xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/scheduler.c (revision 2891f2d5)
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_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 intel_vgpu *vgpu = container_of(nb,
134 			struct intel_vgpu, shadow_ctx_notifier_block);
135 	struct drm_i915_gem_request *req =
136 		(struct drm_i915_gem_request *)data;
137 	struct intel_gvt_workload_scheduler *scheduler =
138 		&vgpu->gvt->scheduler;
139 	struct intel_vgpu_workload *workload =
140 		scheduler->current_workload[req->engine->id];
141 
142 	if (unlikely(!workload))
143 		return NOTIFY_OK;
144 
145 	switch (action) {
146 	case INTEL_CONTEXT_SCHEDULE_IN:
147 		intel_gvt_load_render_mmio(workload->vgpu,
148 					   workload->ring_id);
149 		atomic_set(&workload->shadow_ctx_active, 1);
150 		break;
151 	case INTEL_CONTEXT_SCHEDULE_OUT:
152 		intel_gvt_restore_render_mmio(workload->vgpu,
153 					      workload->ring_id);
154 		/* If the status is -EINPROGRESS means this workload
155 		 * doesn't meet any issue during dispatching so when
156 		 * get the SCHEDULE_OUT set the status to be zero for
157 		 * good. If the status is NOT -EINPROGRESS means there
158 		 * is something wrong happened during dispatching and
159 		 * the status should not be set to zero
160 		 */
161 		if (workload->status == -EINPROGRESS)
162 			workload->status = 0;
163 		atomic_set(&workload->shadow_ctx_active, 0);
164 		break;
165 	default:
166 		WARN_ON(1);
167 		return NOTIFY_OK;
168 	}
169 	wake_up(&workload->shadow_ctx_status_wq);
170 	return NOTIFY_OK;
171 }
172 
173 static int dispatch_workload(struct intel_vgpu_workload *workload)
174 {
175 	int ring_id = workload->ring_id;
176 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
177 	struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
178 	struct drm_i915_gem_request *rq;
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 	rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
191 	if (IS_ERR(rq)) {
192 		gvt_err("fail to allocate gem request\n");
193 		ret = PTR_ERR(rq);
194 		goto out;
195 	}
196 
197 	gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
198 
199 	workload->req = i915_gem_request_get(rq);
200 
201 	ret = intel_gvt_scan_and_shadow_workload(workload);
202 	if (ret)
203 		goto out;
204 
205 	ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
206 	if (ret)
207 		goto out;
208 
209 	ret = populate_shadow_context(workload);
210 	if (ret)
211 		goto out;
212 
213 	if (workload->prepare) {
214 		ret = workload->prepare(workload);
215 		if (ret)
216 			goto out;
217 	}
218 
219 	gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
220 			ring_id, workload->req);
221 
222 	ret = 0;
223 	workload->dispatched = true;
224 out:
225 	if (ret)
226 		workload->status = ret;
227 
228 	if (!IS_ERR_OR_NULL(rq))
229 		i915_add_request_no_flush(rq);
230 	mutex_unlock(&dev_priv->drm.struct_mutex);
231 	return ret;
232 }
233 
234 static struct intel_vgpu_workload *pick_next_workload(
235 		struct intel_gvt *gvt, int ring_id)
236 {
237 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
238 	struct intel_vgpu_workload *workload = NULL;
239 
240 	mutex_lock(&gvt->lock);
241 
242 	/*
243 	 * no current vgpu / will be scheduled out / no workload
244 	 * bail out
245 	 */
246 	if (!scheduler->current_vgpu) {
247 		gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
248 		goto out;
249 	}
250 
251 	if (scheduler->need_reschedule) {
252 		gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
253 		goto out;
254 	}
255 
256 	if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
257 		gvt_dbg_sched("ring id %d stop - no available workload\n",
258 				ring_id);
259 		goto out;
260 	}
261 
262 	/*
263 	 * still have current workload, maybe the workload disptacher
264 	 * fail to submit it for some reason, resubmit it.
265 	 */
266 	if (scheduler->current_workload[ring_id]) {
267 		workload = scheduler->current_workload[ring_id];
268 		gvt_dbg_sched("ring id %d still have current workload %p\n",
269 				ring_id, workload);
270 		goto out;
271 	}
272 
273 	/*
274 	 * pick a workload as current workload
275 	 * once current workload is set, schedule policy routines
276 	 * will wait the current workload is finished when trying to
277 	 * schedule out a vgpu.
278 	 */
279 	scheduler->current_workload[ring_id] = container_of(
280 			workload_q_head(scheduler->current_vgpu, ring_id)->next,
281 			struct intel_vgpu_workload, list);
282 
283 	workload = scheduler->current_workload[ring_id];
284 
285 	gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
286 
287 	atomic_inc(&workload->vgpu->running_workload_num);
288 out:
289 	mutex_unlock(&gvt->lock);
290 	return workload;
291 }
292 
293 static void update_guest_context(struct intel_vgpu_workload *workload)
294 {
295 	struct intel_vgpu *vgpu = workload->vgpu;
296 	struct intel_gvt *gvt = vgpu->gvt;
297 	int ring_id = workload->ring_id;
298 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
299 	struct drm_i915_gem_object *ctx_obj =
300 		shadow_ctx->engine[ring_id].state->obj;
301 	struct execlist_ring_context *shadow_ring_context;
302 	struct page *page;
303 	void *src;
304 	unsigned long context_gpa, context_page_num;
305 	int i;
306 
307 	gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
308 			workload->ctx_desc.lrca);
309 
310 	context_page_num = intel_lr_context_size(
311 			gvt->dev_priv->engine[ring_id]);
312 
313 	context_page_num = context_page_num >> PAGE_SHIFT;
314 
315 	if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
316 		context_page_num = 19;
317 
318 	i = 2;
319 
320 	while (i < context_page_num) {
321 		context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
322 				(u32)((workload->ctx_desc.lrca + i) <<
323 					GTT_PAGE_SHIFT));
324 		if (context_gpa == INTEL_GVT_INVALID_ADDR) {
325 			gvt_err("invalid guest context descriptor\n");
326 			return;
327 		}
328 
329 		page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
330 		src = kmap(page);
331 		intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
332 				GTT_PAGE_SIZE);
333 		kunmap(page);
334 		i++;
335 	}
336 
337 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
338 		RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
339 
340 	page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
341 	shadow_ring_context = kmap(page);
342 
343 #define COPY_REG(name) \
344 	intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
345 		RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
346 
347 	COPY_REG(ctx_ctrl);
348 	COPY_REG(ctx_timestamp);
349 
350 #undef COPY_REG
351 
352 	intel_gvt_hypervisor_write_gpa(vgpu,
353 			workload->ring_context_gpa +
354 			sizeof(*shadow_ring_context),
355 			(void *)shadow_ring_context +
356 			sizeof(*shadow_ring_context),
357 			GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
358 
359 	kunmap(page);
360 }
361 
362 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
363 {
364 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
365 	struct intel_vgpu_workload *workload;
366 	struct intel_vgpu *vgpu;
367 	int event;
368 
369 	mutex_lock(&gvt->lock);
370 
371 	workload = scheduler->current_workload[ring_id];
372 	vgpu = workload->vgpu;
373 
374 	/* For the workload w/ request, needs to wait for the context
375 	 * switch to make sure request is completed.
376 	 * For the workload w/o request, directly complete the workload.
377 	 */
378 	if (workload->req) {
379 		wait_event(workload->shadow_ctx_status_wq,
380 			   !atomic_read(&workload->shadow_ctx_active));
381 
382 		i915_gem_request_put(fetch_and_zero(&workload->req));
383 
384 		if (!workload->status && !vgpu->resetting) {
385 			update_guest_context(workload);
386 
387 			for_each_set_bit(event, workload->pending_events,
388 					 INTEL_GVT_EVENT_MAX)
389 				intel_vgpu_trigger_virtual_event(vgpu, event);
390 		}
391 	}
392 
393 	gvt_dbg_sched("ring id %d complete workload %p status %d\n",
394 			ring_id, workload, workload->status);
395 
396 	scheduler->current_workload[ring_id] = NULL;
397 
398 	list_del_init(&workload->list);
399 	workload->complete(workload);
400 
401 	atomic_dec(&vgpu->running_workload_num);
402 	wake_up(&scheduler->workload_complete_wq);
403 	mutex_unlock(&gvt->lock);
404 }
405 
406 struct workload_thread_param {
407 	struct intel_gvt *gvt;
408 	int ring_id;
409 };
410 
411 static DEFINE_MUTEX(scheduler_mutex);
412 
413 static int workload_thread(void *priv)
414 {
415 	struct workload_thread_param *p = (struct workload_thread_param *)priv;
416 	struct intel_gvt *gvt = p->gvt;
417 	int ring_id = p->ring_id;
418 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
419 	struct intel_vgpu_workload *workload = NULL;
420 	int ret;
421 	bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
422 	DEFINE_WAIT_FUNC(wait, woken_wake_function);
423 
424 	kfree(p);
425 
426 	gvt_dbg_core("workload thread for ring %d started\n", ring_id);
427 
428 	while (!kthread_should_stop()) {
429 		add_wait_queue(&scheduler->waitq[ring_id], &wait);
430 		do {
431 			workload = pick_next_workload(gvt, ring_id);
432 			if (workload)
433 				break;
434 			wait_woken(&wait, TASK_INTERRUPTIBLE,
435 				   MAX_SCHEDULE_TIMEOUT);
436 		} while (!kthread_should_stop());
437 		remove_wait_queue(&scheduler->waitq[ring_id], &wait);
438 
439 		if (!workload)
440 			break;
441 
442 		mutex_lock(&scheduler_mutex);
443 
444 		gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
445 				workload->ring_id, workload,
446 				workload->vgpu->id);
447 
448 		intel_runtime_pm_get(gvt->dev_priv);
449 
450 		gvt_dbg_sched("ring id %d will dispatch workload %p\n",
451 				workload->ring_id, workload);
452 
453 		if (need_force_wake)
454 			intel_uncore_forcewake_get(gvt->dev_priv,
455 					FORCEWAKE_ALL);
456 
457 		mutex_lock(&gvt->lock);
458 		ret = dispatch_workload(workload);
459 		mutex_unlock(&gvt->lock);
460 
461 		if (ret) {
462 			gvt_err("fail to dispatch workload, skip\n");
463 			goto complete;
464 		}
465 
466 		gvt_dbg_sched("ring id %d wait workload %p\n",
467 				workload->ring_id, workload);
468 retry:
469 		i915_wait_request(workload->req,
470 					 0, MAX_SCHEDULE_TIMEOUT);
471 		/* I915 has replay mechanism and a request will be replayed
472 		 * if there is i915 reset. So the seqno will be updated anyway.
473 		 * If the seqno is not updated yet after waiting, which means
474 		 * the replay may still be in progress and we can wait again.
475 		 */
476 		if (!i915_gem_request_completed(workload->req)) {
477 			gvt_dbg_sched("workload %p not completed, wait again\n",
478 					workload);
479 			goto retry;
480 		}
481 
482 complete:
483 		gvt_dbg_sched("will complete workload %p, status: %d\n",
484 				workload, workload->status);
485 
486 		complete_current_workload(gvt, ring_id);
487 
488 		if (need_force_wake)
489 			intel_uncore_forcewake_put(gvt->dev_priv,
490 					FORCEWAKE_ALL);
491 
492 		intel_runtime_pm_put(gvt->dev_priv);
493 
494 		mutex_unlock(&scheduler_mutex);
495 
496 	}
497 	return 0;
498 }
499 
500 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
501 {
502 	struct intel_gvt *gvt = vgpu->gvt;
503 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
504 
505 	if (atomic_read(&vgpu->running_workload_num)) {
506 		gvt_dbg_sched("wait vgpu idle\n");
507 
508 		wait_event(scheduler->workload_complete_wq,
509 				!atomic_read(&vgpu->running_workload_num));
510 	}
511 }
512 
513 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
514 {
515 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
516 	int i;
517 
518 	gvt_dbg_core("clean workload scheduler\n");
519 
520 	for (i = 0; i < I915_NUM_ENGINES; i++) {
521 		if (scheduler->thread[i]) {
522 			kthread_stop(scheduler->thread[i]);
523 			scheduler->thread[i] = NULL;
524 		}
525 	}
526 }
527 
528 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
529 {
530 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
531 	struct workload_thread_param *param = NULL;
532 	int ret;
533 	int i;
534 
535 	gvt_dbg_core("init workload scheduler\n");
536 
537 	init_waitqueue_head(&scheduler->workload_complete_wq);
538 
539 	for (i = 0; i < I915_NUM_ENGINES; i++) {
540 		/* check ring mask at init time */
541 		if (!HAS_ENGINE(gvt->dev_priv, i))
542 			continue;
543 
544 		init_waitqueue_head(&scheduler->waitq[i]);
545 
546 		param = kzalloc(sizeof(*param), GFP_KERNEL);
547 		if (!param) {
548 			ret = -ENOMEM;
549 			goto err;
550 		}
551 
552 		param->gvt = gvt;
553 		param->ring_id = i;
554 
555 		scheduler->thread[i] = kthread_run(workload_thread, param,
556 			"gvt workload %d", i);
557 		if (IS_ERR(scheduler->thread[i])) {
558 			gvt_err("fail to create workload thread\n");
559 			ret = PTR_ERR(scheduler->thread[i]);
560 			goto err;
561 		}
562 	}
563 	return 0;
564 err:
565 	intel_gvt_clean_workload_scheduler(gvt);
566 	kfree(param);
567 	param = NULL;
568 	return ret;
569 }
570 
571 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
572 {
573 	atomic_notifier_chain_unregister(&vgpu->shadow_ctx->status_notifier,
574 			&vgpu->shadow_ctx_notifier_block);
575 
576 	i915_gem_context_put_unlocked(vgpu->shadow_ctx);
577 }
578 
579 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
580 {
581 	atomic_set(&vgpu->running_workload_num, 0);
582 
583 	vgpu->shadow_ctx = i915_gem_context_create_gvt(
584 			&vgpu->gvt->dev_priv->drm);
585 	if (IS_ERR(vgpu->shadow_ctx))
586 		return PTR_ERR(vgpu->shadow_ctx);
587 
588 	vgpu->shadow_ctx->engine[RCS].initialised = true;
589 
590 	vgpu->shadow_ctx_notifier_block.notifier_call =
591 		shadow_context_status_change;
592 
593 	atomic_notifier_chain_register(&vgpu->shadow_ctx->status_notifier,
594 				       &vgpu->shadow_ctx_notifier_block);
595 	return 0;
596 }
597