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 	struct device_node *np, *mem_np;
30 	struct resource r;
31 	phys_addr_t mem_phys;
32 	ssize_t mem_size;
33 	void *mem_region = NULL;
34 	int ret;
35 
36 	if (!IS_ENABLED(CONFIG_ARCH_QCOM)) {
37 		zap_available = false;
38 		return -EINVAL;
39 	}
40 
41 	np = of_get_child_by_name(dev->of_node, "zap-shader");
42 	if (!np) {
43 		zap_available = false;
44 		return -ENODEV;
45 	}
46 
47 	mem_np = of_parse_phandle(np, "memory-region", 0);
48 	of_node_put(np);
49 	if (!mem_np) {
50 		zap_available = false;
51 		return -EINVAL;
52 	}
53 
54 	ret = of_address_to_resource(mem_np, 0, &r);
55 	of_node_put(mem_np);
56 	if (ret)
57 		return ret;
58 
59 	mem_phys = r.start;
60 
61 	/* Request the MDT file for the firmware */
62 	fw = adreno_request_fw(to_adreno_gpu(gpu), fwname);
63 	if (IS_ERR(fw)) {
64 		DRM_DEV_ERROR(dev, "Unable to load %s\n", fwname);
65 		return PTR_ERR(fw);
66 	}
67 
68 	/* Figure out how much memory we need */
69 	mem_size = qcom_mdt_get_size(fw);
70 	if (mem_size < 0) {
71 		ret = mem_size;
72 		goto out;
73 	}
74 
75 	if (mem_size > resource_size(&r)) {
76 		DRM_DEV_ERROR(dev,
77 			"memory region is too small to load the MDT\n");
78 		ret = -E2BIG;
79 		goto out;
80 	}
81 
82 	/* Allocate memory for the firmware image */
83 	mem_region = memremap(mem_phys, mem_size,  MEMREMAP_WC);
84 	if (!mem_region) {
85 		ret = -ENOMEM;
86 		goto out;
87 	}
88 
89 	/*
90 	 * Load the rest of the MDT
91 	 *
92 	 * Note that we could be dealing with two different paths, since
93 	 * with upstream linux-firmware it would be in a qcom/ subdir..
94 	 * adreno_request_fw() handles this, but qcom_mdt_load() does
95 	 * not.  But since we've already gotten through adreno_request_fw()
96 	 * we know which of the two cases it is:
97 	 */
98 	if (to_adreno_gpu(gpu)->fwloc == FW_LOCATION_LEGACY) {
99 		ret = qcom_mdt_load(dev, fw, fwname, pasid,
100 				mem_region, mem_phys, mem_size, NULL);
101 	} else {
102 		char *newname;
103 
104 		newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
105 
106 		ret = qcom_mdt_load(dev, fw, newname, pasid,
107 				mem_region, mem_phys, mem_size, NULL);
108 		kfree(newname);
109 	}
110 	if (ret)
111 		goto out;
112 
113 	/* Send the image to the secure world */
114 	ret = qcom_scm_pas_auth_and_reset(pasid);
115 
116 	/*
117 	 * If the scm call returns -EOPNOTSUPP we assume that this target
118 	 * doesn't need/support the zap shader so quietly fail
119 	 */
120 	if (ret == -EOPNOTSUPP)
121 		zap_available = false;
122 	else if (ret)
123 		DRM_DEV_ERROR(dev, "Unable to authorize the image\n");
124 
125 out:
126 	if (mem_region)
127 		memunmap(mem_region);
128 
129 	release_firmware(fw);
130 
131 	return ret;
132 }
133 
134 int adreno_zap_shader_load(struct msm_gpu *gpu, u32 pasid)
135 {
136 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
137 	struct platform_device *pdev = gpu->pdev;
138 
139 	/* Short cut if we determine the zap shader isn't available/needed */
140 	if (!zap_available)
141 		return -ENODEV;
142 
143 	/* We need SCM to be able to load the firmware */
144 	if (!qcom_scm_is_available()) {
145 		DRM_DEV_ERROR(&pdev->dev, "SCM is not available\n");
146 		return -EPROBE_DEFER;
147 	}
148 
149 	/* Each GPU has a target specific zap shader firmware name to use */
150 	if (!adreno_gpu->info->zapfw) {
151 		zap_available = false;
152 		DRM_DEV_ERROR(&pdev->dev,
153 			"Zap shader firmware file not specified for this target\n");
154 		return -ENODEV;
155 	}
156 
157 	return zap_shader_load_mdt(gpu, adreno_gpu->info->zapfw, pasid);
158 }
159 
160 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
161 {
162 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
163 
164 	switch (param) {
165 	case MSM_PARAM_GPU_ID:
166 		*value = adreno_gpu->info->revn;
167 		return 0;
168 	case MSM_PARAM_GMEM_SIZE:
169 		*value = adreno_gpu->gmem;
170 		return 0;
171 	case MSM_PARAM_GMEM_BASE:
172 		*value = 0x100000;
173 		return 0;
174 	case MSM_PARAM_CHIP_ID:
175 		*value = adreno_gpu->rev.patchid |
176 				(adreno_gpu->rev.minor << 8) |
177 				(adreno_gpu->rev.major << 16) |
178 				(adreno_gpu->rev.core << 24);
179 		return 0;
180 	case MSM_PARAM_MAX_FREQ:
181 		*value = adreno_gpu->base.fast_rate;
182 		return 0;
183 	case MSM_PARAM_TIMESTAMP:
184 		if (adreno_gpu->funcs->get_timestamp) {
185 			int ret;
186 
187 			pm_runtime_get_sync(&gpu->pdev->dev);
188 			ret = adreno_gpu->funcs->get_timestamp(gpu, value);
189 			pm_runtime_put_autosuspend(&gpu->pdev->dev);
190 
191 			return ret;
192 		}
193 		return -EINVAL;
194 	case MSM_PARAM_NR_RINGS:
195 		*value = gpu->nr_rings;
196 		return 0;
197 	case MSM_PARAM_PP_PGTABLE:
198 		*value = 0;
199 		return 0;
200 	case MSM_PARAM_FAULTS:
201 		*value = gpu->global_faults;
202 		return 0;
203 	default:
204 		DBG("%s: invalid param: %u", gpu->name, param);
205 		return -EINVAL;
206 	}
207 }
208 
209 const struct firmware *
210 adreno_request_fw(struct adreno_gpu *adreno_gpu, const char *fwname)
211 {
212 	struct drm_device *drm = adreno_gpu->base.dev;
213 	const struct firmware *fw = NULL;
214 	char *newname;
215 	int ret;
216 
217 	newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
218 	if (!newname)
219 		return ERR_PTR(-ENOMEM);
220 
221 	/*
222 	 * Try first to load from qcom/$fwfile using a direct load (to avoid
223 	 * a potential timeout waiting for usermode helper)
224 	 */
225 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
226 	    (adreno_gpu->fwloc == FW_LOCATION_NEW)) {
227 
228 		ret = request_firmware_direct(&fw, newname, drm->dev);
229 		if (!ret) {
230 			DRM_DEV_INFO(drm->dev, "loaded %s from new location\n",
231 				newname);
232 			adreno_gpu->fwloc = FW_LOCATION_NEW;
233 			goto out;
234 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
235 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
236 				newname, ret);
237 			fw = ERR_PTR(ret);
238 			goto out;
239 		}
240 	}
241 
242 	/*
243 	 * Then try the legacy location without qcom/ prefix
244 	 */
245 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
246 	    (adreno_gpu->fwloc == FW_LOCATION_LEGACY)) {
247 
248 		ret = request_firmware_direct(&fw, fwname, drm->dev);
249 		if (!ret) {
250 			DRM_DEV_INFO(drm->dev, "loaded %s from legacy location\n",
251 				newname);
252 			adreno_gpu->fwloc = FW_LOCATION_LEGACY;
253 			goto out;
254 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
255 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
256 				fwname, ret);
257 			fw = ERR_PTR(ret);
258 			goto out;
259 		}
260 	}
261 
262 	/*
263 	 * Finally fall back to request_firmware() for cases where the
264 	 * usermode helper is needed (I think mainly android)
265 	 */
266 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
267 	    (adreno_gpu->fwloc == FW_LOCATION_HELPER)) {
268 
269 		ret = request_firmware(&fw, newname, drm->dev);
270 		if (!ret) {
271 			DRM_DEV_INFO(drm->dev, "loaded %s with helper\n",
272 				newname);
273 			adreno_gpu->fwloc = FW_LOCATION_HELPER;
274 			goto out;
275 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
276 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
277 				newname, ret);
278 			fw = ERR_PTR(ret);
279 			goto out;
280 		}
281 	}
282 
283 	DRM_DEV_ERROR(drm->dev, "failed to load %s\n", fwname);
284 	fw = ERR_PTR(-ENOENT);
285 out:
286 	kfree(newname);
287 	return fw;
288 }
289 
290 int adreno_load_fw(struct adreno_gpu *adreno_gpu)
291 {
292 	int i;
293 
294 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++) {
295 		const struct firmware *fw;
296 
297 		if (!adreno_gpu->info->fw[i])
298 			continue;
299 
300 		/* Skip if the firmware has already been loaded */
301 		if (adreno_gpu->fw[i])
302 			continue;
303 
304 		fw = adreno_request_fw(adreno_gpu, adreno_gpu->info->fw[i]);
305 		if (IS_ERR(fw))
306 			return PTR_ERR(fw);
307 
308 		adreno_gpu->fw[i] = fw;
309 	}
310 
311 	return 0;
312 }
313 
314 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu,
315 		const struct firmware *fw, u64 *iova)
316 {
317 	struct drm_gem_object *bo;
318 	void *ptr;
319 
320 	ptr = msm_gem_kernel_new_locked(gpu->dev, fw->size - 4,
321 		MSM_BO_UNCACHED | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
322 
323 	if (IS_ERR(ptr))
324 		return ERR_CAST(ptr);
325 
326 	memcpy(ptr, &fw->data[4], fw->size - 4);
327 
328 	msm_gem_put_vaddr(bo);
329 
330 	return bo;
331 }
332 
333 int adreno_hw_init(struct msm_gpu *gpu)
334 {
335 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
336 	int ret, i;
337 
338 	DBG("%s", gpu->name);
339 
340 	ret = adreno_load_fw(adreno_gpu);
341 	if (ret)
342 		return ret;
343 
344 	for (i = 0; i < gpu->nr_rings; i++) {
345 		struct msm_ringbuffer *ring = gpu->rb[i];
346 
347 		if (!ring)
348 			continue;
349 
350 		ring->cur = ring->start;
351 		ring->next = ring->start;
352 
353 		/* reset completed fence seqno: */
354 		ring->memptrs->fence = ring->seqno;
355 		ring->memptrs->rptr = 0;
356 	}
357 
358 	/*
359 	 * Setup REG_CP_RB_CNTL.  The same value is used across targets (with
360 	 * the excpetion of A430 that disables the RPTR shadow) - the cacluation
361 	 * for the ringbuffer size and block size is moved to msm_gpu.h for the
362 	 * pre-processor to deal with and the A430 variant is ORed in here
363 	 */
364 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL,
365 		MSM_GPU_RB_CNTL_DEFAULT |
366 		(adreno_is_a430(adreno_gpu) ? AXXX_CP_RB_CNTL_NO_UPDATE : 0));
367 
368 	/* Setup ringbuffer address - use ringbuffer[0] for GPU init */
369 	adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_BASE,
370 		REG_ADRENO_CP_RB_BASE_HI, gpu->rb[0]->iova);
371 
372 	if (!adreno_is_a430(adreno_gpu)) {
373 		adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR,
374 			REG_ADRENO_CP_RB_RPTR_ADDR_HI,
375 			rbmemptr(gpu->rb[0], rptr));
376 	}
377 
378 	return 0;
379 }
380 
381 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
382 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu,
383 		struct msm_ringbuffer *ring)
384 {
385 	if (adreno_is_a430(adreno_gpu))
386 		return ring->memptrs->rptr = adreno_gpu_read(
387 			adreno_gpu, REG_ADRENO_CP_RB_RPTR);
388 	else
389 		return ring->memptrs->rptr;
390 }
391 
392 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu)
393 {
394 	return gpu->rb[0];
395 }
396 
397 void adreno_recover(struct msm_gpu *gpu)
398 {
399 	struct drm_device *dev = gpu->dev;
400 	int ret;
401 
402 	// XXX pm-runtime??  we *need* the device to be off after this
403 	// so maybe continuing to call ->pm_suspend/resume() is better?
404 
405 	gpu->funcs->pm_suspend(gpu);
406 	gpu->funcs->pm_resume(gpu);
407 
408 	ret = msm_gpu_hw_init(gpu);
409 	if (ret) {
410 		DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
411 		/* hmm, oh well? */
412 	}
413 }
414 
415 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
416 		struct msm_file_private *ctx)
417 {
418 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
419 	struct msm_drm_private *priv = gpu->dev->dev_private;
420 	struct msm_ringbuffer *ring = submit->ring;
421 	unsigned i;
422 
423 	for (i = 0; i < submit->nr_cmds; i++) {
424 		switch (submit->cmd[i].type) {
425 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
426 			/* ignore IB-targets */
427 			break;
428 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
429 			/* ignore if there has not been a ctx switch: */
430 			if (priv->lastctx == ctx)
431 				break;
432 			/* fall-thru */
433 		case MSM_SUBMIT_CMD_BUF:
434 			OUT_PKT3(ring, adreno_is_a430(adreno_gpu) ?
435 				CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
436 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
437 			OUT_RING(ring, submit->cmd[i].size);
438 			OUT_PKT2(ring);
439 			break;
440 		}
441 	}
442 
443 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
444 	OUT_RING(ring, submit->seqno);
445 
446 	if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) {
447 		/* Flush HLSQ lazy updates to make sure there is nothing
448 		 * pending for indirect loads after the timestamp has
449 		 * passed:
450 		 */
451 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
452 		OUT_RING(ring, HLSQ_FLUSH);
453 	}
454 
455 	/* wait for idle before cache flush/interrupt */
456 	OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
457 	OUT_RING(ring, 0x00000000);
458 
459 	if (!adreno_is_a2xx(adreno_gpu)) {
460 		/* BIT(31) of CACHE_FLUSH_TS triggers CACHE_FLUSH_TS IRQ from GPU */
461 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
462 		OUT_RING(ring, CACHE_FLUSH_TS | BIT(31));
463 		OUT_RING(ring, rbmemptr(ring, fence));
464 		OUT_RING(ring, submit->seqno);
465 	} else {
466 		/* BIT(31) means something else on a2xx */
467 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
468 		OUT_RING(ring, CACHE_FLUSH_TS);
469 		OUT_RING(ring, rbmemptr(ring, fence));
470 		OUT_RING(ring, submit->seqno);
471 		OUT_PKT3(ring, CP_INTERRUPT, 1);
472 		OUT_RING(ring, 0x80000000);
473 	}
474 
475 #if 0
476 	if (adreno_is_a3xx(adreno_gpu)) {
477 		/* Dummy set-constant to trigger context rollover */
478 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
479 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
480 		OUT_RING(ring, 0x00000000);
481 	}
482 #endif
483 
484 	gpu->funcs->flush(gpu, ring);
485 }
486 
487 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
488 {
489 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
490 	uint32_t wptr;
491 
492 	/* Copy the shadow to the actual register */
493 	ring->cur = ring->next;
494 
495 	/*
496 	 * Mask wptr value that we calculate to fit in the HW range. This is
497 	 * to account for the possibility that the last command fit exactly into
498 	 * the ringbuffer and rb->next hasn't wrapped to zero yet
499 	 */
500 	wptr = get_wptr(ring);
501 
502 	/* ensure writes to ringbuffer have hit system memory: */
503 	mb();
504 
505 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr);
506 }
507 
508 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
509 {
510 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
511 	uint32_t wptr = get_wptr(ring);
512 
513 	/* wait for CP to drain ringbuffer: */
514 	if (!spin_until(get_rptr(adreno_gpu, ring) == wptr))
515 		return true;
516 
517 	/* TODO maybe we need to reset GPU here to recover from hang? */
518 	DRM_ERROR("%s: timeout waiting to drain ringbuffer %d rptr/wptr = %X/%X\n",
519 		gpu->name, ring->id, get_rptr(adreno_gpu, ring), wptr);
520 
521 	return false;
522 }
523 
524 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state)
525 {
526 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
527 	int i, count = 0;
528 
529 	kref_init(&state->ref);
530 
531 	ktime_get_real_ts64(&state->time);
532 
533 	for (i = 0; i < gpu->nr_rings; i++) {
534 		int size = 0, j;
535 
536 		state->ring[i].fence = gpu->rb[i]->memptrs->fence;
537 		state->ring[i].iova = gpu->rb[i]->iova;
538 		state->ring[i].seqno = gpu->rb[i]->seqno;
539 		state->ring[i].rptr = get_rptr(adreno_gpu, gpu->rb[i]);
540 		state->ring[i].wptr = get_wptr(gpu->rb[i]);
541 
542 		/* Copy at least 'wptr' dwords of the data */
543 		size = state->ring[i].wptr;
544 
545 		/* After wptr find the last non zero dword to save space */
546 		for (j = state->ring[i].wptr; j < MSM_GPU_RINGBUFFER_SZ >> 2; j++)
547 			if (gpu->rb[i]->start[j])
548 				size = j + 1;
549 
550 		if (size) {
551 			state->ring[i].data = kvmalloc(size << 2, GFP_KERNEL);
552 			if (state->ring[i].data) {
553 				memcpy(state->ring[i].data, gpu->rb[i]->start, size << 2);
554 				state->ring[i].data_size = size << 2;
555 			}
556 		}
557 	}
558 
559 	/* Some targets prefer to collect their own registers */
560 	if (!adreno_gpu->registers)
561 		return 0;
562 
563 	/* Count the number of registers */
564 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2)
565 		count += adreno_gpu->registers[i + 1] -
566 			adreno_gpu->registers[i] + 1;
567 
568 	state->registers = kcalloc(count * 2, sizeof(u32), GFP_KERNEL);
569 	if (state->registers) {
570 		int pos = 0;
571 
572 		for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
573 			u32 start = adreno_gpu->registers[i];
574 			u32 end   = adreno_gpu->registers[i + 1];
575 			u32 addr;
576 
577 			for (addr = start; addr <= end; addr++) {
578 				state->registers[pos++] = addr;
579 				state->registers[pos++] = gpu_read(gpu, addr);
580 			}
581 		}
582 
583 		state->nr_registers = count;
584 	}
585 
586 	return 0;
587 }
588 
589 void adreno_gpu_state_destroy(struct msm_gpu_state *state)
590 {
591 	int i;
592 
593 	for (i = 0; i < ARRAY_SIZE(state->ring); i++)
594 		kvfree(state->ring[i].data);
595 
596 	for (i = 0; state->bos && i < state->nr_bos; i++)
597 		kvfree(state->bos[i].data);
598 
599 	kfree(state->bos);
600 	kfree(state->comm);
601 	kfree(state->cmd);
602 	kfree(state->registers);
603 }
604 
605 static void adreno_gpu_state_kref_destroy(struct kref *kref)
606 {
607 	struct msm_gpu_state *state = container_of(kref,
608 		struct msm_gpu_state, ref);
609 
610 	adreno_gpu_state_destroy(state);
611 	kfree(state);
612 }
613 
614 int adreno_gpu_state_put(struct msm_gpu_state *state)
615 {
616 	if (IS_ERR_OR_NULL(state))
617 		return 1;
618 
619 	return kref_put(&state->ref, adreno_gpu_state_kref_destroy);
620 }
621 
622 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
623 
624 static char *adreno_gpu_ascii85_encode(u32 *src, size_t len)
625 {
626 	void *buf;
627 	size_t buf_itr = 0, buffer_size;
628 	char out[ASCII85_BUFSZ];
629 	long l;
630 	int i;
631 
632 	if (!src || !len)
633 		return NULL;
634 
635 	l = ascii85_encode_len(len);
636 
637 	/*
638 	 * Ascii85 outputs either a 5 byte string or a 1 byte string. So we
639 	 * account for the worst case of 5 bytes per dword plus the 1 for '\0'
640 	 */
641 	buffer_size = (l * 5) + 1;
642 
643 	buf = kvmalloc(buffer_size, GFP_KERNEL);
644 	if (!buf)
645 		return NULL;
646 
647 	for (i = 0; i < l; i++)
648 		buf_itr += snprintf(buf + buf_itr, buffer_size - buf_itr, "%s",
649 				ascii85_encode(src[i], out));
650 
651 	return buf;
652 }
653 
654 /* len is expected to be in bytes */
655 static void adreno_show_object(struct drm_printer *p, void **ptr, int len,
656 		bool *encoded)
657 {
658 	if (!*ptr || !len)
659 		return;
660 
661 	if (!*encoded) {
662 		long datalen, i;
663 		u32 *buf = *ptr;
664 
665 		/*
666 		 * Only dump the non-zero part of the buffer - rarely will
667 		 * any data completely fill the entire allocated size of
668 		 * the buffer.
669 		 */
670 		for (datalen = 0, i = 0; i < len >> 2; i++)
671 			if (buf[i])
672 				datalen = ((i + 1) << 2);
673 
674 		/*
675 		 * If we reach here, then the originally captured binary buffer
676 		 * will be replaced with the ascii85 encoded string
677 		 */
678 		*ptr = adreno_gpu_ascii85_encode(buf, datalen);
679 
680 		kvfree(buf);
681 
682 		*encoded = true;
683 	}
684 
685 	if (!*ptr)
686 		return;
687 
688 	drm_puts(p, "    data: !!ascii85 |\n");
689 	drm_puts(p, "     ");
690 
691 	drm_puts(p, *ptr);
692 
693 	drm_puts(p, "\n");
694 }
695 
696 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
697 		struct drm_printer *p)
698 {
699 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
700 	int i;
701 
702 	if (IS_ERR_OR_NULL(state))
703 		return;
704 
705 	drm_printf(p, "revision: %d (%d.%d.%d.%d)\n",
706 			adreno_gpu->info->revn, adreno_gpu->rev.core,
707 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
708 			adreno_gpu->rev.patchid);
709 
710 	drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
711 
712 	drm_puts(p, "ringbuffer:\n");
713 
714 	for (i = 0; i < gpu->nr_rings; i++) {
715 		drm_printf(p, "  - id: %d\n", i);
716 		drm_printf(p, "    iova: 0x%016llx\n", state->ring[i].iova);
717 		drm_printf(p, "    last-fence: %d\n", state->ring[i].seqno);
718 		drm_printf(p, "    retired-fence: %d\n", state->ring[i].fence);
719 		drm_printf(p, "    rptr: %d\n", state->ring[i].rptr);
720 		drm_printf(p, "    wptr: %d\n", state->ring[i].wptr);
721 		drm_printf(p, "    size: %d\n", MSM_GPU_RINGBUFFER_SZ);
722 
723 		adreno_show_object(p, &state->ring[i].data,
724 			state->ring[i].data_size, &state->ring[i].encoded);
725 	}
726 
727 	if (state->bos) {
728 		drm_puts(p, "bos:\n");
729 
730 		for (i = 0; i < state->nr_bos; i++) {
731 			drm_printf(p, "  - iova: 0x%016llx\n",
732 				state->bos[i].iova);
733 			drm_printf(p, "    size: %zd\n", state->bos[i].size);
734 
735 			adreno_show_object(p, &state->bos[i].data,
736 				state->bos[i].size, &state->bos[i].encoded);
737 		}
738 	}
739 
740 	if (state->nr_registers) {
741 		drm_puts(p, "registers:\n");
742 
743 		for (i = 0; i < state->nr_registers; i++) {
744 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
745 				state->registers[i * 2] << 2,
746 				state->registers[(i * 2) + 1]);
747 		}
748 	}
749 }
750 #endif
751 
752 /* Dump common gpu status and scratch registers on any hang, to make
753  * the hangcheck logs more useful.  The scratch registers seem always
754  * safe to read when GPU has hung (unlike some other regs, depending
755  * on how the GPU hung), and they are useful to match up to cmdstream
756  * dumps when debugging hangs:
757  */
758 void adreno_dump_info(struct msm_gpu *gpu)
759 {
760 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
761 	int i;
762 
763 	printk("revision: %d (%d.%d.%d.%d)\n",
764 			adreno_gpu->info->revn, adreno_gpu->rev.core,
765 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
766 			adreno_gpu->rev.patchid);
767 
768 	for (i = 0; i < gpu->nr_rings; i++) {
769 		struct msm_ringbuffer *ring = gpu->rb[i];
770 
771 		printk("rb %d: fence:    %d/%d\n", i,
772 			ring->memptrs->fence,
773 			ring->seqno);
774 
775 		printk("rptr:     %d\n", get_rptr(adreno_gpu, ring));
776 		printk("rb wptr:  %d\n", get_wptr(ring));
777 	}
778 }
779 
780 /* would be nice to not have to duplicate the _show() stuff with printk(): */
781 void adreno_dump(struct msm_gpu *gpu)
782 {
783 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
784 	int i;
785 
786 	if (!adreno_gpu->registers)
787 		return;
788 
789 	/* dump these out in a form that can be parsed by demsm: */
790 	printk("IO:region %s 00000000 00020000\n", gpu->name);
791 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
792 		uint32_t start = adreno_gpu->registers[i];
793 		uint32_t end   = adreno_gpu->registers[i+1];
794 		uint32_t addr;
795 
796 		for (addr = start; addr <= end; addr++) {
797 			uint32_t val = gpu_read(gpu, addr);
798 			printk("IO:R %08x %08x\n", addr<<2, val);
799 		}
800 	}
801 }
802 
803 static uint32_t ring_freewords(struct msm_ringbuffer *ring)
804 {
805 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(ring->gpu);
806 	uint32_t size = MSM_GPU_RINGBUFFER_SZ >> 2;
807 	/* Use ring->next to calculate free size */
808 	uint32_t wptr = ring->next - ring->start;
809 	uint32_t rptr = get_rptr(adreno_gpu, ring);
810 	return (rptr + (size - 1) - wptr) % size;
811 }
812 
813 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords)
814 {
815 	if (spin_until(ring_freewords(ring) >= ndwords))
816 		DRM_DEV_ERROR(ring->gpu->dev->dev,
817 			"timeout waiting for space in ringbuffer %d\n",
818 			ring->id);
819 }
820 
821 /* Get legacy powerlevels from qcom,gpu-pwrlevels and populate the opp table */
822 static int adreno_get_legacy_pwrlevels(struct device *dev)
823 {
824 	struct device_node *child, *node;
825 	int ret;
826 
827 	node = of_get_compatible_child(dev->of_node, "qcom,gpu-pwrlevels");
828 	if (!node) {
829 		DRM_DEV_ERROR(dev, "Could not find the GPU powerlevels\n");
830 		return -ENXIO;
831 	}
832 
833 	for_each_child_of_node(node, child) {
834 		unsigned int val;
835 
836 		ret = of_property_read_u32(child, "qcom,gpu-freq", &val);
837 		if (ret)
838 			continue;
839 
840 		/*
841 		 * Skip the intentionally bogus clock value found at the bottom
842 		 * of most legacy frequency tables
843 		 */
844 		if (val != 27000000)
845 			dev_pm_opp_add(dev, val, 0);
846 	}
847 
848 	of_node_put(node);
849 
850 	return 0;
851 }
852 
853 static int adreno_get_pwrlevels(struct device *dev,
854 		struct msm_gpu *gpu)
855 {
856 	unsigned long freq = ULONG_MAX;
857 	struct dev_pm_opp *opp;
858 	int ret;
859 
860 	gpu->fast_rate = 0;
861 
862 	/* You down with OPP? */
863 	if (!of_find_property(dev->of_node, "operating-points-v2", NULL))
864 		ret = adreno_get_legacy_pwrlevels(dev);
865 	else {
866 		ret = dev_pm_opp_of_add_table(dev);
867 		if (ret)
868 			DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
869 	}
870 
871 	if (!ret) {
872 		/* Find the fastest defined rate */
873 		opp = dev_pm_opp_find_freq_floor(dev, &freq);
874 		if (!IS_ERR(opp)) {
875 			gpu->fast_rate = freq;
876 			dev_pm_opp_put(opp);
877 		}
878 	}
879 
880 	if (!gpu->fast_rate) {
881 		dev_warn(dev,
882 			"Could not find a clock rate. Using a reasonable default\n");
883 		/* Pick a suitably safe clock speed for any target */
884 		gpu->fast_rate = 200000000;
885 	}
886 
887 	DBG("fast_rate=%u, slow_rate=27000000", gpu->fast_rate);
888 
889 	/* Check for an interconnect path for the bus */
890 	gpu->icc_path = of_icc_get(dev, NULL);
891 	if (IS_ERR(gpu->icc_path))
892 		gpu->icc_path = NULL;
893 
894 	return 0;
895 }
896 
897 int adreno_gpu_ocmem_init(struct device *dev, struct adreno_gpu *adreno_gpu,
898 			  struct adreno_ocmem *adreno_ocmem)
899 {
900 	struct ocmem_buf *ocmem_hdl;
901 	struct ocmem *ocmem;
902 
903 	ocmem = of_get_ocmem(dev);
904 	if (IS_ERR(ocmem)) {
905 		if (PTR_ERR(ocmem) == -ENODEV) {
906 			/*
907 			 * Return success since either the ocmem property was
908 			 * not specified in device tree, or ocmem support is
909 			 * not compiled into the kernel.
910 			 */
911 			return 0;
912 		}
913 
914 		return PTR_ERR(ocmem);
915 	}
916 
917 	ocmem_hdl = ocmem_allocate(ocmem, OCMEM_GRAPHICS, adreno_gpu->gmem);
918 	if (IS_ERR(ocmem_hdl))
919 		return PTR_ERR(ocmem_hdl);
920 
921 	adreno_ocmem->ocmem = ocmem;
922 	adreno_ocmem->base = ocmem_hdl->addr;
923 	adreno_ocmem->hdl = ocmem_hdl;
924 	adreno_gpu->gmem = ocmem_hdl->len;
925 
926 	return 0;
927 }
928 
929 void adreno_gpu_ocmem_cleanup(struct adreno_ocmem *adreno_ocmem)
930 {
931 	if (adreno_ocmem && adreno_ocmem->base)
932 		ocmem_free(adreno_ocmem->ocmem, OCMEM_GRAPHICS,
933 			   adreno_ocmem->hdl);
934 }
935 
936 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
937 		struct adreno_gpu *adreno_gpu,
938 		const struct adreno_gpu_funcs *funcs, int nr_rings)
939 {
940 	struct adreno_platform_config *config = pdev->dev.platform_data;
941 	struct msm_gpu_config adreno_gpu_config  = { 0 };
942 	struct msm_gpu *gpu = &adreno_gpu->base;
943 
944 	adreno_gpu->funcs = funcs;
945 	adreno_gpu->info = adreno_info(config->rev);
946 	adreno_gpu->gmem = adreno_gpu->info->gmem;
947 	adreno_gpu->revn = adreno_gpu->info->revn;
948 	adreno_gpu->rev = config->rev;
949 
950 	adreno_gpu_config.ioname = "kgsl_3d0_reg_memory";
951 
952 	adreno_gpu_config.va_start = SZ_16M;
953 	adreno_gpu_config.va_end = 0xffffffff;
954 	/* maximum range of a2xx mmu */
955 	if (adreno_is_a2xx(adreno_gpu))
956 		adreno_gpu_config.va_end = SZ_16M + 0xfff * SZ_64K;
957 
958 	adreno_gpu_config.nr_rings = nr_rings;
959 
960 	adreno_get_pwrlevels(&pdev->dev, gpu);
961 
962 	pm_runtime_set_autosuspend_delay(&pdev->dev,
963 		adreno_gpu->info->inactive_period);
964 	pm_runtime_use_autosuspend(&pdev->dev);
965 	pm_runtime_enable(&pdev->dev);
966 
967 	return msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
968 			adreno_gpu->info->name, &adreno_gpu_config);
969 }
970 
971 void adreno_gpu_cleanup(struct adreno_gpu *adreno_gpu)
972 {
973 	struct msm_gpu *gpu = &adreno_gpu->base;
974 	unsigned int i;
975 
976 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++)
977 		release_firmware(adreno_gpu->fw[i]);
978 
979 	icc_put(gpu->icc_path);
980 
981 	msm_gpu_cleanup(&adreno_gpu->base);
982 }
983