netlink refactoring and various fixes (#2112)

* nbd netlink refactoring (#1963)

- fix context leak when nl_socket_modify_cb fails
- use clojure instead of a static method + custom context as a socket callback
- move TNetlinkDevice methods outside of the class definition to reduce nesting
- use nl_send_auto + nl_wait_for_ack instead of nl_send_sync to differentiate between send and recv errors
- remove "Device" suffixes from TNetlinkDevice methods

* nbd netlink refactoring part 2 (#1971)

- rename Timeout -> RequestTimeout
- rename DeadConnectionTimeout -> ConnectionTimeout
- move Send method from TNetlinkMessage to TNetlinkSocket
- clarify comments regarding ioctl device timeout
- get rid of ShouldStop atomic
- close client socket regardless of deleteDevice flag
- replace hard-coded 1 day ioctl device timeout with connection timeout
  which serves the same purpose
- make Start idempotent

* replace ExtactValue with GetValue (#1984)

* do not call Start on already started nbd endpoint (#1985)

cloud/blockstore/config/server.proto

@@ -183,7 +183,7 @@ message TServerConfig
     // NBD request timeout in seconds, only makes sense if using netlink
     optional uint32 NbdRequestTimeout = 112;
 
-    // NBD connection timeout in seconds, only makes sense if using netlink
+    // NBD connection timeout in seconds
     optional uint32 NbdConnectionTimeout = 113;
 
     // Endpoint Proxy unix socket path. Triggers proxy device factory usage

cloud/blockstore/libs/daemon/common/bootstrap.cpp

@@ -552,10 +552,13 @@ void TBootstrapBase::Init()
             EndpointProxyClient);
     }
 
+    // The only case we want kernel to retry requests is when the socket is dead
+    // due to nbd server restart. And since we can't configure ioctl device to
+    // use a new socket, request timeout effectively becomes connection timeout
     if (!nbdDeviceFactory) {
         nbdDeviceFactory = NBD::CreateDeviceFactory(
             Logging,
-            TDuration::Days(1));    // timeout
+            Configs->ServerConfig->GetNbdConnectionTimeout());  // timeout
     }
 
     EndpointManager = CreateEndpointManager(

cloud/blockstore/libs/endpoint_proxy/server/server.cpp

@@ -575,8 +575,6 @@ struct TServer: IEndpointProxyServer
             Scheduler,
             ep.RequestStats,
             volumeStats);
-
-        ep.Client->Start();
         STORAGE_INFO(request.ShortDebugString().Quote()
             << " - Started DurableClient");
 
@@ -635,18 +633,19 @@ struct TServer: IEndpointProxyServer
                 TDuration::Days(1),         // connection timeout
                 true);                      // reconfigure device if exists
         } else {
-            // For netlink devices we have to balance request timeout between
-            // time it takes to fail request for good and resend it if socket
-            // is dead due to proxy restart. We can't configure ioctl device
-            // to use a fresh socket, so there is no point configuring it
+            // The only case we want kernel to retry requests is when the socket
+            // is dead due to nbd server restart. And since we can't configure
+            // ioctl device to use a new socket, request timeout effectively
+            // becomes connection timeout
             ep.NbdDevice = NBD::CreateDevice(
                 Logging,
                 *ep.ListenAddress,
                 request.GetNbdDevice(),
-                TDuration::Days(1));    // request timeout
+                TDuration::Days(1));        // timeout
         }
 
-        auto status = ep.NbdDevice->Start().ExtractValue();
+        auto future = ep.NbdDevice->Start();
+        const auto& status = future.GetValue();
         if (HasError(status)) {
             STORAGE_ERROR(request.ShortDebugString().Quote()
                 << " - Unable to start nbd device: "