Suppose a process has an open TCP connection to e.g. a newsserver
somewhere across the 'net. Suppose that the interface through which
this connection is made has its IP address changed. Then the process
can try to send all the data it wants without getting anywhere, but in
linux-2.4.0, at least, no errors from send(2) are returned.
This is probably an annoying FAQ, but shouldn't EPIPE or EBADF be
returned?
FWIW, this happens to me a lot. I keep getting programs stuck (no
errors returned or anything) because my ISP disconnected my after 5
minutes of inactivity or 120 minutes online, and when I reconnected
(pppd demand) I got a different IP address.
For example,
ppp1 Link encap:Point-to-Point Protocol
inet addr:213.104.76.251 P-t-P:10.112.112.113 Mask:255.255.255.255
UP POINTOPOINT RUNNING NOARP MULTICAST MTU:1500 Metric:1
RX packets:3822 errors:3 dropped:0 overruns:0 frame:0
TX packets:4084 errors:0 dropped:65 overruns:0 carrier:0
collisions:0 txqueuelen:3
But this connection is still on an old IP address -
Active Internet connections (w/o servers)
Proto Recv-Q Send-Q Local Address Foreign Address State
tcp 0 72 213.104.68.125:1328 62.253.162.107:119 ESTABLISHED
And here is the strace from the process (telnet) on this connection
when I force it to send some stuff:
send(3, "return an error you cheesy os\r\n", 31, 0) = 31
select(4, [0 3], [], [3], {0, 0}) = 0 (Timeout)
select(4, [0 3], [], [3], NULL
Fooling around in the source, I saw that I most probably want to catch
NETDEV_CHANGEADDR and NETDEV_DOWN as thrown by
linux/net/ipv4/devinet.c and scan the list of active connections
somehow or other and kill (tcp_write_err?) the ones that have the old
local address. Is there any reason not to do this - i.e. should it be
a sysctl?
--
http://www.penguinpowered.com/~vii
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@vger.kernel.org
