1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014 Intel Corporation
4  */
5 
6 #include "gen8_engine_cs.h"
7 #include "i915_drv.h"
8 #include "intel_lrc.h"
9 #include "intel_gpu_commands.h"
10 #include "intel_ring.h"
11 
12 int gen8_emit_flush_rcs(struct i915_request *rq, u32 mode)
13 {
14 	bool vf_flush_wa = false, dc_flush_wa = false;
15 	u32 *cs, flags = 0;
16 	int len;
17 
18 	flags |= PIPE_CONTROL_CS_STALL;
19 
20 	if (mode & EMIT_FLUSH) {
21 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
22 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
23 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
24 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
25 	}
26 
27 	if (mode & EMIT_INVALIDATE) {
28 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
29 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
30 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
31 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
32 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
33 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
34 		flags |= PIPE_CONTROL_QW_WRITE;
35 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
36 
37 		/*
38 		 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
39 		 * pipe control.
40 		 */
41 		if (IS_GEN(rq->engine->i915, 9))
42 			vf_flush_wa = true;
43 
44 		/* WaForGAMHang:kbl */
45 		if (IS_KBL_GT_REVID(rq->engine->i915, 0, KBL_REVID_B0))
46 			dc_flush_wa = true;
47 	}
48 
49 	len = 6;
50 
51 	if (vf_flush_wa)
52 		len += 6;
53 
54 	if (dc_flush_wa)
55 		len += 12;
56 
57 	cs = intel_ring_begin(rq, len);
58 	if (IS_ERR(cs))
59 		return PTR_ERR(cs);
60 
61 	if (vf_flush_wa)
62 		cs = gen8_emit_pipe_control(cs, 0, 0);
63 
64 	if (dc_flush_wa)
65 		cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_DC_FLUSH_ENABLE,
66 					    0);
67 
68 	cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
69 
70 	if (dc_flush_wa)
71 		cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_CS_STALL, 0);
72 
73 	intel_ring_advance(rq, cs);
74 
75 	return 0;
76 }
77 
78 int gen8_emit_flush_xcs(struct i915_request *rq, u32 mode)
79 {
80 	u32 cmd, *cs;
81 
82 	cs = intel_ring_begin(rq, 4);
83 	if (IS_ERR(cs))
84 		return PTR_ERR(cs);
85 
86 	cmd = MI_FLUSH_DW + 1;
87 
88 	/*
89 	 * We always require a command barrier so that subsequent
90 	 * commands, such as breadcrumb interrupts, are strictly ordered
91 	 * wrt the contents of the write cache being flushed to memory
92 	 * (and thus being coherent from the CPU).
93 	 */
94 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
95 
96 	if (mode & EMIT_INVALIDATE) {
97 		cmd |= MI_INVALIDATE_TLB;
98 		if (rq->engine->class == VIDEO_DECODE_CLASS)
99 			cmd |= MI_INVALIDATE_BSD;
100 	}
101 
102 	*cs++ = cmd;
103 	*cs++ = LRC_PPHWSP_SCRATCH_ADDR;
104 	*cs++ = 0; /* upper addr */
105 	*cs++ = 0; /* value */
106 	intel_ring_advance(rq, cs);
107 
108 	return 0;
109 }
110 
111 int gen11_emit_flush_rcs(struct i915_request *rq, u32 mode)
112 {
113 	if (mode & EMIT_FLUSH) {
114 		u32 *cs;
115 		u32 flags = 0;
116 
117 		flags |= PIPE_CONTROL_CS_STALL;
118 
119 		flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
120 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
121 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
122 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
123 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
124 		flags |= PIPE_CONTROL_QW_WRITE;
125 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
126 
127 		cs = intel_ring_begin(rq, 6);
128 		if (IS_ERR(cs))
129 			return PTR_ERR(cs);
130 
131 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
132 		intel_ring_advance(rq, cs);
133 	}
134 
135 	if (mode & EMIT_INVALIDATE) {
136 		u32 *cs;
137 		u32 flags = 0;
138 
139 		flags |= PIPE_CONTROL_CS_STALL;
140 
141 		flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
142 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
143 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
144 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
145 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
146 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
147 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
148 		flags |= PIPE_CONTROL_QW_WRITE;
149 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
150 
151 		cs = intel_ring_begin(rq, 6);
152 		if (IS_ERR(cs))
153 			return PTR_ERR(cs);
154 
155 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
156 		intel_ring_advance(rq, cs);
157 	}
158 
159 	return 0;
160 }
161 
162 static u32 preparser_disable(bool state)
163 {
164 	return MI_ARB_CHECK | 1 << 8 | state;
165 }
166 
167 static i915_reg_t aux_inv_reg(const struct intel_engine_cs *engine)
168 {
169 	static const i915_reg_t vd[] = {
170 		GEN12_VD0_AUX_NV,
171 		GEN12_VD1_AUX_NV,
172 		GEN12_VD2_AUX_NV,
173 		GEN12_VD3_AUX_NV,
174 	};
175 
176 	static const i915_reg_t ve[] = {
177 		GEN12_VE0_AUX_NV,
178 		GEN12_VE1_AUX_NV,
179 	};
180 
181 	if (engine->class == VIDEO_DECODE_CLASS)
182 		return vd[engine->instance];
183 
184 	if (engine->class == VIDEO_ENHANCEMENT_CLASS)
185 		return ve[engine->instance];
186 
187 	GEM_BUG_ON("unknown aux_inv reg\n");
188 	return INVALID_MMIO_REG;
189 }
190 
191 static u32 *gen12_emit_aux_table_inv(const i915_reg_t inv_reg, u32 *cs)
192 {
193 	*cs++ = MI_LOAD_REGISTER_IMM(1);
194 	*cs++ = i915_mmio_reg_offset(inv_reg);
195 	*cs++ = AUX_INV;
196 	*cs++ = MI_NOOP;
197 
198 	return cs;
199 }
200 
201 int gen12_emit_flush_rcs(struct i915_request *rq, u32 mode)
202 {
203 	if (mode & EMIT_FLUSH) {
204 		u32 flags = 0;
205 		u32 *cs;
206 
207 		flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
208 		flags |= PIPE_CONTROL_FLUSH_L3;
209 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
210 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
211 		/* Wa_1409600907:tgl */
212 		flags |= PIPE_CONTROL_DEPTH_STALL;
213 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
214 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
215 
216 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
217 		flags |= PIPE_CONTROL_QW_WRITE;
218 
219 		flags |= PIPE_CONTROL_CS_STALL;
220 
221 		cs = intel_ring_begin(rq, 6);
222 		if (IS_ERR(cs))
223 			return PTR_ERR(cs);
224 
225 		cs = gen12_emit_pipe_control(cs,
226 					     PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
227 					     flags, LRC_PPHWSP_SCRATCH_ADDR);
228 		intel_ring_advance(rq, cs);
229 	}
230 
231 	if (mode & EMIT_INVALIDATE) {
232 		u32 flags = 0;
233 		u32 *cs;
234 
235 		flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
236 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
237 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
238 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
239 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
240 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
241 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
242 
243 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
244 		flags |= PIPE_CONTROL_QW_WRITE;
245 
246 		flags |= PIPE_CONTROL_CS_STALL;
247 
248 		cs = intel_ring_begin(rq, 8 + 4);
249 		if (IS_ERR(cs))
250 			return PTR_ERR(cs);
251 
252 		/*
253 		 * Prevent the pre-parser from skipping past the TLB
254 		 * invalidate and loading a stale page for the batch
255 		 * buffer / request payload.
256 		 */
257 		*cs++ = preparser_disable(true);
258 
259 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
260 
261 		/* hsdes: 1809175790 */
262 		cs = gen12_emit_aux_table_inv(GEN12_GFX_CCS_AUX_NV, cs);
263 
264 		*cs++ = preparser_disable(false);
265 		intel_ring_advance(rq, cs);
266 	}
267 
268 	return 0;
269 }
270 
271 int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
272 {
273 	intel_engine_mask_t aux_inv = 0;
274 	u32 cmd, *cs;
275 
276 	cmd = 4;
277 	if (mode & EMIT_INVALIDATE)
278 		cmd += 2;
279 	if (mode & EMIT_INVALIDATE)
280 		aux_inv = rq->engine->mask & ~BIT(BCS0);
281 	if (aux_inv)
282 		cmd += 2 * hweight8(aux_inv) + 2;
283 
284 	cs = intel_ring_begin(rq, cmd);
285 	if (IS_ERR(cs))
286 		return PTR_ERR(cs);
287 
288 	if (mode & EMIT_INVALIDATE)
289 		*cs++ = preparser_disable(true);
290 
291 	cmd = MI_FLUSH_DW + 1;
292 
293 	/*
294 	 * We always require a command barrier so that subsequent
295 	 * commands, such as breadcrumb interrupts, are strictly ordered
296 	 * wrt the contents of the write cache being flushed to memory
297 	 * (and thus being coherent from the CPU).
298 	 */
299 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
300 
301 	if (mode & EMIT_INVALIDATE) {
302 		cmd |= MI_INVALIDATE_TLB;
303 		if (rq->engine->class == VIDEO_DECODE_CLASS)
304 			cmd |= MI_INVALIDATE_BSD;
305 	}
306 
307 	*cs++ = cmd;
308 	*cs++ = LRC_PPHWSP_SCRATCH_ADDR;
309 	*cs++ = 0; /* upper addr */
310 	*cs++ = 0; /* value */
311 
312 	if (aux_inv) { /* hsdes: 1809175790 */
313 		struct intel_engine_cs *engine;
314 		unsigned int tmp;
315 
316 		*cs++ = MI_LOAD_REGISTER_IMM(hweight8(aux_inv));
317 		for_each_engine_masked(engine, rq->engine->gt,
318 				       aux_inv, tmp) {
319 			*cs++ = i915_mmio_reg_offset(aux_inv_reg(engine));
320 			*cs++ = AUX_INV;
321 		}
322 		*cs++ = MI_NOOP;
323 	}
324 
325 	if (mode & EMIT_INVALIDATE)
326 		*cs++ = preparser_disable(false);
327 
328 	intel_ring_advance(rq, cs);
329 
330 	return 0;
331 }
332 
333 static u32 preempt_address(struct intel_engine_cs *engine)
334 {
335 	return (i915_ggtt_offset(engine->status_page.vma) +
336 		I915_GEM_HWS_PREEMPT_ADDR);
337 }
338 
339 static u32 hwsp_offset(const struct i915_request *rq)
340 {
341 	const struct intel_timeline_cacheline *cl;
342 
343 	/* Before the request is executed, the timeline/cachline is fixed */
344 
345 	cl = rcu_dereference_protected(rq->hwsp_cacheline, 1);
346 	if (cl)
347 		return cl->ggtt_offset;
348 
349 	return rcu_dereference_protected(rq->timeline, 1)->hwsp_offset;
350 }
351 
352 int gen8_emit_init_breadcrumb(struct i915_request *rq)
353 {
354 	u32 *cs;
355 
356 	GEM_BUG_ON(i915_request_has_initial_breadcrumb(rq));
357 	if (!i915_request_timeline(rq)->has_initial_breadcrumb)
358 		return 0;
359 
360 	cs = intel_ring_begin(rq, 6);
361 	if (IS_ERR(cs))
362 		return PTR_ERR(cs);
363 
364 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
365 	*cs++ = hwsp_offset(rq);
366 	*cs++ = 0;
367 	*cs++ = rq->fence.seqno - 1;
368 
369 	/*
370 	 * Check if we have been preempted before we even get started.
371 	 *
372 	 * After this point i915_request_started() reports true, even if
373 	 * we get preempted and so are no longer running.
374 	 *
375 	 * i915_request_started() is used during preemption processing
376 	 * to decide if the request is currently inside the user payload
377 	 * or spinning on a kernel semaphore (or earlier). For no-preemption
378 	 * requests, we do allow preemption on the semaphore before the user
379 	 * payload, but do not allow preemption once the request is started.
380 	 *
381 	 * i915_request_started() is similarly used during GPU hangs to
382 	 * determine if the user's payload was guilty, and if so, the
383 	 * request is banned. Before the request is started, it is assumed
384 	 * to be unharmed and an innocent victim of another's hang.
385 	 */
386 	*cs++ = MI_NOOP;
387 	*cs++ = MI_ARB_CHECK;
388 
389 	intel_ring_advance(rq, cs);
390 
391 	/* Record the updated position of the request's payload */
392 	rq->infix = intel_ring_offset(rq, cs);
393 
394 	__set_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags);
395 
396 	return 0;
397 }
398 
399 int gen8_emit_bb_start_noarb(struct i915_request *rq,
400 			     u64 offset, u32 len,
401 			     const unsigned int flags)
402 {
403 	u32 *cs;
404 
405 	cs = intel_ring_begin(rq, 4);
406 	if (IS_ERR(cs))
407 		return PTR_ERR(cs);
408 
409 	/*
410 	 * WaDisableCtxRestoreArbitration:bdw,chv
411 	 *
412 	 * We don't need to perform MI_ARB_ENABLE as often as we do (in
413 	 * particular all the gen that do not need the w/a at all!), if we
414 	 * took care to make sure that on every switch into this context
415 	 * (both ordinary and for preemption) that arbitrartion was enabled
416 	 * we would be fine.  However, for gen8 there is another w/a that
417 	 * requires us to not preempt inside GPGPU execution, so we keep
418 	 * arbitration disabled for gen8 batches. Arbitration will be
419 	 * re-enabled before we close the request
420 	 * (engine->emit_fini_breadcrumb).
421 	 */
422 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
423 
424 	/* FIXME(BDW+): Address space and security selectors. */
425 	*cs++ = MI_BATCH_BUFFER_START_GEN8 |
426 		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
427 	*cs++ = lower_32_bits(offset);
428 	*cs++ = upper_32_bits(offset);
429 
430 	intel_ring_advance(rq, cs);
431 
432 	return 0;
433 }
434 
435 int gen8_emit_bb_start(struct i915_request *rq,
436 		       u64 offset, u32 len,
437 		       const unsigned int flags)
438 {
439 	u32 *cs;
440 
441 	if (unlikely(i915_request_has_nopreempt(rq)))
442 		return gen8_emit_bb_start_noarb(rq, offset, len, flags);
443 
444 	cs = intel_ring_begin(rq, 6);
445 	if (IS_ERR(cs))
446 		return PTR_ERR(cs);
447 
448 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
449 
450 	*cs++ = MI_BATCH_BUFFER_START_GEN8 |
451 		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
452 	*cs++ = lower_32_bits(offset);
453 	*cs++ = upper_32_bits(offset);
454 
455 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
456 	*cs++ = MI_NOOP;
457 
458 	intel_ring_advance(rq, cs);
459 
460 	return 0;
461 }
462 
463 static void assert_request_valid(struct i915_request *rq)
464 {
465 	struct intel_ring *ring __maybe_unused = rq->ring;
466 
467 	/* Can we unwind this request without appearing to go forwards? */
468 	GEM_BUG_ON(intel_ring_direction(ring, rq->wa_tail, rq->head) <= 0);
469 }
470 
471 /*
472  * Reserve space for 2 NOOPs at the end of each request to be
473  * used as a workaround for not being allowed to do lite
474  * restore with HEAD==TAIL (WaIdleLiteRestore).
475  */
476 static u32 *gen8_emit_wa_tail(struct i915_request *rq, u32 *cs)
477 {
478 	/* Ensure there's always at least one preemption point per-request. */
479 	*cs++ = MI_ARB_CHECK;
480 	*cs++ = MI_NOOP;
481 	rq->wa_tail = intel_ring_offset(rq, cs);
482 
483 	/* Check that entire request is less than half the ring */
484 	assert_request_valid(rq);
485 
486 	return cs;
487 }
488 
489 static u32 *emit_preempt_busywait(struct i915_request *rq, u32 *cs)
490 {
491 	*cs++ = MI_ARB_CHECK; /* trigger IDLE->ACTIVE first */
492 	*cs++ = MI_SEMAPHORE_WAIT |
493 		MI_SEMAPHORE_GLOBAL_GTT |
494 		MI_SEMAPHORE_POLL |
495 		MI_SEMAPHORE_SAD_EQ_SDD;
496 	*cs++ = 0;
497 	*cs++ = preempt_address(rq->engine);
498 	*cs++ = 0;
499 	*cs++ = MI_NOOP;
500 
501 	return cs;
502 }
503 
504 static __always_inline u32*
505 gen8_emit_fini_breadcrumb_tail(struct i915_request *rq, u32 *cs)
506 {
507 	*cs++ = MI_USER_INTERRUPT;
508 
509 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
510 	if (intel_engine_has_semaphores(rq->engine))
511 		cs = emit_preempt_busywait(rq, cs);
512 
513 	rq->tail = intel_ring_offset(rq, cs);
514 	assert_ring_tail_valid(rq->ring, rq->tail);
515 
516 	return gen8_emit_wa_tail(rq, cs);
517 }
518 
519 static u32 *emit_xcs_breadcrumb(struct i915_request *rq, u32 *cs)
520 {
521 	return gen8_emit_ggtt_write(cs, rq->fence.seqno, hwsp_offset(rq), 0);
522 }
523 
524 u32 *gen8_emit_fini_breadcrumb_xcs(struct i915_request *rq, u32 *cs)
525 {
526 	return gen8_emit_fini_breadcrumb_tail(rq, emit_xcs_breadcrumb(rq, cs));
527 }
528 
529 u32 *gen8_emit_fini_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
530 {
531 	cs = gen8_emit_pipe_control(cs,
532 				    PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
533 				    PIPE_CONTROL_DEPTH_CACHE_FLUSH |
534 				    PIPE_CONTROL_DC_FLUSH_ENABLE,
535 				    0);
536 
537 	/* XXX flush+write+CS_STALL all in one upsets gem_concurrent_blt:kbl */
538 	cs = gen8_emit_ggtt_write_rcs(cs,
539 				      rq->fence.seqno,
540 				      hwsp_offset(rq),
541 				      PIPE_CONTROL_FLUSH_ENABLE |
542 				      PIPE_CONTROL_CS_STALL);
543 
544 	return gen8_emit_fini_breadcrumb_tail(rq, cs);
545 }
546 
547 u32 *gen11_emit_fini_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
548 {
549 	cs = gen8_emit_ggtt_write_rcs(cs,
550 				      rq->fence.seqno,
551 				      hwsp_offset(rq),
552 				      PIPE_CONTROL_CS_STALL |
553 				      PIPE_CONTROL_TILE_CACHE_FLUSH |
554 				      PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
555 				      PIPE_CONTROL_DEPTH_CACHE_FLUSH |
556 				      PIPE_CONTROL_DC_FLUSH_ENABLE |
557 				      PIPE_CONTROL_FLUSH_ENABLE);
558 
559 	return gen8_emit_fini_breadcrumb_tail(rq, cs);
560 }
561 
562 /*
563  * Note that the CS instruction pre-parser will not stall on the breadcrumb
564  * flush and will continue pre-fetching the instructions after it before the
565  * memory sync is completed. On pre-gen12 HW, the pre-parser will stop at
566  * BB_START/END instructions, so, even though we might pre-fetch the pre-amble
567  * of the next request before the memory has been flushed, we're guaranteed that
568  * we won't access the batch itself too early.
569  * However, on gen12+ the parser can pre-fetch across the BB_START/END commands,
570  * so, if the current request is modifying an instruction in the next request on
571  * the same intel_context, we might pre-fetch and then execute the pre-update
572  * instruction. To avoid this, the users of self-modifying code should either
573  * disable the parser around the code emitting the memory writes, via a new flag
574  * added to MI_ARB_CHECK, or emit the writes from a different intel_context. For
575  * the in-kernel use-cases we've opted to use a separate context, see
576  * reloc_gpu() as an example.
577  * All the above applies only to the instructions themselves. Non-inline data
578  * used by the instructions is not pre-fetched.
579  */
580 
581 static u32 *gen12_emit_preempt_busywait(struct i915_request *rq, u32 *cs)
582 {
583 	*cs++ = MI_ARB_CHECK; /* trigger IDLE->ACTIVE first */
584 	*cs++ = MI_SEMAPHORE_WAIT_TOKEN |
585 		MI_SEMAPHORE_GLOBAL_GTT |
586 		MI_SEMAPHORE_POLL |
587 		MI_SEMAPHORE_SAD_EQ_SDD;
588 	*cs++ = 0;
589 	*cs++ = preempt_address(rq->engine);
590 	*cs++ = 0;
591 	*cs++ = 0;
592 
593 	return cs;
594 }
595 
596 static __always_inline u32*
597 gen12_emit_fini_breadcrumb_tail(struct i915_request *rq, u32 *cs)
598 {
599 	*cs++ = MI_USER_INTERRUPT;
600 
601 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
602 	if (intel_engine_has_semaphores(rq->engine))
603 		cs = gen12_emit_preempt_busywait(rq, cs);
604 
605 	rq->tail = intel_ring_offset(rq, cs);
606 	assert_ring_tail_valid(rq->ring, rq->tail);
607 
608 	return gen8_emit_wa_tail(rq, cs);
609 }
610 
611 u32 *gen12_emit_fini_breadcrumb_xcs(struct i915_request *rq, u32 *cs)
612 {
613 	/* XXX Stalling flush before seqno write; post-sync not */
614 	cs = emit_xcs_breadcrumb(rq, __gen8_emit_flush_dw(cs, 0, 0, 0));
615 	return gen12_emit_fini_breadcrumb_tail(rq, cs);
616 }
617 
618 u32 *gen12_emit_fini_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
619 {
620 	cs = gen12_emit_ggtt_write_rcs(cs,
621 				       rq->fence.seqno,
622 				       hwsp_offset(rq),
623 				       PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
624 				       PIPE_CONTROL_CS_STALL |
625 				       PIPE_CONTROL_TILE_CACHE_FLUSH |
626 				       PIPE_CONTROL_FLUSH_L3 |
627 				       PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
628 				       PIPE_CONTROL_DEPTH_CACHE_FLUSH |
629 				       /* Wa_1409600907:tgl */
630 				       PIPE_CONTROL_DEPTH_STALL |
631 				       PIPE_CONTROL_DC_FLUSH_ENABLE |
632 				       PIPE_CONTROL_FLUSH_ENABLE);
633 
634 	return gen12_emit_fini_breadcrumb_tail(rq, cs);
635 }
636