later fd_close() call. This also removes fd_map() since fd_open w/FDCLOSE_NONE
now does that.
* If you use fd_socket() or fd_accept(), then no change.
When fd_close() is called we call close() on *NIX and closesocket() on Win.
* If you use fd_fileopen(), then no change.
When fd_close() is called we will call close() on both *NIX and Win.
* If you used fd_open() and then fd_unmap() because you didn't want us
to close the socket, then use fd_open() with FDCLOSE_NONE and
just call fd_close() instead of fd_unmap().
We will not actually close the fd in fd_close() (FDCLOSE_NONE).
* If you called fd_open() with other intentions then either specify a
FDCLOSE_SOCKET / FDCLOSE_FILE as the last argument, or more likely:
don't use fd_open() at all and use fd_socket() or fd_fileopen() instead.
For reasons on this change, see previous patch. This way is more sane and
makes it harder to make mistakes even beyond Windows-specific issues.
This fixes a file descriptor leak in Windows that happened in the
logging code. The most visible effect of this was if you had a
log::maxsize set then on Windows you would see:
"Max file size reached, starting new log file"
Every other line, forever (and not actually starting a new log).
fd_close() previously did not close the file descriptor of a file
on Windows because on Windows it needs to call close() for a file
and closesocket() for a socket, and it always did the latter.
On *NIX it's more easy and you can just always close() any fd.
need this and it slows things down for servers.
For clients it's not much of an issue, since traffic rates are low.
However, for server-to-server links it is an entirely different matter.
It is (only) noticeable if you have lots of traffic, such as when there
is a lot to sync while linking two servers, and especially when the two
servers are geographically further apart.
Tested with 100,000 G-lines on both sides being synced (20MB traffic):
* 20ms RTT (same country/state): speed up of x3
* 200ms RTT (transpacific): speed up of x6
Bram Matthys