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 	/*
104 	 * This assumes that if we're switching to 2D, we're switching
105 	 * away from 3D, and vice versa.  Hence, if we're switching to
106 	 * the 2D core, we need to flush the 3D depth and color caches,
107 	 * otherwise we need to flush the 2D pixel engine cache.
108 	 */
109 	if (gpu->exec_state == ETNA_PIPE_2D)
110 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
111 	else if (gpu->exec_state == ETNA_PIPE_3D)
112 		flush = VIVS_GL_FLUSH_CACHE_DEPTH | VIVS_GL_FLUSH_CACHE_COLOR;
113 
114 	CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
115 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
116 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
117 
118 	CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
119 		       VIVS_GL_PIPE_SELECT_PIPE(pipe));
120 }
121 
122 static void etnaviv_buffer_dump(struct etnaviv_gpu *gpu,
123 	struct etnaviv_cmdbuf *buf, u32 off, u32 len)
124 {
125 	u32 size = buf->size;
126 	u32 *ptr = buf->vaddr + off;
127 
128 	dev_info(gpu->dev, "virt %p phys 0x%08x free 0x%08x\n",
129 			ptr, etnaviv_cmdbuf_get_va(buf) + off, size - len * 4 - off);
130 
131 	print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
132 			ptr, len * 4, 0);
133 }
134 
135 /*
136  * Safely replace the WAIT of a waitlink with a new command and argument.
137  * The GPU may be executing this WAIT while we're modifying it, so we have
138  * to write it in a specific order to avoid the GPU branching to somewhere
139  * else.  'wl_offset' is the offset to the first byte of the WAIT command.
140  */
141 static void etnaviv_buffer_replace_wait(struct etnaviv_cmdbuf *buffer,
142 	unsigned int wl_offset, u32 cmd, u32 arg)
143 {
144 	u32 *lw = buffer->vaddr + wl_offset;
145 
146 	lw[1] = arg;
147 	mb();
148 	lw[0] = cmd;
149 	mb();
150 }
151 
152 /*
153  * Ensure that there is space in the command buffer to contiguously write
154  * 'cmd_dwords' 64-bit words into the buffer, wrapping if necessary.
155  */
156 static u32 etnaviv_buffer_reserve(struct etnaviv_gpu *gpu,
157 	struct etnaviv_cmdbuf *buffer, unsigned int cmd_dwords)
158 {
159 	if (buffer->user_size + cmd_dwords * sizeof(u64) > buffer->size)
160 		buffer->user_size = 0;
161 
162 	return etnaviv_cmdbuf_get_va(buffer) + buffer->user_size;
163 }
164 
165 u16 etnaviv_buffer_init(struct etnaviv_gpu *gpu)
166 {
167 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
168 
169 	/* initialize buffer */
170 	buffer->user_size = 0;
171 
172 	CMD_WAIT(buffer);
173 	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
174 		 buffer->user_size - 4);
175 
176 	return buffer->user_size / 8;
177 }
178 
179 u16 etnaviv_buffer_config_mmuv2(struct etnaviv_gpu *gpu, u32 mtlb_addr, u32 safe_addr)
180 {
181 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
182 
183 	buffer->user_size = 0;
184 
185 	if (gpu->identity.features & chipFeatures_PIPE_3D) {
186 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
187 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_3D));
188 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
189 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
190 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
191 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
192 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
193 	}
194 
195 	if (gpu->identity.features & chipFeatures_PIPE_2D) {
196 		CMD_LOAD_STATE(buffer, VIVS_GL_PIPE_SELECT,
197 			       VIVS_GL_PIPE_SELECT_PIPE(ETNA_PIPE_2D));
198 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
199 			mtlb_addr | VIVS_MMUv2_CONFIGURATION_MODE_MODE4_K);
200 		CMD_LOAD_STATE(buffer, VIVS_MMUv2_SAFE_ADDRESS, safe_addr);
201 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
202 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
203 	}
204 
205 	CMD_END(buffer);
206 
207 	buffer->user_size = ALIGN(buffer->user_size, 8);
208 
209 	return buffer->user_size / 8;
210 }
211 
212 void etnaviv_buffer_end(struct etnaviv_gpu *gpu)
213 {
214 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
215 	unsigned int waitlink_offset = buffer->user_size - 16;
216 	u32 link_target, flush = 0;
217 
218 	if (gpu->exec_state == ETNA_PIPE_2D)
219 		flush = VIVS_GL_FLUSH_CACHE_PE2D;
220 	else if (gpu->exec_state == ETNA_PIPE_3D)
221 		flush = VIVS_GL_FLUSH_CACHE_DEPTH |
222 			VIVS_GL_FLUSH_CACHE_COLOR |
223 			VIVS_GL_FLUSH_CACHE_TEXTURE |
224 			VIVS_GL_FLUSH_CACHE_TEXTUREVS |
225 			VIVS_GL_FLUSH_CACHE_SHADER_L2;
226 
227 	if (flush) {
228 		unsigned int dwords = 7;
229 
230 		link_target = etnaviv_buffer_reserve(gpu, buffer, dwords);
231 
232 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
233 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
234 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE, flush);
235 		if (gpu->exec_state == ETNA_PIPE_3D)
236 			CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
237 				       VIVS_TS_FLUSH_CACHE_FLUSH);
238 		CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
239 		CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
240 		CMD_END(buffer);
241 
242 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
243 					    VIV_FE_LINK_HEADER_OP_LINK |
244 					    VIV_FE_LINK_HEADER_PREFETCH(dwords),
245 					    link_target);
246 	} else {
247 		/* Replace the last link-wait with an "END" command */
248 		etnaviv_buffer_replace_wait(buffer, waitlink_offset,
249 					    VIV_FE_END_HEADER_OP_END, 0);
250 	}
251 }
252 
253 /* Append a command buffer to the ring buffer. */
254 void etnaviv_buffer_queue(struct etnaviv_gpu *gpu, unsigned int event,
255 	struct etnaviv_cmdbuf *cmdbuf)
256 {
257 	struct etnaviv_cmdbuf *buffer = gpu->buffer;
258 	unsigned int waitlink_offset = buffer->user_size - 16;
259 	u32 return_target, return_dwords;
260 	u32 link_target, link_dwords;
261 
262 	if (drm_debug & DRM_UT_DRIVER)
263 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
264 
265 	link_target = etnaviv_cmdbuf_get_va(cmdbuf);
266 	link_dwords = cmdbuf->size / 8;
267 
268 	/*
269 	 * If we need maintanence prior to submitting this buffer, we will
270 	 * need to append a mmu flush load state, followed by a new
271 	 * link to this buffer - a total of four additional words.
272 	 */
273 	if (gpu->mmu->need_flush || gpu->switch_context) {
274 		u32 target, extra_dwords;
275 
276 		/* link command */
277 		extra_dwords = 1;
278 
279 		/* flush command */
280 		if (gpu->mmu->need_flush) {
281 			if (gpu->mmu->version == ETNAVIV_IOMMU_V1)
282 				extra_dwords += 1;
283 			else
284 				extra_dwords += 3;
285 		}
286 
287 		/* pipe switch commands */
288 		if (gpu->switch_context)
289 			extra_dwords += 4;
290 
291 		target = etnaviv_buffer_reserve(gpu, buffer, extra_dwords);
292 
293 		if (gpu->mmu->need_flush) {
294 			/* Add the MMU flush */
295 			if (gpu->mmu->version == ETNAVIV_IOMMU_V1) {
296 				CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_MMU,
297 					       VIVS_GL_FLUSH_MMU_FLUSH_FEMMU |
298 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK1 |
299 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK2 |
300 					       VIVS_GL_FLUSH_MMU_FLUSH_PEMMU |
301 					       VIVS_GL_FLUSH_MMU_FLUSH_UNK4);
302 			} else {
303 				CMD_LOAD_STATE(buffer, VIVS_MMUv2_CONFIGURATION,
304 					VIVS_MMUv2_CONFIGURATION_MODE_MASK |
305 					VIVS_MMUv2_CONFIGURATION_ADDRESS_MASK |
306 					VIVS_MMUv2_CONFIGURATION_FLUSH_FLUSH);
307 				CMD_SEM(buffer, SYNC_RECIPIENT_FE,
308 					SYNC_RECIPIENT_PE);
309 				CMD_STALL(buffer, SYNC_RECIPIENT_FE,
310 					SYNC_RECIPIENT_PE);
311 			}
312 
313 			gpu->mmu->need_flush = false;
314 		}
315 
316 		if (gpu->switch_context) {
317 			etnaviv_cmd_select_pipe(gpu, buffer, cmdbuf->exec_state);
318 			gpu->exec_state = cmdbuf->exec_state;
319 			gpu->switch_context = false;
320 		}
321 
322 		/* And the link to the submitted buffer */
323 		CMD_LINK(buffer, link_dwords, link_target);
324 
325 		/* Update the link target to point to above instructions */
326 		link_target = target;
327 		link_dwords = extra_dwords;
328 	}
329 
330 	/*
331 	 * Append a LINK to the submitted command buffer to return to
332 	 * the ring buffer.  return_target is the ring target address.
333 	 * We need at most 7 dwords in the return target: 2 cache flush +
334 	 * 2 semaphore stall + 1 event + 1 wait + 1 link.
335 	 */
336 	return_dwords = 7;
337 	return_target = etnaviv_buffer_reserve(gpu, buffer, return_dwords);
338 	CMD_LINK(cmdbuf, return_dwords, return_target);
339 
340 	/*
341 	 * Append a cache flush, stall, event, wait and link pointing back to
342 	 * the wait command to the ring buffer.
343 	 */
344 	if (gpu->exec_state == ETNA_PIPE_2D) {
345 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
346 				       VIVS_GL_FLUSH_CACHE_PE2D);
347 	} else {
348 		CMD_LOAD_STATE(buffer, VIVS_GL_FLUSH_CACHE,
349 				       VIVS_GL_FLUSH_CACHE_DEPTH |
350 				       VIVS_GL_FLUSH_CACHE_COLOR);
351 		CMD_LOAD_STATE(buffer, VIVS_TS_FLUSH_CACHE,
352 				       VIVS_TS_FLUSH_CACHE_FLUSH);
353 	}
354 	CMD_SEM(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
355 	CMD_STALL(buffer, SYNC_RECIPIENT_FE, SYNC_RECIPIENT_PE);
356 	CMD_LOAD_STATE(buffer, VIVS_GL_EVENT, VIVS_GL_EVENT_EVENT_ID(event) |
357 		       VIVS_GL_EVENT_FROM_PE);
358 	CMD_WAIT(buffer);
359 	CMD_LINK(buffer, 2, etnaviv_cmdbuf_get_va(buffer) +
360 			    buffer->user_size - 4);
361 
362 	if (drm_debug & DRM_UT_DRIVER)
363 		pr_info("stream link to 0x%08x @ 0x%08x %p\n",
364 			return_target, etnaviv_cmdbuf_get_va(cmdbuf),
365 			cmdbuf->vaddr);
366 
367 	if (drm_debug & DRM_UT_DRIVER) {
368 		print_hex_dump(KERN_INFO, "cmd ", DUMP_PREFIX_OFFSET, 16, 4,
369 			       cmdbuf->vaddr, cmdbuf->size, 0);
370 
371 		pr_info("link op: %p\n", buffer->vaddr + waitlink_offset);
372 		pr_info("addr: 0x%08x\n", link_target);
373 		pr_info("back: 0x%08x\n", return_target);
374 		pr_info("event: %d\n", event);
375 	}
376 
377 	/*
378 	 * Kick off the submitted command by replacing the previous
379 	 * WAIT with a link to the address in the ring buffer.
380 	 */
381 	etnaviv_buffer_replace_wait(buffer, waitlink_offset,
382 				    VIV_FE_LINK_HEADER_OP_LINK |
383 				    VIV_FE_LINK_HEADER_PREFETCH(link_dwords),
384 				    link_target);
385 
386 	if (drm_debug & DRM_UT_DRIVER)
387 		etnaviv_buffer_dump(gpu, buffer, 0, 0x50);
388 }
389