1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  *
6  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
7  */
8 
9 #include <linux/ascii85.h>
10 #include <linux/interconnect.h>
11 #include <linux/qcom_scm.h>
12 #include <linux/kernel.h>
13 #include <linux/of_address.h>
14 #include <linux/pm_opp.h>
15 #include <linux/slab.h>
16 #include <linux/soc/qcom/mdt_loader.h>
17 #include <soc/qcom/ocmem.h>
18 #include "adreno_gpu.h"
19 #include "msm_gem.h"
20 #include "msm_mmu.h"
21 
22 static bool zap_available = true;
23 
24 static int zap_shader_load_mdt(struct msm_gpu *gpu, const char *fwname,
25 		u32 pasid)
26 {
27 	struct device *dev = &gpu->pdev->dev;
28 	const struct firmware *fw;
29 	const char *signed_fwname = NULL;
30 	struct device_node *np, *mem_np;
31 	struct resource r;
32 	phys_addr_t mem_phys;
33 	ssize_t mem_size;
34 	void *mem_region = NULL;
35 	int ret;
36 
37 	if (!IS_ENABLED(CONFIG_ARCH_QCOM)) {
38 		zap_available = false;
39 		return -EINVAL;
40 	}
41 
42 	np = of_get_child_by_name(dev->of_node, "zap-shader");
43 	if (!np) {
44 		zap_available = false;
45 		return -ENODEV;
46 	}
47 
48 	mem_np = of_parse_phandle(np, "memory-region", 0);
49 	of_node_put(np);
50 	if (!mem_np) {
51 		zap_available = false;
52 		return -EINVAL;
53 	}
54 
55 	ret = of_address_to_resource(mem_np, 0, &r);
56 	of_node_put(mem_np);
57 	if (ret)
58 		return ret;
59 
60 	mem_phys = r.start;
61 
62 	/*
63 	 * Check for a firmware-name property.  This is the new scheme
64 	 * to handle firmware that may be signed with device specific
65 	 * keys, allowing us to have a different zap fw path for different
66 	 * devices.
67 	 *
68 	 * If the firmware-name property is found, we bypass the
69 	 * adreno_request_fw() mechanism, because we don't need to handle
70 	 * the /lib/firmware/qcom/... vs /lib/firmware/... case.
71 	 *
72 	 * If the firmware-name property is not found, for backwards
73 	 * compatibility we fall back to the fwname from the gpulist
74 	 * table.
75 	 */
76 	of_property_read_string_index(np, "firmware-name", 0, &signed_fwname);
77 	if (signed_fwname) {
78 		fwname = signed_fwname;
79 		ret = request_firmware_direct(&fw, fwname, gpu->dev->dev);
80 		if (ret)
81 			fw = ERR_PTR(ret);
82 	} else if (fwname) {
83 		/* Request the MDT file from the default location: */
84 		fw = adreno_request_fw(to_adreno_gpu(gpu), fwname);
85 	} else {
86 		/*
87 		 * For new targets, we require the firmware-name property,
88 		 * if a zap-shader is required, rather than falling back
89 		 * to a firmware name specified in gpulist.
90 		 *
91 		 * Because the firmware is signed with a (potentially)
92 		 * device specific key, having the name come from gpulist
93 		 * was a bad idea, and is only provided for backwards
94 		 * compatibility for older targets.
95 		 */
96 		return -ENODEV;
97 	}
98 
99 	if (IS_ERR(fw)) {
100 		DRM_DEV_ERROR(dev, "Unable to load %s\n", fwname);
101 		return PTR_ERR(fw);
102 	}
103 
104 	/* Figure out how much memory we need */
105 	mem_size = qcom_mdt_get_size(fw);
106 	if (mem_size < 0) {
107 		ret = mem_size;
108 		goto out;
109 	}
110 
111 	if (mem_size > resource_size(&r)) {
112 		DRM_DEV_ERROR(dev,
113 			"memory region is too small to load the MDT\n");
114 		ret = -E2BIG;
115 		goto out;
116 	}
117 
118 	/* Allocate memory for the firmware image */
119 	mem_region = memremap(mem_phys, mem_size,  MEMREMAP_WC);
120 	if (!mem_region) {
121 		ret = -ENOMEM;
122 		goto out;
123 	}
124 
125 	/*
126 	 * Load the rest of the MDT
127 	 *
128 	 * Note that we could be dealing with two different paths, since
129 	 * with upstream linux-firmware it would be in a qcom/ subdir..
130 	 * adreno_request_fw() handles this, but qcom_mdt_load() does
131 	 * not.  But since we've already gotten through adreno_request_fw()
132 	 * we know which of the two cases it is:
133 	 */
134 	if (signed_fwname || (to_adreno_gpu(gpu)->fwloc == FW_LOCATION_LEGACY)) {
135 		ret = qcom_mdt_load(dev, fw, fwname, pasid,
136 				mem_region, mem_phys, mem_size, NULL);
137 	} else {
138 		char *newname;
139 
140 		newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
141 
142 		ret = qcom_mdt_load(dev, fw, newname, pasid,
143 				mem_region, mem_phys, mem_size, NULL);
144 		kfree(newname);
145 	}
146 	if (ret)
147 		goto out;
148 
149 	/* Send the image to the secure world */
150 	ret = qcom_scm_pas_auth_and_reset(pasid);
151 
152 	/*
153 	 * If the scm call returns -EOPNOTSUPP we assume that this target
154 	 * doesn't need/support the zap shader so quietly fail
155 	 */
156 	if (ret == -EOPNOTSUPP)
157 		zap_available = false;
158 	else if (ret)
159 		DRM_DEV_ERROR(dev, "Unable to authorize the image\n");
160 
161 out:
162 	if (mem_region)
163 		memunmap(mem_region);
164 
165 	release_firmware(fw);
166 
167 	return ret;
168 }
169 
170 int adreno_zap_shader_load(struct msm_gpu *gpu, u32 pasid)
171 {
172 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
173 	struct platform_device *pdev = gpu->pdev;
174 
175 	/* Short cut if we determine the zap shader isn't available/needed */
176 	if (!zap_available)
177 		return -ENODEV;
178 
179 	/* We need SCM to be able to load the firmware */
180 	if (!qcom_scm_is_available()) {
181 		DRM_DEV_ERROR(&pdev->dev, "SCM is not available\n");
182 		return -EPROBE_DEFER;
183 	}
184 
185 	return zap_shader_load_mdt(gpu, adreno_gpu->info->zapfw, pasid);
186 }
187 
188 struct msm_gem_address_space *
189 adreno_iommu_create_address_space(struct msm_gpu *gpu,
190 		struct platform_device *pdev)
191 {
192 	struct iommu_domain *iommu = iommu_domain_alloc(&platform_bus_type);
193 	struct msm_mmu *mmu = msm_iommu_new(&pdev->dev, iommu);
194 	struct msm_gem_address_space *aspace;
195 
196 	aspace = msm_gem_address_space_create(mmu, "gpu", SZ_16M,
197 		0xfffffff);
198 
199 	if (IS_ERR(aspace) && !IS_ERR(mmu))
200 		mmu->funcs->destroy(mmu);
201 
202 	return aspace;
203 }
204 
205 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
206 {
207 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
208 
209 	switch (param) {
210 	case MSM_PARAM_GPU_ID:
211 		*value = adreno_gpu->info->revn;
212 		return 0;
213 	case MSM_PARAM_GMEM_SIZE:
214 		*value = adreno_gpu->gmem;
215 		return 0;
216 	case MSM_PARAM_GMEM_BASE:
217 		*value = !adreno_is_a650(adreno_gpu) ? 0x100000 : 0;
218 		return 0;
219 	case MSM_PARAM_CHIP_ID:
220 		*value = adreno_gpu->rev.patchid |
221 				(adreno_gpu->rev.minor << 8) |
222 				(adreno_gpu->rev.major << 16) |
223 				(adreno_gpu->rev.core << 24);
224 		return 0;
225 	case MSM_PARAM_MAX_FREQ:
226 		*value = adreno_gpu->base.fast_rate;
227 		return 0;
228 	case MSM_PARAM_TIMESTAMP:
229 		if (adreno_gpu->funcs->get_timestamp) {
230 			int ret;
231 
232 			pm_runtime_get_sync(&gpu->pdev->dev);
233 			ret = adreno_gpu->funcs->get_timestamp(gpu, value);
234 			pm_runtime_put_autosuspend(&gpu->pdev->dev);
235 
236 			return ret;
237 		}
238 		return -EINVAL;
239 	case MSM_PARAM_NR_RINGS:
240 		*value = gpu->nr_rings;
241 		return 0;
242 	case MSM_PARAM_PP_PGTABLE:
243 		*value = 0;
244 		return 0;
245 	case MSM_PARAM_FAULTS:
246 		*value = gpu->global_faults;
247 		return 0;
248 	default:
249 		DBG("%s: invalid param: %u", gpu->name, param);
250 		return -EINVAL;
251 	}
252 }
253 
254 const struct firmware *
255 adreno_request_fw(struct adreno_gpu *adreno_gpu, const char *fwname)
256 {
257 	struct drm_device *drm = adreno_gpu->base.dev;
258 	const struct firmware *fw = NULL;
259 	char *newname;
260 	int ret;
261 
262 	newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
263 	if (!newname)
264 		return ERR_PTR(-ENOMEM);
265 
266 	/*
267 	 * Try first to load from qcom/$fwfile using a direct load (to avoid
268 	 * a potential timeout waiting for usermode helper)
269 	 */
270 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
271 	    (adreno_gpu->fwloc == FW_LOCATION_NEW)) {
272 
273 		ret = request_firmware_direct(&fw, newname, drm->dev);
274 		if (!ret) {
275 			DRM_DEV_INFO(drm->dev, "loaded %s from new location\n",
276 				newname);
277 			adreno_gpu->fwloc = FW_LOCATION_NEW;
278 			goto out;
279 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
280 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
281 				newname, ret);
282 			fw = ERR_PTR(ret);
283 			goto out;
284 		}
285 	}
286 
287 	/*
288 	 * Then try the legacy location without qcom/ prefix
289 	 */
290 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
291 	    (adreno_gpu->fwloc == FW_LOCATION_LEGACY)) {
292 
293 		ret = request_firmware_direct(&fw, fwname, drm->dev);
294 		if (!ret) {
295 			DRM_DEV_INFO(drm->dev, "loaded %s from legacy location\n",
296 				newname);
297 			adreno_gpu->fwloc = FW_LOCATION_LEGACY;
298 			goto out;
299 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
300 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
301 				fwname, ret);
302 			fw = ERR_PTR(ret);
303 			goto out;
304 		}
305 	}
306 
307 	/*
308 	 * Finally fall back to request_firmware() for cases where the
309 	 * usermode helper is needed (I think mainly android)
310 	 */
311 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
312 	    (adreno_gpu->fwloc == FW_LOCATION_HELPER)) {
313 
314 		ret = request_firmware(&fw, newname, drm->dev);
315 		if (!ret) {
316 			DRM_DEV_INFO(drm->dev, "loaded %s with helper\n",
317 				newname);
318 			adreno_gpu->fwloc = FW_LOCATION_HELPER;
319 			goto out;
320 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
321 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
322 				newname, ret);
323 			fw = ERR_PTR(ret);
324 			goto out;
325 		}
326 	}
327 
328 	DRM_DEV_ERROR(drm->dev, "failed to load %s\n", fwname);
329 	fw = ERR_PTR(-ENOENT);
330 out:
331 	kfree(newname);
332 	return fw;
333 }
334 
335 int adreno_load_fw(struct adreno_gpu *adreno_gpu)
336 {
337 	int i;
338 
339 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++) {
340 		const struct firmware *fw;
341 
342 		if (!adreno_gpu->info->fw[i])
343 			continue;
344 
345 		/* Skip if the firmware has already been loaded */
346 		if (adreno_gpu->fw[i])
347 			continue;
348 
349 		fw = adreno_request_fw(adreno_gpu, adreno_gpu->info->fw[i]);
350 		if (IS_ERR(fw))
351 			return PTR_ERR(fw);
352 
353 		adreno_gpu->fw[i] = fw;
354 	}
355 
356 	return 0;
357 }
358 
359 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu,
360 		const struct firmware *fw, u64 *iova)
361 {
362 	struct drm_gem_object *bo;
363 	void *ptr;
364 
365 	ptr = msm_gem_kernel_new_locked(gpu->dev, fw->size - 4,
366 		MSM_BO_UNCACHED | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
367 
368 	if (IS_ERR(ptr))
369 		return ERR_CAST(ptr);
370 
371 	memcpy(ptr, &fw->data[4], fw->size - 4);
372 
373 	msm_gem_put_vaddr(bo);
374 
375 	return bo;
376 }
377 
378 int adreno_hw_init(struct msm_gpu *gpu)
379 {
380 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
381 	int ret, i;
382 
383 	DBG("%s", gpu->name);
384 
385 	ret = adreno_load_fw(adreno_gpu);
386 	if (ret)
387 		return ret;
388 
389 	for (i = 0; i < gpu->nr_rings; i++) {
390 		struct msm_ringbuffer *ring = gpu->rb[i];
391 
392 		if (!ring)
393 			continue;
394 
395 		ring->cur = ring->start;
396 		ring->next = ring->start;
397 
398 		/* reset completed fence seqno: */
399 		ring->memptrs->fence = ring->seqno;
400 		ring->memptrs->rptr = 0;
401 	}
402 
403 	/*
404 	 * Setup REG_CP_RB_CNTL.  The same value is used across targets (with
405 	 * the excpetion of A430 that disables the RPTR shadow) - the cacluation
406 	 * for the ringbuffer size and block size is moved to msm_gpu.h for the
407 	 * pre-processor to deal with and the A430 variant is ORed in here
408 	 */
409 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL,
410 		MSM_GPU_RB_CNTL_DEFAULT |
411 		(adreno_is_a430(adreno_gpu) ? AXXX_CP_RB_CNTL_NO_UPDATE : 0));
412 
413 	/* Setup ringbuffer address - use ringbuffer[0] for GPU init */
414 	adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_BASE,
415 		REG_ADRENO_CP_RB_BASE_HI, gpu->rb[0]->iova);
416 
417 	if (!adreno_is_a430(adreno_gpu)) {
418 		adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR,
419 			REG_ADRENO_CP_RB_RPTR_ADDR_HI,
420 			rbmemptr(gpu->rb[0], rptr));
421 	}
422 
423 	return 0;
424 }
425 
426 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
427 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu,
428 		struct msm_ringbuffer *ring)
429 {
430 	if (adreno_is_a430(adreno_gpu))
431 		return ring->memptrs->rptr = adreno_gpu_read(
432 			adreno_gpu, REG_ADRENO_CP_RB_RPTR);
433 	else
434 		return ring->memptrs->rptr;
435 }
436 
437 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu)
438 {
439 	return gpu->rb[0];
440 }
441 
442 void adreno_recover(struct msm_gpu *gpu)
443 {
444 	struct drm_device *dev = gpu->dev;
445 	int ret;
446 
447 	// XXX pm-runtime??  we *need* the device to be off after this
448 	// so maybe continuing to call ->pm_suspend/resume() is better?
449 
450 	gpu->funcs->pm_suspend(gpu);
451 	gpu->funcs->pm_resume(gpu);
452 
453 	ret = msm_gpu_hw_init(gpu);
454 	if (ret) {
455 		DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
456 		/* hmm, oh well? */
457 	}
458 }
459 
460 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
461 		struct msm_file_private *ctx)
462 {
463 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
464 	struct msm_drm_private *priv = gpu->dev->dev_private;
465 	struct msm_ringbuffer *ring = submit->ring;
466 	unsigned i;
467 
468 	for (i = 0; i < submit->nr_cmds; i++) {
469 		switch (submit->cmd[i].type) {
470 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
471 			/* ignore IB-targets */
472 			break;
473 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
474 			/* ignore if there has not been a ctx switch: */
475 			if (priv->lastctx == ctx)
476 				break;
477 			/* fall-thru */
478 		case MSM_SUBMIT_CMD_BUF:
479 			OUT_PKT3(ring, adreno_is_a4xx(adreno_gpu) ?
480 				CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
481 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
482 			OUT_RING(ring, submit->cmd[i].size);
483 			OUT_PKT2(ring);
484 			break;
485 		}
486 	}
487 
488 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
489 	OUT_RING(ring, submit->seqno);
490 
491 	if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) {
492 		/* Flush HLSQ lazy updates to make sure there is nothing
493 		 * pending for indirect loads after the timestamp has
494 		 * passed:
495 		 */
496 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
497 		OUT_RING(ring, HLSQ_FLUSH);
498 	}
499 
500 	/* wait for idle before cache flush/interrupt */
501 	OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
502 	OUT_RING(ring, 0x00000000);
503 
504 	if (!adreno_is_a2xx(adreno_gpu)) {
505 		/* BIT(31) of CACHE_FLUSH_TS triggers CACHE_FLUSH_TS IRQ from GPU */
506 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
507 		OUT_RING(ring, CACHE_FLUSH_TS | BIT(31));
508 		OUT_RING(ring, rbmemptr(ring, fence));
509 		OUT_RING(ring, submit->seqno);
510 	} else {
511 		/* BIT(31) means something else on a2xx */
512 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
513 		OUT_RING(ring, CACHE_FLUSH_TS);
514 		OUT_RING(ring, rbmemptr(ring, fence));
515 		OUT_RING(ring, submit->seqno);
516 		OUT_PKT3(ring, CP_INTERRUPT, 1);
517 		OUT_RING(ring, 0x80000000);
518 	}
519 
520 #if 0
521 	if (adreno_is_a3xx(adreno_gpu)) {
522 		/* Dummy set-constant to trigger context rollover */
523 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
524 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
525 		OUT_RING(ring, 0x00000000);
526 	}
527 #endif
528 
529 	gpu->funcs->flush(gpu, ring);
530 }
531 
532 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
533 {
534 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
535 	uint32_t wptr;
536 
537 	/* Copy the shadow to the actual register */
538 	ring->cur = ring->next;
539 
540 	/*
541 	 * Mask wptr value that we calculate to fit in the HW range. This is
542 	 * to account for the possibility that the last command fit exactly into
543 	 * the ringbuffer and rb->next hasn't wrapped to zero yet
544 	 */
545 	wptr = get_wptr(ring);
546 
547 	/* ensure writes to ringbuffer have hit system memory: */
548 	mb();
549 
550 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr);
551 }
552 
553 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
554 {
555 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
556 	uint32_t wptr = get_wptr(ring);
557 
558 	/* wait for CP to drain ringbuffer: */
559 	if (!spin_until(get_rptr(adreno_gpu, ring) == wptr))
560 		return true;
561 
562 	/* TODO maybe we need to reset GPU here to recover from hang? */
563 	DRM_ERROR("%s: timeout waiting to drain ringbuffer %d rptr/wptr = %X/%X\n",
564 		gpu->name, ring->id, get_rptr(adreno_gpu, ring), wptr);
565 
566 	return false;
567 }
568 
569 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state)
570 {
571 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
572 	int i, count = 0;
573 
574 	kref_init(&state->ref);
575 
576 	ktime_get_real_ts64(&state->time);
577 
578 	for (i = 0; i < gpu->nr_rings; i++) {
579 		int size = 0, j;
580 
581 		state->ring[i].fence = gpu->rb[i]->memptrs->fence;
582 		state->ring[i].iova = gpu->rb[i]->iova;
583 		state->ring[i].seqno = gpu->rb[i]->seqno;
584 		state->ring[i].rptr = get_rptr(adreno_gpu, gpu->rb[i]);
585 		state->ring[i].wptr = get_wptr(gpu->rb[i]);
586 
587 		/* Copy at least 'wptr' dwords of the data */
588 		size = state->ring[i].wptr;
589 
590 		/* After wptr find the last non zero dword to save space */
591 		for (j = state->ring[i].wptr; j < MSM_GPU_RINGBUFFER_SZ >> 2; j++)
592 			if (gpu->rb[i]->start[j])
593 				size = j + 1;
594 
595 		if (size) {
596 			state->ring[i].data = kvmalloc(size << 2, GFP_KERNEL);
597 			if (state->ring[i].data) {
598 				memcpy(state->ring[i].data, gpu->rb[i]->start, size << 2);
599 				state->ring[i].data_size = size << 2;
600 			}
601 		}
602 	}
603 
604 	/* Some targets prefer to collect their own registers */
605 	if (!adreno_gpu->registers)
606 		return 0;
607 
608 	/* Count the number of registers */
609 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2)
610 		count += adreno_gpu->registers[i + 1] -
611 			adreno_gpu->registers[i] + 1;
612 
613 	state->registers = kcalloc(count * 2, sizeof(u32), GFP_KERNEL);
614 	if (state->registers) {
615 		int pos = 0;
616 
617 		for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
618 			u32 start = adreno_gpu->registers[i];
619 			u32 end   = adreno_gpu->registers[i + 1];
620 			u32 addr;
621 
622 			for (addr = start; addr <= end; addr++) {
623 				state->registers[pos++] = addr;
624 				state->registers[pos++] = gpu_read(gpu, addr);
625 			}
626 		}
627 
628 		state->nr_registers = count;
629 	}
630 
631 	return 0;
632 }
633 
634 void adreno_gpu_state_destroy(struct msm_gpu_state *state)
635 {
636 	int i;
637 
638 	for (i = 0; i < ARRAY_SIZE(state->ring); i++)
639 		kvfree(state->ring[i].data);
640 
641 	for (i = 0; state->bos && i < state->nr_bos; i++)
642 		kvfree(state->bos[i].data);
643 
644 	kfree(state->bos);
645 	kfree(state->comm);
646 	kfree(state->cmd);
647 	kfree(state->registers);
648 }
649 
650 static void adreno_gpu_state_kref_destroy(struct kref *kref)
651 {
652 	struct msm_gpu_state *state = container_of(kref,
653 		struct msm_gpu_state, ref);
654 
655 	adreno_gpu_state_destroy(state);
656 	kfree(state);
657 }
658 
659 int adreno_gpu_state_put(struct msm_gpu_state *state)
660 {
661 	if (IS_ERR_OR_NULL(state))
662 		return 1;
663 
664 	return kref_put(&state->ref, adreno_gpu_state_kref_destroy);
665 }
666 
667 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
668 
669 static char *adreno_gpu_ascii85_encode(u32 *src, size_t len)
670 {
671 	void *buf;
672 	size_t buf_itr = 0, buffer_size;
673 	char out[ASCII85_BUFSZ];
674 	long l;
675 	int i;
676 
677 	if (!src || !len)
678 		return NULL;
679 
680 	l = ascii85_encode_len(len);
681 
682 	/*
683 	 * Ascii85 outputs either a 5 byte string or a 1 byte string. So we
684 	 * account for the worst case of 5 bytes per dword plus the 1 for '\0'
685 	 */
686 	buffer_size = (l * 5) + 1;
687 
688 	buf = kvmalloc(buffer_size, GFP_KERNEL);
689 	if (!buf)
690 		return NULL;
691 
692 	for (i = 0; i < l; i++)
693 		buf_itr += scnprintf(buf + buf_itr, buffer_size - buf_itr, "%s",
694 				ascii85_encode(src[i], out));
695 
696 	return buf;
697 }
698 
699 /* len is expected to be in bytes */
700 static void adreno_show_object(struct drm_printer *p, void **ptr, int len,
701 		bool *encoded)
702 {
703 	if (!*ptr || !len)
704 		return;
705 
706 	if (!*encoded) {
707 		long datalen, i;
708 		u32 *buf = *ptr;
709 
710 		/*
711 		 * Only dump the non-zero part of the buffer - rarely will
712 		 * any data completely fill the entire allocated size of
713 		 * the buffer.
714 		 */
715 		for (datalen = 0, i = 0; i < len >> 2; i++)
716 			if (buf[i])
717 				datalen = ((i + 1) << 2);
718 
719 		/*
720 		 * If we reach here, then the originally captured binary buffer
721 		 * will be replaced with the ascii85 encoded string
722 		 */
723 		*ptr = adreno_gpu_ascii85_encode(buf, datalen);
724 
725 		kvfree(buf);
726 
727 		*encoded = true;
728 	}
729 
730 	if (!*ptr)
731 		return;
732 
733 	drm_puts(p, "    data: !!ascii85 |\n");
734 	drm_puts(p, "     ");
735 
736 	drm_puts(p, *ptr);
737 
738 	drm_puts(p, "\n");
739 }
740 
741 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
742 		struct drm_printer *p)
743 {
744 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
745 	int i;
746 
747 	if (IS_ERR_OR_NULL(state))
748 		return;
749 
750 	drm_printf(p, "revision: %d (%d.%d.%d.%d)\n",
751 			adreno_gpu->info->revn, adreno_gpu->rev.core,
752 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
753 			adreno_gpu->rev.patchid);
754 
755 	drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
756 
757 	drm_puts(p, "ringbuffer:\n");
758 
759 	for (i = 0; i < gpu->nr_rings; i++) {
760 		drm_printf(p, "  - id: %d\n", i);
761 		drm_printf(p, "    iova: 0x%016llx\n", state->ring[i].iova);
762 		drm_printf(p, "    last-fence: %d\n", state->ring[i].seqno);
763 		drm_printf(p, "    retired-fence: %d\n", state->ring[i].fence);
764 		drm_printf(p, "    rptr: %d\n", state->ring[i].rptr);
765 		drm_printf(p, "    wptr: %d\n", state->ring[i].wptr);
766 		drm_printf(p, "    size: %d\n", MSM_GPU_RINGBUFFER_SZ);
767 
768 		adreno_show_object(p, &state->ring[i].data,
769 			state->ring[i].data_size, &state->ring[i].encoded);
770 	}
771 
772 	if (state->bos) {
773 		drm_puts(p, "bos:\n");
774 
775 		for (i = 0; i < state->nr_bos; i++) {
776 			drm_printf(p, "  - iova: 0x%016llx\n",
777 				state->bos[i].iova);
778 			drm_printf(p, "    size: %zd\n", state->bos[i].size);
779 
780 			adreno_show_object(p, &state->bos[i].data,
781 				state->bos[i].size, &state->bos[i].encoded);
782 		}
783 	}
784 
785 	if (state->nr_registers) {
786 		drm_puts(p, "registers:\n");
787 
788 		for (i = 0; i < state->nr_registers; i++) {
789 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
790 				state->registers[i * 2] << 2,
791 				state->registers[(i * 2) + 1]);
792 		}
793 	}
794 }
795 #endif
796 
797 /* Dump common gpu status and scratch registers on any hang, to make
798  * the hangcheck logs more useful.  The scratch registers seem always
799  * safe to read when GPU has hung (unlike some other regs, depending
800  * on how the GPU hung), and they are useful to match up to cmdstream
801  * dumps when debugging hangs:
802  */
803 void adreno_dump_info(struct msm_gpu *gpu)
804 {
805 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
806 	int i;
807 
808 	printk("revision: %d (%d.%d.%d.%d)\n",
809 			adreno_gpu->info->revn, adreno_gpu->rev.core,
810 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
811 			adreno_gpu->rev.patchid);
812 
813 	for (i = 0; i < gpu->nr_rings; i++) {
814 		struct msm_ringbuffer *ring = gpu->rb[i];
815 
816 		printk("rb %d: fence:    %d/%d\n", i,
817 			ring->memptrs->fence,
818 			ring->seqno);
819 
820 		printk("rptr:     %d\n", get_rptr(adreno_gpu, ring));
821 		printk("rb wptr:  %d\n", get_wptr(ring));
822 	}
823 }
824 
825 /* would be nice to not have to duplicate the _show() stuff with printk(): */
826 void adreno_dump(struct msm_gpu *gpu)
827 {
828 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
829 	int i;
830 
831 	if (!adreno_gpu->registers)
832 		return;
833 
834 	/* dump these out in a form that can be parsed by demsm: */
835 	printk("IO:region %s 00000000 00020000\n", gpu->name);
836 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
837 		uint32_t start = adreno_gpu->registers[i];
838 		uint32_t end   = adreno_gpu->registers[i+1];
839 		uint32_t addr;
840 
841 		for (addr = start; addr <= end; addr++) {
842 			uint32_t val = gpu_read(gpu, addr);
843 			printk("IO:R %08x %08x\n", addr<<2, val);
844 		}
845 	}
846 }
847 
848 static uint32_t ring_freewords(struct msm_ringbuffer *ring)
849 {
850 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(ring->gpu);
851 	uint32_t size = MSM_GPU_RINGBUFFER_SZ >> 2;
852 	/* Use ring->next to calculate free size */
853 	uint32_t wptr = ring->next - ring->start;
854 	uint32_t rptr = get_rptr(adreno_gpu, ring);
855 	return (rptr + (size - 1) - wptr) % size;
856 }
857 
858 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords)
859 {
860 	if (spin_until(ring_freewords(ring) >= ndwords))
861 		DRM_DEV_ERROR(ring->gpu->dev->dev,
862 			"timeout waiting for space in ringbuffer %d\n",
863 			ring->id);
864 }
865 
866 /* Get legacy powerlevels from qcom,gpu-pwrlevels and populate the opp table */
867 static int adreno_get_legacy_pwrlevels(struct device *dev)
868 {
869 	struct device_node *child, *node;
870 	int ret;
871 
872 	node = of_get_compatible_child(dev->of_node, "qcom,gpu-pwrlevels");
873 	if (!node) {
874 		DRM_DEV_DEBUG(dev, "Could not find the GPU powerlevels\n");
875 		return -ENXIO;
876 	}
877 
878 	for_each_child_of_node(node, child) {
879 		unsigned int val;
880 
881 		ret = of_property_read_u32(child, "qcom,gpu-freq", &val);
882 		if (ret)
883 			continue;
884 
885 		/*
886 		 * Skip the intentionally bogus clock value found at the bottom
887 		 * of most legacy frequency tables
888 		 */
889 		if (val != 27000000)
890 			dev_pm_opp_add(dev, val, 0);
891 	}
892 
893 	of_node_put(node);
894 
895 	return 0;
896 }
897 
898 static int adreno_get_pwrlevels(struct device *dev,
899 		struct msm_gpu *gpu)
900 {
901 	unsigned long freq = ULONG_MAX;
902 	struct dev_pm_opp *opp;
903 	int ret;
904 
905 	gpu->fast_rate = 0;
906 
907 	/* You down with OPP? */
908 	if (!of_find_property(dev->of_node, "operating-points-v2", NULL))
909 		ret = adreno_get_legacy_pwrlevels(dev);
910 	else {
911 		ret = dev_pm_opp_of_add_table(dev);
912 		if (ret)
913 			DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
914 	}
915 
916 	if (!ret) {
917 		/* Find the fastest defined rate */
918 		opp = dev_pm_opp_find_freq_floor(dev, &freq);
919 		if (!IS_ERR(opp)) {
920 			gpu->fast_rate = freq;
921 			dev_pm_opp_put(opp);
922 		}
923 	}
924 
925 	if (!gpu->fast_rate) {
926 		dev_warn(dev,
927 			"Could not find a clock rate. Using a reasonable default\n");
928 		/* Pick a suitably safe clock speed for any target */
929 		gpu->fast_rate = 200000000;
930 	}
931 
932 	DBG("fast_rate=%u, slow_rate=27000000", gpu->fast_rate);
933 
934 	/* Check for an interconnect path for the bus */
935 	gpu->icc_path = of_icc_get(dev, "gfx-mem");
936 	if (!gpu->icc_path) {
937 		/*
938 		 * Keep compatbility with device trees that don't have an
939 		 * interconnect-names property.
940 		 */
941 		gpu->icc_path = of_icc_get(dev, NULL);
942 	}
943 	if (IS_ERR(gpu->icc_path))
944 		gpu->icc_path = NULL;
945 
946 	gpu->ocmem_icc_path = of_icc_get(dev, "ocmem");
947 	if (IS_ERR(gpu->ocmem_icc_path))
948 		gpu->ocmem_icc_path = NULL;
949 
950 	return 0;
951 }
952 
953 int adreno_gpu_ocmem_init(struct device *dev, struct adreno_gpu *adreno_gpu,
954 			  struct adreno_ocmem *adreno_ocmem)
955 {
956 	struct ocmem_buf *ocmem_hdl;
957 	struct ocmem *ocmem;
958 
959 	ocmem = of_get_ocmem(dev);
960 	if (IS_ERR(ocmem)) {
961 		if (PTR_ERR(ocmem) == -ENODEV) {
962 			/*
963 			 * Return success since either the ocmem property was
964 			 * not specified in device tree, or ocmem support is
965 			 * not compiled into the kernel.
966 			 */
967 			return 0;
968 		}
969 
970 		return PTR_ERR(ocmem);
971 	}
972 
973 	ocmem_hdl = ocmem_allocate(ocmem, OCMEM_GRAPHICS, adreno_gpu->gmem);
974 	if (IS_ERR(ocmem_hdl))
975 		return PTR_ERR(ocmem_hdl);
976 
977 	adreno_ocmem->ocmem = ocmem;
978 	adreno_ocmem->base = ocmem_hdl->addr;
979 	adreno_ocmem->hdl = ocmem_hdl;
980 	adreno_gpu->gmem = ocmem_hdl->len;
981 
982 	return 0;
983 }
984 
985 void adreno_gpu_ocmem_cleanup(struct adreno_ocmem *adreno_ocmem)
986 {
987 	if (adreno_ocmem && adreno_ocmem->base)
988 		ocmem_free(adreno_ocmem->ocmem, OCMEM_GRAPHICS,
989 			   adreno_ocmem->hdl);
990 }
991 
992 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
993 		struct adreno_gpu *adreno_gpu,
994 		const struct adreno_gpu_funcs *funcs, int nr_rings)
995 {
996 	struct adreno_platform_config *config = pdev->dev.platform_data;
997 	struct msm_gpu_config adreno_gpu_config  = { 0 };
998 	struct msm_gpu *gpu = &adreno_gpu->base;
999 
1000 	adreno_gpu->funcs = funcs;
1001 	adreno_gpu->info = adreno_info(config->rev);
1002 	adreno_gpu->gmem = adreno_gpu->info->gmem;
1003 	adreno_gpu->revn = adreno_gpu->info->revn;
1004 	adreno_gpu->rev = config->rev;
1005 
1006 	adreno_gpu_config.ioname = "kgsl_3d0_reg_memory";
1007 
1008 	adreno_gpu_config.nr_rings = nr_rings;
1009 
1010 	adreno_get_pwrlevels(&pdev->dev, gpu);
1011 
1012 	pm_runtime_set_autosuspend_delay(&pdev->dev,
1013 		adreno_gpu->info->inactive_period);
1014 	pm_runtime_use_autosuspend(&pdev->dev);
1015 	pm_runtime_enable(&pdev->dev);
1016 
1017 	return msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
1018 			adreno_gpu->info->name, &adreno_gpu_config);
1019 }
1020 
1021 void adreno_gpu_cleanup(struct adreno_gpu *adreno_gpu)
1022 {
1023 	struct msm_gpu *gpu = &adreno_gpu->base;
1024 	unsigned int i;
1025 
1026 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++)
1027 		release_firmware(adreno_gpu->fw[i]);
1028 
1029 	icc_put(gpu->icc_path);
1030 	icc_put(gpu->ocmem_icc_path);
1031 
1032 	msm_gpu_cleanup(&adreno_gpu->base);
1033 }
1034