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