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_gpu_commands.h"
9 #include "intel_gt_regs.h"
10 #include "intel_lrc.h"
11 #include "intel_ring.h"
12 
13 int gen8_emit_flush_rcs(struct i915_request *rq, u32 mode)
14 {
15 	bool vf_flush_wa = false, dc_flush_wa = false;
16 	u32 *cs, flags = 0;
17 	int len;
18 
19 	flags |= PIPE_CONTROL_CS_STALL;
20 
21 	if (mode & EMIT_FLUSH) {
22 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
23 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
24 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
25 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
26 	}
27 
28 	if (mode & EMIT_INVALIDATE) {
29 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
30 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
31 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
32 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
33 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
34 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
35 		flags |= PIPE_CONTROL_QW_WRITE;
36 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
37 
38 		/*
39 		 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
40 		 * pipe control.
41 		 */
42 		if (GRAPHICS_VER(rq->engine->i915) == 9)
43 			vf_flush_wa = true;
44 
45 		/* WaForGAMHang:kbl */
46 		if (IS_KBL_GRAPHICS_STEP(rq->engine->i915, 0, STEP_C0))
47 			dc_flush_wa = true;
48 	}
49 
50 	len = 6;
51 
52 	if (vf_flush_wa)
53 		len += 6;
54 
55 	if (dc_flush_wa)
56 		len += 12;
57 
58 	cs = intel_ring_begin(rq, len);
59 	if (IS_ERR(cs))
60 		return PTR_ERR(cs);
61 
62 	if (vf_flush_wa)
63 		cs = gen8_emit_pipe_control(cs, 0, 0);
64 
65 	if (dc_flush_wa)
66 		cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_DC_FLUSH_ENABLE,
67 					    0);
68 
69 	cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
70 
71 	if (dc_flush_wa)
72 		cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_CS_STALL, 0);
73 
74 	intel_ring_advance(rq, cs);
75 
76 	return 0;
77 }
78 
79 int gen8_emit_flush_xcs(struct i915_request *rq, u32 mode)
80 {
81 	u32 cmd, *cs;
82 
83 	cs = intel_ring_begin(rq, 4);
84 	if (IS_ERR(cs))
85 		return PTR_ERR(cs);
86 
87 	cmd = MI_FLUSH_DW + 1;
88 
89 	/*
90 	 * We always require a command barrier so that subsequent
91 	 * commands, such as breadcrumb interrupts, are strictly ordered
92 	 * wrt the contents of the write cache being flushed to memory
93 	 * (and thus being coherent from the CPU).
94 	 */
95 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
96 
97 	if (mode & EMIT_INVALIDATE) {
98 		cmd |= MI_INVALIDATE_TLB;
99 		if (rq->engine->class == VIDEO_DECODE_CLASS)
100 			cmd |= MI_INVALIDATE_BSD;
101 	}
102 
103 	*cs++ = cmd;
104 	*cs++ = LRC_PPHWSP_SCRATCH_ADDR;
105 	*cs++ = 0; /* upper addr */
106 	*cs++ = 0; /* value */
107 	intel_ring_advance(rq, cs);
108 
109 	return 0;
110 }
111 
112 int gen11_emit_flush_rcs(struct i915_request *rq, u32 mode)
113 {
114 	if (mode & EMIT_FLUSH) {
115 		u32 *cs;
116 		u32 flags = 0;
117 
118 		flags |= PIPE_CONTROL_CS_STALL;
119 
120 		flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
121 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
122 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
123 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
124 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
125 		flags |= PIPE_CONTROL_QW_WRITE;
126 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
127 
128 		cs = intel_ring_begin(rq, 6);
129 		if (IS_ERR(cs))
130 			return PTR_ERR(cs);
131 
132 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
133 		intel_ring_advance(rq, cs);
134 	}
135 
136 	if (mode & EMIT_INVALIDATE) {
137 		u32 *cs;
138 		u32 flags = 0;
139 
140 		flags |= PIPE_CONTROL_CS_STALL;
141 
142 		flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
143 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
144 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
145 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
146 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
147 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
148 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
149 		flags |= PIPE_CONTROL_QW_WRITE;
150 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
151 
152 		cs = intel_ring_begin(rq, 6);
153 		if (IS_ERR(cs))
154 			return PTR_ERR(cs);
155 
156 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
157 		intel_ring_advance(rq, cs);
158 	}
159 
160 	return 0;
161 }
162 
163 static u32 preparser_disable(bool state)
164 {
165 	return MI_ARB_CHECK | 1 << 8 | state;
166 }
167 
168 static i915_reg_t aux_inv_reg(const struct intel_engine_cs *engine)
169 {
170 	static const i915_reg_t vd[] = {
171 		GEN12_VD0_AUX_NV,
172 		GEN12_VD1_AUX_NV,
173 		GEN12_VD2_AUX_NV,
174 		GEN12_VD3_AUX_NV,
175 	};
176 
177 	static const i915_reg_t ve[] = {
178 		GEN12_VE0_AUX_NV,
179 		GEN12_VE1_AUX_NV,
180 	};
181 
182 	if (engine->class == VIDEO_DECODE_CLASS)
183 		return vd[engine->instance];
184 
185 	if (engine->class == VIDEO_ENHANCEMENT_CLASS)
186 		return ve[engine->instance];
187 
188 	GEM_BUG_ON("unknown aux_inv reg\n");
189 	return INVALID_MMIO_REG;
190 }
191 
192 static u32 *gen12_emit_aux_table_inv(const i915_reg_t inv_reg, u32 *cs)
193 {
194 	*cs++ = MI_LOAD_REGISTER_IMM(1);
195 	*cs++ = i915_mmio_reg_offset(inv_reg);
196 	*cs++ = AUX_INV;
197 	*cs++ = MI_NOOP;
198 
199 	return cs;
200 }
201 
202 int gen12_emit_flush_rcs(struct i915_request *rq, u32 mode)
203 {
204 	if (mode & EMIT_FLUSH) {
205 		u32 flags = 0;
206 		u32 *cs;
207 
208 		flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
209 		flags |= PIPE_CONTROL_FLUSH_L3;
210 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
211 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
212 		/* Wa_1409600907:tgl,adl-p */
213 		flags |= PIPE_CONTROL_DEPTH_STALL;
214 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
215 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
216 
217 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
218 		flags |= PIPE_CONTROL_QW_WRITE;
219 
220 		flags |= PIPE_CONTROL_CS_STALL;
221 
222 		cs = intel_ring_begin(rq, 6);
223 		if (IS_ERR(cs))
224 			return PTR_ERR(cs);
225 
226 		cs = gen12_emit_pipe_control(cs,
227 					     PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
228 					     flags, LRC_PPHWSP_SCRATCH_ADDR);
229 		intel_ring_advance(rq, cs);
230 	}
231 
232 	if (mode & EMIT_INVALIDATE) {
233 		u32 flags = 0;
234 		u32 *cs;
235 
236 		flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
237 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
238 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
239 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
240 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
241 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
242 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
243 
244 		flags |= PIPE_CONTROL_STORE_DATA_INDEX;
245 		flags |= PIPE_CONTROL_QW_WRITE;
246 
247 		flags |= PIPE_CONTROL_CS_STALL;
248 
249 		cs = intel_ring_begin(rq, 8 + 4);
250 		if (IS_ERR(cs))
251 			return PTR_ERR(cs);
252 
253 		/*
254 		 * Prevent the pre-parser from skipping past the TLB
255 		 * invalidate and loading a stale page for the batch
256 		 * buffer / request payload.
257 		 */
258 		*cs++ = preparser_disable(true);
259 
260 		cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
261 
262 		/* hsdes: 1809175790 */
263 		cs = gen12_emit_aux_table_inv(GEN12_GFX_CCS_AUX_NV, cs);
264 
265 		*cs++ = preparser_disable(false);
266 		intel_ring_advance(rq, cs);
267 	}
268 
269 	return 0;
270 }
271 
272 int gen12_emit_flush_xcs(struct i915_request *rq, u32 mode)
273 {
274 	intel_engine_mask_t aux_inv = 0;
275 	u32 cmd, *cs;
276 
277 	cmd = 4;
278 	if (mode & EMIT_INVALIDATE)
279 		cmd += 2;
280 	if (mode & EMIT_INVALIDATE)
281 		aux_inv = rq->engine->mask & ~BIT(BCS0);
282 	if (aux_inv)
283 		cmd += 2 * hweight32(aux_inv) + 2;
284 
285 	cs = intel_ring_begin(rq, cmd);
286 	if (IS_ERR(cs))
287 		return PTR_ERR(cs);
288 
289 	if (mode & EMIT_INVALIDATE)
290 		*cs++ = preparser_disable(true);
291 
292 	cmd = MI_FLUSH_DW + 1;
293 
294 	/*
295 	 * We always require a command barrier so that subsequent
296 	 * commands, such as breadcrumb interrupts, are strictly ordered
297 	 * wrt the contents of the write cache being flushed to memory
298 	 * (and thus being coherent from the CPU).
299 	 */
300 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
301 
302 	if (mode & EMIT_INVALIDATE) {
303 		cmd |= MI_INVALIDATE_TLB;
304 		if (rq->engine->class == VIDEO_DECODE_CLASS)
305 			cmd |= MI_INVALIDATE_BSD;
306 	}
307 
308 	*cs++ = cmd;
309 	*cs++ = LRC_PPHWSP_SCRATCH_ADDR;
310 	*cs++ = 0; /* upper addr */
311 	*cs++ = 0; /* value */
312 
313 	if (aux_inv) { /* hsdes: 1809175790 */
314 		struct intel_engine_cs *engine;
315 		unsigned int tmp;
316 
317 		*cs++ = MI_LOAD_REGISTER_IMM(hweight32(aux_inv));
318 		for_each_engine_masked(engine, rq->engine->gt, 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 *tl;
342 
343 	/* Before the request is executed, the timeline is fixed */
344 	tl = rcu_dereference_protected(rq->timeline,
345 				       !i915_request_signaled(rq));
346 
347 	/* See the comment in i915_request_active_seqno(). */
348 	return page_mask_bits(tl->hwsp_offset) + offset_in_page(rq->hwsp_seqno);
349 }
350 
351 int gen8_emit_init_breadcrumb(struct i915_request *rq)
352 {
353 	u32 *cs;
354 
355 	GEM_BUG_ON(i915_request_has_initial_breadcrumb(rq));
356 	if (!i915_request_timeline(rq)->has_initial_breadcrumb)
357 		return 0;
358 
359 	cs = intel_ring_begin(rq, 6);
360 	if (IS_ERR(cs))
361 		return PTR_ERR(cs);
362 
363 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
364 	*cs++ = hwsp_offset(rq);
365 	*cs++ = 0;
366 	*cs++ = rq->fence.seqno - 1;
367 
368 	/*
369 	 * Check if we have been preempted before we even get started.
370 	 *
371 	 * After this point i915_request_started() reports true, even if
372 	 * we get preempted and so are no longer running.
373 	 *
374 	 * i915_request_started() is used during preemption processing
375 	 * to decide if the request is currently inside the user payload
376 	 * or spinning on a kernel semaphore (or earlier). For no-preemption
377 	 * requests, we do allow preemption on the semaphore before the user
378 	 * payload, but do not allow preemption once the request is started.
379 	 *
380 	 * i915_request_started() is similarly used during GPU hangs to
381 	 * determine if the user's payload was guilty, and if so, the
382 	 * request is banned. Before the request is started, it is assumed
383 	 * to be unharmed and an innocent victim of another's hang.
384 	 */
385 	*cs++ = MI_NOOP;
386 	*cs++ = MI_ARB_CHECK;
387 
388 	intel_ring_advance(rq, cs);
389 
390 	/* Record the updated position of the request's payload */
391 	rq->infix = intel_ring_offset(rq, cs);
392 
393 	__set_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags);
394 
395 	return 0;
396 }
397 
398 int gen8_emit_bb_start_noarb(struct i915_request *rq,
399 			     u64 offset, u32 len,
400 			     const unsigned int flags)
401 {
402 	u32 *cs;
403 
404 	cs = intel_ring_begin(rq, 4);
405 	if (IS_ERR(cs))
406 		return PTR_ERR(cs);
407 
408 	/*
409 	 * WaDisableCtxRestoreArbitration:bdw,chv
410 	 *
411 	 * We don't need to perform MI_ARB_ENABLE as often as we do (in
412 	 * particular all the gen that do not need the w/a at all!), if we
413 	 * took care to make sure that on every switch into this context
414 	 * (both ordinary and for preemption) that arbitrartion was enabled
415 	 * we would be fine.  However, for gen8 there is another w/a that
416 	 * requires us to not preempt inside GPGPU execution, so we keep
417 	 * arbitration disabled for gen8 batches. Arbitration will be
418 	 * re-enabled before we close the request
419 	 * (engine->emit_fini_breadcrumb).
420 	 */
421 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
422 
423 	/* FIXME(BDW+): Address space and security selectors. */
424 	*cs++ = MI_BATCH_BUFFER_START_GEN8 |
425 		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
426 	*cs++ = lower_32_bits(offset);
427 	*cs++ = upper_32_bits(offset);
428 
429 	intel_ring_advance(rq, cs);
430 
431 	return 0;
432 }
433 
434 int gen8_emit_bb_start(struct i915_request *rq,
435 		       u64 offset, u32 len,
436 		       const unsigned int flags)
437 {
438 	u32 *cs;
439 
440 	if (unlikely(i915_request_has_nopreempt(rq)))
441 		return gen8_emit_bb_start_noarb(rq, offset, len, flags);
442 
443 	cs = intel_ring_begin(rq, 6);
444 	if (IS_ERR(cs))
445 		return PTR_ERR(cs);
446 
447 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
448 
449 	*cs++ = MI_BATCH_BUFFER_START_GEN8 |
450 		(flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
451 	*cs++ = lower_32_bits(offset);
452 	*cs++ = upper_32_bits(offset);
453 
454 	*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
455 	*cs++ = MI_NOOP;
456 
457 	intel_ring_advance(rq, cs);
458 
459 	return 0;
460 }
461 
462 static void assert_request_valid(struct i915_request *rq)
463 {
464 	struct intel_ring *ring __maybe_unused = rq->ring;
465 
466 	/* Can we unwind this request without appearing to go forwards? */
467 	GEM_BUG_ON(intel_ring_direction(ring, rq->wa_tail, rq->head) <= 0);
468 }
469 
470 /*
471  * Reserve space for 2 NOOPs at the end of each request to be
472  * used as a workaround for not being allowed to do lite
473  * restore with HEAD==TAIL (WaIdleLiteRestore).
474  */
475 static u32 *gen8_emit_wa_tail(struct i915_request *rq, u32 *cs)
476 {
477 	/* Ensure there's always at least one preemption point per-request. */
478 	*cs++ = MI_ARB_CHECK;
479 	*cs++ = MI_NOOP;
480 	rq->wa_tail = intel_ring_offset(rq, cs);
481 
482 	/* Check that entire request is less than half the ring */
483 	assert_request_valid(rq);
484 
485 	return cs;
486 }
487 
488 static u32 *emit_preempt_busywait(struct i915_request *rq, u32 *cs)
489 {
490 	*cs++ = MI_ARB_CHECK; /* trigger IDLE->ACTIVE first */
491 	*cs++ = MI_SEMAPHORE_WAIT |
492 		MI_SEMAPHORE_GLOBAL_GTT |
493 		MI_SEMAPHORE_POLL |
494 		MI_SEMAPHORE_SAD_EQ_SDD;
495 	*cs++ = 0;
496 	*cs++ = preempt_address(rq->engine);
497 	*cs++ = 0;
498 	*cs++ = MI_NOOP;
499 
500 	return cs;
501 }
502 
503 static __always_inline u32*
504 gen8_emit_fini_breadcrumb_tail(struct i915_request *rq, u32 *cs)
505 {
506 	*cs++ = MI_USER_INTERRUPT;
507 
508 	*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
509 	if (intel_engine_has_semaphores(rq->engine) &&
510 	    !intel_uc_uses_guc_submission(&rq->engine->gt->uc))
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 	    !intel_uc_uses_guc_submission(&rq->engine->gt->uc))
604 		cs = gen12_emit_preempt_busywait(rq, cs);
605 
606 	rq->tail = intel_ring_offset(rq, cs);
607 	assert_ring_tail_valid(rq->ring, rq->tail);
608 
609 	return gen8_emit_wa_tail(rq, cs);
610 }
611 
612 u32 *gen12_emit_fini_breadcrumb_xcs(struct i915_request *rq, u32 *cs)
613 {
614 	/* XXX Stalling flush before seqno write; post-sync not */
615 	cs = emit_xcs_breadcrumb(rq, __gen8_emit_flush_dw(cs, 0, 0, 0));
616 	return gen12_emit_fini_breadcrumb_tail(rq, cs);
617 }
618 
619 u32 *gen12_emit_fini_breadcrumb_rcs(struct i915_request *rq, u32 *cs)
620 {
621 	cs = gen12_emit_ggtt_write_rcs(cs,
622 				       rq->fence.seqno,
623 				       hwsp_offset(rq),
624 				       PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
625 				       PIPE_CONTROL_CS_STALL |
626 				       PIPE_CONTROL_TILE_CACHE_FLUSH |
627 				       PIPE_CONTROL_FLUSH_L3 |
628 				       PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
629 				       PIPE_CONTROL_DEPTH_CACHE_FLUSH |
630 				       /* Wa_1409600907:tgl */
631 				       PIPE_CONTROL_DEPTH_STALL |
632 				       PIPE_CONTROL_DC_FLUSH_ENABLE |
633 				       PIPE_CONTROL_FLUSH_ENABLE);
634 
635 	return gen12_emit_fini_breadcrumb_tail(rq, cs);
636 }
637