xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/scheduler.c (revision 232b0b08)
1 /*
2  * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
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 FROM,
20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21  * SOFTWARE.
22  *
23  * Authors:
24  *    Zhi Wang <zhi.a.wang@intel.com>
25  *
26  * Contributors:
27  *    Ping Gao <ping.a.gao@intel.com>
28  *    Tina Zhang <tina.zhang@intel.com>
29  *    Chanbin Du <changbin.du@intel.com>
30  *    Min He <min.he@intel.com>
31  *    Bing Niu <bing.niu@intel.com>
32  *    Zhenyu Wang <zhenyuw@linux.intel.com>
33  *
34  */
35 
36 #include <linux/kthread.h>
37 
38 #include "i915_drv.h"
39 #include "gvt.h"
40 
41 #define RING_CTX_OFF(x) \
42 	offsetof(struct execlist_ring_context, x)
43 
44 static void set_context_pdp_root_pointer(
45 		struct execlist_ring_context *ring_context,
46 		u32 pdp[8])
47 {
48 	struct execlist_mmio_pair *pdp_pair = &ring_context->pdp3_UDW;
49 	int i;
50 
51 	for (i = 0; i < 8; i++)
52 		pdp_pair[i].val = pdp[7 - i];
53 }
54 
55 static int populate_shadow_context(struct intel_vgpu_workload *workload)
56 {
57 	struct intel_vgpu *vgpu = workload->vgpu;
58 	struct intel_gvt *gvt = vgpu->gvt;
59 	int ring_id = workload->ring_id;
60 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
61 	struct drm_i915_gem_object *ctx_obj =
62 		shadow_ctx->engine[ring_id].state->obj;
63 	struct execlist_ring_context *shadow_ring_context;
64 	struct page *page;
65 	void *dst;
66 	unsigned long context_gpa, context_page_num;
67 	int i;
68 
69 	gvt_dbg_sched("ring id %d workload lrca %x", ring_id,
70 			workload->ctx_desc.lrca);
71 
72 	context_page_num = intel_lr_context_size(
73 			gvt->dev_priv->engine[ring_id]);
74 
75 	context_page_num = context_page_num >> PAGE_SHIFT;
76 
77 	if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
78 		context_page_num = 19;
79 
80 	i = 2;
81 
82 	while (i < context_page_num) {
83 		context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
84 				(u32)((workload->ctx_desc.lrca + i) <<
85 				GTT_PAGE_SHIFT));
86 		if (context_gpa == INTEL_GVT_INVALID_ADDR) {
87 			gvt_vgpu_err("Invalid guest context descriptor\n");
88 			return -EINVAL;
89 		}
90 
91 		page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
92 		dst = kmap(page);
93 		intel_gvt_hypervisor_read_gpa(vgpu, context_gpa, dst,
94 				GTT_PAGE_SIZE);
95 		kunmap(page);
96 		i++;
97 	}
98 
99 	page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
100 	shadow_ring_context = kmap(page);
101 
102 #define COPY_REG(name) \
103 	intel_gvt_hypervisor_read_gpa(vgpu, workload->ring_context_gpa \
104 		+ RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
105 
106 	COPY_REG(ctx_ctrl);
107 	COPY_REG(ctx_timestamp);
108 
109 	if (ring_id == RCS) {
110 		COPY_REG(bb_per_ctx_ptr);
111 		COPY_REG(rcs_indirect_ctx);
112 		COPY_REG(rcs_indirect_ctx_offset);
113 	}
114 #undef COPY_REG
115 
116 	set_context_pdp_root_pointer(shadow_ring_context,
117 				     workload->shadow_mm->shadow_page_table);
118 
119 	intel_gvt_hypervisor_read_gpa(vgpu,
120 			workload->ring_context_gpa +
121 			sizeof(*shadow_ring_context),
122 			(void *)shadow_ring_context +
123 			sizeof(*shadow_ring_context),
124 			GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
125 
126 	kunmap(page);
127 	return 0;
128 }
129 
130 static int shadow_context_status_change(struct notifier_block *nb,
131 		unsigned long action, void *data)
132 {
133 	struct drm_i915_gem_request *req = (struct drm_i915_gem_request *)data;
134 	struct intel_gvt *gvt = container_of(nb, struct intel_gvt,
135 				shadow_ctx_notifier_block[req->engine->id]);
136 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
137 	struct intel_vgpu_workload *workload =
138 		scheduler->current_workload[req->engine->id];
139 
140 	if (unlikely(!workload))
141 		return NOTIFY_OK;
142 
143 	switch (action) {
144 	case INTEL_CONTEXT_SCHEDULE_IN:
145 		intel_gvt_load_render_mmio(workload->vgpu,
146 					   workload->ring_id);
147 		atomic_set(&workload->shadow_ctx_active, 1);
148 		break;
149 	case INTEL_CONTEXT_SCHEDULE_OUT:
150 		intel_gvt_restore_render_mmio(workload->vgpu,
151 					      workload->ring_id);
152 		/* If the status is -EINPROGRESS means this workload
153 		 * doesn't meet any issue during dispatching so when
154 		 * get the SCHEDULE_OUT set the status to be zero for
155 		 * good. If the status is NOT -EINPROGRESS means there
156 		 * is something wrong happened during dispatching and
157 		 * the status should not be set to zero
158 		 */
159 		if (workload->status == -EINPROGRESS)
160 			workload->status = 0;
161 		atomic_set(&workload->shadow_ctx_active, 0);
162 		break;
163 	default:
164 		WARN_ON(1);
165 		return NOTIFY_OK;
166 	}
167 	wake_up(&workload->shadow_ctx_status_wq);
168 	return NOTIFY_OK;
169 }
170 
171 static int dispatch_workload(struct intel_vgpu_workload *workload)
172 {
173 	int ring_id = workload->ring_id;
174 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
175 	struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
176 	struct intel_engine_cs *engine = dev_priv->engine[ring_id];
177 	struct drm_i915_gem_request *rq;
178 	struct intel_vgpu *vgpu = workload->vgpu;
179 	int ret;
180 
181 	gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
182 		ring_id, workload);
183 
184 	shadow_ctx->desc_template &= ~(0x3 << GEN8_CTX_ADDRESSING_MODE_SHIFT);
185 	shadow_ctx->desc_template |= workload->ctx_desc.addressing_mode <<
186 				    GEN8_CTX_ADDRESSING_MODE_SHIFT;
187 
188 	mutex_lock(&dev_priv->drm.struct_mutex);
189 
190 	/* pin shadow context by gvt even the shadow context will be pinned
191 	 * when i915 alloc request. That is because gvt will update the guest
192 	 * context from shadow context when workload is completed, and at that
193 	 * moment, i915 may already unpined the shadow context to make the
194 	 * shadow_ctx pages invalid. So gvt need to pin itself. After update
195 	 * the guest context, gvt can unpin the shadow_ctx safely.
196 	 */
197 	ret = engine->context_pin(engine, shadow_ctx);
198 	if (ret) {
199 		gvt_vgpu_err("fail to pin shadow context\n");
200 		workload->status = ret;
201 		mutex_unlock(&dev_priv->drm.struct_mutex);
202 		return ret;
203 	}
204 
205 	rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
206 	if (IS_ERR(rq)) {
207 		gvt_vgpu_err("fail to allocate gem request\n");
208 		ret = PTR_ERR(rq);
209 		goto out;
210 	}
211 
212 	gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
213 
214 	workload->req = i915_gem_request_get(rq);
215 
216 	ret = intel_gvt_scan_and_shadow_workload(workload);
217 	if (ret)
218 		goto out;
219 
220 	if ((workload->ring_id == RCS) &&
221 	    (workload->wa_ctx.indirect_ctx.size != 0)) {
222 		ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
223 		if (ret)
224 			goto out;
225 	}
226 
227 	ret = populate_shadow_context(workload);
228 	if (ret)
229 		goto out;
230 
231 	if (workload->prepare) {
232 		ret = workload->prepare(workload);
233 		if (ret)
234 			goto out;
235 	}
236 
237 	gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
238 			ring_id, workload->req);
239 
240 	ret = 0;
241 	workload->dispatched = true;
242 out:
243 	if (ret)
244 		workload->status = ret;
245 
246 	if (!IS_ERR_OR_NULL(rq))
247 		i915_add_request_no_flush(rq);
248 	else
249 		engine->context_unpin(engine, shadow_ctx);
250 
251 	mutex_unlock(&dev_priv->drm.struct_mutex);
252 	return ret;
253 }
254 
255 static struct intel_vgpu_workload *pick_next_workload(
256 		struct intel_gvt *gvt, int ring_id)
257 {
258 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
259 	struct intel_vgpu_workload *workload = NULL;
260 
261 	mutex_lock(&gvt->lock);
262 
263 	/*
264 	 * no current vgpu / will be scheduled out / no workload
265 	 * bail out
266 	 */
267 	if (!scheduler->current_vgpu) {
268 		gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
269 		goto out;
270 	}
271 
272 	if (scheduler->need_reschedule) {
273 		gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
274 		goto out;
275 	}
276 
277 	if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
278 		gvt_dbg_sched("ring id %d stop - no available workload\n",
279 				ring_id);
280 		goto out;
281 	}
282 
283 	/*
284 	 * still have current workload, maybe the workload disptacher
285 	 * fail to submit it for some reason, resubmit it.
286 	 */
287 	if (scheduler->current_workload[ring_id]) {
288 		workload = scheduler->current_workload[ring_id];
289 		gvt_dbg_sched("ring id %d still have current workload %p\n",
290 				ring_id, workload);
291 		goto out;
292 	}
293 
294 	/*
295 	 * pick a workload as current workload
296 	 * once current workload is set, schedule policy routines
297 	 * will wait the current workload is finished when trying to
298 	 * schedule out a vgpu.
299 	 */
300 	scheduler->current_workload[ring_id] = container_of(
301 			workload_q_head(scheduler->current_vgpu, ring_id)->next,
302 			struct intel_vgpu_workload, list);
303 
304 	workload = scheduler->current_workload[ring_id];
305 
306 	gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
307 
308 	atomic_inc(&workload->vgpu->running_workload_num);
309 out:
310 	mutex_unlock(&gvt->lock);
311 	return workload;
312 }
313 
314 static void update_guest_context(struct intel_vgpu_workload *workload)
315 {
316 	struct intel_vgpu *vgpu = workload->vgpu;
317 	struct intel_gvt *gvt = vgpu->gvt;
318 	int ring_id = workload->ring_id;
319 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
320 	struct drm_i915_gem_object *ctx_obj =
321 		shadow_ctx->engine[ring_id].state->obj;
322 	struct execlist_ring_context *shadow_ring_context;
323 	struct page *page;
324 	void *src;
325 	unsigned long context_gpa, context_page_num;
326 	int i;
327 
328 	gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
329 			workload->ctx_desc.lrca);
330 
331 	context_page_num = intel_lr_context_size(
332 			gvt->dev_priv->engine[ring_id]);
333 
334 	context_page_num = context_page_num >> PAGE_SHIFT;
335 
336 	if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
337 		context_page_num = 19;
338 
339 	i = 2;
340 
341 	while (i < context_page_num) {
342 		context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
343 				(u32)((workload->ctx_desc.lrca + i) <<
344 					GTT_PAGE_SHIFT));
345 		if (context_gpa == INTEL_GVT_INVALID_ADDR) {
346 			gvt_vgpu_err("invalid guest context descriptor\n");
347 			return;
348 		}
349 
350 		page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
351 		src = kmap(page);
352 		intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
353 				GTT_PAGE_SIZE);
354 		kunmap(page);
355 		i++;
356 	}
357 
358 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
359 		RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
360 
361 	page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
362 	shadow_ring_context = kmap(page);
363 
364 #define COPY_REG(name) \
365 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
366 		RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
367 
368 	COPY_REG(ctx_ctrl);
369 	COPY_REG(ctx_timestamp);
370 
371 #undef COPY_REG
372 
373 	intel_gvt_hypervisor_write_gpa(vgpu,
374 			workload->ring_context_gpa +
375 			sizeof(*shadow_ring_context),
376 			(void *)shadow_ring_context +
377 			sizeof(*shadow_ring_context),
378 			GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
379 
380 	kunmap(page);
381 }
382 
383 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
384 {
385 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
386 	struct intel_vgpu_workload *workload;
387 	struct intel_vgpu *vgpu;
388 	int event;
389 
390 	mutex_lock(&gvt->lock);
391 
392 	workload = scheduler->current_workload[ring_id];
393 	vgpu = workload->vgpu;
394 
395 	/* For the workload w/ request, needs to wait for the context
396 	 * switch to make sure request is completed.
397 	 * For the workload w/o request, directly complete the workload.
398 	 */
399 	if (workload->req) {
400 		struct drm_i915_private *dev_priv =
401 			workload->vgpu->gvt->dev_priv;
402 		struct intel_engine_cs *engine =
403 			dev_priv->engine[workload->ring_id];
404 		wait_event(workload->shadow_ctx_status_wq,
405 			   !atomic_read(&workload->shadow_ctx_active));
406 
407 		i915_gem_request_put(fetch_and_zero(&workload->req));
408 
409 		if (!workload->status && !vgpu->resetting) {
410 			update_guest_context(workload);
411 
412 			for_each_set_bit(event, workload->pending_events,
413 					 INTEL_GVT_EVENT_MAX)
414 				intel_vgpu_trigger_virtual_event(vgpu, event);
415 		}
416 		mutex_lock(&dev_priv->drm.struct_mutex);
417 		/* unpin shadow ctx as the shadow_ctx update is done */
418 		engine->context_unpin(engine, workload->vgpu->shadow_ctx);
419 		mutex_unlock(&dev_priv->drm.struct_mutex);
420 	}
421 
422 	gvt_dbg_sched("ring id %d complete workload %p status %d\n",
423 			ring_id, workload, workload->status);
424 
425 	scheduler->current_workload[ring_id] = NULL;
426 
427 	list_del_init(&workload->list);
428 	workload->complete(workload);
429 
430 	atomic_dec(&vgpu->running_workload_num);
431 	wake_up(&scheduler->workload_complete_wq);
432 	mutex_unlock(&gvt->lock);
433 }
434 
435 struct workload_thread_param {
436 	struct intel_gvt *gvt;
437 	int ring_id;
438 };
439 
440 static DEFINE_MUTEX(scheduler_mutex);
441 
442 static int workload_thread(void *priv)
443 {
444 	struct workload_thread_param *p = (struct workload_thread_param *)priv;
445 	struct intel_gvt *gvt = p->gvt;
446 	int ring_id = p->ring_id;
447 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
448 	struct intel_vgpu_workload *workload = NULL;
449 	struct intel_vgpu *vgpu = NULL;
450 	int ret;
451 	bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
452 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
453 
454 	kfree(p);
455 
456 	gvt_dbg_core("workload thread for ring %d started\n", ring_id);
457 
458 	while (!kthread_should_stop()) {
459 		add_wait_queue(&scheduler->waitq[ring_id], &wait);
460 		do {
461 			workload = pick_next_workload(gvt, ring_id);
462 			if (workload)
463 				break;
464 			wait_woken(&wait, TASK_INTERRUPTIBLE,
465 				   MAX_SCHEDULE_TIMEOUT);
466 		} while (!kthread_should_stop());
467 		remove_wait_queue(&scheduler->waitq[ring_id], &wait);
468 
469 		if (!workload)
470 			break;
471 
472 		mutex_lock(&scheduler_mutex);
473 
474 		gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
475 				workload->ring_id, workload,
476 				workload->vgpu->id);
477 
478 		intel_runtime_pm_get(gvt->dev_priv);
479 
480 		gvt_dbg_sched("ring id %d will dispatch workload %p\n",
481 				workload->ring_id, workload);
482 
483 		if (need_force_wake)
484 			intel_uncore_forcewake_get(gvt->dev_priv,
485 					FORCEWAKE_ALL);
486 
487 		mutex_lock(&gvt->lock);
488 		ret = dispatch_workload(workload);
489 		mutex_unlock(&gvt->lock);
490 
491 		if (ret) {
492 			vgpu = workload->vgpu;
493 			gvt_vgpu_err("fail to dispatch workload, skip\n");
494 			goto complete;
495 		}
496 
497 		gvt_dbg_sched("ring id %d wait workload %p\n",
498 				workload->ring_id, workload);
499 		i915_wait_request(workload->req, 0, MAX_SCHEDULE_TIMEOUT);
500 
501 complete:
502 		gvt_dbg_sched("will complete workload %p, status: %d\n",
503 				workload, workload->status);
504 
505 		complete_current_workload(gvt, ring_id);
506 
507 		if (need_force_wake)
508 			intel_uncore_forcewake_put(gvt->dev_priv,
509 					FORCEWAKE_ALL);
510 
511 		intel_runtime_pm_put(gvt->dev_priv);
512 
513 		mutex_unlock(&scheduler_mutex);
514 
515 	}
516 	return 0;
517 }
518 
519 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
520 {
521 	struct intel_gvt *gvt = vgpu->gvt;
522 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
523 
524 	if (atomic_read(&vgpu->running_workload_num)) {
525 		gvt_dbg_sched("wait vgpu idle\n");
526 
527 		wait_event(scheduler->workload_complete_wq,
528 				!atomic_read(&vgpu->running_workload_num));
529 	}
530 }
531 
532 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
533 {
534 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
535 	struct intel_engine_cs *engine;
536 	enum intel_engine_id i;
537 
538 	gvt_dbg_core("clean workload scheduler\n");
539 
540 	for_each_engine(engine, gvt->dev_priv, i) {
541 		atomic_notifier_chain_unregister(
542 					&engine->context_status_notifier,
543 					&gvt->shadow_ctx_notifier_block[i]);
544 		kthread_stop(scheduler->thread[i]);
545 	}
546 }
547 
548 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
549 {
550 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
551 	struct workload_thread_param *param = NULL;
552 	struct intel_engine_cs *engine;
553 	enum intel_engine_id i;
554 	int ret;
555 
556 	gvt_dbg_core("init workload scheduler\n");
557 
558 	init_waitqueue_head(&scheduler->workload_complete_wq);
559 
560 	for_each_engine(engine, gvt->dev_priv, i) {
561 		init_waitqueue_head(&scheduler->waitq[i]);
562 
563 		param = kzalloc(sizeof(*param), GFP_KERNEL);
564 		if (!param) {
565 			ret = -ENOMEM;
566 			goto err;
567 		}
568 
569 		param->gvt = gvt;
570 		param->ring_id = i;
571 
572 		scheduler->thread[i] = kthread_run(workload_thread, param,
573 			"gvt workload %d", i);
574 		if (IS_ERR(scheduler->thread[i])) {
575 			gvt_err("fail to create workload thread\n");
576 			ret = PTR_ERR(scheduler->thread[i]);
577 			goto err;
578 		}
579 
580 		gvt->shadow_ctx_notifier_block[i].notifier_call =
581 					shadow_context_status_change;
582 		atomic_notifier_chain_register(&engine->context_status_notifier,
583 					&gvt->shadow_ctx_notifier_block[i]);
584 	}
585 	return 0;
586 err:
587 	intel_gvt_clean_workload_scheduler(gvt);
588 	kfree(param);
589 	param = NULL;
590 	return ret;
591 }
592 
593 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
594 {
595 	i915_gem_context_put_unlocked(vgpu->shadow_ctx);
596 }
597 
598 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
599 {
600 	atomic_set(&vgpu->running_workload_num, 0);
601 
602 	vgpu->shadow_ctx = i915_gem_context_create_gvt(
603 			&vgpu->gvt->dev_priv->drm);
604 	if (IS_ERR(vgpu->shadow_ctx))
605 		return PTR_ERR(vgpu->shadow_ctx);
606 
607 	vgpu->shadow_ctx->engine[RCS].initialised = true;
608 
609 	return 0;
610 }
611