Commit | Line | Data |
---|---|---|
f6ffbd4f | 1 | // SPDX-License-Identifier: GPL-2.0 |
a8c21a54 | 2 | /* |
f6ffbd4f | 3 | * Copyright (C) 2015-2018 Etnaviv Project |
a8c21a54 T |
4 | */ |
5 | ||
f9d255f4 | 6 | #include <linux/clk.h> |
a8c21a54 | 7 | #include <linux/component.h> |
6eae41fe | 8 | #include <linux/delay.h> |
f54d1867 | 9 | #include <linux/dma-fence.h> |
6eae41fe | 10 | #include <linux/dma-mapping.h> |
2e737e52 | 11 | #include <linux/module.h> |
a8c21a54 | 12 | #include <linux/of_device.h> |
2e737e52 LS |
13 | #include <linux/platform_device.h> |
14 | #include <linux/pm_runtime.h> | |
f9d255f4 | 15 | #include <linux/regulator/consumer.h> |
bcdfb5e5 | 16 | #include <linux/thermal.h> |
ea1f5729 LS |
17 | |
18 | #include "etnaviv_cmdbuf.h" | |
a8c21a54 T |
19 | #include "etnaviv_dump.h" |
20 | #include "etnaviv_gpu.h" | |
21 | #include "etnaviv_gem.h" | |
22 | #include "etnaviv_mmu.h" | |
357713ce | 23 | #include "etnaviv_perfmon.h" |
e93b6dee | 24 | #include "etnaviv_sched.h" |
a8c21a54 T |
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 | ||
a8c21a54 T |
35 | /* |
36 | * Driver functions: | |
37 | */ | |
38 | ||
39 | int etnaviv_gpu_get_param(struct etnaviv_gpu *gpu, u32 param, u64 *value) | |
40 | { | |
088880dd LS |
41 | struct etnaviv_drm_private *priv = gpu->drm->dev_private; |
42 | ||
a8c21a54 T |
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 | ||
602eb489 RK |
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 | ||
0538aaf9 LS |
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 | ||
a8c21a54 T |
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 | ||
602eb489 RK |
144 | case ETNAVIV_PARAM_GPU_NUM_VARYINGS: |
145 | *value = gpu->identity.varyings_count; | |
146 | break; | |
147 | ||
088880dd LS |
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 | ||
1ff79a4a CG |
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 | ||
a8c21a54 T |
167 | default: |
168 | DBG("%s: invalid param: %u", dev_name(gpu->dev), param); | |
169 | return -EINVAL; | |
170 | } | |
171 | ||
172 | return 0; | |
173 | } | |
174 | ||
472f79dc RK |
175 | |
176 | #define etnaviv_is_model_rev(gpu, mod, rev) \ | |
177 | ((gpu)->identity.model == chipModel_##mod && \ | |
178 | (gpu)->identity.revision == rev) | |
52f36ba1 RK |
179 | #define etnaviv_field(val, field) \ |
180 | (((val) & field##__MASK) >> field##__SHIFT) | |
181 | ||
a8c21a54 T |
182 | static void etnaviv_hw_specs(struct etnaviv_gpu *gpu) |
183 | { | |
184 | if (gpu->identity.minor_features0 & | |
185 | chipMinorFeatures0_MORE_MINOR_FEATURES) { | |
602eb489 RK |
186 | u32 specs[4]; |
187 | unsigned int streams; | |
a8c21a54 T |
188 | |
189 | specs[0] = gpu_read(gpu, VIVS_HI_CHIP_SPECS); | |
190 | specs[1] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_2); | |
602eb489 RK |
191 | specs[2] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_3); |
192 | specs[3] = gpu_read(gpu, VIVS_HI_CHIP_SPECS_4); | |
a8c21a54 | 193 | |
52f36ba1 RK |
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); | |
a8c21a54 | 206 | gpu->identity.vertex_output_buffer_size = |
52f36ba1 RK |
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); | |
602eb489 RK |
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; | |
a8c21a54 T |
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; | |
507f8991 | 238 | else if (gpu->identity.model == chipModel_GC400) |
a8c21a54 T |
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; | |
507f8991 | 246 | else if (gpu->identity.model == chipModel_GC400) |
a8c21a54 | 247 | gpu->identity.thread_count = 64; |
507f8991 RK |
248 | else if (gpu->identity.model == chipModel_GC500 || |
249 | gpu->identity.model == chipModel_GC530) | |
a8c21a54 T |
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; | |
507f8991 | 271 | } else if (gpu->identity.model == chipModel_GC400) { |
a8c21a54 T |
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: | |
472f79dc | 284 | if (etnaviv_is_model_rev(gpu, GC2000, 0x5108) || |
507f8991 | 285 | gpu->identity.model == chipModel_GC880) |
a8c21a54 T |
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; | |
602eb489 RK |
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; | |
a8c21a54 T |
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. */ | |
52f36ba1 | 339 | if (etnaviv_field(chipIdentity, VIVS_HI_CHIP_IDENTITY_FAMILY) == 0x01) { |
507f8991 | 340 | gpu->identity.model = chipModel_GC500; |
52f36ba1 RK |
341 | gpu->identity.revision = etnaviv_field(chipIdentity, |
342 | VIVS_HI_CHIP_IDENTITY_REVISION); | |
a8c21a54 | 343 | } else { |
815e45bb | 344 | u32 chipDate = gpu_read(gpu, VIVS_HI_CHIP_DATE); |
a8c21a54 T |
345 | |
346 | gpu->identity.model = gpu_read(gpu, VIVS_HI_CHIP_MODEL); | |
347 | gpu->identity.revision = gpu_read(gpu, VIVS_HI_CHIP_REV); | |
815e45bb | 348 | gpu->identity.customer_id = gpu_read(gpu, VIVS_HI_CHIP_CUSTOMER_ID); |
2c5bf028 CG |
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 | } | |
a8c21a54 T |
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 && | |
507f8991 | 366 | gpu->identity.model != chipModel_GC420) { |
a8c21a54 T |
367 | gpu->identity.model = gpu->identity.model & 0x0400; |
368 | } | |
369 | ||
370 | /* Another special case */ | |
472f79dc | 371 | if (etnaviv_is_model_rev(gpu, GC300, 0x2201)) { |
a8c21a54 T |
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 | } | |
12ff4bde LS |
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 | } | |
815e45bb CG |
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; | |
a8c21a54 T |
400 | } |
401 | ||
402 | dev_info(gpu->dev, "model: GC%x, revision: %x\n", | |
403 | gpu->identity.model, gpu->identity.revision); | |
404 | ||
2b76f5be | 405 | gpu->idle_mask = ~VIVS_HI_IDLE_STATE_AXI_LP; |
681c19c8 LS |
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 | ||
a8c21a54 T |
413 | gpu->identity.features = gpu_read(gpu, VIVS_HI_CHIP_FEATURE); |
414 | ||
415 | /* Disable fast clear on GC700. */ | |
507f8991 | 416 | if (gpu->identity.model == chipModel_GC700) |
a8c21a54 T |
417 | gpu->identity.features &= ~chipFeatures_FAST_CLEAR; |
418 | ||
cc7d3fb4 DB |
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 | ||
507f8991 RK |
425 | if ((gpu->identity.model == chipModel_GC500 && |
426 | gpu->identity.revision < 2) || | |
427 | (gpu->identity.model == chipModel_GC300 && | |
428 | gpu->identity.revision < 0x2000)) { | |
a8c21a54 T |
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; | |
602eb489 RK |
438 | gpu->identity.minor_features4 = 0; |
439 | gpu->identity.minor_features5 = 0; | |
a8c21a54 T |
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); | |
602eb489 RK |
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); | |
a8c21a54 T |
456 | } |
457 | ||
cc7d3fb4 DB |
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) | |
a8c21a54 T |
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; | |
a8c21a54 T |
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, clock | | |
476 | VIVS_HI_CLOCK_CONTROL_FSCALE_CMD_LOAD); | |
477 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, clock); | |
478 | } | |
479 | ||
bcdfb5e5 RK |
480 | static void etnaviv_gpu_update_clock(struct etnaviv_gpu *gpu) |
481 | { | |
d79fd1cc LS |
482 | if (gpu->identity.minor_features2 & |
483 | chipMinorFeatures2_DYNAMIC_FREQUENCY_SCALING) { | |
484 | clk_set_rate(gpu->clk_core, | |
485 | gpu->base_rate_core >> gpu->freq_scale); | |
486 | clk_set_rate(gpu->clk_shader, | |
487 | gpu->base_rate_shader >> gpu->freq_scale); | |
488 | } else { | |
489 | unsigned int fscale = 1 << (6 - gpu->freq_scale); | |
6eb3ecc3 | 490 | u32 clock = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); |
bcdfb5e5 | 491 | |
6eb3ecc3 LS |
492 | clock &= ~VIVS_HI_CLOCK_CONTROL_FSCALE_VAL__MASK; |
493 | clock |= VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); | |
d79fd1cc LS |
494 | etnaviv_gpu_load_clock(gpu, clock); |
495 | } | |
295b6c02 LS |
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); | |
bcdfb5e5 RK |
504 | } |
505 | ||
a8c21a54 T |
506 | static int etnaviv_hw_reset(struct etnaviv_gpu *gpu) |
507 | { | |
508 | u32 control, idle; | |
509 | unsigned long timeout; | |
510 | bool failed = true; | |
511 | ||
a8c21a54 T |
512 | /* We hope that the GPU resets in under one second */ |
513 | timeout = jiffies + msecs_to_jiffies(1000); | |
514 | ||
515 | while (time_is_after_jiffies(timeout)) { | |
a8c21a54 | 516 | /* enable clock */ |
6eb3ecc3 LS |
517 | unsigned int fscale = 1 << (6 - gpu->freq_scale); |
518 | control = VIVS_HI_CLOCK_CONTROL_FSCALE_VAL(fscale); | |
519 | etnaviv_gpu_load_clock(gpu, control); | |
a8c21a54 T |
520 | |
521 | /* isolate the GPU. */ | |
522 | control |= VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; | |
523 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); | |
524 | ||
c997c3df LS |
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, control); | |
532 | } | |
a8c21a54 T |
533 | |
534 | /* wait for reset. */ | |
40462179 | 535 | usleep_range(10, 20); |
a8c21a54 T |
536 | |
537 | /* reset soft reset bit. */ | |
538 | control &= ~VIVS_HI_CLOCK_CONTROL_SOFT_RESET; | |
539 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); | |
540 | ||
541 | /* reset GPU isolation. */ | |
542 | control &= ~VIVS_HI_CLOCK_CONTROL_ISOLATE_GPU; | |
543 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, control); | |
544 | ||
545 | /* read idle register. */ | |
546 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); | |
547 | ||
ea4ed4a5 | 548 | /* try resetting again if FE is not idle */ |
a8c21a54 T |
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 | ||
6eb3ecc3 LS |
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, control); | |
567 | ||
a8c21a54 T |
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 */ | |
bcdfb5e5 | 585 | etnaviv_gpu_update_clock(gpu); |
a8c21a54 | 586 | |
647d817d | 587 | gpu->state = ETNA_GPU_STATE_RESET; |
725cbc78 | 588 | gpu->exec_state = -1; |
f978a530 LS |
589 | if (gpu->mmu_context) |
590 | etnaviv_iommu_context_put(gpu->mmu_context); | |
725cbc78 | 591 | gpu->mmu_context = NULL; |
23e0f5a5 | 592 | |
a8c21a54 T |
593 | return 0; |
594 | } | |
595 | ||
7d0c6e71 RK |
596 | static void etnaviv_gpu_enable_mlcg(struct etnaviv_gpu *gpu) |
597 | { | |
598 | u32 pmc, ppc; | |
599 | ||
600 | /* enable clock gating */ | |
61a6920b | 601 | ppc = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
7d0c6e71 RK |
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 | ||
61a6920b | 609 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, ppc); |
7d0c6e71 | 610 | |
61a6920b | 611 | pmc = gpu_read_power(gpu, VIVS_PM_MODULE_CONTROLS); |
7d0c6e71 | 612 | |
7cef6004 | 613 | /* Disable PA clock gating for GC400+ without bugfix except for GC420 */ |
7d0c6e71 | 614 | if (gpu->identity.model >= chipModel_GC400 && |
7cef6004 LS |
615 | gpu->identity.model != chipModel_GC420 && |
616 | !(gpu->identity.minor_features3 & chipMinorFeatures3_BUG_FIXES12)) | |
7d0c6e71 RK |
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 | ||
7cef6004 LS |
632 | /* Disable TX clock gating on affected core revisions. */ |
633 | if (etnaviv_is_model_rev(gpu, GC4000, 0x5222) || | |
4bce2442 MF |
634 | etnaviv_is_model_rev(gpu, GC2000, 0x5108) || |
635 | etnaviv_is_model_rev(gpu, GC2000, 0x6202) || | |
636 | etnaviv_is_model_rev(gpu, GC2000, 0x6203)) | |
7cef6004 LS |
637 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_TX; |
638 | ||
4bce2442 | 639 | /* Disable SE and RA clock gating on affected core revisions. */ |
432f51e7 MW |
640 | if (etnaviv_is_model_rev(gpu, GC7000, 0x6202)) |
641 | pmc |= VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_SE | | |
4bce2442 | 642 | VIVS_PM_MODULE_CONTROLS_DISABLE_MODULE_CLOCK_GATING_RA; |
432f51e7 | 643 | |
7d0c6e71 RK |
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 | ||
61a6920b | 647 | gpu_write_power(gpu, VIVS_PM_MODULE_CONTROLS, pmc); |
7d0c6e71 RK |
648 | } |
649 | ||
229855b6 LS |
650 | void etnaviv_gpu_start_fe(struct etnaviv_gpu *gpu, u32 address, u16 prefetch) |
651 | { | |
652 | gpu_write(gpu, VIVS_FE_COMMAND_ADDRESS, address); | |
653 | gpu_write(gpu, VIVS_FE_COMMAND_CONTROL, | |
654 | VIVS_FE_COMMAND_CONTROL_ENABLE | | |
655 | VIVS_FE_COMMAND_CONTROL_PREFETCH(prefetch)); | |
c997c3df LS |
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 | } | |
229855b6 LS |
662 | } |
663 | ||
d6408538 LS |
664 | static void etnaviv_gpu_start_fe_idleloop(struct etnaviv_gpu *gpu, |
665 | struct etnaviv_iommu_context *context) | |
d80d842a | 666 | { |
d80d842a | 667 | u16 prefetch; |
d6408538 | 668 | u32 address; |
d80d842a | 669 | |
647d817d LS |
670 | WARN_ON(gpu->state != ETNA_GPU_STATE_INITIALIZED); |
671 | ||
d80d842a | 672 | /* setup the MMU */ |
d6408538 | 673 | etnaviv_iommu_restore(gpu, context); |
d80d842a LS |
674 | |
675 | /* Start command processor */ | |
676 | prefetch = etnaviv_buffer_init(gpu); | |
d6408538 LS |
677 | address = etnaviv_cmdbuf_get_va(&gpu->buffer, |
678 | &gpu->mmu_context->cmdbuf_mapping); | |
d80d842a LS |
679 | |
680 | etnaviv_gpu_start_fe(gpu, address, prefetch); | |
647d817d LS |
681 | |
682 | gpu->state = ETNA_GPU_STATE_RUNNING; | |
d80d842a LS |
683 | } |
684 | ||
e17a0ded WL |
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 */ | |
61a6920b | 709 | pulse_eater = gpu_read_power(gpu, VIVS_PM_PULSE_EATER); |
e17a0ded WL |
710 | pulse_eater |= BIT(18); |
711 | } | |
712 | ||
61a6920b | 713 | gpu_write_power(gpu, VIVS_PM_PULSE_EATER, pulse_eater); |
e17a0ded WL |
714 | } |
715 | ||
a8c21a54 T |
716 | static void etnaviv_gpu_hw_init(struct etnaviv_gpu *gpu) |
717 | { | |
647d817d LS |
718 | WARN_ON(!(gpu->state == ETNA_GPU_STATE_IDENTIFIED || |
719 | gpu->state == ETNA_GPU_STATE_RESET)); | |
720 | ||
472f79dc RK |
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) { | |
a8c21a54 T |
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, mc_memory_debug); | |
734 | } | |
735 | ||
7d0c6e71 RK |
736 | /* enable module-level clock gating */ |
737 | etnaviv_gpu_enable_mlcg(gpu); | |
738 | ||
a8c21a54 T |
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 */ | |
472f79dc | 748 | if (etnaviv_is_model_rev(gpu, GC2000, 0x5108)) { |
a8c21a54 T |
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, bus_config); | |
755 | } | |
756 | ||
c997c3df LS |
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, val); | |
761 | } | |
762 | ||
e17a0ded WL |
763 | /* setup the pulse eater */ |
764 | etnaviv_gpu_setup_pulse_eater(gpu); | |
765 | ||
a8c21a54 | 766 | gpu_write(gpu, VIVS_HI_INTR_ENBL, ~0U); |
647d817d LS |
767 | |
768 | gpu->state = ETNA_GPU_STATE_INITIALIZED; | |
a8c21a54 T |
769 | } |
770 | ||
771 | int etnaviv_gpu_init(struct etnaviv_gpu *gpu) | |
772 | { | |
bffe5db8 | 773 | struct etnaviv_drm_private *priv = gpu->drm->dev_private; |
4bfdd2aa | 774 | dma_addr_t cmdbuf_paddr; |
a8c21a54 | 775 | int ret, i; |
a8c21a54 T |
776 | |
777 | ret = pm_runtime_get_sync(gpu->dev); | |
1409df04 LS |
778 | if (ret < 0) { |
779 | dev_err(gpu->dev, "Failed to enable GPU power domain\n"); | |
c5d5a32e | 780 | goto pm_put; |
1409df04 | 781 | } |
a8c21a54 T |
782 | |
783 | etnaviv_hw_identify(gpu); | |
784 | ||
785 | if (gpu->identity.model == 0) { | |
786 | dev_err(gpu->dev, "Unknown GPU model\n"); | |
f6427760 RK |
787 | ret = -ENXIO; |
788 | goto fail; | |
a8c21a54 T |
789 | } |
790 | ||
d801e6f4 TV |
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 | ||
b98c6688 RK |
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 | ||
c997c3df LS |
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 | ||
647d817d LS |
811 | gpu->state = ETNA_GPU_STATE_IDENTIFIED; |
812 | ||
a8c21a54 | 813 | ret = etnaviv_hw_reset(gpu); |
1409df04 LS |
814 | if (ret) { |
815 | dev_err(gpu->dev, "GPU reset failed\n"); | |
a8c21a54 | 816 | goto fail; |
1409df04 | 817 | } |
a8c21a54 | 818 | |
27b67278 LS |
819 | ret = etnaviv_iommu_global_init(gpu); |
820 | if (ret) | |
a8c21a54 | 821 | goto fail; |
27b67278 | 822 | |
b72af445 LS |
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(gpu->dev)) | |
829 | priv->shm_gfp_mask |= GFP_DMA32; | |
830 | ||
a8c21a54 | 831 | /* Create buffer: */ |
bffe5db8 | 832 | ret = etnaviv_cmdbuf_init(priv->cmdbuf_suballoc, &gpu->buffer, |
2f9225db LS |
833 | PAGE_SIZE); |
834 | if (ret) { | |
a8c21a54 | 835 | dev_err(gpu->dev, "could not create command buffer\n"); |
17e4660a | 836 | goto fail; |
a8c21a54 T |
837 | } |
838 | ||
4bfdd2aa LS |
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 | ||
a8c21a54 T |
864 | /* Setup event management */ |
865 | spin_lock_init(&gpu->event_spinlock); | |
866 | init_completion(&gpu->event_free); | |
355502e0 CG |
867 | bitmap_zero(gpu->event_bitmap, ETNA_NR_EVENTS); |
868 | for (i = 0; i < ARRAY_SIZE(gpu->event); i++) | |
a8c21a54 | 869 | complete(&gpu->event_free); |
a8c21a54 T |
870 | |
871 | /* Now program the hardware */ | |
872 | mutex_lock(&gpu->lock); | |
873 | etnaviv_gpu_hw_init(gpu); | |
874 | mutex_unlock(&gpu->lock); | |
875 | ||
876 | pm_runtime_mark_last_busy(gpu->dev); | |
877 | pm_runtime_put_autosuspend(gpu->dev); | |
878 | ||
db41fe7d LS |
879 | gpu->initialized = true; |
880 | ||
a8c21a54 T |
881 | return 0; |
882 | ||
a8c21a54 T |
883 | fail: |
884 | pm_runtime_mark_last_busy(gpu->dev); | |
c5d5a32e | 885 | pm_put: |
a8c21a54 T |
886 | pm_runtime_put_autosuspend(gpu->dev); |
887 | ||
888 | return ret; | |
889 | } | |
890 | ||
891 | #ifdef CONFIG_DEBUG_FS | |
892 | struct dma_debug { | |
893 | u32 address[2]; | |
894 | u32 state[2]; | |
895 | }; | |
896 | ||
897 | static void verify_dma(struct etnaviv_gpu *gpu, struct dma_debug *debug) | |
898 | { | |
899 | u32 i; | |
900 | ||
901 | debug->address[0] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); | |
902 | debug->state[0] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); | |
903 | ||
904 | for (i = 0; i < 500; i++) { | |
905 | debug->address[1] = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); | |
906 | debug->state[1] = gpu_read(gpu, VIVS_FE_DMA_DEBUG_STATE); | |
907 | ||
908 | if (debug->address[0] != debug->address[1]) | |
909 | break; | |
910 | ||
911 | if (debug->state[0] != debug->state[1]) | |
912 | break; | |
913 | } | |
914 | } | |
915 | ||
916 | int etnaviv_gpu_debugfs(struct etnaviv_gpu *gpu, struct seq_file *m) | |
917 | { | |
918 | struct dma_debug debug; | |
919 | u32 dma_lo, dma_hi, axi, idle; | |
920 | int ret; | |
921 | ||
922 | seq_printf(m, "%s Status:\n", dev_name(gpu->dev)); | |
923 | ||
924 | ret = pm_runtime_get_sync(gpu->dev); | |
925 | if (ret < 0) | |
c5d5a32e | 926 | goto pm_put; |
a8c21a54 T |
927 | |
928 | dma_lo = gpu_read(gpu, VIVS_FE_DMA_LOW); | |
929 | dma_hi = gpu_read(gpu, VIVS_FE_DMA_HIGH); | |
930 | axi = gpu_read(gpu, VIVS_HI_AXI_STATUS); | |
931 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE); | |
932 | ||
933 | verify_dma(gpu, &debug); | |
934 | ||
00080663 CG |
935 | seq_puts(m, "\tidentity\n"); |
936 | seq_printf(m, "\t model: 0x%x\n", gpu->identity.model); | |
937 | seq_printf(m, "\t revision: 0x%x\n", gpu->identity.revision); | |
938 | seq_printf(m, "\t product_id: 0x%x\n", gpu->identity.product_id); | |
939 | seq_printf(m, "\t customer_id: 0x%x\n", gpu->identity.customer_id); | |
940 | seq_printf(m, "\t eco_id: 0x%x\n", gpu->identity.eco_id); | |
941 | ||
a8c21a54 | 942 | seq_puts(m, "\tfeatures\n"); |
3d9fc642 LS |
943 | seq_printf(m, "\t major_features: 0x%08x\n", |
944 | gpu->identity.features); | |
a8c21a54 T |
945 | seq_printf(m, "\t minor_features0: 0x%08x\n", |
946 | gpu->identity.minor_features0); | |
947 | seq_printf(m, "\t minor_features1: 0x%08x\n", | |
948 | gpu->identity.minor_features1); | |
949 | seq_printf(m, "\t minor_features2: 0x%08x\n", | |
950 | gpu->identity.minor_features2); | |
951 | seq_printf(m, "\t minor_features3: 0x%08x\n", | |
952 | gpu->identity.minor_features3); | |
602eb489 RK |
953 | seq_printf(m, "\t minor_features4: 0x%08x\n", |
954 | gpu->identity.minor_features4); | |
955 | seq_printf(m, "\t minor_features5: 0x%08x\n", | |
956 | gpu->identity.minor_features5); | |
0538aaf9 LS |
957 | seq_printf(m, "\t minor_features6: 0x%08x\n", |
958 | gpu->identity.minor_features6); | |
959 | seq_printf(m, "\t minor_features7: 0x%08x\n", | |
960 | gpu->identity.minor_features7); | |
961 | seq_printf(m, "\t minor_features8: 0x%08x\n", | |
962 | gpu->identity.minor_features8); | |
963 | seq_printf(m, "\t minor_features9: 0x%08x\n", | |
964 | gpu->identity.minor_features9); | |
965 | seq_printf(m, "\t minor_features10: 0x%08x\n", | |
966 | gpu->identity.minor_features10); | |
967 | seq_printf(m, "\t minor_features11: 0x%08x\n", | |
968 | gpu->identity.minor_features11); | |
a8c21a54 T |
969 | |
970 | seq_puts(m, "\tspecs\n"); | |
971 | seq_printf(m, "\t stream_count: %d\n", | |
972 | gpu->identity.stream_count); | |
973 | seq_printf(m, "\t register_max: %d\n", | |
974 | gpu->identity.register_max); | |
975 | seq_printf(m, "\t thread_count: %d\n", | |
976 | gpu->identity.thread_count); | |
977 | seq_printf(m, "\t vertex_cache_size: %d\n", | |
978 | gpu->identity.vertex_cache_size); | |
979 | seq_printf(m, "\t shader_core_count: %d\n", | |
980 | gpu->identity.shader_core_count); | |
4c22c61e LS |
981 | seq_printf(m, "\t nn_core_count: %d\n", |
982 | gpu->identity.nn_core_count); | |
a8c21a54 T |
983 | seq_printf(m, "\t pixel_pipes: %d\n", |
984 | gpu->identity.pixel_pipes); | |
985 | seq_printf(m, "\t vertex_output_buffer_size: %d\n", | |
986 | gpu->identity.vertex_output_buffer_size); | |
987 | seq_printf(m, "\t buffer_size: %d\n", | |
988 | gpu->identity.buffer_size); | |
989 | seq_printf(m, "\t instruction_count: %d\n", | |
990 | gpu->identity.instruction_count); | |
991 | seq_printf(m, "\t num_constants: %d\n", | |
992 | gpu->identity.num_constants); | |
602eb489 RK |
993 | seq_printf(m, "\t varyings_count: %d\n", |
994 | gpu->identity.varyings_count); | |
a8c21a54 T |
995 | |
996 | seq_printf(m, "\taxi: 0x%08x\n", axi); | |
997 | seq_printf(m, "\tidle: 0x%08x\n", idle); | |
998 | idle |= ~gpu->idle_mask & ~VIVS_HI_IDLE_STATE_AXI_LP; | |
999 | if ((idle & VIVS_HI_IDLE_STATE_FE) == 0) | |
1000 | seq_puts(m, "\t FE is not idle\n"); | |
1001 | if ((idle & VIVS_HI_IDLE_STATE_DE) == 0) | |
1002 | seq_puts(m, "\t DE is not idle\n"); | |
1003 | if ((idle & VIVS_HI_IDLE_STATE_PE) == 0) | |
1004 | seq_puts(m, "\t PE is not idle\n"); | |
1005 | if ((idle & VIVS_HI_IDLE_STATE_SH) == 0) | |
1006 | seq_puts(m, "\t SH is not idle\n"); | |
1007 | if ((idle & VIVS_HI_IDLE_STATE_PA) == 0) | |
1008 | seq_puts(m, "\t PA is not idle\n"); | |
1009 | if ((idle & VIVS_HI_IDLE_STATE_SE) == 0) | |
1010 | seq_puts(m, "\t SE is not idle\n"); | |
1011 | if ((idle & VIVS_HI_IDLE_STATE_RA) == 0) | |
1012 | seq_puts(m, "\t RA is not idle\n"); | |
1013 | if ((idle & VIVS_HI_IDLE_STATE_TX) == 0) | |
1014 | seq_puts(m, "\t TX is not idle\n"); | |
1015 | if ((idle & VIVS_HI_IDLE_STATE_VG) == 0) | |
1016 | seq_puts(m, "\t VG is not idle\n"); | |
1017 | if ((idle & VIVS_HI_IDLE_STATE_IM) == 0) | |
1018 | seq_puts(m, "\t IM is not idle\n"); | |
1019 | if ((idle & VIVS_HI_IDLE_STATE_FP) == 0) | |
1020 | seq_puts(m, "\t FP is not idle\n"); | |
1021 | if ((idle & VIVS_HI_IDLE_STATE_TS) == 0) | |
1022 | seq_puts(m, "\t TS is not idle\n"); | |
b1704551 GG |
1023 | if ((idle & VIVS_HI_IDLE_STATE_BL) == 0) |
1024 | seq_puts(m, "\t BL is not idle\n"); | |
1025 | if ((idle & VIVS_HI_IDLE_STATE_ASYNCFE) == 0) | |
1026 | seq_puts(m, "\t ASYNCFE is not idle\n"); | |
1027 | if ((idle & VIVS_HI_IDLE_STATE_MC) == 0) | |
1028 | seq_puts(m, "\t MC is not idle\n"); | |
1029 | if ((idle & VIVS_HI_IDLE_STATE_PPA) == 0) | |
1030 | seq_puts(m, "\t PPA is not idle\n"); | |
1031 | if ((idle & VIVS_HI_IDLE_STATE_WD) == 0) | |
1032 | seq_puts(m, "\t WD is not idle\n"); | |
1033 | if ((idle & VIVS_HI_IDLE_STATE_NN) == 0) | |
1034 | seq_puts(m, "\t NN is not idle\n"); | |
1035 | if ((idle & VIVS_HI_IDLE_STATE_TP) == 0) | |
1036 | seq_puts(m, "\t TP is not idle\n"); | |
a8c21a54 T |
1037 | if (idle & VIVS_HI_IDLE_STATE_AXI_LP) |
1038 | seq_puts(m, "\t AXI low power mode\n"); | |
1039 | ||
1040 | if (gpu->identity.features & chipFeatures_DEBUG_MODE) { | |
1041 | u32 read0 = gpu_read(gpu, VIVS_MC_DEBUG_READ0); | |
1042 | u32 read1 = gpu_read(gpu, VIVS_MC_DEBUG_READ1); | |
1043 | u32 write = gpu_read(gpu, VIVS_MC_DEBUG_WRITE); | |
1044 | ||
1045 | seq_puts(m, "\tMC\n"); | |
1046 | seq_printf(m, "\t read0: 0x%08x\n", read0); | |
1047 | seq_printf(m, "\t read1: 0x%08x\n", read1); | |
1048 | seq_printf(m, "\t write: 0x%08x\n", write); | |
1049 | } | |
1050 | ||
1051 | seq_puts(m, "\tDMA "); | |
1052 | ||
1053 | if (debug.address[0] == debug.address[1] && | |
1054 | debug.state[0] == debug.state[1]) { | |
1055 | seq_puts(m, "seems to be stuck\n"); | |
1056 | } else if (debug.address[0] == debug.address[1]) { | |
c01e0159 | 1057 | seq_puts(m, "address is constant\n"); |
a8c21a54 | 1058 | } else { |
c01e0159 | 1059 | seq_puts(m, "is running\n"); |
a8c21a54 T |
1060 | } |
1061 | ||
1062 | seq_printf(m, "\t address 0: 0x%08x\n", debug.address[0]); | |
1063 | seq_printf(m, "\t address 1: 0x%08x\n", debug.address[1]); | |
1064 | seq_printf(m, "\t state 0: 0x%08x\n", debug.state[0]); | |
1065 | seq_printf(m, "\t state 1: 0x%08x\n", debug.state[1]); | |
1066 | seq_printf(m, "\t last fetch 64 bit word: 0x%08x 0x%08x\n", | |
1067 | dma_lo, dma_hi); | |
1068 | ||
1069 | ret = 0; | |
1070 | ||
1071 | pm_runtime_mark_last_busy(gpu->dev); | |
c5d5a32e | 1072 | pm_put: |
a8c21a54 T |
1073 | pm_runtime_put_autosuspend(gpu->dev); |
1074 | ||
1075 | return ret; | |
1076 | } | |
1077 | #endif | |
1078 | ||
a8c21a54 T |
1079 | /* fence object management */ |
1080 | struct etnaviv_fence { | |
1081 | struct etnaviv_gpu *gpu; | |
f54d1867 | 1082 | struct dma_fence base; |
a8c21a54 T |
1083 | }; |
1084 | ||
f54d1867 | 1085 | static inline struct etnaviv_fence *to_etnaviv_fence(struct dma_fence *fence) |
a8c21a54 T |
1086 | { |
1087 | return container_of(fence, struct etnaviv_fence, base); | |
1088 | } | |
1089 | ||
f54d1867 | 1090 | static const char *etnaviv_fence_get_driver_name(struct dma_fence *fence) |
a8c21a54 T |
1091 | { |
1092 | return "etnaviv"; | |
1093 | } | |
1094 | ||
f54d1867 | 1095 | static const char *etnaviv_fence_get_timeline_name(struct dma_fence *fence) |
a8c21a54 T |
1096 | { |
1097 | struct etnaviv_fence *f = to_etnaviv_fence(fence); | |
1098 | ||
1099 | return dev_name(f->gpu->dev); | |
1100 | } | |
1101 | ||
f54d1867 | 1102 | static bool etnaviv_fence_signaled(struct dma_fence *fence) |
a8c21a54 T |
1103 | { |
1104 | struct etnaviv_fence *f = to_etnaviv_fence(fence); | |
1105 | ||
3283ee77 | 1106 | return (s32)(f->gpu->completed_fence - f->base.seqno) >= 0; |
a8c21a54 T |
1107 | } |
1108 | ||
f54d1867 | 1109 | static void etnaviv_fence_release(struct dma_fence *fence) |
a8c21a54 T |
1110 | { |
1111 | struct etnaviv_fence *f = to_etnaviv_fence(fence); | |
1112 | ||
1113 | kfree_rcu(f, base.rcu); | |
1114 | } | |
1115 | ||
f54d1867 | 1116 | static const struct dma_fence_ops etnaviv_fence_ops = { |
a8c21a54 T |
1117 | .get_driver_name = etnaviv_fence_get_driver_name, |
1118 | .get_timeline_name = etnaviv_fence_get_timeline_name, | |
a8c21a54 | 1119 | .signaled = etnaviv_fence_signaled, |
a8c21a54 T |
1120 | .release = etnaviv_fence_release, |
1121 | }; | |
1122 | ||
f54d1867 | 1123 | static struct dma_fence *etnaviv_gpu_fence_alloc(struct etnaviv_gpu *gpu) |
a8c21a54 T |
1124 | { |
1125 | struct etnaviv_fence *f; | |
1126 | ||
b27734c2 LS |
1127 | /* |
1128 | * GPU lock must already be held, otherwise fence completion order might | |
1129 | * not match the seqno order assigned here. | |
1130 | */ | |
1131 | lockdep_assert_held(&gpu->lock); | |
1132 | ||
a8c21a54 T |
1133 | f = kzalloc(sizeof(*f), GFP_KERNEL); |
1134 | if (!f) | |
1135 | return NULL; | |
1136 | ||
1137 | f->gpu = gpu; | |
1138 | ||
f54d1867 CW |
1139 | dma_fence_init(&f->base, &etnaviv_fence_ops, &gpu->fence_spinlock, |
1140 | gpu->fence_context, ++gpu->next_fence); | |
a8c21a54 T |
1141 | |
1142 | return &f->base; | |
1143 | } | |
1144 | ||
3283ee77 LS |
1145 | /* returns true if fence a comes after fence b */ |
1146 | static inline bool fence_after(u32 a, u32 b) | |
1147 | { | |
1148 | return (s32)(a - b) > 0; | |
1149 | } | |
1150 | ||
a8c21a54 T |
1151 | /* |
1152 | * event management: | |
1153 | */ | |
1154 | ||
95a428c1 CG |
1155 | static int event_alloc(struct etnaviv_gpu *gpu, unsigned nr_events, |
1156 | unsigned int *events) | |
a8c21a54 | 1157 | { |
5a23144c | 1158 | unsigned long timeout = msecs_to_jiffies(10 * 10000); |
f098f9b8 LS |
1159 | unsigned i, acquired = 0, rpm_count = 0; |
1160 | int ret; | |
a8c21a54 | 1161 | |
95a428c1 CG |
1162 | for (i = 0; i < nr_events; i++) { |
1163 | unsigned long ret; | |
a8c21a54 | 1164 | |
95a428c1 | 1165 | ret = wait_for_completion_timeout(&gpu->event_free, timeout); |
a8c21a54 | 1166 | |
95a428c1 CG |
1167 | if (!ret) { |
1168 | dev_err(gpu->dev, "wait_for_completion_timeout failed"); | |
f098f9b8 | 1169 | ret = -EBUSY; |
95a428c1 | 1170 | goto out; |
a8c21a54 | 1171 | } |
95a428c1 CG |
1172 | |
1173 | acquired++; | |
1174 | timeout = ret; | |
1175 | } | |
a8c21a54 | 1176 | |
5a23144c | 1177 | spin_lock(&gpu->event_spinlock); |
a8c21a54 | 1178 | |
95a428c1 CG |
1179 | for (i = 0; i < nr_events; i++) { |
1180 | int event = find_first_zero_bit(gpu->event_bitmap, ETNA_NR_EVENTS); | |
1181 | ||
1182 | events[i] = event; | |
547d340d | 1183 | memset(&gpu->event[event], 0, sizeof(struct etnaviv_event)); |
355502e0 | 1184 | set_bit(event, gpu->event_bitmap); |
a8c21a54 T |
1185 | } |
1186 | ||
5a23144c | 1187 | spin_unlock(&gpu->event_spinlock); |
a8c21a54 | 1188 | |
f098f9b8 LS |
1189 | for (i = 0; i < nr_events; i++) { |
1190 | ret = pm_runtime_resume_and_get(gpu->dev); | |
1191 | if (ret) | |
1192 | goto out_rpm; | |
1193 | rpm_count++; | |
1194 | } | |
1195 | ||
95a428c1 CG |
1196 | return 0; |
1197 | ||
f098f9b8 LS |
1198 | out_rpm: |
1199 | for (i = 0; i < rpm_count; i++) | |
1200 | pm_runtime_put_autosuspend(gpu->dev); | |
95a428c1 CG |
1201 | out: |
1202 | for (i = 0; i < acquired; i++) | |
1203 | complete(&gpu->event_free); | |
1204 | ||
f098f9b8 | 1205 | return ret; |
a8c21a54 T |
1206 | } |
1207 | ||
1208 | static void event_free(struct etnaviv_gpu *gpu, unsigned int event) | |
1209 | { | |
355502e0 | 1210 | if (!test_bit(event, gpu->event_bitmap)) { |
a8c21a54 T |
1211 | dev_warn(gpu->dev, "event %u is already marked as free", |
1212 | event); | |
a8c21a54 | 1213 | } else { |
355502e0 | 1214 | clear_bit(event, gpu->event_bitmap); |
a8c21a54 T |
1215 | complete(&gpu->event_free); |
1216 | } | |
f098f9b8 LS |
1217 | |
1218 | pm_runtime_put_autosuspend(gpu->dev); | |
a8c21a54 T |
1219 | } |
1220 | ||
1221 | /* | |
1222 | * Cmdstream submission/retirement: | |
1223 | */ | |
a8c21a54 | 1224 | int etnaviv_gpu_wait_fence_interruptible(struct etnaviv_gpu *gpu, |
38c4a4cf | 1225 | u32 id, struct drm_etnaviv_timespec *timeout) |
a8c21a54 | 1226 | { |
8bc4d885 | 1227 | struct dma_fence *fence; |
a8c21a54 T |
1228 | int ret; |
1229 | ||
8bc4d885 | 1230 | /* |
e93b6dee | 1231 | * Look up the fence and take a reference. We might still find a fence |
8bc4d885 LS |
1232 | * whose refcount has already dropped to zero. dma_fence_get_rcu |
1233 | * pretends we didn't find a fence in that case. | |
1234 | */ | |
e93b6dee | 1235 | rcu_read_lock(); |
764be123 | 1236 | fence = xa_load(&gpu->user_fences, id); |
8bc4d885 LS |
1237 | if (fence) |
1238 | fence = dma_fence_get_rcu(fence); | |
e93b6dee | 1239 | rcu_read_unlock(); |
8bc4d885 LS |
1240 | |
1241 | if (!fence) | |
1242 | return 0; | |
a8c21a54 T |
1243 | |
1244 | if (!timeout) { | |
1245 | /* No timeout was requested: just test for completion */ | |
8bc4d885 | 1246 | ret = dma_fence_is_signaled(fence) ? 0 : -EBUSY; |
a8c21a54 T |
1247 | } else { |
1248 | unsigned long remaining = etnaviv_timeout_to_jiffies(timeout); | |
1249 | ||
8bc4d885 LS |
1250 | ret = dma_fence_wait_timeout(fence, true, remaining); |
1251 | if (ret == 0) | |
a8c21a54 | 1252 | ret = -ETIMEDOUT; |
8bc4d885 | 1253 | else if (ret != -ERESTARTSYS) |
a8c21a54 | 1254 | ret = 0; |
8bc4d885 | 1255 | |
a8c21a54 T |
1256 | } |
1257 | ||
8bc4d885 | 1258 | dma_fence_put(fence); |
a8c21a54 T |
1259 | return ret; |
1260 | } | |
1261 | ||
1262 | /* | |
1263 | * Wait for an object to become inactive. This, on it's own, is not race | |
e93b6dee | 1264 | * free: the object is moved by the scheduler off the active list, and |
a8c21a54 T |
1265 | * then the iova is put. Moreover, the object could be re-submitted just |
1266 | * after we notice that it's become inactive. | |
1267 | * | |
1268 | * Although the retirement happens under the gpu lock, we don't want to hold | |
1269 | * that lock in this function while waiting. | |
1270 | */ | |
1271 | int etnaviv_gpu_wait_obj_inactive(struct etnaviv_gpu *gpu, | |
38c4a4cf AB |
1272 | struct etnaviv_gem_object *etnaviv_obj, |
1273 | struct drm_etnaviv_timespec *timeout) | |
a8c21a54 T |
1274 | { |
1275 | unsigned long remaining; | |
1276 | long ret; | |
1277 | ||
1278 | if (!timeout) | |
1279 | return !is_active(etnaviv_obj) ? 0 : -EBUSY; | |
1280 | ||
1281 | remaining = etnaviv_timeout_to_jiffies(timeout); | |
1282 | ||
1283 | ret = wait_event_interruptible_timeout(gpu->fence_event, | |
1284 | !is_active(etnaviv_obj), | |
1285 | remaining); | |
fa67ac84 | 1286 | if (ret > 0) |
a8c21a54 | 1287 | return 0; |
fa67ac84 | 1288 | else if (ret == -ERESTARTSYS) |
a8c21a54 | 1289 | return -ERESTARTSYS; |
fa67ac84 | 1290 | else |
a8c21a54 | 1291 | return -ETIMEDOUT; |
a8c21a54 T |
1292 | } |
1293 | ||
68dc0b29 CG |
1294 | static void sync_point_perfmon_sample(struct etnaviv_gpu *gpu, |
1295 | struct etnaviv_event *event, unsigned int flags) | |
1296 | { | |
ef146c00 | 1297 | const struct etnaviv_gem_submit *submit = event->submit; |
68dc0b29 CG |
1298 | unsigned int i; |
1299 | ||
ef146c00 LS |
1300 | for (i = 0; i < submit->nr_pmrs; i++) { |
1301 | const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; | |
68dc0b29 CG |
1302 | |
1303 | if (pmr->flags == flags) | |
7a9c0fe2 | 1304 | etnaviv_perfmon_process(gpu, pmr, submit->exec_state); |
68dc0b29 CG |
1305 | } |
1306 | } | |
1307 | ||
1308 | static void sync_point_perfmon_sample_pre(struct etnaviv_gpu *gpu, | |
1309 | struct etnaviv_event *event) | |
1310 | { | |
2c8b0c5a CG |
1311 | u32 val; |
1312 | ||
1313 | /* disable clock gating */ | |
61a6920b | 1314 | val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
2c8b0c5a | 1315 | val &= ~VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; |
61a6920b | 1316 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val); |
2c8b0c5a | 1317 | |
04a7d18d CG |
1318 | /* enable debug register */ |
1319 | val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); | |
1320 | val &= ~VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; | |
1321 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val); | |
1322 | ||
68dc0b29 CG |
1323 | sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_PRE); |
1324 | } | |
1325 | ||
1326 | static void sync_point_perfmon_sample_post(struct etnaviv_gpu *gpu, | |
1327 | struct etnaviv_event *event) | |
1328 | { | |
ef146c00 | 1329 | const struct etnaviv_gem_submit *submit = event->submit; |
68dc0b29 | 1330 | unsigned int i; |
2c8b0c5a | 1331 | u32 val; |
68dc0b29 CG |
1332 | |
1333 | sync_point_perfmon_sample(gpu, event, ETNA_PM_PROCESS_POST); | |
1334 | ||
ef146c00 LS |
1335 | for (i = 0; i < submit->nr_pmrs; i++) { |
1336 | const struct etnaviv_perfmon_request *pmr = submit->pmrs + i; | |
68dc0b29 CG |
1337 | |
1338 | *pmr->bo_vma = pmr->sequence; | |
1339 | } | |
2c8b0c5a | 1340 | |
04a7d18d CG |
1341 | /* disable debug register */ |
1342 | val = gpu_read(gpu, VIVS_HI_CLOCK_CONTROL); | |
1343 | val |= VIVS_HI_CLOCK_CONTROL_DISABLE_DEBUG_REGISTERS; | |
1344 | gpu_write(gpu, VIVS_HI_CLOCK_CONTROL, val); | |
1345 | ||
2c8b0c5a | 1346 | /* enable clock gating */ |
61a6920b | 1347 | val = gpu_read_power(gpu, VIVS_PM_POWER_CONTROLS); |
2c8b0c5a | 1348 | val |= VIVS_PM_POWER_CONTROLS_ENABLE_MODULE_CLOCK_GATING; |
61a6920b | 1349 | gpu_write_power(gpu, VIVS_PM_POWER_CONTROLS, val); |
68dc0b29 CG |
1350 | } |
1351 | ||
1352 | ||
a8c21a54 | 1353 | /* add bo's to gpu's ring, and kick gpu: */ |
e93b6dee | 1354 | struct dma_fence *etnaviv_gpu_submit(struct etnaviv_gem_submit *submit) |
a8c21a54 | 1355 | { |
e93b6dee LS |
1356 | struct etnaviv_gpu *gpu = submit->gpu; |
1357 | struct dma_fence *gpu_fence; | |
68dc0b29 | 1358 | unsigned int i, nr_events = 1, event[3]; |
a8c21a54 T |
1359 | int ret; |
1360 | ||
68dc0b29 CG |
1361 | /* |
1362 | * if there are performance monitor requests we need to have | |
1363 | * - a sync point to re-configure gpu and process ETNA_PM_PROCESS_PRE | |
1364 | * requests. | |
1365 | * - a sync point to re-configure gpu, process ETNA_PM_PROCESS_POST requests | |
1366 | * and update the sequence number for userspace. | |
1367 | */ | |
ef146c00 | 1368 | if (submit->nr_pmrs) |
68dc0b29 CG |
1369 | nr_events = 3; |
1370 | ||
1371 | ret = event_alloc(gpu, nr_events, event); | |
95a428c1 | 1372 | if (ret) { |
68dc0b29 | 1373 | DRM_ERROR("no free events\n"); |
c5d5a32e | 1374 | pm_runtime_put_noidle(gpu->dev); |
e93b6dee | 1375 | return NULL; |
a8c21a54 T |
1376 | } |
1377 | ||
f3cd1b06 LS |
1378 | mutex_lock(&gpu->lock); |
1379 | ||
e93b6dee LS |
1380 | gpu_fence = etnaviv_gpu_fence_alloc(gpu); |
1381 | if (!gpu_fence) { | |
68dc0b29 CG |
1382 | for (i = 0; i < nr_events; i++) |
1383 | event_free(gpu, event[i]); | |
1384 | ||
45abdf35 | 1385 | goto out_unlock; |
a8c21a54 T |
1386 | } |
1387 | ||
647d817d | 1388 | if (gpu->state == ETNA_GPU_STATE_INITIALIZED) |
d6408538 LS |
1389 | etnaviv_gpu_start_fe_idleloop(gpu, submit->mmu_context); |
1390 | ||
1391 | if (submit->prev_mmu_context) | |
1392 | etnaviv_iommu_context_put(submit->prev_mmu_context); | |
1393 | submit->prev_mmu_context = etnaviv_iommu_context_get(gpu->mmu_context); | |
17e4660a | 1394 | |
ef146c00 | 1395 | if (submit->nr_pmrs) { |
68dc0b29 | 1396 | gpu->event[event[1]].sync_point = &sync_point_perfmon_sample_pre; |
ef146c00 LS |
1397 | kref_get(&submit->refcount); |
1398 | gpu->event[event[1]].submit = submit; | |
68dc0b29 CG |
1399 | etnaviv_sync_point_queue(gpu, event[1]); |
1400 | } | |
1401 | ||
e93b6dee | 1402 | gpu->event[event[0]].fence = gpu_fence; |
6d7a20c0 | 1403 | submit->cmdbuf.user_size = submit->cmdbuf.size - 8; |
17e4660a LS |
1404 | etnaviv_buffer_queue(gpu, submit->exec_state, submit->mmu_context, |
1405 | event[0], &submit->cmdbuf); | |
68dc0b29 | 1406 | |
ef146c00 | 1407 | if (submit->nr_pmrs) { |
68dc0b29 | 1408 | gpu->event[event[2]].sync_point = &sync_point_perfmon_sample_post; |
ef146c00 LS |
1409 | kref_get(&submit->refcount); |
1410 | gpu->event[event[2]].submit = submit; | |
68dc0b29 CG |
1411 | etnaviv_sync_point_queue(gpu, event[2]); |
1412 | } | |
a8c21a54 | 1413 | |
45abdf35 | 1414 | out_unlock: |
a8c21a54 T |
1415 | mutex_unlock(&gpu->lock); |
1416 | ||
e93b6dee | 1417 | return gpu_fence; |
a8c21a54 T |
1418 | } |
1419 | ||
357713ce CG |
1420 | static void sync_point_worker(struct work_struct *work) |
1421 | { | |
1422 | struct etnaviv_gpu *gpu = container_of(work, struct etnaviv_gpu, | |
1423 | sync_point_work); | |
b9a48aa7 LS |
1424 | struct etnaviv_event *event = &gpu->event[gpu->sync_point_event]; |
1425 | u32 addr = gpu_read(gpu, VIVS_FE_DMA_ADDRESS); | |
357713ce | 1426 | |
b9a48aa7 | 1427 | event->sync_point(gpu, event); |
ef146c00 | 1428 | etnaviv_submit_put(event->submit); |
357713ce | 1429 | event_free(gpu, gpu->sync_point_event); |
b9a48aa7 LS |
1430 | |
1431 | /* restart FE last to avoid GPU and IRQ racing against this worker */ | |
1432 | etnaviv_gpu_start_fe(gpu, addr + 2, 2); | |
357713ce CG |
1433 | } |
1434 | ||
5cefcf9f LS |
1435 | void etnaviv_gpu_recover_hang(struct etnaviv_gem_submit *submit) |
1436 | { | |
1437 | struct etnaviv_gpu *gpu = submit->gpu; | |
1438 | char *comm = NULL, *cmd = NULL; | |
1439 | struct task_struct *task; | |
1440 | unsigned int i; | |
1441 | ||
1442 | dev_err(gpu->dev, "recover hung GPU!\n"); | |
1443 | ||
1444 | task = get_pid_task(submit->pid, PIDTYPE_PID); | |
1445 | if (task) { | |
1446 | comm = kstrdup(task->comm, GFP_KERNEL); | |
1447 | cmd = kstrdup_quotable_cmdline(task, GFP_KERNEL); | |
1448 | put_task_struct(task); | |
1449 | } | |
1450 | ||
1451 | if (comm && cmd) | |
1452 | dev_err(gpu->dev, "offending task: %s (%s)\n", comm, cmd); | |
1453 | ||
1454 | kfree(cmd); | |
1455 | kfree(comm); | |
1456 | ||
1457 | if (pm_runtime_get_sync(gpu->dev) < 0) | |
1458 | goto pm_put; | |
1459 | ||
1460 | mutex_lock(&gpu->lock); | |
1461 | ||
1462 | etnaviv_hw_reset(gpu); | |
1463 | ||
1464 | /* complete all events, the GPU won't do it after the reset */ | |
1465 | spin_lock(&gpu->event_spinlock); | |
1466 | for_each_set_bit(i, gpu->event_bitmap, ETNA_NR_EVENTS) | |
80f6b63e | 1467 | event_free(gpu, i); |
5cefcf9f LS |
1468 | spin_unlock(&gpu->event_spinlock); |
1469 | ||
1470 | etnaviv_gpu_hw_init(gpu); | |
1471 | ||
1472 | mutex_unlock(&gpu->lock); | |
1473 | pm_runtime_mark_last_busy(gpu->dev); | |
1474 | pm_put: | |
1475 | pm_runtime_put_autosuspend(gpu->dev); | |
1476 | } | |
1477 | ||
4df3000e LS |
1478 | static void dump_mmu_fault(struct etnaviv_gpu *gpu) |
1479 | { | |
b4bc0e74 CG |
1480 | static const char *fault_reasons[] = { |
1481 | "slave not present", | |
1482 | "page not present", | |
1483 | "write violation", | |
1484 | "out of bounds", | |
1485 | "read security violation", | |
1486 | "write security violation", | |
1487 | }; | |
1488 | ||
c997c3df | 1489 | u32 status_reg, status; |
4df3000e LS |
1490 | int i; |
1491 | ||
c997c3df LS |
1492 | if (gpu->sec_mode == ETNA_SEC_NONE) |
1493 | status_reg = VIVS_MMUv2_STATUS; | |
1494 | else | |
1495 | status_reg = VIVS_MMUv2_SEC_STATUS; | |
1496 | ||
1497 | status = gpu_read(gpu, status_reg); | |
4df3000e LS |
1498 | dev_err_ratelimited(gpu->dev, "MMU fault status 0x%08x\n", status); |
1499 | ||
1500 | for (i = 0; i < 4; i++) { | |
b4bc0e74 | 1501 | const char *reason = "unknown"; |
c997c3df | 1502 | u32 address_reg; |
b4bc0e74 | 1503 | u32 mmu_status; |
c997c3df | 1504 | |
b4bc0e74 CG |
1505 | mmu_status = (status >> (i * 4)) & VIVS_MMUv2_STATUS_EXCEPTION0__MASK; |
1506 | if (!mmu_status) | |
4df3000e LS |
1507 | continue; |
1508 | ||
b4bc0e74 CG |
1509 | if ((mmu_status - 1) < ARRAY_SIZE(fault_reasons)) |
1510 | reason = fault_reasons[mmu_status - 1]; | |
1511 | ||
c997c3df LS |
1512 | if (gpu->sec_mode == ETNA_SEC_NONE) |
1513 | address_reg = VIVS_MMUv2_EXCEPTION_ADDR(i); | |
1514 | else | |
1515 | address_reg = VIVS_MMUv2_SEC_EXCEPTION_ADDR; | |
1516 | ||
b4bc0e74 CG |
1517 | dev_err_ratelimited(gpu->dev, |
1518 | "MMU %d fault (%s) addr 0x%08x\n", | |
1519 | i, reason, gpu_read(gpu, address_reg)); | |
4df3000e LS |
1520 | } |
1521 | } | |
1522 | ||
a8c21a54 T |
1523 | static irqreturn_t irq_handler(int irq, void *data) |
1524 | { | |
1525 | struct etnaviv_gpu *gpu = data; | |
1526 | irqreturn_t ret = IRQ_NONE; | |
1527 | ||
1528 | u32 intr = gpu_read(gpu, VIVS_HI_INTR_ACKNOWLEDGE); | |
1529 | ||
1530 | if (intr != 0) { | |
1531 | int event; | |
1532 | ||
1533 | pm_runtime_mark_last_busy(gpu->dev); | |
1534 | ||
1535 | dev_dbg(gpu->dev, "intr 0x%08x\n", intr); | |
1536 | ||
1537 | if (intr & VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR) { | |
1538 | dev_err(gpu->dev, "AXI bus error\n"); | |
1539 | intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_AXI_BUS_ERROR; | |
1540 | } | |
1541 | ||
128a9b1d | 1542 | if (intr & VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION) { |
4df3000e | 1543 | dump_mmu_fault(gpu); |
128a9b1d LS |
1544 | intr &= ~VIVS_HI_INTR_ACKNOWLEDGE_MMU_EXCEPTION; |
1545 | } | |
1546 | ||
a8c21a54 | 1547 | while ((event = ffs(intr)) != 0) { |
f54d1867 | 1548 | struct dma_fence *fence; |
a8c21a54 T |
1549 | |
1550 | event -= 1; | |
1551 | ||
1552 | intr &= ~(1 << event); | |
1553 | ||
1554 | dev_dbg(gpu->dev, "event %u\n", event); | |
1555 | ||
357713ce CG |
1556 | if (gpu->event[event].sync_point) { |
1557 | gpu->sync_point_event = event; | |
a7790d78 | 1558 | queue_work(gpu->wq, &gpu->sync_point_work); |
357713ce CG |
1559 | } |
1560 | ||
a8c21a54 | 1561 | fence = gpu->event[event].fence; |
68dc0b29 CG |
1562 | if (!fence) |
1563 | continue; | |
1564 | ||
a8c21a54 | 1565 | gpu->event[event].fence = NULL; |
a8c21a54 T |
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(fence->seqno, gpu->completed_fence)) | |
1577 | gpu->completed_fence = fence->seqno; | |
8bc4d885 | 1578 | dma_fence_signal(fence); |
a8c21a54 T |
1579 | |
1580 | event_free(gpu, event); | |
a8c21a54 T |
1581 | } |
1582 | ||
a8c21a54 T |
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 | ||
6a5ef3b2 LR |
1593 | ret = clk_prepare_enable(gpu->clk_reg); |
1594 | if (ret) | |
1595 | return ret; | |
65f037e8 | 1596 | |
6a5ef3b2 LR |
1597 | ret = clk_prepare_enable(gpu->clk_bus); |
1598 | if (ret) | |
1599 | goto disable_clk_reg; | |
a8c21a54 | 1600 | |
6a5ef3b2 LR |
1601 | ret = clk_prepare_enable(gpu->clk_core); |
1602 | if (ret) | |
1603 | goto disable_clk_bus; | |
9c7310c0 | 1604 | |
6a5ef3b2 LR |
1605 | ret = clk_prepare_enable(gpu->clk_shader); |
1606 | if (ret) | |
1607 | goto disable_clk_core; | |
a8c21a54 T |
1608 | |
1609 | return 0; | |
9c7310c0 LS |
1610 | |
1611 | disable_clk_core: | |
6a5ef3b2 | 1612 | clk_disable_unprepare(gpu->clk_core); |
9c7310c0 | 1613 | disable_clk_bus: |
6a5ef3b2 | 1614 | clk_disable_unprepare(gpu->clk_bus); |
f8794fea | 1615 | disable_clk_reg: |
6a5ef3b2 | 1616 | clk_disable_unprepare(gpu->clk_reg); |
9c7310c0 LS |
1617 | |
1618 | return ret; | |
a8c21a54 T |
1619 | } |
1620 | ||
1621 | static int etnaviv_gpu_clk_disable(struct etnaviv_gpu *gpu) | |
1622 | { | |
6a5ef3b2 LR |
1623 | clk_disable_unprepare(gpu->clk_shader); |
1624 | clk_disable_unprepare(gpu->clk_core); | |
1625 | clk_disable_unprepare(gpu->clk_bus); | |
1626 | clk_disable_unprepare(gpu->clk_reg); | |
a8c21a54 T |
1627 | |
1628 | return 0; | |
1629 | } | |
1630 | ||
b88163e3 LS |
1631 | int etnaviv_gpu_wait_idle(struct etnaviv_gpu *gpu, unsigned int timeout_ms) |
1632 | { | |
1633 | unsigned long timeout = jiffies + msecs_to_jiffies(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 | ||
7cb54494 | 1652 | static void etnaviv_gpu_hw_suspend(struct etnaviv_gpu *gpu) |
a8c21a54 | 1653 | { |
647d817d | 1654 | if (gpu->state == ETNA_GPU_STATE_RUNNING) { |
a8c21a54 | 1655 | /* Replace the last WAIT with END */ |
40c27bde | 1656 | mutex_lock(&gpu->lock); |
a8c21a54 | 1657 | etnaviv_buffer_end(gpu); |
40c27bde | 1658 | mutex_unlock(&gpu->lock); |
a8c21a54 T |
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 | */ | |
b88163e3 | 1665 | etnaviv_gpu_wait_idle(gpu, 100); |
17e4660a | 1666 | |
647d817d | 1667 | gpu->state = ETNA_GPU_STATE_INITIALIZED; |
a8c21a54 T |
1668 | } |
1669 | ||
17e4660a | 1670 | gpu->exec_state = -1; |
a8c21a54 T |
1671 | } |
1672 | ||
a8c21a54 T |
1673 | static int etnaviv_gpu_hw_resume(struct etnaviv_gpu *gpu) |
1674 | { | |
a8c21a54 T |
1675 | int ret; |
1676 | ||
1677 | ret = mutex_lock_killable(&gpu->lock); | |
1678 | if (ret) | |
1679 | return ret; | |
1680 | ||
bcdfb5e5 | 1681 | etnaviv_gpu_update_clock(gpu); |
a8c21a54 T |
1682 | etnaviv_gpu_hw_init(gpu); |
1683 | ||
a8c21a54 T |
1684 | mutex_unlock(&gpu->lock); |
1685 | ||
1686 | return 0; | |
1687 | } | |
a8c21a54 | 1688 | |
bcdfb5e5 RK |
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(gpu->dev)) | |
1718 | etnaviv_gpu_update_clock(gpu); | |
1719 | mutex_unlock(&gpu->lock); | |
1720 | ||
1721 | return 0; | |
1722 | } | |
1723 | ||
96894b79 | 1724 | static const struct thermal_cooling_device_ops cooling_ops = { |
bcdfb5e5 RK |
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 | ||
a8c21a54 T |
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 | ||
49b82c38 | 1738 | if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) { |
5247e2aa | 1739 | gpu->cooling = thermal_of_cooling_device_register(dev->of_node, |
bcdfb5e5 | 1740 | (char *)dev_name(dev), gpu, &cooling_ops); |
5247e2aa LS |
1741 | if (IS_ERR(gpu->cooling)) |
1742 | return PTR_ERR(gpu->cooling); | |
1743 | } | |
bcdfb5e5 | 1744 | |
a7790d78 LS |
1745 | gpu->wq = alloc_ordered_workqueue(dev_name(dev), 0); |
1746 | if (!gpu->wq) { | |
e93b6dee LS |
1747 | ret = -ENOMEM; |
1748 | goto out_thermal; | |
a7790d78 LS |
1749 | } |
1750 | ||
e93b6dee LS |
1751 | ret = etnaviv_sched_init(gpu); |
1752 | if (ret) | |
1753 | goto out_workqueue; | |
1754 | ||
448406ea | 1755 | if (!IS_ENABLED(CONFIG_PM)) { |
6b05266a | 1756 | ret = etnaviv_gpu_clk_enable(gpu); |
448406ea LS |
1757 | if (ret < 0) |
1758 | goto out_sched; | |
1759 | } | |
a8c21a54 T |
1760 | |
1761 | gpu->drm = drm; | |
f54d1867 | 1762 | gpu->fence_context = dma_fence_context_alloc(1); |
764be123 | 1763 | xa_init_flags(&gpu->user_fences, XA_FLAGS_ALLOC); |
a8c21a54 T |
1764 | spin_lock_init(&gpu->fence_spinlock); |
1765 | ||
357713ce | 1766 | INIT_WORK(&gpu->sync_point_work, sync_point_worker); |
a8c21a54 T |
1767 | init_waitqueue_head(&gpu->fence_event); |
1768 | ||
a8c21a54 T |
1769 | priv->gpu[priv->num_gpus++] = gpu; |
1770 | ||
a8c21a54 | 1771 | return 0; |
e93b6dee LS |
1772 | |
1773 | out_sched: | |
1774 | etnaviv_sched_fini(gpu); | |
1775 | ||
1776 | out_workqueue: | |
1777 | destroy_workqueue(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; | |
a8c21a54 T |
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 | ||
a7790d78 LS |
1793 | destroy_workqueue(gpu->wq); |
1794 | ||
e93b6dee LS |
1795 | etnaviv_sched_fini(gpu); |
1796 | ||
6b05266a PC |
1797 | if (IS_ENABLED(CONFIG_PM)) { |
1798 | pm_runtime_get_sync(gpu->dev); | |
1799 | pm_runtime_put_sync_suspend(gpu->dev); | |
1800 | } else { | |
1801 | etnaviv_gpu_hw_suspend(gpu); | |
7cb54494 | 1802 | etnaviv_gpu_clk_disable(gpu); |
6b05266a | 1803 | } |
a8c21a54 | 1804 | |
8f3eea9d LS |
1805 | if (gpu->mmu_context) |
1806 | etnaviv_iommu_context_put(gpu->mmu_context); | |
1807 | ||
db41fe7d | 1808 | if (gpu->initialized) { |
2f9225db | 1809 | etnaviv_cmdbuf_free(&gpu->buffer); |
27b67278 | 1810 | etnaviv_iommu_global_fini(gpu); |
db41fe7d | 1811 | gpu->initialized = false; |
a8c21a54 T |
1812 | } |
1813 | ||
1814 | gpu->drm = NULL; | |
764be123 | 1815 | xa_destroy(&gpu->user_fences); |
bcdfb5e5 | 1816 | |
49b82c38 PZ |
1817 | if (IS_ENABLED(CONFIG_DRM_ETNAVIV_THERMAL)) |
1818 | thermal_cooling_device_unregister(gpu->cooling); | |
bcdfb5e5 | 1819 | gpu->cooling = NULL; |
a8c21a54 T |
1820 | } |
1821 | ||
1822 | static const struct component_ops gpu_ops = { | |
1823 | .bind = etnaviv_gpu_bind, | |
1824 | .unbind = etnaviv_gpu_unbind, | |
1825 | }; | |
1826 | ||
1827 | static const struct of_device_id etnaviv_gpu_match[] = { | |
1828 | { | |
1829 | .compatible = "vivante,gc" | |
1830 | }, | |
1831 | { /* sentinel */ } | |
1832 | }; | |
246774d1 | 1833 | MODULE_DEVICE_TABLE(of, etnaviv_gpu_match); |
a8c21a54 T |
1834 | |
1835 | static int etnaviv_gpu_platform_probe(struct platform_device *pdev) | |
1836 | { | |
1837 | struct device *dev = &pdev->dev; | |
1838 | struct etnaviv_gpu *gpu; | |
dc227890 | 1839 | int err; |
a8c21a54 T |
1840 | |
1841 | gpu = devm_kzalloc(dev, sizeof(*gpu), GFP_KERNEL); | |
1842 | if (!gpu) | |
1843 | return -ENOMEM; | |
1844 | ||
1845 | gpu->dev = &pdev->dev; | |
1846 | mutex_init(&gpu->lock); | |
2cd5bd98 | 1847 | mutex_init(&gpu->sched_lock); |
a8c21a54 | 1848 | |
a8c21a54 | 1849 | /* Map registers: */ |
facb180d | 1850 | gpu->mmio = devm_platform_ioremap_resource(pdev, 0); |
a8c21a54 T |
1851 | if (IS_ERR(gpu->mmio)) |
1852 | return PTR_ERR(gpu->mmio); | |
1853 | ||
1854 | /* Get Interrupt: */ | |
1855 | gpu->irq = platform_get_irq(pdev, 0); | |
0e63302d | 1856 | if (gpu->irq < 0) |
db60eda3 | 1857 | return gpu->irq; |
a8c21a54 T |
1858 | |
1859 | err = devm_request_irq(&pdev->dev, gpu->irq, irq_handler, 0, | |
1860 | dev_name(gpu->dev), gpu); | |
1861 | if (err) { | |
1862 | dev_err(dev, "failed to request IRQ%u: %d\n", gpu->irq, err); | |
db60eda3 | 1863 | return err; |
a8c21a54 T |
1864 | } |
1865 | ||
1866 | /* Get Clocks: */ | |
f76fc5ff | 1867 | gpu->clk_reg = devm_clk_get_optional(&pdev->dev, "reg"); |
65f037e8 LS |
1868 | DBG("clk_reg: %p", gpu->clk_reg); |
1869 | if (IS_ERR(gpu->clk_reg)) | |
f76fc5ff | 1870 | return PTR_ERR(gpu->clk_reg); |
65f037e8 | 1871 | |
f76fc5ff | 1872 | gpu->clk_bus = devm_clk_get_optional(&pdev->dev, "bus"); |
a8c21a54 T |
1873 | DBG("clk_bus: %p", gpu->clk_bus); |
1874 | if (IS_ERR(gpu->clk_bus)) | |
f76fc5ff | 1875 | return PTR_ERR(gpu->clk_bus); |
a8c21a54 | 1876 | |
a59052d2 | 1877 | gpu->clk_core = devm_clk_get(&pdev->dev, "core"); |
a8c21a54 T |
1878 | DBG("clk_core: %p", gpu->clk_core); |
1879 | if (IS_ERR(gpu->clk_core)) | |
f76fc5ff | 1880 | return PTR_ERR(gpu->clk_core); |
d79fd1cc | 1881 | gpu->base_rate_core = clk_get_rate(gpu->clk_core); |
a8c21a54 | 1882 | |
f76fc5ff | 1883 | gpu->clk_shader = devm_clk_get_optional(&pdev->dev, "shader"); |
a8c21a54 T |
1884 | DBG("clk_shader: %p", gpu->clk_shader); |
1885 | if (IS_ERR(gpu->clk_shader)) | |
f76fc5ff | 1886 | return PTR_ERR(gpu->clk_shader); |
d79fd1cc | 1887 | gpu->base_rate_shader = clk_get_rate(gpu->clk_shader); |
a8c21a54 T |
1888 | |
1889 | /* TODO: figure out max mapped size */ | |
1890 | dev_set_drvdata(dev, gpu); | |
1891 | ||
1892 | /* | |
1893 | * We treat the device as initially suspended. The runtime PM | |
1894 | * autosuspend delay is rather arbitary: no measurements have | |
1895 | * yet been performed to determine an appropriate value. | |
1896 | */ | |
1897 | pm_runtime_use_autosuspend(gpu->dev); | |
1898 | pm_runtime_set_autosuspend_delay(gpu->dev, 200); | |
1899 | pm_runtime_enable(gpu->dev); | |
1900 | ||
1901 | err = component_add(&pdev->dev, &gpu_ops); | |
1902 | if (err < 0) { | |
1903 | dev_err(&pdev->dev, "failed to register component: %d\n", err); | |
db60eda3 | 1904 | return err; |
a8c21a54 T |
1905 | } |
1906 | ||
1907 | return 0; | |
a8c21a54 T |
1908 | } |
1909 | ||
1910 | static int etnaviv_gpu_platform_remove(struct platform_device *pdev) | |
1911 | { | |
1912 | component_del(&pdev->dev, &gpu_ops); | |
1913 | pm_runtime_disable(&pdev->dev); | |
1914 | return 0; | |
1915 | } | |
1916 | ||
a8c21a54 T |
1917 | static int etnaviv_gpu_rpm_suspend(struct device *dev) |
1918 | { | |
1919 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); | |
1920 | u32 idle, mask; | |
1921 | ||
f4163814 LS |
1922 | /* If there are any jobs in the HW queue, we're not idle */ |
1923 | if (atomic_read(&gpu->sched.hw_rq_count)) | |
a8c21a54 T |
1924 | return -EBUSY; |
1925 | ||
1a910c11 GG |
1926 | /* Check whether the hardware (except FE and MC) is idle */ |
1927 | mask = gpu->idle_mask & ~(VIVS_HI_IDLE_STATE_FE | | |
1928 | VIVS_HI_IDLE_STATE_MC); | |
a8c21a54 | 1929 | idle = gpu_read(gpu, VIVS_HI_IDLE_STATE) & mask; |
78f2bfa3 GG |
1930 | if (idle != mask) { |
1931 | dev_warn_ratelimited(dev, "GPU not yet idle, mask: 0x%08x\n", | |
1932 | idle); | |
a8c21a54 | 1933 | return -EBUSY; |
78f2bfa3 | 1934 | } |
a8c21a54 | 1935 | |
7cb54494 LS |
1936 | etnaviv_gpu_hw_suspend(gpu); |
1937 | ||
647d817d LS |
1938 | gpu->state = ETNA_GPU_STATE_IDENTIFIED; |
1939 | ||
7cb54494 | 1940 | return etnaviv_gpu_clk_disable(gpu); |
a8c21a54 T |
1941 | } |
1942 | ||
1943 | static int etnaviv_gpu_rpm_resume(struct device *dev) | |
1944 | { | |
1945 | struct etnaviv_gpu *gpu = dev_get_drvdata(dev); | |
1946 | int ret; | |
1947 | ||
1948 | ret = etnaviv_gpu_clk_enable(gpu); | |
1949 | if (ret) | |
1950 | return ret; | |
1951 | ||
1952 | /* Re-initialise the basic hardware state */ | |
647d817d | 1953 | if (gpu->state == ETNA_GPU_STATE_IDENTIFIED) { |
a8c21a54 T |
1954 | ret = etnaviv_gpu_hw_resume(gpu); |
1955 | if (ret) { | |
1956 | etnaviv_gpu_clk_disable(gpu); | |
1957 | return ret; | |
1958 | } | |
1959 | } | |
1960 | ||
1961 | return 0; | |
1962 | } | |
a8c21a54 T |
1963 | |
1964 | static const struct dev_pm_ops etnaviv_gpu_pm_ops = { | |
6b05266a | 1965 | RUNTIME_PM_OPS(etnaviv_gpu_rpm_suspend, etnaviv_gpu_rpm_resume, NULL) |
a8c21a54 T |
1966 | }; |
1967 | ||
1968 | struct platform_driver etnaviv_gpu_driver = { | |
1969 | .driver = { | |
1970 | .name = "etnaviv-gpu", | |
1971 | .owner = THIS_MODULE, | |
6b05266a | 1972 | .pm = pm_ptr(&etnaviv_gpu_pm_ops), |
a8c21a54 T |
1973 | .of_match_table = etnaviv_gpu_match, |
1974 | }, | |
1975 | .probe = etnaviv_gpu_platform_probe, | |
1976 | .remove = etnaviv_gpu_platform_remove, | |
1977 | .id_table = gpu_ids, | |
1978 | }; |