xref: /linux/drivers/gpu/drm/i915/gt/intel_ring_submission.c (revision 15a1fbdcfb519c2bd291ed01c6c94e0b89537a77)
1 /*
2  * Copyright © 2008-2010 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *    Zou Nan hai <nanhai.zou@intel.com>
26  *    Xiang Hai hao<haihao.xiang@intel.com>
27  *
28  */
29 
30 #include <linux/log2.h>
31 
32 #include <drm/i915_drm.h>
33 
34 #include "gem/i915_gem_context.h"
35 
36 #include "gen6_ppgtt.h"
37 #include "i915_drv.h"
38 #include "i915_trace.h"
39 #include "intel_context.h"
40 #include "intel_gt.h"
41 #include "intel_gt_irq.h"
42 #include "intel_gt_pm_irq.h"
43 #include "intel_reset.h"
44 #include "intel_ring.h"
45 #include "intel_workarounds.h"
46 
47 /* Rough estimate of the typical request size, performing a flush,
48  * set-context and then emitting the batch.
49  */
50 #define LEGACY_REQUEST_SIZE 200
51 
52 static int
53 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
54 {
55 	unsigned int num_store_dw;
56 	u32 cmd, *cs;
57 
58 	cmd = MI_FLUSH;
59 	num_store_dw = 0;
60 	if (mode & EMIT_INVALIDATE)
61 		cmd |= MI_READ_FLUSH;
62 	if (mode & EMIT_FLUSH)
63 		num_store_dw = 4;
64 
65 	cs = intel_ring_begin(rq, 2 + 3 * num_store_dw);
66 	if (IS_ERR(cs))
67 		return PTR_ERR(cs);
68 
69 	*cs++ = cmd;
70 	while (num_store_dw--) {
71 		*cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
72 		*cs++ = intel_gt_scratch_offset(rq->engine->gt,
73 						INTEL_GT_SCRATCH_FIELD_DEFAULT);
74 		*cs++ = 0;
75 	}
76 	*cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH;
77 
78 	intel_ring_advance(rq, cs);
79 
80 	return 0;
81 }
82 
83 static int
84 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
85 {
86 	u32 cmd, *cs;
87 	int i;
88 
89 	/*
90 	 * read/write caches:
91 	 *
92 	 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
93 	 * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
94 	 * also flushed at 2d versus 3d pipeline switches.
95 	 *
96 	 * read-only caches:
97 	 *
98 	 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
99 	 * MI_READ_FLUSH is set, and is always flushed on 965.
100 	 *
101 	 * I915_GEM_DOMAIN_COMMAND may not exist?
102 	 *
103 	 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
104 	 * invalidated when MI_EXE_FLUSH is set.
105 	 *
106 	 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
107 	 * invalidated with every MI_FLUSH.
108 	 *
109 	 * TLBs:
110 	 *
111 	 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
112 	 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
113 	 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
114 	 * are flushed at any MI_FLUSH.
115 	 */
116 
117 	cmd = MI_FLUSH;
118 	if (mode & EMIT_INVALIDATE) {
119 		cmd |= MI_EXE_FLUSH;
120 		if (IS_G4X(rq->i915) || IS_GEN(rq->i915, 5))
121 			cmd |= MI_INVALIDATE_ISP;
122 	}
123 
124 	i = 2;
125 	if (mode & EMIT_INVALIDATE)
126 		i += 20;
127 
128 	cs = intel_ring_begin(rq, i);
129 	if (IS_ERR(cs))
130 		return PTR_ERR(cs);
131 
132 	*cs++ = cmd;
133 
134 	/*
135 	 * A random delay to let the CS invalidate take effect? Without this
136 	 * delay, the GPU relocation path fails as the CS does not see
137 	 * the updated contents. Just as important, if we apply the flushes
138 	 * to the EMIT_FLUSH branch (i.e. immediately after the relocation
139 	 * write and before the invalidate on the next batch), the relocations
140 	 * still fail. This implies that is a delay following invalidation
141 	 * that is required to reset the caches as opposed to a delay to
142 	 * ensure the memory is written.
143 	 */
144 	if (mode & EMIT_INVALIDATE) {
145 		*cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
146 		*cs++ = intel_gt_scratch_offset(rq->engine->gt,
147 						INTEL_GT_SCRATCH_FIELD_DEFAULT) |
148 			PIPE_CONTROL_GLOBAL_GTT;
149 		*cs++ = 0;
150 		*cs++ = 0;
151 
152 		for (i = 0; i < 12; i++)
153 			*cs++ = MI_FLUSH;
154 
155 		*cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
156 		*cs++ = intel_gt_scratch_offset(rq->engine->gt,
157 						INTEL_GT_SCRATCH_FIELD_DEFAULT) |
158 			PIPE_CONTROL_GLOBAL_GTT;
159 		*cs++ = 0;
160 		*cs++ = 0;
161 	}
162 
163 	*cs++ = cmd;
164 
165 	intel_ring_advance(rq, cs);
166 
167 	return 0;
168 }
169 
170 /*
171  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
172  * implementing two workarounds on gen6.  From section 1.4.7.1
173  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
174  *
175  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
176  * produced by non-pipelined state commands), software needs to first
177  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
178  * 0.
179  *
180  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
181  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
182  *
183  * And the workaround for these two requires this workaround first:
184  *
185  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
186  * BEFORE the pipe-control with a post-sync op and no write-cache
187  * flushes.
188  *
189  * And this last workaround is tricky because of the requirements on
190  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
191  * volume 2 part 1:
192  *
193  *     "1 of the following must also be set:
194  *      - Render Target Cache Flush Enable ([12] of DW1)
195  *      - Depth Cache Flush Enable ([0] of DW1)
196  *      - Stall at Pixel Scoreboard ([1] of DW1)
197  *      - Depth Stall ([13] of DW1)
198  *      - Post-Sync Operation ([13] of DW1)
199  *      - Notify Enable ([8] of DW1)"
200  *
201  * The cache flushes require the workaround flush that triggered this
202  * one, so we can't use it.  Depth stall would trigger the same.
203  * Post-sync nonzero is what triggered this second workaround, so we
204  * can't use that one either.  Notify enable is IRQs, which aren't
205  * really our business.  That leaves only stall at scoreboard.
206  */
207 static int
208 gen6_emit_post_sync_nonzero_flush(struct i915_request *rq)
209 {
210 	u32 scratch_addr =
211 		intel_gt_scratch_offset(rq->engine->gt,
212 					INTEL_GT_SCRATCH_FIELD_RENDER_FLUSH);
213 	u32 *cs;
214 
215 	cs = intel_ring_begin(rq, 6);
216 	if (IS_ERR(cs))
217 		return PTR_ERR(cs);
218 
219 	*cs++ = GFX_OP_PIPE_CONTROL(5);
220 	*cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
221 	*cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
222 	*cs++ = 0; /* low dword */
223 	*cs++ = 0; /* high dword */
224 	*cs++ = MI_NOOP;
225 	intel_ring_advance(rq, cs);
226 
227 	cs = intel_ring_begin(rq, 6);
228 	if (IS_ERR(cs))
229 		return PTR_ERR(cs);
230 
231 	*cs++ = GFX_OP_PIPE_CONTROL(5);
232 	*cs++ = PIPE_CONTROL_QW_WRITE;
233 	*cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
234 	*cs++ = 0;
235 	*cs++ = 0;
236 	*cs++ = MI_NOOP;
237 	intel_ring_advance(rq, cs);
238 
239 	return 0;
240 }
241 
242 static int
243 gen6_render_ring_flush(struct i915_request *rq, u32 mode)
244 {
245 	u32 scratch_addr =
246 		intel_gt_scratch_offset(rq->engine->gt,
247 					INTEL_GT_SCRATCH_FIELD_RENDER_FLUSH);
248 	u32 *cs, flags = 0;
249 	int ret;
250 
251 	/* Force SNB workarounds for PIPE_CONTROL flushes */
252 	ret = gen6_emit_post_sync_nonzero_flush(rq);
253 	if (ret)
254 		return ret;
255 
256 	/* Just flush everything.  Experiments have shown that reducing the
257 	 * number of bits based on the write domains has little performance
258 	 * impact.
259 	 */
260 	if (mode & EMIT_FLUSH) {
261 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
262 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
263 		/*
264 		 * Ensure that any following seqno writes only happen
265 		 * when the render cache is indeed flushed.
266 		 */
267 		flags |= PIPE_CONTROL_CS_STALL;
268 	}
269 	if (mode & EMIT_INVALIDATE) {
270 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
271 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
272 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
273 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
274 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
275 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
276 		/*
277 		 * TLB invalidate requires a post-sync write.
278 		 */
279 		flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
280 	}
281 
282 	cs = intel_ring_begin(rq, 4);
283 	if (IS_ERR(cs))
284 		return PTR_ERR(cs);
285 
286 	*cs++ = GFX_OP_PIPE_CONTROL(4);
287 	*cs++ = flags;
288 	*cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
289 	*cs++ = 0;
290 	intel_ring_advance(rq, cs);
291 
292 	return 0;
293 }
294 
295 static u32 *gen6_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
296 {
297 	/* First we do the gen6_emit_post_sync_nonzero_flush w/a */
298 	*cs++ = GFX_OP_PIPE_CONTROL(4);
299 	*cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
300 	*cs++ = 0;
301 	*cs++ = 0;
302 
303 	*cs++ = GFX_OP_PIPE_CONTROL(4);
304 	*cs++ = PIPE_CONTROL_QW_WRITE;
305 	*cs++ = intel_gt_scratch_offset(rq->engine->gt,
306 					INTEL_GT_SCRATCH_FIELD_DEFAULT) |
307 		PIPE_CONTROL_GLOBAL_GTT;
308 	*cs++ = 0;
309 
310 	/* Finally we can flush and with it emit the breadcrumb */
311 	*cs++ = GFX_OP_PIPE_CONTROL(4);
312 	*cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
313 		 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
314 		 PIPE_CONTROL_DC_FLUSH_ENABLE |
315 		 PIPE_CONTROL_QW_WRITE |
316 		 PIPE_CONTROL_CS_STALL);
317 	*cs++ = i915_request_active_timeline(rq)->hwsp_offset |
318 		PIPE_CONTROL_GLOBAL_GTT;
319 	*cs++ = rq->fence.seqno;
320 
321 	*cs++ = MI_USER_INTERRUPT;
322 	*cs++ = MI_NOOP;
323 
324 	rq->tail = intel_ring_offset(rq, cs);
325 	assert_ring_tail_valid(rq->ring, rq->tail);
326 
327 	return cs;
328 }
329 
330 static int
331 gen7_render_ring_cs_stall_wa(struct i915_request *rq)
332 {
333 	u32 *cs;
334 
335 	cs = intel_ring_begin(rq, 4);
336 	if (IS_ERR(cs))
337 		return PTR_ERR(cs);
338 
339 	*cs++ = GFX_OP_PIPE_CONTROL(4);
340 	*cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
341 	*cs++ = 0;
342 	*cs++ = 0;
343 	intel_ring_advance(rq, cs);
344 
345 	return 0;
346 }
347 
348 static int
349 gen7_render_ring_flush(struct i915_request *rq, u32 mode)
350 {
351 	u32 scratch_addr =
352 		intel_gt_scratch_offset(rq->engine->gt,
353 					INTEL_GT_SCRATCH_FIELD_RENDER_FLUSH);
354 	u32 *cs, flags = 0;
355 
356 	/*
357 	 * Ensure that any following seqno writes only happen when the render
358 	 * cache is indeed flushed.
359 	 *
360 	 * Workaround: 4th PIPE_CONTROL command (except the ones with only
361 	 * read-cache invalidate bits set) must have the CS_STALL bit set. We
362 	 * don't try to be clever and just set it unconditionally.
363 	 */
364 	flags |= PIPE_CONTROL_CS_STALL;
365 
366 	/*
367 	 * CS_STALL suggests at least a post-sync write.
368 	 */
369 	flags |= PIPE_CONTROL_QW_WRITE;
370 	flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
371 
372 	/* Just flush everything.  Experiments have shown that reducing the
373 	 * number of bits based on the write domains has little performance
374 	 * impact.
375 	 */
376 	if (mode & EMIT_FLUSH) {
377 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
378 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
379 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
380 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
381 	}
382 	if (mode & EMIT_INVALIDATE) {
383 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
384 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
385 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
386 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
387 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
388 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
389 		flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
390 
391 		/* Workaround: we must issue a pipe_control with CS-stall bit
392 		 * set before a pipe_control command that has the state cache
393 		 * invalidate bit set. */
394 		gen7_render_ring_cs_stall_wa(rq);
395 	}
396 
397 	cs = intel_ring_begin(rq, 4);
398 	if (IS_ERR(cs))
399 		return PTR_ERR(cs);
400 
401 	*cs++ = GFX_OP_PIPE_CONTROL(4);
402 	*cs++ = flags;
403 	*cs++ = scratch_addr;
404 	*cs++ = 0;
405 	intel_ring_advance(rq, cs);
406 
407 	return 0;
408 }
409 
410 static u32 *gen7_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
411 {
412 	*cs++ = GFX_OP_PIPE_CONTROL(4);
413 	*cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
414 		 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
415 		 PIPE_CONTROL_DC_FLUSH_ENABLE |
416 		 PIPE_CONTROL_FLUSH_ENABLE |
417 		 PIPE_CONTROL_QW_WRITE |
418 		 PIPE_CONTROL_GLOBAL_GTT_IVB |
419 		 PIPE_CONTROL_CS_STALL);
420 	*cs++ = i915_request_active_timeline(rq)->hwsp_offset;
421 	*cs++ = rq->fence.seqno;
422 
423 	*cs++ = MI_USER_INTERRUPT;
424 	*cs++ = MI_NOOP;
425 
426 	rq->tail = intel_ring_offset(rq, cs);
427 	assert_ring_tail_valid(rq->ring, rq->tail);
428 
429 	return cs;
430 }
431 
432 static u32 *gen6_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
433 {
434 	GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
435 	GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
436 
437 	*cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
438 	*cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
439 	*cs++ = rq->fence.seqno;
440 
441 	*cs++ = MI_USER_INTERRUPT;
442 
443 	rq->tail = intel_ring_offset(rq, cs);
444 	assert_ring_tail_valid(rq->ring, rq->tail);
445 
446 	return cs;
447 }
448 
449 #define GEN7_XCS_WA 32
450 static u32 *gen7_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
451 {
452 	int i;
453 
454 	GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
455 	GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
456 
457 	*cs++ = MI_FLUSH_DW | MI_INVALIDATE_TLB |
458 		MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
459 	*cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
460 	*cs++ = rq->fence.seqno;
461 
462 	for (i = 0; i < GEN7_XCS_WA; i++) {
463 		*cs++ = MI_STORE_DWORD_INDEX;
464 		*cs++ = I915_GEM_HWS_SEQNO_ADDR;
465 		*cs++ = rq->fence.seqno;
466 	}
467 
468 	*cs++ = MI_FLUSH_DW;
469 	*cs++ = 0;
470 	*cs++ = 0;
471 
472 	*cs++ = MI_USER_INTERRUPT;
473 	*cs++ = MI_NOOP;
474 
475 	rq->tail = intel_ring_offset(rq, cs);
476 	assert_ring_tail_valid(rq->ring, rq->tail);
477 
478 	return cs;
479 }
480 #undef GEN7_XCS_WA
481 
482 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
483 {
484 	/*
485 	 * Keep the render interrupt unmasked as this papers over
486 	 * lost interrupts following a reset.
487 	 */
488 	if (engine->class == RENDER_CLASS) {
489 		if (INTEL_GEN(engine->i915) >= 6)
490 			mask &= ~BIT(0);
491 		else
492 			mask &= ~I915_USER_INTERRUPT;
493 	}
494 
495 	intel_engine_set_hwsp_writemask(engine, mask);
496 }
497 
498 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
499 {
500 	u32 addr;
501 
502 	addr = lower_32_bits(phys);
503 	if (INTEL_GEN(engine->i915) >= 4)
504 		addr |= (phys >> 28) & 0xf0;
505 
506 	intel_uncore_write(engine->uncore, HWS_PGA, addr);
507 }
508 
509 static struct page *status_page(struct intel_engine_cs *engine)
510 {
511 	struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
512 
513 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
514 	return sg_page(obj->mm.pages->sgl);
515 }
516 
517 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
518 {
519 	set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
520 	set_hwstam(engine, ~0u);
521 }
522 
523 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
524 {
525 	i915_reg_t hwsp;
526 
527 	/*
528 	 * The ring status page addresses are no longer next to the rest of
529 	 * the ring registers as of gen7.
530 	 */
531 	if (IS_GEN(engine->i915, 7)) {
532 		switch (engine->id) {
533 		/*
534 		 * No more rings exist on Gen7. Default case is only to shut up
535 		 * gcc switch check warning.
536 		 */
537 		default:
538 			GEM_BUG_ON(engine->id);
539 			/* fallthrough */
540 		case RCS0:
541 			hwsp = RENDER_HWS_PGA_GEN7;
542 			break;
543 		case BCS0:
544 			hwsp = BLT_HWS_PGA_GEN7;
545 			break;
546 		case VCS0:
547 			hwsp = BSD_HWS_PGA_GEN7;
548 			break;
549 		case VECS0:
550 			hwsp = VEBOX_HWS_PGA_GEN7;
551 			break;
552 		}
553 	} else if (IS_GEN(engine->i915, 6)) {
554 		hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
555 	} else {
556 		hwsp = RING_HWS_PGA(engine->mmio_base);
557 	}
558 
559 	intel_uncore_write(engine->uncore, hwsp, offset);
560 	intel_uncore_posting_read(engine->uncore, hwsp);
561 }
562 
563 static void flush_cs_tlb(struct intel_engine_cs *engine)
564 {
565 	struct drm_i915_private *dev_priv = engine->i915;
566 
567 	if (!IS_GEN_RANGE(dev_priv, 6, 7))
568 		return;
569 
570 	/* ring should be idle before issuing a sync flush*/
571 	drm_WARN_ON(&dev_priv->drm,
572 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
573 
574 	ENGINE_WRITE(engine, RING_INSTPM,
575 		     _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
576 					INSTPM_SYNC_FLUSH));
577 	if (intel_wait_for_register(engine->uncore,
578 				    RING_INSTPM(engine->mmio_base),
579 				    INSTPM_SYNC_FLUSH, 0,
580 				    1000))
581 		DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
582 			  engine->name);
583 }
584 
585 static void ring_setup_status_page(struct intel_engine_cs *engine)
586 {
587 	set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
588 	set_hwstam(engine, ~0u);
589 
590 	flush_cs_tlb(engine);
591 }
592 
593 static bool stop_ring(struct intel_engine_cs *engine)
594 {
595 	struct drm_i915_private *dev_priv = engine->i915;
596 
597 	if (INTEL_GEN(dev_priv) > 2) {
598 		ENGINE_WRITE(engine,
599 			     RING_MI_MODE, _MASKED_BIT_ENABLE(STOP_RING));
600 		if (intel_wait_for_register(engine->uncore,
601 					    RING_MI_MODE(engine->mmio_base),
602 					    MODE_IDLE,
603 					    MODE_IDLE,
604 					    1000)) {
605 			DRM_ERROR("%s : timed out trying to stop ring\n",
606 				  engine->name);
607 
608 			/*
609 			 * Sometimes we observe that the idle flag is not
610 			 * set even though the ring is empty. So double
611 			 * check before giving up.
612 			 */
613 			if (ENGINE_READ(engine, RING_HEAD) !=
614 			    ENGINE_READ(engine, RING_TAIL))
615 				return false;
616 		}
617 	}
618 
619 	ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
620 
621 	ENGINE_WRITE(engine, RING_HEAD, 0);
622 	ENGINE_WRITE(engine, RING_TAIL, 0);
623 
624 	/* The ring must be empty before it is disabled */
625 	ENGINE_WRITE(engine, RING_CTL, 0);
626 
627 	return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
628 }
629 
630 static struct i915_address_space *vm_alias(struct i915_address_space *vm)
631 {
632 	if (i915_is_ggtt(vm))
633 		vm = &i915_vm_to_ggtt(vm)->alias->vm;
634 
635 	return vm;
636 }
637 
638 static void set_pp_dir(struct intel_engine_cs *engine)
639 {
640 	struct i915_address_space *vm = vm_alias(engine->gt->vm);
641 
642 	if (vm) {
643 		struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
644 
645 		ENGINE_WRITE(engine, RING_PP_DIR_DCLV, PP_DIR_DCLV_2G);
646 		ENGINE_WRITE(engine, RING_PP_DIR_BASE,
647 			     px_base(ppgtt->pd)->ggtt_offset << 10);
648 	}
649 }
650 
651 static int xcs_resume(struct intel_engine_cs *engine)
652 {
653 	struct drm_i915_private *dev_priv = engine->i915;
654 	struct intel_ring *ring = engine->legacy.ring;
655 	int ret = 0;
656 
657 	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
658 		     ring->head, ring->tail);
659 
660 	intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
661 
662 	/* WaClearRingBufHeadRegAtInit:ctg,elk */
663 	if (!stop_ring(engine)) {
664 		/* G45 ring initialization often fails to reset head to zero */
665 		DRM_DEBUG_DRIVER("%s head not reset to zero "
666 				"ctl %08x head %08x tail %08x start %08x\n",
667 				engine->name,
668 				ENGINE_READ(engine, RING_CTL),
669 				ENGINE_READ(engine, RING_HEAD),
670 				ENGINE_READ(engine, RING_TAIL),
671 				ENGINE_READ(engine, RING_START));
672 
673 		if (!stop_ring(engine)) {
674 			DRM_ERROR("failed to set %s head to zero "
675 				  "ctl %08x head %08x tail %08x start %08x\n",
676 				  engine->name,
677 				  ENGINE_READ(engine, RING_CTL),
678 				  ENGINE_READ(engine, RING_HEAD),
679 				  ENGINE_READ(engine, RING_TAIL),
680 				  ENGINE_READ(engine, RING_START));
681 			ret = -EIO;
682 			goto out;
683 		}
684 	}
685 
686 	if (HWS_NEEDS_PHYSICAL(dev_priv))
687 		ring_setup_phys_status_page(engine);
688 	else
689 		ring_setup_status_page(engine);
690 
691 	intel_engine_reset_breadcrumbs(engine);
692 
693 	/* Enforce ordering by reading HEAD register back */
694 	ENGINE_POSTING_READ(engine, RING_HEAD);
695 
696 	/*
697 	 * Initialize the ring. This must happen _after_ we've cleared the ring
698 	 * registers with the above sequence (the readback of the HEAD registers
699 	 * also enforces ordering), otherwise the hw might lose the new ring
700 	 * register values.
701 	 */
702 	ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
703 
704 	/* Check that the ring offsets point within the ring! */
705 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
706 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
707 	intel_ring_update_space(ring);
708 
709 	set_pp_dir(engine);
710 
711 	/* First wake the ring up to an empty/idle ring */
712 	ENGINE_WRITE(engine, RING_HEAD, ring->head);
713 	ENGINE_WRITE(engine, RING_TAIL, ring->head);
714 	ENGINE_POSTING_READ(engine, RING_TAIL);
715 
716 	ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
717 
718 	/* If the head is still not zero, the ring is dead */
719 	if (intel_wait_for_register(engine->uncore,
720 				    RING_CTL(engine->mmio_base),
721 				    RING_VALID, RING_VALID,
722 				    50)) {
723 		DRM_ERROR("%s initialization failed "
724 			  "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
725 			  engine->name,
726 			  ENGINE_READ(engine, RING_CTL),
727 			  ENGINE_READ(engine, RING_CTL) & RING_VALID,
728 			  ENGINE_READ(engine, RING_HEAD), ring->head,
729 			  ENGINE_READ(engine, RING_TAIL), ring->tail,
730 			  ENGINE_READ(engine, RING_START),
731 			  i915_ggtt_offset(ring->vma));
732 		ret = -EIO;
733 		goto out;
734 	}
735 
736 	if (INTEL_GEN(dev_priv) > 2)
737 		ENGINE_WRITE(engine,
738 			     RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
739 
740 	/* Now awake, let it get started */
741 	if (ring->tail != ring->head) {
742 		ENGINE_WRITE(engine, RING_TAIL, ring->tail);
743 		ENGINE_POSTING_READ(engine, RING_TAIL);
744 	}
745 
746 	/* Papering over lost _interrupts_ immediately following the restart */
747 	intel_engine_signal_breadcrumbs(engine);
748 out:
749 	intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
750 
751 	return ret;
752 }
753 
754 static void reset_prepare(struct intel_engine_cs *engine)
755 {
756 	struct intel_uncore *uncore = engine->uncore;
757 	const u32 base = engine->mmio_base;
758 
759 	/*
760 	 * We stop engines, otherwise we might get failed reset and a
761 	 * dead gpu (on elk). Also as modern gpu as kbl can suffer
762 	 * from system hang if batchbuffer is progressing when
763 	 * the reset is issued, regardless of READY_TO_RESET ack.
764 	 * Thus assume it is best to stop engines on all gens
765 	 * where we have a gpu reset.
766 	 *
767 	 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
768 	 *
769 	 * WaMediaResetMainRingCleanup:ctg,elk (presumably)
770 	 *
771 	 * FIXME: Wa for more modern gens needs to be validated
772 	 */
773 	ENGINE_TRACE(engine, "\n");
774 
775 	if (intel_engine_stop_cs(engine))
776 		ENGINE_TRACE(engine, "timed out on STOP_RING\n");
777 
778 	intel_uncore_write_fw(uncore,
779 			      RING_HEAD(base),
780 			      intel_uncore_read_fw(uncore, RING_TAIL(base)));
781 	intel_uncore_posting_read_fw(uncore, RING_HEAD(base)); /* paranoia */
782 
783 	intel_uncore_write_fw(uncore, RING_HEAD(base), 0);
784 	intel_uncore_write_fw(uncore, RING_TAIL(base), 0);
785 	intel_uncore_posting_read_fw(uncore, RING_TAIL(base));
786 
787 	/* The ring must be empty before it is disabled */
788 	intel_uncore_write_fw(uncore, RING_CTL(base), 0);
789 
790 	/* Check acts as a post */
791 	if (intel_uncore_read_fw(uncore, RING_HEAD(base)))
792 		ENGINE_TRACE(engine, "ring head [%x] not parked\n",
793 			     intel_uncore_read_fw(uncore, RING_HEAD(base)));
794 }
795 
796 static void reset_rewind(struct intel_engine_cs *engine, bool stalled)
797 {
798 	struct i915_request *pos, *rq;
799 	unsigned long flags;
800 	u32 head;
801 
802 	rq = NULL;
803 	spin_lock_irqsave(&engine->active.lock, flags);
804 	list_for_each_entry(pos, &engine->active.requests, sched.link) {
805 		if (!i915_request_completed(pos)) {
806 			rq = pos;
807 			break;
808 		}
809 	}
810 
811 	/*
812 	 * The guilty request will get skipped on a hung engine.
813 	 *
814 	 * Users of client default contexts do not rely on logical
815 	 * state preserved between batches so it is safe to execute
816 	 * queued requests following the hang. Non default contexts
817 	 * rely on preserved state, so skipping a batch loses the
818 	 * evolution of the state and it needs to be considered corrupted.
819 	 * Executing more queued batches on top of corrupted state is
820 	 * risky. But we take the risk by trying to advance through
821 	 * the queued requests in order to make the client behaviour
822 	 * more predictable around resets, by not throwing away random
823 	 * amount of batches it has prepared for execution. Sophisticated
824 	 * clients can use gem_reset_stats_ioctl and dma fence status
825 	 * (exported via sync_file info ioctl on explicit fences) to observe
826 	 * when it loses the context state and should rebuild accordingly.
827 	 *
828 	 * The context ban, and ultimately the client ban, mechanism are safety
829 	 * valves if client submission ends up resulting in nothing more than
830 	 * subsequent hangs.
831 	 */
832 
833 	if (rq) {
834 		/*
835 		 * Try to restore the logical GPU state to match the
836 		 * continuation of the request queue. If we skip the
837 		 * context/PD restore, then the next request may try to execute
838 		 * assuming that its context is valid and loaded on the GPU and
839 		 * so may try to access invalid memory, prompting repeated GPU
840 		 * hangs.
841 		 *
842 		 * If the request was guilty, we still restore the logical
843 		 * state in case the next request requires it (e.g. the
844 		 * aliasing ppgtt), but skip over the hung batch.
845 		 *
846 		 * If the request was innocent, we try to replay the request
847 		 * with the restored context.
848 		 */
849 		__i915_request_reset(rq, stalled);
850 
851 		GEM_BUG_ON(rq->ring != engine->legacy.ring);
852 		head = rq->head;
853 	} else {
854 		head = engine->legacy.ring->tail;
855 	}
856 	engine->legacy.ring->head = intel_ring_wrap(engine->legacy.ring, head);
857 
858 	spin_unlock_irqrestore(&engine->active.lock, flags);
859 }
860 
861 static void reset_finish(struct intel_engine_cs *engine)
862 {
863 }
864 
865 static int rcs_resume(struct intel_engine_cs *engine)
866 {
867 	struct drm_i915_private *i915 = engine->i915;
868 	struct intel_uncore *uncore = engine->uncore;
869 
870 	/*
871 	 * Disable CONSTANT_BUFFER before it is loaded from the context
872 	 * image. For as it is loaded, it is executed and the stored
873 	 * address may no longer be valid, leading to a GPU hang.
874 	 *
875 	 * This imposes the requirement that userspace reload their
876 	 * CONSTANT_BUFFER on every batch, fortunately a requirement
877 	 * they are already accustomed to from before contexts were
878 	 * enabled.
879 	 */
880 	if (IS_GEN(i915, 4))
881 		intel_uncore_write(uncore, ECOSKPD,
882 			   _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE));
883 
884 	if (IS_GEN_RANGE(i915, 6, 7))
885 		intel_uncore_write(uncore, INSTPM,
886 				   _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
887 
888 	return xcs_resume(engine);
889 }
890 
891 static void reset_cancel(struct intel_engine_cs *engine)
892 {
893 	struct i915_request *request;
894 	unsigned long flags;
895 
896 	spin_lock_irqsave(&engine->active.lock, flags);
897 
898 	/* Mark all submitted requests as skipped. */
899 	list_for_each_entry(request, &engine->active.requests, sched.link) {
900 		if (!i915_request_signaled(request))
901 			dma_fence_set_error(&request->fence, -EIO);
902 
903 		i915_request_mark_complete(request);
904 	}
905 
906 	/* Remaining _unready_ requests will be nop'ed when submitted */
907 
908 	spin_unlock_irqrestore(&engine->active.lock, flags);
909 }
910 
911 static void i9xx_submit_request(struct i915_request *request)
912 {
913 	i915_request_submit(request);
914 	wmb(); /* paranoid flush writes out of the WCB before mmio */
915 
916 	ENGINE_WRITE(request->engine, RING_TAIL,
917 		     intel_ring_set_tail(request->ring, request->tail));
918 }
919 
920 static u32 *i9xx_emit_breadcrumb(struct i915_request *rq, u32 *cs)
921 {
922 	GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
923 	GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
924 
925 	*cs++ = MI_FLUSH;
926 
927 	*cs++ = MI_STORE_DWORD_INDEX;
928 	*cs++ = I915_GEM_HWS_SEQNO_ADDR;
929 	*cs++ = rq->fence.seqno;
930 
931 	*cs++ = MI_USER_INTERRUPT;
932 	*cs++ = MI_NOOP;
933 
934 	rq->tail = intel_ring_offset(rq, cs);
935 	assert_ring_tail_valid(rq->ring, rq->tail);
936 
937 	return cs;
938 }
939 
940 #define GEN5_WA_STORES 8 /* must be at least 1! */
941 static u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs)
942 {
943 	int i;
944 
945 	GEM_BUG_ON(i915_request_active_timeline(rq)->hwsp_ggtt != rq->engine->status_page.vma);
946 	GEM_BUG_ON(offset_in_page(i915_request_active_timeline(rq)->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
947 
948 	*cs++ = MI_FLUSH;
949 
950 	BUILD_BUG_ON(GEN5_WA_STORES < 1);
951 	for (i = 0; i < GEN5_WA_STORES; i++) {
952 		*cs++ = MI_STORE_DWORD_INDEX;
953 		*cs++ = I915_GEM_HWS_SEQNO_ADDR;
954 		*cs++ = rq->fence.seqno;
955 	}
956 
957 	*cs++ = MI_USER_INTERRUPT;
958 
959 	rq->tail = intel_ring_offset(rq, cs);
960 	assert_ring_tail_valid(rq->ring, rq->tail);
961 
962 	return cs;
963 }
964 #undef GEN5_WA_STORES
965 
966 static void
967 gen5_irq_enable(struct intel_engine_cs *engine)
968 {
969 	gen5_gt_enable_irq(engine->gt, engine->irq_enable_mask);
970 }
971 
972 static void
973 gen5_irq_disable(struct intel_engine_cs *engine)
974 {
975 	gen5_gt_disable_irq(engine->gt, engine->irq_enable_mask);
976 }
977 
978 static void
979 i9xx_irq_enable(struct intel_engine_cs *engine)
980 {
981 	engine->i915->irq_mask &= ~engine->irq_enable_mask;
982 	intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
983 	intel_uncore_posting_read_fw(engine->uncore, GEN2_IMR);
984 }
985 
986 static void
987 i9xx_irq_disable(struct intel_engine_cs *engine)
988 {
989 	engine->i915->irq_mask |= engine->irq_enable_mask;
990 	intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
991 }
992 
993 static void
994 i8xx_irq_enable(struct intel_engine_cs *engine)
995 {
996 	struct drm_i915_private *i915 = engine->i915;
997 
998 	i915->irq_mask &= ~engine->irq_enable_mask;
999 	intel_uncore_write16(&i915->uncore, GEN2_IMR, i915->irq_mask);
1000 	ENGINE_POSTING_READ16(engine, RING_IMR);
1001 }
1002 
1003 static void
1004 i8xx_irq_disable(struct intel_engine_cs *engine)
1005 {
1006 	struct drm_i915_private *i915 = engine->i915;
1007 
1008 	i915->irq_mask |= engine->irq_enable_mask;
1009 	intel_uncore_write16(&i915->uncore, GEN2_IMR, i915->irq_mask);
1010 }
1011 
1012 static int
1013 bsd_ring_flush(struct i915_request *rq, u32 mode)
1014 {
1015 	u32 *cs;
1016 
1017 	cs = intel_ring_begin(rq, 2);
1018 	if (IS_ERR(cs))
1019 		return PTR_ERR(cs);
1020 
1021 	*cs++ = MI_FLUSH;
1022 	*cs++ = MI_NOOP;
1023 	intel_ring_advance(rq, cs);
1024 	return 0;
1025 }
1026 
1027 static void
1028 gen6_irq_enable(struct intel_engine_cs *engine)
1029 {
1030 	ENGINE_WRITE(engine, RING_IMR,
1031 		     ~(engine->irq_enable_mask | engine->irq_keep_mask));
1032 
1033 	/* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1034 	ENGINE_POSTING_READ(engine, RING_IMR);
1035 
1036 	gen5_gt_enable_irq(engine->gt, engine->irq_enable_mask);
1037 }
1038 
1039 static void
1040 gen6_irq_disable(struct intel_engine_cs *engine)
1041 {
1042 	ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
1043 	gen5_gt_disable_irq(engine->gt, engine->irq_enable_mask);
1044 }
1045 
1046 static void
1047 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
1048 {
1049 	ENGINE_WRITE(engine, RING_IMR, ~engine->irq_enable_mask);
1050 
1051 	/* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1052 	ENGINE_POSTING_READ(engine, RING_IMR);
1053 
1054 	gen6_gt_pm_unmask_irq(engine->gt, engine->irq_enable_mask);
1055 }
1056 
1057 static void
1058 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
1059 {
1060 	ENGINE_WRITE(engine, RING_IMR, ~0);
1061 	gen6_gt_pm_mask_irq(engine->gt, engine->irq_enable_mask);
1062 }
1063 
1064 static int
1065 i965_emit_bb_start(struct i915_request *rq,
1066 		   u64 offset, u32 length,
1067 		   unsigned int dispatch_flags)
1068 {
1069 	u32 *cs;
1070 
1071 	cs = intel_ring_begin(rq, 2);
1072 	if (IS_ERR(cs))
1073 		return PTR_ERR(cs);
1074 
1075 	*cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | (dispatch_flags &
1076 		I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965);
1077 	*cs++ = offset;
1078 	intel_ring_advance(rq, cs);
1079 
1080 	return 0;
1081 }
1082 
1083 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1084 #define I830_BATCH_LIMIT SZ_256K
1085 #define I830_TLB_ENTRIES (2)
1086 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1087 static int
1088 i830_emit_bb_start(struct i915_request *rq,
1089 		   u64 offset, u32 len,
1090 		   unsigned int dispatch_flags)
1091 {
1092 	u32 *cs, cs_offset =
1093 		intel_gt_scratch_offset(rq->engine->gt,
1094 					INTEL_GT_SCRATCH_FIELD_DEFAULT);
1095 
1096 	GEM_BUG_ON(rq->engine->gt->scratch->size < I830_WA_SIZE);
1097 
1098 	cs = intel_ring_begin(rq, 6);
1099 	if (IS_ERR(cs))
1100 		return PTR_ERR(cs);
1101 
1102 	/* Evict the invalid PTE TLBs */
1103 	*cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
1104 	*cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
1105 	*cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
1106 	*cs++ = cs_offset;
1107 	*cs++ = 0xdeadbeef;
1108 	*cs++ = MI_NOOP;
1109 	intel_ring_advance(rq, cs);
1110 
1111 	if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
1112 		if (len > I830_BATCH_LIMIT)
1113 			return -ENOSPC;
1114 
1115 		cs = intel_ring_begin(rq, 6 + 2);
1116 		if (IS_ERR(cs))
1117 			return PTR_ERR(cs);
1118 
1119 		/* Blit the batch (which has now all relocs applied) to the
1120 		 * stable batch scratch bo area (so that the CS never
1121 		 * stumbles over its tlb invalidation bug) ...
1122 		 */
1123 		*cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA | (6 - 2);
1124 		*cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
1125 		*cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
1126 		*cs++ = cs_offset;
1127 		*cs++ = 4096;
1128 		*cs++ = offset;
1129 
1130 		*cs++ = MI_FLUSH;
1131 		*cs++ = MI_NOOP;
1132 		intel_ring_advance(rq, cs);
1133 
1134 		/* ... and execute it. */
1135 		offset = cs_offset;
1136 	}
1137 
1138 	cs = intel_ring_begin(rq, 2);
1139 	if (IS_ERR(cs))
1140 		return PTR_ERR(cs);
1141 
1142 	*cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1143 	*cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1144 		MI_BATCH_NON_SECURE);
1145 	intel_ring_advance(rq, cs);
1146 
1147 	return 0;
1148 }
1149 
1150 static int
1151 i915_emit_bb_start(struct i915_request *rq,
1152 		   u64 offset, u32 len,
1153 		   unsigned int dispatch_flags)
1154 {
1155 	u32 *cs;
1156 
1157 	cs = intel_ring_begin(rq, 2);
1158 	if (IS_ERR(cs))
1159 		return PTR_ERR(cs);
1160 
1161 	*cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1162 	*cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1163 		MI_BATCH_NON_SECURE);
1164 	intel_ring_advance(rq, cs);
1165 
1166 	return 0;
1167 }
1168 
1169 static void __ring_context_fini(struct intel_context *ce)
1170 {
1171 	i915_vma_put(ce->state);
1172 }
1173 
1174 static void ring_context_destroy(struct kref *ref)
1175 {
1176 	struct intel_context *ce = container_of(ref, typeof(*ce), ref);
1177 
1178 	GEM_BUG_ON(intel_context_is_pinned(ce));
1179 
1180 	if (ce->state)
1181 		__ring_context_fini(ce);
1182 
1183 	intel_context_fini(ce);
1184 	intel_context_free(ce);
1185 }
1186 
1187 static int __context_pin_ppgtt(struct intel_context *ce)
1188 {
1189 	struct i915_address_space *vm;
1190 	int err = 0;
1191 
1192 	vm = vm_alias(ce->vm);
1193 	if (vm)
1194 		err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)));
1195 
1196 	return err;
1197 }
1198 
1199 static void __context_unpin_ppgtt(struct intel_context *ce)
1200 {
1201 	struct i915_address_space *vm;
1202 
1203 	vm = vm_alias(ce->vm);
1204 	if (vm)
1205 		gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
1206 }
1207 
1208 static void ring_context_unpin(struct intel_context *ce)
1209 {
1210 	__context_unpin_ppgtt(ce);
1211 }
1212 
1213 static struct i915_vma *
1214 alloc_context_vma(struct intel_engine_cs *engine)
1215 {
1216 	struct drm_i915_private *i915 = engine->i915;
1217 	struct drm_i915_gem_object *obj;
1218 	struct i915_vma *vma;
1219 	int err;
1220 
1221 	obj = i915_gem_object_create_shmem(i915, engine->context_size);
1222 	if (IS_ERR(obj))
1223 		return ERR_CAST(obj);
1224 
1225 	/*
1226 	 * Try to make the context utilize L3 as well as LLC.
1227 	 *
1228 	 * On VLV we don't have L3 controls in the PTEs so we
1229 	 * shouldn't touch the cache level, especially as that
1230 	 * would make the object snooped which might have a
1231 	 * negative performance impact.
1232 	 *
1233 	 * Snooping is required on non-llc platforms in execlist
1234 	 * mode, but since all GGTT accesses use PAT entry 0 we
1235 	 * get snooping anyway regardless of cache_level.
1236 	 *
1237 	 * This is only applicable for Ivy Bridge devices since
1238 	 * later platforms don't have L3 control bits in the PTE.
1239 	 */
1240 	if (IS_IVYBRIDGE(i915))
1241 		i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
1242 
1243 	if (engine->default_state) {
1244 		void *defaults, *vaddr;
1245 
1246 		vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1247 		if (IS_ERR(vaddr)) {
1248 			err = PTR_ERR(vaddr);
1249 			goto err_obj;
1250 		}
1251 
1252 		defaults = i915_gem_object_pin_map(engine->default_state,
1253 						   I915_MAP_WB);
1254 		if (IS_ERR(defaults)) {
1255 			err = PTR_ERR(defaults);
1256 			goto err_map;
1257 		}
1258 
1259 		memcpy(vaddr, defaults, engine->context_size);
1260 		i915_gem_object_unpin_map(engine->default_state);
1261 
1262 		i915_gem_object_flush_map(obj);
1263 		i915_gem_object_unpin_map(obj);
1264 	}
1265 
1266 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
1267 	if (IS_ERR(vma)) {
1268 		err = PTR_ERR(vma);
1269 		goto err_obj;
1270 	}
1271 
1272 	return vma;
1273 
1274 err_map:
1275 	i915_gem_object_unpin_map(obj);
1276 err_obj:
1277 	i915_gem_object_put(obj);
1278 	return ERR_PTR(err);
1279 }
1280 
1281 static int ring_context_alloc(struct intel_context *ce)
1282 {
1283 	struct intel_engine_cs *engine = ce->engine;
1284 
1285 	/* One ringbuffer to rule them all */
1286 	GEM_BUG_ON(!engine->legacy.ring);
1287 	ce->ring = engine->legacy.ring;
1288 	ce->timeline = intel_timeline_get(engine->legacy.timeline);
1289 
1290 	GEM_BUG_ON(ce->state);
1291 	if (engine->context_size) {
1292 		struct i915_vma *vma;
1293 
1294 		vma = alloc_context_vma(engine);
1295 		if (IS_ERR(vma))
1296 			return PTR_ERR(vma);
1297 
1298 		ce->state = vma;
1299 		if (engine->default_state)
1300 			__set_bit(CONTEXT_VALID_BIT, &ce->flags);
1301 	}
1302 
1303 	return 0;
1304 }
1305 
1306 static int ring_context_pin(struct intel_context *ce)
1307 {
1308 	return __context_pin_ppgtt(ce);
1309 }
1310 
1311 static void ring_context_reset(struct intel_context *ce)
1312 {
1313 	intel_ring_reset(ce->ring, ce->ring->emit);
1314 }
1315 
1316 static const struct intel_context_ops ring_context_ops = {
1317 	.alloc = ring_context_alloc,
1318 
1319 	.pin = ring_context_pin,
1320 	.unpin = ring_context_unpin,
1321 
1322 	.enter = intel_context_enter_engine,
1323 	.exit = intel_context_exit_engine,
1324 
1325 	.reset = ring_context_reset,
1326 	.destroy = ring_context_destroy,
1327 };
1328 
1329 static int load_pd_dir(struct i915_request *rq,
1330 		       const struct i915_ppgtt *ppgtt,
1331 		       u32 valid)
1332 {
1333 	const struct intel_engine_cs * const engine = rq->engine;
1334 	u32 *cs;
1335 
1336 	cs = intel_ring_begin(rq, 12);
1337 	if (IS_ERR(cs))
1338 		return PTR_ERR(cs);
1339 
1340 	*cs++ = MI_LOAD_REGISTER_IMM(1);
1341 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
1342 	*cs++ = valid;
1343 
1344 	*cs++ = MI_LOAD_REGISTER_IMM(1);
1345 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1346 	*cs++ = px_base(ppgtt->pd)->ggtt_offset << 10;
1347 
1348 	/* Stall until the page table load is complete? */
1349 	*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1350 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1351 	*cs++ = intel_gt_scratch_offset(engine->gt,
1352 					INTEL_GT_SCRATCH_FIELD_DEFAULT);
1353 
1354 	*cs++ = MI_LOAD_REGISTER_IMM(1);
1355 	*cs++ = i915_mmio_reg_offset(RING_INSTPM(engine->mmio_base));
1356 	*cs++ = _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE);
1357 
1358 	intel_ring_advance(rq, cs);
1359 
1360 	return rq->engine->emit_flush(rq, EMIT_FLUSH);
1361 }
1362 
1363 static inline int mi_set_context(struct i915_request *rq, u32 flags)
1364 {
1365 	struct drm_i915_private *i915 = rq->i915;
1366 	struct intel_engine_cs *engine = rq->engine;
1367 	enum intel_engine_id id;
1368 	const int num_engines =
1369 		IS_HASWELL(i915) ? RUNTIME_INFO(i915)->num_engines - 1 : 0;
1370 	bool force_restore = false;
1371 	int len;
1372 	u32 *cs;
1373 
1374 	len = 4;
1375 	if (IS_GEN(i915, 7))
1376 		len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
1377 	else if (IS_GEN(i915, 5))
1378 		len += 2;
1379 	if (flags & MI_FORCE_RESTORE) {
1380 		GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
1381 		flags &= ~MI_FORCE_RESTORE;
1382 		force_restore = true;
1383 		len += 2;
1384 	}
1385 
1386 	cs = intel_ring_begin(rq, len);
1387 	if (IS_ERR(cs))
1388 		return PTR_ERR(cs);
1389 
1390 	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
1391 	if (IS_GEN(i915, 7)) {
1392 		*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1393 		if (num_engines) {
1394 			struct intel_engine_cs *signaller;
1395 
1396 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1397 			for_each_engine(signaller, engine->gt, id) {
1398 				if (signaller == engine)
1399 					continue;
1400 
1401 				*cs++ = i915_mmio_reg_offset(
1402 					   RING_PSMI_CTL(signaller->mmio_base));
1403 				*cs++ = _MASKED_BIT_ENABLE(
1404 						GEN6_PSMI_SLEEP_MSG_DISABLE);
1405 			}
1406 		}
1407 	} else if (IS_GEN(i915, 5)) {
1408 		/*
1409 		 * This w/a is only listed for pre-production ilk a/b steppings,
1410 		 * but is also mentioned for programming the powerctx. To be
1411 		 * safe, just apply the workaround; we do not use SyncFlush so
1412 		 * this should never take effect and so be a no-op!
1413 		 */
1414 		*cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
1415 	}
1416 
1417 	if (force_restore) {
1418 		/*
1419 		 * The HW doesn't handle being told to restore the current
1420 		 * context very well. Quite often it likes goes to go off and
1421 		 * sulk, especially when it is meant to be reloading PP_DIR.
1422 		 * A very simple fix to force the reload is to simply switch
1423 		 * away from the current context and back again.
1424 		 *
1425 		 * Note that the kernel_context will contain random state
1426 		 * following the INHIBIT_RESTORE. We accept this since we
1427 		 * never use the kernel_context state; it is merely a
1428 		 * placeholder we use to flush other contexts.
1429 		 */
1430 		*cs++ = MI_SET_CONTEXT;
1431 		*cs++ = i915_ggtt_offset(engine->kernel_context->state) |
1432 			MI_MM_SPACE_GTT |
1433 			MI_RESTORE_INHIBIT;
1434 	}
1435 
1436 	*cs++ = MI_NOOP;
1437 	*cs++ = MI_SET_CONTEXT;
1438 	*cs++ = i915_ggtt_offset(rq->context->state) | flags;
1439 	/*
1440 	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
1441 	 * WaMiSetContext_Hang:snb,ivb,vlv
1442 	 */
1443 	*cs++ = MI_NOOP;
1444 
1445 	if (IS_GEN(i915, 7)) {
1446 		if (num_engines) {
1447 			struct intel_engine_cs *signaller;
1448 			i915_reg_t last_reg = {}; /* keep gcc quiet */
1449 
1450 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1451 			for_each_engine(signaller, engine->gt, id) {
1452 				if (signaller == engine)
1453 					continue;
1454 
1455 				last_reg = RING_PSMI_CTL(signaller->mmio_base);
1456 				*cs++ = i915_mmio_reg_offset(last_reg);
1457 				*cs++ = _MASKED_BIT_DISABLE(
1458 						GEN6_PSMI_SLEEP_MSG_DISABLE);
1459 			}
1460 
1461 			/* Insert a delay before the next switch! */
1462 			*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1463 			*cs++ = i915_mmio_reg_offset(last_reg);
1464 			*cs++ = intel_gt_scratch_offset(engine->gt,
1465 							INTEL_GT_SCRATCH_FIELD_DEFAULT);
1466 			*cs++ = MI_NOOP;
1467 		}
1468 		*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1469 	} else if (IS_GEN(i915, 5)) {
1470 		*cs++ = MI_SUSPEND_FLUSH;
1471 	}
1472 
1473 	intel_ring_advance(rq, cs);
1474 
1475 	return 0;
1476 }
1477 
1478 static int remap_l3_slice(struct i915_request *rq, int slice)
1479 {
1480 	u32 *cs, *remap_info = rq->i915->l3_parity.remap_info[slice];
1481 	int i;
1482 
1483 	if (!remap_info)
1484 		return 0;
1485 
1486 	cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
1487 	if (IS_ERR(cs))
1488 		return PTR_ERR(cs);
1489 
1490 	/*
1491 	 * Note: We do not worry about the concurrent register cacheline hang
1492 	 * here because no other code should access these registers other than
1493 	 * at initialization time.
1494 	 */
1495 	*cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
1496 	for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
1497 		*cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
1498 		*cs++ = remap_info[i];
1499 	}
1500 	*cs++ = MI_NOOP;
1501 	intel_ring_advance(rq, cs);
1502 
1503 	return 0;
1504 }
1505 
1506 static int remap_l3(struct i915_request *rq)
1507 {
1508 	struct i915_gem_context *ctx = i915_request_gem_context(rq);
1509 	int i, err;
1510 
1511 	if (!ctx || !ctx->remap_slice)
1512 		return 0;
1513 
1514 	for (i = 0; i < MAX_L3_SLICES; i++) {
1515 		if (!(ctx->remap_slice & BIT(i)))
1516 			continue;
1517 
1518 		err = remap_l3_slice(rq, i);
1519 		if (err)
1520 			return err;
1521 	}
1522 
1523 	ctx->remap_slice = 0;
1524 	return 0;
1525 }
1526 
1527 static int switch_mm(struct i915_request *rq, struct i915_address_space *vm)
1528 {
1529 	int ret;
1530 
1531 	if (!vm)
1532 		return 0;
1533 
1534 	ret = rq->engine->emit_flush(rq, EMIT_FLUSH);
1535 	if (ret)
1536 		return ret;
1537 
1538 	/*
1539 	 * Not only do we need a full barrier (post-sync write) after
1540 	 * invalidating the TLBs, but we need to wait a little bit
1541 	 * longer. Whether this is merely delaying us, or the
1542 	 * subsequent flush is a key part of serialising with the
1543 	 * post-sync op, this extra pass appears vital before a
1544 	 * mm switch!
1545 	 */
1546 	ret = load_pd_dir(rq, i915_vm_to_ppgtt(vm), PP_DIR_DCLV_2G);
1547 	if (ret)
1548 		return ret;
1549 
1550 	return rq->engine->emit_flush(rq, EMIT_INVALIDATE);
1551 }
1552 
1553 static int switch_context(struct i915_request *rq)
1554 {
1555 	struct intel_context *ce = rq->context;
1556 	int ret;
1557 
1558 	GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
1559 
1560 	ret = switch_mm(rq, vm_alias(ce->vm));
1561 	if (ret)
1562 		return ret;
1563 
1564 	if (ce->state) {
1565 		u32 flags;
1566 
1567 		GEM_BUG_ON(rq->engine->id != RCS0);
1568 
1569 		/* For resource streamer on HSW+ and power context elsewhere */
1570 		BUILD_BUG_ON(HSW_MI_RS_SAVE_STATE_EN != MI_SAVE_EXT_STATE_EN);
1571 		BUILD_BUG_ON(HSW_MI_RS_RESTORE_STATE_EN != MI_RESTORE_EXT_STATE_EN);
1572 
1573 		flags = MI_SAVE_EXT_STATE_EN | MI_MM_SPACE_GTT;
1574 		if (test_bit(CONTEXT_VALID_BIT, &ce->flags))
1575 			flags |= MI_RESTORE_EXT_STATE_EN;
1576 		else
1577 			flags |= MI_RESTORE_INHIBIT;
1578 
1579 		ret = mi_set_context(rq, flags);
1580 		if (ret)
1581 			return ret;
1582 	}
1583 
1584 	ret = remap_l3(rq);
1585 	if (ret)
1586 		return ret;
1587 
1588 	return 0;
1589 }
1590 
1591 static int ring_request_alloc(struct i915_request *request)
1592 {
1593 	int ret;
1594 
1595 	GEM_BUG_ON(!intel_context_is_pinned(request->context));
1596 	GEM_BUG_ON(i915_request_timeline(request)->has_initial_breadcrumb);
1597 
1598 	/*
1599 	 * Flush enough space to reduce the likelihood of waiting after
1600 	 * we start building the request - in which case we will just
1601 	 * have to repeat work.
1602 	 */
1603 	request->reserved_space += LEGACY_REQUEST_SIZE;
1604 
1605 	/* Unconditionally invalidate GPU caches and TLBs. */
1606 	ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
1607 	if (ret)
1608 		return ret;
1609 
1610 	ret = switch_context(request);
1611 	if (ret)
1612 		return ret;
1613 
1614 	request->reserved_space -= LEGACY_REQUEST_SIZE;
1615 	return 0;
1616 }
1617 
1618 static void gen6_bsd_submit_request(struct i915_request *request)
1619 {
1620 	struct intel_uncore *uncore = request->engine->uncore;
1621 
1622 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1623 
1624        /* Every tail move must follow the sequence below */
1625 
1626 	/* Disable notification that the ring is IDLE. The GT
1627 	 * will then assume that it is busy and bring it out of rc6.
1628 	 */
1629 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1630 			      _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1631 
1632 	/* Clear the context id. Here be magic! */
1633 	intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
1634 
1635 	/* Wait for the ring not to be idle, i.e. for it to wake up. */
1636 	if (__intel_wait_for_register_fw(uncore,
1637 					 GEN6_BSD_SLEEP_PSMI_CONTROL,
1638 					 GEN6_BSD_SLEEP_INDICATOR,
1639 					 0,
1640 					 1000, 0, NULL))
1641 		drm_err(&uncore->i915->drm,
1642 			"timed out waiting for the BSD ring to wake up\n");
1643 
1644 	/* Now that the ring is fully powered up, update the tail */
1645 	i9xx_submit_request(request);
1646 
1647 	/* Let the ring send IDLE messages to the GT again,
1648 	 * and so let it sleep to conserve power when idle.
1649 	 */
1650 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1651 			      _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1652 
1653 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1654 }
1655 
1656 static int mi_flush_dw(struct i915_request *rq, u32 flags)
1657 {
1658 	u32 cmd, *cs;
1659 
1660 	cs = intel_ring_begin(rq, 4);
1661 	if (IS_ERR(cs))
1662 		return PTR_ERR(cs);
1663 
1664 	cmd = MI_FLUSH_DW;
1665 
1666 	/*
1667 	 * We always require a command barrier so that subsequent
1668 	 * commands, such as breadcrumb interrupts, are strictly ordered
1669 	 * wrt the contents of the write cache being flushed to memory
1670 	 * (and thus being coherent from the CPU).
1671 	 */
1672 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1673 
1674 	/*
1675 	 * Bspec vol 1c.3 - blitter engine command streamer:
1676 	 * "If ENABLED, all TLBs will be invalidated once the flush
1677 	 * operation is complete. This bit is only valid when the
1678 	 * Post-Sync Operation field is a value of 1h or 3h."
1679 	 */
1680 	cmd |= flags;
1681 
1682 	*cs++ = cmd;
1683 	*cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1684 	*cs++ = 0;
1685 	*cs++ = MI_NOOP;
1686 
1687 	intel_ring_advance(rq, cs);
1688 
1689 	return 0;
1690 }
1691 
1692 static int gen6_flush_dw(struct i915_request *rq, u32 mode, u32 invflags)
1693 {
1694 	return mi_flush_dw(rq, mode & EMIT_INVALIDATE ? invflags : 0);
1695 }
1696 
1697 static int gen6_bsd_ring_flush(struct i915_request *rq, u32 mode)
1698 {
1699 	return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB | MI_INVALIDATE_BSD);
1700 }
1701 
1702 static int
1703 hsw_emit_bb_start(struct i915_request *rq,
1704 		  u64 offset, u32 len,
1705 		  unsigned int dispatch_flags)
1706 {
1707 	u32 *cs;
1708 
1709 	cs = intel_ring_begin(rq, 2);
1710 	if (IS_ERR(cs))
1711 		return PTR_ERR(cs);
1712 
1713 	*cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1714 		0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW);
1715 	/* bit0-7 is the length on GEN6+ */
1716 	*cs++ = offset;
1717 	intel_ring_advance(rq, cs);
1718 
1719 	return 0;
1720 }
1721 
1722 static int
1723 gen6_emit_bb_start(struct i915_request *rq,
1724 		   u64 offset, u32 len,
1725 		   unsigned int dispatch_flags)
1726 {
1727 	u32 *cs;
1728 
1729 	cs = intel_ring_begin(rq, 2);
1730 	if (IS_ERR(cs))
1731 		return PTR_ERR(cs);
1732 
1733 	*cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1734 		0 : MI_BATCH_NON_SECURE_I965);
1735 	/* bit0-7 is the length on GEN6+ */
1736 	*cs++ = offset;
1737 	intel_ring_advance(rq, cs);
1738 
1739 	return 0;
1740 }
1741 
1742 /* Blitter support (SandyBridge+) */
1743 
1744 static int gen6_ring_flush(struct i915_request *rq, u32 mode)
1745 {
1746 	return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB);
1747 }
1748 
1749 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
1750 {
1751 	engine->submit_request = i9xx_submit_request;
1752 
1753 	engine->park = NULL;
1754 	engine->unpark = NULL;
1755 }
1756 
1757 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
1758 {
1759 	i9xx_set_default_submission(engine);
1760 	engine->submit_request = gen6_bsd_submit_request;
1761 }
1762 
1763 static void ring_release(struct intel_engine_cs *engine)
1764 {
1765 	struct drm_i915_private *dev_priv = engine->i915;
1766 
1767 	drm_WARN_ON(&dev_priv->drm, INTEL_GEN(dev_priv) > 2 &&
1768 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1769 
1770 	intel_engine_cleanup_common(engine);
1771 
1772 	intel_ring_unpin(engine->legacy.ring);
1773 	intel_ring_put(engine->legacy.ring);
1774 
1775 	intel_timeline_unpin(engine->legacy.timeline);
1776 	intel_timeline_put(engine->legacy.timeline);
1777 }
1778 
1779 static void setup_irq(struct intel_engine_cs *engine)
1780 {
1781 	struct drm_i915_private *i915 = engine->i915;
1782 
1783 	if (INTEL_GEN(i915) >= 6) {
1784 		engine->irq_enable = gen6_irq_enable;
1785 		engine->irq_disable = gen6_irq_disable;
1786 	} else if (INTEL_GEN(i915) >= 5) {
1787 		engine->irq_enable = gen5_irq_enable;
1788 		engine->irq_disable = gen5_irq_disable;
1789 	} else if (INTEL_GEN(i915) >= 3) {
1790 		engine->irq_enable = i9xx_irq_enable;
1791 		engine->irq_disable = i9xx_irq_disable;
1792 	} else {
1793 		engine->irq_enable = i8xx_irq_enable;
1794 		engine->irq_disable = i8xx_irq_disable;
1795 	}
1796 }
1797 
1798 static void setup_common(struct intel_engine_cs *engine)
1799 {
1800 	struct drm_i915_private *i915 = engine->i915;
1801 
1802 	/* gen8+ are only supported with execlists */
1803 	GEM_BUG_ON(INTEL_GEN(i915) >= 8);
1804 
1805 	setup_irq(engine);
1806 
1807 	engine->resume = xcs_resume;
1808 	engine->reset.prepare = reset_prepare;
1809 	engine->reset.rewind = reset_rewind;
1810 	engine->reset.cancel = reset_cancel;
1811 	engine->reset.finish = reset_finish;
1812 
1813 	engine->cops = &ring_context_ops;
1814 	engine->request_alloc = ring_request_alloc;
1815 
1816 	/*
1817 	 * Using a global execution timeline; the previous final breadcrumb is
1818 	 * equivalent to our next initial bread so we can elide
1819 	 * engine->emit_init_breadcrumb().
1820 	 */
1821 	engine->emit_fini_breadcrumb = i9xx_emit_breadcrumb;
1822 	if (IS_GEN(i915, 5))
1823 		engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
1824 
1825 	engine->set_default_submission = i9xx_set_default_submission;
1826 
1827 	if (INTEL_GEN(i915) >= 6)
1828 		engine->emit_bb_start = gen6_emit_bb_start;
1829 	else if (INTEL_GEN(i915) >= 4)
1830 		engine->emit_bb_start = i965_emit_bb_start;
1831 	else if (IS_I830(i915) || IS_I845G(i915))
1832 		engine->emit_bb_start = i830_emit_bb_start;
1833 	else
1834 		engine->emit_bb_start = i915_emit_bb_start;
1835 }
1836 
1837 static void setup_rcs(struct intel_engine_cs *engine)
1838 {
1839 	struct drm_i915_private *i915 = engine->i915;
1840 
1841 	if (HAS_L3_DPF(i915))
1842 		engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1843 
1844 	engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1845 
1846 	if (INTEL_GEN(i915) >= 7) {
1847 		engine->emit_flush = gen7_render_ring_flush;
1848 		engine->emit_fini_breadcrumb = gen7_rcs_emit_breadcrumb;
1849 	} else if (IS_GEN(i915, 6)) {
1850 		engine->emit_flush = gen6_render_ring_flush;
1851 		engine->emit_fini_breadcrumb = gen6_rcs_emit_breadcrumb;
1852 	} else if (IS_GEN(i915, 5)) {
1853 		engine->emit_flush = gen4_render_ring_flush;
1854 	} else {
1855 		if (INTEL_GEN(i915) < 4)
1856 			engine->emit_flush = gen2_render_ring_flush;
1857 		else
1858 			engine->emit_flush = gen4_render_ring_flush;
1859 		engine->irq_enable_mask = I915_USER_INTERRUPT;
1860 	}
1861 
1862 	if (IS_HASWELL(i915))
1863 		engine->emit_bb_start = hsw_emit_bb_start;
1864 
1865 	engine->resume = rcs_resume;
1866 }
1867 
1868 static void setup_vcs(struct intel_engine_cs *engine)
1869 {
1870 	struct drm_i915_private *i915 = engine->i915;
1871 
1872 	if (INTEL_GEN(i915) >= 6) {
1873 		/* gen6 bsd needs a special wa for tail updates */
1874 		if (IS_GEN(i915, 6))
1875 			engine->set_default_submission = gen6_bsd_set_default_submission;
1876 		engine->emit_flush = gen6_bsd_ring_flush;
1877 		engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1878 
1879 		if (IS_GEN(i915, 6))
1880 			engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
1881 		else
1882 			engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
1883 	} else {
1884 		engine->emit_flush = bsd_ring_flush;
1885 		if (IS_GEN(i915, 5))
1886 			engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1887 		else
1888 			engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1889 	}
1890 }
1891 
1892 static void setup_bcs(struct intel_engine_cs *engine)
1893 {
1894 	struct drm_i915_private *i915 = engine->i915;
1895 
1896 	engine->emit_flush = gen6_ring_flush;
1897 	engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1898 
1899 	if (IS_GEN(i915, 6))
1900 		engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
1901 	else
1902 		engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
1903 }
1904 
1905 static void setup_vecs(struct intel_engine_cs *engine)
1906 {
1907 	struct drm_i915_private *i915 = engine->i915;
1908 
1909 	GEM_BUG_ON(INTEL_GEN(i915) < 7);
1910 
1911 	engine->emit_flush = gen6_ring_flush;
1912 	engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
1913 	engine->irq_enable = hsw_vebox_irq_enable;
1914 	engine->irq_disable = hsw_vebox_irq_disable;
1915 
1916 	engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
1917 }
1918 
1919 int intel_ring_submission_setup(struct intel_engine_cs *engine)
1920 {
1921 	struct intel_timeline *timeline;
1922 	struct intel_ring *ring;
1923 	int err;
1924 
1925 	setup_common(engine);
1926 
1927 	switch (engine->class) {
1928 	case RENDER_CLASS:
1929 		setup_rcs(engine);
1930 		break;
1931 	case VIDEO_DECODE_CLASS:
1932 		setup_vcs(engine);
1933 		break;
1934 	case COPY_ENGINE_CLASS:
1935 		setup_bcs(engine);
1936 		break;
1937 	case VIDEO_ENHANCEMENT_CLASS:
1938 		setup_vecs(engine);
1939 		break;
1940 	default:
1941 		MISSING_CASE(engine->class);
1942 		return -ENODEV;
1943 	}
1944 
1945 	timeline = intel_timeline_create(engine->gt, engine->status_page.vma);
1946 	if (IS_ERR(timeline)) {
1947 		err = PTR_ERR(timeline);
1948 		goto err;
1949 	}
1950 	GEM_BUG_ON(timeline->has_initial_breadcrumb);
1951 
1952 	err = intel_timeline_pin(timeline);
1953 	if (err)
1954 		goto err_timeline;
1955 
1956 	ring = intel_engine_create_ring(engine, SZ_16K);
1957 	if (IS_ERR(ring)) {
1958 		err = PTR_ERR(ring);
1959 		goto err_timeline_unpin;
1960 	}
1961 
1962 	err = intel_ring_pin(ring);
1963 	if (err)
1964 		goto err_ring;
1965 
1966 	GEM_BUG_ON(engine->legacy.ring);
1967 	engine->legacy.ring = ring;
1968 	engine->legacy.timeline = timeline;
1969 
1970 	GEM_BUG_ON(timeline->hwsp_ggtt != engine->status_page.vma);
1971 
1972 	/* Finally, take ownership and responsibility for cleanup! */
1973 	engine->release = ring_release;
1974 
1975 	return 0;
1976 
1977 err_ring:
1978 	intel_ring_put(ring);
1979 err_timeline_unpin:
1980 	intel_timeline_unpin(timeline);
1981 err_timeline:
1982 	intel_timeline_put(timeline);
1983 err:
1984 	intel_engine_cleanup_common(engine);
1985 	return err;
1986 }
1987