xref: /linux/drivers/gpu/drm/etnaviv/etnaviv_buffer.c (revision 815e260a18a3af4dab59025ee99a7156c0e8b5e0)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2014-2018 Etnaviv Project
4  */
5 
6 #include <drm/drm_drv.h>
7 #include <drm/drm_print.h>
8 
9 #include "etnaviv_cmdbuf.h"
10 #include "etnaviv_gpu.h"
11 #include "etnaviv_gem.h"
12 #include "etnaviv_mmu.h"
13 
14 #include "common.xml.h"
15 #include "state.xml.h"
16 #include "state_blt.xml.h"
17 #include "state_hi.xml.h"
18 #include "state_3d.xml.h"
19 #include "cmdstream.xml.h"
20 
21 /*
22  * Command Buffer helper:
23  */
24 
25 
26 static inline void OUT(struct etnaviv_cmdbuf *buffer, u32 data)
27 {
28 	u32 *vaddr = (u32 *)buffer->vaddr;
29 
30 	BUG_ON(buffer->user_size >= buffer->size);
31 
32 	vaddr[buffer->user_size / 4] = data;
33 	buffer->user_size += 4;
34 }
35 
36 static inline void CMD_LOAD_STATE(struct etnaviv_cmdbuf *buffer,
37 	u32 reg, u32 value)
38 {
39 	u32 index = reg >> VIV_FE_LOAD_STATE_HEADER_OFFSET__SHR;
40 
41 	buffer->user_size = ALIGN(buffer->user_size, 8);
42 
43 	/* write a register via cmd stream */
44 	OUT(buffer, VIV_FE_LOAD_STATE_HEADER_OP_LOAD_STATE |
45 		    VIV_FE_LOAD_STATE_HEADER_COUNT(1) |
46 		    VIV_FE_LOAD_STATE_HEADER_OFFSET(index));
47 	OUT(buffer, value);
48 }
49 
50 static inline void CMD_END(struct etnaviv_cmdbuf *buffer)
51 {
52 	buffer->user_size = ALIGN(buffer->user_size, 8);
53 
54 	OUT(buffer, VIV_FE_END_HEADER_OP_END);
55 }
56 
57 static inline void CMD_WAIT(struct etnaviv_cmdbuf *buffer,
58 			    unsigned int waitcycles)
59 {
60 	buffer->user_size = ALIGN(buffer->user_size, 8);
61 
62 	OUT(buffer, VIV_FE_WAIT_HEADER_OP_WAIT | waitcycles);
63 }
64 
65 static inline void CMD_LINK(struct etnaviv_cmdbuf *buffer,
66 	u16 prefetch, u32 address)
67 {
68 	buffer->user_size = ALIGN(buffer->user_size, 8);
69 
70 	OUT(buffer, VIV_FE_LINK_HEADER_OP_LINK |
71 		    VIV_FE_LINK_HEADER_PREFETCH(prefetch));
72 	OUT(buffer, address);
73 }
74 
75 static inline void CMD_STALL(struct etnaviv_cmdbuf *buffer,
76 	u32 from, u32 to)
77 {
78 	buffer->user_size = ALIGN(buffer->user_size, 8);
79 
80 	OUT(buffer, VIV_FE_STALL_HEADER_OP_STALL);
81 	OUT(buffer, VIV_FE_STALL_TOKEN_FROM(from) | VIV_FE_STALL_TOKEN_TO(to));
82 }
83 
84 static inline void CMD_SEM(struct etnaviv_cmdbuf *buffer, u32 from, u32 to)
85 {
86 	CMD_LOAD_STATE(buffer, VIVS_GL_SEMAPHORE_TOKEN,
87 		       VIVS_GL_SEMAPHORE_TOKEN_FROM(from) |
88 		       VIVS_GL_SEMAPHORE_TOKEN_TO(to));
89 }
90 
91 static void etnaviv_cmd_select_pipe(struct etnaviv_gpu *gpu,
92 	struct etnaviv_cmdbuf *buffer, u8 pipe)
93 {
94 	u32 flush = 0;
95 
96 	lockdep_assert_held(&gpu->lock);
97 
98 	/*
99 	 * This assumes that if we're switching to 2D, we're switching
100 	 * away from 3D, and vice versa.  Hence, if we're switching to
101 	 * the 2D core, we need to flush the 3D depth and color caches,
102 	 * otherwise we need to flush the 2D pixel engine cache.
103 	 */
104 	if (gpu->exec_state == ETNA_PIPE_2D)
105 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
106 	else if (gpu->exec_state == ETNA_PIPE_3D)
107 		flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
108 
109 	CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
110 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
111 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
112 
113 	CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
114 		       VIVS_GL_PIPE_SELECT_PIPE(pipe));
115 }
116 
117 static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
118 	struct etnaviv_cmdbuf *buf, u32 off, u32 len)
119 {
120 	u32 size = buf->size;
121 	u32 *ptr = buf->vaddr + off;
122 
123 	dev_info(gpu->dev, "virt %p phys 0x%08x free 0x%08x\n",
124 			ptr, etnaviv_cmdbuf_get_va(buf,
125 			&gpu->mmu_context->cmdbuf_mapping) +
126 			off, size - len * 4 - off);
127 
128 	print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
129 			ptr, len * 4, 0);
130 }
131 
132 /*
133  * Safely replace the WAIT of a waitlink with a new command and argument.
134  * The GPU may be executing this WAIT while we're modifying it, so we have
135  * to write it in a specific order to avoid the GPU branching to somewhere
136  * else.  'wl_offset' is the offset to the first byte of the WAIT command.
137  */
138 static void etnaviv_buffer_replace_wait(struct etnaviv_cmdbuf *buffer,
139 	unsigned int wl_offset, u32 cmd, u32 arg)
140 {
141 	u32 *lw = buffer->vaddr + wl_offset;
142 
143 	lw[1] = arg;
144 	mb();
145 	lw[0] = cmd;
146 	mb();
147 }
148 
149 /*
150  * Ensure that there is space in the command buffer to contiguously write
151  * 'cmd_dwords' 64-bit words into the buffer, wrapping if necessary.
152  */
153 static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
154 	struct etnaviv_cmdbuf *buffer, unsigned int cmd_dwords)
155 {
156 	if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
157 		buffer->user_size = 0;
158 
159 	return etnaviv_cmdbuf_get_va(buffer,
160 				     &gpu->mmu_context->cmdbuf_mapping) +
161 	       buffer->user_size;
162 }
163 
164 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
165 {
166 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
167 
168 	lockdep_assert_held(&gpu->lock);
169 
170 	/* initialize buffer */
171 	buffer->user_size = 0;
172 
173 	CMD_WAIT(buffer, gpu->fe_waitcycles);
174 	CMD_LINK(buffer, 2,
175 		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
176 		 + buffer->user_size - 4);
177 
178 	return buffer->user_size / 8;
179 }
180 
181 u16 etnaviv_buffer_config_mmuv2(struct etnaviv_gpu *gpu, u32 mtlb_addr, u32 safe_addr)
182 {
183 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
184 
185 	lockdep_assert_held(&gpu->lock);
186 
187 	buffer->user_size = 0;
188 
189 	if (gpu->identity.features & chipFeatures_PIPE_3D) {
190 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
191 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_3D));
192 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
193 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
194 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
195 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
196 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
197 	}
198 
199 	if (gpu->identity.features & chipFeatures_PIPE_2D) {
200 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
201 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_2D));
202 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
203 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
204 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
205 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
206 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
207 	}
208 
209 	CMD_END(buffer);
210 
211 	buffer->user_size = ALIGN(buffer->user_size, 8);
212 
213 	return buffer->user_size / 8;
214 }
215 
216 u16 etnaviv_buffer_config_pta(struct etnaviv_gpu *gpu, unsigned short id)
217 {
218 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
219 
220 	lockdep_assert_held(&gpu->lock);
221 
222 	buffer->user_size = 0;
223 
224 	CMD_LOAD_STATE(buffer, VIVS_MMUv2_PTA_CONFIG,
225 		       VIVS_MMUv2_PTA_CONFIG_INDEX(id));
226 
227 	CMD_END(buffer);
228 
229 	buffer->user_size = ALIGN(buffer->user_size, 8);
230 
231 	return buffer->user_size / 8;
232 }
233 
234 void etnaviv_buffer_end(struct etnaviv_gpu *gpu)
235 {
236 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
237 	unsigned int waitlink_offset = buffer->user_size - 16;
238 	u32 link_target, flush = 0;
239 	bool has_blt = !!(gpu->identity.minor_features5 &
240 			  chipMinorFeatures5_BLT_ENGINE);
241 
242 	lockdep_assert_held(&gpu->lock);
243 
244 	if (gpu->exec_state == ETNA_PIPE_2D)
245 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
246 	else if (gpu->exec_state == ETNA_PIPE_3D)
247 		flush = VIVS_GL_FLUSH_CACHE_DEPTH |
248 			VIVS_GL_FLUSH_CACHE_COLOR |
249 			VIVS_GL_FLUSH_CACHE_TEXTURE |
250 			VIVS_GL_FLUSH_CACHE_TEXTUREVS |
251 			VIVS_GL_FLUSH_CACHE_SHADER_L2;
252 
253 	if (flush) {
254 		unsigned int dwords = 7;
255 
256 		if (has_blt)
257 			dwords += 10;
258 
259 		link_target = etnaviv_buffer_reserve(gpu, buffer, dwords);
260 
261 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
262 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
263 		if (has_blt) {
264 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x1);
265 			CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
266 			CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
267 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x0);
268 		}
269 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
270 		if (gpu->exec_state == ETNA_PIPE_3D) {
271 			if (has_blt) {
272 				CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x1);
273 				CMD_LOAD_STATE(buffer, VIVS_BLT_SET_COMMAND, 0x1);
274 				CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x0);
275 			} else {
276 				CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
277 					       VIVS_TS_FLUSH_CACHE_FLUSH);
278 			}
279 		}
280 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
281 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
282 		if (has_blt) {
283 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x1);
284 			CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
285 			CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
286 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x0);
287 		}
288 		CMD_END(buffer);
289 
290 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
291 					    VIV_FE_LINK_HEADER_OP_LINK |
292 					    VIV_FE_LINK_HEADER_PREFETCH(dwords),
293 					    link_target);
294 	} else {
295 		/* Replace the last link-wait with an "END" command */
296 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
297 					    VIV_FE_END_HEADER_OP_END, 0);
298 	}
299 }
300 
301 /* Append a 'sync point' to the ring buffer. */
302 void etnaviv_sync_point_queue(struct etnaviv_gpu *gpu, unsigned int event)
303 {
304 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
305 	unsigned int waitlink_offset = buffer->user_size - 16;
306 	u32 dwords, target;
307 
308 	lockdep_assert_held(&gpu->lock);
309 
310 	/*
311 	 * We need at most 3 dwords in the return target:
312 	 * 1 event + 1 end + 1 wait + 1 link.
313 	 */
314 	dwords = 4;
315 	target = etnaviv_buffer_reserve(gpu, buffer, dwords);
316 
317 	/* Signal sync point event */
318 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
319 		       VIVS_GL_EVENT_FROM_PE);
320 
321 	/* Stop the FE to 'pause' the GPU */
322 	CMD_END(buffer);
323 
324 	/* Append waitlink */
325 	CMD_WAIT(buffer, gpu->fe_waitcycles);
326 	CMD_LINK(buffer, 2,
327 		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
328 		 + buffer->user_size - 4);
329 
330 	/*
331 	 * Kick off the 'sync point' command by replacing the previous
332 	 * WAIT with a link to the address in the ring buffer.
333 	 */
334 	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
335 				    VIV_FE_LINK_HEADER_OP_LINK |
336 				    VIV_FE_LINK_HEADER_PREFETCH(dwords),
337 				    target);
338 }
339 
340 /* Append a command buffer to the ring buffer. */
341 void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, u32 exec_state,
342 	struct etnaviv_iommu_context *mmu_context, unsigned int event,
343 	struct etnaviv_cmdbuf *cmdbuf)
344 {
345 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
346 	unsigned int waitlink_offset = buffer->user_size - 16;
347 	u32 return_target, return_dwords;
348 	u32 link_target, link_dwords;
349 	bool switch_context = gpu->exec_state != exec_state;
350 	bool switch_mmu_context = gpu->mmu_context != mmu_context;
351 	unsigned int new_flush_seq = READ_ONCE(gpu->mmu_context->flush_seq);
352 	bool need_flush = switch_mmu_context || gpu->flush_seq != new_flush_seq;
353 	bool has_blt = !!(gpu->identity.minor_features5 &
354 			  chipMinorFeatures5_BLT_ENGINE);
355 
356 	lockdep_assert_held(&gpu->lock);
357 
358 	if (drm_debug_enabled(DRM_UT_DRIVER))
359 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
360 
361 	link_target = etnaviv_cmdbuf_get_va(cmdbuf,
362 					    &gpu->mmu_context->cmdbuf_mapping);
363 	link_dwords = cmdbuf->size / 8;
364 
365 	/*
366 	 * If we need maintenance prior to submitting this buffer, we will
367 	 * need to append a mmu flush load state, followed by a new
368 	 * link to this buffer - a total of four additional words.
369 	 */
370 	if (need_flush || switch_context) {
371 		u32 target, extra_dwords;
372 
373 		/* link command */
374 		extra_dwords = 1;
375 
376 		/* flush command */
377 		if (need_flush) {
378 			if (gpu->mmu_context->global->version == ETNAVIV_IOMMU_V1)
379 				extra_dwords += 1;
380 			else
381 				extra_dwords += 3;
382 		}
383 
384 		/* pipe switch commands */
385 		if (switch_context)
386 			extra_dwords += 4;
387 
388 		/* PTA load command */
389 		if (switch_mmu_context && gpu->sec_mode == ETNA_SEC_KERNEL)
390 			extra_dwords += 1;
391 
392 		target = etnaviv_buffer_reserve(gpu, buffer, extra_dwords);
393 		/*
394 		 * Switch MMU context if necessary. Must be done after the
395 		 * link target has been calculated, as the jump forward in the
396 		 * kernel ring still uses the last active MMU context before
397 		 * the switch.
398 		 */
399 		if (switch_mmu_context) {
400 			struct etnaviv_iommu_context *old_context = gpu->mmu_context;
401 
402 			gpu->mmu_context = etnaviv_iommu_context_get(mmu_context);
403 			etnaviv_iommu_context_put(old_context);
404 		}
405 
406 		if (need_flush) {
407 			/* Add the MMU flush */
408 			if (gpu->mmu_context->global->version == ETNAVIV_IOMMU_V1) {
409 				CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_MMU,
410 					       VIVS_GL_FLUSH_MMU_FLUSH_FEMMU |
411 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK1 |
412 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK2 |
413 					       VIVS_GL_FLUSH_MMU_FLUSH_PEMMU |
414 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK4);
415 			} else {
416 				u32 flush = VIVS_MMUv2_CONFIGURATION_MODE_MASK |
417 					    VIVS_MMUv2_CONFIGURATION_FLUSH_FLUSH;
418 
419 				if (switch_mmu_context &&
420 				    gpu->sec_mode == ETNA_SEC_KERNEL) {
421 					unsigned short id =
422 						etnaviv_iommuv2_get_pta_id(gpu->mmu_context);
423 					CMD_LOAD_STATE(buffer,
424 						VIVS_MMUv2_PTA_CONFIG,
425 						VIVS_MMUv2_PTA_CONFIG_INDEX(id));
426 				}
427 
428 				if (gpu->sec_mode == ETNA_SEC_NONE)
429 					flush |= etnaviv_iommuv2_get_mtlb_addr(gpu->mmu_context);
430 
431 				CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
432 					       flush);
433 				CMD_SEM(buffer, SYNC_RECIPIENT_FE,
434 					SYNC_RECIPIENT_PE);
435 				CMD_STALL(buffer, SYNC_RECIPIENT_FE,
436 					SYNC_RECIPIENT_PE);
437 			}
438 
439 			gpu->flush_seq = new_flush_seq;
440 		}
441 
442 		if (switch_context) {
443 			etnaviv_cmd_select_pipe(gpu, buffer, exec_state);
444 			gpu->exec_state = exec_state;
445 		}
446 
447 		/* And the link to the submitted buffer */
448 		link_target = etnaviv_cmdbuf_get_va(cmdbuf,
449 					&gpu->mmu_context->cmdbuf_mapping);
450 		CMD_LINK(buffer, link_dwords, link_target);
451 
452 		/* Update the link target to point to above instructions */
453 		link_target = target;
454 		link_dwords = extra_dwords;
455 	}
456 
457 	/*
458 	 * Append a LINK to the submitted command buffer to return to
459 	 * the ring buffer.  return_target is the ring target address.
460 	 * We need at most 7 dwords in the return target: 2 cache flush +
461 	 * 2 semaphore stall + 1 event + 1 wait + 1 link.
462 	 */
463 	return_dwords = 7;
464 
465 	/*
466 	 * When the BLT engine is present we need 6 more dwords in the return
467 	 * target: 3 enable/flush/disable + 4 enable/semaphore stall/disable,
468 	 * but we don't need the normal TS flush state.
469 	 */
470 	if (has_blt)
471 		return_dwords += 6;
472 
473 	return_target = etnaviv_buffer_reserve(gpu, buffer, return_dwords);
474 	CMD_LINK(cmdbuf, return_dwords, return_target);
475 
476 	/*
477 	 * Append a cache flush, stall, event, wait and link pointing back to
478 	 * the wait command to the ring buffer.
479 	 */
480 	if (gpu->exec_state == ETNA_PIPE_2D) {
481 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
482 				       VIVS_GL_FLUSH_CACHE_PE2D);
483 	} else {
484 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
485 				       VIVS_GL_FLUSH_CACHE_DEPTH |
486 				       VIVS_GL_FLUSH_CACHE_COLOR |
487 				       VIVS_GL_FLUSH_CACHE_SHADER_L1);
488 		if (has_blt) {
489 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x1);
490 			CMD_LOAD_STATE(buffer, VIVS_BLT_SET_COMMAND, 0x1);
491 			CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x0);
492 		} else {
493 			CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
494 					       VIVS_TS_FLUSH_CACHE_FLUSH);
495 		}
496 	}
497 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
498 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
499 
500 	if (has_blt) {
501 		CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x1);
502 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
503 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_BLT);
504 		CMD_LOAD_STATE(buffer, VIVS_BLT_ENABLE, 0x0);
505 	}
506 
507 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
508 		       VIVS_GL_EVENT_FROM_PE);
509 	CMD_WAIT(buffer, gpu->fe_waitcycles);
510 	CMD_LINK(buffer, 2,
511 		 etnaviv_cmdbuf_get_va(buffer, &gpu->mmu_context->cmdbuf_mapping)
512 		 + buffer->user_size - 4);
513 
514 	if (drm_debug_enabled(DRM_UT_DRIVER))
515 		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
516 			return_target,
517 			etnaviv_cmdbuf_get_va(cmdbuf, &gpu->mmu_context->cmdbuf_mapping),
518 			cmdbuf->vaddr);
519 
520 	if (drm_debug_enabled(DRM_UT_DRIVER)) {
521 		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
522 			       cmdbuf->vaddr, cmdbuf->size, 0);
523 
524 		pr_info("link op: %p\n", buffer->vaddr + waitlink_offset);
525 		pr_info("addr: 0x%08x\n", link_target);
526 		pr_info("back: 0x%08x\n", return_target);
527 		pr_info("event: %d\n", event);
528 	}
529 
530 	/*
531 	 * Kick off the submitted command by replacing the previous
532 	 * WAIT with a link to the address in the ring buffer.
533 	 */
534 	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
535 				    VIV_FE_LINK_HEADER_OP_LINK |
536 				    VIV_FE_LINK_HEADER_PREFETCH(link_dwords),
537 				    link_target);
538 
539 	if (drm_debug_enabled(DRM_UT_DRIVER))
540 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
541 }
542