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/kernel.h>
22 #include <linux/pm_opp.h>
23 #include <linux/slab.h>
24 #include "adreno_gpu.h"
25 #include "msm_gem.h"
26 #include "msm_mmu.h"
27 
28 int adreno_get_param(struct msm_gpu *gpu, uint32_t param, uint64_t *value)
29 {
30 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
31 
32 	switch (param) {
33 	case MSM_PARAM_GPU_ID:
34 		*value = adreno_gpu->info->revn;
35 		return 0;
36 	case MSM_PARAM_GMEM_SIZE:
37 		*value = adreno_gpu->gmem;
38 		return 0;
39 	case MSM_PARAM_GMEM_BASE:
40 		*value = 0x100000;
41 		return 0;
42 	case MSM_PARAM_CHIP_ID:
43 		*value = adreno_gpu->rev.patchid |
44 				(adreno_gpu->rev.minor << 8) |
45 				(adreno_gpu->rev.major << 16) |
46 				(adreno_gpu->rev.core << 24);
47 		return 0;
48 	case MSM_PARAM_MAX_FREQ:
49 		*value = adreno_gpu->base.fast_rate;
50 		return 0;
51 	case MSM_PARAM_TIMESTAMP:
52 		if (adreno_gpu->funcs->get_timestamp) {
53 			int ret;
54 
55 			pm_runtime_get_sync(&gpu->pdev->dev);
56 			ret = adreno_gpu->funcs->get_timestamp(gpu, value);
57 			pm_runtime_put_autosuspend(&gpu->pdev->dev);
58 
59 			return ret;
60 		}
61 		return -EINVAL;
62 	case MSM_PARAM_NR_RINGS:
63 		*value = gpu->nr_rings;
64 		return 0;
65 	default:
66 		DBG("%s: invalid param: %u", gpu->name, param);
67 		return -EINVAL;
68 	}
69 }
70 
71 const struct firmware *
72 adreno_request_fw(struct adreno_gpu *adreno_gpu, const char *fwname)
73 {
74 	struct drm_device *drm = adreno_gpu->base.dev;
75 	const struct firmware *fw = NULL;
76 	char *newname;
77 	int ret;
78 
79 	newname = kasprintf(GFP_KERNEL, "qcom/%s", fwname);
80 	if (!newname)
81 		return ERR_PTR(-ENOMEM);
82 
83 	/*
84 	 * Try first to load from qcom/$fwfile using a direct load (to avoid
85 	 * a potential timeout waiting for usermode helper)
86 	 */
87 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
88 	    (adreno_gpu->fwloc == FW_LOCATION_NEW)) {
89 
90 		ret = request_firmware_direct(&fw, newname, drm->dev);
91 		if (!ret) {
92 			DRM_DEV_INFO(drm->dev, "loaded %s from new location\n",
93 				newname);
94 			adreno_gpu->fwloc = FW_LOCATION_NEW;
95 			goto out;
96 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
97 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
98 				newname, ret);
99 			fw = ERR_PTR(ret);
100 			goto out;
101 		}
102 	}
103 
104 	/*
105 	 * Then try the legacy location without qcom/ prefix
106 	 */
107 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
108 	    (adreno_gpu->fwloc == FW_LOCATION_LEGACY)) {
109 
110 		ret = request_firmware_direct(&fw, fwname, drm->dev);
111 		if (!ret) {
112 			DRM_DEV_INFO(drm->dev, "loaded %s from legacy location\n",
113 				newname);
114 			adreno_gpu->fwloc = FW_LOCATION_LEGACY;
115 			goto out;
116 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
117 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
118 				fwname, ret);
119 			fw = ERR_PTR(ret);
120 			goto out;
121 		}
122 	}
123 
124 	/*
125 	 * Finally fall back to request_firmware() for cases where the
126 	 * usermode helper is needed (I think mainly android)
127 	 */
128 	if ((adreno_gpu->fwloc == FW_LOCATION_UNKNOWN) ||
129 	    (adreno_gpu->fwloc == FW_LOCATION_HELPER)) {
130 
131 		ret = request_firmware(&fw, newname, drm->dev);
132 		if (!ret) {
133 			DRM_DEV_INFO(drm->dev, "loaded %s with helper\n",
134 				newname);
135 			adreno_gpu->fwloc = FW_LOCATION_HELPER;
136 			goto out;
137 		} else if (adreno_gpu->fwloc != FW_LOCATION_UNKNOWN) {
138 			DRM_DEV_ERROR(drm->dev, "failed to load %s: %d\n",
139 				newname, ret);
140 			fw = ERR_PTR(ret);
141 			goto out;
142 		}
143 	}
144 
145 	DRM_DEV_ERROR(drm->dev, "failed to load %s\n", fwname);
146 	fw = ERR_PTR(-ENOENT);
147 out:
148 	kfree(newname);
149 	return fw;
150 }
151 
152 int adreno_load_fw(struct adreno_gpu *adreno_gpu)
153 {
154 	int i;
155 
156 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++) {
157 		const struct firmware *fw;
158 
159 		if (!adreno_gpu->info->fw[i])
160 			continue;
161 
162 		/* Skip if the firmware has already been loaded */
163 		if (adreno_gpu->fw[i])
164 			continue;
165 
166 		fw = adreno_request_fw(adreno_gpu, adreno_gpu->info->fw[i]);
167 		if (IS_ERR(fw))
168 			return PTR_ERR(fw);
169 
170 		adreno_gpu->fw[i] = fw;
171 	}
172 
173 	return 0;
174 }
175 
176 struct drm_gem_object *adreno_fw_create_bo(struct msm_gpu *gpu,
177 		const struct firmware *fw, u64 *iova)
178 {
179 	struct drm_gem_object *bo;
180 	void *ptr;
181 
182 	ptr = msm_gem_kernel_new_locked(gpu->dev, fw->size - 4,
183 		MSM_BO_UNCACHED | MSM_BO_GPU_READONLY, gpu->aspace, &bo, iova);
184 
185 	if (IS_ERR(ptr))
186 		return ERR_CAST(ptr);
187 
188 	memcpy(ptr, &fw->data[4], fw->size - 4);
189 
190 	msm_gem_put_vaddr(bo);
191 
192 	return bo;
193 }
194 
195 int adreno_hw_init(struct msm_gpu *gpu)
196 {
197 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
198 	int ret, i;
199 
200 	DBG("%s", gpu->name);
201 
202 	ret = adreno_load_fw(adreno_gpu);
203 	if (ret)
204 		return ret;
205 
206 	for (i = 0; i < gpu->nr_rings; i++) {
207 		struct msm_ringbuffer *ring = gpu->rb[i];
208 
209 		if (!ring)
210 			continue;
211 
212 		ring->cur = ring->start;
213 		ring->next = ring->start;
214 
215 		/* reset completed fence seqno: */
216 		ring->memptrs->fence = ring->seqno;
217 		ring->memptrs->rptr = 0;
218 	}
219 
220 	/*
221 	 * Setup REG_CP_RB_CNTL.  The same value is used across targets (with
222 	 * the excpetion of A430 that disables the RPTR shadow) - the cacluation
223 	 * for the ringbuffer size and block size is moved to msm_gpu.h for the
224 	 * pre-processor to deal with and the A430 variant is ORed in here
225 	 */
226 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_CNTL,
227 		MSM_GPU_RB_CNTL_DEFAULT |
228 		(adreno_is_a430(adreno_gpu) ? AXXX_CP_RB_CNTL_NO_UPDATE : 0));
229 
230 	/* Setup ringbuffer address - use ringbuffer[0] for GPU init */
231 	adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_BASE,
232 		REG_ADRENO_CP_RB_BASE_HI, gpu->rb[0]->iova);
233 
234 	if (!adreno_is_a430(adreno_gpu)) {
235 		adreno_gpu_write64(adreno_gpu, REG_ADRENO_CP_RB_RPTR_ADDR,
236 			REG_ADRENO_CP_RB_RPTR_ADDR_HI,
237 			rbmemptr(gpu->rb[0], rptr));
238 	}
239 
240 	return 0;
241 }
242 
243 /* Use this helper to read rptr, since a430 doesn't update rptr in memory */
244 static uint32_t get_rptr(struct adreno_gpu *adreno_gpu,
245 		struct msm_ringbuffer *ring)
246 {
247 	if (adreno_is_a430(adreno_gpu))
248 		return ring->memptrs->rptr = adreno_gpu_read(
249 			adreno_gpu, REG_ADRENO_CP_RB_RPTR);
250 	else
251 		return ring->memptrs->rptr;
252 }
253 
254 struct msm_ringbuffer *adreno_active_ring(struct msm_gpu *gpu)
255 {
256 	return gpu->rb[0];
257 }
258 
259 void adreno_recover(struct msm_gpu *gpu)
260 {
261 	struct drm_device *dev = gpu->dev;
262 	int ret;
263 
264 	// XXX pm-runtime??  we *need* the device to be off after this
265 	// so maybe continuing to call ->pm_suspend/resume() is better?
266 
267 	gpu->funcs->pm_suspend(gpu);
268 	gpu->funcs->pm_resume(gpu);
269 
270 	ret = msm_gpu_hw_init(gpu);
271 	if (ret) {
272 		DRM_DEV_ERROR(dev->dev, "gpu hw init failed: %d\n", ret);
273 		/* hmm, oh well? */
274 	}
275 }
276 
277 void adreno_submit(struct msm_gpu *gpu, struct msm_gem_submit *submit,
278 		struct msm_file_private *ctx)
279 {
280 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
281 	struct msm_drm_private *priv = gpu->dev->dev_private;
282 	struct msm_ringbuffer *ring = submit->ring;
283 	unsigned i;
284 
285 	for (i = 0; i < submit->nr_cmds; i++) {
286 		switch (submit->cmd[i].type) {
287 		case MSM_SUBMIT_CMD_IB_TARGET_BUF:
288 			/* ignore IB-targets */
289 			break;
290 		case MSM_SUBMIT_CMD_CTX_RESTORE_BUF:
291 			/* ignore if there has not been a ctx switch: */
292 			if (priv->lastctx == ctx)
293 				break;
294 		case MSM_SUBMIT_CMD_BUF:
295 			OUT_PKT3(ring, adreno_is_a430(adreno_gpu) ?
296 				CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
297 			OUT_RING(ring, lower_32_bits(submit->cmd[i].iova));
298 			OUT_RING(ring, submit->cmd[i].size);
299 			OUT_PKT2(ring);
300 			break;
301 		}
302 	}
303 
304 	OUT_PKT0(ring, REG_AXXX_CP_SCRATCH_REG2, 1);
305 	OUT_RING(ring, submit->seqno);
306 
307 	if (adreno_is_a3xx(adreno_gpu) || adreno_is_a4xx(adreno_gpu)) {
308 		/* Flush HLSQ lazy updates to make sure there is nothing
309 		 * pending for indirect loads after the timestamp has
310 		 * passed:
311 		 */
312 		OUT_PKT3(ring, CP_EVENT_WRITE, 1);
313 		OUT_RING(ring, HLSQ_FLUSH);
314 	}
315 
316 	/* wait for idle before cache flush/interrupt */
317 	OUT_PKT3(ring, CP_WAIT_FOR_IDLE, 1);
318 	OUT_RING(ring, 0x00000000);
319 
320 	if (!adreno_is_a2xx(adreno_gpu)) {
321 		/* BIT(31) of CACHE_FLUSH_TS triggers CACHE_FLUSH_TS IRQ from GPU */
322 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
323 		OUT_RING(ring, CACHE_FLUSH_TS | BIT(31));
324 		OUT_RING(ring, rbmemptr(ring, fence));
325 		OUT_RING(ring, submit->seqno);
326 	} else {
327 		/* BIT(31) means something else on a2xx */
328 		OUT_PKT3(ring, CP_EVENT_WRITE, 3);
329 		OUT_RING(ring, CACHE_FLUSH_TS);
330 		OUT_RING(ring, rbmemptr(ring, fence));
331 		OUT_RING(ring, submit->seqno);
332 		OUT_PKT3(ring, CP_INTERRUPT, 1);
333 		OUT_RING(ring, 0x80000000);
334 	}
335 
336 #if 0
337 	if (adreno_is_a3xx(adreno_gpu)) {
338 		/* Dummy set-constant to trigger context rollover */
339 		OUT_PKT3(ring, CP_SET_CONSTANT, 2);
340 		OUT_RING(ring, CP_REG(REG_A3XX_HLSQ_CL_KERNEL_GROUP_X_REG));
341 		OUT_RING(ring, 0x00000000);
342 	}
343 #endif
344 
345 	gpu->funcs->flush(gpu, ring);
346 }
347 
348 void adreno_flush(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
349 {
350 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
351 	uint32_t wptr;
352 
353 	/* Copy the shadow to the actual register */
354 	ring->cur = ring->next;
355 
356 	/*
357 	 * Mask wptr value that we calculate to fit in the HW range. This is
358 	 * to account for the possibility that the last command fit exactly into
359 	 * the ringbuffer and rb->next hasn't wrapped to zero yet
360 	 */
361 	wptr = get_wptr(ring);
362 
363 	/* ensure writes to ringbuffer have hit system memory: */
364 	mb();
365 
366 	adreno_gpu_write(adreno_gpu, REG_ADRENO_CP_RB_WPTR, wptr);
367 }
368 
369 bool adreno_idle(struct msm_gpu *gpu, struct msm_ringbuffer *ring)
370 {
371 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
372 	uint32_t wptr = get_wptr(ring);
373 
374 	/* wait for CP to drain ringbuffer: */
375 	if (!spin_until(get_rptr(adreno_gpu, ring) == wptr))
376 		return true;
377 
378 	/* TODO maybe we need to reset GPU here to recover from hang? */
379 	DRM_ERROR("%s: timeout waiting to drain ringbuffer %d rptr/wptr = %X/%X\n",
380 		gpu->name, ring->id, get_rptr(adreno_gpu, ring), wptr);
381 
382 	return false;
383 }
384 
385 int adreno_gpu_state_get(struct msm_gpu *gpu, struct msm_gpu_state *state)
386 {
387 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
388 	int i, count = 0;
389 
390 	kref_init(&state->ref);
391 
392 	ktime_get_real_ts64(&state->time);
393 
394 	for (i = 0; i < gpu->nr_rings; i++) {
395 		int size = 0, j;
396 
397 		state->ring[i].fence = gpu->rb[i]->memptrs->fence;
398 		state->ring[i].iova = gpu->rb[i]->iova;
399 		state->ring[i].seqno = gpu->rb[i]->seqno;
400 		state->ring[i].rptr = get_rptr(adreno_gpu, gpu->rb[i]);
401 		state->ring[i].wptr = get_wptr(gpu->rb[i]);
402 
403 		/* Copy at least 'wptr' dwords of the data */
404 		size = state->ring[i].wptr;
405 
406 		/* After wptr find the last non zero dword to save space */
407 		for (j = state->ring[i].wptr; j < MSM_GPU_RINGBUFFER_SZ >> 2; j++)
408 			if (gpu->rb[i]->start[j])
409 				size = j + 1;
410 
411 		if (size) {
412 			state->ring[i].data = kvmalloc(size << 2, GFP_KERNEL);
413 			if (state->ring[i].data) {
414 				memcpy(state->ring[i].data, gpu->rb[i]->start, size << 2);
415 				state->ring[i].data_size = size << 2;
416 			}
417 		}
418 	}
419 
420 	/* Some targets prefer to collect their own registers */
421 	if (!adreno_gpu->registers)
422 		return 0;
423 
424 	/* Count the number of registers */
425 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2)
426 		count += adreno_gpu->registers[i + 1] -
427 			adreno_gpu->registers[i] + 1;
428 
429 	state->registers = kcalloc(count * 2, sizeof(u32), GFP_KERNEL);
430 	if (state->registers) {
431 		int pos = 0;
432 
433 		for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
434 			u32 start = adreno_gpu->registers[i];
435 			u32 end   = adreno_gpu->registers[i + 1];
436 			u32 addr;
437 
438 			for (addr = start; addr <= end; addr++) {
439 				state->registers[pos++] = addr;
440 				state->registers[pos++] = gpu_read(gpu, addr);
441 			}
442 		}
443 
444 		state->nr_registers = count;
445 	}
446 
447 	return 0;
448 }
449 
450 void adreno_gpu_state_destroy(struct msm_gpu_state *state)
451 {
452 	int i;
453 
454 	for (i = 0; i < ARRAY_SIZE(state->ring); i++)
455 		kvfree(state->ring[i].data);
456 
457 	for (i = 0; state->bos && i < state->nr_bos; i++)
458 		kvfree(state->bos[i].data);
459 
460 	kfree(state->bos);
461 	kfree(state->comm);
462 	kfree(state->cmd);
463 	kfree(state->registers);
464 }
465 
466 static void adreno_gpu_state_kref_destroy(struct kref *kref)
467 {
468 	struct msm_gpu_state *state = container_of(kref,
469 		struct msm_gpu_state, ref);
470 
471 	adreno_gpu_state_destroy(state);
472 	kfree(state);
473 }
474 
475 int adreno_gpu_state_put(struct msm_gpu_state *state)
476 {
477 	if (IS_ERR_OR_NULL(state))
478 		return 1;
479 
480 	return kref_put(&state->ref, adreno_gpu_state_kref_destroy);
481 }
482 
483 #if defined(CONFIG_DEBUG_FS) || defined(CONFIG_DEV_COREDUMP)
484 
485 static char *adreno_gpu_ascii85_encode(u32 *src, size_t len)
486 {
487 	void *buf;
488 	size_t buf_itr = 0, buffer_size;
489 	char out[ASCII85_BUFSZ];
490 	long l;
491 	int i;
492 
493 	if (!src || !len)
494 		return NULL;
495 
496 	l = ascii85_encode_len(len);
497 
498 	/*
499 	 * Ascii85 outputs either a 5 byte string or a 1 byte string. So we
500 	 * account for the worst case of 5 bytes per dword plus the 1 for '\0'
501 	 */
502 	buffer_size = (l * 5) + 1;
503 
504 	buf = kvmalloc(buffer_size, GFP_KERNEL);
505 	if (!buf)
506 		return NULL;
507 
508 	for (i = 0; i < l; i++)
509 		buf_itr += snprintf(buf + buf_itr, buffer_size - buf_itr, "%s",
510 				ascii85_encode(src[i], out));
511 
512 	return buf;
513 }
514 
515 /* len is expected to be in bytes */
516 static void adreno_show_object(struct drm_printer *p, void **ptr, int len,
517 		bool *encoded)
518 {
519 	if (!*ptr || !len)
520 		return;
521 
522 	if (!*encoded) {
523 		long datalen, i;
524 		u32 *buf = *ptr;
525 
526 		/*
527 		 * Only dump the non-zero part of the buffer - rarely will
528 		 * any data completely fill the entire allocated size of
529 		 * the buffer.
530 		 */
531 		for (datalen = 0, i = 0; i < len >> 2; i++)
532 			if (buf[i])
533 				datalen = ((i + 1) << 2);
534 
535 		/*
536 		 * If we reach here, then the originally captured binary buffer
537 		 * will be replaced with the ascii85 encoded string
538 		 */
539 		*ptr = adreno_gpu_ascii85_encode(buf, datalen);
540 
541 		kvfree(buf);
542 
543 		*encoded = true;
544 	}
545 
546 	if (!*ptr)
547 		return;
548 
549 	drm_puts(p, "    data: !!ascii85 |\n");
550 	drm_puts(p, "     ");
551 
552 	drm_puts(p, *ptr);
553 
554 	drm_puts(p, "\n");
555 }
556 
557 void adreno_show(struct msm_gpu *gpu, struct msm_gpu_state *state,
558 		struct drm_printer *p)
559 {
560 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
561 	int i;
562 
563 	if (IS_ERR_OR_NULL(state))
564 		return;
565 
566 	drm_printf(p, "revision: %d (%d.%d.%d.%d)\n",
567 			adreno_gpu->info->revn, adreno_gpu->rev.core,
568 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
569 			adreno_gpu->rev.patchid);
570 
571 	drm_printf(p, "rbbm-status: 0x%08x\n", state->rbbm_status);
572 
573 	drm_puts(p, "ringbuffer:\n");
574 
575 	for (i = 0; i < gpu->nr_rings; i++) {
576 		drm_printf(p, "  - id: %d\n", i);
577 		drm_printf(p, "    iova: 0x%016llx\n", state->ring[i].iova);
578 		drm_printf(p, "    last-fence: %d\n", state->ring[i].seqno);
579 		drm_printf(p, "    retired-fence: %d\n", state->ring[i].fence);
580 		drm_printf(p, "    rptr: %d\n", state->ring[i].rptr);
581 		drm_printf(p, "    wptr: %d\n", state->ring[i].wptr);
582 		drm_printf(p, "    size: %d\n", MSM_GPU_RINGBUFFER_SZ);
583 
584 		adreno_show_object(p, &state->ring[i].data,
585 			state->ring[i].data_size, &state->ring[i].encoded);
586 	}
587 
588 	if (state->bos) {
589 		drm_puts(p, "bos:\n");
590 
591 		for (i = 0; i < state->nr_bos; i++) {
592 			drm_printf(p, "  - iova: 0x%016llx\n",
593 				state->bos[i].iova);
594 			drm_printf(p, "    size: %zd\n", state->bos[i].size);
595 
596 			adreno_show_object(p, &state->bos[i].data,
597 				state->bos[i].size, &state->bos[i].encoded);
598 		}
599 	}
600 
601 	if (state->nr_registers) {
602 		drm_puts(p, "registers:\n");
603 
604 		for (i = 0; i < state->nr_registers; i++) {
605 			drm_printf(p, "  - { offset: 0x%04x, value: 0x%08x }\n",
606 				state->registers[i * 2] << 2,
607 				state->registers[(i * 2) + 1]);
608 		}
609 	}
610 }
611 #endif
612 
613 /* Dump common gpu status and scratch registers on any hang, to make
614  * the hangcheck logs more useful.  The scratch registers seem always
615  * safe to read when GPU has hung (unlike some other regs, depending
616  * on how the GPU hung), and they are useful to match up to cmdstream
617  * dumps when debugging hangs:
618  */
619 void adreno_dump_info(struct msm_gpu *gpu)
620 {
621 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
622 	int i;
623 
624 	printk("revision: %d (%d.%d.%d.%d)\n",
625 			adreno_gpu->info->revn, adreno_gpu->rev.core,
626 			adreno_gpu->rev.major, adreno_gpu->rev.minor,
627 			adreno_gpu->rev.patchid);
628 
629 	for (i = 0; i < gpu->nr_rings; i++) {
630 		struct msm_ringbuffer *ring = gpu->rb[i];
631 
632 		printk("rb %d: fence:    %d/%d\n", i,
633 			ring->memptrs->fence,
634 			ring->seqno);
635 
636 		printk("rptr:     %d\n", get_rptr(adreno_gpu, ring));
637 		printk("rb wptr:  %d\n", get_wptr(ring));
638 	}
639 }
640 
641 /* would be nice to not have to duplicate the _show() stuff with printk(): */
642 void adreno_dump(struct msm_gpu *gpu)
643 {
644 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(gpu);
645 	int i;
646 
647 	if (!adreno_gpu->registers)
648 		return;
649 
650 	/* dump these out in a form that can be parsed by demsm: */
651 	printk("IO:region %s 00000000 00020000\n", gpu->name);
652 	for (i = 0; adreno_gpu->registers[i] != ~0; i += 2) {
653 		uint32_t start = adreno_gpu->registers[i];
654 		uint32_t end   = adreno_gpu->registers[i+1];
655 		uint32_t addr;
656 
657 		for (addr = start; addr <= end; addr++) {
658 			uint32_t val = gpu_read(gpu, addr);
659 			printk("IO:R %08x %08x\n", addr<<2, val);
660 		}
661 	}
662 }
663 
664 static uint32_t ring_freewords(struct msm_ringbuffer *ring)
665 {
666 	struct adreno_gpu *adreno_gpu = to_adreno_gpu(ring->gpu);
667 	uint32_t size = MSM_GPU_RINGBUFFER_SZ >> 2;
668 	/* Use ring->next to calculate free size */
669 	uint32_t wptr = ring->next - ring->start;
670 	uint32_t rptr = get_rptr(adreno_gpu, ring);
671 	return (rptr + (size - 1) - wptr) % size;
672 }
673 
674 void adreno_wait_ring(struct msm_ringbuffer *ring, uint32_t ndwords)
675 {
676 	if (spin_until(ring_freewords(ring) >= ndwords))
677 		DRM_DEV_ERROR(ring->gpu->dev->dev,
678 			"timeout waiting for space in ringbuffer %d\n",
679 			ring->id);
680 }
681 
682 /* Get legacy powerlevels from qcom,gpu-pwrlevels and populate the opp table */
683 static int adreno_get_legacy_pwrlevels(struct device *dev)
684 {
685 	struct device_node *child, *node;
686 	int ret;
687 
688 	node = of_get_compatible_child(dev->of_node, "qcom,gpu-pwrlevels");
689 	if (!node) {
690 		DRM_DEV_ERROR(dev, "Could not find the GPU powerlevels\n");
691 		return -ENXIO;
692 	}
693 
694 	for_each_child_of_node(node, child) {
695 		unsigned int val;
696 
697 		ret = of_property_read_u32(child, "qcom,gpu-freq", &val);
698 		if (ret)
699 			continue;
700 
701 		/*
702 		 * Skip the intentionally bogus clock value found at the bottom
703 		 * of most legacy frequency tables
704 		 */
705 		if (val != 27000000)
706 			dev_pm_opp_add(dev, val, 0);
707 	}
708 
709 	of_node_put(node);
710 
711 	return 0;
712 }
713 
714 static int adreno_get_pwrlevels(struct device *dev,
715 		struct msm_gpu *gpu)
716 {
717 	unsigned long freq = ULONG_MAX;
718 	struct dev_pm_opp *opp;
719 	int ret;
720 
721 	gpu->fast_rate = 0;
722 
723 	/* You down with OPP? */
724 	if (!of_find_property(dev->of_node, "operating-points-v2", NULL))
725 		ret = adreno_get_legacy_pwrlevels(dev);
726 	else {
727 		ret = dev_pm_opp_of_add_table(dev);
728 		if (ret)
729 			DRM_DEV_ERROR(dev, "Unable to set the OPP table\n");
730 	}
731 
732 	if (!ret) {
733 		/* Find the fastest defined rate */
734 		opp = dev_pm_opp_find_freq_floor(dev, &freq);
735 		if (!IS_ERR(opp)) {
736 			gpu->fast_rate = freq;
737 			dev_pm_opp_put(opp);
738 		}
739 	}
740 
741 	if (!gpu->fast_rate) {
742 		dev_warn(dev,
743 			"Could not find a clock rate. Using a reasonable default\n");
744 		/* Pick a suitably safe clock speed for any target */
745 		gpu->fast_rate = 200000000;
746 	}
747 
748 	DBG("fast_rate=%u, slow_rate=27000000", gpu->fast_rate);
749 
750 	return 0;
751 }
752 
753 int adreno_gpu_init(struct drm_device *drm, struct platform_device *pdev,
754 		struct adreno_gpu *adreno_gpu,
755 		const struct adreno_gpu_funcs *funcs, int nr_rings)
756 {
757 	struct adreno_platform_config *config = pdev->dev.platform_data;
758 	struct msm_gpu_config adreno_gpu_config  = { 0 };
759 	struct msm_gpu *gpu = &adreno_gpu->base;
760 
761 	adreno_gpu->funcs = funcs;
762 	adreno_gpu->info = adreno_info(config->rev);
763 	adreno_gpu->gmem = adreno_gpu->info->gmem;
764 	adreno_gpu->revn = adreno_gpu->info->revn;
765 	adreno_gpu->rev = config->rev;
766 
767 	adreno_gpu_config.ioname = "kgsl_3d0_reg_memory";
768 	adreno_gpu_config.irqname = "kgsl_3d0_irq";
769 
770 	adreno_gpu_config.va_start = SZ_16M;
771 	adreno_gpu_config.va_end = 0xffffffff;
772 	/* maximum range of a2xx mmu */
773 	if (adreno_is_a2xx(adreno_gpu))
774 		adreno_gpu_config.va_end = SZ_16M + 0xfff * SZ_64K;
775 
776 	adreno_gpu_config.nr_rings = nr_rings;
777 
778 	adreno_get_pwrlevels(&pdev->dev, gpu);
779 
780 	pm_runtime_set_autosuspend_delay(&pdev->dev,
781 		adreno_gpu->info->inactive_period);
782 	pm_runtime_use_autosuspend(&pdev->dev);
783 	pm_runtime_enable(&pdev->dev);
784 
785 	return msm_gpu_init(drm, pdev, &adreno_gpu->base, &funcs->base,
786 			adreno_gpu->info->name, &adreno_gpu_config);
787 }
788 
789 void adreno_gpu_cleanup(struct adreno_gpu *adreno_gpu)
790 {
791 	unsigned int i;
792 
793 	for (i = 0; i < ARRAY_SIZE(adreno_gpu->info->fw); i++)
794 		release_firmware(adreno_gpu->fw[i]);
795 
796 	msm_gpu_cleanup(&adreno_gpu->base);
797 }
798