1 /*
2  * Copyright (C) 2013 Red Hat
3  * Author: Rob Clark <robdclark@gmail.com>
4  *
5  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
6  *
7  * This program is free software; you can redistribute it and/or modify it
8  * under the terms of the GNU General Public License version 2 as published by
9  * the Free Software Foundation.
10  *
11  * This program is distributed in the hope that it will be useful, but WITHOUT
12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
14  * more details.
15  *
16  * You should have received a copy of the GNU General Public License along with
17  * this program.  If not, see <http://www.gnu.org/licenses/>.
18  */
19 
20 #include <linux/ascii85.h>
21 #include <linux/interconnect.h>
22 #include <linux/kernel.h>
23 #include <linux/pm_opp.h>
24 #include <linux/slab.h>
25 #include "adreno_gpu.h"
26 #include "msm_gem.h"
27 #include "msm_mmu.h"
28 
29 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
30 {
31 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
32 
33 	switch (param) {
34 	case MSM_PARAM_GPU_ID:
35 		*value = adreno_gpu->info->revn;
36 		return 0;
37 	case MSM_PARAM_GMEM_SIZE:
38 		*value = adreno_gpu->gmem;
39 		return 0;
40 	case MSM_PARAM_GMEM_BASE:
41 		*value = 0x100000;
42 		return 0;
43 	case MSM_PARAM_CHIP_ID:
44 		*value = adreno_gpu->rev.patchid |
45 				(adreno_gpu->rev.minor << 8) |
46 				(adreno_gpu->rev.major << 16) |
47 				(adreno_gpu->rev.core << 24);
48 		return 0;
49 	case MSM_PARAM_MAX_FREQ:
50 		*value = adreno_gpu->base.fast_rate;
51 		return 0;
52 	case MSM_PARAM_TIMESTAMP:
53 		if (adreno_gpu->funcs->get_timestamp) {
54 			int ret;
55 
56 			pm_runtime_get_sync(&gpu->pdev->dev);
57 			ret = adreno_gpu->funcs->get_timestamp(gpu, value);
58 			pm_runtime_put_autosuspend(&gpu->pdev->dev);
59 
60 			return ret;
61 		}
62 		return -EINVAL;
63 	case MSM_PARAM_NR_RINGS:
64 		*value = gpu->nr_rings;
65 		return 0;
66 	default:
67 		DBG("%s: invalid param: %u", gpu->name, param);
68 		return -EINVAL;
69 	}
70 }
71 
72 const struct firmware *
73 adreno_request_fw(struct adreno_gpu *adreno_gpu, const char *fwname)
74 {
75 	struct drm_device *drm = adreno_gpu->base.dev;
76 	const struct firmware *fw = NULL;
77 	char *newname;
78 	int ret;
79 
80 	newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
81 	if (!newname)
82 		return ERR_PTR(-ENOMEM);
83 
84 	/*
85 	 * Try first to load from qcom/$fwfile using a direct load (to avoid
86 	 * a potential timeout waiting for usermode helper)
87 	 */
88 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
89 	    (adreno_gpu->fwloc == FW_LOCATION_NEW)) {
90 
91 		ret = request_firmware_direct(&fw, newname, drm->dev);
92 		if (!ret) {
93 			DRM_DEV_INFO(drm->dev, "loaded %s from new location\n",
94 				newname);
95 			adreno_gpu->fwloc = FW_LOCATION_NEW;
96 			goto out;
97 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
98 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
99 				newname, ret);
100 			fw = ERR_PTR(ret);
101 			goto out;
102 		}
103 	}
104 
105 	/*
106 	 * Then try the legacy location without qcom/ prefix
107 	 */
108 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
109 	    (adreno_gpu->fwloc == FW_LOCATION_LEGACY)) {
110 
111 		ret = request_firmware_direct(&fw, fwname, drm->dev);
112 		if (!ret) {
113 			DRM_DEV_INFO(drm->dev, "loaded %s from legacy location\n",
114 				newname);
115 			adreno_gpu->fwloc = FW_LOCATION_LEGACY;
116 			goto out;
117 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
118 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
119 				fwname, ret);
120 			fw = ERR_PTR(ret);
121 			goto out;
122 		}
123 	}
124 
125 	/*
126 	 * Finally fall back to request_firmware() for cases where the
127 	 * usermode helper is needed (I think mainly android)
128 	 */
129 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
130 	    (adreno_gpu->fwloc == FW_LOCATION_HELPER)) {
131 
132 		ret = request_firmware(&fw, newname, drm->dev);
133 		if (!ret) {
134 			DRM_DEV_INFO(drm->dev, "loaded %s with helper\n",
135 				newname);
136 			adreno_gpu->fwloc = FW_LOCATION_HELPER;
137 			goto out;
138 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
139 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
140 				newname, ret);
141 			fw = ERR_PTR(ret);
142 			goto out;
143 		}
144 	}
145 
146 	DRM_DEV_ERROR(drm->dev, "failed to load %s\n", fwname);
147 	fw = ERR_PTR(-ENOENT);
148 out:
149 	kfree(newname);
150 	return fw;
151 }
152 
153 int adreno_load_fw(struct adreno_gpu *adreno_gpu)
154 {
155 	int i;
156 
157 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++) {
158 		const struct firmware *fw;
159 
160 		if (!adreno_gpu->info->fw[i])
161 			continue;
162 
163 		/* Skip if the firmware has already been loaded */
164 		if (adreno_gpu->fw[i])
165 			continue;
166 
167 		fw = adreno_request_fw(adreno_gpu, adreno_gpu->info->fw[i]);
168 		if (IS_ERR(fw))
169 			return PTR_ERR(fw);
170 
171 		adreno_gpu->fw[i] = fw;
172 	}
173 
174 	return 0;
175 }
176 
177 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu,
178 		const struct firmware *fw, u64 *iova)
179 {
180 	struct drm_gem_object *bo;
181 	void *ptr;
182 
183 	ptr = msm_gem_kernel_new_locked(gpu->dev, fw->size - 4,
184 		MSM_BO_UNCACHED | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
185 
186 	if (IS_ERR(ptr))
187 		return ERR_CAST(ptr);
188 
189 	memcpy(ptr, &fw->data[4], fw->size - 4);
190 
191 	msm_gem_put_vaddr(bo);
192 
193 	return bo;
194 }
195 
196 int adreno_hw_init(struct msm_gpu *gpu)
197 {
198 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
199 	int ret, i;
200 
201 	DBG("%s", gpu->name);
202 
203 	ret = adreno_load_fw(adreno_gpu);
204 	if (ret)
205 		return ret;
206 
207 	for (i = 0; i < gpu->nr_rings; i++) {
208 		struct msm_ringbuffer *ring = gpu->rb[i];
209 
210 		if (!ring)
211 			continue;
212 
213 		ring->cur = ring->start;
214 		ring->next = ring->start;
215 
216 		/* reset completed fence seqno: */
217 		ring->memptrs->fence = ring->seqno;
218 		ring->memptrs->rptr = 0;
219 	}
220 
221 	/*
222 	 * Setup REG_CP_RB_CNTL.  The same value is used across targets (with
223 	 * the excpetion of A430 that disables the RPTR shadow) - the cacluation
224 	 * for the ringbuffer size and block size is moved to msm_gpu.h for the
225 	 * pre-processor to deal with and the A430 variant is ORed in here
226 	 */
227 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL,
228 		MSM_GPU_RB_CNTL_DEFAULT |
229 		(adreno_is_a430(adreno_gpu) ? AXXX_CP_RB_CNTL_NO_UPDATE : 0));
230 
231 	/* Setup ringbuffer address - use ringbuffer[0] for GPU init */
232 	adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_BASE,
233 		REG_ADRENO_CP_RB_BASE_HI, gpu->rb[0]->iova);
234 
235 	if (!adreno_is_a430(adreno_gpu)) {
236 		adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR,
237 			REG_ADRENO_CP_RB_RPTR_ADDR_HI,
238 			rbmemptr(gpu->rb[0], rptr));
239 	}
240 
241 	return 0;
242 }
243 
244 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
245 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu,
246 		struct msm_ringbuffer *ring)
247 {
248 	if (adreno_is_a430(adreno_gpu))
249 		return ring->memptrs->rptr = adreno_gpu_read(
250 			adreno_gpu, REG_ADRENO_CP_RB_RPTR);
251 	else
252 		return ring->memptrs->rptr;
253 }
254 
255 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu)
256 {
257 	return gpu->rb[0];
258 }
259 
260 void adreno_recover(struct msm_gpu *gpu)
261 {
262 	struct drm_device *dev = gpu->dev;
263 	int ret;
264 
265 	// XXX pm-runtime??  we *need* the device to be off after this
266 	// so maybe continuing to call ->pm_suspend/resume() is better?
267 
268 	gpu->funcs->pm_suspend(gpu);
269 	gpu->funcs->pm_resume(gpu);
270 
271 	ret = msm_gpu_hw_init(gpu);
272 	if (ret) {
273 		DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
274 		/* hmm, oh well? */
275 	}
276 }
277 
278 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
279 		struct msm_file_private *ctx)
280 {
281 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
282 	struct msm_drm_private *priv = gpu->dev->dev_private;
283 	struct msm_ringbuffer *ring = submit->ring;
284 	unsigned i;
285 
286 	for (i = 0; i < submit->nr_cmds; i++) {
287 		switch (submit->cmd[i].type) {
288 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
289 			/* ignore IB-targets */
290 			break;
291 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
292 			/* ignore if there has not been a ctx switch: */
293 			if (priv->lastctx == ctx)
294 				break;
295 		case MSM_SUBMIT_CMD_BUF:
296 			OUT_PKT3(ring, adreno_is_a430(adreno_gpu) ?
297 				CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
298 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
299 			OUT_RING(ring, submit->cmd[i].size);
300 			OUT_PKT2(ring);
301 			break;
302 		}
303 	}
304 
305 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
306 	OUT_RING(ring, submit->seqno);
307 
308 	if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) {
309 		/* Flush HLSQ lazy updates to make sure there is nothing
310 		 * pending for indirect loads after the timestamp has
311 		 * passed:
312 		 */
313 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
314 		OUT_RING(ring, HLSQ_FLUSH);
315 	}
316 
317 	/* wait for idle before cache flush/interrupt */
318 	OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
319 	OUT_RING(ring, 0x00000000);
320 
321 	if (!adreno_is_a2xx(adreno_gpu)) {
322 		/* BIT(31) of CACHE_FLUSH_TS triggers CACHE_FLUSH_TS IRQ from GPU */
323 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
324 		OUT_RING(ring, CACHE_FLUSH_TS | BIT(31));
325 		OUT_RING(ring, rbmemptr(ring, fence));
326 		OUT_RING(ring, submit->seqno);
327 	} else {
328 		/* BIT(31) means something else on a2xx */
329 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
330 		OUT_RING(ring, CACHE_FLUSH_TS);
331 		OUT_RING(ring, rbmemptr(ring, fence));
332 		OUT_RING(ring, submit->seqno);
333 		OUT_PKT3(ring, CP_INTERRUPT, 1);
334 		OUT_RING(ring, 0x80000000);
335 	}
336 
337 #if 0
338 	if (adreno_is_a3xx(adreno_gpu)) {
339 		/* Dummy set-constant to trigger context rollover */
340 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
341 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
342 		OUT_RING(ring, 0x00000000);
343 	}
344 #endif
345 
346 	gpu->funcs->flush(gpu, ring);
347 }
348 
349 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
350 {
351 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
352 	uint32_t wptr;
353 
354 	/* Copy the shadow to the actual register */
355 	ring->cur = ring->next;
356 
357 	/*
358 	 * Mask wptr value that we calculate to fit in the HW range. This is
359 	 * to account for the possibility that the last command fit exactly into
360 	 * the ringbuffer and rb->next hasn't wrapped to zero yet
361 	 */
362 	wptr = get_wptr(ring);
363 
364 	/* ensure writes to ringbuffer have hit system memory: */
365 	mb();
366 
367 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr);
368 }
369 
370 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
371 {
372 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
373 	uint32_t wptr = get_wptr(ring);
374 
375 	/* wait for CP to drain ringbuffer: */
376 	if (!spin_until(get_rptr(adreno_gpu, ring) == wptr))
377 		return true;
378 
379 	/* TODO maybe we need to reset GPU here to recover from hang? */
380 	DRM_ERROR("%s: timeout waiting to drain ringbuffer %d rptr/wptr = %X/%X\n",
381 		gpu->name, ring->id, get_rptr(adreno_gpu, ring), wptr);
382 
383 	return false;
384 }
385 
386 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state)
387 {
388 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
389 	int i, count = 0;
390 
391 	kref_init(&state->ref);
392 
393 	ktime_get_real_ts64(&state->time);
394 
395 	for (i = 0; i < gpu->nr_rings; i++) {
396 		int size = 0, j;
397 
398 		state->ring[i].fence = gpu->rb[i]->memptrs->fence;
399 		state->ring[i].iova = gpu->rb[i]->iova;
400 		state->ring[i].seqno = gpu->rb[i]->seqno;
401 		state->ring[i].rptr = get_rptr(adreno_gpu, gpu->rb[i]);
402 		state->ring[i].wptr = get_wptr(gpu->rb[i]);
403 
404 		/* Copy at least 'wptr' dwords of the data */
405 		size = state->ring[i].wptr;
406 
407 		/* After wptr find the last non zero dword to save space */
408 		for (j = state->ring[i].wptr; j < MSM_GPU_RINGBUFFER_SZ >> 2; j++)
409 			if (gpu->rb[i]->start[j])
410 				size = j + 1;
411 
412 		if (size) {
413 			state->ring[i].data = kvmalloc(size << 2, GFP_KERNEL);
414 			if (state->ring[i].data) {
415 				memcpy(state->ring[i].data, gpu->rb[i]->start, size << 2);
416 				state->ring[i].data_size = size << 2;
417 			}
418 		}
419 	}
420 
421 	/* Some targets prefer to collect their own registers */
422 	if (!adreno_gpu->registers)
423 		return 0;
424 
425 	/* Count the number of registers */
426 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2)
427 		count += adreno_gpu->registers[i + 1] -
428 			adreno_gpu->registers[i] + 1;
429 
430 	state->registers = kcalloc(count * 2, sizeof(u32), GFP_KERNEL);
431 	if (state->registers) {
432 		int pos = 0;
433 
434 		for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
435 			u32 start = adreno_gpu->registers[i];
436 			u32 end   = adreno_gpu->registers[i + 1];
437 			u32 addr;
438 
439 			for (addr = start; addr <= end; addr++) {
440 				state->registers[pos++] = addr;
441 				state->registers[pos++] = gpu_read(gpu, addr);
442 			}
443 		}
444 
445 		state->nr_registers = count;
446 	}
447 
448 	return 0;
449 }
450 
451 void adreno_gpu_state_destroy(struct msm_gpu_state *state)
452 {
453 	int i;
454 
455 	for (i = 0; i < ARRAY_SIZE(state->ring); i++)
456 		kvfree(state->ring[i].data);
457 
458 	for (i = 0; state->bos && i < state->nr_bos; i++)
459 		kvfree(state->bos[i].data);
460 
461 	kfree(state->bos);
462 	kfree(state->comm);
463 	kfree(state->cmd);
464 	kfree(state->registers);
465 }
466 
467 static void adreno_gpu_state_kref_destroy(struct kref *kref)
468 {
469 	struct msm_gpu_state *state = container_of(kref,
470 		struct msm_gpu_state, ref);
471 
472 	adreno_gpu_state_destroy(state);
473 	kfree(state);
474 }
475 
476 int adreno_gpu_state_put(struct msm_gpu_state *state)
477 {
478 	if (IS_ERR_OR_NULL(state))
479 		return 1;
480 
481 	return kref_put(&state->ref, adreno_gpu_state_kref_destroy);
482 }
483 
484 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
485 
486 static char *adreno_gpu_ascii85_encode(u32 *src, size_t len)
487 {
488 	void *buf;
489 	size_t buf_itr = 0, buffer_size;
490 	char out[ASCII85_BUFSZ];
491 	long l;
492 	int i;
493 
494 	if (!src || !len)
495 		return NULL;
496 
497 	l = ascii85_encode_len(len);
498 
499 	/*
500 	 * Ascii85 outputs either a 5 byte string or a 1 byte string. So we
501 	 * account for the worst case of 5 bytes per dword plus the 1 for '\0'
502 	 */
503 	buffer_size = (l * 5) + 1;
504 
505 	buf = kvmalloc(buffer_size, GFP_KERNEL);
506 	if (!buf)
507 		return NULL;
508 
509 	for (i = 0; i < l; i++)
510 		buf_itr += snprintf(buf + buf_itr, buffer_size - buf_itr, "%s",
511 				ascii85_encode(src[i], out));
512 
513 	return buf;
514 }
515 
516 /* len is expected to be in bytes */
517 static void adreno_show_object(struct drm_printer *p, void **ptr, int len,
518 		bool *encoded)
519 {
520 	if (!*ptr || !len)
521 		return;
522 
523 	if (!*encoded) {
524 		long datalen, i;
525 		u32 *buf = *ptr;
526 
527 		/*
528 		 * Only dump the non-zero part of the buffer - rarely will
529 		 * any data completely fill the entire allocated size of
530 		 * the buffer.
531 		 */
532 		for (datalen = 0, i = 0; i < len >> 2; i++)
533 			if (buf[i])
534 				datalen = ((i + 1) << 2);
535 
536 		/*
537 		 * If we reach here, then the originally captured binary buffer
538 		 * will be replaced with the ascii85 encoded string
539 		 */
540 		*ptr = adreno_gpu_ascii85_encode(buf, datalen);
541 
542 		kvfree(buf);
543 
544 		*encoded = true;
545 	}
546 
547 	if (!*ptr)
548 		return;
549 
550 	drm_puts(p, "    data: !!ascii85 |\n");
551 	drm_puts(p, "     ");
552 
553 	drm_puts(p, *ptr);
554 
555 	drm_puts(p, "\n");
556 }
557 
558 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
559 		struct drm_printer *p)
560 {
561 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
562 	int i;
563 
564 	if (IS_ERR_OR_NULL(state))
565 		return;
566 
567 	drm_printf(p, "revision: %d (%d.%d.%d.%d)\n",
568 			adreno_gpu->info->revn, adreno_gpu->rev.core,
569 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
570 			adreno_gpu->rev.patchid);
571 
572 	drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
573 
574 	drm_puts(p, "ringbuffer:\n");
575 
576 	for (i = 0; i < gpu->nr_rings; i++) {
577 		drm_printf(p, "  - id: %d\n", i);
578 		drm_printf(p, "    iova: 0x%016llx\n", state->ring[i].iova);
579 		drm_printf(p, "    last-fence: %d\n", state->ring[i].seqno);
580 		drm_printf(p, "    retired-fence: %d\n", state->ring[i].fence);
581 		drm_printf(p, "    rptr: %d\n", state->ring[i].rptr);
582 		drm_printf(p, "    wptr: %d\n", state->ring[i].wptr);
583 		drm_printf(p, "    size: %d\n", MSM_GPU_RINGBUFFER_SZ);
584 
585 		adreno_show_object(p, &state->ring[i].data,
586 			state->ring[i].data_size, &state->ring[i].encoded);
587 	}
588 
589 	if (state->bos) {
590 		drm_puts(p, "bos:\n");
591 
592 		for (i = 0; i < state->nr_bos; i++) {
593 			drm_printf(p, "  - iova: 0x%016llx\n",
594 				state->bos[i].iova);
595 			drm_printf(p, "    size: %zd\n", state->bos[i].size);
596 
597 			adreno_show_object(p, &state->bos[i].data,
598 				state->bos[i].size, &state->bos[i].encoded);
599 		}
600 	}
601 
602 	if (state->nr_registers) {
603 		drm_puts(p, "registers:\n");
604 
605 		for (i = 0; i < state->nr_registers; i++) {
606 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
607 				state->registers[i * 2] << 2,
608 				state->registers[(i * 2) + 1]);
609 		}
610 	}
611 }
612 #endif
613 
614 /* Dump common gpu status and scratch registers on any hang, to make
615  * the hangcheck logs more useful.  The scratch registers seem always
616  * safe to read when GPU has hung (unlike some other regs, depending
617  * on how the GPU hung), and they are useful to match up to cmdstream
618  * dumps when debugging hangs:
619  */
620 void adreno_dump_info(struct msm_gpu *gpu)
621 {
622 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
623 	int i;
624 
625 	printk("revision: %d (%d.%d.%d.%d)\n",
626 			adreno_gpu->info->revn, adreno_gpu->rev.core,
627 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
628 			adreno_gpu->rev.patchid);
629 
630 	for (i = 0; i < gpu->nr_rings; i++) {
631 		struct msm_ringbuffer *ring = gpu->rb[i];
632 
633 		printk("rb %d: fence:    %d/%d\n", i,
634 			ring->memptrs->fence,
635 			ring->seqno);
636 
637 		printk("rptr:     %d\n", get_rptr(adreno_gpu, ring));
638 		printk("rb wptr:  %d\n", get_wptr(ring));
639 	}
640 }
641 
642 /* would be nice to not have to duplicate the _show() stuff with printk(): */
643 void adreno_dump(struct msm_gpu *gpu)
644 {
645 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
646 	int i;
647 
648 	if (!adreno_gpu->registers)
649 		return;
650 
651 	/* dump these out in a form that can be parsed by demsm: */
652 	printk("IO:region %s 00000000 00020000\n", gpu->name);
653 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
654 		uint32_t start = adreno_gpu->registers[i];
655 		uint32_t end   = adreno_gpu->registers[i+1];
656 		uint32_t addr;
657 
658 		for (addr = start; addr <= end; addr++) {
659 			uint32_t val = gpu_read(gpu, addr);
660 			printk("IO:R %08x %08x\n", addr<<2, val);
661 		}
662 	}
663 }
664 
665 static uint32_t ring_freewords(struct msm_ringbuffer *ring)
666 {
667 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(ring->gpu);
668 	uint32_t size = MSM_GPU_RINGBUFFER_SZ >> 2;
669 	/* Use ring->next to calculate free size */
670 	uint32_t wptr = ring->next - ring->start;
671 	uint32_t rptr = get_rptr(adreno_gpu, ring);
672 	return (rptr + (size - 1) - wptr) % size;
673 }
674 
675 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords)
676 {
677 	if (spin_until(ring_freewords(ring) >= ndwords))
678 		DRM_DEV_ERROR(ring->gpu->dev->dev,
679 			"timeout waiting for space in ringbuffer %d\n",
680 			ring->id);
681 }
682 
683 /* Get legacy powerlevels from qcom,gpu-pwrlevels and populate the opp table */
684 static int adreno_get_legacy_pwrlevels(struct device *dev)
685 {
686 	struct device_node *child, *node;
687 	int ret;
688 
689 	node = of_get_compatible_child(dev->of_node, "qcom,gpu-pwrlevels");
690 	if (!node) {
691 		DRM_DEV_ERROR(dev, "Could not find the GPU powerlevels\n");
692 		return -ENXIO;
693 	}
694 
695 	for_each_child_of_node(node, child) {
696 		unsigned int val;
697 
698 		ret = of_property_read_u32(child, "qcom,gpu-freq", &val);
699 		if (ret)
700 			continue;
701 
702 		/*
703 		 * Skip the intentionally bogus clock value found at the bottom
704 		 * of most legacy frequency tables
705 		 */
706 		if (val != 27000000)
707 			dev_pm_opp_add(dev, val, 0);
708 	}
709 
710 	of_node_put(node);
711 
712 	return 0;
713 }
714 
715 static int adreno_get_pwrlevels(struct device *dev,
716 		struct msm_gpu *gpu)
717 {
718 	unsigned long freq = ULONG_MAX;
719 	struct dev_pm_opp *opp;
720 	int ret;
721 
722 	gpu->fast_rate = 0;
723 
724 	/* You down with OPP? */
725 	if (!of_find_property(dev->of_node, "operating-points-v2", NULL))
726 		ret = adreno_get_legacy_pwrlevels(dev);
727 	else {
728 		ret = dev_pm_opp_of_add_table(dev);
729 		if (ret)
730 			DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
731 	}
732 
733 	if (!ret) {
734 		/* Find the fastest defined rate */
735 		opp = dev_pm_opp_find_freq_floor(dev, &freq);
736 		if (!IS_ERR(opp)) {
737 			gpu->fast_rate = freq;
738 			dev_pm_opp_put(opp);
739 		}
740 	}
741 
742 	if (!gpu->fast_rate) {
743 		dev_warn(dev,
744 			"Could not find a clock rate. Using a reasonable default\n");
745 		/* Pick a suitably safe clock speed for any target */
746 		gpu->fast_rate = 200000000;
747 	}
748 
749 	DBG("fast_rate=%u, slow_rate=27000000", gpu->fast_rate);
750 
751 	/* Check for an interconnect path for the bus */
752 	gpu->icc_path = of_icc_get(dev, NULL);
753 	if (IS_ERR(gpu->icc_path))
754 		gpu->icc_path = NULL;
755 
756 	return 0;
757 }
758 
759 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
760 		struct adreno_gpu *adreno_gpu,
761 		const struct adreno_gpu_funcs *funcs, int nr_rings)
762 {
763 	struct adreno_platform_config *config = pdev->dev.platform_data;
764 	struct msm_gpu_config adreno_gpu_config  = { 0 };
765 	struct msm_gpu *gpu = &adreno_gpu->base;
766 
767 	adreno_gpu->funcs = funcs;
768 	adreno_gpu->info = adreno_info(config->rev);
769 	adreno_gpu->gmem = adreno_gpu->info->gmem;
770 	adreno_gpu->revn = adreno_gpu->info->revn;
771 	adreno_gpu->rev = config->rev;
772 
773 	adreno_gpu_config.ioname = "kgsl_3d0_reg_memory";
774 
775 	adreno_gpu_config.va_start = SZ_16M;
776 	adreno_gpu_config.va_end = 0xffffffff;
777 	/* maximum range of a2xx mmu */
778 	if (adreno_is_a2xx(adreno_gpu))
779 		adreno_gpu_config.va_end = SZ_16M + 0xfff * SZ_64K;
780 
781 	adreno_gpu_config.nr_rings = nr_rings;
782 
783 	adreno_get_pwrlevels(&pdev->dev, gpu);
784 
785 	pm_runtime_set_autosuspend_delay(&pdev->dev,
786 		adreno_gpu->info->inactive_period);
787 	pm_runtime_use_autosuspend(&pdev->dev);
788 	pm_runtime_enable(&pdev->dev);
789 
790 	return msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
791 			adreno_gpu->info->name, &adreno_gpu_config);
792 }
793 
794 void adreno_gpu_cleanup(struct adreno_gpu *adreno_gpu)
795 {
796 	struct msm_gpu *gpu = &adreno_gpu->base;
797 	unsigned int i;
798 
799 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++)
800 		release_firmware(adreno_gpu->fw[i]);
801 
802 	icc_put(gpu->icc_path);
803 
804 	msm_gpu_cleanup(&adreno_gpu->base);
805 }
806