xref: /linux/drivers/gpu/drm/xe/xe_ring_ops.c (revision 1b5d39e6672fdee158c3306f5cb2df8975c77e5a)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2022 Intel Corporation
4  */
5 
6 #include "xe_ring_ops.h"
7 
8 #include <generated/xe_wa_oob.h>
9 
10 #include "instructions/xe_gpu_commands.h"
11 #include "instructions/xe_mi_commands.h"
12 #include "regs/xe_engine_regs.h"
13 #include "regs/xe_gt_regs.h"
14 #include "regs/xe_lrc_layout.h"
15 #include "xe_exec_queue.h"
16 #include "xe_gt.h"
17 #include "xe_lrc.h"
18 #include "xe_macros.h"
19 #include "xe_sched_job.h"
20 #include "xe_sriov.h"
21 #include "xe_vm_types.h"
22 #include "xe_vm.h"
23 #include "xe_wa.h"
24 
25 /*
26  * 3D-related flags that can't be set on _engines_ that lack access to the 3D
27  * pipeline (i.e., CCS engines).
28  */
29 #define PIPE_CONTROL_3D_ENGINE_FLAGS (\
30 		PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH | \
31 		PIPE_CONTROL_DEPTH_CACHE_FLUSH | \
32 		PIPE_CONTROL_TILE_CACHE_FLUSH | \
33 		PIPE_CONTROL_DEPTH_STALL | \
34 		PIPE_CONTROL_STALL_AT_SCOREBOARD | \
35 		PIPE_CONTROL_PSD_SYNC | \
36 		PIPE_CONTROL_AMFS_FLUSH | \
37 		PIPE_CONTROL_VF_CACHE_INVALIDATE | \
38 		PIPE_CONTROL_GLOBAL_SNAPSHOT_RESET)
39 
40 /* 3D-related flags that can't be set on _platforms_ that lack a 3D pipeline */
41 #define PIPE_CONTROL_3D_ARCH_FLAGS ( \
42 		PIPE_CONTROL_3D_ENGINE_FLAGS | \
43 		PIPE_CONTROL_INDIRECT_STATE_DISABLE | \
44 		PIPE_CONTROL_FLUSH_ENABLE | \
45 		PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE | \
46 		PIPE_CONTROL_DC_FLUSH_ENABLE)
47 
48 static u32 preparser_disable(bool state)
49 {
50 	return MI_ARB_CHECK | BIT(8) | state;
51 }
52 
53 static int emit_aux_table_inv(struct xe_gt *gt, struct xe_reg reg,
54 			      u32 *dw, int i)
55 {
56 	dw[i++] = MI_LOAD_REGISTER_IMM | MI_LRI_NUM_REGS(1) | MI_LRI_MMIO_REMAP_EN;
57 	dw[i++] = reg.addr + gt->mmio.adj_offset;
58 	dw[i++] = AUX_INV;
59 	dw[i++] = MI_NOOP;
60 
61 	return i;
62 }
63 
64 static int emit_user_interrupt(u32 *dw, int i)
65 {
66 	dw[i++] = MI_USER_INTERRUPT;
67 	dw[i++] = MI_ARB_ON_OFF | MI_ARB_ENABLE;
68 	dw[i++] = MI_ARB_CHECK;
69 
70 	return i;
71 }
72 
73 static int emit_store_imm_ggtt(u32 addr, u32 value, u32 *dw, int i)
74 {
75 	dw[i++] = MI_STORE_DATA_IMM | MI_SDI_GGTT | MI_SDI_NUM_DW(1);
76 	dw[i++] = addr;
77 	dw[i++] = 0;
78 	dw[i++] = value;
79 
80 	return i;
81 }
82 
83 static int emit_flush_dw(u32 *dw, int i)
84 {
85 	dw[i++] = MI_FLUSH_DW | MI_FLUSH_IMM_DW;
86 	dw[i++] = 0;
87 	dw[i++] = 0;
88 	dw[i++] = 0;
89 
90 	return i;
91 }
92 
93 static int emit_flush_imm_ggtt(u32 addr, u32 value, u32 flags, u32 *dw, int i)
94 {
95 	dw[i++] = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_IMM_DW |
96 		  flags;
97 	dw[i++] = addr | MI_FLUSH_DW_USE_GTT;
98 	dw[i++] = 0;
99 	dw[i++] = value;
100 
101 	return i;
102 }
103 
104 static int emit_bb_start(u64 batch_addr, u32 ppgtt_flag, u32 *dw, int i)
105 {
106 	dw[i++] = MI_BATCH_BUFFER_START | ppgtt_flag | XE_INSTR_NUM_DW(3);
107 	dw[i++] = lower_32_bits(batch_addr);
108 	dw[i++] = upper_32_bits(batch_addr);
109 
110 	return i;
111 }
112 
113 static int emit_flush_invalidate(u32 addr, u32 val, u32 flush_flags, u32 *dw, int i)
114 {
115 	dw[i++] = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW |
116 		  MI_FLUSH_IMM_DW | (flush_flags & MI_INVALIDATE_TLB) ?: 0;
117 
118 	dw[i++] = addr | MI_FLUSH_DW_USE_GTT;
119 	dw[i++] = 0;
120 	dw[i++] = val;
121 
122 	return i;
123 }
124 
125 static int
126 emit_pipe_control(u32 *dw, int i, u32 bit_group_0, u32 bit_group_1, u32 offset, u32 value)
127 {
128 	dw[i++] = GFX_OP_PIPE_CONTROL(6) | bit_group_0;
129 	dw[i++] = bit_group_1;
130 	dw[i++] = offset;
131 	dw[i++] = 0;
132 	dw[i++] = value;
133 	dw[i++] = 0;
134 
135 	return i;
136 }
137 
138 static int emit_pipe_invalidate(struct xe_exec_queue *q, u32 mask_flags,
139 				bool invalidate_tlb, u32 *dw, int i)
140 {
141 	u32 flags0 = 0;
142 	u32 flags1 = PIPE_CONTROL_COMMAND_CACHE_INVALIDATE |
143 		PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE |
144 		PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
145 		PIPE_CONTROL_VF_CACHE_INVALIDATE |
146 		PIPE_CONTROL_CONST_CACHE_INVALIDATE |
147 		PIPE_CONTROL_STATE_CACHE_INVALIDATE |
148 		PIPE_CONTROL_QW_WRITE |
149 		PIPE_CONTROL_STORE_DATA_INDEX;
150 
151 	if (invalidate_tlb)
152 		flags1 |= PIPE_CONTROL_TLB_INVALIDATE;
153 
154 	if (xe_exec_queue_is_multi_queue(q))
155 		flags0 |= PIPE_CONTROL0_QUEUE_DRAIN_MODE;
156 	else
157 		flags1 |= PIPE_CONTROL_CS_STALL;
158 
159 	flags1 &= ~mask_flags;
160 
161 	if (flags1 & PIPE_CONTROL_VF_CACHE_INVALIDATE)
162 		flags0 |= PIPE_CONTROL0_L3_READ_ONLY_CACHE_INVALIDATE;
163 
164 	return emit_pipe_control(dw, i, flags0, flags1,
165 				 LRC_PPHWSP_FLUSH_INVAL_SCRATCH_ADDR, 0);
166 }
167 
168 static int emit_store_imm_ppgtt_posted(u64 addr, u64 value,
169 				       u32 *dw, int i)
170 {
171 	dw[i++] = MI_STORE_DATA_IMM | MI_SDI_NUM_QW(1);
172 	dw[i++] = lower_32_bits(addr);
173 	dw[i++] = upper_32_bits(addr);
174 	dw[i++] = lower_32_bits(value);
175 	dw[i++] = upper_32_bits(value);
176 
177 	return i;
178 }
179 
180 static int emit_render_cache_flush(struct xe_sched_job *job, u32 *dw, int i)
181 {
182 	struct xe_exec_queue *q = job->q;
183 	struct xe_gt *gt = q->gt;
184 	bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
185 	u32 flags0, flags1;
186 
187 	if (XE_GT_WA(gt, 14016712196))
188 		i = emit_pipe_control(dw, i, 0, PIPE_CONTROL_DEPTH_CACHE_FLUSH,
189 				      LRC_PPHWSP_FLUSH_INVAL_SCRATCH_ADDR, 0);
190 
191 	flags0 = PIPE_CONTROL0_HDC_PIPELINE_FLUSH;
192 	flags1 = (PIPE_CONTROL_TILE_CACHE_FLUSH |
193 		 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
194 		 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
195 		 PIPE_CONTROL_DC_FLUSH_ENABLE |
196 		 PIPE_CONTROL_FLUSH_ENABLE);
197 
198 	if (XE_GT_WA(gt, 1409600907))
199 		flags1 |= PIPE_CONTROL_DEPTH_STALL;
200 
201 	if (lacks_render)
202 		flags1 &= ~PIPE_CONTROL_3D_ARCH_FLAGS;
203 	else if (job->q->class == XE_ENGINE_CLASS_COMPUTE)
204 		flags1 &= ~PIPE_CONTROL_3D_ENGINE_FLAGS;
205 
206 	if (xe_exec_queue_is_multi_queue(q))
207 		flags0 |= PIPE_CONTROL0_QUEUE_DRAIN_MODE;
208 	else
209 		flags1 |= PIPE_CONTROL_CS_STALL;
210 
211 	return emit_pipe_control(dw, i, flags0, flags1, 0, 0);
212 }
213 
214 static int emit_pipe_control_to_ring_end(struct xe_exec_queue *q, u32 *dw, int i)
215 {
216 	struct xe_hw_engine *hwe = q->hwe;
217 
218 	if (hwe->class != XE_ENGINE_CLASS_RENDER)
219 		return i;
220 
221 	xe_gt_assert(q->gt, !xe_exec_queue_is_multi_queue(q));
222 
223 	if (XE_GT_WA(hwe->gt, 16020292621))
224 		i = emit_pipe_control(dw, i, 0, PIPE_CONTROL_LRI_POST_SYNC,
225 				      RING_NOPID(hwe->mmio_base).addr, 0);
226 
227 	return i;
228 }
229 
230 static int emit_pipe_imm_ggtt(struct xe_exec_queue *q, u32 addr, u32 value,
231 			      bool stall_only, u32 *dw, int i)
232 {
233 	u32 flags0 = 0, flags1 = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_QW_WRITE;
234 
235 	if (!stall_only)
236 		flags1 |= PIPE_CONTROL_FLUSH_ENABLE;
237 
238 	if (xe_exec_queue_is_multi_queue(q))
239 		flags0 |= PIPE_CONTROL0_QUEUE_DRAIN_MODE;
240 	else
241 		flags1 |= PIPE_CONTROL_CS_STALL;
242 
243 	return emit_pipe_control(dw, i, flags0, flags1, addr, value);
244 }
245 
246 static u32 get_ppgtt_flag(struct xe_sched_job *job)
247 {
248 	if (job->q->vm && !job->ggtt)
249 		return BIT(8);
250 
251 	return 0;
252 }
253 
254 static int emit_copy_timestamp(struct xe_lrc *lrc, u32 *dw, int i)
255 {
256 	dw[i++] = MI_STORE_REGISTER_MEM | MI_SRM_USE_GGTT | MI_SRM_ADD_CS_OFFSET;
257 	dw[i++] = RING_CTX_TIMESTAMP(0).addr;
258 	dw[i++] = xe_lrc_ctx_job_timestamp_ggtt_addr(lrc);
259 	dw[i++] = 0;
260 
261 	return i;
262 }
263 
264 /* for engines that don't require any special HW handling (no EUs, no aux inval, etc) */
265 static void __emit_job_gen12_simple(struct xe_sched_job *job, struct xe_lrc *lrc,
266 				    u64 batch_addr, u32 *head, u32 seqno)
267 {
268 	u32 dw[MAX_JOB_SIZE_DW], i = 0;
269 	u32 ppgtt_flag = get_ppgtt_flag(job);
270 	struct xe_gt *gt = job->q->gt;
271 
272 	*head = lrc->ring.tail;
273 
274 	i = emit_copy_timestamp(lrc, dw, i);
275 
276 	if (job->ring_ops_flush_tlb) {
277 		dw[i++] = preparser_disable(true);
278 		i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
279 					seqno, MI_INVALIDATE_TLB, dw, i);
280 		dw[i++] = preparser_disable(false);
281 	} else {
282 		i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
283 					seqno, dw, i);
284 	}
285 
286 	i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
287 
288 	if (job->user_fence.used) {
289 		i = emit_flush_dw(dw, i);
290 		i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
291 						job->user_fence.value,
292 						dw, i);
293 	}
294 
295 	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno, 0, dw, i);
296 
297 	i = emit_user_interrupt(dw, i);
298 
299 	xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
300 
301 	xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
302 }
303 
304 static bool has_aux_ccs(struct xe_device *xe)
305 {
306 	/*
307 	 * PVC is a special case that has no compression of either type
308 	 * (FlatCCS or AuxCCS).  Also, AuxCCS is no longer used from Xe2
309 	 * onward, so any future platforms with no FlatCCS will not have
310 	 * AuxCCS either.
311 	 */
312 	if (GRAPHICS_VER(xe) >= 20 || xe->info.platform == XE_PVC)
313 		return false;
314 
315 	return !xe->info.has_flat_ccs;
316 }
317 
318 static void __emit_job_gen12_video(struct xe_sched_job *job, struct xe_lrc *lrc,
319 				   u64 batch_addr, u32 *head, u32 seqno)
320 {
321 	u32 dw[MAX_JOB_SIZE_DW], i = 0;
322 	u32 ppgtt_flag = get_ppgtt_flag(job);
323 	struct xe_gt *gt = job->q->gt;
324 	struct xe_device *xe = gt_to_xe(gt);
325 	bool decode = job->q->class == XE_ENGINE_CLASS_VIDEO_DECODE;
326 
327 	*head = lrc->ring.tail;
328 
329 	i = emit_copy_timestamp(lrc, dw, i);
330 
331 	dw[i++] = preparser_disable(true);
332 
333 	/* hsdes: 1809175790 */
334 	if (has_aux_ccs(xe)) {
335 		if (decode)
336 			i = emit_aux_table_inv(gt, VD0_AUX_INV, dw, i);
337 		else
338 			i = emit_aux_table_inv(gt, VE0_AUX_INV, dw, i);
339 	}
340 
341 	if (job->ring_ops_flush_tlb)
342 		i = emit_flush_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
343 					seqno, MI_INVALIDATE_TLB, dw, i);
344 
345 	dw[i++] = preparser_disable(false);
346 
347 	if (!job->ring_ops_flush_tlb)
348 		i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
349 					seqno, dw, i);
350 
351 	i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
352 
353 	if (job->user_fence.used) {
354 		i = emit_flush_dw(dw, i);
355 		i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
356 						job->user_fence.value,
357 						dw, i);
358 	}
359 
360 	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno, 0, dw, i);
361 
362 	i = emit_user_interrupt(dw, i);
363 
364 	xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
365 
366 	xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
367 }
368 
369 static void __emit_job_gen12_render_compute(struct xe_sched_job *job,
370 					    struct xe_lrc *lrc,
371 					    u64 batch_addr, u32 *head,
372 					    u32 seqno)
373 {
374 	u32 dw[MAX_JOB_SIZE_DW], i = 0;
375 	u32 ppgtt_flag = get_ppgtt_flag(job);
376 	struct xe_gt *gt = job->q->gt;
377 	struct xe_device *xe = gt_to_xe(gt);
378 	bool lacks_render = !(gt->info.engine_mask & XE_HW_ENGINE_RCS_MASK);
379 	u32 mask_flags = 0;
380 
381 	*head = lrc->ring.tail;
382 
383 	i = emit_copy_timestamp(lrc, dw, i);
384 
385 	dw[i++] = preparser_disable(true);
386 	if (lacks_render)
387 		mask_flags = PIPE_CONTROL_3D_ARCH_FLAGS;
388 	else if (job->q->class == XE_ENGINE_CLASS_COMPUTE)
389 		mask_flags = PIPE_CONTROL_3D_ENGINE_FLAGS;
390 
391 	/* See __xe_pt_bind_vma() for a discussion on TLB invalidations. */
392 	i = emit_pipe_invalidate(job->q, mask_flags, job->ring_ops_flush_tlb, dw, i);
393 
394 	/* hsdes: 1809175790 */
395 	if (has_aux_ccs(xe))
396 		i = emit_aux_table_inv(gt, CCS_AUX_INV, dw, i);
397 
398 	dw[i++] = preparser_disable(false);
399 
400 	i = emit_store_imm_ggtt(xe_lrc_start_seqno_ggtt_addr(lrc),
401 				seqno, dw, i);
402 
403 	i = emit_bb_start(batch_addr, ppgtt_flag, dw, i);
404 
405 	i = emit_render_cache_flush(job, dw, i);
406 
407 	if (job->user_fence.used)
408 		i = emit_store_imm_ppgtt_posted(job->user_fence.addr,
409 						job->user_fence.value,
410 						dw, i);
411 
412 	i = emit_pipe_imm_ggtt(job->q, xe_lrc_seqno_ggtt_addr(lrc), seqno, lacks_render, dw, i);
413 
414 	i = emit_user_interrupt(dw, i);
415 
416 	i = emit_pipe_control_to_ring_end(job->q, dw, i);
417 
418 	xe_gt_assert(gt, i <= MAX_JOB_SIZE_DW);
419 
420 	xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
421 }
422 
423 static void emit_migration_job_gen12(struct xe_sched_job *job,
424 				     struct xe_lrc *lrc, u32 *head,
425 				     u32 seqno)
426 {
427 	u32 saddr = xe_lrc_start_seqno_ggtt_addr(lrc);
428 	u32 dw[MAX_JOB_SIZE_DW], i = 0;
429 
430 	*head = lrc->ring.tail;
431 
432 	i = emit_copy_timestamp(lrc, dw, i);
433 
434 	i = emit_store_imm_ggtt(saddr, seqno, dw, i);
435 
436 	dw[i++] = MI_ARB_ON_OFF | MI_ARB_DISABLE; /* Enabled again below */
437 
438 	i = emit_bb_start(job->ptrs[0].batch_addr, BIT(8), dw, i);
439 
440 	dw[i++] = preparser_disable(true);
441 	i = emit_flush_invalidate(saddr, seqno, job->migrate_flush_flags, dw, i);
442 	dw[i++] = preparser_disable(false);
443 
444 	i = emit_bb_start(job->ptrs[1].batch_addr, BIT(8), dw, i);
445 
446 	i = emit_flush_imm_ggtt(xe_lrc_seqno_ggtt_addr(lrc), seqno,
447 				job->migrate_flush_flags,
448 				dw, i);
449 
450 	i = emit_user_interrupt(dw, i);
451 
452 	xe_gt_assert(job->q->gt, i <= MAX_JOB_SIZE_DW);
453 
454 	xe_lrc_write_ring(lrc, dw, i * sizeof(*dw));
455 }
456 
457 static void emit_job_gen12_gsc(struct xe_sched_job *job)
458 {
459 	struct xe_gt *gt = job->q->gt;
460 
461 	xe_gt_assert(gt, job->q->width <= 1); /* no parallel submission for GSCCS */
462 
463 	__emit_job_gen12_simple(job, job->q->lrc[0],
464 				job->ptrs[0].batch_addr,
465 				&job->ptrs[0].head,
466 				xe_sched_job_lrc_seqno(job));
467 }
468 
469 static void emit_job_gen12_copy(struct xe_sched_job *job)
470 {
471 	int i;
472 
473 	if (xe_sched_job_is_migration(job->q)) {
474 		emit_migration_job_gen12(job, job->q->lrc[0],
475 					 &job->ptrs[0].head,
476 					 xe_sched_job_lrc_seqno(job));
477 		return;
478 	}
479 
480 	for (i = 0; i < job->q->width; ++i)
481 		__emit_job_gen12_simple(job, job->q->lrc[i],
482 					job->ptrs[i].batch_addr,
483 					&job->ptrs[i].head,
484 					xe_sched_job_lrc_seqno(job));
485 }
486 
487 static void emit_job_gen12_video(struct xe_sched_job *job)
488 {
489 	int i;
490 
491 	/* FIXME: Not doing parallel handshake for now */
492 	for (i = 0; i < job->q->width; ++i)
493 		__emit_job_gen12_video(job, job->q->lrc[i],
494 				       job->ptrs[i].batch_addr,
495 				       &job->ptrs[i].head,
496 				       xe_sched_job_lrc_seqno(job));
497 }
498 
499 static void emit_job_gen12_render_compute(struct xe_sched_job *job)
500 {
501 	int i;
502 
503 	for (i = 0; i < job->q->width; ++i)
504 		__emit_job_gen12_render_compute(job, job->q->lrc[i],
505 						job->ptrs[i].batch_addr,
506 						&job->ptrs[i].head,
507 						xe_sched_job_lrc_seqno(job));
508 }
509 
510 static const struct xe_ring_ops ring_ops_gen12_gsc = {
511 	.emit_job = emit_job_gen12_gsc,
512 };
513 
514 static const struct xe_ring_ops ring_ops_gen12_copy = {
515 	.emit_job = emit_job_gen12_copy,
516 };
517 
518 static const struct xe_ring_ops ring_ops_gen12_video = {
519 	.emit_job = emit_job_gen12_video,
520 };
521 
522 static const struct xe_ring_ops ring_ops_gen12_render_compute = {
523 	.emit_job = emit_job_gen12_render_compute,
524 };
525 
526 const struct xe_ring_ops *
527 xe_ring_ops_get(struct xe_gt *gt, enum xe_engine_class class)
528 {
529 	switch (class) {
530 	case XE_ENGINE_CLASS_OTHER:
531 		return &ring_ops_gen12_gsc;
532 	case XE_ENGINE_CLASS_COPY:
533 		return &ring_ops_gen12_copy;
534 	case XE_ENGINE_CLASS_VIDEO_DECODE:
535 	case XE_ENGINE_CLASS_VIDEO_ENHANCE:
536 		return &ring_ops_gen12_video;
537 	case XE_ENGINE_CLASS_RENDER:
538 	case XE_ENGINE_CLASS_COMPUTE:
539 		return &ring_ops_gen12_render_compute;
540 	default:
541 		return NULL;
542 	}
543 }
544