1/*
2 * Copyright 2010 Jerome Glisse <glisse@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 * on the rights to use, copy, modify, merge, publish, distribute, sub
8 * license, and/or sell copies of the Software, and to permit persons to whom
9 * the 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 NON-INFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHOR(S) AND/OR THEIR SUPPLIERS BE LIABLE FOR ANY CLAIM,
19 * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
20 * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
21 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 *
23 * Authors:
24 *      Jerome Glisse
25 */
26#include "r600_pipe.h"
27#include "r600d.h"
28#include "util/u_memory.h"
29#include <errno.h>
30#include <unistd.h>
31
32
33void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw,
34			boolean count_draw_in, unsigned num_atomics)
35{
36	/* Flush the DMA IB if it's not empty. */
37	if (radeon_emitted(&ctx->b.dma.cs, 0))
38		ctx->b.dma.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
39
40	if (!radeon_cs_memory_below_limit(ctx->b.screen, &ctx->b.gfx.cs,
41					  ctx->b.vram, ctx->b.gtt)) {
42		ctx->b.gtt = 0;
43		ctx->b.vram = 0;
44		ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
45		return;
46	}
47	/* all will be accounted once relocation are emitted */
48	ctx->b.gtt = 0;
49	ctx->b.vram = 0;
50
51	/* Check available space in CS. */
52	if (count_draw_in) {
53		uint64_t mask;
54
55		/* The number of dwords all the dirty states would take. */
56		mask = ctx->dirty_atoms;
57		while (mask != 0)
58			num_dw += ctx->atoms[u_bit_scan64(&mask)]->num_dw;
59
60		/* The upper-bound of how much space a draw command would take. */
61		num_dw += R600_MAX_FLUSH_CS_DWORDS + R600_MAX_DRAW_CS_DWORDS;
62	}
63
64	/* add atomic counters, 8 pre + 8 post per counter + 16 post if any counters */
65	num_dw += (num_atomics * 16) + (num_atomics ? 16 : 0);
66
67	/* Count in r600_suspend_queries. */
68	num_dw += ctx->b.num_cs_dw_queries_suspend;
69
70	/* Count in streamout_end at the end of CS. */
71	if (ctx->b.streamout.begin_emitted) {
72		num_dw += ctx->b.streamout.num_dw_for_end;
73	}
74
75	/* SX_MISC */
76	if (ctx->b.chip_class == R600) {
77		num_dw += 3;
78	}
79
80	/* Count in framebuffer cache flushes at the end of CS. */
81	num_dw += R600_MAX_FLUSH_CS_DWORDS;
82
83	/* The fence at the end of CS. */
84	num_dw += 10;
85
86	/* Flush if there's not enough space. */
87	if (!ctx->b.ws->cs_check_space(&ctx->b.gfx.cs, num_dw, false)) {
88		ctx->b.gfx.flush(ctx, PIPE_FLUSH_ASYNC, NULL);
89	}
90}
91
92void r600_flush_emit(struct r600_context *rctx)
93{
94	struct radeon_cmdbuf *cs = &rctx->b.gfx.cs;
95	unsigned cp_coher_cntl = 0;
96	unsigned wait_until = 0;
97
98	if (!rctx->b.flags) {
99		return;
100	}
101
102	/* Ensure coherency between streamout and shaders. */
103	if (rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH)
104		rctx->b.flags |= r600_get_flush_flags(R600_COHERENCY_SHADER);
105
106	if (rctx->b.flags & R600_CONTEXT_WAIT_3D_IDLE) {
107		wait_until |= S_008040_WAIT_3D_IDLE(1);
108	}
109	if (rctx->b.flags & R600_CONTEXT_WAIT_CP_DMA_IDLE) {
110		wait_until |= S_008040_WAIT_CP_DMA_IDLE(1);
111	}
112
113	if (wait_until) {
114		/* Use of WAIT_UNTIL is deprecated on Cayman+ */
115		if (rctx->b.family >= CHIP_CAYMAN) {
116			/* emit a PS partial flush on Cayman/TN */
117			rctx->b.flags |= R600_CONTEXT_PS_PARTIAL_FLUSH;
118		}
119	}
120
121	/* Wait packets must be executed first, because SURFACE_SYNC doesn't
122	 * wait for shaders if it's not flushing CB or DB.
123	 */
124	if (rctx->b.flags & R600_CONTEXT_PS_PARTIAL_FLUSH) {
125		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
126		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) | EVENT_INDEX(4));
127	}
128
129	if (rctx->b.flags & R600_CONTEXT_CS_PARTIAL_FLUSH) {
130		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
131		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CS_PARTIAL_FLUSH) | EVENT_INDEX(4));
132	}
133
134	if (wait_until) {
135		/* Use of WAIT_UNTIL is deprecated on Cayman+ */
136		if (rctx->b.family < CHIP_CAYMAN) {
137			/* wait for things to settle */
138			radeon_set_config_reg(cs, R_008040_WAIT_UNTIL, wait_until);
139		}
140	}
141
142	if (rctx->b.chip_class >= R700 &&
143	    (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_CB_META)) {
144		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
145		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_FLUSH_AND_INV_CB_META) | EVENT_INDEX(0));
146	}
147
148	if (rctx->b.chip_class >= R700 &&
149	    (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_DB_META)) {
150		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
151		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_FLUSH_AND_INV_DB_META) | EVENT_INDEX(0));
152
153		/* Set FULL_CACHE_ENA for DB META flushes on r7xx and later.
154		 *
155		 * This hack predates use of FLUSH_AND_INV_DB_META, so it's
156		 * unclear whether it's still needed or even whether it has
157		 * any effect.
158		 */
159		cp_coher_cntl |= S_0085F0_FULL_CACHE_ENA(1);
160	}
161
162	if (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV ||
163	    (rctx->b.chip_class == R600 && rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH)) {
164		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
165		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0));
166	}
167
168	if (rctx->b.flags & R600_CONTEXT_INV_CONST_CACHE) {
169		/* Direct constant addressing uses the shader cache.
170		 * Indirect contant addressing uses the vertex cache. */
171		cp_coher_cntl |= S_0085F0_SH_ACTION_ENA(1) |
172				 (rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1)
173							 : S_0085F0_TC_ACTION_ENA(1));
174	}
175	if (rctx->b.flags & R600_CONTEXT_INV_VERTEX_CACHE) {
176		cp_coher_cntl |= rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1)
177							: S_0085F0_TC_ACTION_ENA(1);
178	}
179	if (rctx->b.flags & R600_CONTEXT_INV_TEX_CACHE) {
180		/* Textures use the texture cache.
181		 * Texture buffer objects use the vertex cache. */
182		cp_coher_cntl |= S_0085F0_TC_ACTION_ENA(1) |
183				 (rctx->has_vertex_cache ? S_0085F0_VC_ACTION_ENA(1) : 0);
184	}
185
186	/* Don't use the DB CP COHER logic on r6xx.
187	 * There are hw bugs.
188	 */
189	if (rctx->b.chip_class >= R700 &&
190	    (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_DB)) {
191		cp_coher_cntl |= S_0085F0_DB_ACTION_ENA(1) |
192				S_0085F0_DB_DEST_BASE_ENA(1) |
193				S_0085F0_SMX_ACTION_ENA(1);
194	}
195
196	/* Don't use the CB CP COHER logic on r6xx.
197	 * There are hw bugs.
198	 */
199	if (rctx->b.chip_class >= R700 &&
200	    (rctx->b.flags & R600_CONTEXT_FLUSH_AND_INV_CB)) {
201		cp_coher_cntl |= S_0085F0_CB_ACTION_ENA(1) |
202				S_0085F0_CB0_DEST_BASE_ENA(1) |
203				S_0085F0_CB1_DEST_BASE_ENA(1) |
204				S_0085F0_CB2_DEST_BASE_ENA(1) |
205				S_0085F0_CB3_DEST_BASE_ENA(1) |
206				S_0085F0_CB4_DEST_BASE_ENA(1) |
207				S_0085F0_CB5_DEST_BASE_ENA(1) |
208				S_0085F0_CB6_DEST_BASE_ENA(1) |
209				S_0085F0_CB7_DEST_BASE_ENA(1) |
210				S_0085F0_SMX_ACTION_ENA(1);
211		if (rctx->b.chip_class >= EVERGREEN)
212			cp_coher_cntl |= S_0085F0_CB8_DEST_BASE_ENA(1) |
213					S_0085F0_CB9_DEST_BASE_ENA(1) |
214					S_0085F0_CB10_DEST_BASE_ENA(1) |
215					S_0085F0_CB11_DEST_BASE_ENA(1);
216	}
217
218	if (rctx->b.chip_class >= R700 &&
219	    rctx->b.flags & R600_CONTEXT_STREAMOUT_FLUSH) {
220		cp_coher_cntl |= S_0085F0_SO0_DEST_BASE_ENA(1) |
221				S_0085F0_SO1_DEST_BASE_ENA(1) |
222				S_0085F0_SO2_DEST_BASE_ENA(1) |
223				S_0085F0_SO3_DEST_BASE_ENA(1) |
224				S_0085F0_SMX_ACTION_ENA(1);
225	}
226
227	/* Workaround for buggy flushing on some R6xx chipsets. */
228	if ((rctx->b.flags & (R600_CONTEXT_FLUSH_AND_INV |
229			      R600_CONTEXT_STREAMOUT_FLUSH)) &&
230	    (rctx->b.family == CHIP_RV670 ||
231	     rctx->b.family == CHIP_RS780 ||
232	     rctx->b.family == CHIP_RS880)) {
233		cp_coher_cntl |=  S_0085F0_CB1_DEST_BASE_ENA(1) |
234				  S_0085F0_DEST_BASE_0_ENA(1);
235	}
236
237	if (cp_coher_cntl) {
238		radeon_emit(cs, PKT3(PKT3_SURFACE_SYNC, 3, 0));
239		radeon_emit(cs, cp_coher_cntl);   /* CP_COHER_CNTL */
240		radeon_emit(cs, 0xffffffff);      /* CP_COHER_SIZE */
241		radeon_emit(cs, 0);               /* CP_COHER_BASE */
242		radeon_emit(cs, 0x0000000A);      /* POLL_INTERVAL */
243	}
244
245	if (rctx->b.flags & R600_CONTEXT_START_PIPELINE_STATS) {
246		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
247		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_START) |
248			        EVENT_INDEX(0));
249	} else if (rctx->b.flags & R600_CONTEXT_STOP_PIPELINE_STATS) {
250		radeon_emit(cs, PKT3(PKT3_EVENT_WRITE, 0, 0));
251		radeon_emit(cs, EVENT_TYPE(EVENT_TYPE_PIPELINESTAT_STOP) |
252			        EVENT_INDEX(0));
253	}
254
255	/* everything is properly flushed */
256	rctx->b.flags = 0;
257}
258
259void r600_context_gfx_flush(void *context, unsigned flags,
260			    struct pipe_fence_handle **fence)
261{
262	struct r600_context *ctx = context;
263	struct radeon_cmdbuf *cs = &ctx->b.gfx.cs;
264	struct radeon_winsys *ws = ctx->b.ws;
265
266	if (!radeon_emitted(cs, ctx->b.initial_gfx_cs_size))
267		return;
268
269	if (r600_check_device_reset(&ctx->b))
270		return;
271
272	r600_preflush_suspend_features(&ctx->b);
273
274	/* flush the framebuffer cache */
275	ctx->b.flags |= R600_CONTEXT_FLUSH_AND_INV |
276		      R600_CONTEXT_FLUSH_AND_INV_CB |
277		      R600_CONTEXT_FLUSH_AND_INV_DB |
278		      R600_CONTEXT_FLUSH_AND_INV_CB_META |
279		      R600_CONTEXT_FLUSH_AND_INV_DB_META |
280		      R600_CONTEXT_WAIT_3D_IDLE |
281		      R600_CONTEXT_WAIT_CP_DMA_IDLE;
282
283	r600_flush_emit(ctx);
284
285	if (ctx->trace_buf)
286		eg_trace_emit(ctx);
287	/* old kernels and userspace don't set SX_MISC, so we must reset it to 0 here */
288	if (ctx->b.chip_class == R600) {
289		radeon_set_context_reg(cs, R_028350_SX_MISC, 0);
290	}
291
292	if (ctx->is_debug) {
293		/* Save the IB for debug contexts. */
294		radeon_clear_saved_cs(&ctx->last_gfx);
295		radeon_save_cs(ws, cs, &ctx->last_gfx, true);
296		r600_resource_reference(&ctx->last_trace_buf, ctx->trace_buf);
297		r600_resource_reference(&ctx->trace_buf, NULL);
298	}
299	/* Flush the CS. */
300	ws->cs_flush(cs, flags, &ctx->b.last_gfx_fence);
301	if (fence)
302		ws->fence_reference(fence, ctx->b.last_gfx_fence);
303	ctx->b.num_gfx_cs_flushes++;
304
305	if (ctx->is_debug) {
306		if (!ws->fence_wait(ws, ctx->b.last_gfx_fence, 10000000)) {
307			const char *fname = getenv("R600_TRACE");
308			if (!fname)
309				exit(-1);
310			FILE *fl = fopen(fname, "w+");
311			if (fl) {
312				eg_dump_debug_state(&ctx->b.b, fl, 0);
313				fclose(fl);
314			} else
315				perror(fname);
316			exit(-1);
317		}
318	}
319	r600_begin_new_cs(ctx);
320}
321
322void r600_begin_new_cs(struct r600_context *ctx)
323{
324	unsigned shader;
325
326	if (ctx->is_debug) {
327		uint32_t zero = 0;
328
329		/* Create a buffer used for writing trace IDs and initialize it to 0. */
330		assert(!ctx->trace_buf);
331		ctx->trace_buf = (struct r600_resource*)
332			pipe_buffer_create(ctx->b.b.screen, 0,
333					   PIPE_USAGE_STAGING, 4);
334		if (ctx->trace_buf)
335			pipe_buffer_write_nooverlap(&ctx->b.b, &ctx->trace_buf->b.b,
336						    0, sizeof(zero), &zero);
337		ctx->trace_id = 0;
338	}
339
340	if (ctx->trace_buf)
341		eg_trace_emit(ctx);
342
343	ctx->b.flags = 0;
344	ctx->b.gtt = 0;
345	ctx->b.vram = 0;
346
347	/* Begin a new CS. */
348	r600_emit_command_buffer(&ctx->b.gfx.cs, &ctx->start_cs_cmd);
349
350	/* Re-emit states. */
351	r600_mark_atom_dirty(ctx, &ctx->alphatest_state.atom);
352	r600_mark_atom_dirty(ctx, &ctx->blend_color.atom);
353	r600_mark_atom_dirty(ctx, &ctx->cb_misc_state.atom);
354	r600_mark_atom_dirty(ctx, &ctx->clip_misc_state.atom);
355	r600_mark_atom_dirty(ctx, &ctx->clip_state.atom);
356	r600_mark_atom_dirty(ctx, &ctx->db_misc_state.atom);
357	r600_mark_atom_dirty(ctx, &ctx->db_state.atom);
358	r600_mark_atom_dirty(ctx, &ctx->framebuffer.atom);
359	if (ctx->b.chip_class >= EVERGREEN) {
360		r600_mark_atom_dirty(ctx, &ctx->fragment_images.atom);
361		r600_mark_atom_dirty(ctx, &ctx->fragment_buffers.atom);
362		r600_mark_atom_dirty(ctx, &ctx->compute_images.atom);
363		r600_mark_atom_dirty(ctx, &ctx->compute_buffers.atom);
364	}
365	r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_PS].atom);
366	r600_mark_atom_dirty(ctx, &ctx->poly_offset_state.atom);
367	r600_mark_atom_dirty(ctx, &ctx->vgt_state.atom);
368	r600_mark_atom_dirty(ctx, &ctx->sample_mask.atom);
369	ctx->b.scissors.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
370	r600_mark_atom_dirty(ctx, &ctx->b.scissors.atom);
371	ctx->b.viewports.dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
372	ctx->b.viewports.depth_range_dirty_mask = (1 << R600_MAX_VIEWPORTS) - 1;
373	r600_mark_atom_dirty(ctx, &ctx->b.viewports.atom);
374	if (ctx->b.chip_class <= EVERGREEN) {
375		r600_mark_atom_dirty(ctx, &ctx->config_state.atom);
376	}
377	r600_mark_atom_dirty(ctx, &ctx->stencil_ref.atom);
378	r600_mark_atom_dirty(ctx, &ctx->vertex_fetch_shader.atom);
379	r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_ES].atom);
380	r600_mark_atom_dirty(ctx, &ctx->shader_stages.atom);
381	if (ctx->gs_shader) {
382		r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_GS].atom);
383		r600_mark_atom_dirty(ctx, &ctx->gs_rings.atom);
384	}
385	if (ctx->tes_shader) {
386		r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[EG_HW_STAGE_HS].atom);
387		r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[EG_HW_STAGE_LS].atom);
388	}
389	r600_mark_atom_dirty(ctx, &ctx->hw_shader_stages[R600_HW_STAGE_VS].atom);
390	r600_mark_atom_dirty(ctx, &ctx->b.streamout.enable_atom);
391	r600_mark_atom_dirty(ctx, &ctx->b.render_cond_atom);
392
393	if (ctx->blend_state.cso)
394		r600_mark_atom_dirty(ctx, &ctx->blend_state.atom);
395	if (ctx->dsa_state.cso)
396		r600_mark_atom_dirty(ctx, &ctx->dsa_state.atom);
397	if (ctx->rasterizer_state.cso)
398		r600_mark_atom_dirty(ctx, &ctx->rasterizer_state.atom);
399
400	if (ctx->b.chip_class <= R700) {
401		r600_mark_atom_dirty(ctx, &ctx->seamless_cube_map.atom);
402	}
403
404	ctx->vertex_buffer_state.dirty_mask = ctx->vertex_buffer_state.enabled_mask;
405	r600_vertex_buffers_dirty(ctx);
406
407	/* Re-emit shader resources. */
408	for (shader = 0; shader < PIPE_SHADER_TYPES; shader++) {
409		struct r600_constbuf_state *constbuf = &ctx->constbuf_state[shader];
410		struct r600_textures_info *samplers = &ctx->samplers[shader];
411
412		constbuf->dirty_mask = constbuf->enabled_mask;
413		samplers->views.dirty_mask = samplers->views.enabled_mask;
414		samplers->states.dirty_mask = samplers->states.enabled_mask;
415
416		r600_constant_buffers_dirty(ctx, constbuf);
417		r600_sampler_views_dirty(ctx, &samplers->views);
418		r600_sampler_states_dirty(ctx, &samplers->states);
419	}
420
421	for (shader = 0; shader < ARRAY_SIZE(ctx->scratch_buffers); shader++) {
422		ctx->scratch_buffers[shader].dirty = true;
423	}
424
425	r600_postflush_resume_features(&ctx->b);
426
427	/* Re-emit the draw state. */
428	ctx->last_primitive_type = -1;
429	ctx->last_start_instance = -1;
430	ctx->last_rast_prim      = -1;
431	ctx->current_rast_prim   = -1;
432
433	assert(!ctx->b.gfx.cs.prev_dw);
434	ctx->b.initial_gfx_cs_size = ctx->b.gfx.cs.current.cdw;
435}
436
437void r600_emit_pfp_sync_me(struct r600_context *rctx)
438{
439	struct radeon_cmdbuf *cs = &rctx->b.gfx.cs;
440
441	if (rctx->b.chip_class >= EVERGREEN &&
442	    rctx->b.screen->info.drm_minor >= 46) {
443		radeon_emit(cs, PKT3(PKT3_PFP_SYNC_ME, 0, 0));
444		radeon_emit(cs, 0);
445	} else {
446		/* Emulate PFP_SYNC_ME by writing a value to memory in ME and
447		 * waiting for it in PFP.
448		 */
449		struct r600_resource *buf = NULL;
450		unsigned offset, reloc;
451		uint64_t va;
452
453		/* 16-byte address alignment is required by WAIT_REG_MEM. */
454		u_suballocator_alloc(&rctx->b.allocator_zeroed_memory, 4, 16,
455				     &offset, (struct pipe_resource**)&buf);
456		if (!buf) {
457			/* This is too heavyweight, but will work. */
458			rctx->b.gfx.flush(rctx, PIPE_FLUSH_ASYNC, NULL);
459			return;
460		}
461
462		reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, buf,
463						  RADEON_USAGE_READWRITE,
464						  RADEON_PRIO_FENCE);
465
466		va = buf->gpu_address + offset;
467		assert(va % 16 == 0);
468
469		/* Write 1 to memory in ME. */
470		radeon_emit(cs, PKT3(PKT3_MEM_WRITE, 3, 0));
471		radeon_emit(cs, va);
472		radeon_emit(cs, ((va >> 32) & 0xff) | MEM_WRITE_32_BITS);
473		radeon_emit(cs, 1);
474		radeon_emit(cs, 0);
475
476		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
477		radeon_emit(cs, reloc);
478
479		/* Wait in PFP (PFP can only do GEQUAL against memory). */
480		radeon_emit(cs, PKT3(PKT3_WAIT_REG_MEM, 5, 0));
481		radeon_emit(cs, WAIT_REG_MEM_GEQUAL |
482			        WAIT_REG_MEM_MEMORY |
483			        WAIT_REG_MEM_PFP);
484		radeon_emit(cs, va);
485		radeon_emit(cs, va >> 32);
486		radeon_emit(cs, 1); /* reference value */
487		radeon_emit(cs, 0xffffffff); /* mask */
488		radeon_emit(cs, 4); /* poll interval */
489
490		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
491		radeon_emit(cs, reloc);
492
493		r600_resource_reference(&buf, NULL);
494	}
495}
496
497/* The max number of bytes to copy per packet. */
498#define CP_DMA_MAX_BYTE_COUNT ((1 << 21) - 8)
499
500void r600_cp_dma_copy_buffer(struct r600_context *rctx,
501			     struct pipe_resource *dst, uint64_t dst_offset,
502			     struct pipe_resource *src, uint64_t src_offset,
503			     unsigned size)
504{
505	struct radeon_cmdbuf *cs = &rctx->b.gfx.cs;
506
507	assert(size);
508	assert(rctx->screen->b.has_cp_dma);
509
510	/* Mark the buffer range of destination as valid (initialized),
511	 * so that transfer_map knows it should wait for the GPU when mapping
512	 * that range. */
513	util_range_add(dst, &r600_resource(dst)->valid_buffer_range, dst_offset,
514		       dst_offset + size);
515
516	dst_offset += r600_resource(dst)->gpu_address;
517	src_offset += r600_resource(src)->gpu_address;
518
519	/* Flush the caches where the resources are bound. */
520	rctx->b.flags |= r600_get_flush_flags(R600_COHERENCY_SHADER) |
521			 R600_CONTEXT_WAIT_3D_IDLE;
522
523	/* There are differences between R700 and EG in CP DMA,
524	 * but we only use the common bits here. */
525	while (size) {
526		unsigned sync = 0;
527		unsigned byte_count = MIN2(size, CP_DMA_MAX_BYTE_COUNT);
528		unsigned src_reloc, dst_reloc;
529
530		r600_need_cs_space(rctx,
531				   10 + (rctx->b.flags ? R600_MAX_FLUSH_CS_DWORDS : 0) +
532				   3 + R600_MAX_PFP_SYNC_ME_DWORDS, FALSE, 0);
533
534		/* Flush the caches for the first copy only. */
535		if (rctx->b.flags) {
536			r600_flush_emit(rctx);
537		}
538
539		/* Do the synchronization after the last copy, so that all data is written to memory. */
540		if (size == byte_count) {
541			sync = PKT3_CP_DMA_CP_SYNC;
542		}
543
544		/* This must be done after r600_need_cs_space. */
545		src_reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, (struct r600_resource*)src,
546						  RADEON_USAGE_READ, RADEON_PRIO_CP_DMA);
547		dst_reloc = radeon_add_to_buffer_list(&rctx->b, &rctx->b.gfx, (struct r600_resource*)dst,
548						  RADEON_USAGE_WRITE, RADEON_PRIO_CP_DMA);
549
550		radeon_emit(cs, PKT3(PKT3_CP_DMA, 4, 0));
551		radeon_emit(cs, src_offset);	/* SRC_ADDR_LO [31:0] */
552		radeon_emit(cs, sync | ((src_offset >> 32) & 0xff));		/* CP_SYNC [31] | SRC_ADDR_HI [7:0] */
553		radeon_emit(cs, dst_offset);	/* DST_ADDR_LO [31:0] */
554		radeon_emit(cs, (dst_offset >> 32) & 0xff);		/* DST_ADDR_HI [7:0] */
555		radeon_emit(cs, byte_count);	/* COMMAND [29:22] | BYTE_COUNT [20:0] */
556
557		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
558		radeon_emit(cs, src_reloc);
559		radeon_emit(cs, PKT3(PKT3_NOP, 0, 0));
560		radeon_emit(cs, dst_reloc);
561
562		size -= byte_count;
563		src_offset += byte_count;
564		dst_offset += byte_count;
565	}
566
567	/* CP_DMA_CP_SYNC doesn't wait for idle on R6xx, but this does. */
568	if (rctx->b.chip_class == R600)
569		radeon_set_config_reg(cs, R_008040_WAIT_UNTIL,
570				      S_008040_WAIT_CP_DMA_IDLE(1));
571
572	/* CP DMA is executed in ME, but index buffers are read by PFP.
573	 * This ensures that ME (CP DMA) is idle before PFP starts fetching
574	 * indices. If we wanted to execute CP DMA in PFP, this packet
575	 * should precede it.
576	 */
577	r600_emit_pfp_sync_me(rctx);
578}
579
580void r600_dma_copy_buffer(struct r600_context *rctx,
581			  struct pipe_resource *dst,
582			  struct pipe_resource *src,
583			  uint64_t dst_offset,
584			  uint64_t src_offset,
585			  uint64_t size)
586{
587	struct radeon_cmdbuf *cs = &rctx->b.dma.cs;
588	unsigned i, ncopy, csize;
589	struct r600_resource *rdst = (struct r600_resource*)dst;
590	struct r600_resource *rsrc = (struct r600_resource*)src;
591
592	/* Mark the buffer range of destination as valid (initialized),
593	 * so that transfer_map knows it should wait for the GPU when mapping
594	 * that range. */
595	util_range_add(&rdst->b.b, &rdst->valid_buffer_range, dst_offset,
596		       dst_offset + size);
597
598	size >>= 2; /* convert to dwords */
599	ncopy = (size / R600_DMA_COPY_MAX_SIZE_DW) + !!(size % R600_DMA_COPY_MAX_SIZE_DW);
600
601	r600_need_dma_space(&rctx->b, ncopy * 5, rdst, rsrc);
602	for (i = 0; i < ncopy; i++) {
603		csize = size < R600_DMA_COPY_MAX_SIZE_DW ? size : R600_DMA_COPY_MAX_SIZE_DW;
604		/* emit reloc before writing cs so that cs is always in consistent state */
605		radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rsrc, RADEON_USAGE_READ, 0);
606		radeon_add_to_buffer_list(&rctx->b, &rctx->b.dma, rdst, RADEON_USAGE_WRITE, 0);
607		radeon_emit(cs, DMA_PACKET(DMA_PACKET_COPY, 0, 0, csize));
608		radeon_emit(cs, dst_offset & 0xfffffffc);
609		radeon_emit(cs, src_offset & 0xfffffffc);
610		radeon_emit(cs, (dst_offset >> 32UL) & 0xff);
611		radeon_emit(cs, (src_offset >> 32UL) & 0xff);
612		dst_offset += csize << 2;
613		src_offset += csize << 2;
614		size -= csize;
615	}
616}
617