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
24 * Zhi Wang <zhi.a.wang@intel.com>
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>
36 #include <linux/kthread.h>
41 #define RING_CTX_OFF(x) \
42 offsetof(struct execlist_ring_context, x)
44 static void set_context_pdp_root_pointer(
45 struct execlist_ring_context *ring_context,
48 struct execlist_mmio_pair *pdp_pair = &ring_context->pdp3_UDW;
51 for (i = 0; i < 8; i++)
52 pdp_pair[i].val = pdp[7 - i];
55 static int populate_shadow_context(struct intel_vgpu_workload *workload)
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;
66 unsigned long context_gpa, context_page_num;
69 gvt_dbg_sched("ring id %d workload lrca %x", ring_id,
70 workload->ctx_desc.lrca);
72 context_page_num = intel_lr_context_size(
73 gvt->dev_priv->engine[ring_id]);
75 context_page_num = context_page_num >> PAGE_SHIFT;
77 if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
78 context_page_num = 19;
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) <<
86 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
87 gvt_err("Invalid guest context descriptor\n");
91 page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
93 intel_gvt_hypervisor_read_gpa(vgpu, context_gpa, dst,
99 page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
100 shadow_ring_context = kmap(page);
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)
107 COPY_REG(ctx_timestamp);
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);
116 set_context_pdp_root_pointer(shadow_ring_context,
117 workload->shadow_mm->shadow_page_table);
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));
130 static int shadow_context_status_change(struct notifier_block *nb,
131 unsigned long action, void *data)
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];
140 if (unlikely(!workload))
144 case INTEL_CONTEXT_SCHEDULE_IN:
145 intel_gvt_load_render_mmio(workload->vgpu,
147 atomic_set(&workload->shadow_ctx_active, 1);
149 case INTEL_CONTEXT_SCHEDULE_OUT:
150 intel_gvt_restore_render_mmio(workload->vgpu,
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
159 if (workload->status == -EINPROGRESS)
160 workload->status = 0;
161 atomic_set(&workload->shadow_ctx_active, 0);
167 wake_up(&workload->shadow_ctx_status_wq);
171 static int dispatch_workload(struct intel_vgpu_workload *workload)
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 drm_i915_gem_request *rq;
179 gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
182 shadow_ctx->desc_template &= ~(0x3 << GEN8_CTX_ADDRESSING_MODE_SHIFT);
183 shadow_ctx->desc_template |= workload->ctx_desc.addressing_mode <<
184 GEN8_CTX_ADDRESSING_MODE_SHIFT;
186 mutex_lock(&dev_priv->drm.struct_mutex);
188 rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
190 gvt_err("fail to allocate gem request\n");
195 gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
197 workload->req = i915_gem_request_get(rq);
199 ret = intel_gvt_scan_and_shadow_workload(workload);
203 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
207 ret = populate_shadow_context(workload);
211 if (workload->prepare) {
212 ret = workload->prepare(workload);
217 gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
218 ring_id, workload->req);
221 workload->dispatched = true;
224 workload->status = ret;
226 if (!IS_ERR_OR_NULL(rq))
227 i915_add_request(rq);
228 mutex_unlock(&dev_priv->drm.struct_mutex);
232 static struct intel_vgpu_workload *pick_next_workload(
233 struct intel_gvt *gvt, int ring_id)
235 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
236 struct intel_vgpu_workload *workload = NULL;
238 mutex_lock(&gvt->lock);
241 * no current vgpu / will be scheduled out / no workload
244 if (!scheduler->current_vgpu) {
245 gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
249 if (scheduler->need_reschedule) {
250 gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
254 if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
255 gvt_dbg_sched("ring id %d stop - no available workload\n",
261 * still have current workload, maybe the workload disptacher
262 * fail to submit it for some reason, resubmit it.
264 if (scheduler->current_workload[ring_id]) {
265 workload = scheduler->current_workload[ring_id];
266 gvt_dbg_sched("ring id %d still have current workload %p\n",
272 * pick a workload as current workload
273 * once current workload is set, schedule policy routines
274 * will wait the current workload is finished when trying to
275 * schedule out a vgpu.
277 scheduler->current_workload[ring_id] = container_of(
278 workload_q_head(scheduler->current_vgpu, ring_id)->next,
279 struct intel_vgpu_workload, list);
281 workload = scheduler->current_workload[ring_id];
283 gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
285 atomic_inc(&workload->vgpu->running_workload_num);
287 mutex_unlock(&gvt->lock);
291 static void update_guest_context(struct intel_vgpu_workload *workload)
293 struct intel_vgpu *vgpu = workload->vgpu;
294 struct intel_gvt *gvt = vgpu->gvt;
295 int ring_id = workload->ring_id;
296 struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
297 struct drm_i915_gem_object *ctx_obj =
298 shadow_ctx->engine[ring_id].state->obj;
299 struct execlist_ring_context *shadow_ring_context;
302 unsigned long context_gpa, context_page_num;
305 gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
306 workload->ctx_desc.lrca);
308 context_page_num = intel_lr_context_size(
309 gvt->dev_priv->engine[ring_id]);
311 context_page_num = context_page_num >> PAGE_SHIFT;
313 if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
314 context_page_num = 19;
318 while (i < context_page_num) {
319 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
320 (u32)((workload->ctx_desc.lrca + i) <<
322 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
323 gvt_err("invalid guest context descriptor\n");
327 page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
329 intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
335 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
336 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
338 page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
339 shadow_ring_context = kmap(page);
341 #define COPY_REG(name) \
342 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
343 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
346 COPY_REG(ctx_timestamp);
350 intel_gvt_hypervisor_write_gpa(vgpu,
351 workload->ring_context_gpa +
352 sizeof(*shadow_ring_context),
353 (void *)shadow_ring_context +
354 sizeof(*shadow_ring_context),
355 GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
360 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
362 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
363 struct intel_vgpu_workload *workload;
364 struct intel_vgpu *vgpu;
367 mutex_lock(&gvt->lock);
369 workload = scheduler->current_workload[ring_id];
370 vgpu = workload->vgpu;
372 /* For the workload w/ request, needs to wait for the context
373 * switch to make sure request is completed.
374 * For the workload w/o request, directly complete the workload.
377 wait_event(workload->shadow_ctx_status_wq,
378 !atomic_read(&workload->shadow_ctx_active));
380 i915_gem_request_put(fetch_and_zero(&workload->req));
382 if (!workload->status && !vgpu->resetting) {
383 update_guest_context(workload);
385 for_each_set_bit(event, workload->pending_events,
387 intel_vgpu_trigger_virtual_event(vgpu, event);
391 gvt_dbg_sched("ring id %d complete workload %p status %d\n",
392 ring_id, workload, workload->status);
394 scheduler->current_workload[ring_id] = NULL;
396 list_del_init(&workload->list);
397 workload->complete(workload);
399 atomic_dec(&vgpu->running_workload_num);
400 wake_up(&scheduler->workload_complete_wq);
401 mutex_unlock(&gvt->lock);
404 struct workload_thread_param {
405 struct intel_gvt *gvt;
409 static DEFINE_MUTEX(scheduler_mutex);
411 static int workload_thread(void *priv)
413 struct workload_thread_param *p = (struct workload_thread_param *)priv;
414 struct intel_gvt *gvt = p->gvt;
415 int ring_id = p->ring_id;
416 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
417 struct intel_vgpu_workload *workload = NULL;
419 bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
420 DEFINE_WAIT_FUNC(wait, woken_wake_function);
424 gvt_dbg_core("workload thread for ring %d started\n", ring_id);
426 while (!kthread_should_stop()) {
427 add_wait_queue(&scheduler->waitq[ring_id], &wait);
429 workload = pick_next_workload(gvt, ring_id);
432 wait_woken(&wait, TASK_INTERRUPTIBLE,
433 MAX_SCHEDULE_TIMEOUT);
434 } while (!kthread_should_stop());
435 remove_wait_queue(&scheduler->waitq[ring_id], &wait);
440 mutex_lock(&scheduler_mutex);
442 gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
443 workload->ring_id, workload,
446 intel_runtime_pm_get(gvt->dev_priv);
448 gvt_dbg_sched("ring id %d will dispatch workload %p\n",
449 workload->ring_id, workload);
452 intel_uncore_forcewake_get(gvt->dev_priv,
455 mutex_lock(&gvt->lock);
456 ret = dispatch_workload(workload);
457 mutex_unlock(&gvt->lock);
460 gvt_err("fail to dispatch workload, skip\n");
464 gvt_dbg_sched("ring id %d wait workload %p\n",
465 workload->ring_id, workload);
467 i915_wait_request(workload->req,
468 0, MAX_SCHEDULE_TIMEOUT);
469 /* I915 has replay mechanism and a request will be replayed
470 * if there is i915 reset. So the seqno will be updated anyway.
471 * If the seqno is not updated yet after waiting, which means
472 * the replay may still be in progress and we can wait again.
474 if (!i915_gem_request_completed(workload->req)) {
475 gvt_dbg_sched("workload %p not completed, wait again\n",
481 gvt_dbg_sched("will complete workload %p, status: %d\n",
482 workload, workload->status);
484 complete_current_workload(gvt, ring_id);
487 intel_uncore_forcewake_put(gvt->dev_priv,
490 intel_runtime_pm_put(gvt->dev_priv);
492 mutex_unlock(&scheduler_mutex);
498 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
500 struct intel_gvt *gvt = vgpu->gvt;
501 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
503 if (atomic_read(&vgpu->running_workload_num)) {
504 gvt_dbg_sched("wait vgpu idle\n");
506 wait_event(scheduler->workload_complete_wq,
507 !atomic_read(&vgpu->running_workload_num));
511 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
513 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
514 struct intel_engine_cs *engine;
515 enum intel_engine_id i;
517 gvt_dbg_core("clean workload scheduler\n");
519 for_each_engine(engine, gvt->dev_priv, i) {
520 atomic_notifier_chain_unregister(
521 &engine->context_status_notifier,
522 &gvt->shadow_ctx_notifier_block[i]);
523 kthread_stop(scheduler->thread[i]);
527 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
529 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
530 struct workload_thread_param *param = NULL;
531 struct intel_engine_cs *engine;
532 enum intel_engine_id i;
535 gvt_dbg_core("init workload scheduler\n");
537 init_waitqueue_head(&scheduler->workload_complete_wq);
539 for_each_engine(engine, gvt->dev_priv, i) {
540 init_waitqueue_head(&scheduler->waitq[i]);
542 param = kzalloc(sizeof(*param), GFP_KERNEL);
551 scheduler->thread[i] = kthread_run(workload_thread, param,
552 "gvt workload %d", i);
553 if (IS_ERR(scheduler->thread[i])) {
554 gvt_err("fail to create workload thread\n");
555 ret = PTR_ERR(scheduler->thread[i]);
559 gvt->shadow_ctx_notifier_block[i].notifier_call =
560 shadow_context_status_change;
561 atomic_notifier_chain_register(&engine->context_status_notifier,
562 &gvt->shadow_ctx_notifier_block[i]);
566 intel_gvt_clean_workload_scheduler(gvt);
572 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
574 i915_gem_context_put_unlocked(vgpu->shadow_ctx);
577 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
579 atomic_set(&vgpu->running_workload_num, 0);
581 vgpu->shadow_ctx = i915_gem_context_create_gvt(
582 &vgpu->gvt->dev_priv->drm);
583 if (IS_ERR(vgpu->shadow_ctx))
584 return PTR_ERR(vgpu->shadow_ctx);
586 vgpu->shadow_ctx->engine[RCS].initialised = true;