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