1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * Copyright (C) 2015-2018 Etnaviv Project |
4 | */ |
5 | |
6 | #include <linux/clk.h> |
7 | #include <linux/component.h> |
8 | #include <linux/delay.h> |
9 | #include <linux/dma-fence.h> |
10 | #include <linux/dma-mapping.h> |
11 | #include <linux/mod_devicetable.h> |
12 | #include <linux/module.h> |
13 | #include <linux/platform_device.h> |
14 | #include <linux/pm_runtime.h> |
15 | #include <linux/regulator/consumer.h> |
16 | #include <linux/thermal.h> |
17 | |
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" |
29 | |
30 | static const struct platform_device_id gpu_ids[] = { |
31 | { .name = "etnaviv-gpu,2d" }, |
32 | { }, |
33 | }; |
34 | |
35 | /* |
36 | * Driver functions: |
37 | */ |
38 | |
39 | int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value) |
40 | { |
41 | struct etnaviv_drm_private *priv = gpu->drm->dev_private; |
42 | |
43 | switch (param) { |
44 | case ETNAVIV_PARAM_GPU_MODEL: |
45 | *value = gpu->identity.model; |
46 | break; |
47 | |
48 | case ETNAVIV_PARAM_GPU_REVISION: |
49 | *value = gpu->identity.revision; |
50 | break; |
51 | |
52 | case ETNAVIV_PARAM_GPU_FEATURES_0: |
53 | *value = gpu->identity.features; |
54 | break; |
55 | |
56 | case ETNAVIV_PARAM_GPU_FEATURES_1: |
57 | *value = gpu->identity.minor_features0; |
58 | break; |
59 | |
60 | case ETNAVIV_PARAM_GPU_FEATURES_2: |
61 | *value = gpu->identity.minor_features1; |
62 | break; |
63 | |
64 | case ETNAVIV_PARAM_GPU_FEATURES_3: |
65 | *value = gpu->identity.minor_features2; |
66 | break; |
67 | |
68 | case ETNAVIV_PARAM_GPU_FEATURES_4: |
69 | *value = gpu->identity.minor_features3; |
70 | break; |
71 | |
72 | case ETNAVIV_PARAM_GPU_FEATURES_5: |
73 | *value = gpu->identity.minor_features4; |
74 | break; |
75 | |
76 | case ETNAVIV_PARAM_GPU_FEATURES_6: |
77 | *value = gpu->identity.minor_features5; |
78 | break; |
79 | |
80 | case ETNAVIV_PARAM_GPU_FEATURES_7: |
81 | *value = gpu->identity.minor_features6; |
82 | break; |
83 | |
84 | case ETNAVIV_PARAM_GPU_FEATURES_8: |
85 | *value = gpu->identity.minor_features7; |
86 | break; |
87 | |
88 | case ETNAVIV_PARAM_GPU_FEATURES_9: |
89 | *value = gpu->identity.minor_features8; |
90 | break; |
91 | |
92 | case ETNAVIV_PARAM_GPU_FEATURES_10: |
93 | *value = gpu->identity.minor_features9; |
94 | break; |
95 | |
96 | case ETNAVIV_PARAM_GPU_FEATURES_11: |
97 | *value = gpu->identity.minor_features10; |
98 | break; |
99 | |
100 | case ETNAVIV_PARAM_GPU_FEATURES_12: |
101 | *value = gpu->identity.minor_features11; |
102 | break; |
103 | |
104 | case ETNAVIV_PARAM_GPU_STREAM_COUNT: |
105 | *value = gpu->identity.stream_count; |
106 | break; |
107 | |
108 | case ETNAVIV_PARAM_GPU_REGISTER_MAX: |
109 | *value = gpu->identity.register_max; |
110 | break; |
111 | |
112 | case ETNAVIV_PARAM_GPU_THREAD_COUNT: |
113 | *value = gpu->identity.thread_count; |
114 | break; |
115 | |
116 | case ETNAVIV_PARAM_GPU_VERTEX_CACHE_SIZE: |
117 | *value = gpu->identity.vertex_cache_size; |
118 | break; |
119 | |
120 | case ETNAVIV_PARAM_GPU_SHADER_CORE_COUNT: |
121 | *value = gpu->identity.shader_core_count; |
122 | break; |
123 | |
124 | case ETNAVIV_PARAM_GPU_PIXEL_PIPES: |
125 | *value = gpu->identity.pixel_pipes; |
126 | break; |
127 | |
128 | case ETNAVIV_PARAM_GPU_VERTEX_OUTPUT_BUFFER_SIZE: |
129 | *value = gpu->identity.vertex_output_buffer_size; |
130 | break; |
131 | |
132 | case ETNAVIV_PARAM_GPU_BUFFER_SIZE: |
133 | *value = gpu->identity.buffer_size; |
134 | break; |
135 | |
136 | case ETNAVIV_PARAM_GPU_INSTRUCTION_COUNT: |
137 | *value = gpu->identity.instruction_count; |
138 | break; |
139 | |
140 | case ETNAVIV_PARAM_GPU_NUM_CONSTANTS: |
141 | *value = gpu->identity.num_constants; |
142 | break; |
143 | |
144 | case ETNAVIV_PARAM_GPU_NUM_VARYINGS: |
145 | *value = gpu->identity.varyings_count; |
146 | break; |
147 | |
148 | case ETNAVIV_PARAM_SOFTPIN_START_ADDR: |
149 | if (priv->mmu_global->version == ETNAVIV_IOMMU_V2) |
150 | *value = ETNAVIV_SOFTPIN_START_ADDRESS; |
151 | else |
152 | *value = ~0ULL; |
153 | break; |
154 | |
155 | case ETNAVIV_PARAM_GPU_PRODUCT_ID: |
156 | *value = gpu->identity.product_id; |
157 | break; |
158 | |
159 | case ETNAVIV_PARAM_GPU_CUSTOMER_ID: |
160 | *value = gpu->identity.customer_id; |
161 | break; |
162 | |
163 | case ETNAVIV_PARAM_GPU_ECO_ID: |
164 | *value = gpu->identity.eco_id; |
165 | break; |
166 | |
167 | default: |
168 | DBG("%s: invalid param: %u" , dev_name(gpu->dev), param); |
169 | return -EINVAL; |
170 | } |
171 | |
172 | return 0; |
173 | } |
174 | |
175 | |
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) |
181 | |
182 | static void etnaviv_hw_specs(struct etnaviv_gpu *gpu) |
183 | { |
184 | if (gpu->identity.minor_features0 & |
185 | chipMinorFeatures0_MORE_MINOR_FEATURES) { |
186 | u32 specs[4]; |
187 | unsigned int streams; |
188 | |
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); |
193 | |
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); |
209 | |
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); |
216 | |
217 | gpu->identity.varyings_count = etnaviv_field(specs[2], |
218 | VIVS_HI_CHIP_SPECS_3_VARYINGS_COUNT); |
219 | |
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); |
223 | if (streams) |
224 | gpu->identity.stream_count = streams; |
225 | } |
226 | |
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; |
231 | else |
232 | gpu->identity.stream_count = 1; |
233 | } |
234 | |
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; |
240 | else |
241 | gpu->identity.register_max = 64; |
242 | |
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; |
251 | else |
252 | gpu->identity.thread_count = 256; |
253 | |
254 | if (gpu->identity.vertex_cache_size == 0) |
255 | gpu->identity.vertex_cache_size = 8; |
256 | |
257 | if (gpu->identity.shader_core_count == 0) { |
258 | if (gpu->identity.model >= 0x1000) |
259 | gpu->identity.shader_core_count = 2; |
260 | else |
261 | gpu->identity.shader_core_count = 1; |
262 | } |
263 | |
264 | if (gpu->identity.pixel_pipes == 0) |
265 | gpu->identity.pixel_pipes = 1; |
266 | |
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; |
276 | else |
277 | gpu->identity.vertex_output_buffer_size = 128; |
278 | } else { |
279 | gpu->identity.vertex_output_buffer_size = 512; |
280 | } |
281 | |
282 | switch (gpu->identity.instruction_count) { |
283 | case 0: |
284 | if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) || |
285 | gpu->identity.model == chipModel_GC880) |
286 | gpu->identity.instruction_count = 512; |
287 | else |
288 | gpu->identity.instruction_count = 256; |
289 | break; |
290 | |
291 | case 1: |
292 | gpu->identity.instruction_count = 1024; |
293 | break; |
294 | |
295 | case 2: |
296 | gpu->identity.instruction_count = 2048; |
297 | break; |
298 | |
299 | default: |
300 | gpu->identity.instruction_count = 256; |
301 | break; |
302 | } |
303 | |
304 | if (gpu->identity.num_constants == 0) |
305 | gpu->identity.num_constants = 168; |
306 | |
307 | if (gpu->identity.varyings_count == 0) { |
308 | if (gpu->identity.minor_features1 & chipMinorFeatures1_HALTI0) |
309 | gpu->identity.varyings_count = 12; |
310 | else |
311 | gpu->identity.varyings_count = 8; |
312 | } |
313 | |
314 | /* |
315 | * For some cores, two varyings are consumed for position, so the |
316 | * maximum varying count needs to be reduced by one. |
317 | */ |
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; |
330 | } |
331 | |
332 | static void etnaviv_hw_identify(struct etnaviv_gpu *gpu) |
333 | { |
334 | u32 chipIdentity; |
335 | |
336 | chipIdentity = gpu_read(gpu, VIVS_HI_CHIP_IDENTITY); |
337 | |
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); |
343 | } else { |
344 | u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE); |
345 | |
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); |
349 | |
350 | /* |
351 | * Reading these two registers on GC600 rev 0x19 result in a |
352 | * unhandled fault: external abort on non-linefetch |
353 | */ |
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); |
357 | } |
358 | |
359 | /* |
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. |
364 | */ |
365 | if ((gpu->identity.model & 0xff00) == 0x0400 && |
366 | gpu->identity.model != chipModel_GC420) { |
367 | gpu->identity.model = gpu->identity.model & 0x0400; |
368 | } |
369 | |
370 | /* Another special case */ |
371 | if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) { |
372 | u32 chipTime = gpu_read(gpu, VIVS_HI_CHIP_TIME); |
373 | |
374 | if (chipDate == 0x20080814 && chipTime == 0x12051100) { |
375 | /* |
376 | * This IP has an ECO; put the correct |
377 | * revision in it. |
378 | */ |
379 | gpu->identity.revision = 0x1051; |
380 | } |
381 | } |
382 | |
383 | /* |
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. |
389 | */ |
390 | if (etnaviv_is_model_rev(gpu, GC2000, 0xffff5450)) { |
391 | gpu->identity.model = chipModel_GC3000; |
392 | gpu->identity.revision &= 0xffff; |
393 | } |
394 | |
395 | if (etnaviv_is_model_rev(gpu, GC1000, 0x5037) && (chipDate == 0x20120617)) |
396 | gpu->identity.eco_id = 1; |
397 | |
398 | if (etnaviv_is_model_rev(gpu, GC320, 0x5303) && (chipDate == 0x20140511)) |
399 | gpu->identity.eco_id = 1; |
400 | } |
401 | |
402 | dev_info(gpu->dev, "model: GC%x, revision: %x\n" , |
403 | gpu->identity.model, gpu->identity.revision); |
404 | |
405 | gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP; |
406 | /* |
407 | * If there is a match in the HWDB, we aren't interested in the |
408 | * remaining register values, as they might be wrong. |
409 | */ |
410 | if (etnaviv_fill_identity_from_hwdb(gpu)) |
411 | return; |
412 | |
413 | gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE); |
414 | |
415 | /* Disable fast clear on GC700. */ |
416 | if (gpu->identity.model == chipModel_GC700) |
417 | gpu->identity.features &= ~chipFeatures_FAST_CLEAR; |
418 | |
419 | /* These models/revisions don't have the 2D pipe bit */ |
420 | if ((gpu->identity.model == chipModel_GC500 && |
421 | gpu->identity.revision <= 2) || |
422 | gpu->identity.model == chipModel_GC300) |
423 | gpu->identity.features |= chipFeatures_PIPE_2D; |
424 | |
425 | if ((gpu->identity.model == chipModel_GC500 && |
426 | gpu->identity.revision < 2) || |
427 | (gpu->identity.model == chipModel_GC300 && |
428 | gpu->identity.revision < 0x2000)) { |
429 | |
430 | /* |
431 | * GC500 rev 1.x and GC300 rev < 2.0 doesn't have these |
432 | * registers. |
433 | */ |
434 | gpu->identity.minor_features0 = 0; |
435 | gpu->identity.minor_features1 = 0; |
436 | gpu->identity.minor_features2 = 0; |
437 | gpu->identity.minor_features3 = 0; |
438 | gpu->identity.minor_features4 = 0; |
439 | gpu->identity.minor_features5 = 0; |
440 | } else |
441 | gpu->identity.minor_features0 = |
442 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_0); |
443 | |
444 | if (gpu->identity.minor_features0 & |
445 | chipMinorFeatures0_MORE_MINOR_FEATURES) { |
446 | gpu->identity.minor_features1 = |
447 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_1); |
448 | gpu->identity.minor_features2 = |
449 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_2); |
450 | gpu->identity.minor_features3 = |
451 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_3); |
452 | gpu->identity.minor_features4 = |
453 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_4); |
454 | gpu->identity.minor_features5 = |
455 | gpu_read(gpu, VIVS_HI_CHIP_MINOR_FEATURE_5); |
456 | } |
457 | |
458 | /* GC600/300 idle register reports zero bits where modules aren't present */ |
459 | if (gpu->identity.model == chipModel_GC600 || |
460 | gpu->identity.model == chipModel_GC300) |
461 | gpu->idle_mask = VIVS_HI_IDLE_STATE_TX | |
462 | VIVS_HI_IDLE_STATE_RA | |
463 | VIVS_HI_IDLE_STATE_SE | |
464 | VIVS_HI_IDLE_STATE_PA | |
465 | VIVS_HI_IDLE_STATE_SH | |
466 | VIVS_HI_IDLE_STATE_PE | |
467 | VIVS_HI_IDLE_STATE_DE | |
468 | VIVS_HI_IDLE_STATE_FE; |
469 | |
470 | etnaviv_hw_specs(gpu); |
471 | } |
472 | |
473 | static void etnaviv_gpu_load_clock(struct etnaviv_gpu *gpu, u32 clock) |
474 | { |
475 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: clock | |
476 | VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD); |
477 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: clock); |
478 | } |
479 | |
480 | static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu) |
481 | { |
482 | if (gpu->identity.minor_features2 & |
483 | chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) { |
484 | clk_set_rate(clk: gpu->clk_core, |
485 | rate: gpu->base_rate_core >> gpu->freq_scale); |
486 | clk_set_rate(clk: gpu->clk_shader, |
487 | rate: gpu->base_rate_shader >> gpu->freq_scale); |
488 | } else { |
489 | unsigned int fscale = 1 << (6 - gpu->freq_scale); |
490 | u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
491 | |
492 | clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK; |
493 | clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); |
494 | etnaviv_gpu_load_clock(gpu, clock); |
495 | } |
496 | |
497 | /* |
498 | * Choose number of wait cycles to target a ~30us (1/32768) max latency |
499 | * until new work is picked up by the FE when it polls in the idle loop. |
500 | * If the GPU base frequency is unknown use 200 wait cycles. |
501 | */ |
502 | gpu->fe_waitcycles = clamp(gpu->base_rate_core >> (15 - gpu->freq_scale), |
503 | 200UL, 0xffffUL); |
504 | } |
505 | |
506 | static int etnaviv_hw_reset(struct etnaviv_gpu *gpu) |
507 | { |
508 | u32 control, idle; |
509 | unsigned long timeout; |
510 | bool failed = true; |
511 | |
512 | /* We hope that the GPU resets in under one second */ |
513 | timeout = jiffies + msecs_to_jiffies(m: 1000); |
514 | |
515 | while (time_is_after_jiffies(timeout)) { |
516 | /* enable clock */ |
517 | unsigned int fscale = 1 << (6 - gpu->freq_scale); |
518 | control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); |
519 | etnaviv_gpu_load_clock(gpu, clock: control); |
520 | |
521 | /* isolate the GPU. */ |
522 | control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; |
523 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: control); |
524 | |
525 | if (gpu->sec_mode == ETNA_SEC_KERNEL) { |
526 | gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, |
527 | VIVS_MMUv2_AHB_CONTROL_RESET); |
528 | } else { |
529 | /* set soft reset. */ |
530 | control |= VIVS_HI_CLOCK_CONTROL_SOFT_RESET; |
531 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: control); |
532 | } |
533 | |
534 | /* wait for reset. */ |
535 | usleep_range(min: 10, max: 20); |
536 | |
537 | /* reset soft reset bit. */ |
538 | control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET; |
539 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: control); |
540 | |
541 | /* reset GPU isolation. */ |
542 | control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; |
543 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: control); |
544 | |
545 | /* read idle register. */ |
546 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); |
547 | |
548 | /* try resetting again if FE is not idle */ |
549 | if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) { |
550 | dev_dbg(gpu->dev, "FE is not idle\n" ); |
551 | continue; |
552 | } |
553 | |
554 | /* read reset register. */ |
555 | control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
556 | |
557 | /* is the GPU idle? */ |
558 | if (((control & VIVS_HI_CLOCK_CONTROL_IDLE_3D) == 0) || |
559 | ((control & VIVS_HI_CLOCK_CONTROL_IDLE_2D) == 0)) { |
560 | dev_dbg(gpu->dev, "GPU is not idle\n" ); |
561 | continue; |
562 | } |
563 | |
564 | /* disable debug registers, as they are not normally needed */ |
565 | control |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; |
566 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: control); |
567 | |
568 | failed = false; |
569 | break; |
570 | } |
571 | |
572 | if (failed) { |
573 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); |
574 | control = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
575 | |
576 | dev_err(gpu->dev, "GPU failed to reset: FE %sidle, 3D %sidle, 2D %sidle\n" , |
577 | idle & VIVS_HI_IDLE_STATE_FE ? "" : "not " , |
578 | control & VIVS_HI_CLOCK_CONTROL_IDLE_3D ? "" : "not " , |
579 | control & VIVS_HI_CLOCK_CONTROL_IDLE_2D ? "" : "not " ); |
580 | |
581 | return -EBUSY; |
582 | } |
583 | |
584 | /* We rely on the GPU running, so program the clock */ |
585 | etnaviv_gpu_update_clock(gpu); |
586 | |
587 | gpu->state = ETNA_GPU_STATE_RESET; |
588 | gpu->exec_state = -1; |
589 | if (gpu->mmu_context) |
590 | etnaviv_iommu_context_put(ctx: gpu->mmu_context); |
591 | gpu->mmu_context = NULL; |
592 | |
593 | return 0; |
594 | } |
595 | |
596 | static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu) |
597 | { |
598 | u32 pmc, ppc; |
599 | |
600 | /* enable clock gating */ |
601 | ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
602 | ppc |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; |
603 | |
604 | /* Disable stall module clock gating for 4.3.0.1 and 4.3.0.2 revs */ |
605 | if (gpu->identity.revision == 0x4301 || |
606 | gpu->identity.revision == 0x4302) |
607 | ppc |= VIVS_PM_POWER_CONTROLS_DISABLE_STALL_MODULE_CLOCK_GATING; |
608 | |
609 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, data: ppc); |
610 | |
611 | pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS); |
612 | |
613 | /* Disable PA clock gating for GC400+ without bugfix except for GC420 */ |
614 | if (gpu->identity.model >= chipModel_GC400 && |
615 | gpu->identity.model != chipModel_GC420 && |
616 | !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12)) |
617 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PA; |
618 | |
619 | /* |
620 | * Disable PE clock gating on revs < 5.0.0.0 when HZ is |
621 | * present without a bug fix. |
622 | */ |
623 | if (gpu->identity.revision < 0x5000 && |
624 | gpu->identity.minor_features0 & chipMinorFeatures0_HZ && |
625 | !(gpu->identity.minor_features1 & |
626 | chipMinorFeatures1_DISABLE_PE_GATING)) |
627 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_PE; |
628 | |
629 | if (gpu->identity.revision < 0x5422) |
630 | pmc |= BIT(15); /* Unknown bit */ |
631 | |
632 | /* Disable TX clock gating on affected core revisions. */ |
633 | if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) || |
634 | etnaviv_is_model_rev(gpu, GC2000, 0x5108) || |
635 | etnaviv_is_model_rev(gpu, GC2000, 0x6202) || |
636 | etnaviv_is_model_rev(gpu, GC2000, 0x6203)) |
637 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX; |
638 | |
639 | /* Disable SE and RA clock gating on affected core revisions. */ |
640 | if (etnaviv_is_model_rev(gpu, GC7000, 0x6202)) |
641 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE | |
642 | VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA; |
643 | |
644 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_HZ; |
645 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA_EZ; |
646 | |
647 | gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, data: pmc); |
648 | } |
649 | |
650 | void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch) |
651 | { |
652 | gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, data: address); |
653 | gpu_write(gpu, VIVS_FE_COMMAND_CONTROL, |
654 | VIVS_FE_COMMAND_CONTROL_ENABLE | |
655 | VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch)); |
656 | |
657 | if (gpu->sec_mode == ETNA_SEC_KERNEL) { |
658 | gpu_write(gpu, VIVS_MMUv2_SEC_COMMAND_CONTROL, |
659 | VIVS_MMUv2_SEC_COMMAND_CONTROL_ENABLE | |
660 | VIVS_MMUv2_SEC_COMMAND_CONTROL_PREFETCH(prefetch)); |
661 | } |
662 | } |
663 | |
664 | static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu, |
665 | struct etnaviv_iommu_context *context) |
666 | { |
667 | u16 prefetch; |
668 | u32 address; |
669 | |
670 | WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED); |
671 | |
672 | /* setup the MMU */ |
673 | etnaviv_iommu_restore(gpu, ctx: context); |
674 | |
675 | /* Start command processor */ |
676 | prefetch = etnaviv_buffer_init(gpu); |
677 | address = etnaviv_cmdbuf_get_va(buf: &gpu->buffer, |
678 | mapping: &gpu->mmu_context->cmdbuf_mapping); |
679 | |
680 | etnaviv_gpu_start_fe(gpu, address, prefetch); |
681 | |
682 | gpu->state = ETNA_GPU_STATE_RUNNING; |
683 | } |
684 | |
685 | static void etnaviv_gpu_setup_pulse_eater(struct etnaviv_gpu *gpu) |
686 | { |
687 | /* |
688 | * Base value for VIVS_PM_PULSE_EATER register on models where it |
689 | * cannot be read, extracted from vivante kernel driver. |
690 | */ |
691 | u32 pulse_eater = 0x01590880; |
692 | |
693 | if (etnaviv_is_model_rev(gpu, GC4000, 0x5208) || |
694 | etnaviv_is_model_rev(gpu, GC4000, 0x5222)) { |
695 | pulse_eater |= BIT(23); |
696 | |
697 | } |
698 | |
699 | if (etnaviv_is_model_rev(gpu, GC1000, 0x5039) || |
700 | etnaviv_is_model_rev(gpu, GC1000, 0x5040)) { |
701 | pulse_eater &= ~BIT(16); |
702 | pulse_eater |= BIT(17); |
703 | } |
704 | |
705 | if ((gpu->identity.revision > 0x5420) && |
706 | (gpu->identity.features & chipFeatures_PIPE_3D)) |
707 | { |
708 | /* Performance fix: disable internal DFS */ |
709 | pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER); |
710 | pulse_eater |= BIT(18); |
711 | } |
712 | |
713 | gpu_write_power(gpu, VIVS_PM_PULSE_EATER, data: pulse_eater); |
714 | } |
715 | |
716 | static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu) |
717 | { |
718 | WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED || |
719 | gpu->state == ETNA_GPU_STATE_RESET)); |
720 | |
721 | if ((etnaviv_is_model_rev(gpu, GC320, 0x5007) || |
722 | etnaviv_is_model_rev(gpu, GC320, 0x5220)) && |
723 | gpu_read(gpu, VIVS_HI_CHIP_TIME) != 0x2062400) { |
724 | u32 mc_memory_debug; |
725 | |
726 | mc_memory_debug = gpu_read(gpu, VIVS_MC_DEBUG_MEMORY) & ~0xff; |
727 | |
728 | if (gpu->identity.revision == 0x5007) |
729 | mc_memory_debug |= 0x0c; |
730 | else |
731 | mc_memory_debug |= 0x08; |
732 | |
733 | gpu_write(gpu, VIVS_MC_DEBUG_MEMORY, data: mc_memory_debug); |
734 | } |
735 | |
736 | /* enable module-level clock gating */ |
737 | etnaviv_gpu_enable_mlcg(gpu); |
738 | |
739 | /* |
740 | * Update GPU AXI cache atttribute to "cacheable, no allocate". |
741 | * This is necessary to prevent the iMX6 SoC locking up. |
742 | */ |
743 | gpu_write(gpu, VIVS_HI_AXI_CONFIG, |
744 | VIVS_HI_AXI_CONFIG_AWCACHE(2) | |
745 | VIVS_HI_AXI_CONFIG_ARCACHE(2)); |
746 | |
747 | /* GC2000 rev 5108 needs a special bus config */ |
748 | if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) { |
749 | u32 bus_config = gpu_read(gpu, VIVS_MC_BUS_CONFIG); |
750 | bus_config &= ~(VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG__MASK | |
751 | VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG__MASK); |
752 | bus_config |= VIVS_MC_BUS_CONFIG_FE_BUS_CONFIG(1) | |
753 | VIVS_MC_BUS_CONFIG_TX_BUS_CONFIG(0); |
754 | gpu_write(gpu, VIVS_MC_BUS_CONFIG, data: bus_config); |
755 | } |
756 | |
757 | if (gpu->sec_mode == ETNA_SEC_KERNEL) { |
758 | u32 val = gpu_read(gpu, VIVS_MMUv2_AHB_CONTROL); |
759 | val |= VIVS_MMUv2_AHB_CONTROL_NONSEC_ACCESS; |
760 | gpu_write(gpu, VIVS_MMUv2_AHB_CONTROL, data: val); |
761 | } |
762 | |
763 | /* setup the pulse eater */ |
764 | etnaviv_gpu_setup_pulse_eater(gpu); |
765 | |
766 | gpu_write(gpu, VIVS_HI_INTR_ENBL, data: ~0U); |
767 | |
768 | gpu->state = ETNA_GPU_STATE_INITIALIZED; |
769 | } |
770 | |
771 | int etnaviv_gpu_init(struct etnaviv_gpu *gpu) |
772 | { |
773 | struct etnaviv_drm_private *priv = gpu->drm->dev_private; |
774 | dma_addr_t cmdbuf_paddr; |
775 | int ret, i; |
776 | |
777 | ret = pm_runtime_get_sync(dev: gpu->dev); |
778 | if (ret < 0) { |
779 | dev_err(gpu->dev, "Failed to enable GPU power domain\n" ); |
780 | goto pm_put; |
781 | } |
782 | |
783 | etnaviv_hw_identify(gpu); |
784 | |
785 | if (gpu->identity.model == 0) { |
786 | dev_err(gpu->dev, "Unknown GPU model\n" ); |
787 | ret = -ENXIO; |
788 | goto fail; |
789 | } |
790 | |
791 | if (gpu->identity.nn_core_count > 0) |
792 | dev_warn(gpu->dev, "etnaviv has been instantiated on a NPU, " |
793 | "for which the UAPI is still experimental\n" ); |
794 | |
795 | /* Exclude VG cores with FE2.0 */ |
796 | if (gpu->identity.features & chipFeatures_PIPE_VG && |
797 | gpu->identity.features & chipFeatures_FE20) { |
798 | dev_info(gpu->dev, "Ignoring GPU with VG and FE2.0\n" ); |
799 | ret = -ENXIO; |
800 | goto fail; |
801 | } |
802 | |
803 | /* |
804 | * On cores with security features supported, we claim control over the |
805 | * security states. |
806 | */ |
807 | if ((gpu->identity.minor_features7 & chipMinorFeatures7_BIT_SECURITY) && |
808 | (gpu->identity.minor_features10 & chipMinorFeatures10_SECURITY_AHB)) |
809 | gpu->sec_mode = ETNA_SEC_KERNEL; |
810 | |
811 | gpu->state = ETNA_GPU_STATE_IDENTIFIED; |
812 | |
813 | ret = etnaviv_hw_reset(gpu); |
814 | if (ret) { |
815 | dev_err(gpu->dev, "GPU reset failed\n" ); |
816 | goto fail; |
817 | } |
818 | |
819 | ret = etnaviv_iommu_global_init(gpu); |
820 | if (ret) |
821 | goto fail; |
822 | |
823 | /* |
824 | * If the GPU is part of a system with DMA addressing limitations, |
825 | * request pages for our SHM backend buffers from the DMA32 zone to |
826 | * hopefully avoid performance killing SWIOTLB bounce buffering. |
827 | */ |
828 | if (dma_addressing_limited(dev: gpu->dev)) |
829 | priv->shm_gfp_mask |= GFP_DMA32; |
830 | |
831 | /* Create buffer: */ |
832 | ret = etnaviv_cmdbuf_init(suballoc: priv->cmdbuf_suballoc, cmdbuf: &gpu->buffer, |
833 | PAGE_SIZE); |
834 | if (ret) { |
835 | dev_err(gpu->dev, "could not create command buffer\n" ); |
836 | goto fail; |
837 | } |
838 | |
839 | /* |
840 | * Set the GPU linear window to cover the cmdbuf region, as the GPU |
841 | * won't be able to start execution otherwise. The alignment to 128M is |
842 | * chosen arbitrarily but helps in debugging, as the MMU offset |
843 | * calculations are much more straight forward this way. |
844 | * |
845 | * On MC1.0 cores the linear window offset is ignored by the TS engine, |
846 | * leading to inconsistent memory views. Avoid using the offset on those |
847 | * cores if possible, otherwise disable the TS feature. |
848 | */ |
849 | cmdbuf_paddr = ALIGN_DOWN(etnaviv_cmdbuf_get_pa(&gpu->buffer), SZ_128M); |
850 | |
851 | if (!(gpu->identity.features & chipFeatures_PIPE_3D) || |
852 | (gpu->identity.minor_features0 & chipMinorFeatures0_MC20)) { |
853 | if (cmdbuf_paddr >= SZ_2G) |
854 | priv->mmu_global->memory_base = SZ_2G; |
855 | else |
856 | priv->mmu_global->memory_base = cmdbuf_paddr; |
857 | } else if (cmdbuf_paddr + SZ_128M >= SZ_2G) { |
858 | dev_info(gpu->dev, |
859 | "Need to move linear window on MC1.0, disabling TS\n" ); |
860 | gpu->identity.features &= ~chipFeatures_FAST_CLEAR; |
861 | priv->mmu_global->memory_base = SZ_2G; |
862 | } |
863 | |
864 | /* Setup event management */ |
865 | spin_lock_init(&gpu->event_spinlock); |
866 | init_completion(x: &gpu->event_free); |
867 | bitmap_zero(dst: gpu->event_bitmap, ETNA_NR_EVENTS); |
868 | for (i = 0; i < ARRAY_SIZE(gpu->event); i++) |
869 | complete(&gpu->event_free); |
870 | |
871 | /* Now program the hardware */ |
872 | mutex_lock(&gpu->lock); |
873 | etnaviv_gpu_hw_init(gpu); |
874 | mutex_unlock(lock: &gpu->lock); |
875 | |
876 | pm_runtime_mark_last_busy(dev: gpu->dev); |
877 | pm_runtime_put_autosuspend(dev: gpu->dev); |
878 | |
879 | return 0; |
880 | |
881 | fail: |
882 | pm_runtime_mark_last_busy(dev: gpu->dev); |
883 | pm_put: |
884 | pm_runtime_put_autosuspend(dev: gpu->dev); |
885 | |
886 | return ret; |
887 | } |
888 | |
889 | #ifdef CONFIG_DEBUG_FS |
890 | struct dma_debug { |
891 | u32 address[2]; |
892 | u32 state[2]; |
893 | }; |
894 | |
895 | static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug) |
896 | { |
897 | u32 i; |
898 | |
899 | debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); |
900 | debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); |
901 | |
902 | for (i = 0; i < 500; i++) { |
903 | debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); |
904 | debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); |
905 | |
906 | if (debug->address[0] != debug->address[1]) |
907 | break; |
908 | |
909 | if (debug->state[0] != debug->state[1]) |
910 | break; |
911 | } |
912 | } |
913 | |
914 | int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m) |
915 | { |
916 | struct dma_debug debug; |
917 | u32 dma_lo, dma_hi, axi, idle; |
918 | int ret; |
919 | |
920 | seq_printf(m, fmt: "%s Status:\n" , dev_name(dev: gpu->dev)); |
921 | |
922 | ret = pm_runtime_get_sync(dev: gpu->dev); |
923 | if (ret < 0) |
924 | goto pm_put; |
925 | |
926 | dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW); |
927 | dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH); |
928 | axi = gpu_read(gpu, VIVS_HI_AXI_STATUS); |
929 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); |
930 | |
931 | verify_dma(gpu, debug: &debug); |
932 | |
933 | seq_puts(m, s: "\tidentity\n" ); |
934 | seq_printf(m, fmt: "\t model: 0x%x\n" , gpu->identity.model); |
935 | seq_printf(m, fmt: "\t revision: 0x%x\n" , gpu->identity.revision); |
936 | seq_printf(m, fmt: "\t product_id: 0x%x\n" , gpu->identity.product_id); |
937 | seq_printf(m, fmt: "\t customer_id: 0x%x\n" , gpu->identity.customer_id); |
938 | seq_printf(m, fmt: "\t eco_id: 0x%x\n" , gpu->identity.eco_id); |
939 | |
940 | seq_puts(m, s: "\tfeatures\n" ); |
941 | seq_printf(m, fmt: "\t major_features: 0x%08x\n" , |
942 | gpu->identity.features); |
943 | seq_printf(m, fmt: "\t minor_features0: 0x%08x\n" , |
944 | gpu->identity.minor_features0); |
945 | seq_printf(m, fmt: "\t minor_features1: 0x%08x\n" , |
946 | gpu->identity.minor_features1); |
947 | seq_printf(m, fmt: "\t minor_features2: 0x%08x\n" , |
948 | gpu->identity.minor_features2); |
949 | seq_printf(m, fmt: "\t minor_features3: 0x%08x\n" , |
950 | gpu->identity.minor_features3); |
951 | seq_printf(m, fmt: "\t minor_features4: 0x%08x\n" , |
952 | gpu->identity.minor_features4); |
953 | seq_printf(m, fmt: "\t minor_features5: 0x%08x\n" , |
954 | gpu->identity.minor_features5); |
955 | seq_printf(m, fmt: "\t minor_features6: 0x%08x\n" , |
956 | gpu->identity.minor_features6); |
957 | seq_printf(m, fmt: "\t minor_features7: 0x%08x\n" , |
958 | gpu->identity.minor_features7); |
959 | seq_printf(m, fmt: "\t minor_features8: 0x%08x\n" , |
960 | gpu->identity.minor_features8); |
961 | seq_printf(m, fmt: "\t minor_features9: 0x%08x\n" , |
962 | gpu->identity.minor_features9); |
963 | seq_printf(m, fmt: "\t minor_features10: 0x%08x\n" , |
964 | gpu->identity.minor_features10); |
965 | seq_printf(m, fmt: "\t minor_features11: 0x%08x\n" , |
966 | gpu->identity.minor_features11); |
967 | |
968 | seq_puts(m, s: "\tspecs\n" ); |
969 | seq_printf(m, fmt: "\t stream_count: %d\n" , |
970 | gpu->identity.stream_count); |
971 | seq_printf(m, fmt: "\t register_max: %d\n" , |
972 | gpu->identity.register_max); |
973 | seq_printf(m, fmt: "\t thread_count: %d\n" , |
974 | gpu->identity.thread_count); |
975 | seq_printf(m, fmt: "\t vertex_cache_size: %d\n" , |
976 | gpu->identity.vertex_cache_size); |
977 | seq_printf(m, fmt: "\t shader_core_count: %d\n" , |
978 | gpu->identity.shader_core_count); |
979 | seq_printf(m, fmt: "\t nn_core_count: %d\n" , |
980 | gpu->identity.nn_core_count); |
981 | seq_printf(m, fmt: "\t pixel_pipes: %d\n" , |
982 | gpu->identity.pixel_pipes); |
983 | seq_printf(m, fmt: "\t vertex_output_buffer_size: %d\n" , |
984 | gpu->identity.vertex_output_buffer_size); |
985 | seq_printf(m, fmt: "\t buffer_size: %d\n" , |
986 | gpu->identity.buffer_size); |
987 | seq_printf(m, fmt: "\t instruction_count: %d\n" , |
988 | gpu->identity.instruction_count); |
989 | seq_printf(m, fmt: "\t num_constants: %d\n" , |
990 | gpu->identity.num_constants); |
991 | seq_printf(m, fmt: "\t varyings_count: %d\n" , |
992 | gpu->identity.varyings_count); |
993 | |
994 | seq_printf(m, fmt: "\taxi: 0x%08x\n" , axi); |
995 | seq_printf(m, fmt: "\tidle: 0x%08x\n" , idle); |
996 | idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP; |
997 | if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) |
998 | seq_puts(m, s: "\t FE is not idle\n" ); |
999 | if ((idle & VIVS_HI_IDLE_STATE_DE) == 0) |
1000 | seq_puts(m, s: "\t DE is not idle\n" ); |
1001 | if ((idle & VIVS_HI_IDLE_STATE_PE) == 0) |
1002 | seq_puts(m, s: "\t PE is not idle\n" ); |
1003 | if ((idle & VIVS_HI_IDLE_STATE_SH) == 0) |
1004 | seq_puts(m, s: "\t SH is not idle\n" ); |
1005 | if ((idle & VIVS_HI_IDLE_STATE_PA) == 0) |
1006 | seq_puts(m, s: "\t PA is not idle\n" ); |
1007 | if ((idle & VIVS_HI_IDLE_STATE_SE) == 0) |
1008 | seq_puts(m, s: "\t SE is not idle\n" ); |
1009 | if ((idle & VIVS_HI_IDLE_STATE_RA) == 0) |
1010 | seq_puts(m, s: "\t RA is not idle\n" ); |
1011 | if ((idle & VIVS_HI_IDLE_STATE_TX) == 0) |
1012 | seq_puts(m, s: "\t TX is not idle\n" ); |
1013 | if ((idle & VIVS_HI_IDLE_STATE_VG) == 0) |
1014 | seq_puts(m, s: "\t VG is not idle\n" ); |
1015 | if ((idle & VIVS_HI_IDLE_STATE_IM) == 0) |
1016 | seq_puts(m, s: "\t IM is not idle\n" ); |
1017 | if ((idle & VIVS_HI_IDLE_STATE_FP) == 0) |
1018 | seq_puts(m, s: "\t FP is not idle\n" ); |
1019 | if ((idle & VIVS_HI_IDLE_STATE_TS) == 0) |
1020 | seq_puts(m, s: "\t TS is not idle\n" ); |
1021 | if ((idle & VIVS_HI_IDLE_STATE_BL) == 0) |
1022 | seq_puts(m, s: "\t BL is not idle\n" ); |
1023 | if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0) |
1024 | seq_puts(m, s: "\t ASYNCFE is not idle\n" ); |
1025 | if ((idle & VIVS_HI_IDLE_STATE_MC) == 0) |
1026 | seq_puts(m, s: "\t MC is not idle\n" ); |
1027 | if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0) |
1028 | seq_puts(m, s: "\t PPA is not idle\n" ); |
1029 | if ((idle & VIVS_HI_IDLE_STATE_WD) == 0) |
1030 | seq_puts(m, s: "\t WD is not idle\n" ); |
1031 | if ((idle & VIVS_HI_IDLE_STATE_NN) == 0) |
1032 | seq_puts(m, s: "\t NN is not idle\n" ); |
1033 | if ((idle & VIVS_HI_IDLE_STATE_TP) == 0) |
1034 | seq_puts(m, s: "\t TP is not idle\n" ); |
1035 | if (idle & VIVS_HI_IDLE_STATE_AXI_LP) |
1036 | seq_puts(m, s: "\t AXI low power mode\n" ); |
1037 | |
1038 | if (gpu->identity.features & chipFeatures_DEBUG_MODE) { |
1039 | u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0); |
1040 | u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1); |
1041 | u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE); |
1042 | |
1043 | seq_puts(m, s: "\tMC\n" ); |
1044 | seq_printf(m, fmt: "\t read0: 0x%08x\n" , read0); |
1045 | seq_printf(m, fmt: "\t read1: 0x%08x\n" , read1); |
1046 | seq_printf(m, fmt: "\t write: 0x%08x\n" , write); |
1047 | } |
1048 | |
1049 | seq_puts(m, s: "\tDMA " ); |
1050 | |
1051 | if (debug.address[0] == debug.address[1] && |
1052 | debug.state[0] == debug.state[1]) { |
1053 | seq_puts(m, s: "seems to be stuck\n" ); |
1054 | } else if (debug.address[0] == debug.address[1]) { |
1055 | seq_puts(m, s: "address is constant\n" ); |
1056 | } else { |
1057 | seq_puts(m, s: "is running\n" ); |
1058 | } |
1059 | |
1060 | seq_printf(m, fmt: "\t address 0: 0x%08x\n" , debug.address[0]); |
1061 | seq_printf(m, fmt: "\t address 1: 0x%08x\n" , debug.address[1]); |
1062 | seq_printf(m, fmt: "\t state 0: 0x%08x\n" , debug.state[0]); |
1063 | seq_printf(m, fmt: "\t state 1: 0x%08x\n" , debug.state[1]); |
1064 | seq_printf(m, fmt: "\t last fetch 64 bit word: 0x%08x 0x%08x\n" , |
1065 | dma_lo, dma_hi); |
1066 | |
1067 | ret = 0; |
1068 | |
1069 | pm_runtime_mark_last_busy(dev: gpu->dev); |
1070 | pm_put: |
1071 | pm_runtime_put_autosuspend(dev: gpu->dev); |
1072 | |
1073 | return ret; |
1074 | } |
1075 | #endif |
1076 | |
1077 | /* fence object management */ |
1078 | struct etnaviv_fence { |
1079 | struct etnaviv_gpu *gpu; |
1080 | struct dma_fence base; |
1081 | }; |
1082 | |
1083 | static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence) |
1084 | { |
1085 | return container_of(fence, struct etnaviv_fence, base); |
1086 | } |
1087 | |
1088 | static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence) |
1089 | { |
1090 | return "etnaviv" ; |
1091 | } |
1092 | |
1093 | static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence) |
1094 | { |
1095 | struct etnaviv_fence *f = to_etnaviv_fence(fence); |
1096 | |
1097 | return dev_name(dev: f->gpu->dev); |
1098 | } |
1099 | |
1100 | static bool etnaviv_fence_signaled(struct dma_fence *fence) |
1101 | { |
1102 | struct etnaviv_fence *f = to_etnaviv_fence(fence); |
1103 | |
1104 | return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0; |
1105 | } |
1106 | |
1107 | static void etnaviv_fence_release(struct dma_fence *fence) |
1108 | { |
1109 | struct etnaviv_fence *f = to_etnaviv_fence(fence); |
1110 | |
1111 | kfree_rcu(f, base.rcu); |
1112 | } |
1113 | |
1114 | static const struct dma_fence_ops etnaviv_fence_ops = { |
1115 | .get_driver_name = etnaviv_fence_get_driver_name, |
1116 | .get_timeline_name = etnaviv_fence_get_timeline_name, |
1117 | .signaled = etnaviv_fence_signaled, |
1118 | .release = etnaviv_fence_release, |
1119 | }; |
1120 | |
1121 | static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu) |
1122 | { |
1123 | struct etnaviv_fence *f; |
1124 | |
1125 | /* |
1126 | * GPU lock must already be held, otherwise fence completion order might |
1127 | * not match the seqno order assigned here. |
1128 | */ |
1129 | lockdep_assert_held(&gpu->lock); |
1130 | |
1131 | f = kzalloc(size: sizeof(*f), GFP_KERNEL); |
1132 | if (!f) |
1133 | return NULL; |
1134 | |
1135 | f->gpu = gpu; |
1136 | |
1137 | dma_fence_init(fence: &f->base, ops: &etnaviv_fence_ops, lock: &gpu->fence_spinlock, |
1138 | context: gpu->fence_context, seqno: ++gpu->next_fence); |
1139 | |
1140 | return &f->base; |
1141 | } |
1142 | |
1143 | /* returns true if fence a comes after fence b */ |
1144 | static inline bool fence_after(u32 a, u32 b) |
1145 | { |
1146 | return (s32)(a - b) > 0; |
1147 | } |
1148 | |
1149 | /* |
1150 | * event management: |
1151 | */ |
1152 | |
1153 | static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events, |
1154 | unsigned int *events) |
1155 | { |
1156 | unsigned long timeout = msecs_to_jiffies(m: 10 * 10000); |
1157 | unsigned i, acquired = 0, rpm_count = 0; |
1158 | int ret; |
1159 | |
1160 | for (i = 0; i < nr_events; i++) { |
1161 | unsigned long remaining; |
1162 | |
1163 | remaining = wait_for_completion_timeout(x: &gpu->event_free, timeout); |
1164 | |
1165 | if (!remaining) { |
1166 | dev_err(gpu->dev, "wait_for_completion_timeout failed" ); |
1167 | ret = -EBUSY; |
1168 | goto out; |
1169 | } |
1170 | |
1171 | acquired++; |
1172 | timeout = remaining; |
1173 | } |
1174 | |
1175 | spin_lock(lock: &gpu->event_spinlock); |
1176 | |
1177 | for (i = 0; i < nr_events; i++) { |
1178 | int event = find_first_zero_bit(addr: gpu->event_bitmap, ETNA_NR_EVENTS); |
1179 | |
1180 | events[i] = event; |
1181 | memset(&gpu->event[event], 0, sizeof(struct etnaviv_event)); |
1182 | set_bit(nr: event, addr: gpu->event_bitmap); |
1183 | } |
1184 | |
1185 | spin_unlock(lock: &gpu->event_spinlock); |
1186 | |
1187 | for (i = 0; i < nr_events; i++) { |
1188 | ret = pm_runtime_resume_and_get(dev: gpu->dev); |
1189 | if (ret) |
1190 | goto out_rpm; |
1191 | rpm_count++; |
1192 | } |
1193 | |
1194 | return 0; |
1195 | |
1196 | out_rpm: |
1197 | for (i = 0; i < rpm_count; i++) |
1198 | pm_runtime_put_autosuspend(dev: gpu->dev); |
1199 | out: |
1200 | for (i = 0; i < acquired; i++) |
1201 | complete(&gpu->event_free); |
1202 | |
1203 | return ret; |
1204 | } |
1205 | |
1206 | static void event_free(struct etnaviv_gpu *gpu, unsigned int event) |
1207 | { |
1208 | if (!test_bit(event, gpu->event_bitmap)) { |
1209 | dev_warn(gpu->dev, "event %u is already marked as free" , |
1210 | event); |
1211 | } else { |
1212 | clear_bit(nr: event, addr: gpu->event_bitmap); |
1213 | complete(&gpu->event_free); |
1214 | } |
1215 | |
1216 | pm_runtime_put_autosuspend(dev: gpu->dev); |
1217 | } |
1218 | |
1219 | /* |
1220 | * Cmdstream submission/retirement: |
1221 | */ |
1222 | int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu, |
1223 | u32 id, struct drm_etnaviv_timespec *timeout) |
1224 | { |
1225 | struct dma_fence *fence; |
1226 | int ret; |
1227 | |
1228 | /* |
1229 | * Look up the fence and take a reference. We might still find a fence |
1230 | * whose refcount has already dropped to zero. dma_fence_get_rcu |
1231 | * pretends we didn't find a fence in that case. |
1232 | */ |
1233 | rcu_read_lock(); |
1234 | fence = xa_load(&gpu->user_fences, index: id); |
1235 | if (fence) |
1236 | fence = dma_fence_get_rcu(fence); |
1237 | rcu_read_unlock(); |
1238 | |
1239 | if (!fence) |
1240 | return 0; |
1241 | |
1242 | if (!timeout) { |
1243 | /* No timeout was requested: just test for completion */ |
1244 | ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY; |
1245 | } else { |
1246 | unsigned long remaining = etnaviv_timeout_to_jiffies(timeout); |
1247 | |
1248 | ret = dma_fence_wait_timeout(fence, intr: true, timeout: remaining); |
1249 | if (ret == 0) |
1250 | ret = -ETIMEDOUT; |
1251 | else if (ret != -ERESTARTSYS) |
1252 | ret = 0; |
1253 | |
1254 | } |
1255 | |
1256 | dma_fence_put(fence); |
1257 | return ret; |
1258 | } |
1259 | |
1260 | /* |
1261 | * Wait for an object to become inactive. This, on it's own, is not race |
1262 | * free: the object is moved by the scheduler off the active list, and |
1263 | * then the iova is put. Moreover, the object could be re-submitted just |
1264 | * after we notice that it's become inactive. |
1265 | * |
1266 | * Although the retirement happens under the gpu lock, we don't want to hold |
1267 | * that lock in this function while waiting. |
1268 | */ |
1269 | int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu, |
1270 | struct etnaviv_gem_object *etnaviv_obj, |
1271 | struct drm_etnaviv_timespec *timeout) |
1272 | { |
1273 | unsigned long remaining; |
1274 | long ret; |
1275 | |
1276 | if (!timeout) |
1277 | return !is_active(etnaviv_obj) ? 0 : -EBUSY; |
1278 | |
1279 | remaining = etnaviv_timeout_to_jiffies(timeout); |
1280 | |
1281 | ret = wait_event_interruptible_timeout(gpu->fence_event, |
1282 | !is_active(etnaviv_obj), |
1283 | remaining); |
1284 | if (ret > 0) |
1285 | return 0; |
1286 | else if (ret == -ERESTARTSYS) |
1287 | return -ERESTARTSYS; |
1288 | else |
1289 | return -ETIMEDOUT; |
1290 | } |
1291 | |
1292 | static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu, |
1293 | struct etnaviv_event *event, unsigned int flags) |
1294 | { |
1295 | const struct etnaviv_gem_submit *submit = event->submit; |
1296 | unsigned int i; |
1297 | |
1298 | for (i = 0; i < submit->nr_pmrs; i++) { |
1299 | const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; |
1300 | |
1301 | if (pmr->flags == flags) |
1302 | etnaviv_perfmon_process(gpu, pmr, exec_state: submit->exec_state); |
1303 | } |
1304 | } |
1305 | |
1306 | static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu, |
1307 | struct etnaviv_event *event) |
1308 | { |
1309 | u32 val; |
1310 | |
1311 | /* disable clock gating */ |
1312 | val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
1313 | val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; |
1314 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, data: val); |
1315 | |
1316 | /* enable debug register */ |
1317 | val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
1318 | val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; |
1319 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: val); |
1320 | |
1321 | sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE); |
1322 | } |
1323 | |
1324 | static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu, |
1325 | struct etnaviv_event *event) |
1326 | { |
1327 | const struct etnaviv_gem_submit *submit = event->submit; |
1328 | unsigned int i; |
1329 | u32 val; |
1330 | |
1331 | sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST); |
1332 | |
1333 | for (i = 0; i < submit->nr_pmrs; i++) { |
1334 | const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; |
1335 | |
1336 | *pmr->bo_vma = pmr->sequence; |
1337 | } |
1338 | |
1339 | /* disable debug register */ |
1340 | val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
1341 | val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; |
1342 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, data: val); |
1343 | |
1344 | /* enable clock gating */ |
1345 | val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
1346 | val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; |
1347 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, data: val); |
1348 | } |
1349 | |
1350 | |
1351 | /* add bo's to gpu's ring, and kick gpu: */ |
1352 | struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit) |
1353 | { |
1354 | struct etnaviv_gpu *gpu = submit->gpu; |
1355 | struct dma_fence *gpu_fence; |
1356 | unsigned int i, nr_events = 1, event[3]; |
1357 | int ret; |
1358 | |
1359 | /* |
1360 | * if there are performance monitor requests we need to have |
1361 | * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE |
1362 | * requests. |
1363 | * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests |
1364 | * and update the sequence number for userspace. |
1365 | */ |
1366 | if (submit->nr_pmrs) |
1367 | nr_events = 3; |
1368 | |
1369 | ret = event_alloc(gpu, nr_events, events: event); |
1370 | if (ret) { |
1371 | DRM_ERROR("no free events\n" ); |
1372 | pm_runtime_put_noidle(dev: gpu->dev); |
1373 | return NULL; |
1374 | } |
1375 | |
1376 | mutex_lock(&gpu->lock); |
1377 | |
1378 | gpu_fence = etnaviv_gpu_fence_alloc(gpu); |
1379 | if (!gpu_fence) { |
1380 | for (i = 0; i < nr_events; i++) |
1381 | event_free(gpu, event: event[i]); |
1382 | |
1383 | goto out_unlock; |
1384 | } |
1385 | |
1386 | if (gpu->state == ETNA_GPU_STATE_INITIALIZED) |
1387 | etnaviv_gpu_start_fe_idleloop(gpu, context: submit->mmu_context); |
1388 | |
1389 | if (submit->prev_mmu_context) |
1390 | etnaviv_iommu_context_put(ctx: submit->prev_mmu_context); |
1391 | submit->prev_mmu_context = etnaviv_iommu_context_get(ctx: gpu->mmu_context); |
1392 | |
1393 | if (submit->nr_pmrs) { |
1394 | gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre; |
1395 | kref_get(kref: &submit->refcount); |
1396 | gpu->event[event[1]].submit = submit; |
1397 | etnaviv_sync_point_queue(gpu, event: event[1]); |
1398 | } |
1399 | |
1400 | gpu->event[event[0]].fence = gpu_fence; |
1401 | submit->cmdbuf.user_size = submit->cmdbuf.size - 8; |
1402 | etnaviv_buffer_queue(gpu, exec_state: submit->exec_state, mmu: submit->mmu_context, |
1403 | event: event[0], cmdbuf: &submit->cmdbuf); |
1404 | |
1405 | if (submit->nr_pmrs) { |
1406 | gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post; |
1407 | kref_get(kref: &submit->refcount); |
1408 | gpu->event[event[2]].submit = submit; |
1409 | etnaviv_sync_point_queue(gpu, event: event[2]); |
1410 | } |
1411 | |
1412 | out_unlock: |
1413 | mutex_unlock(lock: &gpu->lock); |
1414 | |
1415 | return gpu_fence; |
1416 | } |
1417 | |
1418 | static void sync_point_worker(struct work_struct *work) |
1419 | { |
1420 | struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu, |
1421 | sync_point_work); |
1422 | struct etnaviv_event *event = &gpu->event[gpu->sync_point_event]; |
1423 | u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); |
1424 | |
1425 | event->sync_point(gpu, event); |
1426 | etnaviv_submit_put(submit: event->submit); |
1427 | event_free(gpu, event: gpu->sync_point_event); |
1428 | |
1429 | /* restart FE last to avoid GPU and IRQ racing against this worker */ |
1430 | etnaviv_gpu_start_fe(gpu, address: addr + 2, prefetch: 2); |
1431 | } |
1432 | |
1433 | void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit) |
1434 | { |
1435 | struct etnaviv_gpu *gpu = submit->gpu; |
1436 | char *comm = NULL, *cmd = NULL; |
1437 | struct task_struct *task; |
1438 | unsigned int i; |
1439 | |
1440 | dev_err(gpu->dev, "recover hung GPU!\n" ); |
1441 | |
1442 | task = get_pid_task(pid: submit->pid, PIDTYPE_PID); |
1443 | if (task) { |
1444 | comm = kstrdup(s: task->comm, GFP_KERNEL); |
1445 | cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL); |
1446 | put_task_struct(t: task); |
1447 | } |
1448 | |
1449 | if (comm && cmd) |
1450 | dev_err(gpu->dev, "offending task: %s (%s)\n" , comm, cmd); |
1451 | |
1452 | kfree(objp: cmd); |
1453 | kfree(objp: comm); |
1454 | |
1455 | if (pm_runtime_get_sync(dev: gpu->dev) < 0) |
1456 | goto pm_put; |
1457 | |
1458 | mutex_lock(&gpu->lock); |
1459 | |
1460 | etnaviv_hw_reset(gpu); |
1461 | |
1462 | /* complete all events, the GPU won't do it after the reset */ |
1463 | spin_lock(lock: &gpu->event_spinlock); |
1464 | for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS) |
1465 | event_free(gpu, event: i); |
1466 | spin_unlock(lock: &gpu->event_spinlock); |
1467 | |
1468 | etnaviv_gpu_hw_init(gpu); |
1469 | |
1470 | mutex_unlock(lock: &gpu->lock); |
1471 | pm_runtime_mark_last_busy(dev: gpu->dev); |
1472 | pm_put: |
1473 | pm_runtime_put_autosuspend(dev: gpu->dev); |
1474 | } |
1475 | |
1476 | static void dump_mmu_fault(struct etnaviv_gpu *gpu) |
1477 | { |
1478 | static const char *fault_reasons[] = { |
1479 | "slave not present" , |
1480 | "page not present" , |
1481 | "write violation" , |
1482 | "out of bounds" , |
1483 | "read security violation" , |
1484 | "write security violation" , |
1485 | }; |
1486 | |
1487 | u32 status_reg, status; |
1488 | int i; |
1489 | |
1490 | if (gpu->sec_mode == ETNA_SEC_NONE) |
1491 | status_reg = VIVS_MMUv2_STATUS; |
1492 | else |
1493 | status_reg = VIVS_MMUv2_SEC_STATUS; |
1494 | |
1495 | status = gpu_read(gpu, reg: status_reg); |
1496 | dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n" , status); |
1497 | |
1498 | for (i = 0; i < 4; i++) { |
1499 | const char *reason = "unknown" ; |
1500 | u32 address_reg; |
1501 | u32 mmu_status; |
1502 | |
1503 | mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK; |
1504 | if (!mmu_status) |
1505 | continue; |
1506 | |
1507 | if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons)) |
1508 | reason = fault_reasons[mmu_status - 1]; |
1509 | |
1510 | if (gpu->sec_mode == ETNA_SEC_NONE) |
1511 | address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i); |
1512 | else |
1513 | address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR; |
1514 | |
1515 | dev_err_ratelimited(gpu->dev, |
1516 | "MMU %d fault (%s) addr 0x%08x\n" , |
1517 | i, reason, gpu_read(gpu, address_reg)); |
1518 | } |
1519 | } |
1520 | |
1521 | static irqreturn_t irq_handler(int irq, void *data) |
1522 | { |
1523 | struct etnaviv_gpu *gpu = data; |
1524 | irqreturn_t ret = IRQ_NONE; |
1525 | |
1526 | u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE); |
1527 | |
1528 | if (intr != 0) { |
1529 | int event; |
1530 | |
1531 | pm_runtime_mark_last_busy(dev: gpu->dev); |
1532 | |
1533 | dev_dbg(gpu->dev, "intr 0x%08x\n" , intr); |
1534 | |
1535 | if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) { |
1536 | dev_err(gpu->dev, "AXI bus error\n" ); |
1537 | intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR; |
1538 | } |
1539 | |
1540 | if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) { |
1541 | dump_mmu_fault(gpu); |
1542 | gpu->state = ETNA_GPU_STATE_FAULT; |
1543 | drm_sched_fault(sched: &gpu->sched); |
1544 | intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION; |
1545 | } |
1546 | |
1547 | while ((event = ffs(intr)) != 0) { |
1548 | struct dma_fence *fence; |
1549 | |
1550 | event -= 1; |
1551 | |
1552 | intr &= ~(1 << event); |
1553 | |
1554 | dev_dbg(gpu->dev, "event %u\n" , event); |
1555 | |
1556 | if (gpu->event[event].sync_point) { |
1557 | gpu->sync_point_event = event; |
1558 | queue_work(wq: gpu->wq, work: &gpu->sync_point_work); |
1559 | } |
1560 | |
1561 | fence = gpu->event[event].fence; |
1562 | if (!fence) |
1563 | continue; |
1564 | |
1565 | gpu->event[event].fence = NULL; |
1566 | |
1567 | /* |
1568 | * Events can be processed out of order. Eg, |
1569 | * - allocate and queue event 0 |
1570 | * - allocate event 1 |
1571 | * - event 0 completes, we process it |
1572 | * - allocate and queue event 0 |
1573 | * - event 1 and event 0 complete |
1574 | * we can end up processing event 0 first, then 1. |
1575 | */ |
1576 | if (fence_after(a: fence->seqno, b: gpu->completed_fence)) |
1577 | gpu->completed_fence = fence->seqno; |
1578 | dma_fence_signal(fence); |
1579 | |
1580 | event_free(gpu, event); |
1581 | } |
1582 | |
1583 | ret = IRQ_HANDLED; |
1584 | } |
1585 | |
1586 | return ret; |
1587 | } |
1588 | |
1589 | static int etnaviv_gpu_clk_enable(struct etnaviv_gpu *gpu) |
1590 | { |
1591 | int ret; |
1592 | |
1593 | ret = clk_prepare_enable(clk: gpu->clk_reg); |
1594 | if (ret) |
1595 | return ret; |
1596 | |
1597 | ret = clk_prepare_enable(clk: gpu->clk_bus); |
1598 | if (ret) |
1599 | goto disable_clk_reg; |
1600 | |
1601 | ret = clk_prepare_enable(clk: gpu->clk_core); |
1602 | if (ret) |
1603 | goto disable_clk_bus; |
1604 | |
1605 | ret = clk_prepare_enable(clk: gpu->clk_shader); |
1606 | if (ret) |
1607 | goto disable_clk_core; |
1608 | |
1609 | return 0; |
1610 | |
1611 | disable_clk_core: |
1612 | clk_disable_unprepare(clk: gpu->clk_core); |
1613 | disable_clk_bus: |
1614 | clk_disable_unprepare(clk: gpu->clk_bus); |
1615 | disable_clk_reg: |
1616 | clk_disable_unprepare(clk: gpu->clk_reg); |
1617 | |
1618 | return ret; |
1619 | } |
1620 | |
1621 | static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu) |
1622 | { |
1623 | clk_disable_unprepare(clk: gpu->clk_shader); |
1624 | clk_disable_unprepare(clk: gpu->clk_core); |
1625 | clk_disable_unprepare(clk: gpu->clk_bus); |
1626 | clk_disable_unprepare(clk: gpu->clk_reg); |
1627 | |
1628 | return 0; |
1629 | } |
1630 | |
1631 | int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms) |
1632 | { |
1633 | unsigned long timeout = jiffies + msecs_to_jiffies(m: timeout_ms); |
1634 | |
1635 | do { |
1636 | u32 idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); |
1637 | |
1638 | if ((idle & gpu->idle_mask) == gpu->idle_mask) |
1639 | return 0; |
1640 | |
1641 | if (time_is_before_jiffies(timeout)) { |
1642 | dev_warn(gpu->dev, |
1643 | "timed out waiting for idle: idle=0x%x\n" , |
1644 | idle); |
1645 | return -ETIMEDOUT; |
1646 | } |
1647 | |
1648 | udelay(5); |
1649 | } while (1); |
1650 | } |
1651 | |
1652 | static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu) |
1653 | { |
1654 | if (gpu->state == ETNA_GPU_STATE_RUNNING) { |
1655 | /* Replace the last WAIT with END */ |
1656 | mutex_lock(&gpu->lock); |
1657 | etnaviv_buffer_end(gpu); |
1658 | mutex_unlock(lock: &gpu->lock); |
1659 | |
1660 | /* |
1661 | * We know that only the FE is busy here, this should |
1662 | * happen quickly (as the WAIT is only 200 cycles). If |
1663 | * we fail, just warn and continue. |
1664 | */ |
1665 | etnaviv_gpu_wait_idle(gpu, timeout_ms: 100); |
1666 | |
1667 | gpu->state = ETNA_GPU_STATE_INITIALIZED; |
1668 | } |
1669 | |
1670 | gpu->exec_state = -1; |
1671 | } |
1672 | |
1673 | static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu) |
1674 | { |
1675 | int ret; |
1676 | |
1677 | ret = mutex_lock_killable(&gpu->lock); |
1678 | if (ret) |
1679 | return ret; |
1680 | |
1681 | etnaviv_gpu_update_clock(gpu); |
1682 | etnaviv_gpu_hw_init(gpu); |
1683 | |
1684 | mutex_unlock(lock: &gpu->lock); |
1685 | |
1686 | return 0; |
1687 | } |
1688 | |
1689 | static int |
1690 | etnaviv_gpu_cooling_get_max_state(struct thermal_cooling_device *cdev, |
1691 | unsigned long *state) |
1692 | { |
1693 | *state = 6; |
1694 | |
1695 | return 0; |
1696 | } |
1697 | |
1698 | static int |
1699 | etnaviv_gpu_cooling_get_cur_state(struct thermal_cooling_device *cdev, |
1700 | unsigned long *state) |
1701 | { |
1702 | struct etnaviv_gpu *gpu = cdev->devdata; |
1703 | |
1704 | *state = gpu->freq_scale; |
1705 | |
1706 | return 0; |
1707 | } |
1708 | |
1709 | static int |
1710 | etnaviv_gpu_cooling_set_cur_state(struct thermal_cooling_device *cdev, |
1711 | unsigned long state) |
1712 | { |
1713 | struct etnaviv_gpu *gpu = cdev->devdata; |
1714 | |
1715 | mutex_lock(&gpu->lock); |
1716 | gpu->freq_scale = state; |
1717 | if (!pm_runtime_suspended(dev: gpu->dev)) |
1718 | etnaviv_gpu_update_clock(gpu); |
1719 | mutex_unlock(lock: &gpu->lock); |
1720 | |
1721 | return 0; |
1722 | } |
1723 | |
1724 | static const struct thermal_cooling_device_ops cooling_ops = { |
1725 | .get_max_state = etnaviv_gpu_cooling_get_max_state, |
1726 | .get_cur_state = etnaviv_gpu_cooling_get_cur_state, |
1727 | .set_cur_state = etnaviv_gpu_cooling_set_cur_state, |
1728 | }; |
1729 | |
1730 | static int etnaviv_gpu_bind(struct device *dev, struct device *master, |
1731 | void *data) |
1732 | { |
1733 | struct drm_device *drm = data; |
1734 | struct etnaviv_drm_private *priv = drm->dev_private; |
1735 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); |
1736 | int ret; |
1737 | |
1738 | if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) { |
1739 | gpu->cooling = thermal_of_cooling_device_register(np: dev->of_node, |
1740 | (char *)dev_name(dev), gpu, &cooling_ops); |
1741 | if (IS_ERR(ptr: gpu->cooling)) |
1742 | return PTR_ERR(ptr: gpu->cooling); |
1743 | } |
1744 | |
1745 | gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0); |
1746 | if (!gpu->wq) { |
1747 | ret = -ENOMEM; |
1748 | goto out_thermal; |
1749 | } |
1750 | |
1751 | ret = etnaviv_sched_init(gpu); |
1752 | if (ret) |
1753 | goto out_workqueue; |
1754 | |
1755 | if (!IS_ENABLED(CONFIG_PM)) { |
1756 | ret = etnaviv_gpu_clk_enable(gpu); |
1757 | if (ret < 0) |
1758 | goto out_sched; |
1759 | } |
1760 | |
1761 | gpu->drm = drm; |
1762 | gpu->fence_context = dma_fence_context_alloc(num: 1); |
1763 | xa_init_flags(xa: &gpu->user_fences, XA_FLAGS_ALLOC); |
1764 | spin_lock_init(&gpu->fence_spinlock); |
1765 | |
1766 | INIT_WORK(&gpu->sync_point_work, sync_point_worker); |
1767 | init_waitqueue_head(&gpu->fence_event); |
1768 | |
1769 | priv->gpu[priv->num_gpus++] = gpu; |
1770 | |
1771 | return 0; |
1772 | |
1773 | out_sched: |
1774 | etnaviv_sched_fini(gpu); |
1775 | |
1776 | out_workqueue: |
1777 | destroy_workqueue(wq: gpu->wq); |
1778 | |
1779 | out_thermal: |
1780 | if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) |
1781 | thermal_cooling_device_unregister(gpu->cooling); |
1782 | |
1783 | return ret; |
1784 | } |
1785 | |
1786 | static void etnaviv_gpu_unbind(struct device *dev, struct device *master, |
1787 | void *data) |
1788 | { |
1789 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); |
1790 | |
1791 | DBG("%s" , dev_name(gpu->dev)); |
1792 | |
1793 | destroy_workqueue(wq: gpu->wq); |
1794 | |
1795 | etnaviv_sched_fini(gpu); |
1796 | |
1797 | if (IS_ENABLED(CONFIG_PM)) { |
1798 | pm_runtime_get_sync(dev: gpu->dev); |
1799 | pm_runtime_put_sync_suspend(dev: gpu->dev); |
1800 | } else { |
1801 | etnaviv_gpu_hw_suspend(gpu); |
1802 | etnaviv_gpu_clk_disable(gpu); |
1803 | } |
1804 | |
1805 | if (gpu->mmu_context) |
1806 | etnaviv_iommu_context_put(ctx: gpu->mmu_context); |
1807 | |
1808 | etnaviv_cmdbuf_free(cmdbuf: &gpu->buffer); |
1809 | etnaviv_iommu_global_fini(gpu); |
1810 | |
1811 | gpu->drm = NULL; |
1812 | xa_destroy(&gpu->user_fences); |
1813 | |
1814 | if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) |
1815 | thermal_cooling_device_unregister(gpu->cooling); |
1816 | gpu->cooling = NULL; |
1817 | } |
1818 | |
1819 | static const struct component_ops gpu_ops = { |
1820 | .bind = etnaviv_gpu_bind, |
1821 | .unbind = etnaviv_gpu_unbind, |
1822 | }; |
1823 | |
1824 | static const struct of_device_id etnaviv_gpu_match[] = { |
1825 | { |
1826 | .compatible = "vivante,gc" |
1827 | }, |
1828 | { /* sentinel */ } |
1829 | }; |
1830 | MODULE_DEVICE_TABLE(of, etnaviv_gpu_match); |
1831 | |
1832 | static int etnaviv_gpu_platform_probe(struct platform_device *pdev) |
1833 | { |
1834 | struct device *dev = &pdev->dev; |
1835 | struct etnaviv_gpu *gpu; |
1836 | int err; |
1837 | |
1838 | gpu = devm_kzalloc(dev, size: sizeof(*gpu), GFP_KERNEL); |
1839 | if (!gpu) |
1840 | return -ENOMEM; |
1841 | |
1842 | gpu->dev = &pdev->dev; |
1843 | mutex_init(&gpu->lock); |
1844 | mutex_init(&gpu->sched_lock); |
1845 | |
1846 | /* Map registers: */ |
1847 | gpu->mmio = devm_platform_ioremap_resource(pdev, index: 0); |
1848 | if (IS_ERR(ptr: gpu->mmio)) |
1849 | return PTR_ERR(ptr: gpu->mmio); |
1850 | |
1851 | /* Get Interrupt: */ |
1852 | gpu->irq = platform_get_irq(pdev, 0); |
1853 | if (gpu->irq < 0) |
1854 | return gpu->irq; |
1855 | |
1856 | err = devm_request_irq(dev: &pdev->dev, irq: gpu->irq, handler: irq_handler, irqflags: 0, |
1857 | devname: dev_name(dev: gpu->dev), dev_id: gpu); |
1858 | if (err) { |
1859 | dev_err(dev, "failed to request IRQ%u: %d\n" , gpu->irq, err); |
1860 | return err; |
1861 | } |
1862 | |
1863 | /* Get Clocks: */ |
1864 | gpu->clk_reg = devm_clk_get_optional(dev: &pdev->dev, id: "reg" ); |
1865 | DBG("clk_reg: %p" , gpu->clk_reg); |
1866 | if (IS_ERR(ptr: gpu->clk_reg)) |
1867 | return PTR_ERR(ptr: gpu->clk_reg); |
1868 | |
1869 | gpu->clk_bus = devm_clk_get_optional(dev: &pdev->dev, id: "bus" ); |
1870 | DBG("clk_bus: %p" , gpu->clk_bus); |
1871 | if (IS_ERR(ptr: gpu->clk_bus)) |
1872 | return PTR_ERR(ptr: gpu->clk_bus); |
1873 | |
1874 | gpu->clk_core = devm_clk_get(dev: &pdev->dev, id: "core" ); |
1875 | DBG("clk_core: %p" , gpu->clk_core); |
1876 | if (IS_ERR(ptr: gpu->clk_core)) |
1877 | return PTR_ERR(ptr: gpu->clk_core); |
1878 | gpu->base_rate_core = clk_get_rate(clk: gpu->clk_core); |
1879 | |
1880 | gpu->clk_shader = devm_clk_get_optional(dev: &pdev->dev, id: "shader" ); |
1881 | DBG("clk_shader: %p" , gpu->clk_shader); |
1882 | if (IS_ERR(ptr: gpu->clk_shader)) |
1883 | return PTR_ERR(ptr: gpu->clk_shader); |
1884 | gpu->base_rate_shader = clk_get_rate(clk: gpu->clk_shader); |
1885 | |
1886 | /* TODO: figure out max mapped size */ |
1887 | dev_set_drvdata(dev, data: gpu); |
1888 | |
1889 | /* |
1890 | * We treat the device as initially suspended. The runtime PM |
1891 | * autosuspend delay is rather arbitary: no measurements have |
1892 | * yet been performed to determine an appropriate value. |
1893 | */ |
1894 | pm_runtime_use_autosuspend(dev: gpu->dev); |
1895 | pm_runtime_set_autosuspend_delay(dev: gpu->dev, delay: 200); |
1896 | pm_runtime_enable(dev: gpu->dev); |
1897 | |
1898 | err = component_add(&pdev->dev, &gpu_ops); |
1899 | if (err < 0) { |
1900 | dev_err(&pdev->dev, "failed to register component: %d\n" , err); |
1901 | return err; |
1902 | } |
1903 | |
1904 | return 0; |
1905 | } |
1906 | |
1907 | static int etnaviv_gpu_platform_remove(struct platform_device *pdev) |
1908 | { |
1909 | component_del(&pdev->dev, &gpu_ops); |
1910 | pm_runtime_disable(dev: &pdev->dev); |
1911 | return 0; |
1912 | } |
1913 | |
1914 | static int etnaviv_gpu_rpm_suspend(struct device *dev) |
1915 | { |
1916 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); |
1917 | u32 idle, mask; |
1918 | |
1919 | /* If there are any jobs in the HW queue, we're not idle */ |
1920 | if (atomic_read(v: &gpu->sched.hw_rq_count)) |
1921 | return -EBUSY; |
1922 | |
1923 | /* Check whether the hardware (except FE and MC) is idle */ |
1924 | mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE | |
1925 | VIVS_HI_IDLE_STATE_MC); |
1926 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask; |
1927 | if (idle != mask) { |
1928 | dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n" , |
1929 | idle); |
1930 | return -EBUSY; |
1931 | } |
1932 | |
1933 | etnaviv_gpu_hw_suspend(gpu); |
1934 | |
1935 | gpu->state = ETNA_GPU_STATE_IDENTIFIED; |
1936 | |
1937 | return etnaviv_gpu_clk_disable(gpu); |
1938 | } |
1939 | |
1940 | static int etnaviv_gpu_rpm_resume(struct device *dev) |
1941 | { |
1942 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); |
1943 | int ret; |
1944 | |
1945 | ret = etnaviv_gpu_clk_enable(gpu); |
1946 | if (ret) |
1947 | return ret; |
1948 | |
1949 | /* Re-initialise the basic hardware state */ |
1950 | if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) { |
1951 | ret = etnaviv_gpu_hw_resume(gpu); |
1952 | if (ret) { |
1953 | etnaviv_gpu_clk_disable(gpu); |
1954 | return ret; |
1955 | } |
1956 | } |
1957 | |
1958 | return 0; |
1959 | } |
1960 | |
1961 | static const struct dev_pm_ops etnaviv_gpu_pm_ops = { |
1962 | RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL) |
1963 | }; |
1964 | |
1965 | struct platform_driver etnaviv_gpu_driver = { |
1966 | .driver = { |
1967 | .name = "etnaviv-gpu" , |
1968 | .owner = THIS_MODULE, |
1969 | .pm = pm_ptr(&etnaviv_gpu_pm_ops), |
1970 | .of_match_table = etnaviv_gpu_match, |
1971 | }, |
1972 | .probe = etnaviv_gpu_platform_probe, |
1973 | .remove = etnaviv_gpu_platform_remove, |
1974 | .id_table = gpu_ids, |
1975 | }; |
1976 | |