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  *    Anhua Xu
25  *    Kevin Tian <kevin.tian@intel.com>
26  *
27  * Contributors:
28  *    Min He <min.he@intel.com>
29  *    Bing Niu <bing.niu@intel.com>
30  *    Zhi Wang <zhi.a.wang@intel.com>
31  *
32  */
33 
34 #include "i915_drv.h"
35 #include "gvt.h"
36 
37 static bool vgpu_has_pending_workload(struct intel_vgpu *vgpu)
38 {
39 	enum intel_engine_id i;
40 	struct intel_engine_cs *engine;
41 
42 	for_each_engine(engine, vgpu->gvt->dev_priv, i) {
43 		if (!list_empty(workload_q_head(vgpu, i)))
44 			return true;
45 	}
46 
47 	return false;
48 }
49 
50 struct vgpu_sched_data {
51 	struct list_head lru_list;
52 	struct intel_vgpu *vgpu;
53 
54 	ktime_t sched_in_time;
55 	ktime_t sched_out_time;
56 	ktime_t sched_time;
57 	ktime_t left_ts;
58 	ktime_t allocated_ts;
59 
60 	struct vgpu_sched_ctl sched_ctl;
61 };
62 
63 struct gvt_sched_data {
64 	struct intel_gvt *gvt;
65 	struct hrtimer timer;
66 	unsigned long period;
67 	struct list_head lru_runq_head;
68 };
69 
70 static void vgpu_update_timeslice(struct intel_vgpu *pre_vgpu)
71 {
72 	ktime_t delta_ts;
73 	struct vgpu_sched_data *vgpu_data = pre_vgpu->sched_data;
74 
75 	delta_ts = vgpu_data->sched_out_time - vgpu_data->sched_in_time;
76 
77 	vgpu_data->sched_time += delta_ts;
78 	vgpu_data->left_ts -= delta_ts;
79 }
80 
81 #define GVT_TS_BALANCE_PERIOD_MS 100
82 #define GVT_TS_BALANCE_STAGE_NUM 10
83 
84 static void gvt_balance_timeslice(struct gvt_sched_data *sched_data)
85 {
86 	struct vgpu_sched_data *vgpu_data;
87 	struct list_head *pos;
88 	static uint64_t stage_check;
89 	int stage = stage_check++ % GVT_TS_BALANCE_STAGE_NUM;
90 
91 	/* The timeslice accumulation reset at stage 0, which is
92 	 * allocated again without adding previous debt.
93 	 */
94 	if (stage == 0) {
95 		int total_weight = 0;
96 		ktime_t fair_timeslice;
97 
98 		list_for_each(pos, &sched_data->lru_runq_head) {
99 			vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
100 			total_weight += vgpu_data->sched_ctl.weight;
101 		}
102 
103 		list_for_each(pos, &sched_data->lru_runq_head) {
104 			vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
105 			fair_timeslice = ms_to_ktime(GVT_TS_BALANCE_PERIOD_MS) *
106 						vgpu_data->sched_ctl.weight /
107 						total_weight;
108 
109 			vgpu_data->allocated_ts = fair_timeslice;
110 			vgpu_data->left_ts = vgpu_data->allocated_ts;
111 		}
112 	} else {
113 		list_for_each(pos, &sched_data->lru_runq_head) {
114 			vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
115 
116 			/* timeslice for next 100ms should add the left/debt
117 			 * slice of previous stages.
118 			 */
119 			vgpu_data->left_ts += vgpu_data->allocated_ts;
120 		}
121 	}
122 }
123 
124 static void try_to_schedule_next_vgpu(struct intel_gvt *gvt)
125 {
126 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
127 	enum intel_engine_id i;
128 	struct intel_engine_cs *engine;
129 	struct vgpu_sched_data *vgpu_data;
130 	ktime_t cur_time;
131 
132 	/* no target to schedule */
133 	if (!scheduler->next_vgpu)
134 		return;
135 
136 	/*
137 	 * after the flag is set, workload dispatch thread will
138 	 * stop dispatching workload for current vgpu
139 	 */
140 	scheduler->need_reschedule = true;
141 
142 	/* still have uncompleted workload? */
143 	for_each_engine(engine, gvt->dev_priv, i) {
144 		if (scheduler->current_workload[i])
145 			return;
146 	}
147 
148 	cur_time = ktime_get();
149 	if (scheduler->current_vgpu) {
150 		vgpu_data = scheduler->current_vgpu->sched_data;
151 		vgpu_data->sched_out_time = cur_time;
152 		vgpu_update_timeslice(scheduler->current_vgpu);
153 	}
154 	vgpu_data = scheduler->next_vgpu->sched_data;
155 	vgpu_data->sched_in_time = cur_time;
156 
157 	/* switch current vgpu */
158 	scheduler->current_vgpu = scheduler->next_vgpu;
159 	scheduler->next_vgpu = NULL;
160 
161 	scheduler->need_reschedule = false;
162 
163 	/* wake up workload dispatch thread */
164 	for_each_engine(engine, gvt->dev_priv, i)
165 		wake_up(&scheduler->waitq[i]);
166 }
167 
168 static struct intel_vgpu *find_busy_vgpu(struct gvt_sched_data *sched_data)
169 {
170 	struct vgpu_sched_data *vgpu_data;
171 	struct intel_vgpu *vgpu = NULL;
172 	struct list_head *head = &sched_data->lru_runq_head;
173 	struct list_head *pos;
174 
175 	/* search a vgpu with pending workload */
176 	list_for_each(pos, head) {
177 
178 		vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
179 		if (!vgpu_has_pending_workload(vgpu_data->vgpu))
180 			continue;
181 
182 		/* Return the vGPU only if it has time slice left */
183 		if (vgpu_data->left_ts > 0) {
184 			vgpu = vgpu_data->vgpu;
185 			break;
186 		}
187 	}
188 
189 	return vgpu;
190 }
191 
192 /* in nanosecond */
193 #define GVT_DEFAULT_TIME_SLICE 1000000
194 
195 static void tbs_sched_func(struct gvt_sched_data *sched_data)
196 {
197 	struct intel_gvt *gvt = sched_data->gvt;
198 	struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
199 	struct vgpu_sched_data *vgpu_data;
200 	struct intel_vgpu *vgpu = NULL;
201 	static uint64_t timer_check;
202 
203 	if (!(timer_check++ % GVT_TS_BALANCE_PERIOD_MS))
204 		gvt_balance_timeslice(sched_data);
205 
206 	/* no active vgpu or has already had a target */
207 	if (list_empty(&sched_data->lru_runq_head) || scheduler->next_vgpu)
208 		goto out;
209 
210 	vgpu = find_busy_vgpu(sched_data);
211 	if (vgpu) {
212 		scheduler->next_vgpu = vgpu;
213 
214 		/* Move the last used vGPU to the tail of lru_list */
215 		vgpu_data = vgpu->sched_data;
216 		list_del_init(&vgpu_data->lru_list);
217 		list_add_tail(&vgpu_data->lru_list,
218 				&sched_data->lru_runq_head);
219 	} else {
220 		scheduler->next_vgpu = gvt->idle_vgpu;
221 	}
222 out:
223 	if (scheduler->next_vgpu)
224 		try_to_schedule_next_vgpu(gvt);
225 }
226 
227 void intel_gvt_schedule(struct intel_gvt *gvt)
228 {
229 	struct gvt_sched_data *sched_data = gvt->scheduler.sched_data;
230 
231 	mutex_lock(&gvt->lock);
232 	tbs_sched_func(sched_data);
233 	mutex_unlock(&gvt->lock);
234 }
235 
236 static enum hrtimer_restart tbs_timer_fn(struct hrtimer *timer_data)
237 {
238 	struct gvt_sched_data *data;
239 
240 	data = container_of(timer_data, struct gvt_sched_data, timer);
241 
242 	intel_gvt_request_service(data->gvt, INTEL_GVT_REQUEST_SCHED);
243 
244 	hrtimer_add_expires_ns(&data->timer, data->period);
245 
246 	return HRTIMER_RESTART;
247 }
248 
249 static int tbs_sched_init(struct intel_gvt *gvt)
250 {
251 	struct intel_gvt_workload_scheduler *scheduler =
252 		&gvt->scheduler;
253 
254 	struct gvt_sched_data *data;
255 
256 	data = kzalloc(sizeof(*data), GFP_KERNEL);
257 	if (!data)
258 		return -ENOMEM;
259 
260 	INIT_LIST_HEAD(&data->lru_runq_head);
261 	hrtimer_init(&data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
262 	data->timer.function = tbs_timer_fn;
263 	data->period = GVT_DEFAULT_TIME_SLICE;
264 	data->gvt = gvt;
265 
266 	scheduler->sched_data = data;
267 
268 	return 0;
269 }
270 
271 static void tbs_sched_clean(struct intel_gvt *gvt)
272 {
273 	struct intel_gvt_workload_scheduler *scheduler =
274 		&gvt->scheduler;
275 	struct gvt_sched_data *data = scheduler->sched_data;
276 
277 	hrtimer_cancel(&data->timer);
278 
279 	kfree(data);
280 	scheduler->sched_data = NULL;
281 }
282 
283 static int tbs_sched_init_vgpu(struct intel_vgpu *vgpu)
284 {
285 	struct vgpu_sched_data *data;
286 
287 	data = kzalloc(sizeof(*data), GFP_KERNEL);
288 	if (!data)
289 		return -ENOMEM;
290 
291 	data->sched_ctl.weight = vgpu->sched_ctl.weight;
292 	data->vgpu = vgpu;
293 	INIT_LIST_HEAD(&data->lru_list);
294 
295 	vgpu->sched_data = data;
296 
297 	return 0;
298 }
299 
300 static void tbs_sched_clean_vgpu(struct intel_vgpu *vgpu)
301 {
302 	kfree(vgpu->sched_data);
303 	vgpu->sched_data = NULL;
304 }
305 
306 static void tbs_sched_start_schedule(struct intel_vgpu *vgpu)
307 {
308 	struct gvt_sched_data *sched_data = vgpu->gvt->scheduler.sched_data;
309 	struct vgpu_sched_data *vgpu_data = vgpu->sched_data;
310 
311 	if (!list_empty(&vgpu_data->lru_list))
312 		return;
313 
314 	list_add_tail(&vgpu_data->lru_list, &sched_data->lru_runq_head);
315 
316 	if (!hrtimer_active(&sched_data->timer))
317 		hrtimer_start(&sched_data->timer, ktime_add_ns(ktime_get(),
318 			sched_data->period), HRTIMER_MODE_ABS);
319 }
320 
321 static void tbs_sched_stop_schedule(struct intel_vgpu *vgpu)
322 {
323 	struct vgpu_sched_data *vgpu_data = vgpu->sched_data;
324 
325 	list_del_init(&vgpu_data->lru_list);
326 }
327 
328 static struct intel_gvt_sched_policy_ops tbs_schedule_ops = {
329 	.init = tbs_sched_init,
330 	.clean = tbs_sched_clean,
331 	.init_vgpu = tbs_sched_init_vgpu,
332 	.clean_vgpu = tbs_sched_clean_vgpu,
333 	.start_schedule = tbs_sched_start_schedule,
334 	.stop_schedule = tbs_sched_stop_schedule,
335 };
336 
337 int intel_gvt_init_sched_policy(struct intel_gvt *gvt)
338 {
339 	gvt->scheduler.sched_ops = &tbs_schedule_ops;
340 
341 	return gvt->scheduler.sched_ops->init(gvt);
342 }
343 
344 void intel_gvt_clean_sched_policy(struct intel_gvt *gvt)
345 {
346 	gvt->scheduler.sched_ops->clean(gvt);
347 }
348 
349 int intel_vgpu_init_sched_policy(struct intel_vgpu *vgpu)
350 {
351 	return vgpu->gvt->scheduler.sched_ops->init_vgpu(vgpu);
352 }
353 
354 void intel_vgpu_clean_sched_policy(struct intel_vgpu *vgpu)
355 {
356 	vgpu->gvt->scheduler.sched_ops->clean_vgpu(vgpu);
357 }
358 
359 void intel_vgpu_start_schedule(struct intel_vgpu *vgpu)
360 {
361 	gvt_dbg_core("vgpu%d: start schedule\n", vgpu->id);
362 
363 	vgpu->gvt->scheduler.sched_ops->start_schedule(vgpu);
364 }
365 
366 void intel_vgpu_stop_schedule(struct intel_vgpu *vgpu)
367 {
368 	struct intel_gvt_workload_scheduler *scheduler =
369 		&vgpu->gvt->scheduler;
370 
371 	gvt_dbg_core("vgpu%d: stop schedule\n", vgpu->id);
372 
373 	scheduler->sched_ops->stop_schedule(vgpu);
374 
375 	if (scheduler->next_vgpu == vgpu)
376 		scheduler->next_vgpu = NULL;
377 
378 	if (scheduler->current_vgpu == vgpu) {
379 		/* stop workload dispatching */
380 		scheduler->need_reschedule = true;
381 		scheduler->current_vgpu = NULL;
382 	}
383 }
384