1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright 2018 Marty E. Plummer <hanetzer@startmail.com> */
3 /* Copyright 2019 Linaro, Ltd., Rob Herring <robh@kernel.org> */
4 /* Copyright 2019 Collabora ltd. */
5 #include <linux/bitfield.h>
6 #include <linux/bitmap.h>
7 #include <linux/delay.h>
8 #include <linux/dma-mapping.h>
9 #include <linux/interrupt.h>
10 #include <linux/io.h>
11 #include <linux/iopoll.h>
12 #include <linux/platform_device.h>
13 #include <linux/pm_runtime.h>
14
15 #include "panfrost_device.h"
16 #include "panfrost_features.h"
17 #include "panfrost_issues.h"
18 #include "panfrost_gpu.h"
19 #include "panfrost_perfcnt.h"
20 #include "panfrost_regs.h"
21
panfrost_gpu_irq_handler(int irq,void * data)22 static irqreturn_t panfrost_gpu_irq_handler(int irq, void *data)
23 {
24 struct panfrost_device *pfdev = data;
25 u32 fault_status, state;
26
27 if (test_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended))
28 return IRQ_NONE;
29
30 fault_status = gpu_read(pfdev, GPU_FAULT_STATUS);
31 state = gpu_read(pfdev, GPU_INT_STAT);
32 if (!state)
33 return IRQ_NONE;
34
35 if (state & GPU_IRQ_MASK_ERROR) {
36 u64 address = (u64) gpu_read(pfdev, GPU_FAULT_ADDRESS_HI) << 32;
37 address |= gpu_read(pfdev, GPU_FAULT_ADDRESS_LO);
38
39 dev_warn(pfdev->dev, "GPU Fault 0x%08x (%s) at 0x%016llx\n",
40 fault_status, panfrost_exception_name(fault_status & 0xFF),
41 address);
42
43 if (state & GPU_IRQ_MULTIPLE_FAULT)
44 dev_warn(pfdev->dev, "There were multiple GPU faults - some have not been reported\n");
45
46 gpu_write(pfdev, GPU_INT_MASK, 0);
47 }
48
49 if (state & GPU_IRQ_PERFCNT_SAMPLE_COMPLETED)
50 panfrost_perfcnt_sample_done(pfdev);
51
52 if (state & GPU_IRQ_CLEAN_CACHES_COMPLETED)
53 panfrost_perfcnt_clean_cache_done(pfdev);
54
55 gpu_write(pfdev, GPU_INT_CLEAR, state);
56
57 return IRQ_HANDLED;
58 }
59
panfrost_gpu_soft_reset(struct panfrost_device * pfdev)60 int panfrost_gpu_soft_reset(struct panfrost_device *pfdev)
61 {
62 int ret;
63 u32 val;
64
65 gpu_write(pfdev, GPU_INT_MASK, 0);
66 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_RESET_COMPLETED);
67
68 clear_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended);
69
70 gpu_write(pfdev, GPU_CMD, GPU_CMD_SOFT_RESET);
71 ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT,
72 val, val & GPU_IRQ_RESET_COMPLETED, 10, 10000);
73
74 if (ret) {
75 dev_err(pfdev->dev, "gpu soft reset timed out, attempting hard reset\n");
76
77 gpu_write(pfdev, GPU_CMD, GPU_CMD_HARD_RESET);
78 ret = readl_relaxed_poll_timeout(pfdev->iomem + GPU_INT_RAWSTAT, val,
79 val & GPU_IRQ_RESET_COMPLETED, 100, 10000);
80 if (ret) {
81 dev_err(pfdev->dev, "gpu hard reset timed out\n");
82 return ret;
83 }
84 }
85
86 gpu_write(pfdev, GPU_INT_CLEAR, GPU_IRQ_MASK_ALL);
87
88 /* Only enable the interrupts we care about */
89 gpu_write(pfdev, GPU_INT_MASK,
90 GPU_IRQ_MASK_ERROR |
91 GPU_IRQ_PERFCNT_SAMPLE_COMPLETED |
92 GPU_IRQ_CLEAN_CACHES_COMPLETED);
93
94 /*
95 * All in-flight jobs should have released their cycle
96 * counter references upon reset, but let us make sure
97 */
98 if (drm_WARN_ON(pfdev->ddev, atomic_read(&pfdev->cycle_counter.use_count) != 0))
99 atomic_set(&pfdev->cycle_counter.use_count, 0);
100
101 return 0;
102 }
103
panfrost_gpu_amlogic_quirk(struct panfrost_device * pfdev)104 void panfrost_gpu_amlogic_quirk(struct panfrost_device *pfdev)
105 {
106 /*
107 * The Amlogic integrated Mali-T820, Mali-G31 & Mali-G52 needs
108 * these undocumented bits in GPU_PWR_OVERRIDE1 to be set in order
109 * to operate correctly.
110 */
111 gpu_write(pfdev, GPU_PWR_KEY, GPU_PWR_KEY_UNLOCK);
112 gpu_write(pfdev, GPU_PWR_OVERRIDE1, 0xfff | (0x20 << 16));
113 }
114
panfrost_gpu_init_quirks(struct panfrost_device * pfdev)115 static void panfrost_gpu_init_quirks(struct panfrost_device *pfdev)
116 {
117 u32 quirks = 0;
118
119 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_8443) ||
120 panfrost_has_hw_issue(pfdev, HW_ISSUE_11035))
121 quirks |= SC_LS_PAUSEBUFFER_DISABLE;
122
123 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10327))
124 quirks |= SC_SDC_DISABLE_OQ_DISCARD;
125
126 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_10797))
127 quirks |= SC_ENABLE_TEXGRD_FLAGS;
128
129 if (!panfrost_has_hw_issue(pfdev, GPUCORE_1619)) {
130 if (panfrost_model_cmp(pfdev, 0x750) < 0) /* T60x, T62x, T72x */
131 quirks |= SC_LS_ATTR_CHECK_DISABLE;
132 else if (panfrost_model_cmp(pfdev, 0x880) <= 0) /* T76x, T8xx */
133 quirks |= SC_LS_ALLOW_ATTR_TYPES;
134 }
135
136 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_TTRX_2968_TTRX_3162))
137 quirks |= SC_VAR_ALGORITHM;
138
139 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_TLS_HASHING))
140 quirks |= SC_TLS_HASH_ENABLE;
141
142 if (quirks)
143 gpu_write(pfdev, GPU_SHADER_CONFIG, quirks);
144
145
146 quirks = gpu_read(pfdev, GPU_TILER_CONFIG);
147
148 /* Set tiler clock gate override if required */
149 if (panfrost_has_hw_issue(pfdev, HW_ISSUE_T76X_3953))
150 quirks |= TC_CLOCK_GATE_OVERRIDE;
151
152 gpu_write(pfdev, GPU_TILER_CONFIG, quirks);
153
154
155 quirks = 0;
156 if ((panfrost_model_eq(pfdev, 0x860) || panfrost_model_eq(pfdev, 0x880)) &&
157 pfdev->features.revision >= 0x2000)
158 quirks |= JM_MAX_JOB_THROTTLE_LIMIT << JM_JOB_THROTTLE_LIMIT_SHIFT;
159 else if (panfrost_model_eq(pfdev, 0x6000) &&
160 pfdev->features.coherency_features == COHERENCY_ACE)
161 quirks |= (COHERENCY_ACE_LITE | COHERENCY_ACE) <<
162 JM_FORCE_COHERENCY_FEATURES_SHIFT;
163
164 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_IDVS_GROUP_SIZE))
165 quirks |= JM_DEFAULT_IDVS_GROUP_SIZE << JM_IDVS_GROUP_SIZE_SHIFT;
166
167 if (quirks)
168 gpu_write(pfdev, GPU_JM_CONFIG, quirks);
169
170 /* Here goes platform specific quirks */
171 if (pfdev->comp->vendor_quirk)
172 pfdev->comp->vendor_quirk(pfdev);
173 }
174
175 #define MAX_HW_REVS 6
176
177 struct panfrost_model {
178 const char *name;
179 u32 id;
180 u32 id_mask;
181 u64 features;
182 u64 issues;
183 struct {
184 u32 revision;
185 u64 issues;
186 } revs[MAX_HW_REVS];
187 };
188
189 #define GPU_MODEL(_name, _id, ...) \
190 {\
191 .name = __stringify(_name), \
192 .id = _id, \
193 .features = hw_features_##_name, \
194 .issues = hw_issues_##_name, \
195 .revs = { __VA_ARGS__ }, \
196 }
197
198 #define GPU_REV_EXT(name, _rev, _p, _s, stat) \
199 {\
200 .revision = (_rev) << 12 | (_p) << 4 | (_s), \
201 .issues = hw_issues_##name##_r##_rev##p##_p##stat, \
202 }
203 #define GPU_REV(name, r, p) GPU_REV_EXT(name, r, p, 0, )
204
205 static const struct panfrost_model gpu_models[] = {
206 /* T60x has an oddball version */
207 GPU_MODEL(t600, 0x600,
208 GPU_REV_EXT(t600, 0, 0, 1, _15dev0)),
209 GPU_MODEL(t620, 0x620,
210 GPU_REV(t620, 0, 1), GPU_REV(t620, 1, 0)),
211 GPU_MODEL(t720, 0x720),
212 GPU_MODEL(t760, 0x750,
213 GPU_REV(t760, 0, 0), GPU_REV(t760, 0, 1),
214 GPU_REV_EXT(t760, 0, 1, 0, _50rel0),
215 GPU_REV(t760, 0, 2), GPU_REV(t760, 0, 3)),
216 GPU_MODEL(t820, 0x820),
217 GPU_MODEL(t830, 0x830),
218 GPU_MODEL(t860, 0x860),
219 GPU_MODEL(t880, 0x880),
220
221 GPU_MODEL(g71, 0x6000,
222 GPU_REV_EXT(g71, 0, 0, 1, _05dev0)),
223 GPU_MODEL(g72, 0x6001),
224 GPU_MODEL(g51, 0x7000),
225 GPU_MODEL(g76, 0x7001),
226 GPU_MODEL(g52, 0x7002),
227 GPU_MODEL(g31, 0x7003,
228 GPU_REV(g31, 1, 0)),
229
230 GPU_MODEL(g57, 0x9001,
231 GPU_REV(g57, 0, 0)),
232
233 /* MediaTek MT8192 has a Mali-G57 with a different GPU ID from the
234 * standard. Arm's driver does not appear to handle this model.
235 * ChromeOS has a hack downstream for it. Treat it as equivalent to
236 * standard Mali-G57 for now.
237 */
238 GPU_MODEL(g57, 0x9003,
239 GPU_REV(g57, 0, 0)),
240 };
241
panfrost_gpu_init_features(struct panfrost_device * pfdev)242 static void panfrost_gpu_init_features(struct panfrost_device *pfdev)
243 {
244 u32 gpu_id, num_js, major, minor, status, rev;
245 const char *name = "unknown";
246 u64 hw_feat = 0;
247 u64 hw_issues = hw_issues_all;
248 const struct panfrost_model *model;
249 int i;
250
251 pfdev->features.l2_features = gpu_read(pfdev, GPU_L2_FEATURES);
252 pfdev->features.core_features = gpu_read(pfdev, GPU_CORE_FEATURES);
253 pfdev->features.tiler_features = gpu_read(pfdev, GPU_TILER_FEATURES);
254 pfdev->features.mem_features = gpu_read(pfdev, GPU_MEM_FEATURES);
255 pfdev->features.mmu_features = gpu_read(pfdev, GPU_MMU_FEATURES);
256 pfdev->features.thread_features = gpu_read(pfdev, GPU_THREAD_FEATURES);
257 pfdev->features.max_threads = gpu_read(pfdev, GPU_THREAD_MAX_THREADS);
258 pfdev->features.thread_max_workgroup_sz = gpu_read(pfdev, GPU_THREAD_MAX_WORKGROUP_SIZE);
259 pfdev->features.thread_max_barrier_sz = gpu_read(pfdev, GPU_THREAD_MAX_BARRIER_SIZE);
260 pfdev->features.coherency_features = gpu_read(pfdev, GPU_COHERENCY_FEATURES);
261 pfdev->features.afbc_features = gpu_read(pfdev, GPU_AFBC_FEATURES);
262 for (i = 0; i < 4; i++)
263 pfdev->features.texture_features[i] = gpu_read(pfdev, GPU_TEXTURE_FEATURES(i));
264
265 pfdev->features.as_present = gpu_read(pfdev, GPU_AS_PRESENT);
266
267 pfdev->features.js_present = gpu_read(pfdev, GPU_JS_PRESENT);
268 num_js = hweight32(pfdev->features.js_present);
269 for (i = 0; i < num_js; i++)
270 pfdev->features.js_features[i] = gpu_read(pfdev, GPU_JS_FEATURES(i));
271
272 pfdev->features.shader_present = gpu_read(pfdev, GPU_SHADER_PRESENT_LO);
273 pfdev->features.shader_present |= (u64)gpu_read(pfdev, GPU_SHADER_PRESENT_HI) << 32;
274
275 pfdev->features.tiler_present = gpu_read(pfdev, GPU_TILER_PRESENT_LO);
276 pfdev->features.tiler_present |= (u64)gpu_read(pfdev, GPU_TILER_PRESENT_HI) << 32;
277
278 pfdev->features.l2_present = gpu_read(pfdev, GPU_L2_PRESENT_LO);
279 pfdev->features.l2_present |= (u64)gpu_read(pfdev, GPU_L2_PRESENT_HI) << 32;
280 pfdev->features.nr_core_groups = hweight64(pfdev->features.l2_present);
281
282 pfdev->features.stack_present = gpu_read(pfdev, GPU_STACK_PRESENT_LO);
283 pfdev->features.stack_present |= (u64)gpu_read(pfdev, GPU_STACK_PRESENT_HI) << 32;
284
285 pfdev->features.thread_tls_alloc = gpu_read(pfdev, GPU_THREAD_TLS_ALLOC);
286
287 gpu_id = gpu_read(pfdev, GPU_ID);
288 pfdev->features.revision = gpu_id & 0xffff;
289 pfdev->features.id = gpu_id >> 16;
290
291 /* The T60x has an oddball ID value. Fix it up to the standard Midgard
292 * format so we (and userspace) don't have to special case it.
293 */
294 if (pfdev->features.id == 0x6956)
295 pfdev->features.id = 0x0600;
296
297 major = (pfdev->features.revision >> 12) & 0xf;
298 minor = (pfdev->features.revision >> 4) & 0xff;
299 status = pfdev->features.revision & 0xf;
300 rev = pfdev->features.revision;
301
302 gpu_id = pfdev->features.id;
303
304 for (model = gpu_models; model->name; model++) {
305 int best = -1;
306
307 if (!panfrost_model_eq(pfdev, model->id))
308 continue;
309
310 name = model->name;
311 hw_feat = model->features;
312 hw_issues |= model->issues;
313 for (i = 0; i < MAX_HW_REVS; i++) {
314 if (model->revs[i].revision == rev) {
315 best = i;
316 break;
317 } else if (model->revs[i].revision == (rev & ~0xf))
318 best = i;
319 }
320
321 if (best >= 0)
322 hw_issues |= model->revs[best].issues;
323
324 break;
325 }
326
327 bitmap_from_u64(pfdev->features.hw_features, hw_feat);
328 bitmap_from_u64(pfdev->features.hw_issues, hw_issues);
329
330 dev_info(pfdev->dev, "mali-%s id 0x%x major 0x%x minor 0x%x status 0x%x",
331 name, gpu_id, major, minor, status);
332 dev_info(pfdev->dev, "features: %64pb, issues: %64pb",
333 pfdev->features.hw_features,
334 pfdev->features.hw_issues);
335
336 dev_info(pfdev->dev, "Features: L2:0x%08x Shader:0x%08x Tiler:0x%08x Mem:0x%0x MMU:0x%08x AS:0x%x JS:0x%x",
337 pfdev->features.l2_features,
338 pfdev->features.core_features,
339 pfdev->features.tiler_features,
340 pfdev->features.mem_features,
341 pfdev->features.mmu_features,
342 pfdev->features.as_present,
343 pfdev->features.js_present);
344
345 dev_info(pfdev->dev, "shader_present=0x%0llx l2_present=0x%0llx",
346 pfdev->features.shader_present, pfdev->features.l2_present);
347 }
348
panfrost_cycle_counter_get(struct panfrost_device * pfdev)349 void panfrost_cycle_counter_get(struct panfrost_device *pfdev)
350 {
351 if (atomic_inc_not_zero(&pfdev->cycle_counter.use_count))
352 return;
353
354 spin_lock(&pfdev->cycle_counter.lock);
355 if (atomic_inc_return(&pfdev->cycle_counter.use_count) == 1)
356 gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_START);
357 spin_unlock(&pfdev->cycle_counter.lock);
358 }
359
panfrost_cycle_counter_put(struct panfrost_device * pfdev)360 void panfrost_cycle_counter_put(struct panfrost_device *pfdev)
361 {
362 if (atomic_add_unless(&pfdev->cycle_counter.use_count, -1, 1))
363 return;
364
365 spin_lock(&pfdev->cycle_counter.lock);
366 if (atomic_dec_return(&pfdev->cycle_counter.use_count) == 0)
367 gpu_write(pfdev, GPU_CMD, GPU_CMD_CYCLE_COUNT_STOP);
368 spin_unlock(&pfdev->cycle_counter.lock);
369 }
370
panfrost_cycle_counter_read(struct panfrost_device * pfdev)371 unsigned long long panfrost_cycle_counter_read(struct panfrost_device *pfdev)
372 {
373 u32 hi, lo;
374
375 do {
376 hi = gpu_read(pfdev, GPU_CYCLE_COUNT_HI);
377 lo = gpu_read(pfdev, GPU_CYCLE_COUNT_LO);
378 } while (hi != gpu_read(pfdev, GPU_CYCLE_COUNT_HI));
379
380 return ((u64)hi << 32) | lo;
381 }
382
panfrost_get_core_mask(struct panfrost_device * pfdev)383 static u64 panfrost_get_core_mask(struct panfrost_device *pfdev)
384 {
385 u64 core_mask;
386
387 if (pfdev->features.l2_present == 1)
388 return U64_MAX;
389
390 /*
391 * Only support one core group now.
392 * ~(l2_present - 1) unsets all bits in l2_present except
393 * the bottom bit. (l2_present - 2) has all the bits in
394 * the first core group set. AND them together to generate
395 * a mask of cores in the first core group.
396 */
397 core_mask = ~(pfdev->features.l2_present - 1) &
398 (pfdev->features.l2_present - 2);
399 dev_info_once(pfdev->dev, "using only 1st core group (%lu cores from %lu)\n",
400 hweight64(core_mask),
401 hweight64(pfdev->features.shader_present));
402
403 return core_mask;
404 }
405
panfrost_gpu_power_on(struct panfrost_device * pfdev)406 void panfrost_gpu_power_on(struct panfrost_device *pfdev)
407 {
408 int ret;
409 u32 val;
410 u64 core_mask;
411
412 panfrost_gpu_init_quirks(pfdev);
413 core_mask = panfrost_get_core_mask(pfdev);
414
415 gpu_write(pfdev, L2_PWRON_LO, pfdev->features.l2_present & core_mask);
416 ret = readl_relaxed_poll_timeout(pfdev->iomem + L2_READY_LO,
417 val, val == (pfdev->features.l2_present & core_mask),
418 10, 20000);
419 if (ret)
420 dev_err(pfdev->dev, "error powering up gpu L2");
421
422 gpu_write(pfdev, SHADER_PWRON_LO,
423 pfdev->features.shader_present & core_mask);
424 ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_READY_LO,
425 val, val == (pfdev->features.shader_present & core_mask),
426 10, 20000);
427 if (ret)
428 dev_err(pfdev->dev, "error powering up gpu shader");
429
430 gpu_write(pfdev, TILER_PWRON_LO, pfdev->features.tiler_present);
431 ret = readl_relaxed_poll_timeout(pfdev->iomem + TILER_READY_LO,
432 val, val == pfdev->features.tiler_present, 10, 1000);
433 if (ret)
434 dev_err(pfdev->dev, "error powering up gpu tiler");
435 }
436
panfrost_gpu_power_off(struct panfrost_device * pfdev)437 void panfrost_gpu_power_off(struct panfrost_device *pfdev)
438 {
439 int ret;
440 u32 val;
441
442 gpu_write(pfdev, SHADER_PWROFF_LO, pfdev->features.shader_present);
443 ret = readl_relaxed_poll_timeout(pfdev->iomem + SHADER_PWRTRANS_LO,
444 val, !val, 1, 2000);
445 if (ret)
446 dev_err(pfdev->dev, "shader power transition timeout");
447
448 gpu_write(pfdev, TILER_PWROFF_LO, pfdev->features.tiler_present);
449 ret = readl_relaxed_poll_timeout(pfdev->iomem + TILER_PWRTRANS_LO,
450 val, !val, 1, 2000);
451 if (ret)
452 dev_err(pfdev->dev, "tiler power transition timeout");
453
454 gpu_write(pfdev, L2_PWROFF_LO, pfdev->features.l2_present);
455 ret = readl_poll_timeout(pfdev->iomem + L2_PWRTRANS_LO,
456 val, !val, 0, 2000);
457 if (ret)
458 dev_err(pfdev->dev, "l2 power transition timeout");
459 }
460
panfrost_gpu_suspend_irq(struct panfrost_device * pfdev)461 void panfrost_gpu_suspend_irq(struct panfrost_device *pfdev)
462 {
463 set_bit(PANFROST_COMP_BIT_GPU, pfdev->is_suspended);
464
465 gpu_write(pfdev, GPU_INT_MASK, 0);
466 synchronize_irq(pfdev->gpu_irq);
467 }
468
panfrost_gpu_init(struct panfrost_device * pfdev)469 int panfrost_gpu_init(struct panfrost_device *pfdev)
470 {
471 int err;
472
473 err = panfrost_gpu_soft_reset(pfdev);
474 if (err)
475 return err;
476
477 panfrost_gpu_init_features(pfdev);
478
479 err = dma_set_mask_and_coherent(pfdev->dev,
480 DMA_BIT_MASK(FIELD_GET(0xff00, pfdev->features.mmu_features)));
481 if (err)
482 return err;
483
484 dma_set_max_seg_size(pfdev->dev, UINT_MAX);
485
486 pfdev->gpu_irq = platform_get_irq_byname(to_platform_device(pfdev->dev), "gpu");
487 if (pfdev->gpu_irq < 0)
488 return pfdev->gpu_irq;
489
490 err = devm_request_irq(pfdev->dev, pfdev->gpu_irq, panfrost_gpu_irq_handler,
491 IRQF_SHARED, KBUILD_MODNAME "-gpu", pfdev);
492 if (err) {
493 dev_err(pfdev->dev, "failed to request gpu irq");
494 return err;
495 }
496
497 panfrost_gpu_power_on(pfdev);
498
499 return 0;
500 }
501
panfrost_gpu_fini(struct panfrost_device * pfdev)502 void panfrost_gpu_fini(struct panfrost_device *pfdev)
503 {
504 panfrost_gpu_power_off(pfdev);
505 }
506
panfrost_gpu_get_latest_flush_id(struct panfrost_device * pfdev)507 u32 panfrost_gpu_get_latest_flush_id(struct panfrost_device *pfdev)
508 {
509 u32 flush_id;
510
511 if (panfrost_has_hw_feature(pfdev, HW_FEATURE_FLUSH_REDUCTION)) {
512 /* Flush reduction only makes sense when the GPU is kept powered on between jobs */
513 if (pm_runtime_get_if_in_use(pfdev->dev)) {
514 flush_id = gpu_read(pfdev, GPU_LATEST_FLUSH_ID);
515 pm_runtime_put(pfdev->dev);
516 return flush_id;
517 }
518 }
519
520 return 0;
521 }
522