Create copies of NBLs to complete them faster

It turns out that waiting for them to come back in the return function
introduces extremely high latency.

We need to eventually stop doing this and move to a proper ring buffer.

Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>

1 files changed, 3 insertions, 3 deletions

diff --git a/wintun.c b/wintun.c
index 3dadffb..3b878ab 100644
--- a/wintun.c
+++ b/wintun.c
@@ -909,7 +909,6 @@ TunDispatchWrite(_Inout_ TUN_CTX *Ctx, _Inout_ IRP *Irp)
 
     InterlockedAdd64(&Ctx->ActiveNBLCount, NblCount);
     InterlockedExchange(IRP_REFCOUNT(Irp), NblCount);
-    IoMarkIrpPending(Irp);
 
     for (EtherTypeIndex Index = EtherTypeIndexStart; Index < EtherTypeIndexEnd; Index++)
     {
@@ -920,12 +919,13 @@ TunDispatchWrite(_Inout_ TUN_CTX *Ctx, _Inout_ IRP *Irp)
             NblQueue[Index].Head,
             NDIS_DEFAULT_PORT_NUMBER,
             NblQueue[Index].Count,
-            NDIS_RECEIVE_FLAGS_SINGLE_ETHER_TYPE | NDIS_RECEIVE_FLAGS_DISPATCH_LEVEL);
+            NDIS_RECEIVE_FLAGS_SINGLE_ETHER_TYPE | NDIS_RECEIVE_FLAGS_DISPATCH_LEVEL | NDIS_RECEIVE_FLAGS_RESOURCES);
+        TunReturnNetBufferLists((NDIS_HANDLE)Ctx, NblQueue[Index].Head, 0);
     }
 
     ExReleaseSpinLockShared(&Ctx->TransitionLock, Irql);
     TunCompletePause(Ctx, TRUE);
-    return STATUS_PENDING;
+    return STATUS_SUCCESS;
 
 cleanup_nbl_queues:
     for (EtherTypeIndex Index = EtherTypeIndexStart; Index < EtherTypeIndexEnd; Index++)