drm/i915/gvt: vGPU schedule policy framework
authorZhi Wang <zhi.a.wang@intel.com>
Sun, 1 May 2016 21:09:58 +0000 (17:09 -0400)
committerZhenyu Wang <zhenyuw@linux.intel.com>
Fri, 14 Oct 2016 10:15:02 +0000 (18:15 +0800)
This patch introduces a vGPU schedule policy framework, with a timer based
schedule policy module for now

Signed-off-by: Zhi Wang <zhi.a.wang@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
drivers/gpu/drm/i915/gvt/Makefile
drivers/gpu/drm/i915/gvt/gvt.c
drivers/gpu/drm/i915/gvt/gvt.h
drivers/gpu/drm/i915/gvt/handlers.c
drivers/gpu/drm/i915/gvt/sched_policy.c [new file with mode: 0644]
drivers/gpu/drm/i915/gvt/sched_policy.h [new file with mode: 0644]
drivers/gpu/drm/i915/gvt/scheduler.h
drivers/gpu/drm/i915/gvt/vgpu.c

index 2d1a5c398a185c243fdfa05270225dbf59b063aa..21000b183b951d3804f80e464fa2af6a1442f97d 100644 (file)
@@ -1,7 +1,7 @@
 GVT_DIR := gvt
 GVT_SOURCE := gvt.o aperture_gm.o handlers.o vgpu.o trace_points.o firmware.o \
        interrupt.o gtt.o cfg_space.o opregion.o mmio.o display.o edid.o \
-       execlist.o scheduler.o
+       execlist.o scheduler.o sched_policy.o
 
 ccflags-y                      += -I$(src) -I$(src)/$(GVT_DIR) -Wall
 i915-y                        += $(addprefix $(GVT_DIR)/, $(GVT_SOURCE))
index 5bf8488beaefedb3a898017e88f8902fc6a4bc19..7f13efbbd93a0bb7c9ed06aa401a9a9ea76c4edd 100644 (file)
@@ -177,6 +177,7 @@ void intel_gvt_clean_device(struct drm_i915_private *dev_priv)
                return;
 
        clean_service_thread(gvt);
+       intel_gvt_clean_sched_policy(gvt);
        intel_gvt_clean_workload_scheduler(gvt);
        intel_gvt_clean_opregion(gvt);
        intel_gvt_clean_gtt(gvt);
@@ -244,14 +245,20 @@ int intel_gvt_init_device(struct drm_i915_private *dev_priv)
        if (ret)
                goto out_clean_opregion;
 
-       ret = init_service_thread(gvt);
+       ret = intel_gvt_init_sched_policy(gvt);
        if (ret)
                goto out_clean_workload_scheduler;
 
+       ret = init_service_thread(gvt);
+       if (ret)
+               goto out_clean_sched_policy;
+
        gvt_dbg_core("gvt device creation is done\n");
        gvt->initialized = true;
        return 0;
 
+out_clean_sched_policy:
+       intel_gvt_clean_sched_policy(gvt);
 out_clean_workload_scheduler:
        intel_gvt_clean_workload_scheduler(gvt);
 out_clean_opregion:
index ae716cf22f693d1c503c6519fb5a983dc3537466..6d37468631b7e973662f6c1a5542a9bf900ac45d 100644 (file)
@@ -43,6 +43,7 @@
 #include "edid.h"
 #include "execlist.h"
 #include "scheduler.h"
+#include "sched_policy.h"
 
 #define GVT_MAX_VGPU 8
 
@@ -139,6 +140,7 @@ struct intel_vgpu {
        unsigned long handle; /* vGPU handle used by hypervisor MPT modules */
        bool active;
        bool resetting;
+       void *sched_data;
 
        struct intel_vgpu_fence fence;
        struct intel_vgpu_gm gm;
index 04da35c2600a8fe27c399819e1622352f545c214..8e8bda584c941a1678f83791b23b22ad8a9d3473 100644 (file)
@@ -235,6 +235,7 @@ static int handle_device_reset(struct intel_vgpu *vgpu, unsigned int offset,
 
        vgpu->resetting = true;
 
+       intel_vgpu_stop_schedule(vgpu);
        if (scheduler->current_vgpu == vgpu) {
                mutex_unlock(&vgpu->gvt->lock);
                intel_gvt_wait_vgpu_idle(vgpu);
@@ -1317,6 +1318,28 @@ static int elsp_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,
        return 0;
 }
 
+static int ring_mode_mmio_write(struct intel_vgpu *vgpu, unsigned int offset,
+               void *p_data, unsigned int bytes)
+{
+       u32 data = *(u32 *)p_data;
+       int ring_id = render_mmio_to_ring_id(vgpu->gvt, offset);
+       bool enable_execlist;
+
+       write_vreg(vgpu, offset, p_data, bytes);
+       if ((data & _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE))
+                       || (data & _MASKED_BIT_DISABLE(GFX_RUN_LIST_ENABLE))) {
+               enable_execlist = !!(data & GFX_RUN_LIST_ENABLE);
+
+               gvt_dbg_core("EXECLIST %s on ring %d\n",
+                               (enable_execlist ? "enabling" : "disabling"),
+                               ring_id);
+
+               if (enable_execlist)
+                       intel_vgpu_start_schedule(vgpu);
+       }
+       return 0;
+}
+
 #define MMIO_F(reg, s, f, am, rm, d, r, w) do { \
        ret = new_mmio_info(gvt, INTEL_GVT_MMIO_OFFSET(reg), \
                f, s, am, rm, d, r, w); \
@@ -1398,7 +1421,7 @@ static int init_generic_mmio_info(struct intel_gvt *gvt)
 
        /* RING MODE */
 #define RING_REG(base) (base + 0x29c)
-       MMIO_RING_DFH(RING_REG, D_ALL, F_MODE_MASK, NULL, NULL);
+       MMIO_RING_DFH(RING_REG, D_ALL, F_MODE_MASK, NULL, ring_mode_mmio_write);
 #undef RING_REG
 
        MMIO_RING_DFH(RING_MI_MODE, D_ALL, F_MODE_MASK, NULL, NULL);
@@ -2213,7 +2236,7 @@ static int init_broadwell_mmio_info(struct intel_gvt *gvt)
        MMIO_D(RING_CTL(GEN8_BSD2_RING_BASE), D_BDW_PLUS);
        MMIO_D(RING_ACTHD(GEN8_BSD2_RING_BASE), D_BDW_PLUS);
        MMIO_D(RING_ACTHD_UDW(GEN8_BSD2_RING_BASE), D_BDW_PLUS);
-       MMIO_DFH(0x1c29c, D_BDW_PLUS, F_MODE_MASK, NULL, NULL);
+       MMIO_DFH(0x1c29c, D_BDW_PLUS, F_MODE_MASK, NULL, ring_mode_mmio_write);
        MMIO_DFH(RING_MI_MODE(GEN8_BSD2_RING_BASE), D_BDW_PLUS, F_MODE_MASK,
                        NULL, NULL);
        MMIO_DFH(RING_INSTPM(GEN8_BSD2_RING_BASE), D_BDW_PLUS, F_MODE_MASK,
diff --git a/drivers/gpu/drm/i915/gvt/sched_policy.c b/drivers/gpu/drm/i915/gvt/sched_policy.c
new file mode 100644 (file)
index 0000000..c607354
--- /dev/null
@@ -0,0 +1,291 @@
+/*
+ * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Anhua Xu
+ *    Kevin Tian <kevin.tian@intel.com>
+ *
+ * Contributors:
+ *    Min He <min.he@intel.com>
+ *    Bing Niu <bing.niu@intel.com>
+ *    Zhi Wang <zhi.a.wang@intel.com>
+ *
+ */
+
+#include "i915_drv.h"
+
+static bool vgpu_has_pending_workload(struct intel_vgpu *vgpu)
+{
+       struct intel_vgpu_execlist *execlist;
+       int i;
+
+       for (i = 0; i < I915_NUM_ENGINES; i++) {
+               execlist = &vgpu->execlist[i];
+               if (!list_empty(workload_q_head(vgpu, i)))
+                       return true;
+       }
+
+       return false;
+}
+
+static void try_to_schedule_next_vgpu(struct intel_gvt *gvt)
+{
+       struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
+       int i;
+
+       /* no target to schedule */
+       if (!scheduler->next_vgpu)
+               return;
+
+       gvt_dbg_sched("try to schedule next vgpu %d\n",
+                       scheduler->next_vgpu->id);
+
+       /*
+        * after the flag is set, workload dispatch thread will
+        * stop dispatching workload for current vgpu
+        */
+       scheduler->need_reschedule = true;
+
+       /* still have uncompleted workload? */
+       for (i = 0; i < I915_NUM_ENGINES; i++) {
+               if (scheduler->current_workload[i]) {
+                       gvt_dbg_sched("still have running workload\n");
+                       return;
+               }
+       }
+
+       gvt_dbg_sched("switch to next vgpu %d\n",
+                       scheduler->next_vgpu->id);
+
+       /* switch current vgpu */
+       scheduler->current_vgpu = scheduler->next_vgpu;
+       scheduler->next_vgpu = NULL;
+
+       scheduler->need_reschedule = false;
+
+       /* wake up workload dispatch thread */
+       for (i = 0; i < I915_NUM_ENGINES; i++)
+               wake_up(&scheduler->waitq[i]);
+}
+
+struct tbs_vgpu_data {
+       struct list_head list;
+       struct intel_vgpu *vgpu;
+       /* put some per-vgpu sched stats here */
+};
+
+struct tbs_sched_data {
+       struct intel_gvt *gvt;
+       struct delayed_work work;
+       unsigned long period;
+       struct list_head runq_head;
+};
+
+#define GVT_DEFAULT_TIME_SLICE (1 * HZ / 1000)
+
+static void tbs_sched_func(struct work_struct *work)
+{
+       struct tbs_sched_data *sched_data = container_of(work,
+                       struct tbs_sched_data, work.work);
+       struct tbs_vgpu_data *vgpu_data;
+
+       struct intel_gvt *gvt = sched_data->gvt;
+       struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
+
+       struct intel_vgpu *vgpu = NULL;
+       struct list_head *pos, *head;
+
+       mutex_lock(&gvt->lock);
+
+       /* no vgpu or has already had a target */
+       if (list_empty(&sched_data->runq_head) || scheduler->next_vgpu)
+               goto out;
+
+       if (scheduler->current_vgpu) {
+               vgpu_data = scheduler->current_vgpu->sched_data;
+               head = &vgpu_data->list;
+       } else {
+               gvt_dbg_sched("no current vgpu search from q head\n");
+               head = &sched_data->runq_head;
+       }
+
+       /* search a vgpu with pending workload */
+       list_for_each(pos, head) {
+               if (pos == &sched_data->runq_head)
+                       continue;
+
+               vgpu_data = container_of(pos, struct tbs_vgpu_data, list);
+               if (!vgpu_has_pending_workload(vgpu_data->vgpu))
+                       continue;
+
+               vgpu = vgpu_data->vgpu;
+               break;
+       }
+
+       if (vgpu) {
+               scheduler->next_vgpu = vgpu;
+               gvt_dbg_sched("pick next vgpu %d\n", vgpu->id);
+       }
+out:
+       if (scheduler->next_vgpu) {
+               gvt_dbg_sched("try to schedule next vgpu %d\n",
+                               scheduler->next_vgpu->id);
+               try_to_schedule_next_vgpu(gvt);
+       }
+
+       /*
+        * still have vgpu on runq
+        * or last schedule haven't finished due to running workload
+        */
+       if (!list_empty(&sched_data->runq_head) || scheduler->next_vgpu)
+               schedule_delayed_work(&sched_data->work, sched_data->period);
+
+       mutex_unlock(&gvt->lock);
+}
+
+static int tbs_sched_init(struct intel_gvt *gvt)
+{
+       struct intel_gvt_workload_scheduler *scheduler =
+               &gvt->scheduler;
+
+       struct tbs_sched_data *data;
+
+       data = kzalloc(sizeof(*data), GFP_KERNEL);
+       if (!data)
+               return -ENOMEM;
+
+       INIT_LIST_HEAD(&data->runq_head);
+       INIT_DELAYED_WORK(&data->work, tbs_sched_func);
+       data->period = GVT_DEFAULT_TIME_SLICE;
+       data->gvt = gvt;
+
+       scheduler->sched_data = data;
+       return 0;
+}
+
+static void tbs_sched_clean(struct intel_gvt *gvt)
+{
+       struct intel_gvt_workload_scheduler *scheduler =
+               &gvt->scheduler;
+       struct tbs_sched_data *data = scheduler->sched_data;
+
+       cancel_delayed_work(&data->work);
+       kfree(data);
+       scheduler->sched_data = NULL;
+}
+
+static int tbs_sched_init_vgpu(struct intel_vgpu *vgpu)
+{
+       struct tbs_vgpu_data *data;
+
+       data = kzalloc(sizeof(*data), GFP_KERNEL);
+       if (!data)
+               return -ENOMEM;
+
+       data->vgpu = vgpu;
+       INIT_LIST_HEAD(&data->list);
+
+       vgpu->sched_data = data;
+       return 0;
+}
+
+static void tbs_sched_clean_vgpu(struct intel_vgpu *vgpu)
+{
+       kfree(vgpu->sched_data);
+       vgpu->sched_data = NULL;
+}
+
+static void tbs_sched_start_schedule(struct intel_vgpu *vgpu)
+{
+       struct tbs_sched_data *sched_data = vgpu->gvt->scheduler.sched_data;
+       struct tbs_vgpu_data *vgpu_data = vgpu->sched_data;
+
+       if (!list_empty(&vgpu_data->list))
+               return;
+
+       list_add_tail(&vgpu_data->list, &sched_data->runq_head);
+       schedule_delayed_work(&sched_data->work, sched_data->period);
+}
+
+static void tbs_sched_stop_schedule(struct intel_vgpu *vgpu)
+{
+       struct tbs_vgpu_data *vgpu_data = vgpu->sched_data;
+
+       list_del_init(&vgpu_data->list);
+}
+
+struct intel_gvt_sched_policy_ops tbs_schedule_ops = {
+       .init = tbs_sched_init,
+       .clean = tbs_sched_clean,
+       .init_vgpu = tbs_sched_init_vgpu,
+       .clean_vgpu = tbs_sched_clean_vgpu,
+       .start_schedule = tbs_sched_start_schedule,
+       .stop_schedule = tbs_sched_stop_schedule,
+};
+
+int intel_gvt_init_sched_policy(struct intel_gvt *gvt)
+{
+       gvt->scheduler.sched_ops = &tbs_schedule_ops;
+
+       return gvt->scheduler.sched_ops->init(gvt);
+}
+
+void intel_gvt_clean_sched_policy(struct intel_gvt *gvt)
+{
+       gvt->scheduler.sched_ops->clean(gvt);
+}
+
+int intel_vgpu_init_sched_policy(struct intel_vgpu *vgpu)
+{
+       return vgpu->gvt->scheduler.sched_ops->init_vgpu(vgpu);
+}
+
+void intel_vgpu_clean_sched_policy(struct intel_vgpu *vgpu)
+{
+       vgpu->gvt->scheduler.sched_ops->clean_vgpu(vgpu);
+}
+
+void intel_vgpu_start_schedule(struct intel_vgpu *vgpu)
+{
+       gvt_dbg_core("vgpu%d: start schedule\n", vgpu->id);
+
+       vgpu->gvt->scheduler.sched_ops->start_schedule(vgpu);
+}
+
+void intel_vgpu_stop_schedule(struct intel_vgpu *vgpu)
+{
+       struct intel_gvt_workload_scheduler *scheduler =
+               &vgpu->gvt->scheduler;
+
+       gvt_dbg_core("vgpu%d: stop schedule\n", vgpu->id);
+
+       scheduler->sched_ops->stop_schedule(vgpu);
+
+       if (scheduler->next_vgpu == vgpu)
+               scheduler->next_vgpu = NULL;
+
+       if (scheduler->current_vgpu == vgpu) {
+               /* stop workload dispatching */
+               scheduler->need_reschedule = true;
+               scheduler->current_vgpu = NULL;
+       }
+}
diff --git a/drivers/gpu/drm/i915/gvt/sched_policy.h b/drivers/gpu/drm/i915/gvt/sched_policy.h
new file mode 100644 (file)
index 0000000..bb8b909
--- /dev/null
@@ -0,0 +1,58 @@
+/*
+ * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ * Authors:
+ *    Anhua Xu
+ *    Kevin Tian <kevin.tian@intel.com>
+ *
+ * Contributors:
+ *    Min He <min.he@intel.com>
+ *    Bing Niu <bing.niu@intel.com>
+ *    Zhi Wang <zhi.a.wang@intel.com>
+ *
+ */
+
+#ifndef __GVT_SCHED_POLICY__
+#define __GVT_SCHED_POLICY__
+
+struct intel_gvt_sched_policy_ops {
+       int (*init)(struct intel_gvt *gvt);
+       void (*clean)(struct intel_gvt *gvt);
+       int (*init_vgpu)(struct intel_vgpu *vgpu);
+       void (*clean_vgpu)(struct intel_vgpu *vgpu);
+       void (*start_schedule)(struct intel_vgpu *vgpu);
+       void (*stop_schedule)(struct intel_vgpu *vgpu);
+};
+
+int intel_gvt_init_sched_policy(struct intel_gvt *gvt);
+
+void intel_gvt_clean_sched_policy(struct intel_gvt *gvt);
+
+int intel_vgpu_init_sched_policy(struct intel_vgpu *vgpu);
+
+void intel_vgpu_clean_sched_policy(struct intel_vgpu *vgpu);
+
+void intel_vgpu_start_schedule(struct intel_vgpu *vgpu);
+
+void intel_vgpu_stop_schedule(struct intel_vgpu *vgpu);
+
+#endif
index 9206cc02c8a883df978199c167a11292ac121293..2500438d7aa7eaa3d0bca57e45405176a10849a1 100644 (file)
@@ -45,6 +45,9 @@ struct intel_gvt_workload_scheduler {
        wait_queue_head_t workload_complete_wq;
        struct task_struct *thread[I915_NUM_ENGINES];
        wait_queue_head_t waitq[I915_NUM_ENGINES];
+
+       void *sched_data;
+       struct intel_gvt_sched_policy_ops *sched_ops;
 };
 
 struct intel_vgpu_workload {
index 609ec55bbf5fb36157e598346448e1116a2d1e8a..c9b8e184f5cbd4a8b310ce3a7847c3ce78926ab4 100644 (file)
@@ -146,6 +146,14 @@ void intel_gvt_destroy_vgpu(struct intel_vgpu *vgpu)
        vgpu->active = false;
        idr_remove(&gvt->vgpu_idr, vgpu->id);
 
+       if (atomic_read(&vgpu->running_workload_num)) {
+               mutex_unlock(&gvt->lock);
+               intel_gvt_wait_vgpu_idle(vgpu);
+               mutex_lock(&gvt->lock);
+       }
+
+       intel_vgpu_stop_schedule(vgpu);
+       intel_vgpu_clean_sched_policy(vgpu);
        intel_vgpu_clean_gvt_context(vgpu);
        intel_vgpu_clean_execlist(vgpu);
        intel_vgpu_clean_display(vgpu);
@@ -231,11 +239,17 @@ struct intel_vgpu *intel_gvt_create_vgpu(struct intel_gvt *gvt,
        if (ret)
                goto out_clean_execlist;
 
+       ret = intel_vgpu_init_sched_policy(vgpu);
+       if (ret)
+               goto out_clean_shadow_ctx;
+
        vgpu->active = true;
        mutex_unlock(&gvt->lock);
 
        return vgpu;
 
+out_clean_shadow_ctx:
+       intel_vgpu_clean_gvt_context(vgpu);
 out_clean_execlist:
        intel_vgpu_clean_execlist(vgpu);
 out_clean_display: