2 * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
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:
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
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
25 * Kevin Tian <kevin.tian@intel.com>
28 * Min He <min.he@intel.com>
29 * Bing Niu <bing.niu@intel.com>
30 * Zhi Wang <zhi.a.wang@intel.com>
37 static bool vgpu_has_pending_workload(struct intel_vgpu *vgpu)
39 enum intel_engine_id i;
40 struct intel_engine_cs *engine;
42 for_each_engine(engine, vgpu->gvt->dev_priv, i) {
43 if (!list_empty(workload_q_head(vgpu, i)))
50 struct vgpu_sched_data {
51 struct list_head lru_list;
52 struct intel_vgpu *vgpu;
54 ktime_t sched_in_time;
55 ktime_t sched_out_time;
60 struct vgpu_sched_ctl sched_ctl;
63 struct gvt_sched_data {
64 struct intel_gvt *gvt;
67 struct list_head lru_runq_head;
70 static void vgpu_update_timeslice(struct intel_vgpu *pre_vgpu)
73 struct vgpu_sched_data *vgpu_data = pre_vgpu->sched_data;
75 delta_ts = vgpu_data->sched_out_time - vgpu_data->sched_in_time;
77 vgpu_data->sched_time += delta_ts;
78 vgpu_data->left_ts -= delta_ts;
81 #define GVT_TS_BALANCE_PERIOD_MS 100
82 #define GVT_TS_BALANCE_STAGE_NUM 10
84 static void gvt_balance_timeslice(struct gvt_sched_data *sched_data)
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;
91 /* The timeslice accumulation reset at stage 0, which is
92 * allocated again without adding previous debt.
96 ktime_t fair_timeslice;
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;
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 /
109 vgpu_data->allocated_ts = fair_timeslice;
110 vgpu_data->left_ts = vgpu_data->allocated_ts;
113 list_for_each(pos, &sched_data->lru_runq_head) {
114 vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
116 /* timeslice for next 100ms should add the left/debt
117 * slice of previous stages.
119 vgpu_data->left_ts += vgpu_data->allocated_ts;
124 static void try_to_schedule_next_vgpu(struct intel_gvt *gvt)
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;
132 /* no need to schedule if next_vgpu is the same with current_vgpu,
133 * let scheduler chose next_vgpu again by setting it to NULL.
135 if (scheduler->next_vgpu == scheduler->current_vgpu) {
136 scheduler->next_vgpu = NULL;
141 * after the flag is set, workload dispatch thread will
142 * stop dispatching workload for current vgpu
144 scheduler->need_reschedule = true;
146 /* still have uncompleted workload? */
147 for_each_engine(engine, gvt->dev_priv, i) {
148 if (scheduler->current_workload[i])
152 cur_time = ktime_get();
153 if (scheduler->current_vgpu) {
154 vgpu_data = scheduler->current_vgpu->sched_data;
155 vgpu_data->sched_out_time = cur_time;
156 vgpu_update_timeslice(scheduler->current_vgpu);
158 vgpu_data = scheduler->next_vgpu->sched_data;
159 vgpu_data->sched_in_time = cur_time;
161 /* switch current vgpu */
162 scheduler->current_vgpu = scheduler->next_vgpu;
163 scheduler->next_vgpu = NULL;
165 scheduler->need_reschedule = false;
167 /* wake up workload dispatch thread */
168 for_each_engine(engine, gvt->dev_priv, i)
169 wake_up(&scheduler->waitq[i]);
172 static struct intel_vgpu *find_busy_vgpu(struct gvt_sched_data *sched_data)
174 struct vgpu_sched_data *vgpu_data;
175 struct intel_vgpu *vgpu = NULL;
176 struct list_head *head = &sched_data->lru_runq_head;
177 struct list_head *pos;
179 /* search a vgpu with pending workload */
180 list_for_each(pos, head) {
182 vgpu_data = container_of(pos, struct vgpu_sched_data, lru_list);
183 if (!vgpu_has_pending_workload(vgpu_data->vgpu))
186 /* Return the vGPU only if it has time slice left */
187 if (vgpu_data->left_ts > 0) {
188 vgpu = vgpu_data->vgpu;
197 #define GVT_DEFAULT_TIME_SLICE 1000000
199 static void tbs_sched_func(struct gvt_sched_data *sched_data)
201 struct intel_gvt *gvt = sched_data->gvt;
202 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
203 struct vgpu_sched_data *vgpu_data;
204 struct intel_vgpu *vgpu = NULL;
205 static uint64_t timer_check;
207 if (!(timer_check++ % GVT_TS_BALANCE_PERIOD_MS))
208 gvt_balance_timeslice(sched_data);
210 /* no active vgpu or has already had a target */
211 if (list_empty(&sched_data->lru_runq_head) || scheduler->next_vgpu)
214 vgpu = find_busy_vgpu(sched_data);
216 scheduler->next_vgpu = vgpu;
218 /* Move the last used vGPU to the tail of lru_list */
219 vgpu_data = vgpu->sched_data;
220 list_del_init(&vgpu_data->lru_list);
221 list_add_tail(&vgpu_data->lru_list,
222 &sched_data->lru_runq_head);
224 scheduler->next_vgpu = gvt->idle_vgpu;
227 if (scheduler->next_vgpu)
228 try_to_schedule_next_vgpu(gvt);
231 void intel_gvt_schedule(struct intel_gvt *gvt)
233 struct gvt_sched_data *sched_data = gvt->scheduler.sched_data;
235 mutex_lock(&gvt->lock);
236 tbs_sched_func(sched_data);
237 mutex_unlock(&gvt->lock);
240 static enum hrtimer_restart tbs_timer_fn(struct hrtimer *timer_data)
242 struct gvt_sched_data *data;
244 data = container_of(timer_data, struct gvt_sched_data, timer);
246 intel_gvt_request_service(data->gvt, INTEL_GVT_REQUEST_SCHED);
248 hrtimer_add_expires_ns(&data->timer, data->period);
250 return HRTIMER_RESTART;
253 static int tbs_sched_init(struct intel_gvt *gvt)
255 struct intel_gvt_workload_scheduler *scheduler =
258 struct gvt_sched_data *data;
260 data = kzalloc(sizeof(*data), GFP_KERNEL);
264 INIT_LIST_HEAD(&data->lru_runq_head);
265 hrtimer_init(&data->timer, CLOCK_MONOTONIC, HRTIMER_MODE_ABS);
266 data->timer.function = tbs_timer_fn;
267 data->period = GVT_DEFAULT_TIME_SLICE;
270 scheduler->sched_data = data;
275 static void tbs_sched_clean(struct intel_gvt *gvt)
277 struct intel_gvt_workload_scheduler *scheduler =
279 struct gvt_sched_data *data = scheduler->sched_data;
281 hrtimer_cancel(&data->timer);
284 scheduler->sched_data = NULL;
287 static int tbs_sched_init_vgpu(struct intel_vgpu *vgpu)
289 struct vgpu_sched_data *data;
291 data = kzalloc(sizeof(*data), GFP_KERNEL);
295 data->sched_ctl.weight = vgpu->sched_ctl.weight;
297 INIT_LIST_HEAD(&data->lru_list);
299 vgpu->sched_data = data;
304 static void tbs_sched_clean_vgpu(struct intel_vgpu *vgpu)
306 kfree(vgpu->sched_data);
307 vgpu->sched_data = NULL;
310 static void tbs_sched_start_schedule(struct intel_vgpu *vgpu)
312 struct gvt_sched_data *sched_data = vgpu->gvt->scheduler.sched_data;
313 struct vgpu_sched_data *vgpu_data = vgpu->sched_data;
315 if (!list_empty(&vgpu_data->lru_list))
318 list_add_tail(&vgpu_data->lru_list, &sched_data->lru_runq_head);
320 if (!hrtimer_active(&sched_data->timer))
321 hrtimer_start(&sched_data->timer, ktime_add_ns(ktime_get(),
322 sched_data->period), HRTIMER_MODE_ABS);
325 static void tbs_sched_stop_schedule(struct intel_vgpu *vgpu)
327 struct vgpu_sched_data *vgpu_data = vgpu->sched_data;
329 list_del_init(&vgpu_data->lru_list);
332 static struct intel_gvt_sched_policy_ops tbs_schedule_ops = {
333 .init = tbs_sched_init,
334 .clean = tbs_sched_clean,
335 .init_vgpu = tbs_sched_init_vgpu,
336 .clean_vgpu = tbs_sched_clean_vgpu,
337 .start_schedule = tbs_sched_start_schedule,
338 .stop_schedule = tbs_sched_stop_schedule,
341 int intel_gvt_init_sched_policy(struct intel_gvt *gvt)
343 gvt->scheduler.sched_ops = &tbs_schedule_ops;
345 return gvt->scheduler.sched_ops->init(gvt);
348 void intel_gvt_clean_sched_policy(struct intel_gvt *gvt)
350 gvt->scheduler.sched_ops->clean(gvt);
353 int intel_vgpu_init_sched_policy(struct intel_vgpu *vgpu)
355 return vgpu->gvt->scheduler.sched_ops->init_vgpu(vgpu);
358 void intel_vgpu_clean_sched_policy(struct intel_vgpu *vgpu)
360 vgpu->gvt->scheduler.sched_ops->clean_vgpu(vgpu);
363 void intel_vgpu_start_schedule(struct intel_vgpu *vgpu)
365 gvt_dbg_core("vgpu%d: start schedule\n", vgpu->id);
367 vgpu->gvt->scheduler.sched_ops->start_schedule(vgpu);
370 void intel_vgpu_stop_schedule(struct intel_vgpu *vgpu)
372 struct intel_gvt_workload_scheduler *scheduler =
373 &vgpu->gvt->scheduler;
375 gvt_dbg_core("vgpu%d: stop schedule\n", vgpu->id);
377 scheduler->sched_ops->stop_schedule(vgpu);
379 if (scheduler->next_vgpu == vgpu)
380 scheduler->next_vgpu = NULL;
382 if (scheduler->current_vgpu == vgpu) {
383 /* stop workload dispatching */
384 scheduler->need_reschedule = true;
385 scheduler->current_vgpu = NULL;