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);
92 dst = kmap_atomic(page);
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_atomic(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));
126 kunmap_atomic(shadow_ring_context);
130 static int shadow_context_status_change(struct notifier_block *nb,
131 unsigned long action, void *data)
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];
143 case INTEL_CONTEXT_SCHEDULE_IN:
144 intel_gvt_load_render_mmio(workload->vgpu,
146 atomic_set(&workload->shadow_ctx_active, 1);
148 case INTEL_CONTEXT_SCHEDULE_OUT:
149 intel_gvt_restore_render_mmio(workload->vgpu,
151 atomic_set(&workload->shadow_ctx_active, 0);
157 wake_up(&workload->shadow_ctx_status_wq);
161 static int dispatch_workload(struct intel_vgpu_workload *workload)
163 struct intel_vgpu *vgpu = workload->vgpu;
164 struct intel_gvt *gvt = vgpu->gvt;
165 int ring_id = workload->ring_id;
166 struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
167 struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
168 struct drm_i915_gem_request *rq;
171 gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
174 shadow_ctx->desc_template = workload->ctx_desc.addressing_mode <<
175 GEN8_CTX_ADDRESSING_MODE_SHIFT;
177 rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
179 gvt_err("fail to allocate gem request\n");
180 workload->status = PTR_ERR(rq);
181 return workload->status;
184 gvt_dbg_sched("ring id %d get i915 gem request %p\n", ring_id, rq);
186 workload->req = i915_gem_request_get(rq);
188 mutex_lock(&gvt->lock);
190 ret = intel_gvt_scan_and_shadow_workload(workload);
194 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
198 ret = populate_shadow_context(workload);
202 if (workload->prepare) {
203 ret = workload->prepare(workload);
208 mutex_unlock(&gvt->lock);
210 gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
211 ring_id, workload->req);
213 i915_add_request_no_flush(rq);
214 workload->dispatched = true;
217 workload->status = ret;
219 mutex_unlock(&gvt->lock);
221 i915_add_request_no_flush(rq);
225 static struct intel_vgpu_workload *pick_next_workload(
226 struct intel_gvt *gvt, int ring_id)
228 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
229 struct intel_vgpu_workload *workload = NULL;
231 mutex_lock(&gvt->lock);
234 * no current vgpu / will be scheduled out / no workload
237 if (!scheduler->current_vgpu) {
238 gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
242 if (scheduler->need_reschedule) {
243 gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
247 if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
248 gvt_dbg_sched("ring id %d stop - no available workload\n",
254 * still have current workload, maybe the workload disptacher
255 * fail to submit it for some reason, resubmit it.
257 if (scheduler->current_workload[ring_id]) {
258 workload = scheduler->current_workload[ring_id];
259 gvt_dbg_sched("ring id %d still have current workload %p\n",
265 * pick a workload as current workload
266 * once current workload is set, schedule policy routines
267 * will wait the current workload is finished when trying to
268 * schedule out a vgpu.
270 scheduler->current_workload[ring_id] = container_of(
271 workload_q_head(scheduler->current_vgpu, ring_id)->next,
272 struct intel_vgpu_workload, list);
274 workload = scheduler->current_workload[ring_id];
276 gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
278 atomic_inc(&workload->vgpu->running_workload_num);
280 mutex_unlock(&gvt->lock);
284 static void update_guest_context(struct intel_vgpu_workload *workload)
286 struct intel_vgpu *vgpu = workload->vgpu;
287 struct intel_gvt *gvt = vgpu->gvt;
288 int ring_id = workload->ring_id;
289 struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
290 struct drm_i915_gem_object *ctx_obj =
291 shadow_ctx->engine[ring_id].state->obj;
292 struct execlist_ring_context *shadow_ring_context;
295 unsigned long context_gpa, context_page_num;
298 gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
299 workload->ctx_desc.lrca);
301 context_page_num = intel_lr_context_size(
302 gvt->dev_priv->engine[ring_id]);
304 context_page_num = context_page_num >> PAGE_SHIFT;
306 if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
307 context_page_num = 19;
311 while (i < context_page_num) {
312 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
313 (u32)((workload->ctx_desc.lrca + i) <<
315 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
316 gvt_err("invalid guest context descriptor\n");
320 page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
321 src = kmap_atomic(page);
322 intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
328 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
329 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
331 page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
332 shadow_ring_context = kmap_atomic(page);
334 #define COPY_REG(name) \
335 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
336 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
339 COPY_REG(ctx_timestamp);
343 intel_gvt_hypervisor_write_gpa(vgpu,
344 workload->ring_context_gpa +
345 sizeof(*shadow_ring_context),
346 (void *)shadow_ring_context +
347 sizeof(*shadow_ring_context),
348 GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
350 kunmap_atomic(shadow_ring_context);
353 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
355 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
356 struct intel_vgpu_workload *workload;
359 mutex_lock(&gvt->lock);
361 workload = scheduler->current_workload[ring_id];
363 if (!workload->status && !workload->vgpu->resetting) {
364 wait_event(workload->shadow_ctx_status_wq,
365 !atomic_read(&workload->shadow_ctx_active));
367 update_guest_context(workload);
369 for_each_set_bit(event, workload->pending_events,
371 intel_vgpu_trigger_virtual_event(workload->vgpu,
375 gvt_dbg_sched("ring id %d complete workload %p status %d\n",
376 ring_id, workload, workload->status);
378 scheduler->current_workload[ring_id] = NULL;
380 atomic_dec(&workload->vgpu->running_workload_num);
382 list_del_init(&workload->list);
383 workload->complete(workload);
385 wake_up(&scheduler->workload_complete_wq);
386 mutex_unlock(&gvt->lock);
389 struct workload_thread_param {
390 struct intel_gvt *gvt;
394 static DEFINE_MUTEX(scheduler_mutex);
396 static int workload_thread(void *priv)
398 struct workload_thread_param *p = (struct workload_thread_param *)priv;
399 struct intel_gvt *gvt = p->gvt;
400 int ring_id = p->ring_id;
401 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
402 struct intel_vgpu_workload *workload = NULL;
404 bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
408 gvt_dbg_core("workload thread for ring %d started\n", ring_id);
410 while (!kthread_should_stop()) {
411 ret = wait_event_interruptible(scheduler->waitq[ring_id],
412 kthread_should_stop() ||
413 (workload = pick_next_workload(gvt, ring_id)));
417 if (kthread_should_stop())
420 mutex_lock(&scheduler_mutex);
422 gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
423 workload->ring_id, workload,
426 intel_runtime_pm_get(gvt->dev_priv);
428 gvt_dbg_sched("ring id %d will dispatch workload %p\n",
429 workload->ring_id, workload);
432 intel_uncore_forcewake_get(gvt->dev_priv,
435 mutex_lock(&gvt->dev_priv->drm.struct_mutex);
436 ret = dispatch_workload(workload);
437 mutex_unlock(&gvt->dev_priv->drm.struct_mutex);
440 gvt_err("fail to dispatch workload, skip\n");
444 gvt_dbg_sched("ring id %d wait workload %p\n",
445 workload->ring_id, workload);
447 workload->status = i915_wait_request(workload->req,
449 if (workload->status != 0)
450 gvt_err("fail to wait workload, skip\n");
453 gvt_dbg_sched("will complete workload %p\n, status: %d\n",
454 workload, workload->status);
456 mutex_lock(&gvt->dev_priv->drm.struct_mutex);
457 complete_current_workload(gvt, ring_id);
458 mutex_unlock(&gvt->dev_priv->drm.struct_mutex);
460 i915_gem_request_put(fetch_and_zero(&workload->req));
463 intel_uncore_forcewake_put(gvt->dev_priv,
466 intel_runtime_pm_put(gvt->dev_priv);
468 mutex_unlock(&scheduler_mutex);
474 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
476 struct intel_gvt *gvt = vgpu->gvt;
477 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
479 if (atomic_read(&vgpu->running_workload_num)) {
480 gvt_dbg_sched("wait vgpu idle\n");
482 wait_event(scheduler->workload_complete_wq,
483 !atomic_read(&vgpu->running_workload_num));
487 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
489 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
492 gvt_dbg_core("clean workload scheduler\n");
494 for (i = 0; i < I915_NUM_ENGINES; i++) {
495 if (scheduler->thread[i]) {
496 kthread_stop(scheduler->thread[i]);
497 scheduler->thread[i] = NULL;
502 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
504 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
505 struct workload_thread_param *param = NULL;
509 gvt_dbg_core("init workload scheduler\n");
511 init_waitqueue_head(&scheduler->workload_complete_wq);
513 for (i = 0; i < I915_NUM_ENGINES; i++) {
514 /* check ring mask at init time */
515 if (!HAS_ENGINE(gvt->dev_priv, i))
518 init_waitqueue_head(&scheduler->waitq[i]);
520 param = kzalloc(sizeof(*param), GFP_KERNEL);
529 scheduler->thread[i] = kthread_run(workload_thread, param,
530 "gvt workload %d", i);
531 if (IS_ERR(scheduler->thread[i])) {
532 gvt_err("fail to create workload thread\n");
533 ret = PTR_ERR(scheduler->thread[i]);
539 intel_gvt_clean_workload_scheduler(gvt);
545 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
547 struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
549 atomic_notifier_chain_unregister(&vgpu->shadow_ctx->status_notifier,
550 &vgpu->shadow_ctx_notifier_block);
552 mutex_lock(&dev_priv->drm.struct_mutex);
554 /* a little hacky to mark as ctx closed */
555 vgpu->shadow_ctx->closed = true;
556 i915_gem_context_put(vgpu->shadow_ctx);
558 mutex_unlock(&dev_priv->drm.struct_mutex);
561 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
563 atomic_set(&vgpu->running_workload_num, 0);
565 vgpu->shadow_ctx = i915_gem_context_create_gvt(
566 &vgpu->gvt->dev_priv->drm);
567 if (IS_ERR(vgpu->shadow_ctx))
568 return PTR_ERR(vgpu->shadow_ctx);
570 vgpu->shadow_ctx->engine[RCS].initialised = true;
572 vgpu->shadow_ctx_notifier_block.notifier_call =
573 shadow_context_status_change;
575 atomic_notifier_chain_register(&vgpu->shadow_ctx->status_notifier,
576 &vgpu->shadow_ctx_notifier_block);