Some OS X users report having a hard time finding peers, even in busy swarms

[charles]

Looks like one problem here is in the "fairness" phase of bandwidth.c's bandwidth allocation code, introduced in 1.41, which goes round-robin to all the open sockets and tries to pump them. Sockets are opened in non-blocking mode, so they may not be done connecting to the peers before bandwidth.c pumps them. In Linux, the read/write calls result in an EINPROGRESS errno, which is handled in our code, but OS X returns a different errno which was treated as an error.

One solution is to add a flag to the peer-io object that denotes whether the socket has finished connecting or not, and to set it to true once the libevent polling says that the socket is ready for IO.

[charles]

fixed in trunk in r7747.

backported to 1.5x branch in 7767.