Push tcp #2

We have a small leak here, that clp sockets keep referenced after the peer closed. Fix.

The design of Donald::Select makes use of Donald::Callback objects, which are objects conainting a sub and call arguments. However, because of the pattern sub do_something_later() { my ($cb_or_sub,@args)=@_; $store_callback_somehere=new Donald::Callback($cb_or_sub,@args); } do_something_later(\&callback,$arg1,$arg2) the caller doesn't have a reference to the Donald::Callback object, which makes its diffucult to identify it, e.g. to cancel the callback. We want to change the design to accept only references to subs as callbacks. Instead of passing arguments, we exepect the caller to make use of closures to pass data to the callback if needed. sub do_something_later() { my ($cb)=@_; $store_callback_somewhere=$cb; } do_something_later(sub{callback($arg1,$args)}); Instead of changing the API of Donald::Select, we import the code directly into clusterd to make the modifications here.

Use the namespace My:: for packages declared inside the main script.

We imported some lines which do no longer have a function now. Remove them.

Non-functional change to bring source into sync with another working branch.

Use refenerces to subs as callback arguments. If the caller wants to pass additional arguments, he can use closures.

This hash is not used but it prevents these sockets to be able to be closed by running out of scope.

The command "clusterd push file..." has a new syntax with a command verb. (as opposed to "clusterd --push file") The push command is distributed via the area routers to the cluster daemons on all nodes. It is intended that the cluster daemons call back over tcp to the originator to pull the file if needed.

When a daemon receices a push command, it checks whether it already has the offered file or not. If not, it calls back to the daemon where the push originated and pulls the file over tcp.

Implement the server side to pull a file over tcp.

While 5 seconds seems to be enough for normal behaviour, we might need more time if the daemon is slowed down .e.g. by strace.

With 234 hosts pulling a file we failed to service a single on with the old setting listen=1.

Push tcp #2

Push tcp #2

Commits on Dec 7, 2017