1 /*
2  * Copyright (C) 2014 Etnaviv Project
3  * Author: Christian Gmeiner <christian.gmeiner@gmail.com>
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms of the GNU General Public License version 2 as published by
7  * the Free Software Foundation.
8  *
9  * This program is distributed in the hope that it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  *
14  * You should have received a copy of the GNU General Public License along with
15  * this program.  If not, see <http://www.gnu.org/licenses/>.
16  */
17 
18 #include "etnaviv_cmdbuf.h"
19 #include "etnaviv_gpu.h"
20 #include "etnaviv_gem.h"
21 #include "etnaviv_mmu.h"
22 
23 #include "common.xml.h"
24 #include "state.xml.h"
25 #include "state_hi.xml.h"
26 #include "state_3d.xml.h"
27 #include "cmdstream.xml.h"
28 
29 /*
30  * Command Buffer helper:
31  */
32 
33 
34 static inline void OUT(struct etnaviv_cmdbuf *buffer, u32 data)
35 {
36 	u32 *vaddr = (u32 *)buffer->vaddr;
37 
38 	BUG_ON(buffer->user_size >= buffer->size);
39 
40 	vaddr[buffer->user_size / 4] = data;
41 	buffer->user_size += 4;
42 }
43 
44 static inline void CMD_LOAD_STATE(struct etnaviv_cmdbuf *buffer,
45 	u32 reg, u32 value)
46 {
47 	u32 index = reg >> VIV_FE_LOAD_STATE_HEADER_OFFSET__SHR;
48 
49 	buffer->user_size = ALIGN(buffer->user_size, 8);
50 
51 	/* write a register via cmd stream */
52 	OUT(buffer, VIV_FE_LOAD_STATE_HEADER_OP_LOAD_STATE |
53 		    VIV_FE_LOAD_STATE_HEADER_COUNT(1) |
54 		    VIV_FE_LOAD_STATE_HEADER_OFFSET(index));
55 	OUT(buffer, value);
56 }
57 
58 static inline void CMD_END(struct etnaviv_cmdbuf *buffer)
59 {
60 	buffer->user_size = ALIGN(buffer->user_size, 8);
61 
62 	OUT(buffer, VIV_FE_END_HEADER_OP_END);
63 }
64 
65 static inline void CMD_WAIT(struct etnaviv_cmdbuf *buffer)
66 {
67 	buffer->user_size = ALIGN(buffer->user_size, 8);
68 
69 	OUT(buffer, VIV_FE_WAIT_HEADER_OP_WAIT | 200);
70 }
71 
72 static inline void CMD_LINK(struct etnaviv_cmdbuf *buffer,
73 	u16 prefetch, u32 address)
74 {
75 	buffer->user_size = ALIGN(buffer->user_size, 8);
76 
77 	OUT(buffer, VIV_FE_LINK_HEADER_OP_LINK |
78 		    VIV_FE_LINK_HEADER_PREFETCH(prefetch));
79 	OUT(buffer, address);
80 }
81 
82 static inline void CMD_STALL(struct etnaviv_cmdbuf *buffer,
83 	u32 from, u32 to)
84 {
85 	buffer->user_size = ALIGN(buffer->user_size, 8);
86 
87 	OUT(buffer, VIV_FE_STALL_HEADER_OP_STALL);
88 	OUT(buffer, VIV_FE_STALL_TOKEN_FROM(from) | VIV_FE_STALL_TOKEN_TO(to));
89 }
90 
91 static inline void CMD_SEM(struct etnaviv_cmdbuf *buffer, u32 from, u32 to)
92 {
93 	CMD_LOAD_STATE(buffer, VIVS_GL_SEMAPHORE_TOKEN,
94 		       VIVS_GL_SEMAPHORE_TOKEN_FROM(from) |
95 		       VIVS_GL_SEMAPHORE_TOKEN_TO(to));
96 }
97 
98 static void etnaviv_cmd_select_pipe(struct etnaviv_gpu *gpu,
99 	struct etnaviv_cmdbuf *buffer, u8 pipe)
100 {
101 	u32 flush = 0;
102 
103 	lockdep_assert_held(&gpu->lock);
104 
105 	/*
106 	 * This assumes that if we're switching to 2D, we're switching
107 	 * away from 3D, and vice versa.  Hence, if we're switching to
108 	 * the 2D core, we need to flush the 3D depth and color caches,
109 	 * otherwise we need to flush the 2D pixel engine cache.
110 	 */
111 	if (gpu->exec_state == ETNA_PIPE_2D)
112 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
113 	else if (gpu->exec_state == ETNA_PIPE_3D)
114 		flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
115 
116 	CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
117 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
118 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
119 
120 	CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
121 		       VIVS_GL_PIPE_SELECT_PIPE(pipe));
122 }
123 
124 static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
125 	struct etnaviv_cmdbuf *buf, u32 off, u32 len)
126 {
127 	u32 size = buf->size;
128 	u32 *ptr = buf->vaddr + off;
129 
130 	dev_info(gpu->dev, "virt %p phys 0x%08x free 0x%08x\n",
131 			ptr, etnaviv_cmdbuf_get_va(buf) + off, size - len * 4 - off);
132 
133 	print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
134 			ptr, len * 4, 0);
135 }
136 
137 /*
138  * Safely replace the WAIT of a waitlink with a new command and argument.
139  * The GPU may be executing this WAIT while we're modifying it, so we have
140  * to write it in a specific order to avoid the GPU branching to somewhere
141  * else.  'wl_offset' is the offset to the first byte of the WAIT command.
142  */
143 static void etnaviv_buffer_replace_wait(struct etnaviv_cmdbuf *buffer,
144 	unsigned int wl_offset, u32 cmd, u32 arg)
145 {
146 	u32 *lw = buffer->vaddr + wl_offset;
147 
148 	lw[1] = arg;
149 	mb();
150 	lw[0] = cmd;
151 	mb();
152 }
153 
154 /*
155  * Ensure that there is space in the command buffer to contiguously write
156  * 'cmd_dwords' 64-bit words into the buffer, wrapping if necessary.
157  */
158 static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
159 	struct etnaviv_cmdbuf *buffer, unsigned int cmd_dwords)
160 {
161 	if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
162 		buffer->user_size = 0;
163 
164 	return etnaviv_cmdbuf_get_va(buffer) + buffer->user_size;
165 }
166 
167 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
168 {
169 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
170 
171 	lockdep_assert_held(&gpu->lock);
172 
173 	/* initialize buffer */
174 	buffer->user_size = 0;
175 
176 	CMD_WAIT(buffer);
177 	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
178 		 buffer->user_size - 4);
179 
180 	return buffer->user_size / 8;
181 }
182 
183 u16 etnaviv_buffer_config_mmuv2(struct etnaviv_gpu *gpu, u32 mtlb_addr, u32 safe_addr)
184 {
185 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
186 
187 	lockdep_assert_held(&gpu->lock);
188 
189 	buffer->user_size = 0;
190 
191 	if (gpu->identity.features & chipFeatures_PIPE_3D) {
192 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
193 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_3D));
194 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
195 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
196 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
197 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
198 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
199 	}
200 
201 	if (gpu->identity.features & chipFeatures_PIPE_2D) {
202 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
203 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_2D));
204 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
205 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
206 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
207 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
208 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
209 	}
210 
211 	CMD_END(buffer);
212 
213 	buffer->user_size = ALIGN(buffer->user_size, 8);
214 
215 	return buffer->user_size / 8;
216 }
217 
218 void etnaviv_buffer_end(struct etnaviv_gpu *gpu)
219 {
220 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
221 	unsigned int waitlink_offset = buffer->user_size - 16;
222 	u32 link_target, flush = 0;
223 
224 	lockdep_assert_held(&gpu->lock);
225 
226 	if (gpu->exec_state == ETNA_PIPE_2D)
227 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
228 	else if (gpu->exec_state == ETNA_PIPE_3D)
229 		flush = VIVS_GL_FLUSH_CACHE_DEPTH |
230 			VIVS_GL_FLUSH_CACHE_COLOR |
231 			VIVS_GL_FLUSH_CACHE_TEXTURE |
232 			VIVS_GL_FLUSH_CACHE_TEXTUREVS |
233 			VIVS_GL_FLUSH_CACHE_SHADER_L2;
234 
235 	if (flush) {
236 		unsigned int dwords = 7;
237 
238 		link_target = etnaviv_buffer_reserve(gpu, buffer, dwords);
239 
240 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
241 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
242 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
243 		if (gpu->exec_state == ETNA_PIPE_3D)
244 			CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
245 				       VIVS_TS_FLUSH_CACHE_FLUSH);
246 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
247 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
248 		CMD_END(buffer);
249 
250 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
251 					    VIV_FE_LINK_HEADER_OP_LINK |
252 					    VIV_FE_LINK_HEADER_PREFETCH(dwords),
253 					    link_target);
254 	} else {
255 		/* Replace the last link-wait with an "END" command */
256 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
257 					    VIV_FE_END_HEADER_OP_END, 0);
258 	}
259 }
260 
261 /* Append a 'sync point' to the ring buffer. */
262 void etnaviv_sync_point_queue(struct etnaviv_gpu *gpu, unsigned int event)
263 {
264 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
265 	unsigned int waitlink_offset = buffer->user_size - 16;
266 	u32 dwords, target;
267 
268 	lockdep_assert_held(&gpu->lock);
269 
270 	/*
271 	 * We need at most 3 dwords in the return target:
272 	 * 1 event + 1 end + 1 wait + 1 link.
273 	 */
274 	dwords = 4;
275 	target = etnaviv_buffer_reserve(gpu, buffer, dwords);
276 
277 	/* Signal sync point event */
278 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
279 		       VIVS_GL_EVENT_FROM_PE);
280 
281 	/* Stop the FE to 'pause' the GPU */
282 	CMD_END(buffer);
283 
284 	/* Append waitlink */
285 	CMD_WAIT(buffer);
286 	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
287 			    buffer->user_size - 4);
288 
289 	/*
290 	 * Kick off the 'sync point' command by replacing the previous
291 	 * WAIT with a link to the address in the ring buffer.
292 	 */
293 	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
294 				    VIV_FE_LINK_HEADER_OP_LINK |
295 				    VIV_FE_LINK_HEADER_PREFETCH(dwords),
296 				    target);
297 }
298 
299 /* Append a command buffer to the ring buffer. */
300 void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, u32 exec_state,
301 	unsigned int event, struct etnaviv_cmdbuf *cmdbuf)
302 {
303 	struct etnaviv_cmdbuf *buffer = &gpu->buffer;
304 	unsigned int waitlink_offset = buffer->user_size - 16;
305 	u32 return_target, return_dwords;
306 	u32 link_target, link_dwords;
307 	bool switch_context = gpu->exec_state != exec_state;
308 
309 	lockdep_assert_held(&gpu->lock);
310 
311 	if (drm_debug & DRM_UT_DRIVER)
312 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
313 
314 	link_target = etnaviv_cmdbuf_get_va(cmdbuf);
315 	link_dwords = cmdbuf->size / 8;
316 
317 	/*
318 	 * If we need maintanence prior to submitting this buffer, we will
319 	 * need to append a mmu flush load state, followed by a new
320 	 * link to this buffer - a total of four additional words.
321 	 */
322 	if (gpu->mmu->need_flush || switch_context) {
323 		u32 target, extra_dwords;
324 
325 		/* link command */
326 		extra_dwords = 1;
327 
328 		/* flush command */
329 		if (gpu->mmu->need_flush) {
330 			if (gpu->mmu->version == ETNAVIV_IOMMU_V1)
331 				extra_dwords += 1;
332 			else
333 				extra_dwords += 3;
334 		}
335 
336 		/* pipe switch commands */
337 		if (switch_context)
338 			extra_dwords += 4;
339 
340 		target = etnaviv_buffer_reserve(gpu, buffer, extra_dwords);
341 
342 		if (gpu->mmu->need_flush) {
343 			/* Add the MMU flush */
344 			if (gpu->mmu->version == ETNAVIV_IOMMU_V1) {
345 				CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_MMU,
346 					       VIVS_GL_FLUSH_MMU_FLUSH_FEMMU |
347 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK1 |
348 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK2 |
349 					       VIVS_GL_FLUSH_MMU_FLUSH_PEMMU |
350 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK4);
351 			} else {
352 				CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
353 					VIVS_MMUv2_CONFIGURATION_MODE_MASK |
354 					VIVS_MMUv2_CONFIGURATION_ADDRESS_MASK |
355 					VIVS_MMUv2_CONFIGURATION_FLUSH_FLUSH);
356 				CMD_SEM(buffer, SYNC_RECIPIENT_FE,
357 					SYNC_RECIPIENT_PE);
358 				CMD_STALL(buffer, SYNC_RECIPIENT_FE,
359 					SYNC_RECIPIENT_PE);
360 			}
361 
362 			gpu->mmu->need_flush = false;
363 		}
364 
365 		if (switch_context) {
366 			etnaviv_cmd_select_pipe(gpu, buffer, exec_state);
367 			gpu->exec_state = exec_state;
368 		}
369 
370 		/* And the link to the submitted buffer */
371 		CMD_LINK(buffer, link_dwords, link_target);
372 
373 		/* Update the link target to point to above instructions */
374 		link_target = target;
375 		link_dwords = extra_dwords;
376 	}
377 
378 	/*
379 	 * Append a LINK to the submitted command buffer to return to
380 	 * the ring buffer.  return_target is the ring target address.
381 	 * We need at most 7 dwords in the return target: 2 cache flush +
382 	 * 2 semaphore stall + 1 event + 1 wait + 1 link.
383 	 */
384 	return_dwords = 7;
385 	return_target = etnaviv_buffer_reserve(gpu, buffer, return_dwords);
386 	CMD_LINK(cmdbuf, return_dwords, return_target);
387 
388 	/*
389 	 * Append a cache flush, stall, event, wait and link pointing back to
390 	 * the wait command to the ring buffer.
391 	 */
392 	if (gpu->exec_state == ETNA_PIPE_2D) {
393 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
394 				       VIVS_GL_FLUSH_CACHE_PE2D);
395 	} else {
396 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
397 				       VIVS_GL_FLUSH_CACHE_DEPTH |
398 				       VIVS_GL_FLUSH_CACHE_COLOR);
399 		CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
400 				       VIVS_TS_FLUSH_CACHE_FLUSH);
401 	}
402 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
403 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
404 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
405 		       VIVS_GL_EVENT_FROM_PE);
406 	CMD_WAIT(buffer);
407 	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
408 			    buffer->user_size - 4);
409 
410 	if (drm_debug & DRM_UT_DRIVER)
411 		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
412 			return_target, etnaviv_cmdbuf_get_va(cmdbuf),
413 			cmdbuf->vaddr);
414 
415 	if (drm_debug & DRM_UT_DRIVER) {
416 		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
417 			       cmdbuf->vaddr, cmdbuf->size, 0);
418 
419 		pr_info("link op: %p\n", buffer->vaddr + waitlink_offset);
420 		pr_info("addr: 0x%08x\n", link_target);
421 		pr_info("back: 0x%08x\n", return_target);
422 		pr_info("event: %d\n", event);
423 	}
424 
425 	/*
426 	 * Kick off the submitted command by replacing the previous
427 	 * WAIT with a link to the address in the ring buffer.
428 	 */
429 	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
430 				    VIV_FE_LINK_HEADER_OP_LINK |
431 				    VIV_FE_LINK_HEADER_PREFETCH(link_dwords),
432 				    link_target);
433 
434 	if (drm_debug & DRM_UT_DRIVER)
435 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
436 
437 	gpu->lastctx = cmdbuf->ctx;
438 }
439