xref: /openbmc/linux/drivers/gpu/drm/i915/gt/intel_lrc.c (revision c34a8052)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2014 Intel Corporation
4  */
5 
6 #include "gem/i915_gem_lmem.h"
7 
8 #include "gen8_engine_cs.h"
9 #include "i915_drv.h"
10 #include "i915_perf.h"
11 #include "i915_reg.h"
12 #include "intel_context.h"
13 #include "intel_engine.h"
14 #include "intel_engine_regs.h"
15 #include "intel_gpu_commands.h"
16 #include "intel_gt.h"
17 #include "intel_gt_regs.h"
18 #include "intel_lrc.h"
19 #include "intel_lrc_reg.h"
20 #include "intel_ring.h"
21 #include "shmem_utils.h"
22 
23 /*
24  * The per-platform tables are u8-encoded in @data. Decode @data and set the
25  * addresses' offset and commands in @regs. The following encoding is used
26  * for each byte. There are 2 steps: decoding commands and decoding addresses.
27  *
28  * Commands:
29  * [7]: create NOPs - number of NOPs are set in lower bits
30  * [6]: When creating MI_LOAD_REGISTER_IMM command, allow to set
31  *      MI_LRI_FORCE_POSTED
32  * [5:0]: Number of NOPs or registers to set values to in case of
33  *        MI_LOAD_REGISTER_IMM
34  *
35  * Addresses: these are decoded after a MI_LOAD_REGISTER_IMM command by "count"
36  * number of registers. They are set by using the REG/REG16 macros: the former
37  * is used for offsets smaller than 0x200 while the latter is for values bigger
38  * than that. Those macros already set all the bits documented below correctly:
39  *
40  * [7]: When a register offset needs more than 6 bits, use additional bytes, to
41  *      follow, for the lower bits
42  * [6:0]: Register offset, without considering the engine base.
43  *
44  * This function only tweaks the commands and register offsets. Values are not
45  * filled out.
46  */
47 static void set_offsets(u32 *regs,
48 			const u8 *data,
49 			const struct intel_engine_cs *engine,
50 			bool close)
51 #define NOP(x) (BIT(7) | (x))
52 #define LRI(count, flags) ((flags) << 6 | (count) | BUILD_BUG_ON_ZERO(count >= BIT(6)))
53 #define POSTED BIT(0)
54 #define REG(x) (((x) >> 2) | BUILD_BUG_ON_ZERO(x >= 0x200))
55 #define REG16(x) \
56 	(((x) >> 9) | BIT(7) | BUILD_BUG_ON_ZERO(x >= 0x10000)), \
57 	(((x) >> 2) & 0x7f)
58 #define END 0
59 {
60 	const u32 base = engine->mmio_base;
61 
62 	while (*data) {
63 		u8 count, flags;
64 
65 		if (*data & BIT(7)) { /* skip */
66 			count = *data++ & ~BIT(7);
67 			regs += count;
68 			continue;
69 		}
70 
71 		count = *data & 0x3f;
72 		flags = *data >> 6;
73 		data++;
74 
75 		*regs = MI_LOAD_REGISTER_IMM(count);
76 		if (flags & POSTED)
77 			*regs |= MI_LRI_FORCE_POSTED;
78 		if (GRAPHICS_VER(engine->i915) >= 11)
79 			*regs |= MI_LRI_LRM_CS_MMIO;
80 		regs++;
81 
82 		GEM_BUG_ON(!count);
83 		do {
84 			u32 offset = 0;
85 			u8 v;
86 
87 			do {
88 				v = *data++;
89 				offset <<= 7;
90 				offset |= v & ~BIT(7);
91 			} while (v & BIT(7));
92 
93 			regs[0] = base + (offset << 2);
94 			regs += 2;
95 		} while (--count);
96 	}
97 
98 	if (close) {
99 		/* Close the batch; used mainly by live_lrc_layout() */
100 		*regs = MI_BATCH_BUFFER_END;
101 		if (GRAPHICS_VER(engine->i915) >= 11)
102 			*regs |= BIT(0);
103 	}
104 }
105 
106 static const u8 gen8_xcs_offsets[] = {
107 	NOP(1),
108 	LRI(11, 0),
109 	REG16(0x244),
110 	REG(0x034),
111 	REG(0x030),
112 	REG(0x038),
113 	REG(0x03c),
114 	REG(0x168),
115 	REG(0x140),
116 	REG(0x110),
117 	REG(0x11c),
118 	REG(0x114),
119 	REG(0x118),
120 
121 	NOP(9),
122 	LRI(9, 0),
123 	REG16(0x3a8),
124 	REG16(0x28c),
125 	REG16(0x288),
126 	REG16(0x284),
127 	REG16(0x280),
128 	REG16(0x27c),
129 	REG16(0x278),
130 	REG16(0x274),
131 	REG16(0x270),
132 
133 	NOP(13),
134 	LRI(2, 0),
135 	REG16(0x200),
136 	REG(0x028),
137 
138 	END
139 };
140 
141 static const u8 gen9_xcs_offsets[] = {
142 	NOP(1),
143 	LRI(14, POSTED),
144 	REG16(0x244),
145 	REG(0x034),
146 	REG(0x030),
147 	REG(0x038),
148 	REG(0x03c),
149 	REG(0x168),
150 	REG(0x140),
151 	REG(0x110),
152 	REG(0x11c),
153 	REG(0x114),
154 	REG(0x118),
155 	REG(0x1c0),
156 	REG(0x1c4),
157 	REG(0x1c8),
158 
159 	NOP(3),
160 	LRI(9, POSTED),
161 	REG16(0x3a8),
162 	REG16(0x28c),
163 	REG16(0x288),
164 	REG16(0x284),
165 	REG16(0x280),
166 	REG16(0x27c),
167 	REG16(0x278),
168 	REG16(0x274),
169 	REG16(0x270),
170 
171 	NOP(13),
172 	LRI(1, POSTED),
173 	REG16(0x200),
174 
175 	NOP(13),
176 	LRI(44, POSTED),
177 	REG(0x028),
178 	REG(0x09c),
179 	REG(0x0c0),
180 	REG(0x178),
181 	REG(0x17c),
182 	REG16(0x358),
183 	REG(0x170),
184 	REG(0x150),
185 	REG(0x154),
186 	REG(0x158),
187 	REG16(0x41c),
188 	REG16(0x600),
189 	REG16(0x604),
190 	REG16(0x608),
191 	REG16(0x60c),
192 	REG16(0x610),
193 	REG16(0x614),
194 	REG16(0x618),
195 	REG16(0x61c),
196 	REG16(0x620),
197 	REG16(0x624),
198 	REG16(0x628),
199 	REG16(0x62c),
200 	REG16(0x630),
201 	REG16(0x634),
202 	REG16(0x638),
203 	REG16(0x63c),
204 	REG16(0x640),
205 	REG16(0x644),
206 	REG16(0x648),
207 	REG16(0x64c),
208 	REG16(0x650),
209 	REG16(0x654),
210 	REG16(0x658),
211 	REG16(0x65c),
212 	REG16(0x660),
213 	REG16(0x664),
214 	REG16(0x668),
215 	REG16(0x66c),
216 	REG16(0x670),
217 	REG16(0x674),
218 	REG16(0x678),
219 	REG16(0x67c),
220 	REG(0x068),
221 
222 	END
223 };
224 
225 static const u8 gen12_xcs_offsets[] = {
226 	NOP(1),
227 	LRI(13, POSTED),
228 	REG16(0x244),
229 	REG(0x034),
230 	REG(0x030),
231 	REG(0x038),
232 	REG(0x03c),
233 	REG(0x168),
234 	REG(0x140),
235 	REG(0x110),
236 	REG(0x1c0),
237 	REG(0x1c4),
238 	REG(0x1c8),
239 	REG(0x180),
240 	REG16(0x2b4),
241 
242 	NOP(5),
243 	LRI(9, POSTED),
244 	REG16(0x3a8),
245 	REG16(0x28c),
246 	REG16(0x288),
247 	REG16(0x284),
248 	REG16(0x280),
249 	REG16(0x27c),
250 	REG16(0x278),
251 	REG16(0x274),
252 	REG16(0x270),
253 
254 	END
255 };
256 
257 static const u8 dg2_xcs_offsets[] = {
258 	NOP(1),
259 	LRI(15, POSTED),
260 	REG16(0x244),
261 	REG(0x034),
262 	REG(0x030),
263 	REG(0x038),
264 	REG(0x03c),
265 	REG(0x168),
266 	REG(0x140),
267 	REG(0x110),
268 	REG(0x1c0),
269 	REG(0x1c4),
270 	REG(0x1c8),
271 	REG(0x180),
272 	REG16(0x2b4),
273 	REG(0x120),
274 	REG(0x124),
275 
276 	NOP(1),
277 	LRI(9, POSTED),
278 	REG16(0x3a8),
279 	REG16(0x28c),
280 	REG16(0x288),
281 	REG16(0x284),
282 	REG16(0x280),
283 	REG16(0x27c),
284 	REG16(0x278),
285 	REG16(0x274),
286 	REG16(0x270),
287 
288 	END
289 };
290 
291 static const u8 gen8_rcs_offsets[] = {
292 	NOP(1),
293 	LRI(14, POSTED),
294 	REG16(0x244),
295 	REG(0x034),
296 	REG(0x030),
297 	REG(0x038),
298 	REG(0x03c),
299 	REG(0x168),
300 	REG(0x140),
301 	REG(0x110),
302 	REG(0x11c),
303 	REG(0x114),
304 	REG(0x118),
305 	REG(0x1c0),
306 	REG(0x1c4),
307 	REG(0x1c8),
308 
309 	NOP(3),
310 	LRI(9, POSTED),
311 	REG16(0x3a8),
312 	REG16(0x28c),
313 	REG16(0x288),
314 	REG16(0x284),
315 	REG16(0x280),
316 	REG16(0x27c),
317 	REG16(0x278),
318 	REG16(0x274),
319 	REG16(0x270),
320 
321 	NOP(13),
322 	LRI(1, 0),
323 	REG(0x0c8),
324 
325 	END
326 };
327 
328 static const u8 gen9_rcs_offsets[] = {
329 	NOP(1),
330 	LRI(14, POSTED),
331 	REG16(0x244),
332 	REG(0x34),
333 	REG(0x30),
334 	REG(0x38),
335 	REG(0x3c),
336 	REG(0x168),
337 	REG(0x140),
338 	REG(0x110),
339 	REG(0x11c),
340 	REG(0x114),
341 	REG(0x118),
342 	REG(0x1c0),
343 	REG(0x1c4),
344 	REG(0x1c8),
345 
346 	NOP(3),
347 	LRI(9, POSTED),
348 	REG16(0x3a8),
349 	REG16(0x28c),
350 	REG16(0x288),
351 	REG16(0x284),
352 	REG16(0x280),
353 	REG16(0x27c),
354 	REG16(0x278),
355 	REG16(0x274),
356 	REG16(0x270),
357 
358 	NOP(13),
359 	LRI(1, 0),
360 	REG(0xc8),
361 
362 	NOP(13),
363 	LRI(44, POSTED),
364 	REG(0x28),
365 	REG(0x9c),
366 	REG(0xc0),
367 	REG(0x178),
368 	REG(0x17c),
369 	REG16(0x358),
370 	REG(0x170),
371 	REG(0x150),
372 	REG(0x154),
373 	REG(0x158),
374 	REG16(0x41c),
375 	REG16(0x600),
376 	REG16(0x604),
377 	REG16(0x608),
378 	REG16(0x60c),
379 	REG16(0x610),
380 	REG16(0x614),
381 	REG16(0x618),
382 	REG16(0x61c),
383 	REG16(0x620),
384 	REG16(0x624),
385 	REG16(0x628),
386 	REG16(0x62c),
387 	REG16(0x630),
388 	REG16(0x634),
389 	REG16(0x638),
390 	REG16(0x63c),
391 	REG16(0x640),
392 	REG16(0x644),
393 	REG16(0x648),
394 	REG16(0x64c),
395 	REG16(0x650),
396 	REG16(0x654),
397 	REG16(0x658),
398 	REG16(0x65c),
399 	REG16(0x660),
400 	REG16(0x664),
401 	REG16(0x668),
402 	REG16(0x66c),
403 	REG16(0x670),
404 	REG16(0x674),
405 	REG16(0x678),
406 	REG16(0x67c),
407 	REG(0x68),
408 
409 	END
410 };
411 
412 static const u8 gen11_rcs_offsets[] = {
413 	NOP(1),
414 	LRI(15, POSTED),
415 	REG16(0x244),
416 	REG(0x034),
417 	REG(0x030),
418 	REG(0x038),
419 	REG(0x03c),
420 	REG(0x168),
421 	REG(0x140),
422 	REG(0x110),
423 	REG(0x11c),
424 	REG(0x114),
425 	REG(0x118),
426 	REG(0x1c0),
427 	REG(0x1c4),
428 	REG(0x1c8),
429 	REG(0x180),
430 
431 	NOP(1),
432 	LRI(9, POSTED),
433 	REG16(0x3a8),
434 	REG16(0x28c),
435 	REG16(0x288),
436 	REG16(0x284),
437 	REG16(0x280),
438 	REG16(0x27c),
439 	REG16(0x278),
440 	REG16(0x274),
441 	REG16(0x270),
442 
443 	LRI(1, POSTED),
444 	REG(0x1b0),
445 
446 	NOP(10),
447 	LRI(1, 0),
448 	REG(0x0c8),
449 
450 	END
451 };
452 
453 static const u8 gen12_rcs_offsets[] = {
454 	NOP(1),
455 	LRI(13, POSTED),
456 	REG16(0x244),
457 	REG(0x034),
458 	REG(0x030),
459 	REG(0x038),
460 	REG(0x03c),
461 	REG(0x168),
462 	REG(0x140),
463 	REG(0x110),
464 	REG(0x1c0),
465 	REG(0x1c4),
466 	REG(0x1c8),
467 	REG(0x180),
468 	REG16(0x2b4),
469 
470 	NOP(5),
471 	LRI(9, POSTED),
472 	REG16(0x3a8),
473 	REG16(0x28c),
474 	REG16(0x288),
475 	REG16(0x284),
476 	REG16(0x280),
477 	REG16(0x27c),
478 	REG16(0x278),
479 	REG16(0x274),
480 	REG16(0x270),
481 
482 	LRI(3, POSTED),
483 	REG(0x1b0),
484 	REG16(0x5a8),
485 	REG16(0x5ac),
486 
487 	NOP(6),
488 	LRI(1, 0),
489 	REG(0x0c8),
490 	NOP(3 + 9 + 1),
491 
492 	LRI(51, POSTED),
493 	REG16(0x588),
494 	REG16(0x588),
495 	REG16(0x588),
496 	REG16(0x588),
497 	REG16(0x588),
498 	REG16(0x588),
499 	REG(0x028),
500 	REG(0x09c),
501 	REG(0x0c0),
502 	REG(0x178),
503 	REG(0x17c),
504 	REG16(0x358),
505 	REG(0x170),
506 	REG(0x150),
507 	REG(0x154),
508 	REG(0x158),
509 	REG16(0x41c),
510 	REG16(0x600),
511 	REG16(0x604),
512 	REG16(0x608),
513 	REG16(0x60c),
514 	REG16(0x610),
515 	REG16(0x614),
516 	REG16(0x618),
517 	REG16(0x61c),
518 	REG16(0x620),
519 	REG16(0x624),
520 	REG16(0x628),
521 	REG16(0x62c),
522 	REG16(0x630),
523 	REG16(0x634),
524 	REG16(0x638),
525 	REG16(0x63c),
526 	REG16(0x640),
527 	REG16(0x644),
528 	REG16(0x648),
529 	REG16(0x64c),
530 	REG16(0x650),
531 	REG16(0x654),
532 	REG16(0x658),
533 	REG16(0x65c),
534 	REG16(0x660),
535 	REG16(0x664),
536 	REG16(0x668),
537 	REG16(0x66c),
538 	REG16(0x670),
539 	REG16(0x674),
540 	REG16(0x678),
541 	REG16(0x67c),
542 	REG(0x068),
543 	REG(0x084),
544 	NOP(1),
545 
546 	END
547 };
548 
549 static const u8 xehp_rcs_offsets[] = {
550 	NOP(1),
551 	LRI(13, POSTED),
552 	REG16(0x244),
553 	REG(0x034),
554 	REG(0x030),
555 	REG(0x038),
556 	REG(0x03c),
557 	REG(0x168),
558 	REG(0x140),
559 	REG(0x110),
560 	REG(0x1c0),
561 	REG(0x1c4),
562 	REG(0x1c8),
563 	REG(0x180),
564 	REG16(0x2b4),
565 
566 	NOP(5),
567 	LRI(9, POSTED),
568 	REG16(0x3a8),
569 	REG16(0x28c),
570 	REG16(0x288),
571 	REG16(0x284),
572 	REG16(0x280),
573 	REG16(0x27c),
574 	REG16(0x278),
575 	REG16(0x274),
576 	REG16(0x270),
577 
578 	LRI(3, POSTED),
579 	REG(0x1b0),
580 	REG16(0x5a8),
581 	REG16(0x5ac),
582 
583 	NOP(6),
584 	LRI(1, 0),
585 	REG(0x0c8),
586 
587 	END
588 };
589 
590 static const u8 dg2_rcs_offsets[] = {
591 	NOP(1),
592 	LRI(15, POSTED),
593 	REG16(0x244),
594 	REG(0x034),
595 	REG(0x030),
596 	REG(0x038),
597 	REG(0x03c),
598 	REG(0x168),
599 	REG(0x140),
600 	REG(0x110),
601 	REG(0x1c0),
602 	REG(0x1c4),
603 	REG(0x1c8),
604 	REG(0x180),
605 	REG16(0x2b4),
606 	REG(0x120),
607 	REG(0x124),
608 
609 	NOP(1),
610 	LRI(9, POSTED),
611 	REG16(0x3a8),
612 	REG16(0x28c),
613 	REG16(0x288),
614 	REG16(0x284),
615 	REG16(0x280),
616 	REG16(0x27c),
617 	REG16(0x278),
618 	REG16(0x274),
619 	REG16(0x270),
620 
621 	LRI(3, POSTED),
622 	REG(0x1b0),
623 	REG16(0x5a8),
624 	REG16(0x5ac),
625 
626 	NOP(6),
627 	LRI(1, 0),
628 	REG(0x0c8),
629 
630 	END
631 };
632 
633 static const u8 mtl_rcs_offsets[] = {
634 	NOP(1),
635 	LRI(15, POSTED),
636 	REG16(0x244),
637 	REG(0x034),
638 	REG(0x030),
639 	REG(0x038),
640 	REG(0x03c),
641 	REG(0x168),
642 	REG(0x140),
643 	REG(0x110),
644 	REG(0x1c0),
645 	REG(0x1c4),
646 	REG(0x1c8),
647 	REG(0x180),
648 	REG16(0x2b4),
649 	REG(0x120),
650 	REG(0x124),
651 
652 	NOP(1),
653 	LRI(9, POSTED),
654 	REG16(0x3a8),
655 	REG16(0x28c),
656 	REG16(0x288),
657 	REG16(0x284),
658 	REG16(0x280),
659 	REG16(0x27c),
660 	REG16(0x278),
661 	REG16(0x274),
662 	REG16(0x270),
663 
664 	NOP(2),
665 	LRI(2, POSTED),
666 	REG16(0x5a8),
667 	REG16(0x5ac),
668 
669 	NOP(6),
670 	LRI(1, 0),
671 	REG(0x0c8),
672 
673 	END
674 };
675 
676 #undef END
677 #undef REG16
678 #undef REG
679 #undef LRI
680 #undef NOP
681 
682 static const u8 *reg_offsets(const struct intel_engine_cs *engine)
683 {
684 	/*
685 	 * The gen12+ lists only have the registers we program in the basic
686 	 * default state. We rely on the context image using relative
687 	 * addressing to automatic fixup the register state between the
688 	 * physical engines for virtual engine.
689 	 */
690 	GEM_BUG_ON(GRAPHICS_VER(engine->i915) >= 12 &&
691 		   !intel_engine_has_relative_mmio(engine));
692 
693 	if (engine->flags & I915_ENGINE_HAS_RCS_REG_STATE) {
694 		if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 70))
695 			return mtl_rcs_offsets;
696 		else if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
697 			return dg2_rcs_offsets;
698 		else if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 50))
699 			return xehp_rcs_offsets;
700 		else if (GRAPHICS_VER(engine->i915) >= 12)
701 			return gen12_rcs_offsets;
702 		else if (GRAPHICS_VER(engine->i915) >= 11)
703 			return gen11_rcs_offsets;
704 		else if (GRAPHICS_VER(engine->i915) >= 9)
705 			return gen9_rcs_offsets;
706 		else
707 			return gen8_rcs_offsets;
708 	} else {
709 		if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 55))
710 			return dg2_xcs_offsets;
711 		else if (GRAPHICS_VER(engine->i915) >= 12)
712 			return gen12_xcs_offsets;
713 		else if (GRAPHICS_VER(engine->i915) >= 9)
714 			return gen9_xcs_offsets;
715 		else
716 			return gen8_xcs_offsets;
717 	}
718 }
719 
720 static int lrc_ring_mi_mode(const struct intel_engine_cs *engine)
721 {
722 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 50))
723 		return 0x70;
724 	else if (GRAPHICS_VER(engine->i915) >= 12)
725 		return 0x60;
726 	else if (GRAPHICS_VER(engine->i915) >= 9)
727 		return 0x54;
728 	else if (engine->class == RENDER_CLASS)
729 		return 0x58;
730 	else
731 		return -1;
732 }
733 
734 static int lrc_ring_bb_offset(const struct intel_engine_cs *engine)
735 {
736 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 50))
737 		return 0x80;
738 	else if (GRAPHICS_VER(engine->i915) >= 12)
739 		return 0x70;
740 	else if (GRAPHICS_VER(engine->i915) >= 9)
741 		return 0x64;
742 	else if (GRAPHICS_VER(engine->i915) >= 8 &&
743 		 engine->class == RENDER_CLASS)
744 		return 0xc4;
745 	else
746 		return -1;
747 }
748 
749 static int lrc_ring_gpr0(const struct intel_engine_cs *engine)
750 {
751 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 50))
752 		return 0x84;
753 	else if (GRAPHICS_VER(engine->i915) >= 12)
754 		return 0x74;
755 	else if (GRAPHICS_VER(engine->i915) >= 9)
756 		return 0x68;
757 	else if (engine->class == RENDER_CLASS)
758 		return 0xd8;
759 	else
760 		return -1;
761 }
762 
763 static int lrc_ring_wa_bb_per_ctx(const struct intel_engine_cs *engine)
764 {
765 	if (GRAPHICS_VER(engine->i915) >= 12)
766 		return 0x12;
767 	else if (GRAPHICS_VER(engine->i915) >= 9 || engine->class == RENDER_CLASS)
768 		return 0x18;
769 	else
770 		return -1;
771 }
772 
773 static int lrc_ring_indirect_ptr(const struct intel_engine_cs *engine)
774 {
775 	int x;
776 
777 	x = lrc_ring_wa_bb_per_ctx(engine);
778 	if (x < 0)
779 		return x;
780 
781 	return x + 2;
782 }
783 
784 static int lrc_ring_indirect_offset(const struct intel_engine_cs *engine)
785 {
786 	int x;
787 
788 	x = lrc_ring_indirect_ptr(engine);
789 	if (x < 0)
790 		return x;
791 
792 	return x + 2;
793 }
794 
795 static int lrc_ring_cmd_buf_cctl(const struct intel_engine_cs *engine)
796 {
797 
798 	if (GRAPHICS_VER_FULL(engine->i915) >= IP_VER(12, 50))
799 		/*
800 		 * Note that the CSFE context has a dummy slot for CMD_BUF_CCTL
801 		 * simply to match the RCS context image layout.
802 		 */
803 		return 0xc6;
804 	else if (engine->class != RENDER_CLASS)
805 		return -1;
806 	else if (GRAPHICS_VER(engine->i915) >= 12)
807 		return 0xb6;
808 	else if (GRAPHICS_VER(engine->i915) >= 11)
809 		return 0xaa;
810 	else
811 		return -1;
812 }
813 
814 static u32
815 lrc_ring_indirect_offset_default(const struct intel_engine_cs *engine)
816 {
817 	if (GRAPHICS_VER(engine->i915) >= 12)
818 		return GEN12_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
819 	else if (GRAPHICS_VER(engine->i915) >= 11)
820 		return GEN11_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
821 	else if (GRAPHICS_VER(engine->i915) >= 9)
822 		return GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
823 	else if (GRAPHICS_VER(engine->i915) >= 8)
824 		return GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
825 
826 	GEM_BUG_ON(GRAPHICS_VER(engine->i915) < 8);
827 
828 	return 0;
829 }
830 
831 static void
832 lrc_setup_indirect_ctx(u32 *regs,
833 		       const struct intel_engine_cs *engine,
834 		       u32 ctx_bb_ggtt_addr,
835 		       u32 size)
836 {
837 	GEM_BUG_ON(!size);
838 	GEM_BUG_ON(!IS_ALIGNED(size, CACHELINE_BYTES));
839 	GEM_BUG_ON(lrc_ring_indirect_ptr(engine) == -1);
840 	regs[lrc_ring_indirect_ptr(engine) + 1] =
841 		ctx_bb_ggtt_addr | (size / CACHELINE_BYTES);
842 
843 	GEM_BUG_ON(lrc_ring_indirect_offset(engine) == -1);
844 	regs[lrc_ring_indirect_offset(engine) + 1] =
845 		lrc_ring_indirect_offset_default(engine) << 6;
846 }
847 
848 static void init_common_regs(u32 * const regs,
849 			     const struct intel_context *ce,
850 			     const struct intel_engine_cs *engine,
851 			     bool inhibit)
852 {
853 	u32 ctl;
854 	int loc;
855 
856 	ctl = _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH);
857 	ctl |= _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
858 	if (inhibit)
859 		ctl |= CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT;
860 	if (GRAPHICS_VER(engine->i915) < 11)
861 		ctl |= _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT |
862 					   CTX_CTRL_RS_CTX_ENABLE);
863 	regs[CTX_CONTEXT_CONTROL] = ctl;
864 
865 	regs[CTX_TIMESTAMP] = ce->stats.runtime.last;
866 
867 	loc = lrc_ring_bb_offset(engine);
868 	if (loc != -1)
869 		regs[loc + 1] = 0;
870 }
871 
872 static void init_wa_bb_regs(u32 * const regs,
873 			    const struct intel_engine_cs *engine)
874 {
875 	const struct i915_ctx_workarounds * const wa_ctx = &engine->wa_ctx;
876 
877 	if (wa_ctx->per_ctx.size) {
878 		const u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
879 
880 		GEM_BUG_ON(lrc_ring_wa_bb_per_ctx(engine) == -1);
881 		regs[lrc_ring_wa_bb_per_ctx(engine) + 1] =
882 			(ggtt_offset + wa_ctx->per_ctx.offset) | 0x01;
883 	}
884 
885 	if (wa_ctx->indirect_ctx.size) {
886 		lrc_setup_indirect_ctx(regs, engine,
887 				       i915_ggtt_offset(wa_ctx->vma) +
888 				       wa_ctx->indirect_ctx.offset,
889 				       wa_ctx->indirect_ctx.size);
890 	}
891 }
892 
893 static void init_ppgtt_regs(u32 *regs, const struct i915_ppgtt *ppgtt)
894 {
895 	if (i915_vm_is_4lvl(&ppgtt->vm)) {
896 		/* 64b PPGTT (48bit canonical)
897 		 * PDP0_DESCRIPTOR contains the base address to PML4 and
898 		 * other PDP Descriptors are ignored.
899 		 */
900 		ASSIGN_CTX_PML4(ppgtt, regs);
901 	} else {
902 		ASSIGN_CTX_PDP(ppgtt, regs, 3);
903 		ASSIGN_CTX_PDP(ppgtt, regs, 2);
904 		ASSIGN_CTX_PDP(ppgtt, regs, 1);
905 		ASSIGN_CTX_PDP(ppgtt, regs, 0);
906 	}
907 }
908 
909 static struct i915_ppgtt *vm_alias(struct i915_address_space *vm)
910 {
911 	if (i915_is_ggtt(vm))
912 		return i915_vm_to_ggtt(vm)->alias;
913 	else
914 		return i915_vm_to_ppgtt(vm);
915 }
916 
917 static void __reset_stop_ring(u32 *regs, const struct intel_engine_cs *engine)
918 {
919 	int x;
920 
921 	x = lrc_ring_mi_mode(engine);
922 	if (x != -1) {
923 		regs[x + 1] &= ~STOP_RING;
924 		regs[x + 1] |= STOP_RING << 16;
925 	}
926 }
927 
928 static void __lrc_init_regs(u32 *regs,
929 			    const struct intel_context *ce,
930 			    const struct intel_engine_cs *engine,
931 			    bool inhibit)
932 {
933 	/*
934 	 * A context is actually a big batch buffer with several
935 	 * MI_LOAD_REGISTER_IMM commands followed by (reg, value) pairs. The
936 	 * values we are setting here are only for the first context restore:
937 	 * on a subsequent save, the GPU will recreate this batchbuffer with new
938 	 * values (including all the missing MI_LOAD_REGISTER_IMM commands that
939 	 * we are not initializing here).
940 	 *
941 	 * Must keep consistent with virtual_update_register_offsets().
942 	 */
943 
944 	if (inhibit)
945 		memset(regs, 0, PAGE_SIZE);
946 
947 	set_offsets(regs, reg_offsets(engine), engine, inhibit);
948 
949 	init_common_regs(regs, ce, engine, inhibit);
950 	init_ppgtt_regs(regs, vm_alias(ce->vm));
951 
952 	init_wa_bb_regs(regs, engine);
953 
954 	__reset_stop_ring(regs, engine);
955 }
956 
957 void lrc_init_regs(const struct intel_context *ce,
958 		   const struct intel_engine_cs *engine,
959 		   bool inhibit)
960 {
961 	__lrc_init_regs(ce->lrc_reg_state, ce, engine, inhibit);
962 }
963 
964 void lrc_reset_regs(const struct intel_context *ce,
965 		    const struct intel_engine_cs *engine)
966 {
967 	__reset_stop_ring(ce->lrc_reg_state, engine);
968 }
969 
970 static void
971 set_redzone(void *vaddr, const struct intel_engine_cs *engine)
972 {
973 	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
974 		return;
975 
976 	vaddr += engine->context_size;
977 
978 	memset(vaddr, CONTEXT_REDZONE, I915_GTT_PAGE_SIZE);
979 }
980 
981 static void
982 check_redzone(const void *vaddr, const struct intel_engine_cs *engine)
983 {
984 	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
985 		return;
986 
987 	vaddr += engine->context_size;
988 
989 	if (memchr_inv(vaddr, CONTEXT_REDZONE, I915_GTT_PAGE_SIZE))
990 		drm_err_once(&engine->i915->drm,
991 			     "%s context redzone overwritten!\n",
992 			     engine->name);
993 }
994 
995 static u32 context_wa_bb_offset(const struct intel_context *ce)
996 {
997 	return PAGE_SIZE * ce->wa_bb_page;
998 }
999 
1000 static u32 *context_indirect_bb(const struct intel_context *ce)
1001 {
1002 	void *ptr;
1003 
1004 	GEM_BUG_ON(!ce->wa_bb_page);
1005 
1006 	ptr = ce->lrc_reg_state;
1007 	ptr -= LRC_STATE_OFFSET; /* back to start of context image */
1008 	ptr += context_wa_bb_offset(ce);
1009 
1010 	return ptr;
1011 }
1012 
1013 void lrc_init_state(struct intel_context *ce,
1014 		    struct intel_engine_cs *engine,
1015 		    void *state)
1016 {
1017 	bool inhibit = true;
1018 
1019 	set_redzone(state, engine);
1020 
1021 	if (engine->default_state) {
1022 		shmem_read(engine->default_state, 0,
1023 			   state, engine->context_size);
1024 		__set_bit(CONTEXT_VALID_BIT, &ce->flags);
1025 		inhibit = false;
1026 	}
1027 
1028 	/* Clear the ppHWSP (inc. per-context counters) */
1029 	memset(state, 0, PAGE_SIZE);
1030 
1031 	/* Clear the indirect wa and storage */
1032 	if (ce->wa_bb_page)
1033 		memset(state + context_wa_bb_offset(ce), 0, PAGE_SIZE);
1034 
1035 	/*
1036 	 * The second page of the context object contains some registers which
1037 	 * must be set up prior to the first execution.
1038 	 */
1039 	__lrc_init_regs(state + LRC_STATE_OFFSET, ce, engine, inhibit);
1040 }
1041 
1042 u32 lrc_indirect_bb(const struct intel_context *ce)
1043 {
1044 	return i915_ggtt_offset(ce->state) + context_wa_bb_offset(ce);
1045 }
1046 
1047 static u32 *setup_predicate_disable_wa(const struct intel_context *ce, u32 *cs)
1048 {
1049 	/* If predication is active, this will be noop'ed */
1050 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT | (4 - 2);
1051 	*cs++ = lrc_indirect_bb(ce) + DG2_PREDICATE_RESULT_WA;
1052 	*cs++ = 0;
1053 	*cs++ = 0; /* No predication */
1054 
1055 	/* predicated end, only terminates if SET_PREDICATE_RESULT:0 is clear */
1056 	*cs++ = MI_BATCH_BUFFER_END | BIT(15);
1057 	*cs++ = MI_SET_PREDICATE | MI_SET_PREDICATE_DISABLE;
1058 
1059 	/* Instructions are no longer predicated (disabled), we can proceed */
1060 	*cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT | (4 - 2);
1061 	*cs++ = lrc_indirect_bb(ce) + DG2_PREDICATE_RESULT_WA;
1062 	*cs++ = 0;
1063 	*cs++ = 1; /* enable predication before the next BB */
1064 
1065 	*cs++ = MI_BATCH_BUFFER_END;
1066 	GEM_BUG_ON(offset_in_page(cs) > DG2_PREDICATE_RESULT_WA);
1067 
1068 	return cs;
1069 }
1070 
1071 static struct i915_vma *
1072 __lrc_alloc_state(struct intel_context *ce, struct intel_engine_cs *engine)
1073 {
1074 	struct drm_i915_gem_object *obj;
1075 	struct i915_vma *vma;
1076 	u32 context_size;
1077 
1078 	context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
1079 
1080 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
1081 		context_size += I915_GTT_PAGE_SIZE; /* for redzone */
1082 
1083 	if (GRAPHICS_VER(engine->i915) >= 12) {
1084 		ce->wa_bb_page = context_size / PAGE_SIZE;
1085 		context_size += PAGE_SIZE;
1086 	}
1087 
1088 	if (intel_context_is_parent(ce) && intel_engine_uses_guc(engine)) {
1089 		ce->parallel.guc.parent_page = context_size / PAGE_SIZE;
1090 		context_size += PARENT_SCRATCH_SIZE;
1091 	}
1092 
1093 	obj = i915_gem_object_create_lmem(engine->i915, context_size,
1094 					  I915_BO_ALLOC_PM_VOLATILE);
1095 	if (IS_ERR(obj)) {
1096 		obj = i915_gem_object_create_shmem(engine->i915, context_size);
1097 		if (IS_ERR(obj))
1098 			return ERR_CAST(obj);
1099 
1100 		/*
1101 		 * Wa_22016122933: For Media version 13.0, all Media GT shared
1102 		 * memory needs to be mapped as WC on CPU side and UC (PAT
1103 		 * index 2) on GPU side.
1104 		 */
1105 		if (intel_gt_needs_wa_22016122933(engine->gt))
1106 			i915_gem_object_set_cache_coherency(obj, I915_CACHE_NONE);
1107 	}
1108 
1109 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
1110 	if (IS_ERR(vma)) {
1111 		i915_gem_object_put(obj);
1112 		return vma;
1113 	}
1114 
1115 	return vma;
1116 }
1117 
1118 static struct intel_timeline *
1119 pinned_timeline(struct intel_context *ce, struct intel_engine_cs *engine)
1120 {
1121 	struct intel_timeline *tl = fetch_and_zero(&ce->timeline);
1122 
1123 	return intel_timeline_create_from_engine(engine, page_unmask_bits(tl));
1124 }
1125 
1126 int lrc_alloc(struct intel_context *ce, struct intel_engine_cs *engine)
1127 {
1128 	struct intel_ring *ring;
1129 	struct i915_vma *vma;
1130 	int err;
1131 
1132 	GEM_BUG_ON(ce->state);
1133 
1134 	vma = __lrc_alloc_state(ce, engine);
1135 	if (IS_ERR(vma))
1136 		return PTR_ERR(vma);
1137 
1138 	ring = intel_engine_create_ring(engine, ce->ring_size);
1139 	if (IS_ERR(ring)) {
1140 		err = PTR_ERR(ring);
1141 		goto err_vma;
1142 	}
1143 
1144 	if (!page_mask_bits(ce->timeline)) {
1145 		struct intel_timeline *tl;
1146 
1147 		/*
1148 		 * Use the static global HWSP for the kernel context, and
1149 		 * a dynamically allocated cacheline for everyone else.
1150 		 */
1151 		if (unlikely(ce->timeline))
1152 			tl = pinned_timeline(ce, engine);
1153 		else
1154 			tl = intel_timeline_create(engine->gt);
1155 		if (IS_ERR(tl)) {
1156 			err = PTR_ERR(tl);
1157 			goto err_ring;
1158 		}
1159 
1160 		ce->timeline = tl;
1161 	}
1162 
1163 	ce->ring = ring;
1164 	ce->state = vma;
1165 
1166 	return 0;
1167 
1168 err_ring:
1169 	intel_ring_put(ring);
1170 err_vma:
1171 	i915_vma_put(vma);
1172 	return err;
1173 }
1174 
1175 void lrc_reset(struct intel_context *ce)
1176 {
1177 	GEM_BUG_ON(!intel_context_is_pinned(ce));
1178 
1179 	intel_ring_reset(ce->ring, ce->ring->emit);
1180 
1181 	/* Scrub away the garbage */
1182 	lrc_init_regs(ce, ce->engine, true);
1183 	ce->lrc.lrca = lrc_update_regs(ce, ce->engine, ce->ring->tail);
1184 }
1185 
1186 int
1187 lrc_pre_pin(struct intel_context *ce,
1188 	    struct intel_engine_cs *engine,
1189 	    struct i915_gem_ww_ctx *ww,
1190 	    void **vaddr)
1191 {
1192 	GEM_BUG_ON(!ce->state);
1193 	GEM_BUG_ON(!i915_vma_is_pinned(ce->state));
1194 
1195 	*vaddr = i915_gem_object_pin_map(ce->state->obj,
1196 					 intel_gt_coherent_map_type(ce->engine->gt,
1197 								    ce->state->obj,
1198 								    false) |
1199 					 I915_MAP_OVERRIDE);
1200 
1201 	return PTR_ERR_OR_ZERO(*vaddr);
1202 }
1203 
1204 int
1205 lrc_pin(struct intel_context *ce,
1206 	struct intel_engine_cs *engine,
1207 	void *vaddr)
1208 {
1209 	ce->lrc_reg_state = vaddr + LRC_STATE_OFFSET;
1210 
1211 	if (!__test_and_set_bit(CONTEXT_INIT_BIT, &ce->flags))
1212 		lrc_init_state(ce, engine, vaddr);
1213 
1214 	ce->lrc.lrca = lrc_update_regs(ce, engine, ce->ring->tail);
1215 	return 0;
1216 }
1217 
1218 void lrc_unpin(struct intel_context *ce)
1219 {
1220 	if (unlikely(ce->parallel.last_rq)) {
1221 		i915_request_put(ce->parallel.last_rq);
1222 		ce->parallel.last_rq = NULL;
1223 	}
1224 	check_redzone((void *)ce->lrc_reg_state - LRC_STATE_OFFSET,
1225 		      ce->engine);
1226 }
1227 
1228 void lrc_post_unpin(struct intel_context *ce)
1229 {
1230 	i915_gem_object_unpin_map(ce->state->obj);
1231 }
1232 
1233 void lrc_fini(struct intel_context *ce)
1234 {
1235 	if (!ce->state)
1236 		return;
1237 
1238 	intel_ring_put(fetch_and_zero(&ce->ring));
1239 	i915_vma_put(fetch_and_zero(&ce->state));
1240 }
1241 
1242 void lrc_destroy(struct kref *kref)
1243 {
1244 	struct intel_context *ce = container_of(kref, typeof(*ce), ref);
1245 
1246 	GEM_BUG_ON(!i915_active_is_idle(&ce->active));
1247 	GEM_BUG_ON(intel_context_is_pinned(ce));
1248 
1249 	lrc_fini(ce);
1250 
1251 	intel_context_fini(ce);
1252 	intel_context_free(ce);
1253 }
1254 
1255 static u32 *
1256 gen12_emit_timestamp_wa(const struct intel_context *ce, u32 *cs)
1257 {
1258 	*cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
1259 		MI_SRM_LRM_GLOBAL_GTT |
1260 		MI_LRI_LRM_CS_MMIO;
1261 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1262 	*cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
1263 		CTX_TIMESTAMP * sizeof(u32);
1264 	*cs++ = 0;
1265 
1266 	*cs++ = MI_LOAD_REGISTER_REG |
1267 		MI_LRR_SOURCE_CS_MMIO |
1268 		MI_LRI_LRM_CS_MMIO;
1269 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1270 	*cs++ = i915_mmio_reg_offset(RING_CTX_TIMESTAMP(0));
1271 
1272 	*cs++ = MI_LOAD_REGISTER_REG |
1273 		MI_LRR_SOURCE_CS_MMIO |
1274 		MI_LRI_LRM_CS_MMIO;
1275 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1276 	*cs++ = i915_mmio_reg_offset(RING_CTX_TIMESTAMP(0));
1277 
1278 	return cs;
1279 }
1280 
1281 static u32 *
1282 gen12_emit_restore_scratch(const struct intel_context *ce, u32 *cs)
1283 {
1284 	GEM_BUG_ON(lrc_ring_gpr0(ce->engine) == -1);
1285 
1286 	*cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
1287 		MI_SRM_LRM_GLOBAL_GTT |
1288 		MI_LRI_LRM_CS_MMIO;
1289 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1290 	*cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
1291 		(lrc_ring_gpr0(ce->engine) + 1) * sizeof(u32);
1292 	*cs++ = 0;
1293 
1294 	return cs;
1295 }
1296 
1297 static u32 *
1298 gen12_emit_cmd_buf_wa(const struct intel_context *ce, u32 *cs)
1299 {
1300 	GEM_BUG_ON(lrc_ring_cmd_buf_cctl(ce->engine) == -1);
1301 
1302 	*cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
1303 		MI_SRM_LRM_GLOBAL_GTT |
1304 		MI_LRI_LRM_CS_MMIO;
1305 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1306 	*cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
1307 		(lrc_ring_cmd_buf_cctl(ce->engine) + 1) * sizeof(u32);
1308 	*cs++ = 0;
1309 
1310 	*cs++ = MI_LOAD_REGISTER_REG |
1311 		MI_LRR_SOURCE_CS_MMIO |
1312 		MI_LRI_LRM_CS_MMIO;
1313 	*cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
1314 	*cs++ = i915_mmio_reg_offset(RING_CMD_BUF_CCTL(0));
1315 
1316 	return cs;
1317 }
1318 
1319 /*
1320  * The bspec's tuning guide asks us to program a vertical watermark value of
1321  * 0x3FF.  However this register is not saved/restored properly by the
1322  * hardware, so we're required to apply the desired value via INDIRECT_CTX
1323  * batch buffer to ensure the value takes effect properly.  All other bits
1324  * in this register should remain at 0 (the hardware default).
1325  */
1326 static u32 *
1327 dg2_emit_draw_watermark_setting(u32 *cs)
1328 {
1329 	*cs++ = MI_LOAD_REGISTER_IMM(1);
1330 	*cs++ = i915_mmio_reg_offset(DRAW_WATERMARK);
1331 	*cs++ = REG_FIELD_PREP(VERT_WM_VAL, 0x3FF);
1332 
1333 	return cs;
1334 }
1335 
1336 static u32 *
1337 gen12_emit_indirect_ctx_rcs(const struct intel_context *ce, u32 *cs)
1338 {
1339 	cs = gen12_emit_timestamp_wa(ce, cs);
1340 	cs = gen12_emit_cmd_buf_wa(ce, cs);
1341 	cs = gen12_emit_restore_scratch(ce, cs);
1342 
1343 	/* Wa_16013000631:dg2 */
1344 	if (IS_DG2_G11(ce->engine->i915))
1345 		cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE, 0);
1346 
1347 	cs = gen12_emit_aux_table_inv(ce->engine, cs);
1348 
1349 	/* Wa_16014892111 */
1350 	if (IS_GFX_GT_IP_STEP(ce->engine->gt, IP_VER(12, 70), STEP_A0, STEP_B0) ||
1351 	    IS_GFX_GT_IP_STEP(ce->engine->gt, IP_VER(12, 71), STEP_A0, STEP_B0) ||
1352 	    IS_DG2(ce->engine->i915))
1353 		cs = dg2_emit_draw_watermark_setting(cs);
1354 
1355 	return cs;
1356 }
1357 
1358 static u32 *
1359 gen12_emit_indirect_ctx_xcs(const struct intel_context *ce, u32 *cs)
1360 {
1361 	cs = gen12_emit_timestamp_wa(ce, cs);
1362 	cs = gen12_emit_restore_scratch(ce, cs);
1363 
1364 	/* Wa_16013000631:dg2 */
1365 	if (IS_DG2_G11(ce->engine->i915))
1366 		if (ce->engine->class == COMPUTE_CLASS)
1367 			cs = gen8_emit_pipe_control(cs,
1368 						    PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE,
1369 						    0);
1370 
1371 	return gen12_emit_aux_table_inv(ce->engine, cs);
1372 }
1373 
1374 static void
1375 setup_indirect_ctx_bb(const struct intel_context *ce,
1376 		      const struct intel_engine_cs *engine,
1377 		      u32 *(*emit)(const struct intel_context *, u32 *))
1378 {
1379 	u32 * const start = context_indirect_bb(ce);
1380 	u32 *cs;
1381 
1382 	cs = emit(ce, start);
1383 	GEM_BUG_ON(cs - start > I915_GTT_PAGE_SIZE / sizeof(*cs));
1384 	while ((unsigned long)cs % CACHELINE_BYTES)
1385 		*cs++ = MI_NOOP;
1386 
1387 	GEM_BUG_ON(cs - start > DG2_PREDICATE_RESULT_BB / sizeof(*start));
1388 	setup_predicate_disable_wa(ce, start + DG2_PREDICATE_RESULT_BB / sizeof(*start));
1389 
1390 	lrc_setup_indirect_ctx(ce->lrc_reg_state, engine,
1391 			       lrc_indirect_bb(ce),
1392 			       (cs - start) * sizeof(*cs));
1393 }
1394 
1395 /*
1396  * The context descriptor encodes various attributes of a context,
1397  * including its GTT address and some flags. Because it's fairly
1398  * expensive to calculate, we'll just do it once and cache the result,
1399  * which remains valid until the context is unpinned.
1400  *
1401  * This is what a descriptor looks like, from LSB to MSB::
1402  *
1403  *      bits  0-11:    flags, GEN8_CTX_* (cached in ctx->desc_template)
1404  *      bits 12-31:    LRCA, GTT address of (the HWSP of) this context
1405  *      bits 32-52:    ctx ID, a globally unique tag (highest bit used by GuC)
1406  *      bits 53-54:    mbz, reserved for use by hardware
1407  *      bits 55-63:    group ID, currently unused and set to 0
1408  *
1409  * Starting from Gen11, the upper dword of the descriptor has a new format:
1410  *
1411  *      bits 32-36:    reserved
1412  *      bits 37-47:    SW context ID
1413  *      bits 48:53:    engine instance
1414  *      bit 54:        mbz, reserved for use by hardware
1415  *      bits 55-60:    SW counter
1416  *      bits 61-63:    engine class
1417  *
1418  * On Xe_HP, the upper dword of the descriptor has a new format:
1419  *
1420  *      bits 32-37:    virtual function number
1421  *      bit 38:        mbz, reserved for use by hardware
1422  *      bits 39-54:    SW context ID
1423  *      bits 55-57:    reserved
1424  *      bits 58-63:    SW counter
1425  *
1426  * engine info, SW context ID and SW counter need to form a unique number
1427  * (Context ID) per lrc.
1428  */
1429 static u32 lrc_descriptor(const struct intel_context *ce)
1430 {
1431 	u32 desc;
1432 
1433 	desc = INTEL_LEGACY_32B_CONTEXT;
1434 	if (i915_vm_is_4lvl(ce->vm))
1435 		desc = INTEL_LEGACY_64B_CONTEXT;
1436 	desc <<= GEN8_CTX_ADDRESSING_MODE_SHIFT;
1437 
1438 	desc |= GEN8_CTX_VALID | GEN8_CTX_PRIVILEGE;
1439 	if (GRAPHICS_VER(ce->vm->i915) == 8)
1440 		desc |= GEN8_CTX_L3LLC_COHERENT;
1441 
1442 	return i915_ggtt_offset(ce->state) | desc;
1443 }
1444 
1445 u32 lrc_update_regs(const struct intel_context *ce,
1446 		    const struct intel_engine_cs *engine,
1447 		    u32 head)
1448 {
1449 	struct intel_ring *ring = ce->ring;
1450 	u32 *regs = ce->lrc_reg_state;
1451 
1452 	GEM_BUG_ON(!intel_ring_offset_valid(ring, head));
1453 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
1454 
1455 	regs[CTX_RING_START] = i915_ggtt_offset(ring->vma);
1456 	regs[CTX_RING_HEAD] = head;
1457 	regs[CTX_RING_TAIL] = ring->tail;
1458 	regs[CTX_RING_CTL] = RING_CTL_SIZE(ring->size) | RING_VALID;
1459 
1460 	/* RPCS */
1461 	if (engine->class == RENDER_CLASS) {
1462 		regs[CTX_R_PWR_CLK_STATE] =
1463 			intel_sseu_make_rpcs(engine->gt, &ce->sseu);
1464 
1465 		i915_oa_init_reg_state(ce, engine);
1466 	}
1467 
1468 	if (ce->wa_bb_page) {
1469 		u32 *(*fn)(const struct intel_context *ce, u32 *cs);
1470 
1471 		fn = gen12_emit_indirect_ctx_xcs;
1472 		if (ce->engine->class == RENDER_CLASS)
1473 			fn = gen12_emit_indirect_ctx_rcs;
1474 
1475 		/* Mutually exclusive wrt to global indirect bb */
1476 		GEM_BUG_ON(engine->wa_ctx.indirect_ctx.size);
1477 		setup_indirect_ctx_bb(ce, engine, fn);
1478 	}
1479 
1480 	return lrc_descriptor(ce) | CTX_DESC_FORCE_RESTORE;
1481 }
1482 
1483 void lrc_update_offsets(struct intel_context *ce,
1484 			struct intel_engine_cs *engine)
1485 {
1486 	set_offsets(ce->lrc_reg_state, reg_offsets(engine), engine, false);
1487 }
1488 
1489 void lrc_check_regs(const struct intel_context *ce,
1490 		    const struct intel_engine_cs *engine,
1491 		    const char *when)
1492 {
1493 	const struct intel_ring *ring = ce->ring;
1494 	u32 *regs = ce->lrc_reg_state;
1495 	bool valid = true;
1496 	int x;
1497 
1498 	if (regs[CTX_RING_START] != i915_ggtt_offset(ring->vma)) {
1499 		pr_err("%s: context submitted with incorrect RING_START [%08x], expected %08x\n",
1500 		       engine->name,
1501 		       regs[CTX_RING_START],
1502 		       i915_ggtt_offset(ring->vma));
1503 		regs[CTX_RING_START] = i915_ggtt_offset(ring->vma);
1504 		valid = false;
1505 	}
1506 
1507 	if ((regs[CTX_RING_CTL] & ~(RING_WAIT | RING_WAIT_SEMAPHORE)) !=
1508 	    (RING_CTL_SIZE(ring->size) | RING_VALID)) {
1509 		pr_err("%s: context submitted with incorrect RING_CTL [%08x], expected %08x\n",
1510 		       engine->name,
1511 		       regs[CTX_RING_CTL],
1512 		       (u32)(RING_CTL_SIZE(ring->size) | RING_VALID));
1513 		regs[CTX_RING_CTL] = RING_CTL_SIZE(ring->size) | RING_VALID;
1514 		valid = false;
1515 	}
1516 
1517 	x = lrc_ring_mi_mode(engine);
1518 	if (x != -1 && regs[x + 1] & (regs[x + 1] >> 16) & STOP_RING) {
1519 		pr_err("%s: context submitted with STOP_RING [%08x] in RING_MI_MODE\n",
1520 		       engine->name, regs[x + 1]);
1521 		regs[x + 1] &= ~STOP_RING;
1522 		regs[x + 1] |= STOP_RING << 16;
1523 		valid = false;
1524 	}
1525 
1526 	WARN_ONCE(!valid, "Invalid lrc state found %s submission\n", when);
1527 }
1528 
1529 /*
1530  * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1531  * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1532  * but there is a slight complication as this is applied in WA batch where the
1533  * values are only initialized once so we cannot take register value at the
1534  * beginning and reuse it further; hence we save its value to memory, upload a
1535  * constant value with bit21 set and then we restore it back with the saved value.
1536  * To simplify the WA, a constant value is formed by using the default value
1537  * of this register. This shouldn't be a problem because we are only modifying
1538  * it for a short period and this batch in non-premptible. We can ofcourse
1539  * use additional instructions that read the actual value of the register
1540  * at that time and set our bit of interest but it makes the WA complicated.
1541  *
1542  * This WA is also required for Gen9 so extracting as a function avoids
1543  * code duplication.
1544  */
1545 static u32 *
1546 gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine, u32 *batch)
1547 {
1548 	/* NB no one else is allowed to scribble over scratch + 256! */
1549 	*batch++ = MI_STORE_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1550 	*batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1551 	*batch++ = intel_gt_scratch_offset(engine->gt,
1552 					   INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA);
1553 	*batch++ = 0;
1554 
1555 	*batch++ = MI_LOAD_REGISTER_IMM(1);
1556 	*batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1557 	*batch++ = 0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES;
1558 
1559 	batch = gen8_emit_pipe_control(batch,
1560 				       PIPE_CONTROL_CS_STALL |
1561 				       PIPE_CONTROL_DC_FLUSH_ENABLE,
1562 				       0);
1563 
1564 	*batch++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
1565 	*batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
1566 	*batch++ = intel_gt_scratch_offset(engine->gt,
1567 					   INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA);
1568 	*batch++ = 0;
1569 
1570 	return batch;
1571 }
1572 
1573 /*
1574  * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1575  * initialized at the beginning and shared across all contexts but this field
1576  * helps us to have multiple batches at different offsets and select them based
1577  * on a criteria. At the moment this batch always start at the beginning of the page
1578  * and at this point we don't have multiple wa_ctx batch buffers.
1579  *
1580  * The number of WA applied are not known at the beginning; we use this field
1581  * to return the no of DWORDS written.
1582  *
1583  * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1584  * so it adds NOOPs as padding to make it cacheline aligned.
1585  * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1586  * makes a complete batch buffer.
1587  */
1588 static u32 *gen8_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1589 {
1590 	/* WaDisableCtxRestoreArbitration:bdw,chv */
1591 	*batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1592 
1593 	/* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1594 	if (IS_BROADWELL(engine->i915))
1595 		batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1596 
1597 	/* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1598 	/* Actual scratch location is at 128 bytes offset */
1599 	batch = gen8_emit_pipe_control(batch,
1600 				       PIPE_CONTROL_FLUSH_L3 |
1601 				       PIPE_CONTROL_STORE_DATA_INDEX |
1602 				       PIPE_CONTROL_CS_STALL |
1603 				       PIPE_CONTROL_QW_WRITE,
1604 				       LRC_PPHWSP_SCRATCH_ADDR);
1605 
1606 	*batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1607 
1608 	/* Pad to end of cacheline */
1609 	while ((unsigned long)batch % CACHELINE_BYTES)
1610 		*batch++ = MI_NOOP;
1611 
1612 	/*
1613 	 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1614 	 * execution depends on the length specified in terms of cache lines
1615 	 * in the register CTX_RCS_INDIRECT_CTX
1616 	 */
1617 
1618 	return batch;
1619 }
1620 
1621 struct lri {
1622 	i915_reg_t reg;
1623 	u32 value;
1624 };
1625 
1626 static u32 *emit_lri(u32 *batch, const struct lri *lri, unsigned int count)
1627 {
1628 	GEM_BUG_ON(!count || count > 63);
1629 
1630 	*batch++ = MI_LOAD_REGISTER_IMM(count);
1631 	do {
1632 		*batch++ = i915_mmio_reg_offset(lri->reg);
1633 		*batch++ = lri->value;
1634 	} while (lri++, --count);
1635 	*batch++ = MI_NOOP;
1636 
1637 	return batch;
1638 }
1639 
1640 static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
1641 {
1642 	static const struct lri lri[] = {
1643 		/* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl,glk */
1644 		{
1645 			COMMON_SLICE_CHICKEN2,
1646 			__MASKED_FIELD(GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE,
1647 				       0),
1648 		},
1649 
1650 		/* BSpec: 11391 */
1651 		{
1652 			FF_SLICE_CHICKEN,
1653 			__MASKED_FIELD(FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX,
1654 				       FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX),
1655 		},
1656 
1657 		/* BSpec: 11299 */
1658 		{
1659 			_3D_CHICKEN3,
1660 			__MASKED_FIELD(_3D_CHICKEN_SF_PROVOKING_VERTEX_FIX,
1661 				       _3D_CHICKEN_SF_PROVOKING_VERTEX_FIX),
1662 		}
1663 	};
1664 
1665 	*batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1666 
1667 	/* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt,glk */
1668 	batch = gen8_emit_flush_coherentl3_wa(engine, batch);
1669 
1670 	/* WaClearSlmSpaceAtContextSwitch:skl,bxt,kbl,glk,cfl */
1671 	batch = gen8_emit_pipe_control(batch,
1672 				       PIPE_CONTROL_FLUSH_L3 |
1673 				       PIPE_CONTROL_STORE_DATA_INDEX |
1674 				       PIPE_CONTROL_CS_STALL |
1675 				       PIPE_CONTROL_QW_WRITE,
1676 				       LRC_PPHWSP_SCRATCH_ADDR);
1677 
1678 	batch = emit_lri(batch, lri, ARRAY_SIZE(lri));
1679 
1680 	/* WaMediaPoolStateCmdInWABB:bxt,glk */
1681 	if (HAS_POOLED_EU(engine->i915)) {
1682 		/*
1683 		 * EU pool configuration is setup along with golden context
1684 		 * during context initialization. This value depends on
1685 		 * device type (2x6 or 3x6) and needs to be updated based
1686 		 * on which subslice is disabled especially for 2x6
1687 		 * devices, however it is safe to load default
1688 		 * configuration of 3x6 device instead of masking off
1689 		 * corresponding bits because HW ignores bits of a disabled
1690 		 * subslice and drops down to appropriate config. Please
1691 		 * see render_state_setup() in i915_gem_render_state.c for
1692 		 * possible configurations, to avoid duplication they are
1693 		 * not shown here again.
1694 		 */
1695 		*batch++ = GEN9_MEDIA_POOL_STATE;
1696 		*batch++ = GEN9_MEDIA_POOL_ENABLE;
1697 		*batch++ = 0x00777000;
1698 		*batch++ = 0;
1699 		*batch++ = 0;
1700 		*batch++ = 0;
1701 	}
1702 
1703 	*batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1704 
1705 	/* Pad to end of cacheline */
1706 	while ((unsigned long)batch % CACHELINE_BYTES)
1707 		*batch++ = MI_NOOP;
1708 
1709 	return batch;
1710 }
1711 
1712 #define CTX_WA_BB_SIZE (PAGE_SIZE)
1713 
1714 static int lrc_create_wa_ctx(struct intel_engine_cs *engine)
1715 {
1716 	struct drm_i915_gem_object *obj;
1717 	struct i915_vma *vma;
1718 	int err;
1719 
1720 	obj = i915_gem_object_create_shmem(engine->i915, CTX_WA_BB_SIZE);
1721 	if (IS_ERR(obj))
1722 		return PTR_ERR(obj);
1723 
1724 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
1725 	if (IS_ERR(vma)) {
1726 		err = PTR_ERR(vma);
1727 		goto err;
1728 	}
1729 
1730 	engine->wa_ctx.vma = vma;
1731 	return 0;
1732 
1733 err:
1734 	i915_gem_object_put(obj);
1735 	return err;
1736 }
1737 
1738 void lrc_fini_wa_ctx(struct intel_engine_cs *engine)
1739 {
1740 	i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1741 }
1742 
1743 typedef u32 *(*wa_bb_func_t)(struct intel_engine_cs *engine, u32 *batch);
1744 
1745 void lrc_init_wa_ctx(struct intel_engine_cs *engine)
1746 {
1747 	struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1748 	struct i915_wa_ctx_bb *wa_bb[] = {
1749 		&wa_ctx->indirect_ctx, &wa_ctx->per_ctx
1750 	};
1751 	wa_bb_func_t wa_bb_fn[ARRAY_SIZE(wa_bb)];
1752 	struct i915_gem_ww_ctx ww;
1753 	void *batch, *batch_ptr;
1754 	unsigned int i;
1755 	int err;
1756 
1757 	if (GRAPHICS_VER(engine->i915) >= 11 ||
1758 	    !(engine->flags & I915_ENGINE_HAS_RCS_REG_STATE))
1759 		return;
1760 
1761 	if (GRAPHICS_VER(engine->i915) == 9) {
1762 		wa_bb_fn[0] = gen9_init_indirectctx_bb;
1763 		wa_bb_fn[1] = NULL;
1764 	} else if (GRAPHICS_VER(engine->i915) == 8) {
1765 		wa_bb_fn[0] = gen8_init_indirectctx_bb;
1766 		wa_bb_fn[1] = NULL;
1767 	}
1768 
1769 	err = lrc_create_wa_ctx(engine);
1770 	if (err) {
1771 		/*
1772 		 * We continue even if we fail to initialize WA batch
1773 		 * because we only expect rare glitches but nothing
1774 		 * critical to prevent us from using GPU
1775 		 */
1776 		drm_err(&engine->i915->drm,
1777 			"Ignoring context switch w/a allocation error:%d\n",
1778 			err);
1779 		return;
1780 	}
1781 
1782 	if (!engine->wa_ctx.vma)
1783 		return;
1784 
1785 	i915_gem_ww_ctx_init(&ww, true);
1786 retry:
1787 	err = i915_gem_object_lock(wa_ctx->vma->obj, &ww);
1788 	if (!err)
1789 		err = i915_ggtt_pin(wa_ctx->vma, &ww, 0, PIN_HIGH);
1790 	if (err)
1791 		goto err;
1792 
1793 	batch = i915_gem_object_pin_map(wa_ctx->vma->obj, I915_MAP_WB);
1794 	if (IS_ERR(batch)) {
1795 		err = PTR_ERR(batch);
1796 		goto err_unpin;
1797 	}
1798 
1799 	/*
1800 	 * Emit the two workaround batch buffers, recording the offset from the
1801 	 * start of the workaround batch buffer object for each and their
1802 	 * respective sizes.
1803 	 */
1804 	batch_ptr = batch;
1805 	for (i = 0; i < ARRAY_SIZE(wa_bb_fn); i++) {
1806 		wa_bb[i]->offset = batch_ptr - batch;
1807 		if (GEM_DEBUG_WARN_ON(!IS_ALIGNED(wa_bb[i]->offset,
1808 						  CACHELINE_BYTES))) {
1809 			err = -EINVAL;
1810 			break;
1811 		}
1812 		if (wa_bb_fn[i])
1813 			batch_ptr = wa_bb_fn[i](engine, batch_ptr);
1814 		wa_bb[i]->size = batch_ptr - (batch + wa_bb[i]->offset);
1815 	}
1816 	GEM_BUG_ON(batch_ptr - batch > CTX_WA_BB_SIZE);
1817 
1818 	__i915_gem_object_flush_map(wa_ctx->vma->obj, 0, batch_ptr - batch);
1819 	__i915_gem_object_release_map(wa_ctx->vma->obj);
1820 
1821 	/* Verify that we can handle failure to setup the wa_ctx */
1822 	if (!err)
1823 		err = i915_inject_probe_error(engine->i915, -ENODEV);
1824 
1825 err_unpin:
1826 	if (err)
1827 		i915_vma_unpin(wa_ctx->vma);
1828 err:
1829 	if (err == -EDEADLK) {
1830 		err = i915_gem_ww_ctx_backoff(&ww);
1831 		if (!err)
1832 			goto retry;
1833 	}
1834 	i915_gem_ww_ctx_fini(&ww);
1835 
1836 	if (err) {
1837 		i915_vma_put(engine->wa_ctx.vma);
1838 
1839 		/* Clear all flags to prevent further use */
1840 		memset(wa_ctx, 0, sizeof(*wa_ctx));
1841 	}
1842 }
1843 
1844 static void st_runtime_underflow(struct intel_context_stats *stats, s32 dt)
1845 {
1846 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1847 	stats->runtime.num_underflow++;
1848 	stats->runtime.max_underflow =
1849 		max_t(u32, stats->runtime.max_underflow, -dt);
1850 #endif
1851 }
1852 
1853 static u32 lrc_get_runtime(const struct intel_context *ce)
1854 {
1855 	/*
1856 	 * We can use either ppHWSP[16] which is recorded before the context
1857 	 * switch (and so excludes the cost of context switches) or use the
1858 	 * value from the context image itself, which is saved/restored earlier
1859 	 * and so includes the cost of the save.
1860 	 */
1861 	return READ_ONCE(ce->lrc_reg_state[CTX_TIMESTAMP]);
1862 }
1863 
1864 void lrc_update_runtime(struct intel_context *ce)
1865 {
1866 	struct intel_context_stats *stats = &ce->stats;
1867 	u32 old;
1868 	s32 dt;
1869 
1870 	old = stats->runtime.last;
1871 	stats->runtime.last = lrc_get_runtime(ce);
1872 	dt = stats->runtime.last - old;
1873 	if (!dt)
1874 		return;
1875 
1876 	if (unlikely(dt < 0)) {
1877 		CE_TRACE(ce, "runtime underflow: last=%u, new=%u, delta=%d\n",
1878 			 old, stats->runtime.last, dt);
1879 		st_runtime_underflow(stats, dt);
1880 		return;
1881 	}
1882 
1883 	ewma_runtime_add(&stats->runtime.avg, dt);
1884 	stats->runtime.total += dt;
1885 }
1886 
1887 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1888 #include "selftest_lrc.c"
1889 #endif
1890