freedreno: fence_server_sync() fixes
[mesa.git] / src / gallium / drivers / freedreno / freedreno_context.c
1 /*
2 * Copyright (C) 2012 Rob Clark <robclark@freedesktop.org>
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21 * SOFTWARE.
22 *
23 * Authors:
24 * Rob Clark <robclark@freedesktop.org>
25 */
26
27 #include "freedreno_context.h"
28 #include "freedreno_blitter.h"
29 #include "freedreno_draw.h"
30 #include "freedreno_fence.h"
31 #include "freedreno_log.h"
32 #include "freedreno_program.h"
33 #include "freedreno_resource.h"
34 #include "freedreno_texture.h"
35 #include "freedreno_state.h"
36 #include "freedreno_gmem.h"
37 #include "freedreno_query.h"
38 #include "freedreno_query_hw.h"
39 #include "freedreno_util.h"
40 #include "util/u_upload_mgr.h"
41
42 #if DETECT_OS_ANDROID
43 #include "util/u_process.h"
44 #include <sys/stat.h>
45 #include <sys/types.h>
46 #endif
47
48 static void
49 fd_context_flush(struct pipe_context *pctx, struct pipe_fence_handle **fencep,
50 unsigned flags)
51 {
52 struct fd_context *ctx = fd_context(pctx);
53 struct pipe_fence_handle *fence = NULL;
54 // TODO we want to lookup batch if it exists, but not create one if not.
55 struct fd_batch *batch = fd_context_batch(ctx);
56
57 DBG("%p: flush: flags=%x\n", ctx->batch, flags);
58
59 /* In some sequence of events, we can end up with a last_fence that is
60 * not an "fd" fence, which results in eglDupNativeFenceFDANDROID()
61 * errors.
62 */
63 if ((flags & PIPE_FLUSH_FENCE_FD) && ctx->last_fence &&
64 !fd_fence_is_fd(ctx->last_fence))
65 fd_fence_ref(&ctx->last_fence, NULL);
66
67 /* if no rendering since last flush, ie. app just decided it needed
68 * a fence, re-use the last one:
69 */
70 if (ctx->last_fence) {
71 fd_fence_ref(&fence, ctx->last_fence);
72 fd_bc_dump(ctx->screen, "%p: reuse last_fence, remaining:\n", ctx);
73 goto out;
74 }
75
76 if (!batch) {
77 fd_bc_dump(ctx->screen, "%p: NULL batch, remaining:\n", ctx);
78 return;
79 }
80
81 /* Take a ref to the batch's fence (batch can be unref'd when flushed: */
82 fd_fence_ref(&fence, batch->fence);
83
84 if (flags & PIPE_FLUSH_FENCE_FD)
85 batch->needs_out_fence_fd = true;
86
87 fd_bc_dump(ctx->screen, "%p: flushing %p<%u>, flags=0x%x, pending:\n",
88 ctx, batch, batch->seqno, flags);
89
90 if (!ctx->screen->reorder) {
91 fd_batch_flush(batch);
92 } else if (flags & PIPE_FLUSH_DEFERRED) {
93 fd_bc_flush_deferred(&ctx->screen->batch_cache, ctx);
94 } else {
95 fd_bc_flush(&ctx->screen->batch_cache, ctx);
96 }
97
98 fd_bc_dump(ctx->screen, "%p: remaining:\n", ctx);
99
100 out:
101 if (fencep)
102 fd_fence_ref(fencep, fence);
103
104 fd_fence_ref(&ctx->last_fence, fence);
105
106 fd_fence_ref(&fence, NULL);
107
108 if (flags & PIPE_FLUSH_END_OF_FRAME)
109 fd_log_eof(ctx);
110 }
111
112 static void
113 fd_texture_barrier(struct pipe_context *pctx, unsigned flags)
114 {
115 if (flags == PIPE_TEXTURE_BARRIER_FRAMEBUFFER) {
116 struct fd_context *ctx = fd_context(pctx);
117
118 if (ctx->framebuffer_barrier) {
119 ctx->framebuffer_barrier(ctx);
120 return;
121 }
122 }
123
124 /* On devices that could sample from GMEM we could possibly do better.
125 * Or if we knew that we were doing GMEM bypass we could just emit a
126 * cache flush, perhaps? But we don't know if future draws would cause
127 * us to use GMEM, and a flush in bypass isn't the end of the world.
128 */
129 fd_context_flush(pctx, NULL, 0);
130 }
131
132 static void
133 fd_memory_barrier(struct pipe_context *pctx, unsigned flags)
134 {
135 if (!(flags & ~PIPE_BARRIER_UPDATE))
136 return;
137
138 fd_context_flush(pctx, NULL, 0);
139 /* TODO do we need to check for persistently mapped buffers and fd_bo_cpu_prep()?? */
140 }
141
142 static void
143 emit_string_tail(struct fd_ringbuffer *ring, const char *string, int len)
144 {
145 const uint32_t *buf = (const void *)string;
146
147 while (len >= 4) {
148 OUT_RING(ring, *buf);
149 buf++;
150 len -= 4;
151 }
152
153 /* copy remainder bytes without reading past end of input string: */
154 if (len > 0) {
155 uint32_t w = 0;
156 memcpy(&w, buf, len);
157 OUT_RING(ring, w);
158 }
159 }
160
161 /* for prior to a5xx: */
162 void
163 fd_emit_string(struct fd_ringbuffer *ring,
164 const char *string, int len)
165 {
166 /* max packet size is 0x3fff+1 dwords: */
167 len = MIN2(len, 0x4000 * 4);
168
169 OUT_PKT3(ring, CP_NOP, align(len, 4) / 4);
170 emit_string_tail(ring, string, len);
171 }
172
173 /* for a5xx+ */
174 void
175 fd_emit_string5(struct fd_ringbuffer *ring,
176 const char *string, int len)
177 {
178 /* max packet size is 0x3fff dwords: */
179 len = MIN2(len, 0x3fff * 4);
180
181 OUT_PKT7(ring, CP_NOP, align(len, 4) / 4);
182 emit_string_tail(ring, string, len);
183 }
184
185 /**
186 * emit marker string as payload of a no-op packet, which can be
187 * decoded by cffdump.
188 */
189 static void
190 fd_emit_string_marker(struct pipe_context *pctx, const char *string, int len)
191 {
192 struct fd_context *ctx = fd_context(pctx);
193
194 if (!ctx->batch)
195 return;
196
197 ctx->batch->needs_flush = true;
198
199 if (ctx->screen->gpu_id >= 500) {
200 fd_emit_string5(ctx->batch->draw, string, len);
201 } else {
202 fd_emit_string(ctx->batch->draw, string, len);
203 }
204 }
205
206 void
207 fd_context_destroy(struct pipe_context *pctx)
208 {
209 struct fd_context *ctx = fd_context(pctx);
210 unsigned i;
211
212 DBG("");
213
214 fd_screen_lock(ctx->screen);
215 list_del(&ctx->node);
216 fd_screen_unlock(ctx->screen);
217
218 fd_log_process(ctx, true);
219 assert(list_is_empty(&ctx->log_chunks));
220
221 fd_fence_ref(&ctx->last_fence, NULL);
222
223 if (ctx->in_fence_fd != -1)
224 close(ctx->in_fence_fd);
225
226 util_copy_framebuffer_state(&ctx->framebuffer, NULL);
227 fd_batch_reference(&ctx->batch, NULL); /* unref current batch */
228 fd_bc_invalidate_context(ctx);
229
230 fd_prog_fini(pctx);
231
232 if (ctx->blitter)
233 util_blitter_destroy(ctx->blitter);
234
235 if (pctx->stream_uploader)
236 u_upload_destroy(pctx->stream_uploader);
237
238 if (ctx->clear_rs_state)
239 pctx->delete_rasterizer_state(pctx, ctx->clear_rs_state);
240
241 if (ctx->primconvert)
242 util_primconvert_destroy(ctx->primconvert);
243
244 slab_destroy_child(&ctx->transfer_pool);
245
246 for (i = 0; i < ARRAY_SIZE(ctx->vsc_pipe_bo); i++) {
247 if (!ctx->vsc_pipe_bo[i])
248 break;
249 fd_bo_del(ctx->vsc_pipe_bo[i]);
250 }
251
252 fd_device_del(ctx->dev);
253 fd_pipe_del(ctx->pipe);
254
255 mtx_destroy(&ctx->gmem_lock);
256
257 if (fd_mesa_debug & (FD_DBG_BSTAT | FD_DBG_MSGS)) {
258 printf("batch_total=%u, batch_sysmem=%u, batch_gmem=%u, batch_nondraw=%u, batch_restore=%u\n",
259 (uint32_t)ctx->stats.batch_total, (uint32_t)ctx->stats.batch_sysmem,
260 (uint32_t)ctx->stats.batch_gmem, (uint32_t)ctx->stats.batch_nondraw,
261 (uint32_t)ctx->stats.batch_restore);
262 }
263 }
264
265 static void
266 fd_set_debug_callback(struct pipe_context *pctx,
267 const struct pipe_debug_callback *cb)
268 {
269 struct fd_context *ctx = fd_context(pctx);
270
271 if (cb)
272 ctx->debug = *cb;
273 else
274 memset(&ctx->debug, 0, sizeof(ctx->debug));
275 }
276
277 static uint32_t
278 fd_get_reset_count(struct fd_context *ctx, bool per_context)
279 {
280 uint64_t val;
281 enum fd_param_id param =
282 per_context ? FD_CTX_FAULTS : FD_GLOBAL_FAULTS;
283 int ret = fd_pipe_get_param(ctx->pipe, param, &val);
284 debug_assert(!ret);
285 return val;
286 }
287
288 static enum pipe_reset_status
289 fd_get_device_reset_status(struct pipe_context *pctx)
290 {
291 struct fd_context *ctx = fd_context(pctx);
292 int context_faults = fd_get_reset_count(ctx, true);
293 int global_faults = fd_get_reset_count(ctx, false);
294 enum pipe_reset_status status;
295
296 if (context_faults != ctx->context_reset_count) {
297 status = PIPE_GUILTY_CONTEXT_RESET;
298 } else if (global_faults != ctx->global_reset_count) {
299 status = PIPE_INNOCENT_CONTEXT_RESET;
300 } else {
301 status = PIPE_NO_RESET;
302 }
303
304 ctx->context_reset_count = context_faults;
305 ctx->global_reset_count = global_faults;
306
307 return status;
308 }
309
310 /* TODO we could combine a few of these small buffers (solid_vbuf,
311 * blit_texcoord_vbuf, and vsc_size_mem, into a single buffer and
312 * save a tiny bit of memory
313 */
314
315 static struct pipe_resource *
316 create_solid_vertexbuf(struct pipe_context *pctx)
317 {
318 static const float init_shader_const[] = {
319 -1.000000, +1.000000, +1.000000,
320 +1.000000, -1.000000, +1.000000,
321 };
322 struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
323 PIPE_BIND_CUSTOM, PIPE_USAGE_IMMUTABLE, sizeof(init_shader_const));
324 pipe_buffer_write(pctx, prsc, 0,
325 sizeof(init_shader_const), init_shader_const);
326 return prsc;
327 }
328
329 static struct pipe_resource *
330 create_blit_texcoord_vertexbuf(struct pipe_context *pctx)
331 {
332 struct pipe_resource *prsc = pipe_buffer_create(pctx->screen,
333 PIPE_BIND_CUSTOM, PIPE_USAGE_DYNAMIC, 16);
334 return prsc;
335 }
336
337 void
338 fd_context_setup_common_vbos(struct fd_context *ctx)
339 {
340 struct pipe_context *pctx = &ctx->base;
341
342 ctx->solid_vbuf = create_solid_vertexbuf(pctx);
343 ctx->blit_texcoord_vbuf = create_blit_texcoord_vertexbuf(pctx);
344
345 /* setup solid_vbuf_state: */
346 ctx->solid_vbuf_state.vtx = pctx->create_vertex_elements_state(
347 pctx, 1, (struct pipe_vertex_element[]){{
348 .vertex_buffer_index = 0,
349 .src_offset = 0,
350 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
351 }});
352 ctx->solid_vbuf_state.vertexbuf.count = 1;
353 ctx->solid_vbuf_state.vertexbuf.vb[0].stride = 12;
354 ctx->solid_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->solid_vbuf;
355
356 /* setup blit_vbuf_state: */
357 ctx->blit_vbuf_state.vtx = pctx->create_vertex_elements_state(
358 pctx, 2, (struct pipe_vertex_element[]){{
359 .vertex_buffer_index = 0,
360 .src_offset = 0,
361 .src_format = PIPE_FORMAT_R32G32_FLOAT,
362 }, {
363 .vertex_buffer_index = 1,
364 .src_offset = 0,
365 .src_format = PIPE_FORMAT_R32G32B32_FLOAT,
366 }});
367 ctx->blit_vbuf_state.vertexbuf.count = 2;
368 ctx->blit_vbuf_state.vertexbuf.vb[0].stride = 8;
369 ctx->blit_vbuf_state.vertexbuf.vb[0].buffer.resource = ctx->blit_texcoord_vbuf;
370 ctx->blit_vbuf_state.vertexbuf.vb[1].stride = 12;
371 ctx->blit_vbuf_state.vertexbuf.vb[1].buffer.resource = ctx->solid_vbuf;
372 }
373
374 void
375 fd_context_cleanup_common_vbos(struct fd_context *ctx)
376 {
377 struct pipe_context *pctx = &ctx->base;
378
379 pctx->delete_vertex_elements_state(pctx, ctx->solid_vbuf_state.vtx);
380 pctx->delete_vertex_elements_state(pctx, ctx->blit_vbuf_state.vtx);
381
382 pipe_resource_reference(&ctx->solid_vbuf, NULL);
383 pipe_resource_reference(&ctx->blit_texcoord_vbuf, NULL);
384 }
385
386 struct pipe_context *
387 fd_context_init(struct fd_context *ctx, struct pipe_screen *pscreen,
388 const uint8_t *primtypes, void *priv, unsigned flags)
389 {
390 struct fd_screen *screen = fd_screen(pscreen);
391 struct pipe_context *pctx;
392 unsigned prio = 1;
393 int i;
394
395 /* lower numerical value == higher priority: */
396 if (fd_mesa_debug & FD_DBG_HIPRIO)
397 prio = 0;
398 else if (flags & PIPE_CONTEXT_HIGH_PRIORITY)
399 prio = 0;
400 else if (flags & PIPE_CONTEXT_LOW_PRIORITY)
401 prio = 2;
402
403 ctx->screen = screen;
404 ctx->pipe = fd_pipe_new2(screen->dev, FD_PIPE_3D, prio);
405
406 ctx->in_fence_fd = -1;
407
408 if (fd_device_version(screen->dev) >= FD_VERSION_ROBUSTNESS) {
409 ctx->context_reset_count = fd_get_reset_count(ctx, true);
410 ctx->global_reset_count = fd_get_reset_count(ctx, false);
411 }
412
413 ctx->primtypes = primtypes;
414 ctx->primtype_mask = 0;
415 for (i = 0; i < PIPE_PRIM_MAX; i++)
416 if (primtypes[i])
417 ctx->primtype_mask |= (1 << i);
418
419 (void) mtx_init(&ctx->gmem_lock, mtx_plain);
420
421 /* need some sane default in case gallium frontends don't
422 * set some state:
423 */
424 ctx->sample_mask = 0xffff;
425 ctx->active_queries = true;
426
427 pctx = &ctx->base;
428 pctx->screen = pscreen;
429 pctx->priv = priv;
430 pctx->flush = fd_context_flush;
431 pctx->emit_string_marker = fd_emit_string_marker;
432 pctx->set_debug_callback = fd_set_debug_callback;
433 pctx->get_device_reset_status = fd_get_device_reset_status;
434 pctx->create_fence_fd = fd_create_fence_fd;
435 pctx->fence_server_sync = fd_fence_server_sync;
436 pctx->fence_server_signal = fd_fence_server_signal;
437 pctx->texture_barrier = fd_texture_barrier;
438 pctx->memory_barrier = fd_memory_barrier;
439
440 pctx->stream_uploader = u_upload_create_default(pctx);
441 if (!pctx->stream_uploader)
442 goto fail;
443 pctx->const_uploader = pctx->stream_uploader;
444
445 slab_create_child(&ctx->transfer_pool, &screen->transfer_pool);
446
447 fd_draw_init(pctx);
448 fd_resource_context_init(pctx);
449 fd_query_context_init(pctx);
450 fd_texture_init(pctx);
451 fd_state_init(pctx);
452
453 ctx->blitter = util_blitter_create(pctx);
454 if (!ctx->blitter)
455 goto fail;
456
457 ctx->primconvert = util_primconvert_create(pctx, ctx->primtype_mask);
458 if (!ctx->primconvert)
459 goto fail;
460
461 list_inithead(&ctx->hw_active_queries);
462 list_inithead(&ctx->acc_active_queries);
463 list_inithead(&ctx->log_chunks);
464
465 fd_screen_lock(ctx->screen);
466 list_add(&ctx->node, &ctx->screen->context_list);
467 fd_screen_unlock(ctx->screen);
468
469 ctx->current_scissor = &ctx->disabled_scissor;
470
471 ctx->log_out = stdout;
472
473 if ((fd_mesa_debug & FD_DBG_LOG) &&
474 !(ctx->record_timestamp && ctx->ts_to_ns)) {
475 printf("logging not supported!\n");
476 fd_mesa_debug &= ~FD_DBG_LOG;
477 }
478
479 #if DETECT_OS_ANDROID
480 if (fd_mesa_debug & FD_DBG_LOG) {
481 static unsigned idx = 0;
482 char *p;
483 asprintf(&p, "/data/fdlog/%s-%d.log", util_get_process_name(), idx++);
484
485 FILE *f = fopen(p, "w");
486 if (f)
487 ctx->log_out = f;
488 }
489 #endif
490
491 return pctx;
492
493 fail:
494 pctx->destroy(pctx);
495 return NULL;
496 }