Bind() methods for TCP sockets

Bind() methods for TCP sockets

BUG=80245
TEST=None

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=87077

Committed: http://src.chromium.org/viewvc/chrome?view=rev&revision=87095

[willchan no longer on Chromium, 2011-05-22 22:42:37]

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:144: return ERR_INVALID_ARGUMENT;
I don't think ERR_INVALID_ARGUMENT is appropriate here. ERR_UNEXPECTED makes
more sense to me.

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:230: if (HANDLE_EINTR(bind(socket_,
addr, addr_len)))
Let's talk it over before you do anything, but I'm not sure how much I like that
the bind() error is only noticed on the Connect(). This is extra problematic if
the bind() fails due to a port conflict and we want to choose a different port.
Getting an ERR_INVALID_ARGUMENT on Connect() is not very clear. What do you
think about adding a new member variable for a created socket that is bound to a
local address?

Also, isn't this broken? CreateSocket() hasn't created the socket yet, right?

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.h (right):

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.h:43: // Binds the socket a local IP
address and port.
to a local

[Sergey Ulanov, 2011-05-23 20:56:03]

Addressed comments. Also fixed error handling for AdoptSocket().

On 2011/05/22 22:42:37, willchan wrote:
>
http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
> File net/socket/tcp_client_socket_libevent.cc (right):
> 
>
http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
> net/socket/tcp_client_socket_libevent.cc:144: return ERR_INVALID_ARGUMENT;
> I don't think ERR_INVALID_ARGUMENT is appropriate here. ERR_UNEXPECTED makes
> more sense to me.
> 
>
http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
> net/socket/tcp_client_socket_libevent.cc:230: if (HANDLE_EINTR(bind(socket_,
> addr, addr_len)))
> Let's talk it over before you do anything, but I'm not sure how much I like
that
> the bind() error is only noticed on the Connect(). This is extra problematic
if
> the bind() fails due to a port conflict and we want to choose a different
port.
> Getting an ERR_INVALID_ARGUMENT on Connect() is not very clear. What do you
> think about adding a new member variable for a created socket that is bound to
a
> local address?
> 
> Also, isn't this broken? CreateSocket() hasn't created the socket yet, right?
> 
>
http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
> File net/socket/tcp_client_socket_libevent.h (right):
> 
>
http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
> net/socket/tcp_client_socket_libevent.h:43: // Binds the socket a local IP
> address and port.
> to a local

[Sergey Ulanov, 2011-05-23 20:57:05]

BTW, do we have unittests for TCPClientSocket? I couldn't find any, otherwise I
would add unittests for Bind().

[Sergey Ulanov, 2011-05-23 20:57:51]

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:144: return ERR_INVALID_ARGUMENT;
On 2011/05/22 22:42:37, willchan wrote:
> I don't think ERR_INVALID_ARGUMENT is appropriate here. ERR_UNEXPECTED makes
> more sense to me.

Done.

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:230: if (HANDLE_EINTR(bind(socket_,
addr, addr_len)))
On 2011/05/22 22:42:37, willchan wrote:
> Let's talk it over before you do anything, but I'm not sure how much I like
that
> the bind() error is only noticed on the Connect(). This is extra problematic
if
> the bind() fails due to a port conflict and we want to choose a different
port.
> Getting an ERR_INVALID_ARGUMENT on Connect() is not very clear. What do you
> think about adding a new member variable for a created socket that is bound to
a
> local address?
Ok, I've added bound_socket_ and that is bound when Bind() is called. We still
need to call bind() in DoConnect() in case there is more than one address in
|addresses_|.

> 
> Also, isn't this broken? CreateSocket() hasn't created the socket yet, right?
Yeah, it was broken in libevent version. Fixed now.

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.h (right):

http://codereview.chromium.org/7004055/diff/3001/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.h:43: // Binds the socket a local IP
address and port.
On 2011/05/22 22:42:37, willchan wrote:
> to a local

Done.

[willchan no longer on Chromium, 2011-05-23 21:03:03]

They got moved to transport_client_socket_unittest.cc I think.

On Mon, May 23, 2011 at 1:57 PM, <[email protected]> wrote:

> BTW, do we have unittests for TCPClientSocket? I couldn't find any,
> otherwise I
> would add unittests for Bind().
>
>
> http://codereview.chromium.org/7004055/
>

[Sergey Ulanov, 2011-05-24 00:28:45]

Ok, I've added unittests in tcp_client_socket_unittest.cc (because Bind() is for
TCP sockets only).

[Sergey Ulanov, 2011-05-24 23:32:54]

ping

[willchan no longer on Chromium, 2011-05-25 14:28:14]

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:148: int
TCPClientSocketLibevent::Bind(const IPEndPoint& address) {
bind() is not allowed to be called for a socket that is already bound. Perhaps
we should return ERR_UNEXPECTED in that case, rather than falling through to
mapping EINVAL?

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:170: return MapConnectError(error);
Is MapConnectError() really what you want here? We should be mapping all the
errors that bind() could return, like EADDRNOTAVAIL.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:269: return MapConnectError(errno);
Again, I'm not sure MapConnectError() is correct here.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:523: close(*socket);
HANDLE_EINTR

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:534: close(socket);
HANDLE_EINTR

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.h (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.h:137: static int CreateSocket(const
struct addrinfo* ai, int* socket);
Now that this isn't a member, perhaps you can take it out of the header file
entirely and put it in an anonymous namespace in the .cc file? Same with
SetupSocket().

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_unittest.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_unittest.cc:54: // external address, verify that
connection fails.
Does this test work? The connect() to 72.14.213.105 could fail for a number of
reasons, not just binding to the wrong interface. Maybe you should lookup the
local ip address for the machine and use that instead?

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_win.h (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_win.h:94: static int SetupSocket(SOCKET socket);
Again, I think you can move these statics to helper functions in the anonymous
namespace in the .cc file.

[Sergey Ulanov, 2011-05-25 20:58:46]

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:148: int
TCPClientSocketLibevent::Bind(const IPEndPoint& address) {
On 2011/05/25 14:28:14, willchan wrote:
> bind() is not allowed to be called for a socket that is already bound. Perhaps
> we should return ERR_UNEXPECTED in that case, rather than falling through to
> mapping EINVAL?

Done.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:170: return MapConnectError(error);
On 2011/05/25 14:28:14, willchan wrote:
> Is MapConnectError() really what you want here? We should be mapping all the
> errors that bind() could return, like EADDRNOTAVAIL.

Done.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:269: return MapConnectError(errno);
On 2011/05/25 14:28:14, willchan wrote:
> Again, I'm not sure MapConnectError() is correct here.
MapConnectError() calls MapSystemError() internally. The only difference in this
particular case is that it would return ERR_NETWORK_ACCESS_DENIED instead of
ERR_ACCESS_DENIED and ERR_CONNECTION_FAILED instead of ERR_FAILED. I don't fully
understand the difference between these error codes, so not sure what's the
right thing to do here. Changed it to MapSystemError().

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:523: close(*socket);
On 2011/05/25 14:28:14, willchan wrote:
> HANDLE_EINTR

Done.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:534: close(socket);
On 2011/05/25 14:28:14, willchan wrote:
> HANDLE_EINTR
Actually we shouldn't close() here. Removed it, as CreateSocket() closes the
socket in case of error.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.h (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.h:137: static int CreateSocket(const
struct addrinfo* ai, int* socket);
On 2011/05/25 14:28:14, willchan wrote:
> Now that this isn't a member, perhaps you can take it out of the header file
> entirely and put it in an anonymous namespace in the .cc file? Same with
> SetupSocket().

Done.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_unittest.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_unittest.cc:54: // external address, verify that
connection fails.
On 2011/05/25 14:28:14, willchan wrote:
> Does this test work? The connect() to 72.14.213.105 could fail for a number of
> reasons, not just binding to the wrong interface. Maybe you should lookup the
> local ip address for the machine and use that instead?
That's exactly what I tried to implement initially, but it doesn't work because,
at least on linux, you can connect from loopback interface to a local interface.
72.14.213.105 is the IP I get for www.google.com. I agree that this solution may
be flaky, but I couldn't think of anything else that would be better, and yet
reliable (i.e. no false negatives). Other suggestions are welcome.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_win.h (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_win.h:94: static int SetupSocket(SOCKET socket);
On 2011/05/25 14:28:14, willchan wrote:
> Again, I think you can move these statics to helper functions in the anonymous
> namespace in the .cc file.

Done.

[Sergey Ulanov, 2011-05-25 21:17:48]

Added one more unittest that connects local interface to loopback interface, and
verifies addresses on both ends.

[Sergey Ulanov, 2011-05-25 23:50:10]

On 2011/05/25 21:17:48, sergeyu wrote:
> Added one more unittest that connects local interface to loopback interface,
and
> verifies addresses on both ends.

Hm, apparently Windows doesn't allow connections from loopback to another local
interface, so I removed BindLoopbackToLocal unittest.

[willchan no longer on Chromium, 2011-05-26 14:25:59]

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_libevent.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_libevent.cc:269: return MapConnectError(errno);
On 2011/05/25 20:58:46, sergeyu wrote:
> On 2011/05/25 14:28:14, willchan wrote:
> > Again, I'm not sure MapConnectError() is correct here.
> MapConnectError() calls MapSystemError() internally. The only difference in
this
> particular case is that it would return ERR_NETWORK_ACCESS_DENIED instead of
> ERR_ACCESS_DENIED and ERR_CONNECTION_FAILED instead of ERR_FAILED. I don't
fully
> understand the difference between these error codes, so not sure what's the
> right thing to do here. Changed it to MapSystemError().

MapSystemError() should be fine.

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
File net/socket/tcp_client_socket_unittest.cc (right):

http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
net/socket/tcp_client_socket_unittest.cc:54: // external address, verify that
connection fails.
On 2011/05/25 20:58:46, sergeyu wrote:
> On 2011/05/25 14:28:14, willchan wrote:
> > Does this test work? The connect() to 72.14.213.105 could fail for a number
of
> > reasons, not just binding to the wrong interface. Maybe you should lookup
the
> > local ip address for the machine and use that instead?
> That's exactly what I tried to implement initially, but it doesn't work
because,
> at least on linux, you can connect from loopback interface to a local
interface.
> 72.14.213.105 is the IP I get for http://www.google.com. I agree that this
solution may
> be flaky, but I couldn't think of anything else that would be better, and yet
> reliable (i.e. no false negatives). Other suggestions are welcome.

How about binding to ipv4 localhost and connecting to ipv6 localhost [::1]?

[Sergey Ulanov, 2011-05-26 19:07:56]

>
http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
> File net/socket/tcp_client_socket_unittest.cc (right):
> 
>
http://codereview.chromium.org/7004055/diff/5004/net/socket/tcp_client_socket...
> net/socket/tcp_client_socket_unittest.cc:54: // external address, verify that
> connection fails.
> On 2011/05/25 20:58:46, sergeyu wrote:
> > On 2011/05/25 14:28:14, willchan wrote:
> > > Does this test work? The connect() to 72.14.213.105 could fail for a
number
> of
> > > reasons, not just binding to the wrong interface. Maybe you should lookup
> the
> > > local ip address for the machine and use that instead?
> > That's exactly what I tried to implement initially, but it doesn't work
> because,
> > at least on linux, you can connect from loopback interface to a local
> interface.
> > 72.14.213.105 is the IP I get for http://www.google.com. I agree that this
> solution may
> > be flaky, but I couldn't think of anything else that would be better, and
yet
> > reliable (i.e. no false negatives). Other suggestions are welcome.
> 
> How about binding to ipv4 localhost and connecting to ipv6 localhost [::1]?

Added this test, also changed the way address family used when |bound_socket_|
is created. The new test isn't perfect ether: it would succeed even if I remove
bind() (because of the different address families). Another problem is that IPv6
is not enabled on all build bots.
Also, the new test revealed a problem with IPv6 in TCPServerSocket - fixed now.

[Sergey Ulanov, 2011-05-27 17:17:02]

ping

[willchan no longer on Chromium, 2011-05-27 17:33:16]

LGTM, sorry for the delays, I have failed you!