drm/i915/gvt: Use vgpu_lock to protect per vgpu access

The patch set splits out 2 small locks from the original big gvt lock:
  - vgpu_lock protects per-vGPU data and logic, especially the vGPU
    trap emulation path.
  - sched_lock protects gvt scheudler structure, context schedule logic
    and vGPU's schedule data.

Use vgpu_lock to replace the gvt big lock. By doing this, the
mmio read/write trap path, vgpu virtual event emulation and other
vgpu related process, would be protected under per vgpu_lock.

v9:
  - Change commit author since the patches are improved a lot compared
    with original version.
    Original author: Pei Zhang <pei.zhang@intel.com>
  - Rebase to latest gvt-staging.
v8:
  - Correct coding and comment style.
  - Rebase to latest gvt-staging.
v7:
  - Remove gtt_lock since already proteced by gvt_lock and vgpu_lock.
  - Fix a typo in intel_gvt_deactivate_vgpu, unlock the wrong lock.
v6:
  - Rebase to latest gvt-staging.
v5:
  - Rebase to latest gvt-staging.
  - intel_vgpu_page_track_handler should use vgpu_lock.
v4:
  - Rebase to latest gvt-staging.
  - Protect vgpu->active access with vgpu_lock.
  - Do not wait gpu idle in vgpu_lock.
v3: update to latest code base
v2: add gvt->lock in function gvt_check_vblank_emulation

Performance comparison on Kabylake platform.
  - Configuration:
    Host: Ubuntu 16.04.
    Guest 1 & 2: Ubuntu 16.04.

glmark2 score comparison:
  - Configuration:
    Host: glxgears.
    Guests: glmark2.
+--------------------------------+-----------------+
| Setup                          | glmark2 score   |
+--------------------------------+-----------------+
| unified lock, iommu=on         | 58~62 (avg. 60) |
+--------------------------------+-----------------+
| unified lock, iommu=igfx_off   | 57~61 (avg. 59) |
+--------------------------------+-----------------+
| per-logic lock, iommu=on       | 60~68 (avg. 64) |
+--------------------------------+-----------------+
| per-logic lock, iommu=igfx_off | 61~67 (avg. 64) |
+--------------------------------+-----------------+

lock_stat comparison:
  - Configuration:
    Stop lock stat immediately after boot up.
    Boot 2 VM Guests.
    Run glmark2 in guests.
    Start perf lock_stat for 20 seconds and stop again.
  - Legend: c - contentions; w - waittime-avg
+------------+-----------------+-----------+---------------+------------+
|            | gvt_lock        |sched_lock | vgpu_lock     | gtt_lock   |
+ lock type; +-----------------+-----------+---------------+------------+
| iommu set  | c     | w       | c  | w    | c    | w      | c   | w    |
+------------+-------+---------+----+------+------+--------+-----+------+
| unified;   | 20697 | 839     |N/A | N/A  | N/A  | N/A    | N/A | N/A  |
| on         |       |         |    |      |      |        |     |      |
+------------+-------+---------+----+------+------+--------+-----+------+
| unified;   | 21838 | 658.15  |N/A | N/A  | N/A  | N/A    | N/A | N/A  |
| igfx_off   |       |         |    |      |      |        |     |      |
+------------+-------+---------+----+------+------+--------+-----+------+
| per-logic; | 1553  | 1599.96 |9458|429.97| 5846 | 274.33 | 0   | 0.00 |
| on         |       |         |    |      |      |        |     |      |
+------------+-------+---------+----+------+------+--------+-----+------+
| per-logic; | 1911  | 1678.32 |8335|445.16| 5451 | 244.80 | 0   | 0.00 |
| igfx_off   |       |         |    |      |      |        |     |      |
+------------+-------+---------+----+------+------+--------+-----+------+

Signed-off-by: Pei Zhang <pei.zhang@intel.com>
Signed-off-by: Colin Xu <colin.xu@intel.com>
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>

1 files changed, 5 insertions, 4 deletions

diff --git a/drivers/gpu/drm/i915/gvt/scheduler.c b/drivers/gpu/drm/i915/gvt/scheduler.c
index c2d183b91500..c0848dcf79c7 100644
--- a/drivers/gpu/drm/i915/gvt/scheduler.c
+++ b/drivers/gpu/drm/i915/gvt/scheduler.c
@@ -680,6 +680,7 @@ static int dispatch_workload(struct intel_vgpu_workload *workload)
 	gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
 		ring_id, workload);
 
+	mutex_lock(&vgpu->vgpu_lock);
 	mutex_lock(&dev_priv->drm.struct_mutex);
 
 	ret = intel_gvt_scan_and_shadow_workload(workload);
@@ -704,6 +705,7 @@ out:
 	}
 
 	mutex_unlock(&dev_priv->drm.struct_mutex);
+	mutex_unlock(&vgpu->vgpu_lock);
 	return ret;
 }
 
@@ -861,14 +863,14 @@ static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
 	int event;
 
 	mutex_lock(&gvt->lock);
+	mutex_lock(&vgpu->vgpu_lock);
 
 	/* For the workload w/ request, needs to wait for the context
 	 * switch to make sure request is completed.
 	 * For the workload w/o request, directly complete the workload.
 	 */
 	if (workload->req) {
-		struct drm_i915_private *dev_priv =
-			workload->vgpu->gvt->dev_priv;
+		struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 		struct intel_engine_cs *engine =
 			dev_priv->engine[workload->ring_id];
 		wait_event(workload->shadow_ctx_status_wq,
@@ -939,6 +941,7 @@ static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
 	if (gvt->scheduler.need_reschedule)
 		intel_gvt_request_service(gvt, INTEL_GVT_REQUEST_EVENT_SCHED);
 
+	mutex_unlock(&vgpu->vgpu_lock);
 	mutex_unlock(&gvt->lock);
 }
 
@@ -991,9 +994,7 @@ static int workload_thread(void *priv)
 			intel_uncore_forcewake_get(gvt->dev_priv,
 					FORCEWAKE_ALL);
 
-		mutex_lock(&gvt->lock);
 		ret = dispatch_workload(workload);
-		mutex_unlock(&gvt->lock);
 
 		if (ret) {
 			vgpu = workload->vgpu;