1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2016 Intel Corporation
4  */
5 
6 #include <drm/drm_print.h>
7 
8 #include "gem/i915_gem_context.h"
9 
10 #include "i915_drv.h"
11 
12 #include "intel_breadcrumbs.h"
13 #include "intel_context.h"
14 #include "intel_engine.h"
15 #include "intel_engine_pm.h"
16 #include "intel_engine_user.h"
17 #include "intel_execlists_submission.h"
18 #include "intel_gt.h"
19 #include "intel_gt_requests.h"
20 #include "intel_gt_pm.h"
21 #include "intel_lrc_reg.h"
22 #include "intel_reset.h"
23 #include "intel_ring.h"
24 #include "uc/intel_guc_submission.h"
25 
26 /* Haswell does have the CXT_SIZE register however it does not appear to be
27  * valid. Now, docs explain in dwords what is in the context object. The full
28  * size is 70720 bytes, however, the power context and execlist context will
29  * never be saved (power context is stored elsewhere, and execlists don't work
30  * on HSW) - so the final size, including the extra state required for the
31  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
32  */
33 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
34 
35 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
36 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
37 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
38 #define GEN10_LR_CONTEXT_RENDER_SIZE	(18 * PAGE_SIZE)
39 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
40 
41 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
42 
43 #define MAX_MMIO_BASES 3
44 struct engine_info {
45 	unsigned int hw_id;
46 	u8 class;
47 	u8 instance;
48 	/* mmio bases table *must* be sorted in reverse graphics_ver order */
49 	struct engine_mmio_base {
50 		u32 graphics_ver : 8;
51 		u32 base : 24;
52 	} mmio_bases[MAX_MMIO_BASES];
53 };
54 
55 static const struct engine_info intel_engines[] = {
56 	[RCS0] = {
57 		.hw_id = RCS0_HW,
58 		.class = RENDER_CLASS,
59 		.instance = 0,
60 		.mmio_bases = {
61 			{ .graphics_ver = 1, .base = RENDER_RING_BASE }
62 		},
63 	},
64 	[BCS0] = {
65 		.hw_id = BCS0_HW,
66 		.class = COPY_ENGINE_CLASS,
67 		.instance = 0,
68 		.mmio_bases = {
69 			{ .graphics_ver = 6, .base = BLT_RING_BASE }
70 		},
71 	},
72 	[VCS0] = {
73 		.hw_id = VCS0_HW,
74 		.class = VIDEO_DECODE_CLASS,
75 		.instance = 0,
76 		.mmio_bases = {
77 			{ .graphics_ver = 11, .base = GEN11_BSD_RING_BASE },
78 			{ .graphics_ver = 6, .base = GEN6_BSD_RING_BASE },
79 			{ .graphics_ver = 4, .base = BSD_RING_BASE }
80 		},
81 	},
82 	[VCS1] = {
83 		.hw_id = VCS1_HW,
84 		.class = VIDEO_DECODE_CLASS,
85 		.instance = 1,
86 		.mmio_bases = {
87 			{ .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE },
88 			{ .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE }
89 		},
90 	},
91 	[VCS2] = {
92 		.hw_id = VCS2_HW,
93 		.class = VIDEO_DECODE_CLASS,
94 		.instance = 2,
95 		.mmio_bases = {
96 			{ .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE }
97 		},
98 	},
99 	[VCS3] = {
100 		.hw_id = VCS3_HW,
101 		.class = VIDEO_DECODE_CLASS,
102 		.instance = 3,
103 		.mmio_bases = {
104 			{ .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE }
105 		},
106 	},
107 	[VECS0] = {
108 		.hw_id = VECS0_HW,
109 		.class = VIDEO_ENHANCEMENT_CLASS,
110 		.instance = 0,
111 		.mmio_bases = {
112 			{ .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE },
113 			{ .graphics_ver = 7, .base = VEBOX_RING_BASE }
114 		},
115 	},
116 	[VECS1] = {
117 		.hw_id = VECS1_HW,
118 		.class = VIDEO_ENHANCEMENT_CLASS,
119 		.instance = 1,
120 		.mmio_bases = {
121 			{ .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE }
122 		},
123 	},
124 };
125 
126 /**
127  * intel_engine_context_size() - return the size of the context for an engine
128  * @gt: the gt
129  * @class: engine class
130  *
131  * Each engine class may require a different amount of space for a context
132  * image.
133  *
134  * Return: size (in bytes) of an engine class specific context image
135  *
136  * Note: this size includes the HWSP, which is part of the context image
137  * in LRC mode, but does not include the "shared data page" used with
138  * GuC submission. The caller should account for this if using the GuC.
139  */
140 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
141 {
142 	struct intel_uncore *uncore = gt->uncore;
143 	u32 cxt_size;
144 
145 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
146 
147 	switch (class) {
148 	case RENDER_CLASS:
149 		switch (GRAPHICS_VER(gt->i915)) {
150 		default:
151 			MISSING_CASE(GRAPHICS_VER(gt->i915));
152 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
153 		case 12:
154 		case 11:
155 			return GEN11_LR_CONTEXT_RENDER_SIZE;
156 		case 10:
157 			return GEN10_LR_CONTEXT_RENDER_SIZE;
158 		case 9:
159 			return GEN9_LR_CONTEXT_RENDER_SIZE;
160 		case 8:
161 			return GEN8_LR_CONTEXT_RENDER_SIZE;
162 		case 7:
163 			if (IS_HASWELL(gt->i915))
164 				return HSW_CXT_TOTAL_SIZE;
165 
166 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
167 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
168 					PAGE_SIZE);
169 		case 6:
170 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
171 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
172 					PAGE_SIZE);
173 		case 5:
174 		case 4:
175 			/*
176 			 * There is a discrepancy here between the size reported
177 			 * by the register and the size of the context layout
178 			 * in the docs. Both are described as authorative!
179 			 *
180 			 * The discrepancy is on the order of a few cachelines,
181 			 * but the total is under one page (4k), which is our
182 			 * minimum allocation anyway so it should all come
183 			 * out in the wash.
184 			 */
185 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
186 			drm_dbg(&gt->i915->drm,
187 				"graphics_ver = %d CXT_SIZE = %d bytes [0x%08x]\n",
188 				GRAPHICS_VER(gt->i915), cxt_size * 64,
189 				cxt_size - 1);
190 			return round_up(cxt_size * 64, PAGE_SIZE);
191 		case 3:
192 		case 2:
193 		/* For the special day when i810 gets merged. */
194 		case 1:
195 			return 0;
196 		}
197 		break;
198 	default:
199 		MISSING_CASE(class);
200 		fallthrough;
201 	case VIDEO_DECODE_CLASS:
202 	case VIDEO_ENHANCEMENT_CLASS:
203 	case COPY_ENGINE_CLASS:
204 		if (GRAPHICS_VER(gt->i915) < 8)
205 			return 0;
206 		return GEN8_LR_CONTEXT_OTHER_SIZE;
207 	}
208 }
209 
210 static u32 __engine_mmio_base(struct drm_i915_private *i915,
211 			      const struct engine_mmio_base *bases)
212 {
213 	int i;
214 
215 	for (i = 0; i < MAX_MMIO_BASES; i++)
216 		if (GRAPHICS_VER(i915) >= bases[i].graphics_ver)
217 			break;
218 
219 	GEM_BUG_ON(i == MAX_MMIO_BASES);
220 	GEM_BUG_ON(!bases[i].base);
221 
222 	return bases[i].base;
223 }
224 
225 static void __sprint_engine_name(struct intel_engine_cs *engine)
226 {
227 	/*
228 	 * Before we know what the uABI name for this engine will be,
229 	 * we still would like to keep track of this engine in the debug logs.
230 	 * We throw in a ' here as a reminder that this isn't its final name.
231 	 */
232 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
233 			     intel_engine_class_repr(engine->class),
234 			     engine->instance) >= sizeof(engine->name));
235 }
236 
237 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
238 {
239 	/*
240 	 * Though they added more rings on g4x/ilk, they did not add
241 	 * per-engine HWSTAM until gen6.
242 	 */
243 	if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
244 		return;
245 
246 	if (INTEL_GEN(engine->i915) >= 3)
247 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
248 	else
249 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
250 }
251 
252 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
253 {
254 	/* Mask off all writes into the unknown HWSP */
255 	intel_engine_set_hwsp_writemask(engine, ~0u);
256 }
257 
258 static void nop_irq_handler(struct intel_engine_cs *engine, u16 iir)
259 {
260 	GEM_DEBUG_WARN_ON(iir);
261 }
262 
263 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
264 {
265 	const struct engine_info *info = &intel_engines[id];
266 	struct drm_i915_private *i915 = gt->i915;
267 	struct intel_engine_cs *engine;
268 
269 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
270 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
271 
272 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
273 		return -EINVAL;
274 
275 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
276 		return -EINVAL;
277 
278 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
279 		return -EINVAL;
280 
281 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
282 		return -EINVAL;
283 
284 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
285 	if (!engine)
286 		return -ENOMEM;
287 
288 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
289 
290 	engine->id = id;
291 	engine->legacy_idx = INVALID_ENGINE;
292 	engine->mask = BIT(id);
293 	engine->i915 = i915;
294 	engine->gt = gt;
295 	engine->uncore = gt->uncore;
296 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
297 	engine->hw_id = info->hw_id;
298 	engine->guc_id = MAKE_GUC_ID(info->class, info->instance);
299 
300 	engine->irq_handler = nop_irq_handler;
301 
302 	engine->class = info->class;
303 	engine->instance = info->instance;
304 	__sprint_engine_name(engine);
305 
306 	engine->props.heartbeat_interval_ms =
307 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
308 	engine->props.max_busywait_duration_ns =
309 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
310 	engine->props.preempt_timeout_ms =
311 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
312 	engine->props.stop_timeout_ms =
313 		CONFIG_DRM_I915_STOP_TIMEOUT;
314 	engine->props.timeslice_duration_ms =
315 		CONFIG_DRM_I915_TIMESLICE_DURATION;
316 
317 	/* Override to uninterruptible for OpenCL workloads. */
318 	if (INTEL_GEN(i915) == 12 && engine->class == RENDER_CLASS)
319 		engine->props.preempt_timeout_ms = 0;
320 
321 	engine->defaults = engine->props; /* never to change again */
322 
323 	engine->context_size = intel_engine_context_size(gt, engine->class);
324 	if (WARN_ON(engine->context_size > BIT(20)))
325 		engine->context_size = 0;
326 	if (engine->context_size)
327 		DRIVER_CAPS(i915)->has_logical_contexts = true;
328 
329 	/* Nothing to do here, execute in order of dependencies */
330 	engine->schedule = NULL;
331 
332 	ewma__engine_latency_init(&engine->latency);
333 	seqcount_init(&engine->stats.lock);
334 
335 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
336 
337 	/* Scrub mmio state on takeover */
338 	intel_engine_sanitize_mmio(engine);
339 
340 	gt->engine_class[info->class][info->instance] = engine;
341 	gt->engine[id] = engine;
342 
343 	return 0;
344 }
345 
346 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
347 {
348 	struct drm_i915_private *i915 = engine->i915;
349 
350 	if (engine->class == VIDEO_DECODE_CLASS) {
351 		/*
352 		 * HEVC support is present on first engine instance
353 		 * before Gen11 and on all instances afterwards.
354 		 */
355 		if (INTEL_GEN(i915) >= 11 ||
356 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
357 			engine->uabi_capabilities |=
358 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
359 
360 		/*
361 		 * SFC block is present only on even logical engine
362 		 * instances.
363 		 */
364 		if ((INTEL_GEN(i915) >= 11 &&
365 		     (engine->gt->info.vdbox_sfc_access &
366 		      BIT(engine->instance))) ||
367 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
368 			engine->uabi_capabilities |=
369 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
370 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
371 		if (INTEL_GEN(i915) >= 9)
372 			engine->uabi_capabilities |=
373 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
374 	}
375 }
376 
377 static void intel_setup_engine_capabilities(struct intel_gt *gt)
378 {
379 	struct intel_engine_cs *engine;
380 	enum intel_engine_id id;
381 
382 	for_each_engine(engine, gt, id)
383 		__setup_engine_capabilities(engine);
384 }
385 
386 /**
387  * intel_engines_release() - free the resources allocated for Command Streamers
388  * @gt: pointer to struct intel_gt
389  */
390 void intel_engines_release(struct intel_gt *gt)
391 {
392 	struct intel_engine_cs *engine;
393 	enum intel_engine_id id;
394 
395 	/*
396 	 * Before we release the resources held by engine, we must be certain
397 	 * that the HW is no longer accessing them -- having the GPU scribble
398 	 * to or read from a page being used for something else causes no end
399 	 * of fun.
400 	 *
401 	 * The GPU should be reset by this point, but assume the worst just
402 	 * in case we aborted before completely initialising the engines.
403 	 */
404 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
405 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
406 		__intel_gt_reset(gt, ALL_ENGINES);
407 
408 	/* Decouple the backend; but keep the layout for late GPU resets */
409 	for_each_engine(engine, gt, id) {
410 		if (!engine->release)
411 			continue;
412 
413 		intel_wakeref_wait_for_idle(&engine->wakeref);
414 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
415 
416 		engine->release(engine);
417 		engine->release = NULL;
418 
419 		memset(&engine->reset, 0, sizeof(engine->reset));
420 	}
421 }
422 
423 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
424 {
425 	if (!engine->request_pool)
426 		return;
427 
428 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
429 }
430 
431 void intel_engines_free(struct intel_gt *gt)
432 {
433 	struct intel_engine_cs *engine;
434 	enum intel_engine_id id;
435 
436 	/* Free the requests! dma-resv keeps fences around for an eternity */
437 	rcu_barrier();
438 
439 	for_each_engine(engine, gt, id) {
440 		intel_engine_free_request_pool(engine);
441 		kfree(engine);
442 		gt->engine[id] = NULL;
443 	}
444 }
445 
446 /*
447  * Determine which engines are fused off in our particular hardware.
448  * Note that we have a catch-22 situation where we need to be able to access
449  * the blitter forcewake domain to read the engine fuses, but at the same time
450  * we need to know which engines are available on the system to know which
451  * forcewake domains are present. We solve this by intializing the forcewake
452  * domains based on the full engine mask in the platform capabilities before
453  * calling this function and pruning the domains for fused-off engines
454  * afterwards.
455  */
456 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
457 {
458 	struct drm_i915_private *i915 = gt->i915;
459 	struct intel_gt_info *info = &gt->info;
460 	struct intel_uncore *uncore = gt->uncore;
461 	unsigned int logical_vdbox = 0;
462 	unsigned int i;
463 	u32 media_fuse;
464 	u16 vdbox_mask;
465 	u16 vebox_mask;
466 
467 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
468 
469 	if (INTEL_GEN(i915) < 11)
470 		return info->engine_mask;
471 
472 	media_fuse = ~intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
473 
474 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
475 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
476 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
477 
478 	for (i = 0; i < I915_MAX_VCS; i++) {
479 		if (!HAS_ENGINE(gt, _VCS(i))) {
480 			vdbox_mask &= ~BIT(i);
481 			continue;
482 		}
483 
484 		if (!(BIT(i) & vdbox_mask)) {
485 			info->engine_mask &= ~BIT(_VCS(i));
486 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
487 			continue;
488 		}
489 
490 		/*
491 		 * In Gen11, only even numbered logical VDBOXes are
492 		 * hooked up to an SFC (Scaler & Format Converter) unit.
493 		 * In TGL each VDBOX has access to an SFC.
494 		 */
495 		if (INTEL_GEN(i915) >= 12 || logical_vdbox++ % 2 == 0)
496 			gt->info.vdbox_sfc_access |= BIT(i);
497 	}
498 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
499 		vdbox_mask, VDBOX_MASK(gt));
500 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
501 
502 	for (i = 0; i < I915_MAX_VECS; i++) {
503 		if (!HAS_ENGINE(gt, _VECS(i))) {
504 			vebox_mask &= ~BIT(i);
505 			continue;
506 		}
507 
508 		if (!(BIT(i) & vebox_mask)) {
509 			info->engine_mask &= ~BIT(_VECS(i));
510 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
511 		}
512 	}
513 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
514 		vebox_mask, VEBOX_MASK(gt));
515 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
516 
517 	return info->engine_mask;
518 }
519 
520 /**
521  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
522  * @gt: pointer to struct intel_gt
523  *
524  * Return: non-zero if the initialization failed.
525  */
526 int intel_engines_init_mmio(struct intel_gt *gt)
527 {
528 	struct drm_i915_private *i915 = gt->i915;
529 	const unsigned int engine_mask = init_engine_mask(gt);
530 	unsigned int mask = 0;
531 	unsigned int i;
532 	int err;
533 
534 	drm_WARN_ON(&i915->drm, engine_mask == 0);
535 	drm_WARN_ON(&i915->drm, engine_mask &
536 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
537 
538 	if (i915_inject_probe_failure(i915))
539 		return -ENODEV;
540 
541 	for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
542 		if (!HAS_ENGINE(gt, i))
543 			continue;
544 
545 		err = intel_engine_setup(gt, i);
546 		if (err)
547 			goto cleanup;
548 
549 		mask |= BIT(i);
550 	}
551 
552 	/*
553 	 * Catch failures to update intel_engines table when the new engines
554 	 * are added to the driver by a warning and disabling the forgotten
555 	 * engines.
556 	 */
557 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
558 		gt->info.engine_mask = mask;
559 
560 	gt->info.num_engines = hweight32(mask);
561 
562 	intel_gt_check_and_clear_faults(gt);
563 
564 	intel_setup_engine_capabilities(gt);
565 
566 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
567 
568 	return 0;
569 
570 cleanup:
571 	intel_engines_free(gt);
572 	return err;
573 }
574 
575 void intel_engine_init_execlists(struct intel_engine_cs *engine)
576 {
577 	struct intel_engine_execlists * const execlists = &engine->execlists;
578 
579 	execlists->port_mask = 1;
580 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
581 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
582 
583 	memset(execlists->pending, 0, sizeof(execlists->pending));
584 	execlists->active =
585 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
586 
587 	execlists->queue_priority_hint = INT_MIN;
588 	execlists->queue = RB_ROOT_CACHED;
589 }
590 
591 static void cleanup_status_page(struct intel_engine_cs *engine)
592 {
593 	struct i915_vma *vma;
594 
595 	/* Prevent writes into HWSP after returning the page to the system */
596 	intel_engine_set_hwsp_writemask(engine, ~0u);
597 
598 	vma = fetch_and_zero(&engine->status_page.vma);
599 	if (!vma)
600 		return;
601 
602 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
603 		i915_vma_unpin(vma);
604 
605 	i915_gem_object_unpin_map(vma->obj);
606 	i915_gem_object_put(vma->obj);
607 }
608 
609 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
610 				struct i915_gem_ww_ctx *ww,
611 				struct i915_vma *vma)
612 {
613 	unsigned int flags;
614 
615 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
616 		/*
617 		 * On g33, we cannot place HWS above 256MiB, so
618 		 * restrict its pinning to the low mappable arena.
619 		 * Though this restriction is not documented for
620 		 * gen4, gen5, or byt, they also behave similarly
621 		 * and hang if the HWS is placed at the top of the
622 		 * GTT. To generalise, it appears that all !llc
623 		 * platforms have issues with us placing the HWS
624 		 * above the mappable region (even though we never
625 		 * actually map it).
626 		 */
627 		flags = PIN_MAPPABLE;
628 	else
629 		flags = PIN_HIGH;
630 
631 	return i915_ggtt_pin(vma, ww, 0, flags);
632 }
633 
634 static int init_status_page(struct intel_engine_cs *engine)
635 {
636 	struct drm_i915_gem_object *obj;
637 	struct i915_gem_ww_ctx ww;
638 	struct i915_vma *vma;
639 	void *vaddr;
640 	int ret;
641 
642 	INIT_LIST_HEAD(&engine->status_page.timelines);
643 
644 	/*
645 	 * Though the HWS register does support 36bit addresses, historically
646 	 * we have had hangs and corruption reported due to wild writes if
647 	 * the HWS is placed above 4G. We only allow objects to be allocated
648 	 * in GFP_DMA32 for i965, and no earlier physical address users had
649 	 * access to more than 4G.
650 	 */
651 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
652 	if (IS_ERR(obj)) {
653 		drm_err(&engine->i915->drm,
654 			"Failed to allocate status page\n");
655 		return PTR_ERR(obj);
656 	}
657 
658 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
659 
660 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
661 	if (IS_ERR(vma)) {
662 		ret = PTR_ERR(vma);
663 		goto err_put;
664 	}
665 
666 	i915_gem_ww_ctx_init(&ww, true);
667 retry:
668 	ret = i915_gem_object_lock(obj, &ww);
669 	if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915))
670 		ret = pin_ggtt_status_page(engine, &ww, vma);
671 	if (ret)
672 		goto err;
673 
674 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
675 	if (IS_ERR(vaddr)) {
676 		ret = PTR_ERR(vaddr);
677 		goto err_unpin;
678 	}
679 
680 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
681 	engine->status_page.vma = vma;
682 
683 err_unpin:
684 	if (ret)
685 		i915_vma_unpin(vma);
686 err:
687 	if (ret == -EDEADLK) {
688 		ret = i915_gem_ww_ctx_backoff(&ww);
689 		if (!ret)
690 			goto retry;
691 	}
692 	i915_gem_ww_ctx_fini(&ww);
693 err_put:
694 	if (ret)
695 		i915_gem_object_put(obj);
696 	return ret;
697 }
698 
699 static int engine_setup_common(struct intel_engine_cs *engine)
700 {
701 	int err;
702 
703 	init_llist_head(&engine->barrier_tasks);
704 
705 	err = init_status_page(engine);
706 	if (err)
707 		return err;
708 
709 	engine->breadcrumbs = intel_breadcrumbs_create(engine);
710 	if (!engine->breadcrumbs) {
711 		err = -ENOMEM;
712 		goto err_status;
713 	}
714 
715 	err = intel_engine_init_cmd_parser(engine);
716 	if (err)
717 		goto err_cmd_parser;
718 
719 	intel_engine_init_active(engine, ENGINE_PHYSICAL);
720 	intel_engine_init_execlists(engine);
721 	intel_engine_init__pm(engine);
722 	intel_engine_init_retire(engine);
723 
724 	/* Use the whole device by default */
725 	engine->sseu =
726 		intel_sseu_from_device_info(&engine->gt->info.sseu);
727 
728 	intel_engine_init_workarounds(engine);
729 	intel_engine_init_whitelist(engine);
730 	intel_engine_init_ctx_wa(engine);
731 
732 	if (INTEL_GEN(engine->i915) >= 12)
733 		engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
734 
735 	return 0;
736 
737 err_cmd_parser:
738 	intel_breadcrumbs_free(engine->breadcrumbs);
739 err_status:
740 	cleanup_status_page(engine);
741 	return err;
742 }
743 
744 struct measure_breadcrumb {
745 	struct i915_request rq;
746 	struct intel_ring ring;
747 	u32 cs[2048];
748 };
749 
750 static int measure_breadcrumb_dw(struct intel_context *ce)
751 {
752 	struct intel_engine_cs *engine = ce->engine;
753 	struct measure_breadcrumb *frame;
754 	int dw;
755 
756 	GEM_BUG_ON(!engine->gt->scratch);
757 
758 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
759 	if (!frame)
760 		return -ENOMEM;
761 
762 	frame->rq.engine = engine;
763 	frame->rq.context = ce;
764 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
765 	frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno;
766 
767 	frame->ring.vaddr = frame->cs;
768 	frame->ring.size = sizeof(frame->cs);
769 	frame->ring.wrap =
770 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
771 	frame->ring.effective_size = frame->ring.size;
772 	intel_ring_update_space(&frame->ring);
773 	frame->rq.ring = &frame->ring;
774 
775 	mutex_lock(&ce->timeline->mutex);
776 	spin_lock_irq(&engine->active.lock);
777 
778 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
779 
780 	spin_unlock_irq(&engine->active.lock);
781 	mutex_unlock(&ce->timeline->mutex);
782 
783 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
784 
785 	kfree(frame);
786 	return dw;
787 }
788 
789 void
790 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
791 {
792 	INIT_LIST_HEAD(&engine->active.requests);
793 	INIT_LIST_HEAD(&engine->active.hold);
794 
795 	spin_lock_init(&engine->active.lock);
796 	lockdep_set_subclass(&engine->active.lock, subclass);
797 
798 	/*
799 	 * Due to an interesting quirk in lockdep's internal debug tracking,
800 	 * after setting a subclass we must ensure the lock is used. Otherwise,
801 	 * nr_unused_locks is incremented once too often.
802 	 */
803 #ifdef CONFIG_DEBUG_LOCK_ALLOC
804 	local_irq_disable();
805 	lock_map_acquire(&engine->active.lock.dep_map);
806 	lock_map_release(&engine->active.lock.dep_map);
807 	local_irq_enable();
808 #endif
809 }
810 
811 static struct intel_context *
812 create_pinned_context(struct intel_engine_cs *engine,
813 		      unsigned int hwsp,
814 		      struct lock_class_key *key,
815 		      const char *name)
816 {
817 	struct intel_context *ce;
818 	int err;
819 
820 	ce = intel_context_create(engine);
821 	if (IS_ERR(ce))
822 		return ce;
823 
824 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
825 	ce->timeline = page_pack_bits(NULL, hwsp);
826 
827 	err = intel_context_pin(ce); /* perma-pin so it is always available */
828 	if (err) {
829 		intel_context_put(ce);
830 		return ERR_PTR(err);
831 	}
832 
833 	/*
834 	 * Give our perma-pinned kernel timelines a separate lockdep class,
835 	 * so that we can use them from within the normal user timelines
836 	 * should we need to inject GPU operations during their request
837 	 * construction.
838 	 */
839 	lockdep_set_class_and_name(&ce->timeline->mutex, key, name);
840 
841 	return ce;
842 }
843 
844 static void destroy_pinned_context(struct intel_context *ce)
845 {
846 	struct intel_engine_cs *engine = ce->engine;
847 	struct i915_vma *hwsp = engine->status_page.vma;
848 
849 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp);
850 
851 	mutex_lock(&hwsp->vm->mutex);
852 	list_del(&ce->timeline->engine_link);
853 	mutex_unlock(&hwsp->vm->mutex);
854 
855 	intel_context_unpin(ce);
856 	intel_context_put(ce);
857 }
858 
859 static struct intel_context *
860 create_kernel_context(struct intel_engine_cs *engine)
861 {
862 	static struct lock_class_key kernel;
863 
864 	return create_pinned_context(engine, I915_GEM_HWS_SEQNO_ADDR,
865 				     &kernel, "kernel_context");
866 }
867 
868 /**
869  * intel_engines_init_common - initialize cengine state which might require hw access
870  * @engine: Engine to initialize.
871  *
872  * Initializes @engine@ structure members shared between legacy and execlists
873  * submission modes which do require hardware access.
874  *
875  * Typcally done at later stages of submission mode specific engine setup.
876  *
877  * Returns zero on success or an error code on failure.
878  */
879 static int engine_init_common(struct intel_engine_cs *engine)
880 {
881 	struct intel_context *ce;
882 	int ret;
883 
884 	engine->set_default_submission(engine);
885 
886 	/*
887 	 * We may need to do things with the shrinker which
888 	 * require us to immediately switch back to the default
889 	 * context. This can cause a problem as pinning the
890 	 * default context also requires GTT space which may not
891 	 * be available. To avoid this we always pin the default
892 	 * context.
893 	 */
894 	ce = create_kernel_context(engine);
895 	if (IS_ERR(ce))
896 		return PTR_ERR(ce);
897 
898 	ret = measure_breadcrumb_dw(ce);
899 	if (ret < 0)
900 		goto err_context;
901 
902 	engine->emit_fini_breadcrumb_dw = ret;
903 	engine->kernel_context = ce;
904 
905 	return 0;
906 
907 err_context:
908 	destroy_pinned_context(ce);
909 	return ret;
910 }
911 
912 int intel_engines_init(struct intel_gt *gt)
913 {
914 	int (*setup)(struct intel_engine_cs *engine);
915 	struct intel_engine_cs *engine;
916 	enum intel_engine_id id;
917 	int err;
918 
919 	if (intel_uc_uses_guc_submission(&gt->uc)) {
920 		gt->submission_method = INTEL_SUBMISSION_GUC;
921 		setup = intel_guc_submission_setup;
922 	} else if (HAS_EXECLISTS(gt->i915)) {
923 		gt->submission_method = INTEL_SUBMISSION_ELSP;
924 		setup = intel_execlists_submission_setup;
925 	} else {
926 		gt->submission_method = INTEL_SUBMISSION_RING;
927 		setup = intel_ring_submission_setup;
928 	}
929 
930 	for_each_engine(engine, gt, id) {
931 		err = engine_setup_common(engine);
932 		if (err)
933 			return err;
934 
935 		err = setup(engine);
936 		if (err)
937 			return err;
938 
939 		err = engine_init_common(engine);
940 		if (err)
941 			return err;
942 
943 		intel_engine_add_user(engine);
944 	}
945 
946 	return 0;
947 }
948 
949 /**
950  * intel_engines_cleanup_common - cleans up the engine state created by
951  *                                the common initiailizers.
952  * @engine: Engine to cleanup.
953  *
954  * This cleans up everything created by the common helpers.
955  */
956 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
957 {
958 	GEM_BUG_ON(!list_empty(&engine->active.requests));
959 	tasklet_kill(&engine->execlists.tasklet); /* flush the callback */
960 
961 	intel_breadcrumbs_free(engine->breadcrumbs);
962 
963 	intel_engine_fini_retire(engine);
964 	intel_engine_cleanup_cmd_parser(engine);
965 
966 	if (engine->default_state)
967 		fput(engine->default_state);
968 
969 	if (engine->kernel_context)
970 		destroy_pinned_context(engine->kernel_context);
971 
972 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
973 	cleanup_status_page(engine);
974 
975 	intel_wa_list_free(&engine->ctx_wa_list);
976 	intel_wa_list_free(&engine->wa_list);
977 	intel_wa_list_free(&engine->whitelist);
978 }
979 
980 /**
981  * intel_engine_resume - re-initializes the HW state of the engine
982  * @engine: Engine to resume.
983  *
984  * Returns zero on success or an error code on failure.
985  */
986 int intel_engine_resume(struct intel_engine_cs *engine)
987 {
988 	intel_engine_apply_workarounds(engine);
989 	intel_engine_apply_whitelist(engine);
990 
991 	return engine->resume(engine);
992 }
993 
994 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
995 {
996 	struct drm_i915_private *i915 = engine->i915;
997 
998 	u64 acthd;
999 
1000 	if (INTEL_GEN(i915) >= 8)
1001 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
1002 	else if (INTEL_GEN(i915) >= 4)
1003 		acthd = ENGINE_READ(engine, RING_ACTHD);
1004 	else
1005 		acthd = ENGINE_READ(engine, ACTHD);
1006 
1007 	return acthd;
1008 }
1009 
1010 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
1011 {
1012 	u64 bbaddr;
1013 
1014 	if (INTEL_GEN(engine->i915) >= 8)
1015 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
1016 	else
1017 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
1018 
1019 	return bbaddr;
1020 }
1021 
1022 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
1023 {
1024 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
1025 		return 0;
1026 
1027 	/*
1028 	 * If we are doing a normal GPU reset, we can take our time and allow
1029 	 * the engine to quiesce. We've stopped submission to the engine, and
1030 	 * if we wait long enough an innocent context should complete and
1031 	 * leave the engine idle. So they should not be caught unaware by
1032 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
1033 	 */
1034 	return READ_ONCE(engine->props.stop_timeout_ms);
1035 }
1036 
1037 static int __intel_engine_stop_cs(struct intel_engine_cs *engine,
1038 				  int fast_timeout_us,
1039 				  int slow_timeout_ms)
1040 {
1041 	struct intel_uncore *uncore = engine->uncore;
1042 	const i915_reg_t mode = RING_MI_MODE(engine->mmio_base);
1043 	int err;
1044 
1045 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
1046 	err = __intel_wait_for_register_fw(engine->uncore, mode,
1047 					   MODE_IDLE, MODE_IDLE,
1048 					   fast_timeout_us,
1049 					   slow_timeout_ms,
1050 					   NULL);
1051 
1052 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1053 	intel_uncore_posting_read_fw(uncore, mode);
1054 	return err;
1055 }
1056 
1057 int intel_engine_stop_cs(struct intel_engine_cs *engine)
1058 {
1059 	int err = 0;
1060 
1061 	if (INTEL_GEN(engine->i915) < 3)
1062 		return -ENODEV;
1063 
1064 	ENGINE_TRACE(engine, "\n");
1065 	if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) {
1066 		ENGINE_TRACE(engine,
1067 			     "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n",
1068 			     ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR,
1069 			     ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR);
1070 
1071 		/*
1072 		 * Sometimes we observe that the idle flag is not
1073 		 * set even though the ring is empty. So double
1074 		 * check before giving up.
1075 		 */
1076 		if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) !=
1077 		    (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR))
1078 			err = -ETIMEDOUT;
1079 	}
1080 
1081 	return err;
1082 }
1083 
1084 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1085 {
1086 	ENGINE_TRACE(engine, "\n");
1087 
1088 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1089 }
1090 
1091 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
1092 {
1093 	switch (type) {
1094 	case I915_CACHE_NONE: return " uncached";
1095 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
1096 	case I915_CACHE_L3_LLC: return " L3+LLC";
1097 	case I915_CACHE_WT: return " WT";
1098 	default: return "";
1099 	}
1100 }
1101 
1102 static u32
1103 read_subslice_reg(const struct intel_engine_cs *engine,
1104 		  int slice, int subslice, i915_reg_t reg)
1105 {
1106 	struct drm_i915_private *i915 = engine->i915;
1107 	struct intel_uncore *uncore = engine->uncore;
1108 	u32 mcr_mask, mcr_ss, mcr, old_mcr, val;
1109 	enum forcewake_domains fw_domains;
1110 
1111 	if (INTEL_GEN(i915) >= 11) {
1112 		mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1113 		mcr_ss = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1114 	} else {
1115 		mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
1116 		mcr_ss = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
1117 	}
1118 
1119 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
1120 						    FW_REG_READ);
1121 	fw_domains |= intel_uncore_forcewake_for_reg(uncore,
1122 						     GEN8_MCR_SELECTOR,
1123 						     FW_REG_READ | FW_REG_WRITE);
1124 
1125 	spin_lock_irq(&uncore->lock);
1126 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
1127 
1128 	old_mcr = mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
1129 
1130 	mcr &= ~mcr_mask;
1131 	mcr |= mcr_ss;
1132 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1133 
1134 	val = intel_uncore_read_fw(uncore, reg);
1135 
1136 	mcr &= ~mcr_mask;
1137 	mcr |= old_mcr & mcr_mask;
1138 
1139 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1140 
1141 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
1142 	spin_unlock_irq(&uncore->lock);
1143 
1144 	return val;
1145 }
1146 
1147 /* NB: please notice the memset */
1148 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1149 			       struct intel_instdone *instdone)
1150 {
1151 	struct drm_i915_private *i915 = engine->i915;
1152 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1153 	struct intel_uncore *uncore = engine->uncore;
1154 	u32 mmio_base = engine->mmio_base;
1155 	int slice;
1156 	int subslice;
1157 
1158 	memset(instdone, 0, sizeof(*instdone));
1159 
1160 	switch (INTEL_GEN(i915)) {
1161 	default:
1162 		instdone->instdone =
1163 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1164 
1165 		if (engine->id != RCS0)
1166 			break;
1167 
1168 		instdone->slice_common =
1169 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1170 		if (INTEL_GEN(i915) >= 12) {
1171 			instdone->slice_common_extra[0] =
1172 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1173 			instdone->slice_common_extra[1] =
1174 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1175 		}
1176 		for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1177 			instdone->sampler[slice][subslice] =
1178 				read_subslice_reg(engine, slice, subslice,
1179 						  GEN7_SAMPLER_INSTDONE);
1180 			instdone->row[slice][subslice] =
1181 				read_subslice_reg(engine, slice, subslice,
1182 						  GEN7_ROW_INSTDONE);
1183 		}
1184 		break;
1185 	case 7:
1186 		instdone->instdone =
1187 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1188 
1189 		if (engine->id != RCS0)
1190 			break;
1191 
1192 		instdone->slice_common =
1193 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1194 		instdone->sampler[0][0] =
1195 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1196 		instdone->row[0][0] =
1197 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1198 
1199 		break;
1200 	case 6:
1201 	case 5:
1202 	case 4:
1203 		instdone->instdone =
1204 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1205 		if (engine->id == RCS0)
1206 			/* HACK: Using the wrong struct member */
1207 			instdone->slice_common =
1208 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1209 		break;
1210 	case 3:
1211 	case 2:
1212 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1213 		break;
1214 	}
1215 }
1216 
1217 static bool ring_is_idle(struct intel_engine_cs *engine)
1218 {
1219 	bool idle = true;
1220 
1221 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1222 		return true;
1223 
1224 	if (!intel_engine_pm_get_if_awake(engine))
1225 		return true;
1226 
1227 	/* First check that no commands are left in the ring */
1228 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1229 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1230 		idle = false;
1231 
1232 	/* No bit for gen2, so assume the CS parser is idle */
1233 	if (INTEL_GEN(engine->i915) > 2 &&
1234 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1235 		idle = false;
1236 
1237 	intel_engine_pm_put(engine);
1238 
1239 	return idle;
1240 }
1241 
1242 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync)
1243 {
1244 	struct tasklet_struct *t = &engine->execlists.tasklet;
1245 
1246 	if (!t->callback)
1247 		return;
1248 
1249 	local_bh_disable();
1250 	if (tasklet_trylock(t)) {
1251 		/* Must wait for any GPU reset in progress. */
1252 		if (__tasklet_is_enabled(t))
1253 			t->callback(t);
1254 		tasklet_unlock(t);
1255 	}
1256 	local_bh_enable();
1257 
1258 	/* Synchronise and wait for the tasklet on another CPU */
1259 	if (sync)
1260 		tasklet_unlock_wait(t);
1261 }
1262 
1263 /**
1264  * intel_engine_is_idle() - Report if the engine has finished process all work
1265  * @engine: the intel_engine_cs
1266  *
1267  * Return true if there are no requests pending, nothing left to be submitted
1268  * to hardware, and that the engine is idle.
1269  */
1270 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1271 {
1272 	/* More white lies, if wedged, hw state is inconsistent */
1273 	if (intel_gt_is_wedged(engine->gt))
1274 		return true;
1275 
1276 	if (!intel_engine_pm_is_awake(engine))
1277 		return true;
1278 
1279 	/* Waiting to drain ELSP? */
1280 	synchronize_hardirq(to_pci_dev(engine->i915->drm.dev)->irq);
1281 	intel_engine_flush_submission(engine);
1282 
1283 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1284 	if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1285 		return false;
1286 
1287 	/* Ring stopped? */
1288 	return ring_is_idle(engine);
1289 }
1290 
1291 bool intel_engines_are_idle(struct intel_gt *gt)
1292 {
1293 	struct intel_engine_cs *engine;
1294 	enum intel_engine_id id;
1295 
1296 	/*
1297 	 * If the driver is wedged, HW state may be very inconsistent and
1298 	 * report that it is still busy, even though we have stopped using it.
1299 	 */
1300 	if (intel_gt_is_wedged(gt))
1301 		return true;
1302 
1303 	/* Already parked (and passed an idleness test); must still be idle */
1304 	if (!READ_ONCE(gt->awake))
1305 		return true;
1306 
1307 	for_each_engine(engine, gt, id) {
1308 		if (!intel_engine_is_idle(engine))
1309 			return false;
1310 	}
1311 
1312 	return true;
1313 }
1314 
1315 void intel_engines_reset_default_submission(struct intel_gt *gt)
1316 {
1317 	struct intel_engine_cs *engine;
1318 	enum intel_engine_id id;
1319 
1320 	for_each_engine(engine, gt, id) {
1321 		if (engine->sanitize)
1322 			engine->sanitize(engine);
1323 
1324 		engine->set_default_submission(engine);
1325 	}
1326 }
1327 
1328 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1329 {
1330 	switch (INTEL_GEN(engine->i915)) {
1331 	case 2:
1332 		return false; /* uses physical not virtual addresses */
1333 	case 3:
1334 		/* maybe only uses physical not virtual addresses */
1335 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1336 	case 4:
1337 		return !IS_I965G(engine->i915); /* who knows! */
1338 	case 6:
1339 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1340 	default:
1341 		return true;
1342 	}
1343 }
1344 
1345 static struct intel_timeline *get_timeline(struct i915_request *rq)
1346 {
1347 	struct intel_timeline *tl;
1348 
1349 	/*
1350 	 * Even though we are holding the engine->active.lock here, there
1351 	 * is no control over the submission queue per-se and we are
1352 	 * inspecting the active state at a random point in time, with an
1353 	 * unknown queue. Play safe and make sure the timeline remains valid.
1354 	 * (Only being used for pretty printing, one extra kref shouldn't
1355 	 * cause a camel stampede!)
1356 	 */
1357 	rcu_read_lock();
1358 	tl = rcu_dereference(rq->timeline);
1359 	if (!kref_get_unless_zero(&tl->kref))
1360 		tl = NULL;
1361 	rcu_read_unlock();
1362 
1363 	return tl;
1364 }
1365 
1366 static int print_ring(char *buf, int sz, struct i915_request *rq)
1367 {
1368 	int len = 0;
1369 
1370 	if (!i915_request_signaled(rq)) {
1371 		struct intel_timeline *tl = get_timeline(rq);
1372 
1373 		len = scnprintf(buf, sz,
1374 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1375 				i915_ggtt_offset(rq->ring->vma),
1376 				tl ? tl->hwsp_offset : 0,
1377 				hwsp_seqno(rq),
1378 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1379 						      1000 * 1000));
1380 
1381 		if (tl)
1382 			intel_timeline_put(tl);
1383 	}
1384 
1385 	return len;
1386 }
1387 
1388 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1389 {
1390 	const size_t rowsize = 8 * sizeof(u32);
1391 	const void *prev = NULL;
1392 	bool skip = false;
1393 	size_t pos;
1394 
1395 	for (pos = 0; pos < len; pos += rowsize) {
1396 		char line[128];
1397 
1398 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1399 			if (!skip) {
1400 				drm_printf(m, "*\n");
1401 				skip = true;
1402 			}
1403 			continue;
1404 		}
1405 
1406 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1407 						rowsize, sizeof(u32),
1408 						line, sizeof(line),
1409 						false) >= sizeof(line));
1410 		drm_printf(m, "[%04zx] %s\n", pos, line);
1411 
1412 		prev = buf + pos;
1413 		skip = false;
1414 	}
1415 }
1416 
1417 static const char *repr_timer(const struct timer_list *t)
1418 {
1419 	if (!READ_ONCE(t->expires))
1420 		return "inactive";
1421 
1422 	if (timer_pending(t))
1423 		return "active";
1424 
1425 	return "expired";
1426 }
1427 
1428 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1429 					 struct drm_printer *m)
1430 {
1431 	struct drm_i915_private *dev_priv = engine->i915;
1432 	struct intel_engine_execlists * const execlists = &engine->execlists;
1433 	u64 addr;
1434 
1435 	if (engine->id == RENDER_CLASS && IS_GEN_RANGE(dev_priv, 4, 7))
1436 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1437 	if (HAS_EXECLISTS(dev_priv)) {
1438 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1439 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1440 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1441 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1442 	}
1443 	drm_printf(m, "\tRING_START: 0x%08x\n",
1444 		   ENGINE_READ(engine, RING_START));
1445 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1446 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1447 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1448 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1449 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1450 		   ENGINE_READ(engine, RING_CTL),
1451 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1452 	if (INTEL_GEN(engine->i915) > 2) {
1453 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1454 			   ENGINE_READ(engine, RING_MI_MODE),
1455 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1456 	}
1457 
1458 	if (INTEL_GEN(dev_priv) >= 6) {
1459 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1460 			   ENGINE_READ(engine, RING_IMR));
1461 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1462 			   ENGINE_READ(engine, RING_ESR));
1463 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1464 			   ENGINE_READ(engine, RING_EMR));
1465 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1466 			   ENGINE_READ(engine, RING_EIR));
1467 	}
1468 
1469 	addr = intel_engine_get_active_head(engine);
1470 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1471 		   upper_32_bits(addr), lower_32_bits(addr));
1472 	addr = intel_engine_get_last_batch_head(engine);
1473 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1474 		   upper_32_bits(addr), lower_32_bits(addr));
1475 	if (INTEL_GEN(dev_priv) >= 8)
1476 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1477 	else if (INTEL_GEN(dev_priv) >= 4)
1478 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1479 	else
1480 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1481 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1482 		   upper_32_bits(addr), lower_32_bits(addr));
1483 	if (INTEL_GEN(dev_priv) >= 4) {
1484 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1485 			   ENGINE_READ(engine, RING_IPEIR));
1486 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1487 			   ENGINE_READ(engine, RING_IPEHR));
1488 	} else {
1489 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1490 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1491 	}
1492 
1493 	if (intel_engine_uses_guc(engine)) {
1494 		/* nothing to print yet */
1495 	} else if (HAS_EXECLISTS(dev_priv)) {
1496 		struct i915_request * const *port, *rq;
1497 		const u32 *hws =
1498 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1499 		const u8 num_entries = execlists->csb_size;
1500 		unsigned int idx;
1501 		u8 read, write;
1502 
1503 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1504 			   yesno(test_bit(TASKLET_STATE_SCHED,
1505 					  &engine->execlists.tasklet.state)),
1506 			   enableddisabled(!atomic_read(&engine->execlists.tasklet.count)),
1507 			   repr_timer(&engine->execlists.preempt),
1508 			   repr_timer(&engine->execlists.timer));
1509 
1510 		read = execlists->csb_head;
1511 		write = READ_ONCE(*execlists->csb_write);
1512 
1513 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1514 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1515 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1516 			   read, write, num_entries);
1517 
1518 		if (read >= num_entries)
1519 			read = 0;
1520 		if (write >= num_entries)
1521 			write = 0;
1522 		if (read > write)
1523 			write += num_entries;
1524 		while (read < write) {
1525 			idx = ++read % num_entries;
1526 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1527 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1528 		}
1529 
1530 		execlists_active_lock_bh(execlists);
1531 		rcu_read_lock();
1532 		for (port = execlists->active; (rq = *port); port++) {
1533 			char hdr[160];
1534 			int len;
1535 
1536 			len = scnprintf(hdr, sizeof(hdr),
1537 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1538 					(int)(port - execlists->active),
1539 					rq->context->lrc.ccid,
1540 					intel_context_is_closed(rq->context) ? "!" : "",
1541 					intel_context_is_banned(rq->context) ? "*" : "");
1542 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1543 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1544 			i915_request_show(m, rq, hdr, 0);
1545 		}
1546 		for (port = execlists->pending; (rq = *port); port++) {
1547 			char hdr[160];
1548 			int len;
1549 
1550 			len = scnprintf(hdr, sizeof(hdr),
1551 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1552 					(int)(port - execlists->pending),
1553 					rq->context->lrc.ccid,
1554 					intel_context_is_closed(rq->context) ? "!" : "",
1555 					intel_context_is_banned(rq->context) ? "*" : "");
1556 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1557 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1558 			i915_request_show(m, rq, hdr, 0);
1559 		}
1560 		rcu_read_unlock();
1561 		execlists_active_unlock_bh(execlists);
1562 	} else if (INTEL_GEN(dev_priv) > 6) {
1563 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1564 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1565 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1566 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1567 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1568 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1569 	}
1570 }
1571 
1572 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1573 {
1574 	void *ring;
1575 	int size;
1576 
1577 	drm_printf(m,
1578 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1579 		   rq->head, rq->postfix, rq->tail,
1580 		   rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1581 		   rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1582 
1583 	size = rq->tail - rq->head;
1584 	if (rq->tail < rq->head)
1585 		size += rq->ring->size;
1586 
1587 	ring = kmalloc(size, GFP_ATOMIC);
1588 	if (ring) {
1589 		const void *vaddr = rq->ring->vaddr;
1590 		unsigned int head = rq->head;
1591 		unsigned int len = 0;
1592 
1593 		if (rq->tail < head) {
1594 			len = rq->ring->size - head;
1595 			memcpy(ring, vaddr + head, len);
1596 			head = 0;
1597 		}
1598 		memcpy(ring + len, vaddr + head, size - len);
1599 
1600 		hexdump(m, ring, size);
1601 		kfree(ring);
1602 	}
1603 }
1604 
1605 static unsigned long list_count(struct list_head *list)
1606 {
1607 	struct list_head *pos;
1608 	unsigned long count = 0;
1609 
1610 	list_for_each(pos, list)
1611 		count++;
1612 
1613 	return count;
1614 }
1615 
1616 static unsigned long read_ul(void *p, size_t x)
1617 {
1618 	return *(unsigned long *)(p + x);
1619 }
1620 
1621 static void print_properties(struct intel_engine_cs *engine,
1622 			     struct drm_printer *m)
1623 {
1624 	static const struct pmap {
1625 		size_t offset;
1626 		const char *name;
1627 	} props[] = {
1628 #define P(x) { \
1629 	.offset = offsetof(typeof(engine->props), x), \
1630 	.name = #x \
1631 }
1632 		P(heartbeat_interval_ms),
1633 		P(max_busywait_duration_ns),
1634 		P(preempt_timeout_ms),
1635 		P(stop_timeout_ms),
1636 		P(timeslice_duration_ms),
1637 
1638 		{},
1639 #undef P
1640 	};
1641 	const struct pmap *p;
1642 
1643 	drm_printf(m, "\tProperties:\n");
1644 	for (p = props; p->name; p++)
1645 		drm_printf(m, "\t\t%s: %lu [default %lu]\n",
1646 			   p->name,
1647 			   read_ul(&engine->props, p->offset),
1648 			   read_ul(&engine->defaults, p->offset));
1649 }
1650 
1651 void intel_engine_dump(struct intel_engine_cs *engine,
1652 		       struct drm_printer *m,
1653 		       const char *header, ...)
1654 {
1655 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1656 	struct i915_request *rq;
1657 	intel_wakeref_t wakeref;
1658 	unsigned long flags;
1659 	ktime_t dummy;
1660 
1661 	if (header) {
1662 		va_list ap;
1663 
1664 		va_start(ap, header);
1665 		drm_vprintf(m, header, &ap);
1666 		va_end(ap);
1667 	}
1668 
1669 	if (intel_gt_is_wedged(engine->gt))
1670 		drm_printf(m, "*** WEDGED ***\n");
1671 
1672 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1673 	drm_printf(m, "\tBarriers?: %s\n",
1674 		   yesno(!llist_empty(&engine->barrier_tasks)));
1675 	drm_printf(m, "\tLatency: %luus\n",
1676 		   ewma__engine_latency_read(&engine->latency));
1677 	if (intel_engine_supports_stats(engine))
1678 		drm_printf(m, "\tRuntime: %llums\n",
1679 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1680 								  &dummy)));
1681 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1682 		   engine->fw_domain, READ_ONCE(engine->fw_active));
1683 
1684 	rcu_read_lock();
1685 	rq = READ_ONCE(engine->heartbeat.systole);
1686 	if (rq)
1687 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1688 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1689 	rcu_read_unlock();
1690 	drm_printf(m, "\tReset count: %d (global %d)\n",
1691 		   i915_reset_engine_count(error, engine),
1692 		   i915_reset_count(error));
1693 	print_properties(engine, m);
1694 
1695 	drm_printf(m, "\tRequests:\n");
1696 
1697 	spin_lock_irqsave(&engine->active.lock, flags);
1698 	rq = intel_engine_find_active_request(engine);
1699 	if (rq) {
1700 		struct intel_timeline *tl = get_timeline(rq);
1701 
1702 		i915_request_show(m, rq, "\t\tactive ", 0);
1703 
1704 		drm_printf(m, "\t\tring->start:  0x%08x\n",
1705 			   i915_ggtt_offset(rq->ring->vma));
1706 		drm_printf(m, "\t\tring->head:   0x%08x\n",
1707 			   rq->ring->head);
1708 		drm_printf(m, "\t\tring->tail:   0x%08x\n",
1709 			   rq->ring->tail);
1710 		drm_printf(m, "\t\tring->emit:   0x%08x\n",
1711 			   rq->ring->emit);
1712 		drm_printf(m, "\t\tring->space:  0x%08x\n",
1713 			   rq->ring->space);
1714 
1715 		if (tl) {
1716 			drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1717 				   tl->hwsp_offset);
1718 			intel_timeline_put(tl);
1719 		}
1720 
1721 		print_request_ring(m, rq);
1722 
1723 		if (rq->context->lrc_reg_state) {
1724 			drm_printf(m, "Logical Ring Context:\n");
1725 			hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1726 		}
1727 	}
1728 	drm_printf(m, "\tOn hold?: %lu\n", list_count(&engine->active.hold));
1729 	spin_unlock_irqrestore(&engine->active.lock, flags);
1730 
1731 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1732 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1733 	if (wakeref) {
1734 		intel_engine_print_registers(engine, m);
1735 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1736 	} else {
1737 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1738 	}
1739 
1740 	intel_execlists_show_requests(engine, m, i915_request_show, 8);
1741 
1742 	drm_printf(m, "HWSP:\n");
1743 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1744 
1745 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1746 
1747 	intel_engine_print_breadcrumbs(engine, m);
1748 }
1749 
1750 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine,
1751 					    ktime_t *now)
1752 {
1753 	ktime_t total = engine->stats.total;
1754 
1755 	/*
1756 	 * If the engine is executing something at the moment
1757 	 * add it to the total.
1758 	 */
1759 	*now = ktime_get();
1760 	if (READ_ONCE(engine->stats.active))
1761 		total = ktime_add(total, ktime_sub(*now, engine->stats.start));
1762 
1763 	return total;
1764 }
1765 
1766 /**
1767  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1768  * @engine: engine to report on
1769  * @now: monotonic timestamp of sampling
1770  *
1771  * Returns accumulated time @engine was busy since engine stats were enabled.
1772  */
1773 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
1774 {
1775 	unsigned int seq;
1776 	ktime_t total;
1777 
1778 	do {
1779 		seq = read_seqcount_begin(&engine->stats.lock);
1780 		total = __intel_engine_get_busy_time(engine, now);
1781 	} while (read_seqcount_retry(&engine->stats.lock, seq));
1782 
1783 	return total;
1784 }
1785 
1786 static bool match_ring(struct i915_request *rq)
1787 {
1788 	u32 ring = ENGINE_READ(rq->engine, RING_START);
1789 
1790 	return ring == i915_ggtt_offset(rq->ring->vma);
1791 }
1792 
1793 struct i915_request *
1794 intel_engine_find_active_request(struct intel_engine_cs *engine)
1795 {
1796 	struct i915_request *request, *active = NULL;
1797 
1798 	/*
1799 	 * We are called by the error capture, reset and to dump engine
1800 	 * state at random points in time. In particular, note that neither is
1801 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1802 	 * and we assume that no more writes can happen (we waited long enough
1803 	 * for all writes that were in transaction to be flushed) - adding an
1804 	 * extra delay for a recent interrupt is pointless. Hence, we do
1805 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
1806 	 * At all other times, we must assume the GPU is still running, but
1807 	 * we only care about the snapshot of this moment.
1808 	 */
1809 	lockdep_assert_held(&engine->active.lock);
1810 
1811 	rcu_read_lock();
1812 	request = execlists_active(&engine->execlists);
1813 	if (request) {
1814 		struct intel_timeline *tl = request->context->timeline;
1815 
1816 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
1817 			if (__i915_request_is_complete(request))
1818 				break;
1819 
1820 			active = request;
1821 		}
1822 	}
1823 	rcu_read_unlock();
1824 	if (active)
1825 		return active;
1826 
1827 	list_for_each_entry(request, &engine->active.requests, sched.link) {
1828 		if (__i915_request_is_complete(request))
1829 			continue;
1830 
1831 		if (!__i915_request_has_started(request))
1832 			continue;
1833 
1834 		/* More than one preemptible request may match! */
1835 		if (!match_ring(request))
1836 			continue;
1837 
1838 		active = request;
1839 		break;
1840 	}
1841 
1842 	return active;
1843 }
1844 
1845 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1846 #include "mock_engine.c"
1847 #include "selftest_engine.c"
1848 #include "selftest_engine_cs.c"
1849 #endif
1850