaviv_sync_point_queue(gpu
, event
[1]);
1321 gpu
->event
[event
[0]].fence
= gpu_fence
;
1322 submit
->cmdbuf
.user_size
= submit
->cmdbuf
.size
- 8;
1323 etnaviv_buffer_queue(gpu
, submit
->exec_state
, event
[0],
1326 if (submit
->nr_pmrs
) {
1327 gpu
->event
[event
[2]].sync_point
= &sync_point_perfmon_sample_post
;
1328 kref_get(&submit
->refcount
);
1329 gpu
->event
[event
[2]].submit
= submit
;
1330 etnaviv_sync_point_queue(gpu
, event
[2]);
1334 mutex_unlock(&gpu
->lock
);
1339 static void sync_point_worker(struct work_struct
*work
)
1341 struct etnaviv_gpu
*gpu
= container_of(work
, struct etnaviv_gpu
,
1343 struct etnaviv_event
*event
= &gpu
->event
[gpu
->sync_point_event
];
1344 u32 addr
= gpu_read(gpu
, VIVS_FE_DMA_ADDRESS
);
1346 event
->sync_point(gpu
, event
);
1347 etnaviv_submit_put(event
->submit
);
1348 event_free(gpu
, gpu
->sync_point_event
);
1350 /* restart FE last to avoid GPU and IRQ racing against this worker */
1351 etnaviv_gpu_start_fe(gpu
, addr
+ 2, 2);
1354 static void dump_mmu_fault(struct etnaviv_gpu
*gpu
)
1356 u32 status_reg
, status
;
1359 if (gpu
->sec_mode
== ETNA_SEC_NONE
)
1360 status_reg
= VIVS_MMUv2_STATUS
;
1362 status_reg
= VIVS_MMUv2_SEC_STATUS
;
1364 status
= gpu_read(gpu
, status_reg
);
1365 dev_err_ratelimited(gpu
->dev
, "MMU fault status 0x%08x\n", status
);
1367 for (i
= 0; i
< 4; i
++) {
1370 if (!(status
& (VIVS_MMUv2_STATUS_EXCEPTION0__MASK
<< (i
* 4))))
1373 if (gpu
->sec_mode
== ETNA_SEC_NONE
)
1374 address_reg
= VIVS_MMUv2_EXCEPTION_ADDR(i
);
1376 address_reg
= VIVS_MMUv2_SEC_EXCEPTION_ADDR
;
1378 dev_err_ratelimited(gpu
->dev
, "MMU %d fault addr 0x%08x\n", i
,
1379 gpu_read(gpu
, address_reg
));
1383 static irqreturn_t
irq_handler(int irq
, void *data
)
1385 struct etnaviv_gpu
*gpu
= data
;
1386 irqreturn_t ret
= IRQ_NONE
;
1388 u32 intr
= gpu_read(gpu
, VIVS_HI_INTR_ACKNOWLEDGE
);
1393 pm_runtime_mark_last_busy(gpu
->dev
);
1395 dev_dbg(gpu
->dev
, "intr 0x%08x\n", intr
);
1397 if (intr
& VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR
) {
1398 dev_err(gpu
->dev
, "AXI bus error\n");
1399 intr
&= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR
;
1402 if (intr
& VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION
) {
1403 dump_mmu_fault(gpu
);
1404 intr
&= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION
;
1407 while ((event
= ffs(intr
)) != 0) {
1408 struct dma_fence
*fence
;
1412 intr
&= ~(1 << event
);
1414 dev_dbg(gpu
->dev
, "event %u\n", event
);
1416 if (gpu
->event
[event
].sync_point
) {
1417 gpu
->sync_point_event
= event
;
1418 queue_work(gpu
->wq
, &gpu
->sync_point_work
);
1421 fence
= gpu
->event
[event
].fence
;
1425 gpu
->event
[event
].fence
= NULL
;
1428 * Events can be processed out of order. Eg,
1429 * - allocate and queue event 0
1430 * - allocate event 1
1431 * - event 0 completes, we process it
1432 * - allocate and queue event 0
1433 * - event 1 and event 0 complete
1434 * we can end up processing event 0 first, then 1.
1436 if (fence_after(fence
->seqno
, gpu
->completed_fence
))
1437 gpu
->completed_fence
= fence
->seqno
;
1438 dma_fence_signal(fence
);
1440 event_free(gpu
, event
);
1449 static int etnaviv_gpu_clk_enable(struct etnaviv_gpu
*gpu
)
1454 ret
= clk_prepare_enable(gpu
->clk_reg
);
1460 ret
= clk_prepare_enable(gpu
->clk_bus
);
1465 if (gpu
->clk_core
) {
1466 ret
= clk_prepare_enable(gpu
->clk_core
);
1468 goto disable_clk_bus
;
1471 if (gpu
->clk_shader
) {
1472 ret
= clk_prepare_enable(gpu
->clk_shader
);
1474 goto disable_clk_core
;
1481 clk_disable_unprepare(gpu
->clk_core
);
1484 clk_disable_unprepare(gpu
->clk_bus
);
1489 static int etnaviv_gpu_clk_disable(struct etnaviv_gpu
*gpu
)
1491 if (gpu
->clk_shader
)
1492 clk_disable_unprepare(gpu
->clk_shader
);
1494 clk_disable_unprepare(gpu
->clk_core
);
1496 clk_disable_unprepare(gpu
->clk_bus
);
1498 clk_disable_unprepare(gpu
->clk_reg
);
1503 int etnaviv_gpu_wait_idle(struct etnaviv_gpu
*gpu
, unsigned int timeout_ms
)
1505 unsigned long timeout
= jiffies
+ msecs_to_jiffies(timeout_ms
);
1508 u32 idle
= gpu_read(gpu
, VIVS_HI_IDLE_STATE
);
1510 if ((idle
& gpu
->idle_mask
) == gpu
->idle_mask
)
1513 if (time_is_before_jiffies(timeout
)) {
1515 "timed out waiting for idle: idle=0x%x\n",
1524 static int etnaviv_gpu_hw_suspend(struct etnaviv_gpu
*gpu
)
1526 if (gpu
->buffer
.suballoc
) {
1527 /* Replace the last WAIT with END */
1528 mutex_lock(&gpu
->lock
);
1529 etnaviv_buffer_end(gpu
);
1530 mutex_unlock(&gpu
->lock
);
1533 * We know that only the FE is busy here, this should
1534 * happen quickly (as the WAIT is only 200 cycles). If
1535 * we fail, just warn and continue.
1537 etnaviv_gpu_wait_idle(gpu
, 100);
1540 return etnaviv_gpu_clk_disable(gpu
);
1544 static int etnaviv_gpu_hw_resume(struct etnaviv_gpu
*gpu
)
1548 ret
= mutex_lock_killable(&gpu
->lock
);
1552 etnaviv_gpu_update_clock(gpu
);
1553 etnaviv_gpu_hw_init(gpu
);
1555 gpu
->lastctx
= NULL
;
1556 gpu
->exec_state
= -1;
1558 mutex_unlock(&gpu
->lock
);
1565 etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device
*cdev
,
1566 unsigned long *state
)
1574 etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device
*cdev
,
1575 unsigned long *state
)
1577 struct etnaviv_gpu
*gpu
= cdev
->devdata
;
1579 *state
= gpu
->freq_scale
;
1585 etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device
*cdev
,
1586 unsigned long state
)
1588 struct etnaviv_gpu
*gpu
= cdev
->devdata
;
1590 mutex_lock(&gpu
->lock
);
1591 gpu
->freq_scale
= state
;
1592 if (!pm_runtime_suspended(gpu
->dev
))
1593 etnaviv_gpu_update_clock(gpu
);
1594 mutex_unlock(&gpu
->lock
);
1599 static struct thermal_cooling_device_ops cooling_ops
= {
1600 .get_max_state
= etnaviv_gpu_cooling_get_max_state
,
1601 .get_cur_state
= etnaviv_gpu_cooling_get_cur_state
,
1602 .set_cur_state
= etnaviv_gpu_cooling_set_cur_state
,
1605 static int etnaviv_gpu_bind(struct device
*dev
, struct device
*master
,
1608 struct drm_device
*drm
= data
;
1609 struct etnaviv_drm_private
*priv
= drm
->dev_private
;
1610 struct etnaviv_gpu
*gpu
= dev_get_drvdata(dev
);
1613 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL
)) {
1614 gpu
->cooling
= thermal_of_cooling_device_register(dev
->of_node
,
1615 (char *)dev_name(dev
), gpu
, &cooling_ops
);
1616 if (IS_ERR(gpu
->cooling
))
1617 return PTR_ERR(gpu
->cooling
);
1620 gpu
->wq
= alloc_ordered_workqueue(dev_name(dev
), 0);
1626 ret
= etnaviv_sched_init(gpu
);
1631 ret
= pm_runtime_get_sync(gpu
->dev
);
1633 ret
= etnaviv_gpu_clk_enable(gpu
);
1640 gpu
->fence_context
= dma_fence_context_alloc(1);
1641 idr_init(&gpu
->fence_idr
);
1642 spin_lock_init(&gpu
->fence_spinlock
);
1644 INIT_WORK(&gpu
->sync_point_work
, sync_point_worker
);
1645 init_waitqueue_head(&gpu
->fence_event
);
1647 priv
->gpu
[priv
->num_gpus
++] = gpu
;
1649 pm_runtime_mark_last_busy(gpu
->dev
);
1650 pm_runtime_put_autosuspend(gpu
->dev
);
1655 etnaviv_sched_fini(gpu
);
1658 destroy_workqueue(gpu
->wq
);
1661 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL
))
1662 thermal_cooling_device_unregister(gpu
->cooling
);
1667 static void etnaviv_gpu_unbind(struct device
*dev
, struct device
*master
,
1670 struct etnaviv_gpu
*gpu
= dev_get_drvdata(dev
);
1672 DBG("%s", dev_name(gpu
->dev
));
1674 flush_workqueue(gpu
->wq
);
1675 destroy_workqueue(gpu
->wq
);
1677 etnaviv_sched_fini(gpu
);
1680 pm_runtime_get_sync(gpu
->dev
);
1681 pm_runtime_put_sync_suspend(gpu
->dev
);
1683 etnaviv_gpu_hw_suspend(gpu
);
1686 if (gpu
->buffer
.suballoc
)
1687 etnaviv_cmdbuf_free(&gpu
->buffer
);
1689 if (gpu
->cmdbuf_suballoc
) {
1690 etnaviv_cmdbuf_suballoc_destroy(gpu
->cmdbuf_suballoc
);
1691 gpu
->cmdbuf_suballoc
= NULL
;
1695 etnaviv_iommu_destroy(gpu
->mmu
);
1700 idr_destroy(&gpu
->fence_idr
);
1702 if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL
))
1703 thermal_cooling_device_unregister(gpu
->cooling
);
1704 gpu
->cooling
= NULL
;
1707 static const struct component_ops gpu_ops
= {
1708 .bind
= etnaviv_gpu_bind
,
1709 .unbind
= etnaviv_gpu_unbind
,
1712 static const struct of_device_id etnaviv_gpu_match
[] = {
1714 .compatible
= "vivante,gc"
1718 MODULE_DEVICE_TABLE(of
, etnaviv_gpu_match
);
1720 static int etnaviv_gpu_platform_probe(struct platform_device
*pdev
)
1722 struct device
*dev
= &pdev
->dev
;
1723 struct etnaviv_gpu
*gpu
;
1724 struct resource
*res
;
1727 gpu
= devm_kzalloc(dev
, sizeof(*gpu
), GFP_KERNEL
);
1731 gpu
->dev
= &pdev
->dev
;
1732 mutex_init(&gpu
->lock
);
1733 mutex_init(&gpu
->fence_lock
);
1735 /* Map registers: */
1736 res
= platform_get_resource(pdev
, IORESOURCE_MEM
, 0);
1737 gpu
->mmio
= devm_ioremap_resource(&pdev
->dev
, res
);
1738 if (IS_ERR(gpu
->mmio
))
1739 return PTR_ERR(gpu
->mmio
);
1741 /* Get Interrupt: */
1742 gpu
->irq
= platform_get_irq(pdev
, 0);
1744 dev_err(dev
, "failed to get irq: %d\n", gpu
->irq
);
1748 err
= devm_request_irq(&pdev
->dev
, gpu
->irq
, irq_handler
, 0,
1749 dev_name(gpu
->dev
), gpu
);
1751 dev_err(dev
, "failed to request IRQ%u: %d\n", gpu
->irq
, err
);
1756 gpu
->clk_reg
= devm_clk_get(&pdev
->dev
, "reg");
1757 DBG("clk_reg: %p", gpu
->clk_reg
);
1758 if (IS_ERR(gpu
->clk_reg
))
1759 gpu
->clk_reg
= NULL
;
1761 gpu
->clk_bus
= devm_clk_get(&pdev
->dev
, "bus");
1762 DBG("clk_bus: %p", gpu
->clk_bus
);
1763 if (IS_ERR(gpu
->clk_bus
))
1764 gpu
->clk_bus
= NULL
;
1766 gpu
->clk_core
= devm_clk_get(&pdev
->dev
, "core");
1767 DBG("clk_core: %p", gpu
->clk_core
);
1768 if (IS_ERR(gpu
->clk_core
))
1769 gpu
->clk_core
= NULL
;
1770 gpu
->base_rate_core
= clk_get_rate(gpu
->clk_core
);
1772 gpu
->clk_shader
= devm_clk_get(&pdev
->dev
, "shader");
1773 DBG("clk_shader: %p", gpu
->clk_shader
);
1774 if (IS_ERR(gpu
->clk_shader
))
1775 gpu
->clk_shader
= NULL
;
1776 gpu
->base_rate_shader
= clk_get_rate(gpu
->clk_shader
);
1778 /* TODO: figure out max mapped size */
1779 dev_set_drvdata(dev
, gpu
);
1782 * We treat the device as initially suspended. The runtime PM
1783 * autosuspend delay is rather arbitary: no measurements have
1784 * yet been performed to determine an appropriate value.
1786 pm_runtime_use_autosuspend(gpu
->dev
);
1787 pm_runtime_set_autosuspend_delay(gpu
->dev
, 200);
1788 pm_runtime_enable(gpu
->dev
);
1790 err
= component_add(&pdev
->dev
, &gpu_ops
);
1792 dev_err(&pdev
->dev
, "failed to register component: %d\n", err
);
1799 static int etnaviv_gpu_platform_remove(struct platform_device
*pdev
)
1801 component_del(&pdev
->dev
, &gpu_ops
);
1802 pm_runtime_disable(&pdev
->dev
);
1807 static int etnaviv_gpu_rpm_suspend(struct device
*dev
)
1809 struct etnaviv_gpu
*gpu
= dev_get_drvdata(dev
);
1812 /* If we have outstanding fences, we're not idle */
1813 if (gpu
->completed_fence
!= gpu
->active_fence
)
1816 /* Check whether the hardware (except FE) is idle */
1817 mask
= gpu
->idle_mask
& ~VIVS_HI_IDLE_STATE_FE
;
1818 idle
= gpu_read(gpu
, VIVS_HI_IDLE_STATE
) & mask
;
1822 return etnaviv_gpu_hw_suspend(gpu
);
1825 static int etnaviv_gpu_rpm_resume(struct device
*dev
)
1827 struct etnaviv_gpu
*gpu
= dev_get_drvdata(dev
);
1830 ret
= etnaviv_gpu_clk_enable(gpu
);
1834 /* Re-initialise the basic hardware state */
1835 if (gpu
->drm
&& gpu
->buffer
.suballoc
) {
1836 ret
= etnaviv_gpu_hw_resume(gpu
);
1838 etnaviv_gpu_clk_disable(gpu
);
1847 static const struct dev_pm_ops etnaviv_gpu_pm_ops
= {
1848 SET_RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend
, etnaviv_gpu_rpm_resume
,
1852 struct platform_driver etnaviv_gpu_driver
= {
1854 .name
= "etnaviv-gpu",
1855 .owner
= THIS_MODULE
,
1856 .pm
= &etnaviv_gpu_pm_ops
,
1857 .of_match_table
= etnaviv_gpu_match
,
1859 .probe
= etnaviv_gpu_platform_probe
,
1860 .remove
= etnaviv_gpu_platform_remove
,
1861 .id_table
= gpu_ids
,