Set TCP keep alive on Linux and Mac.

Set TCP keep alive on Linux and Mac.

On Linux, also set the timeouts.

BUG=27400
TEST=Start wireshark and capture packets.  Start chrome.  Open www.facebook.com.  Let it sit idle for awhile.  Wait for TCP keep alive packets to be sent out after 45 seconds on Linux.

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=71482

[willchan no longer on Chromium, 2011-01-14 16:33:35]

eroman: General review
agl: Please review my Linux specific (non-portable) changes.

[agl, 2011-01-14 16:37:48]

LGTM

http://codereview.chromium.org/6162005/diff/1/net/socket/tcp_client_socket_li...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/6162005/diff/1/net/socket/tcp_client_socket_li...
net/socket/tcp_client_socket_libevent.cc:64: optval = 45;
seems redundant, although you could argue that it's easier to read. Your call.

[willchan no longer on Chromium, 2011-01-14 16:40:55]

http://codereview.chromium.org/6162005/diff/1/net/socket/tcp_client_socket_li...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/6162005/diff/1/net/socket/tcp_client_socket_li...
net/socket/tcp_client_socket_libevent.cc:64: optval = 45;
oops, unintentional copy/paste error.

On 2011/01/14 16:37:48, agl wrote:
> seems redundant, although you could argue that it's easier to read. Your call.

[cbentzel, 2011-01-14 16:42:41]

driveby question [the code looks fine]

Is this mainly being used to early-detect dead peers? Are you concerned about it
keeping stale connections alive longer than it should?

[willchan no longer on Chromium, 2011-01-14 16:52:44]

Actually, it's because NAT routers (or whatever other intermediaries there
are) are whacking our idle TCP connections.  What's worse, when they do
this, they simply blackhole the packets.  So, we don't get any TCP RSTs and
enter congestion control behavior, doing exponential backoff on retransmits.
 This results in user page hangs without getting any timeouts.

Setting TCP keepalives is an attempt to prevent the NAT router from dropping
the TCP connection from its tables.

On Fri, Jan 14, 2011 at 8:42 AM, <cbentzel@chromium.org> wrote:

> driveby question [the code looks fine]
>
> Is this mainly being used to early-detect dead peers? Are you concerned
> about it
> keeping stale connections alive longer than it should?
>
>
> http://codereview.chromium.org/6162005/
>

[wtc, 2011-01-18 20:22:48]

http://codereview.chromium.org/6162005/diff/5001/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/6162005/diff/5001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:519: SetTCPKeepAlive(socket_);
Please add a comment to explain why we need to set the
SO_KEEPALIVE socket option.

[jeremya, 2012-04-30 05:27:32]

This is a rather belated message, but I'm doing some unrelated investigations
into keepalives on various platforms, and it seems that setting SO_KEEPALIVE to
1 on mac (10.6) doesn't cause keepalive packets to be sent with any regularity.
It looks to me from
http://developer.apple.com/library/mac/#DOCUMENTATION/Darwin/Reference/ManPag...
that on OS X 10.6 and greater we should be able to set TCP_KEEPALIVE to the same
thing we set TCP_KEEPIDLE to on linux.

[wtc, 2012-04-30 19:45:33]

jeremya:  Thank you for your comment.  Could you give it a try and
see if it works?

I have two related questions:

1. Do you know what is the default value of TCP_KEEPALIVE on Mac OS X?

2. In the system header files <netinet/tcp_var.h>, I also found
TCPCTL_KEEPIDLE and TCPCTL_KEEPINTVL, which look like the Linux
TCP_KEEPIDLE and TCP_KEEPINTVL options.  Should we use those
TCPCTL_KEEPxxx options instead?