xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/scheduler.c (revision 1c2dd16a)
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 inline bool is_gvt_request(struct drm_i915_gem_request *req)
131 {
132 	return i915_gem_context_force_single_submission(req->ctx);
133 }
134 
135 static int shadow_context_status_change(struct notifier_block *nb,
136 		unsigned long action, void *data)
137 {
138 	struct drm_i915_gem_request *req = (struct drm_i915_gem_request *)data;
139 	struct intel_gvt *gvt = container_of(nb, struct intel_gvt,
140 				shadow_ctx_notifier_block[req->engine->id]);
141 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
142 	struct intel_vgpu_workload *workload =
143 		scheduler->current_workload[req->engine->id];
144 
145 	if (!is_gvt_request(req) || unlikely(!workload))
146 		return NOTIFY_OK;
147 
148 	switch (action) {
149 	case INTEL_CONTEXT_SCHEDULE_IN:
150 		intel_gvt_load_render_mmio(workload->vgpu,
151 					   workload->ring_id);
152 		atomic_set(&workload->shadow_ctx_active, 1);
153 		break;
154 	case INTEL_CONTEXT_SCHEDULE_OUT:
155 		intel_gvt_restore_render_mmio(workload->vgpu,
156 					      workload->ring_id);
157 		/* If the status is -EINPROGRESS means this workload
158 		 * doesn't meet any issue during dispatching so when
159 		 * get the SCHEDULE_OUT set the status to be zero for
160 		 * good. If the status is NOT -EINPROGRESS means there
161 		 * is something wrong happened during dispatching and
162 		 * the status should not be set to zero
163 		 */
164 		if (workload->status == -EINPROGRESS)
165 			workload->status = 0;
166 		atomic_set(&workload->shadow_ctx_active, 0);
167 		break;
168 	default:
169 		WARN_ON(1);
170 		return NOTIFY_OK;
171 	}
172 	wake_up(&workload->shadow_ctx_status_wq);
173 	return NOTIFY_OK;
174 }
175 
176 static int dispatch_workload(struct intel_vgpu_workload *workload)
177 {
178 	int ring_id = workload->ring_id;
179 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
180 	struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
181 	struct intel_engine_cs *engine = dev_priv->engine[ring_id];
182 	struct drm_i915_gem_request *rq;
183 	struct intel_vgpu *vgpu = workload->vgpu;
184 	int ret;
185 
186 	gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
187 		ring_id, workload);
188 
189 	shadow_ctx->desc_template &= ~(0x3 << GEN8_CTX_ADDRESSING_MODE_SHIFT);
190 	shadow_ctx->desc_template |= workload->ctx_desc.addressing_mode <<
191 				    GEN8_CTX_ADDRESSING_MODE_SHIFT;
192 
193 	mutex_lock(&dev_priv->drm.struct_mutex);
194 
195 	/* pin shadow context by gvt even the shadow context will be pinned
196 	 * when i915 alloc request. That is because gvt will update the guest
197 	 * context from shadow context when workload is completed, and at that
198 	 * moment, i915 may already unpined the shadow context to make the
199 	 * shadow_ctx pages invalid. So gvt need to pin itself. After update
200 	 * the guest context, gvt can unpin the shadow_ctx safely.
201 	 */
202 	ret = engine->context_pin(engine, shadow_ctx);
203 	if (ret) {
204 		gvt_vgpu_err("fail to pin shadow context\n");
205 		workload->status = ret;
206 		mutex_unlock(&dev_priv->drm.struct_mutex);
207 		return ret;
208 	}
209 
210 	rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
211 	if (IS_ERR(rq)) {
212 		gvt_vgpu_err("fail to allocate gem request\n");
213 		ret = PTR_ERR(rq);
214 		goto out;
215 	}
216 
217 	gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
218 
219 	workload->req = i915_gem_request_get(rq);
220 
221 	ret = intel_gvt_scan_and_shadow_workload(workload);
222 	if (ret)
223 		goto out;
224 
225 	if ((workload->ring_id == RCS) &&
226 	    (workload->wa_ctx.indirect_ctx.size != 0)) {
227 		ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
228 		if (ret)
229 			goto out;
230 	}
231 
232 	ret = populate_shadow_context(workload);
233 	if (ret)
234 		goto out;
235 
236 	if (workload->prepare) {
237 		ret = workload->prepare(workload);
238 		if (ret)
239 			goto out;
240 	}
241 
242 	gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
243 			ring_id, workload->req);
244 
245 	ret = 0;
246 	workload->dispatched = true;
247 out:
248 	if (ret)
249 		workload->status = ret;
250 
251 	if (!IS_ERR_OR_NULL(rq))
252 		i915_add_request_no_flush(rq);
253 	else
254 		engine->context_unpin(engine, shadow_ctx);
255 
256 	mutex_unlock(&dev_priv->drm.struct_mutex);
257 	return ret;
258 }
259 
260 static struct intel_vgpu_workload *pick_next_workload(
261 		struct intel_gvt *gvt, int ring_id)
262 {
263 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
264 	struct intel_vgpu_workload *workload = NULL;
265 
266 	mutex_lock(&gvt->lock);
267 
268 	/*
269 	 * no current vgpu / will be scheduled out / no workload
270 	 * bail out
271 	 */
272 	if (!scheduler->current_vgpu) {
273 		gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
274 		goto out;
275 	}
276 
277 	if (scheduler->need_reschedule) {
278 		gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
279 		goto out;
280 	}
281 
282 	if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
283 		gvt_dbg_sched("ring id %d stop - no available workload\n",
284 				ring_id);
285 		goto out;
286 	}
287 
288 	/*
289 	 * still have current workload, maybe the workload disptacher
290 	 * fail to submit it for some reason, resubmit it.
291 	 */
292 	if (scheduler->current_workload[ring_id]) {
293 		workload = scheduler->current_workload[ring_id];
294 		gvt_dbg_sched("ring id %d still have current workload %p\n",
295 				ring_id, workload);
296 		goto out;
297 	}
298 
299 	/*
300 	 * pick a workload as current workload
301 	 * once current workload is set, schedule policy routines
302 	 * will wait the current workload is finished when trying to
303 	 * schedule out a vgpu.
304 	 */
305 	scheduler->current_workload[ring_id] = container_of(
306 			workload_q_head(scheduler->current_vgpu, ring_id)->next,
307 			struct intel_vgpu_workload, list);
308 
309 	workload = scheduler->current_workload[ring_id];
310 
311 	gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
312 
313 	atomic_inc(&workload->vgpu->running_workload_num);
314 out:
315 	mutex_unlock(&gvt->lock);
316 	return workload;
317 }
318 
319 static void update_guest_context(struct intel_vgpu_workload *workload)
320 {
321 	struct intel_vgpu *vgpu = workload->vgpu;
322 	struct intel_gvt *gvt = vgpu->gvt;
323 	int ring_id = workload->ring_id;
324 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
325 	struct drm_i915_gem_object *ctx_obj =
326 		shadow_ctx->engine[ring_id].state->obj;
327 	struct execlist_ring_context *shadow_ring_context;
328 	struct page *page;
329 	void *src;
330 	unsigned long context_gpa, context_page_num;
331 	int i;
332 
333 	gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
334 			workload->ctx_desc.lrca);
335 
336 	context_page_num = intel_lr_context_size(
337 			gvt->dev_priv->engine[ring_id]);
338 
339 	context_page_num = context_page_num >> PAGE_SHIFT;
340 
341 	if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
342 		context_page_num = 19;
343 
344 	i = 2;
345 
346 	while (i < context_page_num) {
347 		context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
348 				(u32)((workload->ctx_desc.lrca + i) <<
349 					GTT_PAGE_SHIFT));
350 		if (context_gpa == INTEL_GVT_INVALID_ADDR) {
351 			gvt_vgpu_err("invalid guest context descriptor\n");
352 			return;
353 		}
354 
355 		page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
356 		src = kmap(page);
357 		intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
358 				GTT_PAGE_SIZE);
359 		kunmap(page);
360 		i++;
361 	}
362 
363 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
364 		RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
365 
366 	page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
367 	shadow_ring_context = kmap(page);
368 
369 #define COPY_REG(name) \
370 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
371 		RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
372 
373 	COPY_REG(ctx_ctrl);
374 	COPY_REG(ctx_timestamp);
375 
376 #undef COPY_REG
377 
378 	intel_gvt_hypervisor_write_gpa(vgpu,
379 			workload->ring_context_gpa +
380 			sizeof(*shadow_ring_context),
381 			(void *)shadow_ring_context +
382 			sizeof(*shadow_ring_context),
383 			GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
384 
385 	kunmap(page);
386 }
387 
388 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
389 {
390 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
391 	struct intel_vgpu_workload *workload;
392 	struct intel_vgpu *vgpu;
393 	int event;
394 
395 	mutex_lock(&gvt->lock);
396 
397 	workload = scheduler->current_workload[ring_id];
398 	vgpu = workload->vgpu;
399 
400 	/* For the workload w/ request, needs to wait for the context
401 	 * switch to make sure request is completed.
402 	 * For the workload w/o request, directly complete the workload.
403 	 */
404 	if (workload->req) {
405 		struct drm_i915_private *dev_priv =
406 			workload->vgpu->gvt->dev_priv;
407 		struct intel_engine_cs *engine =
408 			dev_priv->engine[workload->ring_id];
409 		wait_event(workload->shadow_ctx_status_wq,
410 			   !atomic_read(&workload->shadow_ctx_active));
411 
412 		i915_gem_request_put(fetch_and_zero(&workload->req));
413 
414 		if (!workload->status && !vgpu->resetting) {
415 			update_guest_context(workload);
416 
417 			for_each_set_bit(event, workload->pending_events,
418 					 INTEL_GVT_EVENT_MAX)
419 				intel_vgpu_trigger_virtual_event(vgpu, event);
420 		}
421 		mutex_lock(&dev_priv->drm.struct_mutex);
422 		/* unpin shadow ctx as the shadow_ctx update is done */
423 		engine->context_unpin(engine, workload->vgpu->shadow_ctx);
424 		mutex_unlock(&dev_priv->drm.struct_mutex);
425 	}
426 
427 	gvt_dbg_sched("ring id %d complete workload %p status %d\n",
428 			ring_id, workload, workload->status);
429 
430 	scheduler->current_workload[ring_id] = NULL;
431 
432 	list_del_init(&workload->list);
433 	workload->complete(workload);
434 
435 	atomic_dec(&vgpu->running_workload_num);
436 	wake_up(&scheduler->workload_complete_wq);
437 	mutex_unlock(&gvt->lock);
438 }
439 
440 struct workload_thread_param {
441 	struct intel_gvt *gvt;
442 	int ring_id;
443 };
444 
445 static DEFINE_MUTEX(scheduler_mutex);
446 
447 static int workload_thread(void *priv)
448 {
449 	struct workload_thread_param *p = (struct workload_thread_param *)priv;
450 	struct intel_gvt *gvt = p->gvt;
451 	int ring_id = p->ring_id;
452 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
453 	struct intel_vgpu_workload *workload = NULL;
454 	struct intel_vgpu *vgpu = NULL;
455 	int ret;
456 	bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
457 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
458 
459 	kfree(p);
460 
461 	gvt_dbg_core("workload thread for ring %d started\n", ring_id);
462 
463 	while (!kthread_should_stop()) {
464 		add_wait_queue(&scheduler->waitq[ring_id], &wait);
465 		do {
466 			workload = pick_next_workload(gvt, ring_id);
467 			if (workload)
468 				break;
469 			wait_woken(&wait, TASK_INTERRUPTIBLE,
470 				   MAX_SCHEDULE_TIMEOUT);
471 		} while (!kthread_should_stop());
472 		remove_wait_queue(&scheduler->waitq[ring_id], &wait);
473 
474 		if (!workload)
475 			break;
476 
477 		mutex_lock(&scheduler_mutex);
478 
479 		gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
480 				workload->ring_id, workload,
481 				workload->vgpu->id);
482 
483 		intel_runtime_pm_get(gvt->dev_priv);
484 
485 		gvt_dbg_sched("ring id %d will dispatch workload %p\n",
486 				workload->ring_id, workload);
487 
488 		if (need_force_wake)
489 			intel_uncore_forcewake_get(gvt->dev_priv,
490 					FORCEWAKE_ALL);
491 
492 		mutex_lock(&gvt->lock);
493 		ret = dispatch_workload(workload);
494 		mutex_unlock(&gvt->lock);
495 
496 		if (ret) {
497 			vgpu = workload->vgpu;
498 			gvt_vgpu_err("fail to dispatch workload, skip\n");
499 			goto complete;
500 		}
501 
502 		gvt_dbg_sched("ring id %d wait workload %p\n",
503 				workload->ring_id, workload);
504 		i915_wait_request(workload->req, 0, MAX_SCHEDULE_TIMEOUT);
505 
506 complete:
507 		gvt_dbg_sched("will complete workload %p, status: %d\n",
508 				workload, workload->status);
509 
510 		complete_current_workload(gvt, ring_id);
511 
512 		if (need_force_wake)
513 			intel_uncore_forcewake_put(gvt->dev_priv,
514 					FORCEWAKE_ALL);
515 
516 		intel_runtime_pm_put(gvt->dev_priv);
517 
518 		mutex_unlock(&scheduler_mutex);
519 
520 	}
521 	return 0;
522 }
523 
524 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
525 {
526 	struct intel_gvt *gvt = vgpu->gvt;
527 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
528 
529 	if (atomic_read(&vgpu->running_workload_num)) {
530 		gvt_dbg_sched("wait vgpu idle\n");
531 
532 		wait_event(scheduler->workload_complete_wq,
533 				!atomic_read(&vgpu->running_workload_num));
534 	}
535 }
536 
537 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
538 {
539 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
540 	struct intel_engine_cs *engine;
541 	enum intel_engine_id i;
542 
543 	gvt_dbg_core("clean workload scheduler\n");
544 
545 	for_each_engine(engine, gvt->dev_priv, i) {
546 		atomic_notifier_chain_unregister(
547 					&engine->context_status_notifier,
548 					&gvt->shadow_ctx_notifier_block[i]);
549 		kthread_stop(scheduler->thread[i]);
550 	}
551 }
552 
553 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
554 {
555 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
556 	struct workload_thread_param *param = NULL;
557 	struct intel_engine_cs *engine;
558 	enum intel_engine_id i;
559 	int ret;
560 
561 	gvt_dbg_core("init workload scheduler\n");
562 
563 	init_waitqueue_head(&scheduler->workload_complete_wq);
564 
565 	for_each_engine(engine, gvt->dev_priv, i) {
566 		init_waitqueue_head(&scheduler->waitq[i]);
567 
568 		param = kzalloc(sizeof(*param), GFP_KERNEL);
569 		if (!param) {
570 			ret = -ENOMEM;
571 			goto err;
572 		}
573 
574 		param->gvt = gvt;
575 		param->ring_id = i;
576 
577 		scheduler->thread[i] = kthread_run(workload_thread, param,
578 			"gvt workload %d", i);
579 		if (IS_ERR(scheduler->thread[i])) {
580 			gvt_err("fail to create workload thread\n");
581 			ret = PTR_ERR(scheduler->thread[i]);
582 			goto err;
583 		}
584 
585 		gvt->shadow_ctx_notifier_block[i].notifier_call =
586 					shadow_context_status_change;
587 		atomic_notifier_chain_register(&engine->context_status_notifier,
588 					&gvt->shadow_ctx_notifier_block[i]);
589 	}
590 	return 0;
591 err:
592 	intel_gvt_clean_workload_scheduler(gvt);
593 	kfree(param);
594 	param = NULL;
595 	return ret;
596 }
597 
598 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
599 {
600 	i915_gem_context_put_unlocked(vgpu->shadow_ctx);
601 }
602 
603 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
604 {
605 	atomic_set(&vgpu->running_workload_num, 0);
606 
607 	vgpu->shadow_ctx = i915_gem_context_create_gvt(
608 			&vgpu->gvt->dev_priv->drm);
609 	if (IS_ERR(vgpu->shadow_ctx))
610 		return PTR_ERR(vgpu->shadow_ctx);
611 
612 	vgpu->shadow_ctx->engine[RCS].initialised = true;
613 
614 	return 0;
615 }
616