]> git.karo-electronics.de Git - linux-beck.git/commitdiff
drm/i915/gvt: fix deadlock in workload_thread
authorPei Zhang <pei.zhang@intel.com>
Mon, 14 Nov 2016 10:02:57 +0000 (18:02 +0800)
committerZhenyu Wang <zhenyuw@linux.intel.com>
Mon, 14 Nov 2016 10:02:57 +0000 (18:02 +0800)
It's a classical abba type deadlock when using 2 mutex objects, which
are gvt.lock(a) and drm.struct_mutex(b). Deadlock happens in threads:
1. intel_gvt_create/destroy_vgpu: P(a)->P(b)
2. workload_thread: P(b)->P(a)

Fix solution is align the lock acquire sequence in both threads. This
patch choose to adjust the sequence in workload_thread function.

This fixed lockup symptom for guest-reboot stress test.

v2: adjust sequence in workload_thread based on zhenyu's suggestion.
    adjust sequence in create/destroy_vgpu function.
v3: fix to still require struct_mutex for dispatch_workload()

Signed-off-by: Pei Zhang <pei.zhang@intel.com>
[zhenyuw: fix unused variables warnings.]
Signed-off-by: Zhenyu Wang <zhenyuw@linux.intel.com>
drivers/gpu/drm/i915/gvt/scheduler.c

index 7d87c43661c58627012bbc6d870b70b9a9854678..f898df38dd9a79f8b6dbc208ed803dba76b78635 100644 (file)
@@ -160,8 +160,6 @@ static int shadow_context_status_change(struct notifier_block *nb,
 
 static int dispatch_workload(struct intel_vgpu_workload *workload)
 {
-       struct intel_vgpu *vgpu = workload->vgpu;
-       struct intel_gvt *gvt = vgpu->gvt;
        int ring_id = workload->ring_id;
        struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
        struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
@@ -174,6 +172,8 @@ static int dispatch_workload(struct intel_vgpu_workload *workload)
        shadow_ctx->desc_template = workload->ctx_desc.addressing_mode <<
                                    GEN8_CTX_ADDRESSING_MODE_SHIFT;
 
+       mutex_lock(&dev_priv->drm.struct_mutex);
+
        rq = i915_gem_request_alloc(dev_priv->engine[ring_id], shadow_ctx);
        if (IS_ERR(rq)) {
                gvt_err("fail to allocate gem request\n");
@@ -185,40 +185,35 @@ static int dispatch_workload(struct intel_vgpu_workload *workload)
 
        workload->req = i915_gem_request_get(rq);
 
-       mutex_lock(&gvt->lock);
-
        ret = intel_gvt_scan_and_shadow_workload(workload);
        if (ret)
-               goto err;
+               goto out;
 
        ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
        if (ret)
-               goto err;
+               goto out;
 
        ret = populate_shadow_context(workload);
        if (ret)
-               goto err;
+               goto out;
 
        if (workload->prepare) {
                ret = workload->prepare(workload);
                if (ret)
-                       goto err;
+                       goto out;
        }
 
-       mutex_unlock(&gvt->lock);
-
        gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
                        ring_id, workload->req);
 
-       i915_add_request_no_flush(rq);
+       ret = 0;
        workload->dispatched = true;
-       return 0;
-err:
-       workload->status = ret;
-
-       mutex_unlock(&gvt->lock);
+out:
+       if (ret)
+               workload->status = ret;
 
        i915_add_request_no_flush(rq);
+       mutex_unlock(&dev_priv->drm.struct_mutex);
        return ret;
 }
 
@@ -438,9 +433,9 @@ static int workload_thread(void *priv)
                        intel_uncore_forcewake_get(gvt->dev_priv,
                                        FORCEWAKE_ALL);
 
-               mutex_lock(&gvt->dev_priv->drm.struct_mutex);
+               mutex_lock(&gvt->lock);
                ret = dispatch_workload(workload);
-               mutex_unlock(&gvt->dev_priv->drm.struct_mutex);
+               mutex_unlock(&gvt->lock);
 
                if (ret) {
                        gvt_err("fail to dispatch workload, skip\n");
@@ -463,9 +458,7 @@ complete:
                gvt_dbg_sched("will complete workload %p\n, status: %d\n",
                                workload, workload->status);
 
-               mutex_lock(&gvt->dev_priv->drm.struct_mutex);
                complete_current_workload(gvt, ring_id);
-               mutex_unlock(&gvt->dev_priv->drm.struct_mutex);
 
                i915_gem_request_put(fetch_and_zero(&workload->req));