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 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 int ring_id = workload->ring_id;
164 struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
165 struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
166 struct drm_i915_gem_request *rq;
169 gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
172 shadow_ctx->desc_template = workload->ctx_desc.addressing_mode <<
173 GEN8_CTX_ADDRESSING_MODE_SHIFT;
175 mutex_lock(&dev_priv->drm.struct_mutex);
177 rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
179 gvt_err("fail to allocate gem request\n");
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 ret = intel_gvt_scan_and_shadow_workload(workload);
192 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
196 ret = populate_shadow_context(workload);
200 if (workload->prepare) {
201 ret = workload->prepare(workload);
206 gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
207 ring_id, workload->req);
210 workload->dispatched = true;
213 workload->status = ret;
215 if (!IS_ERR_OR_NULL(rq))
216 i915_add_request_no_flush(rq);
217 mutex_unlock(&dev_priv->drm.struct_mutex);
221 static struct intel_vgpu_workload *pick_next_workload(
222 struct intel_gvt *gvt, int ring_id)
224 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
225 struct intel_vgpu_workload *workload = NULL;
227 mutex_lock(&gvt->lock);
230 * no current vgpu / will be scheduled out / no workload
233 if (!scheduler->current_vgpu) {
234 gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
238 if (scheduler->need_reschedule) {
239 gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
243 if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id))) {
244 gvt_dbg_sched("ring id %d stop - no available workload\n",
250 * still have current workload, maybe the workload disptacher
251 * fail to submit it for some reason, resubmit it.
253 if (scheduler->current_workload[ring_id]) {
254 workload = scheduler->current_workload[ring_id];
255 gvt_dbg_sched("ring id %d still have current workload %p\n",
261 * pick a workload as current workload
262 * once current workload is set, schedule policy routines
263 * will wait the current workload is finished when trying to
264 * schedule out a vgpu.
266 scheduler->current_workload[ring_id] = container_of(
267 workload_q_head(scheduler->current_vgpu, ring_id)->next,
268 struct intel_vgpu_workload, list);
270 workload = scheduler->current_workload[ring_id];
272 gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
274 atomic_inc(&workload->vgpu->running_workload_num);
276 mutex_unlock(&gvt->lock);
280 static void update_guest_context(struct intel_vgpu_workload *workload)
282 struct intel_vgpu *vgpu = workload->vgpu;
283 struct intel_gvt *gvt = vgpu->gvt;
284 int ring_id = workload->ring_id;
285 struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
286 struct drm_i915_gem_object *ctx_obj =
287 shadow_ctx->engine[ring_id].state->obj;
288 struct execlist_ring_context *shadow_ring_context;
291 unsigned long context_gpa, context_page_num;
294 gvt_dbg_sched("ring id %d workload lrca %x\n", ring_id,
295 workload->ctx_desc.lrca);
297 context_page_num = intel_lr_context_size(
298 gvt->dev_priv->engine[ring_id]);
300 context_page_num = context_page_num >> PAGE_SHIFT;
302 if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
303 context_page_num = 19;
307 while (i < context_page_num) {
308 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
309 (u32)((workload->ctx_desc.lrca + i) <<
311 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
312 gvt_err("invalid guest context descriptor\n");
316 page = i915_gem_object_get_page(ctx_obj, LRC_PPHWSP_PN + i);
318 intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
324 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
325 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
327 page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
328 shadow_ring_context = kmap(page);
330 #define COPY_REG(name) \
331 intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
332 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
335 COPY_REG(ctx_timestamp);
339 intel_gvt_hypervisor_write_gpa(vgpu,
340 workload->ring_context_gpa +
341 sizeof(*shadow_ring_context),
342 (void *)shadow_ring_context +
343 sizeof(*shadow_ring_context),
344 GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
349 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
351 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
352 struct intel_vgpu_workload *workload;
353 struct intel_vgpu *vgpu;
356 mutex_lock(&gvt->lock);
358 workload = scheduler->current_workload[ring_id];
359 vgpu = workload->vgpu;
361 if (!workload->status && !vgpu->resetting) {
362 wait_event(workload->shadow_ctx_status_wq,
363 !atomic_read(&workload->shadow_ctx_active));
365 update_guest_context(workload);
367 for_each_set_bit(event, workload->pending_events,
369 intel_vgpu_trigger_virtual_event(vgpu, event);
372 gvt_dbg_sched("ring id %d complete workload %p status %d\n",
373 ring_id, workload, workload->status);
375 scheduler->current_workload[ring_id] = NULL;
377 list_del_init(&workload->list);
378 workload->complete(workload);
380 atomic_dec(&vgpu->running_workload_num);
381 wake_up(&scheduler->workload_complete_wq);
382 mutex_unlock(&gvt->lock);
385 struct workload_thread_param {
386 struct intel_gvt *gvt;
390 static DEFINE_MUTEX(scheduler_mutex);
392 static int workload_thread(void *priv)
394 struct workload_thread_param *p = (struct workload_thread_param *)priv;
395 struct intel_gvt *gvt = p->gvt;
396 int ring_id = p->ring_id;
397 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
398 struct intel_vgpu_workload *workload = NULL;
401 bool need_force_wake = IS_SKYLAKE(gvt->dev_priv);
402 DEFINE_WAIT_FUNC(wait, woken_wake_function);
406 gvt_dbg_core("workload thread for ring %d started\n", ring_id);
408 while (!kthread_should_stop()) {
409 add_wait_queue(&scheduler->waitq[ring_id], &wait);
411 workload = pick_next_workload(gvt, ring_id);
414 wait_woken(&wait, TASK_INTERRUPTIBLE,
415 MAX_SCHEDULE_TIMEOUT);
416 } while (!kthread_should_stop());
417 remove_wait_queue(&scheduler->waitq[ring_id], &wait);
422 mutex_lock(&scheduler_mutex);
424 gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
425 workload->ring_id, workload,
428 intel_runtime_pm_get(gvt->dev_priv);
430 gvt_dbg_sched("ring id %d will dispatch workload %p\n",
431 workload->ring_id, workload);
434 intel_uncore_forcewake_get(gvt->dev_priv,
437 mutex_lock(&gvt->lock);
438 ret = dispatch_workload(workload);
439 mutex_unlock(&gvt->lock);
442 gvt_err("fail to dispatch workload, skip\n");
446 gvt_dbg_sched("ring id %d wait workload %p\n",
447 workload->ring_id, workload);
449 lret = i915_wait_request(workload->req,
450 0, MAX_SCHEDULE_TIMEOUT);
452 workload->status = lret;
453 gvt_err("fail to wait workload, skip\n");
455 workload->status = 0;
459 gvt_dbg_sched("will complete workload %p\n, status: %d\n",
460 workload, workload->status);
463 i915_gem_request_put(fetch_and_zero(&workload->req));
465 complete_current_workload(gvt, ring_id);
468 intel_uncore_forcewake_put(gvt->dev_priv,
471 intel_runtime_pm_put(gvt->dev_priv);
473 mutex_unlock(&scheduler_mutex);
479 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
481 struct intel_gvt *gvt = vgpu->gvt;
482 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
484 if (atomic_read(&vgpu->running_workload_num)) {
485 gvt_dbg_sched("wait vgpu idle\n");
487 wait_event(scheduler->workload_complete_wq,
488 !atomic_read(&vgpu->running_workload_num));
492 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
494 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
497 gvt_dbg_core("clean workload scheduler\n");
499 for (i = 0; i < I915_NUM_ENGINES; i++) {
500 if (scheduler->thread[i]) {
501 kthread_stop(scheduler->thread[i]);
502 scheduler->thread[i] = NULL;
507 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
509 struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
510 struct workload_thread_param *param = NULL;
514 gvt_dbg_core("init workload scheduler\n");
516 init_waitqueue_head(&scheduler->workload_complete_wq);
518 for (i = 0; i < I915_NUM_ENGINES; i++) {
519 /* check ring mask at init time */
520 if (!HAS_ENGINE(gvt->dev_priv, i))
523 init_waitqueue_head(&scheduler->waitq[i]);
525 param = kzalloc(sizeof(*param), GFP_KERNEL);
534 scheduler->thread[i] = kthread_run(workload_thread, param,
535 "gvt workload %d", i);
536 if (IS_ERR(scheduler->thread[i])) {
537 gvt_err("fail to create workload thread\n");
538 ret = PTR_ERR(scheduler->thread[i]);
544 intel_gvt_clean_workload_scheduler(gvt);
550 void intel_vgpu_clean_gvt_context(struct intel_vgpu *vgpu)
552 struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
554 atomic_notifier_chain_unregister(&vgpu->shadow_ctx->status_notifier,
555 &vgpu->shadow_ctx_notifier_block);
557 mutex_lock(&dev_priv->drm.struct_mutex);
559 /* a little hacky to mark as ctx closed */
560 vgpu->shadow_ctx->closed = true;
561 i915_gem_context_put(vgpu->shadow_ctx);
563 mutex_unlock(&dev_priv->drm.struct_mutex);
566 int intel_vgpu_init_gvt_context(struct intel_vgpu *vgpu)
568 atomic_set(&vgpu->running_workload_num, 0);
570 vgpu->shadow_ctx = i915_gem_context_create_gvt(
571 &vgpu->gvt->dev_priv->drm);
572 if (IS_ERR(vgpu->shadow_ctx))
573 return PTR_ERR(vgpu->shadow_ctx);
575 vgpu->shadow_ctx->engine[RCS].initialised = true;
577 vgpu->shadow_ctx_notifier_block.notifier_call =
578 shadow_context_status_change;
580 atomic_notifier_chain_register(&vgpu->shadow_ctx->status_notifier,
581 &vgpu->shadow_ctx_notifier_block);